1// Copyright 2020 Google LLC 2// 3// This source code is licensed under the BSD-style license found in the 4// LICENSE file in the root directory of this source tree. 5 6$assert DATATYPE in ["QS8", "QU8"] 7$assert BATCH_TILE % 8 == 0 8$assert BATCH_TILE >= 8 9$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ" 10#include <assert.h> 11 12#include <immintrin.h> 13 14#include <xnnpack/intrinsics-polyfill.h> 15#include <xnnpack/vadd.h> 16 17 18$XINT8_T = {"QS8": "int8_t", "QU8": "uint8_t"}[DATATYPE] 19$_MM256_CVTEPX8_EPI32 = {"QS8": "_mm256_cvtepi8_epi32", "QU8": "_mm256_cvtepu8_epi32"}[DATATYPE] 20$_MM_PACKXS_EPI16 = {"QS8": "_mm_packs_epi16", "QU8": "_mm_packus_epi16"}[DATATYPE] 21$_MM_MIN_EPX8 = {"QS8": "_mm_min_epi8", "QU8": "_mm_min_epu8"}[DATATYPE] 22$_MM_MAX_EPX8 = {"QS8": "_mm_max_epi8", "QU8": "_mm_max_epu8"}[DATATYPE] 23void xnn_${DATATYPE.lower()}_vadd_minmax_ukernel__avx2_mul32_ld64_x${BATCH_TILE}( 24 size_t n, 25 const ${XINT8_T}* input_a, 26 const ${XINT8_T}* input_b, 27 ${XINT8_T}* output, 28 const union xnn_${DATATYPE.lower()}_add_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS 29{ 30 const __m256i vbias = _mm256_load_si256((const __m256i*) params->avx2.bias); 31 const __m256i va_multiplier = _mm256_load_si256((const __m256i*) params->avx2.a_multiplier); 32 const __m256i vb_multiplier = _mm256_load_si256((const __m256i*) params->avx2.b_multiplier); 33 const __m128i vshift = _mm_load_si128((const __m128i*) params->avx2.shift); 34 $if BATCH_TILE > 8: 35 const __m256i voutput_zero_point = _mm256_load_si256((const __m256i*) params->avx2.output_zero_point); 36 $else: 37 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->avx2.output_zero_point); 38 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->avx2.output_min); 39 const __m128i voutput_max = _mm_load_si128((const __m128i*) params->avx2.output_max); 40 41 for (; n >= ${BATCH_TILE} * sizeof(${XINT8_T}); n -= ${BATCH_TILE} * sizeof(${XINT8_T})) { 42 const __m256i va${ABC[0:8]} = ${_MM256_CVTEPX8_EPI32}(_mm_loadl_epi64((const __m128i*) input_a)); 43 const __m256i vb${ABC[0:8]} = ${_MM256_CVTEPX8_EPI32}(_mm_loadl_epi64((const __m128i*) input_b)); 44 $for N in range(8, BATCH_TILE, 8): 45 const __m256i va${ABC[N:N+8]} = ${_MM256_CVTEPX8_EPI32}(_mm_loadl_epi64((const __m128i*) (input_a + ${N}))); 46 const __m256i vb${ABC[N:N+8]} = ${_MM256_CVTEPX8_EPI32}(_mm_loadl_epi64((const __m128i*) (input_b + ${N}))); 47 input_a += ${BATCH_TILE}; 48 input_b += ${BATCH_TILE}; 49 50 $for N in range(0, BATCH_TILE, 8): 51 __m256i vacc${ABC[N:N+8]} = _mm256_add_epi32(vbias, _mm256_mullo_epi32(va${ABC[N:N+8]}, va_multiplier)); 52 53 $for N in range(0, BATCH_TILE, 8): 54 vacc${ABC[N:N+8]} = _mm256_add_epi32(vacc${ABC[N:N+8]}, _mm256_mullo_epi32(vb${ABC[N:N+8]}, vb_multiplier)); 55 56 $for N in range(0, BATCH_TILE, 8): 57 vacc${ABC[N:N+8]} = _mm256_sra_epi32(vacc${ABC[N:N+8]}, vshift); 58 59 $for N in range(0, BATCH_TILE, 16): 60 $if N + 8 < BATCH_TILE: 61 __m256i vout${ABC[N:N+4]}${ABC[N+8:N+12]}${ABC[N+4:N+8]}${ABC[N+12:N+16]} = _mm256_adds_epi16(_mm256_packs_epi32(vacc${ABC[N:N+8]}, vacc${ABC[N+8:N+16]}), voutput_zero_point); 62 $elif BATCH_TILE > 8: 63 __m128i vout${ABC[N:N+8]} = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc${ABC[N:N+8]}), _mm256_extracti128_si256(vacc${ABC[N:N+8]}, 1)), _mm256_castsi256_si128(voutput_zero_point)); 64 $else: 65 __m128i vout${ABC[N:N+8]} = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc${ABC[N:N+8]}), _mm256_extracti128_si256(vacc${ABC[N:N+8]}, 1)), voutput_zero_point); 66 67 $for N in range(0, BATCH_TILE, 16): 68 $if N + 8 < BATCH_TILE: 69 __m128i vout${ABC[N:N+16]} = _mm_shuffle_epi32(${_MM_PACKXS_EPI16}(_mm256_castsi256_si128(vout${ABC[N:N+4]}${ABC[N+8:N+12]}${ABC[N+4:N+8]}${ABC[N+12:N+16]}), _mm256_extracti128_si256(vout${ABC[N:N+4]}${ABC[N+8:N+12]}${ABC[N+4:N+8]}${ABC[N+12:N+16]}, 1)), _MM_SHUFFLE(3, 1, 2, 0)); 70 $else: 71 __m128i vout${ABC[N:N+8]}${ABC[N:N+8]} = ${_MM_PACKXS_EPI16}(vout${ABC[N:N+8]}, vout${ABC[N:N+8]}); 72 73 $for N in range(0, BATCH_TILE, 16): 74 $if N + 8 < BATCH_TILE: 75 vout${ABC[N:N+16]} = ${_MM_MAX_EPX8}(vout${ABC[N:N+16]}, voutput_min); 76 $else: 77 vout${ABC[N:N+8]}${ABC[N:N+8]} = ${_MM_MAX_EPX8}(vout${ABC[N:N+8]}${ABC[N:N+8]}, voutput_min); 78 79 $for N in range(0, BATCH_TILE, 16): 80 $if N + 8 < BATCH_TILE: 81 vout${ABC[N:N+16]} = ${_MM_MIN_EPX8}(vout${ABC[N:N+16]}, voutput_max); 82 $else: 83 vout${ABC[N:N+8]}${ABC[N:N+8]} = ${_MM_MIN_EPX8}(vout${ABC[N:N+8]}${ABC[N:N+8]}, voutput_max); 84 85 $if BATCH_TILE >= 16: 86 _mm_storeu_si128((__m128i*) output, vout${ABC[0:16]}); 87 $else: 88 _mm_storel_epi64((__m128i*) output, vout${ABC[0:8]}${ABC[0:8]}); 89 $for N in range(16, BATCH_TILE, 16): 90 $if N + 8 < BATCH_TILE: 91 _mm_storeu_si128((__m128i*) (output + ${N}), vout${ABC[N:N+16]}); 92 $else: 93 _mm_storel_epi64((__m128i*) (output + ${N}), vout${ABC[N:N+8]}${ABC[N:N+8]}); 94 output += ${BATCH_TILE}; 95 } 96 if XNN_UNLIKELY(n != 0) { 97 ${"do " if BATCH_TILE > 8 else ""}{ 98 const __m256i va${ABC[0:8]} = ${_MM256_CVTEPX8_EPI32}(_mm_loadl_epi64((const __m128i*) input_a)); 99 const __m256i vb${ABC[0:8]} = ${_MM256_CVTEPX8_EPI32}(_mm_loadl_epi64((const __m128i*) input_b)); 100 $if BATCH_TILE > 8: 101 input_a += 8; 102 input_b += 8; 103 104 __m256i vacc${ABC[0:8]} = _mm256_add_epi32(vbias, _mm256_mullo_epi32(va${ABC[0:8]}, va_multiplier)); 105 106 vacc${ABC[0:8]} = _mm256_add_epi32(vacc${ABC[0:8]}, _mm256_mullo_epi32(vb${ABC[0:8]}, vb_multiplier)); 107 108 vacc${ABC[0:8]} = _mm256_sra_epi32(vacc${ABC[0:8]}, vshift); 109 110 $if BATCH_TILE > 8: 111 __m128i vout${ABC[0:8]} = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc${ABC[0:8]}), _mm256_extracti128_si256(vacc${ABC[0:8]}, 1)), _mm256_castsi256_si128(voutput_zero_point)); 112 $else: 113 __m128i vout${ABC[0:8]} = _mm_adds_epi16(_mm_packs_epi32(_mm256_castsi256_si128(vacc${ABC[0:8]}), _mm256_extracti128_si256(vacc${ABC[0:8]}, 1)), voutput_zero_point); 114 __m128i vout${ABC[0:8]}${ABC[0:8]} = ${_MM_PACKXS_EPI16}(vout${ABC[0:8]}, vout${ABC[0:8]}); 115 vout${ABC[0:8]}${ABC[0:8]} = ${_MM_MAX_EPX8}(vout${ABC[0:8]}${ABC[0:8]}, voutput_min); 116 vout${ABC[0:8]}${ABC[0:8]} = ${_MM_MIN_EPX8}(vout${ABC[0:8]}${ABC[0:8]}, voutput_max); 117 118 $if BATCH_TILE > 8: 119 if XNN_LIKELY(n >= (8 * sizeof(${XINT8_T}))) { 120 _mm_storel_epi64((__m128i*) output, vout${ABC[0:8]}${ABC[0:8]}); 121 output += 8; 122 n -= 8 * sizeof(${XINT8_T}); 123 } else { 124 if (n & (4 * sizeof(${XINT8_T}))) { 125 _mm_storeu_si32(output, vout${ABC[0:8]}${ABC[0:8]}); 126 vout${ABC[0:8]}${ABC[0:8]} = _mm_srli_epi64(vout${ABC[0:8]}${ABC[0:8]}, 32); 127 output += 4; 128 } 129 if (n & (2 * sizeof(${XINT8_T}))) { 130 _mm_storeu_si16(output, vout${ABC[0:8]}${ABC[0:8]}); 131 vout${ABC[0:8]}${ABC[0:8]} = _mm_srli_epi32(vout${ABC[0:8]}${ABC[0:8]}, 16); 132 output += 2; 133 } 134 if (n & (1 * sizeof(${XINT8_T}))) { 135 *output = (${XINT8_T}) _mm_extract_epi8(vout${ABC[0:8]}${ABC[0:8]}, 0); 136 } 137 n = 0; 138 } 139 $else: 140 if (n & (4 * sizeof(${XINT8_T}))) { 141 _mm_storeu_si32(output, vout${ABC[0:8]}${ABC[0:8]}); 142 vout${ABC[0:8]}${ABC[0:8]} = _mm_srli_epi64(vout${ABC[0:8]}${ABC[0:8]}, 32); 143 output += 4; 144 } 145 if (n & (2 * sizeof(${XINT8_T}))) { 146 _mm_storeu_si16(output, vout${ABC[0:8]}${ABC[0:8]}); 147 vout${ABC[0:8]}${ABC[0:8]} = _mm_srli_epi32(vout${ABC[0:8]}${ABC[0:8]}, 16); 148 output += 2; 149 } 150 if (n & (1 * sizeof(${XINT8_T}))) { 151 *output = (${XINT8_T}) _mm_extract_epi8(vout${ABC[0:8]}${ABC[0:8]}, 0); 152 } 153 }${" while (n != 0);" if BATCH_TILE > 8 else ""} 154 } 155} 156