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$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ" 9#include <assert.h> 10 11#include <immintrin.h> 12 13#include <xnnpack/common.h> 14#include <xnnpack/vunary.h> 15 16 17void xnn_f32_vclamp_ukernel__avx_x${BATCH_TILE}( 18 size_t n, 19 const float* x, 20 float* y, 21 const union xnn_f32_minmax_params params[restrict XNN_MIN_ELEMENTS(1)]) 22{ 23 assert(n != 0); 24 assert(n % sizeof(float) == 0); 25 assert(x != NULL); 26 assert(y != NULL); 27 28 const __m256 vy_min = _mm256_load_ps(params->avx.min); 29 const __m256 vy_max = _mm256_load_ps(params->avx.max); 30 31 for (; n >= ${BATCH_TILE} * sizeof(float); n -= ${BATCH_TILE} * sizeof(float)) { 32 __m256 vacc${ABC[0:8]} = _mm256_loadu_ps(x); 33 $for N in range(8, BATCH_TILE, 8): 34 __m256 vacc${ABC[N:N+8]} = _mm256_loadu_ps(x + ${N}); 35 x += ${BATCH_TILE}; 36 37 $for N in range(0, BATCH_TILE, 8): 38 vacc${ABC[N:N+8]} = _mm256_max_ps(vacc${ABC[N:N+8]}, vy_min); 39 40 $for N in range(0, BATCH_TILE, 8): 41 vacc${ABC[N:N+8]} = _mm256_min_ps(vacc${ABC[N:N+8]}, vy_max); 42 43 _mm256_storeu_ps(y, vacc${ABC[0:8]}); 44 $for N in range(8, BATCH_TILE, 8): 45 _mm256_storeu_ps(y + ${N}, vacc${ABC[N:N+8]}); 46 y += ${BATCH_TILE}; 47 } 48 $if BATCH_TILE > 8: 49 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) { 50 __m256 vacc = _mm256_loadu_ps(x); 51 x += 8; 52 53 vacc = _mm256_max_ps(vacc, vy_min); 54 vacc = _mm256_min_ps(vacc, vy_max); 55 56 _mm256_storeu_ps(y, vacc); 57 y += 8; 58 } 59 if XNN_UNLIKELY(n != 0) { 60 assert(n >= 1 * sizeof(float)); 61 assert(n <= 7 * sizeof(float)); 62 const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) ¶ms->avx.mask_table[7] - n)); 63 64 __m256 vacc = _mm256_maskload_ps(x, vmask); 65 vacc = _mm256_max_ps(vacc, vy_min); 66 vacc = _mm256_min_ps(vacc, vy_max); 67 68 __m128 vacc_lo = _mm256_castps256_ps128(vacc); 69 if (n & (4 * sizeof(float))) { 70 _mm_storeu_ps(y, vacc_lo); 71 vacc_lo = _mm256_extractf128_ps(vacc, 1); 72 y += 4; 73 } 74 if (n & (2 * sizeof(float))) { 75 _mm_storel_pi((__m64*) y, vacc_lo); 76 vacc_lo = _mm_movehl_ps(vacc_lo, vacc_lo); 77 y += 2; 78 } 79 if (n & (1 * sizeof(float))) { 80 _mm_store_ss(y, vacc_lo); 81 } 82 } 83} 84