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$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ" 9#include <assert.h> 10 11#include <immintrin.h> 12 13#include <xnnpack/common.h> 14#include <xnnpack/vunary.h> 15 16 17$ISA = {0: "avx", 3: "fma3"}[FMA] 18void xnn_f32_vhswish_ukernel__${ISA}_x${BATCH_TILE}( 19 size_t n, 20 const float* x, 21 float* y, 22 const union xnn_f32_hswish_params params[restrict XNN_MIN_ELEMENTS(1)]) 23{ 24 assert(n != 0); 25 assert(n % sizeof(float) == 0); 26 27 const __m256 vsixth = _mm256_load_ps(params->avx.sixth); 28 const __m256 vhalf = _mm256_load_ps(params->avx.half); 29 const __m256 vone = _mm256_load_ps(params->avx.one); 30 const __m256 vzero = _mm256_setzero_ps(); 31 32 $if BATCH_TILE > 8: 33 for (; n >= ${BATCH_TILE} * sizeof(float); n -= ${BATCH_TILE} * sizeof(float)) { 34 const __m256 vx${ABC[0:8]} = _mm256_loadu_ps(x); 35 $for N in range(8, BATCH_TILE, 8): 36 const __m256 vx${ABC[N:N+8]} = _mm256_loadu_ps(x + ${N}); 37 x += ${BATCH_TILE}; 38 39 $if FMA == 3: 40 $for N in range(0, BATCH_TILE, 8): 41 __m256 vacc${ABC[N:N+8]} = _mm256_fmadd_ps(vx${ABC[N:N+8]}, vsixth, vhalf); 42 $else: 43 $for N in range(0, BATCH_TILE, 8): 44 __m256 vacc${ABC[N:N+8]} = _mm256_mul_ps(vx${ABC[N:N+8]}, vsixth); 45 46 $for N in range(0, BATCH_TILE, 8): 47 vacc${ABC[N:N+8]} = _mm256_add_ps(vacc${ABC[N:N+8]}, vhalf); 48 49 $for N in range(0, BATCH_TILE, 8): 50 vacc${ABC[N:N+8]} = _mm256_max_ps(vacc${ABC[N:N+8]}, vzero); 51 52 $for N in range(0, BATCH_TILE, 8): 53 vacc${ABC[N:N+8]} = _mm256_min_ps(vacc${ABC[N:N+8]}, vone); 54 55 $for N in range(0, BATCH_TILE, 8): 56 vacc${ABC[N:N+8]} = _mm256_mul_ps(vacc${ABC[N:N+8]}, vx${ABC[N:N+8]}); 57 58 _mm256_storeu_ps(y, vacc${ABC[0:8]}); 59 $for N in range(8, BATCH_TILE, 8): 60 _mm256_storeu_ps(y + ${N}, vacc${ABC[N:N+8]}); 61 y += ${BATCH_TILE}; 62 } 63 for (; n >= 8 * sizeof(float); n -= 8 * sizeof(float)) { 64 const __m256 vx = _mm256_loadu_ps(x); 65 x += 8; 66 $if FMA == 3: 67 __m256 vacc = _mm256_fmadd_ps(vx, vsixth, vhalf); 68 $else: 69 __m256 vacc = _mm256_mul_ps(vx, vsixth); 70 vacc = _mm256_add_ps(vacc, vhalf); 71 vacc = _mm256_max_ps(vacc, vzero); 72 vacc = _mm256_min_ps(vacc, vone); 73 vacc = _mm256_mul_ps(vacc, vx); 74 _mm256_storeu_ps(y, vacc); 75 y += 8; 76 } 77 if XNN_UNLIKELY(n != 0) { 78 assert(n >= 1 * sizeof(float)); 79 assert(n <= 7 * sizeof(float)); 80 const __m256i vmask = _mm256_loadu_si256((const __m256i*) ((uintptr_t) ¶ms->avx.mask_table[7] - n)); 81 82 const __m256 vx = _mm256_maskload_ps(x, vmask); 83 $if FMA == 3: 84 __m256 vacc = _mm256_fmadd_ps(vx, vsixth, vhalf); 85 $else: 86 __m256 vacc = _mm256_mul_ps(vx, vsixth); 87 vacc = _mm256_add_ps(vacc, vhalf); 88 vacc = _mm256_max_ps(vacc, vzero); 89 vacc = _mm256_min_ps(vacc, vone); 90 vacc = _mm256_mul_ps(vacc, vx); 91 92 __m128 vacc_lo = _mm256_castps256_ps128(vacc); 93 if (n & (4 * sizeof(float))) { 94 _mm_storeu_ps(y, vacc_lo); 95 vacc_lo = _mm256_extractf128_ps(vacc, 1); 96 y += 4; 97 } 98 if (n & (2 * sizeof(float))) { 99 _mm_storel_pi((__m64*) y, vacc_lo); 100 vacc_lo = _mm_movehl_ps(vacc_lo, vacc_lo); 101 y += 2; 102 } 103 if (n & (1 * sizeof(float))) { 104 _mm_store_ss(y, vacc_lo); 105 } 106 } 107} 108