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$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ" 9#include <assert.h> 10 11#include <immintrin.h> 12 13#include <xnnpack/common.h> 14#include <xnnpack/vunary.h> 15 16 17void xnn_f32_vsqrt_ukernel__avx_sqrt_x${BATCH_TILE}( 18 size_t n, 19 const float* x, 20 float* y, 21 const union xnn_f32_sqrt_params params[restrict XNN_MIN_ELEMENTS(1)]) 22{ 23 assert(n != 0); 24 assert(n % sizeof(float) == 0); 25 26 $if BATCH_TILE > 8: 27 for (; n >= ${BATCH_TILE} * sizeof(float); n -= ${BATCH_TILE} * sizeof(float)) { 28 const __m256 vx${ABC[0:8]} = _mm256_loadu_ps(x); 29 $for N in range(8, BATCH_TILE, 8): 30 const __m256 vx${ABC[N:N+8]} = _mm256_loadu_ps(x + ${N}); 31 x += ${BATCH_TILE}; 32 33 $for N in range(0, BATCH_TILE, 8): 34 const __m256 vy${ABC[N:N+8]} = _mm256_sqrt_ps(vx${ABC[N:N+8]}); 35 36 _mm256_storeu_ps(y, vy${ABC[0:8]}); 37 $for N in range(8, BATCH_TILE, 8): 38 _mm256_storeu_ps(y + ${N}, vy${ABC[N:N+8]}); 39 y += ${BATCH_TILE}; 40 } 41 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) { 42 const __m256 vx = _mm256_loadu_ps(x); 43 x += 8; 44 const __m256 vy = _mm256_sqrt_ps(vx); 45 _mm256_storeu_ps(y, vy); 46 y += 8; 47 } 48 if XNN_UNLIKELY(n != 0) { 49 assert(n >= 1 * sizeof(float)); 50 assert(n <= 7 * sizeof(float)); 51 const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) ¶ms->avx.mask_table[7] - n)); 52 53 const __m256 vx = _mm256_maskload_ps(x, vmask); 54 const __m256 vy = _mm256_sqrt_ps(vx); 55 56 __m128 vy_lo = _mm256_castps256_ps128(vy); 57 if (n & (4 * sizeof(float))) { 58 _mm_storeu_ps(y, vy_lo); 59 vy_lo = _mm256_extractf128_ps(vy, 1); 60 y += 4; 61 } 62 if (n & (2 * sizeof(float))) { 63 _mm_storel_pi((__m64*) y, vy_lo); 64 vy_lo = _mm_movehl_ps(vy_lo, vy_lo); 65 y += 2; 66 } 67 if (n & (1 * sizeof(float))) { 68 _mm_store_ss(y, vy_lo); 69 } 70 } 71} 72