// Auto-generated file. Do not edit! // Template: src/qs8-gemm/MRx4c8-ssevnni.c.in // Generator: tools/xngen // // Copyright 2024 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/gemm.h" #include "xnnpack/intrinsics-polyfill.h" #include "xnnpack/math.h" #include "xnnpack/unaligned.h" void xnn_qd8_f32_qc4w_gemm_minmax_ukernel_3x4c8__ssse3_madd( 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_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 <= 3); assert(nc != 0); assert(kc != 0); assert(kc % sizeof(int8_t) == 0); assert(a != NULL); assert(w != NULL); assert(c != NULL); kc = round_up_po2(kc, 8 * 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 __m128i vsign_mask = _mm_set1_epi8(0x80); XNN_FORCE_REALIZATION(vsign_mask); const __m128i vinput_zero_point0 = _mm_set1_epi32((int) quantization_params[0].zero_point + 128); const __m128i vinput_zero_point1 = _mm_set1_epi32((int) quantization_params[1].zero_point + 128); const __m128i vinput_zero_point2 = _mm_set1_epi32((int) quantization_params[2].zero_point + 128); const __m128 voutput_min = _mm_set1_ps(params->scalar.min); const __m128 voutput_max = _mm_set1_ps(params->scalar.max); const __m128i vmask = _mm_set1_epi8(0x0F); XNN_FORCE_REALIZATION(vmask); do { const __m128i vksum0123 = _mm_load_si128(w); const __m128i vksum13 = _mm_shuffle_epi32(vksum0123, 0xF5); const __m128i vsum0x02 = _mm_mul_epu32(vksum0123, vinput_zero_point0); const __m128i vsum0x13 = _mm_mul_epu32(vksum13, vinput_zero_point0); const __m128i vsum0x01 = _mm_unpacklo_epi32(vsum0x02, vsum0x13); const __m128i vsum0x23 = _mm_unpackhi_epi32(vsum0x02, vsum0x13); __m128i vacc0x01 = _mm_unpacklo_epi32(vsum0x01, _mm_setzero_si128()); __m128i vacc0x23 = _mm_unpacklo_epi32(vsum0x23, _mm_setzero_si128()); const __m128i vsum1x02 = _mm_mul_epu32(vksum0123, vinput_zero_point1); const __m128i vsum1x13 = _mm_mul_epu32(vksum13, vinput_zero_point1); const __m128i vsum1x01 = _mm_unpacklo_epi32(vsum1x02, vsum1x13); const __m128i vsum1x23 = _mm_unpackhi_epi32(vsum1x02, vsum1x13); __m128i vacc1x01 = _mm_unpacklo_epi32(vsum1x01, _mm_setzero_si128()); __m128i vacc1x23 = _mm_unpacklo_epi32(vsum1x23, _mm_setzero_si128()); const __m128i vsum2x02 = _mm_mul_epu32(vksum0123, vinput_zero_point2); const __m128i vsum2x13 = _mm_mul_epu32(vksum13, vinput_zero_point2); const __m128i vsum2x01 = _mm_unpacklo_epi32(vsum2x02, vsum2x13); const __m128i vsum2x23 = _mm_unpackhi_epi32(vsum2x02, vsum2x13); __m128i vacc2x01 = _mm_unpacklo_epi32(vsum2x01, _mm_setzero_si128()); __m128i vacc2x23 = _mm_unpacklo_epi32(vsum2x23, _mm_setzero_si128()); w = (const int32_t*) w + 4; size_t k = kc; while (k >= 16 * sizeof(int8_t)) { const __m128i va0x01234567 = _mm_xor_si128(_mm_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask); const __m128i va0x89ABCDEF = _mm_xor_si128(_mm_set1_epi64x((int64_t) unaligned_load_u64(a0 + 8)), vsign_mask); a0 += 16; const __m128i va1x01234567 = _mm_xor_si128(_mm_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask); const __m128i va1x89ABCDEF = _mm_xor_si128(_mm_set1_epi64x((int64_t) unaligned_load_u64(a1 + 8)), vsign_mask); a1 += 16; const __m128i va2x01234567 = _mm_xor_si128(_mm_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask); const __m128i va2x89ABCDEF = _mm_xor_si128(_mm_set1_epi64x((int64_t) unaligned_load_u64(a2 + 8)), vsign_mask); a2 += 16; const __m128i vbb01234567x0123 = _mm_load_si128(w); const __m128i vbb89ABCDEFx0123 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16)); const __m128i vbs01234567x01 = _mm_srli_epi32(vbb01234567x0123, 4); const __m128i vbs89ABCDEFx01 = _mm_srli_epi32(vbb89ABCDEFx0123, 4); const __m128i vb01234567x01 = _mm_and_si128(vbb01234567x0123, vmask); const __m128i vb89ABCDEFx01 = _mm_and_si128(vbb89ABCDEFx0123, vmask); const __m128i vb01234567x23 = _mm_and_si128(vbs01234567x01, vmask); const __m128i vb89ABCDEFx23 = _mm_and_si128(vbs89ABCDEFx01, vmask); vacc0x01 = _mm_dpbusd_epi32_madd(vacc0x01, va0x01234567, vb01234567x01); vacc0x23 = _mm_dpbusd_epi32_madd(vacc0x23, va0x01234567, vb89ABCDEFx01); vacc1x01 = _mm_dpbusd_epi32_madd(vacc1x01, va1x01234567, vb01234567x01); vacc1x23 = _mm_dpbusd_epi32_madd(vacc1x23, va1x01234567, vb89ABCDEFx01); vacc2x01 = _mm_dpbusd_epi32_madd(vacc2x01, va2x01234567, vb01234567x01); vacc2x23 = _mm_dpbusd_epi32_madd(vacc2x23, va2x01234567, vb89ABCDEFx01); vacc0x01 = _mm_dpbusd_epi32_madd(vacc0x01, va0x89ABCDEF, vb01234567x23); vacc0x23 = _mm_dpbusd_epi32_madd(vacc0x23, va0x89ABCDEF, vb89ABCDEFx23); vacc1x01 = _mm_dpbusd_epi32_madd(vacc1x01, va1x89ABCDEF, vb01234567x23); vacc1x23 = _mm_dpbusd_epi32_madd(vacc1x23, va1x89ABCDEF, vb89ABCDEFx23); vacc2x01 = _mm_dpbusd_epi32_madd(vacc2x01, va2x89ABCDEF, vb01234567x23); vacc2x23 = _mm_dpbusd_epi32_madd(vacc2x23, va2x89ABCDEF, vb89ABCDEFx23); w = (const int8_t*) w + 32; k -= 16 * sizeof(int8_t); } if (k != 0) { const __m128i va0x01234567 = _mm_xor_si128(_mm_set1_epi64x((int64_t) unaligned_load_u64(a0)), vsign_mask); a0 += 8; const __m128i va1x01234567 = _mm_xor_si128(_mm_set1_epi64x((int64_t) unaligned_load_u64(a1)), vsign_mask); a1 += 8; const __m128i va2x01234567 = _mm_xor_si128(_mm_set1_epi64x((int64_t) unaligned_load_u64(a2)), vsign_mask); a2 += 8; const __m128i vbb01234567x0123 = _mm_load_si128(w); const __m128i vbb89ABCDEFx0123 = _mm_load_si128((const __m128i*) ((const int8_t*) w + 16)); const __m128i vb01234567x01 = _mm_and_si128(vbb01234567x0123, vmask); const __m128i vb89ABCDEFx01 = _mm_and_si128(vbb89ABCDEFx0123, vmask); vacc0x01 = _mm_dpbusd_epi32_madd(vacc0x01, va0x01234567, vb01234567x01); vacc0x23 = _mm_dpbusd_epi32_madd(vacc0x23, va0x01234567, vb89ABCDEFx01); vacc1x01 = _mm_dpbusd_epi32_madd(vacc1x01, va1x01234567, vb01234567x01); vacc1x23 = _mm_dpbusd_epi32_madd(vacc1x23, va1x01234567, vb89ABCDEFx01); vacc2x01 = _mm_dpbusd_epi32_madd(vacc2x01, va2x01234567, vb01234567x01); vacc2x23 = _mm_dpbusd_epi32_madd(vacc2x23, va2x01234567, vb89ABCDEFx01); w = (const int8_t*) w + 32; k -= 8 * sizeof(int8_t); } // Add adjacent pairs __m128i vacc0x0123 = _mm_hadd_epi32(vacc0x01, vacc0x23); __m128i vacc1x0123 = _mm_hadd_epi32(vacc1x01, vacc1x23); __m128i vacc2x0123 = _mm_hadd_epi32(vacc2x01, vacc2x23); __m128 vout0x0123 = _mm_cvtepi32_ps(vacc0x0123); __m128 vout1x0123 = _mm_cvtepi32_ps(vacc1x0123); __m128 vout2x0123 = _mm_cvtepi32_ps(vacc2x0123); vout0x0123 = _mm_mul_ps(vout0x0123, _mm_set1_ps(quantization_params[0].inv_scale)); vout1x0123 = _mm_mul_ps(vout1x0123, _mm_set1_ps(quantization_params[1].inv_scale)); vout2x0123 = _mm_mul_ps(vout2x0123, _mm_set1_ps(quantization_params[2].inv_scale)); const __m128 vfilter_output_scale0123 = _mm_load_ps((const float*) w); const __m128 vbias0123 = _mm_load_ps((const float*) w + 4); w = (const float*) w + 8; vout0x0123 = _mm_add_ps(_mm_mul_ps(vout0x0123, vfilter_output_scale0123), vbias0123); vout1x0123 = _mm_add_ps(_mm_mul_ps(vout1x0123, vfilter_output_scale0123), vbias0123); vout2x0123 = _mm_add_ps(_mm_mul_ps(vout2x0123, vfilter_output_scale0123), vbias0123); vout0x0123 = _mm_max_ps(vout0x0123, voutput_min); vout1x0123 = _mm_max_ps(vout1x0123, voutput_min); vout2x0123 = _mm_max_ps(vout2x0123, voutput_min); vout0x0123 = _mm_min_ps(vout0x0123, voutput_max); vout1x0123 = _mm_min_ps(vout1x0123, voutput_max); vout2x0123 = _mm_min_ps(vout2x0123, voutput_max); if XNN_LIKELY(nc >= 4) { _mm_storeu_ps(c0, vout0x0123); a0 = (const int8_t*) ((uintptr_t) a0 - kc); c0 = (float*) ((uintptr_t) c0 + cn_stride); _mm_storeu_ps(c1, vout1x0123); a1 = (const int8_t*) ((uintptr_t) a1 - kc); c1 = (float*) ((uintptr_t) c1 + cn_stride); _mm_storeu_ps(c2, vout2x0123); a2 = (const int8_t*) ((uintptr_t) a2 - kc); c2 = (float*) ((uintptr_t) c2 + cn_stride); nc -= 4; } else { if (nc & 2) { _mm_storel_pi((__m64*) c0, vout0x0123); c0 += 2; vout0x0123 = _mm_movehl_ps(vout0x0123, vout0x0123); _mm_storel_pi((__m64*) c1, vout1x0123); c1 += 2; vout1x0123 = _mm_movehl_ps(vout1x0123, vout1x0123); _mm_storel_pi((__m64*) c2, vout2x0123); c2 += 2; vout2x0123 = _mm_movehl_ps(vout2x0123, vout2x0123); } if (nc & 1) { _mm_store_ss(c0, vout0x0123); _mm_store_ss(c1, vout1x0123); _mm_store_ss(c2, vout2x0123); } nc = 0; } } while (nc != 0); }