// Auto-generated file. Do not edit! // Template: src/qs8-gemm/MRx16c8-avx512vnni.c.in // Generator: tools/xngen // // Copyright 2023 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_f32_qb4w_gemm_minmax_ukernel_9x16c8__avx512vnni( size_t mr, size_t nc, size_t kc, const int8_t* restrict a, size_t a_stride, const void* restrict w, float* restrict c, size_t cm_stride, size_t cn_stride, const struct xnn_f32_qb4w_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 <= 9); 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; float* c0 = c; const int8_t* a1 = (const int8_t*) ((uintptr_t) a0 + a_stride); float* c1 = (float*) ((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); float* c2 = (float*) ((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); float* c3 = (float*) ((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); float* c4 = (float*) ((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); float* c5 = (float*) ((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); float* c6 = (float*) ((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); float* c7 = (float*) ((uintptr_t) c6 + cm_stride); if XNN_UNPREDICTABLE(mr < 8) { a7 = a6; c7 = c6; } const int8_t* a8 = (const int8_t*) ((uintptr_t) a7 + a_stride); float* c8 = (float*) ((uintptr_t) c7 + cm_stride); if XNN_UNPREDICTABLE(mr <= 8) { a8 = a7; c8 = c7; } size_t bl = params->scalar.blocksize; assert(bl != 0); assert(bl <= kc); assert(kc % bl == 0); assert(bl % 32 == 0); const __m512i vsign_mask = _mm512_set1_epi8(0x80); XNN_FORCE_REALIZATION(vsign_mask); const __m512 vinput_zero_point0 = _mm512_set1_ps((float) quantization_params[0].zero_point + 128); const __m512 vinput_zero_point1 = _mm512_set1_ps((float) quantization_params[1].zero_point + 128); const __m512 vinput_zero_point2 = _mm512_set1_ps((float) quantization_params[2].zero_point + 128); const __m512 vinput_zero_point3 = _mm512_set1_ps((float) quantization_params[3].zero_point + 128); const __m512 vinput_zero_point4 = _mm512_set1_ps((float) quantization_params[4].zero_point + 128); const __m512 vinput_zero_point5 = _mm512_set1_ps((float) quantization_params[5].zero_point + 128); const __m512 vinput_zero_point6 = _mm512_set1_ps((float) quantization_params[6].zero_point + 128); const __m512 vinput_zero_point7 = _mm512_set1_ps((float) quantization_params[7].zero_point + 128); const __m512 vinput_zero_point8 = _mm512_set1_ps((float) quantization_params[8].zero_point + 128); const __m512 voutput_min = _mm512_set1_ps(params->scalar.min); const __m512 voutput_max = _mm512_set1_ps(params->scalar.max); // XNN_FORCE_REALIZATION(voutput_min); // XNN_FORCE_REALIZATION(voutput_max); const __m512i vmask = _mm512_set1_epi8(0xF0); XNN_FORCE_REALIZATION(vmask); do { const __m512 vksum0123456789ABCDEF = _mm512_loadu_ps(w); __m512 vscaled0x0123456789ABCDEF = _mm512_mul_ps(vksum0123456789ABCDEF, vinput_zero_point0); __m512 vscaled1x0123456789ABCDEF = _mm512_mul_ps(vksum0123456789ABCDEF, vinput_zero_point1); __m512 vscaled2x0123456789ABCDEF = _mm512_mul_ps(vksum0123456789ABCDEF, vinput_zero_point2); __m512 vscaled3x0123456789ABCDEF = _mm512_mul_ps(vksum0123456789ABCDEF, vinput_zero_point3); __m512 vscaled4x0123456789ABCDEF = _mm512_mul_ps(vksum0123456789ABCDEF, vinput_zero_point4); __m512 vscaled5x0123456789ABCDEF = _mm512_mul_ps(vksum0123456789ABCDEF, vinput_zero_point5); __m512 vscaled6x0123456789ABCDEF = _mm512_mul_ps(vksum0123456789ABCDEF, vinput_zero_point6); __m512 vscaled7x0123456789ABCDEF = _mm512_mul_ps(vksum0123456789ABCDEF, vinput_zero_point7); __m512 vscaled8x0123456789ABCDEF = _mm512_mul_ps(vksum0123456789ABCDEF, vinput_zero_point8); w = (const int32_t*) w + 16; for (size_t kb=0; kb < kc; kb+=bl) { __m512i vacc0x01234567 = _mm512_setzero_epi32(); __m512i vacc0x89ABCDEF = _mm512_setzero_epi32(); __m512i vacc1x01234567 = _mm512_setzero_epi32(); __m512i vacc1x89ABCDEF = _mm512_setzero_epi32(); __m512i vacc2x01234567 = _mm512_setzero_epi32(); __m512i vacc2x89ABCDEF = _mm512_setzero_epi32(); __m512i vacc3x01234567 = _mm512_setzero_epi32(); __m512i vacc3x89ABCDEF = _mm512_setzero_epi32(); __m512i vacc4x01234567 = _mm512_setzero_epi32(); __m512i vacc4x89ABCDEF = _mm512_setzero_epi32(); __m512i vacc5x01234567 = _mm512_setzero_epi32(); __m512i vacc5x89ABCDEF = _mm512_setzero_epi32(); __m512i vacc6x01234567 = _mm512_setzero_epi32(); __m512i vacc6x89ABCDEF = _mm512_setzero_epi32(); __m512i vacc7x01234567 = _mm512_setzero_epi32(); __m512i vacc7x89ABCDEF = _mm512_setzero_epi32(); __m512i vacc8x01234567 = _mm512_setzero_epi32(); __m512i vacc8x89ABCDEF = _mm512_setzero_epi32(); size_t k = bl; while (k >= 16 * sizeof(int8_t)) { const __m512i va0x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a0)), vsign_mask); const __m512i va0x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a0 + 8)), vsign_mask); a0 += 16; const __m512i va1x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a1)), vsign_mask); const __m512i va1x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a1 + 8)), vsign_mask); a1 += 16; const __m512i va2x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a2)), vsign_mask); const __m512i va2x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a2 + 8)), vsign_mask); a2 += 16; const __m512i va3x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a3)), vsign_mask); const __m512i va3x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a3 + 8)), vsign_mask); a3 += 16; const __m512i va4x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a4)), vsign_mask); const __m512i va4x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a4 + 8)), vsign_mask); a4 += 16; const __m512i va5x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a5)), vsign_mask); const __m512i va5x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a5 + 8)), vsign_mask); a5 += 16; const __m512i va6x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a6)), vsign_mask); const __m512i va6x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a6 + 8)), vsign_mask); a6 += 16; const __m512i va7x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a7)), vsign_mask); const __m512i va7x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a7 + 8)), vsign_mask); a7 += 16; const __m512i va8x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a8)), vsign_mask); const __m512i va8x89ABCDEF = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a8 + 8)), vsign_mask); a8 += 16; const __m512i vbb01234567x01234567 = _mm512_loadu_si512(w); const __m512i vbb89ABCDEFx01234567 = _mm512_loadu_si512((const int8_t*) w + 64); const __m512i vbs01234567x01234567 = _mm512_slli_epi32(vbb01234567x01234567, 4); const __m512i vbs89ABCDEFx01234567 = _mm512_slli_epi32(vbb89ABCDEFx01234567, 4); const __m512i vb01234567x89ABCDEF = _mm512_and_si512(vbb01234567x01234567, vmask); const __m512i vb89ABCDEFx89ABCDEF = _mm512_and_si512(vbb89ABCDEFx01234567, vmask); const __m512i vb01234567x01234567 = _mm512_and_si512(vbs01234567x01234567, vmask); const __m512i vb89ABCDEFx01234567 = _mm512_and_si512(vbs89ABCDEFx01234567, vmask); vacc0x01234567 = _mm512_dpbusd_epi32(vacc0x01234567, va0x01234567, vb01234567x01234567); vacc0x89ABCDEF = _mm512_dpbusd_epi32(vacc0x89ABCDEF, va0x01234567, vb89ABCDEFx01234567); vacc1x01234567 = _mm512_dpbusd_epi32(vacc1x01234567, va1x01234567, vb01234567x01234567); vacc1x89ABCDEF = _mm512_dpbusd_epi32(vacc1x89ABCDEF, va1x01234567, vb89ABCDEFx01234567); vacc2x01234567 = _mm512_dpbusd_epi32(vacc2x01234567, va2x01234567, vb01234567x01234567); vacc2x89ABCDEF = _mm512_dpbusd_epi32(vacc2x89ABCDEF, va2x01234567, vb89ABCDEFx01234567); vacc3x01234567 = _mm512_dpbusd_epi32(vacc3x01234567, va3x01234567, vb01234567x01234567); vacc3x89ABCDEF = _mm512_dpbusd_epi32(vacc3x89ABCDEF, va3x01234567, vb89ABCDEFx01234567); vacc4x01234567 = _mm512_dpbusd_epi32(vacc4x01234567, va4x01234567, vb01234567x01234567); vacc4x89ABCDEF = _mm512_dpbusd_epi32(vacc4x89ABCDEF, va4x01234567, vb89ABCDEFx01234567); vacc5x01234567 = _mm512_dpbusd_epi32(vacc5x01234567, va5x01234567, vb01234567x01234567); vacc5x89ABCDEF = _mm512_dpbusd_epi32(vacc5x89ABCDEF, va5x01234567, vb89ABCDEFx01234567); vacc6x01234567 = _mm512_dpbusd_epi32(vacc6x01234567, va6x01234567, vb01234567x01234567); vacc6x89ABCDEF = _mm512_dpbusd_epi32(vacc6x89ABCDEF, va6x01234567, vb89ABCDEFx01234567); vacc7x01234567 = _mm512_dpbusd_epi32(vacc7x01234567, va7x01234567, vb01234567x01234567); vacc7x89ABCDEF = _mm512_dpbusd_epi32(vacc7x89ABCDEF, va7x01234567, vb89ABCDEFx01234567); vacc8x01234567 = _mm512_dpbusd_epi32(vacc8x01234567, va8x01234567, vb01234567x01234567); vacc8x89ABCDEF = _mm512_dpbusd_epi32(vacc8x89ABCDEF, va8x01234567, vb89ABCDEFx01234567); vacc0x01234567 = _mm512_dpbusd_epi32(vacc0x01234567, va0x89ABCDEF, vb01234567x89ABCDEF); vacc0x89ABCDEF = _mm512_dpbusd_epi32(vacc0x89ABCDEF, va0x89ABCDEF, vb89ABCDEFx89ABCDEF); vacc1x01234567 = _mm512_dpbusd_epi32(vacc1x01234567, va1x89ABCDEF, vb01234567x89ABCDEF); vacc1x89ABCDEF = _mm512_dpbusd_epi32(vacc1x89ABCDEF, va1x89ABCDEF, vb89ABCDEFx89ABCDEF); vacc2x01234567 = _mm512_dpbusd_epi32(vacc2x01234567, va2x89ABCDEF, vb01234567x89ABCDEF); vacc2x89ABCDEF = _mm512_dpbusd_epi32(vacc2x89ABCDEF, va2x89ABCDEF, vb89ABCDEFx89ABCDEF); vacc3x01234567 = _mm512_dpbusd_epi32(vacc3x01234567, va3x89ABCDEF, vb01234567x89ABCDEF); vacc3x89ABCDEF = _mm512_dpbusd_epi32(vacc3x89ABCDEF, va3x89ABCDEF, vb89ABCDEFx89ABCDEF); vacc4x01234567 = _mm512_dpbusd_epi32(vacc4x01234567, va4x89ABCDEF, vb01234567x89ABCDEF); vacc4x89ABCDEF = _mm512_dpbusd_epi32(vacc4x89ABCDEF, va4x89ABCDEF, vb89ABCDEFx89ABCDEF); vacc5x01234567 = _mm512_dpbusd_epi32(vacc5x01234567, va5x89ABCDEF, vb01234567x89ABCDEF); vacc5x89ABCDEF = _mm512_dpbusd_epi32(vacc5x89ABCDEF, va5x89ABCDEF, vb89ABCDEFx89ABCDEF); vacc6x01234567 = _mm512_dpbusd_epi32(vacc6x01234567, va6x89ABCDEF, vb01234567x89ABCDEF); vacc6x89ABCDEF = _mm512_dpbusd_epi32(vacc6x89ABCDEF, va6x89ABCDEF, vb89ABCDEFx89ABCDEF); vacc7x01234567 = _mm512_dpbusd_epi32(vacc7x01234567, va7x89ABCDEF, vb01234567x89ABCDEF); vacc7x89ABCDEF = _mm512_dpbusd_epi32(vacc7x89ABCDEF, va7x89ABCDEF, vb89ABCDEFx89ABCDEF); vacc8x01234567 = _mm512_dpbusd_epi32(vacc8x01234567, va8x89ABCDEF, vb01234567x89ABCDEF); vacc8x89ABCDEF = _mm512_dpbusd_epi32(vacc8x89ABCDEF, va8x89ABCDEF, vb89ABCDEFx89ABCDEF); w = (const int8_t*) w + 128; k -= 16 * sizeof(int8_t); } if (k != 0) { const __m512i va0x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a0)), vsign_mask); a0 += 8; const __m512i va1x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a1)), vsign_mask); a1 += 8; const __m512i va2x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a2)), vsign_mask); a2 += 8; const __m512i va3x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a3)), vsign_mask); a3 += 8; const __m512i va4x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a4)), vsign_mask); a4 += 8; const __m512i va5x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a5)), vsign_mask); a5 += 8; const __m512i va6x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a6)), vsign_mask); a6 += 8; const __m512i va7x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a7)), vsign_mask); a7 += 8; const __m512i va8x01234567 = _mm512_xor_epi64(_mm512_set1_epi64((int64_t) unaligned_load_u64(a8)), vsign_mask); a8 += 8; const __m512i vbb01234567x01234567 = _mm512_load_si512(w); const __m512i vbb89ABCDEFx01234567 = _mm512_load_si512((const int8_t*) w + 64); const __m512i vb01234567x01234567 = _mm512_slli_epi32(vbb01234567x01234567, 4); const __m512i vb89ABCDEFx01234567 = _mm512_slli_epi32(vbb89ABCDEFx01234567, 4); vacc0x01234567 = _mm512_dpbusd_epi32(vacc0x01234567, va0x01234567, vb01234567x01234567); vacc0x89ABCDEF = _mm512_dpbusd_epi32(vacc0x89ABCDEF, va0x01234567, vb89ABCDEFx01234567); vacc1x01234567 = _mm512_dpbusd_epi32(vacc1x01234567, va1x01234567, vb01234567x01234567); vacc1x89ABCDEF = _mm512_dpbusd_epi32(vacc1x89ABCDEF, va1x01234567, vb89ABCDEFx01234567); vacc2x01234567 = _mm512_dpbusd_epi32(vacc2x01234567, va2x01234567, vb01234567x01234567); vacc2x89ABCDEF = _mm512_dpbusd_epi32(vacc2x89ABCDEF, va2x01234567, vb89ABCDEFx01234567); vacc3x01234567 = _mm512_dpbusd_epi32(vacc3x01234567, va3x01234567, vb01234567x01234567); vacc3x89ABCDEF = _mm512_dpbusd_epi32(vacc3x89ABCDEF, va3x01234567, vb89ABCDEFx01234567); vacc4x01234567 = _mm512_dpbusd_epi32(vacc4x01234567, va4x01234567, vb01234567x01234567); vacc4x89ABCDEF = _mm512_dpbusd_epi32(vacc4x89ABCDEF, va4x01234567, vb89ABCDEFx01234567); vacc5x01234567 = _mm512_dpbusd_epi32(vacc5x01234567, va5x01234567, vb01234567x01234567); vacc5x89ABCDEF = _mm512_dpbusd_epi32(vacc5x89ABCDEF, va5x01234567, vb89ABCDEFx01234567); vacc6x01234567 = _mm512_dpbusd_epi32(vacc6x01234567, va6x01234567, vb01234567x01234567); vacc6x89ABCDEF = _mm512_dpbusd_epi32(vacc6x89ABCDEF, va6x01234567, vb89ABCDEFx01234567); vacc7x01234567 = _mm512_dpbusd_epi32(vacc7x01234567, va7x01234567, vb01234567x01234567); vacc7x89ABCDEF = _mm512_dpbusd_epi32(vacc7x89ABCDEF, va7x01234567, vb89ABCDEFx01234567); vacc8x01234567 = _mm512_dpbusd_epi32(vacc8x01234567, va8x01234567, vb01234567x01234567); vacc8x89ABCDEF = _mm512_dpbusd_epi32(vacc8x89ABCDEF, va8x01234567, vb89ABCDEFx01234567); w = (const int8_t*) w + 128; k -= 8 * sizeof(int8_t); } const __m512 vfilter_output_scale0123456789ABCDEF = _mm512_castsi512_ps(_mm512_slli_epi32( _mm512_cvtepu16_epi32(_mm256_load_si256((const __m256i*) w)), 16)); w = (const uint16_t*) w + 16; // Add adjacent pairs const __m512i vidx = _mm512_set_epi32(30, 28, 26, 24, 22, 20, 18, 16, 14, 12, 10, 8, 6, 4, 2, 0); const __m512i vsum0x01234567 = _mm512_add_epi32(vacc0x01234567, _mm512_srli_epi64(vacc0x01234567, 32)); const __m512i vsum0x89ABCDEF = _mm512_add_epi32(vacc0x89ABCDEF, _mm512_srli_epi64(vacc0x89ABCDEF, 32)); __m512i vacc0x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum0x01234567, vidx, vsum0x89ABCDEF); const __m512i vsum1x01234567 = _mm512_add_epi32(vacc1x01234567, _mm512_srli_epi64(vacc1x01234567, 32)); const __m512i vsum1x89ABCDEF = _mm512_add_epi32(vacc1x89ABCDEF, _mm512_srli_epi64(vacc1x89ABCDEF, 32)); __m512i vacc1x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum1x01234567, vidx, vsum1x89ABCDEF); const __m512i vsum2x01234567 = _mm512_add_epi32(vacc2x01234567, _mm512_srli_epi64(vacc2x01234567, 32)); const __m512i vsum2x89ABCDEF = _mm512_add_epi32(vacc2x89ABCDEF, _mm512_srli_epi64(vacc2x89ABCDEF, 32)); __m512i vacc2x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum2x01234567, vidx, vsum2x89ABCDEF); const __m512i vsum3x01234567 = _mm512_add_epi32(vacc3x01234567, _mm512_srli_epi64(vacc3x01234567, 32)); const __m512i vsum3x89ABCDEF = _mm512_add_epi32(vacc3x89ABCDEF, _mm512_srli_epi64(vacc3x89ABCDEF, 32)); __m512i vacc3x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum3x01234567, vidx, vsum3x89ABCDEF); const __m512i vsum4x01234567 = _mm512_add_epi32(vacc4x01234567, _mm512_srli_epi64(vacc4x01234567, 32)); const __m512i vsum4x89ABCDEF = _mm512_add_epi32(vacc4x89ABCDEF, _mm512_srli_epi64(vacc4x89ABCDEF, 32)); __m512i vacc4x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum4x01234567, vidx, vsum4x89ABCDEF); const __m512i vsum5x01234567 = _mm512_add_epi32(vacc5x01234567, _mm512_srli_epi64(vacc5x01234567, 32)); const __m512i vsum5x89ABCDEF = _mm512_add_epi32(vacc5x89ABCDEF, _mm512_srli_epi64(vacc5x89ABCDEF, 32)); __m512i vacc5x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum5x01234567, vidx, vsum5x89ABCDEF); const __m512i vsum6x01234567 = _mm512_add_epi32(vacc6x01234567, _mm512_srli_epi64(vacc6x01234567, 32)); const __m512i vsum6x89ABCDEF = _mm512_add_epi32(vacc6x89ABCDEF, _mm512_srli_epi64(vacc6x89ABCDEF, 32)); __m512i vacc6x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum6x01234567, vidx, vsum6x89ABCDEF); const __m512i vsum7x01234567 = _mm512_add_epi32(vacc7x01234567, _mm512_srli_epi64(vacc7x01234567, 32)); const __m512i vsum7x89ABCDEF = _mm512_add_epi32(vacc7x89ABCDEF, _mm512_srli_epi64(vacc7x89ABCDEF, 32)); __m512i vacc7x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum7x01234567, vidx, vsum7x89ABCDEF); const __m512i vsum8x01234567 = _mm512_add_epi32(vacc8x01234567, _mm512_srli_epi64(vacc8x01234567, 32)); const __m512i vsum8x89ABCDEF = _mm512_add_epi32(vacc8x89ABCDEF, _mm512_srli_epi64(vacc8x89ABCDEF, 32)); __m512i vacc8x0123456789ABCDEF = _mm512_permutex2var_epi32(vsum8x01234567, vidx, vsum8x89ABCDEF); __m512 vf0x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc0x0123456789ABCDEF); __m512 vf1x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc1x0123456789ABCDEF); __m512 vf2x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc2x0123456789ABCDEF); __m512 vf3x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc3x0123456789ABCDEF); __m512 vf4x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc4x0123456789ABCDEF); __m512 vf5x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc5x0123456789ABCDEF); __m512 vf6x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc6x0123456789ABCDEF); __m512 vf7x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc7x0123456789ABCDEF); __m512 vf8x0123456789ABCDEF = _mm512_cvtepi32_ps(vacc8x0123456789ABCDEF); vscaled0x0123456789ABCDEF = _mm512_fmadd_ps(vf0x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vscaled0x0123456789ABCDEF); vscaled1x0123456789ABCDEF = _mm512_fmadd_ps(vf1x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vscaled1x0123456789ABCDEF); vscaled2x0123456789ABCDEF = _mm512_fmadd_ps(vf2x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vscaled2x0123456789ABCDEF); vscaled3x0123456789ABCDEF = _mm512_fmadd_ps(vf3x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vscaled3x0123456789ABCDEF); vscaled4x0123456789ABCDEF = _mm512_fmadd_ps(vf4x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vscaled4x0123456789ABCDEF); vscaled5x0123456789ABCDEF = _mm512_fmadd_ps(vf5x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vscaled5x0123456789ABCDEF); vscaled6x0123456789ABCDEF = _mm512_fmadd_ps(vf6x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vscaled6x0123456789ABCDEF); vscaled7x0123456789ABCDEF = _mm512_fmadd_ps(vf7x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vscaled7x0123456789ABCDEF); vscaled8x0123456789ABCDEF = _mm512_fmadd_ps(vf8x0123456789ABCDEF, vfilter_output_scale0123456789ABCDEF, vscaled8x0123456789ABCDEF); } vscaled0x0123456789ABCDEF = _mm512_mul_ps(vscaled0x0123456789ABCDEF, _mm512_set1_ps(quantization_params[0].inv_scale)); vscaled1x0123456789ABCDEF = _mm512_mul_ps(vscaled1x0123456789ABCDEF, _mm512_set1_ps(quantization_params[1].inv_scale)); vscaled2x0123456789ABCDEF = _mm512_mul_ps(vscaled2x0123456789ABCDEF, _mm512_set1_ps(quantization_params[2].inv_scale)); vscaled3x0123456789ABCDEF = _mm512_mul_ps(vscaled3x0123456789ABCDEF, _mm512_set1_ps(quantization_params[3].inv_scale)); vscaled4x0123456789ABCDEF = _mm512_mul_ps(vscaled4x0123456789ABCDEF, _mm512_set1_ps(quantization_params[4].inv_scale)); vscaled5x0123456789ABCDEF = _mm512_mul_ps(vscaled5x0123456789ABCDEF, _mm512_set1_ps(quantization_params[5].inv_scale)); vscaled6x0123456789ABCDEF = _mm512_mul_ps(vscaled6x0123456789ABCDEF, _mm512_set1_ps(quantization_params[6].inv_scale)); vscaled7x0123456789ABCDEF = _mm512_mul_ps(vscaled7x0123456789ABCDEF, _mm512_set1_ps(quantization_params[7].inv_scale)); vscaled8x0123456789ABCDEF = _mm512_mul_ps(vscaled8x0123456789ABCDEF, _mm512_set1_ps(quantization_params[8].inv_scale)); const __m512 vbias0123456789ABCDEF = _mm512_loadu_ps((const float*) w); w = (const float*) w + 16; vscaled0x0123456789ABCDEF = _mm512_add_ps(vscaled0x0123456789ABCDEF, vbias0123456789ABCDEF); vscaled1x0123456789ABCDEF = _mm512_add_ps(vscaled1x0123456789ABCDEF, vbias0123456789ABCDEF); vscaled2x0123456789ABCDEF = _mm512_add_ps(vscaled2x0123456789ABCDEF, vbias0123456789ABCDEF); vscaled3x0123456789ABCDEF = _mm512_add_ps(vscaled3x0123456789ABCDEF, vbias0123456789ABCDEF); vscaled4x0123456789ABCDEF = _mm512_add_ps(vscaled4x0123456789ABCDEF, vbias0123456789ABCDEF); vscaled5x0123456789ABCDEF = _mm512_add_ps(vscaled5x0123456789ABCDEF, vbias0123456789ABCDEF); vscaled6x0123456789ABCDEF = _mm512_add_ps(vscaled6x0123456789ABCDEF, vbias0123456789ABCDEF); vscaled7x0123456789ABCDEF = _mm512_add_ps(vscaled7x0123456789ABCDEF, vbias0123456789ABCDEF); vscaled8x0123456789ABCDEF = _mm512_add_ps(vscaled8x0123456789ABCDEF, vbias0123456789ABCDEF); vscaled0x0123456789ABCDEF = _mm512_max_ps(vscaled0x0123456789ABCDEF, voutput_min); vscaled1x0123456789ABCDEF = _mm512_max_ps(vscaled1x0123456789ABCDEF, voutput_min); vscaled2x0123456789ABCDEF = _mm512_max_ps(vscaled2x0123456789ABCDEF, voutput_min); vscaled3x0123456789ABCDEF = _mm512_max_ps(vscaled3x0123456789ABCDEF, voutput_min); vscaled4x0123456789ABCDEF = _mm512_max_ps(vscaled4x0123456789ABCDEF, voutput_min); vscaled5x0123456789ABCDEF = _mm512_max_ps(vscaled5x0123456789ABCDEF, voutput_min); vscaled6x0123456789ABCDEF = _mm512_max_ps(vscaled6x0123456789ABCDEF, voutput_min); vscaled7x0123456789ABCDEF = _mm512_max_ps(vscaled7x0123456789ABCDEF, voutput_min); vscaled8x0123456789ABCDEF = _mm512_max_ps(vscaled8x0123456789ABCDEF, voutput_min); vscaled0x0123456789ABCDEF = _mm512_min_ps(vscaled0x0123456789ABCDEF, voutput_max); vscaled1x0123456789ABCDEF = _mm512_min_ps(vscaled1x0123456789ABCDEF, voutput_max); vscaled2x0123456789ABCDEF = _mm512_min_ps(vscaled2x0123456789ABCDEF, voutput_max); vscaled3x0123456789ABCDEF = _mm512_min_ps(vscaled3x0123456789ABCDEF, voutput_max); vscaled4x0123456789ABCDEF = _mm512_min_ps(vscaled4x0123456789ABCDEF, voutput_max); vscaled5x0123456789ABCDEF = _mm512_min_ps(vscaled5x0123456789ABCDEF, voutput_max); vscaled6x0123456789ABCDEF = _mm512_min_ps(vscaled6x0123456789ABCDEF, voutput_max); vscaled7x0123456789ABCDEF = _mm512_min_ps(vscaled7x0123456789ABCDEF, voutput_max); vscaled8x0123456789ABCDEF = _mm512_min_ps(vscaled8x0123456789ABCDEF, voutput_max); if XNN_LIKELY(nc >= 16) { _mm512_storeu_ps(c0, vscaled0x0123456789ABCDEF); _mm512_storeu_ps(c1, vscaled1x0123456789ABCDEF); _mm512_storeu_ps(c2, vscaled2x0123456789ABCDEF); _mm512_storeu_ps(c3, vscaled3x0123456789ABCDEF); _mm512_storeu_ps(c4, vscaled4x0123456789ABCDEF); _mm512_storeu_ps(c5, vscaled5x0123456789ABCDEF); _mm512_storeu_ps(c6, vscaled6x0123456789ABCDEF); _mm512_storeu_ps(c7, vscaled7x0123456789ABCDEF); _mm512_storeu_ps(c8, vscaled8x0123456789ABCDEF); a0 = (const int8_t*) ((uintptr_t) a0 - kc); a1 = (const int8_t*) ((uintptr_t) a1 - kc); a2 = (const int8_t*) ((uintptr_t) a2 - kc); a3 = (const int8_t*) ((uintptr_t) a3 - kc); a4 = (const int8_t*) ((uintptr_t) a4 - kc); a5 = (const int8_t*) ((uintptr_t) a5 - kc); a6 = (const int8_t*) ((uintptr_t) a6 - kc); a7 = (const int8_t*) ((uintptr_t) a7 - kc); a8 = (const int8_t*) ((uintptr_t) a8 - kc); c0 = (float*) ((uintptr_t) c0 + cn_stride); c1 = (float*) ((uintptr_t) c1 + cn_stride); c2 = (float*) ((uintptr_t) c2 + cn_stride); c3 = (float*) ((uintptr_t) c3 + cn_stride); c4 = (float*) ((uintptr_t) c4 + cn_stride); c5 = (float*) ((uintptr_t) c5 + cn_stride); c6 = (float*) ((uintptr_t) c6 + cn_stride); c7 = (float*) ((uintptr_t) c7 + cn_stride); c8 = (float*) ((uintptr_t) c8 + cn_stride); nc -= 16; } else { // Prepare mask for valid 32-bit elements (depends on nc). const __mmask16 vmask = _cvtu32_mask16((UINT32_C(1) << nc) - 1); _mm512_mask_storeu_ps(c0, vmask, vscaled0x0123456789ABCDEF); _mm512_mask_storeu_ps(c1, vmask, vscaled1x0123456789ABCDEF); _mm512_mask_storeu_ps(c2, vmask, vscaled2x0123456789ABCDEF); _mm512_mask_storeu_ps(c3, vmask, vscaled3x0123456789ABCDEF); _mm512_mask_storeu_ps(c4, vmask, vscaled4x0123456789ABCDEF); _mm512_mask_storeu_ps(c5, vmask, vscaled5x0123456789ABCDEF); _mm512_mask_storeu_ps(c6, vmask, vscaled6x0123456789ABCDEF); _mm512_mask_storeu_ps(c7, vmask, vscaled7x0123456789ABCDEF); _mm512_mask_storeu_ps(c8, vmask, vscaled8x0123456789ABCDEF); nc = 0; } } while (nc != 0); }