CoCalc -- benchmark

GitHub Repository: emscripten-core/emscripten
Path: blob/main/test/benchmark/benchmark_sse1.cpp
⁶¹⁷⁴ views
1
/*
2
 * Copyright 2020 The Emscripten Authors.  All rights reserved.
3
 * Emscripten is available under two separate licenses, the MIT license and the
4
 * University of Illinois/NCSA Open Source License.  Both these licenses can be
5
 * found in the LICENSE file.
6
 */
7
#include <xmmintrin.h>
8
#include "benchmark_sse.h"
9

10
int main() {
11
  printf ("{ \"workload\": %u, \"results\": [\n", N);
12

13
  float *src_flt = alloc_float_buffer();
14
  float *src2_flt = alloc_float_buffer();
15
  float *dst_flt = alloc_float_buffer();
16
  for(int i = 0; i < N; ++i) src_flt[i] = (float)(1.0 + (double)rand() / RAND_MAX);
17
  for(int i = 0; i < N; ++i) src2_flt[i] = (float)(1.0 + (double)rand() / RAND_MAX);
18

19
  double *src_dbl = alloc_double_buffer();
20
  double *src2_dbl = alloc_double_buffer();
21
  double *dst_dbl = alloc_double_buffer();
22
  for(int i = 0; i < N; ++i) src_dbl[i] = 1.0 + (double)rand() / RAND_MAX;
23
  for(int i = 0; i < N; ++i) src2_dbl[i] = 1.0 + (double)rand() / RAND_MAX;
24

25
  int *src_int = alloc_int_buffer();
26
  int *src2_int = alloc_int_buffer();
27
  int *dst_int = alloc_int_buffer();
28
  int dst_int_scalar = 0;
29
  for(int i = 0; i < N; ++i) src_int[i] = rand();
30
  for(int i = 0; i < N; ++i) src2_int[i] = rand();
31

32
  float scalarTime = 0.f;
33

34
  // Benchmarks start:
35
  SETCHART("load");
36
  START();
37
    for(int i = 0; i < N; ++i)
38
      dst_flt[i] = src_flt[i];
39
  ENDSCALAR(checksum_dst(dst_flt), "scalar");
40

41
  LOAD_STORE_F("_mm_load_ps", _mm_load_ps, 0, _mm_store_ps, float*, 0, 4);
42
  LOAD_STORE_F("_mm_load_ps1", _mm_load_ps1, 1, _mm_store_ps, float*, 0, 4);
43
  LOAD_STORE_F("_mm_load_ss", _mm_load_ss, 1, _mm_store_ps, float*, 0, 4);
44
  LOAD_STORE_F("_mm_load1_ps", _mm_load1_ps, 1, _mm_store_ps, float*, 0, 4);
45

46
  __m128 tempReg = _mm_set_ps(1.f, 2.f, 3.f, 4.f);
47
  LOAD_STORE_M64("_mm_loadh_pi", tempReg, _mm_loadh_pi, __m64*, 1, _mm_store_ps, float*, 0, 4);
48
  LOAD_STORE_M64("_mm_loadl_pi", tempReg, _mm_loadh_pi, __m64*, 1, _mm_store_ps, float*, 0, 4);
49

50
  LOAD_STORE_F("_mm_loadr_ps", _mm_loadr_ps, 0, _mm_store_ps, float*, 0, 4);
51
  LOAD_STORE_F("_mm_loadu_ps", _mm_loadu_ps, 1, _mm_store_ps, float*, 0, 4);
52

53
  SETCHART("set");
54
  SET_STORE_F("_mm_set_ps", _mm_set_ps(src_flt[i+2], src_flt[i+1], src_flt[i+5], src_flt[i+0]));
55
  SET_STORE_F("_mm_set_ps1", _mm_set_ps1(src_flt[i]));
56
  SET_STORE_F("_mm_set_ss", _mm_set_ss(src_flt[i]));
57
  SET_STORE_F("_mm_set1_ps", _mm_set1_ps(src_flt[i]));
58
  SET_STORE_F("_mm_setr_ps", _mm_set_ps(src_flt[i+2], src_flt[i+1], src_flt[i+5], src_flt[i+0]));
59
  SET_STORE_F("_mm_setzero_ps", _mm_setzero_ps());
60

61
  SETCHART("move");
62
  SET_STORE_F("_mm_move_ss", _mm_move_ss(_mm_load_ps(src_flt+i), _mm_load_ps(src2_flt+i)));
63
  SET_STORE_F("_mm_movehl_ps", _mm_movehl_ps(_mm_load_ps(src_flt+i), _mm_load_ps(src2_flt+i)));
64
  SET_STORE_F("_mm_movelh_ps", _mm_movelh_ps(_mm_load_ps(src_flt+i), _mm_load_ps(src2_flt+i)));
65

66
  SETCHART("movemask");
67
  START(); for(int i = 0; i < N; i += 4) { int movemask = ((unsigned int)src_flt[i] >> 31) | (((unsigned int)src_flt[i+1] >> 30) & 2)  | (((unsigned int)src_flt[i+2] >> 29) & 4)  | (((unsigned int)src_flt[i+3] >> 28) & 8); dst_int_scalar += movemask; } ENDSCALAR(dst_int_scalar, "scalar movemask");
68
  UNARYOP_i_F("_mm_movemask_ps", _mm_movemask_ps(_mm_load_ps(src_flt+i)));
69

70
  SETCHART("store");
71
  LOAD_STORE_F("_mm_store_ps", _mm_load_ps, 0, _mm_store_ps, float*, 0, 4);
72
  LOAD_STORE_F("_mm_store_ps1", _mm_load_ps, 0, _mm_store_ps1, float*, 0, 4);
73
  LOAD_STORE_F("_mm_store_ss", _mm_load_ps, 0, _mm_store_ss, float*, 1, 4);
74
  LOAD_STORE_64_F("_mm_storeh_pi", _mm_load_ps, 0, _mm_storeh_pi, 1, 4);
75
  LOAD_STORE_64_F("_mm_storel_pi", _mm_load_ps, 0, _mm_storel_pi, 1, 4);
76
  LOAD_STORE_F("_mm_storer_ps", _mm_load_ps, 0, _mm_storer_ps, float*, 0, 4);
77
  LOAD_STORE_F("_mm_storeu_ps", _mm_load_ps, 0, _mm_storeu_ps, float*, 1, 4);
78
  LOAD_STORE_F("_mm_stream_ps", _mm_load_ps, 0, _mm_stream_ps, float*, 0, 4);
79

80
  SETCHART("arithmetic");
81
  START(); dst_flt[0] = src_flt[0]; dst_flt[1] = src_flt[1]; dst_flt[2] = src_flt[2]; dst_flt[3] = src_flt[3]; for(int i = 0; i < N; ++i) { dst_flt[0] += src2_flt[0]; dst_flt[1] += src2_flt[1]; dst_flt[2] += src2_flt[2]; dst_flt[3] += src2_flt[3]; } ENDSCALAR(checksum_dst(dst_flt), "scalar add");
82
  BINARYOP_F_FF("_mm_add_ps", _mm_add_ps, _mm_load_ps(src_flt), _mm_load_ps(src2_flt));
83
  BINARYOP_F_FF("_mm_add_ss", _mm_add_ss, _mm_load_ps(src_flt), _mm_load_ps(src2_flt));
84
  START(); dst_flt[0] = src_flt[0]; dst_flt[1] = src_flt[1]; dst_flt[2] = src_flt[2]; dst_flt[3] = src_flt[3]; for(int i = 0; i < N; ++i) { dst_flt[0] /= src2_flt[0]; dst_flt[1] /= src2_flt[1]; dst_flt[2] /= src2_flt[2]; dst_flt[3] /= src2_flt[3]; } ENDSCALAR(checksum_dst(dst_flt), "scalar div");
85
  BINARYOP_F_FF("_mm_div_ps", _mm_div_ps, _mm_load_ps(src_flt), _mm_load_ps(src2_flt));
86
  BINARYOP_F_FF("_mm_div_ss", _mm_div_ss, _mm_load_ps(src_flt), _mm_load_ps(src2_flt));
87
  START(); dst_flt[0] = src_flt[0]; dst_flt[1] = src_flt[1]; dst_flt[2] = src_flt[2]; dst_flt[3] = src_flt[3]; for(int i = 0; i < N; ++i) { dst_flt[0] *= src2_flt[0]; dst_flt[1] *= src2_flt[1]; dst_flt[2] *= src2_flt[2]; dst_flt[3] *= src2_flt[3]; } ENDSCALAR(checksum_dst(dst_flt), "scalar mul");
88
  BINARYOP_F_FF("_mm_mul_ps", _mm_mul_ps, _mm_load_ps(src_flt), _mm_load_ps(src2_flt));
89
  BINARYOP_F_FF("_mm_mul_ss", _mm_mul_ss, _mm_load_ps(src_flt), _mm_load_ps(src2_flt));
90
  START(); dst_flt[0] = src_flt[0]; dst_flt[1] = src_flt[1]; dst_flt[2] = src_flt[2]; dst_flt[3] = src_flt[3]; for(int i = 0; i < N; ++i) { dst_flt[0] -= src2_flt[0]; dst_flt[1] -= src2_flt[1]; dst_flt[2] -= src2_flt[2]; dst_flt[3] -= src2_flt[3]; } ENDSCALAR(checksum_dst(dst_flt), "scalar sub");
91
  BINARYOP_F_FF("_mm_sub_ps", _mm_sub_ps, _mm_load_ps(src_flt), _mm_load_ps(src2_flt));
92
  BINARYOP_F_FF("_mm_sub_ss", _mm_sub_ss, _mm_load_ps(src_flt), _mm_load_ps(src2_flt));
93

94
  SETCHART("roots");
95
  START(); dst_flt[0] = src_flt[0]; dst_flt[1] = src_flt[1]; dst_flt[2] = src_flt[2]; dst_flt[3] = src_flt[3]; for(int i = 0; i < N; ++i) { dst_flt[0] = 1.f / dst_flt[0]; dst_flt[1] = 1.f / dst_flt[1]; dst_flt[2] = 1.f / dst_flt[2]; dst_flt[3] = 1.f / dst_flt[3]; } ENDSCALAR(checksum_dst(dst_flt), "scalar rcp");
96
  UNARYOP_F_F("_mm_rcp_ps", _mm_rcp_ps, _mm_load_ps(src_flt));
97
  UNARYOP_F_F("_mm_rcp_ss", _mm_rcp_ss, _mm_load_ps(src_flt));
98

99
  START(); dst_flt[0] = src_flt[0]; dst_flt[1] = src_flt[1]; dst_flt[2] = src_flt[2]; dst_flt[3] = src_flt[3]; for(int i = 0; i < N; ++i) { dst_flt[0] = 1.f / sqrtf(dst_flt[0]); dst_flt[1] = 1.f / sqrtf(dst_flt[1]); dst_flt[2] = 1.f / sqrtf(dst_flt[2]); dst_flt[3] = 1.f / sqrtf(dst_flt[3]); } ENDSCALAR(checksum_dst(dst_flt), "scalar rsqrt");
100
  UNARYOP_F_F("_mm_rsqrt_ps", _mm_rsqrt_ps, _mm_load_ps(src_flt));
101
  UNARYOP_F_F("_mm_rsqrt_ss", _mm_rsqrt_ss, _mm_load_ps(src_flt));
102

103
  START(); dst_flt[0] = src_flt[0]; dst_flt[1] = src_flt[1]; dst_flt[2] = src_flt[2]; dst_flt[3] = src_flt[3]; for(int i = 0; i < N; ++i) { dst_flt[0] = sqrtf(dst_flt[0]); dst_flt[1] = sqrtf(dst_flt[1]); dst_flt[2] = sqrtf(dst_flt[2]); dst_flt[3] = sqrtf(dst_flt[3]); } ENDSCALAR(checksum_dst(dst_flt), "scalar sqrt");
104
  UNARYOP_F_F("_mm_sqrt_ps", _mm_sqrt_ps, _mm_load_ps(src_flt));
105
  UNARYOP_F_F("_mm_sqrt_ss", _mm_sqrt_ss, _mm_load_ps(src_flt));
106

107
  SETCHART("logical");
108
  START(); dst_flt[0] = src_flt[0]; dst_flt[1] = src_flt[1]; dst_flt[2] = src_flt[2]; dst_flt[3] = src_flt[3]; for(int i = 0; i < N; ++i) { dst_flt[0] = ucastf(fcastu(dst_flt[0]) & fcastu(src2_flt[0])); dst_flt[1] = ucastf(fcastu(dst_flt[1]) & fcastu(src2_flt[1])); dst_flt[2] = ucastf(fcastu(dst_flt[2]) & fcastu(src2_flt[2])); dst_flt[3] = ucastf(fcastu(dst_flt[3]) & fcastu(src2_flt[3])); } ENDSCALAR(checksum_dst(dst_flt), "scalar and");
109
  BINARYOP_F_FF("_mm_and_ps", _mm_and_ps, _mm_load_ps(src_flt), _mm_load_ps(src2_flt));
110
  START(); dst_flt[0] = src_flt[0]; dst_flt[1] = src_flt[1]; dst_flt[2] = src_flt[2]; dst_flt[3] = src_flt[3]; for(int i = 0; i < N; ++i) { dst_flt[0] = ucastf((~fcastu(dst_flt[0])) & fcastu(src2_flt[0])); dst_flt[1] = ucastf((~fcastu(dst_flt[1])) & fcastu(src2_flt[1])); dst_flt[2] = ucastf((~fcastu(dst_flt[2])) & fcastu(src2_flt[2])); dst_flt[3] = ucastf((~fcastu(dst_flt[3])) & fcastu(src2_flt[3])); } ENDSCALAR(checksum_dst(dst_flt), "scalar andnot");
111
  BINARYOP_F_FF("_mm_andnot_ps", _mm_andnot_ps, _mm_load_ps(src_flt), _mm_load_ps(src2_flt));
112
  START(); dst_flt[0] = src_flt[0]; dst_flt[1] = src_flt[1]; dst_flt[2] = src_flt[2]; dst_flt[3] = src_flt[3]; for(int i = 0; i < N; ++i) { dst_flt[0] = ucastf(fcastu(dst_flt[0]) | fcastu(src2_flt[0])); dst_flt[1] = ucastf(fcastu(dst_flt[1]) | fcastu(src2_flt[1])); dst_flt[2] = ucastf(fcastu(dst_flt[2]) | fcastu(src2_flt[2])); dst_flt[3] = ucastf(fcastu(dst_flt[3]) | fcastu(src2_flt[3])); } ENDSCALAR(checksum_dst(dst_flt), "scalar or");
113
  BINARYOP_F_FF("_mm_or_ps", _mm_or_ps, _mm_load_ps(src_flt), _mm_load_ps(src2_flt));
114
  START(); dst_flt[0] = src_flt[0]; dst_flt[1] = src_flt[1]; dst_flt[2] = src_flt[2]; dst_flt[3] = src_flt[3]; for(int i = 0; i < N; ++i) { dst_flt[0] = ucastf(fcastu(dst_flt[0]) ^ fcastu(src2_flt[0])); dst_flt[1] = ucastf(fcastu(dst_flt[1]) ^ fcastu(src2_flt[1])); dst_flt[2] = ucastf(fcastu(dst_flt[2]) ^ fcastu(src2_flt[2])); dst_flt[3] = ucastf(fcastu(dst_flt[3]) ^ fcastu(src2_flt[3])); } ENDSCALAR(checksum_dst(dst_flt), "scalar xor");
115
  BINARYOP_F_FF("_mm_xor_ps", _mm_xor_ps, _mm_load_ps(src_flt), _mm_load_ps(src2_flt));
116

117
  SETCHART("cmp");
118
  START(); dst_flt[0] = src_flt[0]; dst_flt[1] = src_flt[1]; dst_flt[2] = src_flt[2]; dst_flt[3] = src_flt[3]; for(int i = 0; i < N; ++i) { dst_flt[0] = (dst_flt[0] == src2_flt[0]) ? ucastf(0xFFFFFFFFU) : 0.f; dst_flt[1] = (dst_flt[1] == src2_flt[1]) ? ucastf(0xFFFFFFFFU) : 0.f; dst_flt[2] = (dst_flt[2] == src2_flt[2]) ? ucastf(0xFFFFFFFFU) : 0.f; dst_flt[3] = (dst_flt[3] == src2_flt[3]) ? ucastf(0xFFFFFFFFU) : 0.f; } ENDSCALAR(checksum_dst(dst_flt), "scalar cmp==");
119
  BINARYOP_F_FF("_mm_cmpeq_ps", _mm_cmpeq_ps, _mm_load_ps(src_flt), _mm_load_ps(src2_flt));
120
  BINARYOP_F_FF("_mm_cmpeq_ss", _mm_cmpeq_ss, _mm_load_ps(src_flt), _mm_load_ps(src2_flt));
121
  START(); dst_flt[0] = src_flt[0]; dst_flt[1] = src_flt[1]; dst_flt[2] = src_flt[2]; dst_flt[3] = src_flt[3]; for(int i = 0; i < N; ++i) { dst_flt[0] = (dst_flt[0] >= src2_flt[0]) ? ucastf(0xFFFFFFFFU) : 0.f; dst_flt[1] = (dst_flt[1] >= src2_flt[1]) ? ucastf(0xFFFFFFFFU) : 0.f; dst_flt[2] = (dst_flt[2] >= src2_flt[2]) ? ucastf(0xFFFFFFFFU) : 0.f; dst_flt[3] = (dst_flt[3] >= src2_flt[3]) ? ucastf(0xFFFFFFFFU) : 0.f; } ENDSCALAR(checksum_dst(dst_flt), "scalar cmp>=");
122
  BINARYOP_F_FF("_mm_cmpge_ps", _mm_cmpge_ps, _mm_load_ps(src_flt), _mm_load_ps(src2_flt));
123
  BINARYOP_F_FF("_mm_cmpge_ss", _mm_cmpge_ss, _mm_load_ps(src_flt), _mm_load_ps(src2_flt));
124
  START(); dst_flt[0] = src_flt[0]; dst_flt[1] = src_flt[1]; dst_flt[2] = src_flt[2]; dst_flt[3] = src_flt[3]; for(int i = 0; i < N; ++i) { dst_flt[0] = (dst_flt[0] > src2_flt[0]) ? ucastf(0xFFFFFFFFU) : 0.f; dst_flt[1] = (dst_flt[1] > src2_flt[1]) ? ucastf(0xFFFFFFFFU) : 0.f; dst_flt[2] = (dst_flt[2] > src2_flt[2]) ? ucastf(0xFFFFFFFFU) : 0.f; dst_flt[3] = (dst_flt[3] > src2_flt[3]) ? ucastf(0xFFFFFFFFU) : 0.f; } ENDSCALAR(checksum_dst(dst_flt), "scalar cmp>");
125
  BINARYOP_F_FF("_mm_cmpgt_ps", _mm_cmpgt_ps, _mm_load_ps(src_flt), _mm_load_ps(src2_flt));
126
  BINARYOP_F_FF("_mm_cmpgt_ss", _mm_cmpgt_ss, _mm_load_ps(src_flt), _mm_load_ps(src2_flt));
127
  START(); dst_flt[0] = src_flt[0]; dst_flt[1] = src_flt[1]; dst_flt[2] = src_flt[2]; dst_flt[3] = src_flt[3]; for(int i = 0; i < N; ++i) { dst_flt[0] = (dst_flt[0] <= src2_flt[0]) ? ucastf(0xFFFFFFFFU) : 0.f; dst_flt[1] = (dst_flt[1] <= src2_flt[1]) ? ucastf(0xFFFFFFFFU) : 0.f; dst_flt[2] = (dst_flt[2] <= src2_flt[2]) ? ucastf(0xFFFFFFFFU) : 0.f; dst_flt[3] = (dst_flt[3] <= src2_flt[3]) ? ucastf(0xFFFFFFFFU) : 0.f; } ENDSCALAR(checksum_dst(dst_flt), "scalar cmp<=");
128
  BINARYOP_F_FF("_mm_cmple_ps", _mm_cmple_ps, _mm_load_ps(src_flt), _mm_load_ps(src2_flt));
129
  BINARYOP_F_FF("_mm_cmple_ss", _mm_cmple_ss, _mm_load_ps(src_flt), _mm_load_ps(src2_flt));
130
  START(); dst_flt[0] = src_flt[0]; dst_flt[1] = src_flt[1]; dst_flt[2] = src_flt[2]; dst_flt[3] = src_flt[3]; for(int i = 0; i < N; ++i) { dst_flt[0] = (dst_flt[0] < src2_flt[0]) ? ucastf(0xFFFFFFFFU) : 0.f; dst_flt[1] = (dst_flt[1] < src2_flt[1]) ? ucastf(0xFFFFFFFFU) : 0.f; dst_flt[2] = (dst_flt[2] < src2_flt[2]) ? ucastf(0xFFFFFFFFU) : 0.f; dst_flt[3] = (dst_flt[3] < src2_flt[3]) ? ucastf(0xFFFFFFFFU) : 0.f; } ENDSCALAR(checksum_dst(dst_flt), "scalar cmp<");
131
  BINARYOP_F_FF("_mm_cmplt_ps", _mm_cmplt_ps, _mm_load_ps(src_flt), _mm_load_ps(src2_flt));
132
  BINARYOP_F_FF("_mm_cmplt_ss", _mm_cmplt_ss, _mm_load_ps(src_flt), _mm_load_ps(src2_flt));
133

134
  START(); dst_flt[0] = src_flt[0]; dst_flt[1] = src_flt[1]; dst_flt[2] = src_flt[2]; dst_flt[3] = src_flt[3]; for(int i = 0; i < N; ++i) { dst_flt[0] = (!Isnan(dst_flt[0]) && !Isnan(src2_flt[0])) ? ucastf(0xFFFFFFFFU) : 0.f; dst_flt[1] = (!Isnan(dst_flt[1]) && !Isnan(src2_flt[1])) ? ucastf(0xFFFFFFFFU) : 0.f; dst_flt[2] = (!Isnan(dst_flt[2]) && !Isnan(src2_flt[2])) ? ucastf(0xFFFFFFFFU) : 0.f; dst_flt[3] = (!Isnan(dst_flt[3]) && !Isnan(src2_flt[3])) ? ucastf(0xFFFFFFFFU) : 0.f; } ENDSCALAR(checksum_dst(dst_flt), "scalar cmpord");
135
  BINARYOP_F_FF("_mm_cmpord_ps", _mm_cmpord_ps, _mm_load_ps(src_flt), _mm_load_ps(src2_flt));
136
  BINARYOP_F_FF("_mm_cmpord_ss", _mm_cmpord_ss, _mm_load_ps(src_flt), _mm_load_ps(src2_flt));
137
  START(); dst_flt[0] = src_flt[0]; dst_flt[1] = src_flt[1]; dst_flt[2] = src_flt[2]; dst_flt[3] = src_flt[3]; for(int i = 0; i < N; ++i) { dst_flt[0] = (Isnan(dst_flt[0]) || Isnan(src2_flt[0])) ? ucastf(0xFFFFFFFFU) : 0.f; dst_flt[1] = (Isnan(dst_flt[1]) || Isnan(src2_flt[1])) ? ucastf(0xFFFFFFFFU) : 0.f; dst_flt[2] = (Isnan(dst_flt[2]) || Isnan(src2_flt[2])) ? ucastf(0xFFFFFFFFU) : 0.f; dst_flt[3] = (Isnan(dst_flt[3]) || Isnan(src2_flt[3])) ? ucastf(0xFFFFFFFFU) : 0.f; } ENDSCALAR(checksum_dst(dst_flt), "scalar cmpunord");
138
  BINARYOP_F_FF("_mm_cmpunord_ps", _mm_cmpunord_ps, _mm_load_ps(src_flt), _mm_load_ps(src2_flt));
139
  BINARYOP_F_FF("_mm_cmpunord_ss", _mm_cmpunord_ss, _mm_load_ps(src_flt), _mm_load_ps(src2_flt));
140

141
  SETCHART("max");
142
  START(); dst_flt[0] = src_flt[0]; dst_flt[1] = src_flt[1]; dst_flt[2] = src_flt[2]; dst_flt[3] = src_flt[3]; for(int i = 0; i < N; ++i) { dst_flt[0] = Max(dst_flt[0], src2_flt[0]); dst_flt[1] = Max(dst_flt[1], src2_flt[1]); dst_flt[2] = Max(dst_flt[2], src2_flt[2]); dst_flt[3] = Max(dst_flt[3], src2_flt[3]); } ENDSCALAR(checksum_dst(dst_flt), "scalar max");
143
  BINARYOP_F_FF("_mm_max_ps", _mm_max_ps, _mm_load_ps(src_flt), _mm_load_ps(src2_flt));
144
  BINARYOP_F_FF("_mm_max_ss", _mm_max_ss, _mm_load_ps(src_flt), _mm_load_ps(src2_flt));
145
  START(); dst_flt[0] = src_flt[0]; dst_flt[1] = src_flt[1]; dst_flt[2] = src_flt[2]; dst_flt[3] = src_flt[3]; for(int i = 0; i < N; ++i) { dst_flt[0] = Min(dst_flt[0], src2_flt[0]); dst_flt[1] = Min(dst_flt[1], src2_flt[1]); dst_flt[2] = Min(dst_flt[2], src2_flt[2]); dst_flt[3] = Min(dst_flt[3], src2_flt[3]); } ENDSCALAR(checksum_dst(dst_flt), "scalar min");
146
  BINARYOP_F_FF("_mm_min_ps", _mm_min_ps, _mm_load_ps(src_flt), _mm_load_ps(src2_flt));
147
  BINARYOP_F_FF("_mm_min_ss", _mm_min_ss, _mm_load_ps(src_flt), _mm_load_ps(src2_flt));
148

149
  SETCHART("shuffle");
150
  START(); dst_flt[0] = src_flt[0]; dst_flt[1] = src_flt[1]; dst_flt[2] = src_flt[2]; dst_flt[3] = src_flt[3]; for(int i = 0; i < N; ++i) { dst_flt[3] = dst_flt[1]; dst_flt[2] = dst_flt[0]; dst_flt[1] = src2_flt[3]; dst_flt[0] = src2_flt[2]; } ENDSCALAR(checksum_dst(dst_flt), "scalar shuffle");
151

152
  START();
153
    __m128 o0 = _mm_load_ps(src_flt);
154
    __m128 o1 = _mm_load_ps(src2_flt);
155
    for(int i = 0; i < N; i += 4)
156
      o0 = _mm_shuffle_ps(o0, o1, _MM_SHUFFLE(1, 0, 3, 2));
157
    _mm_store_ps(dst_flt, o0);
158
  END(checksum_dst(dst_flt), "_mm_shuffle_ps");
159

160
  START(); dst_flt[0] = src_flt[0]; dst_flt[1] = src_flt[1]; dst_flt[2] = src_flt[2]; dst_flt[3] = src_flt[3]; for(int i = 0; i < N; ++i) { dst_flt[0] = dst_flt[2]; dst_flt[1] = src2_flt[2]; dst_flt[2] = dst_flt[3]; dst_flt[3] = src2_flt[3]; } ENDSCALAR(checksum_dst(dst_flt), "scalar unpackhi_ps");
161
  BINARYOP_F_FF("_mm_unpackhi_ps", _mm_unpackhi_ps, _mm_load_ps(src_flt), _mm_load_ps(src2_flt));
162
  START(); dst_flt[0] = src_flt[0]; dst_flt[1] = src_flt[1]; dst_flt[2] = src_flt[2]; dst_flt[3] = src_flt[3]; for(int i = 0; i < N; ++i) { dst_flt[2] = dst_flt[1]; dst_flt[1] = dst_flt[0]; dst_flt[0] = src2_flt[0]; dst_flt[3] = src2_flt[1]; } ENDSCALAR(checksum_dst(dst_flt), "scalar unpacklo_ps");
163
  BINARYOP_F_FF("_mm_unpacklo_ps", _mm_unpacklo_ps, _mm_load_ps(src_flt), _mm_load_ps(src2_flt));
164

165
  // Benchmarks end:
166
  printf("]}\n");
167
}
168

169
Product

Resources

Company
