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", "RDIV", "MAX", "MIN", "MUL", "SUB", "RSUB", "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: "_mm256_add_ps(%s, vb)" % x, 21$ "DIV": lambda x: "_mm256_div_ps(%s, vb)" % x, 22$ "RDIV": lambda x: "_mm256_div_ps(vb, %s)" % x, 23$ "MAX": lambda x: "_mm256_max_ps(%s, vb)" % x, 24$ "MIN": lambda x: "_mm256_min_ps(%s, vb)" % x, 25$ "MUL": lambda x: "_mm256_mul_ps(%s, vb)" % x, 26$ "SUB": lambda x: "_mm256_sub_ps(%s, vb)" % x, 27$ "RSUB": lambda x: "_mm256_sub_ps(vb, %s)" % x, 28$ "SQRDIFF": lambda x: "_mm256_sub_ps(%s, vb)" % x, 29$}[OP] 30$SUFFIX = {"LINEAR": "", "MINMAX": "_minmax"}[ACTIVATION] 31$PARAMS = {"LINEAR": "xnn_f32_default_params", "MINMAX": "xnn_f32_minmax_params"}[ACTIVATION] 32void xnn_f32_v${OP.lower()}c${SUFFIX}_ukernel__avx_x${BATCH_TILE}( 33 size_t n, 34 const float* a, 35 const float* b, 36 float* y, 37 const union ${PARAMS} params[restrict XNN_MIN_ELEMENTS(1)]) 38{ 39 assert(n != 0); 40 assert(n % sizeof(float) == 0); 41 assert(a != NULL); 42 assert(b != NULL); 43 assert(y != NULL); 44 45 $if ACTIVATION == "MINMAX": 46 const __m256 vy_min = _mm256_load_ps(params->avx.min); 47 const __m256 vy_max = _mm256_load_ps(params->avx.max); 48 49 const __m256 vb = _mm256_broadcast_ss(b); 50 for (; n >= ${BATCH_TILE} * sizeof(float); n -= ${BATCH_TILE} * sizeof(float)) { 51 const __m256 va${ABC[0:8]} = _mm256_loadu_ps(a); 52 $for N in range(8, BATCH_TILE, 8): 53 const __m256 va${ABC[N:N+8]} = _mm256_loadu_ps(a + ${N}); 54 a += ${BATCH_TILE}; 55 56 $for N in range(0, BATCH_TILE, 8): 57 __m256 vy${ABC[N:N+8]} = ${_MM256_OP_PS("va" + ABC[N:N+8])}; 58 59 $if OP == "SQRDIFF": 60 $for N in range(0, BATCH_TILE, 8): 61 vy${ABC[N:N+8]} = _mm256_mul_ps(vy${ABC[N:N+8]}, vy${ABC[N:N+8]}); 62 63 $if ACTIVATION == "MINMAX": 64 $for N in range(0, BATCH_TILE, 8): 65 vy${ABC[N:N+8]} = _mm256_max_ps(vy${ABC[N:N+8]}, vy_min); 66 67 $for N in range(0, BATCH_TILE, 8): 68 vy${ABC[N:N+8]} = _mm256_min_ps(vy${ABC[N:N+8]}, vy_max); 69 70 _mm256_storeu_ps(y, vy${ABC[0:8]}); 71 $for N in range(8, BATCH_TILE, 8): 72 _mm256_storeu_ps(y + ${N}, vy${ABC[N:N+8]}); 73 y += ${BATCH_TILE}; 74 } 75 $if BATCH_TILE > 8: 76 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) { 77 const __m256 va = _mm256_loadu_ps(a); 78 a += 8; 79 80 __m256 vy = ${_MM256_OP_PS("va")}; 81 $if OP == "SQRDIFF": 82 vy = _mm256_mul_ps(vy, vy); 83 $if ACTIVATION == "MINMAX": 84 vy = _mm256_max_ps(vy, vy_min); 85 vy = _mm256_min_ps(vy, vy_max); 86 _mm256_storeu_ps(y, vy); 87 y += 8; 88 } 89 if XNN_UNLIKELY(n != 0) { 90 assert(n >= 1 * sizeof(float)); 91 assert(n <= 7 * sizeof(float)); 92 const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) ¶ms->avx.mask_table[7] - n)); 93 94 const __m256 va = _mm256_maskload_ps(a, vmask); 95 96 __m256 vy = ${_MM256_OP_PS("va")}; 97 $if OP == "SQRDIFF": 98 vy = _mm256_mul_ps(vy, vy); 99 $if ACTIVATION == "MINMAX": 100 vy = _mm256_max_ps(vy, vy_min); 101 vy = _mm256_min_ps(vy, vy_max); 102 103 __m128 vy_lo = _mm256_castps256_ps128(vy); 104 if (n & (4 * sizeof(float))) { 105 _mm_storeu_ps(y, vy_lo); 106 vy_lo = _mm256_extractf128_ps(vy, 1); 107 y += 4; 108 } 109 if (n & (2 * sizeof(float))) { 110 _mm_storel_pi((__m64*) y, vy_lo); 111 vy_lo = _mm_movehl_ps(vy_lo, vy_lo); 112 y += 2; 113 } 114 if (n & (1 * sizeof(float))) { 115 _mm_store_ss(y, vy_lo); 116 } 117 } 118} 119