1// Copyright 2021 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 % 16 == 0 8$assert BATCH_TILE >= 16 9$SIMD_TILE = BATCH_TILE // 4 10$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ" 11#include <assert.h> 12 13#include <immintrin.h> 14 15#include <xnnpack/common.h> 16#include <xnnpack/intrinsics-polyfill.h> 17#include <xnnpack/vcvt.h> 18 19 20$XINT8_T = {"QS8": "int8_t", "QU8": "uint8_t"}[DATATYPE] 21$_MM256_PACKXS_EPI16 = {"QS8": "_mm256_packs_epi16", "QU8": "_mm256_packus_epi16"}[DATATYPE] 22$_MM_PACKXS_EPI16 = {"QS8": "_mm_packs_epi16", "QU8": "_mm_packus_epi16"}[DATATYPE] 23$_MM256_MAX_EPX8 = {"QS8": "_mm256_max_epi8", "QU8": "_mm256_max_epu8"}[DATATYPE] 24$_MM_MAX_EPX8 = {"QS8": "_mm_max_epi8", "QU8": "_mm_max_epu8"}[DATATYPE] 25void xnn_f32_${DATATYPE.lower()}_vcvt_ukernel__avx2_x${BATCH_TILE}( 26 size_t n, 27 const float* x, 28 ${XINT8_T}* y, 29 const union xnn_f32_${DATATYPE.lower()}_cvt_params params[restrict XNN_MIN_ELEMENTS(1)]) 30{ 31 assert(n != 0); 32 assert(n % sizeof(float) == 0); 33 assert(x != NULL); 34 assert(y != NULL); 35 36 const __m256 vscale = _mm256_load_ps(params->avx2.scale); 37 const __m256 voutput_max_less_zero_point = _mm256_load_ps(params->avx2.output_max_less_zero_point); 38 const __m256i voutput_zero_point = _mm256_load_si256((const __m256i*) params->avx2.output_zero_point); 39 $if BATCH_TILE > 16: 40 const __m256i vshuffle_mask = _mm256_load_si256((const __m256i*) params->avx2.shuffle_mask); 41 const __m256i voutput_min = _mm256_load_si256((const __m256i*) params->avx2.output_min); 42 $else: 43 const __m128i voutput_min = _mm_load_si128((const __m128i*) params->avx2.output_min); 44 45 for (; n >= ${BATCH_TILE} * sizeof(float); n -= ${BATCH_TILE} * sizeof(float)) { 46 __m256 vx${ABC[0:2]} = _mm256_loadu_ps(x); 47 $for N in range(2, SIMD_TILE, 2): 48 __m256 vx${ABC[N:N+2]} = _mm256_loadu_ps(x + ${N * 4}); 49 x += ${BATCH_TILE}; 50 51 $for N in range(0, SIMD_TILE, 2): 52 vx${ABC[N:N+2]} = _mm256_mul_ps(vx${ABC[N:N+2]}, vscale); 53 54 $for N in range(0, SIMD_TILE, 2): 55 vx${ABC[N:N+2]} = _mm256_min_ps(vx${ABC[N:N+2]}, voutput_max_less_zero_point); 56 57 $for N in range(0, SIMD_TILE, 2): 58 const __m256i vacc${ABC[N:N+2]} = _mm256_cvtps_epi32(vx${ABC[N:N+2]}); 59 60 $for N in range(0, SIMD_TILE, 4): 61 __m256i vacc${ABC[N]}${ABC[N+2]}${ABC[N+1]}${ABC[N+3]} = _mm256_packs_epi32(vacc${ABC[N:N+2]}, vacc${ABC[N+2:N+4]}); 62 63 $for N in range(0, SIMD_TILE, 4): 64 vacc${ABC[N]}${ABC[N+2]}${ABC[N+1]}${ABC[N+3]} = _mm256_adds_epi16(vacc${ABC[N]}${ABC[N+2]}${ABC[N+1]}${ABC[N+3]}, voutput_zero_point); 65 66 $for N in range(0, SIMD_TILE, 8): 67 $if N + 4 < SIMD_TILE: 68 const __m256i vy${ABC[N]}${ABC[N+2]}${ABC[N+4]}${ABC[N+6]}${ABC[N+1]}${ABC[N+3]}${ABC[N+5]}${ABC[N+7]} = ${_MM256_PACKXS_EPI16}(vacc${ABC[N]}${ABC[N+2]}${ABC[N+1]}${ABC[N+3]}, vacc${ABC[N+4]}${ABC[N+6]}${ABC[N+5]}${ABC[N+7]}); 69 $else: 70 const __m128i vy${ABC[N]}${ABC[N+2]}${ABC[N+1]}${ABC[N+3]} = ${_MM_PACKXS_EPI16}(_mm256_castsi256_si128(vacc${ABC[N]}${ABC[N+2]}${ABC[N+1]}${ABC[N+3]}), _mm256_extracti128_si256(vacc${ABC[N]}${ABC[N+2]}${ABC[N+1]}${ABC[N+3]}, 1)); 71 72 $for N in range(0, SIMD_TILE, 8): 73 $if N + 4 < SIMD_TILE: 74 __m256i vy${ABC[N:N+8]} = _mm256_permutevar8x32_epi32(vy${ABC[N]}${ABC[N+2]}${ABC[N+4]}${ABC[N+6]}${ABC[N+1]}${ABC[N+3]}${ABC[N+5]}${ABC[N+7]}, vshuffle_mask); 75 $else: 76 __m128i vy${ABC[N:N+4]} = _mm_shuffle_epi32(vy${ABC[N]}${ABC[N+2]}${ABC[N+1]}${ABC[N+3]}, _MM_SHUFFLE(3, 1, 2, 0)); 77 78 $for N in range(0, SIMD_TILE, 8): 79 $if N + 4 < SIMD_TILE: 80 vy${ABC[N:N+8]} = ${_MM256_MAX_EPX8}(vy${ABC[N:N+8]}, voutput_min); 81 $elif BATCH_TILE > 16: 82 vy${ABC[N:N+4]} = ${_MM_MAX_EPX8}(vy${ABC[N:N+4]}, _mm256_castsi256_si128(voutput_min)); 83 $else: 84 vy${ABC[N:N+4]} = ${_MM_MAX_EPX8}(vy${ABC[N:N+4]}, voutput_min); 85 86 $if SIMD_TILE > 4: 87 _mm256_storeu_si256((__m256i*) y, vy${ABC[0:8]}); 88 $else: 89 _mm_storeu_si128((__m128i*) y, vy${ABC[0:4]}); 90 $for N in range(8, SIMD_TILE, 8): 91 $if N + 4 < SIMD_TILE: 92 _mm256_storeu_si256((__m256i*) (y + ${N * 4}), vy${ABC[N:N+8]}); 93 $else: 94 _mm_storeu_si128((__m128i*) (y + ${N * 4}), vy${ABC[N:N+4]}); 95 y += ${BATCH_TILE}; 96 } 97 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) { 98 __m256 vx = _mm256_loadu_ps(x); 99 vx = _mm256_mul_ps(vx, vscale); 100 vx = _mm256_min_ps(vx, voutput_max_less_zero_point); 101 x += 8; 102 103 const __m256i vacc = _mm256_cvtps_epi32(vx); 104 105 __m128i vy = _mm_packs_epi32(_mm256_castsi256_si128(vacc), _mm256_extracti128_si256(vacc, 1)); 106 vy = _mm_adds_epi16(vy, _mm256_castsi256_si128(voutput_zero_point)); 107 vy = ${_MM_PACKXS_EPI16}(vy, vy); 108 $if BATCH_TILE > 16: 109 vy = ${_MM_MAX_EPX8}(vy, _mm256_castsi256_si128(voutput_min)); 110 $else: 111 vy = ${_MM_MAX_EPX8}(vy, voutput_min); 112 113 _mm_storel_epi64((__m128i*) y, vy); 114 y += 8; 115 } 116 if XNN_UNLIKELY(n != 0) { 117 assert(n >= 1 * sizeof(float)); 118 assert(n <= 7 * sizeof(float)); 119 const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) ¶ms->avx2.mask_table[7] - n)); 120 121 __m256 vx = _mm256_maskload_ps(x, vmask); 122 vx = _mm256_mul_ps(vx, vscale); 123 vx = _mm256_min_ps(vx, voutput_max_less_zero_point); 124 125 const __m256i vacc = _mm256_cvtps_epi32(vx); 126 127 __m128i vy = _mm_packs_epi32(_mm256_castsi256_si128(vacc), _mm256_extracti128_si256(vacc, 1)); 128 vy = _mm_adds_epi16(vy, _mm256_castsi256_si128(voutput_zero_point)); 129 vy = ${_MM_PACKXS_EPI16}(vy, vy); 130 $if BATCH_TILE > 16: 131 vy = ${_MM_MAX_EPX8}(vy, _mm256_castsi256_si128(voutput_min)); 132 $else: 133 vy = ${_MM_MAX_EPX8}(vy, voutput_min); 134 135 if (n & (4 * sizeof(float))) { 136 _mm_storeu_si32(y, vy); 137 y += 4; 138 vy = _mm_srli_epi64(vy, 32); 139 } 140 if (n & (2 * sizeof(float))) { 141 _mm_storeu_si16(y, vy); 142 y += 2; 143 vy = _mm_srli_epi32(vy, 16); 144 } 145 if (n & (1 * sizeof(float))) { 146 *y = (${XINT8_T}) _mm_extract_epi8(vy, 0); 147 } 148 } 149} 150