NEON implementations of QS8 2x8c16 GEMM and IGEMM assembly microkernels
PiperOrigin-RevId: 382565189
diff --git a/BUILD.bazel b/BUILD.bazel
index e3ce51a..86b3df3 100644
--- a/BUILD.bazel
+++ b/BUILD.bazel
@@ -1489,69 +1489,69 @@
"src/math/roundz-wasmsimd-cvt.c",
"src/math/sigmoid-wasmsimd-rr2-lut64-p2-div.c",
"src/math/sigmoid-wasmsimd-rr2-p5-div.c",
- "src/qc8-dwconv/gen/up16x25-minmax-fp32-wasmsimd-mul16.c",
- "src/qc8-dwconv/gen/up16x9-minmax-fp32-wasmsimd-mul16.c",
- "src/qc8-dwconv/gen/up24x25-minmax-fp32-wasmsimd-mul16.c",
- "src/qc8-dwconv/gen/up24x9-minmax-fp32-wasmsimd-mul16.c",
- "src/qc8-dwconv/gen/up8x25-minmax-fp32-wasmsimd-mul16.c",
"src/qc8-dwconv/gen/up8x9-minmax-fp32-wasmsimd-mul16.c",
- "src/qc8-gemm/gen/1x4c8-minmax-fp32-wasmsimd-ld128.c",
+ "src/qc8-dwconv/gen/up8x25-minmax-fp32-wasmsimd-mul16.c",
+ "src/qc8-dwconv/gen/up16x9-minmax-fp32-wasmsimd-mul16.c",
+ "src/qc8-dwconv/gen/up16x25-minmax-fp32-wasmsimd-mul16.c",
+ "src/qc8-dwconv/gen/up24x9-minmax-fp32-wasmsimd-mul16.c",
+ "src/qc8-dwconv/gen/up24x25-minmax-fp32-wasmsimd-mul16.c",
"src/qc8-gemm/gen/1x4c8-minmax-fp32-wasmsimd-ld64.c",
- "src/qc8-gemm/gen/2x4c8-minmax-fp32-wasmsimd-ld128.c",
+ "src/qc8-gemm/gen/1x4c8-minmax-fp32-wasmsimd-ld128.c",
"src/qc8-gemm/gen/2x4c8-minmax-fp32-wasmsimd-ld64.c",
- "src/qc8-gemm/gen/3x4c8-minmax-fp32-wasmsimd-ld128.c",
+ "src/qc8-gemm/gen/2x4c8-minmax-fp32-wasmsimd-ld128.c",
"src/qc8-gemm/gen/3x4c8-minmax-fp32-wasmsimd-ld64.c",
- "src/qc8-igemm/gen/1x4c8-minmax-fp32-wasmsimd-ld128.c",
+ "src/qc8-gemm/gen/3x4c8-minmax-fp32-wasmsimd-ld128.c",
"src/qc8-igemm/gen/1x4c8-minmax-fp32-wasmsimd-ld64.c",
- "src/qc8-igemm/gen/2x4c8-minmax-fp32-wasmsimd-ld128.c",
+ "src/qc8-igemm/gen/1x4c8-minmax-fp32-wasmsimd-ld128.c",
"src/qc8-igemm/gen/2x4c8-minmax-fp32-wasmsimd-ld64.c",
- "src/qc8-igemm/gen/3x4c8-minmax-fp32-wasmsimd-ld128.c",
+ "src/qc8-igemm/gen/2x4c8-minmax-fp32-wasmsimd-ld128.c",
"src/qc8-igemm/gen/3x4c8-minmax-fp32-wasmsimd-ld64.c",
- "src/qs8-dwconv/gen/up16x25-minmax-fp32-wasmsimd-mul16.c",
- "src/qs8-dwconv/gen/up16x25-minmax-gemmlowp-wasmsimd-mul16.c",
- "src/qs8-dwconv/gen/up16x9-minmax-fp32-wasmsimd-mul16.c",
- "src/qs8-dwconv/gen/up16x9-minmax-gemmlowp-wasmsimd-mul16.c",
- "src/qs8-dwconv/gen/up24x25-minmax-fp32-wasmsimd-mul16.c",
- "src/qs8-dwconv/gen/up24x25-minmax-gemmlowp-wasmsimd-mul16.c",
- "src/qs8-dwconv/gen/up24x9-minmax-fp32-wasmsimd-mul16.c",
- "src/qs8-dwconv/gen/up24x9-minmax-gemmlowp-wasmsimd-mul16.c",
- "src/qs8-dwconv/gen/up8x25-minmax-fp32-wasmsimd-mul16.c",
- "src/qs8-dwconv/gen/up8x25-minmax-gemmlowp-wasmsimd-mul16.c",
+ "src/qc8-igemm/gen/3x4c8-minmax-fp32-wasmsimd-ld128.c",
"src/qs8-dwconv/gen/up8x9-minmax-fp32-wasmsimd-mul16.c",
"src/qs8-dwconv/gen/up8x9-minmax-gemmlowp-wasmsimd-mul16.c",
+ "src/qs8-dwconv/gen/up8x25-minmax-fp32-wasmsimd-mul16.c",
+ "src/qs8-dwconv/gen/up8x25-minmax-gemmlowp-wasmsimd-mul16.c",
+ "src/qs8-dwconv/gen/up16x9-minmax-fp32-wasmsimd-mul16.c",
+ "src/qs8-dwconv/gen/up16x9-minmax-gemmlowp-wasmsimd-mul16.c",
+ "src/qs8-dwconv/gen/up16x25-minmax-fp32-wasmsimd-mul16.c",
+ "src/qs8-dwconv/gen/up16x25-minmax-gemmlowp-wasmsimd-mul16.c",
+ "src/qs8-dwconv/gen/up24x9-minmax-fp32-wasmsimd-mul16.c",
+ "src/qs8-dwconv/gen/up24x9-minmax-gemmlowp-wasmsimd-mul16.c",
+ "src/qs8-dwconv/gen/up24x25-minmax-fp32-wasmsimd-mul16.c",
+ "src/qs8-dwconv/gen/up24x25-minmax-gemmlowp-wasmsimd-mul16.c",
"src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c8-acc2.c",
"src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c16-acc2.c",
"src/qs8-gavgpool/gen/7p7x-minmax-wasmsimd-c24-acc2.c",
"src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c8-acc2.c",
"src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c16-acc2.c",
"src/qs8-gavgpool/gen/7x-minmax-wasmsimd-c24-acc2.c",
- "src/qs8-gemm/gen/1x4c8-minmax-fp32-wasmsimd-ld128.c",
"src/qs8-gemm/gen/1x4c8-minmax-fp32-wasmsimd-ld64.c",
- "src/qs8-gemm/gen/1x4c8-minmax-gemmlowp-wasmsimd-ld128.c",
+ "src/qs8-gemm/gen/1x4c8-minmax-fp32-wasmsimd-ld128.c",
"src/qs8-gemm/gen/1x4c8-minmax-gemmlowp-wasmsimd-ld64.c",
+ "src/qs8-gemm/gen/1x4c8-minmax-gemmlowp-wasmsimd-ld128.c",
"src/qs8-gemm/gen/1x4c8-xw-minmax-gemmlowp-wasmsimd.c",
- "src/qs8-gemm/gen/2x4c8-minmax-fp32-wasmsimd-ld128.c",
"src/qs8-gemm/gen/2x4c8-minmax-fp32-wasmsimd-ld64.c",
- "src/qs8-gemm/gen/2x4c8-minmax-gemmlowp-wasmsimd-ld128.c",
+ "src/qs8-gemm/gen/2x4c8-minmax-fp32-wasmsimd-ld128.c",
"src/qs8-gemm/gen/2x4c8-minmax-gemmlowp-wasmsimd-ld64.c",
+ "src/qs8-gemm/gen/2x4c8-minmax-gemmlowp-wasmsimd-ld128.c",
"src/qs8-gemm/gen/2x4c8-xw-minmax-gemmlowp-wasmsimd.c",
- "src/qs8-gemm/gen/3x4c8-minmax-fp32-wasmsimd-ld128.c",
"src/qs8-gemm/gen/3x4c8-minmax-fp32-wasmsimd-ld64.c",
- "src/qs8-gemm/gen/3x4c8-minmax-gemmlowp-wasmsimd-ld128.c",
+ "src/qs8-gemm/gen/3x4c8-minmax-fp32-wasmsimd-ld128.c",
"src/qs8-gemm/gen/3x4c8-minmax-gemmlowp-wasmsimd-ld64.c",
+ "src/qs8-gemm/gen/3x4c8-minmax-gemmlowp-wasmsimd-ld128.c",
"src/qs8-gemm/gen/3x4c8-xw-minmax-gemmlowp-wasmsimd.c",
- "src/qs8-igemm/gen/1x4c8-minmax-fp32-wasmsimd-ld128.c",
"src/qs8-igemm/gen/1x4c8-minmax-fp32-wasmsimd-ld64.c",
- "src/qs8-igemm/gen/1x4c8-minmax-gemmlowp-wasmsimd-ld128.c",
+ "src/qs8-igemm/gen/1x4c8-minmax-fp32-wasmsimd-ld128.c",
"src/qs8-igemm/gen/1x4c8-minmax-gemmlowp-wasmsimd-ld64.c",
- "src/qs8-igemm/gen/2x4c8-minmax-fp32-wasmsimd-ld128.c",
+ "src/qs8-igemm/gen/1x4c8-minmax-gemmlowp-wasmsimd-ld128.c",
"src/qs8-igemm/gen/2x4c8-minmax-fp32-wasmsimd-ld64.c",
- "src/qs8-igemm/gen/2x4c8-minmax-gemmlowp-wasmsimd-ld128.c",
+ "src/qs8-igemm/gen/2x4c8-minmax-fp32-wasmsimd-ld128.c",
"src/qs8-igemm/gen/2x4c8-minmax-gemmlowp-wasmsimd-ld64.c",
- "src/qs8-igemm/gen/3x4c8-minmax-fp32-wasmsimd-ld128.c",
+ "src/qs8-igemm/gen/2x4c8-minmax-gemmlowp-wasmsimd-ld128.c",
"src/qs8-igemm/gen/3x4c8-minmax-fp32-wasmsimd-ld64.c",
- "src/qs8-igemm/gen/3x4c8-minmax-gemmlowp-wasmsimd-ld128.c",
+ "src/qs8-igemm/gen/3x4c8-minmax-fp32-wasmsimd-ld128.c",
"src/qs8-igemm/gen/3x4c8-minmax-gemmlowp-wasmsimd-ld64.c",
+ "src/qs8-igemm/gen/3x4c8-minmax-gemmlowp-wasmsimd-ld128.c",
"src/qs8-requantization/fp32-wasmsimd.c",
"src/qs8-requantization/gemmlowp-wasmsimd.c",
"src/qs8-vadd/gen/minmax-wasmsimd-x8.c",
@@ -4386,12 +4386,14 @@
"src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-prfm-cortex-a75.S",
"src/qc8-gemm/gen/1x16c4-minmax-fp32-aarch64-neondot-ld32.S",
"src/qc8-gemm/gen/1x16c4-minmax-fp32-aarch64-neondot-ld64.S",
+ "src/qc8-gemm/gen/2x8c16-minmax-fp32-aarch64-neon-mlal-padal.S",
"src/qc8-gemm/gen/4x16-minmax-fp32-aarch64-neon-mlal-lane-cortex-a53.S",
"src/qc8-gemm/gen/4x16-minmax-fp32-aarch64-neon-mlal-lane-prfm-cortex-a53.S",
"src/qc8-gemm/gen/4x16c4-minmax-fp32-aarch64-neondot-cortex-a55.S",
"src/qc8-gemm/gen/4x16c4-minmax-fp32-aarch64-neondot-ld32.S",
"src/qc8-gemm/gen/4x16c4-minmax-fp32-aarch64-neondot-ld64.S",
"src/qc8-gemm/gen/4x16c4-minmax-fp32-aarch64-neondot-ld128.S",
+ "src/qc8-igemm/gen/2x8c16-minmax-fp32-aarch64-neon-mlal-padal.S",
"src/qc8-igemm/gen/4x16-minmax-fp32-aarch64-neon-mlal-lane-cortex-a53.S",
"src/qc8-igemm/gen/4x16-minmax-fp32-aarch64-neon-mlal-lane-prfm-cortex-a53.S",
"src/qc8-igemm/gen/4x16c4-minmax-fp32-aarch64-neondot-cortex-a55.S",
@@ -4410,6 +4412,7 @@
"src/qs8-gemm/gen/2x8c8-minmax-gemmlowp-aarch64-neon-mlal-padal-prfm.S",
"src/qs8-gemm/gen/2x8c8-minmax-gemmlowp-aarch64-neon-mlal-padal.S",
"src/qs8-gemm/gen/2x8c8-minmax-gemmlowp-aarch64-neon-mull-padal.S",
+ "src/qs8-gemm/gen/2x8c16-minmax-fp32-aarch64-neon-mlal-padal.S",
"src/qs8-gemm/gen/2x8c16-minmax-gemmlowp-aarch64-neon-mlal-padal.S",
"src/qs8-gemm/gen/4x16-minmax-fp32-aarch64-neon-mlal-lane-cortex-a53.S",
"src/qs8-gemm/gen/4x16-minmax-fp32-aarch64-neon-mlal-lane-prfm-cortex-a53.S",
@@ -4431,6 +4434,7 @@
"src/qs8-igemm/gen/2x8c8-minmax-gemmlowp-aarch64-neon-mlal-padal-prfm-cortex-a53.S",
"src/qs8-igemm/gen/2x8c8-minmax-gemmlowp-aarch64-neon-mlal-padal-prfm.S",
"src/qs8-igemm/gen/2x8c8-minmax-gemmlowp-aarch64-neon-mlal-padal.S",
+ "src/qs8-igemm/gen/2x8c16-minmax-fp32-aarch64-neon-mlal-padal.S",
"src/qs8-igemm/gen/2x8c16-minmax-gemmlowp-aarch64-neon-mlal-padal.S",
"src/qs8-igemm/gen/4x16-minmax-fp32-aarch64-neon-mlal-lane-cortex-a53.S",
"src/qs8-igemm/gen/4x16-minmax-fp32-aarch64-neon-mlal-lane-prfm-cortex-a53.S",
diff --git a/CMakeLists.txt b/CMakeLists.txt
index 2a2a430..a862b30 100755
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -3572,12 +3572,14 @@
src/f32-igemm/gen/6x8-minmax-aarch64-neonfma-prfm-cortex-a75.S
src/qc8-gemm/gen/1x16c4-minmax-fp32-aarch64-neondot-ld32.S
src/qc8-gemm/gen/1x16c4-minmax-fp32-aarch64-neondot-ld64.S
+ src/qc8-gemm/gen/2x8c16-minmax-fp32-aarch64-neon-mlal-padal.S
src/qc8-gemm/gen/4x16-minmax-fp32-aarch64-neon-mlal-lane-cortex-a53.S
src/qc8-gemm/gen/4x16-minmax-fp32-aarch64-neon-mlal-lane-prfm-cortex-a53.S
src/qc8-gemm/gen/4x16c4-minmax-fp32-aarch64-neondot-cortex-a55.S
src/qc8-gemm/gen/4x16c4-minmax-fp32-aarch64-neondot-ld32.S
src/qc8-gemm/gen/4x16c4-minmax-fp32-aarch64-neondot-ld64.S
src/qc8-gemm/gen/4x16c4-minmax-fp32-aarch64-neondot-ld128.S
+ src/qc8-igemm/gen/2x8c16-minmax-fp32-aarch64-neon-mlal-padal.S
src/qc8-igemm/gen/4x16-minmax-fp32-aarch64-neon-mlal-lane-cortex-a53.S
src/qc8-igemm/gen/4x16-minmax-fp32-aarch64-neon-mlal-lane-prfm-cortex-a53.S
src/qc8-igemm/gen/4x16c4-minmax-fp32-aarch64-neondot-cortex-a55.S
@@ -3596,6 +3598,7 @@
src/qs8-gemm/gen/2x8c8-minmax-gemmlowp-aarch64-neon-mlal-padal-prfm.S
src/qs8-gemm/gen/2x8c8-minmax-gemmlowp-aarch64-neon-mlal-padal.S
src/qs8-gemm/gen/2x8c8-minmax-gemmlowp-aarch64-neon-mull-padal.S
+ src/qs8-gemm/gen/2x8c16-minmax-fp32-aarch64-neon-mlal-padal.S
src/qs8-gemm/gen/2x8c16-minmax-gemmlowp-aarch64-neon-mlal-padal.S
src/qs8-gemm/gen/4x16-minmax-fp32-aarch64-neon-mlal-lane-cortex-a53.S
src/qs8-gemm/gen/4x16-minmax-fp32-aarch64-neon-mlal-lane-prfm-cortex-a53.S
@@ -3617,6 +3620,7 @@
src/qs8-igemm/gen/2x8c8-minmax-gemmlowp-aarch64-neon-mlal-padal-prfm-cortex-a53.S
src/qs8-igemm/gen/2x8c8-minmax-gemmlowp-aarch64-neon-mlal-padal-prfm.S
src/qs8-igemm/gen/2x8c8-minmax-gemmlowp-aarch64-neon-mlal-padal.S
+ src/qs8-igemm/gen/2x8c16-minmax-fp32-aarch64-neon-mlal-padal.S
src/qs8-igemm/gen/2x8c16-minmax-gemmlowp-aarch64-neon-mlal-padal.S
src/qs8-igemm/gen/4x16-minmax-fp32-aarch64-neon-mlal-lane-cortex-a53.S
src/qs8-igemm/gen/4x16-minmax-fp32-aarch64-neon-mlal-lane-prfm-cortex-a53.S
diff --git a/scripts/generate-qs8-gemm.sh b/scripts/generate-qs8-gemm.sh
index de492eb..8135073 100755
--- a/scripts/generate-qs8-gemm.sh
+++ b/scripts/generate-qs8-gemm.sh
@@ -232,11 +232,6 @@
############################### AArch64 assembly ##############################
# Cortex A53 micro-kernel
-tools/xngen src/qs8-gemm/1x8c8-aarch64-neon-mlal-padal-cortex-a53.S.in -D PREFETCH=0 -o src/qs8-gemm/gen/1x8c8-minmax-gemmlowp-aarch64-neon-mlal-padal-cortex-a53.S
-tools/xngen src/qs8-gemm/1x8c8-aarch64-neon-mlal-padal-cortex-a53.S.in -D PREFETCH=1 -o src/qs8-gemm/gen/1x8c8-minmax-gemmlowp-aarch64-neon-mlal-padal-prfm-cortex-a53.S
-tools/xngen src/qs8-gemm/2x8c8-aarch64-neon-mlal-padal-cortex-a53.S.in -D PREFETCH=0 -o src/qs8-gemm/gen/2x8c8-minmax-gemmlowp-aarch64-neon-mlal-padal-cortex-a53.S
-tools/xngen src/qs8-gemm/2x8c8-aarch64-neon-mlal-padal-cortex-a53.S.in -D PREFETCH=1 -o src/qs8-gemm/gen/2x8c8-minmax-gemmlowp-aarch64-neon-mlal-padal-prfm-cortex-a53.S
-
tools/xngen src/qs8-gemm/4x16-aarch64-neon-mlal-lane-cortex-a53.S.in -D PREFETCH=0 -D REQUANTIZATION=GEMMLOWP -D CHANNELWISE=0 -o src/qs8-gemm/gen/4x16-minmax-gemmlowp-aarch64-neon-mlal-lane-cortex-a53.S
tools/xngen src/qs8-gemm/4x16-aarch64-neon-mlal-lane-cortex-a53.S.in -D PREFETCH=1 -D REQUANTIZATION=GEMMLOWP -D CHANNELWISE=0 -o src/qs8-gemm/gen/4x16-minmax-gemmlowp-aarch64-neon-mlal-lane-prfm-cortex-a53.S
@@ -246,15 +241,6 @@
tools/xngen src/qs8-gemm/4x16-aarch64-neon-mlal-lane-cortex-a53.S.in -D PREFETCH=0 -D REQUANTIZATION=FP32 -D CHANNELWISE=1 -o src/qc8-gemm/gen/4x16-minmax-fp32-aarch64-neon-mlal-lane-cortex-a53.S
tools/xngen src/qs8-gemm/4x16-aarch64-neon-mlal-lane-cortex-a53.S.in -D PREFETCH=1 -D REQUANTIZATION=FP32 -D CHANNELWISE=1 -o src/qc8-gemm/gen/4x16-minmax-fp32-aarch64-neon-mlal-lane-prfm-cortex-a53.S
-# Cortex A75 micro-kernel
-tools/xngen src/qs8-gemm/1x8c8-aarch64-neon-mlal-padal.S.in -D PREFETCH=0 -o src/qs8-gemm/gen/1x8c8-minmax-gemmlowp-aarch64-neon-mlal-padal.S
-tools/xngen src/qs8-gemm/1x8c8-aarch64-neon-mlal-padal.S.in -D PREFETCH=1 -o src/qs8-gemm/gen/1x8c8-minmax-gemmlowp-aarch64-neon-mlal-padal-prfm.S
-tools/xngen src/qs8-gemm/2x8c8-aarch64-neon-mlal-padal.S.in -D PREFETCH=0 -o src/qs8-gemm/gen/2x8c8-minmax-gemmlowp-aarch64-neon-mlal-padal.S
-tools/xngen src/qs8-gemm/2x8c8-aarch64-neon-mlal-padal.S.in -D PREFETCH=1 -o src/qs8-gemm/gen/2x8c8-minmax-gemmlowp-aarch64-neon-mlal-padal-prfm.S
-
-tools/xngen src/qs8-gemm/2x8c16-aarch64-neon-mlal-padal.S.in -o src/qs8-gemm/gen/2x8c16-minmax-gemmlowp-aarch64-neon-mlal-padal.S
-tools/xngen src/qs8-gemm/2x8c8-aarch64-neon-mull-padal.S.in -o src/qs8-gemm/gen/2x8c8-minmax-gemmlowp-aarch64-neon-mull-padal.S
-
### C4 micro-kernels
tools/xngen src/qs8-gemm/1x16c4-aarch64-neondot-ld32.S.in -D REQUANTIZATION=GEMMLOWP -D CHANNELWISE=0 -o src/qs8-gemm/gen/1x16c4-minmax-gemmlowp-aarch64-neondot-ld32.S
tools/xngen src/qs8-gemm/1x16c4-aarch64-neondot-ld64.S.in -D REQUANTIZATION=GEMMLOWP -D CHANNELWISE=0 -o src/qs8-gemm/gen/1x16c4-minmax-gemmlowp-aarch64-neondot-ld64.S
@@ -277,6 +263,23 @@
tools/xngen src/qs8-gemm/4x16c4-aarch64-neondot-ld64.S.in -D REQUANTIZATION=FP32 -D CHANNELWISE=1 -o src/qc8-gemm/gen/4x16c4-minmax-fp32-aarch64-neondot-ld64.S
tools/xngen src/qs8-gemm/4x16c4-aarch64-neondot-ld128.S.in -D REQUANTIZATION=FP32 -D CHANNELWISE=1 -o src/qc8-gemm/gen/4x16c4-minmax-fp32-aarch64-neondot-ld128.S
+### C8 / C16 micro-kernels
+tools/xngen src/qs8-gemm/1x8c8-aarch64-neon-mlal-padal-cortex-a53.S.in -D PREFETCH=0 -D REQUANTIZATION=GEMMLOWP -D CHANNELWISE=0 -o src/qs8-gemm/gen/1x8c8-minmax-gemmlowp-aarch64-neon-mlal-padal-cortex-a53.S
+tools/xngen src/qs8-gemm/1x8c8-aarch64-neon-mlal-padal-cortex-a53.S.in -D PREFETCH=1 -D REQUANTIZATION=GEMMLOWP -D CHANNELWISE=0 -o src/qs8-gemm/gen/1x8c8-minmax-gemmlowp-aarch64-neon-mlal-padal-prfm-cortex-a53.S
+tools/xngen src/qs8-gemm/2x8c8-aarch64-neon-mlal-padal-cortex-a53.S.in -D PREFETCH=0 -D REQUANTIZATION=GEMMLOWP -D CHANNELWISE=0 -o src/qs8-gemm/gen/2x8c8-minmax-gemmlowp-aarch64-neon-mlal-padal-cortex-a53.S
+tools/xngen src/qs8-gemm/2x8c8-aarch64-neon-mlal-padal-cortex-a53.S.in -D PREFETCH=1 -D REQUANTIZATION=GEMMLOWP -D CHANNELWISE=0 -o src/qs8-gemm/gen/2x8c8-minmax-gemmlowp-aarch64-neon-mlal-padal-prfm-cortex-a53.S
+tools/xngen src/qs8-gemm/1x8c8-aarch64-neon-mlal-padal.S.in -D PREFETCH=0 -D REQUANTIZATION=GEMMLOWP -D CHANNELWISE=0 -o src/qs8-gemm/gen/1x8c8-minmax-gemmlowp-aarch64-neon-mlal-padal.S
+tools/xngen src/qs8-gemm/1x8c8-aarch64-neon-mlal-padal.S.in -D PREFETCH=1 -D REQUANTIZATION=GEMMLOWP -D CHANNELWISE=0 -o src/qs8-gemm/gen/1x8c8-minmax-gemmlowp-aarch64-neon-mlal-padal-prfm.S
+tools/xngen src/qs8-gemm/2x8c8-aarch64-neon-mlal-padal.S.in -D PREFETCH=0 -D REQUANTIZATION=GEMMLOWP -D CHANNELWISE=0 -o src/qs8-gemm/gen/2x8c8-minmax-gemmlowp-aarch64-neon-mlal-padal.S
+tools/xngen src/qs8-gemm/2x8c8-aarch64-neon-mlal-padal.S.in -D PREFETCH=1 -D REQUANTIZATION=GEMMLOWP -D CHANNELWISE=0 -o src/qs8-gemm/gen/2x8c8-minmax-gemmlowp-aarch64-neon-mlal-padal-prfm.S
+tools/xngen src/qs8-gemm/2x8c8-aarch64-neon-mull-padal.S.in -D REQUANTIZATION=GEMMLOWP -D CHANNELWISE=0 -o src/qs8-gemm/gen/2x8c8-minmax-gemmlowp-aarch64-neon-mull-padal.S
+
+tools/xngen src/qs8-gemm/2x8c16-aarch64-neon-mlal-padal.S.in -D REQUANTIZATION=GEMMLOWP -D CHANNELWISE=0 -o src/qs8-gemm/gen/2x8c16-minmax-gemmlowp-aarch64-neon-mlal-padal.S
+
+tools/xngen src/qs8-gemm/2x8c16-aarch64-neon-mlal-padal.S.in -D REQUANTIZATION=FP32 -D CHANNELWISE=0 -o src/qs8-gemm/gen/2x8c16-minmax-fp32-aarch64-neon-mlal-padal.S
+
+tools/xngen src/qs8-gemm/2x8c16-aarch64-neon-mlal-padal.S.in -D REQUANTIZATION=FP32 -D CHANNELWISE=1 -o src/qc8-gemm/gen/2x8c16-minmax-fp32-aarch64-neon-mlal-padal.S
+
################################### x86 SSE ###################################
### C2 micro-kernels
tools/xngen src/qs8-gemm/MRx4c2-sse.c.in -D MR=1 -D SSE=2 -D AVX=0 -D XOP=0 -D CHANNELWISE=0 -D REQUANTIZATION=GEMMLOWP -D VARIANT=LD64 -o src/qs8-gemm/gen/1x4c2-minmax-gemmlowp-sse2-ld64.c
diff --git a/scripts/generate-qs8-igemm.sh b/scripts/generate-qs8-igemm.sh
index ee46633..1b5a661 100755
--- a/scripts/generate-qs8-igemm.sh
+++ b/scripts/generate-qs8-igemm.sh
@@ -229,11 +229,6 @@
############################### AArch64 assembly ##############################
# Cortex A53 micro-kernel
-tools/xngen src/qs8-igemm/1x8c8-aarch64-neon-mlal-padal-cortex-a53.S.in -D PREFETCH=0 -o src/qs8-igemm/gen/1x8c8-minmax-gemmlowp-aarch64-neon-mlal-padal-cortex-a53.S
-tools/xngen src/qs8-igemm/1x8c8-aarch64-neon-mlal-padal-cortex-a53.S.in -D PREFETCH=1 -o src/qs8-igemm/gen/1x8c8-minmax-gemmlowp-aarch64-neon-mlal-padal-prfm-cortex-a53.S
-tools/xngen src/qs8-igemm/2x8c8-aarch64-neon-mlal-padal-cortex-a53.S.in -D PREFETCH=0 -o src/qs8-igemm/gen/2x8c8-minmax-gemmlowp-aarch64-neon-mlal-padal-cortex-a53.S
-tools/xngen src/qs8-igemm/2x8c8-aarch64-neon-mlal-padal-cortex-a53.S.in -D PREFETCH=1 -o src/qs8-igemm/gen/2x8c8-minmax-gemmlowp-aarch64-neon-mlal-padal-prfm-cortex-a53.S
-
tools/xngen src/qs8-igemm/4x16-aarch64-neon-mlal-lane-cortex-a53.S.in -D PREFETCH=0 -D REQUANTIZATION=GEMMLOWP -D CHANNELWISE=0 -o src/qs8-igemm/gen/4x16-minmax-gemmlowp-aarch64-neon-mlal-lane-cortex-a53.S
tools/xngen src/qs8-igemm/4x16-aarch64-neon-mlal-lane-cortex-a53.S.in -D PREFETCH=1 -D REQUANTIZATION=GEMMLOWP -D CHANNELWISE=0 -o src/qs8-igemm/gen/4x16-minmax-gemmlowp-aarch64-neon-mlal-lane-prfm-cortex-a53.S
@@ -243,14 +238,6 @@
tools/xngen src/qs8-igemm/4x16-aarch64-neon-mlal-lane-cortex-a53.S.in -D PREFETCH=0 -D REQUANTIZATION=FP32 -D CHANNELWISE=1 -o src/qc8-igemm/gen/4x16-minmax-fp32-aarch64-neon-mlal-lane-cortex-a53.S
tools/xngen src/qs8-igemm/4x16-aarch64-neon-mlal-lane-cortex-a53.S.in -D PREFETCH=1 -D REQUANTIZATION=FP32 -D CHANNELWISE=1 -o src/qc8-igemm/gen/4x16-minmax-fp32-aarch64-neon-mlal-lane-prfm-cortex-a53.S
-# Cortex A75 micro-kernel
-tools/xngen src/qs8-igemm/1x8c8-aarch64-neon-mlal-padal.S.in -D PREFETCH=0 -o src/qs8-igemm/gen/1x8c8-minmax-gemmlowp-aarch64-neon-mlal-padal.S
-tools/xngen src/qs8-igemm/1x8c8-aarch64-neon-mlal-padal.S.in -D PREFETCH=1 -o src/qs8-igemm/gen/1x8c8-minmax-gemmlowp-aarch64-neon-mlal-padal-prfm.S
-tools/xngen src/qs8-igemm/2x8c8-aarch64-neon-mlal-padal.S.in -D PREFETCH=0 -o src/qs8-igemm/gen/2x8c8-minmax-gemmlowp-aarch64-neon-mlal-padal.S
-tools/xngen src/qs8-igemm/2x8c8-aarch64-neon-mlal-padal.S.in -D PREFETCH=1 -o src/qs8-igemm/gen/2x8c8-minmax-gemmlowp-aarch64-neon-mlal-padal-prfm.S
-
-tools/xngen src/qs8-igemm/2x8c16-aarch64-neon-mlal-padal.S.in -o src/qs8-igemm/gen/2x8c16-minmax-gemmlowp-aarch64-neon-mlal-padal.S
-
### C4 micro-kernels
tools/xngen src/qs8-igemm/4x16c4-aarch64-neondot-cortex-a55.S.in -D REQUANTIZATION=GEMMLOWP -D CHANNELWISE=0 -o src/qs8-igemm/gen/4x16c4-minmax-gemmlowp-aarch64-neondot-cortex-a55.S
tools/xngen src/qs8-igemm/4x16c4-aarch64-neondot-ld64.S.in -D REQUANTIZATION=GEMMLOWP -D CHANNELWISE=0 -o src/qs8-igemm/gen/4x16c4-minmax-gemmlowp-aarch64-neondot-ld64.S
@@ -264,6 +251,22 @@
tools/xngen src/qs8-igemm/4x16c4-aarch64-neondot-ld64.S.in -D REQUANTIZATION=FP32 -D CHANNELWISE=1 -o src/qc8-igemm/gen/4x16c4-minmax-fp32-aarch64-neondot-ld64.S
tools/xngen src/qs8-igemm/4x16c4-aarch64-neondot-ld128.S.in -D REQUANTIZATION=FP32 -D CHANNELWISE=1 -o src/qc8-igemm/gen/4x16c4-minmax-fp32-aarch64-neondot-ld128.S
+### C8 / C16 micro-kernels
+tools/xngen src/qs8-igemm/1x8c8-aarch64-neon-mlal-padal-cortex-a53.S.in -D PREFETCH=0 -D REQUANTIZATION=GEMMLOWP -D CHANNELWISE=0 -o src/qs8-igemm/gen/1x8c8-minmax-gemmlowp-aarch64-neon-mlal-padal-cortex-a53.S
+tools/xngen src/qs8-igemm/1x8c8-aarch64-neon-mlal-padal-cortex-a53.S.in -D PREFETCH=1 -D REQUANTIZATION=GEMMLOWP -D CHANNELWISE=0 -o src/qs8-igemm/gen/1x8c8-minmax-gemmlowp-aarch64-neon-mlal-padal-prfm-cortex-a53.S
+tools/xngen src/qs8-igemm/2x8c8-aarch64-neon-mlal-padal-cortex-a53.S.in -D PREFETCH=0 -D REQUANTIZATION=GEMMLOWP -D CHANNELWISE=0 -o src/qs8-igemm/gen/2x8c8-minmax-gemmlowp-aarch64-neon-mlal-padal-cortex-a53.S
+tools/xngen src/qs8-igemm/2x8c8-aarch64-neon-mlal-padal-cortex-a53.S.in -D PREFETCH=1 -D REQUANTIZATION=GEMMLOWP -D CHANNELWISE=0 -o src/qs8-igemm/gen/2x8c8-minmax-gemmlowp-aarch64-neon-mlal-padal-prfm-cortex-a53.S
+tools/xngen src/qs8-igemm/1x8c8-aarch64-neon-mlal-padal.S.in -D PREFETCH=0 -D REQUANTIZATION=GEMMLOWP -D CHANNELWISE=0 -o src/qs8-igemm/gen/1x8c8-minmax-gemmlowp-aarch64-neon-mlal-padal.S
+tools/xngen src/qs8-igemm/1x8c8-aarch64-neon-mlal-padal.S.in -D PREFETCH=1 -D REQUANTIZATION=GEMMLOWP -D CHANNELWISE=0 -o src/qs8-igemm/gen/1x8c8-minmax-gemmlowp-aarch64-neon-mlal-padal-prfm.S
+tools/xngen src/qs8-igemm/2x8c8-aarch64-neon-mlal-padal.S.in -D PREFETCH=0 -D REQUANTIZATION=GEMMLOWP -D CHANNELWISE=0 -o src/qs8-igemm/gen/2x8c8-minmax-gemmlowp-aarch64-neon-mlal-padal.S
+tools/xngen src/qs8-igemm/2x8c8-aarch64-neon-mlal-padal.S.in -D PREFETCH=1 -D REQUANTIZATION=GEMMLOWP -D CHANNELWISE=0 -o src/qs8-igemm/gen/2x8c8-minmax-gemmlowp-aarch64-neon-mlal-padal-prfm.S
+tools/xngen src/qs8-igemm/2x8c16-aarch64-neon-mlal-padal.S.in -D REQUANTIZATION=GEMMLOWP -D CHANNELWISE=0 -o src/qs8-igemm/gen/2x8c16-minmax-gemmlowp-aarch64-neon-mlal-padal.S
+
+tools/xngen src/qs8-igemm/2x8c16-aarch64-neon-mlal-padal.S.in -D REQUANTIZATION=FP32 -D CHANNELWISE=0 -o src/qs8-igemm/gen/2x8c16-minmax-fp32-aarch64-neon-mlal-padal.S
+
+tools/xngen src/qs8-igemm/2x8c16-aarch64-neon-mlal-padal.S.in -D REQUANTIZATION=FP32 -D CHANNELWISE=1 -o src/qc8-igemm/gen/2x8c16-minmax-fp32-aarch64-neon-mlal-padal.S
+
+
################################### x86 SSE ###################################
### C2 micro-kernels
tools/xngen src/qs8-igemm/MRx4c2-sse.c.in -D MR=1 -D SSE=2 -D AVX=0 -D XOP=0 -D CHANNELWISE=0 -D REQUANTIZATION=GEMMLOWP -D VARIANT=LD64 -o src/qs8-igemm/gen/1x4c2-minmax-gemmlowp-sse2-ld64.c
diff --git a/src/qc8-gemm/gen/2x8c16-minmax-fp32-aarch64-neon-mlal-padal.S b/src/qc8-gemm/gen/2x8c16-minmax-fp32-aarch64-neon-mlal-padal.S
new file mode 100644
index 0000000..205df60
--- /dev/null
+++ b/src/qc8-gemm/gen/2x8c16-minmax-fp32-aarch64-neon-mlal-padal.S
@@ -0,0 +1,220 @@
+// Auto-generated file. Do not edit!
+// Template: src/qs8-gemm/2x8c16-aarch64-neon-mlal-padal.S.in
+// Generator: tools/xngen
+//
+// Copyright 2021 Google LLC
+//
+// This source code is licensed under the BSD-style license found in the
+// LICENSE file in the root directory of this source tree.
+
+
+#include <xnnpack/assembly.h>
+
+# void xnn_qc8_gemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal(
+# size_t mr, x0
+# size_t nc, x1
+# size_t kc, x2 / x0
+# const int8_t* restrict a, x3
+# size_t a_stride, x4
+# const void* restrict w, x5
+# int8_t* restrict c, x6
+# size_t cm_stride, x7
+# size_t cn_stride, [sp] -> x10
+# const union xnn_qs8_minmax_params params) [sp + 8] -> x11
+
+# d8-d15, x19-x30 need to be preserved if used. x18 is reserved by the OS.
+
+# Register usage
+# A0 x3 v0
+# A1 x4 v1
+# B x5 v4 v5 v6 v7
+# C0 x7 v16 v18 v20 v22 v24 v26 v28 v30
+# C1 x8 v17 v19 v21 v23 v25 v27 v29 v31
+# temp0 v2 v10 v12 v14
+# temp1 v3 v11 v13 v15
+# unused v8 v9
+
+BEGIN_FUNCTION xnn_qc8_gemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal
+
+ # Clamp A and C pointers
+ CMP x0, 2 // if mr < 2
+ STP d10, d11, [sp, -48]!
+ ADD x4, x3, x4 // a1 = a0 + a_stride
+ STP d12, d13, [sp, 16]
+ ADD x7, x6, x7 // c1 = c0 + cm_stride
+ STP d14, d15, [sp, 32]
+ CSEL x4, x3, x4, LO // a1 = a0
+ ADD x2, x2, 15 // kc = (kc + 15) & ~15
+ CSEL x7, x6, x7, LO // c1 = c0
+ BIC x2, x2, 15
+
+ .p2align 3
+0:
+ # Load initial bias from w into accumulators
+ MOV x0, x2 // k = kc
+ LDP s16, s18, [x5], 8
+ MOV v17.16b, v16.16b
+ MOV v19.16b, v18.16b
+ LDP s20, s22, [x5], 8
+ MOV v21.16b, v20.16b
+ MOV v23.16b, v22.16b
+ LDP s24, s26, [x5], 8
+ MOV v25.16b, v24.16b
+ MOV v27.16b, v26.16b
+ LDP s28, s30, [x5], 8
+ MOV v29.16b, v28.16b
+ LDP x10, x11, [sp, 48] // cn_stride, params
+ MOV v31.16b, v30.16b
+
+ # Main loop - 16 bytes of A
+ .p2align 3
+1:
+ LDR q0, [x3], 16
+ LDP q4, q5, [x5]
+ LDR q1, [x4], 16
+ LDP q6, q7, [x5, 32]
+ SMULL v2.8h, v4.8b, v0.8b
+ SMULL v3.8h, v4.8b, v1.8b
+ SMULL v10.8h, v5.8b, v0.8b
+ SMULL v11.8h, v5.8b, v1.8b
+ SMLAL2 v2.8h, v4.16b, v0.16b
+ SMLAL2 v3.8h, v4.16b, v1.16b
+ SMLAL2 v10.8h, v5.16b, v0.16b
+ SMLAL2 v11.8h, v5.16b, v1.16b
+ SMULL v12.8h, v6.8b, v0.8b
+ SADALP v16.4s, v2.8h
+ SMULL v13.8h, v6.8b, v1.8b
+ SADALP v17.4s, v3.8h
+ SMULL v14.8h, v7.8b, v0.8b
+ SADALP v18.4s, v10.8h
+ SMULL v15.8h, v7.8b, v1.8b
+ SADALP v19.4s, v11.8h
+ LDP q4, q5, [x5, 64]
+ SMLAL2 v12.8h, v6.16b, v0.16b
+ SMLAL2 v13.8h, v6.16b, v1.16b
+ SMLAL2 v14.8h, v7.16b, v0.16b
+ SMLAL2 v15.8h, v7.16b, v1.16b
+ SMULL v2.8h, v4.8b, v0.8b
+ SADALP v20.4s, v12.8h
+ SMULL v3.8h, v4.8b, v1.8b
+ SADALP v21.4s, v13.8h
+ SMULL v10.8h, v5.8b, v0.8b
+ SADALP v22.4s, v14.8h
+ SMULL v11.8h, v5.8b, v1.8b
+ SADALP v23.4s, v15.8h
+ LDP q6, q7, [x5, 96]
+
+ SMLAL2 v2.8h, v4.16b, v0.16b
+ SMLAL2 v3.8h, v4.16b, v1.16b
+ SMLAL2 v10.8h, v5.16b, v0.16b
+ SMLAL2 v11.8h, v5.16b, v1.16b
+ ADD x5, x5, 128
+ SMULL v12.8h, v6.8b, v0.8b
+ SADALP v24.4s, v2.8h
+ SMULL v13.8h, v6.8b, v1.8b
+ SADALP v25.4s, v3.8h
+ SMULL v14.8h, v7.8b, v0.8b
+ SADALP v26.4s, v10.8h
+ SMULL v15.8h, v7.8b, v1.8b
+ SADALP v27.4s, v11.8h
+ SUBS x0, x0, 16
+ SMLAL2 v12.8h, v6.16b, v0.16b
+ SMLAL2 v13.8h, v6.16b, v1.16b
+ SMLAL2 v14.8h, v7.16b, v0.16b
+ SMLAL2 v15.8h, v7.16b, v1.16b
+ SADALP v28.4s, v12.8h
+ SADALP v29.4s, v13.8h
+ SADALP v30.4s, v14.8h
+ SADALP v31.4s, v15.8h
+ B.HI 1b
+
+ # Add columns
+ ADDP v16.4s, v16.4s, v18.4s
+ ADDP v20.4s, v20.4s, v22.4s
+ ADDP v24.4s, v24.4s, v26.4s
+ ADDP v28.4s, v28.4s, v30.4s
+ ADDP v17.4s, v17.4s, v19.4s
+ ADDP v21.4s, v21.4s, v23.4s
+ ADDP v25.4s, v25.4s, v27.4s
+ ADDP v29.4s, v29.4s, v31.4s
+ ADDP v0.4s, v16.4s, v20.4s
+ ADDP v1.4s, v24.4s, v28.4s
+ ADDP v2.4s, v17.4s, v21.4s
+ ADDP v3.4s, v25.4s, v29.4s
+
+ # Load per channel scale values from weights
+ SCVTF v0.4s, v0.4s
+ LDR q4, [x5], 16
+ SCVTF v1.4s, v1.4s
+ LDR q5, [x5], 16
+ SCVTF v2.4s, v2.4s
+ SCVTF v3.4s, v3.4s
+ FMUL v0.4s, v0.4s, v4.4s
+ FMUL v1.4s, v1.4s, v5.4s
+ FMUL v2.4s, v2.4s, v4.4s
+ FMUL v3.4s, v3.4s, v5.4s
+
+ FCVTNS v0.4s, v0.4s
+ FCVTNS v1.4s, v1.4s
+ FCVTNS v2.4s, v2.4s
+ FCVTNS v3.4s, v3.4s
+
+ LD1R {v5.8h}, [x11], 2
+ SQXTN v0.4h, v0.4s
+ SQXTN v2.4h, v2.4s
+ SQXTN2 v0.8h, v1.4s
+ SQXTN2 v2.8h, v3.4s
+ SUBS x1, x1, 8
+ SQADD v0.8h, v0.8h, v5.8h
+ SQADD v1.8h, v2.8h, v5.8h
+ SQXTN v0.8b, v0.8h
+ SQXTN2 v0.16b, v1.8h
+ LD1R {v1.16b}, [x11], 1
+ LD1R {v2.16b}, [x11]
+ SMAX v0.16b, v0.16b, v1.16b
+ SMIN v0.16b, v0.16b, v2.16b
+ B.LO 2f
+
+ # Store full 2 x 8
+ ST1 {v0.8b}, [x6], x10
+ SUB x3, x3, x2 // a0 -= kc
+ ST1 {v0.d}[1], [x7], x10
+ SUB x4, x4, x2 // a1 -= kc
+ B.HI 0b
+
+ # Restore d10-d15 from stack
+ LDP d14, d15, [sp, 32]
+ LDP d12, d13, [sp, 16]
+ LDP d10, d11, [sp], 48
+ RET
+
+ # Store odd width
+ .p2align 3
+2:
+ TBZ x1, 2, 3f
+ STR s0, [x6], 4
+ ST1 {v0.s}[2], [x7], 4
+ EXT v0.16b, v0.16b, v0.16b, 4
+
+3:
+ TBZ x1, 1, 4f
+ ST1 {v0.h}[0], [x6], 2
+ ST1 {v0.h}[4], [x7], 2
+ EXT v0.16b, v0.16b, v0.16b, 2
+4:
+ TBZ x1, 0, 5f
+ ST1 {v0.b}[0], [x6]
+ ST1 {v0.b}[8], [x7]
+5:
+ # Restore d10-d15 from stack
+ LDP d14, d15, [sp, 32]
+ LDP d12, d13, [sp, 16]
+ LDP d10, d11, [sp], 48
+ RET
+
+END_FUNCTION xnn_qc8_gemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal
+
+#ifdef __ELF__
+.section ".note.GNU-stack","",%progbits
+#endif
+
diff --git a/src/qc8-igemm/gen/2x8c16-minmax-fp32-aarch64-neon-mlal-padal.S b/src/qc8-igemm/gen/2x8c16-minmax-fp32-aarch64-neon-mlal-padal.S
new file mode 100644
index 0000000..c14a4c6
--- /dev/null
+++ b/src/qc8-igemm/gen/2x8c16-minmax-fp32-aarch64-neon-mlal-padal.S
@@ -0,0 +1,241 @@
+// Auto-generated file. Do not edit!
+// Template: src/qs8-igemm/2x8c16-aarch64-neon-mlal-padal.S.in
+// Generator: tools/xngen
+//
+// Copyright 2021 Google LLC
+//
+// This source code is licensed under the BSD-style license found in the
+// LICENSE file in the root directory of this source tree.
+
+
+#include <xnnpack/assembly.h>
+
+# void xnn_qc8_igemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal(
+# size_t mr, x0
+# size_t nc, x1
+# size_t kc, x2 / x0
+# size_t ks, x3 / x9
+# const int8_t**restrict a, x4
+# const int8_t* restrict w, x5
+# int8_t* restrict c, x6
+# size_t cm_stride, x7
+# size_t cn_stride, [sp] -> x10
+# size_t a_offset, [sp + 8] -> x8
+# const float* zero, [sp + 16] -> x12
+# const union xnn_qs8_minmax_params params [sp + 24] -> x11
+
+# d8-d15, x19-x30 need to be preserved if used. x18 is reserved by the OS.
+
+# Register usage
+# A0 x13 v0
+# A1 x15 v1
+# B x5 v4 v5 v6 v7
+# C0 x6 v16 v18 v20 v22 v24 v26 v28 v30
+# C1 x7 v17 v19 v21 v23 v25 v27 v29 v31
+# temp0 v2 v10 v12 v14
+# temp1 v3 v11 v13 v15
+# unused v8 v9
+
+BEGIN_FUNCTION xnn_qc8_igemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal
+
+ # Clamp C pointers
+ LDP x10, x8, [sp] // Load cn_stride, a_offset
+ CMP x0, 2 // if mr < 2
+ LDP x12, x11, [sp, 16] // Load zero, params pointer
+ ADD x7, x6, x7 // c1 = c0 + cm_stride
+ STP d10, d11, [sp, -48]!
+ ADD x2, x2, 15 // kc = (kc + 15) & ~15
+ STP d12, d13, [sp, 16]
+ CSEL x7, x6, x7, LO // c1 = c0
+ STP d14, d15, [sp, 32]
+ BIC x2, x2, 15
+
+ .p2align 3
+0:
+ # Load initial bias from w into accumulators
+ LDP s16, s18, [x5], 8
+ MOV v17.16b, v16.16b
+ MOV v19.16b, v18.16b
+ LDP s20, s22, [x5], 8
+ MOV v21.16b, v20.16b
+ MOV v23.16b, v22.16b
+ LDP s24, s26, [x5], 8
+ MOV v25.16b, v24.16b
+ MOV v27.16b, v26.16b
+ LDP s28, s30, [x5], 8
+ MOV v29.16b, v28.16b
+ MOV v31.16b, v30.16b
+ MOV x9, x3 // p = ks
+
+ .p2align 3
+1:
+ # Load next 2 A pointers
+ LDP x13, x15, [x4], 16
+
+ CMP x13, x12 // if a0 == zero
+ ADD x13, x13, x8 // a0 += a_offset
+ CSEL x13, x12, x13, EQ // a0 = zero, else += a0 + a_offset
+ CMP x15, x12 // if a1 == zero
+ ADD x15, x15, x8 // a1 += a_offset
+ CSEL x15, x12, x15, EQ // a1 = zero, else += a1 + a_offset
+
+ MOV x0, x2 // k = kc
+
+ # Main loop - 16 bytes of A
+ .p2align 3
+2:
+ LDR q0, [x13], 16
+ LDP q4, q5, [x5]
+ LDR q1, [x15], 16
+ LDP q6, q7, [x5, 32]
+ SMULL v2.8h, v4.8b, v0.8b
+ SMULL v3.8h, v4.8b, v1.8b
+ SMULL v10.8h, v5.8b, v0.8b
+ SMULL v11.8h, v5.8b, v1.8b
+ SMLAL2 v2.8h, v4.16b, v0.16b
+ SMLAL2 v3.8h, v4.16b, v1.16b
+ SMLAL2 v10.8h, v5.16b, v0.16b
+ SMLAL2 v11.8h, v5.16b, v1.16b
+ SMULL v12.8h, v6.8b, v0.8b
+ SADALP v16.4s, v2.8h
+ SMULL v13.8h, v6.8b, v1.8b
+ SADALP v17.4s, v3.8h
+ SMULL v14.8h, v7.8b, v0.8b
+ SADALP v18.4s, v10.8h
+ SMULL v15.8h, v7.8b, v1.8b
+ SADALP v19.4s, v11.8h
+ LDP q4, q5, [x5, 64]
+ SMLAL2 v12.8h, v6.16b, v0.16b
+ SMLAL2 v13.8h, v6.16b, v1.16b
+ SMLAL2 v14.8h, v7.16b, v0.16b
+ SMLAL2 v15.8h, v7.16b, v1.16b
+ SMULL v2.8h, v4.8b, v0.8b
+ SADALP v20.4s, v12.8h
+ SMULL v3.8h, v4.8b, v1.8b
+ SADALP v21.4s, v13.8h
+ SMULL v10.8h, v5.8b, v0.8b
+ SADALP v22.4s, v14.8h
+ SMULL v11.8h, v5.8b, v1.8b
+ SADALP v23.4s, v15.8h
+ LDP q6, q7, [x5, 96]
+
+ SMLAL2 v2.8h, v4.16b, v0.16b
+ SMLAL2 v3.8h, v4.16b, v1.16b
+ SMLAL2 v10.8h, v5.16b, v0.16b
+ SMLAL2 v11.8h, v5.16b, v1.16b
+ ADD x5, x5, 128
+ SMULL v12.8h, v6.8b, v0.8b
+ SADALP v24.4s, v2.8h
+ SMULL v13.8h, v6.8b, v1.8b
+ SADALP v25.4s, v3.8h
+ SMULL v14.8h, v7.8b, v0.8b
+ SADALP v26.4s, v10.8h
+ SMULL v15.8h, v7.8b, v1.8b
+ SADALP v27.4s, v11.8h
+ SUBS x0, x0, 16
+ SMLAL2 v12.8h, v6.16b, v0.16b
+ SMLAL2 v13.8h, v6.16b, v1.16b
+ SMLAL2 v14.8h, v7.16b, v0.16b
+ SMLAL2 v15.8h, v7.16b, v1.16b
+ SADALP v28.4s, v12.8h
+ SADALP v29.4s, v13.8h
+ SADALP v30.4s, v14.8h
+ SADALP v31.4s, v15.8h
+ B.HI 2b
+
+ # ks loop
+ SUBS x9, x9, 16 // ks -= MR * sizeof(int8_t*)
+ B.HI 1b
+
+ # Add columns
+ ADDP v16.4s, v16.4s, v18.4s
+ ADDP v20.4s, v20.4s, v22.4s
+ ADDP v24.4s, v24.4s, v26.4s
+ ADDP v28.4s, v28.4s, v30.4s
+ ADDP v17.4s, v17.4s, v19.4s
+ ADDP v21.4s, v21.4s, v23.4s
+ ADDP v25.4s, v25.4s, v27.4s
+ ADDP v29.4s, v29.4s, v31.4s
+ ADDP v0.4s, v16.4s, v20.4s
+ ADDP v1.4s, v24.4s, v28.4s
+ ADDP v2.4s, v17.4s, v21.4s
+ ADDP v3.4s, v25.4s, v29.4s
+
+ # Load per channel scale values from weights
+ SCVTF v0.4s, v0.4s
+ LDR q4, [x5], 16
+ SCVTF v1.4s, v1.4s
+ LDR q5, [x5], 16
+ SCVTF v2.4s, v2.4s
+ SCVTF v3.4s, v3.4s
+ FMUL v0.4s, v0.4s, v4.4s
+ FMUL v1.4s, v1.4s, v5.4s
+ FMUL v2.4s, v2.4s, v4.4s
+ FMUL v3.4s, v3.4s, v5.4s
+
+ FCVTNS v0.4s, v0.4s
+ FCVTNS v1.4s, v1.4s
+ FCVTNS v2.4s, v2.4s
+ FCVTNS v3.4s, v3.4s
+
+ LD1R {v5.8h}, [x11], 2
+ SQXTN v0.4h, v0.4s
+ SQXTN v2.4h, v2.4s
+ SQXTN2 v0.8h, v1.4s
+ SQXTN2 v2.8h, v3.4s
+ SUBS x1, x1, 8
+ SQADD v0.8h, v0.8h, v5.8h
+ SQADD v1.8h, v2.8h, v5.8h
+ SQXTN v0.8b, v0.8h
+ SQXTN2 v0.16b, v1.8h
+ LD1R {v1.16b}, [x11], 1
+ LD1R {v2.16b}, [x11]
+ SMAX v0.16b, v0.16b, v1.16b
+ SUB x11, x11, 3 // rewind params pointer
+ SMIN v0.16b, v0.16b, v2.16b
+ B.LO 3f
+
+ # Store full 2 x 8
+ ST1 {v0.d}[1], [x7], x10
+ SUB x4, x4, x3 // a -= ks
+ ST1 {v0.8b}, [x6], x10
+
+ # nc loop
+ B.HI 0b
+
+ # Restore d10-d15 from stack
+ LDP d14, d15, [sp, 32]
+ LDP d12, d13, [sp, 16]
+ LDP d10, d11, [sp], 48
+ RET
+
+ # Store odd width
+ .p2align 3
+3:
+ TBZ x1, 2, 4f
+ ST1 {v0.s}[2], [x7], 4
+ STR s0, [x6], 4
+ EXT v0.16b, v0.16b, v0.16b, 4
+
+4:
+ TBZ x1, 1, 5f
+ ST1 {v0.h}[4], [x7], 2
+ ST1 {v0.h}[0], [x6], 2
+ EXT v0.16b, v0.16b, v0.16b, 2
+5:
+ TBZ x1, 0, 6f
+ ST1 {v0.b}[8], [x7]
+ ST1 {v0.b}[0], [x6]
+6:
+ # Restore d10-d15 from stack
+ LDP d14, d15, [sp, 32]
+ LDP d12, d13, [sp, 16]
+ LDP d10, d11, [sp], 48
+ RET
+
+END_FUNCTION xnn_qc8_igemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal
+
+#ifdef __ELF__
+.section ".note.GNU-stack","",%progbits
+#endif
+
diff --git a/src/qs8-gemm/2x8c16-aarch64-neon-mlal-padal.S.in b/src/qs8-gemm/2x8c16-aarch64-neon-mlal-padal.S.in
index d021f8b..7f3cf72 100644
--- a/src/qs8-gemm/2x8c16-aarch64-neon-mlal-padal.S.in
+++ b/src/qs8-gemm/2x8c16-aarch64-neon-mlal-padal.S.in
@@ -3,9 +3,14 @@
// This source code is licensed under the BSD-style license found in the
// LICENSE file in the root directory of this source tree.
+$assert REQUANTIZATION in ["GEMMLOWP", "FP32"]
+$assert not CHANNELWISE or REQUANTIZATION == "FP32"
+
#include <xnnpack/assembly.h>
-# void xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal(
+$DATATYPE = "qc8" if CHANNELWISE else "qs8"
+$CONV_PARAMS = "xnn_qs8_minmax_params" if CHANNELWISE else "xnn_qs8_conv_minmax_params"
+# void xnn_${DATATYPE}_gemm_minmax_${REQUANTIZATION.lower()}_ukernel_2x8c16__aarch64_neon_mlal_padal(
# size_t mr, x0
# size_t nc, x1
# size_t kc, x2 / x0
@@ -15,7 +20,7 @@
# int8_t* restrict c, x6
# size_t cm_stride, x7
# size_t cn_stride, [sp] -> x10
-# const union xnn_qs8_conv_minmax_params params) [sp + 8] -> x11
+# const union ${CONV_PARAMS} params) [sp + 8] -> x11
# d8-d15, x19-x30 need to be preserved if used. x18 is reserved by the OS.
@@ -29,7 +34,7 @@
# temp1 v3 v11 v13 v15
# unused v8 v9
-BEGIN_FUNCTION xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal
+BEGIN_FUNCTION xnn_${DATATYPE}_gemm_minmax_${REQUANTIZATION.lower()}_ukernel_2x8c16__aarch64_neon_mlal_padal
# Clamp A and C pointers
CMP x0, 2 // if mr < 2
@@ -126,10 +131,12 @@
# Add columns
ADDP v16.4s, v16.4s, v18.4s
ADDP v20.4s, v20.4s, v22.4s
- LD1R {v4.4s}, [x11], 4
+ $if REQUANTIZATION == "GEMMLOWP":
+ LD1R {v4.4s}, [x11], 4
ADDP v24.4s, v24.4s, v26.4s
ADDP v28.4s, v28.4s, v30.4s
- LD1R {v7.4s}, [x11], 4
+ $if REQUANTIZATION == "GEMMLOWP":
+ LD1R {v7.4s}, [x11], 4
ADDP v17.4s, v17.4s, v19.4s
ADDP v21.4s, v21.4s, v23.4s
ADDP v25.4s, v25.4s, v27.4s
@@ -139,25 +146,56 @@
ADDP v2.4s, v17.4s, v21.4s
ADDP v3.4s, v25.4s, v29.4s
- # Apply params - scale, shift, bias and clamp
- SQRDMULH v0.4s, v0.4s, v4.4s
- SQRDMULH v1.4s, v1.4s, v4.4s
- SQRDMULH v2.4s, v2.4s, v4.4s
- SQRDMULH v3.4s, v3.4s, v4.4s
- CMEQ v4.4s, v7.4s, 0
+ $if REQUANTIZATION == "GEMMLOWP":
+ # Apply params - scale, shift, bias and clamp
+ SQRDMULH v0.4s, v0.4s, v4.4s
+ SQRDMULH v1.4s, v1.4s, v4.4s
+ SQRDMULH v2.4s, v2.4s, v4.4s
+ SQRDMULH v3.4s, v3.4s, v4.4s
+ CMEQ v4.4s, v7.4s, 0
+ BIC v6.16b, v0.16b, v4.16b
+ BIC v16.16b, v1.16b, v4.16b
+ BIC v17.16b, v2.16b, v4.16b
+ BIC v4.16b, v3.16b, v4.16b
+ SSRA v0.4s, v6.4s, 31
+ SSRA v1.4s, v16.4s, 31
+ SSRA v2.4s, v17.4s, 31
+ SSRA v3.4s, v4.4s, 31
+ SRSHL v0.4s, v0.4s, v7.4s
+ SRSHL v1.4s, v1.4s, v7.4s
+ SRSHL v2.4s, v2.4s, v7.4s
+ SRSHL v3.4s, v3.4s, v7.4s
+ $elif REQUANTIZATION == "FP32":
+ $if not CHANNELWISE:
+ # Apply params - scale, bias and clamp
+ SCVTF v0.4s, v0.4s
+ LD1R {v4.4s}, [x11], 4
+ SCVTF v1.4s, v1.4s
+ SCVTF v2.4s, v2.4s
+ SCVTF v3.4s, v3.4s
+ FMUL v0.4s, v0.4s, v4.4s
+ FMUL v1.4s, v1.4s, v4.4s
+ FMUL v2.4s, v2.4s, v4.4s
+ FMUL v3.4s, v3.4s, v4.4s
+ $else:
+ # Load per channel scale values from weights
+ SCVTF v0.4s, v0.4s
+ LDR q4, [x5], 16
+ SCVTF v1.4s, v1.4s
+ LDR q5, [x5], 16
+ SCVTF v2.4s, v2.4s
+ SCVTF v3.4s, v3.4s
+ FMUL v0.4s, v0.4s, v4.4s
+ FMUL v1.4s, v1.4s, v5.4s
+ FMUL v2.4s, v2.4s, v4.4s
+ FMUL v3.4s, v3.4s, v5.4s
+
+ FCVTNS v0.4s, v0.4s
+ FCVTNS v1.4s, v1.4s
+ FCVTNS v2.4s, v2.4s
+ FCVTNS v3.4s, v3.4s
+
LD1R {v5.8h}, [x11], 2
- BIC v6.16b, v0.16b, v4.16b
- BIC v16.16b, v1.16b, v4.16b
- BIC v17.16b, v2.16b, v4.16b
- BIC v4.16b, v3.16b, v4.16b
- SSRA v0.4s, v6.4s, 31
- SSRA v1.4s, v16.4s, 31
- SSRA v2.4s, v17.4s, 31
- SSRA v3.4s, v4.4s, 31
- SRSHL v0.4s, v0.4s, v7.4s
- SRSHL v1.4s, v1.4s, v7.4s
- SRSHL v2.4s, v2.4s, v7.4s
- SRSHL v3.4s, v3.4s, v7.4s
SQXTN v0.4h, v0.4s
SQXTN v2.4h, v2.4s
SQXTN2 v0.8h, v1.4s
@@ -210,7 +248,7 @@
LDP d10, d11, [sp], 48
RET
-END_FUNCTION xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal
+END_FUNCTION xnn_${DATATYPE}_gemm_minmax_${REQUANTIZATION.lower()}_ukernel_2x8c16__aarch64_neon_mlal_padal
#ifdef __ELF__
.section ".note.GNU-stack","",%progbits
diff --git a/src/qs8-gemm/gen/2x8c16-minmax-fp32-aarch64-neon-mlal-padal.S b/src/qs8-gemm/gen/2x8c16-minmax-fp32-aarch64-neon-mlal-padal.S
new file mode 100644
index 0000000..d0eb12f
--- /dev/null
+++ b/src/qs8-gemm/gen/2x8c16-minmax-fp32-aarch64-neon-mlal-padal.S
@@ -0,0 +1,219 @@
+// Auto-generated file. Do not edit!
+// Template: src/qs8-gemm/2x8c16-aarch64-neon-mlal-padal.S.in
+// Generator: tools/xngen
+//
+// Copyright 2021 Google LLC
+//
+// This source code is licensed under the BSD-style license found in the
+// LICENSE file in the root directory of this source tree.
+
+
+#include <xnnpack/assembly.h>
+
+# void xnn_qs8_gemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal(
+# size_t mr, x0
+# size_t nc, x1
+# size_t kc, x2 / x0
+# const int8_t* restrict a, x3
+# size_t a_stride, x4
+# const void* restrict w, x5
+# int8_t* restrict c, x6
+# size_t cm_stride, x7
+# size_t cn_stride, [sp] -> x10
+# const union xnn_qs8_conv_minmax_params params) [sp + 8] -> x11
+
+# d8-d15, x19-x30 need to be preserved if used. x18 is reserved by the OS.
+
+# Register usage
+# A0 x3 v0
+# A1 x4 v1
+# B x5 v4 v5 v6 v7
+# C0 x7 v16 v18 v20 v22 v24 v26 v28 v30
+# C1 x8 v17 v19 v21 v23 v25 v27 v29 v31
+# temp0 v2 v10 v12 v14
+# temp1 v3 v11 v13 v15
+# unused v8 v9
+
+BEGIN_FUNCTION xnn_qs8_gemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal
+
+ # Clamp A and C pointers
+ CMP x0, 2 // if mr < 2
+ STP d10, d11, [sp, -48]!
+ ADD x4, x3, x4 // a1 = a0 + a_stride
+ STP d12, d13, [sp, 16]
+ ADD x7, x6, x7 // c1 = c0 + cm_stride
+ STP d14, d15, [sp, 32]
+ CSEL x4, x3, x4, LO // a1 = a0
+ ADD x2, x2, 15 // kc = (kc + 15) & ~15
+ CSEL x7, x6, x7, LO // c1 = c0
+ BIC x2, x2, 15
+
+ .p2align 3
+0:
+ # Load initial bias from w into accumulators
+ MOV x0, x2 // k = kc
+ LDP s16, s18, [x5], 8
+ MOV v17.16b, v16.16b
+ MOV v19.16b, v18.16b
+ LDP s20, s22, [x5], 8
+ MOV v21.16b, v20.16b
+ MOV v23.16b, v22.16b
+ LDP s24, s26, [x5], 8
+ MOV v25.16b, v24.16b
+ MOV v27.16b, v26.16b
+ LDP s28, s30, [x5], 8
+ MOV v29.16b, v28.16b
+ LDP x10, x11, [sp, 48] // cn_stride, params
+ MOV v31.16b, v30.16b
+
+ # Main loop - 16 bytes of A
+ .p2align 3
+1:
+ LDR q0, [x3], 16
+ LDP q4, q5, [x5]
+ LDR q1, [x4], 16
+ LDP q6, q7, [x5, 32]
+ SMULL v2.8h, v4.8b, v0.8b
+ SMULL v3.8h, v4.8b, v1.8b
+ SMULL v10.8h, v5.8b, v0.8b
+ SMULL v11.8h, v5.8b, v1.8b
+ SMLAL2 v2.8h, v4.16b, v0.16b
+ SMLAL2 v3.8h, v4.16b, v1.16b
+ SMLAL2 v10.8h, v5.16b, v0.16b
+ SMLAL2 v11.8h, v5.16b, v1.16b
+ SMULL v12.8h, v6.8b, v0.8b
+ SADALP v16.4s, v2.8h
+ SMULL v13.8h, v6.8b, v1.8b
+ SADALP v17.4s, v3.8h
+ SMULL v14.8h, v7.8b, v0.8b
+ SADALP v18.4s, v10.8h
+ SMULL v15.8h, v7.8b, v1.8b
+ SADALP v19.4s, v11.8h
+ LDP q4, q5, [x5, 64]
+ SMLAL2 v12.8h, v6.16b, v0.16b
+ SMLAL2 v13.8h, v6.16b, v1.16b
+ SMLAL2 v14.8h, v7.16b, v0.16b
+ SMLAL2 v15.8h, v7.16b, v1.16b
+ SMULL v2.8h, v4.8b, v0.8b
+ SADALP v20.4s, v12.8h
+ SMULL v3.8h, v4.8b, v1.8b
+ SADALP v21.4s, v13.8h
+ SMULL v10.8h, v5.8b, v0.8b
+ SADALP v22.4s, v14.8h
+ SMULL v11.8h, v5.8b, v1.8b
+ SADALP v23.4s, v15.8h
+ LDP q6, q7, [x5, 96]
+
+ SMLAL2 v2.8h, v4.16b, v0.16b
+ SMLAL2 v3.8h, v4.16b, v1.16b
+ SMLAL2 v10.8h, v5.16b, v0.16b
+ SMLAL2 v11.8h, v5.16b, v1.16b
+ ADD x5, x5, 128
+ SMULL v12.8h, v6.8b, v0.8b
+ SADALP v24.4s, v2.8h
+ SMULL v13.8h, v6.8b, v1.8b
+ SADALP v25.4s, v3.8h
+ SMULL v14.8h, v7.8b, v0.8b
+ SADALP v26.4s, v10.8h
+ SMULL v15.8h, v7.8b, v1.8b
+ SADALP v27.4s, v11.8h
+ SUBS x0, x0, 16
+ SMLAL2 v12.8h, v6.16b, v0.16b
+ SMLAL2 v13.8h, v6.16b, v1.16b
+ SMLAL2 v14.8h, v7.16b, v0.16b
+ SMLAL2 v15.8h, v7.16b, v1.16b
+ SADALP v28.4s, v12.8h
+ SADALP v29.4s, v13.8h
+ SADALP v30.4s, v14.8h
+ SADALP v31.4s, v15.8h
+ B.HI 1b
+
+ # Add columns
+ ADDP v16.4s, v16.4s, v18.4s
+ ADDP v20.4s, v20.4s, v22.4s
+ ADDP v24.4s, v24.4s, v26.4s
+ ADDP v28.4s, v28.4s, v30.4s
+ ADDP v17.4s, v17.4s, v19.4s
+ ADDP v21.4s, v21.4s, v23.4s
+ ADDP v25.4s, v25.4s, v27.4s
+ ADDP v29.4s, v29.4s, v31.4s
+ ADDP v0.4s, v16.4s, v20.4s
+ ADDP v1.4s, v24.4s, v28.4s
+ ADDP v2.4s, v17.4s, v21.4s
+ ADDP v3.4s, v25.4s, v29.4s
+
+ # Apply params - scale, bias and clamp
+ SCVTF v0.4s, v0.4s
+ LD1R {v4.4s}, [x11], 4
+ SCVTF v1.4s, v1.4s
+ SCVTF v2.4s, v2.4s
+ SCVTF v3.4s, v3.4s
+ FMUL v0.4s, v0.4s, v4.4s
+ FMUL v1.4s, v1.4s, v4.4s
+ FMUL v2.4s, v2.4s, v4.4s
+ FMUL v3.4s, v3.4s, v4.4s
+
+ FCVTNS v0.4s, v0.4s
+ FCVTNS v1.4s, v1.4s
+ FCVTNS v2.4s, v2.4s
+ FCVTNS v3.4s, v3.4s
+
+ LD1R {v5.8h}, [x11], 2
+ SQXTN v0.4h, v0.4s
+ SQXTN v2.4h, v2.4s
+ SQXTN2 v0.8h, v1.4s
+ SQXTN2 v2.8h, v3.4s
+ SUBS x1, x1, 8
+ SQADD v0.8h, v0.8h, v5.8h
+ SQADD v1.8h, v2.8h, v5.8h
+ SQXTN v0.8b, v0.8h
+ SQXTN2 v0.16b, v1.8h
+ LD1R {v1.16b}, [x11], 1
+ LD1R {v2.16b}, [x11]
+ SMAX v0.16b, v0.16b, v1.16b
+ SMIN v0.16b, v0.16b, v2.16b
+ B.LO 2f
+
+ # Store full 2 x 8
+ ST1 {v0.8b}, [x6], x10
+ SUB x3, x3, x2 // a0 -= kc
+ ST1 {v0.d}[1], [x7], x10
+ SUB x4, x4, x2 // a1 -= kc
+ B.HI 0b
+
+ # Restore d10-d15 from stack
+ LDP d14, d15, [sp, 32]
+ LDP d12, d13, [sp, 16]
+ LDP d10, d11, [sp], 48
+ RET
+
+ # Store odd width
+ .p2align 3
+2:
+ TBZ x1, 2, 3f
+ STR s0, [x6], 4
+ ST1 {v0.s}[2], [x7], 4
+ EXT v0.16b, v0.16b, v0.16b, 4
+
+3:
+ TBZ x1, 1, 4f
+ ST1 {v0.h}[0], [x6], 2
+ ST1 {v0.h}[4], [x7], 2
+ EXT v0.16b, v0.16b, v0.16b, 2
+4:
+ TBZ x1, 0, 5f
+ ST1 {v0.b}[0], [x6]
+ ST1 {v0.b}[8], [x7]
+5:
+ # Restore d10-d15 from stack
+ LDP d14, d15, [sp, 32]
+ LDP d12, d13, [sp, 16]
+ LDP d10, d11, [sp], 48
+ RET
+
+END_FUNCTION xnn_qs8_gemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal
+
+#ifdef __ELF__
+.section ".note.GNU-stack","",%progbits
+#endif
+
diff --git a/src/qs8-gemm/gen/2x8c16-minmax-gemmlowp-aarch64-neon-mlal-padal.S b/src/qs8-gemm/gen/2x8c16-minmax-gemmlowp-aarch64-neon-mlal-padal.S
index 670bc2f..b77500e 100644
--- a/src/qs8-gemm/gen/2x8c16-minmax-gemmlowp-aarch64-neon-mlal-padal.S
+++ b/src/qs8-gemm/gen/2x8c16-minmax-gemmlowp-aarch64-neon-mlal-padal.S
@@ -7,6 +7,7 @@
// This source code is licensed under the BSD-style license found in the
// LICENSE file in the root directory of this source tree.
+
#include <xnnpack/assembly.h>
# void xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal(
@@ -149,7 +150,6 @@
SQRDMULH v2.4s, v2.4s, v4.4s
SQRDMULH v3.4s, v3.4s, v4.4s
CMEQ v4.4s, v7.4s, 0
- LD1R {v5.8h}, [x11], 2
BIC v6.16b, v0.16b, v4.16b
BIC v16.16b, v1.16b, v4.16b
BIC v17.16b, v2.16b, v4.16b
@@ -162,6 +162,8 @@
SRSHL v1.4s, v1.4s, v7.4s
SRSHL v2.4s, v2.4s, v7.4s
SRSHL v3.4s, v3.4s, v7.4s
+
+ LD1R {v5.8h}, [x11], 2
SQXTN v0.4h, v0.4s
SQXTN v2.4h, v2.4s
SQXTN2 v0.8h, v1.4s
diff --git a/src/qs8-igemm/2x8c16-aarch64-neon-mlal-padal.S.in b/src/qs8-igemm/2x8c16-aarch64-neon-mlal-padal.S.in
index ad9ea6f..bb30160 100644
--- a/src/qs8-igemm/2x8c16-aarch64-neon-mlal-padal.S.in
+++ b/src/qs8-igemm/2x8c16-aarch64-neon-mlal-padal.S.in
@@ -3,9 +3,14 @@
// This source code is licensed under the BSD-style license found in the
// LICENSE file in the root directory of this source tree.
+$assert REQUANTIZATION in ["GEMMLOWP", "FP32"]
+$assert not CHANNELWISE or REQUANTIZATION == "FP32"
+
#include <xnnpack/assembly.h>
-# void xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal(
+$DATATYPE = "qc8" if CHANNELWISE else "qs8"
+$CONV_PARAMS = "xnn_qs8_minmax_params" if CHANNELWISE else "xnn_qs8_conv_minmax_params"
+# void xnn_${DATATYPE}_igemm_minmax_${REQUANTIZATION.lower()}_ukernel_2x8c16__aarch64_neon_mlal_padal(
# size_t mr, x0
# size_t nc, x1
# size_t kc, x2 / x0
@@ -17,7 +22,7 @@
# size_t cn_stride, [sp] -> x10
# size_t a_offset, [sp + 8] -> x8
# const float* zero, [sp + 16] -> x12
-# const xnn_qs8_conv_minmax_params params [sp + 24] -> x11
+# const union ${CONV_PARAMS} params [sp + 24] -> x11
# d8-d15, x19-x30 need to be preserved if used. x18 is reserved by the OS.
@@ -31,7 +36,7 @@
# temp1 v3 v11 v13 v15
# unused v8 v9
-BEGIN_FUNCTION xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal
+BEGIN_FUNCTION xnn_${DATATYPE}_igemm_minmax_${REQUANTIZATION.lower()}_ukernel_2x8c16__aarch64_neon_mlal_padal
# Clamp C pointers
LDP x10, x8, [sp] // Load cn_stride, a_offset
@@ -145,10 +150,12 @@
# Add columns
ADDP v16.4s, v16.4s, v18.4s
ADDP v20.4s, v20.4s, v22.4s
- LD1R {v4.4s}, [x11], 4
+ $if REQUANTIZATION == "GEMMLOWP":
+ LD1R {v4.4s}, [x11], 4
ADDP v24.4s, v24.4s, v26.4s
ADDP v28.4s, v28.4s, v30.4s
- LD1R {v7.4s}, [x11], 4
+ $if REQUANTIZATION == "GEMMLOWP":
+ LD1R {v7.4s}, [x11], 4
ADDP v17.4s, v17.4s, v19.4s
ADDP v21.4s, v21.4s, v23.4s
ADDP v25.4s, v25.4s, v27.4s
@@ -158,25 +165,56 @@
ADDP v2.4s, v17.4s, v21.4s
ADDP v3.4s, v25.4s, v29.4s
- # Apply params - scale, shift, bias and clamp
- SQRDMULH v0.4s, v0.4s, v4.4s
- SQRDMULH v1.4s, v1.4s, v4.4s
- SQRDMULH v2.4s, v2.4s, v4.4s
- SQRDMULH v3.4s, v3.4s, v4.4s
- CMEQ v4.4s, v7.4s, 0
+ $if REQUANTIZATION == "GEMMLOWP":
+ # Apply params - scale, shift, bias and clamp
+ SQRDMULH v0.4s, v0.4s, v4.4s
+ SQRDMULH v1.4s, v1.4s, v4.4s
+ SQRDMULH v2.4s, v2.4s, v4.4s
+ SQRDMULH v3.4s, v3.4s, v4.4s
+ CMEQ v4.4s, v7.4s, 0
+ BIC v6.16b, v0.16b, v4.16b
+ BIC v16.16b, v1.16b, v4.16b
+ BIC v17.16b, v2.16b, v4.16b
+ BIC v4.16b, v3.16b, v4.16b
+ SSRA v0.4s, v6.4s, 31
+ SSRA v1.4s, v16.4s, 31
+ SSRA v2.4s, v17.4s, 31
+ SSRA v3.4s, v4.4s, 31
+ SRSHL v0.4s, v0.4s, v7.4s
+ SRSHL v1.4s, v1.4s, v7.4s
+ SRSHL v2.4s, v2.4s, v7.4s
+ SRSHL v3.4s, v3.4s, v7.4s
+ $elif REQUANTIZATION == "FP32":
+ $if not CHANNELWISE:
+ # Apply params - scale, bias and clamp
+ SCVTF v0.4s, v0.4s
+ LD1R {v4.4s}, [x11], 4
+ SCVTF v1.4s, v1.4s
+ SCVTF v2.4s, v2.4s
+ SCVTF v3.4s, v3.4s
+ FMUL v0.4s, v0.4s, v4.4s
+ FMUL v1.4s, v1.4s, v4.4s
+ FMUL v2.4s, v2.4s, v4.4s
+ FMUL v3.4s, v3.4s, v4.4s
+ $else:
+ # Load per channel scale values from weights
+ SCVTF v0.4s, v0.4s
+ LDR q4, [x5], 16
+ SCVTF v1.4s, v1.4s
+ LDR q5, [x5], 16
+ SCVTF v2.4s, v2.4s
+ SCVTF v3.4s, v3.4s
+ FMUL v0.4s, v0.4s, v4.4s
+ FMUL v1.4s, v1.4s, v5.4s
+ FMUL v2.4s, v2.4s, v4.4s
+ FMUL v3.4s, v3.4s, v5.4s
+
+ FCVTNS v0.4s, v0.4s
+ FCVTNS v1.4s, v1.4s
+ FCVTNS v2.4s, v2.4s
+ FCVTNS v3.4s, v3.4s
+
LD1R {v5.8h}, [x11], 2
- BIC v6.16b, v0.16b, v4.16b
- BIC v16.16b, v1.16b, v4.16b
- BIC v17.16b, v2.16b, v4.16b
- BIC v4.16b, v3.16b, v4.16b
- SSRA v0.4s, v6.4s, 31
- SSRA v1.4s, v16.4s, 31
- SSRA v2.4s, v17.4s, 31
- SSRA v3.4s, v4.4s, 31
- SRSHL v0.4s, v0.4s, v7.4s
- SRSHL v1.4s, v1.4s, v7.4s
- SRSHL v2.4s, v2.4s, v7.4s
- SRSHL v3.4s, v3.4s, v7.4s
SQXTN v0.4h, v0.4s
SQXTN v2.4h, v2.4s
SQXTN2 v0.8h, v1.4s
@@ -189,7 +227,13 @@
LD1R {v1.16b}, [x11], 1
LD1R {v2.16b}, [x11]
SMAX v0.16b, v0.16b, v1.16b
- SUB x11, x11, 11 // rewind params pointer
+ $if REQUANTIZATION == "GEMMLOWP":
+ SUB x11, x11, 11 // rewind params pointer
+ $elif REQUANTIZATION == "FP32":
+ $if CHANNELWISE:
+ SUB x11, x11, 3 // rewind params pointer
+ $else:
+ SUB x11, x11, 7 // rewind params pointer
SMIN v0.16b, v0.16b, v2.16b
B.LO 3f
@@ -231,7 +275,7 @@
LDP d10, d11, [sp], 48
RET
-END_FUNCTION xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal
+END_FUNCTION xnn_${DATATYPE}_igemm_minmax_${REQUANTIZATION.lower()}_ukernel_2x8c16__aarch64_neon_mlal_padal
#ifdef __ELF__
.section ".note.GNU-stack","",%progbits
diff --git a/src/qs8-igemm/gen/2x8c16-minmax-fp32-aarch64-neon-mlal-padal.S b/src/qs8-igemm/gen/2x8c16-minmax-fp32-aarch64-neon-mlal-padal.S
new file mode 100644
index 0000000..33057f9
--- /dev/null
+++ b/src/qs8-igemm/gen/2x8c16-minmax-fp32-aarch64-neon-mlal-padal.S
@@ -0,0 +1,240 @@
+// Auto-generated file. Do not edit!
+// Template: src/qs8-igemm/2x8c16-aarch64-neon-mlal-padal.S.in
+// Generator: tools/xngen
+//
+// Copyright 2021 Google LLC
+//
+// This source code is licensed under the BSD-style license found in the
+// LICENSE file in the root directory of this source tree.
+
+
+#include <xnnpack/assembly.h>
+
+# void xnn_qs8_igemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal(
+# size_t mr, x0
+# size_t nc, x1
+# size_t kc, x2 / x0
+# size_t ks, x3 / x9
+# const int8_t**restrict a, x4
+# const int8_t* restrict w, x5
+# int8_t* restrict c, x6
+# size_t cm_stride, x7
+# size_t cn_stride, [sp] -> x10
+# size_t a_offset, [sp + 8] -> x8
+# const float* zero, [sp + 16] -> x12
+# const union xnn_qs8_conv_minmax_params params [sp + 24] -> x11
+
+# d8-d15, x19-x30 need to be preserved if used. x18 is reserved by the OS.
+
+# Register usage
+# A0 x13 v0
+# A1 x15 v1
+# B x5 v4 v5 v6 v7
+# C0 x6 v16 v18 v20 v22 v24 v26 v28 v30
+# C1 x7 v17 v19 v21 v23 v25 v27 v29 v31
+# temp0 v2 v10 v12 v14
+# temp1 v3 v11 v13 v15
+# unused v8 v9
+
+BEGIN_FUNCTION xnn_qs8_igemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal
+
+ # Clamp C pointers
+ LDP x10, x8, [sp] // Load cn_stride, a_offset
+ CMP x0, 2 // if mr < 2
+ LDP x12, x11, [sp, 16] // Load zero, params pointer
+ ADD x7, x6, x7 // c1 = c0 + cm_stride
+ STP d10, d11, [sp, -48]!
+ ADD x2, x2, 15 // kc = (kc + 15) & ~15
+ STP d12, d13, [sp, 16]
+ CSEL x7, x6, x7, LO // c1 = c0
+ STP d14, d15, [sp, 32]
+ BIC x2, x2, 15
+
+ .p2align 3
+0:
+ # Load initial bias from w into accumulators
+ LDP s16, s18, [x5], 8
+ MOV v17.16b, v16.16b
+ MOV v19.16b, v18.16b
+ LDP s20, s22, [x5], 8
+ MOV v21.16b, v20.16b
+ MOV v23.16b, v22.16b
+ LDP s24, s26, [x5], 8
+ MOV v25.16b, v24.16b
+ MOV v27.16b, v26.16b
+ LDP s28, s30, [x5], 8
+ MOV v29.16b, v28.16b
+ MOV v31.16b, v30.16b
+ MOV x9, x3 // p = ks
+
+ .p2align 3
+1:
+ # Load next 2 A pointers
+ LDP x13, x15, [x4], 16
+
+ CMP x13, x12 // if a0 == zero
+ ADD x13, x13, x8 // a0 += a_offset
+ CSEL x13, x12, x13, EQ // a0 = zero, else += a0 + a_offset
+ CMP x15, x12 // if a1 == zero
+ ADD x15, x15, x8 // a1 += a_offset
+ CSEL x15, x12, x15, EQ // a1 = zero, else += a1 + a_offset
+
+ MOV x0, x2 // k = kc
+
+ # Main loop - 16 bytes of A
+ .p2align 3
+2:
+ LDR q0, [x13], 16
+ LDP q4, q5, [x5]
+ LDR q1, [x15], 16
+ LDP q6, q7, [x5, 32]
+ SMULL v2.8h, v4.8b, v0.8b
+ SMULL v3.8h, v4.8b, v1.8b
+ SMULL v10.8h, v5.8b, v0.8b
+ SMULL v11.8h, v5.8b, v1.8b
+ SMLAL2 v2.8h, v4.16b, v0.16b
+ SMLAL2 v3.8h, v4.16b, v1.16b
+ SMLAL2 v10.8h, v5.16b, v0.16b
+ SMLAL2 v11.8h, v5.16b, v1.16b
+ SMULL v12.8h, v6.8b, v0.8b
+ SADALP v16.4s, v2.8h
+ SMULL v13.8h, v6.8b, v1.8b
+ SADALP v17.4s, v3.8h
+ SMULL v14.8h, v7.8b, v0.8b
+ SADALP v18.4s, v10.8h
+ SMULL v15.8h, v7.8b, v1.8b
+ SADALP v19.4s, v11.8h
+ LDP q4, q5, [x5, 64]
+ SMLAL2 v12.8h, v6.16b, v0.16b
+ SMLAL2 v13.8h, v6.16b, v1.16b
+ SMLAL2 v14.8h, v7.16b, v0.16b
+ SMLAL2 v15.8h, v7.16b, v1.16b
+ SMULL v2.8h, v4.8b, v0.8b
+ SADALP v20.4s, v12.8h
+ SMULL v3.8h, v4.8b, v1.8b
+ SADALP v21.4s, v13.8h
+ SMULL v10.8h, v5.8b, v0.8b
+ SADALP v22.4s, v14.8h
+ SMULL v11.8h, v5.8b, v1.8b
+ SADALP v23.4s, v15.8h
+ LDP q6, q7, [x5, 96]
+
+ SMLAL2 v2.8h, v4.16b, v0.16b
+ SMLAL2 v3.8h, v4.16b, v1.16b
+ SMLAL2 v10.8h, v5.16b, v0.16b
+ SMLAL2 v11.8h, v5.16b, v1.16b
+ ADD x5, x5, 128
+ SMULL v12.8h, v6.8b, v0.8b
+ SADALP v24.4s, v2.8h
+ SMULL v13.8h, v6.8b, v1.8b
+ SADALP v25.4s, v3.8h
+ SMULL v14.8h, v7.8b, v0.8b
+ SADALP v26.4s, v10.8h
+ SMULL v15.8h, v7.8b, v1.8b
+ SADALP v27.4s, v11.8h
+ SUBS x0, x0, 16
+ SMLAL2 v12.8h, v6.16b, v0.16b
+ SMLAL2 v13.8h, v6.16b, v1.16b
+ SMLAL2 v14.8h, v7.16b, v0.16b
+ SMLAL2 v15.8h, v7.16b, v1.16b
+ SADALP v28.4s, v12.8h
+ SADALP v29.4s, v13.8h
+ SADALP v30.4s, v14.8h
+ SADALP v31.4s, v15.8h
+ B.HI 2b
+
+ # ks loop
+ SUBS x9, x9, 16 // ks -= MR * sizeof(int8_t*)
+ B.HI 1b
+
+ # Add columns
+ ADDP v16.4s, v16.4s, v18.4s
+ ADDP v20.4s, v20.4s, v22.4s
+ ADDP v24.4s, v24.4s, v26.4s
+ ADDP v28.4s, v28.4s, v30.4s
+ ADDP v17.4s, v17.4s, v19.4s
+ ADDP v21.4s, v21.4s, v23.4s
+ ADDP v25.4s, v25.4s, v27.4s
+ ADDP v29.4s, v29.4s, v31.4s
+ ADDP v0.4s, v16.4s, v20.4s
+ ADDP v1.4s, v24.4s, v28.4s
+ ADDP v2.4s, v17.4s, v21.4s
+ ADDP v3.4s, v25.4s, v29.4s
+
+ # Apply params - scale, bias and clamp
+ SCVTF v0.4s, v0.4s
+ LD1R {v4.4s}, [x11], 4
+ SCVTF v1.4s, v1.4s
+ SCVTF v2.4s, v2.4s
+ SCVTF v3.4s, v3.4s
+ FMUL v0.4s, v0.4s, v4.4s
+ FMUL v1.4s, v1.4s, v4.4s
+ FMUL v2.4s, v2.4s, v4.4s
+ FMUL v3.4s, v3.4s, v4.4s
+
+ FCVTNS v0.4s, v0.4s
+ FCVTNS v1.4s, v1.4s
+ FCVTNS v2.4s, v2.4s
+ FCVTNS v3.4s, v3.4s
+
+ LD1R {v5.8h}, [x11], 2
+ SQXTN v0.4h, v0.4s
+ SQXTN v2.4h, v2.4s
+ SQXTN2 v0.8h, v1.4s
+ SQXTN2 v2.8h, v3.4s
+ SUBS x1, x1, 8
+ SQADD v0.8h, v0.8h, v5.8h
+ SQADD v1.8h, v2.8h, v5.8h
+ SQXTN v0.8b, v0.8h
+ SQXTN2 v0.16b, v1.8h
+ LD1R {v1.16b}, [x11], 1
+ LD1R {v2.16b}, [x11]
+ SMAX v0.16b, v0.16b, v1.16b
+ SUB x11, x11, 7 // rewind params pointer
+ SMIN v0.16b, v0.16b, v2.16b
+ B.LO 3f
+
+ # Store full 2 x 8
+ ST1 {v0.d}[1], [x7], x10
+ SUB x4, x4, x3 // a -= ks
+ ST1 {v0.8b}, [x6], x10
+
+ # nc loop
+ B.HI 0b
+
+ # Restore d10-d15 from stack
+ LDP d14, d15, [sp, 32]
+ LDP d12, d13, [sp, 16]
+ LDP d10, d11, [sp], 48
+ RET
+
+ # Store odd width
+ .p2align 3
+3:
+ TBZ x1, 2, 4f
+ ST1 {v0.s}[2], [x7], 4
+ STR s0, [x6], 4
+ EXT v0.16b, v0.16b, v0.16b, 4
+
+4:
+ TBZ x1, 1, 5f
+ ST1 {v0.h}[4], [x7], 2
+ ST1 {v0.h}[0], [x6], 2
+ EXT v0.16b, v0.16b, v0.16b, 2
+5:
+ TBZ x1, 0, 6f
+ ST1 {v0.b}[8], [x7]
+ ST1 {v0.b}[0], [x6]
+6:
+ # Restore d10-d15 from stack
+ LDP d14, d15, [sp, 32]
+ LDP d12, d13, [sp, 16]
+ LDP d10, d11, [sp], 48
+ RET
+
+END_FUNCTION xnn_qs8_igemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal
+
+#ifdef __ELF__
+.section ".note.GNU-stack","",%progbits
+#endif
+
diff --git a/src/qs8-igemm/gen/2x8c16-minmax-gemmlowp-aarch64-neon-mlal-padal.S b/src/qs8-igemm/gen/2x8c16-minmax-gemmlowp-aarch64-neon-mlal-padal.S
index c5455d4..8fa98c9 100644
--- a/src/qs8-igemm/gen/2x8c16-minmax-gemmlowp-aarch64-neon-mlal-padal.S
+++ b/src/qs8-igemm/gen/2x8c16-minmax-gemmlowp-aarch64-neon-mlal-padal.S
@@ -7,6 +7,7 @@
// This source code is licensed under the BSD-style license found in the
// LICENSE file in the root directory of this source tree.
+
#include <xnnpack/assembly.h>
# void xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal(
@@ -21,7 +22,7 @@
# size_t cn_stride, [sp] -> x10
# size_t a_offset, [sp + 8] -> x8
# const float* zero, [sp + 16] -> x12
-# const xnn_qs8_conv_minmax_params params [sp + 24] -> x11
+# const union xnn_qs8_conv_minmax_params params [sp + 24] -> x11
# d8-d15, x19-x30 need to be preserved if used. x18 is reserved by the OS.
@@ -168,7 +169,6 @@
SQRDMULH v2.4s, v2.4s, v4.4s
SQRDMULH v3.4s, v3.4s, v4.4s
CMEQ v4.4s, v7.4s, 0
- LD1R {v5.8h}, [x11], 2
BIC v6.16b, v0.16b, v4.16b
BIC v16.16b, v1.16b, v4.16b
BIC v17.16b, v2.16b, v4.16b
@@ -181,6 +181,8 @@
SRSHL v1.4s, v1.4s, v7.4s
SRSHL v2.4s, v2.4s, v7.4s
SRSHL v3.4s, v3.4s, v7.4s
+
+ LD1R {v5.8h}, [x11], 2
SQXTN v0.4h, v0.4s
SQXTN v2.4h, v2.4s
SQXTN2 v0.8h, v1.4s
diff --git a/src/xnnpack/gemm.h b/src/xnnpack/gemm.h
index 7c743e7..a8c7f59 100644
--- a/src/xnnpack/gemm.h
+++ b/src/xnnpack/gemm.h
@@ -651,6 +651,8 @@
DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53)
DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal)
+DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal)
+
DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_gemmlowp_ukernel_4x16__aarch64_neon_mlal_lane_cortex_a53)
DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_gemmlowp_ukernel_4x16__aarch64_neon_mlal_lane_prfm_cortex_a53)
@@ -1004,6 +1006,8 @@
DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_6x16c4__neondot)
DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_8x16c4__neondot)
+DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal)
+
DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_4x16__aarch64_neon_mlal_lane_cortex_a53)
DECLARE_QC8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_gemm_minmax_fp32_ukernel_4x16__aarch64_neon_mlal_lane_prfm_cortex_a53)
diff --git a/src/xnnpack/igemm.h b/src/xnnpack/igemm.h
index aeafe6d..9989252 100644
--- a/src/xnnpack/igemm.h
+++ b/src/xnnpack/igemm.h
@@ -461,6 +461,8 @@
DECLARE_QS8_IGEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53)
DECLARE_QS8_IGEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal)
+DECLARE_QS8_IGEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_igemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal)
+
DECLARE_QS8_IGEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_igemm_minmax_gemmlowp_ukernel_4x16__aarch64_neon_mlal_lane_cortex_a53)
DECLARE_QS8_IGEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_igemm_minmax_gemmlowp_ukernel_4x16__aarch64_neon_mlal_lane_prfm_cortex_a53)
@@ -753,6 +755,8 @@
DECLARE_QC8_IGEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_igemm_minmax_fp32_ukernel_6x16c4__neondot)
DECLARE_QC8_IGEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_igemm_minmax_fp32_ukernel_8x16c4__neondot)
+DECLARE_QC8_IGEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_igemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal)
+
DECLARE_QC8_IGEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_igemm_minmax_fp32_ukernel_4x16__aarch64_neon_mlal_lane_cortex_a53)
DECLARE_QC8_IGEMM_MINMAX_UKERNEL_FUNCTION(xnn_qc8_igemm_minmax_fp32_ukernel_4x16__aarch64_neon_mlal_lane_prfm_cortex_a53)
diff --git a/test/qc8-gemm-minmax-fp32.cc b/test/qc8-gemm-minmax-fp32.cc
index 42e82bb..e27b48e 100644
--- a/test/qc8-gemm-minmax-fp32.cc
+++ b/test/qc8-gemm-minmax-fp32.cc
@@ -23,6 +23,462 @@
#if XNN_ARCH_ARM64
+ TEST(QC8_GEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, k_eq_16) {
+ TEST_REQUIRES_ARM_NEON;
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(16)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_minmax_neon_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+
+ TEST(QC8_GEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, strided_cn) {
+ TEST_REQUIRES_ARM_NEON;
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(16)
+ .cn_stride(11)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_minmax_neon_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+
+ TEST(QC8_GEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, k_eq_16_strided_a) {
+ TEST_REQUIRES_ARM_NEON;
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(16)
+ .a_stride(19)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_minmax_neon_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+
+ TEST(QC8_GEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, k_eq_16_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(16)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_minmax_neon_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+
+ TEST(QC8_GEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, k_eq_16_subtile_m) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(m)
+ .n(8)
+ .k(16)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_minmax_neon_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+
+ TEST(QC8_GEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, k_eq_16_subtile_n) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(16)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_minmax_neon_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+
+ TEST(QC8_GEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, k_lt_16) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 1; k < 16; k++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_minmax_neon_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+
+ TEST(QC8_GEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, k_lt_16_strided_a) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 1; k < 16; k++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .a_stride(19)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_minmax_neon_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+
+ TEST(QC8_GEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, k_lt_16_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 1; k < 16; k++) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_minmax_neon_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+ }
+
+ TEST(QC8_GEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, k_gt_16) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 17; k < 32; k++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_minmax_neon_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+
+ TEST(QC8_GEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, k_gt_16_strided_a) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 17; k < 32; k++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .a_stride(37)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_minmax_neon_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+
+ TEST(QC8_GEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, k_gt_16_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 17; k < 32; k++) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_minmax_neon_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+ }
+
+ TEST(QC8_GEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, k_div_16) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 32; k <= 160; k += 16) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_minmax_neon_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+
+ TEST(QC8_GEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, k_div_16_strided_a) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 32; k <= 160; k += 16) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .a_stride(163)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_minmax_neon_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+
+ TEST(QC8_GEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, k_div_16_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 32; k <= 160; k += 16) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_minmax_neon_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+ }
+
+ TEST(QC8_GEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, n_gt_8) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 9; n < 16; n++) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_minmax_neon_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+
+ TEST(QC8_GEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, n_gt_8_strided_cn) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 9; n < 16; n++) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .cn_stride(11)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_minmax_neon_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+
+ TEST(QC8_GEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, n_gt_8_strided_a) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 9; n < 16; n++) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(k)
+ .a_stride(83)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_minmax_neon_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+
+ TEST(QC8_GEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, n_gt_8_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 9; n < 16; n++) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_minmax_neon_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+ }
+
+ TEST(QC8_GEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, n_div_8) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 16; n <= 24; n += 8) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_minmax_neon_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+
+ TEST(QC8_GEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, n_div_8_strided_cn) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 16; n <= 24; n += 8) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(k)
+ .cn_stride(11)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_minmax_neon_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+
+ TEST(QC8_GEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, n_div_8_strided_a) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 16; n <= 24; n += 8) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(k)
+ .a_stride(83)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_minmax_neon_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+
+ TEST(QC8_GEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, n_div_8_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 16; n <= 24; n += 8) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_minmax_neon_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+ }
+
+ TEST(QC8_GEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, strided_cm_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 1; k <= 80; k += 17) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(11)
+ .iterations(1)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_minmax_neon_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+ }
+
+ TEST(QC8_GEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, qmin) {
+ TEST_REQUIRES_ARM_NEON;
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(16)
+ .qmin(128)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_minmax_neon_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+
+ TEST(QC8_GEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, qmax) {
+ TEST_REQUIRES_ARM_NEON;
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(16)
+ .qmax(128)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_minmax_neon_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+
+ TEST(QC8_GEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, strided_cm) {
+ TEST_REQUIRES_ARM_NEON;
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(16)
+ .cm_stride(11)
+ .Test(xnn_qc8_gemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_minmax_neon_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+#endif // XNN_ARCH_ARM64
+
+
+#if XNN_ARCH_ARM64
TEST(QC8_GEMM_MINMAX_FP32_4X16__AARCH64_NEON_MLAL_LANE_PRFM_CORTEX_A53, k_eq_8) {
TEST_REQUIRES_ARM_NEON;
GemmMicrokernelTester()
diff --git a/test/qc8-gemm-minmax-fp32.yaml b/test/qc8-gemm-minmax-fp32.yaml
index fc4a252..0440f76 100644
--- a/test/qc8-gemm-minmax-fp32.yaml
+++ b/test/qc8-gemm-minmax-fp32.yaml
@@ -3,6 +3,9 @@
# This source code is licensed under the BSD-style license found in the
# LICENSE file in the root directory of this source tree.
+- name: xnn_qc8_gemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal
+ init: xnn_init_qs8_minmax_neon_params
+ k-block: 16
- name: xnn_qc8_gemm_minmax_fp32_ukernel_4x16__aarch64_neon_mlal_lane_prfm_cortex_a53
init: xnn_init_qs8_minmax_neon_params
k-block: 8
diff --git a/test/qc8-igemm-minmax-fp32.cc b/test/qc8-igemm-minmax-fp32.cc
index 8f517b4..5237802 100644
--- a/test/qc8-igemm-minmax-fp32.cc
+++ b/test/qc8-igemm-minmax-fp32.cc
@@ -23,6 +23,474 @@
#if XNN_ARCH_ARM64
+ TEST(QC8_IGEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, k_eq_16) {
+ TEST_REQUIRES_ARM_NEON;
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(16)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_minmax_neon_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+
+ TEST(QC8_IGEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, strided_cn) {
+ TEST_REQUIRES_ARM_NEON;
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(16)
+ .cn_stride(11)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_minmax_neon_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+
+ TEST(QC8_IGEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, k_eq_16_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(16)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_minmax_neon_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+
+ TEST(QC8_IGEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, k_eq_16_subtile_m) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(m)
+ .n(8)
+ .k(16)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_minmax_neon_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+
+ TEST(QC8_IGEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, k_eq_16_subtile_n) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(16)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_minmax_neon_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+
+ TEST(QC8_IGEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, k_lt_16) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 1; k < 16; k++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_minmax_neon_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+
+ TEST(QC8_IGEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, k_lt_16_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 1; k < 16; k++) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_minmax_neon_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+ }
+
+ TEST(QC8_IGEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, k_gt_16) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 17; k < 32; k++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_minmax_neon_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+
+ TEST(QC8_IGEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, k_gt_16_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 17; k < 32; k++) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_minmax_neon_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+ }
+
+ TEST(QC8_IGEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, k_div_16) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 32; k <= 160; k += 16) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_minmax_neon_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+
+ TEST(QC8_IGEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, k_div_16_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 32; k <= 160; k += 16) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_minmax_neon_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+ }
+
+ TEST(QC8_IGEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, n_gt_8) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 9; n < 16; n++) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_minmax_neon_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+
+ TEST(QC8_IGEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, n_gt_8_strided_cn) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 9; n < 16; n++) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .cn_stride(11)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_minmax_neon_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+
+ TEST(QC8_IGEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, n_gt_8_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 9; n < 16; n++) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_minmax_neon_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+ }
+
+ TEST(QC8_IGEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, n_div_8) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 16; n <= 24; n += 8) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_minmax_neon_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+
+ TEST(QC8_IGEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, n_div_8_strided_cn) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 16; n <= 24; n += 8) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(k)
+ .cn_stride(11)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_minmax_neon_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+
+ TEST(QC8_IGEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, n_div_8_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 16; n <= 24; n += 8) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_minmax_neon_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+ }
+
+ TEST(QC8_IGEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, small_kernel) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_minmax_neon_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+
+ TEST(QC8_IGEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, small_kernel_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 1; k <= 80; k += 17) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .ks(3)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_minmax_neon_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+ }
+
+ TEST(QC8_IGEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, n_gt_8_small_kernel) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 9; n < 16; n++) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_minmax_neon_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+
+ TEST(QC8_IGEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, n_div_8_small_kernel) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 16; n <= 24; n += 8) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_minmax_neon_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+
+ TEST(QC8_IGEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, strided_cm_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 1; k <= 80; k += 17) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(11)
+ .iterations(1)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_minmax_neon_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+ }
+
+ TEST(QC8_IGEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, a_offset) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .ks(3)
+ .a_offset(163)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_minmax_neon_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+
+ TEST(QC8_IGEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, zero) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t mz = 0; mz < 2; mz++) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .ks(3)
+ .a_offset(163)
+ .zero_index(mz)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_minmax_neon_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+
+ TEST(QC8_IGEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, qmin) {
+ TEST_REQUIRES_ARM_NEON;
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(16)
+ .qmin(128)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_minmax_neon_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+
+ TEST(QC8_IGEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, qmax) {
+ TEST_REQUIRES_ARM_NEON;
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(16)
+ .qmax(128)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_minmax_neon_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+
+ TEST(QC8_IGEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, strided_cm) {
+ TEST_REQUIRES_ARM_NEON;
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(16)
+ .cm_stride(11)
+ .Test(xnn_qc8_igemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_minmax_neon_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+#endif // XNN_ARCH_ARM64
+
+
+#if XNN_ARCH_ARM64
TEST(QC8_IGEMM_MINMAX_FP32_4X16__AARCH64_NEON_MLAL_LANE_PRFM_CORTEX_A53, k_eq_8) {
TEST_REQUIRES_ARM_NEON;
GemmMicrokernelTester()
diff --git a/test/qc8-igemm-minmax-fp32.yaml b/test/qc8-igemm-minmax-fp32.yaml
index 0ab6b5e..d2e34c2 100644
--- a/test/qc8-igemm-minmax-fp32.yaml
+++ b/test/qc8-igemm-minmax-fp32.yaml
@@ -3,6 +3,9 @@
# This source code is licensed under the BSD-style license found in the
# LICENSE file in the root directory of this source tree.
+- name: xnn_qc8_igemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal
+ init: xnn_init_qs8_minmax_neon_params
+ k-block: 16
- name: xnn_qc8_igemm_minmax_fp32_ukernel_4x16__aarch64_neon_mlal_lane_prfm_cortex_a53
init: xnn_init_qs8_minmax_neon_params
k-block: 8
diff --git a/test/qs8-gemm-minmax-fp32.cc b/test/qs8-gemm-minmax-fp32.cc
index a7bf802..6269a55 100644
--- a/test/qs8-gemm-minmax-fp32.cc
+++ b/test/qs8-gemm-minmax-fp32.cc
@@ -23,6 +23,462 @@
#if XNN_ARCH_ARM64
+ TEST(QS8_GEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, k_eq_16) {
+ TEST_REQUIRES_ARM_NEON;
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(16)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_fp32_neonv8_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+
+ TEST(QS8_GEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, strided_cn) {
+ TEST_REQUIRES_ARM_NEON;
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(16)
+ .cn_stride(11)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_fp32_neonv8_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+
+ TEST(QS8_GEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, k_eq_16_strided_a) {
+ TEST_REQUIRES_ARM_NEON;
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(16)
+ .a_stride(19)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_fp32_neonv8_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+
+ TEST(QS8_GEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, k_eq_16_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(16)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_fp32_neonv8_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+
+ TEST(QS8_GEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, k_eq_16_subtile_m) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(m)
+ .n(8)
+ .k(16)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_fp32_neonv8_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+
+ TEST(QS8_GEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, k_eq_16_subtile_n) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(16)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_fp32_neonv8_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+
+ TEST(QS8_GEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, k_lt_16) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 1; k < 16; k++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_fp32_neonv8_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+
+ TEST(QS8_GEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, k_lt_16_strided_a) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 1; k < 16; k++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .a_stride(19)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_fp32_neonv8_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+
+ TEST(QS8_GEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, k_lt_16_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 1; k < 16; k++) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_fp32_neonv8_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+ }
+
+ TEST(QS8_GEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, k_gt_16) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 17; k < 32; k++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_fp32_neonv8_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+
+ TEST(QS8_GEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, k_gt_16_strided_a) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 17; k < 32; k++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .a_stride(37)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_fp32_neonv8_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+
+ TEST(QS8_GEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, k_gt_16_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 17; k < 32; k++) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_fp32_neonv8_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+ }
+
+ TEST(QS8_GEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, k_div_16) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 32; k <= 160; k += 16) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_fp32_neonv8_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+
+ TEST(QS8_GEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, k_div_16_strided_a) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 32; k <= 160; k += 16) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .a_stride(163)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_fp32_neonv8_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+
+ TEST(QS8_GEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, k_div_16_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 32; k <= 160; k += 16) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_fp32_neonv8_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+ }
+
+ TEST(QS8_GEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, n_gt_8) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 9; n < 16; n++) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_fp32_neonv8_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+
+ TEST(QS8_GEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, n_gt_8_strided_cn) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 9; n < 16; n++) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .cn_stride(11)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_fp32_neonv8_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+
+ TEST(QS8_GEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, n_gt_8_strided_a) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 9; n < 16; n++) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(k)
+ .a_stride(83)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_fp32_neonv8_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+
+ TEST(QS8_GEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, n_gt_8_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 9; n < 16; n++) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_fp32_neonv8_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+ }
+
+ TEST(QS8_GEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, n_div_8) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 16; n <= 24; n += 8) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_fp32_neonv8_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+
+ TEST(QS8_GEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, n_div_8_strided_cn) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 16; n <= 24; n += 8) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(k)
+ .cn_stride(11)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_fp32_neonv8_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+
+ TEST(QS8_GEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, n_div_8_strided_a) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 16; n <= 24; n += 8) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(k)
+ .a_stride(83)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_fp32_neonv8_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+
+ TEST(QS8_GEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, n_div_8_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 16; n <= 24; n += 8) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_fp32_neonv8_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+ }
+
+ TEST(QS8_GEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, strided_cm_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 1; k <= 80; k += 17) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(11)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_fp32_neonv8_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+ }
+
+ TEST(QS8_GEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, qmin) {
+ TEST_REQUIRES_ARM_NEON;
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(16)
+ .qmin(128)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_fp32_neonv8_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+
+ TEST(QS8_GEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, qmax) {
+ TEST_REQUIRES_ARM_NEON;
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(16)
+ .qmax(128)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_fp32_neonv8_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+
+ TEST(QS8_GEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, strided_cm) {
+ TEST_REQUIRES_ARM_NEON;
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(16)
+ .cm_stride(11)
+ .Test(xnn_qs8_gemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_fp32_neonv8_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+#endif // XNN_ARCH_ARM64
+
+
+#if XNN_ARCH_ARM64
TEST(QS8_GEMM_MINMAX_FP32_4X16__AARCH64_NEON_MLAL_LANE_PRFM_CORTEX_A53, k_eq_8) {
TEST_REQUIRES_ARM_NEON;
GemmMicrokernelTester()
diff --git a/test/qs8-gemm-minmax-fp32.yaml b/test/qs8-gemm-minmax-fp32.yaml
index e13389f..945a1fe 100644
--- a/test/qs8-gemm-minmax-fp32.yaml
+++ b/test/qs8-gemm-minmax-fp32.yaml
@@ -3,6 +3,9 @@
# This source code is licensed under the BSD-style license found in the
# LICENSE file in the root directory of this source tree.
+- name: xnn_qs8_gemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal
+ init: xnn_init_qs8_conv_minmax_fp32_neonv8_params
+ k-block: 16
- name: xnn_qs8_gemm_minmax_fp32_ukernel_4x16__aarch64_neon_mlal_lane_prfm_cortex_a53
init: xnn_init_qs8_conv_minmax_fp32_neonv8_params
k-block: 8
diff --git a/test/qs8-gemm-minmax-gemmlowp.cc b/test/qs8-gemm-minmax-gemmlowp.cc
index 709167f..9d20437 100644
--- a/test/qs8-gemm-minmax-gemmlowp.cc
+++ b/test/qs8-gemm-minmax-gemmlowp.cc
@@ -23,6 +23,462 @@
#if XNN_ARCH_ARM64
+ TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, k_eq_16) {
+ TEST_REQUIRES_ARM_NEON;
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(16)
+ .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+
+ TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, strided_cn) {
+ TEST_REQUIRES_ARM_NEON;
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(16)
+ .cn_stride(11)
+ .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+
+ TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, k_eq_16_strided_a) {
+ TEST_REQUIRES_ARM_NEON;
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(16)
+ .a_stride(19)
+ .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+
+ TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, k_eq_16_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(16)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+
+ TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, k_eq_16_subtile_m) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(m)
+ .n(8)
+ .k(16)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+
+ TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, k_eq_16_subtile_n) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(16)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+
+ TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, k_lt_16) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 1; k < 16; k++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+
+ TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, k_lt_16_strided_a) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 1; k < 16; k++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .a_stride(19)
+ .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+
+ TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, k_lt_16_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 1; k < 16; k++) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+ }
+
+ TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, k_gt_16) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 17; k < 32; k++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+
+ TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, k_gt_16_strided_a) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 17; k < 32; k++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .a_stride(37)
+ .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+
+ TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, k_gt_16_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 17; k < 32; k++) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+ }
+
+ TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, k_div_16) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 32; k <= 160; k += 16) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+
+ TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, k_div_16_strided_a) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 32; k <= 160; k += 16) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .a_stride(163)
+ .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+
+ TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, k_div_16_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 32; k <= 160; k += 16) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+ }
+
+ TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, n_gt_8) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 9; n < 16; n++) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+
+ TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, n_gt_8_strided_cn) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 9; n < 16; n++) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .cn_stride(11)
+ .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+
+ TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, n_gt_8_strided_a) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 9; n < 16; n++) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(k)
+ .a_stride(83)
+ .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+
+ TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, n_gt_8_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 9; n < 16; n++) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+ }
+
+ TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, n_div_8) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 16; n <= 24; n += 8) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+
+ TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, n_div_8_strided_cn) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 16; n <= 24; n += 8) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(k)
+ .cn_stride(11)
+ .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+
+ TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, n_div_8_strided_a) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 16; n <= 24; n += 8) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(k)
+ .a_stride(83)
+ .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+
+ TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, n_div_8_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 16; n <= 24; n += 8) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+ }
+
+ TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, strided_cm_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 1; k <= 80; k += 17) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(11)
+ .iterations(1)
+ .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+ }
+
+ TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, qmin) {
+ TEST_REQUIRES_ARM_NEON;
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(16)
+ .qmin(128)
+ .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+
+ TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, qmax) {
+ TEST_REQUIRES_ARM_NEON;
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(16)
+ .qmax(128)
+ .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+
+ TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, strided_cm) {
+ TEST_REQUIRES_ARM_NEON;
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(16)
+ .cm_stride(11)
+ .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+#endif // XNN_ARCH_ARM64
+
+
+#if XNN_ARCH_ARM64
TEST(QS8_GEMM_MINMAX_GEMMLOWP_4X16__AARCH64_NEON_MLAL_LANE_PRFM_CORTEX_A53, k_eq_8) {
TEST_REQUIRES_ARM_NEON;
GemmMicrokernelTester()
@@ -3670,4110 +4126,6 @@
#endif // XNN_ARCH_ARM64
-#if XNN_ARCH_ARM64
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL, k_eq_16) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(16)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL, strided_cn) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(16)
- .cn_stride(11)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL, k_eq_16_strided_a) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(16)
- .a_stride(19)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL, k_eq_16_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t m = 1; m <= 1; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(16)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL, k_eq_16_subtile_m) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t m = 1; m <= 1; m++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(8)
- .k(16)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL, k_eq_16_subtile_n) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(n)
- .k(16)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL, k_lt_16) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 1; k < 16; k++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(k)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL, k_lt_16_strided_a) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 1; k < 16; k++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(k)
- .a_stride(19)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL, k_lt_16_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 1; k < 16; k++) {
- for (uint32_t m = 1; m <= 1; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL, k_gt_16) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 17; k < 32; k++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(k)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL, k_gt_16_strided_a) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 17; k < 32; k++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(k)
- .a_stride(37)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL, k_gt_16_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 17; k < 32; k++) {
- for (uint32_t m = 1; m <= 1; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL, k_div_16) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 32; k <= 160; k += 16) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(k)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL, k_div_16_strided_a) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 32; k <= 160; k += 16) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(k)
- .a_stride(163)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL, k_div_16_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 32; k <= 160; k += 16) {
- for (uint32_t m = 1; m <= 1; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL, n_gt_8) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 9; n < 16; n++) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(k)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL, n_gt_8_strided_cn) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 9; n < 16; n++) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(k)
- .cn_stride(11)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL, n_gt_8_strided_a) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 9; n < 16; n++) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(n)
- .k(k)
- .a_stride(83)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL, n_gt_8_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 9; n < 16; n++) {
- for (size_t k = 1; k <= 80; k += 17) {
- for (uint32_t m = 1; m <= 1; m++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL, n_div_8) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 16; n <= 24; n += 8) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(k)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL, n_div_8_strided_cn) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 16; n <= 24; n += 8) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(n)
- .k(k)
- .cn_stride(11)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL, n_div_8_strided_a) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 16; n <= 24; n += 8) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(n)
- .k(k)
- .a_stride(83)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL, n_div_8_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 16; n <= 24; n += 8) {
- for (size_t k = 1; k <= 80; k += 17) {
- for (uint32_t m = 1; m <= 1; m++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL, strided_cm_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 1; k <= 80; k += 17) {
- for (uint32_t m = 1; m <= 1; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .cm_stride(11)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL, qmin) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(16)
- .qmin(128)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL, qmax) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(16)
- .qmax(128)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL, strided_cm) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(16)
- .cm_stride(11)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-#endif // XNN_ARCH_ARM64
-
-
-#if XNN_ARCH_ARM64
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, k_eq_16) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(16)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, strided_cn) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(16)
- .cn_stride(11)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, k_eq_16_strided_a) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(16)
- .a_stride(19)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, k_eq_16_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t m = 1; m <= 1; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(16)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, k_eq_16_subtile_m) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t m = 1; m <= 1; m++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(8)
- .k(16)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, k_eq_16_subtile_n) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(n)
- .k(16)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, k_lt_16) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 1; k < 16; k++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(k)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, k_lt_16_strided_a) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 1; k < 16; k++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(k)
- .a_stride(19)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, k_lt_16_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 1; k < 16; k++) {
- for (uint32_t m = 1; m <= 1; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, k_gt_16) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 17; k < 32; k++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(k)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, k_gt_16_strided_a) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 17; k < 32; k++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(k)
- .a_stride(37)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, k_gt_16_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 17; k < 32; k++) {
- for (uint32_t m = 1; m <= 1; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, k_div_16) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 32; k <= 160; k += 16) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(k)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, k_div_16_strided_a) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 32; k <= 160; k += 16) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(k)
- .a_stride(163)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, k_div_16_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 32; k <= 160; k += 16) {
- for (uint32_t m = 1; m <= 1; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, n_gt_8) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 9; n < 16; n++) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(k)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, n_gt_8_strided_cn) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 9; n < 16; n++) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(k)
- .cn_stride(11)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, n_gt_8_strided_a) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 9; n < 16; n++) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(n)
- .k(k)
- .a_stride(83)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, n_gt_8_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 9; n < 16; n++) {
- for (size_t k = 1; k <= 80; k += 17) {
- for (uint32_t m = 1; m <= 1; m++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, n_div_8) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 16; n <= 24; n += 8) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(k)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, n_div_8_strided_cn) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 16; n <= 24; n += 8) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(n)
- .k(k)
- .cn_stride(11)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, n_div_8_strided_a) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 16; n <= 24; n += 8) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(n)
- .k(k)
- .a_stride(83)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, n_div_8_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 16; n <= 24; n += 8) {
- for (size_t k = 1; k <= 80; k += 17) {
- for (uint32_t m = 1; m <= 1; m++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, strided_cm_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 1; k <= 80; k += 17) {
- for (uint32_t m = 1; m <= 1; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .cm_stride(11)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, qmin) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(16)
- .qmin(128)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, qmax) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(16)
- .qmax(128)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, strided_cm) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(16)
- .cm_stride(11)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-#endif // XNN_ARCH_ARM64
-
-
-#if XNN_ARCH_ARM64
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, k_eq_16) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(16)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, strided_cn) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(16)
- .cn_stride(11)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, k_eq_16_strided_a) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(16)
- .a_stride(19)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, k_eq_16_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t m = 1; m <= 1; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(16)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, k_eq_16_subtile_m) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t m = 1; m <= 1; m++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(8)
- .k(16)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, k_eq_16_subtile_n) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(n)
- .k(16)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, k_lt_16) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 1; k < 16; k++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(k)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, k_lt_16_strided_a) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 1; k < 16; k++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(k)
- .a_stride(19)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, k_lt_16_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 1; k < 16; k++) {
- for (uint32_t m = 1; m <= 1; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, k_gt_16) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 17; k < 32; k++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(k)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, k_gt_16_strided_a) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 17; k < 32; k++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(k)
- .a_stride(37)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, k_gt_16_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 17; k < 32; k++) {
- for (uint32_t m = 1; m <= 1; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, k_div_16) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 32; k <= 160; k += 16) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(k)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, k_div_16_strided_a) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 32; k <= 160; k += 16) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(k)
- .a_stride(163)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, k_div_16_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 32; k <= 160; k += 16) {
- for (uint32_t m = 1; m <= 1; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, n_gt_8) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 9; n < 16; n++) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(k)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, n_gt_8_strided_cn) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 9; n < 16; n++) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(k)
- .cn_stride(11)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, n_gt_8_strided_a) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 9; n < 16; n++) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(n)
- .k(k)
- .a_stride(83)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, n_gt_8_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 9; n < 16; n++) {
- for (size_t k = 1; k <= 80; k += 17) {
- for (uint32_t m = 1; m <= 1; m++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, n_div_8) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 16; n <= 24; n += 8) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(k)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, n_div_8_strided_cn) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 16; n <= 24; n += 8) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(n)
- .k(k)
- .cn_stride(11)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, n_div_8_strided_a) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 16; n <= 24; n += 8) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(n)
- .k(k)
- .a_stride(83)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, n_div_8_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 16; n <= 24; n += 8) {
- for (size_t k = 1; k <= 80; k += 17) {
- for (uint32_t m = 1; m <= 1; m++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, strided_cm_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 1; k <= 80; k += 17) {
- for (uint32_t m = 1; m <= 1; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .cm_stride(11)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, qmin) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(16)
- .qmin(128)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, qmax) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(16)
- .qmax(128)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, strided_cm) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(16)
- .cm_stride(11)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-#endif // XNN_ARCH_ARM64
-
-
-#if XNN_ARCH_ARM64
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, k_eq_16) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(16)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, strided_cn) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(16)
- .cn_stride(11)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, k_eq_16_strided_a) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(16)
- .a_stride(19)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, k_eq_16_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t m = 1; m <= 1; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(16)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, k_eq_16_subtile_m) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t m = 1; m <= 1; m++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(8)
- .k(16)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, k_eq_16_subtile_n) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(n)
- .k(16)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, k_lt_16) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 1; k < 16; k++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(k)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, k_lt_16_strided_a) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 1; k < 16; k++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(k)
- .a_stride(19)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, k_lt_16_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 1; k < 16; k++) {
- for (uint32_t m = 1; m <= 1; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, k_gt_16) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 17; k < 32; k++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(k)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, k_gt_16_strided_a) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 17; k < 32; k++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(k)
- .a_stride(37)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, k_gt_16_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 17; k < 32; k++) {
- for (uint32_t m = 1; m <= 1; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, k_div_16) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 32; k <= 160; k += 16) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(k)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, k_div_16_strided_a) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 32; k <= 160; k += 16) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(k)
- .a_stride(163)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, k_div_16_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 32; k <= 160; k += 16) {
- for (uint32_t m = 1; m <= 1; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, n_gt_8) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 9; n < 16; n++) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(k)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, n_gt_8_strided_cn) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 9; n < 16; n++) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(k)
- .cn_stride(11)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, n_gt_8_strided_a) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 9; n < 16; n++) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(n)
- .k(k)
- .a_stride(83)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, n_gt_8_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 9; n < 16; n++) {
- for (size_t k = 1; k <= 80; k += 17) {
- for (uint32_t m = 1; m <= 1; m++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, n_div_8) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 16; n <= 24; n += 8) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(k)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, n_div_8_strided_cn) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 16; n <= 24; n += 8) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(n)
- .k(k)
- .cn_stride(11)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, n_div_8_strided_a) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 16; n <= 24; n += 8) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(n)
- .k(k)
- .a_stride(83)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, n_div_8_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 16; n <= 24; n += 8) {
- for (size_t k = 1; k <= 80; k += 17) {
- for (uint32_t m = 1; m <= 1; m++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, strided_cm_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 1; k <= 80; k += 17) {
- for (uint32_t m = 1; m <= 1; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .cm_stride(11)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, qmin) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(16)
- .qmin(128)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, qmax) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(16)
- .qmax(128)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, strided_cm) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(16)
- .cm_stride(11)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-#endif // XNN_ARCH_ARM64
-
-
-#if XNN_ARCH_ARM64
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, k_eq_16) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(16)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, strided_cn) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(16)
- .cn_stride(11)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, k_eq_16_strided_a) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(16)
- .a_stride(19)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, k_eq_16_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(16)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, k_eq_16_subtile_m) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t m = 1; m <= 2; m++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(8)
- .k(16)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, k_eq_16_subtile_n) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(n)
- .k(16)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, k_lt_16) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 1; k < 16; k++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, k_lt_16_strided_a) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 1; k < 16; k++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .a_stride(19)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, k_lt_16_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 1; k < 16; k++) {
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, k_gt_16) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 17; k < 32; k++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, k_gt_16_strided_a) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 17; k < 32; k++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .a_stride(37)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, k_gt_16_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 17; k < 32; k++) {
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, k_div_16) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 32; k <= 160; k += 16) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, k_div_16_strided_a) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 32; k <= 160; k += 16) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .a_stride(163)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, k_div_16_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 32; k <= 160; k += 16) {
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, n_gt_8) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 9; n < 16; n++) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, n_gt_8_strided_cn) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 9; n < 16; n++) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .cn_stride(11)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, n_gt_8_strided_a) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 9; n < 16; n++) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(n)
- .k(k)
- .a_stride(83)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, n_gt_8_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 9; n < 16; n++) {
- for (size_t k = 1; k <= 80; k += 17) {
- for (uint32_t m = 1; m <= 2; m++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, n_div_8) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 16; n <= 24; n += 8) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, n_div_8_strided_cn) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 16; n <= 24; n += 8) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(n)
- .k(k)
- .cn_stride(11)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, n_div_8_strided_a) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 16; n <= 24; n += 8) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(n)
- .k(k)
- .a_stride(83)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, n_div_8_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 16; n <= 24; n += 8) {
- for (size_t k = 1; k <= 80; k += 17) {
- for (uint32_t m = 1; m <= 2; m++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, strided_cm_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 1; k <= 80; k += 17) {
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .cm_stride(11)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, qmin) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(16)
- .qmin(128)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, qmax) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(16)
- .qmax(128)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, strided_cm) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(16)
- .cm_stride(11)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-#endif // XNN_ARCH_ARM64
-
-
-#if XNN_ARCH_ARM64
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, k_eq_16) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(16)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, strided_cn) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(16)
- .cn_stride(11)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, k_eq_16_strided_a) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(16)
- .a_stride(19)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, k_eq_16_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(16)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, k_eq_16_subtile_m) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t m = 1; m <= 2; m++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(8)
- .k(16)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, k_eq_16_subtile_n) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(n)
- .k(16)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, k_lt_16) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 1; k < 16; k++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, k_lt_16_strided_a) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 1; k < 16; k++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .a_stride(19)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, k_lt_16_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 1; k < 16; k++) {
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, k_gt_16) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 17; k < 32; k++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, k_gt_16_strided_a) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 17; k < 32; k++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .a_stride(37)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, k_gt_16_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 17; k < 32; k++) {
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, k_div_16) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 32; k <= 160; k += 16) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, k_div_16_strided_a) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 32; k <= 160; k += 16) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .a_stride(163)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, k_div_16_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 32; k <= 160; k += 16) {
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, n_gt_8) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 9; n < 16; n++) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, n_gt_8_strided_cn) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 9; n < 16; n++) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .cn_stride(11)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, n_gt_8_strided_a) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 9; n < 16; n++) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(n)
- .k(k)
- .a_stride(83)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, n_gt_8_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 9; n < 16; n++) {
- for (size_t k = 1; k <= 80; k += 17) {
- for (uint32_t m = 1; m <= 2; m++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, n_div_8) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 16; n <= 24; n += 8) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, n_div_8_strided_cn) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 16; n <= 24; n += 8) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(n)
- .k(k)
- .cn_stride(11)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, n_div_8_strided_a) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 16; n <= 24; n += 8) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(n)
- .k(k)
- .a_stride(83)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, n_div_8_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 16; n <= 24; n += 8) {
- for (size_t k = 1; k <= 80; k += 17) {
- for (uint32_t m = 1; m <= 2; m++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, strided_cm_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 1; k <= 80; k += 17) {
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .cm_stride(11)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, qmin) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(16)
- .qmin(128)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, qmax) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(16)
- .qmax(128)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, strided_cm) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(16)
- .cm_stride(11)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-#endif // XNN_ARCH_ARM64
-
-
-#if XNN_ARCH_ARM64
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL, k_eq_16) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(16)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL, strided_cn) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(16)
- .cn_stride(11)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL, k_eq_16_strided_a) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(16)
- .a_stride(19)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL, k_eq_16_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(16)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL, k_eq_16_subtile_m) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t m = 1; m <= 2; m++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(8)
- .k(16)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL, k_eq_16_subtile_n) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(n)
- .k(16)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL, k_lt_16) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 1; k < 16; k++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL, k_lt_16_strided_a) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 1; k < 16; k++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .a_stride(19)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL, k_lt_16_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 1; k < 16; k++) {
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL, k_gt_16) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 17; k < 32; k++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL, k_gt_16_strided_a) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 17; k < 32; k++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .a_stride(37)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL, k_gt_16_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 17; k < 32; k++) {
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL, k_div_16) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 32; k <= 160; k += 16) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL, k_div_16_strided_a) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 32; k <= 160; k += 16) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .a_stride(163)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL, k_div_16_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 32; k <= 160; k += 16) {
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL, n_gt_8) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 9; n < 16; n++) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL, n_gt_8_strided_cn) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 9; n < 16; n++) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .cn_stride(11)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL, n_gt_8_strided_a) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 9; n < 16; n++) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(n)
- .k(k)
- .a_stride(83)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL, n_gt_8_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 9; n < 16; n++) {
- for (size_t k = 1; k <= 80; k += 17) {
- for (uint32_t m = 1; m <= 2; m++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL, n_div_8) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 16; n <= 24; n += 8) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL, n_div_8_strided_cn) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 16; n <= 24; n += 8) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(n)
- .k(k)
- .cn_stride(11)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL, n_div_8_strided_a) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 16; n <= 24; n += 8) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(n)
- .k(k)
- .a_stride(83)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL, n_div_8_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 16; n <= 24; n += 8) {
- for (size_t k = 1; k <= 80; k += 17) {
- for (uint32_t m = 1; m <= 2; m++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL, strided_cm_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 1; k <= 80; k += 17) {
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .cm_stride(11)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL, qmin) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(16)
- .qmin(128)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL, qmax) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(16)
- .qmax(128)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL, strided_cm) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(16)
- .cm_stride(11)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-#endif // XNN_ARCH_ARM64
-
-
-#if XNN_ARCH_ARM64
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, k_eq_16) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(16)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, strided_cn) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(16)
- .cn_stride(11)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, k_eq_16_strided_a) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(16)
- .a_stride(19)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, k_eq_16_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(16)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, k_eq_16_subtile_m) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t m = 1; m <= 2; m++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(8)
- .k(16)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, k_eq_16_subtile_n) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(n)
- .k(16)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, k_lt_16) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 1; k < 16; k++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, k_lt_16_strided_a) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 1; k < 16; k++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .a_stride(19)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, k_lt_16_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 1; k < 16; k++) {
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, k_gt_16) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 17; k < 32; k++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, k_gt_16_strided_a) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 17; k < 32; k++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .a_stride(37)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, k_gt_16_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 17; k < 32; k++) {
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, k_div_16) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 32; k <= 160; k += 16) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, k_div_16_strided_a) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 32; k <= 160; k += 16) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .a_stride(163)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, k_div_16_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 32; k <= 160; k += 16) {
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, n_gt_8) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 9; n < 16; n++) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, n_gt_8_strided_cn) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 9; n < 16; n++) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .cn_stride(11)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, n_gt_8_strided_a) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 9; n < 16; n++) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(n)
- .k(k)
- .a_stride(83)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, n_gt_8_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 9; n < 16; n++) {
- for (size_t k = 1; k <= 80; k += 17) {
- for (uint32_t m = 1; m <= 2; m++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, n_div_8) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 16; n <= 24; n += 8) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, n_div_8_strided_cn) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 16; n <= 24; n += 8) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(n)
- .k(k)
- .cn_stride(11)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, n_div_8_strided_a) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 16; n <= 24; n += 8) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(n)
- .k(k)
- .a_stride(83)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, n_div_8_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 16; n <= 24; n += 8) {
- for (size_t k = 1; k <= 80; k += 17) {
- for (uint32_t m = 1; m <= 2; m++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, strided_cm_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 1; k <= 80; k += 17) {
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .cm_stride(11)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, qmin) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(16)
- .qmin(128)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, qmax) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(16)
- .qmax(128)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, strided_cm) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(16)
- .cm_stride(11)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-#endif // XNN_ARCH_ARM64
-
-
-#if XNN_ARCH_ARM64
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MULL_PADAL, k_eq_8) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(8)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mull_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MULL_PADAL, strided_cn) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(8)
- .cn_stride(11)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mull_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MULL_PADAL, k_eq_8_strided_a) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(8)
- .a_stride(11)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mull_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MULL_PADAL, k_eq_8_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(8)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mull_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MULL_PADAL, k_eq_8_subtile_m) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t m = 1; m <= 2; m++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(8)
- .k(8)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mull_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MULL_PADAL, k_eq_8_subtile_n) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(n)
- .k(8)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mull_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MULL_PADAL, k_lt_8) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 1; k < 8; k++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mull_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MULL_PADAL, k_lt_8_strided_a) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 1; k < 8; k++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .a_stride(11)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mull_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MULL_PADAL, k_lt_8_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 1; k < 8; k++) {
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mull_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MULL_PADAL, k_gt_8) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 9; k < 16; k++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mull_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MULL_PADAL, k_gt_8_strided_a) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 9; k < 16; k++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .a_stride(19)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mull_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MULL_PADAL, k_gt_8_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 9; k < 16; k++) {
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mull_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MULL_PADAL, k_div_8) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 16; k <= 80; k += 8) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mull_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MULL_PADAL, k_div_8_strided_a) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 16; k <= 80; k += 8) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .a_stride(83)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mull_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MULL_PADAL, k_div_8_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 16; k <= 80; k += 8) {
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mull_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MULL_PADAL, n_gt_8) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 9; n < 16; n++) {
- for (size_t k = 1; k <= 40; k += 9) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mull_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MULL_PADAL, n_gt_8_strided_cn) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 9; n < 16; n++) {
- for (size_t k = 1; k <= 40; k += 9) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .cn_stride(11)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mull_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MULL_PADAL, n_gt_8_strided_a) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 9; n < 16; n++) {
- for (size_t k = 1; k <= 40; k += 9) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(n)
- .k(k)
- .a_stride(43)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mull_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MULL_PADAL, n_gt_8_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 9; n < 16; n++) {
- for (size_t k = 1; k <= 40; k += 9) {
- for (uint32_t m = 1; m <= 2; m++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mull_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MULL_PADAL, n_div_8) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 16; n <= 24; n += 8) {
- for (size_t k = 1; k <= 40; k += 9) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mull_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MULL_PADAL, n_div_8_strided_cn) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 16; n <= 24; n += 8) {
- for (size_t k = 1; k <= 40; k += 9) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(n)
- .k(k)
- .cn_stride(11)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mull_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MULL_PADAL, n_div_8_strided_a) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 16; n <= 24; n += 8) {
- for (size_t k = 1; k <= 40; k += 9) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(n)
- .k(k)
- .a_stride(43)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mull_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MULL_PADAL, n_div_8_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 16; n <= 24; n += 8) {
- for (size_t k = 1; k <= 40; k += 9) {
- for (uint32_t m = 1; m <= 2; m++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mull_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MULL_PADAL, strided_cm_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 1; k <= 40; k += 9) {
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .cm_stride(11)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mull_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MULL_PADAL, qmin) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(8)
- .qmin(128)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mull_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MULL_PADAL, qmax) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(8)
- .qmax(128)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mull_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MULL_PADAL, strided_cm) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(8)
- .cm_stride(11)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mull_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-#endif // XNN_ARCH_ARM64
-
-
#if XNN_ARCH_ARM || XNN_ARCH_ARM64
TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X8__NEON_MLAL_LANE, k_eq_8) {
TEST_REQUIRES_ARM_NEON;
@@ -42430,462 +38782,6 @@
#endif // XNN_ARCH_ARM && !XNN_PLATFORM_IOS || XNN_ARCH_ARM64
-#if XNN_ARCH_ARM64
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, k_eq_16) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(16)
- .sr(1)
- .m(2)
- .n(8)
- .k(16)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, strided_cn) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(16)
- .sr(1)
- .m(2)
- .n(8)
- .k(16)
- .cn_stride(11)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, k_eq_16_strided_a) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(16)
- .sr(1)
- .m(2)
- .n(8)
- .k(16)
- .a_stride(19)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, k_eq_16_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(16)
- .sr(1)
- .m(m)
- .n(n)
- .k(16)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, k_eq_16_subtile_m) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t m = 1; m <= 2; m++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(16)
- .sr(1)
- .m(m)
- .n(8)
- .k(16)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, k_eq_16_subtile_n) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(16)
- .sr(1)
- .m(2)
- .n(n)
- .k(16)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, k_lt_16) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 1; k < 16; k++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(16)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, k_lt_16_strided_a) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 1; k < 16; k++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(16)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .a_stride(19)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, k_lt_16_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 1; k < 16; k++) {
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(16)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, k_gt_16) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 17; k < 32; k++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(16)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, k_gt_16_strided_a) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 17; k < 32; k++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(16)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .a_stride(37)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, k_gt_16_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 17; k < 32; k++) {
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(16)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, k_div_16) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 32; k <= 160; k += 16) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(16)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, k_div_16_strided_a) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 32; k <= 160; k += 16) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(16)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .a_stride(163)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, k_div_16_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 32; k <= 160; k += 16) {
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(16)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, n_gt_8) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 9; n < 16; n++) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(16)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, n_gt_8_strided_cn) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 9; n < 16; n++) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(16)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .cn_stride(11)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, n_gt_8_strided_a) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 9; n < 16; n++) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(16)
- .sr(1)
- .m(2)
- .n(n)
- .k(k)
- .a_stride(83)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, n_gt_8_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 9; n < 16; n++) {
- for (size_t k = 1; k <= 80; k += 17) {
- for (uint32_t m = 1; m <= 2; m++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(16)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, n_div_8) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 16; n <= 24; n += 8) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(16)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, n_div_8_strided_cn) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 16; n <= 24; n += 8) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(16)
- .sr(1)
- .m(2)
- .n(n)
- .k(k)
- .cn_stride(11)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, n_div_8_strided_a) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 16; n <= 24; n += 8) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(16)
- .sr(1)
- .m(2)
- .n(n)
- .k(k)
- .a_stride(83)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, n_div_8_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 16; n <= 24; n += 8) {
- for (size_t k = 1; k <= 80; k += 17) {
- for (uint32_t m = 1; m <= 2; m++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(16)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, strided_cm_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 1; k <= 80; k += 17) {
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(16)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .cm_stride(11)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, qmin) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(16)
- .sr(1)
- .m(2)
- .n(8)
- .k(16)
- .qmin(128)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, qmax) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(16)
- .sr(1)
- .m(2)
- .n(8)
- .k(16)
- .qmax(128)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-
- TEST(QS8_GEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, strided_cm) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(16)
- .sr(1)
- .m(2)
- .n(8)
- .k(16)
- .cm_stride(11)
- .Test(xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-#endif // XNN_ARCH_ARM64
-
-
#if XNN_ARCH_X86 || XNN_ARCH_X86_64
TEST(QS8_GEMM_MINMAX_GEMMLOWP_1X4C2__SSE2_LD64, k_eq_8) {
TEST_REQUIRES_X86_SSE2;
diff --git a/test/qs8-gemm-minmax-gemmlowp.yaml b/test/qs8-gemm-minmax-gemmlowp.yaml
index 9748274..3053ae8 100644
--- a/test/qs8-gemm-minmax-gemmlowp.yaml
+++ b/test/qs8-gemm-minmax-gemmlowp.yaml
@@ -3,6 +3,9 @@
# This source code is licensed under the BSD-style license found in the
# LICENSE file in the root directory of this source tree.
+- name: xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal
+ init: xnn_init_qs8_conv_minmax_gemmlowp_neon_params
+ k-block: 16
- name: xnn_qs8_gemm_minmax_gemmlowp_ukernel_4x16__aarch64_neon_mlal_lane_prfm_cortex_a53
init: xnn_init_qs8_conv_minmax_gemmlowp_neon_params
k-block: 8
@@ -27,33 +30,6 @@
- name: xnn_qs8_gemm_minmax_gemmlowp_ukernel_4x16c4__aarch64_neondot_cortex_a55
init: xnn_init_qs8_conv_minmax_gemmlowp_neon_params
k-block: 16
-- name: xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal
- init: xnn_init_qs8_conv_minmax_gemmlowp_neon_params
- k-block: 16
-- name: xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm
- init: xnn_init_qs8_conv_minmax_gemmlowp_neon_params
- k-block: 16
-- name: xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_cortex_a53
- init: xnn_init_qs8_conv_minmax_gemmlowp_neon_params
- k-block: 16
-- name: xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53
- init: xnn_init_qs8_conv_minmax_gemmlowp_neon_params
- k-block: 16
-- name: xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_cortex_a53
- init: xnn_init_qs8_conv_minmax_gemmlowp_neon_params
- k-block: 16
-- name: xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53
- init: xnn_init_qs8_conv_minmax_gemmlowp_neon_params
- k-block: 16
-- name: xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal
- init: xnn_init_qs8_conv_minmax_gemmlowp_neon_params
- k-block: 16
-- name: xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm
- init: xnn_init_qs8_conv_minmax_gemmlowp_neon_params
- k-block: 16
-- name: xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mull_padal
- init: xnn_init_qs8_conv_minmax_gemmlowp_neon_params
- k-block: 8
- name: xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x8__neon_mlal_lane
init: xnn_init_qs8_conv_minmax_gemmlowp_neon_params
k-block: 8
@@ -282,9 +258,6 @@
- name: xnn_qs8_gemm_minmax_gemmlowp_ukernel_8x16c4__neondot
init: xnn_init_qs8_conv_minmax_gemmlowp_neon_params
k-block: 8
-- name: xnn_qs8_gemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal
- init: xnn_init_qs8_conv_minmax_gemmlowp_neon_params
- k-block: 16
- name: xnn_qs8_gemm_minmax_gemmlowp_ukernel_1x4c2__sse2_ld64
init: xnn_init_qs8_conv_minmax_gemmlowp_sse2_params
k-block: 8
diff --git a/test/qs8-igemm-minmax-fp32.cc b/test/qs8-igemm-minmax-fp32.cc
index c60a0c0..9ac4d84 100644
--- a/test/qs8-igemm-minmax-fp32.cc
+++ b/test/qs8-igemm-minmax-fp32.cc
@@ -23,6 +23,474 @@
#if XNN_ARCH_ARM64
+ TEST(QS8_IGEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, k_eq_16) {
+ TEST_REQUIRES_ARM_NEON;
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(16)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_fp32_neonv8_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+
+ TEST(QS8_IGEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, strided_cn) {
+ TEST_REQUIRES_ARM_NEON;
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(16)
+ .cn_stride(11)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_fp32_neonv8_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+
+ TEST(QS8_IGEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, k_eq_16_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(16)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_fp32_neonv8_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, k_eq_16_subtile_m) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(m)
+ .n(8)
+ .k(16)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_fp32_neonv8_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, k_eq_16_subtile_n) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(16)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_fp32_neonv8_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, k_lt_16) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 1; k < 16; k++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_fp32_neonv8_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, k_lt_16_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 1; k < 16; k++) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_fp32_neonv8_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, k_gt_16) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 17; k < 32; k++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_fp32_neonv8_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, k_gt_16_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 17; k < 32; k++) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_fp32_neonv8_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, k_div_16) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 32; k <= 160; k += 16) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_fp32_neonv8_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, k_div_16_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 32; k <= 160; k += 16) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_fp32_neonv8_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, n_gt_8) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 9; n < 16; n++) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_fp32_neonv8_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, n_gt_8_strided_cn) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 9; n < 16; n++) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .cn_stride(11)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_fp32_neonv8_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, n_gt_8_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 9; n < 16; n++) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_fp32_neonv8_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, n_div_8) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 16; n <= 24; n += 8) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_fp32_neonv8_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, n_div_8_strided_cn) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 16; n <= 24; n += 8) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(k)
+ .cn_stride(11)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_fp32_neonv8_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, n_div_8_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 16; n <= 24; n += 8) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_fp32_neonv8_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, small_kernel) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_fp32_neonv8_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, small_kernel_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 1; k <= 80; k += 17) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .ks(3)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_fp32_neonv8_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, n_gt_8_small_kernel) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 9; n < 16; n++) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_fp32_neonv8_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, n_div_8_small_kernel) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 16; n <= 24; n += 8) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_fp32_neonv8_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, strided_cm_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 1; k <= 80; k += 17) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(11)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_fp32_neonv8_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, a_offset) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .ks(3)
+ .a_offset(163)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_fp32_neonv8_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, zero) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t mz = 0; mz < 2; mz++) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .ks(3)
+ .a_offset(163)
+ .zero_index(mz)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_fp32_neonv8_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, qmin) {
+ TEST_REQUIRES_ARM_NEON;
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(16)
+ .qmin(128)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_fp32_neonv8_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+
+ TEST(QS8_IGEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, qmax) {
+ TEST_REQUIRES_ARM_NEON;
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(16)
+ .qmax(128)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_fp32_neonv8_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+
+ TEST(QS8_IGEMM_MINMAX_FP32_2X8C16__AARCH64_NEON_MLAL_PADAL, strided_cm) {
+ TEST_REQUIRES_ARM_NEON;
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(16)
+ .cm_stride(11)
+ .Test(xnn_qs8_igemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_fp32_neonv8_params, xnn_init_qs8_requantization_fp32_params, xnn_qs8_requantize_fp32);
+ }
+#endif // XNN_ARCH_ARM64
+
+
+#if XNN_ARCH_ARM64
TEST(QS8_IGEMM_MINMAX_FP32_4X16__AARCH64_NEON_MLAL_LANE_PRFM_CORTEX_A53, k_eq_8) {
TEST_REQUIRES_ARM_NEON;
GemmMicrokernelTester()
diff --git a/test/qs8-igemm-minmax-fp32.yaml b/test/qs8-igemm-minmax-fp32.yaml
index c30c31b..b2ecf87 100644
--- a/test/qs8-igemm-minmax-fp32.yaml
+++ b/test/qs8-igemm-minmax-fp32.yaml
@@ -3,6 +3,9 @@
# This source code is licensed under the BSD-style license found in the
# LICENSE file in the root directory of this source tree.
+- name: xnn_qs8_igemm_minmax_fp32_ukernel_2x8c16__aarch64_neon_mlal_padal
+ init: xnn_init_qs8_conv_minmax_fp32_neonv8_params
+ k-block: 16
- name: xnn_qs8_igemm_minmax_fp32_ukernel_4x16__aarch64_neon_mlal_lane_prfm_cortex_a53
init: xnn_init_qs8_conv_minmax_fp32_neonv8_params
k-block: 8
diff --git a/test/qs8-igemm-minmax-gemmlowp.cc b/test/qs8-igemm-minmax-gemmlowp.cc
index 6ab8d9c..bbb87a5 100644
--- a/test/qs8-igemm-minmax-gemmlowp.cc
+++ b/test/qs8-igemm-minmax-gemmlowp.cc
@@ -23,6 +23,4218 @@
#if XNN_ARCH_ARM64
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL, k_eq_16) {
+ TEST_REQUIRES_ARM_NEON;
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(16)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL, strided_cn) {
+ TEST_REQUIRES_ARM_NEON;
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(16)
+ .cn_stride(11)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL, k_eq_16_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(16)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL, k_eq_16_subtile_m) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(m)
+ .n(8)
+ .k(16)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL, k_eq_16_subtile_n) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(16)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL, k_lt_16) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 1; k < 16; k++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL, k_lt_16_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 1; k < 16; k++) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL, k_gt_16) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 17; k < 32; k++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL, k_gt_16_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 17; k < 32; k++) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL, k_div_16) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 32; k <= 160; k += 16) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL, k_div_16_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 32; k <= 160; k += 16) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL, n_gt_8) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 9; n < 16; n++) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL, n_gt_8_strided_cn) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 9; n < 16; n++) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .cn_stride(11)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL, n_gt_8_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 9; n < 16; n++) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL, n_div_8) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 16; n <= 24; n += 8) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL, n_div_8_strided_cn) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 16; n <= 24; n += 8) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(k)
+ .cn_stride(11)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL, n_div_8_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 16; n <= 24; n += 8) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL, small_kernel) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL, small_kernel_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 1; k <= 80; k += 17) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .ks(3)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL, n_gt_8_small_kernel) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 9; n < 16; n++) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL, n_div_8_small_kernel) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 16; n <= 24; n += 8) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL, strided_cm_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 1; k <= 80; k += 17) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(11)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL, a_offset) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .ks(3)
+ .a_offset(163)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL, zero) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t mz = 0; mz < 2; mz++) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .ks(3)
+ .a_offset(163)
+ .zero_index(mz)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL, qmin) {
+ TEST_REQUIRES_ARM_NEON;
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(16)
+ .qmin(128)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL, qmax) {
+ TEST_REQUIRES_ARM_NEON;
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(16)
+ .qmax(128)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL, strided_cm) {
+ TEST_REQUIRES_ARM_NEON;
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(16)
+ .cm_stride(11)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+#endif // XNN_ARCH_ARM64
+
+
+#if XNN_ARCH_ARM64
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, k_eq_16) {
+ TEST_REQUIRES_ARM_NEON;
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(16)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, strided_cn) {
+ TEST_REQUIRES_ARM_NEON;
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(16)
+ .cn_stride(11)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, k_eq_16_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(16)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, k_eq_16_subtile_m) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(m)
+ .n(8)
+ .k(16)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, k_eq_16_subtile_n) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(16)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, k_lt_16) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 1; k < 16; k++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, k_lt_16_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 1; k < 16; k++) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, k_gt_16) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 17; k < 32; k++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, k_gt_16_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 17; k < 32; k++) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, k_div_16) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 32; k <= 160; k += 16) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, k_div_16_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 32; k <= 160; k += 16) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, n_gt_8) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 9; n < 16; n++) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, n_gt_8_strided_cn) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 9; n < 16; n++) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .cn_stride(11)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, n_gt_8_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 9; n < 16; n++) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, n_div_8) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 16; n <= 24; n += 8) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, n_div_8_strided_cn) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 16; n <= 24; n += 8) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(k)
+ .cn_stride(11)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, n_div_8_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 16; n <= 24; n += 8) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, small_kernel) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, small_kernel_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 1; k <= 80; k += 17) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .ks(3)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, n_gt_8_small_kernel) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 9; n < 16; n++) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, n_div_8_small_kernel) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 16; n <= 24; n += 8) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, strided_cm_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 1; k <= 80; k += 17) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(11)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, a_offset) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .ks(3)
+ .a_offset(163)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, zero) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t mz = 0; mz < 2; mz++) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .ks(3)
+ .a_offset(163)
+ .zero_index(mz)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, qmin) {
+ TEST_REQUIRES_ARM_NEON;
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(16)
+ .qmin(128)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, qmax) {
+ TEST_REQUIRES_ARM_NEON;
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(16)
+ .qmax(128)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, strided_cm) {
+ TEST_REQUIRES_ARM_NEON;
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(16)
+ .cm_stride(11)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+#endif // XNN_ARCH_ARM64
+
+
+#if XNN_ARCH_ARM64
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, k_eq_16) {
+ TEST_REQUIRES_ARM_NEON;
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(16)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, strided_cn) {
+ TEST_REQUIRES_ARM_NEON;
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(16)
+ .cn_stride(11)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, k_eq_16_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(16)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, k_eq_16_subtile_m) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(m)
+ .n(8)
+ .k(16)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, k_eq_16_subtile_n) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(16)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, k_lt_16) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 1; k < 16; k++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, k_lt_16_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 1; k < 16; k++) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, k_gt_16) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 17; k < 32; k++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, k_gt_16_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 17; k < 32; k++) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, k_div_16) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 32; k <= 160; k += 16) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, k_div_16_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 32; k <= 160; k += 16) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, n_gt_8) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 9; n < 16; n++) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, n_gt_8_strided_cn) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 9; n < 16; n++) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .cn_stride(11)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, n_gt_8_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 9; n < 16; n++) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, n_div_8) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 16; n <= 24; n += 8) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, n_div_8_strided_cn) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 16; n <= 24; n += 8) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(k)
+ .cn_stride(11)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, n_div_8_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 16; n <= 24; n += 8) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, small_kernel) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, small_kernel_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 1; k <= 80; k += 17) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .ks(3)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, n_gt_8_small_kernel) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 9; n < 16; n++) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, n_div_8_small_kernel) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 16; n <= 24; n += 8) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, strided_cm_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 1; k <= 80; k += 17) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(11)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, a_offset) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .ks(3)
+ .a_offset(163)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, zero) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t mz = 0; mz < 2; mz++) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .ks(3)
+ .a_offset(163)
+ .zero_index(mz)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, qmin) {
+ TEST_REQUIRES_ARM_NEON;
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(16)
+ .qmin(128)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, qmax) {
+ TEST_REQUIRES_ARM_NEON;
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(16)
+ .qmax(128)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, strided_cm) {
+ TEST_REQUIRES_ARM_NEON;
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(16)
+ .cm_stride(11)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+#endif // XNN_ARCH_ARM64
+
+
+#if XNN_ARCH_ARM64
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, k_eq_16) {
+ TEST_REQUIRES_ARM_NEON;
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(16)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, strided_cn) {
+ TEST_REQUIRES_ARM_NEON;
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(16)
+ .cn_stride(11)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, k_eq_16_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(16)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, k_eq_16_subtile_m) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(m)
+ .n(8)
+ .k(16)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, k_eq_16_subtile_n) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(16)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, k_lt_16) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 1; k < 16; k++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, k_lt_16_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 1; k < 16; k++) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, k_gt_16) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 17; k < 32; k++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, k_gt_16_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 17; k < 32; k++) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, k_div_16) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 32; k <= 160; k += 16) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, k_div_16_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 32; k <= 160; k += 16) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, n_gt_8) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 9; n < 16; n++) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, n_gt_8_strided_cn) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 9; n < 16; n++) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .cn_stride(11)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, n_gt_8_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 9; n < 16; n++) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, n_div_8) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 16; n <= 24; n += 8) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, n_div_8_strided_cn) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 16; n <= 24; n += 8) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(k)
+ .cn_stride(11)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, n_div_8_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 16; n <= 24; n += 8) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, small_kernel) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, small_kernel_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 1; k <= 80; k += 17) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .ks(3)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, n_gt_8_small_kernel) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 9; n < 16; n++) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, n_div_8_small_kernel) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 16; n <= 24; n += 8) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, strided_cm_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 1; k <= 80; k += 17) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(11)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, a_offset) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .ks(3)
+ .a_offset(163)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, zero) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t mz = 0; mz < 2; mz++) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .ks(3)
+ .a_offset(163)
+ .zero_index(mz)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, qmin) {
+ TEST_REQUIRES_ARM_NEON;
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(16)
+ .qmin(128)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, qmax) {
+ TEST_REQUIRES_ARM_NEON;
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(16)
+ .qmax(128)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, strided_cm) {
+ TEST_REQUIRES_ARM_NEON;
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(16)
+ .cm_stride(11)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+#endif // XNN_ARCH_ARM64
+
+
+#if XNN_ARCH_ARM64
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, k_eq_16) {
+ TEST_REQUIRES_ARM_NEON;
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(16)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, strided_cn) {
+ TEST_REQUIRES_ARM_NEON;
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(16)
+ .cn_stride(11)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, k_eq_16_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(16)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, k_eq_16_subtile_m) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(m)
+ .n(8)
+ .k(16)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, k_eq_16_subtile_n) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(16)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, k_lt_16) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 1; k < 16; k++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, k_lt_16_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 1; k < 16; k++) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, k_gt_16) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 17; k < 32; k++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, k_gt_16_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 17; k < 32; k++) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, k_div_16) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 32; k <= 160; k += 16) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, k_div_16_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 32; k <= 160; k += 16) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, n_gt_8) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 9; n < 16; n++) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, n_gt_8_strided_cn) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 9; n < 16; n++) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .cn_stride(11)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, n_gt_8_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 9; n < 16; n++) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, n_div_8) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 16; n <= 24; n += 8) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, n_div_8_strided_cn) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 16; n <= 24; n += 8) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(n)
+ .k(k)
+ .cn_stride(11)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, n_div_8_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 16; n <= 24; n += 8) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, small_kernel) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, small_kernel_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 1; k <= 80; k += 17) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .ks(3)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, n_gt_8_small_kernel) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 9; n < 16; n++) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, n_div_8_small_kernel) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 16; n <= 24; n += 8) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, strided_cm_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 1; k <= 80; k += 17) {
+ for (uint32_t m = 1; m <= 2; m++) {
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(11)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, a_offset) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .ks(3)
+ .a_offset(163)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, zero) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t mz = 0; mz < 2; mz++) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(k)
+ .ks(3)
+ .a_offset(163)
+ .zero_index(mz)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, qmin) {
+ TEST_REQUIRES_ARM_NEON;
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(16)
+ .qmin(128)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, qmax) {
+ TEST_REQUIRES_ARM_NEON;
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(16)
+ .qmax(128)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, strided_cm) {
+ TEST_REQUIRES_ARM_NEON;
+ GemmMicrokernelTester()
+ .mr(2)
+ .nr(8)
+ .kr(16)
+ .sr(1)
+ .m(2)
+ .n(8)
+ .k(16)
+ .cm_stride(11)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+#endif // XNN_ARCH_ARM64
+
+
+#if XNN_ARCH_ARM64
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL, k_eq_16) {
+ TEST_REQUIRES_ARM_NEON;
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(1)
+ .n(8)
+ .k(16)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL, strided_cn) {
+ TEST_REQUIRES_ARM_NEON;
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(1)
+ .n(8)
+ .k(16)
+ .cn_stride(11)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL, k_eq_16_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(16)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL, k_eq_16_subtile_m) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t m = 1; m <= 1; m++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(m)
+ .n(8)
+ .k(16)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL, k_eq_16_subtile_n) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(1)
+ .n(n)
+ .k(16)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL, k_lt_16) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 1; k < 16; k++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(1)
+ .n(8)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL, k_lt_16_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 1; k < 16; k++) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL, k_gt_16) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 17; k < 32; k++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(1)
+ .n(8)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL, k_gt_16_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 17; k < 32; k++) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL, k_div_16) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 32; k <= 160; k += 16) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(1)
+ .n(8)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL, k_div_16_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 32; k <= 160; k += 16) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL, n_gt_8) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 9; n < 16; n++) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(1)
+ .n(8)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL, n_gt_8_strided_cn) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 9; n < 16; n++) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(1)
+ .n(8)
+ .k(k)
+ .cn_stride(11)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL, n_gt_8_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 9; n < 16; n++) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL, n_div_8) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 16; n <= 24; n += 8) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(1)
+ .n(8)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL, n_div_8_strided_cn) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 16; n <= 24; n += 8) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(1)
+ .n(n)
+ .k(k)
+ .cn_stride(11)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL, n_div_8_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 16; n <= 24; n += 8) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL, small_kernel) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(1)
+ .n(8)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL, small_kernel_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 1; k <= 80; k += 17) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .ks(3)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL, n_gt_8_small_kernel) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 9; n < 16; n++) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(1)
+ .n(8)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL, n_div_8_small_kernel) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 16; n <= 24; n += 8) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(1)
+ .n(8)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL, strided_cm_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 1; k <= 80; k += 17) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(11)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL, a_offset) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(1)
+ .n(8)
+ .k(k)
+ .ks(3)
+ .a_offset(83)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL, zero) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t mz = 0; mz < 1; mz++) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(1)
+ .n(8)
+ .k(k)
+ .ks(3)
+ .a_offset(83)
+ .zero_index(mz)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL, qmin) {
+ TEST_REQUIRES_ARM_NEON;
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(1)
+ .n(8)
+ .k(16)
+ .qmin(128)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL, qmax) {
+ TEST_REQUIRES_ARM_NEON;
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(1)
+ .n(8)
+ .k(16)
+ .qmax(128)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL, strided_cm) {
+ TEST_REQUIRES_ARM_NEON;
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(1)
+ .n(8)
+ .k(16)
+ .cm_stride(11)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+#endif // XNN_ARCH_ARM64
+
+
+#if XNN_ARCH_ARM64
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, k_eq_16) {
+ TEST_REQUIRES_ARM_NEON;
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(1)
+ .n(8)
+ .k(16)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, strided_cn) {
+ TEST_REQUIRES_ARM_NEON;
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(1)
+ .n(8)
+ .k(16)
+ .cn_stride(11)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, k_eq_16_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(16)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, k_eq_16_subtile_m) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t m = 1; m <= 1; m++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(m)
+ .n(8)
+ .k(16)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, k_eq_16_subtile_n) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(1)
+ .n(n)
+ .k(16)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, k_lt_16) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 1; k < 16; k++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(1)
+ .n(8)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, k_lt_16_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 1; k < 16; k++) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, k_gt_16) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 17; k < 32; k++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(1)
+ .n(8)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, k_gt_16_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 17; k < 32; k++) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, k_div_16) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 32; k <= 160; k += 16) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(1)
+ .n(8)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, k_div_16_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 32; k <= 160; k += 16) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, n_gt_8) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 9; n < 16; n++) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(1)
+ .n(8)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, n_gt_8_strided_cn) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 9; n < 16; n++) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(1)
+ .n(8)
+ .k(k)
+ .cn_stride(11)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, n_gt_8_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 9; n < 16; n++) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, n_div_8) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 16; n <= 24; n += 8) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(1)
+ .n(8)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, n_div_8_strided_cn) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 16; n <= 24; n += 8) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(1)
+ .n(n)
+ .k(k)
+ .cn_stride(11)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, n_div_8_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 16; n <= 24; n += 8) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, small_kernel) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(1)
+ .n(8)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, small_kernel_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 1; k <= 80; k += 17) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .ks(3)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, n_gt_8_small_kernel) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 9; n < 16; n++) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(1)
+ .n(8)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, n_div_8_small_kernel) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 16; n <= 24; n += 8) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(1)
+ .n(8)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, strided_cm_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 1; k <= 80; k += 17) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(11)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, a_offset) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(1)
+ .n(8)
+ .k(k)
+ .ks(3)
+ .a_offset(83)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, zero) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t mz = 0; mz < 1; mz++) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(1)
+ .n(8)
+ .k(k)
+ .ks(3)
+ .a_offset(83)
+ .zero_index(mz)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, qmin) {
+ TEST_REQUIRES_ARM_NEON;
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(1)
+ .n(8)
+ .k(16)
+ .qmin(128)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, qmax) {
+ TEST_REQUIRES_ARM_NEON;
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(1)
+ .n(8)
+ .k(16)
+ .qmax(128)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, strided_cm) {
+ TEST_REQUIRES_ARM_NEON;
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(1)
+ .n(8)
+ .k(16)
+ .cm_stride(11)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+#endif // XNN_ARCH_ARM64
+
+
+#if XNN_ARCH_ARM64
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, k_eq_16) {
+ TEST_REQUIRES_ARM_NEON;
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(1)
+ .n(8)
+ .k(16)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, strided_cn) {
+ TEST_REQUIRES_ARM_NEON;
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(1)
+ .n(8)
+ .k(16)
+ .cn_stride(11)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, k_eq_16_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(16)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, k_eq_16_subtile_m) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t m = 1; m <= 1; m++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(m)
+ .n(8)
+ .k(16)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, k_eq_16_subtile_n) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(1)
+ .n(n)
+ .k(16)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, k_lt_16) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 1; k < 16; k++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(1)
+ .n(8)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, k_lt_16_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 1; k < 16; k++) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, k_gt_16) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 17; k < 32; k++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(1)
+ .n(8)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, k_gt_16_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 17; k < 32; k++) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, k_div_16) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 32; k <= 160; k += 16) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(1)
+ .n(8)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, k_div_16_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 32; k <= 160; k += 16) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, n_gt_8) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 9; n < 16; n++) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(1)
+ .n(8)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, n_gt_8_strided_cn) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 9; n < 16; n++) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(1)
+ .n(8)
+ .k(k)
+ .cn_stride(11)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, n_gt_8_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 9; n < 16; n++) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, n_div_8) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 16; n <= 24; n += 8) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(1)
+ .n(8)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, n_div_8_strided_cn) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 16; n <= 24; n += 8) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(1)
+ .n(n)
+ .k(k)
+ .cn_stride(11)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, n_div_8_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 16; n <= 24; n += 8) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, small_kernel) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(1)
+ .n(8)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, small_kernel_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 1; k <= 80; k += 17) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .ks(3)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, n_gt_8_small_kernel) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 9; n < 16; n++) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(1)
+ .n(8)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, n_div_8_small_kernel) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 16; n <= 24; n += 8) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(1)
+ .n(8)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, strided_cm_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 1; k <= 80; k += 17) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(11)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, a_offset) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(1)
+ .n(8)
+ .k(k)
+ .ks(3)
+ .a_offset(83)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, zero) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t mz = 0; mz < 1; mz++) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(1)
+ .n(8)
+ .k(k)
+ .ks(3)
+ .a_offset(83)
+ .zero_index(mz)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, qmin) {
+ TEST_REQUIRES_ARM_NEON;
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(1)
+ .n(8)
+ .k(16)
+ .qmin(128)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, qmax) {
+ TEST_REQUIRES_ARM_NEON;
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(1)
+ .n(8)
+ .k(16)
+ .qmax(128)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, strided_cm) {
+ TEST_REQUIRES_ARM_NEON;
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(1)
+ .n(8)
+ .k(16)
+ .cm_stride(11)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+#endif // XNN_ARCH_ARM64
+
+
+#if XNN_ARCH_ARM64
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, k_eq_16) {
+ TEST_REQUIRES_ARM_NEON;
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(1)
+ .n(8)
+ .k(16)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, strided_cn) {
+ TEST_REQUIRES_ARM_NEON;
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(1)
+ .n(8)
+ .k(16)
+ .cn_stride(11)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, k_eq_16_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(16)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, k_eq_16_subtile_m) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t m = 1; m <= 1; m++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(m)
+ .n(8)
+ .k(16)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, k_eq_16_subtile_n) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(1)
+ .n(n)
+ .k(16)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, k_lt_16) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 1; k < 16; k++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(1)
+ .n(8)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, k_lt_16_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 1; k < 16; k++) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, k_gt_16) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 17; k < 32; k++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(1)
+ .n(8)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, k_gt_16_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 17; k < 32; k++) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, k_div_16) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 32; k <= 160; k += 16) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(1)
+ .n(8)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, k_div_16_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 32; k <= 160; k += 16) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, n_gt_8) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 9; n < 16; n++) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(1)
+ .n(8)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, n_gt_8_strided_cn) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 9; n < 16; n++) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(1)
+ .n(8)
+ .k(k)
+ .cn_stride(11)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, n_gt_8_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 9; n < 16; n++) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, n_div_8) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 16; n <= 24; n += 8) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(1)
+ .n(8)
+ .k(k)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, n_div_8_strided_cn) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 16; n <= 24; n += 8) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(1)
+ .n(n)
+ .k(k)
+ .cn_stride(11)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, n_div_8_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 16; n <= 24; n += 8) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, small_kernel) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(1)
+ .n(8)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, small_kernel_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 1; k <= 80; k += 17) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .ks(3)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, n_gt_8_small_kernel) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 9; n < 16; n++) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(1)
+ .n(8)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, n_div_8_small_kernel) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t n = 16; n <= 24; n += 8) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(1)
+ .n(8)
+ .k(k)
+ .ks(3)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, strided_cm_subtile) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 1; k <= 80; k += 17) {
+ for (uint32_t m = 1; m <= 1; m++) {
+ for (uint32_t n = 1; n <= 8; n++) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(m)
+ .n(n)
+ .k(k)
+ .cm_stride(11)
+ .iterations(1)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, a_offset) {
+ TEST_REQUIRES_ARM_NEON;
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(1)
+ .n(8)
+ .k(k)
+ .ks(3)
+ .a_offset(83)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, zero) {
+ TEST_REQUIRES_ARM_NEON;
+ for (uint32_t mz = 0; mz < 1; mz++) {
+ for (size_t k = 1; k <= 80; k += 17) {
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(1)
+ .n(8)
+ .k(k)
+ .ks(3)
+ .a_offset(83)
+ .zero_index(mz)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+ }
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, qmin) {
+ TEST_REQUIRES_ARM_NEON;
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(1)
+ .n(8)
+ .k(16)
+ .qmin(128)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, qmax) {
+ TEST_REQUIRES_ARM_NEON;
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(1)
+ .n(8)
+ .k(16)
+ .qmax(128)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+
+ TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, strided_cm) {
+ TEST_REQUIRES_ARM_NEON;
+ GemmMicrokernelTester()
+ .mr(1)
+ .nr(8)
+ .kr(8)
+ .sr(1)
+ .m(1)
+ .n(8)
+ .k(16)
+ .cm_stride(11)
+ .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
+ }
+#endif // XNN_ARCH_ARM64
+
+
+#if XNN_ARCH_ARM64
TEST(QS8_IGEMM_MINMAX_GEMMLOWP_4X16__AARCH64_NEON_MLAL_LANE_PRFM_CORTEX_A53, k_eq_8) {
TEST_REQUIRES_ARM_NEON;
GemmMicrokernelTester()
@@ -2362,4218 +6574,6 @@
#endif // XNN_ARCH_ARM64
-#if XNN_ARCH_ARM64
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL, k_eq_16) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(16)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL, strided_cn) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(16)
- .cn_stride(11)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL, k_eq_16_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t m = 1; m <= 1; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(16)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL, k_eq_16_subtile_m) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t m = 1; m <= 1; m++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(8)
- .k(16)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL, k_eq_16_subtile_n) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(n)
- .k(16)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL, k_lt_16) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 1; k < 16; k++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(k)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL, k_lt_16_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 1; k < 16; k++) {
- for (uint32_t m = 1; m <= 1; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL, k_gt_16) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 17; k < 32; k++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(k)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL, k_gt_16_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 17; k < 32; k++) {
- for (uint32_t m = 1; m <= 1; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL, k_div_16) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 32; k <= 160; k += 16) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(k)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL, k_div_16_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 32; k <= 160; k += 16) {
- for (uint32_t m = 1; m <= 1; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL, n_gt_8) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 9; n < 16; n++) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(k)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL, n_gt_8_strided_cn) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 9; n < 16; n++) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(k)
- .cn_stride(11)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL, n_gt_8_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 9; n < 16; n++) {
- for (size_t k = 1; k <= 80; k += 17) {
- for (uint32_t m = 1; m <= 1; m++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL, n_div_8) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 16; n <= 24; n += 8) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(k)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL, n_div_8_strided_cn) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 16; n <= 24; n += 8) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(n)
- .k(k)
- .cn_stride(11)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL, n_div_8_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 16; n <= 24; n += 8) {
- for (size_t k = 1; k <= 80; k += 17) {
- for (uint32_t m = 1; m <= 1; m++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL, small_kernel) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(k)
- .ks(3)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL, small_kernel_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 1; k <= 80; k += 17) {
- for (uint32_t m = 1; m <= 1; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .ks(3)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL, n_gt_8_small_kernel) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 9; n < 16; n++) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(k)
- .ks(3)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL, n_div_8_small_kernel) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 16; n <= 24; n += 8) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(k)
- .ks(3)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL, strided_cm_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 1; k <= 80; k += 17) {
- for (uint32_t m = 1; m <= 1; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .cm_stride(11)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL, a_offset) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(k)
- .ks(3)
- .a_offset(83)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL, zero) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t mz = 0; mz < 1; mz++) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(k)
- .ks(3)
- .a_offset(83)
- .zero_index(mz)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL, qmin) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(16)
- .qmin(128)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL, qmax) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(16)
- .qmax(128)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL, strided_cm) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(16)
- .cm_stride(11)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-#endif // XNN_ARCH_ARM64
-
-
-#if XNN_ARCH_ARM64
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, k_eq_16) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(16)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, strided_cn) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(16)
- .cn_stride(11)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, k_eq_16_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t m = 1; m <= 1; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(16)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, k_eq_16_subtile_m) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t m = 1; m <= 1; m++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(8)
- .k(16)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, k_eq_16_subtile_n) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(n)
- .k(16)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, k_lt_16) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 1; k < 16; k++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(k)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, k_lt_16_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 1; k < 16; k++) {
- for (uint32_t m = 1; m <= 1; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, k_gt_16) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 17; k < 32; k++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(k)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, k_gt_16_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 17; k < 32; k++) {
- for (uint32_t m = 1; m <= 1; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, k_div_16) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 32; k <= 160; k += 16) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(k)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, k_div_16_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 32; k <= 160; k += 16) {
- for (uint32_t m = 1; m <= 1; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, n_gt_8) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 9; n < 16; n++) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(k)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, n_gt_8_strided_cn) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 9; n < 16; n++) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(k)
- .cn_stride(11)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, n_gt_8_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 9; n < 16; n++) {
- for (size_t k = 1; k <= 80; k += 17) {
- for (uint32_t m = 1; m <= 1; m++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, n_div_8) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 16; n <= 24; n += 8) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(k)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, n_div_8_strided_cn) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 16; n <= 24; n += 8) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(n)
- .k(k)
- .cn_stride(11)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, n_div_8_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 16; n <= 24; n += 8) {
- for (size_t k = 1; k <= 80; k += 17) {
- for (uint32_t m = 1; m <= 1; m++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, small_kernel) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(k)
- .ks(3)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, small_kernel_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 1; k <= 80; k += 17) {
- for (uint32_t m = 1; m <= 1; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .ks(3)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, n_gt_8_small_kernel) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 9; n < 16; n++) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(k)
- .ks(3)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, n_div_8_small_kernel) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 16; n <= 24; n += 8) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(k)
- .ks(3)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, strided_cm_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 1; k <= 80; k += 17) {
- for (uint32_t m = 1; m <= 1; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .cm_stride(11)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, a_offset) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(k)
- .ks(3)
- .a_offset(83)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, zero) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t mz = 0; mz < 1; mz++) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(k)
- .ks(3)
- .a_offset(83)
- .zero_index(mz)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, qmin) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(16)
- .qmin(128)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, qmax) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(16)
- .qmax(128)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, strided_cm) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(16)
- .cm_stride(11)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-#endif // XNN_ARCH_ARM64
-
-
-#if XNN_ARCH_ARM64
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, k_eq_16) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(16)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, strided_cn) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(16)
- .cn_stride(11)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, k_eq_16_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t m = 1; m <= 1; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(16)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, k_eq_16_subtile_m) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t m = 1; m <= 1; m++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(8)
- .k(16)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, k_eq_16_subtile_n) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(n)
- .k(16)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, k_lt_16) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 1; k < 16; k++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(k)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, k_lt_16_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 1; k < 16; k++) {
- for (uint32_t m = 1; m <= 1; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, k_gt_16) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 17; k < 32; k++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(k)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, k_gt_16_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 17; k < 32; k++) {
- for (uint32_t m = 1; m <= 1; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, k_div_16) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 32; k <= 160; k += 16) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(k)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, k_div_16_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 32; k <= 160; k += 16) {
- for (uint32_t m = 1; m <= 1; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, n_gt_8) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 9; n < 16; n++) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(k)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, n_gt_8_strided_cn) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 9; n < 16; n++) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(k)
- .cn_stride(11)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, n_gt_8_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 9; n < 16; n++) {
- for (size_t k = 1; k <= 80; k += 17) {
- for (uint32_t m = 1; m <= 1; m++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, n_div_8) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 16; n <= 24; n += 8) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(k)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, n_div_8_strided_cn) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 16; n <= 24; n += 8) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(n)
- .k(k)
- .cn_stride(11)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, n_div_8_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 16; n <= 24; n += 8) {
- for (size_t k = 1; k <= 80; k += 17) {
- for (uint32_t m = 1; m <= 1; m++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, small_kernel) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(k)
- .ks(3)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, small_kernel_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 1; k <= 80; k += 17) {
- for (uint32_t m = 1; m <= 1; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .ks(3)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, n_gt_8_small_kernel) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 9; n < 16; n++) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(k)
- .ks(3)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, n_div_8_small_kernel) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 16; n <= 24; n += 8) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(k)
- .ks(3)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, strided_cm_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 1; k <= 80; k += 17) {
- for (uint32_t m = 1; m <= 1; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .cm_stride(11)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, a_offset) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(k)
- .ks(3)
- .a_offset(83)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, zero) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t mz = 0; mz < 1; mz++) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(k)
- .ks(3)
- .a_offset(83)
- .zero_index(mz)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, qmin) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(16)
- .qmin(128)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, qmax) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(16)
- .qmax(128)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, strided_cm) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(16)
- .cm_stride(11)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-#endif // XNN_ARCH_ARM64
-
-
-#if XNN_ARCH_ARM64
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, k_eq_16) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(16)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, strided_cn) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(16)
- .cn_stride(11)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, k_eq_16_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t m = 1; m <= 1; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(16)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, k_eq_16_subtile_m) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t m = 1; m <= 1; m++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(8)
- .k(16)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, k_eq_16_subtile_n) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(n)
- .k(16)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, k_lt_16) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 1; k < 16; k++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(k)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, k_lt_16_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 1; k < 16; k++) {
- for (uint32_t m = 1; m <= 1; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, k_gt_16) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 17; k < 32; k++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(k)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, k_gt_16_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 17; k < 32; k++) {
- for (uint32_t m = 1; m <= 1; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, k_div_16) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 32; k <= 160; k += 16) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(k)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, k_div_16_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 32; k <= 160; k += 16) {
- for (uint32_t m = 1; m <= 1; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, n_gt_8) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 9; n < 16; n++) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(k)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, n_gt_8_strided_cn) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 9; n < 16; n++) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(k)
- .cn_stride(11)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, n_gt_8_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 9; n < 16; n++) {
- for (size_t k = 1; k <= 80; k += 17) {
- for (uint32_t m = 1; m <= 1; m++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, n_div_8) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 16; n <= 24; n += 8) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(k)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, n_div_8_strided_cn) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 16; n <= 24; n += 8) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(n)
- .k(k)
- .cn_stride(11)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, n_div_8_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 16; n <= 24; n += 8) {
- for (size_t k = 1; k <= 80; k += 17) {
- for (uint32_t m = 1; m <= 1; m++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, small_kernel) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(k)
- .ks(3)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, small_kernel_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 1; k <= 80; k += 17) {
- for (uint32_t m = 1; m <= 1; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .ks(3)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, n_gt_8_small_kernel) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 9; n < 16; n++) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(k)
- .ks(3)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, n_div_8_small_kernel) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 16; n <= 24; n += 8) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(k)
- .ks(3)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, strided_cm_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 1; k <= 80; k += 17) {
- for (uint32_t m = 1; m <= 1; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .cm_stride(11)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, a_offset) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(k)
- .ks(3)
- .a_offset(83)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, zero) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t mz = 0; mz < 1; mz++) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(k)
- .ks(3)
- .a_offset(83)
- .zero_index(mz)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, qmin) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(16)
- .qmin(128)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, qmax) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(16)
- .qmax(128)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, strided_cm) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(1)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(1)
- .n(8)
- .k(16)
- .cm_stride(11)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-#endif // XNN_ARCH_ARM64
-
-
-#if XNN_ARCH_ARM64
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, k_eq_16) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(16)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, strided_cn) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(16)
- .cn_stride(11)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, k_eq_16_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(16)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, k_eq_16_subtile_m) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t m = 1; m <= 2; m++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(8)
- .k(16)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, k_eq_16_subtile_n) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(n)
- .k(16)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, k_lt_16) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 1; k < 16; k++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, k_lt_16_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 1; k < 16; k++) {
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, k_gt_16) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 17; k < 32; k++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, k_gt_16_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 17; k < 32; k++) {
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, k_div_16) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 32; k <= 160; k += 16) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, k_div_16_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 32; k <= 160; k += 16) {
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, n_gt_8) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 9; n < 16; n++) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, n_gt_8_strided_cn) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 9; n < 16; n++) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .cn_stride(11)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, n_gt_8_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 9; n < 16; n++) {
- for (size_t k = 1; k <= 80; k += 17) {
- for (uint32_t m = 1; m <= 2; m++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, n_div_8) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 16; n <= 24; n += 8) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, n_div_8_strided_cn) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 16; n <= 24; n += 8) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(n)
- .k(k)
- .cn_stride(11)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, n_div_8_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 16; n <= 24; n += 8) {
- for (size_t k = 1; k <= 80; k += 17) {
- for (uint32_t m = 1; m <= 2; m++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, small_kernel) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .ks(3)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, small_kernel_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 1; k <= 80; k += 17) {
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .ks(3)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, n_gt_8_small_kernel) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 9; n < 16; n++) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .ks(3)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, n_div_8_small_kernel) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 16; n <= 24; n += 8) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .ks(3)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, strided_cm_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 1; k <= 80; k += 17) {
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .cm_stride(11)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, a_offset) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .ks(3)
- .a_offset(163)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, zero) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t mz = 0; mz < 2; mz++) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .ks(3)
- .a_offset(163)
- .zero_index(mz)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, qmin) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(16)
- .qmin(128)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, qmax) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(16)
- .qmax(128)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_CORTEX_A53, strided_cm) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(16)
- .cm_stride(11)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-#endif // XNN_ARCH_ARM64
-
-
-#if XNN_ARCH_ARM64
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, k_eq_16) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(16)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, strided_cn) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(16)
- .cn_stride(11)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, k_eq_16_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(16)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, k_eq_16_subtile_m) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t m = 1; m <= 2; m++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(8)
- .k(16)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, k_eq_16_subtile_n) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(n)
- .k(16)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, k_lt_16) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 1; k < 16; k++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, k_lt_16_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 1; k < 16; k++) {
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, k_gt_16) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 17; k < 32; k++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, k_gt_16_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 17; k < 32; k++) {
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, k_div_16) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 32; k <= 160; k += 16) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, k_div_16_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 32; k <= 160; k += 16) {
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, n_gt_8) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 9; n < 16; n++) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, n_gt_8_strided_cn) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 9; n < 16; n++) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .cn_stride(11)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, n_gt_8_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 9; n < 16; n++) {
- for (size_t k = 1; k <= 80; k += 17) {
- for (uint32_t m = 1; m <= 2; m++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, n_div_8) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 16; n <= 24; n += 8) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, n_div_8_strided_cn) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 16; n <= 24; n += 8) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(n)
- .k(k)
- .cn_stride(11)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, n_div_8_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 16; n <= 24; n += 8) {
- for (size_t k = 1; k <= 80; k += 17) {
- for (uint32_t m = 1; m <= 2; m++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, small_kernel) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .ks(3)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, small_kernel_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 1; k <= 80; k += 17) {
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .ks(3)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, n_gt_8_small_kernel) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 9; n < 16; n++) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .ks(3)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, n_div_8_small_kernel) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 16; n <= 24; n += 8) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .ks(3)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, strided_cm_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 1; k <= 80; k += 17) {
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .cm_stride(11)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, a_offset) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .ks(3)
- .a_offset(163)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, zero) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t mz = 0; mz < 2; mz++) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .ks(3)
- .a_offset(163)
- .zero_index(mz)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, qmin) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(16)
- .qmin(128)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, qmax) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(16)
- .qmax(128)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM_CORTEX_A53, strided_cm) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(16)
- .cm_stride(11)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-#endif // XNN_ARCH_ARM64
-
-
-#if XNN_ARCH_ARM64
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL, k_eq_16) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(16)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL, strided_cn) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(16)
- .cn_stride(11)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL, k_eq_16_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(16)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL, k_eq_16_subtile_m) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t m = 1; m <= 2; m++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(8)
- .k(16)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL, k_eq_16_subtile_n) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(n)
- .k(16)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL, k_lt_16) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 1; k < 16; k++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL, k_lt_16_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 1; k < 16; k++) {
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL, k_gt_16) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 17; k < 32; k++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL, k_gt_16_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 17; k < 32; k++) {
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL, k_div_16) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 32; k <= 160; k += 16) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL, k_div_16_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 32; k <= 160; k += 16) {
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL, n_gt_8) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 9; n < 16; n++) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL, n_gt_8_strided_cn) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 9; n < 16; n++) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .cn_stride(11)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL, n_gt_8_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 9; n < 16; n++) {
- for (size_t k = 1; k <= 80; k += 17) {
- for (uint32_t m = 1; m <= 2; m++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL, n_div_8) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 16; n <= 24; n += 8) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL, n_div_8_strided_cn) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 16; n <= 24; n += 8) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(n)
- .k(k)
- .cn_stride(11)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL, n_div_8_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 16; n <= 24; n += 8) {
- for (size_t k = 1; k <= 80; k += 17) {
- for (uint32_t m = 1; m <= 2; m++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL, small_kernel) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .ks(3)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL, small_kernel_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 1; k <= 80; k += 17) {
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .ks(3)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL, n_gt_8_small_kernel) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 9; n < 16; n++) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .ks(3)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL, n_div_8_small_kernel) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 16; n <= 24; n += 8) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .ks(3)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL, strided_cm_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 1; k <= 80; k += 17) {
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .cm_stride(11)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL, a_offset) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .ks(3)
- .a_offset(163)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL, zero) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t mz = 0; mz < 2; mz++) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .ks(3)
- .a_offset(163)
- .zero_index(mz)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL, qmin) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(16)
- .qmin(128)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL, qmax) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(16)
- .qmax(128)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL, strided_cm) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(16)
- .cm_stride(11)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-#endif // XNN_ARCH_ARM64
-
-
-#if XNN_ARCH_ARM64
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, k_eq_16) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(16)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, strided_cn) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(16)
- .cn_stride(11)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, k_eq_16_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(16)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, k_eq_16_subtile_m) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t m = 1; m <= 2; m++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(8)
- .k(16)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, k_eq_16_subtile_n) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(n)
- .k(16)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, k_lt_16) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 1; k < 16; k++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, k_lt_16_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 1; k < 16; k++) {
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, k_gt_16) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 17; k < 32; k++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, k_gt_16_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 17; k < 32; k++) {
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, k_div_16) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 32; k <= 160; k += 16) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, k_div_16_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 32; k <= 160; k += 16) {
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, n_gt_8) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 9; n < 16; n++) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, n_gt_8_strided_cn) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 9; n < 16; n++) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .cn_stride(11)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, n_gt_8_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 9; n < 16; n++) {
- for (size_t k = 1; k <= 80; k += 17) {
- for (uint32_t m = 1; m <= 2; m++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, n_div_8) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 16; n <= 24; n += 8) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, n_div_8_strided_cn) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 16; n <= 24; n += 8) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(n)
- .k(k)
- .cn_stride(11)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, n_div_8_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 16; n <= 24; n += 8) {
- for (size_t k = 1; k <= 80; k += 17) {
- for (uint32_t m = 1; m <= 2; m++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, small_kernel) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .ks(3)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, small_kernel_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 1; k <= 80; k += 17) {
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .ks(3)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, n_gt_8_small_kernel) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 9; n < 16; n++) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .ks(3)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, n_div_8_small_kernel) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 16; n <= 24; n += 8) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .ks(3)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, strided_cm_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 1; k <= 80; k += 17) {
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .cm_stride(11)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, a_offset) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .ks(3)
- .a_offset(163)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, zero) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t mz = 0; mz < 2; mz++) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .ks(3)
- .a_offset(163)
- .zero_index(mz)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, qmin) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(16)
- .qmin(128)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, qmax) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(16)
- .qmax(128)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C8__AARCH64_NEON_MLAL_PADAL_PRFM, strided_cm) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(8)
- .sr(1)
- .m(2)
- .n(8)
- .k(16)
- .cm_stride(11)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-#endif // XNN_ARCH_ARM64
-
-
-#if XNN_ARCH_ARM64
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, k_eq_16) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(16)
- .sr(1)
- .m(2)
- .n(8)
- .k(16)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, strided_cn) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(16)
- .sr(1)
- .m(2)
- .n(8)
- .k(16)
- .cn_stride(11)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, k_eq_16_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(16)
- .sr(1)
- .m(m)
- .n(n)
- .k(16)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, k_eq_16_subtile_m) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t m = 1; m <= 2; m++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(16)
- .sr(1)
- .m(m)
- .n(8)
- .k(16)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, k_eq_16_subtile_n) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(16)
- .sr(1)
- .m(2)
- .n(n)
- .k(16)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, k_lt_16) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 1; k < 16; k++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(16)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, k_lt_16_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 1; k < 16; k++) {
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(16)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, k_gt_16) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 17; k < 32; k++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(16)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, k_gt_16_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 17; k < 32; k++) {
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(16)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, k_div_16) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 32; k <= 160; k += 16) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(16)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, k_div_16_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 32; k <= 160; k += 16) {
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(16)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, n_gt_8) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 9; n < 16; n++) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(16)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, n_gt_8_strided_cn) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 9; n < 16; n++) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(16)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .cn_stride(11)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, n_gt_8_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 9; n < 16; n++) {
- for (size_t k = 1; k <= 80; k += 17) {
- for (uint32_t m = 1; m <= 2; m++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(16)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, n_div_8) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 16; n <= 24; n += 8) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(16)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, n_div_8_strided_cn) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 16; n <= 24; n += 8) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(16)
- .sr(1)
- .m(2)
- .n(n)
- .k(k)
- .cn_stride(11)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, n_div_8_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 16; n <= 24; n += 8) {
- for (size_t k = 1; k <= 80; k += 17) {
- for (uint32_t m = 1; m <= 2; m++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(16)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, small_kernel) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(16)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .ks(3)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, small_kernel_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 1; k <= 80; k += 17) {
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(16)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .ks(3)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, n_gt_8_small_kernel) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 9; n < 16; n++) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(16)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .ks(3)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, n_div_8_small_kernel) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t n = 16; n <= 24; n += 8) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(16)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .ks(3)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, strided_cm_subtile) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 1; k <= 80; k += 17) {
- for (uint32_t m = 1; m <= 2; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(16)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .cm_stride(11)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, a_offset) {
- TEST_REQUIRES_ARM_NEON;
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(16)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .ks(3)
- .a_offset(163)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, zero) {
- TEST_REQUIRES_ARM_NEON;
- for (uint32_t mz = 0; mz < 2; mz++) {
- for (size_t k = 1; k <= 80; k += 17) {
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(16)
- .sr(1)
- .m(2)
- .n(8)
- .k(k)
- .ks(3)
- .a_offset(163)
- .zero_index(mz)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, qmin) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(16)
- .sr(1)
- .m(2)
- .n(8)
- .k(16)
- .qmin(128)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, qmax) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(16)
- .sr(1)
- .m(2)
- .n(8)
- .k(16)
- .qmax(128)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-
- TEST(QS8_IGEMM_MINMAX_GEMMLOWP_2X8C16__AARCH64_NEON_MLAL_PADAL, strided_cm) {
- TEST_REQUIRES_ARM_NEON;
- GemmMicrokernelTester()
- .mr(2)
- .nr(8)
- .kr(16)
- .sr(1)
- .m(2)
- .n(8)
- .k(16)
- .cm_stride(11)
- .Test(xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal, xnn_init_qs8_conv_minmax_gemmlowp_neon_params, xnn_init_qs8_requantization_gemmlowp_params, xnn_qs8_requantize_gemmlowp);
- }
-#endif // XNN_ARCH_ARM64
-
-
#if XNN_ARCH_ARM || XNN_ARCH_ARM64
TEST(QS8_IGEMM_MINMAX_GEMMLOWP_1X8__NEON_MLAL_LANE, k_eq_8) {
TEST_REQUIRES_ARM_NEON;
diff --git a/test/qs8-igemm-minmax-gemmlowp.yaml b/test/qs8-igemm-minmax-gemmlowp.yaml
index 5cc6a7d..d516a1d 100644
--- a/test/qs8-igemm-minmax-gemmlowp.yaml
+++ b/test/qs8-igemm-minmax-gemmlowp.yaml
@@ -3,19 +3,19 @@
# This source code is licensed under the BSD-style license found in the
# LICENSE file in the root directory of this source tree.
-- name: xnn_qs8_igemm_minmax_gemmlowp_ukernel_4x16__aarch64_neon_mlal_lane_prfm_cortex_a53
- init: xnn_init_qs8_conv_minmax_gemmlowp_neon_params
- k-block: 8
-- name: xnn_qs8_igemm_minmax_gemmlowp_ukernel_4x16__aarch64_neon_mlal_lane_cortex_a53
- init: xnn_init_qs8_conv_minmax_gemmlowp_neon_params
- k-block: 8
-- name: xnn_qs8_igemm_minmax_gemmlowp_ukernel_4x16c4__aarch64_neondot_cortex_a55
+- name: xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal
init: xnn_init_qs8_conv_minmax_gemmlowp_neon_params
k-block: 16
-- name: xnn_qs8_igemm_minmax_gemmlowp_ukernel_4x16c4__aarch64_neondot_ld64
+- name: xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm
init: xnn_init_qs8_conv_minmax_gemmlowp_neon_params
- k-block: 8
-- name: xnn_qs8_igemm_minmax_gemmlowp_ukernel_4x16c4__aarch64_neondot_ld128
+ k-block: 16
+- name: xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_cortex_a53
+ init: xnn_init_qs8_conv_minmax_gemmlowp_neon_params
+ k-block: 16
+- name: xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53
+ init: xnn_init_qs8_conv_minmax_gemmlowp_neon_params
+ k-block: 16
+- name: xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal
init: xnn_init_qs8_conv_minmax_gemmlowp_neon_params
k-block: 16
- name: xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal
@@ -30,19 +30,19 @@
- name: xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53
init: xnn_init_qs8_conv_minmax_gemmlowp_neon_params
k-block: 16
-- name: xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_cortex_a53
+- name: xnn_qs8_igemm_minmax_gemmlowp_ukernel_4x16__aarch64_neon_mlal_lane_prfm_cortex_a53
+ init: xnn_init_qs8_conv_minmax_gemmlowp_neon_params
+ k-block: 8
+- name: xnn_qs8_igemm_minmax_gemmlowp_ukernel_4x16__aarch64_neon_mlal_lane_cortex_a53
+ init: xnn_init_qs8_conv_minmax_gemmlowp_neon_params
+ k-block: 8
+- name: xnn_qs8_igemm_minmax_gemmlowp_ukernel_4x16c4__aarch64_neondot_cortex_a55
init: xnn_init_qs8_conv_minmax_gemmlowp_neon_params
k-block: 16
-- name: xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm_cortex_a53
+- name: xnn_qs8_igemm_minmax_gemmlowp_ukernel_4x16c4__aarch64_neondot_ld64
init: xnn_init_qs8_conv_minmax_gemmlowp_neon_params
- k-block: 16
-- name: xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal
- init: xnn_init_qs8_conv_minmax_gemmlowp_neon_params
- k-block: 16
-- name: xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c8__aarch64_neon_mlal_padal_prfm
- init: xnn_init_qs8_conv_minmax_gemmlowp_neon_params
- k-block: 16
-- name: xnn_qs8_igemm_minmax_gemmlowp_ukernel_2x8c16__aarch64_neon_mlal_padal
+ k-block: 8
+- name: xnn_qs8_igemm_minmax_gemmlowp_ukernel_4x16c4__aarch64_neondot_ld128
init: xnn_init_qs8_conv_minmax_gemmlowp_neon_params
k-block: 16
- name: xnn_qs8_igemm_minmax_gemmlowp_ukernel_1x8__neon_mlal_lane