Path: blob/main/test/benchmark/benchmark_sse2.cpp
6174 views
/*1* Copyright 2020 The Emscripten Authors. All rights reserved.2* Emscripten is available under two separate licenses, the MIT license and the3* University of Illinois/NCSA Open Source License. Both these licenses can be4* found in the LICENSE file.5*/6#include <emmintrin.h>7#include "benchmark_sse.h"89int main() {10printf ("{ \"workload\": %u, \"results\": [\n", N);1112float *src_flt = alloc_float_buffer();13float *src2_flt = alloc_float_buffer();14float *dst_flt = alloc_float_buffer();15for(int i = 0; i < N; ++i) src_flt[i] = (float)(1.0 + (double)rand() / RAND_MAX);16for(int i = 0; i < N; ++i) src2_flt[i] = (float)(1.0 + (double)rand() / RAND_MAX);1718double *src_dbl = alloc_double_buffer();19double *src2_dbl = alloc_double_buffer();20double *dst_dbl = alloc_double_buffer();21for(int i = 0; i < N; ++i) src_dbl[i] = 1.0 + (double)rand() / RAND_MAX;22for(int i = 0; i < N; ++i) src2_dbl[i] = 1.0 + (double)rand() / RAND_MAX;2324int *src_int = alloc_int_buffer();25int *src2_int = alloc_int_buffer();26int *dst_int = alloc_int_buffer();27for(int i = 0; i < N; ++i) src_int[i] = rand();28for(int i = 0; i < N; ++i) src2_int[i] = rand();2930float scalarTime = 0.f;3132// Benchmarks start:33SETCHART("load");34START();35for(int i = 0; i < N; ++i)36dst_dbl[i] = src_dbl[i];37ENDSCALAR(checksum_dst(dst_dbl), "scalar");3839LOAD_STORE_D("_mm_load_pd", _mm_load_pd, 0, _mm_store_pd, double*, 0, 2);40LOAD_STORE_D("_mm_load_pd1", _mm_load_pd1, 1, _mm_store_pd, double*, 0, 2);41LOAD_STORE_D("_mm_load_sd", _mm_load_sd, 1, _mm_store_pd, double*, 0, 2);42// _mm_load_si12843LOAD_STORE_D("_mm_load1_pd", _mm_load1_pd, 1, _mm_store_pd, double*, 0, 2);4445__m128d tempReg = _mm_set_pd(1.0, 2.0);46LOAD_STORE_M64("_mm_loadh_pd", tempReg, _mm_loadh_pd, double*, 1, _mm_store_pd, double*, 0, 2);47// _mm_loadl_epi6448LOAD_STORE_M64("_mm_loadl_pd", tempReg, _mm_loadh_pd, double*, 1, _mm_store_pd, double*, 0, 2);4950LOAD_STORE_D("_mm_loadr_pd", _mm_loadr_pd, 0, _mm_store_pd, double*, 0, 2);51LOAD_STORE_D("_mm_loadu_pd", _mm_loadu_pd, 1, _mm_store_pd, double*, 0, 2);52// _mm_loadu_si1285354SETCHART("set");55/* _mm_set_epi1656_mm_set_epi3257_mm_set_epi6458_mm_set_epi64x59_mm_set_epi8 */60SET_STORE_D("_mm_set_pd", _mm_set_pd(src_dbl[i+2], src_dbl[i+0]));61//SET_STORE_D("_mm_set_pd1", _mm_set_pd1(src_dbl[i]));62SET_STORE_D("_mm_set_sd", _mm_set_sd(src_dbl[i]));63/* _mm_set1_epi1664_mm_set1_epi3265_mm_set1_epi6466_mm_set1_epi64x67_mm_set1_epi8 */68SET_STORE_D("_mm_set1_pd", _mm_set1_pd(src_dbl[i]));69/* _mm_setr_epi1670_mm_setr_epi3271_mm_setr_epi6472_mm_setr_epi8 */73SET_STORE_D("_mm_setr_pd", _mm_set_pd(src_dbl[i+2], src_dbl[i+0]));74SET_STORE_D("_mm_setzero_pd", _mm_setzero_pd());75// _mm_setzero_si1287677SETCHART("move");78// _mm_move_epi6479SET_STORE_D("_mm_move_sd", _mm_move_sd(_mm_load_pd(src_dbl+i), _mm_load_pd(src2_dbl+i)));8081SETCHART("store");82// _mm_maskmoveu_si12883LOAD_STORE_D("_mm_store_pd", _mm_load_pd, 0, _mm_store_pd, double*, 0, 2);84// LOAD_STORE_D("_mm_store_pd1", _mm_load_pd, 0, _mm_store_pd1, double*, 0);85LOAD_STORE_D("_mm_store_sd", _mm_load_pd, 0, _mm_store_sd, double*, 1, 2);86// _mm_store_si12887// _mm_store1_pd88LOAD_STORE_64_D("_mm_storeh_pi", _mm_load_pd, 0, _mm_storeh_pi, 1, 2);89// _mm_storel_epi6490LOAD_STORE_64_D("_mm_storel_pi", _mm_load_pd, 0, _mm_storel_pi, 1, 2);91LOAD_STORE_D("_mm_storer_pd", _mm_load_pd, 0, _mm_storer_pd, double*, 0, 2);92LOAD_STORE_D("_mm_storeu_pd", _mm_load_pd, 0, _mm_storeu_pd, double*, 1, 2);93// _mm_storeu_si12894LOAD_STORE_D("_mm_stream_pd", _mm_load_pd, 0, _mm_stream_pd, double*, 0, 2);95// _mm_stream_si12896// _mm_stream_si3297// _mm_stream_si649899SETCHART("arithmetic");100// _mm_add_epi16101// _mm_add_epi32102// _mm_add_epi64103// _mm_add_epi8104START(); dst_dbl[0] = src_dbl[0]; dst_dbl[1] = src_dbl[1]; dst_dbl[2] = src_dbl[2]; dst_dbl[3] = src_dbl[3]; for(int i = 0; i < N; ++i) { dst_dbl[0] += src2_dbl[0]; dst_dbl[1] += src2_dbl[1]; dst_dbl[2] += src2_dbl[2]; dst_dbl[3] += src2_dbl[3]; } ENDSCALAR(checksum_dst(dst_dbl), "scalar add");105BINARYOP_D_DD("_mm_add_pd", _mm_add_pd, _mm_load_pd(src_dbl), _mm_load_pd(src2_dbl));106BINARYOP_D_DD("_mm_add_sd", _mm_add_sd, _mm_load_pd(src_dbl), _mm_load_pd(src2_dbl));107// _mm_adds_epi16108// _mm_adds_epi8109// _mm_adds_epu16110// _mm_adds_epu8111START(); dst_dbl[0] = src_dbl[0]; dst_dbl[1] = src_dbl[1]; dst_dbl[2] = src_dbl[2]; dst_dbl[3] = src_dbl[3]; for(int i = 0; i < N; ++i) { dst_dbl[0] /= src2_dbl[0]; dst_dbl[1] /= src2_dbl[1]; dst_dbl[2] /= src2_dbl[2]; dst_dbl[3] /= src2_dbl[3]; } ENDSCALAR(checksum_dst(dst_dbl), "scalar div");112BINARYOP_D_DD("_mm_div_pd", _mm_div_pd, _mm_load_pd(src_dbl), _mm_load_pd(src2_dbl));113BINARYOP_D_DD("_mm_div_sd", _mm_div_sd, _mm_load_pd(src_dbl), _mm_load_pd(src2_dbl));114// _mm_madd_epi16115// _mm_mul_epu32116START(); dst_dbl[0] = src_dbl[0]; dst_dbl[1] = src_dbl[1]; dst_dbl[2] = src_dbl[2]; dst_dbl[3] = src_dbl[3]; for(int i = 0; i < N; ++i) { dst_dbl[0] *= src2_dbl[0]; dst_dbl[1] *= src2_dbl[1]; dst_dbl[2] *= src2_dbl[2]; dst_dbl[3] *= src2_dbl[3]; } ENDSCALAR(checksum_dst(dst_dbl), "scalar mul");117BINARYOP_D_DD("_mm_mul_pd", _mm_mul_pd, _mm_load_pd(src_dbl), _mm_load_pd(src2_dbl));118BINARYOP_D_DD("_mm_mul_sd", _mm_mul_sd, _mm_load_pd(src_dbl), _mm_load_pd(src2_dbl));119// _mm_mulhi_epi16120// _mm_mulhi_epu16121// _mm_mullo_epi16122// _mm_sad_epu8123// _mm_sub_epi16124// _mm_sub_epi32125// _mm_sub_epi64126// _mm_sub_epi8127START(); dst_dbl[0] = src_dbl[0]; dst_dbl[1] = src_dbl[1]; dst_dbl[2] = src_dbl[2]; dst_dbl[3] = src_dbl[3]; for(int i = 0; i < N; ++i) { dst_dbl[0] -= src2_dbl[0]; dst_dbl[1] -= src2_dbl[1]; dst_dbl[2] -= src2_dbl[2]; dst_dbl[3] -= src2_dbl[3]; } ENDSCALAR(checksum_dst(dst_dbl), "scalar sub");128BINARYOP_D_DD("_mm_sub_pd", _mm_sub_pd, _mm_load_pd(src_dbl), _mm_load_pd(src2_dbl));129BINARYOP_D_DD("_mm_sub_sd", _mm_sub_sd, _mm_load_pd(src_dbl), _mm_load_pd(src2_dbl));130// _mm_subs_epi16131// _mm_subs_epi8132// _mm_subs_epu16133// _mm_subs_epu8134135SETCHART("roots");136START(); dst_dbl[0] = src_dbl[0]; dst_dbl[1] = src_dbl[1]; dst_dbl[2] = src_dbl[2]; dst_dbl[3] = src_dbl[3]; for(int i = 0; i < N; ++i) { dst_dbl[0] = sqrt(dst_dbl[0]); dst_dbl[1] = sqrt(dst_dbl[1]); dst_dbl[2] = sqrt(dst_dbl[2]); dst_dbl[3] = sqrt(dst_dbl[3]); } ENDSCALAR(checksum_dst(dst_dbl), "scalar sqrt");137UNARYOP_D_D("_mm_sqrt_pd", _mm_sqrt_pd, _mm_load_pd(src_dbl));138// UNARYOP_D_D("_mm_sqrt_sd", _mm_sqrt_sd, _mm_load_pd(src_dbl));139140SETCHART("logical");141START(); dst_dbl[0] = src_dbl[0]; dst_dbl[1] = src_dbl[1]; dst_dbl[2] = src_dbl[2]; dst_dbl[3] = src_dbl[3]; for(int i = 0; i < N; ++i) { dst_dbl[0] = ucastd(dcastu(dst_dbl[0]) & dcastu(src2_dbl[0])); dst_dbl[1] = ucastd(dcastu(dst_dbl[1]) & dcastu(src2_dbl[1])); dst_dbl[2] = ucastd(dcastu(dst_dbl[2]) & dcastu(src2_dbl[2])); dst_dbl[3] = ucastd(dcastu(dst_dbl[3]) & dcastu(src2_dbl[3])); } ENDSCALAR(checksum_dst(dst_dbl), "scalar and");142BINARYOP_D_DD("_mm_and_pd", _mm_and_pd, _mm_load_pd(src_dbl), _mm_load_pd(src2_dbl));143// _mm_and_si128144START(); dst_dbl[0] = src_dbl[0]; dst_dbl[1] = src_dbl[1]; dst_dbl[2] = src_dbl[2]; dst_dbl[3] = src_dbl[3]; for(int i = 0; i < N; ++i) { dst_dbl[0] = ucastd((~dcastu(dst_dbl[0])) & dcastu(src2_dbl[0])); dst_dbl[1] = ucastd((~dcastu(dst_dbl[1])) & dcastu(src2_dbl[1])); dst_dbl[2] = ucastd((~dcastu(dst_dbl[2])) & dcastu(src2_dbl[2])); dst_dbl[3] = ucastd((~dcastu(dst_dbl[3])) & dcastu(src2_dbl[3])); } ENDSCALAR(checksum_dst(dst_dbl), "scalar andnot");145BINARYOP_D_DD("_mm_andnot_pd", _mm_andnot_pd, _mm_load_pd(src_dbl), _mm_load_pd(src2_dbl));146// _mm_andnot_si128147START(); dst_dbl[0] = src_dbl[0]; dst_dbl[1] = src_dbl[1]; dst_dbl[2] = src_dbl[2]; dst_dbl[3] = src_dbl[3]; for(int i = 0; i < N; ++i) { dst_dbl[0] = ucastd(dcastu(dst_dbl[0]) | dcastu(src2_dbl[0])); dst_dbl[1] = ucastd(dcastu(dst_dbl[1]) | dcastu(src2_dbl[1])); dst_dbl[2] = ucastd(dcastu(dst_dbl[2]) | dcastu(src2_dbl[2])); dst_dbl[3] = ucastd(dcastu(dst_dbl[3]) | dcastu(src2_dbl[3])); } ENDSCALAR(checksum_dst(dst_dbl), "scalar or");148BINARYOP_D_DD("_mm_or_pd", _mm_or_pd, _mm_load_pd(src_dbl), _mm_load_pd(src2_dbl));149// _mm_or_si128150START(); dst_dbl[0] = src_dbl[0]; dst_dbl[1] = src_dbl[1]; dst_dbl[2] = src_dbl[2]; dst_dbl[3] = src_dbl[3]; for(int i = 0; i < N; ++i) { dst_dbl[0] = ucastd(dcastu(dst_dbl[0]) ^ dcastu(src2_dbl[0])); dst_dbl[1] = ucastd(dcastu(dst_dbl[1]) ^ dcastu(src2_dbl[1])); dst_dbl[2] = ucastd(dcastu(dst_dbl[2]) ^ dcastu(src2_dbl[2])); dst_dbl[3] = ucastd(dcastu(dst_dbl[3]) ^ dcastu(src2_dbl[3])); } ENDSCALAR(checksum_dst(dst_dbl), "scalar xor");151BINARYOP_D_DD("_mm_xor_pd", _mm_xor_pd, _mm_load_pd(src_dbl), _mm_load_pd(src2_dbl));152// _mm_xor_si128153154SETCHART("cmp");155// _mm_cmpeq_epi16156// _mm_cmpeq_epi32157// _mm_cmpeq_epi8158START(); dst_dbl[0] = src_dbl[0]; dst_dbl[1] = src_dbl[1]; dst_dbl[2] = src_dbl[2]; dst_dbl[3] = src_dbl[3]; for(int i = 0; i < N; ++i) { dst_dbl[0] = (dst_dbl[0] == src2_dbl[0]) ? ucastd(0xFFFFFFFFU) : 0.f; dst_dbl[1] = (dst_dbl[1] == src2_dbl[1]) ? ucastd(0xFFFFFFFFU) : 0.f; dst_dbl[2] = (dst_dbl[2] == src2_dbl[2]) ? ucastd(0xFFFFFFFFU) : 0.f; dst_dbl[3] = (dst_dbl[3] == src2_dbl[3]) ? ucastd(0xFFFFFFFFU) : 0.f; } ENDSCALAR(checksum_dst(dst_dbl), "scalar cmp==");159BINARYOP_D_DD("_mm_cmpeq_pd", _mm_cmpeq_pd, _mm_load_pd(src_dbl), _mm_load_pd(src2_dbl));160BINARYOP_D_DD("_mm_cmpeq_sd", _mm_cmpeq_sd, _mm_load_pd(src_dbl), _mm_load_pd(src2_dbl));161START(); dst_dbl[0] = src_dbl[0]; dst_dbl[1] = src_dbl[1]; dst_dbl[2] = src_dbl[2]; dst_dbl[3] = src_dbl[3]; for(int i = 0; i < N; ++i) { dst_dbl[0] = (dst_dbl[0] >= src2_dbl[0]) ? ucastd(0xFFFFFFFFU) : 0.f; dst_dbl[1] = (dst_dbl[1] >= src2_dbl[1]) ? ucastd(0xFFFFFFFFU) : 0.f; dst_dbl[2] = (dst_dbl[2] >= src2_dbl[2]) ? ucastd(0xFFFFFFFFU) : 0.f; dst_dbl[3] = (dst_dbl[3] >= src2_dbl[3]) ? ucastd(0xFFFFFFFFU) : 0.f; } ENDSCALAR(checksum_dst(dst_dbl), "scalar cmp>=");162BINARYOP_D_DD("_mm_cmpge_pd", _mm_cmpge_pd, _mm_load_pd(src_dbl), _mm_load_pd(src2_dbl));163BINARYOP_D_DD("_mm_cmpge_sd", _mm_cmpge_sd, _mm_load_pd(src_dbl), _mm_load_pd(src2_dbl));164// _mm_cmpgt_epi16165// _mm_cmpgt_epi32166// _mm_cmpgt_epi8167START(); dst_dbl[0] = src_dbl[0]; dst_dbl[1] = src_dbl[1]; dst_dbl[2] = src_dbl[2]; dst_dbl[3] = src_dbl[3]; for(int i = 0; i < N; ++i) { dst_dbl[0] = (dst_dbl[0] > src2_dbl[0]) ? ucastd(0xFFFFFFFFU) : 0.f; dst_dbl[1] = (dst_dbl[1] > src2_dbl[1]) ? ucastd(0xFFFFFFFFU) : 0.f; dst_dbl[2] = (dst_dbl[2] > src2_dbl[2]) ? ucastd(0xFFFFFFFFU) : 0.f; dst_dbl[3] = (dst_dbl[3] > src2_dbl[3]) ? ucastd(0xFFFFFFFFU) : 0.f; } ENDSCALAR(checksum_dst(dst_dbl), "scalar cmp>");168BINARYOP_D_DD("_mm_cmpgt_pd", _mm_cmpgt_pd, _mm_load_pd(src_dbl), _mm_load_pd(src2_dbl));169BINARYOP_D_DD("_mm_cmpgt_sd", _mm_cmpgt_sd, _mm_load_pd(src_dbl), _mm_load_pd(src2_dbl));170START(); dst_dbl[0] = src_dbl[0]; dst_dbl[1] = src_dbl[1]; dst_dbl[2] = src_dbl[2]; dst_dbl[3] = src_dbl[3]; for(int i = 0; i < N; ++i) { dst_dbl[0] = (dst_dbl[0] <= src2_dbl[0]) ? ucastd(0xFFFFFFFFU) : 0.f; dst_dbl[1] = (dst_dbl[1] <= src2_dbl[1]) ? ucastd(0xFFFFFFFFU) : 0.f; dst_dbl[2] = (dst_dbl[2] <= src2_dbl[2]) ? ucastd(0xFFFFFFFFU) : 0.f; dst_dbl[3] = (dst_dbl[3] <= src2_dbl[3]) ? ucastd(0xFFFFFFFFU) : 0.f; } ENDSCALAR(checksum_dst(dst_dbl), "scalar cmp<=");171BINARYOP_D_DD("_mm_cmple_pd", _mm_cmple_pd, _mm_load_pd(src_dbl), _mm_load_pd(src2_dbl));172BINARYOP_D_DD("_mm_cmple_sd", _mm_cmple_sd, _mm_load_pd(src_dbl), _mm_load_pd(src2_dbl));173// _mm_cmplt_epi16174// _mm_cmplt_epi32175// _mm_cmplt_epi8176START(); dst_dbl[0] = src_dbl[0]; dst_dbl[1] = src_dbl[1]; dst_dbl[2] = src_dbl[2]; dst_dbl[3] = src_dbl[3]; for(int i = 0; i < N; ++i) { dst_dbl[0] = (dst_dbl[0] < src2_dbl[0]) ? ucastd(0xFFFFFFFFU) : 0.f; dst_dbl[1] = (dst_dbl[1] < src2_dbl[1]) ? ucastd(0xFFFFFFFFU) : 0.f; dst_dbl[2] = (dst_dbl[2] < src2_dbl[2]) ? ucastd(0xFFFFFFFFU) : 0.f; dst_dbl[3] = (dst_dbl[3] < src2_dbl[3]) ? ucastd(0xFFFFFFFFU) : 0.f; } ENDSCALAR(checksum_dst(dst_dbl), "scalar cmp<");177BINARYOP_D_DD("_mm_cmplt_pd", _mm_cmplt_pd, _mm_load_pd(src_dbl), _mm_load_pd(src2_dbl));178BINARYOP_D_DD("_mm_cmplt_sd", _mm_cmplt_sd, _mm_load_pd(src_dbl), _mm_load_pd(src2_dbl));179180/*_mm_cmpneq_pd181_mm_cmpneq_sd182_mm_cmpnge_pd183_mm_cmpnge_sd184_mm_cmpngt_pd185_mm_cmpngt_sd186_mm_cmpnle_pd187_mm_cmpnle_sd188_mm_cmpnlt_pd189_mm_cmpnlt_sd*/190191START(); dst_dbl[0] = src_dbl[0]; dst_dbl[1] = src_dbl[1]; dst_dbl[2] = src_dbl[2]; dst_dbl[3] = src_dbl[3]; for(int i = 0; i < N; ++i) { dst_dbl[0] = (!Isnan(dst_dbl[0]) && !Isnan(src2_dbl[0])) ? ucastd(0xFFFFFFFFU) : 0.f; dst_dbl[1] = (!Isnan(dst_dbl[1]) && !Isnan(src2_dbl[1])) ? ucastd(0xFFFFFFFFU) : 0.f; dst_dbl[2] = (!Isnan(dst_dbl[2]) && !Isnan(src2_dbl[2])) ? ucastd(0xFFFFFFFFU) : 0.f; dst_dbl[3] = (!Isnan(dst_dbl[3]) && !Isnan(src2_dbl[3])) ? ucastd(0xFFFFFFFFU) : 0.f; } ENDSCALAR(checksum_dst(dst_dbl), "scalar cmpord");192BINARYOP_D_DD("_mm_cmpord_pd", _mm_cmpord_pd, _mm_load_pd(src_dbl), _mm_load_pd(src2_dbl));193BINARYOP_D_DD("_mm_cmpord_sd", _mm_cmpord_sd, _mm_load_pd(src_dbl), _mm_load_pd(src2_dbl));194START(); dst_dbl[0] = src_dbl[0]; dst_dbl[1] = src_dbl[1]; dst_dbl[2] = src_dbl[2]; dst_dbl[3] = src_dbl[3]; for(int i = 0; i < N; ++i) { dst_dbl[0] = (Isnan(dst_dbl[0]) || Isnan(src2_dbl[0])) ? ucastd(0xFFFFFFFFU) : 0.f; dst_dbl[1] = (Isnan(dst_dbl[1]) || Isnan(src2_dbl[1])) ? ucastd(0xFFFFFFFFU) : 0.f; dst_dbl[2] = (Isnan(dst_dbl[2]) || Isnan(src2_dbl[2])) ? ucastd(0xFFFFFFFFU) : 0.f; dst_dbl[3] = (Isnan(dst_dbl[3]) || Isnan(src2_dbl[3])) ? ucastd(0xFFFFFFFFU) : 0.f; } ENDSCALAR(checksum_dst(dst_dbl), "scalar cmpunord");195BINARYOP_D_DD("_mm_cmpunord_pd", _mm_cmpunord_pd, _mm_load_pd(src_dbl), _mm_load_pd(src2_dbl));196BINARYOP_D_DD("_mm_cmpunord_sd", _mm_cmpunord_sd, _mm_load_pd(src_dbl), _mm_load_pd(src2_dbl));197198SETCHART("max");199// _mm_max_epi16200// _mm_max_epu8201START(); dst_dbl[0] = src_dbl[0]; dst_dbl[1] = src_dbl[1]; dst_dbl[2] = src_dbl[2]; dst_dbl[3] = src_dbl[3]; for(int i = 0; i < N; ++i) { dst_dbl[0] = Max(dst_dbl[0], src2_dbl[0]); dst_dbl[1] = Max(dst_dbl[1], src2_dbl[1]); dst_dbl[2] = Max(dst_dbl[2], src2_dbl[2]); dst_dbl[3] = Max(dst_dbl[3], src2_dbl[3]); } ENDSCALAR(checksum_dst(dst_dbl), "scalar max");202BINARYOP_D_DD("_mm_max_pd", _mm_max_pd, _mm_load_pd(src_dbl), _mm_load_pd(src2_dbl));203BINARYOP_D_DD("_mm_max_sd", _mm_max_sd, _mm_load_pd(src_dbl), _mm_load_pd(src2_dbl));204// _mm_min_epi16205// _mm_min_epu8206START(); dst_dbl[0] = src_dbl[0]; dst_dbl[1] = src_dbl[1]; dst_dbl[2] = src_dbl[2]; dst_dbl[3] = src_dbl[3]; for(int i = 0; i < N; ++i) { dst_dbl[0] = Min(dst_dbl[0], src2_dbl[0]); dst_dbl[1] = Min(dst_dbl[1], src2_dbl[1]); dst_dbl[2] = Min(dst_dbl[2], src2_dbl[2]); dst_dbl[3] = Min(dst_dbl[3], src2_dbl[3]); } ENDSCALAR(checksum_dst(dst_dbl), "scalar min");207BINARYOP_D_DD("_mm_min_pd", _mm_min_pd, _mm_load_pd(src_dbl), _mm_load_pd(src2_dbl));208BINARYOP_D_DD("_mm_min_sd", _mm_min_sd, _mm_load_pd(src_dbl), _mm_load_pd(src2_dbl));209210SETCHART("shuffle");211// _mm_extract_epi16212// _mm_insert_epi16213// _mm_shuffle_epi32214START(); dst_dbl[0] = src_dbl[0]; dst_dbl[1] = src_dbl[1]; dst_dbl[2] = src_dbl[2]; dst_dbl[3] = src_dbl[3]; for(int i = 0; i < N; ++i) { dst_dbl[3] = dst_dbl[1]; dst_dbl[2] = dst_dbl[0]; dst_dbl[1] = src2_dbl[3]; dst_dbl[0] = src2_dbl[2]; } ENDSCALAR(checksum_dst(dst_dbl), "scalar shuffle");215// BINARYOP_D_DD("_mm_shuffle_pd", _mm_shuffle_pd, _mm_load_pd(src_dbl), _mm_load_pd(src2_dbl));216START();217__m128 o0 = _mm_load_pd(src_dbl);218__m128 o1 = _mm_load_pd(src2_dbl);219for(int i = 0; i < N; i += 4)220o0 = _mm_shuffle_pd(o0, o1, _MM_SHUFFLE2(1, 0));221_mm_store_pd(dst_dbl, o0);222END(checksum_dst(dst_dbl), "_mm_shuffle_pd");223224// _mm_shufflehi_epi16225// _mm_shufflelo_epi16226// _mm_unpackhi_epi16227// _mm_unpackhi_epi32228// _mm_unpackhi_epi64229// _mm_unpackhi_epi8230START(); dst_dbl[0] = src_dbl[0]; dst_dbl[1] = src_dbl[1]; dst_dbl[2] = src_dbl[2]; dst_dbl[3] = src_dbl[3]; for(int i = 0; i < N; ++i) { dst_dbl[0] = dst_dbl[2]; dst_dbl[1] = src2_dbl[2]; dst_dbl[2] = dst_dbl[3]; dst_dbl[3] = src2_dbl[3]; } ENDSCALAR(checksum_dst(dst_dbl), "scalar unpackhi_pd");231BINARYOP_D_DD("_mm_unpackhi_pd", _mm_unpackhi_pd, _mm_load_pd(src_dbl), _mm_load_pd(src2_dbl));232// _mm_unpacklo_epi16233// _mm_unpacklo_epi32234// _mm_unpacklo_epi64235// _mm_unpacklo_epi8236START(); dst_dbl[0] = src_dbl[0]; dst_dbl[1] = src_dbl[1]; dst_dbl[2] = src_dbl[2]; dst_dbl[3] = src_dbl[3]; for(int i = 0; i < N; ++i) { dst_dbl[2] = dst_dbl[1]; dst_dbl[1] = dst_dbl[0]; dst_dbl[0] = src2_dbl[0]; dst_dbl[3] = src2_dbl[1]; } ENDSCALAR(checksum_dst(dst_dbl), "scalar unpacklo_pd");237BINARYOP_D_DD("_mm_unpacklo_pd", _mm_unpacklo_pd, _mm_load_pd(src_dbl), _mm_load_pd(src2_dbl));238239// Benchmarks end:240printf("]}\n");241}242243244