// Auto-generated file. Do not edit! // Template: src/qs8-dwconv/unipass-avx512skx-mul32.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/dwconv.h" #include "xnnpack/intrinsics-polyfill.h" void xnn_qs8_dwconv_minmax_fp32_ukernel_25p32c__avx512skx_mul32( size_t channels, size_t output_width, const int8_t** input, const void* weights, int8_t* output, intptr_t input_stride, size_t output_increment, size_t input_offset, const int8_t* zero, const union xnn_qs8_conv_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS { assert(channels != 0); assert(output_width != 0); const __m512 vscale = _mm512_set1_ps(params->fp32_scalar.scale); XNN_FORCE_REALIZATION(vscale); const __m512 voutput_max_less_zero_point = _mm512_set1_ps((int32_t) params->fp32_scalar.output_max - (int32_t) params->fp32_scalar.output_zero_point); // XNN_FORCE_REALIZATION(voutput_max_less_zero_point); const __m512i voutput_zero_point = _mm512_set1_epi16((int16_t) params->fp32_scalar.output_zero_point); const __m256i voutput_min = _mm256_set1_epi8(params->fp32_scalar.output_min); const __m256i vpermute_mask = _mm256_set_epi32(7, 3, 5, 1, 6, 2, 4, 0); // XNN_FORCE_REALIZATION(voutput_zero_point); // XNN_FORCE_REALIZATION(voutput_min); do { const int8_t* i0 = input[0]; assert(i0 != NULL); if XNN_UNPREDICTABLE(i0 != zero) { i0 = (const int8_t*) ((uintptr_t) i0 + input_offset); } const int8_t* i1 = input[1]; assert(i1 != NULL); if XNN_UNPREDICTABLE(i1 != zero) { i1 = (const int8_t*) ((uintptr_t) i1 + input_offset); } const int8_t* i2 = input[2]; assert(i2 != NULL); if XNN_UNPREDICTABLE(i2 != zero) { i2 = (const int8_t*) ((uintptr_t) i2 + input_offset); } const int8_t* i3 = input[3]; assert(i3 != NULL); if XNN_UNPREDICTABLE(i3 != zero) { i3 = (const int8_t*) ((uintptr_t) i3 + input_offset); } const int8_t* i4 = input[4]; assert(i4 != NULL); if XNN_UNPREDICTABLE(i4 != zero) { i4 = (const int8_t*) ((uintptr_t) i4 + input_offset); } const int8_t* i5 = input[5]; assert(i5 != NULL); if XNN_UNPREDICTABLE(i5 != zero) { i5 = (const int8_t*) ((uintptr_t) i5 + input_offset); } const int8_t* i6 = input[6]; assert(i6 != NULL); if XNN_UNPREDICTABLE(i6 != zero) { i6 = (const int8_t*) ((uintptr_t) i6 + input_offset); } const int8_t* i7 = input[7]; assert(i7 != NULL); if XNN_UNPREDICTABLE(i7 != zero) { i7 = (const int8_t*) ((uintptr_t) i7 + input_offset); } const int8_t* i8 = input[8]; assert(i8 != NULL); if XNN_UNPREDICTABLE(i8 != zero) { i8 = (const int8_t*) ((uintptr_t) i8 + input_offset); } const int8_t* i9 = input[9]; assert(i9 != NULL); if XNN_UNPREDICTABLE(i9 != zero) { i9 = (const int8_t*) ((uintptr_t) i9 + input_offset); } const int8_t* i10 = input[10]; assert(i10 != NULL); if XNN_UNPREDICTABLE(i10 != zero) { i10 = (const int8_t*) ((uintptr_t) i10 + input_offset); } const int8_t* i11 = input[11]; assert(i11 != NULL); if XNN_UNPREDICTABLE(i11 != zero) { i11 = (const int8_t*) ((uintptr_t) i11 + input_offset); } const int8_t* i12 = input[12]; assert(i12 != NULL); if XNN_UNPREDICTABLE(i12 != zero) { i12 = (const int8_t*) ((uintptr_t) i12 + input_offset); } const int8_t* i13 = input[13]; assert(i13 != NULL); if XNN_UNPREDICTABLE(i13 != zero) { i13 = (const int8_t*) ((uintptr_t) i13 + input_offset); } const int8_t* i14 = input[14]; assert(i14 != NULL); if XNN_UNPREDICTABLE(i14 != zero) { i14 = (const int8_t*) ((uintptr_t) i14 + input_offset); } const int8_t* i15 = input[15]; assert(i15 != NULL); if XNN_UNPREDICTABLE(i15 != zero) { i15 = (const int8_t*) ((uintptr_t) i15 + input_offset); } const int8_t* i16 = input[16]; assert(i16 != NULL); if XNN_UNPREDICTABLE(i16 != zero) { i16 = (const int8_t*) ((uintptr_t) i16 + input_offset); } const int8_t* i17 = input[17]; assert(i17 != NULL); if XNN_UNPREDICTABLE(i17 != zero) { i17 = (const int8_t*) ((uintptr_t) i17 + input_offset); } const int8_t* i18 = input[18]; assert(i18 != NULL); if XNN_UNPREDICTABLE(i18 != zero) { i18 = (const int8_t*) ((uintptr_t) i18 + input_offset); } const int8_t* i19 = input[19]; assert(i19 != NULL); if XNN_UNPREDICTABLE(i19 != zero) { i19 = (const int8_t*) ((uintptr_t) i19 + input_offset); } const int8_t* i20 = input[20]; assert(i20 != NULL); if XNN_UNPREDICTABLE(i20 != zero) { i20 = (const int8_t*) ((uintptr_t) i20 + input_offset); } const int8_t* i21 = input[21]; assert(i21 != NULL); if XNN_UNPREDICTABLE(i21 != zero) { i21 = (const int8_t*) ((uintptr_t) i21 + input_offset); } const int8_t* i22 = input[22]; assert(i22 != NULL); if XNN_UNPREDICTABLE(i22 != zero) { i22 = (const int8_t*) ((uintptr_t) i22 + input_offset); } const int8_t* i23 = input[23]; assert(i23 != NULL); if XNN_UNPREDICTABLE(i23 != zero) { i23 = (const int8_t*) ((uintptr_t) i23 + input_offset); } const int8_t* i24 = input[24]; assert(i24 != NULL); if XNN_UNPREDICTABLE(i24 != zero) { i24 = (const int8_t*) ((uintptr_t) i24 + input_offset); } input = (const int8_t**) ((uintptr_t) input + input_stride); size_t c = channels; const void* w = weights; for (; c >= 32; c -= 32) { __m512i vacc0123456789ABCDEF = _mm512_loadu_si512(w); __m512i vaccGHIJKLMNOPQRSTUV = _mm512_loadu_si512((const void*) ((uintptr_t) w + 16 * sizeof(int32_t))); const __m512i vi0x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i0)); const __m512i vk0x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 0 * sizeof(int8_t)))); const __m512i vi0xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (i0 + 16))); const __m512i vk0xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 16 * sizeof(int8_t)))); i0 += 32; vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF)); vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vaccGHIJKLMNOPQRSTUV, _mm512_mullo_epi32(vi0xGHIJKLMNOPQRSTUV, vk0xGHIJKLMNOPQRSTUV)); const __m512i vi1x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i1)); const __m512i vk1x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 32 * sizeof(int8_t)))); const __m512i vi1xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (i1 + 16))); const __m512i vk1xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 48 * sizeof(int8_t)))); i1 += 32; vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF)); vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vaccGHIJKLMNOPQRSTUV, _mm512_mullo_epi32(vi1xGHIJKLMNOPQRSTUV, vk1xGHIJKLMNOPQRSTUV)); const __m512i vi2x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i2)); const __m512i vk2x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 64 * sizeof(int8_t)))); const __m512i vi2xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (i2 + 16))); const __m512i vk2xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 80 * sizeof(int8_t)))); i2 += 32; vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF)); vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vaccGHIJKLMNOPQRSTUV, _mm512_mullo_epi32(vi2xGHIJKLMNOPQRSTUV, vk2xGHIJKLMNOPQRSTUV)); const __m512i vi3x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i3)); const __m512i vk3x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 96 * sizeof(int8_t)))); const __m512i vi3xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (i3 + 16))); const __m512i vk3xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 112 * sizeof(int8_t)))); i3 += 32; vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF)); vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vaccGHIJKLMNOPQRSTUV, _mm512_mullo_epi32(vi3xGHIJKLMNOPQRSTUV, vk3xGHIJKLMNOPQRSTUV)); const __m512i vi4x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i4)); const __m512i vk4x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 128 * sizeof(int8_t)))); const __m512i vi4xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (i4 + 16))); const __m512i vk4xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 144 * sizeof(int8_t)))); i4 += 32; vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF)); vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vaccGHIJKLMNOPQRSTUV, _mm512_mullo_epi32(vi4xGHIJKLMNOPQRSTUV, vk4xGHIJKLMNOPQRSTUV)); const __m512i vi5x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i5)); const __m512i vk5x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 160 * sizeof(int8_t)))); const __m512i vi5xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (i5 + 16))); const __m512i vk5xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 176 * sizeof(int8_t)))); i5 += 32; vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi5x0123456789ABCDEF, vk5x0123456789ABCDEF)); vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vaccGHIJKLMNOPQRSTUV, _mm512_mullo_epi32(vi5xGHIJKLMNOPQRSTUV, vk5xGHIJKLMNOPQRSTUV)); const __m512i vi6x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i6)); const __m512i vk6x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 192 * sizeof(int8_t)))); const __m512i vi6xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (i6 + 16))); const __m512i vk6xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 208 * sizeof(int8_t)))); i6 += 32; vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi6x0123456789ABCDEF, vk6x0123456789ABCDEF)); vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vaccGHIJKLMNOPQRSTUV, _mm512_mullo_epi32(vi6xGHIJKLMNOPQRSTUV, vk6xGHIJKLMNOPQRSTUV)); const __m512i vi7x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i7)); const __m512i vk7x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 224 * sizeof(int8_t)))); const __m512i vi7xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (i7 + 16))); const __m512i vk7xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 240 * sizeof(int8_t)))); i7 += 32; vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi7x0123456789ABCDEF, vk7x0123456789ABCDEF)); vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vaccGHIJKLMNOPQRSTUV, _mm512_mullo_epi32(vi7xGHIJKLMNOPQRSTUV, vk7xGHIJKLMNOPQRSTUV)); const __m512i vi8x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i8)); const __m512i vk8x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 256 * sizeof(int8_t)))); const __m512i vi8xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (i8 + 16))); const __m512i vk8xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 272 * sizeof(int8_t)))); i8 += 32; vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi8x0123456789ABCDEF, vk8x0123456789ABCDEF)); vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vaccGHIJKLMNOPQRSTUV, _mm512_mullo_epi32(vi8xGHIJKLMNOPQRSTUV, vk8xGHIJKLMNOPQRSTUV)); const __m512i vi9x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i9)); const __m512i vk9x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 288 * sizeof(int8_t)))); const __m512i vi9xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (i9 + 16))); const __m512i vk9xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 304 * sizeof(int8_t)))); i9 += 32; vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi9x0123456789ABCDEF, vk9x0123456789ABCDEF)); vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vaccGHIJKLMNOPQRSTUV, _mm512_mullo_epi32(vi9xGHIJKLMNOPQRSTUV, vk9xGHIJKLMNOPQRSTUV)); const __m512i vi10x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i10)); const __m512i vk10x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 320 * sizeof(int8_t)))); const __m512i vi10xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (i10 + 16))); const __m512i vk10xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 336 * sizeof(int8_t)))); i10 += 32; vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi10x0123456789ABCDEF, vk10x0123456789ABCDEF)); vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vaccGHIJKLMNOPQRSTUV, _mm512_mullo_epi32(vi10xGHIJKLMNOPQRSTUV, vk10xGHIJKLMNOPQRSTUV)); const __m512i vi11x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i11)); const __m512i vk11x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 352 * sizeof(int8_t)))); const __m512i vi11xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (i11 + 16))); const __m512i vk11xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 368 * sizeof(int8_t)))); i11 += 32; vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi11x0123456789ABCDEF, vk11x0123456789ABCDEF)); vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vaccGHIJKLMNOPQRSTUV, _mm512_mullo_epi32(vi11xGHIJKLMNOPQRSTUV, vk11xGHIJKLMNOPQRSTUV)); const __m512i vi12x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i12)); const __m512i vk12x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 384 * sizeof(int8_t)))); const __m512i vi12xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (i12 + 16))); const __m512i vk12xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 400 * sizeof(int8_t)))); i12 += 32; vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi12x0123456789ABCDEF, vk12x0123456789ABCDEF)); vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vaccGHIJKLMNOPQRSTUV, _mm512_mullo_epi32(vi12xGHIJKLMNOPQRSTUV, vk12xGHIJKLMNOPQRSTUV)); const __m512i vi13x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i13)); const __m512i vk13x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 416 * sizeof(int8_t)))); const __m512i vi13xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (i13 + 16))); const __m512i vk13xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 432 * sizeof(int8_t)))); i13 += 32; vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi13x0123456789ABCDEF, vk13x0123456789ABCDEF)); vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vaccGHIJKLMNOPQRSTUV, _mm512_mullo_epi32(vi13xGHIJKLMNOPQRSTUV, vk13xGHIJKLMNOPQRSTUV)); const __m512i vi14x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i14)); const __m512i vk14x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 448 * sizeof(int8_t)))); const __m512i vi14xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (i14 + 16))); const __m512i vk14xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 464 * sizeof(int8_t)))); i14 += 32; vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi14x0123456789ABCDEF, vk14x0123456789ABCDEF)); vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vaccGHIJKLMNOPQRSTUV, _mm512_mullo_epi32(vi14xGHIJKLMNOPQRSTUV, vk14xGHIJKLMNOPQRSTUV)); const __m512i vi15x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i15)); const __m512i vk15x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 480 * sizeof(int8_t)))); const __m512i vi15xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (i15 + 16))); const __m512i vk15xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 496 * sizeof(int8_t)))); i15 += 32; vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi15x0123456789ABCDEF, vk15x0123456789ABCDEF)); vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vaccGHIJKLMNOPQRSTUV, _mm512_mullo_epi32(vi15xGHIJKLMNOPQRSTUV, vk15xGHIJKLMNOPQRSTUV)); const __m512i vi16x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i16)); const __m512i vk16x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 512 * sizeof(int8_t)))); const __m512i vi16xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (i16 + 16))); const __m512i vk16xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 528 * sizeof(int8_t)))); i16 += 32; vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi16x0123456789ABCDEF, vk16x0123456789ABCDEF)); vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vaccGHIJKLMNOPQRSTUV, _mm512_mullo_epi32(vi16xGHIJKLMNOPQRSTUV, vk16xGHIJKLMNOPQRSTUV)); const __m512i vi17x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i17)); const __m512i vk17x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 544 * sizeof(int8_t)))); const __m512i vi17xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (i17 + 16))); const __m512i vk17xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 560 * sizeof(int8_t)))); i17 += 32; vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi17x0123456789ABCDEF, vk17x0123456789ABCDEF)); vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vaccGHIJKLMNOPQRSTUV, _mm512_mullo_epi32(vi17xGHIJKLMNOPQRSTUV, vk17xGHIJKLMNOPQRSTUV)); const __m512i vi18x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i18)); const __m512i vk18x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 576 * sizeof(int8_t)))); const __m512i vi18xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (i18 + 16))); const __m512i vk18xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 592 * sizeof(int8_t)))); i18 += 32; vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi18x0123456789ABCDEF, vk18x0123456789ABCDEF)); vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vaccGHIJKLMNOPQRSTUV, _mm512_mullo_epi32(vi18xGHIJKLMNOPQRSTUV, vk18xGHIJKLMNOPQRSTUV)); const __m512i vi19x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i19)); const __m512i vk19x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 608 * sizeof(int8_t)))); const __m512i vi19xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (i19 + 16))); const __m512i vk19xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 624 * sizeof(int8_t)))); i19 += 32; vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi19x0123456789ABCDEF, vk19x0123456789ABCDEF)); vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vaccGHIJKLMNOPQRSTUV, _mm512_mullo_epi32(vi19xGHIJKLMNOPQRSTUV, vk19xGHIJKLMNOPQRSTUV)); const __m512i vi20x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i20)); const __m512i vk20x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 640 * sizeof(int8_t)))); const __m512i vi20xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (i20 + 16))); const __m512i vk20xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 656 * sizeof(int8_t)))); i20 += 32; vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi20x0123456789ABCDEF, vk20x0123456789ABCDEF)); vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vaccGHIJKLMNOPQRSTUV, _mm512_mullo_epi32(vi20xGHIJKLMNOPQRSTUV, vk20xGHIJKLMNOPQRSTUV)); const __m512i vi21x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i21)); const __m512i vk21x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 672 * sizeof(int8_t)))); const __m512i vi21xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (i21 + 16))); const __m512i vk21xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 688 * sizeof(int8_t)))); i21 += 32; vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi21x0123456789ABCDEF, vk21x0123456789ABCDEF)); vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vaccGHIJKLMNOPQRSTUV, _mm512_mullo_epi32(vi21xGHIJKLMNOPQRSTUV, vk21xGHIJKLMNOPQRSTUV)); const __m512i vi22x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i22)); const __m512i vk22x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 704 * sizeof(int8_t)))); const __m512i vi22xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (i22 + 16))); const __m512i vk22xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 720 * sizeof(int8_t)))); i22 += 32; vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi22x0123456789ABCDEF, vk22x0123456789ABCDEF)); vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vaccGHIJKLMNOPQRSTUV, _mm512_mullo_epi32(vi22xGHIJKLMNOPQRSTUV, vk22xGHIJKLMNOPQRSTUV)); const __m512i vi23x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i23)); const __m512i vk23x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 736 * sizeof(int8_t)))); const __m512i vi23xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (i23 + 16))); const __m512i vk23xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 752 * sizeof(int8_t)))); i23 += 32; vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi23x0123456789ABCDEF, vk23x0123456789ABCDEF)); vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vaccGHIJKLMNOPQRSTUV, _mm512_mullo_epi32(vi23xGHIJKLMNOPQRSTUV, vk23xGHIJKLMNOPQRSTUV)); const __m512i vi24x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i24)); const __m512i vk24x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 768 * sizeof(int8_t)))); const __m512i vi24xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (i24 + 16))); const __m512i vk24xGHIJKLMNOPQRSTUV = _mm512_cvtepi8_epi32(_mm_load_si128((const __m128i*) ((uintptr_t) w + 32 * sizeof(int32_t) + 784 * sizeof(int8_t)))); i24 += 32; vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi24x0123456789ABCDEF, vk24x0123456789ABCDEF)); vaccGHIJKLMNOPQRSTUV = _mm512_add_epi32(vaccGHIJKLMNOPQRSTUV, _mm512_mullo_epi32(vi24xGHIJKLMNOPQRSTUV, vk24xGHIJKLMNOPQRSTUV)); w = (const void*) ((uintptr_t) w + 32 * sizeof(int32_t) + 800 * sizeof(int8_t)); __m512 vscaled0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0123456789ABCDEF); __m512 vscaledGHIJKLMNOPQRSTUV = _mm512_cvtepi32_ps(vaccGHIJKLMNOPQRSTUV); vscaled0123456789ABCDEF = _mm512_mul_ps(vscaled0123456789ABCDEF, vscale); vscaledGHIJKLMNOPQRSTUV = _mm512_mul_ps(vscaledGHIJKLMNOPQRSTUV, vscale); vscaled0123456789ABCDEF = _mm512_min_ps(vscaled0123456789ABCDEF, voutput_max_less_zero_point); vscaledGHIJKLMNOPQRSTUV = _mm512_min_ps(vscaledGHIJKLMNOPQRSTUV, voutput_max_less_zero_point); vacc0123456789ABCDEF = _mm512_cvtps_epi32(vscaled0123456789ABCDEF); vaccGHIJKLMNOPQRSTUV = _mm512_cvtps_epi32(vscaledGHIJKLMNOPQRSTUV); __m512i vout0123GHIJ4567KLMN89ABOPQRCDEFSTUV = _mm512_adds_epi16(_mm512_packs_epi32(vacc0123456789ABCDEF, vaccGHIJKLMNOPQRSTUV), voutput_zero_point); __m256i voutGHIJOPQRKLMNSTUV = _mm256_adds_epi16(_mm256_packs_epi32(_mm512_castsi512_si256(vaccGHIJKLMNOPQRSTUV), _mm512_extracti32x8_epi32(vaccGHIJKLMNOPQRSTUV, 1)), _mm512_castsi512_si256(voutput_zero_point)); const __m256i vout0123GHIJ4567KLMN = _mm512_castsi512_si256(vout0123GHIJ4567KLMN89ABOPQRCDEFSTUV); const __m256i vout89ABOPQRCDEFSTUV = _mm512_extracti32x8_epi32(vout0123GHIJ4567KLMN89ABOPQRCDEFSTUV, 1); const __m256i vout0123GHIJ89ABOPQR4567KLMNCDEFSTUV = _mm256_packs_epi16(vout0123GHIJ4567KLMN, vout89ABOPQRCDEFSTUV); __m256i vout0123456789ABCDEFGHIJKLMNOPQRSTUV = _mm256_permutevar8x32_epi32(vout0123GHIJ89ABOPQR4567KLMNCDEFSTUV, vpermute_mask); const __m128i voutGHIJOPQR = _mm256_castsi256_si128(voutGHIJOPQRKLMNSTUV); const __m128i voutKLMNSTUV = _mm256_extracti128_si256(voutGHIJOPQRKLMNSTUV, 1); __m128i voutGHIJKLMNOPQRSTUV = _mm_shuffle_epi32(_mm_packs_epi16(voutGHIJOPQR, voutKLMNSTUV), _MM_SHUFFLE(3, 1, 2, 0)); vout0123456789ABCDEFGHIJKLMNOPQRSTUV = _mm256_max_epi8(vout0123456789ABCDEFGHIJKLMNOPQRSTUV, voutput_min); voutGHIJKLMNOPQRSTUV = _mm_max_epi8(voutGHIJKLMNOPQRSTUV, _mm256_castsi256_si128(voutput_min)); _mm256_storeu_si256((__m256i*) output, vout0123456789ABCDEFGHIJKLMNOPQRSTUV); _mm_storeu_si128((__m128i*) (output + 16), voutGHIJKLMNOPQRSTUV); output += 32; } if XNN_UNLIKELY(c != 0) { // Prepare mask for valid 8-bit elements (depends on nc). const __mmask16 vmask = _cvtu32_mask16((uint32_t) ((UINT32_C(1) << (c & 15)) - UINT32_C(1))); const int8_t* k = (const int8_t*) ((uintptr_t) w + 32 * sizeof(int32_t)); do { __m512i vacc0123456789ABCDEF = _mm512_loadu_si512(w); const __m512i vi0x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i0)); const __m512i vk0x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) k)); i0 += 16; vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi0x0123456789ABCDEF, vk0x0123456789ABCDEF)); const __m512i vi1x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i1)); const __m512i vk1x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (k + 32))); i1 += 16; vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi1x0123456789ABCDEF, vk1x0123456789ABCDEF)); const __m512i vi2x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i2)); const __m512i vk2x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (k + 64))); i2 += 16; vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi2x0123456789ABCDEF, vk2x0123456789ABCDEF)); const __m512i vi3x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i3)); const __m512i vk3x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (k + 96))); i3 += 16; vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi3x0123456789ABCDEF, vk3x0123456789ABCDEF)); const __m512i vi4x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i4)); const __m512i vk4x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (k + 128))); i4 += 16; vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi4x0123456789ABCDEF, vk4x0123456789ABCDEF)); const __m512i vi5x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i5)); const __m512i vk5x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (k + 160))); i5 += 16; vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi5x0123456789ABCDEF, vk5x0123456789ABCDEF)); const __m512i vi6x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i6)); const __m512i vk6x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (k + 192))); i6 += 16; vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi6x0123456789ABCDEF, vk6x0123456789ABCDEF)); const __m512i vi7x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i7)); const __m512i vk7x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (k + 224))); i7 += 16; vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi7x0123456789ABCDEF, vk7x0123456789ABCDEF)); const __m512i vi8x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i8)); const __m512i vk8x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (k + 256))); i8 += 16; vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi8x0123456789ABCDEF, vk8x0123456789ABCDEF)); const __m512i vi9x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i9)); const __m512i vk9x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (k + 288))); i9 += 16; vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi9x0123456789ABCDEF, vk9x0123456789ABCDEF)); const __m512i vi10x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i10)); const __m512i vk10x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (k + 320))); i10 += 16; vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi10x0123456789ABCDEF, vk10x0123456789ABCDEF)); const __m512i vi11x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i11)); const __m512i vk11x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (k + 352))); i11 += 16; vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi11x0123456789ABCDEF, vk11x0123456789ABCDEF)); const __m512i vi12x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i12)); const __m512i vk12x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (k + 384))); i12 += 16; vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi12x0123456789ABCDEF, vk12x0123456789ABCDEF)); const __m512i vi13x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i13)); const __m512i vk13x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (k + 416))); i13 += 16; vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi13x0123456789ABCDEF, vk13x0123456789ABCDEF)); const __m512i vi14x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i14)); const __m512i vk14x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (k + 448))); i14 += 16; vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi14x0123456789ABCDEF, vk14x0123456789ABCDEF)); const __m512i vi15x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i15)); const __m512i vk15x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (k + 480))); i15 += 16; vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi15x0123456789ABCDEF, vk15x0123456789ABCDEF)); const __m512i vi16x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i16)); const __m512i vk16x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (k + 512))); i16 += 16; vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi16x0123456789ABCDEF, vk16x0123456789ABCDEF)); const __m512i vi17x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i17)); const __m512i vk17x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (k + 544))); i17 += 16; vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi17x0123456789ABCDEF, vk17x0123456789ABCDEF)); const __m512i vi18x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i18)); const __m512i vk18x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (k + 576))); i18 += 16; vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi18x0123456789ABCDEF, vk18x0123456789ABCDEF)); const __m512i vi19x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i19)); const __m512i vk19x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (k + 608))); i19 += 16; vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi19x0123456789ABCDEF, vk19x0123456789ABCDEF)); const __m512i vi20x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i20)); const __m512i vk20x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (k + 640))); i20 += 16; vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi20x0123456789ABCDEF, vk20x0123456789ABCDEF)); const __m512i vi21x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i21)); const __m512i vk21x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (k + 672))); i21 += 16; vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi21x0123456789ABCDEF, vk21x0123456789ABCDEF)); const __m512i vi22x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i22)); const __m512i vk22x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (k + 704))); i22 += 16; vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi22x0123456789ABCDEF, vk22x0123456789ABCDEF)); const __m512i vi23x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i23)); const __m512i vk23x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (k + 736))); i23 += 16; vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi23x0123456789ABCDEF, vk23x0123456789ABCDEF)); const __m512i vi24x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) i24)); const __m512i vk24x0123456789ABCDEF = _mm512_cvtepi8_epi32(_mm_loadu_si128((const __m128i*) (k + 768))); i24 += 16; vacc0123456789ABCDEF = _mm512_add_epi32(vacc0123456789ABCDEF, _mm512_mullo_epi32(vi24x0123456789ABCDEF, vk24x0123456789ABCDEF)); k += 16; __m512 vscaled0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0123456789ABCDEF); vscaled0123456789ABCDEF = _mm512_mul_ps(vscaled0123456789ABCDEF, vscale); vscaled0123456789ABCDEF = _mm512_min_ps(vscaled0123456789ABCDEF, voutput_max_less_zero_point); vacc0123456789ABCDEF = _mm512_cvtps_epi32(vscaled0123456789ABCDEF); w = (const void*) ((uintptr_t) w + 16 * sizeof(int32_t)); __m256i vout012389AB4567CDEF = _mm256_adds_epi16(_mm256_packs_epi32(_mm512_castsi512_si256(vacc0123456789ABCDEF), _mm512_extracti32x8_epi32(vacc0123456789ABCDEF, 1)), _mm512_castsi512_si256(voutput_zero_point)); const __m128i vout012389AB = _mm256_castsi256_si128(vout012389AB4567CDEF); const __m128i vout4567CDEF = _mm256_extracti128_si256(vout012389AB4567CDEF, 1); __m128i vout0123456789ABCDEF = _mm_shuffle_epi32(_mm_packs_epi16(vout012389AB, vout4567CDEF), _MM_SHUFFLE(3, 1, 2, 0)); vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, _mm256_castsi256_si128(voutput_min)); if XNN_LIKELY(c >= 16) { _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF); output += 16; c -= 16; } else { _mm_mask_storeu_epi8(output, vmask, vout0123456789ABCDEF); output = (int8_t*) ((uintptr_t) output + c); c = 0; } } while (c != 0); } output = (int8_t*) ((uintptr_t) output + output_increment); } while (--output_width != 0); }