| // Auto-generated file. Do not edit! |
| // Template: src/qs8-vaddc/avx2-mul32-ld64.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 <assert.h> |
| |
| #include <immintrin.h> |
| |
| #include <xnnpack/intrinsics-polyfill.h> |
| #include <xnnpack/vadd.h> |
| |
| |
| void xnn_qs8_vaddc_minmax_ukernel__avx2_mul32_ld64_x24( |
| size_t n, |
| const int8_t* input_a, |
| const int8_t* input_b, |
| int8_t* output, |
| const union xnn_qs8_add_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_DISABLE_TSAN |
| { |
| const __m256i va_multiplier = _mm256_load_si256((const __m256i*) params->avx2.a_multiplier); |
| const __m256i vrounding = _mm256_load_si256((const __m256i*) params->avx2.rounding); |
| const __m128i vshift = _mm_loadu_si32(params->avx2.shift); |
| const __m256i voutput_zero_point = _mm256_load_si256((const __m256i*) params->avx2.output_zero_point); |
| const __m128i voutput_min = _mm_load_si128((const __m128i*) params->avx2.output_min); |
| const __m128i voutput_max = _mm_load_si128((const __m128i*) params->avx2.output_max); |
| |
| __m256i vbias = _mm256_add_epi32( |
| _mm256_broadcastd_epi32(_mm_cvtsi32_si128(params->avx2.b_multiplier[0] * (int32_t) *input_b)), |
| _mm256_load_si256((const __m256i*) params->avx2.bias)); |
| for (; n >= 24 * sizeof(int8_t); n -= 24 * sizeof(int8_t)) { |
| const __m256i va01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) input_a)); |
| const __m256i va89ABCDEF = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (input_a + 8))); |
| const __m256i vaGHIJKLMN = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) (input_a + 16))); |
| input_a += 24; |
| |
| __m256i vacc01234567 = _mm256_add_epi32(vbias, _mm256_mullo_epi32(va01234567, va_multiplier)); |
| __m256i vacc89ABCDEF = _mm256_add_epi32(vbias, _mm256_mullo_epi32(va89ABCDEF, va_multiplier)); |
| __m256i vaccGHIJKLMN = _mm256_add_epi32(vbias, _mm256_mullo_epi32(vaGHIJKLMN, va_multiplier)); |
| |
| vacc01234567 = _mm256_sra_epi32(_mm256_add_epi32(vacc01234567, vrounding), vshift); |
| vacc89ABCDEF = _mm256_sra_epi32(_mm256_add_epi32(vacc89ABCDEF, vrounding), vshift); |
| vaccGHIJKLMN = _mm256_sra_epi32(_mm256_add_epi32(vaccGHIJKLMN, vrounding), vshift); |
| |
| __m256i vout012389AB4567CDEF = _mm256_adds_epi16(_mm256_packs_epi32(vacc01234567, vacc89ABCDEF), voutput_zero_point); |
| __m128i voutGHIJKLMN = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vaccGHIJKLMN), _mm256_extracti128_si256(vaccGHIJKLMN, 1)), _mm256_castsi256_si128(voutput_zero_point)); |
| |
| __m128i vout0123456789ABCDEF = _mm_shuffle_epi32(_mm_packs_epi16(_mm256_castsi256_si128(vout012389AB4567CDEF), _mm256_extracti128_si256(vout012389AB4567CDEF, 1)), _MM_SHUFFLE(3, 1, 2, 0)); |
| __m128i voutGHIJKLMNGHIJKLMN = _mm_packs_epi16(voutGHIJKLMN, voutGHIJKLMN); |
| |
| vout0123456789ABCDEF = _mm_max_epi8(vout0123456789ABCDEF, voutput_min); |
| voutGHIJKLMNGHIJKLMN = _mm_max_epi8(voutGHIJKLMNGHIJKLMN, voutput_min); |
| |
| vout0123456789ABCDEF = _mm_min_epi8(vout0123456789ABCDEF, voutput_max); |
| voutGHIJKLMNGHIJKLMN = _mm_min_epi8(voutGHIJKLMNGHIJKLMN, voutput_max); |
| |
| _mm_storeu_si128((__m128i*) output, vout0123456789ABCDEF); |
| _mm_storel_epi64((__m128i*) (output + 16), voutGHIJKLMNGHIJKLMN); |
| output += 24; |
| } |
| if XNN_UNLIKELY(n != 0) { |
| do { |
| const __m256i va01234567 = _mm256_cvtepi8_epi32(_mm_loadl_epi64((const __m128i*) input_a)); |
| input_a += 8; |
| |
| __m256i vacc01234567 = _mm256_add_epi32(vbias, _mm256_mullo_epi32(va01234567, va_multiplier)); |
| |
| vacc01234567 = _mm256_sra_epi32(_mm256_add_epi32(vacc01234567, vrounding), vshift); |
| |
| __m128i vout01234567 = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc01234567), _mm256_extracti128_si256(vacc01234567, 1)), _mm256_castsi256_si128(voutput_zero_point)); |
| __m128i vout0123456701234567 = _mm_packs_epi16(vout01234567, vout01234567); |
| vout0123456701234567 = _mm_max_epi8(vout0123456701234567, voutput_min); |
| vout0123456701234567 = _mm_min_epi8(vout0123456701234567, voutput_max); |
| |
| if XNN_LIKELY(n >= (8 * sizeof(int8_t))) { |
| _mm_storel_epi64((__m128i*) output, vout0123456701234567); |
| output += 8; |
| n -= 8 * sizeof(int8_t); |
| } else { |
| if (n & (4 * sizeof(int8_t))) { |
| *((uint32_t*) output) = (uint32_t) _mm_cvtsi128_si32(vout0123456701234567); |
| vout0123456701234567 = _mm_srli_epi64(vout0123456701234567, 32); |
| output += 4; |
| } |
| if (n & (2 * sizeof(int8_t))) { |
| *((uint16_t*) output) = (uint16_t) _mm_extract_epi16(vout0123456701234567, 0); |
| vout0123456701234567 = _mm_srli_epi32(vout0123456701234567, 16); |
| output += 2; |
| } |
| if (n & (1 * sizeof(int8_t))) { |
| *output = (int8_t) _mm_extract_epi8(vout0123456701234567, 0); |
| } |
| n = 0; |
| } |
| } while (n != 0); |
| } |
| } |