1// Copyright 2022 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$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ" 9#include <assert.h> 10 11#include <immintrin.h> 12 13#include <xnnpack/common.h> 14#include <xnnpack/vunary.h> 15 16 17void xnn_f16_vhswish_ukernel__f16c_x${BATCH_TILE}( 18 size_t n, 19 const void* restrict x_ptr, 20 void* restrict y_ptr, 21 const union xnn_f16_hswish_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS 22{ 23 assert(n != 0); 24 assert(n % sizeof(uint16_t) == 0); 25 26 const uint16_t* x = (const uint16_t*) x_ptr; 27 uint16_t* y = (uint16_t*) y_ptr; 28 29 const __m256 vsixth = _mm256_load_ps(params->avx.sixth); 30 const __m256 vthree = _mm256_load_ps(params->avx.three); 31 const __m128i vsix = _mm_load_si128((const __m128i*) params->avx.six); 32 const __m128i vzero = _mm_setzero_si128(); 33 34 $if BATCH_TILE > 8: 35 for (; n >= ${BATCH_TILE} * sizeof(uint16_t); n -= ${BATCH_TILE} * sizeof(uint16_t)) { 36 __m256 vx${ABC[0:8]} = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) x)); 37 $for N in range(8, BATCH_TILE, 8): 38 __m256 vx${ABC[N:N+8]} = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (x + ${N}))); 39 x += ${BATCH_TILE}; 40 41 $for N in range(0, BATCH_TILE, 8): 42 __m128i vacc${ABC[N:N+8]} = _mm256_cvtps_ph(_mm256_add_ps(vx${ABC[N:N+8]}, vthree), _MM_FROUND_NO_EXC); 43 vx${ABC[N:N+8]} = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_mul_ps(vx${ABC[N:N+8]}, vsixth), _MM_FROUND_NO_EXC)); 44 45 $for N in range(0, BATCH_TILE, 8): 46 vacc${ABC[N:N+8]} = _mm_max_epi16(vacc${ABC[N:N+8]}, vzero); 47 48 $for N in range(0, BATCH_TILE, 8): 49 vacc${ABC[N:N+8]} = _mm_min_epi16(vacc${ABC[N:N+8]}, vsix); 50 51 $for N in range(0, BATCH_TILE, 8): 52 vacc${ABC[N:N+8]} = _mm256_cvtps_ph(_mm256_mul_ps(_mm256_cvtph_ps(vacc${ABC[N:N+8]}), vx${ABC[N:N+8]}), _MM_FROUND_NO_EXC); 53 54 _mm_storeu_si128((__m128i*) y, vacc${ABC[0:8]}); 55 $for N in range(8, BATCH_TILE, 8): 56 _mm_storeu_si128((__m128i*) (y + ${N}), vacc${ABC[N:N+8]}); 57 y += ${BATCH_TILE}; 58 } 59 for (; n >= 8 * sizeof(uint16_t); n -= 8 * sizeof(uint16_t)) { 60 __m256 vx = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) x)); 61 x += 8; 62 __m128i vacc = _mm256_cvtps_ph(_mm256_add_ps(vx, vthree), _MM_FROUND_NO_EXC); 63 vx = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_mul_ps(vx, vsixth), _MM_FROUND_NO_EXC)); 64 vacc = _mm_max_epi16(vacc, vzero); 65 vacc = _mm_min_epi16(vacc, vsix); 66 vacc = _mm256_cvtps_ph(_mm256_mul_ps(_mm256_cvtph_ps(vacc), vx), _MM_FROUND_NO_EXC); 67 _mm_storeu_si128((__m128i*) y, vacc); 68 y += 8; 69 } 70 if XNN_UNLIKELY(n != 0) { 71 __m256 vx = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) x)); 72 __m128i vacc = _mm256_cvtps_ph(_mm256_add_ps(vx, vthree), _MM_FROUND_NO_EXC); 73 vx = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_mul_ps(vx, vsixth), _MM_FROUND_NO_EXC)); 74 vacc = _mm_max_epi16(vacc, vzero); 75 vacc = _mm_min_epi16(vacc, vsix); 76 vacc = _mm256_cvtps_ph(_mm256_mul_ps(_mm256_cvtph_ps(vacc), vx), _MM_FROUND_NO_EXC); 77 78 if (n & (4 * sizeof(uint16_t))) { 79 _mm_storel_epi64((__m128i*) y, vacc); 80 vacc = _mm_unpackhi_epi64(vacc, vacc); 81 y += 4; 82 } 83 if (n & (2 * sizeof(uint16_t))) { 84 *((uint32_t*) y) = (uint32_t) _mm_cvtsi128_si32(vacc); 85 vacc = _mm_srli_epi64(vacc, 32); 86 y += 2; 87 } 88 if (n & (1 * sizeof(uint16_t))) { 89 *y = (uint16_t) _mm_extract_epi16(vacc, 0); 90 } 91 } 92} 93