S8 MAXPOOL microkernels for all architectures
- S8 MAXPOOL microkernels for SSE2, SSE4.1, NEON, WAsm SIMD, and scalar
architectures
- Parameters and parameter initialization functions
- Unit tests
PiperOrigin-RevId: 391040942
diff --git a/BUILD.bazel b/BUILD.bazel
index 5a8f845..bd79371 100644
--- a/BUILD.bazel
+++ b/BUILD.bazel
@@ -281,6 +281,7 @@
"src/qu8-vaddc/gen/minmax-scalar-x4.c",
"src/qu8-vmul/gen/minmax-fp32-scalar-x4.c",
"src/qu8-vmulc/gen/minmax-fp32-scalar-x4.c",
+ "src/s8-maxpool/9p8x-minmax-scalar-c1.c",
"src/u8-lut32norm/scalar.c",
"src/u8-maxpool/9p8x-minmax-scalar-c1.c",
"src/u8-rmax/scalar.c",
@@ -891,6 +892,7 @@
"src/qu8-vmulc/gen/minmax-fp32-scalar-x1.c",
"src/qu8-vmulc/gen/minmax-fp32-scalar-x2.c",
"src/qu8-vmulc/gen/minmax-fp32-scalar-x4.c",
+ "src/s8-maxpool/9p8x-minmax-scalar-c1.c",
"src/u8-lut32norm/scalar.c",
"src/u8-maxpool/9p8x-minmax-scalar-c1.c",
"src/u8-rmax/scalar.c",
@@ -1830,6 +1832,7 @@
"src/qu8-vmul/gen/minmax-fp32-wasmsimd-mul32-ld64-x16.c",
"src/qu8-vmulc/gen/minmax-fp32-wasmsimd-mul32-ld64-x8.c",
"src/qu8-vmulc/gen/minmax-fp32-wasmsimd-mul32-ld64-x16.c",
+ "src/s8-maxpool/9p8x-minmax-wasmsimd-c16.c",
"src/u8-maxpool/9p8x-minmax-wasmsimd-c16.c",
"src/x32-packx/x4-wasmsimd.c",
"src/x32-unpool/wasmsimd.c",
@@ -1940,6 +1943,7 @@
"src/qu8-vaddc/gen/minmax-neon-ld64-x8.c",
"src/qu8-vmul/gen/minmax-fp32-neon-ld64-x16.c",
"src/qu8-vmulc/gen/minmax-fp32-neon-ld64-x16.c",
+ "src/s8-maxpool/9p8x-minmax-neon-c16.c",
"src/u8-maxpool/9p8x-minmax-neon-c16.c",
"src/u8-rmax/neon.c",
"src/u8-vclamp/neon-x64.c",
@@ -2511,6 +2515,7 @@
"src/qu8-vmulc/gen/minmax-fp32-neon-ld64-x8.c",
"src/qu8-vmulc/gen/minmax-fp32-neon-ld64-x16.c",
"src/qu8-vmulc/gen/minmax-fp32-neon-ld128-x16.c",
+ "src/s8-maxpool/9p8x-minmax-neon-c16.c",
"src/u8-maxpool/9p8x-minmax-neon-c16.c",
"src/u8-rmax/neon.c",
"src/u8-vclamp/neon-x64.c",
@@ -3471,6 +3476,7 @@
"src/qu8-vaddc/gen/minmax-sse2-mul16-ld64-x8.c",
"src/qu8-vmul/gen/minmax-fp32-sse2-mul16-ld64-x8.c",
"src/qu8-vmulc/gen/minmax-fp32-sse2-mul16-ld64-x8.c",
+ "src/s8-maxpool/9p8x-minmax-sse2-c16.c",
"src/u8-maxpool/9p8x-minmax-sse2-c16.c",
"src/u8-rmax/sse2.c",
"src/u8-vclamp/sse2-x64.c",
@@ -3731,6 +3737,7 @@
"src/qu8-vmul/gen/minmax-fp32-sse2-mul16-ld64-x16.c",
"src/qu8-vmulc/gen/minmax-fp32-sse2-mul16-ld64-x8.c",
"src/qu8-vmulc/gen/minmax-fp32-sse2-mul16-ld64-x16.c",
+ "src/s8-maxpool/9p8x-minmax-sse2-c16.c",
"src/u8-maxpool/9p8x-minmax-sse2-c16.c",
"src/u8-rmax/sse2.c",
"src/u8-vclamp/sse2-x64.c",
@@ -3842,6 +3849,7 @@
"src/qu8-vaddc/gen/minmax-sse41-mul16-ld64-x8.c",
"src/qu8-vmul/gen/minmax-fp32-sse41-mul16-ld64-x16.c",
"src/qu8-vmulc/gen/minmax-fp32-sse41-mul16-ld64-x16.c",
+ "src/s8-maxpool/9p8x-minmax-sse41-c16.c",
]
ALL_SSE41_MICROKERNEL_SRCS = [
@@ -4081,6 +4089,7 @@
"src/qu8-vmul/gen/minmax-fp32-sse41-mul16-ld64-x16.c",
"src/qu8-vmulc/gen/minmax-fp32-sse41-mul16-ld64-x8.c",
"src/qu8-vmulc/gen/minmax-fp32-sse41-mul16-ld64-x16.c",
+ "src/s8-maxpool/9p8x-minmax-sse41-c16.c",
]
PROD_AVX_MICROKERNEL_SRCS = [
@@ -7158,6 +7167,7 @@
"//conditions:default": [],
}),
defines = [
+ "XNN_NO_S8_OPERATORS",
"XNN_NO_U8_OPERATORS",
"XNN_NO_F16_OPERATORS",
"XNN_NO_X16_OPERATORS",
@@ -7222,6 +7232,7 @@
defines = [
"XNN_NO_QS8_OPERATORS",
"XNN_NO_QU8_OPERATORS",
+ "XNN_NO_S8_OPERATORS",
"XNN_NO_U8_OPERATORS",
"XNN_NO_X8_OPERATORS",
"XNN_NO_NCHW_OPERATORS",
@@ -9545,6 +9556,15 @@
)
xnnpack_unit_test(
+ name = "s8_maxpool_minmax_test",
+ srcs = [
+ "test/s8-maxpool-minmax.cc",
+ "test/maxpool-microkernel-tester.h",
+ ] + MICROKERNEL_TEST_HDRS,
+ deps = MICROKERNEL_TEST_DEPS,
+)
+
+xnnpack_unit_test(
name = "u8_lut32norm_test",
srcs = [
"test/u8-lut32norm.cc",