Remove 12x8 QS8 GEMM and IGEMM Neon dotproduct microkernels.
PiperOrigin-RevId: 361204308
diff --git a/BUILD.bazel b/BUILD.bazel
index fd27340..dc22870 100644
--- a/BUILD.bazel
+++ b/BUILD.bazel
@@ -2276,7 +2276,6 @@
"src/qs8-gemm/gen/6x16c4-minmax-neondot.c",
"src/qs8-gemm/gen/8x8c4-minmax-neondot.c",
"src/qs8-gemm/gen/8x16c4-minmax-neondot.c",
- "src/qs8-gemm/gen/12x8c4-minmax-neondot.c",
"src/qs8-igemm/gen/1x8c4-minmax-neondot.c",
"src/qs8-igemm/gen/1x16c4-minmax-neondot.c",
"src/qs8-igemm/gen/4x8c4-minmax-neondot.c",
@@ -2285,7 +2284,6 @@
"src/qs8-igemm/gen/6x16c4-minmax-neondot.c",
"src/qs8-igemm/gen/8x8c4-minmax-neondot.c",
"src/qs8-igemm/gen/8x16c4-minmax-neondot.c",
- "src/qs8-igemm/gen/12x8c4-minmax-neondot.c",
]
SSE_UKERNELS = [
diff --git a/CMakeLists.txt b/CMakeLists.txt
index 4e8f774..8d26e08 100755
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -1520,7 +1520,6 @@
src/qs8-gemm/gen/6x16c4-minmax-neondot.c
src/qs8-gemm/gen/8x8c4-minmax-neondot.c
src/qs8-gemm/gen/8x16c4-minmax-neondot.c
- src/qs8-gemm/gen/12x8c4-minmax-neondot.c
src/qs8-igemm/gen/1x8c4-minmax-neondot.c
src/qs8-igemm/gen/1x16c4-minmax-neondot.c
src/qs8-igemm/gen/4x8c4-minmax-neondot.c
@@ -1528,8 +1527,7 @@
src/qs8-igemm/gen/6x8c4-minmax-neondot.c
src/qs8-igemm/gen/6x16c4-minmax-neondot.c
src/qs8-igemm/gen/8x8c4-minmax-neondot.c
- src/qs8-igemm/gen/8x16c4-minmax-neondot.c
- src/qs8-igemm/gen/12x8c4-minmax-neondot.c)
+ src/qs8-igemm/gen/8x16c4-minmax-neondot.c)
SET(XNNPACK_SSE_MICROKERNEL_SRCS
src/f32-avgpool/9p8x-minmax-sse-c4.c
diff --git a/bench/qs8-gemm-e2e.cc b/bench/qs8-gemm-e2e.cc
index 80d60bb..41f5c3d 100644
--- a/bench/qs8-gemm-e2e.cc
+++ b/bench/qs8-gemm-e2e.cc
@@ -544,16 +544,6 @@
benchmark::utils::CheckNEONDOT);
}
- static void qs8_gemm_minmax_ukernel_12x8c4__neondot(benchmark::State& state, models::ExecutionPlanFactory model) {
- GEMMEnd2EndBenchmark(state, model,
- xnn_qs8_gemm_minmax_ukernel_12x8c4__neondot,
- xnn_qs8_igemm_minmax_ukernel_12x8c4__neondot,
- xnn_qs8_gemm_minmax_ukernel_1x8c4__neondot,
- xnn_qs8_igemm_minmax_ukernel_1x8c4__neondot,
- 12 /* mr */, 8 /* nr */, 2 /* log2_kr */, 0 /* log2_sr */,
- benchmark::utils::CheckNEONDOT);
- }
-
#if XNN_ENABLE_FULL_BENCHMARKS
static void qs8_gemm_minmax_ukernel_1x8c8__neon_mull_padal(benchmark::State& state, models::ExecutionPlanFactory model) {
GEMMEnd2EndBenchmark(state, model,
@@ -810,7 +800,6 @@
BENCHMARK_QS8_END2END(qs8_gemm_minmax_ukernel_6x16c4__neondot);
BENCHMARK_QS8_END2END(qs8_gemm_minmax_ukernel_8x8c4__neondot);
BENCHMARK_QS8_END2END(qs8_gemm_minmax_ukernel_8x16c4__neondot);
- BENCHMARK_QS8_END2END(qs8_gemm_minmax_ukernel_12x8c4__neondot);
#if XNN_ENABLE_FULL_BENCHMARKS
BENCHMARK_QS8_END2END(qs8_gemm_minmax_ukernel_1x8c8__neon_mlal_padal);
diff --git a/bench/qs8-gemm.cc b/bench/qs8-gemm.cc
index 524b784..cb1067b 100644
--- a/bench/qs8-gemm.cc
+++ b/bench/qs8-gemm.cc
@@ -464,9 +464,6 @@
static void qs8_gemm_8x8c4__neondot(benchmark::State& state, const char* net) {
GEMMBenchmark(state, xnn_qs8_gemm_minmax_ukernel_8x8c4__neondot, 8, 8, 4, 1, benchmark::utils::CheckNEONDOT);
}
- static void qs8_gemm_12x8c4__neondot(benchmark::State& state, const char* net) {
- GEMMBenchmark(state, xnn_qs8_gemm_minmax_ukernel_12x8c4__neondot, 12, 8, 4, 1, benchmark::utils::CheckNEONDOT);
- }
static void qs8_gemm_1x16c4__neondot(benchmark::State& state, const char* net) {
GEMMBenchmark(state, xnn_qs8_gemm_minmax_ukernel_1x16c4__neondot, 1, 16, 4, 1, benchmark::utils::CheckNEONDOT);
}
@@ -540,7 +537,6 @@
BENCHMARK_GEMM(qs8_gemm_4x8c4__neondot)
BENCHMARK_GEMM(qs8_gemm_6x8c4__neondot)
BENCHMARK_GEMM(qs8_gemm_8x8c4__neondot)
- BENCHMARK_GEMM(qs8_gemm_12x8c4__neondot)
BENCHMARK_GEMM(qs8_gemm_1x16c4__neondot)
BENCHMARK_GEMM(qs8_gemm_4x16c4__neondot)
BENCHMARK_GEMM(qs8_gemm_6x16c4__neondot)
diff --git a/scripts/generate-qs8-gemm.sh b/scripts/generate-qs8-gemm.sh
index 25e5cf3..5f071ac 100755
--- a/scripts/generate-qs8-gemm.sh
+++ b/scripts/generate-qs8-gemm.sh
@@ -92,7 +92,6 @@
tools/xngen src/qs8-gemm/MRxNRc4-neondot.c.in -D MR=4 -D NR=8 -o src/qs8-gemm/gen/4x8c4-minmax-neondot.c
tools/xngen src/qs8-gemm/MRxNRc4-neondot.c.in -D MR=6 -D NR=8 -o src/qs8-gemm/gen/6x8c4-minmax-neondot.c
tools/xngen src/qs8-gemm/MRxNRc4-neondot.c.in -D MR=8 -D NR=8 -o src/qs8-gemm/gen/8x8c4-minmax-neondot.c
-tools/xngen src/qs8-gemm/MRxNRc4-neondot.c.in -D MR=12 -D NR=8 -o src/qs8-gemm/gen/12x8c4-minmax-neondot.c
tools/xngen src/qs8-gemm/MRxNRc4-neondot.c.in -D MR=1 -D NR=16 -o src/qs8-gemm/gen/1x16c4-minmax-neondot.c
tools/xngen src/qs8-gemm/MRxNRc4-neondot.c.in -D MR=4 -D NR=16 -o src/qs8-gemm/gen/4x16c4-minmax-neondot.c
tools/xngen src/qs8-gemm/MRxNRc4-neondot.c.in -D MR=6 -D NR=16 -o src/qs8-gemm/gen/6x16c4-minmax-neondot.c
diff --git a/scripts/generate-qs8-igemm.sh b/scripts/generate-qs8-igemm.sh
index 8b93963..e2a1960 100755
--- a/scripts/generate-qs8-igemm.sh
+++ b/scripts/generate-qs8-igemm.sh
@@ -88,7 +88,6 @@
tools/xngen src/qs8-igemm/MRxNRc4-neondot.c.in -D MR=4 -D NR=8 -o src/qs8-igemm/gen/4x8c4-minmax-neondot.c
tools/xngen src/qs8-igemm/MRxNRc4-neondot.c.in -D MR=6 -D NR=8 -o src/qs8-igemm/gen/6x8c4-minmax-neondot.c
tools/xngen src/qs8-igemm/MRxNRc4-neondot.c.in -D MR=8 -D NR=8 -o src/qs8-igemm/gen/8x8c4-minmax-neondot.c
-tools/xngen src/qs8-igemm/MRxNRc4-neondot.c.in -D MR=12 -D NR=8 -o src/qs8-igemm/gen/12x8c4-minmax-neondot.c
tools/xngen src/qs8-igemm/MRxNRc4-neondot.c.in -D MR=1 -D NR=16 -o src/qs8-igemm/gen/1x16c4-minmax-neondot.c
tools/xngen src/qs8-igemm/MRxNRc4-neondot.c.in -D MR=4 -D NR=16 -o src/qs8-igemm/gen/4x16c4-minmax-neondot.c
tools/xngen src/qs8-igemm/MRxNRc4-neondot.c.in -D MR=6 -D NR=16 -o src/qs8-igemm/gen/6x16c4-minmax-neondot.c
diff --git a/src/qs8-gemm/gen/12x8c4-minmax-neondot.c b/src/qs8-gemm/gen/12x8c4-minmax-neondot.c
deleted file mode 100644
index 71829b0..0000000
--- a/src/qs8-gemm/gen/12x8c4-minmax-neondot.c
+++ /dev/null
@@ -1,501 +0,0 @@
-// Auto-generated file. Do not edit!
-// Template: src/qs8-gemm/MRxNRc4-neondot.c.in
-// Generator: tools/xngen
-//
-// Copyright 2020 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 <assert.h>
-
-#include <arm_neon.h>
-
-#include <xnnpack/gemm.h>
-#include <xnnpack/math.h>
-
-
-void xnn_qs8_gemm_minmax_ukernel_12x8c4__neondot(
- size_t mr,
- size_t nc,
- size_t kc,
- const int8_t* restrict a,
- size_t a_stride,
- const void* restrict w,
- int8_t* restrict c,
- size_t cm_stride,
- size_t cn_stride,
- const union xnn_qs8_gemm_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN {
- assert(mr != 0);
- assert(mr <= 12);
- assert(nc != 0);
- assert(kc != 0);
- assert(kc % sizeof(int8_t) == 0);
- assert(a != NULL);
- assert(w != NULL);
- assert(c != NULL);
-
- kc = round_up_po2(kc, 4);
- const int8_t* a0 = a;
- int8_t* c0 = c;
- const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride);
- int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
- if XNN_UNPREDICTABLE(mr < 2) {
- a1 = a0;
- c1 = c0;
- }
- const int8_t* a2 = (const int8_t*) ((uintptr_t) a1 + a_stride);
- int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
- if XNN_UNPREDICTABLE(mr <= 2) {
- a2 = a1;
- c2 = c1;
- }
- const int8_t* a3 = (const int8_t*) ((uintptr_t) a2 + a_stride);
- int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride);
- if XNN_UNPREDICTABLE(mr < 4) {
- a3 = a2;
- c3 = c2;
- }
- const int8_t* a4 = (const int8_t*) ((uintptr_t) a3 + a_stride);
- int8_t* c4 = (int8_t*) ((uintptr_t) c3 + cm_stride);
- if XNN_UNPREDICTABLE(mr <= 4) {
- a4 = a3;
- c4 = c3;
- }
- const int8_t* a5 = (const int8_t*) ((uintptr_t) a4 + a_stride);
- int8_t* c5 = (int8_t*) ((uintptr_t) c4 + cm_stride);
- if XNN_UNPREDICTABLE(mr < 6) {
- a5 = a4;
- c5 = c4;
- }
- const int8_t* a6 = (const int8_t*) ((uintptr_t) a5 + a_stride);
- int8_t* c6 = (int8_t*) ((uintptr_t) c5 + cm_stride);
- if XNN_UNPREDICTABLE(mr <= 6) {
- a6 = a5;
- c6 = c5;
- }
- const int8_t* a7 = (const int8_t*) ((uintptr_t) a6 + a_stride);
- int8_t* c7 = (int8_t*) ((uintptr_t) c6 + cm_stride);
- if XNN_UNPREDICTABLE(mr < 8) {
- a7 = a6;
- c7 = c6;
- }
- const int8_t* a8 = (const int8_t*) ((uintptr_t) a7 + a_stride);
- int8_t* c8 = (int8_t*) ((uintptr_t) c7 + cm_stride);
- if XNN_UNPREDICTABLE(mr <= 8) {
- a8 = a7;
- c8 = c7;
- }
- const int8_t* a9 = (const int8_t*) ((uintptr_t) a8 + a_stride);
- int8_t* c9 = (int8_t*) ((uintptr_t) c8 + cm_stride);
- if XNN_UNPREDICTABLE(mr < 10) {
- a9 = a8;
- c9 = c8;
- }
- const int8_t* a10 = (const int8_t*) ((uintptr_t) a9 + a_stride);
- int8_t* c10 = (int8_t*) ((uintptr_t) c9 + cm_stride);
- if XNN_UNPREDICTABLE(mr <= 10) {
- a10 = a9;
- c10 = c9;
- }
- const int8_t* a11 = (const int8_t*) ((uintptr_t) a10 + a_stride);
- int8_t* c11 = (int8_t*) ((uintptr_t) c10 + cm_stride);
- if XNN_UNPREDICTABLE(mr != 12) {
- a11 = a10;
- c11 = c10;
- }
-
- // Loop over groups of 8 columns.
- do {
- // Initialize accumulators with bias. 8 bias values are loaded from the
- // weight matrix, at the start of the group of 8 columns.
- int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
- int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
- int32x4_t vacc1x0123 = vacc0x0123;
- int32x4_t vacc1x4567 = vacc0x4567;
- int32x4_t vacc2x0123 = vacc0x0123;
- int32x4_t vacc2x4567 = vacc0x4567;
- int32x4_t vacc3x0123 = vacc0x0123;
- int32x4_t vacc3x4567 = vacc0x4567;
- int32x4_t vacc4x0123 = vacc0x0123;
- int32x4_t vacc4x4567 = vacc0x4567;
- int32x4_t vacc5x0123 = vacc0x0123;
- int32x4_t vacc5x4567 = vacc0x4567;
- int32x4_t vacc6x0123 = vacc0x0123;
- int32x4_t vacc6x4567 = vacc0x4567;
- int32x4_t vacc7x0123 = vacc0x0123;
- int32x4_t vacc7x4567 = vacc0x4567;
- int32x4_t vacc8x0123 = vacc0x0123;
- int32x4_t vacc8x4567 = vacc0x4567;
- int32x4_t vacc9x0123 = vacc0x0123;
- int32x4_t vacc9x4567 = vacc0x4567;
- int32x4_t vacc10x0123 = vacc0x0123;
- int32x4_t vacc10x4567 = vacc0x4567;
- int32x4_t vacc11x0123 = vacc0x0123;
- int32x4_t vacc11x4567 = vacc0x4567;
-
- // Inner accumulation loop along the 8 columns.
- size_t k = kc;
- // 2x partial unrolled loop to load 8 bytes at a time.
- while (k >= 8 * sizeof(int8_t)) {
- // Load a 12x8 block of activations.
- const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 8;
- const int8x8_t va1x01234567 = vld1_s8(a1); a1 += 8;
- const int8x8_t va2x01234567 = vld1_s8(a2); a2 += 8;
- const int8x8_t va3x01234567 = vld1_s8(a3); a3 += 8;
- const int8x8_t va4x01234567 = vld1_s8(a4); a4 += 8;
- const int8x8_t va5x01234567 = vld1_s8(a5); a5 += 8;
- const int8x8_t va6x01234567 = vld1_s8(a6); a6 += 8;
- const int8x8_t va7x01234567 = vld1_s8(a7); a7 += 8;
- const int8x8_t va8x01234567 = vld1_s8(a8); a8 += 8;
- const int8x8_t va9x01234567 = vld1_s8(a9); a9 += 8;
- const int8x8_t va10x01234567 = vld1_s8(a10); a10 += 8;
- const int8x8_t va11x01234567 = vld1_s8(a11); a11 += 8;
-
- // Load a 8x8 block of weights.
- const int8x16_t vb0123x0123 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
- const int8x16_t vb0123x4567 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
- const int8x16_t vb4567x0123 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
- const int8x16_t vb4567x4567 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
-
- // Multiply-accumulate: 12x8 * 8x8 --> 12x8.
- vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
- vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
- vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb0123x0123, va1x01234567, 0);
- vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb0123x4567, va1x01234567, 0);
- vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb0123x0123, va2x01234567, 0);
- vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb0123x4567, va2x01234567, 0);
- vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb0123x0123, va3x01234567, 0);
- vacc3x4567 = vdotq_lane_s32(vacc3x4567, vb0123x4567, va3x01234567, 0);
- vacc4x0123 = vdotq_lane_s32(vacc4x0123, vb0123x0123, va4x01234567, 0);
- vacc4x4567 = vdotq_lane_s32(vacc4x4567, vb0123x4567, va4x01234567, 0);
- vacc5x0123 = vdotq_lane_s32(vacc5x0123, vb0123x0123, va5x01234567, 0);
- vacc5x4567 = vdotq_lane_s32(vacc5x4567, vb0123x4567, va5x01234567, 0);
- vacc6x0123 = vdotq_lane_s32(vacc6x0123, vb0123x0123, va6x01234567, 0);
- vacc6x4567 = vdotq_lane_s32(vacc6x4567, vb0123x4567, va6x01234567, 0);
- vacc7x0123 = vdotq_lane_s32(vacc7x0123, vb0123x0123, va7x01234567, 0);
- vacc7x4567 = vdotq_lane_s32(vacc7x4567, vb0123x4567, va7x01234567, 0);
- vacc8x0123 = vdotq_lane_s32(vacc8x0123, vb0123x0123, va8x01234567, 0);
- vacc8x4567 = vdotq_lane_s32(vacc8x4567, vb0123x4567, va8x01234567, 0);
- vacc9x0123 = vdotq_lane_s32(vacc9x0123, vb0123x0123, va9x01234567, 0);
- vacc9x4567 = vdotq_lane_s32(vacc9x4567, vb0123x4567, va9x01234567, 0);
- vacc10x0123 = vdotq_lane_s32(vacc10x0123, vb0123x0123, va10x01234567, 0);
- vacc10x4567 = vdotq_lane_s32(vacc10x4567, vb0123x4567, va10x01234567, 0);
- vacc11x0123 = vdotq_lane_s32(vacc11x0123, vb0123x0123, va11x01234567, 0);
- vacc11x4567 = vdotq_lane_s32(vacc11x4567, vb0123x4567, va11x01234567, 0);
- vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb4567x0123, va0x01234567, 1);
- vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb4567x4567, va0x01234567, 1);
- vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb4567x0123, va1x01234567, 1);
- vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb4567x4567, va1x01234567, 1);
- vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb4567x0123, va2x01234567, 1);
- vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb4567x4567, va2x01234567, 1);
- vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb4567x0123, va3x01234567, 1);
- vacc3x4567 = vdotq_lane_s32(vacc3x4567, vb4567x4567, va3x01234567, 1);
- vacc4x0123 = vdotq_lane_s32(vacc4x0123, vb4567x0123, va4x01234567, 1);
- vacc4x4567 = vdotq_lane_s32(vacc4x4567, vb4567x4567, va4x01234567, 1);
- vacc5x0123 = vdotq_lane_s32(vacc5x0123, vb4567x0123, va5x01234567, 1);
- vacc5x4567 = vdotq_lane_s32(vacc5x4567, vb4567x4567, va5x01234567, 1);
- vacc6x0123 = vdotq_lane_s32(vacc6x0123, vb4567x0123, va6x01234567, 1);
- vacc6x4567 = vdotq_lane_s32(vacc6x4567, vb4567x4567, va6x01234567, 1);
- vacc7x0123 = vdotq_lane_s32(vacc7x0123, vb4567x0123, va7x01234567, 1);
- vacc7x4567 = vdotq_lane_s32(vacc7x4567, vb4567x4567, va7x01234567, 1);
- vacc8x0123 = vdotq_lane_s32(vacc8x0123, vb4567x0123, va8x01234567, 1);
- vacc8x4567 = vdotq_lane_s32(vacc8x4567, vb4567x4567, va8x01234567, 1);
- vacc9x0123 = vdotq_lane_s32(vacc9x0123, vb4567x0123, va9x01234567, 1);
- vacc9x4567 = vdotq_lane_s32(vacc9x4567, vb4567x4567, va9x01234567, 1);
- vacc10x0123 = vdotq_lane_s32(vacc10x0123, vb4567x0123, va10x01234567, 1);
- vacc10x4567 = vdotq_lane_s32(vacc10x4567, vb4567x4567, va10x01234567, 1);
- vacc11x0123 = vdotq_lane_s32(vacc11x0123, vb4567x0123, va11x01234567, 1);
- vacc11x4567 = vdotq_lane_s32(vacc11x4567, vb4567x4567, va11x01234567, 1);
-
- k -= 8 * sizeof(int8_t);
- }
- // Handle up to 4 final positions of `k`
- if XNN_UNLIKELY(k != 0) {
- // Load a 12x4 block of activations.
- const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 4;
- const int8x8_t va1x01234567 = vld1_s8(a1); a1 += 4;
- const int8x8_t va2x01234567 = vld1_s8(a2); a2 += 4;
- const int8x8_t va3x01234567 = vld1_s8(a3); a3 += 4;
- const int8x8_t va4x01234567 = vld1_s8(a4); a4 += 4;
- const int8x8_t va5x01234567 = vld1_s8(a5); a5 += 4;
- const int8x8_t va6x01234567 = vld1_s8(a6); a6 += 4;
- const int8x8_t va7x01234567 = vld1_s8(a7); a7 += 4;
- const int8x8_t va8x01234567 = vld1_s8(a8); a8 += 4;
- const int8x8_t va9x01234567 = vld1_s8(a9); a9 += 4;
- const int8x8_t va10x01234567 = vld1_s8(a10); a10 += 4;
- const int8x8_t va11x01234567 = vld1_s8(a11); a11 += 4;
-
- // Load a 4x8 block of weights.
- const int8x16_t vb0123x0123 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
- const int8x16_t vb0123x4567 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
-
- // Multiply-accumulate: 12x4 * 4x8 --> 12x8.
- vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
- vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
- vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb0123x0123, va1x01234567, 0);
- vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb0123x4567, va1x01234567, 0);
- vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb0123x0123, va2x01234567, 0);
- vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb0123x4567, va2x01234567, 0);
- vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb0123x0123, va3x01234567, 0);
- vacc3x4567 = vdotq_lane_s32(vacc3x4567, vb0123x4567, va3x01234567, 0);
- vacc4x0123 = vdotq_lane_s32(vacc4x0123, vb0123x0123, va4x01234567, 0);
- vacc4x4567 = vdotq_lane_s32(vacc4x4567, vb0123x4567, va4x01234567, 0);
- vacc5x0123 = vdotq_lane_s32(vacc5x0123, vb0123x0123, va5x01234567, 0);
- vacc5x4567 = vdotq_lane_s32(vacc5x4567, vb0123x4567, va5x01234567, 0);
- vacc6x0123 = vdotq_lane_s32(vacc6x0123, vb0123x0123, va6x01234567, 0);
- vacc6x4567 = vdotq_lane_s32(vacc6x4567, vb0123x4567, va6x01234567, 0);
- vacc7x0123 = vdotq_lane_s32(vacc7x0123, vb0123x0123, va7x01234567, 0);
- vacc7x4567 = vdotq_lane_s32(vacc7x4567, vb0123x4567, va7x01234567, 0);
- vacc8x0123 = vdotq_lane_s32(vacc8x0123, vb0123x0123, va8x01234567, 0);
- vacc8x4567 = vdotq_lane_s32(vacc8x4567, vb0123x4567, va8x01234567, 0);
- vacc9x0123 = vdotq_lane_s32(vacc9x0123, vb0123x0123, va9x01234567, 0);
- vacc9x4567 = vdotq_lane_s32(vacc9x4567, vb0123x4567, va9x01234567, 0);
- vacc10x0123 = vdotq_lane_s32(vacc10x0123, vb0123x0123, va10x01234567, 0);
- vacc10x4567 = vdotq_lane_s32(vacc10x4567, vb0123x4567, va10x01234567, 0);
- vacc11x0123 = vdotq_lane_s32(vacc11x0123, vb0123x0123, va11x01234567, 0);
- vacc11x4567 = vdotq_lane_s32(vacc11x4567, vb0123x4567, va11x01234567, 0);
- }
-
- // Post-accumulation work
- const int32x4_t vright_shift = vld1q_dup_s32(¶ms->neon.right_shift);
- const int32x4_t vzero_shift_mask = vreinterpretq_s32_u32(vceqq_s32(vright_shift, vmovq_n_s32(0)));
-
- const int32x4_t vproduct0x0123 = vqrdmulhq_n_s32(vacc0x0123, params->neon.multiplier);
- const int32x4_t vproduct0x4567 = vqrdmulhq_n_s32(vacc0x4567, params->neon.multiplier);
- const int32x4_t vproduct1x0123 = vqrdmulhq_n_s32(vacc1x0123, params->neon.multiplier);
- const int32x4_t vproduct1x4567 = vqrdmulhq_n_s32(vacc1x4567, params->neon.multiplier);
- const int32x4_t vproduct2x0123 = vqrdmulhq_n_s32(vacc2x0123, params->neon.multiplier);
- const int32x4_t vproduct2x4567 = vqrdmulhq_n_s32(vacc2x4567, params->neon.multiplier);
- const int32x4_t vproduct3x0123 = vqrdmulhq_n_s32(vacc3x0123, params->neon.multiplier);
- const int32x4_t vproduct3x4567 = vqrdmulhq_n_s32(vacc3x4567, params->neon.multiplier);
- const int32x4_t vproduct4x0123 = vqrdmulhq_n_s32(vacc4x0123, params->neon.multiplier);
- const int32x4_t vproduct4x4567 = vqrdmulhq_n_s32(vacc4x4567, params->neon.multiplier);
- const int32x4_t vproduct5x0123 = vqrdmulhq_n_s32(vacc5x0123, params->neon.multiplier);
- const int32x4_t vproduct5x4567 = vqrdmulhq_n_s32(vacc5x4567, params->neon.multiplier);
- const int32x4_t vproduct6x0123 = vqrdmulhq_n_s32(vacc6x0123, params->neon.multiplier);
- const int32x4_t vproduct6x4567 = vqrdmulhq_n_s32(vacc6x4567, params->neon.multiplier);
- const int32x4_t vproduct7x0123 = vqrdmulhq_n_s32(vacc7x0123, params->neon.multiplier);
- const int32x4_t vproduct7x4567 = vqrdmulhq_n_s32(vacc7x4567, params->neon.multiplier);
- const int32x4_t vproduct8x0123 = vqrdmulhq_n_s32(vacc8x0123, params->neon.multiplier);
- const int32x4_t vproduct8x4567 = vqrdmulhq_n_s32(vacc8x4567, params->neon.multiplier);
- const int32x4_t vproduct9x0123 = vqrdmulhq_n_s32(vacc9x0123, params->neon.multiplier);
- const int32x4_t vproduct9x4567 = vqrdmulhq_n_s32(vacc9x4567, params->neon.multiplier);
- const int32x4_t vproduct10x0123 = vqrdmulhq_n_s32(vacc10x0123, params->neon.multiplier);
- const int32x4_t vproduct10x4567 = vqrdmulhq_n_s32(vacc10x4567, params->neon.multiplier);
- const int32x4_t vproduct11x0123 = vqrdmulhq_n_s32(vacc11x0123, params->neon.multiplier);
- const int32x4_t vproduct11x4567 = vqrdmulhq_n_s32(vacc11x4567, params->neon.multiplier);
-
- vacc0x0123 = vsraq_n_s32(vproduct0x0123, vbicq_s32(vacc0x0123, vzero_shift_mask), 31);
- vacc0x4567 = vsraq_n_s32(vproduct0x4567, vbicq_s32(vacc0x4567, vzero_shift_mask), 31);
- vacc1x0123 = vsraq_n_s32(vproduct1x0123, vbicq_s32(vacc1x0123, vzero_shift_mask), 31);
- vacc1x4567 = vsraq_n_s32(vproduct1x4567, vbicq_s32(vacc1x4567, vzero_shift_mask), 31);
- vacc2x0123 = vsraq_n_s32(vproduct2x0123, vbicq_s32(vacc2x0123, vzero_shift_mask), 31);
- vacc2x4567 = vsraq_n_s32(vproduct2x4567, vbicq_s32(vacc2x4567, vzero_shift_mask), 31);
- vacc3x0123 = vsraq_n_s32(vproduct3x0123, vbicq_s32(vacc3x0123, vzero_shift_mask), 31);
- vacc3x4567 = vsraq_n_s32(vproduct3x4567, vbicq_s32(vacc3x4567, vzero_shift_mask), 31);
- vacc4x0123 = vsraq_n_s32(vproduct4x0123, vbicq_s32(vacc4x0123, vzero_shift_mask), 31);
- vacc4x4567 = vsraq_n_s32(vproduct4x4567, vbicq_s32(vacc4x4567, vzero_shift_mask), 31);
- vacc5x0123 = vsraq_n_s32(vproduct5x0123, vbicq_s32(vacc5x0123, vzero_shift_mask), 31);
- vacc5x4567 = vsraq_n_s32(vproduct5x4567, vbicq_s32(vacc5x4567, vzero_shift_mask), 31);
- vacc6x0123 = vsraq_n_s32(vproduct6x0123, vbicq_s32(vacc6x0123, vzero_shift_mask), 31);
- vacc6x4567 = vsraq_n_s32(vproduct6x4567, vbicq_s32(vacc6x4567, vzero_shift_mask), 31);
- vacc7x0123 = vsraq_n_s32(vproduct7x0123, vbicq_s32(vacc7x0123, vzero_shift_mask), 31);
- vacc7x4567 = vsraq_n_s32(vproduct7x4567, vbicq_s32(vacc7x4567, vzero_shift_mask), 31);
- vacc8x0123 = vsraq_n_s32(vproduct8x0123, vbicq_s32(vacc8x0123, vzero_shift_mask), 31);
- vacc8x4567 = vsraq_n_s32(vproduct8x4567, vbicq_s32(vacc8x4567, vzero_shift_mask), 31);
- vacc9x0123 = vsraq_n_s32(vproduct9x0123, vbicq_s32(vacc9x0123, vzero_shift_mask), 31);
- vacc9x4567 = vsraq_n_s32(vproduct9x4567, vbicq_s32(vacc9x4567, vzero_shift_mask), 31);
- vacc10x0123 = vsraq_n_s32(vproduct10x0123, vbicq_s32(vacc10x0123, vzero_shift_mask), 31);
- vacc10x4567 = vsraq_n_s32(vproduct10x4567, vbicq_s32(vacc10x4567, vzero_shift_mask), 31);
- vacc11x0123 = vsraq_n_s32(vproduct11x0123, vbicq_s32(vacc11x0123, vzero_shift_mask), 31);
- vacc11x4567 = vsraq_n_s32(vproduct11x4567, vbicq_s32(vacc11x4567, vzero_shift_mask), 31);
-
- vacc0x0123 = vrshlq_s32(vacc0x0123, vright_shift);
- vacc0x4567 = vrshlq_s32(vacc0x4567, vright_shift);
- vacc1x0123 = vrshlq_s32(vacc1x0123, vright_shift);
- vacc1x4567 = vrshlq_s32(vacc1x4567, vright_shift);
- vacc2x0123 = vrshlq_s32(vacc2x0123, vright_shift);
- vacc2x4567 = vrshlq_s32(vacc2x4567, vright_shift);
- vacc3x0123 = vrshlq_s32(vacc3x0123, vright_shift);
- vacc3x4567 = vrshlq_s32(vacc3x4567, vright_shift);
- vacc4x0123 = vrshlq_s32(vacc4x0123, vright_shift);
- vacc4x4567 = vrshlq_s32(vacc4x4567, vright_shift);
- vacc5x0123 = vrshlq_s32(vacc5x0123, vright_shift);
- vacc5x4567 = vrshlq_s32(vacc5x4567, vright_shift);
- vacc6x0123 = vrshlq_s32(vacc6x0123, vright_shift);
- vacc6x4567 = vrshlq_s32(vacc6x4567, vright_shift);
- vacc7x0123 = vrshlq_s32(vacc7x0123, vright_shift);
- vacc7x4567 = vrshlq_s32(vacc7x4567, vright_shift);
- vacc8x0123 = vrshlq_s32(vacc8x0123, vright_shift);
- vacc8x4567 = vrshlq_s32(vacc8x4567, vright_shift);
- vacc9x0123 = vrshlq_s32(vacc9x0123, vright_shift);
- vacc9x4567 = vrshlq_s32(vacc9x4567, vright_shift);
- vacc10x0123 = vrshlq_s32(vacc10x0123, vright_shift);
- vacc10x4567 = vrshlq_s32(vacc10x4567, vright_shift);
- vacc11x0123 = vrshlq_s32(vacc11x0123, vright_shift);
- vacc11x4567 = vrshlq_s32(vacc11x4567, vright_shift);
-
- const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->neon.output_zero_point);
-#if XNN_ARCH_ARM64
- const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
- const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), voutput_zero_point);
- const int16x8_t vacc2x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567), voutput_zero_point);
- const int16x8_t vacc3x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567), voutput_zero_point);
- const int16x8_t vacc4x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc4x0123), vacc4x4567), voutput_zero_point);
- const int16x8_t vacc5x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc5x0123), vacc5x4567), voutput_zero_point);
- const int16x8_t vacc6x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc6x0123), vacc6x4567), voutput_zero_point);
- const int16x8_t vacc7x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc7x0123), vacc7x4567), voutput_zero_point);
- const int16x8_t vacc8x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc8x0123), vacc8x4567), voutput_zero_point);
- const int16x8_t vacc9x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc9x0123), vacc9x4567), voutput_zero_point);
- const int16x8_t vacc10x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc10x0123), vacc10x4567), voutput_zero_point);
- const int16x8_t vacc11x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc11x0123), vacc11x4567), voutput_zero_point);
-
- int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
- int8x16_t vout2x01234567_3x01234567 = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc3x01234567);
- int8x16_t vout4x01234567_5x01234567 = vqmovn_high_s16(vqmovn_s16(vacc4x01234567), vacc5x01234567);
- int8x16_t vout6x01234567_7x01234567 = vqmovn_high_s16(vqmovn_s16(vacc6x01234567), vacc7x01234567);
- int8x16_t vout8x01234567_9x01234567 = vqmovn_high_s16(vqmovn_s16(vacc8x01234567), vacc9x01234567);
- int8x16_t vout10x01234567_11x01234567 = vqmovn_high_s16(vqmovn_s16(vacc10x01234567), vacc11x01234567);
-#else
- const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
- const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567)), voutput_zero_point);
- const int16x8_t vacc2x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567)), voutput_zero_point);
- const int16x8_t vacc3x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x4567)), voutput_zero_point);
- const int16x8_t vacc4x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc4x0123), vqmovn_s32(vacc4x4567)), voutput_zero_point);
- const int16x8_t vacc5x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc5x0123), vqmovn_s32(vacc5x4567)), voutput_zero_point);
- const int16x8_t vacc6x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc6x0123), vqmovn_s32(vacc6x4567)), voutput_zero_point);
- const int16x8_t vacc7x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc7x0123), vqmovn_s32(vacc7x4567)), voutput_zero_point);
- const int16x8_t vacc8x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc8x0123), vqmovn_s32(vacc8x4567)), voutput_zero_point);
- const int16x8_t vacc9x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc9x0123), vqmovn_s32(vacc9x4567)), voutput_zero_point);
- const int16x8_t vacc10x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc10x0123), vqmovn_s32(vacc10x4567)), voutput_zero_point);
- const int16x8_t vacc11x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc11x0123), vqmovn_s32(vacc11x4567)), voutput_zero_point);
-
- int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
- int8x16_t vout2x01234567_3x01234567 = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc3x01234567));
- int8x16_t vout4x01234567_5x01234567 = vcombine_s8(vqmovn_s16(vacc4x01234567), vqmovn_s16(vacc5x01234567));
- int8x16_t vout6x01234567_7x01234567 = vcombine_s8(vqmovn_s16(vacc6x01234567), vqmovn_s16(vacc7x01234567));
- int8x16_t vout8x01234567_9x01234567 = vcombine_s8(vqmovn_s16(vacc8x01234567), vqmovn_s16(vacc9x01234567));
- int8x16_t vout10x01234567_11x01234567 = vcombine_s8(vqmovn_s16(vacc10x01234567), vqmovn_s16(vacc11x01234567));
-#endif
- const int8x16_t voutput_min = vld1q_dup_s8(¶ms->neon.output_min);
- const int8x16_t voutput_max = vld1q_dup_s8(¶ms->neon.output_max);
-
- vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
- vout2x01234567_3x01234567 = vmaxq_s8(vout2x01234567_3x01234567, voutput_min);
- vout4x01234567_5x01234567 = vmaxq_s8(vout4x01234567_5x01234567, voutput_min);
- vout6x01234567_7x01234567 = vmaxq_s8(vout6x01234567_7x01234567, voutput_min);
- vout8x01234567_9x01234567 = vmaxq_s8(vout8x01234567_9x01234567, voutput_min);
- vout10x01234567_11x01234567 = vmaxq_s8(vout10x01234567_11x01234567, voutput_min);
-
- vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
- vout2x01234567_3x01234567 = vminq_s8(vout2x01234567_3x01234567, voutput_max);
- vout4x01234567_5x01234567 = vminq_s8(vout4x01234567_5x01234567, voutput_max);
- vout6x01234567_7x01234567 = vminq_s8(vout6x01234567_7x01234567, voutput_max);
- vout8x01234567_9x01234567 = vminq_s8(vout8x01234567_9x01234567, voutput_max);
- vout10x01234567_11x01234567 = vminq_s8(vout10x01234567_11x01234567, voutput_max);
-
- if (nc >= 8) {
- // Main case where there the 8 columns fit in the destination.
- vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
- vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
- vst1_s8(c2 + 0, vget_low_s8(vout2x01234567_3x01234567));
- vst1_s8(c3 + 0, vget_high_s8(vout2x01234567_3x01234567));
- vst1_s8(c4 + 0, vget_low_s8(vout4x01234567_5x01234567));
- vst1_s8(c5 + 0, vget_high_s8(vout4x01234567_5x01234567));
- vst1_s8(c6 + 0, vget_low_s8(vout6x01234567_7x01234567));
- vst1_s8(c7 + 0, vget_high_s8(vout6x01234567_7x01234567));
- vst1_s8(c8 + 0, vget_low_s8(vout8x01234567_9x01234567));
- vst1_s8(c9 + 0, vget_high_s8(vout8x01234567_9x01234567));
- vst1_s8(c10 + 0, vget_low_s8(vout10x01234567_11x01234567));
- vst1_s8(c11 + 0, vget_high_s8(vout10x01234567_11x01234567));
-
- // Advance to the next 8 columns.
- c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
- c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
- c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
- c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
- c4 = (int8_t*) ((uintptr_t) c4 + cn_stride);
- c5 = (int8_t*) ((uintptr_t) c5 + cn_stride);
- c6 = (int8_t*) ((uintptr_t) c6 + cn_stride);
- c7 = (int8_t*) ((uintptr_t) c7 + cn_stride);
- c8 = (int8_t*) ((uintptr_t) c8 + cn_stride);
- c9 = (int8_t*) ((uintptr_t) c9 + cn_stride);
- c10 = (int8_t*) ((uintptr_t) c10 + cn_stride);
- c11 = (int8_t*) ((uintptr_t) c11 + cn_stride);
-
- a0 = (const int8_t*) ((uintptr_t) a0 - kc);
- a1 = (const int8_t*) ((uintptr_t) a1 - kc);
- a2 = (const int8_t*) ((uintptr_t) a2 - kc);
- a3 = (const int8_t*) ((uintptr_t) a3 - kc);
- a4 = (const int8_t*) ((uintptr_t) a4 - kc);
- a5 = (const int8_t*) ((uintptr_t) a5 - kc);
- a6 = (const int8_t*) ((uintptr_t) a6 - kc);
- a7 = (const int8_t*) ((uintptr_t) a7 - kc);
- a8 = (const int8_t*) ((uintptr_t) a8 - kc);
- a9 = (const int8_t*) ((uintptr_t) a9 - kc);
- a10 = (const int8_t*) ((uintptr_t) a10 - kc);
- a11 = (const int8_t*) ((uintptr_t) a11 - kc);
-
- nc -= 8;
- } else {
- // Final case where not all of the 8 columns fit in the destination.
- if (nc & 4) {
- vst1q_lane_u32(__builtin_assume_aligned(c0, 1), vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
- vst1q_lane_u32(__builtin_assume_aligned(c1, 1), vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
- vst1q_lane_u32(__builtin_assume_aligned(c2, 1), vreinterpretq_u32_s8(vout2x01234567_3x01234567), 0); c2 += 4;
- vst1q_lane_u32(__builtin_assume_aligned(c3, 1), vreinterpretq_u32_s8(vout2x01234567_3x01234567), 2); c3 += 4;
- vst1q_lane_u32(__builtin_assume_aligned(c4, 1), vreinterpretq_u32_s8(vout4x01234567_5x01234567), 0); c4 += 4;
- vst1q_lane_u32(__builtin_assume_aligned(c5, 1), vreinterpretq_u32_s8(vout4x01234567_5x01234567), 2); c5 += 4;
- vst1q_lane_u32(__builtin_assume_aligned(c6, 1), vreinterpretq_u32_s8(vout6x01234567_7x01234567), 0); c6 += 4;
- vst1q_lane_u32(__builtin_assume_aligned(c7, 1), vreinterpretq_u32_s8(vout6x01234567_7x01234567), 2); c7 += 4;
- vst1q_lane_u32(__builtin_assume_aligned(c8, 1), vreinterpretq_u32_s8(vout8x01234567_9x01234567), 0); c8 += 4;
- vst1q_lane_u32(__builtin_assume_aligned(c9, 1), vreinterpretq_u32_s8(vout8x01234567_9x01234567), 2); c9 += 4;
- vst1q_lane_u32(__builtin_assume_aligned(c10, 1), vreinterpretq_u32_s8(vout10x01234567_11x01234567), 0); c10 += 4;
- vst1q_lane_u32(__builtin_assume_aligned(c11, 1), vreinterpretq_u32_s8(vout10x01234567_11x01234567), 2); c11 += 4;
- vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
- vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 4);
- vout4x01234567_5x01234567 = vextq_s8(vout4x01234567_5x01234567, vout4x01234567_5x01234567, 4);
- vout6x01234567_7x01234567 = vextq_s8(vout6x01234567_7x01234567, vout6x01234567_7x01234567, 4);
- vout8x01234567_9x01234567 = vextq_s8(vout8x01234567_9x01234567, vout8x01234567_9x01234567, 4);
- vout10x01234567_11x01234567 = vextq_s8(vout10x01234567_11x01234567, vout10x01234567_11x01234567, 4);
- }
- if (nc & 2) {
- vst1q_lane_u16(__builtin_assume_aligned(c0, 1), vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
- vst1q_lane_u16(__builtin_assume_aligned(c1, 1), vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
- vst1q_lane_u16(__builtin_assume_aligned(c2, 1), vreinterpretq_u16_s8(vout2x01234567_3x01234567), 0); c2 += 2;
- vst1q_lane_u16(__builtin_assume_aligned(c3, 1), vreinterpretq_u16_s8(vout2x01234567_3x01234567), 4); c3 += 2;
- vst1q_lane_u16(__builtin_assume_aligned(c4, 1), vreinterpretq_u16_s8(vout4x01234567_5x01234567), 0); c4 += 2;
- vst1q_lane_u16(__builtin_assume_aligned(c5, 1), vreinterpretq_u16_s8(vout4x01234567_5x01234567), 4); c5 += 2;
- vst1q_lane_u16(__builtin_assume_aligned(c6, 1), vreinterpretq_u16_s8(vout6x01234567_7x01234567), 0); c6 += 2;
- vst1q_lane_u16(__builtin_assume_aligned(c7, 1), vreinterpretq_u16_s8(vout6x01234567_7x01234567), 4); c7 += 2;
- vst1q_lane_u16(__builtin_assume_aligned(c8, 1), vreinterpretq_u16_s8(vout8x01234567_9x01234567), 0); c8 += 2;
- vst1q_lane_u16(__builtin_assume_aligned(c9, 1), vreinterpretq_u16_s8(vout8x01234567_9x01234567), 4); c9 += 2;
- vst1q_lane_u16(__builtin_assume_aligned(c10, 1), vreinterpretq_u16_s8(vout10x01234567_11x01234567), 0); c10 += 2;
- vst1q_lane_u16(__builtin_assume_aligned(c11, 1), vreinterpretq_u16_s8(vout10x01234567_11x01234567), 4); c11 += 2;
- vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
- vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 2);
- vout4x01234567_5x01234567 = vextq_s8(vout4x01234567_5x01234567, vout4x01234567_5x01234567, 2);
- vout6x01234567_7x01234567 = vextq_s8(vout6x01234567_7x01234567, vout6x01234567_7x01234567, 2);
- vout8x01234567_9x01234567 = vextq_s8(vout8x01234567_9x01234567, vout8x01234567_9x01234567, 2);
- vout10x01234567_11x01234567 = vextq_s8(vout10x01234567_11x01234567, vout10x01234567_11x01234567, 2);
- }
- if (nc & 1) {
- vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
- vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
- vst1q_lane_s8(c2, vout2x01234567_3x01234567, 0);
- vst1q_lane_s8(c3, vout2x01234567_3x01234567, 8);
- vst1q_lane_s8(c4, vout4x01234567_5x01234567, 0);
- vst1q_lane_s8(c5, vout4x01234567_5x01234567, 8);
- vst1q_lane_s8(c6, vout6x01234567_7x01234567, 0);
- vst1q_lane_s8(c7, vout6x01234567_7x01234567, 8);
- vst1q_lane_s8(c8, vout8x01234567_9x01234567, 0);
- vst1q_lane_s8(c9, vout8x01234567_9x01234567, 8);
- vst1q_lane_s8(c10, vout10x01234567_11x01234567, 0);
- vst1q_lane_s8(c11, vout10x01234567_11x01234567, 8);
- }
-
- nc = 0;
- }
- } while (nc != 0);
-}
diff --git a/src/qs8-igemm/gen/12x8c4-minmax-neondot.c b/src/qs8-igemm/gen/12x8c4-minmax-neondot.c
deleted file mode 100644
index 0f0953b..0000000
--- a/src/qs8-igemm/gen/12x8c4-minmax-neondot.c
+++ /dev/null
@@ -1,519 +0,0 @@
-// Auto-generated file. Do not edit!
-// Template: src/qs8-igemm/MRxNRc4-neondot.c.in
-// Generator: tools/xngen
-//
-// Copyright 2020 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 <assert.h>
-
-#include <arm_neon.h>
-
-#include <xnnpack/igemm.h>
-#include <xnnpack/math.h>
-
-
-void xnn_qs8_igemm_minmax_ukernel_12x8c4__neondot(
- size_t mr,
- size_t nc,
- size_t kc,
- size_t ks,
- const int8_t** restrict a,
- const void* restrict w,
- int8_t* restrict c,
- size_t cm_stride,
- size_t cn_stride,
- size_t a_offset,
- const int8_t* zero,
- const union xnn_qs8_gemm_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN
-{
- assert(mr != 0);
- assert(mr <= 12);
- assert(nc != 0);
- assert(kc != 0);
- assert(ks != 0);
- assert(ks % (12 * sizeof(void*)) == 0);
- assert(a_offset % sizeof(int8_t) == 0);
- assert(a != NULL);
- assert(w != NULL);
- assert(c != NULL);
-
- kc = round_up_po2(kc, 4);
- int8_t* c0 = c;
- int8_t* c1 = (int8_t*) ((uintptr_t) c0 + cm_stride);
- if XNN_UNPREDICTABLE(mr < 2) {
- c1 = c0;
- }
- int8_t* c2 = (int8_t*) ((uintptr_t) c1 + cm_stride);
- if XNN_UNPREDICTABLE(mr <= 2) {
- c2 = c1;
- }
- int8_t* c3 = (int8_t*) ((uintptr_t) c2 + cm_stride);
- if XNN_UNPREDICTABLE(mr < 4) {
- c3 = c2;
- }
- int8_t* c4 = (int8_t*) ((uintptr_t) c3 + cm_stride);
- if XNN_UNPREDICTABLE(mr <= 4) {
- c4 = c3;
- }
- int8_t* c5 = (int8_t*) ((uintptr_t) c4 + cm_stride);
- if XNN_UNPREDICTABLE(mr < 6) {
- c5 = c4;
- }
- int8_t* c6 = (int8_t*) ((uintptr_t) c5 + cm_stride);
- if XNN_UNPREDICTABLE(mr <= 6) {
- c6 = c5;
- }
- int8_t* c7 = (int8_t*) ((uintptr_t) c6 + cm_stride);
- if XNN_UNPREDICTABLE(mr < 8) {
- c7 = c6;
- }
- int8_t* c8 = (int8_t*) ((uintptr_t) c7 + cm_stride);
- if XNN_UNPREDICTABLE(mr <= 8) {
- c8 = c7;
- }
- int8_t* c9 = (int8_t*) ((uintptr_t) c8 + cm_stride);
- if XNN_UNPREDICTABLE(mr < 10) {
- c9 = c8;
- }
- int8_t* c10 = (int8_t*) ((uintptr_t) c9 + cm_stride);
- if XNN_UNPREDICTABLE(mr <= 10) {
- c10 = c9;
- }
- int8_t* c11 = (int8_t*) ((uintptr_t) c10 + cm_stride);
- if XNN_UNPREDICTABLE(mr != 12) {
- c11 = c10;
- }
-
- do {
- int32x4_t vacc0x0123 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
- int32x4_t vacc0x4567 = vld1q_s32(w); w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
- int32x4_t vacc1x0123 = vacc0x0123;
- int32x4_t vacc1x4567 = vacc0x4567;
- int32x4_t vacc2x0123 = vacc0x0123;
- int32x4_t vacc2x4567 = vacc0x4567;
- int32x4_t vacc3x0123 = vacc0x0123;
- int32x4_t vacc3x4567 = vacc0x4567;
- int32x4_t vacc4x0123 = vacc0x0123;
- int32x4_t vacc4x4567 = vacc0x4567;
- int32x4_t vacc5x0123 = vacc0x0123;
- int32x4_t vacc5x4567 = vacc0x4567;
- int32x4_t vacc6x0123 = vacc0x0123;
- int32x4_t vacc6x4567 = vacc0x4567;
- int32x4_t vacc7x0123 = vacc0x0123;
- int32x4_t vacc7x4567 = vacc0x4567;
- int32x4_t vacc8x0123 = vacc0x0123;
- int32x4_t vacc8x4567 = vacc0x4567;
- int32x4_t vacc9x0123 = vacc0x0123;
- int32x4_t vacc9x4567 = vacc0x4567;
- int32x4_t vacc10x0123 = vacc0x0123;
- int32x4_t vacc10x4567 = vacc0x4567;
- int32x4_t vacc11x0123 = vacc0x0123;
- int32x4_t vacc11x4567 = vacc0x4567;
-
- size_t p = ks;
- do {
- const int8_t* restrict a0 = a[0];
- if XNN_UNPREDICTABLE(a0 != zero) {
- a0 = (const int8_t*) ((uintptr_t) a0 + a_offset);
- }
- const int8_t* restrict a1 = a[1];
- if XNN_UNPREDICTABLE(a1 != zero) {
- a1 = (const int8_t*) ((uintptr_t) a1 + a_offset);
- }
- const int8_t* restrict a2 = a[2];
- if XNN_UNPREDICTABLE(a2 != zero) {
- a2 = (const int8_t*) ((uintptr_t) a2 + a_offset);
- }
- const int8_t* restrict a3 = a[3];
- if XNN_UNPREDICTABLE(a3 != zero) {
- a3 = (const int8_t*) ((uintptr_t) a3 + a_offset);
- }
- const int8_t* restrict a4 = a[4];
- if XNN_UNPREDICTABLE(a4 != zero) {
- a4 = (const int8_t*) ((uintptr_t) a4 + a_offset);
- }
- const int8_t* restrict a5 = a[5];
- if XNN_UNPREDICTABLE(a5 != zero) {
- a5 = (const int8_t*) ((uintptr_t) a5 + a_offset);
- }
- const int8_t* restrict a6 = a[6];
- if XNN_UNPREDICTABLE(a6 != zero) {
- a6 = (const int8_t*) ((uintptr_t) a6 + a_offset);
- }
- const int8_t* restrict a7 = a[7];
- if XNN_UNPREDICTABLE(a7 != zero) {
- a7 = (const int8_t*) ((uintptr_t) a7 + a_offset);
- }
- const int8_t* restrict a8 = a[8];
- if XNN_UNPREDICTABLE(a8 != zero) {
- a8 = (const int8_t*) ((uintptr_t) a8 + a_offset);
- }
- const int8_t* restrict a9 = a[9];
- if XNN_UNPREDICTABLE(a9 != zero) {
- a9 = (const int8_t*) ((uintptr_t) a9 + a_offset);
- }
- const int8_t* restrict a10 = a[10];
- if XNN_UNPREDICTABLE(a10 != zero) {
- a10 = (const int8_t*) ((uintptr_t) a10 + a_offset);
- }
- const int8_t* restrict a11 = a[11];
- if XNN_UNPREDICTABLE(a11 != zero) {
- a11 = (const int8_t*) ((uintptr_t) a11 + a_offset);
- }
- a += 12;
-
- // Inner accumulation loop along the 8 columns.
- size_t k = kc;
- // 2x partial unrolled loop to load 8 bytes at a time.
- while (k >= 8 * sizeof(int8_t)) {
- // Load a 12x8 block of activations.
- const int8x8_t va0x01234567 = vld1_s8(a0); a0 += 8;
- const int8x8_t va1x01234567 = vld1_s8(a1); a1 += 8;
- const int8x8_t va2x01234567 = vld1_s8(a2); a2 += 8;
- const int8x8_t va3x01234567 = vld1_s8(a3); a3 += 8;
- const int8x8_t va4x01234567 = vld1_s8(a4); a4 += 8;
- const int8x8_t va5x01234567 = vld1_s8(a5); a5 += 8;
- const int8x8_t va6x01234567 = vld1_s8(a6); a6 += 8;
- const int8x8_t va7x01234567 = vld1_s8(a7); a7 += 8;
- const int8x8_t va8x01234567 = vld1_s8(a8); a8 += 8;
- const int8x8_t va9x01234567 = vld1_s8(a9); a9 += 8;
- const int8x8_t va10x01234567 = vld1_s8(a10); a10 += 8;
- const int8x8_t va11x01234567 = vld1_s8(a11); a11 += 8;
-
- // Load a 8x8 block of weights.
- const int8x16_t vb0123x0123 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
- const int8x16_t vb0123x4567 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
- const int8x16_t vb4567x0123 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
- const int8x16_t vb4567x4567 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
-
- // Multiply-accumulate: 12x8 * 8x8 --> 12x8.
- vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
- vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
- vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb0123x0123, va1x01234567, 0);
- vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb0123x4567, va1x01234567, 0);
- vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb0123x0123, va2x01234567, 0);
- vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb0123x4567, va2x01234567, 0);
- vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb0123x0123, va3x01234567, 0);
- vacc3x4567 = vdotq_lane_s32(vacc3x4567, vb0123x4567, va3x01234567, 0);
- vacc4x0123 = vdotq_lane_s32(vacc4x0123, vb0123x0123, va4x01234567, 0);
- vacc4x4567 = vdotq_lane_s32(vacc4x4567, vb0123x4567, va4x01234567, 0);
- vacc5x0123 = vdotq_lane_s32(vacc5x0123, vb0123x0123, va5x01234567, 0);
- vacc5x4567 = vdotq_lane_s32(vacc5x4567, vb0123x4567, va5x01234567, 0);
- vacc6x0123 = vdotq_lane_s32(vacc6x0123, vb0123x0123, va6x01234567, 0);
- vacc6x4567 = vdotq_lane_s32(vacc6x4567, vb0123x4567, va6x01234567, 0);
- vacc7x0123 = vdotq_lane_s32(vacc7x0123, vb0123x0123, va7x01234567, 0);
- vacc7x4567 = vdotq_lane_s32(vacc7x4567, vb0123x4567, va7x01234567, 0);
- vacc8x0123 = vdotq_lane_s32(vacc8x0123, vb0123x0123, va8x01234567, 0);
- vacc8x4567 = vdotq_lane_s32(vacc8x4567, vb0123x4567, va8x01234567, 0);
- vacc9x0123 = vdotq_lane_s32(vacc9x0123, vb0123x0123, va9x01234567, 0);
- vacc9x4567 = vdotq_lane_s32(vacc9x4567, vb0123x4567, va9x01234567, 0);
- vacc10x0123 = vdotq_lane_s32(vacc10x0123, vb0123x0123, va10x01234567, 0);
- vacc10x4567 = vdotq_lane_s32(vacc10x4567, vb0123x4567, va10x01234567, 0);
- vacc11x0123 = vdotq_lane_s32(vacc11x0123, vb0123x0123, va11x01234567, 0);
- vacc11x4567 = vdotq_lane_s32(vacc11x4567, vb0123x4567, va11x01234567, 0);
- vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb4567x0123, va0x01234567, 1);
- vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb4567x4567, va0x01234567, 1);
- vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb4567x0123, va1x01234567, 1);
- vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb4567x4567, va1x01234567, 1);
- vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb4567x0123, va2x01234567, 1);
- vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb4567x4567, va2x01234567, 1);
- vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb4567x0123, va3x01234567, 1);
- vacc3x4567 = vdotq_lane_s32(vacc3x4567, vb4567x4567, va3x01234567, 1);
- vacc4x0123 = vdotq_lane_s32(vacc4x0123, vb4567x0123, va4x01234567, 1);
- vacc4x4567 = vdotq_lane_s32(vacc4x4567, vb4567x4567, va4x01234567, 1);
- vacc5x0123 = vdotq_lane_s32(vacc5x0123, vb4567x0123, va5x01234567, 1);
- vacc5x4567 = vdotq_lane_s32(vacc5x4567, vb4567x4567, va5x01234567, 1);
- vacc6x0123 = vdotq_lane_s32(vacc6x0123, vb4567x0123, va6x01234567, 1);
- vacc6x4567 = vdotq_lane_s32(vacc6x4567, vb4567x4567, va6x01234567, 1);
- vacc7x0123 = vdotq_lane_s32(vacc7x0123, vb4567x0123, va7x01234567, 1);
- vacc7x4567 = vdotq_lane_s32(vacc7x4567, vb4567x4567, va7x01234567, 1);
- vacc8x0123 = vdotq_lane_s32(vacc8x0123, vb4567x0123, va8x01234567, 1);
- vacc8x4567 = vdotq_lane_s32(vacc8x4567, vb4567x4567, va8x01234567, 1);
- vacc9x0123 = vdotq_lane_s32(vacc9x0123, vb4567x0123, va9x01234567, 1);
- vacc9x4567 = vdotq_lane_s32(vacc9x4567, vb4567x4567, va9x01234567, 1);
- vacc10x0123 = vdotq_lane_s32(vacc10x0123, vb4567x0123, va10x01234567, 1);
- vacc10x4567 = vdotq_lane_s32(vacc10x4567, vb4567x4567, va10x01234567, 1);
- vacc11x0123 = vdotq_lane_s32(vacc11x0123, vb4567x0123, va11x01234567, 1);
- vacc11x4567 = vdotq_lane_s32(vacc11x4567, vb4567x4567, va11x01234567, 1);
-
- k -= 8 * sizeof(int8_t);
- }
- // Handle up to 4 final positions of `k`
- if XNN_UNLIKELY(k != 0) {
- // Load a 12x4 block of activations.
- const int8x8_t va0x01234567 = vld1_s8(a0);
- const int8x8_t va1x01234567 = vld1_s8(a1);
- const int8x8_t va2x01234567 = vld1_s8(a2);
- const int8x8_t va3x01234567 = vld1_s8(a3);
- const int8x8_t va4x01234567 = vld1_s8(a4);
- const int8x8_t va5x01234567 = vld1_s8(a5);
- const int8x8_t va6x01234567 = vld1_s8(a6);
- const int8x8_t va7x01234567 = vld1_s8(a7);
- const int8x8_t va8x01234567 = vld1_s8(a8);
- const int8x8_t va9x01234567 = vld1_s8(a9);
- const int8x8_t va10x01234567 = vld1_s8(a10);
- const int8x8_t va11x01234567 = vld1_s8(a11);
-
- // Load a 4x8 block of weights.
- const int8x16_t vb0123x0123 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
- const int8x16_t vb0123x4567 = vld1q_s8(w); w = (const void*) ((const int8_t*) w + 16);
-
- // Multiply-accumulate: 12x4 * 4x8 --> 12x8.
- vacc0x0123 = vdotq_lane_s32(vacc0x0123, vb0123x0123, va0x01234567, 0);
- vacc0x4567 = vdotq_lane_s32(vacc0x4567, vb0123x4567, va0x01234567, 0);
- vacc1x0123 = vdotq_lane_s32(vacc1x0123, vb0123x0123, va1x01234567, 0);
- vacc1x4567 = vdotq_lane_s32(vacc1x4567, vb0123x4567, va1x01234567, 0);
- vacc2x0123 = vdotq_lane_s32(vacc2x0123, vb0123x0123, va2x01234567, 0);
- vacc2x4567 = vdotq_lane_s32(vacc2x4567, vb0123x4567, va2x01234567, 0);
- vacc3x0123 = vdotq_lane_s32(vacc3x0123, vb0123x0123, va3x01234567, 0);
- vacc3x4567 = vdotq_lane_s32(vacc3x4567, vb0123x4567, va3x01234567, 0);
- vacc4x0123 = vdotq_lane_s32(vacc4x0123, vb0123x0123, va4x01234567, 0);
- vacc4x4567 = vdotq_lane_s32(vacc4x4567, vb0123x4567, va4x01234567, 0);
- vacc5x0123 = vdotq_lane_s32(vacc5x0123, vb0123x0123, va5x01234567, 0);
- vacc5x4567 = vdotq_lane_s32(vacc5x4567, vb0123x4567, va5x01234567, 0);
- vacc6x0123 = vdotq_lane_s32(vacc6x0123, vb0123x0123, va6x01234567, 0);
- vacc6x4567 = vdotq_lane_s32(vacc6x4567, vb0123x4567, va6x01234567, 0);
- vacc7x0123 = vdotq_lane_s32(vacc7x0123, vb0123x0123, va7x01234567, 0);
- vacc7x4567 = vdotq_lane_s32(vacc7x4567, vb0123x4567, va7x01234567, 0);
- vacc8x0123 = vdotq_lane_s32(vacc8x0123, vb0123x0123, va8x01234567, 0);
- vacc8x4567 = vdotq_lane_s32(vacc8x4567, vb0123x4567, va8x01234567, 0);
- vacc9x0123 = vdotq_lane_s32(vacc9x0123, vb0123x0123, va9x01234567, 0);
- vacc9x4567 = vdotq_lane_s32(vacc9x4567, vb0123x4567, va9x01234567, 0);
- vacc10x0123 = vdotq_lane_s32(vacc10x0123, vb0123x0123, va10x01234567, 0);
- vacc10x4567 = vdotq_lane_s32(vacc10x4567, vb0123x4567, va10x01234567, 0);
- vacc11x0123 = vdotq_lane_s32(vacc11x0123, vb0123x0123, va11x01234567, 0);
- vacc11x4567 = vdotq_lane_s32(vacc11x4567, vb0123x4567, va11x01234567, 0);
- }
- p -= 12 * sizeof(void*);
- } while (p != 0);
-
- const int32x4_t vmultiplier = vld1q_dup_s32(¶ms->neon.multiplier);
- vacc0x0123 = vqrdmulhq_s32(vacc0x0123, vmultiplier);
- vacc0x4567 = vqrdmulhq_s32(vacc0x4567, vmultiplier);
- vacc1x0123 = vqrdmulhq_s32(vacc1x0123, vmultiplier);
- vacc1x4567 = vqrdmulhq_s32(vacc1x4567, vmultiplier);
- vacc2x0123 = vqrdmulhq_s32(vacc2x0123, vmultiplier);
- vacc2x4567 = vqrdmulhq_s32(vacc2x4567, vmultiplier);
- vacc3x0123 = vqrdmulhq_s32(vacc3x0123, vmultiplier);
- vacc3x4567 = vqrdmulhq_s32(vacc3x4567, vmultiplier);
- vacc4x0123 = vqrdmulhq_s32(vacc4x0123, vmultiplier);
- vacc4x4567 = vqrdmulhq_s32(vacc4x4567, vmultiplier);
- vacc5x0123 = vqrdmulhq_s32(vacc5x0123, vmultiplier);
- vacc5x4567 = vqrdmulhq_s32(vacc5x4567, vmultiplier);
- vacc6x0123 = vqrdmulhq_s32(vacc6x0123, vmultiplier);
- vacc6x4567 = vqrdmulhq_s32(vacc6x4567, vmultiplier);
- vacc7x0123 = vqrdmulhq_s32(vacc7x0123, vmultiplier);
- vacc7x4567 = vqrdmulhq_s32(vacc7x4567, vmultiplier);
- vacc8x0123 = vqrdmulhq_s32(vacc8x0123, vmultiplier);
- vacc8x4567 = vqrdmulhq_s32(vacc8x4567, vmultiplier);
- vacc9x0123 = vqrdmulhq_s32(vacc9x0123, vmultiplier);
- vacc9x4567 = vqrdmulhq_s32(vacc9x4567, vmultiplier);
- vacc10x0123 = vqrdmulhq_s32(vacc10x0123, vmultiplier);
- vacc10x4567 = vqrdmulhq_s32(vacc10x4567, vmultiplier);
- vacc11x0123 = vqrdmulhq_s32(vacc11x0123, vmultiplier);
- vacc11x4567 = vqrdmulhq_s32(vacc11x4567, vmultiplier);
-
- const int32x4_t vright_shift = vld1q_dup_s32(¶ms->neon.right_shift);
- const int32x4_t vzero_shift_mask = vreinterpretq_s32_u32(vceqq_s32(vright_shift, vmovq_n_s32(0)));
- vacc0x0123 = vsraq_n_s32(vacc0x0123, vbicq_s32(vacc0x0123, vzero_shift_mask), 31);
- vacc0x4567 = vsraq_n_s32(vacc0x4567, vbicq_s32(vacc0x4567, vzero_shift_mask), 31);
- vacc1x0123 = vsraq_n_s32(vacc1x0123, vbicq_s32(vacc1x0123, vzero_shift_mask), 31);
- vacc1x4567 = vsraq_n_s32(vacc1x4567, vbicq_s32(vacc1x4567, vzero_shift_mask), 31);
- vacc2x0123 = vsraq_n_s32(vacc2x0123, vbicq_s32(vacc2x0123, vzero_shift_mask), 31);
- vacc2x4567 = vsraq_n_s32(vacc2x4567, vbicq_s32(vacc2x4567, vzero_shift_mask), 31);
- vacc3x0123 = vsraq_n_s32(vacc3x0123, vbicq_s32(vacc3x0123, vzero_shift_mask), 31);
- vacc3x4567 = vsraq_n_s32(vacc3x4567, vbicq_s32(vacc3x4567, vzero_shift_mask), 31);
- vacc4x0123 = vsraq_n_s32(vacc4x0123, vbicq_s32(vacc4x0123, vzero_shift_mask), 31);
- vacc4x4567 = vsraq_n_s32(vacc4x4567, vbicq_s32(vacc4x4567, vzero_shift_mask), 31);
- vacc5x0123 = vsraq_n_s32(vacc5x0123, vbicq_s32(vacc5x0123, vzero_shift_mask), 31);
- vacc5x4567 = vsraq_n_s32(vacc5x4567, vbicq_s32(vacc5x4567, vzero_shift_mask), 31);
- vacc6x0123 = vsraq_n_s32(vacc6x0123, vbicq_s32(vacc6x0123, vzero_shift_mask), 31);
- vacc6x4567 = vsraq_n_s32(vacc6x4567, vbicq_s32(vacc6x4567, vzero_shift_mask), 31);
- vacc7x0123 = vsraq_n_s32(vacc7x0123, vbicq_s32(vacc7x0123, vzero_shift_mask), 31);
- vacc7x4567 = vsraq_n_s32(vacc7x4567, vbicq_s32(vacc7x4567, vzero_shift_mask), 31);
- vacc8x0123 = vsraq_n_s32(vacc8x0123, vbicq_s32(vacc8x0123, vzero_shift_mask), 31);
- vacc8x4567 = vsraq_n_s32(vacc8x4567, vbicq_s32(vacc8x4567, vzero_shift_mask), 31);
- vacc9x0123 = vsraq_n_s32(vacc9x0123, vbicq_s32(vacc9x0123, vzero_shift_mask), 31);
- vacc9x4567 = vsraq_n_s32(vacc9x4567, vbicq_s32(vacc9x4567, vzero_shift_mask), 31);
- vacc10x0123 = vsraq_n_s32(vacc10x0123, vbicq_s32(vacc10x0123, vzero_shift_mask), 31);
- vacc10x4567 = vsraq_n_s32(vacc10x4567, vbicq_s32(vacc10x4567, vzero_shift_mask), 31);
- vacc11x0123 = vsraq_n_s32(vacc11x0123, vbicq_s32(vacc11x0123, vzero_shift_mask), 31);
- vacc11x4567 = vsraq_n_s32(vacc11x4567, vbicq_s32(vacc11x4567, vzero_shift_mask), 31);
-
- vacc0x0123 = vrshlq_s32(vacc0x0123, vright_shift);
- vacc0x4567 = vrshlq_s32(vacc0x4567, vright_shift);
- vacc1x0123 = vrshlq_s32(vacc1x0123, vright_shift);
- vacc1x4567 = vrshlq_s32(vacc1x4567, vright_shift);
- vacc2x0123 = vrshlq_s32(vacc2x0123, vright_shift);
- vacc2x4567 = vrshlq_s32(vacc2x4567, vright_shift);
- vacc3x0123 = vrshlq_s32(vacc3x0123, vright_shift);
- vacc3x4567 = vrshlq_s32(vacc3x4567, vright_shift);
- vacc4x0123 = vrshlq_s32(vacc4x0123, vright_shift);
- vacc4x4567 = vrshlq_s32(vacc4x4567, vright_shift);
- vacc5x0123 = vrshlq_s32(vacc5x0123, vright_shift);
- vacc5x4567 = vrshlq_s32(vacc5x4567, vright_shift);
- vacc6x0123 = vrshlq_s32(vacc6x0123, vright_shift);
- vacc6x4567 = vrshlq_s32(vacc6x4567, vright_shift);
- vacc7x0123 = vrshlq_s32(vacc7x0123, vright_shift);
- vacc7x4567 = vrshlq_s32(vacc7x4567, vright_shift);
- vacc8x0123 = vrshlq_s32(vacc8x0123, vright_shift);
- vacc8x4567 = vrshlq_s32(vacc8x4567, vright_shift);
- vacc9x0123 = vrshlq_s32(vacc9x0123, vright_shift);
- vacc9x4567 = vrshlq_s32(vacc9x4567, vright_shift);
- vacc10x0123 = vrshlq_s32(vacc10x0123, vright_shift);
- vacc10x4567 = vrshlq_s32(vacc10x4567, vright_shift);
- vacc11x0123 = vrshlq_s32(vacc11x0123, vright_shift);
- vacc11x4567 = vrshlq_s32(vacc11x4567, vright_shift);
-
- const int16x8_t voutput_zero_point = vld1q_dup_s16(¶ms->neon.output_zero_point);
-#if XNN_ARCH_ARM64
- const int16x8_t vacc0x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc0x0123), vacc0x4567), voutput_zero_point);
- const int16x8_t vacc1x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc1x0123), vacc1x4567), voutput_zero_point);
- const int16x8_t vacc2x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc2x0123), vacc2x4567), voutput_zero_point);
- const int16x8_t vacc3x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc3x0123), vacc3x4567), voutput_zero_point);
- const int16x8_t vacc4x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc4x0123), vacc4x4567), voutput_zero_point);
- const int16x8_t vacc5x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc5x0123), vacc5x4567), voutput_zero_point);
- const int16x8_t vacc6x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc6x0123), vacc6x4567), voutput_zero_point);
- const int16x8_t vacc7x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc7x0123), vacc7x4567), voutput_zero_point);
- const int16x8_t vacc8x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc8x0123), vacc8x4567), voutput_zero_point);
- const int16x8_t vacc9x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc9x0123), vacc9x4567), voutput_zero_point);
- const int16x8_t vacc10x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc10x0123), vacc10x4567), voutput_zero_point);
- const int16x8_t vacc11x01234567 = vqaddq_s16(vqmovn_high_s32(vqmovn_s32(vacc11x0123), vacc11x4567), voutput_zero_point);
-
- int8x16_t vout0x01234567_1x01234567 = vqmovn_high_s16(vqmovn_s16(vacc0x01234567), vacc1x01234567);
- int8x16_t vout2x01234567_3x01234567 = vqmovn_high_s16(vqmovn_s16(vacc2x01234567), vacc3x01234567);
- int8x16_t vout4x01234567_5x01234567 = vqmovn_high_s16(vqmovn_s16(vacc4x01234567), vacc5x01234567);
- int8x16_t vout6x01234567_7x01234567 = vqmovn_high_s16(vqmovn_s16(vacc6x01234567), vacc7x01234567);
- int8x16_t vout8x01234567_9x01234567 = vqmovn_high_s16(vqmovn_s16(vacc8x01234567), vacc9x01234567);
- int8x16_t vout10x01234567_11x01234567 = vqmovn_high_s16(vqmovn_s16(vacc10x01234567), vacc11x01234567);
-#else
- const int16x8_t vacc0x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc0x0123), vqmovn_s32(vacc0x4567)), voutput_zero_point);
- const int16x8_t vacc1x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc1x0123), vqmovn_s32(vacc1x4567)), voutput_zero_point);
- const int16x8_t vacc2x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc2x0123), vqmovn_s32(vacc2x4567)), voutput_zero_point);
- const int16x8_t vacc3x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc3x0123), vqmovn_s32(vacc3x4567)), voutput_zero_point);
- const int16x8_t vacc4x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc4x0123), vqmovn_s32(vacc4x4567)), voutput_zero_point);
- const int16x8_t vacc5x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc5x0123), vqmovn_s32(vacc5x4567)), voutput_zero_point);
- const int16x8_t vacc6x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc6x0123), vqmovn_s32(vacc6x4567)), voutput_zero_point);
- const int16x8_t vacc7x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc7x0123), vqmovn_s32(vacc7x4567)), voutput_zero_point);
- const int16x8_t vacc8x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc8x0123), vqmovn_s32(vacc8x4567)), voutput_zero_point);
- const int16x8_t vacc9x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc9x0123), vqmovn_s32(vacc9x4567)), voutput_zero_point);
- const int16x8_t vacc10x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc10x0123), vqmovn_s32(vacc10x4567)), voutput_zero_point);
- const int16x8_t vacc11x01234567 = vqaddq_s16(vcombine_s16(vqmovn_s32(vacc11x0123), vqmovn_s32(vacc11x4567)), voutput_zero_point);
-
- int8x16_t vout0x01234567_1x01234567 = vcombine_s8(vqmovn_s16(vacc0x01234567), vqmovn_s16(vacc1x01234567));
- int8x16_t vout2x01234567_3x01234567 = vcombine_s8(vqmovn_s16(vacc2x01234567), vqmovn_s16(vacc3x01234567));
- int8x16_t vout4x01234567_5x01234567 = vcombine_s8(vqmovn_s16(vacc4x01234567), vqmovn_s16(vacc5x01234567));
- int8x16_t vout6x01234567_7x01234567 = vcombine_s8(vqmovn_s16(vacc6x01234567), vqmovn_s16(vacc7x01234567));
- int8x16_t vout8x01234567_9x01234567 = vcombine_s8(vqmovn_s16(vacc8x01234567), vqmovn_s16(vacc9x01234567));
- int8x16_t vout10x01234567_11x01234567 = vcombine_s8(vqmovn_s16(vacc10x01234567), vqmovn_s16(vacc11x01234567));
-#endif
- const int8x16_t voutput_min = vld1q_dup_s8(¶ms->neon.output_min);
- const int8x16_t voutput_max = vld1q_dup_s8(¶ms->neon.output_max);
-
- vout10x01234567_11x01234567 = vmaxq_s8(vout10x01234567_11x01234567, voutput_min);
- vout8x01234567_9x01234567 = vmaxq_s8(vout8x01234567_9x01234567, voutput_min);
- vout6x01234567_7x01234567 = vmaxq_s8(vout6x01234567_7x01234567, voutput_min);
- vout4x01234567_5x01234567 = vmaxq_s8(vout4x01234567_5x01234567, voutput_min);
- vout2x01234567_3x01234567 = vmaxq_s8(vout2x01234567_3x01234567, voutput_min);
- vout0x01234567_1x01234567 = vmaxq_s8(vout0x01234567_1x01234567, voutput_min);
-
- vout10x01234567_11x01234567 = vminq_s8(vout10x01234567_11x01234567, voutput_max);
- vout8x01234567_9x01234567 = vminq_s8(vout8x01234567_9x01234567, voutput_max);
- vout6x01234567_7x01234567 = vminq_s8(vout6x01234567_7x01234567, voutput_max);
- vout4x01234567_5x01234567 = vminq_s8(vout4x01234567_5x01234567, voutput_max);
- vout2x01234567_3x01234567 = vminq_s8(vout2x01234567_3x01234567, voutput_max);
- vout0x01234567_1x01234567 = vminq_s8(vout0x01234567_1x01234567, voutput_max);
-
- if (nc >= 8) {
- vst1_s8(c11 + 0, vget_high_s8(vout10x01234567_11x01234567));
- vst1_s8(c10 + 0, vget_low_s8(vout10x01234567_11x01234567));
- vst1_s8(c9 + 0, vget_high_s8(vout8x01234567_9x01234567));
- vst1_s8(c8 + 0, vget_low_s8(vout8x01234567_9x01234567));
- vst1_s8(c7 + 0, vget_high_s8(vout6x01234567_7x01234567));
- vst1_s8(c6 + 0, vget_low_s8(vout6x01234567_7x01234567));
- vst1_s8(c5 + 0, vget_high_s8(vout4x01234567_5x01234567));
- vst1_s8(c4 + 0, vget_low_s8(vout4x01234567_5x01234567));
- vst1_s8(c3 + 0, vget_high_s8(vout2x01234567_3x01234567));
- vst1_s8(c2 + 0, vget_low_s8(vout2x01234567_3x01234567));
- vst1_s8(c1 + 0, vget_high_s8(vout0x01234567_1x01234567));
- vst1_s8(c0 + 0, vget_low_s8(vout0x01234567_1x01234567));
-
- c11 = (int8_t*) ((uintptr_t) c11 + cn_stride);
- c10 = (int8_t*) ((uintptr_t) c10 + cn_stride);
- c9 = (int8_t*) ((uintptr_t) c9 + cn_stride);
- c8 = (int8_t*) ((uintptr_t) c8 + cn_stride);
- c7 = (int8_t*) ((uintptr_t) c7 + cn_stride);
- c6 = (int8_t*) ((uintptr_t) c6 + cn_stride);
- c5 = (int8_t*) ((uintptr_t) c5 + cn_stride);
- c4 = (int8_t*) ((uintptr_t) c4 + cn_stride);
- c3 = (int8_t*) ((uintptr_t) c3 + cn_stride);
- c2 = (int8_t*) ((uintptr_t) c2 + cn_stride);
- c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
- c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
-
- a = (const int8_t**restrict) ((uintptr_t) a - ks);
-
- nc -= 8;
- } else {
- if (nc & 4) {
- vst1q_lane_u32(__builtin_assume_aligned(c11, 1), vreinterpretq_u32_s8(vout10x01234567_11x01234567), 2); c11 += 4;
- vst1q_lane_u32(__builtin_assume_aligned(c10, 1), vreinterpretq_u32_s8(vout10x01234567_11x01234567), 0); c10 += 4;
- vst1q_lane_u32(__builtin_assume_aligned(c9, 1), vreinterpretq_u32_s8(vout8x01234567_9x01234567), 2); c9 += 4;
- vst1q_lane_u32(__builtin_assume_aligned(c8, 1), vreinterpretq_u32_s8(vout8x01234567_9x01234567), 0); c8 += 4;
- vst1q_lane_u32(__builtin_assume_aligned(c7, 1), vreinterpretq_u32_s8(vout6x01234567_7x01234567), 2); c7 += 4;
- vst1q_lane_u32(__builtin_assume_aligned(c6, 1), vreinterpretq_u32_s8(vout6x01234567_7x01234567), 0); c6 += 4;
- vst1q_lane_u32(__builtin_assume_aligned(c5, 1), vreinterpretq_u32_s8(vout4x01234567_5x01234567), 2); c5 += 4;
- vst1q_lane_u32(__builtin_assume_aligned(c4, 1), vreinterpretq_u32_s8(vout4x01234567_5x01234567), 0); c4 += 4;
- vst1q_lane_u32(__builtin_assume_aligned(c3, 1), vreinterpretq_u32_s8(vout2x01234567_3x01234567), 2); c3 += 4;
- vst1q_lane_u32(__builtin_assume_aligned(c2, 1), vreinterpretq_u32_s8(vout2x01234567_3x01234567), 0); c2 += 4;
- vst1q_lane_u32(__builtin_assume_aligned(c1, 1), vreinterpretq_u32_s8(vout0x01234567_1x01234567), 2); c1 += 4;
- vst1q_lane_u32(__builtin_assume_aligned(c0, 1), vreinterpretq_u32_s8(vout0x01234567_1x01234567), 0); c0 += 4;
- vout10x01234567_11x01234567 = vextq_s8(vout10x01234567_11x01234567, vout10x01234567_11x01234567, 4);
- vout8x01234567_9x01234567 = vextq_s8(vout8x01234567_9x01234567, vout8x01234567_9x01234567, 4);
- vout6x01234567_7x01234567 = vextq_s8(vout6x01234567_7x01234567, vout6x01234567_7x01234567, 4);
- vout4x01234567_5x01234567 = vextq_s8(vout4x01234567_5x01234567, vout4x01234567_5x01234567, 4);
- vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 4);
- vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 4);
- }
- if (nc & 2) {
- vst1q_lane_u16(__builtin_assume_aligned(c11, 1), vreinterpretq_u16_s8(vout10x01234567_11x01234567), 4); c11 += 2;
- vst1q_lane_u16(__builtin_assume_aligned(c10, 1), vreinterpretq_u16_s8(vout10x01234567_11x01234567), 0); c10 += 2;
- vst1q_lane_u16(__builtin_assume_aligned(c9, 1), vreinterpretq_u16_s8(vout8x01234567_9x01234567), 4); c9 += 2;
- vst1q_lane_u16(__builtin_assume_aligned(c8, 1), vreinterpretq_u16_s8(vout8x01234567_9x01234567), 0); c8 += 2;
- vst1q_lane_u16(__builtin_assume_aligned(c7, 1), vreinterpretq_u16_s8(vout6x01234567_7x01234567), 4); c7 += 2;
- vst1q_lane_u16(__builtin_assume_aligned(c6, 1), vreinterpretq_u16_s8(vout6x01234567_7x01234567), 0); c6 += 2;
- vst1q_lane_u16(__builtin_assume_aligned(c5, 1), vreinterpretq_u16_s8(vout4x01234567_5x01234567), 4); c5 += 2;
- vst1q_lane_u16(__builtin_assume_aligned(c4, 1), vreinterpretq_u16_s8(vout4x01234567_5x01234567), 0); c4 += 2;
- vst1q_lane_u16(__builtin_assume_aligned(c3, 1), vreinterpretq_u16_s8(vout2x01234567_3x01234567), 4); c3 += 2;
- vst1q_lane_u16(__builtin_assume_aligned(c2, 1), vreinterpretq_u16_s8(vout2x01234567_3x01234567), 0); c2 += 2;
- vst1q_lane_u16(__builtin_assume_aligned(c1, 1), vreinterpretq_u16_s8(vout0x01234567_1x01234567), 4); c1 += 2;
- vst1q_lane_u16(__builtin_assume_aligned(c0, 1), vreinterpretq_u16_s8(vout0x01234567_1x01234567), 0); c0 += 2;
- vout10x01234567_11x01234567 = vextq_s8(vout10x01234567_11x01234567, vout10x01234567_11x01234567, 2);
- vout8x01234567_9x01234567 = vextq_s8(vout8x01234567_9x01234567, vout8x01234567_9x01234567, 2);
- vout6x01234567_7x01234567 = vextq_s8(vout6x01234567_7x01234567, vout6x01234567_7x01234567, 2);
- vout4x01234567_5x01234567 = vextq_s8(vout4x01234567_5x01234567, vout4x01234567_5x01234567, 2);
- vout2x01234567_3x01234567 = vextq_s8(vout2x01234567_3x01234567, vout2x01234567_3x01234567, 2);
- vout0x01234567_1x01234567 = vextq_s8(vout0x01234567_1x01234567, vout0x01234567_1x01234567, 2);
- }
- if (nc & 1) {
- vst1q_lane_s8(c11, vout10x01234567_11x01234567, 8);
- vst1q_lane_s8(c10, vout10x01234567_11x01234567, 0);
- vst1q_lane_s8(c9, vout8x01234567_9x01234567, 8);
- vst1q_lane_s8(c8, vout8x01234567_9x01234567, 0);
- vst1q_lane_s8(c7, vout6x01234567_7x01234567, 8);
- vst1q_lane_s8(c6, vout6x01234567_7x01234567, 0);
- vst1q_lane_s8(c5, vout4x01234567_5x01234567, 8);
- vst1q_lane_s8(c4, vout4x01234567_5x01234567, 0);
- vst1q_lane_s8(c3, vout2x01234567_3x01234567, 8);
- vst1q_lane_s8(c2, vout2x01234567_3x01234567, 0);
- vst1q_lane_s8(c1, vout0x01234567_1x01234567, 8);
- vst1q_lane_s8(c0, vout0x01234567_1x01234567, 0);
- }
-
- nc = 0;
- }
- } while (nc != 0);
-}
diff --git a/src/xnnpack/gemm.h b/src/xnnpack/gemm.h
index bb94b8a..f0da3b5 100644
--- a/src/xnnpack/gemm.h
+++ b/src/xnnpack/gemm.h
@@ -599,7 +599,6 @@
DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_ukernel_4x8c4__neondot)
DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_ukernel_6x8c4__neondot)
DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_ukernel_8x8c4__neondot)
-DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_ukernel_12x8c4__neondot)
DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_ukernel_1x16c4__neondot)
DECLARE_QS8_GEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot)
diff --git a/src/xnnpack/igemm.h b/src/xnnpack/igemm.h
index fa418a5..219bd23 100644
--- a/src/xnnpack/igemm.h
+++ b/src/xnnpack/igemm.h
@@ -401,7 +401,6 @@
DECLARE_QS8_IGEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_igemm_minmax_ukernel_4x8c4__neondot)
DECLARE_QS8_IGEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_igemm_minmax_ukernel_6x8c4__neondot)
DECLARE_QS8_IGEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_igemm_minmax_ukernel_8x8c4__neondot)
-DECLARE_QS8_IGEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_igemm_minmax_ukernel_12x8c4__neondot)
DECLARE_QS8_IGEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_igemm_minmax_ukernel_1x16c4__neondot)
DECLARE_QS8_IGEMM_MINMAX_UKERNEL_FUNCTION(xnn_qs8_igemm_minmax_ukernel_4x16c4__neondot)
diff --git a/test/qs8-gemm-minmax.cc b/test/qs8-gemm-minmax.cc
index 68cd1b1..4b8a491 100644
--- a/test/qs8-gemm-minmax.cc
+++ b/test/qs8-gemm-minmax.cc
@@ -27383,462 +27383,6 @@
#if XNN_ARCH_ARM || XNN_ARCH_ARM64
- TEST(QS8_GEMM_MINMAX_12X8C4__NEONDOT, k_eq_8) {
- TEST_REQUIRES_ARM_NEON_DOT;
- GemmMicrokernelTester()
- .mr(12)
- .nr(8)
- .kr(4)
- .sr(1)
- .m(12)
- .n(8)
- .k(8)
- .Test(xnn_qs8_gemm_minmax_ukernel_12x8c4__neondot);
- }
-
- TEST(QS8_GEMM_MINMAX_12X8C4__NEONDOT, strided_cn) {
- TEST_REQUIRES_ARM_NEON_DOT;
- GemmMicrokernelTester()
- .mr(12)
- .nr(8)
- .kr(4)
- .sr(1)
- .m(12)
- .n(8)
- .k(8)
- .cn_stride(11)
- .Test(xnn_qs8_gemm_minmax_ukernel_12x8c4__neondot);
- }
-
- TEST(QS8_GEMM_MINMAX_12X8C4__NEONDOT, k_eq_8_strided_a) {
- TEST_REQUIRES_ARM_NEON_DOT;
- GemmMicrokernelTester()
- .mr(12)
- .nr(8)
- .kr(4)
- .sr(1)
- .m(12)
- .n(8)
- .k(8)
- .a_stride(11)
- .Test(xnn_qs8_gemm_minmax_ukernel_12x8c4__neondot);
- }
-
- TEST(QS8_GEMM_MINMAX_12X8C4__NEONDOT, k_eq_8_subtile) {
- TEST_REQUIRES_ARM_NEON_DOT;
- for (uint32_t m = 1; m <= 12; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(12)
- .nr(8)
- .kr(4)
- .sr(1)
- .m(m)
- .n(n)
- .k(8)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_ukernel_12x8c4__neondot);
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_12X8C4__NEONDOT, k_eq_8_subtile_m) {
- TEST_REQUIRES_ARM_NEON_DOT;
- for (uint32_t m = 1; m <= 12; m++) {
- GemmMicrokernelTester()
- .mr(12)
- .nr(8)
- .kr(4)
- .sr(1)
- .m(m)
- .n(8)
- .k(8)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_ukernel_12x8c4__neondot);
- }
- }
-
- TEST(QS8_GEMM_MINMAX_12X8C4__NEONDOT, k_eq_8_subtile_n) {
- TEST_REQUIRES_ARM_NEON_DOT;
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(12)
- .nr(8)
- .kr(4)
- .sr(1)
- .m(12)
- .n(n)
- .k(8)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_ukernel_12x8c4__neondot);
- }
- }
-
- TEST(QS8_GEMM_MINMAX_12X8C4__NEONDOT, k_lt_8) {
- TEST_REQUIRES_ARM_NEON_DOT;
- for (size_t k = 1; k < 8; k++) {
- GemmMicrokernelTester()
- .mr(12)
- .nr(8)
- .kr(4)
- .sr(1)
- .m(12)
- .n(8)
- .k(k)
- .Test(xnn_qs8_gemm_minmax_ukernel_12x8c4__neondot);
- }
- }
-
- TEST(QS8_GEMM_MINMAX_12X8C4__NEONDOT, k_lt_8_strided_a) {
- TEST_REQUIRES_ARM_NEON_DOT;
- for (size_t k = 1; k < 8; k++) {
- GemmMicrokernelTester()
- .mr(12)
- .nr(8)
- .kr(4)
- .sr(1)
- .m(12)
- .n(8)
- .k(k)
- .a_stride(11)
- .Test(xnn_qs8_gemm_minmax_ukernel_12x8c4__neondot);
- }
- }
-
- TEST(QS8_GEMM_MINMAX_12X8C4__NEONDOT, k_lt_8_subtile) {
- TEST_REQUIRES_ARM_NEON_DOT;
- for (size_t k = 1; k < 8; k++) {
- for (uint32_t m = 1; m <= 12; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(12)
- .nr(8)
- .kr(4)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_ukernel_12x8c4__neondot);
- }
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_12X8C4__NEONDOT, k_gt_8) {
- TEST_REQUIRES_ARM_NEON_DOT;
- for (size_t k = 9; k < 16; k++) {
- GemmMicrokernelTester()
- .mr(12)
- .nr(8)
- .kr(4)
- .sr(1)
- .m(12)
- .n(8)
- .k(k)
- .Test(xnn_qs8_gemm_minmax_ukernel_12x8c4__neondot);
- }
- }
-
- TEST(QS8_GEMM_MINMAX_12X8C4__NEONDOT, k_gt_8_strided_a) {
- TEST_REQUIRES_ARM_NEON_DOT;
- for (size_t k = 9; k < 16; k++) {
- GemmMicrokernelTester()
- .mr(12)
- .nr(8)
- .kr(4)
- .sr(1)
- .m(12)
- .n(8)
- .k(k)
- .a_stride(19)
- .Test(xnn_qs8_gemm_minmax_ukernel_12x8c4__neondot);
- }
- }
-
- TEST(QS8_GEMM_MINMAX_12X8C4__NEONDOT, k_gt_8_subtile) {
- TEST_REQUIRES_ARM_NEON_DOT;
- for (size_t k = 9; k < 16; k++) {
- for (uint32_t m = 1; m <= 12; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(12)
- .nr(8)
- .kr(4)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_ukernel_12x8c4__neondot);
- }
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_12X8C4__NEONDOT, k_div_8) {
- TEST_REQUIRES_ARM_NEON_DOT;
- for (size_t k = 16; k <= 80; k += 8) {
- GemmMicrokernelTester()
- .mr(12)
- .nr(8)
- .kr(4)
- .sr(1)
- .m(12)
- .n(8)
- .k(k)
- .Test(xnn_qs8_gemm_minmax_ukernel_12x8c4__neondot);
- }
- }
-
- TEST(QS8_GEMM_MINMAX_12X8C4__NEONDOT, k_div_8_strided_a) {
- TEST_REQUIRES_ARM_NEON_DOT;
- for (size_t k = 16; k <= 80; k += 8) {
- GemmMicrokernelTester()
- .mr(12)
- .nr(8)
- .kr(4)
- .sr(1)
- .m(12)
- .n(8)
- .k(k)
- .a_stride(83)
- .Test(xnn_qs8_gemm_minmax_ukernel_12x8c4__neondot);
- }
- }
-
- TEST(QS8_GEMM_MINMAX_12X8C4__NEONDOT, k_div_8_subtile) {
- TEST_REQUIRES_ARM_NEON_DOT;
- for (size_t k = 16; k <= 80; k += 8) {
- for (uint32_t m = 1; m <= 12; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(12)
- .nr(8)
- .kr(4)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_ukernel_12x8c4__neondot);
- }
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_12X8C4__NEONDOT, n_gt_8) {
- TEST_REQUIRES_ARM_NEON_DOT;
- for (uint32_t n = 9; n < 16; n++) {
- for (size_t k = 1; k <= 40; k += 9) {
- GemmMicrokernelTester()
- .mr(12)
- .nr(8)
- .kr(4)
- .sr(1)
- .m(12)
- .n(8)
- .k(k)
- .Test(xnn_qs8_gemm_minmax_ukernel_12x8c4__neondot);
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_12X8C4__NEONDOT, n_gt_8_strided_cn) {
- TEST_REQUIRES_ARM_NEON_DOT;
- for (uint32_t n = 9; n < 16; n++) {
- for (size_t k = 1; k <= 40; k += 9) {
- GemmMicrokernelTester()
- .mr(12)
- .nr(8)
- .kr(4)
- .sr(1)
- .m(12)
- .n(8)
- .k(k)
- .cn_stride(11)
- .Test(xnn_qs8_gemm_minmax_ukernel_12x8c4__neondot);
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_12X8C4__NEONDOT, n_gt_8_strided_a) {
- TEST_REQUIRES_ARM_NEON_DOT;
- for (uint32_t n = 9; n < 16; n++) {
- for (size_t k = 1; k <= 40; k += 9) {
- GemmMicrokernelTester()
- .mr(12)
- .nr(8)
- .kr(4)
- .sr(1)
- .m(12)
- .n(n)
- .k(k)
- .a_stride(43)
- .Test(xnn_qs8_gemm_minmax_ukernel_12x8c4__neondot);
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_12X8C4__NEONDOT, n_gt_8_subtile) {
- TEST_REQUIRES_ARM_NEON_DOT;
- for (uint32_t n = 9; n < 16; n++) {
- for (size_t k = 1; k <= 40; k += 9) {
- for (uint32_t m = 1; m <= 12; m++) {
- GemmMicrokernelTester()
- .mr(12)
- .nr(8)
- .kr(4)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_ukernel_12x8c4__neondot);
- }
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_12X8C4__NEONDOT, n_div_8) {
- TEST_REQUIRES_ARM_NEON_DOT;
- for (uint32_t n = 16; n <= 24; n += 8) {
- for (size_t k = 1; k <= 40; k += 9) {
- GemmMicrokernelTester()
- .mr(12)
- .nr(8)
- .kr(4)
- .sr(1)
- .m(12)
- .n(8)
- .k(k)
- .Test(xnn_qs8_gemm_minmax_ukernel_12x8c4__neondot);
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_12X8C4__NEONDOT, n_div_8_strided_cn) {
- TEST_REQUIRES_ARM_NEON_DOT;
- for (uint32_t n = 16; n <= 24; n += 8) {
- for (size_t k = 1; k <= 40; k += 9) {
- GemmMicrokernelTester()
- .mr(12)
- .nr(8)
- .kr(4)
- .sr(1)
- .m(12)
- .n(n)
- .k(k)
- .cn_stride(11)
- .Test(xnn_qs8_gemm_minmax_ukernel_12x8c4__neondot);
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_12X8C4__NEONDOT, n_div_8_strided_a) {
- TEST_REQUIRES_ARM_NEON_DOT;
- for (uint32_t n = 16; n <= 24; n += 8) {
- for (size_t k = 1; k <= 40; k += 9) {
- GemmMicrokernelTester()
- .mr(12)
- .nr(8)
- .kr(4)
- .sr(1)
- .m(12)
- .n(n)
- .k(k)
- .a_stride(43)
- .Test(xnn_qs8_gemm_minmax_ukernel_12x8c4__neondot);
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_12X8C4__NEONDOT, n_div_8_subtile) {
- TEST_REQUIRES_ARM_NEON_DOT;
- for (uint32_t n = 16; n <= 24; n += 8) {
- for (size_t k = 1; k <= 40; k += 9) {
- for (uint32_t m = 1; m <= 12; m++) {
- GemmMicrokernelTester()
- .mr(12)
- .nr(8)
- .kr(4)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_ukernel_12x8c4__neondot);
- }
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_12X8C4__NEONDOT, strided_cm_subtile) {
- TEST_REQUIRES_ARM_NEON_DOT;
- for (size_t k = 1; k <= 40; k += 9) {
- for (uint32_t m = 1; m <= 12; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(12)
- .nr(8)
- .kr(4)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .cm_stride(11)
- .iterations(1)
- .Test(xnn_qs8_gemm_minmax_ukernel_12x8c4__neondot);
- }
- }
- }
- }
-
- TEST(QS8_GEMM_MINMAX_12X8C4__NEONDOT, qmin) {
- TEST_REQUIRES_ARM_NEON_DOT;
- GemmMicrokernelTester()
- .mr(12)
- .nr(8)
- .kr(4)
- .sr(1)
- .m(12)
- .n(8)
- .k(8)
- .qmin(128)
- .Test(xnn_qs8_gemm_minmax_ukernel_12x8c4__neondot);
- }
-
- TEST(QS8_GEMM_MINMAX_12X8C4__NEONDOT, qmax) {
- TEST_REQUIRES_ARM_NEON_DOT;
- GemmMicrokernelTester()
- .mr(12)
- .nr(8)
- .kr(4)
- .sr(1)
- .m(12)
- .n(8)
- .k(8)
- .qmax(128)
- .Test(xnn_qs8_gemm_minmax_ukernel_12x8c4__neondot);
- }
-
- TEST(QS8_GEMM_MINMAX_12X8C4__NEONDOT, strided_cm) {
- TEST_REQUIRES_ARM_NEON_DOT;
- GemmMicrokernelTester()
- .mr(12)
- .nr(8)
- .kr(4)
- .sr(1)
- .m(12)
- .n(8)
- .k(8)
- .cm_stride(11)
- .Test(xnn_qs8_gemm_minmax_ukernel_12x8c4__neondot);
- }
-#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
-
-
-#if XNN_ARCH_ARM || XNN_ARCH_ARM64
TEST(QS8_GEMM_MINMAX_1X16C4__NEONDOT, k_eq_8) {
TEST_REQUIRES_ARM_NEON_DOT;
GemmMicrokernelTester()
diff --git a/test/qs8-gemm-minmax.yaml b/test/qs8-gemm-minmax.yaml
index f2d98c1..e632257 100644
--- a/test/qs8-gemm-minmax.yaml
+++ b/test/qs8-gemm-minmax.yaml
@@ -122,8 +122,6 @@
k-block: 8
- name: xnn_qs8_gemm_minmax_ukernel_8x8c4__neondot
k-block: 8
-- name: xnn_qs8_gemm_minmax_ukernel_12x8c4__neondot
- k-block: 8
- name: xnn_qs8_gemm_minmax_ukernel_1x16c4__neondot
k-block: 8
- name: xnn_qs8_gemm_minmax_ukernel_4x16c4__neondot
diff --git a/test/qs8-igemm-minmax.cc b/test/qs8-igemm-minmax.cc
index 325379e..7314980 100644
--- a/test/qs8-igemm-minmax.cc
+++ b/test/qs8-igemm-minmax.cc
@@ -28103,474 +28103,6 @@
#if XNN_ARCH_ARM || XNN_ARCH_ARM64
- TEST(QS8_IGEMM_MINMAX_12X8C4__NEONDOT, k_eq_8) {
- TEST_REQUIRES_ARM_NEON_DOT;
- GemmMicrokernelTester()
- .mr(12)
- .nr(8)
- .kr(4)
- .sr(1)
- .m(12)
- .n(8)
- .k(8)
- .Test(xnn_qs8_igemm_minmax_ukernel_12x8c4__neondot);
- }
-
- TEST(QS8_IGEMM_MINMAX_12X8C4__NEONDOT, strided_cn) {
- TEST_REQUIRES_ARM_NEON_DOT;
- GemmMicrokernelTester()
- .mr(12)
- .nr(8)
- .kr(4)
- .sr(1)
- .m(12)
- .n(8)
- .k(8)
- .cn_stride(11)
- .Test(xnn_qs8_igemm_minmax_ukernel_12x8c4__neondot);
- }
-
- TEST(QS8_IGEMM_MINMAX_12X8C4__NEONDOT, k_eq_8_subtile) {
- TEST_REQUIRES_ARM_NEON_DOT;
- for (uint32_t m = 1; m <= 12; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(12)
- .nr(8)
- .kr(4)
- .sr(1)
- .m(m)
- .n(n)
- .k(8)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_ukernel_12x8c4__neondot);
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_12X8C4__NEONDOT, k_eq_8_subtile_m) {
- TEST_REQUIRES_ARM_NEON_DOT;
- for (uint32_t m = 1; m <= 12; m++) {
- GemmMicrokernelTester()
- .mr(12)
- .nr(8)
- .kr(4)
- .sr(1)
- .m(m)
- .n(8)
- .k(8)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_ukernel_12x8c4__neondot);
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_12X8C4__NEONDOT, k_eq_8_subtile_n) {
- TEST_REQUIRES_ARM_NEON_DOT;
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(12)
- .nr(8)
- .kr(4)
- .sr(1)
- .m(12)
- .n(n)
- .k(8)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_ukernel_12x8c4__neondot);
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_12X8C4__NEONDOT, k_lt_8) {
- TEST_REQUIRES_ARM_NEON_DOT;
- for (size_t k = 1; k < 8; k++) {
- GemmMicrokernelTester()
- .mr(12)
- .nr(8)
- .kr(4)
- .sr(1)
- .m(12)
- .n(8)
- .k(k)
- .Test(xnn_qs8_igemm_minmax_ukernel_12x8c4__neondot);
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_12X8C4__NEONDOT, k_lt_8_subtile) {
- TEST_REQUIRES_ARM_NEON_DOT;
- for (size_t k = 1; k < 8; k++) {
- for (uint32_t m = 1; m <= 12; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(12)
- .nr(8)
- .kr(4)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_ukernel_12x8c4__neondot);
- }
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_12X8C4__NEONDOT, k_gt_8) {
- TEST_REQUIRES_ARM_NEON_DOT;
- for (size_t k = 9; k < 16; k++) {
- GemmMicrokernelTester()
- .mr(12)
- .nr(8)
- .kr(4)
- .sr(1)
- .m(12)
- .n(8)
- .k(k)
- .Test(xnn_qs8_igemm_minmax_ukernel_12x8c4__neondot);
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_12X8C4__NEONDOT, k_gt_8_subtile) {
- TEST_REQUIRES_ARM_NEON_DOT;
- for (size_t k = 9; k < 16; k++) {
- for (uint32_t m = 1; m <= 12; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(12)
- .nr(8)
- .kr(4)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_ukernel_12x8c4__neondot);
- }
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_12X8C4__NEONDOT, k_div_8) {
- TEST_REQUIRES_ARM_NEON_DOT;
- for (size_t k = 16; k <= 80; k += 8) {
- GemmMicrokernelTester()
- .mr(12)
- .nr(8)
- .kr(4)
- .sr(1)
- .m(12)
- .n(8)
- .k(k)
- .Test(xnn_qs8_igemm_minmax_ukernel_12x8c4__neondot);
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_12X8C4__NEONDOT, k_div_8_subtile) {
- TEST_REQUIRES_ARM_NEON_DOT;
- for (size_t k = 16; k <= 80; k += 8) {
- for (uint32_t m = 1; m <= 12; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(12)
- .nr(8)
- .kr(4)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_ukernel_12x8c4__neondot);
- }
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_12X8C4__NEONDOT, n_gt_8) {
- TEST_REQUIRES_ARM_NEON_DOT;
- for (uint32_t n = 9; n < 16; n++) {
- for (size_t k = 1; k <= 40; k += 9) {
- GemmMicrokernelTester()
- .mr(12)
- .nr(8)
- .kr(4)
- .sr(1)
- .m(12)
- .n(8)
- .k(k)
- .Test(xnn_qs8_igemm_minmax_ukernel_12x8c4__neondot);
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_12X8C4__NEONDOT, n_gt_8_strided_cn) {
- TEST_REQUIRES_ARM_NEON_DOT;
- for (uint32_t n = 9; n < 16; n++) {
- for (size_t k = 1; k <= 40; k += 9) {
- GemmMicrokernelTester()
- .mr(12)
- .nr(8)
- .kr(4)
- .sr(1)
- .m(12)
- .n(8)
- .k(k)
- .cn_stride(11)
- .Test(xnn_qs8_igemm_minmax_ukernel_12x8c4__neondot);
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_12X8C4__NEONDOT, n_gt_8_subtile) {
- TEST_REQUIRES_ARM_NEON_DOT;
- for (uint32_t n = 9; n < 16; n++) {
- for (size_t k = 1; k <= 40; k += 9) {
- for (uint32_t m = 1; m <= 12; m++) {
- GemmMicrokernelTester()
- .mr(12)
- .nr(8)
- .kr(4)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_ukernel_12x8c4__neondot);
- }
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_12X8C4__NEONDOT, n_div_8) {
- TEST_REQUIRES_ARM_NEON_DOT;
- for (uint32_t n = 16; n <= 24; n += 8) {
- for (size_t k = 1; k <= 40; k += 9) {
- GemmMicrokernelTester()
- .mr(12)
- .nr(8)
- .kr(4)
- .sr(1)
- .m(12)
- .n(8)
- .k(k)
- .Test(xnn_qs8_igemm_minmax_ukernel_12x8c4__neondot);
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_12X8C4__NEONDOT, n_div_8_strided_cn) {
- TEST_REQUIRES_ARM_NEON_DOT;
- for (uint32_t n = 16; n <= 24; n += 8) {
- for (size_t k = 1; k <= 40; k += 9) {
- GemmMicrokernelTester()
- .mr(12)
- .nr(8)
- .kr(4)
- .sr(1)
- .m(12)
- .n(n)
- .k(k)
- .cn_stride(11)
- .Test(xnn_qs8_igemm_minmax_ukernel_12x8c4__neondot);
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_12X8C4__NEONDOT, n_div_8_subtile) {
- TEST_REQUIRES_ARM_NEON_DOT;
- for (uint32_t n = 16; n <= 24; n += 8) {
- for (size_t k = 1; k <= 40; k += 9) {
- for (uint32_t m = 1; m <= 12; m++) {
- GemmMicrokernelTester()
- .mr(12)
- .nr(8)
- .kr(4)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_ukernel_12x8c4__neondot);
- }
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_12X8C4__NEONDOT, small_kernel) {
- TEST_REQUIRES_ARM_NEON_DOT;
- for (size_t k = 1; k <= 40; k += 9) {
- GemmMicrokernelTester()
- .mr(12)
- .nr(8)
- .kr(4)
- .sr(1)
- .m(12)
- .n(8)
- .k(k)
- .ks(3)
- .Test(xnn_qs8_igemm_minmax_ukernel_12x8c4__neondot);
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_12X8C4__NEONDOT, small_kernel_subtile) {
- TEST_REQUIRES_ARM_NEON_DOT;
- for (size_t k = 1; k <= 40; k += 9) {
- for (uint32_t m = 1; m <= 12; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(12)
- .nr(8)
- .kr(4)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .ks(3)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_ukernel_12x8c4__neondot);
- }
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_12X8C4__NEONDOT, n_gt_8_small_kernel) {
- TEST_REQUIRES_ARM_NEON_DOT;
- for (uint32_t n = 9; n < 16; n++) {
- for (size_t k = 1; k <= 40; k += 9) {
- GemmMicrokernelTester()
- .mr(12)
- .nr(8)
- .kr(4)
- .sr(1)
- .m(12)
- .n(8)
- .k(k)
- .ks(3)
- .Test(xnn_qs8_igemm_minmax_ukernel_12x8c4__neondot);
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_12X8C4__NEONDOT, n_div_8_small_kernel) {
- TEST_REQUIRES_ARM_NEON_DOT;
- for (uint32_t n = 16; n <= 24; n += 8) {
- for (size_t k = 1; k <= 40; k += 9) {
- GemmMicrokernelTester()
- .mr(12)
- .nr(8)
- .kr(4)
- .sr(1)
- .m(12)
- .n(8)
- .k(k)
- .ks(3)
- .Test(xnn_qs8_igemm_minmax_ukernel_12x8c4__neondot);
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_12X8C4__NEONDOT, strided_cm_subtile) {
- TEST_REQUIRES_ARM_NEON_DOT;
- for (size_t k = 1; k <= 40; k += 9) {
- for (uint32_t m = 1; m <= 12; m++) {
- for (uint32_t n = 1; n <= 8; n++) {
- GemmMicrokernelTester()
- .mr(12)
- .nr(8)
- .kr(4)
- .sr(1)
- .m(m)
- .n(n)
- .k(k)
- .cm_stride(11)
- .iterations(1)
- .Test(xnn_qs8_igemm_minmax_ukernel_12x8c4__neondot);
- }
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_12X8C4__NEONDOT, a_offset) {
- TEST_REQUIRES_ARM_NEON_DOT;
- for (size_t k = 1; k <= 40; k += 9) {
- GemmMicrokernelTester()
- .mr(12)
- .nr(8)
- .kr(4)
- .sr(1)
- .m(12)
- .n(8)
- .k(k)
- .ks(3)
- .a_offset(487)
- .Test(xnn_qs8_igemm_minmax_ukernel_12x8c4__neondot);
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_12X8C4__NEONDOT, zero) {
- TEST_REQUIRES_ARM_NEON_DOT;
- for (uint32_t mz = 0; mz < 12; mz++) {
- for (size_t k = 1; k <= 40; k += 9) {
- GemmMicrokernelTester()
- .mr(12)
- .nr(8)
- .kr(4)
- .sr(1)
- .m(12)
- .n(8)
- .k(k)
- .ks(3)
- .a_offset(487)
- .zero_index(mz)
- .Test(xnn_qs8_igemm_minmax_ukernel_12x8c4__neondot);
- }
- }
- }
-
- TEST(QS8_IGEMM_MINMAX_12X8C4__NEONDOT, qmin) {
- TEST_REQUIRES_ARM_NEON_DOT;
- GemmMicrokernelTester()
- .mr(12)
- .nr(8)
- .kr(4)
- .sr(1)
- .m(12)
- .n(8)
- .k(8)
- .qmin(128)
- .Test(xnn_qs8_igemm_minmax_ukernel_12x8c4__neondot);
- }
-
- TEST(QS8_IGEMM_MINMAX_12X8C4__NEONDOT, qmax) {
- TEST_REQUIRES_ARM_NEON_DOT;
- GemmMicrokernelTester()
- .mr(12)
- .nr(8)
- .kr(4)
- .sr(1)
- .m(12)
- .n(8)
- .k(8)
- .qmax(128)
- .Test(xnn_qs8_igemm_minmax_ukernel_12x8c4__neondot);
- }
-
- TEST(QS8_IGEMM_MINMAX_12X8C4__NEONDOT, strided_cm) {
- TEST_REQUIRES_ARM_NEON_DOT;
- GemmMicrokernelTester()
- .mr(12)
- .nr(8)
- .kr(4)
- .sr(1)
- .m(12)
- .n(8)
- .k(8)
- .cm_stride(11)
- .Test(xnn_qs8_igemm_minmax_ukernel_12x8c4__neondot);
- }
-#endif // XNN_ARCH_ARM || XNN_ARCH_ARM64
-
-
-#if XNN_ARCH_ARM || XNN_ARCH_ARM64
TEST(QS8_IGEMM_MINMAX_1X16C4__NEONDOT, k_eq_8) {
TEST_REQUIRES_ARM_NEON_DOT;
GemmMicrokernelTester()
diff --git a/test/qs8-igemm-minmax.yaml b/test/qs8-igemm-minmax.yaml
index b0e0f57..26f13bb 100644
--- a/test/qs8-igemm-minmax.yaml
+++ b/test/qs8-igemm-minmax.yaml
@@ -122,8 +122,6 @@
k-block: 8
- name: xnn_qs8_igemm_minmax_ukernel_8x8c4__neondot
k-block: 8
-- name: xnn_qs8_igemm_minmax_ukernel_12x8c4__neondot
- k-block: 8
- name: xnn_qs8_igemm_minmax_ukernel_1x16c4__neondot
k-block: 8
- name: xnn_qs8_igemm_minmax_ukernel_4x16c4__neondot