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$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ" 9#include <assert.h> 10 11#include <immintrin.h> 12 13#include <xnnpack/common.h> 14#include <xnnpack/intrinsics-polyfill.h> 15#include <xnnpack/vunary.h> 16 17 18void xnn_f16_vhswish_ukernel__f16c_x${BATCH_TILE}( 19 size_t n, 20 const void* restrict x_ptr, 21 void* restrict y_ptr, 22 const union xnn_f16_hswish_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS 23{ 24 assert(n != 0); 25 assert(n % sizeof(uint16_t) == 0); 26 27 const uint16_t* x = (const uint16_t*) x_ptr; 28 uint16_t* y = (uint16_t*) y_ptr; 29 30 const __m256 vsixth = _mm256_load_ps(params->avx.sixth); 31 const __m256 vthree = _mm256_load_ps(params->avx.three); 32 const __m128i vsix = _mm_load_si128((const __m128i*) params->avx.six); 33 const __m128i vzero = _mm_setzero_si128(); 34 35 $if BATCH_TILE > 8: 36 for (; n >= ${BATCH_TILE} * sizeof(uint16_t); n -= ${BATCH_TILE} * sizeof(uint16_t)) { 37 __m256 vx${ABC[0:8]} = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) x)); 38 $for N in range(8, BATCH_TILE, 8): 39 __m256 vx${ABC[N:N+8]} = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) (x + ${N}))); 40 x += ${BATCH_TILE}; 41 42 $for N in range(0, BATCH_TILE, 8): 43 __m128i vacc${ABC[N:N+8]} = _mm256_cvtps_ph(_mm256_add_ps(vx${ABC[N:N+8]}, vthree), _MM_FROUND_NO_EXC); 44 vx${ABC[N:N+8]} = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_mul_ps(vx${ABC[N:N+8]}, vsixth), _MM_FROUND_NO_EXC)); 45 46 $for N in range(0, BATCH_TILE, 8): 47 vacc${ABC[N:N+8]} = _mm_max_epi16(vacc${ABC[N:N+8]}, vzero); 48 49 $for N in range(0, BATCH_TILE, 8): 50 vacc${ABC[N:N+8]} = _mm_min_epi16(vacc${ABC[N:N+8]}, vsix); 51 52 $for N in range(0, BATCH_TILE, 8): 53 vacc${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); 54 55 _mm_storeu_si128((__m128i*) y, vacc${ABC[0:8]}); 56 $for N in range(8, BATCH_TILE, 8): 57 _mm_storeu_si128((__m128i*) (y + ${N}), vacc${ABC[N:N+8]}); 58 y += ${BATCH_TILE}; 59 } 60 for (; n >= 8 * sizeof(uint16_t); n -= 8 * sizeof(uint16_t)) { 61 __m256 vx = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) x)); 62 x += 8; 63 __m128i vacc = _mm256_cvtps_ph(_mm256_add_ps(vx, vthree), _MM_FROUND_NO_EXC); 64 vx = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_mul_ps(vx, vsixth), _MM_FROUND_NO_EXC)); 65 vacc = _mm_max_epi16(vacc, vzero); 66 vacc = _mm_min_epi16(vacc, vsix); 67 vacc = _mm256_cvtps_ph(_mm256_mul_ps(_mm256_cvtph_ps(vacc), vx), _MM_FROUND_NO_EXC); 68 _mm_storeu_si128((__m128i*) y, vacc); 69 y += 8; 70 } 71 if XNN_UNLIKELY(n != 0) { 72 __m256 vx = _mm256_cvtph_ps(_mm_loadu_si128((const __m128i*) x)); 73 __m128i vacc = _mm256_cvtps_ph(_mm256_add_ps(vx, vthree), _MM_FROUND_NO_EXC); 74 vx = _mm256_cvtph_ps(_mm256_cvtps_ph(_mm256_mul_ps(vx, vsixth), _MM_FROUND_NO_EXC)); 75 vacc = _mm_max_epi16(vacc, vzero); 76 vacc = _mm_min_epi16(vacc, vsix); 77 vacc = _mm256_cvtps_ph(_mm256_mul_ps(_mm256_cvtph_ps(vacc), vx), _MM_FROUND_NO_EXC); 78 79 if (n & (4 * sizeof(uint16_t))) { 80 _mm_storel_epi64((__m128i*) y, vacc); 81 vacc = _mm_unpackhi_epi64(vacc, vacc); 82 y += 4; 83 } 84 if (n & (2 * sizeof(uint16_t))) { 85 _mm_storeu_si32(y, vacc); 86 vacc = _mm_srli_epi64(vacc, 32); 87 y += 2; 88 } 89 if (n & (1 * sizeof(uint16_t))) { 90 *y = (uint16_t) _mm_extract_epi16(vacc, 0); 91 } 92 } 93} 94