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/vcvt.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_${DATATYPE.lower()}_vcvt_ukernel__${ISA}_x${BATCH_TILE}( 24 size_t n, 25 const ${XINT8_T}* x, 26 ${XINT8_T}* y, 27 const union xnn_${DATATYPE.lower()}_cvt_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 const __m128i vinput_zero_point = _mm_load_si128((const __m128i*) params->ssse3.input_zero_point); 35 const __m128i vmultiplier = _mm_load_si128((const __m128i*) params->ssse3.multiplier); 36 const __m128i voutput_zero_point = _mm_load_si128((const __m128i*) params->ssse3.output_zero_point); 37 $if BATCH_TILE > 8: 38 for (; n >= ${BATCH_TILE} * sizeof(${XINT8_T}); n -= ${BATCH_TILE} * sizeof(${XINT8_T})) { 39 __m128i vacc${ABC[0]} = ${_MM_CVTEPX8_EPI16}(_mm_loadl_epi64((const __m128i*) x)); 40 $for N in range(1, 2*SIMD_TILE): 41 __m128i vacc${ABC[N]} = ${_MM_CVTEPX8_EPI16}(_mm_loadl_epi64((const __m128i*) (x + ${N * 8}))); 42 x += ${BATCH_TILE}; 43 44 $for N in range(2*SIMD_TILE): 45 vacc${ABC[N]} = _mm_sub_epi16(vinput_zero_point, vacc${ABC[N]}); 46 47 $for N in range(2*SIMD_TILE): 48 vacc${ABC[N]} = _mm_slli_epi16(vacc${ABC[N]}, 7); 49 50 $for N in range(2*SIMD_TILE): 51 vacc${ABC[N]} = _mm_mulhrs_epi16(vacc${ABC[N]}, vmultiplier); 52 53 $for N in range(2*SIMD_TILE): 54 vacc${ABC[N]} = _mm_adds_epi16(vacc${ABC[N]}, voutput_zero_point); 55 56 $for N in range(SIMD_TILE): 57 const __m128i vy${ABC[N]} = ${_MM_PACKXS_EPI16}(vacc${ABC[2*N]}, vacc${ABC[2*N+1]}); 58 59 _mm_storeu_si128((__m128i*) y, vy${ABC[0]}); 60 $for N in range(1, SIMD_TILE): 61 _mm_storeu_si128((__m128i*) (y + ${N * 16}), vy${ABC[N]}); 62 y += ${BATCH_TILE}; 63 } 64 for (; n >= 8 * sizeof(${XINT8_T}); n -= 8 * sizeof(${XINT8_T})) { 65 __m128i vacc = ${_MM_CVTEPX8_EPI16}(_mm_loadl_epi64((const __m128i*) x)); 66 vacc = _mm_sub_epi16(vinput_zero_point, vacc); 67 vacc = _mm_slli_epi16(vacc, 7); 68 vacc = _mm_mulhrs_epi16(vacc, vmultiplier); 69 vacc = _mm_adds_epi16(vacc, voutput_zero_point); 70 x += 8; 71 72 const __m128i vy = ${_MM_PACKXS_EPI16}(vacc, vacc); 73 _mm_storel_epi64((__m128i*) y, vy); 74 y += 8; 75 } 76 if XNN_UNLIKELY(n != 0) { 77 assert(n >= 1 * sizeof(${XINT8_T})); 78 assert(n <= 7 * sizeof(${XINT8_T})); 79 80 __m128i vacc = ${_MM_CVTEPX8_EPI16}(_mm_loadl_epi64((const __m128i*) x)); 81 vacc = _mm_sub_epi16(vinput_zero_point, vacc); 82 vacc = _mm_slli_epi16(vacc, 7); 83 vacc = _mm_mulhrs_epi16(vacc, vmultiplier); 84 vacc = _mm_adds_epi16(vacc, voutput_zero_point); 85 86 __m128i vy = ${_MM_PACKXS_EPI16}(vacc, vacc); 87 if (n & (4 * sizeof(${XINT8_T}))) { 88 _mm_storeu_si32(y, vy); 89 vy = _mm_srli_epi64(vy, 32); 90 y += 4; 91 } 92 if (n & (2 * sizeof(${XINT8_T}))) { 93 _mm_storeu_si16(y, vy); 94 vy = _mm_srli_epi32(vy, 16); 95 y += 2; 96 } 97 if (n & (1 * sizeof(${XINT8_T}))) { 98 *y = (${XINT8_T}) _mm_extract_epi8(vy, 0); 99 } 100 } 101} 102