// Auto-generated file. Do not edit! // Template: src/qs8-gemm/MRx8c8-avx2.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 #include #include "xnnpack/common.h" #include "xnnpack/gemm.h" #include "xnnpack/intrinsics-polyfill.h" #include "xnnpack/math.h" #include "xnnpack/unaligned.h" void xnn_qd8_f16_qc4w_gemm_minmax_ukernel_8x8c8__avx256skx( size_t mr, size_t nc, size_t kc, const int8_t* restrict a, size_t a_stride, const void* restrict w, xnn_float16* restrict c, size_t cm_stride, size_t cn_stride, const struct xnn_f16_qc4w_minmax_params params[restrict XNN_MIN_ELEMENTS(1)], const struct xnn_qd8_quantization_params quantization_params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS { assert(mr != 0); assert(mr <= 8); 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 * sizeof(int8_t)); const int8_t* a0 = a; uint16_t* c0 = (uint16_t*) c; const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride); uint16_t* c1 = (uint16_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); uint16_t* c2 = (uint16_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); uint16_t* c3 = (uint16_t*) ((uintptr_t) c2 + cm_stride); if XNN_UNPREDICTABLE(mr < 4) { a3 = a2; c3 = c2; } const int8_t* a4 = (const int8_t*) ((uintptr_t) a3 + a_stride); uint16_t* c4 = (uint16_t*) ((uintptr_t) c3 + cm_stride); if XNN_UNPREDICTABLE(mr <= 4) { a4 = a3; c4 = c3; } const int8_t* a5 = (const int8_t*) ((uintptr_t) a4 + a_stride); uint16_t* c5 = (uint16_t*) ((uintptr_t) c4 + cm_stride); if XNN_UNPREDICTABLE(mr < 6) { a5 = a4; c5 = c4; } const int8_t* a6 = (const int8_t*) ((uintptr_t) a5 + a_stride); uint16_t* c6 = (uint16_t*) ((uintptr_t) c5 + cm_stride); if XNN_UNPREDICTABLE(mr <= 6) { a6 = a5; c6 = c5; } const int8_t* a7 = (const int8_t*) ((uintptr_t) a6 + a_stride); uint16_t* c7 = (uint16_t*) ((uintptr_t) c6 + cm_stride); if XNN_UNPREDICTABLE(mr != 8) { a7 = a6; c7 = c6; } const __m128i vmask = _mm_set1_epi8(0xF0); XNN_FORCE_REALIZATION(vmask); const __m256 vmin = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) ¶ms->scalar.min)); const __m256 vmax = _mm256_cvtph_ps(_mm_set1_epi16(*(const uint16_t*) ¶ms->scalar.max)); XNN_FORCE_REALIZATION(vmin); XNN_FORCE_REALIZATION(vmax); do { const __m128i vinit0 = _mm_cvtsi32_si128(((const int*) w)[0]); const __m128i vinit1 = _mm_cvtsi32_si128(((const int*) w)[1]); const __m256i vinit01 = _mm256_inserti128_si256(_mm256_castsi128_si256(vinit0), vinit1, 1); const __m128i vinit2 = _mm_cvtsi32_si128(((const int*) w)[2]); const __m128i vinit3 = _mm_cvtsi32_si128(((const int*) w)[3]); const __m256i vinit23 = _mm256_inserti128_si256(_mm256_castsi128_si256(vinit2), vinit3, 1); const __m128i vinit4 = _mm_cvtsi32_si128(((const int*) w)[4]); const __m128i vinit5 = _mm_cvtsi32_si128(((const int*) w)[5]); const __m256i vinit45 = _mm256_inserti128_si256(_mm256_castsi128_si256(vinit4), vinit5, 1); const __m128i vinit6 = _mm_cvtsi32_si128(((const int*) w)[6]); const __m128i vinit7 = _mm_cvtsi32_si128(((const int*) w)[7]); const __m256i vinit67 = _mm256_inserti128_si256(_mm256_castsi128_si256(vinit6), vinit7, 1); const __m256i vinput_zero_point0 = _mm256_set1_epi32((int) quantization_params[0].zero_point); __m256i vacc0x01 = _mm256_mullo_epi32(vinit01, vinput_zero_point0); __m256i vacc0x23 = _mm256_mullo_epi32(vinit23, vinput_zero_point0); __m256i vacc0x45 = _mm256_mullo_epi32(vinit45, vinput_zero_point0); __m256i vacc0x67 = _mm256_mullo_epi32(vinit67, vinput_zero_point0); const __m256i vinput_zero_point1 = _mm256_set1_epi32((int) quantization_params[1].zero_point); __m256i vacc1x01 = _mm256_mullo_epi32(vinit01, vinput_zero_point1); __m256i vacc1x23 = _mm256_mullo_epi32(vinit23, vinput_zero_point1); __m256i vacc1x45 = _mm256_mullo_epi32(vinit45, vinput_zero_point1); __m256i vacc1x67 = _mm256_mullo_epi32(vinit67, vinput_zero_point1); const __m256i vinput_zero_point2 = _mm256_set1_epi32((int) quantization_params[2].zero_point); __m256i vacc2x01 = _mm256_mullo_epi32(vinit01, vinput_zero_point2); __m256i vacc2x23 = _mm256_mullo_epi32(vinit23, vinput_zero_point2); __m256i vacc2x45 = _mm256_mullo_epi32(vinit45, vinput_zero_point2); __m256i vacc2x67 = _mm256_mullo_epi32(vinit67, vinput_zero_point2); const __m256i vinput_zero_point3 = _mm256_set1_epi32((int) quantization_params[3].zero_point); __m256i vacc3x01 = _mm256_mullo_epi32(vinit01, vinput_zero_point3); __m256i vacc3x23 = _mm256_mullo_epi32(vinit23, vinput_zero_point3); __m256i vacc3x45 = _mm256_mullo_epi32(vinit45, vinput_zero_point3); __m256i vacc3x67 = _mm256_mullo_epi32(vinit67, vinput_zero_point3); const __m256i vinput_zero_point4 = _mm256_set1_epi32((int) quantization_params[4].zero_point); __m256i vacc4x01 = _mm256_mullo_epi32(vinit01, vinput_zero_point4); __m256i vacc4x23 = _mm256_mullo_epi32(vinit23, vinput_zero_point4); __m256i vacc4x45 = _mm256_mullo_epi32(vinit45, vinput_zero_point4); __m256i vacc4x67 = _mm256_mullo_epi32(vinit67, vinput_zero_point4); const __m256i vinput_zero_point5 = _mm256_set1_epi32((int) quantization_params[5].zero_point); __m256i vacc5x01 = _mm256_mullo_epi32(vinit01, vinput_zero_point5); __m256i vacc5x23 = _mm256_mullo_epi32(vinit23, vinput_zero_point5); __m256i vacc5x45 = _mm256_mullo_epi32(vinit45, vinput_zero_point5); __m256i vacc5x67 = _mm256_mullo_epi32(vinit67, vinput_zero_point5); const __m256i vinput_zero_point6 = _mm256_set1_epi32((int) quantization_params[6].zero_point); __m256i vacc6x01 = _mm256_mullo_epi32(vinit01, vinput_zero_point6); __m256i vacc6x23 = _mm256_mullo_epi32(vinit23, vinput_zero_point6); __m256i vacc6x45 = _mm256_mullo_epi32(vinit45, vinput_zero_point6); __m256i vacc6x67 = _mm256_mullo_epi32(vinit67, vinput_zero_point6); const __m256i vinput_zero_point7 = _mm256_set1_epi32((int) quantization_params[7].zero_point); __m256i vacc7x01 = _mm256_mullo_epi32(vinit01, vinput_zero_point7); __m256i vacc7x23 = _mm256_mullo_epi32(vinit23, vinput_zero_point7); __m256i vacc7x45 = _mm256_mullo_epi32(vinit45, vinput_zero_point7); __m256i vacc7x67 = _mm256_mullo_epi32(vinit67, vinput_zero_point7); w = (const int32_t*) w + 8; size_t k = kc; while (k >= 16 * sizeof(int8_t)) { __m128i va0 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a0)); __m256i vxa0 = _mm256_cvtepi8_epi16(va0); a0 += 8; __m128i va1 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a1)); __m256i vxa1 = _mm256_cvtepi8_epi16(va1); a1 += 8; __m128i va2 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a2)); __m256i vxa2 = _mm256_cvtepi8_epi16(va2); a2 += 8; __m128i va3 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a3)); __m256i vxa3 = _mm256_cvtepi8_epi16(va3); a3 += 8; __m128i va4 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a4)); __m256i vxa4 = _mm256_cvtepi8_epi16(va4); a4 += 8; __m128i va5 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a5)); __m256i vxa5 = _mm256_cvtepi8_epi16(va5); a5 += 8; __m128i va6 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a6)); __m256i vxa6 = _mm256_cvtepi8_epi16(va6); a6 += 8; __m128i va7 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a7)); __m256i vxa7 = _mm256_cvtepi8_epi16(va7); a7 += 8; __m128i vb01 = _mm_load_si128((const __m128i*) w); __m128i vbs01 = _mm_slli_epi32(vb01, 4); __m128i vbm01 = _mm_and_si128(vbs01, vmask); __m256i vxb01 = _mm256_cvtepi8_epi16(vbm01); vacc0x01 = _mm256_add_epi32(vacc0x01, _mm256_madd_epi16(vxa0, vxb01)); vacc1x01 = _mm256_add_epi32(vacc1x01, _mm256_madd_epi16(vxa1, vxb01)); vacc2x01 = _mm256_add_epi32(vacc2x01, _mm256_madd_epi16(vxa2, vxb01)); vacc3x01 = _mm256_add_epi32(vacc3x01, _mm256_madd_epi16(vxa3, vxb01)); vacc4x01 = _mm256_add_epi32(vacc4x01, _mm256_madd_epi16(vxa4, vxb01)); vacc5x01 = _mm256_add_epi32(vacc5x01, _mm256_madd_epi16(vxa5, vxb01)); vacc6x01 = _mm256_add_epi32(vacc6x01, _mm256_madd_epi16(vxa6, vxb01)); vacc7x01 = _mm256_add_epi32(vacc7x01, _mm256_madd_epi16(vxa7, vxb01)); __m128i vb23 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16)); __m128i vbs23 = _mm_slli_epi32(vb23, 4); __m128i vbm23 = _mm_and_si128(vbs23, vmask); __m256i vxb23 = _mm256_cvtepi8_epi16(vbm23); vacc0x23 = _mm256_add_epi32(vacc0x23, _mm256_madd_epi16(vxa0, vxb23)); vacc1x23 = _mm256_add_epi32(vacc1x23, _mm256_madd_epi16(vxa1, vxb23)); vacc2x23 = _mm256_add_epi32(vacc2x23, _mm256_madd_epi16(vxa2, vxb23)); vacc3x23 = _mm256_add_epi32(vacc3x23, _mm256_madd_epi16(vxa3, vxb23)); vacc4x23 = _mm256_add_epi32(vacc4x23, _mm256_madd_epi16(vxa4, vxb23)); vacc5x23 = _mm256_add_epi32(vacc5x23, _mm256_madd_epi16(vxa5, vxb23)); vacc6x23 = _mm256_add_epi32(vacc6x23, _mm256_madd_epi16(vxa6, vxb23)); vacc7x23 = _mm256_add_epi32(vacc7x23, _mm256_madd_epi16(vxa7, vxb23)); __m128i vb45 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 32)); __m128i vbs45 = _mm_slli_epi32(vb45, 4); __m128i vbm45 = _mm_and_si128(vbs45, vmask); __m256i vxb45 = _mm256_cvtepi8_epi16(vbm45); vacc0x45 = _mm256_add_epi32(vacc0x45, _mm256_madd_epi16(vxa0, vxb45)); vacc1x45 = _mm256_add_epi32(vacc1x45, _mm256_madd_epi16(vxa1, vxb45)); vacc2x45 = _mm256_add_epi32(vacc2x45, _mm256_madd_epi16(vxa2, vxb45)); vacc3x45 = _mm256_add_epi32(vacc3x45, _mm256_madd_epi16(vxa3, vxb45)); vacc4x45 = _mm256_add_epi32(vacc4x45, _mm256_madd_epi16(vxa4, vxb45)); vacc5x45 = _mm256_add_epi32(vacc5x45, _mm256_madd_epi16(vxa5, vxb45)); vacc6x45 = _mm256_add_epi32(vacc6x45, _mm256_madd_epi16(vxa6, vxb45)); vacc7x45 = _mm256_add_epi32(vacc7x45, _mm256_madd_epi16(vxa7, vxb45)); __m128i vb67 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 48)); __m128i vbs67 = _mm_slli_epi32(vb67, 4); __m128i vbm67 = _mm_and_si128(vbs67, vmask); __m256i vxb67 = _mm256_cvtepi8_epi16(vbm67); vacc0x67 = _mm256_add_epi32(vacc0x67, _mm256_madd_epi16(vxa0, vxb67)); vacc1x67 = _mm256_add_epi32(vacc1x67, _mm256_madd_epi16(vxa1, vxb67)); vacc2x67 = _mm256_add_epi32(vacc2x67, _mm256_madd_epi16(vxa2, vxb67)); vacc3x67 = _mm256_add_epi32(vacc3x67, _mm256_madd_epi16(vxa3, vxb67)); vacc4x67 = _mm256_add_epi32(vacc4x67, _mm256_madd_epi16(vxa4, vxb67)); vacc5x67 = _mm256_add_epi32(vacc5x67, _mm256_madd_epi16(vxa5, vxb67)); vacc6x67 = _mm256_add_epi32(vacc6x67, _mm256_madd_epi16(vxa6, vxb67)); vacc7x67 = _mm256_add_epi32(vacc7x67, _mm256_madd_epi16(vxa7, vxb67)); va0 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a0)); vxa0 = _mm256_cvtepi8_epi16(va0); a0 += 8; va1 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a1)); vxa1 = _mm256_cvtepi8_epi16(va1); a1 += 8; va2 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a2)); vxa2 = _mm256_cvtepi8_epi16(va2); a2 += 8; va3 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a3)); vxa3 = _mm256_cvtepi8_epi16(va3); a3 += 8; va4 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a4)); vxa4 = _mm256_cvtepi8_epi16(va4); a4 += 8; va5 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a5)); vxa5 = _mm256_cvtepi8_epi16(va5); a5 += 8; va6 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a6)); vxa6 = _mm256_cvtepi8_epi16(va6); a6 += 8; va7 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a7)); vxa7 = _mm256_cvtepi8_epi16(va7); a7 += 8; vbm01 = _mm_and_si128(vb01, vmask); vxb01 = _mm256_cvtepi8_epi16(vbm01); vacc0x01 = _mm256_add_epi32(vacc0x01, _mm256_madd_epi16(vxa0, vxb01)); vacc1x01 = _mm256_add_epi32(vacc1x01, _mm256_madd_epi16(vxa1, vxb01)); vacc2x01 = _mm256_add_epi32(vacc2x01, _mm256_madd_epi16(vxa2, vxb01)); vacc3x01 = _mm256_add_epi32(vacc3x01, _mm256_madd_epi16(vxa3, vxb01)); vacc4x01 = _mm256_add_epi32(vacc4x01, _mm256_madd_epi16(vxa4, vxb01)); vacc5x01 = _mm256_add_epi32(vacc5x01, _mm256_madd_epi16(vxa5, vxb01)); vacc6x01 = _mm256_add_epi32(vacc6x01, _mm256_madd_epi16(vxa6, vxb01)); vacc7x01 = _mm256_add_epi32(vacc7x01, _mm256_madd_epi16(vxa7, vxb01)); vbm23 = _mm_and_si128(vb23, vmask); vxb23 = _mm256_cvtepi8_epi16(vbm23); vacc0x23 = _mm256_add_epi32(vacc0x23, _mm256_madd_epi16(vxa0, vxb23)); vacc1x23 = _mm256_add_epi32(vacc1x23, _mm256_madd_epi16(vxa1, vxb23)); vacc2x23 = _mm256_add_epi32(vacc2x23, _mm256_madd_epi16(vxa2, vxb23)); vacc3x23 = _mm256_add_epi32(vacc3x23, _mm256_madd_epi16(vxa3, vxb23)); vacc4x23 = _mm256_add_epi32(vacc4x23, _mm256_madd_epi16(vxa4, vxb23)); vacc5x23 = _mm256_add_epi32(vacc5x23, _mm256_madd_epi16(vxa5, vxb23)); vacc6x23 = _mm256_add_epi32(vacc6x23, _mm256_madd_epi16(vxa6, vxb23)); vacc7x23 = _mm256_add_epi32(vacc7x23, _mm256_madd_epi16(vxa7, vxb23)); vbm45 = _mm_and_si128(vb45, vmask); vxb45 = _mm256_cvtepi8_epi16(vbm45); vacc0x45 = _mm256_add_epi32(vacc0x45, _mm256_madd_epi16(vxa0, vxb45)); vacc1x45 = _mm256_add_epi32(vacc1x45, _mm256_madd_epi16(vxa1, vxb45)); vacc2x45 = _mm256_add_epi32(vacc2x45, _mm256_madd_epi16(vxa2, vxb45)); vacc3x45 = _mm256_add_epi32(vacc3x45, _mm256_madd_epi16(vxa3, vxb45)); vacc4x45 = _mm256_add_epi32(vacc4x45, _mm256_madd_epi16(vxa4, vxb45)); vacc5x45 = _mm256_add_epi32(vacc5x45, _mm256_madd_epi16(vxa5, vxb45)); vacc6x45 = _mm256_add_epi32(vacc6x45, _mm256_madd_epi16(vxa6, vxb45)); vacc7x45 = _mm256_add_epi32(vacc7x45, _mm256_madd_epi16(vxa7, vxb45)); vbm67 = _mm_and_si128(vb67, vmask); vxb67 = _mm256_cvtepi8_epi16(vbm67); vacc0x67 = _mm256_add_epi32(vacc0x67, _mm256_madd_epi16(vxa0, vxb67)); vacc1x67 = _mm256_add_epi32(vacc1x67, _mm256_madd_epi16(vxa1, vxb67)); vacc2x67 = _mm256_add_epi32(vacc2x67, _mm256_madd_epi16(vxa2, vxb67)); vacc3x67 = _mm256_add_epi32(vacc3x67, _mm256_madd_epi16(vxa3, vxb67)); vacc4x67 = _mm256_add_epi32(vacc4x67, _mm256_madd_epi16(vxa4, vxb67)); vacc5x67 = _mm256_add_epi32(vacc5x67, _mm256_madd_epi16(vxa5, vxb67)); vacc6x67 = _mm256_add_epi32(vacc6x67, _mm256_madd_epi16(vxa6, vxb67)); vacc7x67 = _mm256_add_epi32(vacc7x67, _mm256_madd_epi16(vxa7, vxb67)); w = (const int8_t*) w + 64; k -= 16 * sizeof(int8_t); } while (k >= 8 * sizeof(int8_t)) { const __m128i va0 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a0)); const __m256i vxa0 = _mm256_cvtepi8_epi16(va0); a0 += 8; const __m128i va1 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a1)); const __m256i vxa1 = _mm256_cvtepi8_epi16(va1); a1 += 8; const __m128i va2 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a2)); const __m256i vxa2 = _mm256_cvtepi8_epi16(va2); a2 += 8; const __m128i va3 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a3)); const __m256i vxa3 = _mm256_cvtepi8_epi16(va3); a3 += 8; const __m128i va4 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a4)); const __m256i vxa4 = _mm256_cvtepi8_epi16(va4); a4 += 8; const __m128i va5 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a5)); const __m256i vxa5 = _mm256_cvtepi8_epi16(va5); a5 += 8; const __m128i va6 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a6)); const __m256i vxa6 = _mm256_cvtepi8_epi16(va6); a6 += 8; const __m128i va7 = _mm_broadcastq_epi64(_mm_loadl_epi64((const __m128i*) a7)); const __m256i vxa7 = _mm256_cvtepi8_epi16(va7); a7 += 8; const __m128i vb01 = _mm_load_si128((const __m128i*) w); const __m128i vbs01 = _mm_slli_epi32(vb01, 4); const __m128i vbm01 = _mm_and_si128(vbs01, vmask); const __m256i vxb01 = _mm256_cvtepi8_epi16(vbm01); vacc0x01 = _mm256_add_epi32(vacc0x01, _mm256_madd_epi16(vxa0, vxb01)); vacc1x01 = _mm256_add_epi32(vacc1x01, _mm256_madd_epi16(vxa1, vxb01)); vacc2x01 = _mm256_add_epi32(vacc2x01, _mm256_madd_epi16(vxa2, vxb01)); vacc3x01 = _mm256_add_epi32(vacc3x01, _mm256_madd_epi16(vxa3, vxb01)); vacc4x01 = _mm256_add_epi32(vacc4x01, _mm256_madd_epi16(vxa4, vxb01)); vacc5x01 = _mm256_add_epi32(vacc5x01, _mm256_madd_epi16(vxa5, vxb01)); vacc6x01 = _mm256_add_epi32(vacc6x01, _mm256_madd_epi16(vxa6, vxb01)); vacc7x01 = _mm256_add_epi32(vacc7x01, _mm256_madd_epi16(vxa7, vxb01)); const __m128i vb23 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16)); const __m128i vbs23 = _mm_slli_epi32(vb23, 4); const __m128i vbm23 = _mm_and_si128(vbs23, vmask); const __m256i vxb23 = _mm256_cvtepi8_epi16(vbm23); vacc0x23 = _mm256_add_epi32(vacc0x23, _mm256_madd_epi16(vxa0, vxb23)); vacc1x23 = _mm256_add_epi32(vacc1x23, _mm256_madd_epi16(vxa1, vxb23)); vacc2x23 = _mm256_add_epi32(vacc2x23, _mm256_madd_epi16(vxa2, vxb23)); vacc3x23 = _mm256_add_epi32(vacc3x23, _mm256_madd_epi16(vxa3, vxb23)); vacc4x23 = _mm256_add_epi32(vacc4x23, _mm256_madd_epi16(vxa4, vxb23)); vacc5x23 = _mm256_add_epi32(vacc5x23, _mm256_madd_epi16(vxa5, vxb23)); vacc6x23 = _mm256_add_epi32(vacc6x23, _mm256_madd_epi16(vxa6, vxb23)); vacc7x23 = _mm256_add_epi32(vacc7x23, _mm256_madd_epi16(vxa7, vxb23)); const __m128i vb45 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 32)); const __m128i vbs45 = _mm_slli_epi32(vb45, 4); const __m128i vbm45 = _mm_and_si128(vbs45, vmask); const __m256i vxb45 = _mm256_cvtepi8_epi16(vbm45); vacc0x45 = _mm256_add_epi32(vacc0x45, _mm256_madd_epi16(vxa0, vxb45)); vacc1x45 = _mm256_add_epi32(vacc1x45, _mm256_madd_epi16(vxa1, vxb45)); vacc2x45 = _mm256_add_epi32(vacc2x45, _mm256_madd_epi16(vxa2, vxb45)); vacc3x45 = _mm256_add_epi32(vacc3x45, _mm256_madd_epi16(vxa3, vxb45)); vacc4x45 = _mm256_add_epi32(vacc4x45, _mm256_madd_epi16(vxa4, vxb45)); vacc5x45 = _mm256_add_epi32(vacc5x45, _mm256_madd_epi16(vxa5, vxb45)); vacc6x45 = _mm256_add_epi32(vacc6x45, _mm256_madd_epi16(vxa6, vxb45)); vacc7x45 = _mm256_add_epi32(vacc7x45, _mm256_madd_epi16(vxa7, vxb45)); const __m128i vb67 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 48)); const __m128i vbs67 = _mm_slli_epi32(vb67, 4); const __m128i vbm67 = _mm_and_si128(vbs67, vmask); const __m256i vxb67 = _mm256_cvtepi8_epi16(vbm67); vacc0x67 = _mm256_add_epi32(vacc0x67, _mm256_madd_epi16(vxa0, vxb67)); vacc1x67 = _mm256_add_epi32(vacc1x67, _mm256_madd_epi16(vxa1, vxb67)); vacc2x67 = _mm256_add_epi32(vacc2x67, _mm256_madd_epi16(vxa2, vxb67)); vacc3x67 = _mm256_add_epi32(vacc3x67, _mm256_madd_epi16(vxa3, vxb67)); vacc4x67 = _mm256_add_epi32(vacc4x67, _mm256_madd_epi16(vxa4, vxb67)); vacc5x67 = _mm256_add_epi32(vacc5x67, _mm256_madd_epi16(vxa5, vxb67)); vacc6x67 = _mm256_add_epi32(vacc6x67, _mm256_madd_epi16(vxa6, vxb67)); vacc7x67 = _mm256_add_epi32(vacc7x67, _mm256_madd_epi16(vxa7, vxb67)); w = (const int8_t*) w + 64; k -= 8 * sizeof(int8_t); } const __m256i vacc0x0213 = _mm256_hadd_epi32(vacc0x01, vacc0x23); const __m256i vacc0x4657 = _mm256_hadd_epi32(vacc0x45, vacc0x67); const __m256i vacc1x0213 = _mm256_hadd_epi32(vacc1x01, vacc1x23); const __m256i vacc1x4657 = _mm256_hadd_epi32(vacc1x45, vacc1x67); const __m256i vacc2x0213 = _mm256_hadd_epi32(vacc2x01, vacc2x23); const __m256i vacc2x4657 = _mm256_hadd_epi32(vacc2x45, vacc2x67); const __m256i vacc3x0213 = _mm256_hadd_epi32(vacc3x01, vacc3x23); const __m256i vacc3x4657 = _mm256_hadd_epi32(vacc3x45, vacc3x67); const __m256i vacc4x0213 = _mm256_hadd_epi32(vacc4x01, vacc4x23); const __m256i vacc4x4657 = _mm256_hadd_epi32(vacc4x45, vacc4x67); const __m256i vacc5x0213 = _mm256_hadd_epi32(vacc5x01, vacc5x23); const __m256i vacc5x4657 = _mm256_hadd_epi32(vacc5x45, vacc5x67); const __m256i vacc6x0213 = _mm256_hadd_epi32(vacc6x01, vacc6x23); const __m256i vacc6x4657 = _mm256_hadd_epi32(vacc6x45, vacc6x67); const __m256i vacc7x0213 = _mm256_hadd_epi32(vacc7x01, vacc7x23); const __m256i vacc7x4657 = _mm256_hadd_epi32(vacc7x45, vacc7x67); const __m256i vacc0x02461357 = _mm256_hadd_epi32(vacc0x0213, vacc0x4657); const __m256i vacc1x02461357 = _mm256_hadd_epi32(vacc1x0213, vacc1x4657); const __m256i vacc2x02461357 = _mm256_hadd_epi32(vacc2x0213, vacc2x4657); const __m256i vacc3x02461357 = _mm256_hadd_epi32(vacc3x0213, vacc3x4657); const __m256i vacc4x02461357 = _mm256_hadd_epi32(vacc4x0213, vacc4x4657); const __m256i vacc5x02461357 = _mm256_hadd_epi32(vacc5x0213, vacc5x4657); const __m256i vacc6x02461357 = _mm256_hadd_epi32(vacc6x0213, vacc6x4657); const __m256i vacc7x02461357 = _mm256_hadd_epi32(vacc7x0213, vacc7x4657); const __m256i vpermute_mask = _mm256_set_epi32(7, 3, 6, 2, 5, 1, 4, 0); __m256i vacc0x01234567 = _mm256_permutevar8x32_epi32(vacc0x02461357, vpermute_mask); __m256i vacc1x01234567 = _mm256_permutevar8x32_epi32(vacc1x02461357, vpermute_mask); __m256i vacc2x01234567 = _mm256_permutevar8x32_epi32(vacc2x02461357, vpermute_mask); __m256i vacc3x01234567 = _mm256_permutevar8x32_epi32(vacc3x02461357, vpermute_mask); __m256i vacc4x01234567 = _mm256_permutevar8x32_epi32(vacc4x02461357, vpermute_mask); __m256i vacc5x01234567 = _mm256_permutevar8x32_epi32(vacc5x02461357, vpermute_mask); __m256i vacc6x01234567 = _mm256_permutevar8x32_epi32(vacc6x02461357, vpermute_mask); __m256i vacc7x01234567 = _mm256_permutevar8x32_epi32(vacc7x02461357, vpermute_mask); vacc0x01234567 = _mm256_srai_epi32(vacc0x01234567, 4); vacc1x01234567 = _mm256_srai_epi32(vacc1x01234567, 4); vacc2x01234567 = _mm256_srai_epi32(vacc2x01234567, 4); vacc3x01234567 = _mm256_srai_epi32(vacc3x01234567, 4); vacc4x01234567 = _mm256_srai_epi32(vacc4x01234567, 4); vacc5x01234567 = _mm256_srai_epi32(vacc5x01234567, 4); vacc6x01234567 = _mm256_srai_epi32(vacc6x01234567, 4); vacc7x01234567 = _mm256_srai_epi32(vacc7x01234567, 4); __m256 vout0x01234567 = _mm256_cvtepi32_ps(vacc0x01234567); const __m256 vinput_scale0 = _mm256_broadcast_ss(&quantization_params[0].inv_scale); __m256 vout1x01234567 = _mm256_cvtepi32_ps(vacc1x01234567); const __m256 vinput_scale1 = _mm256_broadcast_ss(&quantization_params[1].inv_scale); __m256 vout2x01234567 = _mm256_cvtepi32_ps(vacc2x01234567); const __m256 vinput_scale2 = _mm256_broadcast_ss(&quantization_params[2].inv_scale); __m256 vout3x01234567 = _mm256_cvtepi32_ps(vacc3x01234567); const __m256 vinput_scale3 = _mm256_broadcast_ss(&quantization_params[3].inv_scale); __m256 vout4x01234567 = _mm256_cvtepi32_ps(vacc4x01234567); const __m256 vinput_scale4 = _mm256_broadcast_ss(&quantization_params[4].inv_scale); __m256 vout5x01234567 = _mm256_cvtepi32_ps(vacc5x01234567); const __m256 vinput_scale5 = _mm256_broadcast_ss(&quantization_params[5].inv_scale); __m256 vout6x01234567 = _mm256_cvtepi32_ps(vacc6x01234567); const __m256 vinput_scale6 = _mm256_broadcast_ss(&quantization_params[6].inv_scale); __m256 vout7x01234567 = _mm256_cvtepi32_ps(vacc7x01234567); const __m256 vinput_scale7 = _mm256_broadcast_ss(&quantization_params[7].inv_scale); vout0x01234567 = _mm256_mul_ps(vout0x01234567, vinput_scale0); vout1x01234567 = _mm256_mul_ps(vout1x01234567, vinput_scale1); vout2x01234567 = _mm256_mul_ps(vout2x01234567, vinput_scale2); vout3x01234567 = _mm256_mul_ps(vout3x01234567, vinput_scale3); vout4x01234567 = _mm256_mul_ps(vout4x01234567, vinput_scale4); vout5x01234567 = _mm256_mul_ps(vout5x01234567, vinput_scale5); vout6x01234567 = _mm256_mul_ps(vout6x01234567, vinput_scale6); vout7x01234567 = _mm256_mul_ps(vout7x01234567, vinput_scale7); const __m256 vfilter_output_scale01234567 = _mm256_load_ps((const float*) w); const __m256 vbias01234567 = _mm256_load_ps((const float*) w + 8); w = (const float*) w + 16; vout0x01234567 = _mm256_fmadd_ps(vout0x01234567, vfilter_output_scale01234567, vbias01234567); vout1x01234567 = _mm256_fmadd_ps(vout1x01234567, vfilter_output_scale01234567, vbias01234567); vout2x01234567 = _mm256_fmadd_ps(vout2x01234567, vfilter_output_scale01234567, vbias01234567); vout3x01234567 = _mm256_fmadd_ps(vout3x01234567, vfilter_output_scale01234567, vbias01234567); vout4x01234567 = _mm256_fmadd_ps(vout4x01234567, vfilter_output_scale01234567, vbias01234567); vout5x01234567 = _mm256_fmadd_ps(vout5x01234567, vfilter_output_scale01234567, vbias01234567); vout6x01234567 = _mm256_fmadd_ps(vout6x01234567, vfilter_output_scale01234567, vbias01234567); vout7x01234567 = _mm256_fmadd_ps(vout7x01234567, vfilter_output_scale01234567, vbias01234567); vout0x01234567 = _mm256_max_ps(vout0x01234567, vmin); vout1x01234567 = _mm256_max_ps(vout1x01234567, vmin); vout2x01234567 = _mm256_max_ps(vout2x01234567, vmin); vout3x01234567 = _mm256_max_ps(vout3x01234567, vmin); vout4x01234567 = _mm256_max_ps(vout4x01234567, vmin); vout5x01234567 = _mm256_max_ps(vout5x01234567, vmin); vout6x01234567 = _mm256_max_ps(vout6x01234567, vmin); vout7x01234567 = _mm256_max_ps(vout7x01234567, vmin); vout0x01234567 = _mm256_min_ps(vout0x01234567, vmax); vout1x01234567 = _mm256_min_ps(vout1x01234567, vmax); vout2x01234567 = _mm256_min_ps(vout2x01234567, vmax); vout3x01234567 = _mm256_min_ps(vout3x01234567, vmax); vout4x01234567 = _mm256_min_ps(vout4x01234567, vmax); vout5x01234567 = _mm256_min_ps(vout5x01234567, vmax); vout6x01234567 = _mm256_min_ps(vout6x01234567, vmax); vout7x01234567 = _mm256_min_ps(vout7x01234567, vmax); __m128i vfp16out0x01234567 = _mm256_cvtps_ph(vout0x01234567, _MM_FROUND_TO_NEAREST_INT); __m128i vfp16out1x01234567 = _mm256_cvtps_ph(vout1x01234567, _MM_FROUND_TO_NEAREST_INT); __m128i vfp16out2x01234567 = _mm256_cvtps_ph(vout2x01234567, _MM_FROUND_TO_NEAREST_INT); __m128i vfp16out3x01234567 = _mm256_cvtps_ph(vout3x01234567, _MM_FROUND_TO_NEAREST_INT); __m128i vfp16out4x01234567 = _mm256_cvtps_ph(vout4x01234567, _MM_FROUND_TO_NEAREST_INT); __m128i vfp16out5x01234567 = _mm256_cvtps_ph(vout5x01234567, _MM_FROUND_TO_NEAREST_INT); __m128i vfp16out6x01234567 = _mm256_cvtps_ph(vout6x01234567, _MM_FROUND_TO_NEAREST_INT); __m128i vfp16out7x01234567 = _mm256_cvtps_ph(vout7x01234567, _MM_FROUND_TO_NEAREST_INT); if XNN_LIKELY(nc >= 8) { _mm_storeu_si128((__m128i*) c0, vfp16out0x01234567); c0 = (uint16_t*) ((uintptr_t) c0 + cn_stride); _mm_storeu_si128((__m128i*) c1, vfp16out1x01234567); c1 = (uint16_t*) ((uintptr_t) c1 + cn_stride); _mm_storeu_si128((__m128i*) c2, vfp16out2x01234567); c2 = (uint16_t*) ((uintptr_t) c2 + cn_stride); _mm_storeu_si128((__m128i*) c3, vfp16out3x01234567); c3 = (uint16_t*) ((uintptr_t) c3 + cn_stride); _mm_storeu_si128((__m128i*) c4, vfp16out4x01234567); c4 = (uint16_t*) ((uintptr_t) c4 + cn_stride); _mm_storeu_si128((__m128i*) c5, vfp16out5x01234567); c5 = (uint16_t*) ((uintptr_t) c5 + cn_stride); _mm_storeu_si128((__m128i*) c6, vfp16out6x01234567); c6 = (uint16_t*) ((uintptr_t) c6 + cn_stride); _mm_storeu_si128((__m128i*) c7, vfp16out7x01234567); c7 = (uint16_t*) ((uintptr_t) c7 + cn_stride); a0 = (const int8_t*) ((uintptr_t) a0 - kc); a1 = (const int8_t*) ((uintptr_t) a1 - kc); a2 = (const int8_t*) ((uintptr_t) a2 - kc); a3 = (const int8_t*) ((uintptr_t) a3 - kc); a4 = (const int8_t*) ((uintptr_t) a4 - kc); a5 = (const int8_t*) ((uintptr_t) a5 - kc); a6 = (const int8_t*) ((uintptr_t) a6 - kc); a7 = (const int8_t*) ((uintptr_t) a7 - kc); nc -= 8; } else { // Prepare mask for valid 16-bit elements (depends on nc). const __mmask8 vmask = _cvtu32_mask8((UINT32_C(1) << nc) - 1); _mm_mask_storeu_epi16(c0, vmask, vfp16out0x01234567); _mm_mask_storeu_epi16(c1, vmask, vfp16out1x01234567); _mm_mask_storeu_epi16(c2, vmask, vfp16out2x01234567); _mm_mask_storeu_epi16(c3, vmask, vfp16out3x01234567); _mm_mask_storeu_epi16(c4, vmask, vfp16out4x01234567); _mm_mask_storeu_epi16(c5, vmask, vfp16out5x01234567); _mm_mask_storeu_epi16(c6, vmask, vfp16out6x01234567); _mm_mask_storeu_epi16(c7, vmask, vfp16out7x01234567); nc = 0; } } while (nc != 0); }