// Auto-generated file. Do not edit! // Template: src/qs8-gemm/MRx4c8-sse.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/math.h" #include "xnnpack/unaligned.h" void xnn_qd8_f32_qb4w_gemm_minmax_ukernel_1x4c8__sse2_ld64( 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 <= 1); assert(nc != 0); assert(kc != 0); assert(kc % sizeof(int8_t) == 0); assert(a != NULL); assert(w != NULL); assert(c != NULL); size_t bl = params->scalar.blocksize; assert(bl <= round_up_po2(kc, 2)); assert(bl != 0); assert(bl % 32 == 0); kc = round_up_po2(kc, 8 * sizeof(int8_t)); const int8_t* a0 = a; float* c0 = c; const __m128 vmin = _mm_set1_ps(params->scalar.min); const __m128 vmax = _mm_set1_ps(params->scalar.max); XNN_FORCE_REALIZATION(vmin); XNN_FORCE_REALIZATION(vmax); const __m128i vmask = _mm_set1_epi8(0xF0); XNN_FORCE_REALIZATION(vmask); do { const __m128 vksum = _mm_load_ps((const float*) w); __m128i vinput_zero_point0 = _mm_cvtsi32_si128(*((const int*) &quantization_params[0].zero_point)); vinput_zero_point0 = _mm_shuffle_epi32(vinput_zero_point0, _MM_SHUFFLE(0, 0, 0, 0)); __m128 vinput_zero_point0_float = _mm_cvtepi32_ps(vinput_zero_point0); __m128 vout0x0123 = _mm_mul_ps(vksum, vinput_zero_point0_float); w = (const int32_t*) w + 4; for (size_t kb=0; kb < kc; kb += bl) { __m128i vacc0x0 = _mm_setzero_si128(); __m128i vacc0x1 = _mm_setzero_si128(); __m128i vacc0x2 = _mm_setzero_si128(); __m128i vacc0x3 = _mm_setzero_si128(); size_t k = bl; while (k >= 16 * sizeof(int8_t)) { const __m128i va0c0 = _mm_loadl_epi64((const __m128i*) a0); const __m128i vxa0c0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0c0, va0c0), 8); a0 += 8; const __m128i vb0c01 = _mm_loadl_epi64((const __m128i*) w); const __m128i vbs0c0 = _mm_slli_epi32(vb0c01, 4); const __m128i vb0c0 = _mm_and_si128(vbs0c0, vmask); const __m128i vxb0c0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0c0, vb0c0), 8); vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0c0, vxb0c0)); const __m128i vb1c01 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8)); const __m128i vbs1c0 = _mm_slli_epi32(vb1c01, 4); const __m128i vb1c0 = _mm_and_si128(vbs1c0, vmask); const __m128i vxb1c0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1c0, vb1c0), 8); vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0c0, vxb1c0)); const __m128i vb2c01 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16)); const __m128i vbs2c0 = _mm_slli_epi32(vb2c01, 4); const __m128i vb2c0 = _mm_and_si128(vbs2c0, vmask); const __m128i vxb2c0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2c0, vb2c0), 8); vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0c0, vxb2c0)); const __m128i vb3c01 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24)); const __m128i vbs3c0 = _mm_slli_epi32(vb3c01, 4); const __m128i vb3c0 = _mm_and_si128(vbs3c0, vmask); const __m128i vxb3c0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb3c0, vb3c0), 8); vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0c0, vxb3c0)); const __m128i va0c1 = _mm_loadl_epi64((const __m128i*) a0); const __m128i vxa0c1 = _mm_srai_epi16(_mm_unpacklo_epi8(va0c1, va0c1), 8); a0 += 8; const __m128i vb0c1 = _mm_and_si128(vb0c01, vmask); const __m128i vxb0c1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0c1, vb0c1), 8); vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0c1, vxb0c1)); const __m128i vb1c1 = _mm_and_si128(vb1c01, vmask); const __m128i vxb1c1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1c1, vb1c1), 8); vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0c1, vxb1c1)); const __m128i vb2c1 = _mm_and_si128(vb2c01, vmask); const __m128i vxb2c1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2c1, vb2c1), 8); vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0c1, vxb2c1)); const __m128i vb3c1 = _mm_and_si128(vb3c01, vmask); const __m128i vxb3c1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb3c1, vb3c1), 8); vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0c1, vxb3c1)); w = (const int8_t*) w + 32; k -= 16 * sizeof(int8_t); } while (k >= 8 * sizeof(int8_t)) { const __m128i va0 = _mm_loadl_epi64((const __m128i*) a0); const __m128i vxa0 = _mm_srai_epi16(_mm_unpacklo_epi8(va0, va0), 8); a0 += 8; __m128i vb0 = _mm_loadl_epi64((const __m128i*) w); vb0 = _mm_slli_epi32(vb0, 4); vb0 = _mm_and_si128(vb0, vmask); const __m128i vxb0 = _mm_srai_epi16(_mm_unpacklo_epi8(vb0, vb0), 8); vacc0x0 = _mm_add_epi32(vacc0x0, _mm_madd_epi16(vxa0, vxb0)); __m128i vb1 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 8)); vb1 = _mm_slli_epi32(vb1, 4); vb1 = _mm_and_si128(vb1, vmask); const __m128i vxb1 = _mm_srai_epi16(_mm_unpacklo_epi8(vb1, vb1), 8); vacc0x1 = _mm_add_epi32(vacc0x1, _mm_madd_epi16(vxa0, vxb1)); __m128i vb2 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 16)); vb2 = _mm_slli_epi32(vb2, 4); vb2 = _mm_and_si128(vb2, vmask); const __m128i vxb2 = _mm_srai_epi16(_mm_unpacklo_epi8(vb2, vb2), 8); vacc0x2 = _mm_add_epi32(vacc0x2, _mm_madd_epi16(vxa0, vxb2)); __m128i vb3 = _mm_loadl_epi64((const __m128i*) ((const int8_t*) w + 24)); vb3 = _mm_slli_epi32(vb3, 4); vb3 = _mm_and_si128(vb3, vmask); const __m128i vxb3 = _mm_srai_epi16(_mm_unpacklo_epi8(vb3, vb3), 8); vacc0x3 = _mm_add_epi32(vacc0x3, _mm_madd_epi16(vxa0, vxb3)); w = (const int8_t*) w + 32; k -= 8 * sizeof(int8_t); } // accumulate float const __m128 vfilter_output_scale0123 = _mm_castsi128_ps(_mm_unpacklo_epi16(_mm_setzero_si128(), _mm_loadl_epi64((const __m128i*) w))); w = (const uint16_t*) w + 4; const __m128i vacc0x02 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x0, vacc0x2), _mm_unpackhi_epi32(vacc0x0, vacc0x2)); const __m128i vacc0x13 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x1, vacc0x3), _mm_unpackhi_epi32(vacc0x1, vacc0x3)); __m128i vacc0x0123 = _mm_add_epi32(_mm_unpacklo_epi32(vacc0x02, vacc0x13), _mm_unpackhi_epi32(vacc0x02, vacc0x13)); vout0x0123 = _mm_add_ps(vout0x0123, _mm_mul_ps(_mm_cvtepi32_ps(vacc0x0123), vfilter_output_scale0123)); } const __m128 vinput_scale0 = _mm_load1_ps(&quantization_params[0].inv_scale); vout0x0123 = _mm_mul_ps(vout0x0123, vinput_scale0); const __m128 vbias0123 = _mm_loadu_ps((const float*) w); w = (const float*) w + 4; vout0x0123 = _mm_add_ps(vout0x0123, vbias0123); vout0x0123 = _mm_max_ps(vout0x0123, vmin); vout0x0123 = _mm_min_ps(vout0x0123, vmax); 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); nc -= 4; } else { if (nc & 2) { _mm_storel_pi((__m64*) c0, vout0x0123); vout0x0123 = _mm_unpackhi_ps(vout0x0123, vout0x0123); c0 += 2; } if (nc & 1) { _mm_store_ss(c0, vout0x0123); } nc = 0; } } while (nc != 0); }