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 18$_MM256_OP_PS = { 19$ "ABS": lambda x: "_mm256_and_ps(%s, vnonsign_mask)" % x, 20$ "NEG": lambda x: "_mm256_xor_ps(%s, vsign_mask)" % x, 21$ "SQR": lambda x: "_mm256_mul_ps(%s, %s)" % (x, x), 22$}[OP] 23$PARAMS = { 24$ "ABS": "xnn_f32_abs_params", 25$ "NEG": "xnn_f32_neg_params", 26$ "SQR": "xnn_f32_default_params", 27$}[OP] 28void xnn_f32_v${OP.lower()}_ukernel__avx_x${BATCH_TILE}( 29 size_t n, 30 const float* x, 31 float* y, 32 const union ${PARAMS} params[restrict XNN_MIN_ELEMENTS(1)]) 33{ 34 assert(n != 0); 35 assert(n % sizeof(float) == 0); 36 assert(x != NULL); 37 assert(y != NULL); 38 39 $if OP == "ABS": 40 const __m256 vnonsign_mask = _mm256_load_ps(params->avx.nonsign_mask); 41 $elif OP == "NEG": 42 const __m256 vsign_mask = _mm256_load_ps(params->sse.sign_mask); 43 for (; n >= ${BATCH_TILE} * sizeof(float); n -= ${BATCH_TILE} * sizeof(float)) { 44 const __m256 vx${ABC[0:8]} = _mm256_loadu_ps(x); 45 $for N in range(8, BATCH_TILE, 8): 46 const __m256 vx${ABC[N:N+8]} = _mm256_loadu_ps(x + ${N}); 47 x += ${BATCH_TILE}; 48 49 $for N in range(0, BATCH_TILE, 8): 50 const __m256 vy${ABC[N:N+8]} = ${_MM256_OP_PS("vx" + ABC[N:N+8])}; 51 52 _mm256_storeu_ps(y, vy${ABC[0:8]}); 53 $for N in range(8, BATCH_TILE, 8): 54 _mm256_storeu_ps(y + ${N}, vy${ABC[N:N+8]}); 55 y += ${BATCH_TILE}; 56 } 57 $if BATCH_TILE > 8: 58 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) { 59 const __m256 vx = _mm256_loadu_ps(x); 60 x += 8; 61 const __m256 vy = ${_MM256_OP_PS("vx")}; 62 _mm256_storeu_ps(y, vy); 63 y += 8; 64 } 65 if XNN_UNLIKELY(n != 0) { 66 assert(n >= 1 * sizeof(float)); 67 assert(n <= 7 * sizeof(float)); 68 const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) ¶ms->avx.mask_table[7] - n)); 69 70 const __m256 vx = _mm256_maskload_ps(x, vmask); 71 const __m256 vy = ${_MM256_OP_PS("vx")}; 72 73 __m128 vy_lo = _mm256_castps256_ps128(vy); 74 if (n & (4 * sizeof(float))) { 75 _mm_storeu_ps(y, vy_lo); 76 vy_lo = _mm256_extractf128_ps(vy, 1); 77 y += 4; 78 } 79 if (n & (2 * sizeof(float))) { 80 _mm_storel_pi((__m64*) y, vy_lo); 81 vy_lo = _mm_movehl_ps(vy_lo, vy_lo); 82 y += 2; 83 } 84 if (n & (1 * sizeof(float))) { 85 _mm_store_ss(y, vy_lo); 86 } 87 } 88} 89