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$assert BATCH_TILE % 8 == 0 7$assert BATCH_TILE >= 8 8$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ" 9$assert OP in ["ABS", "NEG", "SQR"] 10#include <assert.h> 11 12#include <immintrin.h> 13 14#include <xnnpack/common.h> 15#include <xnnpack/vunary.h> 16 17 18static const int32_t mask_table[14] = {-1, -1, -1, -1, -1, -1, -1, 0, 0, 0, 0, 0, 0, 0}; 19 20$_MM256_OP_PS = { 21$ "ABS": lambda x: "_mm256_and_ps(%s, vnonsign_mask)" % x, 22$ "NEG": lambda x: "_mm256_xor_ps(%s, vsign_mask)" % x, 23$ "SQR": lambda x: "_mm256_mul_ps(%s, %s)" % (x, x), 24$}[OP] 25$PARAMS = { 26$ "ABS": "const union xnn_f32_abs_params params[restrict XNN_MIN_ELEMENTS(1)]", 27$ "NEG": "const union xnn_f32_neg_params params[restrict XNN_MIN_ELEMENTS(1)]", 28$ "SQR": "const void* params", 29$}[OP] 30void xnn_f32_v${OP.lower()}_ukernel__avx_x${BATCH_TILE}( 31 size_t n, 32 const float* x, 33 float* y, 34 ${PARAMS}) 35{ 36 assert(n != 0); 37 assert(n % sizeof(float) == 0); 38 assert(x != NULL); 39 assert(y != NULL); 40 41 $if OP == "ABS": 42 const __m256 vnonsign_mask = _mm256_broadcast_ps((const __m128*) params->sse.nonsign_mask); 43 $elif OP == "NEG": 44 const __m256 vsign_mask = _mm256_broadcast_ps((const __m128*) params->sse.sign_mask); 45 for (; n >= ${BATCH_TILE} * sizeof(float); n -= ${BATCH_TILE} * sizeof(float)) { 46 const __m256 vx${ABC[0:8]} = _mm256_loadu_ps(x); 47 $for N in range(8, BATCH_TILE, 8): 48 const __m256 vx${ABC[N:N+8]} = _mm256_loadu_ps(x + ${N}); 49 x += ${BATCH_TILE}; 50 51 $for N in range(0, BATCH_TILE, 8): 52 const __m256 vy${ABC[N:N+8]} = ${_MM256_OP_PS("vx" + ABC[N:N+8])}; 53 54 _mm256_storeu_ps(y, vy${ABC[0:8]}); 55 $for N in range(8, BATCH_TILE, 8): 56 _mm256_storeu_ps(y + ${N}, vy${ABC[N:N+8]}); 57 y += ${BATCH_TILE}; 58 } 59 $if BATCH_TILE > 8: 60 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) { 61 const __m256 vx = _mm256_loadu_ps(x); 62 x += 8; 63 const __m256 vy = ${_MM256_OP_PS("vx")}; 64 _mm256_storeu_ps(y, vy); 65 y += 8; 66 } 67 if XNN_UNLIKELY(n != 0) { 68 assert(n >= 1 * sizeof(float)); 69 assert(n <= 7 * sizeof(float)); 70 __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) &mask_table[7] - n)); 71 72 const __m256 vx = _mm256_maskload_ps(x, vmask); 73 const __m256 vy = ${_MM256_OP_PS("vx")}; 74 75 // _mm256_maskstore_ps(y, vmask, vy) could be used here, but triggers msan failures (probably an msan bug). 76 __m128 vy_lo = _mm256_castps256_ps128(vy); 77 if (n & (4 * sizeof(float))) { 78 _mm_storeu_ps(y, vy_lo); 79 vy_lo = _mm256_extractf128_ps(vy, 1); 80 y += 4; 81 } 82 if (n & (2 * sizeof(float))) { 83 _mm_storel_pi((__m64*) y, vy_lo); 84 vy_lo = _mm_movehl_ps(vy_lo, vy_lo); 85 y += 2; 86 } 87 if (n & (1 * sizeof(float))) { 88 _mm_store_ss(y, vy_lo); 89 } 90 } 91} 92