QC8 GEMM/IGEMM microkernels for SSE/AVX/XOP
PiperOrigin-RevId: 377345679
diff --git a/src/qc8-gemm/gen/1x4c2-minmax-fp32-avx-ld128.c b/src/qc8-gemm/gen/1x4c2-minmax-fp32-avx-ld128.c
new file mode 100644
index 0000000..f2e24a7
--- /dev/null
+++ b/src/qc8-gemm/gen/1x4c2-minmax-fp32-avx-ld128.c
@@ -0,0 +1,147 @@
+// Auto-generated file. Do not edit!
+// Template: src/qs8-gemm/MRx4c2-sse.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 <smmintrin.h>
+
+#include <xnnpack/gemm.h>
+#include <xnnpack/math.h>
+
+
+
+void xnn_qc8_gemm_minmax_fp32_ukernel_1x4c2__avx_ld128(
+ 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_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN XNN_DISABLE_MSAN
+{
+ assert(mr != 0);
+ assert(mr <= 1);
+ 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, 2);
+ const int8_t* a0 = a;
+ int8_t* c0 = c;
+
+ do {
+ __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+
+ size_t k = kc;
+ while (k >= 8 * sizeof(int8_t)) {
+ const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
+ const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
+ a0 += 8;
+
+ const __m128i vb01 = _mm_loadu_si128((const __m128i*) w);
+ const __m128i vsb01 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb01);
+ const __m128i vxb0 = _mm_unpacklo_epi8(vb01, vsb01);
+ const __m128i vxb1 = _mm_unpackhi_epi8(vb01, vsb01);
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
+ const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
+ const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
+ const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
+
+ w = (const void*) ((uintptr_t) w + 32 * sizeof(int8_t));
+ k -= 8 * sizeof(int8_t);
+ }
+ if (k != 0) {
+ const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
+ const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
+ a0 = (const int8_t*) ((uintptr_t) a0 + k);
+
+ const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+
+ if (k > 2 * sizeof(int8_t)) {
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+
+ if (k > 4 * sizeof(int8_t)) {
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ }
+ }
+ }
+
+ __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
+
+ const __m128 vscale0123 = _mm_loadu_ps((const float*) w);
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(float));
+ vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123);
+
+ vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
+
+ const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4.output_zero_point);
+ __m128i vacc00x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc0x0123), voutput_zero_point);
+
+
+ __m128i vout = _mm_packs_epi16(vacc00x0123, vacc00x0123);
+
+ vout = _mm_max_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_min));
+ vout = _mm_min_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_max));
+
+ if (nc >= 4) {
+ *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
+
+ c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
+
+ a0 = (const int8_t*) ((uintptr_t) a0 - kc);
+
+ nc -= 4;
+ } else {
+ if (nc & 2) {
+ *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
+ c0 += 2;
+ vout = _mm_srli_epi32(vout, 16);
+ }
+ if (nc & 1) {
+ *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ }
+
+ nc = 0;
+ }
+ } while (nc != 0);
+}
diff --git a/src/qc8-gemm/gen/1x4c2-minmax-fp32-avx-ld64.c b/src/qc8-gemm/gen/1x4c2-minmax-fp32-avx-ld64.c
new file mode 100644
index 0000000..2a04fc8
--- /dev/null
+++ b/src/qc8-gemm/gen/1x4c2-minmax-fp32-avx-ld64.c
@@ -0,0 +1,147 @@
+// Auto-generated file. Do not edit!
+// Template: src/qs8-gemm/MRx4c2-sse.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 <smmintrin.h>
+
+#include <xnnpack/gemm.h>
+#include <xnnpack/math.h>
+
+
+
+void xnn_qc8_gemm_minmax_fp32_ukernel_1x4c2__avx_ld64(
+ 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_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN XNN_DISABLE_MSAN
+{
+ assert(mr != 0);
+ assert(mr <= 1);
+ 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, 2);
+ const int8_t* a0 = a;
+ int8_t* c0 = c;
+
+ do {
+ __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+
+ size_t k = kc;
+ while (k >= 8 * sizeof(int8_t)) {
+ const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
+ const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
+ a0 += 8;
+
+ const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int8_t)));
+ const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
+ const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int8_t)));
+ const __m128i vxb3 = _mm_cvtepi8_epi16(vb3);
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
+
+ w = (const void*) ((uintptr_t) w + 32 * sizeof(int8_t));
+ k -= 8 * sizeof(int8_t);
+ }
+ if (k != 0) {
+ const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
+ const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
+ a0 = (const int8_t*) ((uintptr_t) a0 + k);
+
+ const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+
+ if (k > 2 * sizeof(int8_t)) {
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+
+ if (k > 4 * sizeof(int8_t)) {
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ }
+ }
+ }
+
+ __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
+
+ const __m128 vscale0123 = _mm_loadu_ps((const float*) w);
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(float));
+ vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123);
+
+ vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
+
+ const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4.output_zero_point);
+ __m128i vacc00x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc0x0123), voutput_zero_point);
+
+
+ __m128i vout = _mm_packs_epi16(vacc00x0123, vacc00x0123);
+
+ vout = _mm_max_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_min));
+ vout = _mm_min_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_max));
+
+ if (nc >= 4) {
+ *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
+
+ c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
+
+ a0 = (const int8_t*) ((uintptr_t) a0 - kc);
+
+ nc -= 4;
+ } else {
+ if (nc & 2) {
+ *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
+ c0 += 2;
+ vout = _mm_srli_epi32(vout, 16);
+ }
+ if (nc & 1) {
+ *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ }
+
+ nc = 0;
+ }
+ } while (nc != 0);
+}
diff --git a/src/qc8-gemm/gen/1x4c2-minmax-fp32-sse2-ld128.c b/src/qc8-gemm/gen/1x4c2-minmax-fp32-sse2-ld128.c
new file mode 100644
index 0000000..617fc79
--- /dev/null
+++ b/src/qc8-gemm/gen/1x4c2-minmax-fp32-sse2-ld128.c
@@ -0,0 +1,148 @@
+// Auto-generated file. Do not edit!
+// Template: src/qs8-gemm/MRx4c2-sse.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 <emmintrin.h>
+
+#include <xnnpack/gemm.h>
+#include <xnnpack/math.h>
+
+
+
+void xnn_qc8_gemm_minmax_fp32_ukernel_1x4c2__sse2_ld128(
+ 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_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN XNN_DISABLE_MSAN
+{
+ assert(mr != 0);
+ assert(mr <= 1);
+ 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, 2);
+ const int8_t* a0 = a;
+ int8_t* c0 = c;
+
+ do {
+ __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+
+ size_t k = kc;
+ while (k >= 8 * sizeof(int8_t)) {
+ const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
+ const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ a0 += 8;
+
+ const __m128i vb01 = _mm_loadu_si128((const __m128i*) w);
+ const __m128i vsb01 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb01);
+ const __m128i vxb0 = _mm_unpacklo_epi8(vb01, vsb01);
+ const __m128i vxb1 = _mm_unpackhi_epi8(vb01, vsb01);
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
+ const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
+ const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
+ const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
+
+ w = (const void*) ((uintptr_t) w + 32 * sizeof(int8_t));
+ k -= 8 * sizeof(int8_t);
+ }
+ if (k != 0) {
+ const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
+ const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ a0 = (const int8_t*) ((uintptr_t) a0 + k);
+
+ const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+
+ if (k > 2 * sizeof(int8_t)) {
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+
+ if (k > 4 * sizeof(int8_t)) {
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ }
+ }
+ }
+
+ __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
+
+ const __m128 vscale0123 = _mm_loadu_ps((const float*) w);
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(float));
+ vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123);
+
+ vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
+
+ const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
+ __m128i vacc00x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc0x0123), voutput_zero_point);
+
+ const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
+ const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max);
+ vacc00x0123 = _mm_min_epi16(_mm_max_epi16(vacc00x0123, voutput_min), voutput_max);
+
+ __m128i vout = _mm_packs_epi16(vacc00x0123, vacc00x0123);
+
+
+ if (nc >= 4) {
+ *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
+
+ c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
+
+ a0 = (const int8_t*) ((uintptr_t) a0 - kc);
+
+ nc -= 4;
+ } else {
+ if (nc & 2) {
+ *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
+ c0 += 2;
+ vout = _mm_srli_epi32(vout, 16);
+ }
+ if (nc & 1) {
+ *((int8_t*) c0) = (int8_t) _mm_cvtsi128_si32(vout);
+ }
+
+ nc = 0;
+ }
+ } while (nc != 0);
+}
diff --git a/src/qc8-gemm/gen/1x4c2-minmax-fp32-sse2-ld64.c b/src/qc8-gemm/gen/1x4c2-minmax-fp32-sse2-ld64.c
new file mode 100644
index 0000000..ddc1095
--- /dev/null
+++ b/src/qc8-gemm/gen/1x4c2-minmax-fp32-sse2-ld64.c
@@ -0,0 +1,148 @@
+// Auto-generated file. Do not edit!
+// Template: src/qs8-gemm/MRx4c2-sse.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 <emmintrin.h>
+
+#include <xnnpack/gemm.h>
+#include <xnnpack/math.h>
+
+
+
+void xnn_qc8_gemm_minmax_fp32_ukernel_1x4c2__sse2_ld64(
+ 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_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN XNN_DISABLE_MSAN
+{
+ assert(mr != 0);
+ assert(mr <= 1);
+ 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, 2);
+ const int8_t* a0 = a;
+ int8_t* c0 = c;
+
+ do {
+ __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+
+ size_t k = kc;
+ while (k >= 8 * sizeof(int8_t)) {
+ const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
+ const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ a0 += 8;
+
+ const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int8_t)));
+ const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
+ const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int8_t)));
+ const __m128i vxb3 = _mm_unpacklo_epi8(vb3, _mm_cmpgt_epi8(_mm_setzero_si128(), vb3));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
+
+ w = (const void*) ((uintptr_t) w + 32 * sizeof(int8_t));
+ k -= 8 * sizeof(int8_t);
+ }
+ if (k != 0) {
+ const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
+ const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ a0 = (const int8_t*) ((uintptr_t) a0 + k);
+
+ const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+
+ if (k > 2 * sizeof(int8_t)) {
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+
+ if (k > 4 * sizeof(int8_t)) {
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ }
+ }
+ }
+
+ __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
+
+ const __m128 vscale0123 = _mm_loadu_ps((const float*) w);
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(float));
+ vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123);
+
+ vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
+
+ const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
+ __m128i vacc00x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc0x0123), voutput_zero_point);
+
+ const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
+ const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max);
+ vacc00x0123 = _mm_min_epi16(_mm_max_epi16(vacc00x0123, voutput_min), voutput_max);
+
+ __m128i vout = _mm_packs_epi16(vacc00x0123, vacc00x0123);
+
+
+ if (nc >= 4) {
+ *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
+
+ c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
+
+ a0 = (const int8_t*) ((uintptr_t) a0 - kc);
+
+ nc -= 4;
+ } else {
+ if (nc & 2) {
+ *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
+ c0 += 2;
+ vout = _mm_srli_epi32(vout, 16);
+ }
+ if (nc & 1) {
+ *((int8_t*) c0) = (int8_t) _mm_cvtsi128_si32(vout);
+ }
+
+ nc = 0;
+ }
+ } while (nc != 0);
+}
diff --git a/src/qc8-gemm/gen/1x4c2-minmax-fp32-sse41-ld128.c b/src/qc8-gemm/gen/1x4c2-minmax-fp32-sse41-ld128.c
new file mode 100644
index 0000000..e213352
--- /dev/null
+++ b/src/qc8-gemm/gen/1x4c2-minmax-fp32-sse41-ld128.c
@@ -0,0 +1,147 @@
+// Auto-generated file. Do not edit!
+// Template: src/qs8-gemm/MRx4c2-sse.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 <smmintrin.h>
+
+#include <xnnpack/gemm.h>
+#include <xnnpack/math.h>
+
+
+
+void xnn_qc8_gemm_minmax_fp32_ukernel_1x4c2__sse41_ld128(
+ 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_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN XNN_DISABLE_MSAN
+{
+ assert(mr != 0);
+ assert(mr <= 1);
+ 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, 2);
+ const int8_t* a0 = a;
+ int8_t* c0 = c;
+
+ do {
+ __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+
+ size_t k = kc;
+ while (k >= 8 * sizeof(int8_t)) {
+ const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
+ const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
+ a0 += 8;
+
+ const __m128i vb01 = _mm_loadu_si128((const __m128i*) w);
+ const __m128i vsb01 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb01);
+ const __m128i vxb0 = _mm_unpacklo_epi8(vb01, vsb01);
+ const __m128i vxb1 = _mm_unpackhi_epi8(vb01, vsb01);
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
+ const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
+ const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
+ const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
+
+ w = (const void*) ((uintptr_t) w + 32 * sizeof(int8_t));
+ k -= 8 * sizeof(int8_t);
+ }
+ if (k != 0) {
+ const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
+ const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
+ a0 = (const int8_t*) ((uintptr_t) a0 + k);
+
+ const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+
+ if (k > 2 * sizeof(int8_t)) {
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+
+ if (k > 4 * sizeof(int8_t)) {
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ }
+ }
+ }
+
+ __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
+
+ const __m128 vscale0123 = _mm_loadu_ps((const float*) w);
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(float));
+ vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123);
+
+ vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
+
+ const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4.output_zero_point);
+ __m128i vacc00x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc0x0123), voutput_zero_point);
+
+
+ __m128i vout = _mm_packs_epi16(vacc00x0123, vacc00x0123);
+
+ vout = _mm_max_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_min));
+ vout = _mm_min_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_max));
+
+ if (nc >= 4) {
+ *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
+
+ c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
+
+ a0 = (const int8_t*) ((uintptr_t) a0 - kc);
+
+ nc -= 4;
+ } else {
+ if (nc & 2) {
+ *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
+ c0 += 2;
+ vout = _mm_srli_epi32(vout, 16);
+ }
+ if (nc & 1) {
+ *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ }
+
+ nc = 0;
+ }
+ } while (nc != 0);
+}
diff --git a/src/qc8-gemm/gen/1x4c2-minmax-fp32-sse41-ld64.c b/src/qc8-gemm/gen/1x4c2-minmax-fp32-sse41-ld64.c
new file mode 100644
index 0000000..048febc
--- /dev/null
+++ b/src/qc8-gemm/gen/1x4c2-minmax-fp32-sse41-ld64.c
@@ -0,0 +1,147 @@
+// Auto-generated file. Do not edit!
+// Template: src/qs8-gemm/MRx4c2-sse.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 <smmintrin.h>
+
+#include <xnnpack/gemm.h>
+#include <xnnpack/math.h>
+
+
+
+void xnn_qc8_gemm_minmax_fp32_ukernel_1x4c2__sse41_ld64(
+ 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_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN XNN_DISABLE_MSAN
+{
+ assert(mr != 0);
+ assert(mr <= 1);
+ 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, 2);
+ const int8_t* a0 = a;
+ int8_t* c0 = c;
+
+ do {
+ __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+
+ size_t k = kc;
+ while (k >= 8 * sizeof(int8_t)) {
+ const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
+ const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
+ a0 += 8;
+
+ const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int8_t)));
+ const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
+ const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int8_t)));
+ const __m128i vxb3 = _mm_cvtepi8_epi16(vb3);
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
+
+ w = (const void*) ((uintptr_t) w + 32 * sizeof(int8_t));
+ k -= 8 * sizeof(int8_t);
+ }
+ if (k != 0) {
+ const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
+ const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
+ a0 = (const int8_t*) ((uintptr_t) a0 + k);
+
+ const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+
+ if (k > 2 * sizeof(int8_t)) {
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+
+ if (k > 4 * sizeof(int8_t)) {
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ }
+ }
+ }
+
+ __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
+
+ const __m128 vscale0123 = _mm_loadu_ps((const float*) w);
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(float));
+ vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123);
+
+ vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
+
+ const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4.output_zero_point);
+ __m128i vacc00x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc0x0123), voutput_zero_point);
+
+
+ __m128i vout = _mm_packs_epi16(vacc00x0123, vacc00x0123);
+
+ vout = _mm_max_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_min));
+ vout = _mm_min_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_max));
+
+ if (nc >= 4) {
+ *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
+
+ c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
+
+ a0 = (const int8_t*) ((uintptr_t) a0 - kc);
+
+ nc -= 4;
+ } else {
+ if (nc & 2) {
+ *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
+ c0 += 2;
+ vout = _mm_srli_epi32(vout, 16);
+ }
+ if (nc & 1) {
+ *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ }
+
+ nc = 0;
+ }
+ } while (nc != 0);
+}
diff --git a/src/qc8-gemm/gen/1x4c2-minmax-fp32-xop-ld128.c b/src/qc8-gemm/gen/1x4c2-minmax-fp32-xop-ld128.c
new file mode 100644
index 0000000..6b65214
--- /dev/null
+++ b/src/qc8-gemm/gen/1x4c2-minmax-fp32-xop-ld128.c
@@ -0,0 +1,152 @@
+// Auto-generated file. Do not edit!
+// Template: src/qs8-gemm/MRx4c2-sse.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>
+
+#if defined(__GNUC__) || defined(__clang__)
+ #include <x86intrin.h>
+#else
+ #include <immintrin.h>
+ #include <ammintrin.h>
+#endif
+
+#include <xnnpack/gemm.h>
+#include <xnnpack/math.h>
+
+
+
+void xnn_qc8_gemm_minmax_fp32_ukernel_1x4c2__xop_ld128(
+ 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_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN XNN_DISABLE_MSAN
+{
+ assert(mr != 0);
+ assert(mr <= 1);
+ 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, 2);
+ const int8_t* a0 = a;
+ int8_t* c0 = c;
+
+ do {
+ __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+
+ size_t k = kc;
+ while (k >= 8 * sizeof(int8_t)) {
+ const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
+ const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
+ a0 += 8;
+
+ const __m128i vb01 = _mm_loadu_si128((const __m128i*) w);
+ const __m128i vsb01 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb01);
+ const __m128i vxb0 = _mm_unpacklo_epi8(vb01, vsb01);
+ const __m128i vxb1 = _mm_unpackhi_epi8(vb01, vsb01);
+
+ vacc0x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
+
+ vacc0x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
+ const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
+ const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
+ const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
+ const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
+
+ vacc0x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
+
+ vacc0x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc0x0123);
+
+ w = (const void*) ((uintptr_t) w + 32 * sizeof(int8_t));
+ k -= 8 * sizeof(int8_t);
+ }
+ if (k != 0) {
+ const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
+ const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
+ a0 = (const int8_t*) ((uintptr_t) a0 + k);
+
+ const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
+
+ if (k > 2 * sizeof(int8_t)) {
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
+
+ if (k > 4 * sizeof(int8_t)) {
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
+ }
+ }
+ }
+
+ __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
+
+ const __m128 vscale0123 = _mm_loadu_ps((const float*) w);
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(float));
+ vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123);
+
+ vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
+
+ const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4.output_zero_point);
+ __m128i vacc00x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc0x0123), voutput_zero_point);
+
+
+ __m128i vout = _mm_packs_epi16(vacc00x0123, vacc00x0123);
+
+ vout = _mm_max_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_min));
+ vout = _mm_min_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_max));
+
+ if (nc >= 4) {
+ *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
+
+ c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
+
+ a0 = (const int8_t*) ((uintptr_t) a0 - kc);
+
+ nc -= 4;
+ } else {
+ if (nc & 2) {
+ *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
+ c0 += 2;
+ vout = _mm_srli_epi32(vout, 16);
+ }
+ if (nc & 1) {
+ *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ }
+
+ nc = 0;
+ }
+ } while (nc != 0);
+}
diff --git a/src/qc8-gemm/gen/1x4c2-minmax-fp32-xop-ld64.c b/src/qc8-gemm/gen/1x4c2-minmax-fp32-xop-ld64.c
new file mode 100644
index 0000000..a182e89
--- /dev/null
+++ b/src/qc8-gemm/gen/1x4c2-minmax-fp32-xop-ld64.c
@@ -0,0 +1,152 @@
+// Auto-generated file. Do not edit!
+// Template: src/qs8-gemm/MRx4c2-sse.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>
+
+#if defined(__GNUC__) || defined(__clang__)
+ #include <x86intrin.h>
+#else
+ #include <immintrin.h>
+ #include <ammintrin.h>
+#endif
+
+#include <xnnpack/gemm.h>
+#include <xnnpack/math.h>
+
+
+
+void xnn_qc8_gemm_minmax_fp32_ukernel_1x4c2__xop_ld64(
+ 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_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN XNN_DISABLE_MSAN
+{
+ assert(mr != 0);
+ assert(mr <= 1);
+ 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, 2);
+ const int8_t* a0 = a;
+ int8_t* c0 = c;
+
+ do {
+ __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+
+ size_t k = kc;
+ while (k >= 8 * sizeof(int8_t)) {
+ const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
+ const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
+ a0 += 8;
+
+ const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
+
+ vacc0x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int8_t)));
+ const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
+
+ vacc0x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
+ const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
+
+ vacc0x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int8_t)));
+ const __m128i vxb3 = _mm_cvtepi8_epi16(vb3);
+
+ vacc0x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc0x0123);
+
+ w = (const void*) ((uintptr_t) w + 32 * sizeof(int8_t));
+ k -= 8 * sizeof(int8_t);
+ }
+ if (k != 0) {
+ const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
+ const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
+ a0 = (const int8_t*) ((uintptr_t) a0 + k);
+
+ const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
+
+ if (k > 2 * sizeof(int8_t)) {
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
+
+ if (k > 4 * sizeof(int8_t)) {
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
+ }
+ }
+ }
+
+ __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
+
+ const __m128 vscale0123 = _mm_loadu_ps((const float*) w);
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(float));
+ vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123);
+
+ vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
+
+ const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4.output_zero_point);
+ __m128i vacc00x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc0x0123), voutput_zero_point);
+
+
+ __m128i vout = _mm_packs_epi16(vacc00x0123, vacc00x0123);
+
+ vout = _mm_max_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_min));
+ vout = _mm_min_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_max));
+
+ if (nc >= 4) {
+ *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
+
+ c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
+
+ a0 = (const int8_t*) ((uintptr_t) a0 - kc);
+
+ nc -= 4;
+ } else {
+ if (nc & 2) {
+ *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
+ c0 += 2;
+ vout = _mm_srli_epi32(vout, 16);
+ }
+ if (nc & 1) {
+ *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ }
+
+ nc = 0;
+ }
+ } while (nc != 0);
+}
diff --git a/src/qc8-gemm/gen/1x4c8-minmax-fp32-avx-ld128.c b/src/qc8-gemm/gen/1x4c8-minmax-fp32-avx-ld128.c
new file mode 100644
index 0000000..82f7100
--- /dev/null
+++ b/src/qc8-gemm/gen/1x4c8-minmax-fp32-avx-ld128.c
@@ -0,0 +1,118 @@
+// Auto-generated file. Do not edit!
+// Template: src/qs8-gemm/MRx4c8-sse.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 <smmintrin.h>
+
+#include <xnnpack/gemm.h>
+#include <xnnpack/math.h>
+
+
+void xnn_qc8_gemm_minmax_fp32_ukernel_1x4c8__avx_ld128(
+ 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_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN XNN_DISABLE_MSAN
+{
+ assert(mr != 0);
+ assert(mr <= 1);
+ 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, 8);
+ const int8_t* a0 = a;
+ int8_t* c0 = c;
+
+ do {
+ __m128i vacc0x0 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[0]);
+ __m128i vacc0x1 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[1]);
+ __m128i vacc0x2 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[2]);
+ __m128i vacc0x3 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[3]);
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+
+ size_t k = 0;
+ while (k < kc) {
+ const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
+ const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
+ a0 += 8;
+
+ const __m128i vb01 = _mm_load_si128((const __m128i*) w);
+ const __m128i vsb01 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb01);
+ const __m128i vxb0 = _mm_unpacklo_epi8(vb01, vsb01);
+ const __m128i vxb1 = _mm_unpackhi_epi8(vb01, vsb01);
+
+ vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
+ vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
+ const __m128i vb23 = _mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
+ const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
+ const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
+ const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
+
+ vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
+ vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
+
+ w = (const void*) ((uintptr_t) w + 32 * sizeof(int8_t));
+ k += 8 * sizeof(int8_t);
+ }
+
+ const __m128i vacc0x01 = _mm_hadd_epi32(vacc0x0, vacc0x1);
+ const __m128i vacc0x23 = _mm_hadd_epi32(vacc0x2, vacc0x3);
+
+ __m128i vacc0x0123 = _mm_hadd_epi32(vacc0x01, vacc0x23);
+
+ __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
+
+ const __m128 vscale0123 = _mm_load_ps((const float*) w);
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(float));
+ vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123);
+
+ vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
+
+ const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4.output_zero_point);
+ __m128i vacc00x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc0x0123), voutput_zero_point);
+
+
+ __m128i vout = _mm_packs_epi16(vacc00x0123, vacc00x0123);
+
+ vout = _mm_max_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_min));
+ vout = _mm_min_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_max));
+
+ if (nc >= 4) {
+ *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
+
+ c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
+
+ a0 = (const int8_t*) ((uintptr_t) a0 - kc);
+
+ nc -= 4;
+ } else {
+ if (nc & 2) {
+ *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
+ c0 += 2;
+ vout = _mm_srli_epi32(vout, 16);
+ }
+ if (nc & 1) {
+ *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ }
+
+ nc = 0;
+ }
+ } while (nc != 0);
+}
diff --git a/src/qc8-gemm/gen/1x4c8-minmax-fp32-avx-ld64.c b/src/qc8-gemm/gen/1x4c8-minmax-fp32-avx-ld64.c
new file mode 100644
index 0000000..709321b
--- /dev/null
+++ b/src/qc8-gemm/gen/1x4c8-minmax-fp32-avx-ld64.c
@@ -0,0 +1,120 @@
+// Auto-generated file. Do not edit!
+// Template: src/qs8-gemm/MRx4c8-sse.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 <smmintrin.h>
+
+#include <xnnpack/gemm.h>
+#include <xnnpack/math.h>
+
+
+void xnn_qc8_gemm_minmax_fp32_ukernel_1x4c8__avx_ld64(
+ 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_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN XNN_DISABLE_MSAN
+{
+ assert(mr != 0);
+ assert(mr <= 1);
+ 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, 8);
+ const int8_t* a0 = a;
+ int8_t* c0 = c;
+
+ do {
+ __m128i vacc0x0 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[0]);
+ __m128i vacc0x1 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[1]);
+ __m128i vacc0x2 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[2]);
+ __m128i vacc0x3 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[3]);
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+
+ size_t k = 0;
+ while (k < kc) {
+ const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
+ const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
+ a0 += 8;
+
+ const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
+
+ vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int8_t)));
+ const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
+
+ vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
+ const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
+
+ vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int8_t)));
+ const __m128i vxb3 = _mm_cvtepi8_epi16(vb3);
+
+ vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
+
+ w = (const void*) ((uintptr_t) w + 32 * sizeof(int8_t));
+ k += 8 * sizeof(int8_t);
+ }
+
+ const __m128i vacc0x01 = _mm_hadd_epi32(vacc0x0, vacc0x1);
+ const __m128i vacc0x23 = _mm_hadd_epi32(vacc0x2, vacc0x3);
+
+ __m128i vacc0x0123 = _mm_hadd_epi32(vacc0x01, vacc0x23);
+
+ __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
+
+ const __m128 vscale0123 = _mm_load_ps((const float*) w);
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(float));
+ vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123);
+
+ vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
+
+ const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4.output_zero_point);
+ __m128i vacc00x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc0x0123), voutput_zero_point);
+
+
+ __m128i vout = _mm_packs_epi16(vacc00x0123, vacc00x0123);
+
+ vout = _mm_max_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_min));
+ vout = _mm_min_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_max));
+
+ if (nc >= 4) {
+ *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
+
+ c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
+
+ a0 = (const int8_t*) ((uintptr_t) a0 - kc);
+
+ nc -= 4;
+ } else {
+ if (nc & 2) {
+ *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
+ c0 += 2;
+ vout = _mm_srli_epi32(vout, 16);
+ }
+ if (nc & 1) {
+ *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ }
+
+ nc = 0;
+ }
+ } while (nc != 0);
+}
diff --git a/src/qc8-gemm/gen/1x4c8-minmax-fp32-sse2-ld128.c b/src/qc8-gemm/gen/1x4c8-minmax-fp32-sse2-ld128.c
new file mode 100644
index 0000000..f82da93
--- /dev/null
+++ b/src/qc8-gemm/gen/1x4c8-minmax-fp32-sse2-ld128.c
@@ -0,0 +1,119 @@
+// Auto-generated file. Do not edit!
+// Template: src/qs8-gemm/MRx4c8-sse.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 <emmintrin.h>
+
+#include <xnnpack/gemm.h>
+#include <xnnpack/math.h>
+
+
+void xnn_qc8_gemm_minmax_fp32_ukernel_1x4c8__sse2_ld128(
+ 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_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN XNN_DISABLE_MSAN
+{
+ assert(mr != 0);
+ assert(mr <= 1);
+ 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, 8);
+ const int8_t* a0 = a;
+ int8_t* c0 = c;
+
+ do {
+ __m128i vacc0x0 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[0]);
+ __m128i vacc0x1 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[1]);
+ __m128i vacc0x2 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[2]);
+ __m128i vacc0x3 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[3]);
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+
+ size_t k = 0;
+ while (k < kc) {
+ const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
+ const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ a0 += 8;
+
+ const __m128i vb01 = _mm_load_si128((const __m128i*) w);
+ const __m128i vsb01 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb01);
+ const __m128i vxb0 = _mm_unpacklo_epi8(vb01, vsb01);
+ const __m128i vxb1 = _mm_unpackhi_epi8(vb01, vsb01);
+
+ vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
+ vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
+ const __m128i vb23 = _mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
+ const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
+ const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
+ const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
+
+ vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
+ vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
+
+ w = (const void*) ((uintptr_t) w + 32 * sizeof(int8_t));
+ k += 8 * sizeof(int8_t);
+ }
+
+ const __m128i vacc0x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x0, vacc0x2), _mm_unpackhi_epi32(vacc0x0, vacc0x2));
+ const __m128i vacc0x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x1, vacc0x3), _mm_unpackhi_epi32(vacc0x1, vacc0x3));
+
+ __m128i vacc0x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x02, vacc0x13), _mm_unpackhi_epi32(vacc0x02, vacc0x13));
+
+ __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
+
+ const __m128 vscale0123 = _mm_load_ps((const float*) w);
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(float));
+ vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123);
+
+ vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
+
+ const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
+ __m128i vacc00x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc0x0123), voutput_zero_point);
+
+ const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
+ const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max);
+ vacc00x0123 = _mm_min_epi16(_mm_max_epi16(vacc00x0123, voutput_min), voutput_max);
+
+ __m128i vout = _mm_packs_epi16(vacc00x0123, vacc00x0123);
+
+
+ if (nc >= 4) {
+ *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
+
+ c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
+
+ a0 = (const int8_t*) ((uintptr_t) a0 - kc);
+
+ nc -= 4;
+ } else {
+ if (nc & 2) {
+ *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
+ c0 += 2;
+ vout = _mm_srli_epi32(vout, 16);
+ }
+ if (nc & 1) {
+ *((int8_t*) c0) = (int8_t) _mm_cvtsi128_si32(vout);
+ }
+
+ nc = 0;
+ }
+ } while (nc != 0);
+}
diff --git a/src/qc8-gemm/gen/1x4c8-minmax-fp32-sse2-ld64.c b/src/qc8-gemm/gen/1x4c8-minmax-fp32-sse2-ld64.c
new file mode 100644
index 0000000..dceacec
--- /dev/null
+++ b/src/qc8-gemm/gen/1x4c8-minmax-fp32-sse2-ld64.c
@@ -0,0 +1,121 @@
+// Auto-generated file. Do not edit!
+// Template: src/qs8-gemm/MRx4c8-sse.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 <emmintrin.h>
+
+#include <xnnpack/gemm.h>
+#include <xnnpack/math.h>
+
+
+void xnn_qc8_gemm_minmax_fp32_ukernel_1x4c8__sse2_ld64(
+ 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_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN XNN_DISABLE_MSAN
+{
+ assert(mr != 0);
+ assert(mr <= 1);
+ 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, 8);
+ const int8_t* a0 = a;
+ int8_t* c0 = c;
+
+ do {
+ __m128i vacc0x0 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[0]);
+ __m128i vacc0x1 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[1]);
+ __m128i vacc0x2 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[2]);
+ __m128i vacc0x3 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[3]);
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+
+ size_t k = 0;
+ while (k < kc) {
+ const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
+ const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ a0 += 8;
+
+ const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+
+ vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int8_t)));
+ const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+
+ vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
+ const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+
+ vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int8_t)));
+ const __m128i vxb3 = _mm_unpacklo_epi8(vb3, _mm_cmpgt_epi8(_mm_setzero_si128(), vb3));
+
+ vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
+
+ w = (const void*) ((uintptr_t) w + 32 * sizeof(int8_t));
+ k += 8 * sizeof(int8_t);
+ }
+
+ const __m128i vacc0x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x0, vacc0x2), _mm_unpackhi_epi32(vacc0x0, vacc0x2));
+ const __m128i vacc0x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x1, vacc0x3), _mm_unpackhi_epi32(vacc0x1, vacc0x3));
+
+ __m128i vacc0x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x02, vacc0x13), _mm_unpackhi_epi32(vacc0x02, vacc0x13));
+
+ __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
+
+ const __m128 vscale0123 = _mm_load_ps((const float*) w);
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(float));
+ vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123);
+
+ vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
+
+ const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
+ __m128i vacc00x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc0x0123), voutput_zero_point);
+
+ const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
+ const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max);
+ vacc00x0123 = _mm_min_epi16(_mm_max_epi16(vacc00x0123, voutput_min), voutput_max);
+
+ __m128i vout = _mm_packs_epi16(vacc00x0123, vacc00x0123);
+
+
+ if (nc >= 4) {
+ *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
+
+ c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
+
+ a0 = (const int8_t*) ((uintptr_t) a0 - kc);
+
+ nc -= 4;
+ } else {
+ if (nc & 2) {
+ *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
+ c0 += 2;
+ vout = _mm_srli_epi32(vout, 16);
+ }
+ if (nc & 1) {
+ *((int8_t*) c0) = (int8_t) _mm_cvtsi128_si32(vout);
+ }
+
+ nc = 0;
+ }
+ } while (nc != 0);
+}
diff --git a/src/qc8-gemm/gen/1x4c8-minmax-fp32-sse41-ld128.c b/src/qc8-gemm/gen/1x4c8-minmax-fp32-sse41-ld128.c
new file mode 100644
index 0000000..4b0bffd
--- /dev/null
+++ b/src/qc8-gemm/gen/1x4c8-minmax-fp32-sse41-ld128.c
@@ -0,0 +1,118 @@
+// Auto-generated file. Do not edit!
+// Template: src/qs8-gemm/MRx4c8-sse.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 <smmintrin.h>
+
+#include <xnnpack/gemm.h>
+#include <xnnpack/math.h>
+
+
+void xnn_qc8_gemm_minmax_fp32_ukernel_1x4c8__sse41_ld128(
+ 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_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN XNN_DISABLE_MSAN
+{
+ assert(mr != 0);
+ assert(mr <= 1);
+ 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, 8);
+ const int8_t* a0 = a;
+ int8_t* c0 = c;
+
+ do {
+ __m128i vacc0x0 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[0]);
+ __m128i vacc0x1 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[1]);
+ __m128i vacc0x2 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[2]);
+ __m128i vacc0x3 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[3]);
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+
+ size_t k = 0;
+ while (k < kc) {
+ const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
+ const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
+ a0 += 8;
+
+ const __m128i vb01 = _mm_load_si128((const __m128i*) w);
+ const __m128i vsb01 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb01);
+ const __m128i vxb0 = _mm_unpacklo_epi8(vb01, vsb01);
+ const __m128i vxb1 = _mm_unpackhi_epi8(vb01, vsb01);
+
+ vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
+ vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
+ const __m128i vb23 = _mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
+ const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
+ const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
+ const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
+
+ vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
+ vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
+
+ w = (const void*) ((uintptr_t) w + 32 * sizeof(int8_t));
+ k += 8 * sizeof(int8_t);
+ }
+
+ const __m128i vacc0x01 = _mm_hadd_epi32(vacc0x0, vacc0x1);
+ const __m128i vacc0x23 = _mm_hadd_epi32(vacc0x2, vacc0x3);
+
+ __m128i vacc0x0123 = _mm_hadd_epi32(vacc0x01, vacc0x23);
+
+ __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
+
+ const __m128 vscale0123 = _mm_load_ps((const float*) w);
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(float));
+ vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123);
+
+ vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
+
+ const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4.output_zero_point);
+ __m128i vacc00x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc0x0123), voutput_zero_point);
+
+
+ __m128i vout = _mm_packs_epi16(vacc00x0123, vacc00x0123);
+
+ vout = _mm_max_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_min));
+ vout = _mm_min_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_max));
+
+ if (nc >= 4) {
+ *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
+
+ c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
+
+ a0 = (const int8_t*) ((uintptr_t) a0 - kc);
+
+ nc -= 4;
+ } else {
+ if (nc & 2) {
+ *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
+ c0 += 2;
+ vout = _mm_srli_epi32(vout, 16);
+ }
+ if (nc & 1) {
+ *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ }
+
+ nc = 0;
+ }
+ } while (nc != 0);
+}
diff --git a/src/qc8-gemm/gen/1x4c8-minmax-fp32-sse41-ld64.c b/src/qc8-gemm/gen/1x4c8-minmax-fp32-sse41-ld64.c
new file mode 100644
index 0000000..0c1870e
--- /dev/null
+++ b/src/qc8-gemm/gen/1x4c8-minmax-fp32-sse41-ld64.c
@@ -0,0 +1,120 @@
+// Auto-generated file. Do not edit!
+// Template: src/qs8-gemm/MRx4c8-sse.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 <smmintrin.h>
+
+#include <xnnpack/gemm.h>
+#include <xnnpack/math.h>
+
+
+void xnn_qc8_gemm_minmax_fp32_ukernel_1x4c8__sse41_ld64(
+ 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_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN XNN_DISABLE_MSAN
+{
+ assert(mr != 0);
+ assert(mr <= 1);
+ 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, 8);
+ const int8_t* a0 = a;
+ int8_t* c0 = c;
+
+ do {
+ __m128i vacc0x0 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[0]);
+ __m128i vacc0x1 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[1]);
+ __m128i vacc0x2 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[2]);
+ __m128i vacc0x3 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[3]);
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+
+ size_t k = 0;
+ while (k < kc) {
+ const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
+ const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
+ a0 += 8;
+
+ const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
+
+ vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int8_t)));
+ const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
+
+ vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
+ const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
+
+ vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int8_t)));
+ const __m128i vxb3 = _mm_cvtepi8_epi16(vb3);
+
+ vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
+
+ w = (const void*) ((uintptr_t) w + 32 * sizeof(int8_t));
+ k += 8 * sizeof(int8_t);
+ }
+
+ const __m128i vacc0x01 = _mm_hadd_epi32(vacc0x0, vacc0x1);
+ const __m128i vacc0x23 = _mm_hadd_epi32(vacc0x2, vacc0x3);
+
+ __m128i vacc0x0123 = _mm_hadd_epi32(vacc0x01, vacc0x23);
+
+ __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
+
+ const __m128 vscale0123 = _mm_load_ps((const float*) w);
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(float));
+ vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123);
+
+ vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
+
+ const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4.output_zero_point);
+ __m128i vacc00x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc0x0123), voutput_zero_point);
+
+
+ __m128i vout = _mm_packs_epi16(vacc00x0123, vacc00x0123);
+
+ vout = _mm_max_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_min));
+ vout = _mm_min_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_max));
+
+ if (nc >= 4) {
+ *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
+
+ c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
+
+ a0 = (const int8_t*) ((uintptr_t) a0 - kc);
+
+ nc -= 4;
+ } else {
+ if (nc & 2) {
+ *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
+ c0 += 2;
+ vout = _mm_srli_epi32(vout, 16);
+ }
+ if (nc & 1) {
+ *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ }
+
+ nc = 0;
+ }
+ } while (nc != 0);
+}
diff --git a/src/qc8-gemm/gen/1x4c8-minmax-fp32-xop-ld128.c b/src/qc8-gemm/gen/1x4c8-minmax-fp32-xop-ld128.c
new file mode 100644
index 0000000..239f15d
--- /dev/null
+++ b/src/qc8-gemm/gen/1x4c8-minmax-fp32-xop-ld128.c
@@ -0,0 +1,123 @@
+// Auto-generated file. Do not edit!
+// Template: src/qs8-gemm/MRx4c8-sse.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>
+
+#if defined(__GNUC__) || defined(__clang__)
+ #include <x86intrin.h>
+#else
+ #include <immintrin.h>
+ #include <ammintrin.h>
+#endif
+
+#include <xnnpack/gemm.h>
+#include <xnnpack/math.h>
+
+
+void xnn_qc8_gemm_minmax_fp32_ukernel_1x4c8__xop_ld128(
+ 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_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN XNN_DISABLE_MSAN
+{
+ assert(mr != 0);
+ assert(mr <= 1);
+ 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, 8);
+ const int8_t* a0 = a;
+ int8_t* c0 = c;
+
+ do {
+ __m128i vacc0x0 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[0]);
+ __m128i vacc0x1 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[1]);
+ __m128i vacc0x2 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[2]);
+ __m128i vacc0x3 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[3]);
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+
+ size_t k = 0;
+ while (k < kc) {
+ const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
+ const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
+ a0 += 8;
+
+ const __m128i vb01 = _mm_load_si128((const __m128i*) w);
+ const __m128i vsb01 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb01);
+ const __m128i vxb0 = _mm_unpacklo_epi8(vb01, vsb01);
+ const __m128i vxb1 = _mm_unpackhi_epi8(vb01, vsb01);
+
+ vacc0x0 = _mm_maddd_epi16(vxa0, vxb0, vacc0x0);
+ vacc0x1 = _mm_maddd_epi16(vxa0, vxb1, vacc0x1);
+ const __m128i vb23 = _mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
+ const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
+ const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
+ const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
+
+ vacc0x2 = _mm_maddd_epi16(vxa0, vxb2, vacc0x2);
+ vacc0x3 = _mm_maddd_epi16(vxa0, vxb3, vacc0x3);
+
+ w = (const void*) ((uintptr_t) w + 32 * sizeof(int8_t));
+ k += 8 * sizeof(int8_t);
+ }
+
+ const __m128i vacc0x01 = _mm_hadd_epi32(vacc0x0, vacc0x1);
+ const __m128i vacc0x23 = _mm_hadd_epi32(vacc0x2, vacc0x3);
+
+ __m128i vacc0x0123 = _mm_hadd_epi32(vacc0x01, vacc0x23);
+
+ __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
+
+ const __m128 vscale0123 = _mm_load_ps((const float*) w);
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(float));
+ vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123);
+
+ vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
+
+ const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4.output_zero_point);
+ __m128i vacc00x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc0x0123), voutput_zero_point);
+
+
+ __m128i vout = _mm_packs_epi16(vacc00x0123, vacc00x0123);
+
+ vout = _mm_max_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_min));
+ vout = _mm_min_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_max));
+
+ if (nc >= 4) {
+ *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
+
+ c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
+
+ a0 = (const int8_t*) ((uintptr_t) a0 - kc);
+
+ nc -= 4;
+ } else {
+ if (nc & 2) {
+ *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
+ c0 += 2;
+ vout = _mm_srli_epi32(vout, 16);
+ }
+ if (nc & 1) {
+ *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ }
+
+ nc = 0;
+ }
+ } while (nc != 0);
+}
diff --git a/src/qc8-gemm/gen/1x4c8-minmax-fp32-xop-ld64.c b/src/qc8-gemm/gen/1x4c8-minmax-fp32-xop-ld64.c
new file mode 100644
index 0000000..9acbca1
--- /dev/null
+++ b/src/qc8-gemm/gen/1x4c8-minmax-fp32-xop-ld64.c
@@ -0,0 +1,125 @@
+// Auto-generated file. Do not edit!
+// Template: src/qs8-gemm/MRx4c8-sse.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>
+
+#if defined(__GNUC__) || defined(__clang__)
+ #include <x86intrin.h>
+#else
+ #include <immintrin.h>
+ #include <ammintrin.h>
+#endif
+
+#include <xnnpack/gemm.h>
+#include <xnnpack/math.h>
+
+
+void xnn_qc8_gemm_minmax_fp32_ukernel_1x4c8__xop_ld64(
+ 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_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN XNN_DISABLE_MSAN
+{
+ assert(mr != 0);
+ assert(mr <= 1);
+ 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, 8);
+ const int8_t* a0 = a;
+ int8_t* c0 = c;
+
+ do {
+ __m128i vacc0x0 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[0]);
+ __m128i vacc0x1 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[1]);
+ __m128i vacc0x2 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[2]);
+ __m128i vacc0x3 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[3]);
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+
+ size_t k = 0;
+ while (k < kc) {
+ const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
+ const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
+ a0 += 8;
+
+ const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
+
+ vacc0x0 = _mm_maddd_epi16(vxa0, vxb0, vacc0x0);
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int8_t)));
+ const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
+
+ vacc0x1 = _mm_maddd_epi16(vxa0, vxb1, vacc0x1);
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
+ const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
+
+ vacc0x2 = _mm_maddd_epi16(vxa0, vxb2, vacc0x2);
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int8_t)));
+ const __m128i vxb3 = _mm_cvtepi8_epi16(vb3);
+
+ vacc0x3 = _mm_maddd_epi16(vxa0, vxb3, vacc0x3);
+
+ w = (const void*) ((uintptr_t) w + 32 * sizeof(int8_t));
+ k += 8 * sizeof(int8_t);
+ }
+
+ const __m128i vacc0x01 = _mm_hadd_epi32(vacc0x0, vacc0x1);
+ const __m128i vacc0x23 = _mm_hadd_epi32(vacc0x2, vacc0x3);
+
+ __m128i vacc0x0123 = _mm_hadd_epi32(vacc0x01, vacc0x23);
+
+ __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
+
+ const __m128 vscale0123 = _mm_load_ps((const float*) w);
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(float));
+ vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123);
+
+ vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
+
+ const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4.output_zero_point);
+ __m128i vacc00x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc0x0123), voutput_zero_point);
+
+
+ __m128i vout = _mm_packs_epi16(vacc00x0123, vacc00x0123);
+
+ vout = _mm_max_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_min));
+ vout = _mm_min_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_max));
+
+ if (nc >= 4) {
+ *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
+
+ c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
+
+ a0 = (const int8_t*) ((uintptr_t) a0 - kc);
+
+ nc -= 4;
+ } else {
+ if (nc & 2) {
+ *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
+ c0 += 2;
+ vout = _mm_srli_epi32(vout, 16);
+ }
+ if (nc & 1) {
+ *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ }
+
+ nc = 0;
+ }
+ } while (nc != 0);
+}
diff --git a/src/qc8-gemm/gen/2x4c2-minmax-fp32-avx-ld128.c b/src/qc8-gemm/gen/2x4c2-minmax-fp32-avx-ld128.c
new file mode 100644
index 0000000..b208c3b
--- /dev/null
+++ b/src/qc8-gemm/gen/2x4c2-minmax-fp32-avx-ld128.c
@@ -0,0 +1,183 @@
+// Auto-generated file. Do not edit!
+// Template: src/qs8-gemm/MRx4c2-sse.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 <smmintrin.h>
+
+#include <xnnpack/gemm.h>
+#include <xnnpack/math.h>
+
+
+
+void xnn_qc8_gemm_minmax_fp32_ukernel_2x4c2__avx_ld128(
+ 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_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN XNN_DISABLE_MSAN
+{
+ assert(mr != 0);
+ assert(mr <= 2);
+ 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, 2);
+ 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;
+ }
+
+ do {
+ __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
+ __m128i vacc1x0123 = vacc0x0123;
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+
+ size_t k = kc;
+ while (k >= 8 * sizeof(int8_t)) {
+ const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
+ const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
+ a0 += 8;
+ const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
+ const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
+ a1 += 8;
+
+ const __m128i vb01 = _mm_loadu_si128((const __m128i*) w);
+ const __m128i vsb01 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb01);
+ const __m128i vxb0 = _mm_unpacklo_epi8(vb01, vsb01);
+ const __m128i vxb1 = _mm_unpackhi_epi8(vb01, vsb01);
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
+ const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
+ const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
+ const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
+
+ w = (const void*) ((uintptr_t) w + 32 * sizeof(int8_t));
+ k -= 8 * sizeof(int8_t);
+ }
+ if (k != 0) {
+ const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
+ const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
+ a0 = (const int8_t*) ((uintptr_t) a0 + k);
+ const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
+ const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
+ a1 = (const int8_t*) ((uintptr_t) a1 + k);
+
+ const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+
+ if (k > 2 * sizeof(int8_t)) {
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+
+ if (k > 4 * sizeof(int8_t)) {
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ }
+ }
+ }
+
+ __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
+ __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
+
+ const __m128 vscale0123 = _mm_loadu_ps((const float*) w);
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(float));
+ vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123);
+ vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale0123);
+
+ vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
+ vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
+
+ const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4.output_zero_point);
+ __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
+
+
+ __m128i vout = _mm_packs_epi16(vacc01x0123, vacc01x0123);
+
+ vout = _mm_max_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_min));
+ vout = _mm_min_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_max));
+
+ if (nc >= 4) {
+ *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
+ *((uint32_t*) c1) = (uint32_t) _mm_extract_epi32(vout, 1);
+
+ c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
+ c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
+
+ a0 = (const int8_t*) ((uintptr_t) a0 - kc);
+ a1 = (const int8_t*) ((uintptr_t) a1 - kc);
+
+ nc -= 4;
+ } else {
+ if (nc & 2) {
+ *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
+ c0 += 2;
+ *((uint16_t*) c1) = (uint16_t) _mm_extract_epi16(vout, 2);
+ c1 += 2;
+ vout = _mm_srli_epi32(vout, 16);
+ }
+ if (nc & 1) {
+ *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
+ }
+
+ nc = 0;
+ }
+ } while (nc != 0);
+}
diff --git a/src/qc8-gemm/gen/2x4c2-minmax-fp32-avx-ld64.c b/src/qc8-gemm/gen/2x4c2-minmax-fp32-avx-ld64.c
new file mode 100644
index 0000000..701c66d
--- /dev/null
+++ b/src/qc8-gemm/gen/2x4c2-minmax-fp32-avx-ld64.c
@@ -0,0 +1,183 @@
+// Auto-generated file. Do not edit!
+// Template: src/qs8-gemm/MRx4c2-sse.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 <smmintrin.h>
+
+#include <xnnpack/gemm.h>
+#include <xnnpack/math.h>
+
+
+
+void xnn_qc8_gemm_minmax_fp32_ukernel_2x4c2__avx_ld64(
+ 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_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN XNN_DISABLE_MSAN
+{
+ assert(mr != 0);
+ assert(mr <= 2);
+ 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, 2);
+ 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;
+ }
+
+ do {
+ __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
+ __m128i vacc1x0123 = vacc0x0123;
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+
+ size_t k = kc;
+ while (k >= 8 * sizeof(int8_t)) {
+ const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
+ const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
+ a0 += 8;
+ const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
+ const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
+ a1 += 8;
+
+ const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int8_t)));
+ const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
+ const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int8_t)));
+ const __m128i vxb3 = _mm_cvtepi8_epi16(vb3);
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
+
+ w = (const void*) ((uintptr_t) w + 32 * sizeof(int8_t));
+ k -= 8 * sizeof(int8_t);
+ }
+ if (k != 0) {
+ const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
+ const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
+ a0 = (const int8_t*) ((uintptr_t) a0 + k);
+ const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
+ const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
+ a1 = (const int8_t*) ((uintptr_t) a1 + k);
+
+ const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+
+ if (k > 2 * sizeof(int8_t)) {
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+
+ if (k > 4 * sizeof(int8_t)) {
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ }
+ }
+ }
+
+ __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
+ __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
+
+ const __m128 vscale0123 = _mm_loadu_ps((const float*) w);
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(float));
+ vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123);
+ vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale0123);
+
+ vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
+ vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
+
+ const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4.output_zero_point);
+ __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
+
+
+ __m128i vout = _mm_packs_epi16(vacc01x0123, vacc01x0123);
+
+ vout = _mm_max_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_min));
+ vout = _mm_min_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_max));
+
+ if (nc >= 4) {
+ *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
+ *((uint32_t*) c1) = (uint32_t) _mm_extract_epi32(vout, 1);
+
+ c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
+ c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
+
+ a0 = (const int8_t*) ((uintptr_t) a0 - kc);
+ a1 = (const int8_t*) ((uintptr_t) a1 - kc);
+
+ nc -= 4;
+ } else {
+ if (nc & 2) {
+ *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
+ c0 += 2;
+ *((uint16_t*) c1) = (uint16_t) _mm_extract_epi16(vout, 2);
+ c1 += 2;
+ vout = _mm_srli_epi32(vout, 16);
+ }
+ if (nc & 1) {
+ *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
+ }
+
+ nc = 0;
+ }
+ } while (nc != 0);
+}
diff --git a/src/qc8-gemm/gen/2x4c2-minmax-fp32-sse2-ld128.c b/src/qc8-gemm/gen/2x4c2-minmax-fp32-sse2-ld128.c
new file mode 100644
index 0000000..f232688
--- /dev/null
+++ b/src/qc8-gemm/gen/2x4c2-minmax-fp32-sse2-ld128.c
@@ -0,0 +1,185 @@
+// Auto-generated file. Do not edit!
+// Template: src/qs8-gemm/MRx4c2-sse.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 <emmintrin.h>
+
+#include <xnnpack/gemm.h>
+#include <xnnpack/math.h>
+
+
+
+void xnn_qc8_gemm_minmax_fp32_ukernel_2x4c2__sse2_ld128(
+ 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_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN XNN_DISABLE_MSAN
+{
+ assert(mr != 0);
+ assert(mr <= 2);
+ 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, 2);
+ 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;
+ }
+
+ do {
+ __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
+ __m128i vacc1x0123 = vacc0x0123;
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+
+ size_t k = kc;
+ while (k >= 8 * sizeof(int8_t)) {
+ const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
+ const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ a0 += 8;
+ const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
+ const __m128i vxa1 = _mm_unpacklo_epi8(va1, _mm_cmpgt_epi8(_mm_setzero_si128(), va1));
+ a1 += 8;
+
+ const __m128i vb01 = _mm_loadu_si128((const __m128i*) w);
+ const __m128i vsb01 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb01);
+ const __m128i vxb0 = _mm_unpacklo_epi8(vb01, vsb01);
+ const __m128i vxb1 = _mm_unpackhi_epi8(vb01, vsb01);
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
+ const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
+ const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
+ const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
+
+ w = (const void*) ((uintptr_t) w + 32 * sizeof(int8_t));
+ k -= 8 * sizeof(int8_t);
+ }
+ if (k != 0) {
+ const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
+ const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ a0 = (const int8_t*) ((uintptr_t) a0 + k);
+ const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
+ const __m128i vxa1 = _mm_unpacklo_epi8(va1, _mm_cmpgt_epi8(_mm_setzero_si128(), va1));
+ a1 = (const int8_t*) ((uintptr_t) a1 + k);
+
+ const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+
+ if (k > 2 * sizeof(int8_t)) {
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+
+ if (k > 4 * sizeof(int8_t)) {
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ }
+ }
+ }
+
+ __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
+ __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
+
+ const __m128 vscale0123 = _mm_loadu_ps((const float*) w);
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(float));
+ vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123);
+ vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale0123);
+
+ vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
+ vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
+
+ const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
+ __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
+
+ const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
+ const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max);
+ vacc01x0123 = _mm_min_epi16(_mm_max_epi16(vacc01x0123, voutput_min), voutput_max);
+
+ __m128i vout = _mm_packs_epi16(vacc01x0123, vacc01x0123);
+
+
+ if (nc >= 4) {
+ *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
+ vout = _mm_srli_si128(vout, 4);
+ *((uint32_t*) c1) = (uint32_t) _mm_cvtsi128_si32(vout);
+
+ c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
+ c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
+
+ a0 = (const int8_t*) ((uintptr_t) a0 - kc);
+ a1 = (const int8_t*) ((uintptr_t) a1 - kc);
+
+ nc -= 4;
+ } else {
+ if (nc & 2) {
+ *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
+ c0 += 2;
+ *((uint16_t*) c1) = (uint16_t) _mm_extract_epi16(vout, 2);
+ c1 += 2;
+ vout = _mm_srli_epi32(vout, 16);
+ }
+ if (nc & 1) {
+ *((int8_t*) c0) = (int8_t) _mm_cvtsi128_si32(vout);
+ *((int8_t*) c1) = (int8_t) _mm_extract_epi16(vout, 2);
+ }
+
+ nc = 0;
+ }
+ } while (nc != 0);
+}
diff --git a/src/qc8-gemm/gen/2x4c2-minmax-fp32-sse2-ld64.c b/src/qc8-gemm/gen/2x4c2-minmax-fp32-sse2-ld64.c
new file mode 100644
index 0000000..cbfea83
--- /dev/null
+++ b/src/qc8-gemm/gen/2x4c2-minmax-fp32-sse2-ld64.c
@@ -0,0 +1,185 @@
+// Auto-generated file. Do not edit!
+// Template: src/qs8-gemm/MRx4c2-sse.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 <emmintrin.h>
+
+#include <xnnpack/gemm.h>
+#include <xnnpack/math.h>
+
+
+
+void xnn_qc8_gemm_minmax_fp32_ukernel_2x4c2__sse2_ld64(
+ 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_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN XNN_DISABLE_MSAN
+{
+ assert(mr != 0);
+ assert(mr <= 2);
+ 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, 2);
+ 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;
+ }
+
+ do {
+ __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
+ __m128i vacc1x0123 = vacc0x0123;
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+
+ size_t k = kc;
+ while (k >= 8 * sizeof(int8_t)) {
+ const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
+ const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ a0 += 8;
+ const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
+ const __m128i vxa1 = _mm_unpacklo_epi8(va1, _mm_cmpgt_epi8(_mm_setzero_si128(), va1));
+ a1 += 8;
+
+ const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int8_t)));
+ const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
+ const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int8_t)));
+ const __m128i vxb3 = _mm_unpacklo_epi8(vb3, _mm_cmpgt_epi8(_mm_setzero_si128(), vb3));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
+
+ w = (const void*) ((uintptr_t) w + 32 * sizeof(int8_t));
+ k -= 8 * sizeof(int8_t);
+ }
+ if (k != 0) {
+ const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
+ const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ a0 = (const int8_t*) ((uintptr_t) a0 + k);
+ const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
+ const __m128i vxa1 = _mm_unpacklo_epi8(va1, _mm_cmpgt_epi8(_mm_setzero_si128(), va1));
+ a1 = (const int8_t*) ((uintptr_t) a1 + k);
+
+ const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+
+ if (k > 2 * sizeof(int8_t)) {
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+
+ if (k > 4 * sizeof(int8_t)) {
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ }
+ }
+ }
+
+ __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
+ __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
+
+ const __m128 vscale0123 = _mm_loadu_ps((const float*) w);
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(float));
+ vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123);
+ vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale0123);
+
+ vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
+ vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
+
+ const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
+ __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
+
+ const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
+ const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max);
+ vacc01x0123 = _mm_min_epi16(_mm_max_epi16(vacc01x0123, voutput_min), voutput_max);
+
+ __m128i vout = _mm_packs_epi16(vacc01x0123, vacc01x0123);
+
+
+ if (nc >= 4) {
+ *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
+ vout = _mm_srli_si128(vout, 4);
+ *((uint32_t*) c1) = (uint32_t) _mm_cvtsi128_si32(vout);
+
+ c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
+ c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
+
+ a0 = (const int8_t*) ((uintptr_t) a0 - kc);
+ a1 = (const int8_t*) ((uintptr_t) a1 - kc);
+
+ nc -= 4;
+ } else {
+ if (nc & 2) {
+ *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
+ c0 += 2;
+ *((uint16_t*) c1) = (uint16_t) _mm_extract_epi16(vout, 2);
+ c1 += 2;
+ vout = _mm_srli_epi32(vout, 16);
+ }
+ if (nc & 1) {
+ *((int8_t*) c0) = (int8_t) _mm_cvtsi128_si32(vout);
+ *((int8_t*) c1) = (int8_t) _mm_extract_epi16(vout, 2);
+ }
+
+ nc = 0;
+ }
+ } while (nc != 0);
+}
diff --git a/src/qc8-gemm/gen/2x4c2-minmax-fp32-sse41-ld128.c b/src/qc8-gemm/gen/2x4c2-minmax-fp32-sse41-ld128.c
new file mode 100644
index 0000000..8ace15a
--- /dev/null
+++ b/src/qc8-gemm/gen/2x4c2-minmax-fp32-sse41-ld128.c
@@ -0,0 +1,183 @@
+// Auto-generated file. Do not edit!
+// Template: src/qs8-gemm/MRx4c2-sse.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 <smmintrin.h>
+
+#include <xnnpack/gemm.h>
+#include <xnnpack/math.h>
+
+
+
+void xnn_qc8_gemm_minmax_fp32_ukernel_2x4c2__sse41_ld128(
+ 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_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN XNN_DISABLE_MSAN
+{
+ assert(mr != 0);
+ assert(mr <= 2);
+ 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, 2);
+ 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;
+ }
+
+ do {
+ __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
+ __m128i vacc1x0123 = vacc0x0123;
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+
+ size_t k = kc;
+ while (k >= 8 * sizeof(int8_t)) {
+ const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
+ const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
+ a0 += 8;
+ const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
+ const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
+ a1 += 8;
+
+ const __m128i vb01 = _mm_loadu_si128((const __m128i*) w);
+ const __m128i vsb01 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb01);
+ const __m128i vxb0 = _mm_unpacklo_epi8(vb01, vsb01);
+ const __m128i vxb1 = _mm_unpackhi_epi8(vb01, vsb01);
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
+ const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
+ const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
+ const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
+
+ w = (const void*) ((uintptr_t) w + 32 * sizeof(int8_t));
+ k -= 8 * sizeof(int8_t);
+ }
+ if (k != 0) {
+ const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
+ const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
+ a0 = (const int8_t*) ((uintptr_t) a0 + k);
+ const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
+ const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
+ a1 = (const int8_t*) ((uintptr_t) a1 + k);
+
+ const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+
+ if (k > 2 * sizeof(int8_t)) {
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+
+ if (k > 4 * sizeof(int8_t)) {
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ }
+ }
+ }
+
+ __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
+ __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
+
+ const __m128 vscale0123 = _mm_loadu_ps((const float*) w);
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(float));
+ vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123);
+ vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale0123);
+
+ vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
+ vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
+
+ const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4.output_zero_point);
+ __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
+
+
+ __m128i vout = _mm_packs_epi16(vacc01x0123, vacc01x0123);
+
+ vout = _mm_max_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_min));
+ vout = _mm_min_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_max));
+
+ if (nc >= 4) {
+ *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
+ *((uint32_t*) c1) = (uint32_t) _mm_extract_epi32(vout, 1);
+
+ c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
+ c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
+
+ a0 = (const int8_t*) ((uintptr_t) a0 - kc);
+ a1 = (const int8_t*) ((uintptr_t) a1 - kc);
+
+ nc -= 4;
+ } else {
+ if (nc & 2) {
+ *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
+ c0 += 2;
+ *((uint16_t*) c1) = (uint16_t) _mm_extract_epi16(vout, 2);
+ c1 += 2;
+ vout = _mm_srli_epi32(vout, 16);
+ }
+ if (nc & 1) {
+ *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
+ }
+
+ nc = 0;
+ }
+ } while (nc != 0);
+}
diff --git a/src/qc8-gemm/gen/2x4c2-minmax-fp32-sse41-ld64.c b/src/qc8-gemm/gen/2x4c2-minmax-fp32-sse41-ld64.c
new file mode 100644
index 0000000..f1c07c1
--- /dev/null
+++ b/src/qc8-gemm/gen/2x4c2-minmax-fp32-sse41-ld64.c
@@ -0,0 +1,183 @@
+// Auto-generated file. Do not edit!
+// Template: src/qs8-gemm/MRx4c2-sse.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 <smmintrin.h>
+
+#include <xnnpack/gemm.h>
+#include <xnnpack/math.h>
+
+
+
+void xnn_qc8_gemm_minmax_fp32_ukernel_2x4c2__sse41_ld64(
+ 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_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN XNN_DISABLE_MSAN
+{
+ assert(mr != 0);
+ assert(mr <= 2);
+ 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, 2);
+ 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;
+ }
+
+ do {
+ __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
+ __m128i vacc1x0123 = vacc0x0123;
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+
+ size_t k = kc;
+ while (k >= 8 * sizeof(int8_t)) {
+ const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
+ const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
+ a0 += 8;
+ const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
+ const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
+ a1 += 8;
+
+ const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int8_t)));
+ const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
+ const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int8_t)));
+ const __m128i vxb3 = _mm_cvtepi8_epi16(vb3);
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
+
+ w = (const void*) ((uintptr_t) w + 32 * sizeof(int8_t));
+ k -= 8 * sizeof(int8_t);
+ }
+ if (k != 0) {
+ const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
+ const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
+ a0 = (const int8_t*) ((uintptr_t) a0 + k);
+ const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
+ const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
+ a1 = (const int8_t*) ((uintptr_t) a1 + k);
+
+ const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+
+ if (k > 2 * sizeof(int8_t)) {
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+
+ if (k > 4 * sizeof(int8_t)) {
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ }
+ }
+ }
+
+ __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
+ __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
+
+ const __m128 vscale0123 = _mm_loadu_ps((const float*) w);
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(float));
+ vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123);
+ vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale0123);
+
+ vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
+ vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
+
+ const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4.output_zero_point);
+ __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
+
+
+ __m128i vout = _mm_packs_epi16(vacc01x0123, vacc01x0123);
+
+ vout = _mm_max_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_min));
+ vout = _mm_min_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_max));
+
+ if (nc >= 4) {
+ *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
+ *((uint32_t*) c1) = (uint32_t) _mm_extract_epi32(vout, 1);
+
+ c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
+ c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
+
+ a0 = (const int8_t*) ((uintptr_t) a0 - kc);
+ a1 = (const int8_t*) ((uintptr_t) a1 - kc);
+
+ nc -= 4;
+ } else {
+ if (nc & 2) {
+ *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
+ c0 += 2;
+ *((uint16_t*) c1) = (uint16_t) _mm_extract_epi16(vout, 2);
+ c1 += 2;
+ vout = _mm_srli_epi32(vout, 16);
+ }
+ if (nc & 1) {
+ *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
+ }
+
+ nc = 0;
+ }
+ } while (nc != 0);
+}
diff --git a/src/qc8-gemm/gen/2x4c2-minmax-fp32-xop-ld128.c b/src/qc8-gemm/gen/2x4c2-minmax-fp32-xop-ld128.c
new file mode 100644
index 0000000..7dcb9d9
--- /dev/null
+++ b/src/qc8-gemm/gen/2x4c2-minmax-fp32-xop-ld128.c
@@ -0,0 +1,188 @@
+// Auto-generated file. Do not edit!
+// Template: src/qs8-gemm/MRx4c2-sse.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>
+
+#if defined(__GNUC__) || defined(__clang__)
+ #include <x86intrin.h>
+#else
+ #include <immintrin.h>
+ #include <ammintrin.h>
+#endif
+
+#include <xnnpack/gemm.h>
+#include <xnnpack/math.h>
+
+
+
+void xnn_qc8_gemm_minmax_fp32_ukernel_2x4c2__xop_ld128(
+ 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_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN XNN_DISABLE_MSAN
+{
+ assert(mr != 0);
+ assert(mr <= 2);
+ 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, 2);
+ 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;
+ }
+
+ do {
+ __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
+ __m128i vacc1x0123 = vacc0x0123;
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+
+ size_t k = kc;
+ while (k >= 8 * sizeof(int8_t)) {
+ const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
+ const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
+ a0 += 8;
+ const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
+ const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
+ a1 += 8;
+
+ const __m128i vb01 = _mm_loadu_si128((const __m128i*) w);
+ const __m128i vsb01 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb01);
+ const __m128i vxb0 = _mm_unpacklo_epi8(vb01, vsb01);
+ const __m128i vxb1 = _mm_unpackhi_epi8(vb01, vsb01);
+
+ vacc0x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
+ vacc1x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc1x0123);
+
+ vacc0x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
+ vacc1x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc1x0123);
+ const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
+ const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
+ const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
+ const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
+
+ vacc0x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
+ vacc1x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc1x0123);
+
+ vacc0x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc0x0123);
+ vacc1x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc1x0123);
+
+ w = (const void*) ((uintptr_t) w + 32 * sizeof(int8_t));
+ k -= 8 * sizeof(int8_t);
+ }
+ if (k != 0) {
+ const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
+ const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
+ a0 = (const int8_t*) ((uintptr_t) a0 + k);
+ const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
+ const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
+ a1 = (const int8_t*) ((uintptr_t) a1 + k);
+
+ const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
+ vacc1x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc1x0123);
+
+ if (k > 2 * sizeof(int8_t)) {
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
+ vacc1x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc1x0123);
+
+ if (k > 4 * sizeof(int8_t)) {
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
+ vacc1x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc1x0123);
+ }
+ }
+ }
+
+ __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
+ __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
+
+ const __m128 vscale0123 = _mm_loadu_ps((const float*) w);
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(float));
+ vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123);
+ vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale0123);
+
+ vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
+ vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
+
+ const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4.output_zero_point);
+ __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
+
+
+ __m128i vout = _mm_packs_epi16(vacc01x0123, vacc01x0123);
+
+ vout = _mm_max_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_min));
+ vout = _mm_min_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_max));
+
+ if (nc >= 4) {
+ *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
+ *((uint32_t*) c1) = (uint32_t) _mm_extract_epi32(vout, 1);
+
+ c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
+ c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
+
+ a0 = (const int8_t*) ((uintptr_t) a0 - kc);
+ a1 = (const int8_t*) ((uintptr_t) a1 - kc);
+
+ nc -= 4;
+ } else {
+ if (nc & 2) {
+ *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
+ c0 += 2;
+ *((uint16_t*) c1) = (uint16_t) _mm_extract_epi16(vout, 2);
+ c1 += 2;
+ vout = _mm_srli_epi32(vout, 16);
+ }
+ if (nc & 1) {
+ *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
+ }
+
+ nc = 0;
+ }
+ } while (nc != 0);
+}
diff --git a/src/qc8-gemm/gen/2x4c2-minmax-fp32-xop-ld64.c b/src/qc8-gemm/gen/2x4c2-minmax-fp32-xop-ld64.c
new file mode 100644
index 0000000..ea2a328
--- /dev/null
+++ b/src/qc8-gemm/gen/2x4c2-minmax-fp32-xop-ld64.c
@@ -0,0 +1,188 @@
+// Auto-generated file. Do not edit!
+// Template: src/qs8-gemm/MRx4c2-sse.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>
+
+#if defined(__GNUC__) || defined(__clang__)
+ #include <x86intrin.h>
+#else
+ #include <immintrin.h>
+ #include <ammintrin.h>
+#endif
+
+#include <xnnpack/gemm.h>
+#include <xnnpack/math.h>
+
+
+
+void xnn_qc8_gemm_minmax_fp32_ukernel_2x4c2__xop_ld64(
+ 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_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN XNN_DISABLE_MSAN
+{
+ assert(mr != 0);
+ assert(mr <= 2);
+ 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, 2);
+ 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;
+ }
+
+ do {
+ __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
+ __m128i vacc1x0123 = vacc0x0123;
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+
+ size_t k = kc;
+ while (k >= 8 * sizeof(int8_t)) {
+ const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
+ const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
+ a0 += 8;
+ const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
+ const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
+ a1 += 8;
+
+ const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
+
+ vacc0x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
+ vacc1x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc1x0123);
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int8_t)));
+ const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
+
+ vacc0x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
+ vacc1x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc1x0123);
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
+ const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
+
+ vacc0x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
+ vacc1x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc1x0123);
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int8_t)));
+ const __m128i vxb3 = _mm_cvtepi8_epi16(vb3);
+
+ vacc0x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc0x0123);
+ vacc1x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc1x0123);
+
+ w = (const void*) ((uintptr_t) w + 32 * sizeof(int8_t));
+ k -= 8 * sizeof(int8_t);
+ }
+ if (k != 0) {
+ const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
+ const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
+ a0 = (const int8_t*) ((uintptr_t) a0 + k);
+ const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
+ const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
+ a1 = (const int8_t*) ((uintptr_t) a1 + k);
+
+ const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
+ vacc1x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc1x0123);
+
+ if (k > 2 * sizeof(int8_t)) {
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
+ vacc1x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc1x0123);
+
+ if (k > 4 * sizeof(int8_t)) {
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
+ vacc1x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc1x0123);
+ }
+ }
+ }
+
+ __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
+ __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
+
+ const __m128 vscale0123 = _mm_loadu_ps((const float*) w);
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(float));
+ vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123);
+ vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale0123);
+
+ vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
+ vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
+
+ const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4.output_zero_point);
+ __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
+
+
+ __m128i vout = _mm_packs_epi16(vacc01x0123, vacc01x0123);
+
+ vout = _mm_max_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_min));
+ vout = _mm_min_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_max));
+
+ if (nc >= 4) {
+ *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
+ *((uint32_t*) c1) = (uint32_t) _mm_extract_epi32(vout, 1);
+
+ c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
+ c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
+
+ a0 = (const int8_t*) ((uintptr_t) a0 - kc);
+ a1 = (const int8_t*) ((uintptr_t) a1 - kc);
+
+ nc -= 4;
+ } else {
+ if (nc & 2) {
+ *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
+ c0 += 2;
+ *((uint16_t*) c1) = (uint16_t) _mm_extract_epi16(vout, 2);
+ c1 += 2;
+ vout = _mm_srli_epi32(vout, 16);
+ }
+ if (nc & 1) {
+ *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
+ }
+
+ nc = 0;
+ }
+ } while (nc != 0);
+}
diff --git a/src/qc8-gemm/gen/2x4c8-minmax-fp32-avx-ld128.c b/src/qc8-gemm/gen/2x4c8-minmax-fp32-avx-ld128.c
new file mode 100644
index 0000000..3595651
--- /dev/null
+++ b/src/qc8-gemm/gen/2x4c8-minmax-fp32-avx-ld128.c
@@ -0,0 +1,147 @@
+// Auto-generated file. Do not edit!
+// Template: src/qs8-gemm/MRx4c8-sse.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 <smmintrin.h>
+
+#include <xnnpack/gemm.h>
+#include <xnnpack/math.h>
+
+
+void xnn_qc8_gemm_minmax_fp32_ukernel_2x4c8__avx_ld128(
+ 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_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN XNN_DISABLE_MSAN
+{
+ assert(mr != 0);
+ assert(mr <= 2);
+ 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, 8);
+ 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;
+ }
+
+ do {
+ __m128i vacc0x0 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[0]);
+ __m128i vacc0x1 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[1]);
+ __m128i vacc0x2 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[2]);
+ __m128i vacc0x3 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[3]);
+ __m128i vacc1x0 = vacc0x0;
+ __m128i vacc1x1 = vacc0x1;
+ __m128i vacc1x2 = vacc0x2;
+ __m128i vacc1x3 = vacc0x3;
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+
+ size_t k = 0;
+ while (k < kc) {
+ const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
+ const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
+ a0 += 8;
+ const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
+ const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
+ a1 += 8;
+
+ const __m128i vb01 = _mm_load_si128((const __m128i*) w);
+ const __m128i vsb01 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb01);
+ const __m128i vxb0 = _mm_unpacklo_epi8(vb01, vsb01);
+ const __m128i vxb1 = _mm_unpackhi_epi8(vb01, vsb01);
+
+ vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
+ vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
+ vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0));
+ vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
+ const __m128i vb23 = _mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
+ const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
+ const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
+ const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
+
+ vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
+ vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
+ vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2));
+ vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3));
+
+ w = (const void*) ((uintptr_t) w + 32 * sizeof(int8_t));
+ k += 8 * sizeof(int8_t);
+ }
+
+ const __m128i vacc0x01 = _mm_hadd_epi32(vacc0x0, vacc0x1);
+ const __m128i vacc0x23 = _mm_hadd_epi32(vacc0x2, vacc0x3);
+ const __m128i vacc1x01 = _mm_hadd_epi32(vacc1x0, vacc1x1);
+ const __m128i vacc1x23 = _mm_hadd_epi32(vacc1x2, vacc1x3);
+
+ __m128i vacc0x0123 = _mm_hadd_epi32(vacc0x01, vacc0x23);
+ __m128i vacc1x0123 = _mm_hadd_epi32(vacc1x01, vacc1x23);
+
+ __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
+ __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
+
+ const __m128 vscale0123 = _mm_load_ps((const float*) w);
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(float));
+ vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123);
+ vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale0123);
+
+ vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
+ vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
+
+ const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4.output_zero_point);
+ __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
+
+
+ __m128i vout = _mm_packs_epi16(vacc01x0123, vacc01x0123);
+
+ vout = _mm_max_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_min));
+ vout = _mm_min_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_max));
+
+ if (nc >= 4) {
+ *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
+ *((uint32_t*) c1) = (uint32_t) _mm_extract_epi32(vout, 1);
+
+ c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
+ c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
+
+ a0 = (const int8_t*) ((uintptr_t) a0 - kc);
+ a1 = (const int8_t*) ((uintptr_t) a1 - kc);
+
+ nc -= 4;
+ } else {
+ if (nc & 2) {
+ *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
+ c0 += 2;
+ *((uint16_t*) c1) = (uint16_t) _mm_extract_epi16(vout, 2);
+ c1 += 2;
+ vout = _mm_srli_epi32(vout, 16);
+ }
+ if (nc & 1) {
+ *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
+ }
+
+ nc = 0;
+ }
+ } while (nc != 0);
+}
diff --git a/src/qc8-gemm/gen/2x4c8-minmax-fp32-avx-ld64.c b/src/qc8-gemm/gen/2x4c8-minmax-fp32-avx-ld64.c
new file mode 100644
index 0000000..d9b0072
--- /dev/null
+++ b/src/qc8-gemm/gen/2x4c8-minmax-fp32-avx-ld64.c
@@ -0,0 +1,149 @@
+// Auto-generated file. Do not edit!
+// Template: src/qs8-gemm/MRx4c8-sse.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 <smmintrin.h>
+
+#include <xnnpack/gemm.h>
+#include <xnnpack/math.h>
+
+
+void xnn_qc8_gemm_minmax_fp32_ukernel_2x4c8__avx_ld64(
+ 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_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN XNN_DISABLE_MSAN
+{
+ assert(mr != 0);
+ assert(mr <= 2);
+ 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, 8);
+ 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;
+ }
+
+ do {
+ __m128i vacc0x0 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[0]);
+ __m128i vacc0x1 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[1]);
+ __m128i vacc0x2 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[2]);
+ __m128i vacc0x3 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[3]);
+ __m128i vacc1x0 = vacc0x0;
+ __m128i vacc1x1 = vacc0x1;
+ __m128i vacc1x2 = vacc0x2;
+ __m128i vacc1x3 = vacc0x3;
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+
+ size_t k = 0;
+ while (k < kc) {
+ const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
+ const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
+ a0 += 8;
+ const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
+ const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
+ a1 += 8;
+
+ const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
+
+ vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
+ vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0));
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int8_t)));
+ const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
+
+ vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
+ vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
+ const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
+
+ vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
+ vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2));
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int8_t)));
+ const __m128i vxb3 = _mm_cvtepi8_epi16(vb3);
+
+ vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
+ vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3));
+
+ w = (const void*) ((uintptr_t) w + 32 * sizeof(int8_t));
+ k += 8 * sizeof(int8_t);
+ }
+
+ const __m128i vacc0x01 = _mm_hadd_epi32(vacc0x0, vacc0x1);
+ const __m128i vacc0x23 = _mm_hadd_epi32(vacc0x2, vacc0x3);
+ const __m128i vacc1x01 = _mm_hadd_epi32(vacc1x0, vacc1x1);
+ const __m128i vacc1x23 = _mm_hadd_epi32(vacc1x2, vacc1x3);
+
+ __m128i vacc0x0123 = _mm_hadd_epi32(vacc0x01, vacc0x23);
+ __m128i vacc1x0123 = _mm_hadd_epi32(vacc1x01, vacc1x23);
+
+ __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
+ __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
+
+ const __m128 vscale0123 = _mm_load_ps((const float*) w);
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(float));
+ vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123);
+ vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale0123);
+
+ vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
+ vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
+
+ const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4.output_zero_point);
+ __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
+
+
+ __m128i vout = _mm_packs_epi16(vacc01x0123, vacc01x0123);
+
+ vout = _mm_max_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_min));
+ vout = _mm_min_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_max));
+
+ if (nc >= 4) {
+ *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
+ *((uint32_t*) c1) = (uint32_t) _mm_extract_epi32(vout, 1);
+
+ c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
+ c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
+
+ a0 = (const int8_t*) ((uintptr_t) a0 - kc);
+ a1 = (const int8_t*) ((uintptr_t) a1 - kc);
+
+ nc -= 4;
+ } else {
+ if (nc & 2) {
+ *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
+ c0 += 2;
+ *((uint16_t*) c1) = (uint16_t) _mm_extract_epi16(vout, 2);
+ c1 += 2;
+ vout = _mm_srli_epi32(vout, 16);
+ }
+ if (nc & 1) {
+ *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
+ }
+
+ nc = 0;
+ }
+ } while (nc != 0);
+}
diff --git a/src/qc8-gemm/gen/2x4c8-minmax-fp32-sse2-ld128.c b/src/qc8-gemm/gen/2x4c8-minmax-fp32-sse2-ld128.c
new file mode 100644
index 0000000..7ab73a6
--- /dev/null
+++ b/src/qc8-gemm/gen/2x4c8-minmax-fp32-sse2-ld128.c
@@ -0,0 +1,149 @@
+// Auto-generated file. Do not edit!
+// Template: src/qs8-gemm/MRx4c8-sse.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 <emmintrin.h>
+
+#include <xnnpack/gemm.h>
+#include <xnnpack/math.h>
+
+
+void xnn_qc8_gemm_minmax_fp32_ukernel_2x4c8__sse2_ld128(
+ 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_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN XNN_DISABLE_MSAN
+{
+ assert(mr != 0);
+ assert(mr <= 2);
+ 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, 8);
+ 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;
+ }
+
+ do {
+ __m128i vacc0x0 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[0]);
+ __m128i vacc0x1 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[1]);
+ __m128i vacc0x2 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[2]);
+ __m128i vacc0x3 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[3]);
+ __m128i vacc1x0 = vacc0x0;
+ __m128i vacc1x1 = vacc0x1;
+ __m128i vacc1x2 = vacc0x2;
+ __m128i vacc1x3 = vacc0x3;
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+
+ size_t k = 0;
+ while (k < kc) {
+ const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
+ const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ a0 += 8;
+ const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
+ const __m128i vxa1 = _mm_unpacklo_epi8(va1, _mm_cmpgt_epi8(_mm_setzero_si128(), va1));
+ a1 += 8;
+
+ const __m128i vb01 = _mm_load_si128((const __m128i*) w);
+ const __m128i vsb01 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb01);
+ const __m128i vxb0 = _mm_unpacklo_epi8(vb01, vsb01);
+ const __m128i vxb1 = _mm_unpackhi_epi8(vb01, vsb01);
+
+ vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
+ vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
+ vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0));
+ vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
+ const __m128i vb23 = _mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
+ const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
+ const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
+ const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
+
+ vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
+ vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
+ vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2));
+ vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3));
+
+ w = (const void*) ((uintptr_t) w + 32 * sizeof(int8_t));
+ k += 8 * sizeof(int8_t);
+ }
+
+ const __m128i vacc0x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x0, vacc0x2), _mm_unpackhi_epi32(vacc0x0, vacc0x2));
+ const __m128i vacc0x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x1, vacc0x3), _mm_unpackhi_epi32(vacc0x1, vacc0x3));
+ const __m128i vacc1x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x0, vacc1x2), _mm_unpackhi_epi32(vacc1x0, vacc1x2));
+ const __m128i vacc1x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x1, vacc1x3), _mm_unpackhi_epi32(vacc1x1, vacc1x3));
+
+ __m128i vacc0x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x02, vacc0x13), _mm_unpackhi_epi32(vacc0x02, vacc0x13));
+ __m128i vacc1x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x02, vacc1x13), _mm_unpackhi_epi32(vacc1x02, vacc1x13));
+
+ __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
+ __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
+
+ const __m128 vscale0123 = _mm_load_ps((const float*) w);
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(float));
+ vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123);
+ vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale0123);
+
+ vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
+ vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
+
+ const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
+ __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
+
+ const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
+ const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max);
+ vacc01x0123 = _mm_min_epi16(_mm_max_epi16(vacc01x0123, voutput_min), voutput_max);
+
+ __m128i vout = _mm_packs_epi16(vacc01x0123, vacc01x0123);
+
+
+ if (nc >= 4) {
+ *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
+ vout = _mm_srli_si128(vout, 4);
+ *((uint32_t*) c1) = (uint32_t) _mm_cvtsi128_si32(vout);
+
+ c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
+ c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
+
+ a0 = (const int8_t*) ((uintptr_t) a0 - kc);
+ a1 = (const int8_t*) ((uintptr_t) a1 - kc);
+
+ nc -= 4;
+ } else {
+ if (nc & 2) {
+ *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
+ c0 += 2;
+ *((uint16_t*) c1) = (uint16_t) _mm_extract_epi16(vout, 2);
+ c1 += 2;
+ vout = _mm_srli_epi32(vout, 16);
+ }
+ if (nc & 1) {
+ *((int8_t*) c0) = (int8_t) _mm_cvtsi128_si32(vout);
+ *((int8_t*) c1) = (int8_t) _mm_extract_epi16(vout, 2);
+ }
+
+ nc = 0;
+ }
+ } while (nc != 0);
+}
diff --git a/src/qc8-gemm/gen/2x4c8-minmax-fp32-sse2-ld64.c b/src/qc8-gemm/gen/2x4c8-minmax-fp32-sse2-ld64.c
new file mode 100644
index 0000000..e205ebe
--- /dev/null
+++ b/src/qc8-gemm/gen/2x4c8-minmax-fp32-sse2-ld64.c
@@ -0,0 +1,151 @@
+// Auto-generated file. Do not edit!
+// Template: src/qs8-gemm/MRx4c8-sse.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 <emmintrin.h>
+
+#include <xnnpack/gemm.h>
+#include <xnnpack/math.h>
+
+
+void xnn_qc8_gemm_minmax_fp32_ukernel_2x4c8__sse2_ld64(
+ 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_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN XNN_DISABLE_MSAN
+{
+ assert(mr != 0);
+ assert(mr <= 2);
+ 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, 8);
+ 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;
+ }
+
+ do {
+ __m128i vacc0x0 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[0]);
+ __m128i vacc0x1 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[1]);
+ __m128i vacc0x2 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[2]);
+ __m128i vacc0x3 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[3]);
+ __m128i vacc1x0 = vacc0x0;
+ __m128i vacc1x1 = vacc0x1;
+ __m128i vacc1x2 = vacc0x2;
+ __m128i vacc1x3 = vacc0x3;
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+
+ size_t k = 0;
+ while (k < kc) {
+ const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
+ const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ a0 += 8;
+ const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
+ const __m128i vxa1 = _mm_unpacklo_epi8(va1, _mm_cmpgt_epi8(_mm_setzero_si128(), va1));
+ a1 += 8;
+
+ const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+
+ vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
+ vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0));
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int8_t)));
+ const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+
+ vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
+ vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
+ const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+
+ vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
+ vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2));
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int8_t)));
+ const __m128i vxb3 = _mm_unpacklo_epi8(vb3, _mm_cmpgt_epi8(_mm_setzero_si128(), vb3));
+
+ vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
+ vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3));
+
+ w = (const void*) ((uintptr_t) w + 32 * sizeof(int8_t));
+ k += 8 * sizeof(int8_t);
+ }
+
+ const __m128i vacc0x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x0, vacc0x2), _mm_unpackhi_epi32(vacc0x0, vacc0x2));
+ const __m128i vacc0x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x1, vacc0x3), _mm_unpackhi_epi32(vacc0x1, vacc0x3));
+ const __m128i vacc1x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x0, vacc1x2), _mm_unpackhi_epi32(vacc1x0, vacc1x2));
+ const __m128i vacc1x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x1, vacc1x3), _mm_unpackhi_epi32(vacc1x1, vacc1x3));
+
+ __m128i vacc0x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x02, vacc0x13), _mm_unpackhi_epi32(vacc0x02, vacc0x13));
+ __m128i vacc1x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x02, vacc1x13), _mm_unpackhi_epi32(vacc1x02, vacc1x13));
+
+ __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
+ __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
+
+ const __m128 vscale0123 = _mm_load_ps((const float*) w);
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(float));
+ vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123);
+ vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale0123);
+
+ vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
+ vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
+
+ const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
+ __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
+
+ const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
+ const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max);
+ vacc01x0123 = _mm_min_epi16(_mm_max_epi16(vacc01x0123, voutput_min), voutput_max);
+
+ __m128i vout = _mm_packs_epi16(vacc01x0123, vacc01x0123);
+
+
+ if (nc >= 4) {
+ *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
+ vout = _mm_srli_si128(vout, 4);
+ *((uint32_t*) c1) = (uint32_t) _mm_cvtsi128_si32(vout);
+
+ c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
+ c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
+
+ a0 = (const int8_t*) ((uintptr_t) a0 - kc);
+ a1 = (const int8_t*) ((uintptr_t) a1 - kc);
+
+ nc -= 4;
+ } else {
+ if (nc & 2) {
+ *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
+ c0 += 2;
+ *((uint16_t*) c1) = (uint16_t) _mm_extract_epi16(vout, 2);
+ c1 += 2;
+ vout = _mm_srli_epi32(vout, 16);
+ }
+ if (nc & 1) {
+ *((int8_t*) c0) = (int8_t) _mm_cvtsi128_si32(vout);
+ *((int8_t*) c1) = (int8_t) _mm_extract_epi16(vout, 2);
+ }
+
+ nc = 0;
+ }
+ } while (nc != 0);
+}
diff --git a/src/qc8-gemm/gen/2x4c8-minmax-fp32-sse41-ld128.c b/src/qc8-gemm/gen/2x4c8-minmax-fp32-sse41-ld128.c
new file mode 100644
index 0000000..7fa45b8
--- /dev/null
+++ b/src/qc8-gemm/gen/2x4c8-minmax-fp32-sse41-ld128.c
@@ -0,0 +1,147 @@
+// Auto-generated file. Do not edit!
+// Template: src/qs8-gemm/MRx4c8-sse.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 <smmintrin.h>
+
+#include <xnnpack/gemm.h>
+#include <xnnpack/math.h>
+
+
+void xnn_qc8_gemm_minmax_fp32_ukernel_2x4c8__sse41_ld128(
+ 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_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN XNN_DISABLE_MSAN
+{
+ assert(mr != 0);
+ assert(mr <= 2);
+ 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, 8);
+ 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;
+ }
+
+ do {
+ __m128i vacc0x0 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[0]);
+ __m128i vacc0x1 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[1]);
+ __m128i vacc0x2 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[2]);
+ __m128i vacc0x3 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[3]);
+ __m128i vacc1x0 = vacc0x0;
+ __m128i vacc1x1 = vacc0x1;
+ __m128i vacc1x2 = vacc0x2;
+ __m128i vacc1x3 = vacc0x3;
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+
+ size_t k = 0;
+ while (k < kc) {
+ const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
+ const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
+ a0 += 8;
+ const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
+ const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
+ a1 += 8;
+
+ const __m128i vb01 = _mm_load_si128((const __m128i*) w);
+ const __m128i vsb01 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb01);
+ const __m128i vxb0 = _mm_unpacklo_epi8(vb01, vsb01);
+ const __m128i vxb1 = _mm_unpackhi_epi8(vb01, vsb01);
+
+ vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
+ vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
+ vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0));
+ vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
+ const __m128i vb23 = _mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
+ const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
+ const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
+ const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
+
+ vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
+ vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
+ vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2));
+ vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3));
+
+ w = (const void*) ((uintptr_t) w + 32 * sizeof(int8_t));
+ k += 8 * sizeof(int8_t);
+ }
+
+ const __m128i vacc0x01 = _mm_hadd_epi32(vacc0x0, vacc0x1);
+ const __m128i vacc0x23 = _mm_hadd_epi32(vacc0x2, vacc0x3);
+ const __m128i vacc1x01 = _mm_hadd_epi32(vacc1x0, vacc1x1);
+ const __m128i vacc1x23 = _mm_hadd_epi32(vacc1x2, vacc1x3);
+
+ __m128i vacc0x0123 = _mm_hadd_epi32(vacc0x01, vacc0x23);
+ __m128i vacc1x0123 = _mm_hadd_epi32(vacc1x01, vacc1x23);
+
+ __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
+ __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
+
+ const __m128 vscale0123 = _mm_load_ps((const float*) w);
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(float));
+ vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123);
+ vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale0123);
+
+ vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
+ vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
+
+ const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4.output_zero_point);
+ __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
+
+
+ __m128i vout = _mm_packs_epi16(vacc01x0123, vacc01x0123);
+
+ vout = _mm_max_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_min));
+ vout = _mm_min_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_max));
+
+ if (nc >= 4) {
+ *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
+ *((uint32_t*) c1) = (uint32_t) _mm_extract_epi32(vout, 1);
+
+ c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
+ c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
+
+ a0 = (const int8_t*) ((uintptr_t) a0 - kc);
+ a1 = (const int8_t*) ((uintptr_t) a1 - kc);
+
+ nc -= 4;
+ } else {
+ if (nc & 2) {
+ *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
+ c0 += 2;
+ *((uint16_t*) c1) = (uint16_t) _mm_extract_epi16(vout, 2);
+ c1 += 2;
+ vout = _mm_srli_epi32(vout, 16);
+ }
+ if (nc & 1) {
+ *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
+ }
+
+ nc = 0;
+ }
+ } while (nc != 0);
+}
diff --git a/src/qc8-gemm/gen/2x4c8-minmax-fp32-sse41-ld64.c b/src/qc8-gemm/gen/2x4c8-minmax-fp32-sse41-ld64.c
new file mode 100644
index 0000000..d758e6c
--- /dev/null
+++ b/src/qc8-gemm/gen/2x4c8-minmax-fp32-sse41-ld64.c
@@ -0,0 +1,149 @@
+// Auto-generated file. Do not edit!
+// Template: src/qs8-gemm/MRx4c8-sse.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 <smmintrin.h>
+
+#include <xnnpack/gemm.h>
+#include <xnnpack/math.h>
+
+
+void xnn_qc8_gemm_minmax_fp32_ukernel_2x4c8__sse41_ld64(
+ 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_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN XNN_DISABLE_MSAN
+{
+ assert(mr != 0);
+ assert(mr <= 2);
+ 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, 8);
+ 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;
+ }
+
+ do {
+ __m128i vacc0x0 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[0]);
+ __m128i vacc0x1 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[1]);
+ __m128i vacc0x2 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[2]);
+ __m128i vacc0x3 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[3]);
+ __m128i vacc1x0 = vacc0x0;
+ __m128i vacc1x1 = vacc0x1;
+ __m128i vacc1x2 = vacc0x2;
+ __m128i vacc1x3 = vacc0x3;
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+
+ size_t k = 0;
+ while (k < kc) {
+ const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
+ const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
+ a0 += 8;
+ const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
+ const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
+ a1 += 8;
+
+ const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
+
+ vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
+ vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0));
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int8_t)));
+ const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
+
+ vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
+ vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
+ const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
+
+ vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
+ vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2));
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int8_t)));
+ const __m128i vxb3 = _mm_cvtepi8_epi16(vb3);
+
+ vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
+ vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3));
+
+ w = (const void*) ((uintptr_t) w + 32 * sizeof(int8_t));
+ k += 8 * sizeof(int8_t);
+ }
+
+ const __m128i vacc0x01 = _mm_hadd_epi32(vacc0x0, vacc0x1);
+ const __m128i vacc0x23 = _mm_hadd_epi32(vacc0x2, vacc0x3);
+ const __m128i vacc1x01 = _mm_hadd_epi32(vacc1x0, vacc1x1);
+ const __m128i vacc1x23 = _mm_hadd_epi32(vacc1x2, vacc1x3);
+
+ __m128i vacc0x0123 = _mm_hadd_epi32(vacc0x01, vacc0x23);
+ __m128i vacc1x0123 = _mm_hadd_epi32(vacc1x01, vacc1x23);
+
+ __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
+ __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
+
+ const __m128 vscale0123 = _mm_load_ps((const float*) w);
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(float));
+ vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123);
+ vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale0123);
+
+ vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
+ vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
+
+ const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4.output_zero_point);
+ __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
+
+
+ __m128i vout = _mm_packs_epi16(vacc01x0123, vacc01x0123);
+
+ vout = _mm_max_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_min));
+ vout = _mm_min_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_max));
+
+ if (nc >= 4) {
+ *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
+ *((uint32_t*) c1) = (uint32_t) _mm_extract_epi32(vout, 1);
+
+ c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
+ c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
+
+ a0 = (const int8_t*) ((uintptr_t) a0 - kc);
+ a1 = (const int8_t*) ((uintptr_t) a1 - kc);
+
+ nc -= 4;
+ } else {
+ if (nc & 2) {
+ *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
+ c0 += 2;
+ *((uint16_t*) c1) = (uint16_t) _mm_extract_epi16(vout, 2);
+ c1 += 2;
+ vout = _mm_srli_epi32(vout, 16);
+ }
+ if (nc & 1) {
+ *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
+ }
+
+ nc = 0;
+ }
+ } while (nc != 0);
+}
diff --git a/src/qc8-gemm/gen/2x4c8-minmax-fp32-xop-ld128.c b/src/qc8-gemm/gen/2x4c8-minmax-fp32-xop-ld128.c
new file mode 100644
index 0000000..33313e4
--- /dev/null
+++ b/src/qc8-gemm/gen/2x4c8-minmax-fp32-xop-ld128.c
@@ -0,0 +1,152 @@
+// Auto-generated file. Do not edit!
+// Template: src/qs8-gemm/MRx4c8-sse.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>
+
+#if defined(__GNUC__) || defined(__clang__)
+ #include <x86intrin.h>
+#else
+ #include <immintrin.h>
+ #include <ammintrin.h>
+#endif
+
+#include <xnnpack/gemm.h>
+#include <xnnpack/math.h>
+
+
+void xnn_qc8_gemm_minmax_fp32_ukernel_2x4c8__xop_ld128(
+ 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_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN XNN_DISABLE_MSAN
+{
+ assert(mr != 0);
+ assert(mr <= 2);
+ 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, 8);
+ 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;
+ }
+
+ do {
+ __m128i vacc0x0 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[0]);
+ __m128i vacc0x1 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[1]);
+ __m128i vacc0x2 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[2]);
+ __m128i vacc0x3 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[3]);
+ __m128i vacc1x0 = vacc0x0;
+ __m128i vacc1x1 = vacc0x1;
+ __m128i vacc1x2 = vacc0x2;
+ __m128i vacc1x3 = vacc0x3;
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+
+ size_t k = 0;
+ while (k < kc) {
+ const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
+ const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
+ a0 += 8;
+ const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
+ const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
+ a1 += 8;
+
+ const __m128i vb01 = _mm_load_si128((const __m128i*) w);
+ const __m128i vsb01 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb01);
+ const __m128i vxb0 = _mm_unpacklo_epi8(vb01, vsb01);
+ const __m128i vxb1 = _mm_unpackhi_epi8(vb01, vsb01);
+
+ vacc0x0 = _mm_maddd_epi16(vxa0, vxb0, vacc0x0);
+ vacc0x1 = _mm_maddd_epi16(vxa0, vxb1, vacc0x1);
+ vacc1x0 = _mm_maddd_epi16(vxa1, vxb0, vacc1x0);
+ vacc1x1 = _mm_maddd_epi16(vxa1, vxb1, vacc1x1);
+ const __m128i vb23 = _mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
+ const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
+ const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
+ const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
+
+ vacc0x2 = _mm_maddd_epi16(vxa0, vxb2, vacc0x2);
+ vacc0x3 = _mm_maddd_epi16(vxa0, vxb3, vacc0x3);
+ vacc1x2 = _mm_maddd_epi16(vxa1, vxb2, vacc1x2);
+ vacc1x3 = _mm_maddd_epi16(vxa1, vxb3, vacc1x3);
+
+ w = (const void*) ((uintptr_t) w + 32 * sizeof(int8_t));
+ k += 8 * sizeof(int8_t);
+ }
+
+ const __m128i vacc0x01 = _mm_hadd_epi32(vacc0x0, vacc0x1);
+ const __m128i vacc0x23 = _mm_hadd_epi32(vacc0x2, vacc0x3);
+ const __m128i vacc1x01 = _mm_hadd_epi32(vacc1x0, vacc1x1);
+ const __m128i vacc1x23 = _mm_hadd_epi32(vacc1x2, vacc1x3);
+
+ __m128i vacc0x0123 = _mm_hadd_epi32(vacc0x01, vacc0x23);
+ __m128i vacc1x0123 = _mm_hadd_epi32(vacc1x01, vacc1x23);
+
+ __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
+ __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
+
+ const __m128 vscale0123 = _mm_load_ps((const float*) w);
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(float));
+ vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123);
+ vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale0123);
+
+ vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
+ vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
+
+ const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4.output_zero_point);
+ __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
+
+
+ __m128i vout = _mm_packs_epi16(vacc01x0123, vacc01x0123);
+
+ vout = _mm_max_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_min));
+ vout = _mm_min_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_max));
+
+ if (nc >= 4) {
+ *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
+ *((uint32_t*) c1) = (uint32_t) _mm_extract_epi32(vout, 1);
+
+ c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
+ c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
+
+ a0 = (const int8_t*) ((uintptr_t) a0 - kc);
+ a1 = (const int8_t*) ((uintptr_t) a1 - kc);
+
+ nc -= 4;
+ } else {
+ if (nc & 2) {
+ *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
+ c0 += 2;
+ *((uint16_t*) c1) = (uint16_t) _mm_extract_epi16(vout, 2);
+ c1 += 2;
+ vout = _mm_srli_epi32(vout, 16);
+ }
+ if (nc & 1) {
+ *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
+ }
+
+ nc = 0;
+ }
+ } while (nc != 0);
+}
diff --git a/src/qc8-gemm/gen/2x4c8-minmax-fp32-xop-ld64.c b/src/qc8-gemm/gen/2x4c8-minmax-fp32-xop-ld64.c
new file mode 100644
index 0000000..7d2cffe
--- /dev/null
+++ b/src/qc8-gemm/gen/2x4c8-minmax-fp32-xop-ld64.c
@@ -0,0 +1,154 @@
+// Auto-generated file. Do not edit!
+// Template: src/qs8-gemm/MRx4c8-sse.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>
+
+#if defined(__GNUC__) || defined(__clang__)
+ #include <x86intrin.h>
+#else
+ #include <immintrin.h>
+ #include <ammintrin.h>
+#endif
+
+#include <xnnpack/gemm.h>
+#include <xnnpack/math.h>
+
+
+void xnn_qc8_gemm_minmax_fp32_ukernel_2x4c8__xop_ld64(
+ 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_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN XNN_DISABLE_MSAN
+{
+ assert(mr != 0);
+ assert(mr <= 2);
+ 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, 8);
+ 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;
+ }
+
+ do {
+ __m128i vacc0x0 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[0]);
+ __m128i vacc0x1 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[1]);
+ __m128i vacc0x2 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[2]);
+ __m128i vacc0x3 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[3]);
+ __m128i vacc1x0 = vacc0x0;
+ __m128i vacc1x1 = vacc0x1;
+ __m128i vacc1x2 = vacc0x2;
+ __m128i vacc1x3 = vacc0x3;
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+
+ size_t k = 0;
+ while (k < kc) {
+ const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
+ const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
+ a0 += 8;
+ const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
+ const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
+ a1 += 8;
+
+ const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
+
+ vacc0x0 = _mm_maddd_epi16(vxa0, vxb0, vacc0x0);
+ vacc1x0 = _mm_maddd_epi16(vxa1, vxb0, vacc1x0);
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int8_t)));
+ const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
+
+ vacc0x1 = _mm_maddd_epi16(vxa0, vxb1, vacc0x1);
+ vacc1x1 = _mm_maddd_epi16(vxa1, vxb1, vacc1x1);
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
+ const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
+
+ vacc0x2 = _mm_maddd_epi16(vxa0, vxb2, vacc0x2);
+ vacc1x2 = _mm_maddd_epi16(vxa1, vxb2, vacc1x2);
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int8_t)));
+ const __m128i vxb3 = _mm_cvtepi8_epi16(vb3);
+
+ vacc0x3 = _mm_maddd_epi16(vxa0, vxb3, vacc0x3);
+ vacc1x3 = _mm_maddd_epi16(vxa1, vxb3, vacc1x3);
+
+ w = (const void*) ((uintptr_t) w + 32 * sizeof(int8_t));
+ k += 8 * sizeof(int8_t);
+ }
+
+ const __m128i vacc0x01 = _mm_hadd_epi32(vacc0x0, vacc0x1);
+ const __m128i vacc0x23 = _mm_hadd_epi32(vacc0x2, vacc0x3);
+ const __m128i vacc1x01 = _mm_hadd_epi32(vacc1x0, vacc1x1);
+ const __m128i vacc1x23 = _mm_hadd_epi32(vacc1x2, vacc1x3);
+
+ __m128i vacc0x0123 = _mm_hadd_epi32(vacc0x01, vacc0x23);
+ __m128i vacc1x0123 = _mm_hadd_epi32(vacc1x01, vacc1x23);
+
+ __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
+ __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
+
+ const __m128 vscale0123 = _mm_load_ps((const float*) w);
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(float));
+ vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123);
+ vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale0123);
+
+ vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
+ vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
+
+ const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4.output_zero_point);
+ __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
+
+
+ __m128i vout = _mm_packs_epi16(vacc01x0123, vacc01x0123);
+
+ vout = _mm_max_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_min));
+ vout = _mm_min_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_max));
+
+ if (nc >= 4) {
+ *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
+ *((uint32_t*) c1) = (uint32_t) _mm_extract_epi32(vout, 1);
+
+ c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
+ c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
+
+ a0 = (const int8_t*) ((uintptr_t) a0 - kc);
+ a1 = (const int8_t*) ((uintptr_t) a1 - kc);
+
+ nc -= 4;
+ } else {
+ if (nc & 2) {
+ *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
+ c0 += 2;
+ *((uint16_t*) c1) = (uint16_t) _mm_extract_epi16(vout, 2);
+ c1 += 2;
+ vout = _mm_srli_epi32(vout, 16);
+ }
+ if (nc & 1) {
+ *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
+ }
+
+ nc = 0;
+ }
+ } while (nc != 0);
+}
diff --git a/src/qc8-gemm/gen/3x4c2-minmax-fp32-avx-ld128.c b/src/qc8-gemm/gen/3x4c2-minmax-fp32-avx-ld128.c
new file mode 100644
index 0000000..1550ecc
--- /dev/null
+++ b/src/qc8-gemm/gen/3x4c2-minmax-fp32-avx-ld128.c
@@ -0,0 +1,220 @@
+// Auto-generated file. Do not edit!
+// Template: src/qs8-gemm/MRx4c2-sse.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 <smmintrin.h>
+
+#include <xnnpack/gemm.h>
+#include <xnnpack/math.h>
+
+
+
+void xnn_qc8_gemm_minmax_fp32_ukernel_3x4c2__avx_ld128(
+ 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_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN XNN_DISABLE_MSAN
+{
+ assert(mr != 0);
+ assert(mr <= 3);
+ 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, 2);
+ 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;
+ }
+
+ do {
+ __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
+ __m128i vacc1x0123 = vacc0x0123;
+ __m128i vacc2x0123 = vacc0x0123;
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+
+ size_t k = kc;
+ while (k >= 8 * sizeof(int8_t)) {
+ const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
+ const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
+ a0 += 8;
+ const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
+ const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
+ a1 += 8;
+ const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
+ const __m128i vxa2 = _mm_cvtepi8_epi16(va2);
+ a2 += 8;
+
+ const __m128i vb01 = _mm_loadu_si128((const __m128i*) w);
+ const __m128i vsb01 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb01);
+ const __m128i vxb0 = _mm_unpacklo_epi8(vb01, vsb01);
+ const __m128i vxb1 = _mm_unpackhi_epi8(vb01, vsb01);
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+ vacc2x0123 = _mm_add_epi32(vacc2x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ vacc2x0123 = _mm_add_epi32(vacc2x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
+ const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
+ const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
+ const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ vacc2x0123 = _mm_add_epi32(vacc2x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
+ vacc2x0123 = _mm_add_epi32(vacc2x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
+
+ w = (const void*) ((uintptr_t) w + 32 * sizeof(int8_t));
+ k -= 8 * sizeof(int8_t);
+ }
+ if (k != 0) {
+ const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
+ const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
+ a0 = (const int8_t*) ((uintptr_t) a0 + k);
+ const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
+ const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
+ a1 = (const int8_t*) ((uintptr_t) a1 + k);
+ const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
+ const __m128i vxa2 = _mm_cvtepi8_epi16(va2);
+ a2 = (const int8_t*) ((uintptr_t) a2 + k);
+
+ const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+ vacc2x0123 = _mm_add_epi32(vacc2x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+
+ if (k > 2 * sizeof(int8_t)) {
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ vacc2x0123 = _mm_add_epi32(vacc2x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+
+ if (k > 4 * sizeof(int8_t)) {
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ vacc2x0123 = _mm_add_epi32(vacc2x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ }
+ }
+ }
+
+ __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
+ __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
+ __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123);
+
+ const __m128 vscale0123 = _mm_loadu_ps((const float*) w);
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(float));
+ vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123);
+ vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale0123);
+ vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale0123);
+
+ vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
+ vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
+ vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123);
+
+ const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4.output_zero_point);
+ __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
+ __m128i vacc22x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc2x0123), voutput_zero_point);
+
+
+ __m128i vout = _mm_packs_epi16(vacc01x0123, vacc22x0123);
+
+ vout = _mm_max_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_min));
+ vout = _mm_min_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_max));
+
+ if (nc >= 4) {
+ *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
+ *((uint32_t*) c1) = (uint32_t) _mm_extract_epi32(vout, 1);
+ *((uint32_t*) c2) = (uint32_t) _mm_extract_epi32(vout, 2);
+
+ c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
+ c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
+ c2 = (int8_t*) ((uintptr_t) c2 + 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);
+
+ nc -= 4;
+ } else {
+ if (nc & 2) {
+ *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
+ c0 += 2;
+ *((uint16_t*) c1) = (uint16_t) _mm_extract_epi16(vout, 2);
+ c1 += 2;
+ *((uint16_t*) c2) = (uint16_t) _mm_extract_epi16(vout, 4);
+ c2 += 2;
+ vout = _mm_srli_epi32(vout, 16);
+ }
+ if (nc & 1) {
+ *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
+ *((int8_t*) c2) = (int8_t) _mm_extract_epi8(vout, 8);
+ }
+
+ nc = 0;
+ }
+ } while (nc != 0);
+}
diff --git a/src/qc8-gemm/gen/3x4c2-minmax-fp32-avx-ld64.c b/src/qc8-gemm/gen/3x4c2-minmax-fp32-avx-ld64.c
new file mode 100644
index 0000000..bf95edb
--- /dev/null
+++ b/src/qc8-gemm/gen/3x4c2-minmax-fp32-avx-ld64.c
@@ -0,0 +1,220 @@
+// Auto-generated file. Do not edit!
+// Template: src/qs8-gemm/MRx4c2-sse.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 <smmintrin.h>
+
+#include <xnnpack/gemm.h>
+#include <xnnpack/math.h>
+
+
+
+void xnn_qc8_gemm_minmax_fp32_ukernel_3x4c2__avx_ld64(
+ 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_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN XNN_DISABLE_MSAN
+{
+ assert(mr != 0);
+ assert(mr <= 3);
+ 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, 2);
+ 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;
+ }
+
+ do {
+ __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
+ __m128i vacc1x0123 = vacc0x0123;
+ __m128i vacc2x0123 = vacc0x0123;
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+
+ size_t k = kc;
+ while (k >= 8 * sizeof(int8_t)) {
+ const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
+ const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
+ a0 += 8;
+ const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
+ const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
+ a1 += 8;
+ const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
+ const __m128i vxa2 = _mm_cvtepi8_epi16(va2);
+ a2 += 8;
+
+ const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+ vacc2x0123 = _mm_add_epi32(vacc2x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int8_t)));
+ const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ vacc2x0123 = _mm_add_epi32(vacc2x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
+ const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ vacc2x0123 = _mm_add_epi32(vacc2x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int8_t)));
+ const __m128i vxb3 = _mm_cvtepi8_epi16(vb3);
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
+ vacc2x0123 = _mm_add_epi32(vacc2x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
+
+ w = (const void*) ((uintptr_t) w + 32 * sizeof(int8_t));
+ k -= 8 * sizeof(int8_t);
+ }
+ if (k != 0) {
+ const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
+ const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
+ a0 = (const int8_t*) ((uintptr_t) a0 + k);
+ const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
+ const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
+ a1 = (const int8_t*) ((uintptr_t) a1 + k);
+ const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
+ const __m128i vxa2 = _mm_cvtepi8_epi16(va2);
+ a2 = (const int8_t*) ((uintptr_t) a2 + k);
+
+ const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+ vacc2x0123 = _mm_add_epi32(vacc2x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+
+ if (k > 2 * sizeof(int8_t)) {
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ vacc2x0123 = _mm_add_epi32(vacc2x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+
+ if (k > 4 * sizeof(int8_t)) {
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ vacc2x0123 = _mm_add_epi32(vacc2x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ }
+ }
+ }
+
+ __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
+ __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
+ __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123);
+
+ const __m128 vscale0123 = _mm_loadu_ps((const float*) w);
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(float));
+ vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123);
+ vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale0123);
+ vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale0123);
+
+ vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
+ vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
+ vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123);
+
+ const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4.output_zero_point);
+ __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
+ __m128i vacc22x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc2x0123), voutput_zero_point);
+
+
+ __m128i vout = _mm_packs_epi16(vacc01x0123, vacc22x0123);
+
+ vout = _mm_max_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_min));
+ vout = _mm_min_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_max));
+
+ if (nc >= 4) {
+ *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
+ *((uint32_t*) c1) = (uint32_t) _mm_extract_epi32(vout, 1);
+ *((uint32_t*) c2) = (uint32_t) _mm_extract_epi32(vout, 2);
+
+ c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
+ c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
+ c2 = (int8_t*) ((uintptr_t) c2 + 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);
+
+ nc -= 4;
+ } else {
+ if (nc & 2) {
+ *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
+ c0 += 2;
+ *((uint16_t*) c1) = (uint16_t) _mm_extract_epi16(vout, 2);
+ c1 += 2;
+ *((uint16_t*) c2) = (uint16_t) _mm_extract_epi16(vout, 4);
+ c2 += 2;
+ vout = _mm_srli_epi32(vout, 16);
+ }
+ if (nc & 1) {
+ *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
+ *((int8_t*) c2) = (int8_t) _mm_extract_epi8(vout, 8);
+ }
+
+ nc = 0;
+ }
+ } while (nc != 0);
+}
diff --git a/src/qc8-gemm/gen/3x4c2-minmax-fp32-sse2-ld128.c b/src/qc8-gemm/gen/3x4c2-minmax-fp32-sse2-ld128.c
new file mode 100644
index 0000000..e337bbc
--- /dev/null
+++ b/src/qc8-gemm/gen/3x4c2-minmax-fp32-sse2-ld128.c
@@ -0,0 +1,224 @@
+// Auto-generated file. Do not edit!
+// Template: src/qs8-gemm/MRx4c2-sse.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 <emmintrin.h>
+
+#include <xnnpack/gemm.h>
+#include <xnnpack/math.h>
+
+
+
+void xnn_qc8_gemm_minmax_fp32_ukernel_3x4c2__sse2_ld128(
+ 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_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN XNN_DISABLE_MSAN
+{
+ assert(mr != 0);
+ assert(mr <= 3);
+ 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, 2);
+ 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;
+ }
+
+ do {
+ __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
+ __m128i vacc1x0123 = vacc0x0123;
+ __m128i vacc2x0123 = vacc0x0123;
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+
+ size_t k = kc;
+ while (k >= 8 * sizeof(int8_t)) {
+ const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
+ const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ a0 += 8;
+ const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
+ const __m128i vxa1 = _mm_unpacklo_epi8(va1, _mm_cmpgt_epi8(_mm_setzero_si128(), va1));
+ a1 += 8;
+ const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
+ const __m128i vxa2 = _mm_unpacklo_epi8(va2, _mm_cmpgt_epi8(_mm_setzero_si128(), va2));
+ a2 += 8;
+
+ const __m128i vb01 = _mm_loadu_si128((const __m128i*) w);
+ const __m128i vsb01 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb01);
+ const __m128i vxb0 = _mm_unpacklo_epi8(vb01, vsb01);
+ const __m128i vxb1 = _mm_unpackhi_epi8(vb01, vsb01);
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+ vacc2x0123 = _mm_add_epi32(vacc2x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ vacc2x0123 = _mm_add_epi32(vacc2x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
+ const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
+ const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
+ const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ vacc2x0123 = _mm_add_epi32(vacc2x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
+ vacc2x0123 = _mm_add_epi32(vacc2x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
+
+ w = (const void*) ((uintptr_t) w + 32 * sizeof(int8_t));
+ k -= 8 * sizeof(int8_t);
+ }
+ if (k != 0) {
+ const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
+ const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ a0 = (const int8_t*) ((uintptr_t) a0 + k);
+ const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
+ const __m128i vxa1 = _mm_unpacklo_epi8(va1, _mm_cmpgt_epi8(_mm_setzero_si128(), va1));
+ a1 = (const int8_t*) ((uintptr_t) a1 + k);
+ const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
+ const __m128i vxa2 = _mm_unpacklo_epi8(va2, _mm_cmpgt_epi8(_mm_setzero_si128(), va2));
+ a2 = (const int8_t*) ((uintptr_t) a2 + k);
+
+ const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+ vacc2x0123 = _mm_add_epi32(vacc2x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+
+ if (k > 2 * sizeof(int8_t)) {
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ vacc2x0123 = _mm_add_epi32(vacc2x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+
+ if (k > 4 * sizeof(int8_t)) {
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ vacc2x0123 = _mm_add_epi32(vacc2x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ }
+ }
+ }
+
+ __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
+ __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
+ __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123);
+
+ const __m128 vscale0123 = _mm_loadu_ps((const float*) w);
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(float));
+ vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123);
+ vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale0123);
+ vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale0123);
+
+ vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
+ vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
+ vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123);
+
+ const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
+ __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
+ __m128i vacc22x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc2x0123), voutput_zero_point);
+
+ const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
+ const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max);
+ vacc01x0123 = _mm_min_epi16(_mm_max_epi16(vacc01x0123, voutput_min), voutput_max);
+ vacc22x0123 = _mm_min_epi16(_mm_max_epi16(vacc22x0123, voutput_min), voutput_max);
+
+ __m128i vout = _mm_packs_epi16(vacc01x0123, vacc22x0123);
+
+
+ if (nc >= 4) {
+ *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
+ vout = _mm_srli_si128(vout, 4);
+ *((uint32_t*) c1) = (uint32_t) _mm_cvtsi128_si32(vout);
+ vout = _mm_srli_si128(vout, 4);
+ *((uint32_t*) c2) = (uint32_t) _mm_cvtsi128_si32(vout);
+
+ c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
+ c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
+ c2 = (int8_t*) ((uintptr_t) c2 + 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);
+
+ nc -= 4;
+ } else {
+ if (nc & 2) {
+ *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
+ c0 += 2;
+ *((uint16_t*) c1) = (uint16_t) _mm_extract_epi16(vout, 2);
+ c1 += 2;
+ *((uint16_t*) c2) = (uint16_t) _mm_extract_epi16(vout, 4);
+ c2 += 2;
+ vout = _mm_srli_epi32(vout, 16);
+ }
+ if (nc & 1) {
+ *((int8_t*) c0) = (int8_t) _mm_cvtsi128_si32(vout);
+ *((int8_t*) c1) = (int8_t) _mm_extract_epi16(vout, 2);
+ *((int8_t*) c2) = (int8_t) _mm_extract_epi16(vout, 4);
+ }
+
+ nc = 0;
+ }
+ } while (nc != 0);
+}
diff --git a/src/qc8-gemm/gen/3x4c2-minmax-fp32-sse2-ld64.c b/src/qc8-gemm/gen/3x4c2-minmax-fp32-sse2-ld64.c
new file mode 100644
index 0000000..bb1730c
--- /dev/null
+++ b/src/qc8-gemm/gen/3x4c2-minmax-fp32-sse2-ld64.c
@@ -0,0 +1,224 @@
+// Auto-generated file. Do not edit!
+// Template: src/qs8-gemm/MRx4c2-sse.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 <emmintrin.h>
+
+#include <xnnpack/gemm.h>
+#include <xnnpack/math.h>
+
+
+
+void xnn_qc8_gemm_minmax_fp32_ukernel_3x4c2__sse2_ld64(
+ 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_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN XNN_DISABLE_MSAN
+{
+ assert(mr != 0);
+ assert(mr <= 3);
+ 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, 2);
+ 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;
+ }
+
+ do {
+ __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
+ __m128i vacc1x0123 = vacc0x0123;
+ __m128i vacc2x0123 = vacc0x0123;
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+
+ size_t k = kc;
+ while (k >= 8 * sizeof(int8_t)) {
+ const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
+ const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ a0 += 8;
+ const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
+ const __m128i vxa1 = _mm_unpacklo_epi8(va1, _mm_cmpgt_epi8(_mm_setzero_si128(), va1));
+ a1 += 8;
+ const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
+ const __m128i vxa2 = _mm_unpacklo_epi8(va2, _mm_cmpgt_epi8(_mm_setzero_si128(), va2));
+ a2 += 8;
+
+ const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+ vacc2x0123 = _mm_add_epi32(vacc2x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int8_t)));
+ const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ vacc2x0123 = _mm_add_epi32(vacc2x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
+ const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ vacc2x0123 = _mm_add_epi32(vacc2x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int8_t)));
+ const __m128i vxb3 = _mm_unpacklo_epi8(vb3, _mm_cmpgt_epi8(_mm_setzero_si128(), vb3));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
+ vacc2x0123 = _mm_add_epi32(vacc2x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
+
+ w = (const void*) ((uintptr_t) w + 32 * sizeof(int8_t));
+ k -= 8 * sizeof(int8_t);
+ }
+ if (k != 0) {
+ const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
+ const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ a0 = (const int8_t*) ((uintptr_t) a0 + k);
+ const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
+ const __m128i vxa1 = _mm_unpacklo_epi8(va1, _mm_cmpgt_epi8(_mm_setzero_si128(), va1));
+ a1 = (const int8_t*) ((uintptr_t) a1 + k);
+ const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
+ const __m128i vxa2 = _mm_unpacklo_epi8(va2, _mm_cmpgt_epi8(_mm_setzero_si128(), va2));
+ a2 = (const int8_t*) ((uintptr_t) a2 + k);
+
+ const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+ vacc2x0123 = _mm_add_epi32(vacc2x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+
+ if (k > 2 * sizeof(int8_t)) {
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ vacc2x0123 = _mm_add_epi32(vacc2x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+
+ if (k > 4 * sizeof(int8_t)) {
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ vacc2x0123 = _mm_add_epi32(vacc2x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ }
+ }
+ }
+
+ __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
+ __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
+ __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123);
+
+ const __m128 vscale0123 = _mm_loadu_ps((const float*) w);
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(float));
+ vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123);
+ vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale0123);
+ vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale0123);
+
+ vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
+ vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
+ vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123);
+
+ const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
+ __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
+ __m128i vacc22x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc2x0123), voutput_zero_point);
+
+ const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
+ const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max);
+ vacc01x0123 = _mm_min_epi16(_mm_max_epi16(vacc01x0123, voutput_min), voutput_max);
+ vacc22x0123 = _mm_min_epi16(_mm_max_epi16(vacc22x0123, voutput_min), voutput_max);
+
+ __m128i vout = _mm_packs_epi16(vacc01x0123, vacc22x0123);
+
+
+ if (nc >= 4) {
+ *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
+ vout = _mm_srli_si128(vout, 4);
+ *((uint32_t*) c1) = (uint32_t) _mm_cvtsi128_si32(vout);
+ vout = _mm_srli_si128(vout, 4);
+ *((uint32_t*) c2) = (uint32_t) _mm_cvtsi128_si32(vout);
+
+ c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
+ c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
+ c2 = (int8_t*) ((uintptr_t) c2 + 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);
+
+ nc -= 4;
+ } else {
+ if (nc & 2) {
+ *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
+ c0 += 2;
+ *((uint16_t*) c1) = (uint16_t) _mm_extract_epi16(vout, 2);
+ c1 += 2;
+ *((uint16_t*) c2) = (uint16_t) _mm_extract_epi16(vout, 4);
+ c2 += 2;
+ vout = _mm_srli_epi32(vout, 16);
+ }
+ if (nc & 1) {
+ *((int8_t*) c0) = (int8_t) _mm_cvtsi128_si32(vout);
+ *((int8_t*) c1) = (int8_t) _mm_extract_epi16(vout, 2);
+ *((int8_t*) c2) = (int8_t) _mm_extract_epi16(vout, 4);
+ }
+
+ nc = 0;
+ }
+ } while (nc != 0);
+}
diff --git a/src/qc8-gemm/gen/3x4c2-minmax-fp32-sse41-ld128.c b/src/qc8-gemm/gen/3x4c2-minmax-fp32-sse41-ld128.c
new file mode 100644
index 0000000..f5f3f1c
--- /dev/null
+++ b/src/qc8-gemm/gen/3x4c2-minmax-fp32-sse41-ld128.c
@@ -0,0 +1,220 @@
+// Auto-generated file. Do not edit!
+// Template: src/qs8-gemm/MRx4c2-sse.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 <smmintrin.h>
+
+#include <xnnpack/gemm.h>
+#include <xnnpack/math.h>
+
+
+
+void xnn_qc8_gemm_minmax_fp32_ukernel_3x4c2__sse41_ld128(
+ 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_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN XNN_DISABLE_MSAN
+{
+ assert(mr != 0);
+ assert(mr <= 3);
+ 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, 2);
+ 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;
+ }
+
+ do {
+ __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
+ __m128i vacc1x0123 = vacc0x0123;
+ __m128i vacc2x0123 = vacc0x0123;
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+
+ size_t k = kc;
+ while (k >= 8 * sizeof(int8_t)) {
+ const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
+ const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
+ a0 += 8;
+ const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
+ const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
+ a1 += 8;
+ const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
+ const __m128i vxa2 = _mm_cvtepi8_epi16(va2);
+ a2 += 8;
+
+ const __m128i vb01 = _mm_loadu_si128((const __m128i*) w);
+ const __m128i vsb01 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb01);
+ const __m128i vxb0 = _mm_unpacklo_epi8(vb01, vsb01);
+ const __m128i vxb1 = _mm_unpackhi_epi8(vb01, vsb01);
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+ vacc2x0123 = _mm_add_epi32(vacc2x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ vacc2x0123 = _mm_add_epi32(vacc2x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
+ const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
+ const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
+ const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ vacc2x0123 = _mm_add_epi32(vacc2x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
+ vacc2x0123 = _mm_add_epi32(vacc2x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
+
+ w = (const void*) ((uintptr_t) w + 32 * sizeof(int8_t));
+ k -= 8 * sizeof(int8_t);
+ }
+ if (k != 0) {
+ const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
+ const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
+ a0 = (const int8_t*) ((uintptr_t) a0 + k);
+ const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
+ const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
+ a1 = (const int8_t*) ((uintptr_t) a1 + k);
+ const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
+ const __m128i vxa2 = _mm_cvtepi8_epi16(va2);
+ a2 = (const int8_t*) ((uintptr_t) a2 + k);
+
+ const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+ vacc2x0123 = _mm_add_epi32(vacc2x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+
+ if (k > 2 * sizeof(int8_t)) {
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ vacc2x0123 = _mm_add_epi32(vacc2x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+
+ if (k > 4 * sizeof(int8_t)) {
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ vacc2x0123 = _mm_add_epi32(vacc2x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ }
+ }
+ }
+
+ __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
+ __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
+ __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123);
+
+ const __m128 vscale0123 = _mm_loadu_ps((const float*) w);
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(float));
+ vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123);
+ vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale0123);
+ vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale0123);
+
+ vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
+ vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
+ vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123);
+
+ const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4.output_zero_point);
+ __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
+ __m128i vacc22x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc2x0123), voutput_zero_point);
+
+
+ __m128i vout = _mm_packs_epi16(vacc01x0123, vacc22x0123);
+
+ vout = _mm_max_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_min));
+ vout = _mm_min_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_max));
+
+ if (nc >= 4) {
+ *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
+ *((uint32_t*) c1) = (uint32_t) _mm_extract_epi32(vout, 1);
+ *((uint32_t*) c2) = (uint32_t) _mm_extract_epi32(vout, 2);
+
+ c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
+ c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
+ c2 = (int8_t*) ((uintptr_t) c2 + 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);
+
+ nc -= 4;
+ } else {
+ if (nc & 2) {
+ *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
+ c0 += 2;
+ *((uint16_t*) c1) = (uint16_t) _mm_extract_epi16(vout, 2);
+ c1 += 2;
+ *((uint16_t*) c2) = (uint16_t) _mm_extract_epi16(vout, 4);
+ c2 += 2;
+ vout = _mm_srli_epi32(vout, 16);
+ }
+ if (nc & 1) {
+ *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
+ *((int8_t*) c2) = (int8_t) _mm_extract_epi8(vout, 8);
+ }
+
+ nc = 0;
+ }
+ } while (nc != 0);
+}
diff --git a/src/qc8-gemm/gen/3x4c2-minmax-fp32-sse41-ld64.c b/src/qc8-gemm/gen/3x4c2-minmax-fp32-sse41-ld64.c
new file mode 100644
index 0000000..042d7b5
--- /dev/null
+++ b/src/qc8-gemm/gen/3x4c2-minmax-fp32-sse41-ld64.c
@@ -0,0 +1,220 @@
+// Auto-generated file. Do not edit!
+// Template: src/qs8-gemm/MRx4c2-sse.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 <smmintrin.h>
+
+#include <xnnpack/gemm.h>
+#include <xnnpack/math.h>
+
+
+
+void xnn_qc8_gemm_minmax_fp32_ukernel_3x4c2__sse41_ld64(
+ 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_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN XNN_DISABLE_MSAN
+{
+ assert(mr != 0);
+ assert(mr <= 3);
+ 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, 2);
+ 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;
+ }
+
+ do {
+ __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
+ __m128i vacc1x0123 = vacc0x0123;
+ __m128i vacc2x0123 = vacc0x0123;
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+
+ size_t k = kc;
+ while (k >= 8 * sizeof(int8_t)) {
+ const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
+ const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
+ a0 += 8;
+ const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
+ const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
+ a1 += 8;
+ const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
+ const __m128i vxa2 = _mm_cvtepi8_epi16(va2);
+ a2 += 8;
+
+ const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+ vacc2x0123 = _mm_add_epi32(vacc2x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int8_t)));
+ const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ vacc2x0123 = _mm_add_epi32(vacc2x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
+ const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ vacc2x0123 = _mm_add_epi32(vacc2x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int8_t)));
+ const __m128i vxb3 = _mm_cvtepi8_epi16(vb3);
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
+ vacc2x0123 = _mm_add_epi32(vacc2x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
+
+ w = (const void*) ((uintptr_t) w + 32 * sizeof(int8_t));
+ k -= 8 * sizeof(int8_t);
+ }
+ if (k != 0) {
+ const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
+ const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
+ a0 = (const int8_t*) ((uintptr_t) a0 + k);
+ const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
+ const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
+ a1 = (const int8_t*) ((uintptr_t) a1 + k);
+ const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
+ const __m128i vxa2 = _mm_cvtepi8_epi16(va2);
+ a2 = (const int8_t*) ((uintptr_t) a2 + k);
+
+ const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+ vacc2x0123 = _mm_add_epi32(vacc2x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+
+ if (k > 2 * sizeof(int8_t)) {
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ vacc2x0123 = _mm_add_epi32(vacc2x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+
+ if (k > 4 * sizeof(int8_t)) {
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ vacc2x0123 = _mm_add_epi32(vacc2x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ }
+ }
+ }
+
+ __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
+ __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
+ __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123);
+
+ const __m128 vscale0123 = _mm_loadu_ps((const float*) w);
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(float));
+ vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123);
+ vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale0123);
+ vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale0123);
+
+ vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
+ vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
+ vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123);
+
+ const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4.output_zero_point);
+ __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
+ __m128i vacc22x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc2x0123), voutput_zero_point);
+
+
+ __m128i vout = _mm_packs_epi16(vacc01x0123, vacc22x0123);
+
+ vout = _mm_max_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_min));
+ vout = _mm_min_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_max));
+
+ if (nc >= 4) {
+ *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
+ *((uint32_t*) c1) = (uint32_t) _mm_extract_epi32(vout, 1);
+ *((uint32_t*) c2) = (uint32_t) _mm_extract_epi32(vout, 2);
+
+ c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
+ c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
+ c2 = (int8_t*) ((uintptr_t) c2 + 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);
+
+ nc -= 4;
+ } else {
+ if (nc & 2) {
+ *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
+ c0 += 2;
+ *((uint16_t*) c1) = (uint16_t) _mm_extract_epi16(vout, 2);
+ c1 += 2;
+ *((uint16_t*) c2) = (uint16_t) _mm_extract_epi16(vout, 4);
+ c2 += 2;
+ vout = _mm_srli_epi32(vout, 16);
+ }
+ if (nc & 1) {
+ *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
+ *((int8_t*) c2) = (int8_t) _mm_extract_epi8(vout, 8);
+ }
+
+ nc = 0;
+ }
+ } while (nc != 0);
+}
diff --git a/src/qc8-gemm/gen/3x4c2-minmax-fp32-xop-ld128.c b/src/qc8-gemm/gen/3x4c2-minmax-fp32-xop-ld128.c
new file mode 100644
index 0000000..def9e5e
--- /dev/null
+++ b/src/qc8-gemm/gen/3x4c2-minmax-fp32-xop-ld128.c
@@ -0,0 +1,225 @@
+// Auto-generated file. Do not edit!
+// Template: src/qs8-gemm/MRx4c2-sse.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>
+
+#if defined(__GNUC__) || defined(__clang__)
+ #include <x86intrin.h>
+#else
+ #include <immintrin.h>
+ #include <ammintrin.h>
+#endif
+
+#include <xnnpack/gemm.h>
+#include <xnnpack/math.h>
+
+
+
+void xnn_qc8_gemm_minmax_fp32_ukernel_3x4c2__xop_ld128(
+ 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_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN XNN_DISABLE_MSAN
+{
+ assert(mr != 0);
+ assert(mr <= 3);
+ 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, 2);
+ 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;
+ }
+
+ do {
+ __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
+ __m128i vacc1x0123 = vacc0x0123;
+ __m128i vacc2x0123 = vacc0x0123;
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+
+ size_t k = kc;
+ while (k >= 8 * sizeof(int8_t)) {
+ const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
+ const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
+ a0 += 8;
+ const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
+ const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
+ a1 += 8;
+ const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
+ const __m128i vxa2 = _mm_cvtepi8_epi16(va2);
+ a2 += 8;
+
+ const __m128i vb01 = _mm_loadu_si128((const __m128i*) w);
+ const __m128i vsb01 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb01);
+ const __m128i vxb0 = _mm_unpacklo_epi8(vb01, vsb01);
+ const __m128i vxb1 = _mm_unpackhi_epi8(vb01, vsb01);
+
+ vacc0x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
+ vacc1x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc1x0123);
+ vacc2x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc2x0123);
+
+ vacc0x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
+ vacc1x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc1x0123);
+ vacc2x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc2x0123);
+ const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
+ const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
+ const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
+ const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
+
+ vacc0x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
+ vacc1x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc1x0123);
+ vacc2x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc2x0123);
+
+ vacc0x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc0x0123);
+ vacc1x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc1x0123);
+ vacc2x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc2x0123);
+
+ w = (const void*) ((uintptr_t) w + 32 * sizeof(int8_t));
+ k -= 8 * sizeof(int8_t);
+ }
+ if (k != 0) {
+ const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
+ const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
+ a0 = (const int8_t*) ((uintptr_t) a0 + k);
+ const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
+ const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
+ a1 = (const int8_t*) ((uintptr_t) a1 + k);
+ const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
+ const __m128i vxa2 = _mm_cvtepi8_epi16(va2);
+ a2 = (const int8_t*) ((uintptr_t) a2 + k);
+
+ const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
+ vacc1x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc1x0123);
+ vacc2x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc2x0123);
+
+ if (k > 2 * sizeof(int8_t)) {
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
+ vacc1x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc1x0123);
+ vacc2x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc2x0123);
+
+ if (k > 4 * sizeof(int8_t)) {
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
+ vacc1x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc1x0123);
+ vacc2x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc2x0123);
+ }
+ }
+ }
+
+ __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
+ __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
+ __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123);
+
+ const __m128 vscale0123 = _mm_loadu_ps((const float*) w);
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(float));
+ vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123);
+ vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale0123);
+ vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale0123);
+
+ vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
+ vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
+ vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123);
+
+ const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4.output_zero_point);
+ __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
+ __m128i vacc22x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc2x0123), voutput_zero_point);
+
+
+ __m128i vout = _mm_packs_epi16(vacc01x0123, vacc22x0123);
+
+ vout = _mm_max_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_min));
+ vout = _mm_min_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_max));
+
+ if (nc >= 4) {
+ *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
+ *((uint32_t*) c1) = (uint32_t) _mm_extract_epi32(vout, 1);
+ *((uint32_t*) c2) = (uint32_t) _mm_extract_epi32(vout, 2);
+
+ c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
+ c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
+ c2 = (int8_t*) ((uintptr_t) c2 + 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);
+
+ nc -= 4;
+ } else {
+ if (nc & 2) {
+ *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
+ c0 += 2;
+ *((uint16_t*) c1) = (uint16_t) _mm_extract_epi16(vout, 2);
+ c1 += 2;
+ *((uint16_t*) c2) = (uint16_t) _mm_extract_epi16(vout, 4);
+ c2 += 2;
+ vout = _mm_srli_epi32(vout, 16);
+ }
+ if (nc & 1) {
+ *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
+ *((int8_t*) c2) = (int8_t) _mm_extract_epi8(vout, 8);
+ }
+
+ nc = 0;
+ }
+ } while (nc != 0);
+}
diff --git a/src/qc8-gemm/gen/3x4c2-minmax-fp32-xop-ld64.c b/src/qc8-gemm/gen/3x4c2-minmax-fp32-xop-ld64.c
new file mode 100644
index 0000000..184754d
--- /dev/null
+++ b/src/qc8-gemm/gen/3x4c2-minmax-fp32-xop-ld64.c
@@ -0,0 +1,225 @@
+// Auto-generated file. Do not edit!
+// Template: src/qs8-gemm/MRx4c2-sse.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>
+
+#if defined(__GNUC__) || defined(__clang__)
+ #include <x86intrin.h>
+#else
+ #include <immintrin.h>
+ #include <ammintrin.h>
+#endif
+
+#include <xnnpack/gemm.h>
+#include <xnnpack/math.h>
+
+
+
+void xnn_qc8_gemm_minmax_fp32_ukernel_3x4c2__xop_ld64(
+ 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_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN XNN_DISABLE_MSAN
+{
+ assert(mr != 0);
+ assert(mr <= 3);
+ 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, 2);
+ 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;
+ }
+
+ do {
+ __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
+ __m128i vacc1x0123 = vacc0x0123;
+ __m128i vacc2x0123 = vacc0x0123;
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+
+ size_t k = kc;
+ while (k >= 8 * sizeof(int8_t)) {
+ const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
+ const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
+ a0 += 8;
+ const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
+ const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
+ a1 += 8;
+ const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
+ const __m128i vxa2 = _mm_cvtepi8_epi16(va2);
+ a2 += 8;
+
+ const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
+
+ vacc0x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
+ vacc1x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc1x0123);
+ vacc2x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc2x0123);
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int8_t)));
+ const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
+
+ vacc0x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
+ vacc1x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc1x0123);
+ vacc2x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc2x0123);
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
+ const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
+
+ vacc0x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
+ vacc1x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc1x0123);
+ vacc2x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc2x0123);
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int8_t)));
+ const __m128i vxb3 = _mm_cvtepi8_epi16(vb3);
+
+ vacc0x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc0x0123);
+ vacc1x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc1x0123);
+ vacc2x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc2x0123);
+
+ w = (const void*) ((uintptr_t) w + 32 * sizeof(int8_t));
+ k -= 8 * sizeof(int8_t);
+ }
+ if (k != 0) {
+ const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
+ const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
+ a0 = (const int8_t*) ((uintptr_t) a0 + k);
+ const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
+ const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
+ a1 = (const int8_t*) ((uintptr_t) a1 + k);
+ const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
+ const __m128i vxa2 = _mm_cvtepi8_epi16(va2);
+ a2 = (const int8_t*) ((uintptr_t) a2 + k);
+
+ const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
+ vacc1x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc1x0123);
+ vacc2x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc2x0123);
+
+ if (k > 2 * sizeof(int8_t)) {
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
+ vacc1x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc1x0123);
+ vacc2x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc2x0123);
+
+ if (k > 4 * sizeof(int8_t)) {
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
+ vacc1x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc1x0123);
+ vacc2x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc2x0123);
+ }
+ }
+ }
+
+ __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
+ __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
+ __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123);
+
+ const __m128 vscale0123 = _mm_loadu_ps((const float*) w);
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(float));
+ vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123);
+ vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale0123);
+ vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale0123);
+
+ vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
+ vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
+ vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123);
+
+ const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4.output_zero_point);
+ __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
+ __m128i vacc22x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc2x0123), voutput_zero_point);
+
+
+ __m128i vout = _mm_packs_epi16(vacc01x0123, vacc22x0123);
+
+ vout = _mm_max_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_min));
+ vout = _mm_min_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_max));
+
+ if (nc >= 4) {
+ *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
+ *((uint32_t*) c1) = (uint32_t) _mm_extract_epi32(vout, 1);
+ *((uint32_t*) c2) = (uint32_t) _mm_extract_epi32(vout, 2);
+
+ c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
+ c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
+ c2 = (int8_t*) ((uintptr_t) c2 + 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);
+
+ nc -= 4;
+ } else {
+ if (nc & 2) {
+ *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
+ c0 += 2;
+ *((uint16_t*) c1) = (uint16_t) _mm_extract_epi16(vout, 2);
+ c1 += 2;
+ *((uint16_t*) c2) = (uint16_t) _mm_extract_epi16(vout, 4);
+ c2 += 2;
+ vout = _mm_srli_epi32(vout, 16);
+ }
+ if (nc & 1) {
+ *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
+ *((int8_t*) c2) = (int8_t) _mm_extract_epi8(vout, 8);
+ }
+
+ nc = 0;
+ }
+ } while (nc != 0);
+}
diff --git a/src/qc8-gemm/gen/3x4c8-minmax-fp32-avx-ld128.c b/src/qc8-gemm/gen/3x4c8-minmax-fp32-avx-ld128.c
new file mode 100644
index 0000000..5744d2a
--- /dev/null
+++ b/src/qc8-gemm/gen/3x4c8-minmax-fp32-avx-ld128.c
@@ -0,0 +1,177 @@
+// Auto-generated file. Do not edit!
+// Template: src/qs8-gemm/MRx4c8-sse.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 <smmintrin.h>
+
+#include <xnnpack/gemm.h>
+#include <xnnpack/math.h>
+
+
+void xnn_qc8_gemm_minmax_fp32_ukernel_3x4c8__avx_ld128(
+ 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_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN XNN_DISABLE_MSAN
+{
+ assert(mr != 0);
+ assert(mr <= 3);
+ 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, 8);
+ 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;
+ }
+
+ do {
+ __m128i vacc0x0 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[0]);
+ __m128i vacc0x1 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[1]);
+ __m128i vacc0x2 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[2]);
+ __m128i vacc0x3 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[3]);
+ __m128i vacc1x0 = vacc0x0;
+ __m128i vacc1x1 = vacc0x1;
+ __m128i vacc1x2 = vacc0x2;
+ __m128i vacc1x3 = vacc0x3;
+ __m128i vacc2x0 = vacc0x0;
+ __m128i vacc2x1 = vacc0x1;
+ __m128i vacc2x2 = vacc0x2;
+ __m128i vacc2x3 = vacc0x3;
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+
+ size_t k = 0;
+ while (k < kc) {
+ const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
+ const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
+ a0 += 8;
+ const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
+ const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
+ a1 += 8;
+ const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
+ const __m128i vxa2 = _mm_cvtepi8_epi16(va2);
+ a2 += 8;
+
+ const __m128i vb01 = _mm_load_si128((const __m128i*) w);
+ const __m128i vsb01 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb01);
+ const __m128i vxb0 = _mm_unpacklo_epi8(vb01, vsb01);
+ const __m128i vxb1 = _mm_unpackhi_epi8(vb01, vsb01);
+
+ vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
+ vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
+ vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0));
+ vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
+ vacc2x0 = _mm_add_epi32(vacc2x0, _mm_madd_epi16(vxa2, vxb0));
+ vacc2x1 = _mm_add_epi32(vacc2x1, _mm_madd_epi16(vxa2, vxb1));
+ const __m128i vb23 = _mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
+ const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
+ const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
+ const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
+
+ vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
+ vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
+ vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2));
+ vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3));
+ vacc2x2 = _mm_add_epi32(vacc2x2, _mm_madd_epi16(vxa2, vxb2));
+ vacc2x3 = _mm_add_epi32(vacc2x3, _mm_madd_epi16(vxa2, vxb3));
+
+ w = (const void*) ((uintptr_t) w + 32 * sizeof(int8_t));
+ k += 8 * sizeof(int8_t);
+ }
+
+ const __m128i vacc0x01 = _mm_hadd_epi32(vacc0x0, vacc0x1);
+ const __m128i vacc0x23 = _mm_hadd_epi32(vacc0x2, vacc0x3);
+ const __m128i vacc1x01 = _mm_hadd_epi32(vacc1x0, vacc1x1);
+ const __m128i vacc1x23 = _mm_hadd_epi32(vacc1x2, vacc1x3);
+ const __m128i vacc2x01 = _mm_hadd_epi32(vacc2x0, vacc2x1);
+ const __m128i vacc2x23 = _mm_hadd_epi32(vacc2x2, vacc2x3);
+
+ __m128i vacc0x0123 = _mm_hadd_epi32(vacc0x01, vacc0x23);
+ __m128i vacc1x0123 = _mm_hadd_epi32(vacc1x01, vacc1x23);
+ __m128i vacc2x0123 = _mm_hadd_epi32(vacc2x01, vacc2x23);
+
+ __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
+ __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
+ __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123);
+
+ const __m128 vscale0123 = _mm_load_ps((const float*) w);
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(float));
+ vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123);
+ vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale0123);
+ vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale0123);
+
+ vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
+ vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
+ vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123);
+
+ const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4.output_zero_point);
+ __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
+ __m128i vacc22x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc2x0123), voutput_zero_point);
+
+
+ __m128i vout = _mm_packs_epi16(vacc01x0123, vacc22x0123);
+
+ vout = _mm_max_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_min));
+ vout = _mm_min_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_max));
+
+ if (nc >= 4) {
+ *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
+ *((uint32_t*) c1) = (uint32_t) _mm_extract_epi32(vout, 1);
+ *((uint32_t*) c2) = (uint32_t) _mm_extract_epi32(vout, 2);
+
+ c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
+ c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
+ c2 = (int8_t*) ((uintptr_t) c2 + 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);
+
+ nc -= 4;
+ } else {
+ if (nc & 2) {
+ *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
+ c0 += 2;
+ *((uint16_t*) c1) = (uint16_t) _mm_extract_epi16(vout, 2);
+ c1 += 2;
+ *((uint16_t*) c2) = (uint16_t) _mm_extract_epi16(vout, 4);
+ c2 += 2;
+ vout = _mm_srli_epi32(vout, 16);
+ }
+ if (nc & 1) {
+ *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
+ *((int8_t*) c2) = (int8_t) _mm_extract_epi8(vout, 8);
+ }
+
+ nc = 0;
+ }
+ } while (nc != 0);
+}
diff --git a/src/qc8-gemm/gen/3x4c8-minmax-fp32-avx-ld64.c b/src/qc8-gemm/gen/3x4c8-minmax-fp32-avx-ld64.c
new file mode 100644
index 0000000..a6246de
--- /dev/null
+++ b/src/qc8-gemm/gen/3x4c8-minmax-fp32-avx-ld64.c
@@ -0,0 +1,179 @@
+// Auto-generated file. Do not edit!
+// Template: src/qs8-gemm/MRx4c8-sse.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 <smmintrin.h>
+
+#include <xnnpack/gemm.h>
+#include <xnnpack/math.h>
+
+
+void xnn_qc8_gemm_minmax_fp32_ukernel_3x4c8__avx_ld64(
+ 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_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN XNN_DISABLE_MSAN
+{
+ assert(mr != 0);
+ assert(mr <= 3);
+ 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, 8);
+ 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;
+ }
+
+ do {
+ __m128i vacc0x0 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[0]);
+ __m128i vacc0x1 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[1]);
+ __m128i vacc0x2 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[2]);
+ __m128i vacc0x3 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[3]);
+ __m128i vacc1x0 = vacc0x0;
+ __m128i vacc1x1 = vacc0x1;
+ __m128i vacc1x2 = vacc0x2;
+ __m128i vacc1x3 = vacc0x3;
+ __m128i vacc2x0 = vacc0x0;
+ __m128i vacc2x1 = vacc0x1;
+ __m128i vacc2x2 = vacc0x2;
+ __m128i vacc2x3 = vacc0x3;
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+
+ size_t k = 0;
+ while (k < kc) {
+ const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
+ const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
+ a0 += 8;
+ const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
+ const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
+ a1 += 8;
+ const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
+ const __m128i vxa2 = _mm_cvtepi8_epi16(va2);
+ a2 += 8;
+
+ const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
+
+ vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
+ vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0));
+ vacc2x0 = _mm_add_epi32(vacc2x0, _mm_madd_epi16(vxa2, vxb0));
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int8_t)));
+ const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
+
+ vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
+ vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
+ vacc2x1 = _mm_add_epi32(vacc2x1, _mm_madd_epi16(vxa2, vxb1));
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
+ const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
+
+ vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
+ vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2));
+ vacc2x2 = _mm_add_epi32(vacc2x2, _mm_madd_epi16(vxa2, vxb2));
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int8_t)));
+ const __m128i vxb3 = _mm_cvtepi8_epi16(vb3);
+
+ vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
+ vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3));
+ vacc2x3 = _mm_add_epi32(vacc2x3, _mm_madd_epi16(vxa2, vxb3));
+
+ w = (const void*) ((uintptr_t) w + 32 * sizeof(int8_t));
+ k += 8 * sizeof(int8_t);
+ }
+
+ const __m128i vacc0x01 = _mm_hadd_epi32(vacc0x0, vacc0x1);
+ const __m128i vacc0x23 = _mm_hadd_epi32(vacc0x2, vacc0x3);
+ const __m128i vacc1x01 = _mm_hadd_epi32(vacc1x0, vacc1x1);
+ const __m128i vacc1x23 = _mm_hadd_epi32(vacc1x2, vacc1x3);
+ const __m128i vacc2x01 = _mm_hadd_epi32(vacc2x0, vacc2x1);
+ const __m128i vacc2x23 = _mm_hadd_epi32(vacc2x2, vacc2x3);
+
+ __m128i vacc0x0123 = _mm_hadd_epi32(vacc0x01, vacc0x23);
+ __m128i vacc1x0123 = _mm_hadd_epi32(vacc1x01, vacc1x23);
+ __m128i vacc2x0123 = _mm_hadd_epi32(vacc2x01, vacc2x23);
+
+ __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
+ __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
+ __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123);
+
+ const __m128 vscale0123 = _mm_load_ps((const float*) w);
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(float));
+ vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123);
+ vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale0123);
+ vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale0123);
+
+ vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
+ vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
+ vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123);
+
+ const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4.output_zero_point);
+ __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
+ __m128i vacc22x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc2x0123), voutput_zero_point);
+
+
+ __m128i vout = _mm_packs_epi16(vacc01x0123, vacc22x0123);
+
+ vout = _mm_max_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_min));
+ vout = _mm_min_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_max));
+
+ if (nc >= 4) {
+ *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
+ *((uint32_t*) c1) = (uint32_t) _mm_extract_epi32(vout, 1);
+ *((uint32_t*) c2) = (uint32_t) _mm_extract_epi32(vout, 2);
+
+ c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
+ c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
+ c2 = (int8_t*) ((uintptr_t) c2 + 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);
+
+ nc -= 4;
+ } else {
+ if (nc & 2) {
+ *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
+ c0 += 2;
+ *((uint16_t*) c1) = (uint16_t) _mm_extract_epi16(vout, 2);
+ c1 += 2;
+ *((uint16_t*) c2) = (uint16_t) _mm_extract_epi16(vout, 4);
+ c2 += 2;
+ vout = _mm_srli_epi32(vout, 16);
+ }
+ if (nc & 1) {
+ *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
+ *((int8_t*) c2) = (int8_t) _mm_extract_epi8(vout, 8);
+ }
+
+ nc = 0;
+ }
+ } while (nc != 0);
+}
diff --git a/src/qc8-gemm/gen/3x4c8-minmax-fp32-sse2-ld128.c b/src/qc8-gemm/gen/3x4c8-minmax-fp32-sse2-ld128.c
new file mode 100644
index 0000000..8275917
--- /dev/null
+++ b/src/qc8-gemm/gen/3x4c8-minmax-fp32-sse2-ld128.c
@@ -0,0 +1,181 @@
+// Auto-generated file. Do not edit!
+// Template: src/qs8-gemm/MRx4c8-sse.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 <emmintrin.h>
+
+#include <xnnpack/gemm.h>
+#include <xnnpack/math.h>
+
+
+void xnn_qc8_gemm_minmax_fp32_ukernel_3x4c8__sse2_ld128(
+ 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_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN XNN_DISABLE_MSAN
+{
+ assert(mr != 0);
+ assert(mr <= 3);
+ 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, 8);
+ 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;
+ }
+
+ do {
+ __m128i vacc0x0 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[0]);
+ __m128i vacc0x1 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[1]);
+ __m128i vacc0x2 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[2]);
+ __m128i vacc0x3 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[3]);
+ __m128i vacc1x0 = vacc0x0;
+ __m128i vacc1x1 = vacc0x1;
+ __m128i vacc1x2 = vacc0x2;
+ __m128i vacc1x3 = vacc0x3;
+ __m128i vacc2x0 = vacc0x0;
+ __m128i vacc2x1 = vacc0x1;
+ __m128i vacc2x2 = vacc0x2;
+ __m128i vacc2x3 = vacc0x3;
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+
+ size_t k = 0;
+ while (k < kc) {
+ const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
+ const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ a0 += 8;
+ const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
+ const __m128i vxa1 = _mm_unpacklo_epi8(va1, _mm_cmpgt_epi8(_mm_setzero_si128(), va1));
+ a1 += 8;
+ const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
+ const __m128i vxa2 = _mm_unpacklo_epi8(va2, _mm_cmpgt_epi8(_mm_setzero_si128(), va2));
+ a2 += 8;
+
+ const __m128i vb01 = _mm_load_si128((const __m128i*) w);
+ const __m128i vsb01 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb01);
+ const __m128i vxb0 = _mm_unpacklo_epi8(vb01, vsb01);
+ const __m128i vxb1 = _mm_unpackhi_epi8(vb01, vsb01);
+
+ vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
+ vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
+ vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0));
+ vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
+ vacc2x0 = _mm_add_epi32(vacc2x0, _mm_madd_epi16(vxa2, vxb0));
+ vacc2x1 = _mm_add_epi32(vacc2x1, _mm_madd_epi16(vxa2, vxb1));
+ const __m128i vb23 = _mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
+ const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
+ const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
+ const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
+
+ vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
+ vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
+ vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2));
+ vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3));
+ vacc2x2 = _mm_add_epi32(vacc2x2, _mm_madd_epi16(vxa2, vxb2));
+ vacc2x3 = _mm_add_epi32(vacc2x3, _mm_madd_epi16(vxa2, vxb3));
+
+ w = (const void*) ((uintptr_t) w + 32 * sizeof(int8_t));
+ k += 8 * sizeof(int8_t);
+ }
+
+ const __m128i vacc0x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x0, vacc0x2), _mm_unpackhi_epi32(vacc0x0, vacc0x2));
+ const __m128i vacc0x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x1, vacc0x3), _mm_unpackhi_epi32(vacc0x1, vacc0x3));
+ const __m128i vacc1x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x0, vacc1x2), _mm_unpackhi_epi32(vacc1x0, vacc1x2));
+ const __m128i vacc1x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x1, vacc1x3), _mm_unpackhi_epi32(vacc1x1, vacc1x3));
+ const __m128i vacc2x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc2x0, vacc2x2), _mm_unpackhi_epi32(vacc2x0, vacc2x2));
+ const __m128i vacc2x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc2x1, vacc2x3), _mm_unpackhi_epi32(vacc2x1, vacc2x3));
+
+ __m128i vacc0x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x02, vacc0x13), _mm_unpackhi_epi32(vacc0x02, vacc0x13));
+ __m128i vacc1x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x02, vacc1x13), _mm_unpackhi_epi32(vacc1x02, vacc1x13));
+ __m128i vacc2x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc2x02, vacc2x13), _mm_unpackhi_epi32(vacc2x02, vacc2x13));
+
+ __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
+ __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
+ __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123);
+
+ const __m128 vscale0123 = _mm_load_ps((const float*) w);
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(float));
+ vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123);
+ vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale0123);
+ vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale0123);
+
+ vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
+ vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
+ vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123);
+
+ const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
+ __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
+ __m128i vacc22x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc2x0123), voutput_zero_point);
+
+ const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
+ const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max);
+ vacc01x0123 = _mm_min_epi16(_mm_max_epi16(vacc01x0123, voutput_min), voutput_max);
+ vacc22x0123 = _mm_min_epi16(_mm_max_epi16(vacc22x0123, voutput_min), voutput_max);
+
+ __m128i vout = _mm_packs_epi16(vacc01x0123, vacc22x0123);
+
+
+ if (nc >= 4) {
+ *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
+ vout = _mm_srli_si128(vout, 4);
+ *((uint32_t*) c1) = (uint32_t) _mm_cvtsi128_si32(vout);
+ vout = _mm_srli_si128(vout, 4);
+ *((uint32_t*) c2) = (uint32_t) _mm_cvtsi128_si32(vout);
+
+ c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
+ c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
+ c2 = (int8_t*) ((uintptr_t) c2 + 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);
+
+ nc -= 4;
+ } else {
+ if (nc & 2) {
+ *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
+ c0 += 2;
+ *((uint16_t*) c1) = (uint16_t) _mm_extract_epi16(vout, 2);
+ c1 += 2;
+ *((uint16_t*) c2) = (uint16_t) _mm_extract_epi16(vout, 4);
+ c2 += 2;
+ vout = _mm_srli_epi32(vout, 16);
+ }
+ if (nc & 1) {
+ *((int8_t*) c0) = (int8_t) _mm_cvtsi128_si32(vout);
+ *((int8_t*) c1) = (int8_t) _mm_extract_epi16(vout, 2);
+ *((int8_t*) c2) = (int8_t) _mm_extract_epi16(vout, 4);
+ }
+
+ nc = 0;
+ }
+ } while (nc != 0);
+}
diff --git a/src/qc8-gemm/gen/3x4c8-minmax-fp32-sse2-ld64.c b/src/qc8-gemm/gen/3x4c8-minmax-fp32-sse2-ld64.c
new file mode 100644
index 0000000..80d112b
--- /dev/null
+++ b/src/qc8-gemm/gen/3x4c8-minmax-fp32-sse2-ld64.c
@@ -0,0 +1,183 @@
+// Auto-generated file. Do not edit!
+// Template: src/qs8-gemm/MRx4c8-sse.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 <emmintrin.h>
+
+#include <xnnpack/gemm.h>
+#include <xnnpack/math.h>
+
+
+void xnn_qc8_gemm_minmax_fp32_ukernel_3x4c8__sse2_ld64(
+ 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_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN XNN_DISABLE_MSAN
+{
+ assert(mr != 0);
+ assert(mr <= 3);
+ 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, 8);
+ 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;
+ }
+
+ do {
+ __m128i vacc0x0 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[0]);
+ __m128i vacc0x1 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[1]);
+ __m128i vacc0x2 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[2]);
+ __m128i vacc0x3 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[3]);
+ __m128i vacc1x0 = vacc0x0;
+ __m128i vacc1x1 = vacc0x1;
+ __m128i vacc1x2 = vacc0x2;
+ __m128i vacc1x3 = vacc0x3;
+ __m128i vacc2x0 = vacc0x0;
+ __m128i vacc2x1 = vacc0x1;
+ __m128i vacc2x2 = vacc0x2;
+ __m128i vacc2x3 = vacc0x3;
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+
+ size_t k = 0;
+ while (k < kc) {
+ const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
+ const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ a0 += 8;
+ const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
+ const __m128i vxa1 = _mm_unpacklo_epi8(va1, _mm_cmpgt_epi8(_mm_setzero_si128(), va1));
+ a1 += 8;
+ const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
+ const __m128i vxa2 = _mm_unpacklo_epi8(va2, _mm_cmpgt_epi8(_mm_setzero_si128(), va2));
+ a2 += 8;
+
+ const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+
+ vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
+ vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0));
+ vacc2x0 = _mm_add_epi32(vacc2x0, _mm_madd_epi16(vxa2, vxb0));
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int8_t)));
+ const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+
+ vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
+ vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
+ vacc2x1 = _mm_add_epi32(vacc2x1, _mm_madd_epi16(vxa2, vxb1));
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
+ const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+
+ vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
+ vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2));
+ vacc2x2 = _mm_add_epi32(vacc2x2, _mm_madd_epi16(vxa2, vxb2));
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int8_t)));
+ const __m128i vxb3 = _mm_unpacklo_epi8(vb3, _mm_cmpgt_epi8(_mm_setzero_si128(), vb3));
+
+ vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
+ vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3));
+ vacc2x3 = _mm_add_epi32(vacc2x3, _mm_madd_epi16(vxa2, vxb3));
+
+ w = (const void*) ((uintptr_t) w + 32 * sizeof(int8_t));
+ k += 8 * sizeof(int8_t);
+ }
+
+ const __m128i vacc0x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x0, vacc0x2), _mm_unpackhi_epi32(vacc0x0, vacc0x2));
+ const __m128i vacc0x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x1, vacc0x3), _mm_unpackhi_epi32(vacc0x1, vacc0x3));
+ const __m128i vacc1x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x0, vacc1x2), _mm_unpackhi_epi32(vacc1x0, vacc1x2));
+ const __m128i vacc1x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x1, vacc1x3), _mm_unpackhi_epi32(vacc1x1, vacc1x3));
+ const __m128i vacc2x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc2x0, vacc2x2), _mm_unpackhi_epi32(vacc2x0, vacc2x2));
+ const __m128i vacc2x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc2x1, vacc2x3), _mm_unpackhi_epi32(vacc2x1, vacc2x3));
+
+ __m128i vacc0x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x02, vacc0x13), _mm_unpackhi_epi32(vacc0x02, vacc0x13));
+ __m128i vacc1x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc1x02, vacc1x13), _mm_unpackhi_epi32(vacc1x02, vacc1x13));
+ __m128i vacc2x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc2x02, vacc2x13), _mm_unpackhi_epi32(vacc2x02, vacc2x13));
+
+ __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
+ __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
+ __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123);
+
+ const __m128 vscale0123 = _mm_load_ps((const float*) w);
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(float));
+ vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123);
+ vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale0123);
+ vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale0123);
+
+ vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
+ vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
+ vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123);
+
+ const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
+ __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
+ __m128i vacc22x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc2x0123), voutput_zero_point);
+
+ const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
+ const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max);
+ vacc01x0123 = _mm_min_epi16(_mm_max_epi16(vacc01x0123, voutput_min), voutput_max);
+ vacc22x0123 = _mm_min_epi16(_mm_max_epi16(vacc22x0123, voutput_min), voutput_max);
+
+ __m128i vout = _mm_packs_epi16(vacc01x0123, vacc22x0123);
+
+
+ if (nc >= 4) {
+ *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
+ vout = _mm_srli_si128(vout, 4);
+ *((uint32_t*) c1) = (uint32_t) _mm_cvtsi128_si32(vout);
+ vout = _mm_srli_si128(vout, 4);
+ *((uint32_t*) c2) = (uint32_t) _mm_cvtsi128_si32(vout);
+
+ c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
+ c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
+ c2 = (int8_t*) ((uintptr_t) c2 + 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);
+
+ nc -= 4;
+ } else {
+ if (nc & 2) {
+ *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
+ c0 += 2;
+ *((uint16_t*) c1) = (uint16_t) _mm_extract_epi16(vout, 2);
+ c1 += 2;
+ *((uint16_t*) c2) = (uint16_t) _mm_extract_epi16(vout, 4);
+ c2 += 2;
+ vout = _mm_srli_epi32(vout, 16);
+ }
+ if (nc & 1) {
+ *((int8_t*) c0) = (int8_t) _mm_cvtsi128_si32(vout);
+ *((int8_t*) c1) = (int8_t) _mm_extract_epi16(vout, 2);
+ *((int8_t*) c2) = (int8_t) _mm_extract_epi16(vout, 4);
+ }
+
+ nc = 0;
+ }
+ } while (nc != 0);
+}
diff --git a/src/qc8-gemm/gen/3x4c8-minmax-fp32-sse41-ld128.c b/src/qc8-gemm/gen/3x4c8-minmax-fp32-sse41-ld128.c
new file mode 100644
index 0000000..b1209ec
--- /dev/null
+++ b/src/qc8-gemm/gen/3x4c8-minmax-fp32-sse41-ld128.c
@@ -0,0 +1,177 @@
+// Auto-generated file. Do not edit!
+// Template: src/qs8-gemm/MRx4c8-sse.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 <smmintrin.h>
+
+#include <xnnpack/gemm.h>
+#include <xnnpack/math.h>
+
+
+void xnn_qc8_gemm_minmax_fp32_ukernel_3x4c8__sse41_ld128(
+ 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_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN XNN_DISABLE_MSAN
+{
+ assert(mr != 0);
+ assert(mr <= 3);
+ 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, 8);
+ 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;
+ }
+
+ do {
+ __m128i vacc0x0 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[0]);
+ __m128i vacc0x1 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[1]);
+ __m128i vacc0x2 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[2]);
+ __m128i vacc0x3 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[3]);
+ __m128i vacc1x0 = vacc0x0;
+ __m128i vacc1x1 = vacc0x1;
+ __m128i vacc1x2 = vacc0x2;
+ __m128i vacc1x3 = vacc0x3;
+ __m128i vacc2x0 = vacc0x0;
+ __m128i vacc2x1 = vacc0x1;
+ __m128i vacc2x2 = vacc0x2;
+ __m128i vacc2x3 = vacc0x3;
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+
+ size_t k = 0;
+ while (k < kc) {
+ const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
+ const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
+ a0 += 8;
+ const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
+ const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
+ a1 += 8;
+ const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
+ const __m128i vxa2 = _mm_cvtepi8_epi16(va2);
+ a2 += 8;
+
+ const __m128i vb01 = _mm_load_si128((const __m128i*) w);
+ const __m128i vsb01 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb01);
+ const __m128i vxb0 = _mm_unpacklo_epi8(vb01, vsb01);
+ const __m128i vxb1 = _mm_unpackhi_epi8(vb01, vsb01);
+
+ vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
+ vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
+ vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0));
+ vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
+ vacc2x0 = _mm_add_epi32(vacc2x0, _mm_madd_epi16(vxa2, vxb0));
+ vacc2x1 = _mm_add_epi32(vacc2x1, _mm_madd_epi16(vxa2, vxb1));
+ const __m128i vb23 = _mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
+ const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
+ const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
+ const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
+
+ vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
+ vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
+ vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2));
+ vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3));
+ vacc2x2 = _mm_add_epi32(vacc2x2, _mm_madd_epi16(vxa2, vxb2));
+ vacc2x3 = _mm_add_epi32(vacc2x3, _mm_madd_epi16(vxa2, vxb3));
+
+ w = (const void*) ((uintptr_t) w + 32 * sizeof(int8_t));
+ k += 8 * sizeof(int8_t);
+ }
+
+ const __m128i vacc0x01 = _mm_hadd_epi32(vacc0x0, vacc0x1);
+ const __m128i vacc0x23 = _mm_hadd_epi32(vacc0x2, vacc0x3);
+ const __m128i vacc1x01 = _mm_hadd_epi32(vacc1x0, vacc1x1);
+ const __m128i vacc1x23 = _mm_hadd_epi32(vacc1x2, vacc1x3);
+ const __m128i vacc2x01 = _mm_hadd_epi32(vacc2x0, vacc2x1);
+ const __m128i vacc2x23 = _mm_hadd_epi32(vacc2x2, vacc2x3);
+
+ __m128i vacc0x0123 = _mm_hadd_epi32(vacc0x01, vacc0x23);
+ __m128i vacc1x0123 = _mm_hadd_epi32(vacc1x01, vacc1x23);
+ __m128i vacc2x0123 = _mm_hadd_epi32(vacc2x01, vacc2x23);
+
+ __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
+ __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
+ __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123);
+
+ const __m128 vscale0123 = _mm_load_ps((const float*) w);
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(float));
+ vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123);
+ vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale0123);
+ vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale0123);
+
+ vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
+ vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
+ vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123);
+
+ const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4.output_zero_point);
+ __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
+ __m128i vacc22x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc2x0123), voutput_zero_point);
+
+
+ __m128i vout = _mm_packs_epi16(vacc01x0123, vacc22x0123);
+
+ vout = _mm_max_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_min));
+ vout = _mm_min_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_max));
+
+ if (nc >= 4) {
+ *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
+ *((uint32_t*) c1) = (uint32_t) _mm_extract_epi32(vout, 1);
+ *((uint32_t*) c2) = (uint32_t) _mm_extract_epi32(vout, 2);
+
+ c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
+ c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
+ c2 = (int8_t*) ((uintptr_t) c2 + 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);
+
+ nc -= 4;
+ } else {
+ if (nc & 2) {
+ *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
+ c0 += 2;
+ *((uint16_t*) c1) = (uint16_t) _mm_extract_epi16(vout, 2);
+ c1 += 2;
+ *((uint16_t*) c2) = (uint16_t) _mm_extract_epi16(vout, 4);
+ c2 += 2;
+ vout = _mm_srli_epi32(vout, 16);
+ }
+ if (nc & 1) {
+ *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
+ *((int8_t*) c2) = (int8_t) _mm_extract_epi8(vout, 8);
+ }
+
+ nc = 0;
+ }
+ } while (nc != 0);
+}
diff --git a/src/qc8-gemm/gen/3x4c8-minmax-fp32-sse41-ld64.c b/src/qc8-gemm/gen/3x4c8-minmax-fp32-sse41-ld64.c
new file mode 100644
index 0000000..6d9a6bd
--- /dev/null
+++ b/src/qc8-gemm/gen/3x4c8-minmax-fp32-sse41-ld64.c
@@ -0,0 +1,179 @@
+// Auto-generated file. Do not edit!
+// Template: src/qs8-gemm/MRx4c8-sse.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 <smmintrin.h>
+
+#include <xnnpack/gemm.h>
+#include <xnnpack/math.h>
+
+
+void xnn_qc8_gemm_minmax_fp32_ukernel_3x4c8__sse41_ld64(
+ 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_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN XNN_DISABLE_MSAN
+{
+ assert(mr != 0);
+ assert(mr <= 3);
+ 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, 8);
+ 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;
+ }
+
+ do {
+ __m128i vacc0x0 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[0]);
+ __m128i vacc0x1 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[1]);
+ __m128i vacc0x2 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[2]);
+ __m128i vacc0x3 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[3]);
+ __m128i vacc1x0 = vacc0x0;
+ __m128i vacc1x1 = vacc0x1;
+ __m128i vacc1x2 = vacc0x2;
+ __m128i vacc1x3 = vacc0x3;
+ __m128i vacc2x0 = vacc0x0;
+ __m128i vacc2x1 = vacc0x1;
+ __m128i vacc2x2 = vacc0x2;
+ __m128i vacc2x3 = vacc0x3;
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+
+ size_t k = 0;
+ while (k < kc) {
+ const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
+ const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
+ a0 += 8;
+ const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
+ const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
+ a1 += 8;
+ const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
+ const __m128i vxa2 = _mm_cvtepi8_epi16(va2);
+ a2 += 8;
+
+ const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
+
+ vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0));
+ vacc1x0 = _mm_add_epi32(vacc1x0, _mm_madd_epi16(vxa1, vxb0));
+ vacc2x0 = _mm_add_epi32(vacc2x0, _mm_madd_epi16(vxa2, vxb0));
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int8_t)));
+ const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
+
+ vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1));
+ vacc1x1 = _mm_add_epi32(vacc1x1, _mm_madd_epi16(vxa1, vxb1));
+ vacc2x1 = _mm_add_epi32(vacc2x1, _mm_madd_epi16(vxa2, vxb1));
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
+ const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
+
+ vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2));
+ vacc1x2 = _mm_add_epi32(vacc1x2, _mm_madd_epi16(vxa1, vxb2));
+ vacc2x2 = _mm_add_epi32(vacc2x2, _mm_madd_epi16(vxa2, vxb2));
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int8_t)));
+ const __m128i vxb3 = _mm_cvtepi8_epi16(vb3);
+
+ vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3));
+ vacc1x3 = _mm_add_epi32(vacc1x3, _mm_madd_epi16(vxa1, vxb3));
+ vacc2x3 = _mm_add_epi32(vacc2x3, _mm_madd_epi16(vxa2, vxb3));
+
+ w = (const void*) ((uintptr_t) w + 32 * sizeof(int8_t));
+ k += 8 * sizeof(int8_t);
+ }
+
+ const __m128i vacc0x01 = _mm_hadd_epi32(vacc0x0, vacc0x1);
+ const __m128i vacc0x23 = _mm_hadd_epi32(vacc0x2, vacc0x3);
+ const __m128i vacc1x01 = _mm_hadd_epi32(vacc1x0, vacc1x1);
+ const __m128i vacc1x23 = _mm_hadd_epi32(vacc1x2, vacc1x3);
+ const __m128i vacc2x01 = _mm_hadd_epi32(vacc2x0, vacc2x1);
+ const __m128i vacc2x23 = _mm_hadd_epi32(vacc2x2, vacc2x3);
+
+ __m128i vacc0x0123 = _mm_hadd_epi32(vacc0x01, vacc0x23);
+ __m128i vacc1x0123 = _mm_hadd_epi32(vacc1x01, vacc1x23);
+ __m128i vacc2x0123 = _mm_hadd_epi32(vacc2x01, vacc2x23);
+
+ __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
+ __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
+ __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123);
+
+ const __m128 vscale0123 = _mm_load_ps((const float*) w);
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(float));
+ vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123);
+ vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale0123);
+ vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale0123);
+
+ vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
+ vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
+ vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123);
+
+ const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4.output_zero_point);
+ __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
+ __m128i vacc22x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc2x0123), voutput_zero_point);
+
+
+ __m128i vout = _mm_packs_epi16(vacc01x0123, vacc22x0123);
+
+ vout = _mm_max_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_min));
+ vout = _mm_min_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_max));
+
+ if (nc >= 4) {
+ *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
+ *((uint32_t*) c1) = (uint32_t) _mm_extract_epi32(vout, 1);
+ *((uint32_t*) c2) = (uint32_t) _mm_extract_epi32(vout, 2);
+
+ c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
+ c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
+ c2 = (int8_t*) ((uintptr_t) c2 + 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);
+
+ nc -= 4;
+ } else {
+ if (nc & 2) {
+ *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
+ c0 += 2;
+ *((uint16_t*) c1) = (uint16_t) _mm_extract_epi16(vout, 2);
+ c1 += 2;
+ *((uint16_t*) c2) = (uint16_t) _mm_extract_epi16(vout, 4);
+ c2 += 2;
+ vout = _mm_srli_epi32(vout, 16);
+ }
+ if (nc & 1) {
+ *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
+ *((int8_t*) c2) = (int8_t) _mm_extract_epi8(vout, 8);
+ }
+
+ nc = 0;
+ }
+ } while (nc != 0);
+}
diff --git a/src/qc8-gemm/gen/3x4c8-minmax-fp32-xop-ld128.c b/src/qc8-gemm/gen/3x4c8-minmax-fp32-xop-ld128.c
new file mode 100644
index 0000000..904263f
--- /dev/null
+++ b/src/qc8-gemm/gen/3x4c8-minmax-fp32-xop-ld128.c
@@ -0,0 +1,182 @@
+// Auto-generated file. Do not edit!
+// Template: src/qs8-gemm/MRx4c8-sse.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>
+
+#if defined(__GNUC__) || defined(__clang__)
+ #include <x86intrin.h>
+#else
+ #include <immintrin.h>
+ #include <ammintrin.h>
+#endif
+
+#include <xnnpack/gemm.h>
+#include <xnnpack/math.h>
+
+
+void xnn_qc8_gemm_minmax_fp32_ukernel_3x4c8__xop_ld128(
+ 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_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN XNN_DISABLE_MSAN
+{
+ assert(mr != 0);
+ assert(mr <= 3);
+ 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, 8);
+ 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;
+ }
+
+ do {
+ __m128i vacc0x0 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[0]);
+ __m128i vacc0x1 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[1]);
+ __m128i vacc0x2 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[2]);
+ __m128i vacc0x3 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[3]);
+ __m128i vacc1x0 = vacc0x0;
+ __m128i vacc1x1 = vacc0x1;
+ __m128i vacc1x2 = vacc0x2;
+ __m128i vacc1x3 = vacc0x3;
+ __m128i vacc2x0 = vacc0x0;
+ __m128i vacc2x1 = vacc0x1;
+ __m128i vacc2x2 = vacc0x2;
+ __m128i vacc2x3 = vacc0x3;
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+
+ size_t k = 0;
+ while (k < kc) {
+ const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
+ const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
+ a0 += 8;
+ const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
+ const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
+ a1 += 8;
+ const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
+ const __m128i vxa2 = _mm_cvtepi8_epi16(va2);
+ a2 += 8;
+
+ const __m128i vb01 = _mm_load_si128((const __m128i*) w);
+ const __m128i vsb01 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb01);
+ const __m128i vxb0 = _mm_unpacklo_epi8(vb01, vsb01);
+ const __m128i vxb1 = _mm_unpackhi_epi8(vb01, vsb01);
+
+ vacc0x0 = _mm_maddd_epi16(vxa0, vxb0, vacc0x0);
+ vacc0x1 = _mm_maddd_epi16(vxa0, vxb1, vacc0x1);
+ vacc1x0 = _mm_maddd_epi16(vxa1, vxb0, vacc1x0);
+ vacc1x1 = _mm_maddd_epi16(vxa1, vxb1, vacc1x1);
+ vacc2x0 = _mm_maddd_epi16(vxa2, vxb0, vacc2x0);
+ vacc2x1 = _mm_maddd_epi16(vxa2, vxb1, vacc2x1);
+ const __m128i vb23 = _mm_load_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
+ const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
+ const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
+ const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
+
+ vacc0x2 = _mm_maddd_epi16(vxa0, vxb2, vacc0x2);
+ vacc0x3 = _mm_maddd_epi16(vxa0, vxb3, vacc0x3);
+ vacc1x2 = _mm_maddd_epi16(vxa1, vxb2, vacc1x2);
+ vacc1x3 = _mm_maddd_epi16(vxa1, vxb3, vacc1x3);
+ vacc2x2 = _mm_maddd_epi16(vxa2, vxb2, vacc2x2);
+ vacc2x3 = _mm_maddd_epi16(vxa2, vxb3, vacc2x3);
+
+ w = (const void*) ((uintptr_t) w + 32 * sizeof(int8_t));
+ k += 8 * sizeof(int8_t);
+ }
+
+ const __m128i vacc0x01 = _mm_hadd_epi32(vacc0x0, vacc0x1);
+ const __m128i vacc0x23 = _mm_hadd_epi32(vacc0x2, vacc0x3);
+ const __m128i vacc1x01 = _mm_hadd_epi32(vacc1x0, vacc1x1);
+ const __m128i vacc1x23 = _mm_hadd_epi32(vacc1x2, vacc1x3);
+ const __m128i vacc2x01 = _mm_hadd_epi32(vacc2x0, vacc2x1);
+ const __m128i vacc2x23 = _mm_hadd_epi32(vacc2x2, vacc2x3);
+
+ __m128i vacc0x0123 = _mm_hadd_epi32(vacc0x01, vacc0x23);
+ __m128i vacc1x0123 = _mm_hadd_epi32(vacc1x01, vacc1x23);
+ __m128i vacc2x0123 = _mm_hadd_epi32(vacc2x01, vacc2x23);
+
+ __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
+ __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
+ __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123);
+
+ const __m128 vscale0123 = _mm_load_ps((const float*) w);
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(float));
+ vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123);
+ vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale0123);
+ vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale0123);
+
+ vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
+ vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
+ vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123);
+
+ const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4.output_zero_point);
+ __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
+ __m128i vacc22x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc2x0123), voutput_zero_point);
+
+
+ __m128i vout = _mm_packs_epi16(vacc01x0123, vacc22x0123);
+
+ vout = _mm_max_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_min));
+ vout = _mm_min_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_max));
+
+ if (nc >= 4) {
+ *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
+ *((uint32_t*) c1) = (uint32_t) _mm_extract_epi32(vout, 1);
+ *((uint32_t*) c2) = (uint32_t) _mm_extract_epi32(vout, 2);
+
+ c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
+ c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
+ c2 = (int8_t*) ((uintptr_t) c2 + 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);
+
+ nc -= 4;
+ } else {
+ if (nc & 2) {
+ *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
+ c0 += 2;
+ *((uint16_t*) c1) = (uint16_t) _mm_extract_epi16(vout, 2);
+ c1 += 2;
+ *((uint16_t*) c2) = (uint16_t) _mm_extract_epi16(vout, 4);
+ c2 += 2;
+ vout = _mm_srli_epi32(vout, 16);
+ }
+ if (nc & 1) {
+ *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
+ *((int8_t*) c2) = (int8_t) _mm_extract_epi8(vout, 8);
+ }
+
+ nc = 0;
+ }
+ } while (nc != 0);
+}
diff --git a/src/qc8-gemm/gen/3x4c8-minmax-fp32-xop-ld64.c b/src/qc8-gemm/gen/3x4c8-minmax-fp32-xop-ld64.c
new file mode 100644
index 0000000..48bec57
--- /dev/null
+++ b/src/qc8-gemm/gen/3x4c8-minmax-fp32-xop-ld64.c
@@ -0,0 +1,184 @@
+// Auto-generated file. Do not edit!
+// Template: src/qs8-gemm/MRx4c8-sse.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>
+
+#if defined(__GNUC__) || defined(__clang__)
+ #include <x86intrin.h>
+#else
+ #include <immintrin.h>
+ #include <ammintrin.h>
+#endif
+
+#include <xnnpack/gemm.h>
+#include <xnnpack/math.h>
+
+
+void xnn_qc8_gemm_minmax_fp32_ukernel_3x4c8__xop_ld64(
+ 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_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN XNN_DISABLE_MSAN
+{
+ assert(mr != 0);
+ assert(mr <= 3);
+ 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, 8);
+ 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;
+ }
+
+ do {
+ __m128i vacc0x0 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[0]);
+ __m128i vacc0x1 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[1]);
+ __m128i vacc0x2 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[2]);
+ __m128i vacc0x3 = _mm_cvtsi32_si128((int) ((const int32_t*) w)[3]);
+ __m128i vacc1x0 = vacc0x0;
+ __m128i vacc1x1 = vacc0x1;
+ __m128i vacc1x2 = vacc0x2;
+ __m128i vacc1x3 = vacc0x3;
+ __m128i vacc2x0 = vacc0x0;
+ __m128i vacc2x1 = vacc0x1;
+ __m128i vacc2x2 = vacc0x2;
+ __m128i vacc2x3 = vacc0x3;
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+
+ size_t k = 0;
+ while (k < kc) {
+ const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
+ const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
+ a0 += 8;
+ const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
+ const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
+ a1 += 8;
+ const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
+ const __m128i vxa2 = _mm_cvtepi8_epi16(va2);
+ a2 += 8;
+
+ const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
+
+ vacc0x0 = _mm_maddd_epi16(vxa0, vxb0, vacc0x0);
+ vacc1x0 = _mm_maddd_epi16(vxa1, vxb0, vacc1x0);
+ vacc2x0 = _mm_maddd_epi16(vxa2, vxb0, vacc2x0);
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int8_t)));
+ const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
+
+ vacc0x1 = _mm_maddd_epi16(vxa0, vxb1, vacc0x1);
+ vacc1x1 = _mm_maddd_epi16(vxa1, vxb1, vacc1x1);
+ vacc2x1 = _mm_maddd_epi16(vxa2, vxb1, vacc2x1);
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
+ const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
+
+ vacc0x2 = _mm_maddd_epi16(vxa0, vxb2, vacc0x2);
+ vacc1x2 = _mm_maddd_epi16(vxa1, vxb2, vacc1x2);
+ vacc2x2 = _mm_maddd_epi16(vxa2, vxb2, vacc2x2);
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int8_t)));
+ const __m128i vxb3 = _mm_cvtepi8_epi16(vb3);
+
+ vacc0x3 = _mm_maddd_epi16(vxa0, vxb3, vacc0x3);
+ vacc1x3 = _mm_maddd_epi16(vxa1, vxb3, vacc1x3);
+ vacc2x3 = _mm_maddd_epi16(vxa2, vxb3, vacc2x3);
+
+ w = (const void*) ((uintptr_t) w + 32 * sizeof(int8_t));
+ k += 8 * sizeof(int8_t);
+ }
+
+ const __m128i vacc0x01 = _mm_hadd_epi32(vacc0x0, vacc0x1);
+ const __m128i vacc0x23 = _mm_hadd_epi32(vacc0x2, vacc0x3);
+ const __m128i vacc1x01 = _mm_hadd_epi32(vacc1x0, vacc1x1);
+ const __m128i vacc1x23 = _mm_hadd_epi32(vacc1x2, vacc1x3);
+ const __m128i vacc2x01 = _mm_hadd_epi32(vacc2x0, vacc2x1);
+ const __m128i vacc2x23 = _mm_hadd_epi32(vacc2x2, vacc2x3);
+
+ __m128i vacc0x0123 = _mm_hadd_epi32(vacc0x01, vacc0x23);
+ __m128i vacc1x0123 = _mm_hadd_epi32(vacc1x01, vacc1x23);
+ __m128i vacc2x0123 = _mm_hadd_epi32(vacc2x01, vacc2x23);
+
+ __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
+ __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
+ __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123);
+
+ const __m128 vscale0123 = _mm_load_ps((const float*) w);
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(float));
+ vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123);
+ vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale0123);
+ vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale0123);
+
+ vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
+ vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
+ vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123);
+
+ const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4.output_zero_point);
+ __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
+ __m128i vacc22x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc2x0123), voutput_zero_point);
+
+
+ __m128i vout = _mm_packs_epi16(vacc01x0123, vacc22x0123);
+
+ vout = _mm_max_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_min));
+ vout = _mm_min_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_max));
+
+ if (nc >= 4) {
+ *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
+ *((uint32_t*) c1) = (uint32_t) _mm_extract_epi32(vout, 1);
+ *((uint32_t*) c2) = (uint32_t) _mm_extract_epi32(vout, 2);
+
+ c0 = (int8_t*) ((uintptr_t) c0 + cn_stride);
+ c1 = (int8_t*) ((uintptr_t) c1 + cn_stride);
+ c2 = (int8_t*) ((uintptr_t) c2 + 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);
+
+ nc -= 4;
+ } else {
+ if (nc & 2) {
+ *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
+ c0 += 2;
+ *((uint16_t*) c1) = (uint16_t) _mm_extract_epi16(vout, 2);
+ c1 += 2;
+ *((uint16_t*) c2) = (uint16_t) _mm_extract_epi16(vout, 4);
+ c2 += 2;
+ vout = _mm_srli_epi32(vout, 16);
+ }
+ if (nc & 1) {
+ *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
+ *((int8_t*) c2) = (int8_t) _mm_extract_epi8(vout, 8);
+ }
+
+ nc = 0;
+ }
+ } while (nc != 0);
+}
diff --git a/src/qc8-gemm/gen/4x4c2-minmax-fp32-avx-ld128.c b/src/qc8-gemm/gen/4x4c2-minmax-fp32-avx-ld128.c
new file mode 100644
index 0000000..e18d5d7
--- /dev/null
+++ b/src/qc8-gemm/gen/4x4c2-minmax-fp32-avx-ld128.c
@@ -0,0 +1,256 @@
+// Auto-generated file. Do not edit!
+// Template: src/qs8-gemm/MRx4c2-sse.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 <smmintrin.h>
+
+#include <xnnpack/gemm.h>
+#include <xnnpack/math.h>
+
+
+
+void xnn_qc8_gemm_minmax_fp32_ukernel_4x4c2__avx_ld128(
+ 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_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN XNN_DISABLE_MSAN
+{
+ assert(mr != 0);
+ assert(mr <= 4);
+ 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, 2);
+ 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;
+ }
+
+ do {
+ __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
+ __m128i vacc1x0123 = vacc0x0123;
+ __m128i vacc2x0123 = vacc0x0123;
+ __m128i vacc3x0123 = vacc0x0123;
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+
+ size_t k = kc;
+ while (k >= 8 * sizeof(int8_t)) {
+ const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
+ const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
+ a0 += 8;
+ const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
+ const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
+ a1 += 8;
+ const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
+ const __m128i vxa2 = _mm_cvtepi8_epi16(va2);
+ a2 += 8;
+ const __m128i va3 = _mm_loadl_epi64((const __m128i*) a3);
+ const __m128i vxa3 = _mm_cvtepi8_epi16(va3);
+ a3 += 8;
+
+ const __m128i vb01 = _mm_loadu_si128((const __m128i*) w);
+ const __m128i vsb01 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb01);
+ const __m128i vxb0 = _mm_unpacklo_epi8(vb01, vsb01);
+ const __m128i vxb1 = _mm_unpackhi_epi8(vb01, vsb01);
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+ vacc2x0123 = _mm_add_epi32(vacc2x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+ vacc3x0123 = _mm_add_epi32(vacc3x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ vacc2x0123 = _mm_add_epi32(vacc2x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ vacc3x0123 = _mm_add_epi32(vacc3x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
+ const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
+ const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
+ const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ vacc2x0123 = _mm_add_epi32(vacc2x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ vacc3x0123 = _mm_add_epi32(vacc3x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
+ vacc2x0123 = _mm_add_epi32(vacc2x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
+ vacc3x0123 = _mm_add_epi32(vacc3x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
+
+ w = (const void*) ((uintptr_t) w + 32 * sizeof(int8_t));
+ k -= 8 * sizeof(int8_t);
+ }
+ if (k != 0) {
+ const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
+ const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
+ a0 = (const int8_t*) ((uintptr_t) a0 + k);
+ const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
+ const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
+ a1 = (const int8_t*) ((uintptr_t) a1 + k);
+ const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
+ const __m128i vxa2 = _mm_cvtepi8_epi16(va2);
+ a2 = (const int8_t*) ((uintptr_t) a2 + k);
+ const __m128i va3 = _mm_loadl_epi64((const __m128i*) a3);
+ const __m128i vxa3 = _mm_cvtepi8_epi16(va3);
+ a3 = (const int8_t*) ((uintptr_t) a3 + k);
+
+ const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+ vacc2x0123 = _mm_add_epi32(vacc2x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+ vacc3x0123 = _mm_add_epi32(vacc3x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+
+ if (k > 2 * sizeof(int8_t)) {
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ vacc2x0123 = _mm_add_epi32(vacc2x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ vacc3x0123 = _mm_add_epi32(vacc3x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+
+ if (k > 4 * sizeof(int8_t)) {
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ vacc2x0123 = _mm_add_epi32(vacc2x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ vacc3x0123 = _mm_add_epi32(vacc3x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ }
+ }
+ }
+
+ __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
+ __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
+ __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123);
+ __m128 vscaled3x0123 = _mm_cvtepi32_ps(vacc3x0123);
+
+ const __m128 vscale0123 = _mm_loadu_ps((const float*) w);
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(float));
+ vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123);
+ vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale0123);
+ vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale0123);
+ vscaled3x0123 = _mm_mul_ps(vscaled3x0123, vscale0123);
+
+ vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
+ vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
+ vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123);
+ vacc3x0123 = _mm_cvtps_epi32(vscaled3x0123);
+
+ const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4.output_zero_point);
+ __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
+ __m128i vacc23x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc3x0123), voutput_zero_point);
+
+
+ __m128i vout = _mm_packs_epi16(vacc01x0123, vacc23x0123);
+
+ vout = _mm_max_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_min));
+ vout = _mm_min_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_max));
+
+ if (nc >= 4) {
+ *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
+ *((uint32_t*) c1) = (uint32_t) _mm_extract_epi32(vout, 1);
+ *((uint32_t*) c2) = (uint32_t) _mm_extract_epi32(vout, 2);
+ *((uint32_t*) c3) = (uint32_t) _mm_extract_epi32(vout, 3);
+
+ 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);
+
+ 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);
+
+ nc -= 4;
+ } else {
+ if (nc & 2) {
+ *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
+ c0 += 2;
+ *((uint16_t*) c1) = (uint16_t) _mm_extract_epi16(vout, 2);
+ c1 += 2;
+ *((uint16_t*) c2) = (uint16_t) _mm_extract_epi16(vout, 4);
+ c2 += 2;
+ *((uint16_t*) c3) = (uint16_t) _mm_extract_epi16(vout, 6);
+ c3 += 2;
+ vout = _mm_srli_epi32(vout, 16);
+ }
+ if (nc & 1) {
+ *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
+ *((int8_t*) c2) = (int8_t) _mm_extract_epi8(vout, 8);
+ *((int8_t*) c3) = (int8_t) _mm_extract_epi8(vout, 12);
+ }
+
+ nc = 0;
+ }
+ } while (nc != 0);
+}
diff --git a/src/qc8-gemm/gen/4x4c2-minmax-fp32-avx-ld64.c b/src/qc8-gemm/gen/4x4c2-minmax-fp32-avx-ld64.c
new file mode 100644
index 0000000..d6f5fc2
--- /dev/null
+++ b/src/qc8-gemm/gen/4x4c2-minmax-fp32-avx-ld64.c
@@ -0,0 +1,256 @@
+// Auto-generated file. Do not edit!
+// Template: src/qs8-gemm/MRx4c2-sse.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 <smmintrin.h>
+
+#include <xnnpack/gemm.h>
+#include <xnnpack/math.h>
+
+
+
+void xnn_qc8_gemm_minmax_fp32_ukernel_4x4c2__avx_ld64(
+ 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_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN XNN_DISABLE_MSAN
+{
+ assert(mr != 0);
+ assert(mr <= 4);
+ 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, 2);
+ 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;
+ }
+
+ do {
+ __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
+ __m128i vacc1x0123 = vacc0x0123;
+ __m128i vacc2x0123 = vacc0x0123;
+ __m128i vacc3x0123 = vacc0x0123;
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+
+ size_t k = kc;
+ while (k >= 8 * sizeof(int8_t)) {
+ const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
+ const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
+ a0 += 8;
+ const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
+ const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
+ a1 += 8;
+ const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
+ const __m128i vxa2 = _mm_cvtepi8_epi16(va2);
+ a2 += 8;
+ const __m128i va3 = _mm_loadl_epi64((const __m128i*) a3);
+ const __m128i vxa3 = _mm_cvtepi8_epi16(va3);
+ a3 += 8;
+
+ const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+ vacc2x0123 = _mm_add_epi32(vacc2x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+ vacc3x0123 = _mm_add_epi32(vacc3x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int8_t)));
+ const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ vacc2x0123 = _mm_add_epi32(vacc2x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ vacc3x0123 = _mm_add_epi32(vacc3x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
+ const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ vacc2x0123 = _mm_add_epi32(vacc2x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ vacc3x0123 = _mm_add_epi32(vacc3x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int8_t)));
+ const __m128i vxb3 = _mm_cvtepi8_epi16(vb3);
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
+ vacc2x0123 = _mm_add_epi32(vacc2x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
+ vacc3x0123 = _mm_add_epi32(vacc3x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
+
+ w = (const void*) ((uintptr_t) w + 32 * sizeof(int8_t));
+ k -= 8 * sizeof(int8_t);
+ }
+ if (k != 0) {
+ const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
+ const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
+ a0 = (const int8_t*) ((uintptr_t) a0 + k);
+ const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
+ const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
+ a1 = (const int8_t*) ((uintptr_t) a1 + k);
+ const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
+ const __m128i vxa2 = _mm_cvtepi8_epi16(va2);
+ a2 = (const int8_t*) ((uintptr_t) a2 + k);
+ const __m128i va3 = _mm_loadl_epi64((const __m128i*) a3);
+ const __m128i vxa3 = _mm_cvtepi8_epi16(va3);
+ a3 = (const int8_t*) ((uintptr_t) a3 + k);
+
+ const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+ vacc2x0123 = _mm_add_epi32(vacc2x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+ vacc3x0123 = _mm_add_epi32(vacc3x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+
+ if (k > 2 * sizeof(int8_t)) {
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ vacc2x0123 = _mm_add_epi32(vacc2x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ vacc3x0123 = _mm_add_epi32(vacc3x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+
+ if (k > 4 * sizeof(int8_t)) {
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ vacc2x0123 = _mm_add_epi32(vacc2x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ vacc3x0123 = _mm_add_epi32(vacc3x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ }
+ }
+ }
+
+ __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
+ __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
+ __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123);
+ __m128 vscaled3x0123 = _mm_cvtepi32_ps(vacc3x0123);
+
+ const __m128 vscale0123 = _mm_loadu_ps((const float*) w);
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(float));
+ vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123);
+ vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale0123);
+ vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale0123);
+ vscaled3x0123 = _mm_mul_ps(vscaled3x0123, vscale0123);
+
+ vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
+ vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
+ vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123);
+ vacc3x0123 = _mm_cvtps_epi32(vscaled3x0123);
+
+ const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4.output_zero_point);
+ __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
+ __m128i vacc23x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc3x0123), voutput_zero_point);
+
+
+ __m128i vout = _mm_packs_epi16(vacc01x0123, vacc23x0123);
+
+ vout = _mm_max_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_min));
+ vout = _mm_min_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_max));
+
+ if (nc >= 4) {
+ *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
+ *((uint32_t*) c1) = (uint32_t) _mm_extract_epi32(vout, 1);
+ *((uint32_t*) c2) = (uint32_t) _mm_extract_epi32(vout, 2);
+ *((uint32_t*) c3) = (uint32_t) _mm_extract_epi32(vout, 3);
+
+ 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);
+
+ 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);
+
+ nc -= 4;
+ } else {
+ if (nc & 2) {
+ *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
+ c0 += 2;
+ *((uint16_t*) c1) = (uint16_t) _mm_extract_epi16(vout, 2);
+ c1 += 2;
+ *((uint16_t*) c2) = (uint16_t) _mm_extract_epi16(vout, 4);
+ c2 += 2;
+ *((uint16_t*) c3) = (uint16_t) _mm_extract_epi16(vout, 6);
+ c3 += 2;
+ vout = _mm_srli_epi32(vout, 16);
+ }
+ if (nc & 1) {
+ *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
+ *((int8_t*) c2) = (int8_t) _mm_extract_epi8(vout, 8);
+ *((int8_t*) c3) = (int8_t) _mm_extract_epi8(vout, 12);
+ }
+
+ nc = 0;
+ }
+ } while (nc != 0);
+}
diff --git a/src/qc8-gemm/gen/4x4c2-minmax-fp32-sse2-ld128.c b/src/qc8-gemm/gen/4x4c2-minmax-fp32-sse2-ld128.c
new file mode 100644
index 0000000..0d9da78
--- /dev/null
+++ b/src/qc8-gemm/gen/4x4c2-minmax-fp32-sse2-ld128.c
@@ -0,0 +1,261 @@
+// Auto-generated file. Do not edit!
+// Template: src/qs8-gemm/MRx4c2-sse.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 <emmintrin.h>
+
+#include <xnnpack/gemm.h>
+#include <xnnpack/math.h>
+
+
+
+void xnn_qc8_gemm_minmax_fp32_ukernel_4x4c2__sse2_ld128(
+ 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_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN XNN_DISABLE_MSAN
+{
+ assert(mr != 0);
+ assert(mr <= 4);
+ 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, 2);
+ 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;
+ }
+
+ do {
+ __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
+ __m128i vacc1x0123 = vacc0x0123;
+ __m128i vacc2x0123 = vacc0x0123;
+ __m128i vacc3x0123 = vacc0x0123;
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+
+ size_t k = kc;
+ while (k >= 8 * sizeof(int8_t)) {
+ const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
+ const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ a0 += 8;
+ const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
+ const __m128i vxa1 = _mm_unpacklo_epi8(va1, _mm_cmpgt_epi8(_mm_setzero_si128(), va1));
+ a1 += 8;
+ const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
+ const __m128i vxa2 = _mm_unpacklo_epi8(va2, _mm_cmpgt_epi8(_mm_setzero_si128(), va2));
+ a2 += 8;
+ const __m128i va3 = _mm_loadl_epi64((const __m128i*) a3);
+ const __m128i vxa3 = _mm_unpacklo_epi8(va3, _mm_cmpgt_epi8(_mm_setzero_si128(), va3));
+ a3 += 8;
+
+ const __m128i vb01 = _mm_loadu_si128((const __m128i*) w);
+ const __m128i vsb01 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb01);
+ const __m128i vxb0 = _mm_unpacklo_epi8(vb01, vsb01);
+ const __m128i vxb1 = _mm_unpackhi_epi8(vb01, vsb01);
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+ vacc2x0123 = _mm_add_epi32(vacc2x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+ vacc3x0123 = _mm_add_epi32(vacc3x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ vacc2x0123 = _mm_add_epi32(vacc2x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ vacc3x0123 = _mm_add_epi32(vacc3x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
+ const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
+ const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
+ const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ vacc2x0123 = _mm_add_epi32(vacc2x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ vacc3x0123 = _mm_add_epi32(vacc3x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
+ vacc2x0123 = _mm_add_epi32(vacc2x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
+ vacc3x0123 = _mm_add_epi32(vacc3x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
+
+ w = (const void*) ((uintptr_t) w + 32 * sizeof(int8_t));
+ k -= 8 * sizeof(int8_t);
+ }
+ if (k != 0) {
+ const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
+ const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ a0 = (const int8_t*) ((uintptr_t) a0 + k);
+ const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
+ const __m128i vxa1 = _mm_unpacklo_epi8(va1, _mm_cmpgt_epi8(_mm_setzero_si128(), va1));
+ a1 = (const int8_t*) ((uintptr_t) a1 + k);
+ const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
+ const __m128i vxa2 = _mm_unpacklo_epi8(va2, _mm_cmpgt_epi8(_mm_setzero_si128(), va2));
+ a2 = (const int8_t*) ((uintptr_t) a2 + k);
+ const __m128i va3 = _mm_loadl_epi64((const __m128i*) a3);
+ const __m128i vxa3 = _mm_unpacklo_epi8(va3, _mm_cmpgt_epi8(_mm_setzero_si128(), va3));
+ a3 = (const int8_t*) ((uintptr_t) a3 + k);
+
+ const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+ vacc2x0123 = _mm_add_epi32(vacc2x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+ vacc3x0123 = _mm_add_epi32(vacc3x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+
+ if (k > 2 * sizeof(int8_t)) {
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ vacc2x0123 = _mm_add_epi32(vacc2x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ vacc3x0123 = _mm_add_epi32(vacc3x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+
+ if (k > 4 * sizeof(int8_t)) {
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ vacc2x0123 = _mm_add_epi32(vacc2x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ vacc3x0123 = _mm_add_epi32(vacc3x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ }
+ }
+ }
+
+ __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
+ __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
+ __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123);
+ __m128 vscaled3x0123 = _mm_cvtepi32_ps(vacc3x0123);
+
+ const __m128 vscale0123 = _mm_loadu_ps((const float*) w);
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(float));
+ vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123);
+ vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale0123);
+ vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale0123);
+ vscaled3x0123 = _mm_mul_ps(vscaled3x0123, vscale0123);
+
+ vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
+ vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
+ vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123);
+ vacc3x0123 = _mm_cvtps_epi32(vscaled3x0123);
+
+ const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
+ __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
+ __m128i vacc23x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc3x0123), voutput_zero_point);
+
+ const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
+ const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max);
+ vacc01x0123 = _mm_min_epi16(_mm_max_epi16(vacc01x0123, voutput_min), voutput_max);
+ vacc23x0123 = _mm_min_epi16(_mm_max_epi16(vacc23x0123, voutput_min), voutput_max);
+
+ __m128i vout = _mm_packs_epi16(vacc01x0123, vacc23x0123);
+
+
+ if (nc >= 4) {
+ *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
+ vout = _mm_srli_si128(vout, 4);
+ *((uint32_t*) c1) = (uint32_t) _mm_cvtsi128_si32(vout);
+ vout = _mm_srli_si128(vout, 4);
+ *((uint32_t*) c2) = (uint32_t) _mm_cvtsi128_si32(vout);
+ vout = _mm_srli_si128(vout, 4);
+ *((uint32_t*) c3) = (uint32_t) _mm_cvtsi128_si32(vout);
+
+ 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);
+
+ 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);
+
+ nc -= 4;
+ } else {
+ if (nc & 2) {
+ *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
+ c0 += 2;
+ *((uint16_t*) c1) = (uint16_t) _mm_extract_epi16(vout, 2);
+ c1 += 2;
+ *((uint16_t*) c2) = (uint16_t) _mm_extract_epi16(vout, 4);
+ c2 += 2;
+ *((uint16_t*) c3) = (uint16_t) _mm_extract_epi16(vout, 6);
+ c3 += 2;
+ vout = _mm_srli_epi32(vout, 16);
+ }
+ if (nc & 1) {
+ *((int8_t*) c0) = (int8_t) _mm_cvtsi128_si32(vout);
+ *((int8_t*) c1) = (int8_t) _mm_extract_epi16(vout, 2);
+ *((int8_t*) c2) = (int8_t) _mm_extract_epi16(vout, 4);
+ *((int8_t*) c3) = (int8_t) _mm_extract_epi16(vout, 6);
+ }
+
+ nc = 0;
+ }
+ } while (nc != 0);
+}
diff --git a/src/qc8-gemm/gen/4x4c2-minmax-fp32-sse2-ld64.c b/src/qc8-gemm/gen/4x4c2-minmax-fp32-sse2-ld64.c
new file mode 100644
index 0000000..81f3f5c
--- /dev/null
+++ b/src/qc8-gemm/gen/4x4c2-minmax-fp32-sse2-ld64.c
@@ -0,0 +1,261 @@
+// Auto-generated file. Do not edit!
+// Template: src/qs8-gemm/MRx4c2-sse.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 <emmintrin.h>
+
+#include <xnnpack/gemm.h>
+#include <xnnpack/math.h>
+
+
+
+void xnn_qc8_gemm_minmax_fp32_ukernel_4x4c2__sse2_ld64(
+ 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_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN XNN_DISABLE_MSAN
+{
+ assert(mr != 0);
+ assert(mr <= 4);
+ 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, 2);
+ 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;
+ }
+
+ do {
+ __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
+ __m128i vacc1x0123 = vacc0x0123;
+ __m128i vacc2x0123 = vacc0x0123;
+ __m128i vacc3x0123 = vacc0x0123;
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+
+ size_t k = kc;
+ while (k >= 8 * sizeof(int8_t)) {
+ const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
+ const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ a0 += 8;
+ const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
+ const __m128i vxa1 = _mm_unpacklo_epi8(va1, _mm_cmpgt_epi8(_mm_setzero_si128(), va1));
+ a1 += 8;
+ const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
+ const __m128i vxa2 = _mm_unpacklo_epi8(va2, _mm_cmpgt_epi8(_mm_setzero_si128(), va2));
+ a2 += 8;
+ const __m128i va3 = _mm_loadl_epi64((const __m128i*) a3);
+ const __m128i vxa3 = _mm_unpacklo_epi8(va3, _mm_cmpgt_epi8(_mm_setzero_si128(), va3));
+ a3 += 8;
+
+ const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+ vacc2x0123 = _mm_add_epi32(vacc2x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+ vacc3x0123 = _mm_add_epi32(vacc3x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int8_t)));
+ const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ vacc2x0123 = _mm_add_epi32(vacc2x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ vacc3x0123 = _mm_add_epi32(vacc3x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
+ const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ vacc2x0123 = _mm_add_epi32(vacc2x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ vacc3x0123 = _mm_add_epi32(vacc3x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int8_t)));
+ const __m128i vxb3 = _mm_unpacklo_epi8(vb3, _mm_cmpgt_epi8(_mm_setzero_si128(), vb3));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
+ vacc2x0123 = _mm_add_epi32(vacc2x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
+ vacc3x0123 = _mm_add_epi32(vacc3x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
+
+ w = (const void*) ((uintptr_t) w + 32 * sizeof(int8_t));
+ k -= 8 * sizeof(int8_t);
+ }
+ if (k != 0) {
+ const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
+ const __m128i vxa0 = _mm_unpacklo_epi8(va0, _mm_cmpgt_epi8(_mm_setzero_si128(), va0));
+ a0 = (const int8_t*) ((uintptr_t) a0 + k);
+ const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
+ const __m128i vxa1 = _mm_unpacklo_epi8(va1, _mm_cmpgt_epi8(_mm_setzero_si128(), va1));
+ a1 = (const int8_t*) ((uintptr_t) a1 + k);
+ const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
+ const __m128i vxa2 = _mm_unpacklo_epi8(va2, _mm_cmpgt_epi8(_mm_setzero_si128(), va2));
+ a2 = (const int8_t*) ((uintptr_t) a2 + k);
+ const __m128i va3 = _mm_loadl_epi64((const __m128i*) a3);
+ const __m128i vxa3 = _mm_unpacklo_epi8(va3, _mm_cmpgt_epi8(_mm_setzero_si128(), va3));
+ a3 = (const int8_t*) ((uintptr_t) a3 + k);
+
+ const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb0 = _mm_unpacklo_epi8(vb0, _mm_cmpgt_epi8(_mm_setzero_si128(), vb0));
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+ vacc2x0123 = _mm_add_epi32(vacc2x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+ vacc3x0123 = _mm_add_epi32(vacc3x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+
+ if (k > 2 * sizeof(int8_t)) {
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb1 = _mm_unpacklo_epi8(vb1, _mm_cmpgt_epi8(_mm_setzero_si128(), vb1));
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ vacc2x0123 = _mm_add_epi32(vacc2x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ vacc3x0123 = _mm_add_epi32(vacc3x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+
+ if (k > 4 * sizeof(int8_t)) {
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb2 = _mm_unpacklo_epi8(vb2, _mm_cmpgt_epi8(_mm_setzero_si128(), vb2));
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ vacc2x0123 = _mm_add_epi32(vacc2x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ vacc3x0123 = _mm_add_epi32(vacc3x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ }
+ }
+ }
+
+ __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
+ __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
+ __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123);
+ __m128 vscaled3x0123 = _mm_cvtepi32_ps(vacc3x0123);
+
+ const __m128 vscale0123 = _mm_loadu_ps((const float*) w);
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(float));
+ vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123);
+ vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale0123);
+ vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale0123);
+ vscaled3x0123 = _mm_mul_ps(vscaled3x0123, vscale0123);
+
+ vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
+ vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
+ vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123);
+ vacc3x0123 = _mm_cvtps_epi32(vscaled3x0123);
+
+ const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point);
+ __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
+ __m128i vacc23x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc3x0123), voutput_zero_point);
+
+ const __m128i voutput_min = _mm_load_si128((const __m128i*) params->sse2.output_min);
+ const __m128i voutput_max = _mm_load_si128((const __m128i*) params->sse2.output_max);
+ vacc01x0123 = _mm_min_epi16(_mm_max_epi16(vacc01x0123, voutput_min), voutput_max);
+ vacc23x0123 = _mm_min_epi16(_mm_max_epi16(vacc23x0123, voutput_min), voutput_max);
+
+ __m128i vout = _mm_packs_epi16(vacc01x0123, vacc23x0123);
+
+
+ if (nc >= 4) {
+ *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
+ vout = _mm_srli_si128(vout, 4);
+ *((uint32_t*) c1) = (uint32_t) _mm_cvtsi128_si32(vout);
+ vout = _mm_srli_si128(vout, 4);
+ *((uint32_t*) c2) = (uint32_t) _mm_cvtsi128_si32(vout);
+ vout = _mm_srli_si128(vout, 4);
+ *((uint32_t*) c3) = (uint32_t) _mm_cvtsi128_si32(vout);
+
+ 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);
+
+ 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);
+
+ nc -= 4;
+ } else {
+ if (nc & 2) {
+ *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
+ c0 += 2;
+ *((uint16_t*) c1) = (uint16_t) _mm_extract_epi16(vout, 2);
+ c1 += 2;
+ *((uint16_t*) c2) = (uint16_t) _mm_extract_epi16(vout, 4);
+ c2 += 2;
+ *((uint16_t*) c3) = (uint16_t) _mm_extract_epi16(vout, 6);
+ c3 += 2;
+ vout = _mm_srli_epi32(vout, 16);
+ }
+ if (nc & 1) {
+ *((int8_t*) c0) = (int8_t) _mm_cvtsi128_si32(vout);
+ *((int8_t*) c1) = (int8_t) _mm_extract_epi16(vout, 2);
+ *((int8_t*) c2) = (int8_t) _mm_extract_epi16(vout, 4);
+ *((int8_t*) c3) = (int8_t) _mm_extract_epi16(vout, 6);
+ }
+
+ nc = 0;
+ }
+ } while (nc != 0);
+}
diff --git a/src/qc8-gemm/gen/4x4c2-minmax-fp32-sse41-ld128.c b/src/qc8-gemm/gen/4x4c2-minmax-fp32-sse41-ld128.c
new file mode 100644
index 0000000..f9b111d
--- /dev/null
+++ b/src/qc8-gemm/gen/4x4c2-minmax-fp32-sse41-ld128.c
@@ -0,0 +1,256 @@
+// Auto-generated file. Do not edit!
+// Template: src/qs8-gemm/MRx4c2-sse.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 <smmintrin.h>
+
+#include <xnnpack/gemm.h>
+#include <xnnpack/math.h>
+
+
+
+void xnn_qc8_gemm_minmax_fp32_ukernel_4x4c2__sse41_ld128(
+ 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_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN XNN_DISABLE_MSAN
+{
+ assert(mr != 0);
+ assert(mr <= 4);
+ 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, 2);
+ 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;
+ }
+
+ do {
+ __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
+ __m128i vacc1x0123 = vacc0x0123;
+ __m128i vacc2x0123 = vacc0x0123;
+ __m128i vacc3x0123 = vacc0x0123;
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+
+ size_t k = kc;
+ while (k >= 8 * sizeof(int8_t)) {
+ const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
+ const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
+ a0 += 8;
+ const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
+ const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
+ a1 += 8;
+ const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
+ const __m128i vxa2 = _mm_cvtepi8_epi16(va2);
+ a2 += 8;
+ const __m128i va3 = _mm_loadl_epi64((const __m128i*) a3);
+ const __m128i vxa3 = _mm_cvtepi8_epi16(va3);
+ a3 += 8;
+
+ const __m128i vb01 = _mm_loadu_si128((const __m128i*) w);
+ const __m128i vsb01 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb01);
+ const __m128i vxb0 = _mm_unpacklo_epi8(vb01, vsb01);
+ const __m128i vxb1 = _mm_unpackhi_epi8(vb01, vsb01);
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+ vacc2x0123 = _mm_add_epi32(vacc2x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+ vacc3x0123 = _mm_add_epi32(vacc3x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ vacc2x0123 = _mm_add_epi32(vacc2x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ vacc3x0123 = _mm_add_epi32(vacc3x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
+ const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
+ const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
+ const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ vacc2x0123 = _mm_add_epi32(vacc2x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ vacc3x0123 = _mm_add_epi32(vacc3x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
+ vacc2x0123 = _mm_add_epi32(vacc2x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
+ vacc3x0123 = _mm_add_epi32(vacc3x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
+
+ w = (const void*) ((uintptr_t) w + 32 * sizeof(int8_t));
+ k -= 8 * sizeof(int8_t);
+ }
+ if (k != 0) {
+ const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
+ const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
+ a0 = (const int8_t*) ((uintptr_t) a0 + k);
+ const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
+ const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
+ a1 = (const int8_t*) ((uintptr_t) a1 + k);
+ const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
+ const __m128i vxa2 = _mm_cvtepi8_epi16(va2);
+ a2 = (const int8_t*) ((uintptr_t) a2 + k);
+ const __m128i va3 = _mm_loadl_epi64((const __m128i*) a3);
+ const __m128i vxa3 = _mm_cvtepi8_epi16(va3);
+ a3 = (const int8_t*) ((uintptr_t) a3 + k);
+
+ const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+ vacc2x0123 = _mm_add_epi32(vacc2x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+ vacc3x0123 = _mm_add_epi32(vacc3x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+
+ if (k > 2 * sizeof(int8_t)) {
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ vacc2x0123 = _mm_add_epi32(vacc2x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ vacc3x0123 = _mm_add_epi32(vacc3x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+
+ if (k > 4 * sizeof(int8_t)) {
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ vacc2x0123 = _mm_add_epi32(vacc2x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ vacc3x0123 = _mm_add_epi32(vacc3x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ }
+ }
+ }
+
+ __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
+ __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
+ __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123);
+ __m128 vscaled3x0123 = _mm_cvtepi32_ps(vacc3x0123);
+
+ const __m128 vscale0123 = _mm_loadu_ps((const float*) w);
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(float));
+ vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123);
+ vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale0123);
+ vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale0123);
+ vscaled3x0123 = _mm_mul_ps(vscaled3x0123, vscale0123);
+
+ vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
+ vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
+ vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123);
+ vacc3x0123 = _mm_cvtps_epi32(vscaled3x0123);
+
+ const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4.output_zero_point);
+ __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
+ __m128i vacc23x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc3x0123), voutput_zero_point);
+
+
+ __m128i vout = _mm_packs_epi16(vacc01x0123, vacc23x0123);
+
+ vout = _mm_max_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_min));
+ vout = _mm_min_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_max));
+
+ if (nc >= 4) {
+ *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
+ *((uint32_t*) c1) = (uint32_t) _mm_extract_epi32(vout, 1);
+ *((uint32_t*) c2) = (uint32_t) _mm_extract_epi32(vout, 2);
+ *((uint32_t*) c3) = (uint32_t) _mm_extract_epi32(vout, 3);
+
+ 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);
+
+ 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);
+
+ nc -= 4;
+ } else {
+ if (nc & 2) {
+ *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
+ c0 += 2;
+ *((uint16_t*) c1) = (uint16_t) _mm_extract_epi16(vout, 2);
+ c1 += 2;
+ *((uint16_t*) c2) = (uint16_t) _mm_extract_epi16(vout, 4);
+ c2 += 2;
+ *((uint16_t*) c3) = (uint16_t) _mm_extract_epi16(vout, 6);
+ c3 += 2;
+ vout = _mm_srli_epi32(vout, 16);
+ }
+ if (nc & 1) {
+ *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
+ *((int8_t*) c2) = (int8_t) _mm_extract_epi8(vout, 8);
+ *((int8_t*) c3) = (int8_t) _mm_extract_epi8(vout, 12);
+ }
+
+ nc = 0;
+ }
+ } while (nc != 0);
+}
diff --git a/src/qc8-gemm/gen/4x4c2-minmax-fp32-sse41-ld64.c b/src/qc8-gemm/gen/4x4c2-minmax-fp32-sse41-ld64.c
new file mode 100644
index 0000000..965ef36
--- /dev/null
+++ b/src/qc8-gemm/gen/4x4c2-minmax-fp32-sse41-ld64.c
@@ -0,0 +1,256 @@
+// Auto-generated file. Do not edit!
+// Template: src/qs8-gemm/MRx4c2-sse.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 <smmintrin.h>
+
+#include <xnnpack/gemm.h>
+#include <xnnpack/math.h>
+
+
+
+void xnn_qc8_gemm_minmax_fp32_ukernel_4x4c2__sse41_ld64(
+ 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_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN XNN_DISABLE_MSAN
+{
+ assert(mr != 0);
+ assert(mr <= 4);
+ 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, 2);
+ 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;
+ }
+
+ do {
+ __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
+ __m128i vacc1x0123 = vacc0x0123;
+ __m128i vacc2x0123 = vacc0x0123;
+ __m128i vacc3x0123 = vacc0x0123;
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+
+ size_t k = kc;
+ while (k >= 8 * sizeof(int8_t)) {
+ const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
+ const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
+ a0 += 8;
+ const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
+ const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
+ a1 += 8;
+ const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
+ const __m128i vxa2 = _mm_cvtepi8_epi16(va2);
+ a2 += 8;
+ const __m128i va3 = _mm_loadl_epi64((const __m128i*) a3);
+ const __m128i vxa3 = _mm_cvtepi8_epi16(va3);
+ a3 += 8;
+
+ const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+ vacc2x0123 = _mm_add_epi32(vacc2x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+ vacc3x0123 = _mm_add_epi32(vacc3x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int8_t)));
+ const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ vacc2x0123 = _mm_add_epi32(vacc2x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ vacc3x0123 = _mm_add_epi32(vacc3x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
+ const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ vacc2x0123 = _mm_add_epi32(vacc2x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ vacc3x0123 = _mm_add_epi32(vacc3x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int8_t)));
+ const __m128i vxb3 = _mm_cvtepi8_epi16(vb3);
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
+ vacc2x0123 = _mm_add_epi32(vacc2x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
+ vacc3x0123 = _mm_add_epi32(vacc3x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(3, 3, 3, 3)), vxb3));
+
+ w = (const void*) ((uintptr_t) w + 32 * sizeof(int8_t));
+ k -= 8 * sizeof(int8_t);
+ }
+ if (k != 0) {
+ const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
+ const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
+ a0 = (const int8_t*) ((uintptr_t) a0 + k);
+ const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
+ const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
+ a1 = (const int8_t*) ((uintptr_t) a1 + k);
+ const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
+ const __m128i vxa2 = _mm_cvtepi8_epi16(va2);
+ a2 = (const int8_t*) ((uintptr_t) a2 + k);
+ const __m128i va3 = _mm_loadl_epi64((const __m128i*) a3);
+ const __m128i vxa3 = _mm_cvtepi8_epi16(va3);
+ a3 = (const int8_t*) ((uintptr_t) a3 + k);
+
+ const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+ vacc2x0123 = _mm_add_epi32(vacc2x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+ vacc3x0123 = _mm_add_epi32(vacc3x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(0, 0, 0, 0)), vxb0));
+
+ if (k > 2 * sizeof(int8_t)) {
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ vacc2x0123 = _mm_add_epi32(vacc2x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+ vacc3x0123 = _mm_add_epi32(vacc3x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(1, 1, 1, 1)), vxb1));
+
+ if (k > 4 * sizeof(int8_t)) {
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_add_epi32(vacc0x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ vacc1x0123 = _mm_add_epi32(vacc1x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ vacc2x0123 = _mm_add_epi32(vacc2x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ vacc3x0123 = _mm_add_epi32(vacc3x0123,
+ _mm_madd_epi16(_mm_shuffle_epi32(vxa3, _MM_SHUFFLE(2, 2, 2, 2)), vxb2));
+ }
+ }
+ }
+
+ __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
+ __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
+ __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123);
+ __m128 vscaled3x0123 = _mm_cvtepi32_ps(vacc3x0123);
+
+ const __m128 vscale0123 = _mm_loadu_ps((const float*) w);
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(float));
+ vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123);
+ vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale0123);
+ vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale0123);
+ vscaled3x0123 = _mm_mul_ps(vscaled3x0123, vscale0123);
+
+ vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
+ vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
+ vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123);
+ vacc3x0123 = _mm_cvtps_epi32(vscaled3x0123);
+
+ const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4.output_zero_point);
+ __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
+ __m128i vacc23x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc3x0123), voutput_zero_point);
+
+
+ __m128i vout = _mm_packs_epi16(vacc01x0123, vacc23x0123);
+
+ vout = _mm_max_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_min));
+ vout = _mm_min_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_max));
+
+ if (nc >= 4) {
+ *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
+ *((uint32_t*) c1) = (uint32_t) _mm_extract_epi32(vout, 1);
+ *((uint32_t*) c2) = (uint32_t) _mm_extract_epi32(vout, 2);
+ *((uint32_t*) c3) = (uint32_t) _mm_extract_epi32(vout, 3);
+
+ 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);
+
+ 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);
+
+ nc -= 4;
+ } else {
+ if (nc & 2) {
+ *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
+ c0 += 2;
+ *((uint16_t*) c1) = (uint16_t) _mm_extract_epi16(vout, 2);
+ c1 += 2;
+ *((uint16_t*) c2) = (uint16_t) _mm_extract_epi16(vout, 4);
+ c2 += 2;
+ *((uint16_t*) c3) = (uint16_t) _mm_extract_epi16(vout, 6);
+ c3 += 2;
+ vout = _mm_srli_epi32(vout, 16);
+ }
+ if (nc & 1) {
+ *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
+ *((int8_t*) c2) = (int8_t) _mm_extract_epi8(vout, 8);
+ *((int8_t*) c3) = (int8_t) _mm_extract_epi8(vout, 12);
+ }
+
+ nc = 0;
+ }
+ } while (nc != 0);
+}
diff --git a/src/qc8-gemm/gen/4x4c2-minmax-fp32-xop-ld128.c b/src/qc8-gemm/gen/4x4c2-minmax-fp32-xop-ld128.c
new file mode 100644
index 0000000..f0b662e
--- /dev/null
+++ b/src/qc8-gemm/gen/4x4c2-minmax-fp32-xop-ld128.c
@@ -0,0 +1,261 @@
+// Auto-generated file. Do not edit!
+// Template: src/qs8-gemm/MRx4c2-sse.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>
+
+#if defined(__GNUC__) || defined(__clang__)
+ #include <x86intrin.h>
+#else
+ #include <immintrin.h>
+ #include <ammintrin.h>
+#endif
+
+#include <xnnpack/gemm.h>
+#include <xnnpack/math.h>
+
+
+
+void xnn_qc8_gemm_minmax_fp32_ukernel_4x4c2__xop_ld128(
+ 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_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN XNN_DISABLE_MSAN
+{
+ assert(mr != 0);
+ assert(mr <= 4);
+ 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, 2);
+ 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;
+ }
+
+ do {
+ __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
+ __m128i vacc1x0123 = vacc0x0123;
+ __m128i vacc2x0123 = vacc0x0123;
+ __m128i vacc3x0123 = vacc0x0123;
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+
+ size_t k = kc;
+ while (k >= 8 * sizeof(int8_t)) {
+ const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
+ const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
+ a0 += 8;
+ const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
+ const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
+ a1 += 8;
+ const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
+ const __m128i vxa2 = _mm_cvtepi8_epi16(va2);
+ a2 += 8;
+ const __m128i va3 = _mm_loadl_epi64((const __m128i*) a3);
+ const __m128i vxa3 = _mm_cvtepi8_epi16(va3);
+ a3 += 8;
+
+ const __m128i vb01 = _mm_loadu_si128((const __m128i*) w);
+ const __m128i vsb01 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb01);
+ const __m128i vxb0 = _mm_unpacklo_epi8(vb01, vsb01);
+ const __m128i vxb1 = _mm_unpackhi_epi8(vb01, vsb01);
+
+ vacc0x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
+ vacc1x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc1x0123);
+ vacc2x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc2x0123);
+ vacc3x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa3, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc3x0123);
+
+ vacc0x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
+ vacc1x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc1x0123);
+ vacc2x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc2x0123);
+ vacc3x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa3, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc3x0123);
+ const __m128i vb23 = _mm_loadu_si128((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
+ const __m128i vsb23 = _mm_cmpgt_epi8(_mm_setzero_si128(), vb23);
+ const __m128i vxb2 = _mm_unpacklo_epi8(vb23, vsb23);
+ const __m128i vxb3 = _mm_unpackhi_epi8(vb23, vsb23);
+
+ vacc0x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
+ vacc1x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc1x0123);
+ vacc2x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc2x0123);
+ vacc3x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa3, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc3x0123);
+
+ vacc0x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc0x0123);
+ vacc1x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc1x0123);
+ vacc2x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc2x0123);
+ vacc3x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa3, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc3x0123);
+
+ w = (const void*) ((uintptr_t) w + 32 * sizeof(int8_t));
+ k -= 8 * sizeof(int8_t);
+ }
+ if (k != 0) {
+ const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
+ const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
+ a0 = (const int8_t*) ((uintptr_t) a0 + k);
+ const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
+ const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
+ a1 = (const int8_t*) ((uintptr_t) a1 + k);
+ const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
+ const __m128i vxa2 = _mm_cvtepi8_epi16(va2);
+ a2 = (const int8_t*) ((uintptr_t) a2 + k);
+ const __m128i va3 = _mm_loadl_epi64((const __m128i*) a3);
+ const __m128i vxa3 = _mm_cvtepi8_epi16(va3);
+ a3 = (const int8_t*) ((uintptr_t) a3 + k);
+
+ const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
+ vacc1x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc1x0123);
+ vacc2x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc2x0123);
+ vacc3x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa3, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc3x0123);
+
+ if (k > 2 * sizeof(int8_t)) {
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
+ vacc1x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc1x0123);
+ vacc2x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc2x0123);
+ vacc3x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa3, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc3x0123);
+
+ if (k > 4 * sizeof(int8_t)) {
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
+ vacc1x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc1x0123);
+ vacc2x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc2x0123);
+ vacc3x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa3, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc3x0123);
+ }
+ }
+ }
+
+ __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
+ __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
+ __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123);
+ __m128 vscaled3x0123 = _mm_cvtepi32_ps(vacc3x0123);
+
+ const __m128 vscale0123 = _mm_loadu_ps((const float*) w);
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(float));
+ vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123);
+ vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale0123);
+ vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale0123);
+ vscaled3x0123 = _mm_mul_ps(vscaled3x0123, vscale0123);
+
+ vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
+ vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
+ vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123);
+ vacc3x0123 = _mm_cvtps_epi32(vscaled3x0123);
+
+ const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4.output_zero_point);
+ __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
+ __m128i vacc23x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc3x0123), voutput_zero_point);
+
+
+ __m128i vout = _mm_packs_epi16(vacc01x0123, vacc23x0123);
+
+ vout = _mm_max_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_min));
+ vout = _mm_min_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_max));
+
+ if (nc >= 4) {
+ *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
+ *((uint32_t*) c1) = (uint32_t) _mm_extract_epi32(vout, 1);
+ *((uint32_t*) c2) = (uint32_t) _mm_extract_epi32(vout, 2);
+ *((uint32_t*) c3) = (uint32_t) _mm_extract_epi32(vout, 3);
+
+ 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);
+
+ 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);
+
+ nc -= 4;
+ } else {
+ if (nc & 2) {
+ *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
+ c0 += 2;
+ *((uint16_t*) c1) = (uint16_t) _mm_extract_epi16(vout, 2);
+ c1 += 2;
+ *((uint16_t*) c2) = (uint16_t) _mm_extract_epi16(vout, 4);
+ c2 += 2;
+ *((uint16_t*) c3) = (uint16_t) _mm_extract_epi16(vout, 6);
+ c3 += 2;
+ vout = _mm_srli_epi32(vout, 16);
+ }
+ if (nc & 1) {
+ *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
+ *((int8_t*) c2) = (int8_t) _mm_extract_epi8(vout, 8);
+ *((int8_t*) c3) = (int8_t) _mm_extract_epi8(vout, 12);
+ }
+
+ nc = 0;
+ }
+ } while (nc != 0);
+}
diff --git a/src/qc8-gemm/gen/4x4c2-minmax-fp32-xop-ld64.c b/src/qc8-gemm/gen/4x4c2-minmax-fp32-xop-ld64.c
new file mode 100644
index 0000000..ca51f86
--- /dev/null
+++ b/src/qc8-gemm/gen/4x4c2-minmax-fp32-xop-ld64.c
@@ -0,0 +1,261 @@
+// Auto-generated file. Do not edit!
+// Template: src/qs8-gemm/MRx4c2-sse.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>
+
+#if defined(__GNUC__) || defined(__clang__)
+ #include <x86intrin.h>
+#else
+ #include <immintrin.h>
+ #include <ammintrin.h>
+#endif
+
+#include <xnnpack/gemm.h>
+#include <xnnpack/math.h>
+
+
+
+void xnn_qc8_gemm_minmax_fp32_ukernel_4x4c2__xop_ld64(
+ 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_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN XNN_DISABLE_MSAN
+{
+ assert(mr != 0);
+ assert(mr <= 4);
+ 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, 2);
+ 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;
+ }
+
+ do {
+ __m128i vacc0x0123 = _mm_loadu_si128((const __m128i*) w);
+ __m128i vacc1x0123 = vacc0x0123;
+ __m128i vacc2x0123 = vacc0x0123;
+ __m128i vacc3x0123 = vacc0x0123;
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(int32_t));
+
+ size_t k = kc;
+ while (k >= 8 * sizeof(int8_t)) {
+ const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
+ const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
+ a0 += 8;
+ const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
+ const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
+ a1 += 8;
+ const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
+ const __m128i vxa2 = _mm_cvtepi8_epi16(va2);
+ a2 += 8;
+ const __m128i va3 = _mm_loadl_epi64((const __m128i*) a3);
+ const __m128i vxa3 = _mm_cvtepi8_epi16(va3);
+ a3 += 8;
+
+ const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
+
+ vacc0x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
+ vacc1x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc1x0123);
+ vacc2x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc2x0123);
+ vacc3x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa3, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc3x0123);
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 8 * sizeof(int8_t)));
+ const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
+
+ vacc0x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
+ vacc1x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc1x0123);
+ vacc2x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc2x0123);
+ vacc3x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa3, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc3x0123);
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 16 * sizeof(int8_t)));
+ const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
+
+ vacc0x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
+ vacc1x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc1x0123);
+ vacc2x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc2x0123);
+ vacc3x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa3, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc3x0123);
+ const __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((uintptr_t) w + 24 * sizeof(int8_t)));
+ const __m128i vxb3 = _mm_cvtepi8_epi16(vb3);
+
+ vacc0x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc0x0123);
+ vacc1x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc1x0123);
+ vacc2x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc2x0123);
+ vacc3x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa3, _MM_SHUFFLE(3, 3, 3, 3)), vxb3, vacc3x0123);
+
+ w = (const void*) ((uintptr_t) w + 32 * sizeof(int8_t));
+ k -= 8 * sizeof(int8_t);
+ }
+ if (k != 0) {
+ const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0);
+ const __m128i vxa0 = _mm_cvtepi8_epi16(va0);
+ a0 = (const int8_t*) ((uintptr_t) a0 + k);
+ const __m128i va1 = _mm_loadl_epi64((const __m128i*) a1);
+ const __m128i vxa1 = _mm_cvtepi8_epi16(va1);
+ a1 = (const int8_t*) ((uintptr_t) a1 + k);
+ const __m128i va2 = _mm_loadl_epi64((const __m128i*) a2);
+ const __m128i vxa2 = _mm_cvtepi8_epi16(va2);
+ a2 = (const int8_t*) ((uintptr_t) a2 + k);
+ const __m128i va3 = _mm_loadl_epi64((const __m128i*) a3);
+ const __m128i vxa3 = _mm_cvtepi8_epi16(va3);
+ a3 = (const int8_t*) ((uintptr_t) a3 + k);
+
+ const __m128i vb0 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb0 = _mm_cvtepi8_epi16(vb0);
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc0x0123);
+ vacc1x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc1x0123);
+ vacc2x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc2x0123);
+ vacc3x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa3, _MM_SHUFFLE(0, 0, 0, 0)), vxb0, vacc3x0123);
+
+ if (k > 2 * sizeof(int8_t)) {
+ const __m128i vb1 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb1 = _mm_cvtepi8_epi16(vb1);
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc0x0123);
+ vacc1x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc1x0123);
+ vacc2x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc2x0123);
+ vacc3x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa3, _MM_SHUFFLE(1, 1, 1, 1)), vxb1, vacc3x0123);
+
+ if (k > 4 * sizeof(int8_t)) {
+ const __m128i vb2 = _mm_loadl_epi64((const __m128i*) w);
+ const __m128i vxb2 = _mm_cvtepi8_epi16(vb2);
+ w = (const void*) ((uintptr_t) w + 8 * sizeof(int8_t));
+
+ vacc0x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa0, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc0x0123);
+ vacc1x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa1, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc1x0123);
+ vacc2x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa2, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc2x0123);
+ vacc3x0123 = _mm_maddd_epi16(
+ _mm_shuffle_epi32(vxa3, _MM_SHUFFLE(2, 2, 2, 2)), vxb2, vacc3x0123);
+ }
+ }
+ }
+
+ __m128 vscaled0x0123 = _mm_cvtepi32_ps(vacc0x0123);
+ __m128 vscaled1x0123 = _mm_cvtepi32_ps(vacc1x0123);
+ __m128 vscaled2x0123 = _mm_cvtepi32_ps(vacc2x0123);
+ __m128 vscaled3x0123 = _mm_cvtepi32_ps(vacc3x0123);
+
+ const __m128 vscale0123 = _mm_loadu_ps((const float*) w);
+ w = (const void*) ((uintptr_t) w + 4 * sizeof(float));
+ vscaled0x0123 = _mm_mul_ps(vscaled0x0123, vscale0123);
+ vscaled1x0123 = _mm_mul_ps(vscaled1x0123, vscale0123);
+ vscaled2x0123 = _mm_mul_ps(vscaled2x0123, vscale0123);
+ vscaled3x0123 = _mm_mul_ps(vscaled3x0123, vscale0123);
+
+ vacc0x0123 = _mm_cvtps_epi32(vscaled0x0123);
+ vacc1x0123 = _mm_cvtps_epi32(vscaled1x0123);
+ vacc2x0123 = _mm_cvtps_epi32(vscaled2x0123);
+ vacc3x0123 = _mm_cvtps_epi32(vscaled3x0123);
+
+ const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse4.output_zero_point);
+ __m128i vacc01x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc0x0123, vacc1x0123), voutput_zero_point);
+ __m128i vacc23x0123 = _mm_adds_epi16(_mm_packs_epi32(vacc2x0123, vacc3x0123), voutput_zero_point);
+
+
+ __m128i vout = _mm_packs_epi16(vacc01x0123, vacc23x0123);
+
+ vout = _mm_max_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_min));
+ vout = _mm_min_epi8(vout, _mm_load_si128((const __m128i*) params->sse4.output_max));
+
+ if (nc >= 4) {
+ *((uint32_t*) c0) = (uint32_t) _mm_cvtsi128_si32(vout);
+ *((uint32_t*) c1) = (uint32_t) _mm_extract_epi32(vout, 1);
+ *((uint32_t*) c2) = (uint32_t) _mm_extract_epi32(vout, 2);
+ *((uint32_t*) c3) = (uint32_t) _mm_extract_epi32(vout, 3);
+
+ 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);
+
+ 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);
+
+ nc -= 4;
+ } else {
+ if (nc & 2) {
+ *((uint16_t*) c0) = (uint16_t) _mm_extract_epi16(vout, 0);
+ c0 += 2;
+ *((uint16_t*) c1) = (uint16_t) _mm_extract_epi16(vout, 2);
+ c1 += 2;
+ *((uint16_t*) c2) = (uint16_t) _mm_extract_epi16(vout, 4);
+ c2 += 2;
+ *((uint16_t*) c3) = (uint16_t) _mm_extract_epi16(vout, 6);
+ c3 += 2;
+ vout = _mm_srli_epi32(vout, 16);
+ }
+ if (nc & 1) {
+ *((int8_t*) c0) = (int8_t) _mm_extract_epi8(vout, 0);
+ *((int8_t*) c1) = (int8_t) _mm_extract_epi8(vout, 4);
+ *((int8_t*) c2) = (int8_t) _mm_extract_epi8(vout, 8);
+ *((int8_t*) c3) = (int8_t) _mm_extract_epi8(vout, 12);
+ }
+
+ nc = 0;
+ }
+ } while (nc != 0);
+}