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$assert BATCH_TILE >= 8 7$assert BATCH_TILE == 8 or BATCH_TILE % 16 == 0 8$SIMD_TILE = BATCH_TILE // 16 9$ABC = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZ" 10#include <assert.h> 11 12#include <immintrin.h> 13 14#include <xnnpack/common.h> 15#include <xnnpack/intrinsics-polyfill.h> 16#include <xnnpack/vlrelu.h> 17 18 19$XINT8_T = {"QS8": "int8_t", "QU8": "uint8_t"}[DATATYPE] 20$_MM_CVTEPX8_EPI16 = {"QS8": "_mm_cvtepi8_epi16", "QU8": "_mm_cvtepu8_epi16"}[DATATYPE] 21$_MM_PACKXS_EPI16 = {"QS8": "_mm_packs_epi16", "QU8": "_mm_packus_epi16"}[DATATYPE] 22$ISA = "avx" if AVX else "sse41" 23void xnn_${DATATYPE.lower()}_vlrelu_ukernel__${ISA}_x${BATCH_TILE}( 24 size_t n, 25 const ${XINT8_T}* x, 26 ${XINT8_T}* y, 27 const union xnn_${DATATYPE.lower()}_lrelu_params params[restrict XNN_MIN_ELEMENTS(1)]) XNN_OOB_READS 28{ 29 assert(n != 0); 30 assert(n % sizeof(${XINT8_T}) == 0); 31 assert(x != NULL); 32 assert(y != NULL); 33 34 $if AVX: 35 const __m128i vinput_zero_point = _mm_load_si128((const __m128i*) params->avx.input_zero_point); 36 const __m128i vpositive_multiplier = _mm_load_si128((const __m128i*) params->avx.positive_multiplier); 37 const __m128i vnegative_multiplier = _mm_load_si128((const __m128i*) params->avx.negative_multiplier); 38 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->avx.output_zero_point); 39 $else: 40 const __m128i vinput_zero_point = _mm_load_si128((const __m128i*) params->sse2.input_zero_point); 41 const __m128i vmultiplier_diff = _mm_load_si128((const __m128i*) params->sse2.multiplier_diff); 42 const __m128i vmultiplier_base = _mm_load_si128((const __m128i*) params->sse2.multiplier_base); 43 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->sse2.output_zero_point); 44 $if BATCH_TILE > 8: 45 for (; n >= ${BATCH_TILE} * sizeof(${XINT8_T}); n -= ${BATCH_TILE} * sizeof(${XINT8_T})) { 46 __m128i vacc${ABC[0]} = ${_MM_CVTEPX8_EPI16}(_mm_loadl_epi64((const __m128i*) x)); 47 $for N in range(1, 2*SIMD_TILE): 48 __m128i vacc${ABC[N]} = ${_MM_CVTEPX8_EPI16}(_mm_loadl_epi64((const __m128i*) (x + ${N * 8}))); 49 x += ${BATCH_TILE}; 50 51 $for N in range(2*SIMD_TILE): 52 __m128i vmultiplier${ABC[N]} = _mm_cmpgt_epi16(vacc${ABC[N]}, vinput_zero_point); 53 vacc${ABC[N]} = _mm_sub_epi16(vinput_zero_point, vacc${ABC[N]}); 54 55 $for N in range(2*SIMD_TILE): 56 $if AVX: 57 vmultiplier${ABC[N]} = _mm_blendv_epi8(vnegative_multiplier, vpositive_multiplier, vmultiplier${ABC[N]}); 58 $else: 59 vmultiplier${ABC[N]} = _mm_and_si128(vmultiplier${ABC[N]}, vmultiplier_diff); 60 vacc${ABC[N]} = _mm_slli_epi16(vacc${ABC[N]}, 7); 61 $if not AVX: 62 vmultiplier${ABC[N]} = _mm_xor_si128(vmultiplier${ABC[N]}, vmultiplier_base); 63 64 $for N in range(2*SIMD_TILE): 65 vacc${ABC[N]} = _mm_mulhrs_epi16(vacc${ABC[N]}, vmultiplier${ABC[N]}); 66 67 $for N in range(2*SIMD_TILE): 68 vacc${ABC[N]} = _mm_adds_epi16(vacc${ABC[N]}, voutput_zero_point); 69 70 $for N in range(SIMD_TILE): 71 const __m128i vy${ABC[N]} = ${_MM_PACKXS_EPI16}(vacc${ABC[2*N]}, vacc${ABC[2*N+1]}); 72 73 _mm_storeu_si128((__m128i*) y, vy${ABC[0]}); 74 $for N in range(1, SIMD_TILE): 75 _mm_storeu_si128((__m128i*) (y + ${N * 16}), vy${ABC[N]}); 76 y += ${BATCH_TILE}; 77 } 78 for (; n >= 8 * sizeof(${XINT8_T}); n -= 8 * sizeof(${XINT8_T})) { 79 __m128i vacc = ${_MM_CVTEPX8_EPI16}(_mm_loadl_epi64((const __m128i*) x)); 80 __m128i vmultiplier = _mm_cmpgt_epi16(vacc, vinput_zero_point); 81 vacc = _mm_sub_epi16(vinput_zero_point, vacc); 82 $if AVX: 83 vmultiplier = _mm_blendv_epi8(vnegative_multiplier, vpositive_multiplier, vmultiplier); 84 $else: 85 vmultiplier = _mm_and_si128(vmultiplier, vmultiplier_diff); 86 vacc = _mm_slli_epi16(vacc, 7); 87 $if not AVX: 88 vmultiplier = _mm_xor_si128(vmultiplier, vmultiplier_base); 89 vacc = _mm_mulhrs_epi16(vacc, vmultiplier); 90 vacc = _mm_adds_epi16(vacc, voutput_zero_point); 91 x += 8; 92 93 const __m128i vy = ${_MM_PACKXS_EPI16}(vacc, vacc); 94 _mm_storel_epi64((__m128i*) y, vy); 95 y += 8; 96 } 97 if XNN_UNLIKELY(n != 0) { 98 assert(n >= 1 * sizeof(${XINT8_T})); 99 assert(n <= 7 * sizeof(${XINT8_T})); 100 101 __m128i vacc = ${_MM_CVTEPX8_EPI16}(_mm_loadl_epi64((const __m128i*) x)); 102 __m128i vmultiplier = _mm_cmpgt_epi16(vacc, vinput_zero_point); 103 vacc = _mm_sub_epi16(vinput_zero_point, vacc); 104 $if AVX: 105 vmultiplier = _mm_blendv_epi8(vnegative_multiplier, vpositive_multiplier, vmultiplier); 106 $else: 107 vmultiplier = _mm_and_si128(vmultiplier, vmultiplier_diff); 108 vacc = _mm_slli_epi16(vacc, 7); 109 $if not AVX: 110 vmultiplier = _mm_xor_si128(vmultiplier, vmultiplier_base); 111 vacc = _mm_mulhrs_epi16(vacc, vmultiplier); 112 vacc = _mm_adds_epi16(vacc, voutput_zero_point); 113 114 __m128i vy = ${_MM_PACKXS_EPI16}(vacc, vacc); 115 if (n & (4 * sizeof(${XINT8_T}))) { 116 _mm_storeu_si32(y, vy); 117 vy = _mm_srli_epi64(vy, 32); 118 y += 4; 119 } 120 if (n & (2 * sizeof(${XINT8_T}))) { 121 _mm_storeu_si16(y, vy); 122 vy = _mm_srli_epi32(vy, 16); 123 y += 2; 124 } 125 if (n & (1 * sizeof(${XINT8_T}))) { 126 *y = (${XINT8_T}) _mm_extract_epi8(vy, 0); 127 } 128 } 129} 130