// 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_qb4w_gemm_minmax_ukernel_2x8c8__avx2( 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_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 <= 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 * sizeof(int8_t)); size_t bl = params->scalar.blocksize; assert(bl <= round_up_po2(kc, 16)); assert(bl != 0); assert(bl % 32 == 0); 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 __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 __m128 vinit0 = _mm_load_ss(&((const float*) w)[0]); const __m128 vinit1 = _mm_load_ss(&((const float*) w)[1]); const __m256 vinit01 = _mm256_insertf128_ps(_mm256_castps128_ps256(vinit0), vinit1, 1); const __m128 vinit2 = _mm_load_ss(&((const float*) w)[2]); const __m128 vinit3 = _mm_load_ss(&((const float*) w)[3]); const __m256 vinit23 = _mm256_insertf128_ps(_mm256_castps128_ps256(vinit2), vinit3, 1); const __m128 vinit4 = _mm_load_ss(&((const float*) w)[4]); const __m128 vinit5 = _mm_load_ss(&((const float*) w)[5]); const __m256 vinit45 = _mm256_insertf128_ps(_mm256_castps128_ps256(vinit4), vinit5, 1); const __m128 vinit6 = _mm_load_ss(&((const float*) w)[6]); const __m128 vinit7 = _mm_load_ss(&((const float*) w)[7]); const __m256 vinit67 = _mm256_insertf128_ps(_mm256_castps128_ps256(vinit6), vinit7, 1); const __m256 vinput_zero_point0 = _mm256_set1_ps((float) quantization_params[0].zero_point); __m256 vout0x01 = _mm256_mul_ps(vinit01, vinput_zero_point0); __m256 vout0x23 = _mm256_mul_ps(vinit23, vinput_zero_point0); __m256 vout0x45 = _mm256_mul_ps(vinit45, vinput_zero_point0); __m256 vout0x67 = _mm256_mul_ps(vinit67, vinput_zero_point0); const __m256 vinput_zero_point1 = _mm256_set1_ps((float) quantization_params[1].zero_point); __m256 vout1x01 = _mm256_mul_ps(vinit01, vinput_zero_point1); __m256 vout1x23 = _mm256_mul_ps(vinit23, vinput_zero_point1); __m256 vout1x45 = _mm256_mul_ps(vinit45, vinput_zero_point1); __m256 vout1x67 = _mm256_mul_ps(vinit67, vinput_zero_point1); w = (const int32_t*) w + 8; for (size_t kb=0; kb < kc; kb += bl) { __m256i vacc0x01 = _mm256_setzero_si256(); __m256i vacc0x23 = _mm256_setzero_si256(); __m256i vacc0x45 = _mm256_setzero_si256(); __m256i vacc0x67 = _mm256_setzero_si256(); __m256i vacc1x01 = _mm256_setzero_si256(); __m256i vacc1x23 = _mm256_setzero_si256(); __m256i vacc1x45 = _mm256_setzero_si256(); __m256i vacc1x67 = _mm256_setzero_si256(); size_t k = bl; 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 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)); __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)); __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)); __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)); 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; 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)); 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)); 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)); 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)); 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 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)); 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)); 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)); 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)); w = (const int8_t*) w + 64; k -= 8 * sizeof(int8_t); } const __m128 vfilter_output_scale0 = _mm_castsi128_ps(_mm_set1_epi32((uint32_t) ((const uint16_t*) w)[0] << 16)); const __m128 vfilter_output_scale1 = _mm_castsi128_ps(_mm_set1_epi32((uint32_t) ((const uint16_t*) w)[1] << 16)); const __m256 vfilter_output_scale01 = _mm256_insertf128_ps( _mm256_castps128_ps256(vfilter_output_scale0), vfilter_output_scale1, 1); vout0x01 = _mm256_fmadd_ps(_mm256_cvtepi32_ps(vacc0x01), vfilter_output_scale01, vout0x01); vout1x01 = _mm256_fmadd_ps(_mm256_cvtepi32_ps(vacc1x01), vfilter_output_scale01, vout1x01); const __m128 vfilter_output_scale2 = _mm_castsi128_ps(_mm_set1_epi32((uint32_t) ((const uint16_t*) w)[2] << 16)); const __m128 vfilter_output_scale3 = _mm_castsi128_ps(_mm_set1_epi32((uint32_t) ((const uint16_t*) w)[3] << 16)); const __m256 vfilter_output_scale23 = _mm256_insertf128_ps( _mm256_castps128_ps256(vfilter_output_scale2), vfilter_output_scale3, 1); vout0x23 = _mm256_fmadd_ps(_mm256_cvtepi32_ps(vacc0x23), vfilter_output_scale23, vout0x23); vout1x23 = _mm256_fmadd_ps(_mm256_cvtepi32_ps(vacc1x23), vfilter_output_scale23, vout1x23); const __m128 vfilter_output_scale4 = _mm_castsi128_ps(_mm_set1_epi32((uint32_t) ((const uint16_t*) w)[4] << 16)); const __m128 vfilter_output_scale5 = _mm_castsi128_ps(_mm_set1_epi32((uint32_t) ((const uint16_t*) w)[5] << 16)); const __m256 vfilter_output_scale45 = _mm256_insertf128_ps( _mm256_castps128_ps256(vfilter_output_scale4), vfilter_output_scale5, 1); vout0x45 = _mm256_fmadd_ps(_mm256_cvtepi32_ps(vacc0x45), vfilter_output_scale45, vout0x45); vout1x45 = _mm256_fmadd_ps(_mm256_cvtepi32_ps(vacc1x45), vfilter_output_scale45, vout1x45); const __m128 vfilter_output_scale6 = _mm_castsi128_ps(_mm_set1_epi32((uint32_t) ((const uint16_t*) w)[6] << 16)); const __m128 vfilter_output_scale7 = _mm_castsi128_ps(_mm_set1_epi32((uint32_t) ((const uint16_t*) w)[7] << 16)); const __m256 vfilter_output_scale67 = _mm256_insertf128_ps( _mm256_castps128_ps256(vfilter_output_scale6), vfilter_output_scale7, 1); vout0x67 = _mm256_fmadd_ps(_mm256_cvtepi32_ps(vacc0x67), vfilter_output_scale67, vout0x67); vout1x67 = _mm256_fmadd_ps(_mm256_cvtepi32_ps(vacc1x67), vfilter_output_scale67, vout1x67); w = (const uint16_t*) w + 8; } const __m256 vout0x0213 = _mm256_hadd_ps(vout0x01, vout0x23); const __m256 vout0x4657 = _mm256_hadd_ps(vout0x45, vout0x67); const __m256 vout1x0213 = _mm256_hadd_ps(vout1x01, vout1x23); const __m256 vout1x4657 = _mm256_hadd_ps(vout1x45, vout1x67); const __m256 vout0x02461357 = _mm256_hadd_ps(vout0x0213, vout0x4657); const __m256 vout1x02461357 = _mm256_hadd_ps(vout1x0213, vout1x4657); const __m256i vpermute_mask = _mm256_set_epi32(7, 3, 6, 2, 5, 1, 4, 0); __m256 vout0x01234567 = _mm256_permutevar8x32_ps(vout0x02461357, vpermute_mask); __m256 vout1x01234567 = _mm256_permutevar8x32_ps(vout1x02461357, vpermute_mask); const __m256 vinput_scale0 = _mm256_broadcast_ss(&quantization_params[0].inv_scale); const __m256 vinput_scale1 = _mm256_broadcast_ss(&quantization_params[1].inv_scale); const __m256 vbias01234567 = _mm256_loadu_ps((const float*) w); w = (const float*) w + 8; vout0x01234567 = _mm256_fmadd_ps(vout0x01234567, vinput_scale0, vbias01234567); vout1x01234567 = _mm256_fmadd_ps(vout1x01234567, vinput_scale1, vbias01234567); vout0x01234567 = _mm256_max_ps(vout0x01234567, vmin); vout1x01234567 = _mm256_max_ps(vout1x01234567, vmin); vout0x01234567 = _mm256_min_ps(vout0x01234567, vmax); vout1x01234567 = _mm256_min_ps(vout1x01234567, vmax); __m128i vfp16out0x01234567 = _mm256_cvtps_ph(vout0x01234567, _MM_FROUND_TO_NEAREST_INT); __m128i vfp16out1x01234567 = _mm256_cvtps_ph(vout1x01234567, _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); a0 = (const int8_t*) ((uintptr_t) a0 - kc); a1 = (const int8_t*) ((uintptr_t) a1 - kc); nc -= 8; } else { if (nc & 4) { _mm_storel_epi64((__m128i*) c0, vfp16out0x01234567); _mm_storel_epi64((__m128i*) c1, vfp16out1x01234567); vfp16out0x01234567 = _mm_unpackhi_epi64(vfp16out0x01234567, vfp16out0x01234567); vfp16out1x01234567 = _mm_unpackhi_epi64(vfp16out1x01234567, vfp16out1x01234567); c0 += 4; c1 += 4; } if (nc & 2) { _mm_storeu_si32(c0, vfp16out0x01234567); _mm_storeu_si32(c1, vfp16out1x01234567); vfp16out0x01234567 = _mm_srli_epi64(vfp16out0x01234567, 32); vfp16out1x01234567 = _mm_srli_epi64(vfp16out1x01234567, 32); c0 += 2; c1 += 2; } if (nc & 1) { *c0 = (uint16_t) _mm_extract_epi16(vfp16out0x01234567, 0); *c1 = (uint16_t) _mm_extract_epi16(vfp16out1x01234567, 0); } nc = 0; } } while (nc != 0); }