1// Copyright 2019 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$assert BATCH_TILE % 8 == 0 7$assert BATCH_TILE >= 8 8$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ" 9$assert OP in ["ADD", "DIV", "MAX", "MIN", "MUL", "SUB", "SQRDIFF"] 10$assert ACTIVATION in ["LINEAR", "MINMAX"] 11#include <assert.h> 12 13#include <immintrin.h> 14 15#include <xnnpack/common.h> 16#include <xnnpack/vbinary.h> 17 18 19$_MM256_OP_PS = { 20$ "ADD": lambda x, y: "_mm256_add_ps(%s, %s)" % (x, y), 21$ "DIV": lambda x, y: "_mm256_div_ps(%s, %s)" % (x, y), 22$ "MAX": lambda x, y: "_mm256_max_ps(%s, %s)" % (x, y), 23$ "MIN": lambda x, y: "_mm256_min_ps(%s, %s)" % (x, y), 24$ "MUL": lambda x, y: "_mm256_mul_ps(%s, %s)" % (x, y), 25$ "SUB": lambda x, y: "_mm256_sub_ps(%s, %s)" % (x, y), 26$ "SQRDIFF": lambda x, y: "_mm256_sub_ps(%s, %s)" % (x, y), 27$}[OP] 28$SUFFIX = {"LINEAR": "", "MINMAX": "_minmax"}[ACTIVATION] 29$PARAMS = {"LINEAR": "xnn_f32_default_params", "MINMAX": "xnn_f32_minmax_params"}[ACTIVATION] 30void xnn_f32_v${OP.lower()}${SUFFIX}_ukernel__avx_x${BATCH_TILE}( 31 size_t n, 32 const float* a, 33 const float* b, 34 float* y, 35 const union ${PARAMS} params[restrict XNN_MIN_ELEMENTS(1)]) 36{ 37 assert(n != 0); 38 assert(n % sizeof(float) == 0); 39 assert(a != NULL); 40 assert(b != NULL); 41 assert(y != NULL); 42 43 $if ACTIVATION == "MINMAX": 44 const __m256 vy_min = _mm256_load_ps(params->avx.min); 45 const __m256 vy_max = _mm256_load_ps(params->avx.max); 46 47 for (; n >= ${BATCH_TILE} * sizeof(float); n -= ${BATCH_TILE} * sizeof(float)) { 48 const __m256 va${ABC[0:8]} = _mm256_loadu_ps(a); 49 $for N in range(8, BATCH_TILE, 8): 50 const __m256 va${ABC[N:N+8]} = _mm256_loadu_ps(a + ${N}); 51 a += ${BATCH_TILE}; 52 53 const __m256 vb${ABC[0:8]} = _mm256_loadu_ps(b); 54 $for N in range(8, BATCH_TILE, 8): 55 const __m256 vb${ABC[N:N+8]} = _mm256_loadu_ps(b + ${N}); 56 b += ${BATCH_TILE}; 57 58 $for N in range(0, BATCH_TILE, 8): 59 __m256 vy${ABC[N:N+8]} = ${_MM256_OP_PS("va" + ABC[N:N+8], "vb" + ABC[N:N+8])}; 60 61 $if OP == "SQRDIFF": 62 $for N in range(0, BATCH_TILE, 8): 63 vy${ABC[N:N+8]} = _mm256_mul_ps(vy${ABC[N:N+8]}, vy${ABC[N:N+8]}); 64 65 $if ACTIVATION == "MINMAX": 66 $for N in range(0, BATCH_TILE, 8): 67 vy${ABC[N:N+8]} = _mm256_max_ps(vy${ABC[N:N+8]}, vy_min); 68 69 $for N in range(0, BATCH_TILE, 8): 70 vy${ABC[N:N+8]} = _mm256_min_ps(vy${ABC[N:N+8]}, vy_max); 71 72 _mm256_storeu_ps(y, vy${ABC[0:8]}); 73 $for N in range(8, BATCH_TILE, 8): 74 _mm256_storeu_ps(y + ${N}, vy${ABC[N:N+8]}); 75 y += ${BATCH_TILE}; 76 } 77 $if BATCH_TILE > 8: 78 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) { 79 const __m256 va = _mm256_loadu_ps(a); 80 a += 8; 81 82 const __m256 vb = _mm256_loadu_ps(b); 83 b += 8; 84 85 __m256 vy = ${_MM256_OP_PS("va", "vb")}; 86 $if OP == "SQRDIFF": 87 vy = _mm256_mul_ps(vy, vy); 88 $if ACTIVATION == "MINMAX": 89 vy = _mm256_max_ps(vy, vy_min); 90 vy = _mm256_min_ps(vy, vy_max); 91 _mm256_storeu_ps(y, vy); 92 y += 8; 93 } 94 if XNN_UNLIKELY(n != 0) { 95 assert(n >= 1 * sizeof(float)); 96 assert(n <= 7 * sizeof(float)); 97 const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) ¶ms->avx.mask_table[7] - n)); 98 99 const __m256 va = _mm256_maskload_ps(a, vmask); 100 const __m256 vb = _mm256_maskload_ps(b, vmask); 101 102 __m256 vy = ${_MM256_OP_PS("va", "vb")}; 103 $if OP == "SQRDIFF": 104 vy = _mm256_mul_ps(vy, vy); 105 $if ACTIVATION == "MINMAX": 106 vy = _mm256_max_ps(vy, vy_min); 107 vy = _mm256_min_ps(vy, vy_max); 108 109 __m128 vy_lo = _mm256_castps256_ps128(vy); 110 if (n & (4 * sizeof(float))) { 111 _mm_storeu_ps(y, vy_lo); 112 vy_lo = _mm256_extractf128_ps(vy, 1); 113 y += 4; 114 } 115 if (n & (2 * sizeof(float))) { 116 _mm_storel_pi((__m64*) y, vy_lo); 117 vy_lo = _mm_movehl_ps(vy_lo, vy_lo); 118 y += 2; 119 } 120 if (n & (1 * sizeof(float))) { 121 _mm_store_ss(y, vy_lo); 122 } 123 } 124} 125