minimap2 0.2.24.3 → 0.2.24.6

Sign up to get free protection for your applications and to get access to all the features.
Files changed (101) hide show
  1. checksums.yaml +4 -4
  2. data/ext/minimap2/lib/simde/CONTRIBUTING.md +114 -0
  3. data/ext/minimap2/lib/simde/COPYING +20 -0
  4. data/ext/minimap2/lib/simde/README.md +333 -0
  5. data/ext/minimap2/lib/simde/amalgamate.py +58 -0
  6. data/ext/minimap2/lib/simde/meson.build +33 -0
  7. data/ext/minimap2/lib/simde/netlify.toml +20 -0
  8. data/ext/minimap2/lib/simde/simde/arm/neon/float32x2.h +140 -0
  9. data/ext/minimap2/lib/simde/simde/arm/neon/float32x4.h +137 -0
  10. data/ext/minimap2/lib/simde/simde/arm/neon/float64x1.h +142 -0
  11. data/ext/minimap2/lib/simde/simde/arm/neon/float64x2.h +145 -0
  12. data/ext/minimap2/lib/simde/simde/arm/neon/int16x4.h +140 -0
  13. data/ext/minimap2/lib/simde/simde/arm/neon/int16x8.h +145 -0
  14. data/ext/minimap2/lib/simde/simde/arm/neon/int32x2.h +140 -0
  15. data/ext/minimap2/lib/simde/simde/arm/neon/int32x4.h +143 -0
  16. data/ext/minimap2/lib/simde/simde/arm/neon/int64x1.h +137 -0
  17. data/ext/minimap2/lib/simde/simde/arm/neon/int64x2.h +141 -0
  18. data/ext/minimap2/lib/simde/simde/arm/neon/int8x16.h +147 -0
  19. data/ext/minimap2/lib/simde/simde/arm/neon/int8x8.h +141 -0
  20. data/ext/minimap2/lib/simde/simde/arm/neon/uint16x4.h +134 -0
  21. data/ext/minimap2/lib/simde/simde/arm/neon/uint16x8.h +138 -0
  22. data/ext/minimap2/lib/simde/simde/arm/neon/uint32x2.h +134 -0
  23. data/ext/minimap2/lib/simde/simde/arm/neon/uint32x4.h +137 -0
  24. data/ext/minimap2/lib/simde/simde/arm/neon/uint64x1.h +131 -0
  25. data/ext/minimap2/lib/simde/simde/arm/neon/uint64x2.h +135 -0
  26. data/ext/minimap2/lib/simde/simde/arm/neon/uint8x16.h +141 -0
  27. data/ext/minimap2/lib/simde/simde/arm/neon/uint8x8.h +135 -0
  28. data/ext/minimap2/lib/simde/simde/arm/neon.h +97 -0
  29. data/ext/minimap2/lib/simde/simde/check.h +267 -0
  30. data/ext/minimap2/lib/simde/simde/debug-trap.h +83 -0
  31. data/ext/minimap2/lib/simde/simde/hedley.h +1899 -0
  32. data/ext/minimap2/lib/simde/simde/simde-arch.h +445 -0
  33. data/ext/minimap2/lib/simde/simde/simde-common.h +697 -0
  34. data/ext/minimap2/lib/simde/simde/x86/avx.h +5385 -0
  35. data/ext/minimap2/lib/simde/simde/x86/avx2.h +2402 -0
  36. data/ext/minimap2/lib/simde/simde/x86/avx512bw.h +391 -0
  37. data/ext/minimap2/lib/simde/simde/x86/avx512f.h +3389 -0
  38. data/ext/minimap2/lib/simde/simde/x86/avx512vl.h +112 -0
  39. data/ext/minimap2/lib/simde/simde/x86/fma.h +659 -0
  40. data/ext/minimap2/lib/simde/simde/x86/mmx.h +2210 -0
  41. data/ext/minimap2/lib/simde/simde/x86/sse.h +3696 -0
  42. data/ext/minimap2/lib/simde/simde/x86/sse2.h +5991 -0
  43. data/ext/minimap2/lib/simde/simde/x86/sse3.h +343 -0
  44. data/ext/minimap2/lib/simde/simde/x86/sse4.1.h +1783 -0
  45. data/ext/minimap2/lib/simde/simde/x86/sse4.2.h +105 -0
  46. data/ext/minimap2/lib/simde/simde/x86/ssse3.h +1053 -0
  47. data/ext/minimap2/lib/simde/simde/x86/svml.h +543 -0
  48. data/ext/minimap2/lib/simde/test/CMakeLists.txt +166 -0
  49. data/ext/minimap2/lib/simde/test/arm/meson.build +4 -0
  50. data/ext/minimap2/lib/simde/test/arm/neon/meson.build +23 -0
  51. data/ext/minimap2/lib/simde/test/arm/neon/skel.c +871 -0
  52. data/ext/minimap2/lib/simde/test/arm/neon/test-neon-internal.h +134 -0
  53. data/ext/minimap2/lib/simde/test/arm/neon/test-neon.c +39 -0
  54. data/ext/minimap2/lib/simde/test/arm/neon/test-neon.h +10 -0
  55. data/ext/minimap2/lib/simde/test/arm/neon/vadd.c +1260 -0
  56. data/ext/minimap2/lib/simde/test/arm/neon/vdup_n.c +873 -0
  57. data/ext/minimap2/lib/simde/test/arm/neon/vmul.c +1084 -0
  58. data/ext/minimap2/lib/simde/test/arm/neon/vsub.c +1260 -0
  59. data/ext/minimap2/lib/simde/test/arm/test-arm-internal.h +18 -0
  60. data/ext/minimap2/lib/simde/test/arm/test-arm.c +20 -0
  61. data/ext/minimap2/lib/simde/test/arm/test-arm.h +8 -0
  62. data/ext/minimap2/lib/simde/test/cmake/AddCompilerFlags.cmake +171 -0
  63. data/ext/minimap2/lib/simde/test/cmake/ExtraWarningFlags.cmake +68 -0
  64. data/ext/minimap2/lib/simde/test/meson.build +64 -0
  65. data/ext/minimap2/lib/simde/test/munit/COPYING +21 -0
  66. data/ext/minimap2/lib/simde/test/munit/Makefile +55 -0
  67. data/ext/minimap2/lib/simde/test/munit/README.md +54 -0
  68. data/ext/minimap2/lib/simde/test/munit/example.c +351 -0
  69. data/ext/minimap2/lib/simde/test/munit/meson.build +37 -0
  70. data/ext/minimap2/lib/simde/test/munit/munit.c +2055 -0
  71. data/ext/minimap2/lib/simde/test/munit/munit.h +535 -0
  72. data/ext/minimap2/lib/simde/test/run-tests.c +20 -0
  73. data/ext/minimap2/lib/simde/test/run-tests.h +260 -0
  74. data/ext/minimap2/lib/simde/test/x86/avx.c +13752 -0
  75. data/ext/minimap2/lib/simde/test/x86/avx2.c +9977 -0
  76. data/ext/minimap2/lib/simde/test/x86/avx512bw.c +2664 -0
  77. data/ext/minimap2/lib/simde/test/x86/avx512f.c +10416 -0
  78. data/ext/minimap2/lib/simde/test/x86/avx512vl.c +210 -0
  79. data/ext/minimap2/lib/simde/test/x86/fma.c +2557 -0
  80. data/ext/minimap2/lib/simde/test/x86/meson.build +33 -0
  81. data/ext/minimap2/lib/simde/test/x86/mmx.c +2878 -0
  82. data/ext/minimap2/lib/simde/test/x86/skel.c +2984 -0
  83. data/ext/minimap2/lib/simde/test/x86/sse.c +5121 -0
  84. data/ext/minimap2/lib/simde/test/x86/sse2.c +9860 -0
  85. data/ext/minimap2/lib/simde/test/x86/sse3.c +486 -0
  86. data/ext/minimap2/lib/simde/test/x86/sse4.1.c +3446 -0
  87. data/ext/minimap2/lib/simde/test/x86/sse4.2.c +101 -0
  88. data/ext/minimap2/lib/simde/test/x86/ssse3.c +2084 -0
  89. data/ext/minimap2/lib/simde/test/x86/svml.c +1545 -0
  90. data/ext/minimap2/lib/simde/test/x86/test-avx.h +16 -0
  91. data/ext/minimap2/lib/simde/test/x86/test-avx512.h +25 -0
  92. data/ext/minimap2/lib/simde/test/x86/test-mmx.h +13 -0
  93. data/ext/minimap2/lib/simde/test/x86/test-sse.h +13 -0
  94. data/ext/minimap2/lib/simde/test/x86/test-sse2.h +13 -0
  95. data/ext/minimap2/lib/simde/test/x86/test-x86-internal.h +196 -0
  96. data/ext/minimap2/lib/simde/test/x86/test-x86.c +48 -0
  97. data/ext/minimap2/lib/simde/test/x86/test-x86.h +8 -0
  98. data/lib/minimap2/aligner.rb +2 -2
  99. data/lib/minimap2/ffi/constants.rb +3 -0
  100. data/lib/minimap2/version.rb +1 -1
  101. metadata +99 -3
@@ -0,0 +1,3446 @@
1
+ /* Copyright (c) 2017 Evan Nemerson <evan@nemerson.com>
2
+ *
3
+ * Permission is hereby granted, free of charge, to any person
4
+ * obtaining a copy of this software and associated documentation
5
+ * files (the "Software"), to deal in the Software without
6
+ * restriction, including without limitation the rights to use, copy,
7
+ * modify, merge, publish, distribute, sublicense, and/or sell copies
8
+ * of the Software, and to permit persons to whom the Software is
9
+ * furnished to do so, subject to the following conditions:
10
+ *
11
+ * The above copyright notice and this permission notice shall be
12
+ * included in all copies or substantial portions of the Software.
13
+ *
14
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
15
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
16
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
17
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
18
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
19
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
20
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
21
+ * SOFTWARE.
22
+ */
23
+
24
+ #define SIMDE_TESTS_CURRENT_ISAX sse4_1
25
+ #include <simde/x86/sse4.1.h>
26
+ #include <test/x86/test-sse2.h>
27
+
28
+ #if defined(SIMDE_SSE4_1_NATIVE) || defined(SIMDE_NO_NATIVE) || defined(SIMDE_ALWAYS_BUILD_NATIVE_TESTS)
29
+
30
+ static MunitResult
31
+ test_simde_mm_blendv_epi8(const MunitParameter params[], void* data) {
32
+ (void) params;
33
+ (void) data;
34
+
35
+ const struct {
36
+ simde__m128i a;
37
+ simde__m128i b;
38
+ simde__m128i mask;
39
+ simde__m128i r;
40
+ } test_vec[8] = {
41
+ { simde_mm_set_epi8(INT8_C( -85), INT8_C( -39), INT8_C(-107), INT8_C( -76),
42
+ INT8_C( -25), INT8_C( -32), INT8_C( -50), INT8_C(-115),
43
+ INT8_C( -54), INT8_C( -26), INT8_C( 56), INT8_C( 16),
44
+ INT8_C( -17), INT8_C( 98), INT8_C( -10), INT8_C( -86)),
45
+ simde_mm_set_epi8(INT8_C( 111), INT8_C( -17), INT8_C( 70), INT8_C( -98),
46
+ INT8_C( 38), INT8_C( 75), INT8_C( 17), INT8_C( -50),
47
+ INT8_C( -64), INT8_C( 25), INT8_C( -4), INT8_C( 42),
48
+ INT8_C( 13), INT8_C( 109), INT8_C( 122), INT8_C(-104)),
49
+ simde_mm_set_epi8(INT8_C( -13), INT8_C( 21), INT8_C( 111), INT8_C( -8),
50
+ INT8_C( 32), INT8_C( -14), INT8_C( 25), INT8_C( 45),
51
+ INT8_C( 34), INT8_C( -43), INT8_C( -81), INT8_C( 26),
52
+ INT8_C(-105), INT8_C( 63), INT8_C( -61), INT8_C( -38)),
53
+ simde_mm_set_epi8(INT8_C( 111), INT8_C( -39), INT8_C(-107), INT8_C( -98),
54
+ INT8_C( -25), INT8_C( 75), INT8_C( -50), INT8_C(-115),
55
+ INT8_C( -54), INT8_C( 25), INT8_C( -4), INT8_C( 16),
56
+ INT8_C( 13), INT8_C( 98), INT8_C( 122), INT8_C(-104)) },
57
+ { simde_mm_set_epi8(INT8_C( -79), INT8_C( 12), INT8_C( 105), INT8_C( -95),
58
+ INT8_C( -27), INT8_C( 97), INT8_C( 123), INT8_C( -2),
59
+ INT8_C( -75), INT8_C( -82), INT8_C( -63), INT8_C( 121),
60
+ INT8_C( 39), INT8_C( -40), INT8_C(-103), INT8_C( -72)),
61
+ simde_mm_set_epi8(INT8_C( -77), INT8_C( -58), INT8_C( -15), INT8_C( -99),
62
+ INT8_C(-120), INT8_C( 124), INT8_C( 40), INT8_C( 82),
63
+ INT8_C( 121), INT8_C( 50), INT8_C(-103), INT8_C( -45),
64
+ INT8_C(-125), INT8_C( -55), INT8_C( -57), INT8_C( -95)),
65
+ simde_mm_set_epi8(INT8_C( -97), INT8_C( -10), INT8_C( 107), INT8_C( 95),
66
+ INT8_C( 99), INT8_C(-103), INT8_C( 112), INT8_C( -80),
67
+ INT8_C( -21), INT8_C( 52), INT8_C(-128), INT8_C( 9),
68
+ INT8_C( 47), INT8_C( -45), INT8_C( -98), INT8_C( 67)),
69
+ simde_mm_set_epi8(INT8_C( -77), INT8_C( -58), INT8_C( 105), INT8_C( -95),
70
+ INT8_C( -27), INT8_C( 124), INT8_C( 123), INT8_C( 82),
71
+ INT8_C( 121), INT8_C( -82), INT8_C(-103), INT8_C( 121),
72
+ INT8_C( 39), INT8_C( -55), INT8_C( -57), INT8_C( -72)) },
73
+ { simde_mm_set_epi8(INT8_C( 65), INT8_C( 33), INT8_C( -24), INT8_C( -68),
74
+ INT8_C( -81), INT8_C( 103), INT8_C( -77), INT8_C( -49),
75
+ INT8_C( 99), INT8_C( 55), INT8_C( 98), INT8_C( 36),
76
+ INT8_C( 31), INT8_C( 86), INT8_C( -79), INT8_C( 106)),
77
+ simde_mm_set_epi8(INT8_C( 101), INT8_C( 91), INT8_C( -24), INT8_C( 33),
78
+ INT8_C( -68), INT8_C(-109), INT8_C( -92), INT8_C( -49),
79
+ INT8_C( -88), INT8_C( 86), INT8_C( 106), INT8_C( 125),
80
+ INT8_C( 81), INT8_C( -60), INT8_C( -32), INT8_C( 38)),
81
+ simde_mm_set_epi8(INT8_C( 3), INT8_C( 58), INT8_C( 40), INT8_C( 64),
82
+ INT8_C( 7), INT8_C(-123), INT8_C( 47), INT8_C( 24),
83
+ INT8_C( -14), INT8_C( 119), INT8_C( 8), INT8_C( 126),
84
+ INT8_C( 123), INT8_C( 25), INT8_C( 42), INT8_C(-118)),
85
+ simde_mm_set_epi8(INT8_C( 65), INT8_C( 33), INT8_C( -24), INT8_C( -68),
86
+ INT8_C( -81), INT8_C(-109), INT8_C( -77), INT8_C( -49),
87
+ INT8_C( -88), INT8_C( 55), INT8_C( 98), INT8_C( 36),
88
+ INT8_C( 31), INT8_C( 86), INT8_C( -79), INT8_C( 38)) },
89
+ { simde_mm_set_epi8(INT8_C( 34), INT8_C(-102), INT8_C( 26), INT8_C( 6),
90
+ INT8_C( -98), INT8_C(-126), INT8_C( 34), INT8_C( 109),
91
+ INT8_C( -65), INT8_C( -47), INT8_C( 54), INT8_C(-102),
92
+ INT8_C( 52), INT8_C( 72), INT8_C( 95), INT8_C(-117)),
93
+ simde_mm_set_epi8(INT8_C( -47), INT8_C( 2), INT8_C( 12), INT8_C( -2),
94
+ INT8_C( 113), INT8_C( 15), INT8_C( -36), INT8_C( 63),
95
+ INT8_C( 39), INT8_C( 47), INT8_C( -65), INT8_C( 45),
96
+ INT8_C( -99), INT8_C( 91), INT8_C(-116), INT8_C( 94)),
97
+ simde_mm_set_epi8(INT8_C( -98), INT8_C( 39), INT8_C( 27), INT8_C(-110),
98
+ INT8_C( -56), INT8_C( 121), INT8_C( 15), INT8_C( 100),
99
+ INT8_C( 94), INT8_C( 8), INT8_C(-121), INT8_C( 116),
100
+ INT8_C( 47), INT8_C(-128), INT8_C( 82), INT8_C( 52)),
101
+ simde_mm_set_epi8(INT8_C( -47), INT8_C(-102), INT8_C( 26), INT8_C( -2),
102
+ INT8_C( 113), INT8_C(-126), INT8_C( 34), INT8_C( 109),
103
+ INT8_C( -65), INT8_C( -47), INT8_C( -65), INT8_C(-102),
104
+ INT8_C( 52), INT8_C( 91), INT8_C( 95), INT8_C(-117)) },
105
+ { simde_mm_set_epi8(INT8_C( 114), INT8_C(-111), INT8_C( -9), INT8_C( -74),
106
+ INT8_C( 56), INT8_C( -88), INT8_C( 102), INT8_C( 9),
107
+ INT8_C( 123), INT8_C(-124), INT8_C( 2), INT8_C( -90),
108
+ INT8_C( -87), INT8_C(-122), INT8_C( -39), INT8_C(-114)),
109
+ simde_mm_set_epi8(INT8_C( -81), INT8_C( -26), INT8_C( 59), INT8_C( 3),
110
+ INT8_C( 113), INT8_C( -96), INT8_C(-108), INT8_C( 77),
111
+ INT8_C( -17), INT8_C( 59), INT8_C( -35), INT8_C( 43),
112
+ INT8_C( 117), INT8_C( 11), INT8_C(-103), INT8_C( -21)),
113
+ simde_mm_set_epi8(INT8_C(-116), INT8_C( 124), INT8_C( -23), INT8_C( -24),
114
+ INT8_C(-105), INT8_C( -39), INT8_C( 45), INT8_C( -32),
115
+ INT8_C( 5), INT8_C( -60), INT8_C( -98), INT8_C( 111),
116
+ INT8_C(-112), INT8_C( -60), INT8_C( 8), INT8_C( 13)),
117
+ simde_mm_set_epi8(INT8_C( -81), INT8_C(-111), INT8_C( 59), INT8_C( 3),
118
+ INT8_C( 113), INT8_C( -96), INT8_C( 102), INT8_C( 77),
119
+ INT8_C( 123), INT8_C( 59), INT8_C( -35), INT8_C( -90),
120
+ INT8_C( 117), INT8_C( 11), INT8_C( -39), INT8_C(-114)) },
121
+ { simde_mm_set_epi8(INT8_C( 53), INT8_C( -97), INT8_C( 79), INT8_C( -93),
122
+ INT8_C( 55), INT8_C(-115), INT8_C( -80), INT8_C( -27),
123
+ INT8_C( 104), INT8_C( 109), INT8_C( -16), INT8_C( -75),
124
+ INT8_C( 125), INT8_C( 6), INT8_C( 38), INT8_C( -55)),
125
+ simde_mm_set_epi8(INT8_C( 48), INT8_C( 25), INT8_C( 24), INT8_C( -17),
126
+ INT8_C(-102), INT8_C( -66), INT8_C( -6), INT8_C( 22),
127
+ INT8_C(-100), INT8_C( -56), INT8_C( 25), INT8_C( -26),
128
+ INT8_C( -37), INT8_C( 49), INT8_C( 8), INT8_C( -6)),
129
+ simde_mm_set_epi8(INT8_C(-114), INT8_C( 27), INT8_C( -68), INT8_C( 3),
130
+ INT8_C( 88), INT8_C( 71), INT8_C( 77), INT8_C( 108),
131
+ INT8_C(-112), INT8_C( 117), INT8_C(-100), INT8_C( -26),
132
+ INT8_C( 86), INT8_C( -48), INT8_C( -34), INT8_C( -71)),
133
+ simde_mm_set_epi8(INT8_C( 48), INT8_C( -97), INT8_C( 24), INT8_C( -93),
134
+ INT8_C( 55), INT8_C(-115), INT8_C( -80), INT8_C( -27),
135
+ INT8_C(-100), INT8_C( 109), INT8_C( 25), INT8_C( -26),
136
+ INT8_C( 125), INT8_C( 49), INT8_C( 8), INT8_C( -6)) },
137
+ { simde_mm_set_epi8(INT8_C(-121), INT8_C( 29), INT8_C( -87), INT8_C( 80),
138
+ INT8_C( 19), INT8_C( 64), INT8_C( 104), INT8_C( 67),
139
+ INT8_C( -22), INT8_C( -16), INT8_C( -57), INT8_C( -78),
140
+ INT8_C( -96), INT8_C(-113), INT8_C( 114), INT8_C( 15)),
141
+ simde_mm_set_epi8(INT8_C( 99), INT8_C( -98), INT8_C( 122), INT8_C( 119),
142
+ INT8_C( 80), INT8_C( -87), INT8_C( -77), INT8_C( -36),
143
+ INT8_C( 10), INT8_C( 99), INT8_C( 114), INT8_C( -46),
144
+ INT8_C( -79), INT8_C( -58), INT8_C( -87), INT8_C( 106)),
145
+ simde_mm_set_epi8(INT8_C( -96), INT8_C( 67), INT8_C( -85), INT8_C( 80),
146
+ INT8_C( 107), INT8_C( 9), INT8_C( 20), INT8_C( 91),
147
+ INT8_C( -58), INT8_C( -6), INT8_C( 120), INT8_C( -24),
148
+ INT8_C( 26), INT8_C( -94), INT8_C( -17), INT8_C( 25)),
149
+ simde_mm_set_epi8(INT8_C( 99), INT8_C( 29), INT8_C( 122), INT8_C( 80),
150
+ INT8_C( 19), INT8_C( 64), INT8_C( 104), INT8_C( 67),
151
+ INT8_C( 10), INT8_C( 99), INT8_C( -57), INT8_C( -46),
152
+ INT8_C( -96), INT8_C( -58), INT8_C( -87), INT8_C( 15)) },
153
+ { simde_mm_set_epi8(INT8_C( -13), INT8_C( 28), INT8_C( -62), INT8_C( 108),
154
+ INT8_C( -15), INT8_C( 17), INT8_C( -48), INT8_C( 13),
155
+ INT8_C( 32), INT8_C( -56), INT8_C( 83), INT8_C( -65),
156
+ INT8_C( 73), INT8_C(-114), INT8_C( 47), INT8_C( 99)),
157
+ simde_mm_set_epi8(INT8_C( 84), INT8_C( -32), INT8_C( -4), INT8_C( 8),
158
+ INT8_C( 66), INT8_C( -71), INT8_C( -69), INT8_C( -41),
159
+ INT8_C( -21), INT8_C( -40), INT8_C( 64), INT8_C( -47),
160
+ INT8_C( 16), INT8_C( -61), INT8_C( 3), INT8_C( -15)),
161
+ simde_mm_set_epi8(INT8_C( 27), INT8_C( 104), INT8_C( -26), INT8_C( -51),
162
+ INT8_C( 114), INT8_C( -9), INT8_C( -56), INT8_C( 125),
163
+ INT8_C( 41), INT8_C( -51), INT8_C( 26), INT8_C( 121),
164
+ INT8_C( -16), INT8_C( -59), INT8_C( 93), INT8_C( -56)),
165
+ simde_mm_set_epi8(INT8_C( -13), INT8_C( 28), INT8_C( -4), INT8_C( 8),
166
+ INT8_C( -15), INT8_C( -71), INT8_C( -69), INT8_C( 13),
167
+ INT8_C( 32), INT8_C( -40), INT8_C( 83), INT8_C( -65),
168
+ INT8_C( 16), INT8_C( -61), INT8_C( 47), INT8_C( -15)) }
169
+ };
170
+
171
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
172
+ simde__m128i r = simde_mm_blendv_epi8(test_vec[i].a, test_vec[i].b, test_vec[i].mask);
173
+ simde_assert_m128i_i8(r, ==, test_vec[i].r);
174
+ }
175
+
176
+ return MUNIT_OK;
177
+ }
178
+
179
+ static MunitResult
180
+ test_simde_x_mm_blendv_epi16(const MunitParameter params[], void* data) {
181
+ (void) params;
182
+ (void) data;
183
+
184
+ const struct {
185
+ simde__m128i a;
186
+ simde__m128i b;
187
+ simde__m128i mask;
188
+ simde__m128i r;
189
+ } test_vec[8] = {
190
+ { simde_mm_set_epi16(INT16_C( -5889), INT16_C( -9240), INT16_C( 13406), INT16_C( 23197),
191
+ INT16_C( 25390), INT16_C( 20613), INT16_C( -3211), INT16_C( -5281)),
192
+ simde_mm_set_epi16(INT16_C( -5787), INT16_C( 260), INT16_C(-28529), INT16_C( -341),
193
+ INT16_C( 26074), INT16_C(-30339), INT16_C( -9562), INT16_C( -4010)),
194
+ simde_mm_set_epi16(INT16_C( 11580), INT16_C(-26297), INT16_C( 8234), INT16_C( -6769),
195
+ INT16_C( -3010), INT16_C( 26987), INT16_C( 26235), INT16_C(-23439)),
196
+ simde_mm_set_epi16(INT16_C( -5889), INT16_C( 260), INT16_C( 13406), INT16_C( -341),
197
+ INT16_C( 26074), INT16_C( 20613), INT16_C( -3211), INT16_C( -4010)) },
198
+ { simde_mm_set_epi16(INT16_C(-10477), INT16_C( 1551), INT16_C(-16362), INT16_C(-15016),
199
+ INT16_C( -3374), INT16_C(-19084), INT16_C( -221), INT16_C(-19089)),
200
+ simde_mm_set_epi16(INT16_C( 8603), INT16_C( 8519), INT16_C( 4642), INT16_C( 1129),
201
+ INT16_C( 6053), INT16_C( 28257), INT16_C( 10987), INT16_C(-17611)),
202
+ simde_mm_set_epi16(INT16_C(-14703), INT16_C( 23402), INT16_C( -5363), INT16_C( 25451),
203
+ INT16_C( 4787), INT16_C( 6143), INT16_C( -8049), INT16_C(-25291)),
204
+ simde_mm_set_epi16(INT16_C( 8603), INT16_C( 1551), INT16_C( 4642), INT16_C(-15016),
205
+ INT16_C( -3374), INT16_C(-19084), INT16_C( 10987), INT16_C(-17611)) },
206
+ { simde_mm_set_epi16(INT16_C(-10958), INT16_C( 239), INT16_C(-29416), INT16_C( 25642),
207
+ INT16_C(-31255), INT16_C( 18557), INT16_C( 17603), INT16_C( 9411)),
208
+ simde_mm_set_epi16(INT16_C(-22204), INT16_C(-12104), INT16_C( 21385), INT16_C( 163),
209
+ INT16_C(-13806), INT16_C( -4673), INT16_C(-31502), INT16_C( 30113)),
210
+ simde_mm_set_epi16(INT16_C( 26590), INT16_C(-27173), INT16_C( 3052), INT16_C( 29423),
211
+ INT16_C( 14159), INT16_C( 16581), INT16_C( 15114), INT16_C( 31350)),
212
+ simde_mm_set_epi16(INT16_C(-10958), INT16_C(-12104), INT16_C(-29416), INT16_C( 25642),
213
+ INT16_C(-31255), INT16_C( 18557), INT16_C( 17603), INT16_C( 9411)) },
214
+ { simde_mm_set_epi16(INT16_C( -8795), INT16_C(-24496), INT16_C(-21018), INT16_C(-23768),
215
+ INT16_C( 25939), INT16_C( -2265), INT16_C( 13801), INT16_C( 28390)),
216
+ simde_mm_set_epi16(INT16_C(-31776), INT16_C( 7886), INT16_C( 19773), INT16_C( 11337),
217
+ INT16_C( -9683), INT16_C(-15212), INT16_C(-11745), INT16_C(-21367)),
218
+ simde_mm_set_epi16(INT16_C(-30437), INT16_C( -4024), INT16_C(-15989), INT16_C( -4852),
219
+ INT16_C( 20067), INT16_C( 28092), INT16_C(-20772), INT16_C( 17444)),
220
+ simde_mm_set_epi16(INT16_C(-31776), INT16_C( 7886), INT16_C( 19773), INT16_C( 11337),
221
+ INT16_C( 25939), INT16_C( -2265), INT16_C(-11745), INT16_C( 28390)) },
222
+ { simde_mm_set_epi16(INT16_C( 11712), INT16_C( 21433), INT16_C( -1195), INT16_C( 25259),
223
+ INT16_C( 18957), INT16_C( 5262), INT16_C( -9688), INT16_C(-14506)),
224
+ simde_mm_set_epi16(INT16_C( 24574), INT16_C( 2777), INT16_C(-18691), INT16_C( 4928),
225
+ INT16_C( 32122), INT16_C(-10541), INT16_C( 2746), INT16_C( 8282)),
226
+ simde_mm_set_epi16(INT16_C( 5301), INT16_C( 12242), INT16_C( 18664), INT16_C( 26110),
227
+ INT16_C(-14367), INT16_C( -9137), INT16_C( -7209), INT16_C( 24768)),
228
+ simde_mm_set_epi16(INT16_C( 11712), INT16_C( 21433), INT16_C( -1195), INT16_C( 25259),
229
+ INT16_C( 32122), INT16_C(-10541), INT16_C( 2746), INT16_C(-14506)) },
230
+ { simde_mm_set_epi16(INT16_C(-20542), INT16_C( 3332), INT16_C( 24354), INT16_C(-23695),
231
+ INT16_C(-10493), INT16_C( 32452), INT16_C(-26923), INT16_C( 16579)),
232
+ simde_mm_set_epi16(INT16_C( 29156), INT16_C( 8347), INT16_C(-32604), INT16_C(-24707),
233
+ INT16_C(-20393), INT16_C(-18566), INT16_C( 17032), INT16_C(-18520)),
234
+ simde_mm_set_epi16(INT16_C(-23886), INT16_C( 30820), INT16_C( 12721), INT16_C(-28893),
235
+ INT16_C( -7688), INT16_C( 11154), INT16_C(-11967), INT16_C( 31969)),
236
+ simde_mm_set_epi16(INT16_C( 29156), INT16_C( 3332), INT16_C( 24354), INT16_C(-24707),
237
+ INT16_C(-20393), INT16_C( 32452), INT16_C( 17032), INT16_C( 16579)) },
238
+ { simde_mm_set_epi16(INT16_C( 22583), INT16_C( 4415), INT16_C(-24878), INT16_C(-15401),
239
+ INT16_C(-29840), INT16_C(-16999), INT16_C(-23406), INT16_C(-32259)),
240
+ simde_mm_set_epi16(INT16_C( 31553), INT16_C( 23166), INT16_C(-15560), INT16_C( 31886),
241
+ INT16_C( 7368), INT16_C(-15822), INT16_C(-17797), INT16_C( 1487)),
242
+ simde_mm_set_epi16(INT16_C(-12147), INT16_C( 2989), INT16_C(-24661), INT16_C(-22664),
243
+ INT16_C(-16476), INT16_C( 408), INT16_C(-13750), INT16_C(-15888)),
244
+ simde_mm_set_epi16(INT16_C( 31553), INT16_C( 4415), INT16_C(-15560), INT16_C( 31886),
245
+ INT16_C( 7368), INT16_C(-16999), INT16_C(-17797), INT16_C( 1487)) },
246
+ { simde_mm_set_epi16(INT16_C( 32299), INT16_C( 7492), INT16_C( 26066), INT16_C( 30675),
247
+ INT16_C( 26958), INT16_C( 8712), INT16_C(-17582), INT16_C(-18062)),
248
+ simde_mm_set_epi16(INT16_C( 25592), INT16_C( 2248), INT16_C(-21774), INT16_C( 7511),
249
+ INT16_C( 6101), INT16_C( 18791), INT16_C(-31731), INT16_C( 22067)),
250
+ simde_mm_set_epi16(INT16_C(-26320), INT16_C( 16537), INT16_C(-11326), INT16_C( 8393),
251
+ INT16_C(-25832), INT16_C( 24810), INT16_C(-17206), INT16_C(-26982)),
252
+ simde_mm_set_epi16(INT16_C( 25592), INT16_C( 7492), INT16_C(-21774), INT16_C( 30675),
253
+ INT16_C( 6101), INT16_C( 8712), INT16_C(-31731), INT16_C( 22067)) }
254
+ };
255
+
256
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
257
+ simde__m128i r = simde_x_mm_blendv_epi16(test_vec[i].a, test_vec[i].b, test_vec[i].mask);
258
+ simde_assert_m128i_i16(r, ==, test_vec[i].r);
259
+ }
260
+
261
+ return MUNIT_OK;
262
+ }
263
+
264
+ static MunitResult
265
+ test_simde_x_mm_blendv_epi32(const MunitParameter params[], void* data) {
266
+ (void) params;
267
+ (void) data;
268
+
269
+ const struct {
270
+ simde__m128i a;
271
+ simde__m128i b;
272
+ simde__m128i mask;
273
+ simde__m128i r;
274
+ } test_vec[8] = {
275
+ { simde_mm_set_epi32(INT32_C( -385885208), INT32_C( 878598813), INT32_C( 1663979653), INT32_C( -210375841)),
276
+ simde_mm_set_epi32(INT32_C( -379256572), INT32_C(-1869611349), INT32_C( 1708820861), INT32_C( -626593706)),
277
+ simde_mm_set_epi32(INT32_C( 758946119), INT32_C( 539682191), INT32_C( -197236373), INT32_C( 1719379057)),
278
+ simde_mm_set_epi32(INT32_C( -385885208), INT32_C( 878598813), INT32_C( 1708820861), INT32_C( -210375841)) },
279
+ { simde_mm_set_epi32(INT32_C( -686619121), INT32_C(-1072249512), INT32_C( -221072012), INT32_C( -14437009)),
280
+ simde_mm_set_epi32(INT32_C( 563814727), INT32_C( 304219241), INT32_C( 396717665), INT32_C( 720091957)),
281
+ simde_mm_set_epi32(INT32_C( -963552406), INT32_C( -351444117), INT32_C( 313726975), INT32_C( -527459019)),
282
+ simde_mm_set_epi32(INT32_C( 563814727), INT32_C( 304219241), INT32_C( -221072012), INT32_C( 720091957)) },
283
+ { simde_mm_set_epi32(INT32_C( -718143249), INT32_C(-1927781334), INT32_C(-2048309123), INT32_C( 1153639619)),
284
+ simde_mm_set_epi32(INT32_C(-1455107912), INT32_C( 1401487523), INT32_C( -904729153), INT32_C(-2064484959)),
285
+ simde_mm_set_epi32(INT32_C( 1742640603), INT32_C( 200045295), INT32_C( 927940805), INT32_C( 990542454)),
286
+ simde_mm_set_epi32(INT32_C( -718143249), INT32_C(-1927781334), INT32_C(-2048309123), INT32_C( 1153639619)) },
287
+ { simde_mm_set_epi32(INT32_C( -576348080), INT32_C(-1377393880), INT32_C( 1700001575), INT32_C( 904490726)),
288
+ simde_mm_set_epi32(INT32_C(-2082464050), INT32_C( 1295854665), INT32_C( -634534764), INT32_C( -769676151)),
289
+ simde_mm_set_epi32(INT32_C(-1994657720), INT32_C(-1047794420), INT32_C( 1315139004), INT32_C(-1361296348)),
290
+ simde_mm_set_epi32(INT32_C(-2082464050), INT32_C( 1295854665), INT32_C( 1700001575), INT32_C( -769676151)) },
291
+ { simde_mm_set_epi32(INT32_C( 767579065), INT32_C( -78290261), INT32_C( 1242371214), INT32_C( -634861738)),
292
+ simde_mm_set_epi32(INT32_C( 1610484441), INT32_C(-1224928448), INT32_C( 2105202387), INT32_C( 179970138)),
293
+ simde_mm_set_epi32(INT32_C( 347418578), INT32_C( 1223190014), INT32_C( -941499313), INT32_C( -472424256)),
294
+ simde_mm_set_epi32(INT32_C( 767579065), INT32_C( -78290261), INT32_C( 2105202387), INT32_C( 179970138)) },
295
+ { simde_mm_set_epi32(INT32_C(-1346237180), INT32_C( 1596105585), INT32_C( -687636796), INT32_C(-1764409149)),
296
+ simde_mm_set_epi32(INT32_C( 1910775963), INT32_C(-2136694915), INT32_C(-1336428678), INT32_C( 1116256168)),
297
+ simde_mm_set_epi32(INT32_C(-1565362076), INT32_C( 833720099), INT32_C( -503829614), INT32_C( -784237343)),
298
+ simde_mm_set_epi32(INT32_C( 1910775963), INT32_C( 1596105585), INT32_C(-1336428678), INT32_C( 1116256168)) },
299
+ { simde_mm_set_epi32(INT32_C( 1480003903), INT32_C(-1630354473), INT32_C(-1955545703), INT32_C(-1533902339)),
300
+ simde_mm_set_epi32(INT32_C( 2067880574), INT32_C(-1019708274), INT32_C( 482918962), INT32_C(-1166342705)),
301
+ simde_mm_set_epi32(INT32_C( -796062803), INT32_C(-1616140424), INT32_C(-1079770728), INT32_C( -901070352)),
302
+ simde_mm_set_epi32(INT32_C( 2067880574), INT32_C(-1019708274), INT32_C( 482918962), INT32_C(-1166342705)) },
303
+ { simde_mm_set_epi32(INT32_C( 2116754756), INT32_C( 1708292051), INT32_C( 1766728200), INT32_C(-1152206478)),
304
+ simde_mm_set_epi32(INT32_C( 1677199560), INT32_C(-1426973353), INT32_C( 399853927), INT32_C(-2079500749)),
305
+ simde_mm_set_epi32(INT32_C(-1724890983), INT32_C( -742252343), INT32_C(-1692901142), INT32_C(-1127573862)),
306
+ simde_mm_set_epi32(INT32_C( 1677199560), INT32_C(-1426973353), INT32_C( 399853927), INT32_C(-2079500749)) }
307
+ };
308
+
309
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
310
+ simde__m128i r = simde_x_mm_blendv_epi32(test_vec[i].a, test_vec[i].b, test_vec[i].mask);
311
+ simde_assert_m128i_i32(r, ==, test_vec[i].r);
312
+ }
313
+
314
+ return MUNIT_OK;
315
+ }
316
+
317
+ static MunitResult
318
+ test_simde_x_mm_blendv_epi64(const MunitParameter params[], void* data) {
319
+ (void) params;
320
+ (void) data;
321
+
322
+ const struct {
323
+ simde__m128i a;
324
+ simde__m128i b;
325
+ simde__m128i mask;
326
+ simde__m128i r;
327
+ } test_vec[8] = {
328
+ { simde_mm_set_epi64x(INT64_C(-1657364347491558755), INT64_C( 7146738194929019743)),
329
+ simde_mm_set_epi64x(INT64_C(-1628894571107713365), INT64_C( 7339329716385935446)),
330
+ simde_mm_set_epi64x(INT64_C( 3259648761070806415), INT64_C( -847123769897278351)),
331
+ simde_mm_set_epi64x(INT64_C(-1657364347491558755), INT64_C( 7339329716385935446)) },
332
+ { simde_mm_set_epi64x(INT64_C(-2949006666280549032), INT64_C( -949497057320389265)),
333
+ simde_mm_set_epi64x(INT64_C( 2421565813772387433), INT64_C( 1703889397640575797)),
334
+ simde_mm_set_epi64x(INT64_C(-4138426067808590997), INT64_C( 1347447101265517877)),
335
+ simde_mm_set_epi64x(INT64_C( 2421565813772387433), INT64_C( -949497057320389265)) },
336
+ { simde_mm_set_epi64x(INT64_C(-3084401765930998742), INT64_C(-8797420694229801789)),
337
+ simde_mm_set_epi64x(INT64_C(-6249640892789358429), INT64_C(-3885782121642297951)),
338
+ simde_mm_set_epi64x(INT64_C( 7484584398766764783), INT64_C( 3985475411089455734)),
339
+ simde_mm_set_epi64x(INT64_C(-3084401765930998742), INT64_C(-8797420694229801789)) },
340
+ { simde_mm_set_epi64x(INT64_C(-2475396151794818264), INT64_C( 7301451168677981926)),
341
+ simde_mm_set_epi64x(INT64_C(-8944114988549854135), INT64_C(-2725306056029786999)),
342
+ simde_mm_set_epi64x(INT64_C(-8566989670866752244), INT64_C( 5648479014807684132)),
343
+ simde_mm_set_epi64x(INT64_C(-8944114988549854135), INT64_C( 7301451168677981926)) },
344
+ { simde_mm_set_epi64x(INT64_C( 3296726985485935275), INT64_C( 5335943737281922902)),
345
+ simde_mm_set_epi64x(INT64_C( 6916978007881880384), INT64_C( 9041775403806105690)),
346
+ simde_mm_set_epi64x(INT64_C( 1492151431756015102), INT64_C(-4043708754718924608)),
347
+ simde_mm_set_epi64x(INT64_C( 3296726985485935275), INT64_C( 9041775403806105690)) },
348
+ { simde_mm_set_epi64x(INT64_C(-5782044659163159695), INT64_C(-2953377547815665469)),
349
+ simde_mm_set_epi64x(INT64_C( 8206720273226178429), INT64_C(-5739917464330258520)),
350
+ simde_mm_set_epi64x(INT64_C(-6723178921984946397), INT64_C(-2163931711375573791)),
351
+ simde_mm_set_epi64x(INT64_C( 8206720273226178429), INT64_C(-5739917464330258520)) },
352
+ { simde_mm_set_epi64x(INT64_C( 6356568364001969111), INT64_C(-8399004837457264131)),
353
+ simde_mm_set_epi64x(INT64_C( 8881479440638966926), INT64_C( 2074121151536891343)),
354
+ simde_mm_set_epi64x(INT64_C(-3419063701768263816), INT64_C(-4637579960544214544)),
355
+ simde_mm_set_epi64x(INT64_C( 8881479440638966926), INT64_C( 2074121151536891343)) },
356
+ { simde_mm_set_epi64x(INT64_C( 9091392452380751827), INT64_C( 7588039843063708018)),
357
+ simde_mm_set_epi64x(INT64_C( 7203517261933583703), INT64_C( 1717359541857637939)),
358
+ simde_mm_set_epi64x(INT64_C(-7408350357597577015), INT64_C(-7270955037083658598)),
359
+ simde_mm_set_epi64x(INT64_C( 7203517261933583703), INT64_C( 1717359541857637939)) }
360
+ };
361
+
362
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
363
+ simde__m128i r = simde_x_mm_blendv_epi64(test_vec[i].a, test_vec[i].b, test_vec[i].mask);
364
+ simde_assert_m128i_i64(r, ==, test_vec[i].r);
365
+ }
366
+
367
+ return MUNIT_OK;
368
+ }
369
+
370
+ static MunitResult
371
+ test_simde_mm_blend_epi16(const MunitParameter params[], void* data) {
372
+ (void) params;
373
+ (void) data;
374
+
375
+ const struct {
376
+ simde__m128i a;
377
+ simde__m128i b;
378
+ simde__m128i r;
379
+ } test_vec[8] = {
380
+ { simde_mm_set_epi16(INT16_C( 13825), INT16_C(-17802), INT16_C( -1598), INT16_C( 25326),
381
+ INT16_C( 14943), INT16_C( -8227), INT16_C( 21745), INT16_C( 9989)),
382
+ simde_mm_set_epi16(INT16_C(-11058), INT16_C( 4192), INT16_C(-28252), INT16_C( 10795),
383
+ INT16_C( 19700), INT16_C( 29704), INT16_C( -7052), INT16_C( 16432)),
384
+ simde_mm_set_epi16(INT16_C( 13825), INT16_C(-17802), INT16_C( -1598), INT16_C( 25326),
385
+ INT16_C( 14943), INT16_C( -8227), INT16_C( -7052), INT16_C( 16432)) },
386
+ { simde_mm_set_epi16(INT16_C(-17914), INT16_C(-32103), INT16_C( 17164), INT16_C( -187),
387
+ INT16_C( 21942), INT16_C(-25306), INT16_C( -7849), INT16_C(-11297)),
388
+ simde_mm_set_epi16(INT16_C( -9027), INT16_C(-14363), INT16_C( 16126), INT16_C( 30921),
389
+ INT16_C( 3730), INT16_C(-20492), INT16_C( -5510), INT16_C( 32723)),
390
+ simde_mm_set_epi16(INT16_C(-17914), INT16_C(-32103), INT16_C( 17164), INT16_C( -187),
391
+ INT16_C( 21942), INT16_C(-25306), INT16_C( -5510), INT16_C( 32723)) },
392
+ { simde_mm_set_epi16(INT16_C(-25152), INT16_C( 30665), INT16_C( 26270), INT16_C( 23820),
393
+ INT16_C(-24322), INT16_C( 706), INT16_C(-26336), INT16_C(-13296)),
394
+ simde_mm_set_epi16(INT16_C(-24671), INT16_C(-19680), INT16_C( 10653), INT16_C( -9500),
395
+ INT16_C(-16899), INT16_C( 495), INT16_C( 26780), INT16_C(-11315)),
396
+ simde_mm_set_epi16(INT16_C(-25152), INT16_C( 30665), INT16_C( 26270), INT16_C( 23820),
397
+ INT16_C(-24322), INT16_C( 706), INT16_C( 26780), INT16_C(-11315)) },
398
+ { simde_mm_set_epi16(INT16_C( 20562), INT16_C(-19834), INT16_C( 5180), INT16_C( -5117),
399
+ INT16_C( 16168), INT16_C( 23520), INT16_C(-16838), INT16_C( 28892)),
400
+ simde_mm_set_epi16(INT16_C( 2860), INT16_C( -6670), INT16_C(-25365), INT16_C( -4954),
401
+ INT16_C( 30912), INT16_C( 6045), INT16_C( 2749), INT16_C( 4401)),
402
+ simde_mm_set_epi16(INT16_C( 20562), INT16_C(-19834), INT16_C( 5180), INT16_C( -5117),
403
+ INT16_C( 16168), INT16_C( 23520), INT16_C( 2749), INT16_C( 4401)) },
404
+ { simde_mm_set_epi16(INT16_C( 25997), INT16_C( -1860), INT16_C( -9126), INT16_C( 12214),
405
+ INT16_C( 7159), INT16_C( -2303), INT16_C( 7182), INT16_C(-20779)),
406
+ simde_mm_set_epi16(INT16_C( -7432), INT16_C( 954), INT16_C(-26453), INT16_C( -2062),
407
+ INT16_C( 14984), INT16_C( -7563), INT16_C( 13369), INT16_C( -4134)),
408
+ simde_mm_set_epi16(INT16_C( 25997), INT16_C( -1860), INT16_C( -9126), INT16_C( 12214),
409
+ INT16_C( 7159), INT16_C( -2303), INT16_C( 13369), INT16_C( -4134)) },
410
+ { simde_mm_set_epi16(INT16_C( -6104), INT16_C(-18564), INT16_C( 5298), INT16_C( -6001),
411
+ INT16_C( -6122), INT16_C( 14804), INT16_C(-32440), INT16_C(-10960)),
412
+ simde_mm_set_epi16(INT16_C( 8197), INT16_C( 14220), INT16_C(-10924), INT16_C( 17196),
413
+ INT16_C( 21768), INT16_C( 7639), INT16_C( -4935), INT16_C( -993)),
414
+ simde_mm_set_epi16(INT16_C( -6104), INT16_C(-18564), INT16_C( 5298), INT16_C( -6001),
415
+ INT16_C( -6122), INT16_C( 14804), INT16_C( -4935), INT16_C( -993)) },
416
+ { simde_mm_set_epi16(INT16_C( 6003), INT16_C( -5880), INT16_C( 31967), INT16_C(-28984),
417
+ INT16_C( -9503), INT16_C( 30320), INT16_C(-19146), INT16_C( 32146)),
418
+ simde_mm_set_epi16(INT16_C( 4317), INT16_C( 22644), INT16_C(-25807), INT16_C(-30524),
419
+ INT16_C( 223), INT16_C( -7098), INT16_C( 9867), INT16_C( -9190)),
420
+ simde_mm_set_epi16(INT16_C( 6003), INT16_C( -5880), INT16_C( 31967), INT16_C(-28984),
421
+ INT16_C( -9503), INT16_C( 30320), INT16_C( 9867), INT16_C( -9190)) },
422
+ { simde_mm_set_epi16(INT16_C( 12033), INT16_C( 14639), INT16_C( 28420), INT16_C( 14334),
423
+ INT16_C( -3029), INT16_C(-20328), INT16_C( 8988), INT16_C( 23547)),
424
+ simde_mm_set_epi16(INT16_C( 9966), INT16_C(-21867), INT16_C( 738), INT16_C(-29688),
425
+ INT16_C( 9349), INT16_C( 2360), INT16_C(-12771), INT16_C(-15402)),
426
+ simde_mm_set_epi16(INT16_C( 12033), INT16_C( 14639), INT16_C( 28420), INT16_C( 14334),
427
+ INT16_C( -3029), INT16_C(-20328), INT16_C(-12771), INT16_C(-15402)) }
428
+ };
429
+
430
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
431
+ simde__m128i r = simde_mm_blend_epi16(test_vec[i].a, test_vec[i].b, 3);
432
+ simde_assert_m128i_i16(r, ==, test_vec[i].r);
433
+ }
434
+
435
+ return MUNIT_OK;
436
+ }
437
+
438
+ static MunitResult
439
+ test_simde_mm_blend_pd(const MunitParameter params[], void* data) {
440
+ (void) params;
441
+ (void) data;
442
+
443
+ const struct {
444
+ simde__m128d a;
445
+ simde__m128d b;
446
+ simde__m128d r;
447
+ } test_vec[8] = {
448
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( 307.09), SIMDE_FLOAT64_C( -991.93)),
449
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -252.92), SIMDE_FLOAT64_C( -56.61)),
450
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 307.09), SIMDE_FLOAT64_C( -56.61)) },
451
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( 263.25), SIMDE_FLOAT64_C( 576.67)),
452
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -398.47), SIMDE_FLOAT64_C( -845.00)),
453
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 263.25), SIMDE_FLOAT64_C( -845.00)) },
454
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( -156.21), SIMDE_FLOAT64_C( 805.98)),
455
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -712.32), SIMDE_FLOAT64_C( -557.85)),
456
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -156.21), SIMDE_FLOAT64_C( -557.85)) },
457
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( 191.78), SIMDE_FLOAT64_C( -309.61)),
458
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -997.82), SIMDE_FLOAT64_C( -493.58)),
459
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 191.78), SIMDE_FLOAT64_C( -493.58)) },
460
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( -933.58), SIMDE_FLOAT64_C( -486.36)),
461
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 880.67), SIMDE_FLOAT64_C( 791.10)),
462
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -933.58), SIMDE_FLOAT64_C( 791.10)) },
463
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( 45.45), SIMDE_FLOAT64_C( -436.38)),
464
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 211.82), SIMDE_FLOAT64_C( 653.98)),
465
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 45.45), SIMDE_FLOAT64_C( 653.98)) },
466
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( 915.51), SIMDE_FLOAT64_C( -445.43)),
467
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -69.14), SIMDE_FLOAT64_C( -345.82)),
468
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 915.51), SIMDE_FLOAT64_C( -345.82)) },
469
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( -603.21), SIMDE_FLOAT64_C( -182.69)),
470
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -463.45), SIMDE_FLOAT64_C( 929.21)),
471
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -603.21), SIMDE_FLOAT64_C( 929.21)) }
472
+ };
473
+
474
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
475
+ simde__m128d r = simde_mm_blend_pd(test_vec[i].a, test_vec[i].b, 1);
476
+ simde_assert_m128d_close(r, test_vec[i].r, 1);
477
+ }
478
+
479
+ return MUNIT_OK;
480
+ }
481
+
482
+ static MunitResult
483
+ test_simde_mm_blend_ps(const MunitParameter params[], void* data) {
484
+ (void) params;
485
+ (void) data;
486
+
487
+ const struct {
488
+ simde__m128 a;
489
+ simde__m128 b;
490
+ simde__m128 r;
491
+ } test_vec[8] = {
492
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( 263.25), SIMDE_FLOAT32_C( 576.67), SIMDE_FLOAT32_C( 307.09), SIMDE_FLOAT32_C( -991.93)),
493
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -398.47), SIMDE_FLOAT32_C( -845.00), SIMDE_FLOAT32_C( -252.92), SIMDE_FLOAT32_C( -56.61)),
494
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 263.25), SIMDE_FLOAT32_C( 576.67), SIMDE_FLOAT32_C( -252.92), SIMDE_FLOAT32_C( -991.93)) },
495
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( 191.78), SIMDE_FLOAT32_C( -309.61), SIMDE_FLOAT32_C( -156.21), SIMDE_FLOAT32_C( 805.98)),
496
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -997.82), SIMDE_FLOAT32_C( -493.58), SIMDE_FLOAT32_C( -712.32), SIMDE_FLOAT32_C( -557.85)),
497
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 191.78), SIMDE_FLOAT32_C( -309.61), SIMDE_FLOAT32_C( -712.32), SIMDE_FLOAT32_C( 805.98)) },
498
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( 45.45), SIMDE_FLOAT32_C( -436.38), SIMDE_FLOAT32_C( -933.58), SIMDE_FLOAT32_C( -486.36)),
499
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 211.82), SIMDE_FLOAT32_C( 653.98), SIMDE_FLOAT32_C( 880.67), SIMDE_FLOAT32_C( 791.10)),
500
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 45.45), SIMDE_FLOAT32_C( -436.38), SIMDE_FLOAT32_C( 880.67), SIMDE_FLOAT32_C( -486.36)) },
501
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( -603.21), SIMDE_FLOAT32_C( -182.69), SIMDE_FLOAT32_C( 915.51), SIMDE_FLOAT32_C( -445.43)),
502
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -463.45), SIMDE_FLOAT32_C( 929.21), SIMDE_FLOAT32_C( -69.14), SIMDE_FLOAT32_C( -345.82)),
503
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -603.21), SIMDE_FLOAT32_C( -182.69), SIMDE_FLOAT32_C( -69.14), SIMDE_FLOAT32_C( -445.43)) },
504
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( 898.89), SIMDE_FLOAT32_C( -75.69), SIMDE_FLOAT32_C( 396.94), SIMDE_FLOAT32_C( -959.19)),
505
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -419.62), SIMDE_FLOAT32_C( -299.33), SIMDE_FLOAT32_C( -147.28), SIMDE_FLOAT32_C( -458.75)),
506
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 898.89), SIMDE_FLOAT32_C( -75.69), SIMDE_FLOAT32_C( -147.28), SIMDE_FLOAT32_C( -959.19)) },
507
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( 480.97), SIMDE_FLOAT32_C( 316.97), SIMDE_FLOAT32_C( -700.74), SIMDE_FLOAT32_C( -81.35)),
508
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -930.08), SIMDE_FLOAT32_C( 58.79), SIMDE_FLOAT32_C( -729.84), SIMDE_FLOAT32_C( 572.31)),
509
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 480.97), SIMDE_FLOAT32_C( 316.97), SIMDE_FLOAT32_C( -729.84), SIMDE_FLOAT32_C( -81.35)) },
510
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( -51.53), SIMDE_FLOAT32_C( 742.30), SIMDE_FLOAT32_C( 440.93), SIMDE_FLOAT32_C( -885.24)),
511
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 883.38), SIMDE_FLOAT32_C( -314.94), SIMDE_FLOAT32_C( -805.26), SIMDE_FLOAT32_C( 843.17)),
512
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -51.53), SIMDE_FLOAT32_C( 742.30), SIMDE_FLOAT32_C( -805.26), SIMDE_FLOAT32_C( -885.24)) },
513
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( 812.23), SIMDE_FLOAT32_C( -335.37), SIMDE_FLOAT32_C( 369.92), SIMDE_FLOAT32_C( -137.19)),
514
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 119.89), SIMDE_FLOAT32_C( 901.62), SIMDE_FLOAT32_C( -234.15), SIMDE_FLOAT32_C( 494.86)),
515
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 812.23), SIMDE_FLOAT32_C( -335.37), SIMDE_FLOAT32_C( -234.15), SIMDE_FLOAT32_C( -137.19)) }
516
+ };
517
+
518
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
519
+ simde__m128 r = simde_mm_blend_ps(test_vec[i].a, test_vec[i].b, 2);
520
+ simde_assert_m128_close(r, test_vec[i].r, 1);
521
+ }
522
+
523
+ return MUNIT_OK;
524
+ }
525
+
526
+ static MunitResult
527
+ test_simde_mm_blendv_pd(const MunitParameter params[], void* data) {
528
+ (void) params;
529
+ (void) data;
530
+
531
+ const struct {
532
+ simde__m128d a;
533
+ simde__m128d b;
534
+ simde__m128i mask;
535
+ simde__m128d r;
536
+ } test_vec[8] = {
537
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( 290.02), SIMDE_FLOAT64_C( -775.12)),
538
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 567.65), SIMDE_FLOAT64_C( 339.87)),
539
+ simde_x_mm_set_epu64x(UINT64_C( 0), UINT64_C( 1)),
540
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 290.02), SIMDE_FLOAT64_C( -775.12)) },
541
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( 702.24), SIMDE_FLOAT64_C( -371.99)),
542
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 899.02), SIMDE_FLOAT64_C( -664.73)),
543
+ simde_x_mm_set_epu64x(UINT64_C( 1), UINT64_C( 1)),
544
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 702.24), SIMDE_FLOAT64_C( -371.99)) },
545
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( 403.27), SIMDE_FLOAT64_C( -740.87)),
546
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 975.40), SIMDE_FLOAT64_C( -480.24)),
547
+ simde_x_mm_set_epu64x(UINT64_C( 0), UINT64_C( 0)),
548
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 403.27), SIMDE_FLOAT64_C( -740.87)) },
549
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( -417.67), SIMDE_FLOAT64_C( 711.92)),
550
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -897.21), SIMDE_FLOAT64_C( -402.59)),
551
+ simde_x_mm_set_epu64x(UINT64_C( 1), UINT64_C( 0)),
552
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -417.67), SIMDE_FLOAT64_C( 711.92)) },
553
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( 578.89), SIMDE_FLOAT64_C( -895.99)),
554
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -800.50), SIMDE_FLOAT64_C( -144.28)),
555
+ simde_x_mm_set_epu64x(UINT64_C( 1), UINT64_C( 0)),
556
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 578.89), SIMDE_FLOAT64_C( -895.99)) },
557
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( 546.78), SIMDE_FLOAT64_C( 883.57)),
558
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 415.01), SIMDE_FLOAT64_C( 773.21)),
559
+ simde_x_mm_set_epu64x(UINT64_C( 1), UINT64_C( 0)),
560
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 546.78), SIMDE_FLOAT64_C( 883.57)) },
561
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( 26.01), SIMDE_FLOAT64_C( -59.50)),
562
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -530.26), SIMDE_FLOAT64_C( 580.89)),
563
+ simde_x_mm_set_epu64x(UINT64_C( 0), UINT64_C( 1)),
564
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 26.01), SIMDE_FLOAT64_C( -59.50)) },
565
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( -344.22), SIMDE_FLOAT64_C( 626.91)),
566
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 752.48), SIMDE_FLOAT64_C( 453.28)),
567
+ simde_x_mm_set_epu64x(UINT64_C( 0), UINT64_C( 1)),
568
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -344.22), SIMDE_FLOAT64_C( 626.91)) }
569
+ };
570
+
571
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
572
+ simde__m128d r = simde_mm_blendv_pd(test_vec[i].a, test_vec[i].b, simde_mm_castsi128_pd(test_vec[i].mask));
573
+ simde_assert_m128d_close(r, test_vec[i].r, 1);
574
+ }
575
+
576
+ return MUNIT_OK;
577
+ }
578
+
579
+ static MunitResult
580
+ test_simde_mm_blendv_ps(const MunitParameter params[], void* data) {
581
+ (void) params;
582
+ (void) data;
583
+
584
+ const struct {
585
+ simde__m128 a;
586
+ simde__m128 b;
587
+ simde__m128i mask;
588
+ simde__m128 r;
589
+ } test_vec[8] = {
590
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( -155.18), SIMDE_FLOAT32_C( -698.08), SIMDE_FLOAT32_C( -535.18), SIMDE_FLOAT32_C( -144.32)),
591
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -362.54), SIMDE_FLOAT32_C( -779.39), SIMDE_FLOAT32_C( 825.01), SIMDE_FLOAT32_C( -777.93)),
592
+ simde_x_mm_set_epu32(UINT32_C( 1), UINT32_C( 0), UINT32_C( 1), UINT32_C( 0)),
593
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -155.18), SIMDE_FLOAT32_C( -698.08), SIMDE_FLOAT32_C( -535.18), SIMDE_FLOAT32_C( -144.32)) },
594
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( -284.42), SIMDE_FLOAT32_C( 692.24), SIMDE_FLOAT32_C( 101.85), SIMDE_FLOAT32_C( -819.28)),
595
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 327.10), SIMDE_FLOAT32_C( 184.40), SIMDE_FLOAT32_C( 440.54), SIMDE_FLOAT32_C( 732.07)),
596
+ simde_x_mm_set_epu32(UINT32_C( 0), UINT32_C( 0), UINT32_C( 0), UINT32_C( 1)),
597
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -284.42), SIMDE_FLOAT32_C( 692.24), SIMDE_FLOAT32_C( 101.85), SIMDE_FLOAT32_C( -819.28)) },
598
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( 264.15), SIMDE_FLOAT32_C( -480.80), SIMDE_FLOAT32_C( -304.32), SIMDE_FLOAT32_C( 862.50)),
599
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 230.55), SIMDE_FLOAT32_C( -649.28), SIMDE_FLOAT32_C( 827.43), SIMDE_FLOAT32_C( -462.60)),
600
+ simde_x_mm_set_epu32(UINT32_C( 0), UINT32_C( 0), UINT32_C( 0), UINT32_C( 0)),
601
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 264.15), SIMDE_FLOAT32_C( -480.80), SIMDE_FLOAT32_C( -304.32), SIMDE_FLOAT32_C( 862.50)) },
602
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( 235.66), SIMDE_FLOAT32_C( 478.16), SIMDE_FLOAT32_C( 480.58), SIMDE_FLOAT32_C( -53.52)),
603
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 982.67), SIMDE_FLOAT32_C( -772.62), SIMDE_FLOAT32_C( -150.10), SIMDE_FLOAT32_C( -737.86)),
604
+ simde_x_mm_set_epu32(UINT32_C( 1), UINT32_C( 1), UINT32_C( 0), UINT32_C( 0)),
605
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 235.66), SIMDE_FLOAT32_C( 478.16), SIMDE_FLOAT32_C( 480.58), SIMDE_FLOAT32_C( -53.52)) },
606
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( 121.52), SIMDE_FLOAT32_C( 798.57), SIMDE_FLOAT32_C( 563.90), SIMDE_FLOAT32_C( 845.42)),
607
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 593.41), SIMDE_FLOAT32_C( 201.43), SIMDE_FLOAT32_C( -263.22), SIMDE_FLOAT32_C( 41.51)),
608
+ simde_x_mm_set_epu32(UINT32_C( 1), UINT32_C( 1), UINT32_C( 1), UINT32_C( 0)),
609
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 121.52), SIMDE_FLOAT32_C( 798.57), SIMDE_FLOAT32_C( 563.90), SIMDE_FLOAT32_C( 845.42)) },
610
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( -938.86), SIMDE_FLOAT32_C( -190.10), SIMDE_FLOAT32_C( -729.05), SIMDE_FLOAT32_C( 956.23)),
611
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 574.98), SIMDE_FLOAT32_C( -82.92), SIMDE_FLOAT32_C( 247.07), SIMDE_FLOAT32_C( -802.43)),
612
+ simde_x_mm_set_epu32(UINT32_C( 0), UINT32_C( 1), UINT32_C( 0), UINT32_C( 0)),
613
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -938.86), SIMDE_FLOAT32_C( -190.10), SIMDE_FLOAT32_C( -729.05), SIMDE_FLOAT32_C( 956.23)) },
614
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( -592.31), SIMDE_FLOAT32_C( -123.91), SIMDE_FLOAT32_C( 105.67), SIMDE_FLOAT32_C( 596.25)),
615
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -99.40), SIMDE_FLOAT32_C( 86.32), SIMDE_FLOAT32_C( -827.82), SIMDE_FLOAT32_C( 443.02)),
616
+ simde_x_mm_set_epu32(UINT32_C( 1), UINT32_C( 1), UINT32_C( 1), UINT32_C( 0)),
617
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -592.31), SIMDE_FLOAT32_C( -123.91), SIMDE_FLOAT32_C( 105.67), SIMDE_FLOAT32_C( 596.25)) },
618
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( 59.00), SIMDE_FLOAT32_C( -93.37), SIMDE_FLOAT32_C( -934.05), SIMDE_FLOAT32_C( 930.21)),
619
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -592.32), SIMDE_FLOAT32_C( 420.31), SIMDE_FLOAT32_C( -235.36), SIMDE_FLOAT32_C( -335.98)),
620
+ simde_x_mm_set_epu32(UINT32_C( 1), UINT32_C( 1), UINT32_C( 1), UINT32_C( 1)),
621
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 59.00), SIMDE_FLOAT32_C( -93.37), SIMDE_FLOAT32_C( -934.05), SIMDE_FLOAT32_C( 930.21)) }
622
+ };
623
+
624
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
625
+ simde__m128 r = simde_mm_blendv_ps(test_vec[i].a, test_vec[i].b, simde_mm_castsi128_ps(test_vec[i].mask));
626
+ simde_assert_m128_close(r, test_vec[i].r, 1);
627
+ }
628
+
629
+ return MUNIT_OK;
630
+ }
631
+
632
+ static MunitResult
633
+ test_simde_mm_ceil_pd(const MunitParameter params[], void* data) {
634
+ (void) params;
635
+ (void) data;
636
+
637
+ const struct {
638
+ simde__m128d a;
639
+ simde__m128d r;
640
+ } test_vec[8] = {
641
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( 241.39), SIMDE_FLOAT64_C( 553.26)),
642
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 242.00), SIMDE_FLOAT64_C( 554.00)) },
643
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( 307.83), SIMDE_FLOAT64_C( 377.37)),
644
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 308.00), SIMDE_FLOAT64_C( 378.00)) },
645
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( 558.16), SIMDE_FLOAT64_C( -404.55)),
646
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 559.00), SIMDE_FLOAT64_C( -404.00)) },
647
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( 219.15), SIMDE_FLOAT64_C( 818.17)),
648
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 220.00), SIMDE_FLOAT64_C( 819.00)) },
649
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( 25.83), SIMDE_FLOAT64_C( -230.94)),
650
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 26.00), SIMDE_FLOAT64_C( -230.00)) },
651
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( 173.52), SIMDE_FLOAT64_C( 64.50)),
652
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 174.00), SIMDE_FLOAT64_C( 65.00)) },
653
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( -453.11), SIMDE_FLOAT64_C( -845.77)),
654
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -453.00), SIMDE_FLOAT64_C( -845.00)) },
655
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( -50.98), SIMDE_FLOAT64_C( -179.55)),
656
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -50.00), SIMDE_FLOAT64_C( -179.00)) }
657
+ };
658
+
659
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
660
+ simde__m128d r = simde_mm_ceil_pd(test_vec[i].a);
661
+ simde_assert_m128d_close(r, test_vec[i].r, 1);
662
+ }
663
+
664
+ return MUNIT_OK;
665
+ }
666
+
667
+ static MunitResult
668
+ test_simde_mm_ceil_ps(const MunitParameter params[], void* data) {
669
+ (void) params;
670
+ (void) data;
671
+
672
+ const struct {
673
+ simde__m128 a;
674
+ simde__m128 r;
675
+ } test_vec[8] = {
676
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( -624.94), SIMDE_FLOAT32_C( -243.25), SIMDE_FLOAT32_C( -61.63), SIMDE_FLOAT32_C( -948.73)),
677
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -624.00), SIMDE_FLOAT32_C( -243.00), SIMDE_FLOAT32_C( -61.00), SIMDE_FLOAT32_C( -948.00)) },
678
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( -906.63), SIMDE_FLOAT32_C( -236.42), SIMDE_FLOAT32_C( 571.51), SIMDE_FLOAT32_C( 624.62)),
679
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -906.00), SIMDE_FLOAT32_C( -236.00), SIMDE_FLOAT32_C( 572.00), SIMDE_FLOAT32_C( 625.00)) },
680
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( 503.85), SIMDE_FLOAT32_C( -317.51), SIMDE_FLOAT32_C( -529.30), SIMDE_FLOAT32_C( -871.89)),
681
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 504.00), SIMDE_FLOAT32_C( -317.00), SIMDE_FLOAT32_C( -529.00), SIMDE_FLOAT32_C( -871.00)) },
682
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( 571.49), SIMDE_FLOAT32_C( -129.77), SIMDE_FLOAT32_C( -739.71), SIMDE_FLOAT32_C( 413.18)),
683
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 572.00), SIMDE_FLOAT32_C( -129.00), SIMDE_FLOAT32_C( -739.00), SIMDE_FLOAT32_C( 414.00)) },
684
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( 438.35), SIMDE_FLOAT32_C( 517.03), SIMDE_FLOAT32_C( 278.46), SIMDE_FLOAT32_C( -656.36)),
685
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 439.00), SIMDE_FLOAT32_C( 518.00), SIMDE_FLOAT32_C( 279.00), SIMDE_FLOAT32_C( -656.00)) },
686
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( -159.40), SIMDE_FLOAT32_C( -802.82), SIMDE_FLOAT32_C( 663.71), SIMDE_FLOAT32_C( 186.86)),
687
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -159.00), SIMDE_FLOAT32_C( -802.00), SIMDE_FLOAT32_C( 664.00), SIMDE_FLOAT32_C( 187.00)) },
688
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( -163.74), SIMDE_FLOAT32_C( 485.54), SIMDE_FLOAT32_C( 138.17), SIMDE_FLOAT32_C( 794.67)),
689
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -163.00), SIMDE_FLOAT32_C( 486.00), SIMDE_FLOAT32_C( 139.00), SIMDE_FLOAT32_C( 795.00)) },
690
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( -469.36), SIMDE_FLOAT32_C( -418.24), SIMDE_FLOAT32_C( -419.25), SIMDE_FLOAT32_C( 390.46)),
691
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -469.00), SIMDE_FLOAT32_C( -418.00), SIMDE_FLOAT32_C( -419.00), SIMDE_FLOAT32_C( 391.00)) }
692
+ };
693
+
694
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
695
+ simde__m128 r = simde_mm_ceil_ps(test_vec[i].a);
696
+ simde_assert_m128_close(r, test_vec[i].r, 1);
697
+ }
698
+
699
+ return MUNIT_OK;
700
+ }
701
+
702
+ static MunitResult
703
+ test_simde_mm_ceil_sd(const MunitParameter params[], void* data) {
704
+ (void) params;
705
+ (void) data;
706
+
707
+ const struct {
708
+ simde__m128d a;
709
+ simde__m128d b;
710
+ simde__m128d r;
711
+ } test_vec[8] = {
712
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( -610.18), SIMDE_FLOAT64_C( 172.21)),
713
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 234.40), SIMDE_FLOAT64_C( -339.84)),
714
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -610.18), SIMDE_FLOAT64_C( -339.00)) },
715
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( -467.52), SIMDE_FLOAT64_C( 904.78)),
716
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -527.39), SIMDE_FLOAT64_C( -115.63)),
717
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -467.52), SIMDE_FLOAT64_C( -115.00)) },
718
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( 191.24), SIMDE_FLOAT64_C( 823.78)),
719
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 835.27), SIMDE_FLOAT64_C( 744.81)),
720
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 191.24), SIMDE_FLOAT64_C( 745.00)) },
721
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( -49.25), SIMDE_FLOAT64_C( -695.95)),
722
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 800.62), SIMDE_FLOAT64_C( 853.84)),
723
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -49.25), SIMDE_FLOAT64_C( 854.00)) },
724
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( 668.83), SIMDE_FLOAT64_C( -182.73)),
725
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -875.46), SIMDE_FLOAT64_C( 752.10)),
726
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 668.83), SIMDE_FLOAT64_C( 753.00)) },
727
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( 242.57), SIMDE_FLOAT64_C( -906.00)),
728
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -816.22), SIMDE_FLOAT64_C( 339.03)),
729
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 242.57), SIMDE_FLOAT64_C( 340.00)) },
730
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( -745.15), SIMDE_FLOAT64_C( -744.04)),
731
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -802.81), SIMDE_FLOAT64_C( 890.79)),
732
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -745.15), SIMDE_FLOAT64_C( 891.00)) },
733
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( 238.48), SIMDE_FLOAT64_C( -167.28)),
734
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -484.51), SIMDE_FLOAT64_C( 175.95)),
735
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 238.48), SIMDE_FLOAT64_C( 176.00)) }
736
+ };
737
+
738
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
739
+ simde__m128d r = simde_mm_ceil_sd(test_vec[i].a, test_vec[i].b);
740
+ simde_assert_m128d_close(r, test_vec[i].r, 1);
741
+ }
742
+
743
+ return MUNIT_OK;
744
+ }
745
+
746
+ static MunitResult
747
+ test_simde_mm_ceil_ss(const MunitParameter params[], void* data) {
748
+ (void) params;
749
+ (void) data;
750
+
751
+ const struct {
752
+ simde__m128 a;
753
+ simde__m128 b;
754
+ simde__m128 r;
755
+ } test_vec[8] = {
756
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( -124.45), SIMDE_FLOAT32_C( 911.12), SIMDE_FLOAT32_C( 900.52), SIMDE_FLOAT32_C( -468.92)),
757
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -31.56), SIMDE_FLOAT32_C( -29.76), SIMDE_FLOAT32_C( -939.35), SIMDE_FLOAT32_C( 459.71)),
758
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -124.45), SIMDE_FLOAT32_C( 911.12), SIMDE_FLOAT32_C( 900.52), SIMDE_FLOAT32_C( 460.00)) },
759
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( 514.22), SIMDE_FLOAT32_C( -972.44), SIMDE_FLOAT32_C( -908.76), SIMDE_FLOAT32_C( -817.52)),
760
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -711.78), SIMDE_FLOAT32_C( 489.62), SIMDE_FLOAT32_C( -109.52), SIMDE_FLOAT32_C( -2.01)),
761
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 514.22), SIMDE_FLOAT32_C( -972.44), SIMDE_FLOAT32_C( -908.76), SIMDE_FLOAT32_C( -2.00)) },
762
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( -269.88), SIMDE_FLOAT32_C( -875.85), SIMDE_FLOAT32_C( -139.27), SIMDE_FLOAT32_C( -841.56)),
763
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 501.05), SIMDE_FLOAT32_C( 280.43), SIMDE_FLOAT32_C( -173.91), SIMDE_FLOAT32_C( 644.42)),
764
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -269.88), SIMDE_FLOAT32_C( -875.85), SIMDE_FLOAT32_C( -139.27), SIMDE_FLOAT32_C( 645.00)) },
765
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( -643.26), SIMDE_FLOAT32_C( 616.48), SIMDE_FLOAT32_C( 835.28), SIMDE_FLOAT32_C( 527.52)),
766
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 636.48), SIMDE_FLOAT32_C( -46.56), SIMDE_FLOAT32_C( -891.84), SIMDE_FLOAT32_C( 948.43)),
767
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -643.26), SIMDE_FLOAT32_C( 616.48), SIMDE_FLOAT32_C( 835.28), SIMDE_FLOAT32_C( 949.00)) },
768
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( 3.13), SIMDE_FLOAT32_C( -654.36), SIMDE_FLOAT32_C( -994.98), SIMDE_FLOAT32_C( -148.95)),
769
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 806.15), SIMDE_FLOAT32_C( 529.29), SIMDE_FLOAT32_C( -804.42), SIMDE_FLOAT32_C( 408.19)),
770
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 3.13), SIMDE_FLOAT32_C( -654.36), SIMDE_FLOAT32_C( -994.98), SIMDE_FLOAT32_C( 409.00)) },
771
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( -138.18), SIMDE_FLOAT32_C( -446.48), SIMDE_FLOAT32_C( -413.95), SIMDE_FLOAT32_C( -252.30)),
772
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 957.68), SIMDE_FLOAT32_C( 373.35), SIMDE_FLOAT32_C( 271.98), SIMDE_FLOAT32_C( -653.93)),
773
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -138.18), SIMDE_FLOAT32_C( -446.48), SIMDE_FLOAT32_C( -413.95), SIMDE_FLOAT32_C( -653.00)) },
774
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( 124.40), SIMDE_FLOAT32_C( -866.66), SIMDE_FLOAT32_C( 288.57), SIMDE_FLOAT32_C( -708.46)),
775
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 685.14), SIMDE_FLOAT32_C( 776.00), SIMDE_FLOAT32_C( 453.56), SIMDE_FLOAT32_C( -112.81)),
776
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 124.40), SIMDE_FLOAT32_C( -866.66), SIMDE_FLOAT32_C( 288.57), SIMDE_FLOAT32_C( -112.00)) },
777
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( -232.75), SIMDE_FLOAT32_C( -137.55), SIMDE_FLOAT32_C( 50.26), SIMDE_FLOAT32_C( -189.71)),
778
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -288.67), SIMDE_FLOAT32_C( -112.95), SIMDE_FLOAT32_C( 792.47), SIMDE_FLOAT32_C( -200.27)),
779
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -232.75), SIMDE_FLOAT32_C( -137.55), SIMDE_FLOAT32_C( 50.26), SIMDE_FLOAT32_C( -200.00)) }
780
+ };
781
+
782
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
783
+ simde__m128 r = simde_mm_ceil_ss(test_vec[i].a, test_vec[i].b);
784
+ simde_assert_m128_close(r, test_vec[i].r, 1);
785
+ }
786
+
787
+ return MUNIT_OK;
788
+ }
789
+
790
+ static MunitResult
791
+ test_simde_mm_cmpeq_epi64(const MunitParameter params[], void* data) {
792
+ (void) params;
793
+ (void) data;
794
+
795
+ const struct {
796
+ simde__m128i a;
797
+ simde__m128i b;
798
+ simde__m128i r;
799
+ } test_vec[8] = {
800
+ { simde_x_mm_set_epu64x(UINT64_C(13269709557652540150), UINT64_C( 5229646237911669536)),
801
+ simde_x_mm_set_epu64x(UINT64_C(13269709557652540150), UINT64_C( 5229646237911669536)),
802
+ simde_x_mm_set_epu64x(UINT64_C(18446744073709551615), UINT64_C(18446744073709551615)) },
803
+ { simde_x_mm_set_epu64x(UINT64_C(14039735832537203737), UINT64_C(18065414586371585158)),
804
+ simde_x_mm_set_epu64x(UINT64_C( 7896134184979754273), UINT64_C(18065414586371585158)),
805
+ simde_x_mm_set_epu64x(UINT64_C( 0), UINT64_C(18446744073709551615)) },
806
+ { simde_x_mm_set_epu64x(UINT64_C( 8088434365645432882), UINT64_C( 2110149590601129498)),
807
+ simde_x_mm_set_epu64x(UINT64_C(12894308138858138936), UINT64_C( 7645444803376971084)),
808
+ simde_x_mm_set_epu64x(UINT64_C( 0), UINT64_C( 0)) },
809
+ { simde_x_mm_set_epu64x(UINT64_C( 6900998750312547586), UINT64_C(11115665826166806875)),
810
+ simde_x_mm_set_epu64x(UINT64_C( 253910074237567150), UINT64_C(11115665826166806875)),
811
+ simde_x_mm_set_epu64x(UINT64_C( 0), UINT64_C(18446744073709551615)) },
812
+ { simde_x_mm_set_epu64x(UINT64_C(15513947591537518317), UINT64_C( 6346622963553980783)),
813
+ simde_x_mm_set_epu64x(UINT64_C( 1894817647097843127), UINT64_C( 9019647896255166506)),
814
+ simde_x_mm_set_epu64x(UINT64_C( 0), UINT64_C( 0)) },
815
+ { simde_x_mm_set_epu64x(UINT64_C( 189073800069710221), UINT64_C( 5830713182261233236)),
816
+ simde_x_mm_set_epu64x(UINT64_C( 189073800069710221), UINT64_C( 8422068610497871432)),
817
+ simde_x_mm_set_epu64x(UINT64_C(18446744073709551615), UINT64_C( 0)) },
818
+ { simde_x_mm_set_epu64x(UINT64_C(17115829668729212386), UINT64_C(11158226665026575121)),
819
+ simde_x_mm_set_epu64x(UINT64_C(17115829668729212386), UINT64_C(11158226665026575121)),
820
+ simde_x_mm_set_epu64x(UINT64_C(18446744073709551615), UINT64_C(18446744073709551615)) },
821
+ { simde_x_mm_set_epu64x(UINT64_C( 9543887299291080046), UINT64_C(14913997234931691871)),
822
+ simde_x_mm_set_epu64x(UINT64_C( 9543887299291080046), UINT64_C(12485648986158046833)),
823
+ simde_x_mm_set_epu64x(UINT64_C(18446744073709551615), UINT64_C( 0)) }
824
+ };
825
+
826
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
827
+ simde__m128i r = simde_mm_cmpeq_epi64(test_vec[i].a, test_vec[i].b);
828
+ simde_assert_m128i_i64(r, ==, test_vec[i].r);
829
+ }
830
+
831
+ return MUNIT_OK;
832
+ }
833
+
834
+ static MunitResult
835
+ test_simde_mm_cvtepi16_epi32(const MunitParameter params[], void* data) {
836
+ (void) params;
837
+ (void) data;
838
+
839
+ const struct {
840
+ simde__m128i a;
841
+ simde__m128i r;
842
+ } test_vec[8] = {
843
+ { simde_mm_set_epi16(-15854, 0, -12807, 5951, 29318, -11078, -2252, 10096),
844
+ simde_mm_set_epi32( 29318, -11078, -2252, 10096) },
845
+ { simde_mm_set_epi16( 24388, 32521, -30980, -12289, 27092, 24521, 12106, -27709),
846
+ simde_mm_set_epi32( 27092, 24521, 12106, -27709) },
847
+ { simde_mm_set_epi16(-11871, 15750, -7986, 4594, 22346, 1400, -21386, 7662),
848
+ simde_mm_set_epi32( 22346, 1400, -21386, 7662) },
849
+ { simde_mm_set_epi16(-15687, -15477, -26077, -28781, -29737, -30311, 16950, 14614),
850
+ simde_mm_set_epi32(-29737, -30311, 16950, 14614) },
851
+ { simde_mm_set_epi16(-30646, 13049, -24166, 637, 7297, -627, 6143, -26200),
852
+ simde_mm_set_epi32( 7297, -627, 6143, -26200) },
853
+ { simde_mm_set_epi16( 30171, 1922, -122, -17991, 30806, 12552, -28515, -9185),
854
+ simde_mm_set_epi32( 30806, 12552, -28515, -9185) },
855
+ { simde_mm_set_epi16( 29385, 21267, -10152, 25553, 20674, 6661, -29222, -19848),
856
+ simde_mm_set_epi32( 20674, 6661, -29222, -19848) },
857
+ { simde_mm_set_epi16( 24527, 18305, 30094, -5437, -26652, 20361, 10408, -4948),
858
+ simde_mm_set_epi32(-26652, 20361, 10408, -4948) }
859
+ };
860
+
861
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
862
+ simde__m128i r = simde_mm_cvtepi16_epi32(test_vec[i].a);
863
+ simde_assert_m128i_i32(r, ==, test_vec[i].r);
864
+ }
865
+
866
+ return MUNIT_OK;
867
+ }
868
+
869
+ static MunitResult
870
+ test_simde_mm_cvtepi16_epi64(const MunitParameter params[], void* data) {
871
+ (void) params;
872
+ (void) data;
873
+
874
+ const struct {
875
+ simde__m128i a;
876
+ simde__m128i r;
877
+ } test_vec[8] = {
878
+ { simde_mm_set_epi16(INT16_C(-25288), INT16_C( 737), INT16_C(-31366), INT16_C(-21360),
879
+ INT16_C( 1927), INT16_C( 32195), INT16_C( -4472), INT16_C( -2085)),
880
+ simde_mm_set_epi64x(INT64_C( -4472), INT64_C( -2085)) },
881
+ { simde_mm_set_epi16(INT16_C(-30817), INT16_C( 5734), INT16_C( -6209), INT16_C(-18846),
882
+ INT16_C( 14970), INT16_C( 30345), INT16_C( -8853), INT16_C(-31121)),
883
+ simde_mm_set_epi64x(INT64_C( -8853), INT64_C( -31121)) },
884
+ { simde_mm_set_epi16(INT16_C( -970), INT16_C(-11861), INT16_C(-11667), INT16_C( 14936),
885
+ INT16_C( 21732), INT16_C(-17006), INT16_C( -2668), INT16_C( 30170)),
886
+ simde_mm_set_epi64x(INT64_C( -2668), INT64_C( 30170)) },
887
+ { simde_mm_set_epi16(INT16_C( 2721), INT16_C(-13375), INT16_C( -1976), INT16_C( 30833),
888
+ INT16_C(-18247), INT16_C( 16242), INT16_C( 4569), INT16_C( 10132)),
889
+ simde_mm_set_epi64x(INT64_C( 4569), INT64_C( 10132)) },
890
+ { simde_mm_set_epi16(INT16_C(-22879), INT16_C(-24842), INT16_C(-29886), INT16_C(-13312),
891
+ INT16_C( 24862), INT16_C(-30679), INT16_C( 6258), INT16_C(-19461)),
892
+ simde_mm_set_epi64x(INT64_C( 6258), INT64_C( -19461)) },
893
+ { simde_mm_set_epi16(INT16_C( 15048), INT16_C( 5334), INT16_C( 21517), INT16_C( 10106),
894
+ INT16_C(-25890), INT16_C( -8759), INT16_C( 21361), INT16_C( 17584)),
895
+ simde_mm_set_epi64x(INT64_C( 21361), INT64_C( 17584)) },
896
+ { simde_mm_set_epi16(INT16_C( 18353), INT16_C( 9848), INT16_C( 20583), INT16_C(-28706),
897
+ INT16_C( 28353), INT16_C( 979), INT16_C(-15312), INT16_C( -4296)),
898
+ simde_mm_set_epi64x(INT64_C( -15312), INT64_C( -4296)) },
899
+ { simde_mm_set_epi16(INT16_C( 21066), INT16_C( 26185), INT16_C( 29520), INT16_C( -9848),
900
+ INT16_C(-20587), INT16_C( -254), INT16_C(-19969), INT16_C( -2307)),
901
+ simde_mm_set_epi64x(INT64_C( -19969), INT64_C( -2307)) }
902
+ };
903
+
904
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
905
+ simde__m128i r = simde_mm_cvtepi16_epi64(test_vec[i].a);
906
+ simde_assert_m128i_i64(r, ==, test_vec[i].r);
907
+ }
908
+
909
+ return MUNIT_OK;
910
+ }
911
+
912
+ static MunitResult
913
+ test_simde_mm_cvtepi32_epi64(const MunitParameter params[], void* data) {
914
+ (void) params;
915
+ (void) data;
916
+
917
+ const struct {
918
+ simde__m128i a;
919
+ simde__m128i r;
920
+ } test_vec[8] = {
921
+ { simde_mm_set_epi32(INT32_C(-1750586859), INT32_C( 1743108010), INT32_C( -56370522), INT32_C( 214732685)),
922
+ simde_mm_set_epi64x(INT64_C( -56370522), INT64_C( 214732685)) },
923
+ { simde_mm_set_epi32(INT32_C( 1019578476), INT32_C( 2006962808), INT32_C( 1356966666), INT32_C( 443191452)),
924
+ simde_mm_set_epi64x(INT64_C( 1356966666), INT64_C( 443191452)) },
925
+ { simde_mm_set_epi32(INT32_C( 811952499), INT32_C( 1388789927), INT32_C( 1562076587), INT32_C( 1374574189)),
926
+ simde_mm_set_epi64x(INT64_C( 1562076587), INT64_C( 1374574189)) },
927
+ { simde_mm_set_epi32(INT32_C(-1130424581), INT32_C( -754503048), INT32_C( -175987570), INT32_C( 252857640)),
928
+ simde_mm_set_epi64x(INT64_C( -175987570), INT64_C( 252857640)) },
929
+ { simde_mm_set_epi32(INT32_C( -662885345), INT32_C( -469769762), INT32_C( 1876427062), INT32_C(-1227144336)),
930
+ simde_mm_set_epi64x(INT64_C( 1876427062), INT64_C( -1227144336)) },
931
+ { simde_mm_set_epi32(INT32_C(-1459145913), INT32_C( -624950308), INT32_C(-1740403350), INT32_C(-2023012954)),
932
+ simde_mm_set_epi64x(INT64_C( -1740403350), INT64_C( -2023012954)) },
933
+ { simde_mm_set_epi32(INT32_C( 807996884), INT32_C( -721517626), INT32_C(-1345085365), INT32_C( 3500292)),
934
+ simde_mm_set_epi64x(INT64_C( -1345085365), INT64_C( 3500292)) },
935
+ { simde_mm_set_epi32(INT32_C( 1631614297), INT32_C(-2082769218), INT32_C( 285385405), INT32_C(-1874296921)),
936
+ simde_mm_set_epi64x(INT64_C( 285385405), INT64_C( -1874296921)) }
937
+ };
938
+
939
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
940
+ simde__m128i r = simde_mm_cvtepi32_epi64(test_vec[i].a);
941
+ simde_assert_m128i_i64(r, ==, test_vec[i].r);
942
+ }
943
+
944
+ return MUNIT_OK;
945
+ }
946
+
947
+ static MunitResult
948
+ test_simde_mm_cvtepi8_epi16(const MunitParameter params[], void* data) {
949
+ (void) params;
950
+ (void) data;
951
+
952
+ const struct {
953
+ simde__m128i a;
954
+ simde__m128i r;
955
+ } test_vec[8] = {
956
+ { simde_mm_set_epi8(INT8_C( -34), INT8_C( 20), INT8_C( 53), INT8_C( 81),
957
+ INT8_C( 37), INT8_C( -97), INT8_C( -59), INT8_C( 127),
958
+ INT8_C( -43), INT8_C( -82), INT8_C( 84), INT8_C(-127),
959
+ INT8_C( 11), INT8_C( 120), INT8_C( -98), INT8_C( -80)),
960
+ simde_mm_set_epi16(INT16_C( -43), INT16_C( -82), INT16_C( 84), INT16_C( -127),
961
+ INT16_C( 11), INT16_C( 120), INT16_C( -98), INT16_C( -80)) },
962
+ { simde_mm_set_epi8(INT8_C( 66), INT8_C( 57), INT8_C( 40), INT8_C(-113),
963
+ INT8_C( 71), INT8_C( -5), INT8_C( 70), INT8_C( 22),
964
+ INT8_C( -70), INT8_C( 14), INT8_C( -76), INT8_C(-101),
965
+ INT8_C( 13), INT8_C( 64), INT8_C( -86), INT8_C( 34)),
966
+ simde_mm_set_epi16(INT16_C( -70), INT16_C( 14), INT16_C( -76), INT16_C( -101),
967
+ INT16_C( 13), INT16_C( 64), INT16_C( -86), INT16_C( 34)) },
968
+ { simde_mm_set_epi8(INT8_C( 87), INT8_C( 115), INT8_C( 108), INT8_C( -88),
969
+ INT8_C( 113), INT8_C( 91), INT8_C( -95), INT8_C( -69),
970
+ INT8_C( 32), INT8_C( 98), INT8_C( -16), INT8_C( -55),
971
+ INT8_C( 11), INT8_C( -23), INT8_C( 72), INT8_C( 121)),
972
+ simde_mm_set_epi16(INT16_C( 32), INT16_C( 98), INT16_C( -16), INT16_C( -55),
973
+ INT16_C( 11), INT16_C( -23), INT16_C( 72), INT16_C( 121)) },
974
+ { simde_mm_set_epi8(INT8_C( 88), INT8_C( 68), INT8_C( -76), INT8_C( 35),
975
+ INT8_C(-119), INT8_C( -67), INT8_C( 52), INT8_C( 58),
976
+ INT8_C( 7), INT8_C( 70), INT8_C( -26), INT8_C( 55),
977
+ INT8_C( 126), INT8_C( 69), INT8_C( 105), INT8_C( 40)),
978
+ simde_mm_set_epi16(INT16_C( 7), INT16_C( 70), INT16_C( -26), INT16_C( 55),
979
+ INT16_C( 126), INT16_C( 69), INT16_C( 105), INT16_C( 40)) },
980
+ { simde_mm_set_epi8(INT8_C( 110), INT8_C( -4), INT8_C( 41), INT8_C( -98),
981
+ INT8_C( -91), INT8_C( 117), INT8_C( -33), INT8_C(-110),
982
+ INT8_C( 126), INT8_C( -63), INT8_C( -65), INT8_C( -86),
983
+ INT8_C(-108), INT8_C( -64), INT8_C( -4), INT8_C( 25)),
984
+ simde_mm_set_epi16(INT16_C( 126), INT16_C( -63), INT16_C( -65), INT16_C( -86),
985
+ INT16_C( -108), INT16_C( -64), INT16_C( -4), INT16_C( 25)) },
986
+ { simde_mm_set_epi8(INT8_C( 105), INT8_C( -31), INT8_C( -57), INT8_C( -74),
987
+ INT8_C( 119), INT8_C( 56), INT8_C( -36), INT8_C(-128),
988
+ INT8_C( 102), INT8_C( 57), INT8_C( -27), INT8_C( 19),
989
+ INT8_C( -50), INT8_C( -15), INT8_C( 36), INT8_C(-124)),
990
+ simde_mm_set_epi16(INT16_C( 102), INT16_C( 57), INT16_C( -27), INT16_C( 19),
991
+ INT16_C( -50), INT16_C( -15), INT16_C( 36), INT16_C( -124)) },
992
+ { simde_mm_set_epi8(INT8_C( 44), INT8_C( -84), INT8_C( -61), INT8_C( 59),
993
+ INT8_C( 18), INT8_C(-118), INT8_C( 6), INT8_C( 56),
994
+ INT8_C( -11), INT8_C( 78), INT8_C( -64), INT8_C( -28),
995
+ INT8_C( -85), INT8_C( -77), INT8_C( -25), INT8_C(-100)),
996
+ simde_mm_set_epi16(INT16_C( -11), INT16_C( 78), INT16_C( -64), INT16_C( -28),
997
+ INT16_C( -85), INT16_C( -77), INT16_C( -25), INT16_C( -100)) },
998
+ { simde_mm_set_epi8(INT8_C( -13), INT8_C( -24), INT8_C( 94), INT8_C( 58),
999
+ INT8_C( 1), INT8_C( -14), INT8_C( -90), INT8_C(-123),
1000
+ INT8_C(-101), INT8_C( 71), INT8_C( -51), INT8_C( 87),
1001
+ INT8_C( -34), INT8_C( 77), INT8_C( 124), INT8_C( -63)),
1002
+ simde_mm_set_epi16(INT16_C( -101), INT16_C( 71), INT16_C( -51), INT16_C( 87),
1003
+ INT16_C( -34), INT16_C( 77), INT16_C( 124), INT16_C( -63)) }
1004
+ };
1005
+
1006
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
1007
+ simde__m128i r = simde_mm_cvtepi8_epi16(test_vec[i].a);
1008
+ simde_assert_m128i_i16(r, ==, test_vec[i].r);
1009
+ }
1010
+
1011
+ return MUNIT_OK;
1012
+ }
1013
+
1014
+ static MunitResult
1015
+ test_simde_mm_cvtepi8_epi32(const MunitParameter params[], void* data) {
1016
+ (void) params;
1017
+ (void) data;
1018
+
1019
+ const struct {
1020
+ simde__m128i a;
1021
+ simde__m128i r;
1022
+ } test_vec[8] = {
1023
+ { simde_mm_set_epi8(INT8_C( -34), INT8_C( 20), INT8_C( 53), INT8_C( 81),
1024
+ INT8_C( 37), INT8_C( -97), INT8_C( -59), INT8_C( 127),
1025
+ INT8_C( -43), INT8_C( -82), INT8_C( 84), INT8_C(-127),
1026
+ INT8_C( 11), INT8_C( 120), INT8_C( -98), INT8_C( -80)),
1027
+ simde_mm_set_epi32(INT32_C( 11), INT32_C( 120), INT32_C( -98), INT32_C( -80)) },
1028
+ { simde_mm_set_epi8(INT8_C( 66), INT8_C( 57), INT8_C( 40), INT8_C(-113),
1029
+ INT8_C( 71), INT8_C( -5), INT8_C( 70), INT8_C( 22),
1030
+ INT8_C( -70), INT8_C( 14), INT8_C( -76), INT8_C(-101),
1031
+ INT8_C( 13), INT8_C( 64), INT8_C( -86), INT8_C( 34)),
1032
+ simde_mm_set_epi32(INT32_C( 13), INT32_C( 64), INT32_C( -86), INT32_C( 34)) },
1033
+ { simde_mm_set_epi8(INT8_C( 87), INT8_C( 115), INT8_C( 108), INT8_C( -88),
1034
+ INT8_C( 113), INT8_C( 91), INT8_C( -95), INT8_C( -69),
1035
+ INT8_C( 32), INT8_C( 98), INT8_C( -16), INT8_C( -55),
1036
+ INT8_C( 11), INT8_C( -23), INT8_C( 72), INT8_C( 121)),
1037
+ simde_mm_set_epi32(INT32_C( 11), INT32_C( -23), INT32_C( 72), INT32_C( 121)) },
1038
+ { simde_mm_set_epi8(INT8_C( 88), INT8_C( 68), INT8_C( -76), INT8_C( 35),
1039
+ INT8_C(-119), INT8_C( -67), INT8_C( 52), INT8_C( 58),
1040
+ INT8_C( 7), INT8_C( 70), INT8_C( -26), INT8_C( 55),
1041
+ INT8_C( 126), INT8_C( 69), INT8_C( 105), INT8_C( 40)),
1042
+ simde_mm_set_epi32(INT32_C( 126), INT32_C( 69), INT32_C( 105), INT32_C( 40)) },
1043
+ { simde_mm_set_epi8(INT8_C( 110), INT8_C( -4), INT8_C( 41), INT8_C( -98),
1044
+ INT8_C( -91), INT8_C( 117), INT8_C( -33), INT8_C(-110),
1045
+ INT8_C( 126), INT8_C( -63), INT8_C( -65), INT8_C( -86),
1046
+ INT8_C(-108), INT8_C( -64), INT8_C( -4), INT8_C( 25)),
1047
+ simde_mm_set_epi32(INT32_C( -108), INT32_C( -64), INT32_C( -4), INT32_C( 25)) },
1048
+ { simde_mm_set_epi8(INT8_C( 105), INT8_C( -31), INT8_C( -57), INT8_C( -74),
1049
+ INT8_C( 119), INT8_C( 56), INT8_C( -36), INT8_C(-128),
1050
+ INT8_C( 102), INT8_C( 57), INT8_C( -27), INT8_C( 19),
1051
+ INT8_C( -50), INT8_C( -15), INT8_C( 36), INT8_C(-124)),
1052
+ simde_mm_set_epi32(INT32_C( -50), INT32_C( -15), INT32_C( 36), INT32_C( -124)) },
1053
+ { simde_mm_set_epi8(INT8_C( 44), INT8_C( -84), INT8_C( -61), INT8_C( 59),
1054
+ INT8_C( 18), INT8_C(-118), INT8_C( 6), INT8_C( 56),
1055
+ INT8_C( -11), INT8_C( 78), INT8_C( -64), INT8_C( -28),
1056
+ INT8_C( -85), INT8_C( -77), INT8_C( -25), INT8_C(-100)),
1057
+ simde_mm_set_epi32(INT32_C( -85), INT32_C( -77), INT32_C( -25), INT32_C( -100)) },
1058
+ { simde_mm_set_epi8(INT8_C( -13), INT8_C( -24), INT8_C( 94), INT8_C( 58),
1059
+ INT8_C( 1), INT8_C( -14), INT8_C( -90), INT8_C(-123),
1060
+ INT8_C(-101), INT8_C( 71), INT8_C( -51), INT8_C( 87),
1061
+ INT8_C( -34), INT8_C( 77), INT8_C( 124), INT8_C( -63)),
1062
+ simde_mm_set_epi32(INT32_C( -34), INT32_C( 77), INT32_C( 124), INT32_C( -63)) }
1063
+ };
1064
+
1065
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
1066
+ simde__m128i r = simde_mm_cvtepi8_epi32(test_vec[i].a);
1067
+ simde_assert_m128i_i32(r, ==, test_vec[i].r);
1068
+ }
1069
+
1070
+ return MUNIT_OK;
1071
+ }
1072
+
1073
+ static MunitResult
1074
+ test_simde_mm_cvtepi8_epi64(const MunitParameter params[], void* data) {
1075
+ (void) params;
1076
+ (void) data;
1077
+
1078
+ const struct {
1079
+ simde__m128i a;
1080
+ simde__m128i r;
1081
+ } test_vec[8] = {
1082
+ { simde_mm_set_epi8(INT8_C( -34), INT8_C( 20), INT8_C( 53), INT8_C( 81),
1083
+ INT8_C( 37), INT8_C( -97), INT8_C( -59), INT8_C( 127),
1084
+ INT8_C( -43), INT8_C( -82), INT8_C( 84), INT8_C(-127),
1085
+ INT8_C( 11), INT8_C( 120), INT8_C( -98), INT8_C( -80)),
1086
+ simde_mm_set_epi64x(INT64_C( -98), INT64_C( -80)) },
1087
+ { simde_mm_set_epi8(INT8_C( 66), INT8_C( 57), INT8_C( 40), INT8_C(-113),
1088
+ INT8_C( 71), INT8_C( -5), INT8_C( 70), INT8_C( 22),
1089
+ INT8_C( -70), INT8_C( 14), INT8_C( -76), INT8_C(-101),
1090
+ INT8_C( 13), INT8_C( 64), INT8_C( -86), INT8_C( 34)),
1091
+ simde_mm_set_epi64x(INT64_C( -86), INT64_C( 34)) },
1092
+ { simde_mm_set_epi8(INT8_C( 87), INT8_C( 115), INT8_C( 108), INT8_C( -88),
1093
+ INT8_C( 113), INT8_C( 91), INT8_C( -95), INT8_C( -69),
1094
+ INT8_C( 32), INT8_C( 98), INT8_C( -16), INT8_C( -55),
1095
+ INT8_C( 11), INT8_C( -23), INT8_C( 72), INT8_C( 121)),
1096
+ simde_mm_set_epi64x(INT64_C( 72), INT64_C( 121)) },
1097
+ { simde_mm_set_epi8(INT8_C( 88), INT8_C( 68), INT8_C( -76), INT8_C( 35),
1098
+ INT8_C(-119), INT8_C( -67), INT8_C( 52), INT8_C( 58),
1099
+ INT8_C( 7), INT8_C( 70), INT8_C( -26), INT8_C( 55),
1100
+ INT8_C( 126), INT8_C( 69), INT8_C( 105), INT8_C( 40)),
1101
+ simde_mm_set_epi64x(INT64_C( 105), INT64_C( 40)) },
1102
+ { simde_mm_set_epi8(INT8_C( 110), INT8_C( -4), INT8_C( 41), INT8_C( -98),
1103
+ INT8_C( -91), INT8_C( 117), INT8_C( -33), INT8_C(-110),
1104
+ INT8_C( 126), INT8_C( -63), INT8_C( -65), INT8_C( -86),
1105
+ INT8_C(-108), INT8_C( -64), INT8_C( -4), INT8_C( 25)),
1106
+ simde_mm_set_epi64x(INT64_C( -4), INT64_C( 25)) },
1107
+ { simde_mm_set_epi8(INT8_C( 105), INT8_C( -31), INT8_C( -57), INT8_C( -74),
1108
+ INT8_C( 119), INT8_C( 56), INT8_C( -36), INT8_C(-128),
1109
+ INT8_C( 102), INT8_C( 57), INT8_C( -27), INT8_C( 19),
1110
+ INT8_C( -50), INT8_C( -15), INT8_C( 36), INT8_C(-124)),
1111
+ simde_mm_set_epi64x(INT64_C( 36), INT64_C( -124)) },
1112
+ { simde_mm_set_epi8(INT8_C( 44), INT8_C( -84), INT8_C( -61), INT8_C( 59),
1113
+ INT8_C( 18), INT8_C(-118), INT8_C( 6), INT8_C( 56),
1114
+ INT8_C( -11), INT8_C( 78), INT8_C( -64), INT8_C( -28),
1115
+ INT8_C( -85), INT8_C( -77), INT8_C( -25), INT8_C(-100)),
1116
+ simde_mm_set_epi64x(INT64_C( -25), INT64_C( -100)) },
1117
+ { simde_mm_set_epi8(INT8_C( -13), INT8_C( -24), INT8_C( 94), INT8_C( 58),
1118
+ INT8_C( 1), INT8_C( -14), INT8_C( -90), INT8_C(-123),
1119
+ INT8_C(-101), INT8_C( 71), INT8_C( -51), INT8_C( 87),
1120
+ INT8_C( -34), INT8_C( 77), INT8_C( 124), INT8_C( -63)),
1121
+ simde_mm_set_epi64x(INT64_C( 124), INT64_C( -63)) }
1122
+ };
1123
+
1124
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
1125
+ simde__m128i r = simde_mm_cvtepi8_epi64(test_vec[i].a);
1126
+ simde_assert_m128i_i64(r, ==, test_vec[i].r);
1127
+ }
1128
+
1129
+ return MUNIT_OK;
1130
+ }
1131
+
1132
+ static MunitResult
1133
+ test_simde_mm_cvtepu8_epi16(const MunitParameter params[], void* data) {
1134
+ (void) params;
1135
+ (void) data;
1136
+
1137
+ const struct {
1138
+ simde__m128i a;
1139
+ simde__m128i r;
1140
+ } test_vec[8] = {
1141
+ { simde_x_mm_set_epu8(UINT8_C(133), UINT8_C( 55), UINT8_C(102), UINT8_C(130),
1142
+ UINT8_C(197), UINT8_C( 52), UINT8_C(225), UINT8_C( 87),
1143
+ UINT8_C(218), UINT8_C( 15), UINT8_C(214), UINT8_C(152),
1144
+ UINT8_C(108), UINT8_C( 18), UINT8_C(104), UINT8_C( 3)),
1145
+ simde_mm_set_epi16(INT16_C( 218), INT16_C( 15), INT16_C( 214), INT16_C( 152),
1146
+ INT16_C( 108), INT16_C( 18), INT16_C( 104), INT16_C( 3)) },
1147
+ { simde_x_mm_set_epu8(UINT8_C( 51), UINT8_C(122), UINT8_C(167), UINT8_C(232),
1148
+ UINT8_C( 3), UINT8_C(167), UINT8_C( 89), UINT8_C(206),
1149
+ UINT8_C( 96), UINT8_C(235), UINT8_C(128), UINT8_C(134),
1150
+ UINT8_C(187), UINT8_C(190), UINT8_C(232), UINT8_C(112)),
1151
+ simde_mm_set_epi16(INT16_C( 96), INT16_C( 235), INT16_C( 128), INT16_C( 134),
1152
+ INT16_C( 187), INT16_C( 190), INT16_C( 232), INT16_C( 112)) },
1153
+ { simde_x_mm_set_epu8(UINT8_C( 15), UINT8_C( 99), UINT8_C(228), UINT8_C( 12),
1154
+ UINT8_C(155), UINT8_C(138), UINT8_C( 87), UINT8_C( 96),
1155
+ UINT8_C( 98), UINT8_C( 66), UINT8_C( 96), UINT8_C( 53),
1156
+ UINT8_C( 65), UINT8_C(187), UINT8_C(183), UINT8_C(114)),
1157
+ simde_mm_set_epi16(INT16_C( 98), INT16_C( 66), INT16_C( 96), INT16_C( 53),
1158
+ INT16_C( 65), INT16_C( 187), INT16_C( 183), INT16_C( 114)) },
1159
+ { simde_x_mm_set_epu8(UINT8_C( 58), UINT8_C( 51), UINT8_C(210), UINT8_C( 42),
1160
+ UINT8_C(132), UINT8_C(100), UINT8_C( 83), UINT8_C( 44),
1161
+ UINT8_C(115), UINT8_C(204), UINT8_C(195), UINT8_C( 11),
1162
+ UINT8_C( 48), UINT8_C( 39), UINT8_C( 37), UINT8_C(183)),
1163
+ simde_mm_set_epi16(INT16_C( 115), INT16_C( 204), INT16_C( 195), INT16_C( 11),
1164
+ INT16_C( 48), INT16_C( 39), INT16_C( 37), INT16_C( 183)) },
1165
+ { simde_x_mm_set_epu8(UINT8_C( 56), UINT8_C(101), UINT8_C(162), UINT8_C(107),
1166
+ UINT8_C( 59), UINT8_C( 91), UINT8_C( 63), UINT8_C(129),
1167
+ UINT8_C(189), UINT8_C( 4), UINT8_C( 9), UINT8_C(155),
1168
+ UINT8_C(150), UINT8_C( 39), UINT8_C(140), UINT8_C( 87)),
1169
+ simde_mm_set_epi16(INT16_C( 189), INT16_C( 4), INT16_C( 9), INT16_C( 155),
1170
+ INT16_C( 150), INT16_C( 39), INT16_C( 140), INT16_C( 87)) },
1171
+ { simde_x_mm_set_epu8(UINT8_C(203), UINT8_C(158), UINT8_C(175), UINT8_C( 0),
1172
+ UINT8_C( 27), UINT8_C( 64), UINT8_C(129), UINT8_C( 41),
1173
+ UINT8_C(208), UINT8_C(180), UINT8_C(174), UINT8_C( 83),
1174
+ UINT8_C(142), UINT8_C(198), UINT8_C( 36), UINT8_C(158)),
1175
+ simde_mm_set_epi16(INT16_C( 208), INT16_C( 180), INT16_C( 174), INT16_C( 83),
1176
+ INT16_C( 142), INT16_C( 198), INT16_C( 36), INT16_C( 158)) },
1177
+ { simde_x_mm_set_epu8(UINT8_C( 48), UINT8_C(139), UINT8_C( 69), UINT8_C( 42),
1178
+ UINT8_C(239), UINT8_C(255), UINT8_C( 84), UINT8_C(226),
1179
+ UINT8_C(180), UINT8_C(204), UINT8_C( 81), UINT8_C(133),
1180
+ UINT8_C(230), UINT8_C( 25), UINT8_C(218), UINT8_C(139)),
1181
+ simde_mm_set_epi16(INT16_C( 180), INT16_C( 204), INT16_C( 81), INT16_C( 133),
1182
+ INT16_C( 230), INT16_C( 25), INT16_C( 218), INT16_C( 139)) },
1183
+ { simde_x_mm_set_epu8(UINT8_C(213), UINT8_C( 5), UINT8_C( 37), UINT8_C(243),
1184
+ UINT8_C(196), UINT8_C(241), UINT8_C( 11), UINT8_C( 32),
1185
+ UINT8_C(137), UINT8_C( 5), UINT8_C(216), UINT8_C(243),
1186
+ UINT8_C( 25), UINT8_C( 23), UINT8_C(212), UINT8_C( 57)),
1187
+ simde_mm_set_epi16(INT16_C( 137), INT16_C( 5), INT16_C( 216), INT16_C( 243),
1188
+ INT16_C( 25), INT16_C( 23), INT16_C( 212), INT16_C( 57)) }
1189
+ };
1190
+
1191
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
1192
+ simde__m128i r = simde_mm_cvtepu8_epi16(test_vec[i].a);
1193
+ simde_assert_m128i_i16(r, ==, test_vec[i].r);
1194
+ }
1195
+
1196
+ return MUNIT_OK;
1197
+ }
1198
+
1199
+ static MunitResult
1200
+ test_simde_mm_cvtepu8_epi32(const MunitParameter params[], void* data) {
1201
+ (void) params;
1202
+ (void) data;
1203
+
1204
+ const struct {
1205
+ simde__m128i a;
1206
+ simde__m128i r;
1207
+ } test_vec[8] = {
1208
+ { simde_x_mm_set_epu8(216, 58, 77, 216, 179, 254, 83, 42,
1209
+ 179, 211, 84, 116, 211, 142, 19, 247),
1210
+ simde_mm_set_epi32(211, 142, 19, 247) },
1211
+ { simde_x_mm_set_epu8(161, 70, 71, 24, 62, 44, 28, 40,
1212
+ 34, 181, 114, 59, 33, 122, 68, 198),
1213
+ simde_mm_set_epi32( 33, 122, 68, 198) },
1214
+ { simde_x_mm_set_epu8(145, 160, 60, 140, 53, 127, 68, 235,
1215
+ 225, 143, 244, 79, 20, 67, 173, 209),
1216
+ simde_mm_set_epi32( 20, 67, 173, 209) },
1217
+ { simde_x_mm_set_epu8(117, 146, 120, 54, 143, 242, 141, 149,
1218
+ 73, 58, 87, 32, 199, 183, 202, 12),
1219
+ simde_mm_set_epi32(199, 183, 202, 12) },
1220
+ { simde_x_mm_set_epu8( 40, 57, 16, 79, 103, 249, 143, 197,
1221
+ 214, 78, 201, 95, 85, 53, 247, 134),
1222
+ simde_mm_set_epi32( 85, 53, 247, 134) },
1223
+ { simde_x_mm_set_epu8( 72, 163, 211, 103, 152, 114, 2, 140,
1224
+ 52, 64, 1, 9, 184, 145, 190, 245),
1225
+ simde_mm_set_epi32(184, 145, 190, 245) },
1226
+ { simde_x_mm_set_epu8( 7, 28, 147, 3, 70, 73, 159, 19,
1227
+ 118, 122, 100, 52, 236, 17, 135, 121),
1228
+ simde_mm_set_epi32(236, 17, 135, 121) },
1229
+ { simde_x_mm_set_epu8(100, 233, 199, 97, 158, 240, 121, 115,
1230
+ 142, 146, 159, 107, 202, 224, 156, 149),
1231
+ simde_mm_set_epi32(202, 224, 156, 149) }
1232
+ };
1233
+
1234
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
1235
+ simde__m128i r = simde_mm_cvtepu8_epi32(test_vec[i].a);
1236
+ simde_assert_m128i_i32(r, ==, test_vec[i].r);
1237
+ }
1238
+
1239
+ return MUNIT_OK;
1240
+ }
1241
+
1242
+ static MunitResult
1243
+ test_simde_mm_cvtepu8_epi64(const MunitParameter params[], void* data) {
1244
+ (void) params;
1245
+ (void) data;
1246
+
1247
+ const struct {
1248
+ simde__m128i a;
1249
+ simde__m128i r;
1250
+ } test_vec[8] = {
1251
+ { simde_x_mm_set_epu8(UINT8_C(165), UINT8_C(138), UINT8_C( 9), UINT8_C(193),
1252
+ UINT8_C(199), UINT8_C(213), UINT8_C(108), UINT8_C(127),
1253
+ UINT8_C( 76), UINT8_C( 44), UINT8_C( 97), UINT8_C( 43),
1254
+ UINT8_C( 10), UINT8_C( 41), UINT8_C(226), UINT8_C(134)),
1255
+ simde_mm_set_epi64x(INT64_C( 226), INT64_C( 134)) },
1256
+ { simde_x_mm_set_epu8(UINT8_C(206), UINT8_C(216), UINT8_C(228), UINT8_C( 95),
1257
+ UINT8_C( 16), UINT8_C(235), UINT8_C(117), UINT8_C(249),
1258
+ UINT8_C(118), UINT8_C(169), UINT8_C( 43), UINT8_C(125),
1259
+ UINT8_C(139), UINT8_C( 7), UINT8_C( 78), UINT8_C( 91)),
1260
+ simde_mm_set_epi64x(INT64_C( 78), INT64_C( 91)) },
1261
+ { simde_x_mm_set_epu8(UINT8_C(217), UINT8_C( 54), UINT8_C( 70), UINT8_C( 24),
1262
+ UINT8_C( 18), UINT8_C(185), UINT8_C( 8), UINT8_C( 83),
1263
+ UINT8_C(138), UINT8_C(127), UINT8_C(171), UINT8_C(232),
1264
+ UINT8_C( 25), UINT8_C( 28), UINT8_C(187), UINT8_C(172)),
1265
+ simde_mm_set_epi64x(INT64_C( 187), INT64_C( 172)) },
1266
+ { simde_x_mm_set_epu8(UINT8_C(139), UINT8_C(164), UINT8_C(130), UINT8_C(202),
1267
+ UINT8_C(244), UINT8_C(121), UINT8_C( 41), UINT8_C(156),
1268
+ UINT8_C(187), UINT8_C( 51), UINT8_C( 98), UINT8_C( 32),
1269
+ UINT8_C(179), UINT8_C(192), UINT8_C(214), UINT8_C(180)),
1270
+ simde_mm_set_epi64x(INT64_C( 214), INT64_C( 180)) },
1271
+ { simde_x_mm_set_epu8(UINT8_C(216), UINT8_C( 40), UINT8_C(253), UINT8_C(179),
1272
+ UINT8_C(134), UINT8_C(199), UINT8_C( 8), UINT8_C( 38),
1273
+ UINT8_C(169), UINT8_C(253), UINT8_C(114), UINT8_C( 51),
1274
+ UINT8_C(228), UINT8_C(119), UINT8_C(103), UINT8_C(155)),
1275
+ simde_mm_set_epi64x(INT64_C( 103), INT64_C( 155)) },
1276
+ { simde_x_mm_set_epu8(UINT8_C(197), UINT8_C(169), UINT8_C( 70), UINT8_C(235),
1277
+ UINT8_C(100), UINT8_C(250), UINT8_C( 64), UINT8_C(127),
1278
+ UINT8_C(118), UINT8_C(162), UINT8_C( 19), UINT8_C(173),
1279
+ UINT8_C(116), UINT8_C(180), UINT8_C( 7), UINT8_C(216)),
1280
+ simde_mm_set_epi64x(INT64_C( 7), INT64_C( 216)) },
1281
+ { simde_x_mm_set_epu8(UINT8_C( 50), UINT8_C( 69), UINT8_C( 69), UINT8_C(227),
1282
+ UINT8_C( 69), UINT8_C( 0), UINT8_C( 34), UINT8_C(232),
1283
+ UINT8_C( 24), UINT8_C( 46), UINT8_C(141), UINT8_C(128),
1284
+ UINT8_C(115), UINT8_C(157), UINT8_C(112), UINT8_C( 71)),
1285
+ simde_mm_set_epi64x(INT64_C( 112), INT64_C( 71)) },
1286
+ { simde_x_mm_set_epu8(UINT8_C(120), UINT8_C( 3), UINT8_C(197), UINT8_C(139),
1287
+ UINT8_C( 31), UINT8_C(200), UINT8_C(250), UINT8_C(106),
1288
+ UINT8_C(219), UINT8_C( 4), UINT8_C(206), UINT8_C( 64),
1289
+ UINT8_C(217), UINT8_C(231), UINT8_C(229), UINT8_C( 14)),
1290
+ simde_mm_set_epi64x(INT64_C( 229), INT64_C( 14)) }
1291
+ };
1292
+
1293
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
1294
+ simde__m128i r = simde_mm_cvtepu8_epi64(test_vec[i].a);
1295
+ simde_assert_m128i_i64(r, ==, test_vec[i].r);
1296
+ }
1297
+
1298
+ return MUNIT_OK;
1299
+ }
1300
+
1301
+ static MunitResult
1302
+ test_simde_mm_cvtepu16_epi32(const MunitParameter params[], void* data) {
1303
+ (void) params;
1304
+ (void) data;
1305
+
1306
+ const struct {
1307
+ simde__m128i a;
1308
+ simde__m128i r;
1309
+ } test_vec[8] = {
1310
+ { simde_x_mm_set_epu16(UINT16_C(45997), UINT16_C(54881), UINT16_C( 8073), UINT16_C(50053),
1311
+ UINT16_C(33513), UINT16_C(60494), UINT16_C(44453), UINT16_C(35449)),
1312
+ simde_mm_set_epi32(INT32_C( 33513), INT32_C( 60494), INT32_C( 44453), INT32_C( 35449)) },
1313
+ { simde_x_mm_set_epu16(UINT16_C(29322), UINT16_C(60312), UINT16_C(23894), UINT16_C( 5611),
1314
+ UINT16_C(43169), UINT16_C(36178), UINT16_C(30857), UINT16_C(27437)),
1315
+ simde_mm_set_epi32(INT32_C( 43169), INT32_C( 36178), INT32_C( 30857), INT32_C( 27437)) },
1316
+ { simde_x_mm_set_epu16(UINT16_C(36165), UINT16_C(32250), UINT16_C(32420), UINT16_C(54083),
1317
+ UINT16_C(37595), UINT16_C( 6323), UINT16_C(24868), UINT16_C(13806)),
1318
+ simde_mm_set_epi32(INT32_C( 37595), INT32_C( 6323), INT32_C( 24868), INT32_C( 13806)) },
1319
+ { simde_x_mm_set_epu16(UINT16_C(47473), UINT16_C(25142), UINT16_C(53974), UINT16_C(29717),
1320
+ UINT16_C(38600), UINT16_C(37432), UINT16_C(33857), UINT16_C( 7799)),
1321
+ simde_mm_set_epi32(INT32_C( 38600), INT32_C( 37432), INT32_C( 33857), INT32_C( 7799)) },
1322
+ { simde_x_mm_set_epu16(UINT16_C( 9476), UINT16_C(49162), UINT16_C(59100), UINT16_C(46196),
1323
+ UINT16_C(39403), UINT16_C(33179), UINT16_C(42032), UINT16_C(16407)),
1324
+ simde_mm_set_epi32(INT32_C( 39403), INT32_C( 33179), INT32_C( 42032), INT32_C( 16407)) },
1325
+ { simde_x_mm_set_epu16(UINT16_C(13947), UINT16_C(14884), UINT16_C(63463), UINT16_C(60885),
1326
+ UINT16_C(44795), UINT16_C(57577), UINT16_C( 9399), UINT16_C(54740)),
1327
+ simde_mm_set_epi32(INT32_C( 44795), INT32_C( 57577), INT32_C( 9399), INT32_C( 54740)) },
1328
+ { simde_x_mm_set_epu16(UINT16_C(16348), UINT16_C(41263), UINT16_C(29933), UINT16_C( 7966),
1329
+ UINT16_C(11747), UINT16_C(45808), UINT16_C(10076), UINT16_C(52634)),
1330
+ simde_mm_set_epi32(INT32_C( 11747), INT32_C( 45808), INT32_C( 10076), INT32_C( 52634)) },
1331
+ { simde_x_mm_set_epu16(UINT16_C(44921), UINT16_C(31426), UINT16_C(50653), UINT16_C(39011),
1332
+ UINT16_C(15796), UINT16_C(28031), UINT16_C(50348), UINT16_C(35430)),
1333
+ simde_mm_set_epi32(INT32_C( 15796), INT32_C( 28031), INT32_C( 50348), INT32_C( 35430)) }
1334
+ };
1335
+
1336
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
1337
+ simde__m128i r = simde_mm_cvtepu16_epi32(test_vec[i].a);
1338
+ simde_assert_m128i_i32(r, ==, test_vec[i].r);
1339
+ }
1340
+
1341
+ return MUNIT_OK;
1342
+ }
1343
+
1344
+ static MunitResult
1345
+ test_simde_mm_cvtepu16_epi64(const MunitParameter params[], void* data) {
1346
+ (void) params;
1347
+ (void) data;
1348
+
1349
+ const struct {
1350
+ simde__m128i a;
1351
+ simde__m128i r;
1352
+ } test_vec[8] = {
1353
+ { simde_x_mm_set_epu16(UINT16_C(36944), UINT16_C(47640), UINT16_C( 2187), UINT16_C(10560),
1354
+ UINT16_C( 1923), UINT16_C(52017), UINT16_C(41190), UINT16_C(51381)),
1355
+ simde_mm_set_epi64x(INT64_C( 41190), INT64_C( 51381)) },
1356
+ { simde_x_mm_set_epu16(UINT16_C(59110), UINT16_C( 781), UINT16_C( 2267), UINT16_C(32518),
1357
+ UINT16_C(16361), UINT16_C(14242), UINT16_C(28197), UINT16_C(27455)),
1358
+ simde_mm_set_epi64x(INT64_C( 28197), INT64_C( 27455)) },
1359
+ { simde_x_mm_set_epu16(UINT16_C(45312), UINT16_C( 9417), UINT16_C(53316), UINT16_C(54914),
1360
+ UINT16_C( 6651), UINT16_C(32993), UINT16_C(51134), UINT16_C(31066)),
1361
+ simde_mm_set_epi64x(INT64_C( 51134), INT64_C( 31066)) },
1362
+ { simde_x_mm_set_epu16(UINT16_C( 6198), UINT16_C(41996), UINT16_C(34257), UINT16_C( 1856),
1363
+ UINT16_C(54306), UINT16_C(57993), UINT16_C(28922), UINT16_C(25456)),
1364
+ simde_mm_set_epi64x(INT64_C( 28922), INT64_C( 25456)) },
1365
+ { simde_x_mm_set_epu16(UINT16_C(52812), UINT16_C(10727), UINT16_C(10105), UINT16_C(31173),
1366
+ UINT16_C( 2196), UINT16_C(38149), UINT16_C(12797), UINT16_C(56721)),
1367
+ simde_mm_set_epi64x(INT64_C( 12797), INT64_C( 56721)) },
1368
+ { simde_x_mm_set_epu16(UINT16_C(43893), UINT16_C(22212), UINT16_C(55173), UINT16_C(39507),
1369
+ UINT16_C(17371), UINT16_C( 1379), UINT16_C(37505), UINT16_C(51758)),
1370
+ simde_mm_set_epi64x(INT64_C( 37505), INT64_C( 51758)) },
1371
+ { simde_x_mm_set_epu16(UINT16_C(10102), UINT16_C(59029), UINT16_C(11878), UINT16_C(57288),
1372
+ UINT16_C(63955), UINT16_C(62443), UINT16_C( 1106), UINT16_C(32843)),
1373
+ simde_mm_set_epi64x(INT64_C( 1106), INT64_C( 32843)) },
1374
+ { simde_x_mm_set_epu16(UINT16_C( 7384), UINT16_C(18925), UINT16_C(61664), UINT16_C(62156),
1375
+ UINT16_C(64229), UINT16_C(12260), UINT16_C(17269), UINT16_C( 165)),
1376
+ simde_mm_set_epi64x(INT64_C( 17269), INT64_C( 165)) }
1377
+ };
1378
+
1379
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
1380
+ simde__m128i r = simde_mm_cvtepu16_epi64(test_vec[i].a);
1381
+ simde_assert_m128i_i64(r, ==, test_vec[i].r);
1382
+ }
1383
+
1384
+ return MUNIT_OK;
1385
+ }
1386
+
1387
+ static MunitResult
1388
+ test_simde_mm_cvtepu32_epi64(const MunitParameter params[], void* data) {
1389
+ (void) params;
1390
+ (void) data;
1391
+
1392
+ const struct {
1393
+ simde__m128i a;
1394
+ simde__m128i r;
1395
+ } test_vec[8] = {
1396
+ { simde_x_mm_set_epu32(UINT32_C(1089356158), UINT32_C( 716812601), UINT32_C(2600352285), UINT32_C( 851998215)),
1397
+ simde_mm_set_epi64x(INT64_C( 2600352285), INT64_C( 851998215)) },
1398
+ { simde_x_mm_set_epu32(UINT32_C(2083471993), UINT32_C(3995663695), UINT32_C( 718227173), UINT32_C(3645469582)),
1399
+ simde_mm_set_epi64x(INT64_C( 718227173), INT64_C( 3645469582)) },
1400
+ { simde_x_mm_set_epu32(UINT32_C(3121544286), UINT32_C( 583537062), UINT32_C( 362422775), UINT32_C(3876080612)),
1401
+ simde_mm_set_epi64x(INT64_C( 362422775), INT64_C( 3876080612)) },
1402
+ { simde_x_mm_set_epu32(UINT32_C( 336299187), UINT32_C(1772325077), UINT32_C(4159021368), UINT32_C(3897849224)),
1403
+ simde_mm_set_epi64x(INT64_C( 4159021368), INT64_C( 3897849224)) },
1404
+ { simde_x_mm_set_epu32(UINT32_C(2614142766), UINT32_C(3735995755), UINT32_C(2880699788), UINT32_C( 410975381)),
1405
+ simde_mm_set_epi64x(INT64_C( 2880699788), INT64_C( 410975381)) },
1406
+ { simde_x_mm_set_epu32(UINT32_C(2584819359), UINT32_C(2278276482), UINT32_C(3087779260), UINT32_C(2573826343)),
1407
+ simde_mm_set_epi64x(INT64_C( 3087779260), INT64_C( 2573826343)) },
1408
+ { simde_x_mm_set_epu32(UINT32_C( 809379705), UINT32_C( 343015345), UINT32_C( 434582428), UINT32_C( 776037113)),
1409
+ simde_mm_set_epi64x(INT64_C( 434582428), INT64_C( 776037113)) },
1410
+ { simde_x_mm_set_epu32(UINT32_C(2474833343), UINT32_C(1439643533), UINT32_C(3207542234), UINT32_C(2907982947)),
1411
+ simde_mm_set_epi64x(INT64_C( 3207542234), INT64_C( 2907982947)) }
1412
+ };
1413
+
1414
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
1415
+ simde__m128i r = simde_mm_cvtepu32_epi64(test_vec[i].a);
1416
+ simde_assert_m128i_i64(r, ==, test_vec[i].r);
1417
+ }
1418
+
1419
+ return MUNIT_OK;
1420
+ }
1421
+
1422
+ static MunitResult
1423
+ test_simde_mm_dp_pd(const MunitParameter params[], void* data) {
1424
+ (void) params;
1425
+ (void) data;
1426
+
1427
+ const struct {
1428
+ simde__m128d a;
1429
+ simde__m128d b;
1430
+ simde__m128d r;
1431
+ } test_vec[8] = {
1432
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( -564.11), SIMDE_FLOAT64_C( 340.88)),
1433
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -778.50), SIMDE_FLOAT64_C( -785.47)),
1434
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 0.00), SIMDE_FLOAT64_C(-267751.01)) },
1435
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( -557.21), SIMDE_FLOAT64_C( -291.36)),
1436
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -799.93), SIMDE_FLOAT64_C( -716.65)),
1437
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 0.00), SIMDE_FLOAT64_C(208803.14)) },
1438
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( -377.58), SIMDE_FLOAT64_C( 860.62)),
1439
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -199.70), SIMDE_FLOAT64_C( -642.81)),
1440
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 0.00), SIMDE_FLOAT64_C(-553215.14)) },
1441
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( -782.59), SIMDE_FLOAT64_C( 223.96)),
1442
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -805.25), SIMDE_FLOAT64_C( 736.23)),
1443
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 0.00), SIMDE_FLOAT64_C(164886.07)) },
1444
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( 472.55), SIMDE_FLOAT64_C( -591.15)),
1445
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 932.48), SIMDE_FLOAT64_C( 630.07)),
1446
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 0.00), SIMDE_FLOAT64_C(-372465.88)) },
1447
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( 484.26), SIMDE_FLOAT64_C( -41.63)),
1448
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 593.03), SIMDE_FLOAT64_C( 955.21)),
1449
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 0.00), SIMDE_FLOAT64_C(-39765.39)) },
1450
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( 820.67), SIMDE_FLOAT64_C( 120.74)),
1451
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 576.08), SIMDE_FLOAT64_C( 524.49)),
1452
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 0.00), SIMDE_FLOAT64_C(63326.92)) },
1453
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( 696.23), SIMDE_FLOAT64_C( -565.44)),
1454
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 313.65), SIMDE_FLOAT64_C( 242.19)),
1455
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 0.00), SIMDE_FLOAT64_C(-136943.91)) }
1456
+ };
1457
+
1458
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
1459
+ simde__m128d r = simde_mm_dp_pd(test_vec[i].a, test_vec[i].b, 85);
1460
+ simde_assert_m128d_close(r, test_vec[i].r, 1);
1461
+ }
1462
+
1463
+ return MUNIT_OK;
1464
+ }
1465
+
1466
+ static MunitResult
1467
+ test_simde_mm_dp_ps(const MunitParameter params[], void* data) {
1468
+ (void) params;
1469
+ (void) data;
1470
+
1471
+ const struct {
1472
+ simde__m128 a;
1473
+ simde__m128 b;
1474
+ simde__m128 r;
1475
+ } test_vec[8] = {
1476
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( 553.37), SIMDE_FLOAT32_C( 120.49), SIMDE_FLOAT32_C( -272.20), SIMDE_FLOAT32_C( -36.55)),
1477
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 115.84), SIMDE_FLOAT32_C( 676.89), SIMDE_FLOAT32_C( -919.81), SIMDE_FLOAT32_C( -927.02)),
1478
+ simde_mm_set_ps(SIMDE_FLOAT32_C(250372.30), SIMDE_FLOAT32_C( 0.00), SIMDE_FLOAT32_C(250372.30), SIMDE_FLOAT32_C( 0.00)) },
1479
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( 329.44), SIMDE_FLOAT32_C( -11.73), SIMDE_FLOAT32_C( -693.65), SIMDE_FLOAT32_C( -971.02)),
1480
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -316.00), SIMDE_FLOAT32_C( -919.35), SIMDE_FLOAT32_C( 956.50), SIMDE_FLOAT32_C( -382.80)),
1481
+ simde_mm_set_ps(SIMDE_FLOAT32_C(-663476.25), SIMDE_FLOAT32_C( 0.00), SIMDE_FLOAT32_C(-663476.25), SIMDE_FLOAT32_C( 0.00)) },
1482
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( 48.87), SIMDE_FLOAT32_C( -212.15), SIMDE_FLOAT32_C( 812.72), SIMDE_FLOAT32_C( 350.56)),
1483
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 612.16), SIMDE_FLOAT32_C( -336.28), SIMDE_FLOAT32_C( -314.44), SIMDE_FLOAT32_C( 725.39)),
1484
+ simde_mm_set_ps(SIMDE_FLOAT32_C(-255551.67), SIMDE_FLOAT32_C( 0.00), SIMDE_FLOAT32_C(-255551.67), SIMDE_FLOAT32_C( 0.00)) },
1485
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( -17.80), SIMDE_FLOAT32_C( 55.32), SIMDE_FLOAT32_C( 517.59), SIMDE_FLOAT32_C( 356.24)),
1486
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 440.41), SIMDE_FLOAT32_C( -728.96), SIMDE_FLOAT32_C( 250.15), SIMDE_FLOAT32_C( -658.46)),
1487
+ simde_mm_set_ps(SIMDE_FLOAT32_C(129475.14), SIMDE_FLOAT32_C( 0.00), SIMDE_FLOAT32_C(129475.14), SIMDE_FLOAT32_C( 0.00)) },
1488
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( 724.19), SIMDE_FLOAT32_C( 892.03), SIMDE_FLOAT32_C( 632.84), SIMDE_FLOAT32_C( 912.13)),
1489
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 816.72), SIMDE_FLOAT32_C( 934.76), SIMDE_FLOAT32_C( -86.66), SIMDE_FLOAT32_C( 780.21)),
1490
+ simde_mm_set_ps(SIMDE_FLOAT32_C(-54841.92), SIMDE_FLOAT32_C( 0.00), SIMDE_FLOAT32_C(-54841.92), SIMDE_FLOAT32_C( 0.00)) },
1491
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( 767.35), SIMDE_FLOAT32_C( 593.67), SIMDE_FLOAT32_C( 295.16), SIMDE_FLOAT32_C( -928.99)),
1492
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 446.17), SIMDE_FLOAT32_C( -781.62), SIMDE_FLOAT32_C( 967.10), SIMDE_FLOAT32_C( -630.50)),
1493
+ simde_mm_set_ps(SIMDE_FLOAT32_C(285449.22), SIMDE_FLOAT32_C( 0.00), SIMDE_FLOAT32_C(285449.22), SIMDE_FLOAT32_C( 0.00)) },
1494
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( 954.16), SIMDE_FLOAT32_C( -247.39), SIMDE_FLOAT32_C( 69.65), SIMDE_FLOAT32_C( -817.12)),
1495
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 330.91), SIMDE_FLOAT32_C( 409.60), SIMDE_FLOAT32_C( 841.84), SIMDE_FLOAT32_C( -474.27)),
1496
+ simde_mm_set_ps(SIMDE_FLOAT32_C(58634.16), SIMDE_FLOAT32_C( 0.00), SIMDE_FLOAT32_C(58634.16), SIMDE_FLOAT32_C( 0.00)) },
1497
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( -783.51), SIMDE_FLOAT32_C( -866.65), SIMDE_FLOAT32_C( -877.85), SIMDE_FLOAT32_C( -911.84)),
1498
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 4.33), SIMDE_FLOAT32_C( 855.54), SIMDE_FLOAT32_C( -839.28), SIMDE_FLOAT32_C( 261.89)),
1499
+ simde_mm_set_ps(SIMDE_FLOAT32_C(736761.94), SIMDE_FLOAT32_C( 0.00), SIMDE_FLOAT32_C(736761.94), SIMDE_FLOAT32_C( 0.00)) }
1500
+ };
1501
+
1502
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
1503
+ simde__m128 r = simde_mm_dp_ps(test_vec[i].a, test_vec[i].b, 42);
1504
+ simde_assert_m128_close(r, test_vec[i].r, 1);
1505
+ }
1506
+
1507
+ return MUNIT_OK;
1508
+ }
1509
+
1510
+ static MunitResult
1511
+ test_simde_mm_extract_epi8(const MunitParameter params[], void* data) {
1512
+ (void) params;
1513
+ (void) data;
1514
+
1515
+ const struct {
1516
+ simde__m128i a;
1517
+ int8_t r;
1518
+ } test_vec[8] = {
1519
+ { simde_mm_set_epi8(INT8_C(-105), INT8_C(-107), INT8_C( -84), INT8_C( -57),
1520
+ INT8_C( 13), INT8_C(-101), INT8_C( -59), INT8_C( -93),
1521
+ INT8_C( 53), INT8_C( 83), INT8_C( -28), INT8_C( 126),
1522
+ INT8_C(-111), INT8_C( -97), INT8_C( 25), INT8_C( -72)),
1523
+ INT8_C(-97) },
1524
+ { simde_mm_set_epi8(INT8_C( -76), INT8_C( -25), INT8_C( 36), INT8_C( -92),
1525
+ INT8_C( -6), INT8_C( 77), INT8_C( -49), INT8_C(-126),
1526
+ INT8_C( -50), INT8_C( 79), INT8_C( -93), INT8_C( -95),
1527
+ INT8_C( -54), INT8_C(-116), INT8_C( -84), INT8_C( 66)),
1528
+ INT8_C(-116) },
1529
+ { simde_mm_set_epi8(INT8_C( 83), INT8_C( 123), INT8_C( 120), INT8_C( -66),
1530
+ INT8_C( -3), INT8_C( 66), INT8_C( 110), INT8_C( 76),
1531
+ INT8_C( 0), INT8_C( 83), INT8_C( 20), INT8_C( -12),
1532
+ INT8_C( 83), INT8_C( 4), INT8_C( -39), INT8_C( 118)),
1533
+ INT8_C(4) },
1534
+ { simde_mm_set_epi8(INT8_C( 32), INT8_C( 30), INT8_C( 127), INT8_C( -79),
1535
+ INT8_C( -10), INT8_C( 64), INT8_C( -45), INT8_C( 101),
1536
+ INT8_C( 44), INT8_C( -25), INT8_C(-110), INT8_C( 105),
1537
+ INT8_C( 83), INT8_C( 27), INT8_C( 15), INT8_C( 3)),
1538
+ INT8_C(27) },
1539
+ { simde_mm_set_epi8(INT8_C( 79), INT8_C( 95), INT8_C(-109), INT8_C( 26),
1540
+ INT8_C( -81), INT8_C( -5), INT8_C( -84), INT8_C( 115),
1541
+ INT8_C( -48), INT8_C( 104), INT8_C( 67), INT8_C( 26),
1542
+ INT8_C( -41), INT8_C( -30), INT8_C( 94), INT8_C( -19)),
1543
+ INT8_C(-30) },
1544
+ { simde_mm_set_epi8(INT8_C( -81), INT8_C( -80), INT8_C(-109), INT8_C( -14),
1545
+ INT8_C( 65), INT8_C( 10), INT8_C( 95), INT8_C( 83),
1546
+ INT8_C( 123), INT8_C( 45), INT8_C( 39), INT8_C( 106),
1547
+ INT8_C( -16), INT8_C(-123), INT8_C( -27), INT8_C( 56)),
1548
+ INT8_C(-123) },
1549
+ { simde_mm_set_epi8(INT8_C( 86), INT8_C( 79), INT8_C( 41), INT8_C( -51),
1550
+ INT8_C( 97), INT8_C( 42), INT8_C( 117), INT8_C(-119),
1551
+ INT8_C( -18), INT8_C(-127), INT8_C( 28), INT8_C( -21),
1552
+ INT8_C(-101), INT8_C( -61), INT8_C( -63), INT8_C( 13)),
1553
+ INT8_C(-61) },
1554
+ { simde_mm_set_epi8(INT8_C( -22), INT8_C( -42), INT8_C( 109), INT8_C( -75),
1555
+ INT8_C( -84), INT8_C( 60), INT8_C( -44), INT8_C( 8),
1556
+ INT8_C( 118), INT8_C( 102), INT8_C( -6), INT8_C( 0),
1557
+ INT8_C( -28), INT8_C( -81), INT8_C( 2), INT8_C( 94)),
1558
+ INT8_C(-81) }
1559
+ };
1560
+
1561
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
1562
+ int8_t r = simde_mm_extract_epi8(test_vec[i].a, 2);
1563
+ munit_assert_int8(r, ==, test_vec[i].r);
1564
+ }
1565
+
1566
+ return MUNIT_OK;
1567
+ }
1568
+
1569
+ static MunitResult
1570
+ test_simde_mm_extract_epi32(const MunitParameter params[], void* data) {
1571
+ (void) params;
1572
+ (void) data;
1573
+
1574
+ const struct {
1575
+ simde__m128i a;
1576
+ int32_t r;
1577
+ } test_vec[8] = {
1578
+ { simde_mm_set_epi32(INT32_C( -989914365), INT32_C( -320568713), INT32_C( 764201252), INT32_C( -823076780)),
1579
+ -320568713 },
1580
+ { simde_mm_set_epi32(INT32_C( 1872661147), INT32_C( 1704933787), INT32_C( -552294669), INT32_C(-1494809408)),
1581
+ 1704933787 },
1582
+ { simde_mm_set_epi32(INT32_C( 1741332044), INT32_C( 170998046), INT32_C( 1368723621), INT32_C(-2071088920)),
1583
+ 170998046 },
1584
+ { simde_mm_set_epi32(INT32_C(-1942216470), INT32_C(-1426359243), INT32_C( 862877610), INT32_C( 898094861)),
1585
+ -1426359243 },
1586
+ { simde_mm_set_epi32(INT32_C( 998704029), INT32_C( 2083244310), INT32_C( 1997200410), INT32_C( -459833332)),
1587
+ 2083244310 },
1588
+ { simde_mm_set_epi32(INT32_C( 1256369906), INT32_C( 953394288), INT32_C( 1499034833), INT32_C( 990189413)),
1589
+ 953394288 },
1590
+ { simde_mm_set_epi32(INT32_C( 1704589159), INT32_C( -719174981), INT32_C( -821402820), INT32_C( -969954348)),
1591
+ -719174981 },
1592
+ { simde_mm_set_epi32(INT32_C( 29062750), INT32_C( 1388055657), INT32_C( 1520861645), INT32_C(-1651761160)),
1593
+ 1388055657 }
1594
+ };
1595
+
1596
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
1597
+ int32_t r = simde_mm_extract_epi32(test_vec[i].a, 2);
1598
+ munit_assert_int32(r, ==, test_vec[i].r);
1599
+ }
1600
+
1601
+ return MUNIT_OK;
1602
+ }
1603
+
1604
+ static MunitResult
1605
+ test_simde_mm_extract_epi64(const MunitParameter params[], void* data) {
1606
+ (void) params;
1607
+ (void) data;
1608
+
1609
+ const struct {
1610
+ simde__m128i a;
1611
+ int64_t r;
1612
+ } test_vec[8] = {
1613
+ { simde_mm_set_epi64x(INT64_C( 3272238239526171606), INT64_C( 1654850679014379440)),
1614
+ INT64_C( 3272238239526171606) },
1615
+ { simde_mm_set_epi64x(INT64_C( -6129313439874067621), INT64_C( 8613384435853876616)),
1616
+ INT64_C( -6129313439874067621) },
1617
+ { simde_mm_set_epi64x(INT64_C( -6049641799095753965), INT64_C( -2810493496227171537)),
1618
+ INT64_C( -6049641799095753965) },
1619
+ { simde_mm_set_epi64x(INT64_C( 1486209521702266950), INT64_C( -7779883419844099734)),
1620
+ INT64_C( 1486209521702266950) },
1621
+ { simde_mm_set_epi64x(INT64_C( -4910022048576872084), INT64_C( -5820696718361059251)),
1622
+ INT64_C( -4910022048576872084) },
1623
+ { simde_mm_set_epi64x(INT64_C( -1923881977407946859), INT64_C( -5589720307210562868)),
1624
+ INT64_C( -1923881977407946859) },
1625
+ { simde_mm_set_epi64x(INT64_C( -3527421069943215520), INT64_C( 4545940692351426251)),
1626
+ INT64_C( -3527421069943215520) },
1627
+ { simde_mm_set_epi64x(INT64_C( 6123229719758625458), INT64_C( 1813312819011147191)),
1628
+ INT64_C( 6123229719758625458) }
1629
+ };
1630
+
1631
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
1632
+ int64_t r = simde_mm_extract_epi64(test_vec[i].a, 1);
1633
+ munit_assert_int64(r, ==, test_vec[i].r);
1634
+ }
1635
+
1636
+ return MUNIT_OK;
1637
+ }
1638
+
1639
+ static MunitResult
1640
+ test_simde_mm_floor_pd(const MunitParameter params[], void* data) {
1641
+ (void) params;
1642
+ (void) data;
1643
+
1644
+ const struct {
1645
+ simde__m128d a;
1646
+ simde__m128d r;
1647
+ } test_vec[8] = {
1648
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( -638.24), SIMDE_FLOAT64_C( 771.01)),
1649
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -639.00), SIMDE_FLOAT64_C( 771.00)) },
1650
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( -79.94), SIMDE_FLOAT64_C( 536.55)),
1651
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -80.00), SIMDE_FLOAT64_C( 536.00)) },
1652
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( 145.74), SIMDE_FLOAT64_C( 124.77)),
1653
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 145.00), SIMDE_FLOAT64_C( 124.00)) },
1654
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( 309.89), SIMDE_FLOAT64_C( -929.92)),
1655
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 309.00), SIMDE_FLOAT64_C( -930.00)) },
1656
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( 138.72), SIMDE_FLOAT64_C( 502.95)),
1657
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 138.00), SIMDE_FLOAT64_C( 502.00)) },
1658
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( 512.22), SIMDE_FLOAT64_C( 187.34)),
1659
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 512.00), SIMDE_FLOAT64_C( 187.00)) },
1660
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( 560.17), SIMDE_FLOAT64_C( 420.10)),
1661
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 560.00), SIMDE_FLOAT64_C( 420.00)) },
1662
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( 367.30), SIMDE_FLOAT64_C( 619.17)),
1663
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 367.00), SIMDE_FLOAT64_C( 619.00)) }
1664
+ };
1665
+
1666
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
1667
+ simde__m128d r = simde_mm_floor_pd(test_vec[i].a);
1668
+ simde_assert_m128d_close(r, test_vec[i].r, 1);
1669
+ }
1670
+
1671
+ return MUNIT_OK;
1672
+ }
1673
+
1674
+ static MunitResult
1675
+ test_simde_mm_floor_ps(const MunitParameter params[], void* data) {
1676
+ (void) params;
1677
+ (void) data;
1678
+
1679
+ const struct {
1680
+ simde__m128 a;
1681
+ simde__m128 r;
1682
+ } test_vec[8] = {
1683
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( 50.80), SIMDE_FLOAT32_C( 459.30), SIMDE_FLOAT32_C( -834.21), SIMDE_FLOAT32_C( 808.45)),
1684
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 50.00), SIMDE_FLOAT32_C( 459.00), SIMDE_FLOAT32_C( -835.00), SIMDE_FLOAT32_C( 808.00)) },
1685
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( -648.20), SIMDE_FLOAT32_C( 768.10), SIMDE_FLOAT32_C( -885.11), SIMDE_FLOAT32_C( -162.62)),
1686
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -649.00), SIMDE_FLOAT32_C( 768.00), SIMDE_FLOAT32_C( -886.00), SIMDE_FLOAT32_C( -163.00)) },
1687
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( -136.33), SIMDE_FLOAT32_C( 361.95), SIMDE_FLOAT32_C( -729.16), SIMDE_FLOAT32_C( -786.42)),
1688
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -137.00), SIMDE_FLOAT32_C( 361.00), SIMDE_FLOAT32_C( -730.00), SIMDE_FLOAT32_C( -787.00)) },
1689
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( -579.89), SIMDE_FLOAT32_C( 55.32), SIMDE_FLOAT32_C( -998.01), SIMDE_FLOAT32_C( 191.65)),
1690
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -580.00), SIMDE_FLOAT32_C( 55.00), SIMDE_FLOAT32_C( -999.00), SIMDE_FLOAT32_C( 191.00)) },
1691
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( 781.10), SIMDE_FLOAT32_C( 884.98), SIMDE_FLOAT32_C( 903.22), SIMDE_FLOAT32_C( 964.51)),
1692
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 781.00), SIMDE_FLOAT32_C( 884.00), SIMDE_FLOAT32_C( 903.00), SIMDE_FLOAT32_C( 964.00)) },
1693
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( 255.11), SIMDE_FLOAT32_C( 929.26), SIMDE_FLOAT32_C( 869.55), SIMDE_FLOAT32_C( -749.47)),
1694
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 255.00), SIMDE_FLOAT32_C( 929.00), SIMDE_FLOAT32_C( 869.00), SIMDE_FLOAT32_C( -750.00)) },
1695
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( -859.43), SIMDE_FLOAT32_C( 380.85), SIMDE_FLOAT32_C( -956.74), SIMDE_FLOAT32_C( -486.36)),
1696
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -860.00), SIMDE_FLOAT32_C( 380.00), SIMDE_FLOAT32_C( -957.00), SIMDE_FLOAT32_C( -487.00)) },
1697
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( 154.98), SIMDE_FLOAT32_C( 341.89), SIMDE_FLOAT32_C( -431.03), SIMDE_FLOAT32_C( -687.74)),
1698
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 154.00), SIMDE_FLOAT32_C( 341.00), SIMDE_FLOAT32_C( -432.00), SIMDE_FLOAT32_C( -688.00)) }
1699
+ };
1700
+
1701
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
1702
+ simde__m128 r = simde_mm_floor_ps(test_vec[i].a);
1703
+ simde_assert_m128_close(r, test_vec[i].r, 1);
1704
+ }
1705
+
1706
+ return MUNIT_OK;
1707
+ }
1708
+
1709
+ static MunitResult
1710
+ test_simde_mm_floor_sd(const MunitParameter params[], void* data) {
1711
+ (void) params;
1712
+ (void) data;
1713
+
1714
+ const struct {
1715
+ simde__m128d a;
1716
+ simde__m128d b;
1717
+ simde__m128d r;
1718
+ } test_vec[8] = {
1719
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( -651.87), SIMDE_FLOAT64_C( 624.42)),
1720
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 634.56), SIMDE_FLOAT64_C( -443.27)),
1721
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -651.87), SIMDE_FLOAT64_C( -444.00)) },
1722
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( -650.57), SIMDE_FLOAT64_C( -388.70)),
1723
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 862.67), SIMDE_FLOAT64_C( -847.32)),
1724
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -650.57), SIMDE_FLOAT64_C( -848.00)) },
1725
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( -429.60), SIMDE_FLOAT64_C( -927.81)),
1726
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 516.52), SIMDE_FLOAT64_C( -109.38)),
1727
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -429.60), SIMDE_FLOAT64_C( -110.00)) },
1728
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( 119.92), SIMDE_FLOAT64_C( 571.26)),
1729
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 15.01), SIMDE_FLOAT64_C( 819.76)),
1730
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 119.92), SIMDE_FLOAT64_C( 819.00)) },
1731
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( -625.27), SIMDE_FLOAT64_C( 998.84)),
1732
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -570.19), SIMDE_FLOAT64_C( -85.39)),
1733
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -625.27), SIMDE_FLOAT64_C( -86.00)) },
1734
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( 105.61), SIMDE_FLOAT64_C( -345.20)),
1735
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -705.02), SIMDE_FLOAT64_C( -93.90)),
1736
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 105.61), SIMDE_FLOAT64_C( -94.00)) },
1737
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( -824.97), SIMDE_FLOAT64_C( -396.57)),
1738
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -840.00), SIMDE_FLOAT64_C( 609.38)),
1739
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -824.97), SIMDE_FLOAT64_C( 609.00)) },
1740
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( 488.09), SIMDE_FLOAT64_C( -64.70)),
1741
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 92.99), SIMDE_FLOAT64_C( 531.79)),
1742
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 488.09), SIMDE_FLOAT64_C( 531.00)) }
1743
+ };
1744
+
1745
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
1746
+ simde__m128d r = simde_mm_floor_sd(test_vec[i].a, test_vec[i].b);
1747
+ simde_assert_m128d_close(r, test_vec[i].r, 1);
1748
+ }
1749
+
1750
+ return MUNIT_OK;
1751
+ }
1752
+
1753
+ static MunitResult
1754
+ test_simde_mm_floor_ss(const MunitParameter params[], void* data) {
1755
+ (void) params;
1756
+ (void) data;
1757
+
1758
+ const struct {
1759
+ simde__m128 a;
1760
+ simde__m128 b;
1761
+ simde__m128 r;
1762
+ } test_vec[8] = {
1763
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( 932.95), SIMDE_FLOAT32_C( 946.85), SIMDE_FLOAT32_C( -866.96), SIMDE_FLOAT32_C( 383.68)),
1764
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 82.96), SIMDE_FLOAT32_C( 541.21), SIMDE_FLOAT32_C( 296.52), SIMDE_FLOAT32_C( 677.59)),
1765
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 932.95), SIMDE_FLOAT32_C( 946.85), SIMDE_FLOAT32_C( -866.96), SIMDE_FLOAT32_C( 677.00)) },
1766
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( -824.32), SIMDE_FLOAT32_C( -924.44), SIMDE_FLOAT32_C( -44.45), SIMDE_FLOAT32_C( 340.62)),
1767
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 771.85), SIMDE_FLOAT32_C( -635.91), SIMDE_FLOAT32_C( 706.67), SIMDE_FLOAT32_C( 477.32)),
1768
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -824.32), SIMDE_FLOAT32_C( -924.44), SIMDE_FLOAT32_C( -44.45), SIMDE_FLOAT32_C( 477.00)) },
1769
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( 991.94), SIMDE_FLOAT32_C( 918.28), SIMDE_FLOAT32_C( -617.36), SIMDE_FLOAT32_C( 932.02)),
1770
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -21.99), SIMDE_FLOAT32_C( 274.28), SIMDE_FLOAT32_C( 819.55), SIMDE_FLOAT32_C( -957.22)),
1771
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 991.94), SIMDE_FLOAT32_C( 918.28), SIMDE_FLOAT32_C( -617.36), SIMDE_FLOAT32_C( -958.00)) },
1772
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( 178.34), SIMDE_FLOAT32_C( -961.13), SIMDE_FLOAT32_C( 949.60), SIMDE_FLOAT32_C( -711.62)),
1773
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -951.61), SIMDE_FLOAT32_C( 53.93), SIMDE_FLOAT32_C( 376.79), SIMDE_FLOAT32_C( -892.19)),
1774
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 178.34), SIMDE_FLOAT32_C( -961.13), SIMDE_FLOAT32_C( 949.60), SIMDE_FLOAT32_C( -893.00)) },
1775
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( -858.80), SIMDE_FLOAT32_C( 721.75), SIMDE_FLOAT32_C( 422.70), SIMDE_FLOAT32_C( -573.55)),
1776
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 473.81), SIMDE_FLOAT32_C( -473.48), SIMDE_FLOAT32_C( 815.69), SIMDE_FLOAT32_C( 520.75)),
1777
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -858.80), SIMDE_FLOAT32_C( 721.75), SIMDE_FLOAT32_C( 422.70), SIMDE_FLOAT32_C( 520.00)) },
1778
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( -849.38), SIMDE_FLOAT32_C( 6.45), SIMDE_FLOAT32_C( -927.45), SIMDE_FLOAT32_C( -143.78)),
1779
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 593.59), SIMDE_FLOAT32_C( 314.52), SIMDE_FLOAT32_C( 115.80), SIMDE_FLOAT32_C( 369.25)),
1780
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -849.38), SIMDE_FLOAT32_C( 6.45), SIMDE_FLOAT32_C( -927.45), SIMDE_FLOAT32_C( 369.00)) },
1781
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( -22.36), SIMDE_FLOAT32_C( 891.64), SIMDE_FLOAT32_C( 742.10), SIMDE_FLOAT32_C( 13.68)),
1782
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 900.76), SIMDE_FLOAT32_C( -733.09), SIMDE_FLOAT32_C( -137.10), SIMDE_FLOAT32_C( -90.09)),
1783
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -22.36), SIMDE_FLOAT32_C( 891.64), SIMDE_FLOAT32_C( 742.10), SIMDE_FLOAT32_C( -91.00)) },
1784
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( -977.22), SIMDE_FLOAT32_C( -98.98), SIMDE_FLOAT32_C( -253.27), SIMDE_FLOAT32_C( 664.01)),
1785
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 252.97), SIMDE_FLOAT32_C( 266.57), SIMDE_FLOAT32_C( -16.11), SIMDE_FLOAT32_C( -404.33)),
1786
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -977.22), SIMDE_FLOAT32_C( -98.98), SIMDE_FLOAT32_C( -253.27), SIMDE_FLOAT32_C( -405.00)) }
1787
+ };
1788
+
1789
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
1790
+ simde__m128 r = simde_mm_floor_ss(test_vec[i].a, test_vec[i].b);
1791
+ simde_assert_m128_close(r, test_vec[i].r, 1);
1792
+ }
1793
+
1794
+ return MUNIT_OK;
1795
+ }
1796
+
1797
+ static MunitResult
1798
+ test_simde_mm_insert_epi8(const MunitParameter params[], void* data) {
1799
+ (void) params;
1800
+ (void) data;
1801
+
1802
+ const struct {
1803
+ simde__m128i a;
1804
+ int b;
1805
+ simde__m128i r;
1806
+ } test_vec[8] = {
1807
+ { simde_mm_set_epi8(INT8_C( 10), INT8_C( -18), INT8_C(-117), INT8_C( 95),
1808
+ INT8_C(-108), INT8_C( 57), INT8_C( 54), INT8_C( -50),
1809
+ INT8_C( 60), INT8_C( -68), INT8_C( 115), INT8_C( 14),
1810
+ INT8_C( -71), INT8_C( -40), INT8_C( 34), INT8_C( 112)),
1811
+ 6,
1812
+ simde_mm_set_epi8(INT8_C( 10), INT8_C( -18), INT8_C(-117), INT8_C( 6),
1813
+ INT8_C(-108), INT8_C( 57), INT8_C( 54), INT8_C( -50),
1814
+ INT8_C( 60), INT8_C( -68), INT8_C( 115), INT8_C( 14),
1815
+ INT8_C( -71), INT8_C( -40), INT8_C( 34), INT8_C( 112)) },
1816
+ { simde_mm_set_epi8(INT8_C(-117), INT8_C( 67), INT8_C(-105), INT8_C( 26),
1817
+ INT8_C( 118), INT8_C( 67), INT8_C( 72), INT8_C( 120),
1818
+ INT8_C( -74), INT8_C( -8), INT8_C(-126), INT8_C( -39),
1819
+ INT8_C( 53), INT8_C( 74), INT8_C( 64), INT8_C( -74)),
1820
+ 13,
1821
+ simde_mm_set_epi8(INT8_C(-117), INT8_C( 67), INT8_C(-105), INT8_C( 13),
1822
+ INT8_C( 118), INT8_C( 67), INT8_C( 72), INT8_C( 120),
1823
+ INT8_C( -74), INT8_C( -8), INT8_C(-126), INT8_C( -39),
1824
+ INT8_C( 53), INT8_C( 74), INT8_C( 64), INT8_C( -74)) },
1825
+ { simde_mm_set_epi8(INT8_C( 10), INT8_C(-114), INT8_C( 116), INT8_C(-108),
1826
+ INT8_C( 36), INT8_C( -74), INT8_C( 76), INT8_C( -2),
1827
+ INT8_C( 118), INT8_C( -60), INT8_C( 3), INT8_C( 48),
1828
+ INT8_C( 76), INT8_C( -18), INT8_C( -69), INT8_C(-110)),
1829
+ 13,
1830
+ simde_mm_set_epi8(INT8_C( 10), INT8_C(-114), INT8_C( 116), INT8_C( 13),
1831
+ INT8_C( 36), INT8_C( -74), INT8_C( 76), INT8_C( -2),
1832
+ INT8_C( 118), INT8_C( -60), INT8_C( 3), INT8_C( 48),
1833
+ INT8_C( 76), INT8_C( -18), INT8_C( -69), INT8_C(-110)) },
1834
+ { simde_mm_set_epi8(INT8_C( -48), INT8_C( -33), INT8_C( -9), INT8_C( -27),
1835
+ INT8_C( -99), INT8_C( 29), INT8_C( 111), INT8_C( 85),
1836
+ INT8_C( 104), INT8_C( -59), INT8_C( -17), INT8_C( 23),
1837
+ INT8_C( 113), INT8_C( 8), INT8_C( 94), INT8_C( -32)),
1838
+ 0,
1839
+ simde_mm_set_epi8(INT8_C( -48), INT8_C( -33), INT8_C( -9), INT8_C( 0),
1840
+ INT8_C( -99), INT8_C( 29), INT8_C( 111), INT8_C( 85),
1841
+ INT8_C( 104), INT8_C( -59), INT8_C( -17), INT8_C( 23),
1842
+ INT8_C( 113), INT8_C( 8), INT8_C( 94), INT8_C( -32)) },
1843
+ { simde_mm_set_epi8(INT8_C( -15), INT8_C(-126), INT8_C( 63), INT8_C( 14),
1844
+ INT8_C( 2), INT8_C( 88), INT8_C( -73), INT8_C( 22),
1845
+ INT8_C( -36), INT8_C( 60), INT8_C( -75), INT8_C( -26),
1846
+ INT8_C( -12), INT8_C(-126), INT8_C( 44), INT8_C( -36)),
1847
+ 8,
1848
+ simde_mm_set_epi8(INT8_C( -15), INT8_C(-126), INT8_C( 63), INT8_C( 8),
1849
+ INT8_C( 2), INT8_C( 88), INT8_C( -73), INT8_C( 22),
1850
+ INT8_C( -36), INT8_C( 60), INT8_C( -75), INT8_C( -26),
1851
+ INT8_C( -12), INT8_C(-126), INT8_C( 44), INT8_C( -36)) },
1852
+ { simde_mm_set_epi8(INT8_C( 41), INT8_C(-102), INT8_C( 115), INT8_C( 65),
1853
+ INT8_C( 22), INT8_C( -32), INT8_C( -29), INT8_C( 85),
1854
+ INT8_C( -61), INT8_C( 79), INT8_C(-107), INT8_C( -9),
1855
+ INT8_C( 89), INT8_C( -3), INT8_C(-124), INT8_C( 71)),
1856
+ 4,
1857
+ simde_mm_set_epi8(INT8_C( 41), INT8_C(-102), INT8_C( 115), INT8_C( 4),
1858
+ INT8_C( 22), INT8_C( -32), INT8_C( -29), INT8_C( 85),
1859
+ INT8_C( -61), INT8_C( 79), INT8_C(-107), INT8_C( -9),
1860
+ INT8_C( 89), INT8_C( -3), INT8_C(-124), INT8_C( 71)) },
1861
+ { simde_mm_set_epi8(INT8_C( 109), INT8_C( -46), INT8_C( -71), INT8_C( 107),
1862
+ INT8_C( 105), INT8_C( 123), INT8_C( -72), INT8_C( 72),
1863
+ INT8_C( -43), INT8_C( 111), INT8_C( 75), INT8_C( 60),
1864
+ INT8_C( 47), INT8_C( 78), INT8_C( -45), INT8_C(-101)),
1865
+ 7,
1866
+ simde_mm_set_epi8(INT8_C( 109), INT8_C( -46), INT8_C( -71), INT8_C( 7),
1867
+ INT8_C( 105), INT8_C( 123), INT8_C( -72), INT8_C( 72),
1868
+ INT8_C( -43), INT8_C( 111), INT8_C( 75), INT8_C( 60),
1869
+ INT8_C( 47), INT8_C( 78), INT8_C( -45), INT8_C(-101)) },
1870
+ { simde_mm_set_epi8(INT8_C( -15), INT8_C( -23), INT8_C( 43), INT8_C( 15),
1871
+ INT8_C(-126), INT8_C( 102), INT8_C( 109), INT8_C( 84),
1872
+ INT8_C( -63), INT8_C( 80), INT8_C(-127), INT8_C( -46),
1873
+ INT8_C(-119), INT8_C( 111), INT8_C( -57), INT8_C( 55)),
1874
+ 15,
1875
+ simde_mm_set_epi8(INT8_C( -15), INT8_C( -23), INT8_C( 43), INT8_C( 15),
1876
+ INT8_C(-126), INT8_C( 102), INT8_C( 109), INT8_C( 84),
1877
+ INT8_C( -63), INT8_C( 80), INT8_C(-127), INT8_C( -46),
1878
+ INT8_C(-119), INT8_C( 111), INT8_C( -57), INT8_C( 55)) }
1879
+ };
1880
+
1881
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
1882
+ simde__m128i r = simde_mm_insert_epi8(test_vec[i].a, test_vec[i].b, 12);
1883
+ simde_assert_m128i_i8(r, ==, test_vec[i].r);
1884
+ }
1885
+
1886
+ return MUNIT_OK;
1887
+ }
1888
+
1889
+ static MunitResult
1890
+ test_simde_mm_insert_epi32(const MunitParameter params[], void* data) {
1891
+ (void) params;
1892
+ (void) data;
1893
+
1894
+ const struct {
1895
+ simde__m128i a;
1896
+ int32_t b;
1897
+ simde__m128i r;
1898
+ } test_vec[8] = {
1899
+ { simde_mm_set_epi32(INT32_C( 1714139829), INT32_C( 980372706), INT32_C(-2140490700), INT32_C( 1671335242)),
1900
+ INT32_C( -570572581),
1901
+ simde_mm_set_epi32(INT32_C( -570572581), INT32_C( 980372706), INT32_C(-2140490700), INT32_C( 1671335242)) },
1902
+ { simde_mm_set_epi32(INT32_C( -385110226), INT32_C( 1048550831), INT32_C( -768686987), INT32_C( -197993270)),
1903
+ INT32_C( 292040766),
1904
+ simde_mm_set_epi32(INT32_C( 292040766), INT32_C( 1048550831), INT32_C( -768686987), INT32_C( -197993270)) },
1905
+ { simde_mm_set_epi32(INT32_C( 1804855856), INT32_C( -940421248), INT32_C(-2104492113), INT32_C( 529461525)),
1906
+ INT32_C( 514665012),
1907
+ simde_mm_set_epi32(INT32_C( 514665012), INT32_C( -940421248), INT32_C(-2104492113), INT32_C( 529461525)) },
1908
+ { simde_mm_set_epi32(INT32_C( -85551175), INT32_C( 885869319), INT32_C( 2003605165), INT32_C( -219227805)),
1909
+ INT32_C( 1214139778),
1910
+ simde_mm_set_epi32(INT32_C( 1214139778), INT32_C( 885869319), INT32_C( 2003605165), INT32_C( -219227805)) },
1911
+ { simde_mm_set_epi32(INT32_C( -682699624), INT32_C( -729535744), INT32_C( -714850455), INT32_C( 418841580)),
1912
+ INT32_C( -606767281),
1913
+ simde_mm_set_epi32(INT32_C( -606767281), INT32_C( -729535744), INT32_C( -714850455), INT32_C( 418841580)) },
1914
+ { simde_mm_set_epi32(INT32_C( 1844877442), INT32_C(-1305620543), INT32_C( -803345723), INT32_C(-1617611400)),
1915
+ INT32_C( 166894477),
1916
+ simde_mm_set_epi32(INT32_C( 166894477), INT32_C(-1305620543), INT32_C( -803345723), INT32_C(-1617611400)) },
1917
+ { simde_mm_set_epi32(INT32_C( 1121577587), INT32_C( 1868325196), INT32_C( 567002451), INT32_C(-2029289287)),
1918
+ INT32_C( 566028624),
1919
+ simde_mm_set_epi32(INT32_C( 566028624), INT32_C( 1868325196), INT32_C( 567002451), INT32_C(-2029289287)) },
1920
+ { simde_mm_set_epi32(INT32_C( 1745932520), INT32_C( 83968460), INT32_C(-1677360074), INT32_C( -664327669)),
1921
+ INT32_C(-2041481189),
1922
+ simde_mm_set_epi32(INT32_C(-2041481189), INT32_C( 83968460), INT32_C(-1677360074), INT32_C( -664327669)) }
1923
+ };
1924
+
1925
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
1926
+ simde__m128i r = simde_mm_insert_epi32(test_vec[i].a, test_vec[i].b, 3);
1927
+ simde_assert_m128i_i32(r, ==, test_vec[i].r);
1928
+ }
1929
+
1930
+ return MUNIT_OK;
1931
+ }
1932
+
1933
+ static MunitResult
1934
+ test_simde_mm_insert_epi64(const MunitParameter params[], void* data) {
1935
+ (void) params;
1936
+ (void) data;
1937
+
1938
+ const struct {
1939
+ simde__m128i a;
1940
+ int64_t b;
1941
+ simde__m128i r;
1942
+ } test_vec[8] = {
1943
+ { simde_mm_set_epi64x(INT64_C( 7362174507306405090), INT64_C(-9193337552220811958)),
1944
+ INT64_C( -850374615753703205),
1945
+ simde_mm_set_epi64x(INT64_C( -850374615753703205), INT64_C(-9193337552220811958)) },
1946
+ { simde_mm_set_epi64x(INT64_C( 1254305542978645806), INT64_C( 4503491530864903285)),
1947
+ INT64_C(-9038724799495474923),
1948
+ simde_mm_set_epi64x(INT64_C(-9038724799495474923), INT64_C( 4503491530864903285)) },
1949
+ { simde_mm_set_epi64x(INT64_C( -941576252334200268), INT64_C( 7751796878868631424)),
1950
+ INT64_C( 3804779755638396589),
1951
+ simde_mm_set_epi64x(INT64_C( 3804779755638396589), INT64_C( 7751796878868631424)) },
1952
+ { simde_mm_set_epi64x(INT64_C(-3070259325336878100), INT64_C( 5214690643492116409)),
1953
+ INT64_C(-2932172554506065152),
1954
+ simde_mm_set_epi64x(INT64_C(-2932172554506065152), INT64_C( 5214690643492116409)) },
1955
+ { simde_mm_set_epi64x(INT64_C(-5607597529679140155), INT64_C(-6947588056948574385)),
1956
+ INT64_C( 716806322442901634),
1957
+ simde_mm_set_epi64x(INT64_C( 716806322442901634), INT64_C(-6947588056948574385)) },
1958
+ { simde_mm_set_epi64x(INT64_C( 4817139057959919948), INT64_C( 2435256986062520505)),
1959
+ INT64_C(-2853265611616884400),
1960
+ simde_mm_set_epi64x(INT64_C(-2853265611616884400), INT64_C( 2435256986062520505)) },
1961
+ { simde_mm_set_epi64x(INT64_C(-8768094940408262424), INT64_C( 360641792213091382)),
1962
+ INT64_C(-4200371864439390341),
1963
+ simde_mm_set_epi64x(INT64_C(-4200371864439390341), INT64_C( 360641792213091382)) },
1964
+ { simde_mm_set_epi64x(INT64_C(-6597763380374804743), INT64_C( 8897544020142746680)),
1965
+ INT64_C( 6543676199726157119),
1966
+ simde_mm_set_epi64x(INT64_C( 6543676199726157119), INT64_C( 8897544020142746680)) }
1967
+ };
1968
+
1969
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
1970
+ simde__m128i r = simde_mm_insert_epi64(test_vec[i].a, test_vec[i].b, 1);
1971
+ simde_assert_m128i_i64(r, ==, test_vec[i].r);
1972
+ }
1973
+
1974
+ return MUNIT_OK;
1975
+ }
1976
+
1977
+ static MunitResult
1978
+ test_simde_mm_insert_ps(const MunitParameter params[], void* data) {
1979
+ (void) params;
1980
+ (void) data;
1981
+
1982
+ const struct {
1983
+ simde__m128 a;
1984
+ simde__m128 b;
1985
+ simde__m128 r;
1986
+ } test_vec[8] = {
1987
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( -841.30), SIMDE_FLOAT32_C( 235.25), SIMDE_FLOAT32_C( -673.83), SIMDE_FLOAT32_C( -700.28)),
1988
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 92.06), SIMDE_FLOAT32_C( 735.95), SIMDE_FLOAT32_C( 171.76), SIMDE_FLOAT32_C( 644.60)),
1989
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -841.30), SIMDE_FLOAT32_C( 235.25), SIMDE_FLOAT32_C( 0.00), SIMDE_FLOAT32_C( 0.00)) },
1990
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( 658.47), SIMDE_FLOAT32_C( -385.89), SIMDE_FLOAT32_C( 916.54), SIMDE_FLOAT32_C( -937.06)),
1991
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -209.24), SIMDE_FLOAT32_C( -655.00), SIMDE_FLOAT32_C( -722.53), SIMDE_FLOAT32_C( -502.41)),
1992
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 658.47), SIMDE_FLOAT32_C( -385.89), SIMDE_FLOAT32_C( 0.00), SIMDE_FLOAT32_C( 0.00)) },
1993
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( 364.58), SIMDE_FLOAT32_C( -402.45), SIMDE_FLOAT32_C( -252.76), SIMDE_FLOAT32_C( 525.23)),
1994
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -670.47), SIMDE_FLOAT32_C( 704.70), SIMDE_FLOAT32_C( -13.52), SIMDE_FLOAT32_C( -328.78)),
1995
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 364.58), SIMDE_FLOAT32_C( -402.45), SIMDE_FLOAT32_C( 0.00), SIMDE_FLOAT32_C( 0.00)) },
1996
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( 313.55), SIMDE_FLOAT32_C( 334.89), SIMDE_FLOAT32_C( -877.87), SIMDE_FLOAT32_C( -974.35)),
1997
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 708.89), SIMDE_FLOAT32_C( 60.29), SIMDE_FLOAT32_C( 416.58), SIMDE_FLOAT32_C( -180.37)),
1998
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 313.55), SIMDE_FLOAT32_C( 334.89), SIMDE_FLOAT32_C( 0.00), SIMDE_FLOAT32_C( 0.00)) },
1999
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( 857.51), SIMDE_FLOAT32_C( 674.21), SIMDE_FLOAT32_C( 882.64), SIMDE_FLOAT32_C( 238.47)),
2000
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 556.45), SIMDE_FLOAT32_C( -280.75), SIMDE_FLOAT32_C( -829.81), SIMDE_FLOAT32_C( 113.59)),
2001
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 857.51), SIMDE_FLOAT32_C( 674.21), SIMDE_FLOAT32_C( 0.00), SIMDE_FLOAT32_C( 0.00)) },
2002
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( 520.41), SIMDE_FLOAT32_C( -251.34), SIMDE_FLOAT32_C( 896.76), SIMDE_FLOAT32_C( -419.46)),
2003
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 87.20), SIMDE_FLOAT32_C( 358.54), SIMDE_FLOAT32_C( 859.02), SIMDE_FLOAT32_C( -278.34)),
2004
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 520.41), SIMDE_FLOAT32_C( -251.34), SIMDE_FLOAT32_C( 0.00), SIMDE_FLOAT32_C( 0.00)) },
2005
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( 533.51), SIMDE_FLOAT32_C( 968.42), SIMDE_FLOAT32_C( -332.89), SIMDE_FLOAT32_C( 137.03)),
2006
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -277.45), SIMDE_FLOAT32_C( 701.40), SIMDE_FLOAT32_C( -500.07), SIMDE_FLOAT32_C( -788.33)),
2007
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 533.51), SIMDE_FLOAT32_C( 968.42), SIMDE_FLOAT32_C( 0.00), SIMDE_FLOAT32_C( 0.00)) },
2008
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( -468.51), SIMDE_FLOAT32_C( -644.45), SIMDE_FLOAT32_C( -325.06), SIMDE_FLOAT32_C( 994.31)),
2009
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 290.72), SIMDE_FLOAT32_C( 585.35), SIMDE_FLOAT32_C( 669.48), SIMDE_FLOAT32_C( 65.07)),
2010
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -468.51), SIMDE_FLOAT32_C( -644.45), SIMDE_FLOAT32_C( 0.00), SIMDE_FLOAT32_C( 0.00)) }
2011
+ };
2012
+
2013
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
2014
+ simde__m128 r = simde_mm_insert_ps(test_vec[i].a, test_vec[i].b, 3);
2015
+ simde_assert_m128_close(r, test_vec[i].r, 1);
2016
+ }
2017
+
2018
+ return MUNIT_OK;
2019
+ }
2020
+
2021
+ static MunitResult
2022
+ test_simde_mm_max_epi8(const MunitParameter params[], void* data) {
2023
+ (void) params;
2024
+ (void) data;
2025
+
2026
+ const struct {
2027
+ simde__m128i a;
2028
+ simde__m128i b;
2029
+ simde__m128i r;
2030
+ } test_vec[8] = {
2031
+ { simde_mm_set_epi8(INT8_C( -83), INT8_C( 29), INT8_C( 3), INT8_C( 84), INT8_C( 22), INT8_C( -7), INT8_C( 112), INT8_C(-107),
2032
+ INT8_C( -53), INT8_C(-125), INT8_C( -41), INT8_C( -27), INT8_C(-118), INT8_C( 43), INT8_C( 44), INT8_C( 16)),
2033
+ simde_mm_set_epi8(INT8_C( -80), INT8_C( -52), INT8_C( 93), INT8_C( 71), INT8_C( 45), INT8_C( -99), INT8_C( -14), INT8_C( -97),
2034
+ INT8_C( 11), INT8_C( 62), INT8_C( 111), INT8_C( 17), INT8_C(-108), INT8_C( -6), INT8_C( 95), INT8_C( 1)),
2035
+ simde_mm_set_epi8(INT8_C( -80), INT8_C( 29), INT8_C( 93), INT8_C( 84), INT8_C( 45), INT8_C( -7), INT8_C( 112), INT8_C( -97),
2036
+ INT8_C( 11), INT8_C( 62), INT8_C( 111), INT8_C( 17), INT8_C(-108), INT8_C( 43), INT8_C( 95), INT8_C( 16)) },
2037
+ { simde_mm_set_epi8(INT8_C( 61), INT8_C(-100), INT8_C( 79), INT8_C( 19), INT8_C( -86), INT8_C( 116), INT8_C( -34), INT8_C( 96),
2038
+ INT8_C( 70), INT8_C( -48), INT8_C( -13), INT8_C( 103), INT8_C(-107), INT8_C( 79), INT8_C( 24), INT8_C( -71)),
2039
+ simde_mm_set_epi8(INT8_C( 39), INT8_C( -53), INT8_C(-117), INT8_C( 65), INT8_C( -55), INT8_C( 21), INT8_C( 49), INT8_C( -22),
2040
+ INT8_C( 118), INT8_C( -71), INT8_C(-113), INT8_C( 2), INT8_C( -83), INT8_C( -74), INT8_C( -11), INT8_C( 85)),
2041
+ simde_mm_set_epi8(INT8_C( 61), INT8_C( -53), INT8_C( 79), INT8_C( 65), INT8_C( -55), INT8_C( 116), INT8_C( 49), INT8_C( 96),
2042
+ INT8_C( 118), INT8_C( -48), INT8_C( -13), INT8_C( 103), INT8_C( -83), INT8_C( 79), INT8_C( 24), INT8_C( 85)) },
2043
+ { simde_mm_set_epi8(INT8_C( -13), INT8_C(-124), INT8_C(-110), INT8_C( 22), INT8_C(-109), INT8_C( 127), INT8_C( -45), INT8_C( -62),
2044
+ INT8_C( 0), INT8_C( 51), INT8_C( -98), INT8_C( -11), INT8_C( -54), INT8_C( -75), INT8_C(-101), INT8_C(-109)),
2045
+ simde_mm_set_epi8(INT8_C( -28), INT8_C(-120), INT8_C( 90), INT8_C( -89), INT8_C( -72), INT8_C( 93), INT8_C( -68), INT8_C(-105),
2046
+ INT8_C( 39), INT8_C( 46), INT8_C( 2), INT8_C( 5), INT8_C( 16), INT8_C( 34), INT8_C( -99), INT8_C( 64)),
2047
+ simde_mm_set_epi8(INT8_C( -13), INT8_C(-120), INT8_C( 90), INT8_C( 22), INT8_C( -72), INT8_C( 127), INT8_C( -45), INT8_C( -62),
2048
+ INT8_C( 39), INT8_C( 51), INT8_C( 2), INT8_C( 5), INT8_C( 16), INT8_C( 34), INT8_C( -99), INT8_C( 64)) },
2049
+ { simde_mm_set_epi8(INT8_C( -43), INT8_C(-102), INT8_C( 111), INT8_C( 97), INT8_C( -69), INT8_C( 104), INT8_C( -24), INT8_C( -6),
2050
+ INT8_C( 29), INT8_C( 43), INT8_C( 65), INT8_C( 47), INT8_C( -60), INT8_C( 52), INT8_C( 2), INT8_C( -40)),
2051
+ simde_mm_set_epi8(INT8_C( -7), INT8_C( 85), INT8_C( -63), INT8_C( 48), INT8_C(-117), INT8_C( 19), INT8_C( 99), INT8_C( 67),
2052
+ INT8_C( -50), INT8_C(-116), INT8_C( 5), INT8_C( 2), INT8_C( -96), INT8_C( 125), INT8_C( -95), INT8_C( 0)),
2053
+ simde_mm_set_epi8(INT8_C( -7), INT8_C( 85), INT8_C( 111), INT8_C( 97), INT8_C( -69), INT8_C( 104), INT8_C( 99), INT8_C( 67),
2054
+ INT8_C( 29), INT8_C( 43), INT8_C( 65), INT8_C( 47), INT8_C( -60), INT8_C( 125), INT8_C( 2), INT8_C( 0)) },
2055
+ { simde_mm_set_epi8(INT8_C( 98), INT8_C( -7), INT8_C( -56), INT8_C( 9), INT8_C( 74), INT8_C( 50), INT8_C( -95), INT8_C( 127),
2056
+ INT8_C( 114), INT8_C( 118), INT8_C( -76), INT8_C( 110), INT8_C( -48), INT8_C( -50), INT8_C( 57), INT8_C( -47)),
2057
+ simde_mm_set_epi8(INT8_C( 47), INT8_C( -82), INT8_C( 66), INT8_C( -76), INT8_C( 6), INT8_C( 117), INT8_C( 80), INT8_C( -13),
2058
+ INT8_C( 80), INT8_C(-109), INT8_C( 126), INT8_C( 70), INT8_C( 60), INT8_C( 40), INT8_C( -54), INT8_C( 3)),
2059
+ simde_mm_set_epi8(INT8_C( 98), INT8_C( -7), INT8_C( 66), INT8_C( 9), INT8_C( 74), INT8_C( 117), INT8_C( 80), INT8_C( 127),
2060
+ INT8_C( 114), INT8_C( 118), INT8_C( 126), INT8_C( 110), INT8_C( 60), INT8_C( 40), INT8_C( 57), INT8_C( 3)) },
2061
+ { simde_mm_set_epi8(INT8_C( -80), INT8_C( -39), INT8_C( 113), INT8_C( -58), INT8_C( 58), INT8_C( -88), INT8_C( 22), INT8_C( -99),
2062
+ INT8_C( 112), INT8_C( -21), INT8_C( 44), INT8_C( -31), INT8_C( -2), INT8_C( 109), INT8_C( 19), INT8_C( -74)),
2063
+ simde_mm_set_epi8(INT8_C( 71), INT8_C( -38), INT8_C( 47), INT8_C( -92), INT8_C( 8), INT8_C( 48), INT8_C( 105), INT8_C( -82),
2064
+ INT8_C( 35), INT8_C( 54), INT8_C( 25), INT8_C( 36), INT8_C( 111), INT8_C( -7), INT8_C( -12), INT8_C( -75)),
2065
+ simde_mm_set_epi8(INT8_C( 71), INT8_C( -38), INT8_C( 113), INT8_C( -58), INT8_C( 58), INT8_C( 48), INT8_C( 105), INT8_C( -82),
2066
+ INT8_C( 112), INT8_C( 54), INT8_C( 44), INT8_C( 36), INT8_C( 111), INT8_C( 109), INT8_C( 19), INT8_C( -74)) },
2067
+ { simde_mm_set_epi8(INT8_C( 18), INT8_C( -41), INT8_C( -74), INT8_C( 60), INT8_C( -63), INT8_C( -55), INT8_C( -71), INT8_C( 118),
2068
+ INT8_C( 46), INT8_C( -19), INT8_C( 19), INT8_C( -7), INT8_C( -55), INT8_C( -75), INT8_C( 125), INT8_C( -82)),
2069
+ simde_mm_set_epi8(INT8_C(-125), INT8_C(-106), INT8_C(-127), INT8_C( -23), INT8_C( -63), INT8_C( 0), INT8_C( 56), INT8_C( -23),
2070
+ INT8_C( -34), INT8_C( -32), INT8_C( 33), INT8_C( 127), INT8_C( 39), INT8_C( -68), INT8_C(-103), INT8_C( 126)),
2071
+ simde_mm_set_epi8(INT8_C( 18), INT8_C( -41), INT8_C( -74), INT8_C( 60), INT8_C( -63), INT8_C( 0), INT8_C( 56), INT8_C( 118),
2072
+ INT8_C( 46), INT8_C( -19), INT8_C( 33), INT8_C( 127), INT8_C( 39), INT8_C( -68), INT8_C( 125), INT8_C( 126)) },
2073
+ { simde_mm_set_epi8(INT8_C( -90), INT8_C( -81), INT8_C( 72), INT8_C( -34), INT8_C( -85), INT8_C( 108), INT8_C( 42), INT8_C( 34),
2074
+ INT8_C( -79), INT8_C( -48), INT8_C( 3), INT8_C( 113), INT8_C( -62), INT8_C( 25), INT8_C( -74), INT8_C( 13)),
2075
+ simde_mm_set_epi8(INT8_C( -44), INT8_C( -22), INT8_C( 26), INT8_C( -52), INT8_C( 29), INT8_C( 96), INT8_C( -97), INT8_C( 88),
2076
+ INT8_C( -11), INT8_C(-125), INT8_C( -60), INT8_C( 88), INT8_C( -55), INT8_C( 40), INT8_C( 76), INT8_C( -80)),
2077
+ simde_mm_set_epi8(INT8_C( -44), INT8_C( -22), INT8_C( 72), INT8_C( -34), INT8_C( 29), INT8_C( 108), INT8_C( 42), INT8_C( 88),
2078
+ INT8_C( -11), INT8_C( -48), INT8_C( 3), INT8_C( 113), INT8_C( -55), INT8_C( 40), INT8_C( 76), INT8_C( 13)) }
2079
+ };
2080
+
2081
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
2082
+ simde__m128i r = simde_mm_max_epi8(test_vec[i].a, test_vec[i].b);
2083
+ simde_assert_m128i_i8(r, ==, test_vec[i].r);
2084
+ }
2085
+
2086
+ return MUNIT_OK;
2087
+ }
2088
+
2089
+ static MunitResult
2090
+ test_simde_mm_max_epi32(const MunitParameter params[], void* data) {
2091
+ (void) params;
2092
+ (void) data;
2093
+
2094
+ const struct {
2095
+ simde__m128i a;
2096
+ simde__m128i b;
2097
+ simde__m128i r;
2098
+ } test_vec[8] = {
2099
+ { simde_mm_set_epi32(INT32_C( 477039853), INT32_C( 1555162887), INT32_C( 960008067), INT32_C(-2078514865)),
2100
+ simde_mm_set_epi32(INT32_C( -346254872), INT32_C( 70690131), INT32_C( -249580279), INT32_C( -297974983)),
2101
+ simde_mm_set_epi32(INT32_C( 477039853), INT32_C( 1555162887), INT32_C( 960008067), INT32_C( -297974983)) },
2102
+ { simde_mm_set_epi32(INT32_C( 1946076095), INT32_C( 1626099315), INT32_C( 367020053), INT32_C( -768900730)),
2103
+ simde_mm_set_epi32(INT32_C( -534331032), INT32_C(-1542646073), INT32_C( 900332023), INT32_C( 1625404528)),
2104
+ simde_mm_set_epi32(INT32_C( 1946076095), INT32_C( 1626099315), INT32_C( 900332023), INT32_C( 1625404528)) },
2105
+ { simde_mm_set_epi32(INT32_C( 1785938855), INT32_C( 217054136), INT32_C( 6277613), INT32_C( 1323071612)),
2106
+ simde_mm_set_epi32(INT32_C(-1842447194), INT32_C( 1202045241), INT32_C( 391997517), INT32_C( 904511960)),
2107
+ simde_mm_set_epi32(INT32_C( 1785938855), INT32_C( 1202045241), INT32_C( 391997517), INT32_C( 1323071612)) },
2108
+ { simde_mm_set_epi32(INT32_C( -303416903), INT32_C( 131247622), INT32_C(-2038528381), INT32_C(-1774779832)),
2109
+ simde_mm_set_epi32(INT32_C( 307942484), INT32_C(-1408988466), INT32_C(-1799017519), INT32_C( 1885804693)),
2110
+ simde_mm_set_epi32(INT32_C( 307942484), INT32_C( 131247622), INT32_C(-1799017519), INT32_C( 1885804693)) },
2111
+ { simde_mm_set_epi32(INT32_C( -424176665), INT32_C( 531131808), INT32_C( -752351935), INT32_C( 650848232)),
2112
+ simde_mm_set_epi32(INT32_C(-1513144536), INT32_C( -875213417), INT32_C(-1592044131), INT32_C(-1714853090)),
2113
+ simde_mm_set_epi32(INT32_C( -424176665), INT32_C( 531131808), INT32_C( -752351935), INT32_C( 650848232)) },
2114
+ { simde_mm_set_epi32(INT32_C( 1901471013), INT32_C(-1928782687), INT32_C(-1204237637), INT32_C(-1735607658)),
2115
+ simde_mm_set_epi32(INT32_C( 1526238167), INT32_C( 1155872155), INT32_C( -981128645), INT32_C(-1655952235)),
2116
+ simde_mm_set_epi32(INT32_C( 1901471013), INT32_C( 1155872155), INT32_C( -981128645), INT32_C(-1655952235)) },
2117
+ { simde_mm_set_epi32(INT32_C( 2067002242), INT32_C( -967820279), INT32_C(-1842332164), INT32_C( 1466140749)),
2118
+ simde_mm_set_epi32(INT32_C( -25857443), INT32_C(-1417095277), INT32_C( -31302876), INT32_C(-1114441571)),
2119
+ simde_mm_set_epi32(INT32_C( 2067002242), INT32_C( -967820279), INT32_C( -31302876), INT32_C( 1466140749)) },
2120
+ { simde_mm_set_epi32(INT32_C( -515864750), INT32_C( 1885657427), INT32_C(-1627411585), INT32_C( -455970760)),
2121
+ simde_mm_set_epi32(INT32_C( -505553594), INT32_C(-1363391302), INT32_C( 1929235496), INT32_C(-1212934452)),
2122
+ simde_mm_set_epi32(INT32_C( -505553594), INT32_C( 1885657427), INT32_C( 1929235496), INT32_C( -455970760)) }
2123
+ };
2124
+
2125
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
2126
+ simde__m128i r = simde_mm_max_epi32(test_vec[i].a, test_vec[i].b);
2127
+ simde_assert_m128i_i32(r, ==, test_vec[i].r);
2128
+ }
2129
+
2130
+ return MUNIT_OK;
2131
+ }
2132
+
2133
+ static MunitResult
2134
+ test_simde_mm_max_epu16(const MunitParameter params[], void* data) {
2135
+ (void) params;
2136
+ (void) data;
2137
+
2138
+ const struct {
2139
+ simde__m128i a;
2140
+ simde__m128i b;
2141
+ simde__m128i r;
2142
+ } test_vec[8] = {
2143
+ { simde_x_mm_set_epu16(UINT16_C(31378), UINT16_C(24380), UINT16_C(40255), UINT16_C(13201),
2144
+ UINT16_C(58870), UINT16_C(49460), UINT16_C(64732), UINT16_C(36111)),
2145
+ simde_x_mm_set_epu16(UINT16_C(64055), UINT16_C(45511), UINT16_C(61004), UINT16_C(49058),
2146
+ UINT16_C(25195), UINT16_C(65082), UINT16_C( 861), UINT16_C( 2148)),
2147
+ simde_x_mm_set_epu16(UINT16_C(64055), UINT16_C(45511), UINT16_C(61004), UINT16_C(49058),
2148
+ UINT16_C(58870), UINT16_C(65082), UINT16_C(64732), UINT16_C(36111)) },
2149
+ { simde_x_mm_set_epu16(UINT16_C(17034), UINT16_C(61660), UINT16_C(37691), UINT16_C(48256),
2150
+ UINT16_C(52466), UINT16_C(59150), UINT16_C(48524), UINT16_C(48932)),
2151
+ simde_x_mm_set_epu16(UINT16_C( 9311), UINT16_C(52954), UINT16_C(59922), UINT16_C(37439),
2152
+ UINT16_C(48521), UINT16_C(44576), UINT16_C(58557), UINT16_C( 4347)),
2153
+ simde_x_mm_set_epu16(UINT16_C(17034), UINT16_C(61660), UINT16_C(59922), UINT16_C(48256),
2154
+ UINT16_C(52466), UINT16_C(59150), UINT16_C(58557), UINT16_C(48932)) },
2155
+ { simde_x_mm_set_epu16(UINT16_C( 3717), UINT16_C(65479), UINT16_C(47591), UINT16_C(37591),
2156
+ UINT16_C(56498), UINT16_C(64377), UINT16_C( 4742), UINT16_C(26061)),
2157
+ simde_x_mm_set_epu16(UINT16_C(53135), UINT16_C(25263), UINT16_C(35893), UINT16_C(43723),
2158
+ UINT16_C(55550), UINT16_C(13729), UINT16_C(44168), UINT16_C(18708)),
2159
+ simde_x_mm_set_epu16(UINT16_C(53135), UINT16_C(65479), UINT16_C(47591), UINT16_C(43723),
2160
+ UINT16_C(56498), UINT16_C(64377), UINT16_C(44168), UINT16_C(26061)) },
2161
+ { simde_x_mm_set_epu16(UINT16_C(32738), UINT16_C(32324), UINT16_C(12257), UINT16_C( 4392),
2162
+ UINT16_C(48941), UINT16_C(15991), UINT16_C(56181), UINT16_C(64839)),
2163
+ simde_x_mm_set_epu16(UINT16_C(54182), UINT16_C(32645), UINT16_C(19389), UINT16_C(64736),
2164
+ UINT16_C(55568), UINT16_C(54741), UINT16_C(47863), UINT16_C(26870)),
2165
+ simde_x_mm_set_epu16(UINT16_C(54182), UINT16_C(32645), UINT16_C(19389), UINT16_C(64736),
2166
+ UINT16_C(55568), UINT16_C(54741), UINT16_C(56181), UINT16_C(64839)) },
2167
+ { simde_x_mm_set_epu16(UINT16_C(25092), UINT16_C( 2337), UINT16_C(41136), UINT16_C(62054),
2168
+ UINT16_C(31692), UINT16_C(47896), UINT16_C(26455), UINT16_C(41876)),
2169
+ simde_x_mm_set_epu16(UINT16_C(57533), UINT16_C( 4966), UINT16_C(21312), UINT16_C(22488),
2170
+ UINT16_C(59621), UINT16_C( 4501), UINT16_C(18372), UINT16_C(62293)),
2171
+ simde_x_mm_set_epu16(UINT16_C(57533), UINT16_C( 4966), UINT16_C(41136), UINT16_C(62054),
2172
+ UINT16_C(59621), UINT16_C(47896), UINT16_C(26455), UINT16_C(62293)) },
2173
+ { simde_x_mm_set_epu16(UINT16_C(48994), UINT16_C(54179), UINT16_C(26228), UINT16_C(61187),
2174
+ UINT16_C(44854), UINT16_C(57084), UINT16_C(16902), UINT16_C( 1404)),
2175
+ simde_x_mm_set_epu16(UINT16_C( 4826), UINT16_C(23694), UINT16_C(32546), UINT16_C(41184),
2176
+ UINT16_C( 725), UINT16_C(27358), UINT16_C(30444), UINT16_C(27027)),
2177
+ simde_x_mm_set_epu16(UINT16_C(48994), UINT16_C(54179), UINT16_C(32546), UINT16_C(61187),
2178
+ UINT16_C(44854), UINT16_C(57084), UINT16_C(30444), UINT16_C(27027)) },
2179
+ { simde_x_mm_set_epu16(UINT16_C(60922), UINT16_C(51825), UINT16_C( 9165), UINT16_C(16534),
2180
+ UINT16_C( 2526), UINT16_C(29939), UINT16_C(61826), UINT16_C(45025)),
2181
+ simde_x_mm_set_epu16(UINT16_C(10446), UINT16_C(28319), UINT16_C(61630), UINT16_C(32182),
2182
+ UINT16_C(17860), UINT16_C(15017), UINT16_C(59067), UINT16_C(19548)),
2183
+ simde_x_mm_set_epu16(UINT16_C(60922), UINT16_C(51825), UINT16_C(61630), UINT16_C(32182),
2184
+ UINT16_C(17860), UINT16_C(29939), UINT16_C(61826), UINT16_C(45025)) },
2185
+ { simde_x_mm_set_epu16(UINT16_C(34011), UINT16_C(36834), UINT16_C(12993), UINT16_C(21384),
2186
+ UINT16_C(62643), UINT16_C(55943), UINT16_C(12296), UINT16_C(62062)),
2187
+ simde_x_mm_set_epu16(UINT16_C(22645), UINT16_C(31782), UINT16_C(42377), UINT16_C(17805),
2188
+ UINT16_C(55262), UINT16_C(50550), UINT16_C(53828), UINT16_C(32401)),
2189
+ simde_x_mm_set_epu16(UINT16_C(34011), UINT16_C(36834), UINT16_C(42377), UINT16_C(21384),
2190
+ UINT16_C(62643), UINT16_C(55943), UINT16_C(53828), UINT16_C(62062)) }
2191
+ };
2192
+
2193
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
2194
+ simde__m128i r = simde_mm_max_epu16(test_vec[i].a, test_vec[i].b);
2195
+ simde_assert_m128i_u16(r, ==, test_vec[i].r);
2196
+ }
2197
+
2198
+ return MUNIT_OK;
2199
+ }
2200
+
2201
+ static MunitResult
2202
+ test_simde_mm_max_epu32(const MunitParameter params[], void* data) {
2203
+ (void) params;
2204
+ (void) data;
2205
+
2206
+ const struct {
2207
+ simde__m128i a;
2208
+ simde__m128i b;
2209
+ simde__m128i r;
2210
+ } test_vec[8] = {
2211
+ { simde_x_mm_set_epu32(UINT32_C(2532888443), UINT32_C( 529927089), UINT32_C(4207752004), UINT32_C(3510207683)),
2212
+ simde_x_mm_set_epu32(UINT32_C( 574626150), UINT32_C(1395025810), UINT32_C(1034901368), UINT32_C(3021400990)),
2213
+ simde_x_mm_set_epu32(UINT32_C(2532888443), UINT32_C(1395025810), UINT32_C(4207752004), UINT32_C(3510207683)) },
2214
+ { simde_x_mm_set_epu32(UINT32_C(2363706693), UINT32_C(3928702803), UINT32_C( 881772043), UINT32_C( 57971440)),
2215
+ simde_x_mm_set_epu32(UINT32_C(4216913656), UINT32_C(3303535239), UINT32_C( 568980555), UINT32_C( 308523788)),
2216
+ simde_x_mm_set_epu32(UINT32_C(4216913656), UINT32_C(3928702803), UINT32_C( 881772043), UINT32_C( 308523788)) },
2217
+ { simde_x_mm_set_epu32(UINT32_C( 548389896), UINT32_C(2976724828), UINT32_C(3723033325), UINT32_C(1050716143)),
2218
+ simde_x_mm_set_epu32(UINT32_C(3591135284), UINT32_C( 856620101), UINT32_C( 833768684), UINT32_C( 203883279)),
2219
+ simde_x_mm_set_epu32(UINT32_C(3591135284), UINT32_C(2976724828), UINT32_C(3723033325), UINT32_C(1050716143)) },
2220
+ { simde_x_mm_set_epu32(UINT32_C(3717602953), UINT32_C(2736045656), UINT32_C(2539339865), UINT32_C(2199825857)),
2221
+ simde_x_mm_set_epu32(UINT32_C(3331243307), UINT32_C(1452685186), UINT32_C(2337560284), UINT32_C(3592613020)),
2222
+ simde_x_mm_set_epu32(UINT32_C(3717602953), UINT32_C(2736045656), UINT32_C(2539339865), UINT32_C(3592613020)) },
2223
+ { simde_x_mm_set_epu32(UINT32_C(2462641628), UINT32_C(4101106468), UINT32_C(2939293383), UINT32_C( 448332726)),
2224
+ simde_x_mm_set_epu32(UINT32_C(4091790581), UINT32_C( 227177150), UINT32_C(4294101921), UINT32_C(1459100736)),
2225
+ simde_x_mm_set_epu32(UINT32_C(4091790581), UINT32_C(4101106468), UINT32_C(4294101921), UINT32_C(1459100736)) },
2226
+ { simde_x_mm_set_epu32(UINT32_C(1973671448), UINT32_C(1498045863), UINT32_C(1898038098), UINT32_C(1965060938)),
2227
+ simde_x_mm_set_epu32(UINT32_C( 29809850), UINT32_C( 861011023), UINT32_C( 63413124), UINT32_C(2176922492)),
2228
+ simde_x_mm_set_epu32(UINT32_C(1973671448), UINT32_C(1498045863), UINT32_C(1898038098), UINT32_C(2176922492)) },
2229
+ { simde_x_mm_set_epu32(UINT32_C( 526896309), UINT32_C( 817143907), UINT32_C(1563490657), UINT32_C(3681037749)),
2230
+ simde_x_mm_set_epu32(UINT32_C(3002623714), UINT32_C(3298584662), UINT32_C(1905485247), UINT32_C( 108377847)),
2231
+ simde_x_mm_set_epu32(UINT32_C(3002623714), UINT32_C(3298584662), UINT32_C(1905485247), UINT32_C(3681037749)) },
2232
+ { simde_x_mm_set_epu32(UINT32_C( 749402761), UINT32_C(3319236993), UINT32_C(3240012333), UINT32_C(1509228324)),
2233
+ simde_x_mm_set_epu32(UINT32_C(3285355124), UINT32_C(3206151128), UINT32_C(2987263147), UINT32_C(3706118926)),
2234
+ simde_x_mm_set_epu32(UINT32_C(3285355124), UINT32_C(3319236993), UINT32_C(3240012333), UINT32_C(3706118926)) }
2235
+ };
2236
+
2237
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
2238
+ simde__m128i r = simde_mm_max_epu32(test_vec[i].a, test_vec[i].b);
2239
+ simde_assert_m128i_u32(r, ==, test_vec[i].r);
2240
+ }
2241
+
2242
+ return MUNIT_OK;
2243
+ }
2244
+
2245
+ static MunitResult
2246
+ test_simde_mm_min_epi8(const MunitParameter params[], void* data) {
2247
+ (void) params;
2248
+ (void) data;
2249
+
2250
+ const struct {
2251
+ simde__m128i a;
2252
+ simde__m128i b;
2253
+ simde__m128i r;
2254
+ } test_vec[8] = {
2255
+ { simde_mm_set_epi8(INT8_C( -85), INT8_C( 106), INT8_C( -97), INT8_C( -26),
2256
+ INT8_C( -5), INT8_C( 83), INT8_C(-106), INT8_C(-116),
2257
+ INT8_C(-113), INT8_C( -34), INT8_C( -81), INT8_C( 109),
2258
+ INT8_C( -25), INT8_C( -16), INT8_C( 28), INT8_C( -76)),
2259
+ simde_mm_set_epi8(INT8_C( -69), INT8_C( -4), INT8_C( 73), INT8_C( -90),
2260
+ INT8_C( 93), INT8_C(-116), INT8_C( 19), INT8_C( 124),
2261
+ INT8_C( -67), INT8_C( 48), INT8_C( 38), INT8_C( 11),
2262
+ INT8_C( 21), INT8_C( 91), INT8_C( 118), INT8_C( 103)),
2263
+ simde_mm_set_epi8(INT8_C( -85), INT8_C( -4), INT8_C( -97), INT8_C( -90),
2264
+ INT8_C( -5), INT8_C(-116), INT8_C(-106), INT8_C(-116),
2265
+ INT8_C(-113), INT8_C( -34), INT8_C( -81), INT8_C( 11),
2266
+ INT8_C( -25), INT8_C( -16), INT8_C( 28), INT8_C( -76)) },
2267
+ { simde_mm_set_epi8(INT8_C( -69), INT8_C( -84), INT8_C( 37), INT8_C( -17),
2268
+ INT8_C( 36), INT8_C( -34), INT8_C(-104), INT8_C( 112),
2269
+ INT8_C( 13), INT8_C( 47), INT8_C( -94), INT8_C( -38),
2270
+ INT8_C( 111), INT8_C( -44), INT8_C( -90), INT8_C( -86)),
2271
+ simde_mm_set_epi8(INT8_C(-125), INT8_C( 55), INT8_C( 5), INT8_C( 113),
2272
+ INT8_C( 100), INT8_C( -35), INT8_C( -4), INT8_C( 29),
2273
+ INT8_C( 60), INT8_C( 120), INT8_C( -36), INT8_C( 72),
2274
+ INT8_C( 92), INT8_C( 31), INT8_C( 11), INT8_C(-106)),
2275
+ simde_mm_set_epi8(INT8_C(-125), INT8_C( -84), INT8_C( 5), INT8_C( -17),
2276
+ INT8_C( 36), INT8_C( -35), INT8_C(-104), INT8_C( 29),
2277
+ INT8_C( 13), INT8_C( 47), INT8_C( -94), INT8_C( -38),
2278
+ INT8_C( 92), INT8_C( -44), INT8_C( -90), INT8_C(-106)) },
2279
+ { simde_mm_set_epi8(INT8_C( 49), INT8_C(-113), INT8_C( -25), INT8_C( -86),
2280
+ INT8_C( 106), INT8_C( 71), INT8_C( 43), INT8_C( 81),
2281
+ INT8_C( -66), INT8_C( 111), INT8_C( 40), INT8_C( 35),
2282
+ INT8_C( 15), INT8_C( 51), INT8_C( 21), INT8_C( 60)),
2283
+ simde_mm_set_epi8(INT8_C( 37), INT8_C( -84), INT8_C( 110), INT8_C( -46),
2284
+ INT8_C( -27), INT8_C( -56), INT8_C( 104), INT8_C( 2),
2285
+ INT8_C(-105), INT8_C( 65), INT8_C( 54), INT8_C( 75),
2286
+ INT8_C( 113), INT8_C( 29), INT8_C( -47), INT8_C( 84)),
2287
+ simde_mm_set_epi8(INT8_C( 37), INT8_C(-113), INT8_C( -25), INT8_C( -86),
2288
+ INT8_C( -27), INT8_C( -56), INT8_C( 43), INT8_C( 2),
2289
+ INT8_C(-105), INT8_C( 65), INT8_C( 40), INT8_C( 35),
2290
+ INT8_C( 15), INT8_C( 29), INT8_C( -47), INT8_C( 60)) },
2291
+ { simde_mm_set_epi8(INT8_C( -9), INT8_C( -51), INT8_C( -8), INT8_C( -23),
2292
+ INT8_C( 42), INT8_C( 26), INT8_C( -68), INT8_C( -93),
2293
+ INT8_C( -78), INT8_C( -45), INT8_C( -36), INT8_C( -34),
2294
+ INT8_C( -79), INT8_C( -94), INT8_C( 80), INT8_C( 99)),
2295
+ simde_mm_set_epi8(INT8_C( 27), INT8_C( 45), INT8_C( 48), INT8_C( -73),
2296
+ INT8_C( -21), INT8_C( 86), INT8_C( 54), INT8_C( -39),
2297
+ INT8_C( -2), INT8_C( 108), INT8_C( -19), INT8_C( -81),
2298
+ INT8_C( -52), INT8_C( 118), INT8_C( 61), INT8_C( -36)),
2299
+ simde_mm_set_epi8(INT8_C( -9), INT8_C( -51), INT8_C( -8), INT8_C( -73),
2300
+ INT8_C( -21), INT8_C( 26), INT8_C( -68), INT8_C( -93),
2301
+ INT8_C( -78), INT8_C( -45), INT8_C( -36), INT8_C( -81),
2302
+ INT8_C( -79), INT8_C( -94), INT8_C( 61), INT8_C( -36)) },
2303
+ { simde_mm_set_epi8(INT8_C( 75), INT8_C( 2), INT8_C( -37), INT8_C( -22),
2304
+ INT8_C( 110), INT8_C(-125), INT8_C( 33), INT8_C( -94),
2305
+ INT8_C( -5), INT8_C( -67), INT8_C( -78), INT8_C( -46),
2306
+ INT8_C( 43), INT8_C( -89), INT8_C( 74), INT8_C( 18)),
2307
+ simde_mm_set_epi8(INT8_C( 84), INT8_C( -44), INT8_C(-115), INT8_C( 29),
2308
+ INT8_C(-121), INT8_C( 22), INT8_C( 43), INT8_C(-123),
2309
+ INT8_C( -63), INT8_C( -73), INT8_C( 36), INT8_C( -60),
2310
+ INT8_C( -96), INT8_C( -58), INT8_C( -42), INT8_C( 78)),
2311
+ simde_mm_set_epi8(INT8_C( 75), INT8_C( -44), INT8_C(-115), INT8_C( -22),
2312
+ INT8_C(-121), INT8_C(-125), INT8_C( 33), INT8_C(-123),
2313
+ INT8_C( -63), INT8_C( -73), INT8_C( -78), INT8_C( -60),
2314
+ INT8_C( -96), INT8_C( -89), INT8_C( -42), INT8_C( 18)) },
2315
+ { simde_mm_set_epi8(INT8_C( -82), INT8_C( -34), INT8_C( 74), INT8_C( 86),
2316
+ INT8_C( -7), INT8_C( 93), INT8_C( -99), INT8_C( -7),
2317
+ INT8_C( 127), INT8_C( 26), INT8_C( -51), INT8_C( 41),
2318
+ INT8_C( 71), INT8_C( 67), INT8_C(-102), INT8_C(-119)),
2319
+ simde_mm_set_epi8(INT8_C( -8), INT8_C( 88), INT8_C( 46), INT8_C( 105),
2320
+ INT8_C( -43), INT8_C(-109), INT8_C( 38), INT8_C( -35),
2321
+ INT8_C( -17), INT8_C( -4), INT8_C(-122), INT8_C(-112),
2322
+ INT8_C( -62), INT8_C( 100), INT8_C( 89), INT8_C( -83)),
2323
+ simde_mm_set_epi8(INT8_C( -82), INT8_C( -34), INT8_C( 46), INT8_C( 86),
2324
+ INT8_C( -43), INT8_C(-109), INT8_C( -99), INT8_C( -35),
2325
+ INT8_C( -17), INT8_C( -4), INT8_C(-122), INT8_C(-112),
2326
+ INT8_C( -62), INT8_C( 67), INT8_C(-102), INT8_C(-119)) },
2327
+ { simde_mm_set_epi8(INT8_C( 21), INT8_C( -83), INT8_C( -9), INT8_C( -66),
2328
+ INT8_C( -50), INT8_C( -47), INT8_C( -30), INT8_C( 4),
2329
+ INT8_C( -60), INT8_C( 23), INT8_C( 126), INT8_C( 69),
2330
+ INT8_C( 56), INT8_C( 4), INT8_C( 32), INT8_C( 31)),
2331
+ simde_mm_set_epi8(INT8_C( 76), INT8_C( -95), INT8_C( 81), INT8_C( 9),
2332
+ INT8_C(-106), INT8_C( -75), INT8_C( 46), INT8_C( 117),
2333
+ INT8_C(-119), INT8_C( -56), INT8_C(-115), INT8_C( -97),
2334
+ INT8_C( -52), INT8_C( -81), INT8_C(-123), INT8_C( 54)),
2335
+ simde_mm_set_epi8(INT8_C( 21), INT8_C( -95), INT8_C( -9), INT8_C( -66),
2336
+ INT8_C(-106), INT8_C( -75), INT8_C( -30), INT8_C( 4),
2337
+ INT8_C(-119), INT8_C( -56), INT8_C(-115), INT8_C( -97),
2338
+ INT8_C( -52), INT8_C( -81), INT8_C(-123), INT8_C( 31)) },
2339
+ { simde_mm_set_epi8(INT8_C( -56), INT8_C(-107), INT8_C( 36), INT8_C( -66),
2340
+ INT8_C( -22), INT8_C(-112), INT8_C( 57), INT8_C( 12),
2341
+ INT8_C( -40), INT8_C(-109), INT8_C( -85), INT8_C( 86),
2342
+ INT8_C( -31), INT8_C( -66), INT8_C( 99), INT8_C( -10)),
2343
+ simde_mm_set_epi8(INT8_C( 57), INT8_C(-109), INT8_C( 45), INT8_C( 32),
2344
+ INT8_C( 94), INT8_C( -27), INT8_C( 11), INT8_C( 67),
2345
+ INT8_C(-114), INT8_C( -40), INT8_C(-125), INT8_C(-106),
2346
+ INT8_C( 72), INT8_C( -34), INT8_C(-102), INT8_C( 18)),
2347
+ simde_mm_set_epi8(INT8_C( -56), INT8_C(-109), INT8_C( 36), INT8_C( -66),
2348
+ INT8_C( -22), INT8_C(-112), INT8_C( 11), INT8_C( 12),
2349
+ INT8_C(-114), INT8_C(-109), INT8_C(-125), INT8_C(-106),
2350
+ INT8_C( -31), INT8_C( -66), INT8_C(-102), INT8_C( -10)) }
2351
+ };
2352
+
2353
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
2354
+ simde__m128i r = simde_mm_min_epi8(test_vec[i].a, test_vec[i].b);
2355
+ simde_assert_m128i_i8(r, ==, test_vec[i].r);
2356
+ }
2357
+
2358
+ return MUNIT_OK;
2359
+ }
2360
+
2361
+ static MunitResult
2362
+ test_simde_mm_min_epi32(const MunitParameter params[], void* data) {
2363
+ (void) params;
2364
+ (void) data;
2365
+
2366
+ const struct {
2367
+ simde__m128i a;
2368
+ simde__m128i b;
2369
+ simde__m128i r;
2370
+ } test_vec[8] = {
2371
+ { simde_mm_set_epi32(INT32_C(-1390607532), INT32_C( 385446037), INT32_C( -880551963), INT32_C(-1976882160)),
2372
+ simde_mm_set_epi32(INT32_C(-1328784057), INT32_C( 765325983), INT32_C( 188641041), INT32_C(-1795531007)),
2373
+ simde_mm_set_epi32(INT32_C(-1390607532), INT32_C( 385446037), INT32_C( -880551963), INT32_C(-1976882160)) },
2374
+ { simde_mm_set_epi32(INT32_C( 1033654035), INT32_C(-1435181472), INT32_C( 1188098919), INT32_C(-1789978439)),
2375
+ simde_mm_set_epi32(INT32_C( 667650881), INT32_C( -921357846), INT32_C( 1991872258), INT32_C(-1380518571)),
2376
+ simde_mm_set_epi32(INT32_C( 667650881), INT32_C(-1435181472), INT32_C( 1188098919), INT32_C(-1789978439)) },
2377
+ { simde_mm_set_epi32(INT32_C( -209415658), INT32_C(-1820339262), INT32_C( 3383029), INT32_C( -894067821)),
2378
+ simde_mm_set_epi32(INT32_C( -460825945), INT32_C(-1201816425), INT32_C( 657326597), INT32_C( 270703936)),
2379
+ simde_mm_set_epi32(INT32_C( -460825945), INT32_C(-1820339262), INT32_C( 3383029), INT32_C( -894067821)) },
2380
+ { simde_mm_set_epi32(INT32_C( -711299231), INT32_C(-1150752518), INT32_C( 489373999), INT32_C(-1003224360)),
2381
+ simde_mm_set_epi32(INT32_C( -111820496), INT32_C(-1961663677), INT32_C( -829684478), INT32_C(-1602379520)),
2382
+ simde_mm_set_epi32(INT32_C( -711299231), INT32_C(-1961663677), INT32_C( -829684478), INT32_C(-1602379520)) },
2383
+ { simde_mm_set_epi32(INT32_C( 1660536841), INT32_C( 1244832127), INT32_C( 1920382062), INT32_C( -791791151)),
2384
+ simde_mm_set_epi32(INT32_C( 799949492), INT32_C( 108351731), INT32_C( 1351843398), INT32_C( 1009306115)),
2385
+ simde_mm_set_epi32(INT32_C( 799949492), INT32_C( 108351731), INT32_C( 1351843398), INT32_C( -791791151)) },
2386
+ { simde_mm_set_epi32(INT32_C(-1327926842), INT32_C( 984094365), INT32_C( 1894460641), INT32_C( -26405962)),
2387
+ simde_mm_set_epi32(INT32_C( 1205481380), INT32_C( 137390510), INT32_C( 590747940), INT32_C( 1878652085)),
2388
+ simde_mm_set_epi32(INT32_C(-1327926842), INT32_C( 137390510), INT32_C( 590747940), INT32_C( -26405962)) },
2389
+ { simde_mm_set_epi32(INT32_C( 316126780), INT32_C(-1043744394), INT32_C( 787289081), INT32_C( -910852690)),
2390
+ simde_mm_set_epi32(INT32_C(-2087288343), INT32_C(-1056950039), INT32_C( -555736705), INT32_C( 666671486)),
2391
+ simde_mm_set_epi32(INT32_C(-2087288343), INT32_C(-1056950039), INT32_C( -555736705), INT32_C( -910852690)) },
2392
+ { simde_mm_set_epi32(INT32_C(-1498461986), INT32_C(-1418974686), INT32_C(-1311767695), INT32_C(-1038502387)),
2393
+ simde_mm_set_epi32(INT32_C( -722855220), INT32_C( 492871512), INT32_C( -175913896), INT32_C( -920105808)),
2394
+ simde_mm_set_epi32(INT32_C(-1498461986), INT32_C(-1418974686), INT32_C(-1311767695), INT32_C(-1038502387)) }
2395
+ };
2396
+
2397
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
2398
+ simde__m128i r = simde_mm_min_epi32(test_vec[i].a, test_vec[i].b);
2399
+ simde_assert_m128i_i32(r, ==, test_vec[i].r);
2400
+ }
2401
+
2402
+ return MUNIT_OK;
2403
+ }
2404
+
2405
+ static MunitResult
2406
+ test_simde_mm_min_epu16(const MunitParameter params[], void* data) {
2407
+ (void) params;
2408
+ (void) data;
2409
+
2410
+ const struct {
2411
+ simde__m128i a;
2412
+ simde__m128i b;
2413
+ simde__m128i r;
2414
+ } test_vec[8] = {
2415
+ { simde_x_mm_set_epu16(UINT16_C(51368), UINT16_C(47248), UINT16_C( 2518), UINT16_C(59587),
2416
+ UINT16_C(64433), UINT16_C(45544), UINT16_C( 8944), UINT16_C(56380)),
2417
+ simde_x_mm_set_epu16(UINT16_C(35287), UINT16_C( 4953), UINT16_C(55160), UINT16_C(11599),
2418
+ UINT16_C( 1282), UINT16_C(22629), UINT16_C( 1908), UINT16_C(57232)),
2419
+ simde_x_mm_set_epu16(UINT16_C(35287), UINT16_C( 4953), UINT16_C( 2518), UINT16_C(11599),
2420
+ UINT16_C( 1282), UINT16_C(22629), UINT16_C( 1908), UINT16_C(56380)) },
2421
+ { simde_x_mm_set_epu16(UINT16_C(51082), UINT16_C(41324), UINT16_C(63856), UINT16_C(49625),
2422
+ UINT16_C(47646), UINT16_C(40071), UINT16_C(43894), UINT16_C( 6845)),
2423
+ simde_x_mm_set_epu16(UINT16_C(29622), UINT16_C(57397), UINT16_C(58908), UINT16_C(15516),
2424
+ UINT16_C(57300), UINT16_C(16770), UINT16_C( 6179), UINT16_C(31976)),
2425
+ simde_x_mm_set_epu16(UINT16_C(29622), UINT16_C(41324), UINT16_C(58908), UINT16_C(15516),
2426
+ UINT16_C(47646), UINT16_C(16770), UINT16_C( 6179), UINT16_C( 6845)) },
2427
+ { simde_x_mm_set_epu16(UINT16_C( 7791), UINT16_C( 9357), UINT16_C(51414), UINT16_C(32085),
2428
+ UINT16_C( 5803), UINT16_C(15112), UINT16_C( 4983), UINT16_C(62284)),
2429
+ simde_x_mm_set_epu16(UINT16_C(19191), UINT16_C(53664), UINT16_C(31350), UINT16_C(59818),
2430
+ UINT16_C(36846), UINT16_C(44912), UINT16_C(44597), UINT16_C(18257)),
2431
+ simde_x_mm_set_epu16(UINT16_C( 7791), UINT16_C( 9357), UINT16_C(31350), UINT16_C(32085),
2432
+ UINT16_C( 5803), UINT16_C(15112), UINT16_C( 4983), UINT16_C(18257)) },
2433
+ { simde_x_mm_set_epu16(UINT16_C(23178), UINT16_C(52616), UINT16_C(54786), UINT16_C(18857),
2434
+ UINT16_C(64323), UINT16_C(44008), UINT16_C(12675), UINT16_C(13653)),
2435
+ simde_x_mm_set_epu16(UINT16_C(16010), UINT16_C(49432), UINT16_C(46679), UINT16_C(33525),
2436
+ UINT16_C(47161), UINT16_C(64878), UINT16_C(40685), UINT16_C(46136)),
2437
+ simde_x_mm_set_epu16(UINT16_C(16010), UINT16_C(49432), UINT16_C(46679), UINT16_C(18857),
2438
+ UINT16_C(47161), UINT16_C(44008), UINT16_C(12675), UINT16_C(13653)) },
2439
+ { simde_x_mm_set_epu16(UINT16_C(33571), UINT16_C(12594), UINT16_C( 3662), UINT16_C(56380),
2440
+ UINT16_C(54926), UINT16_C(32164), UINT16_C( 3421), UINT16_C(38714)),
2441
+ simde_x_mm_set_epu16(UINT16_C(65175), UINT16_C(47536), UINT16_C(21879), UINT16_C(12881),
2442
+ UINT16_C(50915), UINT16_C(31304), UINT16_C(10941), UINT16_C(57984)),
2443
+ simde_x_mm_set_epu16(UINT16_C(33571), UINT16_C(12594), UINT16_C( 3662), UINT16_C(12881),
2444
+ UINT16_C(50915), UINT16_C(31304), UINT16_C( 3421), UINT16_C(38714)) },
2445
+ { simde_x_mm_set_epu16(UINT16_C(17540), UINT16_C(62844), UINT16_C(46558), UINT16_C(45098),
2446
+ UINT16_C(32389), UINT16_C(20250), UINT16_C(19285), UINT16_C(40862)),
2447
+ simde_x_mm_set_epu16(UINT16_C(17361), UINT16_C( 8131), UINT16_C(50246), UINT16_C(30481),
2448
+ UINT16_C(30265), UINT16_C(57439), UINT16_C(42358), UINT16_C(24996)),
2449
+ simde_x_mm_set_epu16(UINT16_C(17361), UINT16_C( 8131), UINT16_C(46558), UINT16_C(30481),
2450
+ UINT16_C(30265), UINT16_C(20250), UINT16_C(19285), UINT16_C(24996)) },
2451
+ { simde_x_mm_set_epu16(UINT16_C(11573), UINT16_C(56754), UINT16_C(38860), UINT16_C(26669),
2452
+ UINT16_C(21209), UINT16_C( 2072), UINT16_C(19855), UINT16_C(11067)),
2453
+ simde_x_mm_set_epu16(UINT16_C( 4235), UINT16_C(30280), UINT16_C(23335), UINT16_C(17584),
2454
+ UINT16_C( 3445), UINT16_C( 1117), UINT16_C(46968), UINT16_C(58704)),
2455
+ simde_x_mm_set_epu16(UINT16_C( 4235), UINT16_C(30280), UINT16_C(23335), UINT16_C(17584),
2456
+ UINT16_C( 3445), UINT16_C( 1117), UINT16_C(19855), UINT16_C(11067)) },
2457
+ { simde_x_mm_set_epu16(UINT16_C( 1826), UINT16_C(34774), UINT16_C(17130), UINT16_C(44175),
2458
+ UINT16_C(60865), UINT16_C(54658), UINT16_C(62552), UINT16_C(62233)),
2459
+ simde_x_mm_set_epu16(UINT16_C(15183), UINT16_C(55598), UINT16_C( 5779), UINT16_C(28145),
2460
+ UINT16_C(24471), UINT16_C(50174), UINT16_C(53029), UINT16_C(43004)),
2461
+ simde_x_mm_set_epu16(UINT16_C( 1826), UINT16_C(34774), UINT16_C( 5779), UINT16_C(28145),
2462
+ UINT16_C(24471), UINT16_C(50174), UINT16_C(53029), UINT16_C(43004)) }
2463
+ };
2464
+
2465
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
2466
+ simde__m128i r = simde_mm_min_epu16(test_vec[i].a, test_vec[i].b);
2467
+ simde_assert_m128i_u16(r, ==, test_vec[i].r);
2468
+ }
2469
+
2470
+ return MUNIT_OK;
2471
+ }
2472
+
2473
+ static MunitResult
2474
+ test_simde_mm_min_epu32(const MunitParameter params[], void* data) {
2475
+ (void) params;
2476
+ (void) data;
2477
+
2478
+ const struct {
2479
+ simde__m128i a;
2480
+ simde__m128i b;
2481
+ simde__m128i r;
2482
+ } test_vec[8] = {
2483
+ { simde_x_mm_set_epu32(UINT32_C( 788201309), UINT32_C(2071601612), UINT32_C(3123609508), UINT32_C(1344608341)),
2484
+ simde_x_mm_set_epu32(UINT32_C(3982697916), UINT32_C( 968486606), UINT32_C(1116860132), UINT32_C(2825138301)),
2485
+ simde_x_mm_set_epu32(UINT32_C( 788201309), UINT32_C( 968486606), UINT32_C(1116860132), UINT32_C(1344608341)) },
2486
+ { simde_x_mm_set_epu32(UINT32_C( 384132137), UINT32_C(1977877021), UINT32_C(4289747748), UINT32_C(2515962526)),
2487
+ simde_x_mm_set_epu32(UINT32_C(3869012810), UINT32_C( 645527240), UINT32_C(4230349530), UINT32_C(1145244296)),
2488
+ simde_x_mm_set_epu32(UINT32_C( 384132137), UINT32_C( 645527240), UINT32_C(4230349530), UINT32_C(1145244296)) },
2489
+ { simde_x_mm_set_epu32(UINT32_C(3974730728), UINT32_C(3242218332), UINT32_C(1720053534), UINT32_C(3159005428)),
2490
+ simde_x_mm_set_epu32(UINT32_C(3856942180), UINT32_C(4205396871), UINT32_C(2272972736), UINT32_C(4113553543)),
2491
+ simde_x_mm_set_epu32(UINT32_C(3856942180), UINT32_C(3242218332), UINT32_C(1720053534), UINT32_C(3159005428)) },
2492
+ { simde_x_mm_set_epu32(UINT32_C(3254877076), UINT32_C(2399148141), UINT32_C(2938842657), UINT32_C(1961058722)),
2493
+ simde_x_mm_set_epu32(UINT32_C(1528033433), UINT32_C( 461232656), UINT32_C(1948639979), UINT32_C(1593317844)),
2494
+ simde_x_mm_set_epu32(UINT32_C(1528033433), UINT32_C( 461232656), UINT32_C(1948639979), UINT32_C(1593317844)) },
2495
+ { simde_x_mm_set_epu32(UINT32_C(1721814381), UINT32_C(3279771329), UINT32_C(1032528892), UINT32_C(1770589845)),
2496
+ simde_x_mm_set_epu32(UINT32_C( 338961476), UINT32_C(1015681646), UINT32_C(2529171351), UINT32_C(1474803551)),
2497
+ simde_x_mm_set_epu32(UINT32_C( 338961476), UINT32_C(1015681646), UINT32_C(1032528892), UINT32_C(1474803551)) },
2498
+ { simde_x_mm_set_epu32(UINT32_C(2865994869), UINT32_C(3359283004), UINT32_C(2376950903), UINT32_C(2025629636)),
2499
+ simde_x_mm_set_epu32(UINT32_C(1089016862), UINT32_C(3251738064), UINT32_C(2393163457), UINT32_C( 147184464)),
2500
+ simde_x_mm_set_epu32(UINT32_C(1089016862), UINT32_C(3251738064), UINT32_C(2376950903), UINT32_C( 147184464)) },
2501
+ { simde_x_mm_set_epu32(UINT32_C(1041113949), UINT32_C(2410530901), UINT32_C( 131686205), UINT32_C(3148398563)),
2502
+ simde_x_mm_set_epu32(UINT32_C(1539533337), UINT32_C(3899801708), UINT32_C(3357338729), UINT32_C( 465532279)),
2503
+ simde_x_mm_set_epu32(UINT32_C(1041113949), UINT32_C(2410530901), UINT32_C( 131686205), UINT32_C( 465532279)) },
2504
+ { simde_x_mm_set_epu32(UINT32_C(2571079475), UINT32_C(4121171595), UINT32_C(3849859490), UINT32_C(2510281154)),
2505
+ simde_x_mm_set_epu32(UINT32_C(2596335814), UINT32_C(2240823510), UINT32_C(2609625051), UINT32_C(1259317116)),
2506
+ simde_x_mm_set_epu32(UINT32_C(2571079475), UINT32_C(2240823510), UINT32_C(2609625051), UINT32_C(1259317116)) }
2507
+ };
2508
+
2509
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
2510
+ simde__m128i r = simde_mm_min_epu32(test_vec[i].a, test_vec[i].b);
2511
+ simde_assert_m128i_u32(r, ==, test_vec[i].r);
2512
+ }
2513
+
2514
+ return MUNIT_OK;
2515
+ }
2516
+
2517
+ static MunitResult
2518
+ test_simde_mm_minpos_epu16(const MunitParameter params[], void* data) {
2519
+ (void) params;
2520
+ (void) data;
2521
+
2522
+ const struct {
2523
+ simde__m128i a;
2524
+ simde__m128i r;
2525
+ } test_vec[8] = {
2526
+ { simde_x_mm_set_epu16(UINT16_C(40818), UINT16_C(13812), UINT16_C(10872), UINT16_C(54124),
2527
+ UINT16_C(15100), UINT16_C(22810), UINT16_C(31798), UINT16_C(58017)),
2528
+ simde_x_mm_set_epu16(UINT16_C( 0), UINT16_C( 0), UINT16_C( 0), UINT16_C( 0),
2529
+ UINT16_C( 0), UINT16_C( 0), UINT16_C( 5), UINT16_C(10872)) },
2530
+ { simde_x_mm_set_epu16(UINT16_C(37656), UINT16_C(38804), UINT16_C(35573), UINT16_C(33536),
2531
+ UINT16_C(16030), UINT16_C(50431), UINT16_C(35836), UINT16_C(37491)),
2532
+ simde_x_mm_set_epu16(UINT16_C( 0), UINT16_C( 0), UINT16_C( 0), UINT16_C( 0),
2533
+ UINT16_C( 0), UINT16_C( 0), UINT16_C( 3), UINT16_C(16030)) },
2534
+ { simde_x_mm_set_epu16(UINT16_C(56214), UINT16_C(56676), UINT16_C(56044), UINT16_C(59619),
2535
+ UINT16_C(21150), UINT16_C(47982), UINT16_C( 92), UINT16_C(46668)),
2536
+ simde_x_mm_set_epu16(UINT16_C( 0), UINT16_C( 0), UINT16_C( 0), UINT16_C( 0),
2537
+ UINT16_C( 0), UINT16_C( 0), UINT16_C( 1), UINT16_C( 92)) },
2538
+ { simde_x_mm_set_epu16(UINT16_C(30841), UINT16_C(41481), UINT16_C(22332), UINT16_C(20377),
2539
+ UINT16_C( 8461), UINT16_C(16174), UINT16_C(46168), UINT16_C(38746)),
2540
+ simde_x_mm_set_epu16(UINT16_C( 0), UINT16_C( 0), UINT16_C( 0), UINT16_C( 0),
2541
+ UINT16_C( 0), UINT16_C( 0), UINT16_C( 3), UINT16_C( 8461)) },
2542
+ { simde_x_mm_set_epu16(UINT16_C(38550), UINT16_C(63031), UINT16_C(12859), UINT16_C(36699),
2543
+ UINT16_C(19386), UINT16_C(49588), UINT16_C(27067), UINT16_C(12812)),
2544
+ simde_x_mm_set_epu16(UINT16_C( 0), UINT16_C( 0), UINT16_C( 0), UINT16_C( 0),
2545
+ UINT16_C( 0), UINT16_C( 0), UINT16_C( 0), UINT16_C(12812)) },
2546
+ { simde_x_mm_set_epu16(UINT16_C(20747), UINT16_C(29576), UINT16_C( 5460), UINT16_C(38860),
2547
+ UINT16_C(52108), UINT16_C( 149), UINT16_C(63981), UINT16_C(19000)),
2548
+ simde_x_mm_set_epu16(UINT16_C( 0), UINT16_C( 0), UINT16_C( 0), UINT16_C( 0),
2549
+ UINT16_C( 0), UINT16_C( 0), UINT16_C( 2), UINT16_C( 149)) },
2550
+ { simde_x_mm_set_epu16(UINT16_C(31686), UINT16_C( 5758), UINT16_C(64324), UINT16_C(55924),
2551
+ UINT16_C(43567), UINT16_C(17467), UINT16_C(18633), UINT16_C(20456)),
2552
+ simde_x_mm_set_epu16(UINT16_C( 0), UINT16_C( 0), UINT16_C( 0), UINT16_C( 0),
2553
+ UINT16_C( 0), UINT16_C( 0), UINT16_C( 6), UINT16_C( 5758)) },
2554
+ { simde_x_mm_set_epu16(UINT16_C(13431), UINT16_C(48627), UINT16_C(23340), UINT16_C(23935),
2555
+ UINT16_C(12927), UINT16_C(48412), UINT16_C(33835), UINT16_C(15901)),
2556
+ simde_x_mm_set_epu16(UINT16_C( 0), UINT16_C( 0), UINT16_C( 0), UINT16_C( 0),
2557
+ UINT16_C( 0), UINT16_C( 0), UINT16_C( 3), UINT16_C(12927)) }
2558
+ };
2559
+
2560
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
2561
+ simde__m128i r = simde_mm_minpos_epu16(test_vec[i].a);
2562
+ simde_assert_m128i_u16(r, ==, test_vec[i].r);
2563
+ }
2564
+
2565
+ return MUNIT_OK;
2566
+ }
2567
+
2568
+ static MunitResult
2569
+ test_simde_mm_mpsadbw_epu8(const MunitParameter params[], void* data) {
2570
+ (void) params;
2571
+ (void) data;
2572
+
2573
+ const struct {
2574
+ simde__m128i a;
2575
+ simde__m128i b;
2576
+ simde__m128i r;
2577
+ } test_vec[8] = {
2578
+ { simde_x_mm_set_epu8(UINT8_C(226), UINT8_C(204), UINT8_C(132), UINT8_C(191),
2579
+ UINT8_C(109), UINT8_C( 5), UINT8_C( 22), UINT8_C(209),
2580
+ UINT8_C(243), UINT8_C(105), UINT8_C(186), UINT8_C(151),
2581
+ UINT8_C( 27), UINT8_C( 30), UINT8_C(223), UINT8_C( 88)),
2582
+ simde_x_mm_set_epu8(UINT8_C(139), UINT8_C(113), UINT8_C( 31), UINT8_C( 64),
2583
+ UINT8_C(155), UINT8_C( 28), UINT8_C(223), UINT8_C(195),
2584
+ UINT8_C(151), UINT8_C( 30), UINT8_C( 86), UINT8_C( 95),
2585
+ UINT8_C(158), UINT8_C( 54), UINT8_C( 10), UINT8_C( 66)),
2586
+ simde_x_mm_set_epu16(UINT16_C( 289), UINT16_C( 222), UINT16_C( 124), UINT16_C( 292),
2587
+ UINT16_C( 582), UINT16_C( 466), UINT16_C( 396), UINT16_C( 354)) },
2588
+ { simde_x_mm_set_epu8(UINT8_C(171), UINT8_C( 80), UINT8_C(108), UINT8_C(236),
2589
+ UINT8_C(239), UINT8_C(191), UINT8_C(219), UINT8_C(208),
2590
+ UINT8_C(165), UINT8_C( 30), UINT8_C(251), UINT8_C(196),
2591
+ UINT8_C( 25), UINT8_C(255), UINT8_C( 78), UINT8_C( 81)),
2592
+ simde_x_mm_set_epu8(UINT8_C( 98), UINT8_C( 67), UINT8_C(141), UINT8_C(239),
2593
+ UINT8_C( 33), UINT8_C(138), UINT8_C(161), UINT8_C(132),
2594
+ UINT8_C( 23), UINT8_C( 84), UINT8_C(170), UINT8_C(230),
2595
+ UINT8_C(214), UINT8_C(115), UINT8_C( 58), UINT8_C(130)),
2596
+ simde_x_mm_set_epu16(UINT16_C( 154), UINT16_C( 325), UINT16_C( 380), UINT16_C( 374),
2597
+ UINT16_C( 386), UINT16_C( 495), UINT16_C( 331), UINT16_C( 257)) },
2598
+ { simde_x_mm_set_epu8(UINT8_C( 36), UINT8_C( 1), UINT8_C( 5), UINT8_C( 22),
2599
+ UINT8_C( 73), UINT8_C(242), UINT8_C( 47), UINT8_C( 2),
2600
+ UINT8_C(204), UINT8_C(245), UINT8_C( 42), UINT8_C( 38),
2601
+ UINT8_C( 36), UINT8_C(136), UINT8_C( 36), UINT8_C(163)),
2602
+ simde_x_mm_set_epu8(UINT8_C(168), UINT8_C(213), UINT8_C( 2), UINT8_C( 30),
2603
+ UINT8_C( 25), UINT8_C(127), UINT8_C( 31), UINT8_C(133),
2604
+ UINT8_C(118), UINT8_C(121), UINT8_C(113), UINT8_C(231),
2605
+ UINT8_C( 34), UINT8_C( 80), UINT8_C(116), UINT8_C( 25)),
2606
+ simde_x_mm_set_epu16(UINT16_C( 438), UINT16_C( 637), UINT16_C( 543), UINT16_C( 197),
2607
+ UINT16_C( 414), UINT16_C( 749), UINT16_C( 430), UINT16_C( 116)) },
2608
+ { simde_x_mm_set_epu8(UINT8_C( 29), UINT8_C( 29), UINT8_C(199), UINT8_C(141),
2609
+ UINT8_C(153), UINT8_C(161), UINT8_C(202), UINT8_C( 57),
2610
+ UINT8_C( 56), UINT8_C( 10), UINT8_C(125), UINT8_C( 91),
2611
+ UINT8_C(212), UINT8_C(168), UINT8_C( 43), UINT8_C(108)),
2612
+ simde_x_mm_set_epu8(UINT8_C( 54), UINT8_C( 56), UINT8_C(193), UINT8_C(184),
2613
+ UINT8_C( 44), UINT8_C( 35), UINT8_C(133), UINT8_C( 78),
2614
+ UINT8_C( 34), UINT8_C(199), UINT8_C(192), UINT8_C(116),
2615
+ UINT8_C( 79), UINT8_C(137), UINT8_C(150), UINT8_C( 19)),
2616
+ simde_x_mm_set_epu16(UINT16_C( 251), UINT16_C( 293), UINT16_C( 234), UINT16_C( 340),
2617
+ UINT16_C( 517), UINT16_C( 460), UINT16_C( 245), UINT16_C( 209)) },
2618
+ { simde_x_mm_set_epu8(UINT8_C(211), UINT8_C(227), UINT8_C(150), UINT8_C( 37),
2619
+ UINT8_C(162), UINT8_C(249), UINT8_C(215), UINT8_C(195),
2620
+ UINT8_C( 63), UINT8_C(183), UINT8_C(196), UINT8_C( 97),
2621
+ UINT8_C( 60), UINT8_C(192), UINT8_C(194), UINT8_C(164)),
2622
+ simde_x_mm_set_epu8(UINT8_C(120), UINT8_C( 70), UINT8_C(238), UINT8_C(106),
2623
+ UINT8_C( 29), UINT8_C( 66), UINT8_C( 80), UINT8_C(174),
2624
+ UINT8_C( 76), UINT8_C(151), UINT8_C( 56), UINT8_C( 77),
2625
+ UINT8_C( 36), UINT8_C(138), UINT8_C(170), UINT8_C(105)),
2626
+ simde_x_mm_set_epu16(UINT16_C( 444), UINT16_C( 282), UINT16_C( 295), UINT16_C( 333),
2627
+ UINT16_C( 360), UINT16_C( 472), UINT16_C( 227), UINT16_C( 221)) },
2628
+ { simde_x_mm_set_epu8(UINT8_C(167), UINT8_C(231), UINT8_C(129), UINT8_C(236),
2629
+ UINT8_C(158), UINT8_C( 34), UINT8_C(219), UINT8_C( 52),
2630
+ UINT8_C(201), UINT8_C( 55), UINT8_C( 88), UINT8_C( 76),
2631
+ UINT8_C( 91), UINT8_C(192), UINT8_C( 80), UINT8_C(141)),
2632
+ simde_x_mm_set_epu8(UINT8_C(215), UINT8_C( 0), UINT8_C( 39), UINT8_C(166),
2633
+ UINT8_C( 94), UINT8_C(231), UINT8_C(224), UINT8_C(140),
2634
+ UINT8_C(153), UINT8_C( 38), UINT8_C(229), UINT8_C(239),
2635
+ UINT8_C(167), UINT8_C( 85), UINT8_C(242), UINT8_C(231)),
2636
+ simde_x_mm_set_epu16(UINT16_C( 350), UINT16_C( 573), UINT16_C( 237), UINT16_C( 385),
2637
+ UINT16_C( 448), UINT16_C( 329), UINT16_C( 458), UINT16_C( 208)) },
2638
+ { simde_x_mm_set_epu8(UINT8_C(125), UINT8_C( 70), UINT8_C(248), UINT8_C(193),
2639
+ UINT8_C(244), UINT8_C( 76), UINT8_C(163), UINT8_C(150),
2640
+ UINT8_C(169), UINT8_C( 24), UINT8_C(114), UINT8_C( 21),
2641
+ UINT8_C(171), UINT8_C( 13), UINT8_C(157), UINT8_C(178)),
2642
+ simde_x_mm_set_epu8(UINT8_C(198), UINT8_C(153), UINT8_C(205), UINT8_C(219),
2643
+ UINT8_C(198), UINT8_C( 83), UINT8_C( 62), UINT8_C( 96),
2644
+ UINT8_C(119), UINT8_C(213), UINT8_C( 90), UINT8_C(147),
2645
+ UINT8_C(232), UINT8_C( 64), UINT8_C(164), UINT8_C(129)),
2646
+ simde_x_mm_set_epu16(UINT16_C( 260), UINT16_C( 272), UINT16_C( 281), UINT16_C( 234),
2647
+ UINT16_C( 237), UINT16_C( 269), UINT16_C( 350), UINT16_C( 447)) },
2648
+ { simde_x_mm_set_epu8(UINT8_C( 35), UINT8_C(118), UINT8_C(198), UINT8_C( 85),
2649
+ UINT8_C( 8), UINT8_C(150), UINT8_C(121), UINT8_C(243),
2650
+ UINT8_C( 16), UINT8_C(221), UINT8_C(227), UINT8_C( 37),
2651
+ UINT8_C( 87), UINT8_C( 66), UINT8_C(216), UINT8_C(177)),
2652
+ simde_x_mm_set_epu8(UINT8_C(148), UINT8_C(250), UINT8_C( 9), UINT8_C(204),
2653
+ UINT8_C(136), UINT8_C(106), UINT8_C( 76), UINT8_C(162),
2654
+ UINT8_C(111), UINT8_C( 61), UINT8_C( 51), UINT8_C(174),
2655
+ UINT8_C(202), UINT8_C(248), UINT8_C( 64), UINT8_C(174)),
2656
+ simde_x_mm_set_epu16(UINT16_C( 354), UINT16_C( 270), UINT16_C( 529), UINT16_C( 391),
2657
+ UINT16_C( 553), UINT16_C( 58), UINT16_C( 564), UINT16_C( 546)) }
2658
+ };
2659
+
2660
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
2661
+ simde__m128i r = simde_mm_mpsadbw_epu8(test_vec[i].a, test_vec[i].b, 7);
2662
+ simde_assert_m128i_u16(r, ==, test_vec[i].r);
2663
+ }
2664
+
2665
+ return MUNIT_OK;
2666
+ }
2667
+
2668
+ static MunitResult
2669
+ test_simde_mm_mul_epi32(const MunitParameter params[], void* data) {
2670
+ (void) params;
2671
+ (void) data;
2672
+
2673
+ const struct {
2674
+ simde__m128i a;
2675
+ simde__m128i b;
2676
+ simde__m128i r;
2677
+ } test_vec[8] = {
2678
+ { simde_mm_set_epi32(INT32_C(-1795240134), INT32_C( -225235276), INT32_C( 260214100), INT32_C(-1437997118)),
2679
+ simde_mm_set_epi32(INT32_C(-1930516827), INT32_C(-2112091397), INT32_C(-2111031164), INT32_C( 1337416900)),
2680
+ simde_mm_set_epi64x(INT64_C( 475717488740520572), INT64_C(-1923201647764494200)) },
2681
+ { simde_mm_set_epi32(INT32_C(-2008734650), INT32_C( 1987866784), INT32_C( 1460586541), INT32_C( -308079008)),
2682
+ simde_mm_set_epi32(INT32_C( 1619497162), INT32_C( 1948444521), INT32_C(-1962326453), INT32_C( 355525826)),
2683
+ simde_mm_set_epi64x(INT64_C( 3873248143762690464), INT64_C( -109530043792460608)) },
2684
+ { simde_mm_set_epi32(INT32_C( -967525151), INT32_C( 610102911), INT32_C( -82193087), INT32_C( 610685272)),
2685
+ simde_mm_set_epi32(INT32_C( -464555473), INT32_C( 1911406284), INT32_C( 1504403418), INT32_C( 1044708348)),
2686
+ simde_mm_set_epi64x(INT64_C( 1166154537972092724), INT64_C( 637988001659050656)) },
2687
+ { simde_mm_set_epi32(INT32_C( 734959886), INT32_C( 727387476), INT32_C(-1178464916), INT32_C( 759159032)),
2688
+ simde_mm_set_epi32(INT32_C( 227151790), INT32_C( 2027589232), INT32_C( -510665157), INT32_C( 664410245)),
2689
+ simde_mm_set_epi64x(INT64_C( 1474843013829258432), INT64_C( 504393038445082840)) },
2690
+ { simde_mm_set_epi32(INT32_C( -981875510), INT32_C(-1517853387), INT32_C( 645890048), INT32_C( 1483649404)),
2691
+ simde_mm_set_epi32(INT32_C(-1176179108), INT32_C( -884770028), INT32_C( 1667209942), INT32_C( -150484050)),
2692
+ simde_mm_set_epi64x(INT64_C( 1342951183715884836), INT64_C( -223265571094006200)) },
2693
+ { simde_mm_set_epi32(INT32_C( -229613172), INT32_C( -220620915), INT32_C( 1178015644), INT32_C(-1087297921)),
2694
+ simde_mm_set_epi32(INT32_C(-1751226172), INT32_C(-1807182523), INT32_C( 804591673), INT32_C( -413406506)),
2695
+ simde_mm_set_epi64x(INT64_C( 398702261796268545), INT64_C( 449496034501674026)) },
2696
+ { simde_mm_set_epi32(INT32_C(-1537480837), INT32_C( -185479936), INT32_C( 386155243), INT32_C( 479622311)),
2697
+ simde_mm_set_epi32(INT32_C( -579116298), INT32_C( -674452315), INT32_C( -142346909), INT32_C(-1350064613)),
2698
+ simde_mm_set_epi64x(INT64_C( 125097372221251840), INT64_C( -647521109686380643)) },
2699
+ { simde_mm_set_epi32(INT32_C( 2146108348), INT32_C( 341648975), INT32_C( 901701854), INT32_C( -45673158)),
2700
+ simde_mm_set_epi32(INT32_C( 723948067), INT32_C( 1076365190), INT32_C( -693273363), INT32_C( 420311114)),
2701
+ simde_mm_set_epi64x(INT64_C( 367739063889180250), INT64_C( -19196935918878012)) }
2702
+ };
2703
+
2704
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
2705
+ simde__m128i r = simde_mm_mul_epi32(test_vec[i].a, test_vec[i].b);
2706
+ simde_assert_m128i_i32(r, ==, test_vec[i].r);
2707
+ }
2708
+
2709
+ return MUNIT_OK;
2710
+ }
2711
+
2712
+ static MunitResult
2713
+ test_simde_mm_mullo_epi32(const MunitParameter params[], void* data) {
2714
+ (void) params;
2715
+ (void) data;
2716
+
2717
+ const struct {
2718
+ simde__m128i a;
2719
+ simde__m128i b;
2720
+ simde__m128i r;
2721
+ } test_vec[8] = {
2722
+ { simde_mm_set_epi32(INT32_C( 1431531471), INT32_C(-1934656662), INT32_C( -638668877), INT32_C( 1966621031)),
2723
+ simde_mm_set_epi32(INT32_C( -646581607), INT32_C( 1232530244), INT32_C( 841047033), INT32_C(-1472432488)),
2724
+ simde_mm_set_epi32(INT32_C( 1865056439), INT32_C(-1849485784), INT32_C( 1840285211), INT32_C( 1940934440)) },
2725
+ { simde_mm_set_epi32(INT32_C( 2119691862), INT32_C( 1656188631), INT32_C( 949015012), INT32_C( -645064268)),
2726
+ simde_mm_set_epi32(INT32_C( 1854664684), INT32_C( 1364588379), INT32_C( -185048322), INT32_C( 1546431794)),
2727
+ simde_mm_set_epi32(INT32_C( -314310328), INT32_C( -410462355), INT32_C( 887719992), INT32_C( -220354264)) },
2728
+ { simde_mm_set_epi32(INT32_C( 390191893), INT32_C( 1246080619), INT32_C( 2032818940), INT32_C(-1664549536)),
2729
+ simde_mm_set_epi32(INT32_C( 1802476280), INT32_C(-1492538899), INT32_C( -866846002), INT32_C(-1929477089)),
2730
+ simde_mm_set_epi32(INT32_C( -357441960), INT32_C(-1318446065), INT32_C(-2019602744), INT32_C(-2076430688)) },
2731
+ { simde_mm_set_epi32(INT32_C( -629988407), INT32_C(-1719063308), INT32_C( -231912988), INT32_C( 1462330866)),
2732
+ simde_mm_set_epi32(INT32_C(-1320599620), INT32_C( -769970788), INT32_C( -625134420), INT32_C( 293199422)),
2733
+ simde_mm_set_epi32(INT32_C( -273826148), INT32_C(-1841674064), INT32_C(-1631324880), INT32_C( 1137740956)) },
2734
+ { simde_mm_set_epi32(INT32_C(-1221152076), INT32_C(-1223458578), INT32_C(-1263311025), INT32_C( 2089857790)),
2735
+ simde_mm_set_epi32(INT32_C(-2068401725), INT32_C( 1531524331), INT32_C( 1064638153), INT32_C( -843878287)),
2736
+ simde_mm_set_epi32(INT32_C( 1009816348), INT32_C( 1660269178), INT32_C( -203954425), INT32_C( -674749922)) },
2737
+ { simde_mm_set_epi32(INT32_C(-1283522478), INT32_C( 524522527), INT32_C( 1152886109), INT32_C( -438302719)),
2738
+ simde_mm_set_epi32(INT32_C(-2012975476), INT32_C( -67951228), INT32_C( 258710699), INT32_C( -25585755)),
2739
+ simde_mm_set_epi32(INT32_C( 1289709784), INT32_C( -508699908), INT32_C( 1615798559), INT32_C( 345423781)) },
2740
+ { simde_mm_set_epi32(INT32_C( 199098300), INT32_C(-1941350427), INT32_C(-1170519288), INT32_C( 2117275260)),
2741
+ simde_mm_set_epi32(INT32_C( 1736533752), INT32_C( 668304389), INT32_C(-1871663356), INT32_C( -231105582)),
2742
+ simde_mm_set_epi32(INT32_C( 1957733920), INT32_C(-1823880327), INT32_C( -994597856), INT32_C( -515705416)) },
2743
+ { simde_mm_set_epi32(INT32_C( 796248112), INT32_C( -74441002), INT32_C( 756194998), INT32_C( 128690934)),
2744
+ simde_mm_set_epi32(INT32_C( -934813484), INT32_C( 1495167051), INT32_C( 549239917), INT32_C( 1315900336)),
2745
+ simde_mm_set_epi32(INT32_C( 1644414912), INT32_C(-1010128718), INT32_C( 723729278), INT32_C(-1157096672)) }
2746
+ };
2747
+
2748
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
2749
+ simde__m128i r = simde_mm_mullo_epi32(test_vec[i].a, test_vec[i].b);
2750
+ simde_assert_m128i_i32(r, ==, test_vec[i].r);
2751
+ }
2752
+
2753
+ return MUNIT_OK;
2754
+ }
2755
+
2756
+ static MunitResult
2757
+ test_simde_mm_packus_epi32(const MunitParameter params[], void* data) {
2758
+ (void) params;
2759
+ (void) data;
2760
+
2761
+ const struct {
2762
+ simde__m128i a;
2763
+ simde__m128i b;
2764
+ simde__m128i r;
2765
+ } test_vec[8] = {
2766
+ { simde_mm_set_epi32(INT32_C(-1174506975), INT32_C(-1622542395), INT32_C( -228565455), INT32_C( 1722092951)),
2767
+ simde_mm_set_epi32(INT32_C( -188735178), INT32_C( -859437923), INT32_C( 40707), INT32_C(-2078308106)),
2768
+ simde_x_mm_set_epu16(UINT16_C( 0), UINT16_C( 0), UINT16_C(40707), UINT16_C( 0),
2769
+ UINT16_C( 0), UINT16_C( 0), UINT16_C( 0), UINT16_C(65535)) },
2770
+ { simde_mm_set_epi32(INT32_C( 279157042), INT32_C( -664033471), INT32_C( 1210432777), INT32_C(-1162534366)),
2771
+ simde_mm_set_epi32(INT32_C( 330673804), INT32_C(-1666667776), INT32_C(-1924642427), INT32_C(-1118465206)),
2772
+ simde_x_mm_set_epu16(UINT16_C(65535), UINT16_C( 0), UINT16_C( 0), UINT16_C( 0),
2773
+ UINT16_C(65535), UINT16_C( 0), UINT16_C(65535), UINT16_C( 0)) },
2774
+ { simde_mm_set_epi32(INT32_C(-1675273454), INT32_C( 252374167), INT32_C(-1898193118), INT32_C(-2121147721)),
2775
+ simde_mm_set_epi32(INT32_C(-1175340609), INT32_C( 1859609595), INT32_C( 305301369), INT32_C( 1936442403)),
2776
+ simde_x_mm_set_epu16(UINT16_C( 0), UINT16_C(65535), UINT16_C(65535), UINT16_C(65535),
2777
+ UINT16_C( 0), UINT16_C(65535), UINT16_C( 0), UINT16_C( 0)) },
2778
+ { simde_mm_set_epi32(INT32_C( -372493069), INT32_C(-1071911212), INT32_C( 122218483), INT32_C( 1644546153)),
2779
+ simde_mm_set_epi32(INT32_C( -603471664), INT32_C( -435236634), INT32_C(-1237577227), INT32_C( 40156)),
2780
+ simde_x_mm_set_epu16(UINT16_C( 0), UINT16_C( 0), UINT16_C( 0), UINT16_C(40156),
2781
+ UINT16_C( 0), UINT16_C( 0), UINT16_C(65535), UINT16_C(65535)) },
2782
+ { simde_mm_set_epi32(INT32_C( -132052641), INT32_C( 1567528732), INT32_C( -1505776), INT32_C( 422872655)),
2783
+ simde_mm_set_epi32(INT32_C( 1525337323), INT32_C(-1409556892), INT32_C( -388396270), INT32_C(-2086963765)),
2784
+ simde_x_mm_set_epu16(UINT16_C(65535), UINT16_C( 0), UINT16_C( 0), UINT16_C( 0),
2785
+ UINT16_C( 0), UINT16_C(65535), UINT16_C( 0), UINT16_C(65535)) },
2786
+ { simde_mm_set_epi32(INT32_C(-1206563380), INT32_C( 1557268555), INT32_C( -872999034), INT32_C( 234221095)),
2787
+ simde_mm_set_epi32(INT32_C( 723455818), INT32_C( 773436940), INT32_C( 326719262), INT32_C( -826281673)),
2788
+ simde_x_mm_set_epu16(UINT16_C(65535), UINT16_C(65535), UINT16_C(65535), UINT16_C( 0),
2789
+ UINT16_C( 0), UINT16_C(65535), UINT16_C( 0), UINT16_C(65535)) },
2790
+ { simde_mm_set_epi32(INT32_C( -738729427), INT32_C( 1061139854), INT32_C(-2031157453), INT32_C( 2064098327)),
2791
+ simde_mm_set_epi32(INT32_C(-1034678142), INT32_C( 97345786), INT32_C( 1187222665), INT32_C( 61643)),
2792
+ simde_x_mm_set_epu16(UINT16_C( 0), UINT16_C(65535), UINT16_C(65535), UINT16_C(61643),
2793
+ UINT16_C( 0), UINT16_C(65535), UINT16_C( 0), UINT16_C(65535)) },
2794
+ { simde_mm_set_epi32(INT32_C(-2037944282), INT32_C( -355400074), INT32_C( -341422900), INT32_C(-1605623445)),
2795
+ simde_mm_set_epi32(INT32_C( 629990551), INT32_C( 397893118), INT32_C(-1310491239), INT32_C(-2050362652)),
2796
+ simde_x_mm_set_epu16(UINT16_C(65535), UINT16_C(65535), UINT16_C( 0), UINT16_C( 0),
2797
+ UINT16_C( 0), UINT16_C( 0), UINT16_C( 0), UINT16_C( 0)) }
2798
+ };
2799
+
2800
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
2801
+ simde__m128i r = simde_mm_packus_epi32(test_vec[i].a, test_vec[i].b);
2802
+ simde_assert_m128i_i32(r, ==, test_vec[i].r);
2803
+ }
2804
+
2805
+ return MUNIT_OK;
2806
+ }
2807
+
2808
+ static MunitResult
2809
+ test_simde_mm_round_pd(const MunitParameter params[], void* data) {
2810
+ (void) params;
2811
+ (void) data;
2812
+
2813
+ const struct {
2814
+ simde__m128d a;
2815
+ int rounding;
2816
+ simde__m128d r;
2817
+ } test_vec[8] = {
2818
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( 959.46), SIMDE_FLOAT64_C( -646.56)),
2819
+ SIMDE_MM_FROUND_TO_NEAREST_INT,
2820
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 959.00), SIMDE_FLOAT64_C( -647.00)) },
2821
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( -813.43), SIMDE_FLOAT64_C( 515.75)),
2822
+ SIMDE_MM_FROUND_TO_NEAREST_INT,
2823
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -813.00), SIMDE_FLOAT64_C( 516.00)) },
2824
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( -365.43), SIMDE_FLOAT64_C( 840.28)),
2825
+ SIMDE_MM_FROUND_TO_NEG_INF,
2826
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -366.00), SIMDE_FLOAT64_C( 840.00)) },
2827
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( -843.69), SIMDE_FLOAT64_C( -641.36)),
2828
+ SIMDE_MM_FROUND_TO_NEG_INF,
2829
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -844.00), SIMDE_FLOAT64_C( -642.00)) },
2830
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( -542.64), SIMDE_FLOAT64_C( -757.34)),
2831
+ SIMDE_MM_FROUND_TO_POS_INF,
2832
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -542.00), SIMDE_FLOAT64_C( -757.00)) },
2833
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( -593.16), SIMDE_FLOAT64_C( 973.59)),
2834
+ SIMDE_MM_FROUND_TO_POS_INF,
2835
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -593.00), SIMDE_FLOAT64_C( 974.00)) },
2836
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( 54.67), SIMDE_FLOAT64_C( -54.55)),
2837
+ SIMDE_MM_FROUND_TO_ZERO,
2838
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 54.00), SIMDE_FLOAT64_C( -54.00)) },
2839
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( -31.89), SIMDE_FLOAT64_C( 136.86)),
2840
+ SIMDE_MM_FROUND_TO_ZERO,
2841
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -31.00), SIMDE_FLOAT64_C( 136.00)) }
2842
+ };
2843
+
2844
+ for (size_t i = 0 ; i < 2 ; i++) {
2845
+ simde__m128d r = simde_mm_round_pd(test_vec[i].a, SIMDE_MM_FROUND_TO_NEAREST_INT);
2846
+ simde_assert_m128d_close(r, test_vec[i].r, 1);
2847
+ }
2848
+ for (size_t i = 2 ; i < 4 ; i++) {
2849
+ simde__m128d r = simde_mm_round_pd(test_vec[i].a, SIMDE_MM_FROUND_TO_NEG_INF);
2850
+ simde_assert_m128d_close(r, test_vec[i].r, 1);
2851
+ }
2852
+ for (size_t i = 4 ; i < 6 ; i++) {
2853
+ simde__m128d r = simde_mm_round_pd(test_vec[i].a, SIMDE_MM_FROUND_TO_POS_INF);
2854
+ simde_assert_m128d_close(r, test_vec[i].r, 1);
2855
+ }
2856
+ for (size_t i = 6 ; i < 8 ; i++) {
2857
+ simde__m128d r = simde_mm_round_pd(test_vec[i].a, SIMDE_MM_FROUND_TO_ZERO);
2858
+ simde_assert_m128d_close(r, test_vec[i].r, 1);
2859
+ }
2860
+
2861
+ return MUNIT_OK;
2862
+ }
2863
+
2864
+ static MunitResult
2865
+ test_simde_mm_round_ps(const MunitParameter params[], void* data) {
2866
+ (void) params;
2867
+ (void) data;
2868
+
2869
+ const struct {
2870
+ simde__m128 a;
2871
+ int rounding;
2872
+ simde__m128 r;
2873
+ } test_vec[8] = {
2874
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( 201.21), SIMDE_FLOAT32_C( -972.67), SIMDE_FLOAT32_C( -880.37), SIMDE_FLOAT32_C( 126.05)),
2875
+ SIMDE_MM_FROUND_TO_NEAREST_INT,
2876
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 201.00), SIMDE_FLOAT32_C( -973.00), SIMDE_FLOAT32_C( -880.00), SIMDE_FLOAT32_C( 126.00)) },
2877
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( -339.94), SIMDE_FLOAT32_C( 81.10), SIMDE_FLOAT32_C( 664.61), SIMDE_FLOAT32_C( 207.74)),
2878
+ SIMDE_MM_FROUND_TO_NEAREST_INT,
2879
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -340.00), SIMDE_FLOAT32_C( 81.00), SIMDE_FLOAT32_C( 665.00), SIMDE_FLOAT32_C( 208.00)) },
2880
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( -476.74), SIMDE_FLOAT32_C( 320.44), SIMDE_FLOAT32_C( -89.85), SIMDE_FLOAT32_C( -531.60)),
2881
+ SIMDE_MM_FROUND_TO_NEG_INF,
2882
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -477.00), SIMDE_FLOAT32_C( 320.00), SIMDE_FLOAT32_C( -90.00), SIMDE_FLOAT32_C( -532.00)) },
2883
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( -802.32), SIMDE_FLOAT32_C( -381.83), SIMDE_FLOAT32_C( -743.96), SIMDE_FLOAT32_C( -180.00)),
2884
+ SIMDE_MM_FROUND_TO_NEG_INF,
2885
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -803.00), SIMDE_FLOAT32_C( -382.00), SIMDE_FLOAT32_C( -744.00), SIMDE_FLOAT32_C( -180.00)) },
2886
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( 231.59), SIMDE_FLOAT32_C( 681.41), SIMDE_FLOAT32_C( 561.95), SIMDE_FLOAT32_C( -598.34)),
2887
+ SIMDE_MM_FROUND_TO_POS_INF,
2888
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 232.00), SIMDE_FLOAT32_C( 682.00), SIMDE_FLOAT32_C( 562.00), SIMDE_FLOAT32_C( -598.00)) },
2889
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( -440.45), SIMDE_FLOAT32_C( 343.18), SIMDE_FLOAT32_C( 88.66), SIMDE_FLOAT32_C( -98.54)),
2890
+ SIMDE_MM_FROUND_TO_POS_INF,
2891
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -440.00), SIMDE_FLOAT32_C( 344.00), SIMDE_FLOAT32_C( 89.00), SIMDE_FLOAT32_C( -98.00)) },
2892
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( -128.87), SIMDE_FLOAT32_C( -558.72), SIMDE_FLOAT32_C( 864.66), SIMDE_FLOAT32_C( -576.19)),
2893
+ SIMDE_MM_FROUND_TO_ZERO,
2894
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -128.00), SIMDE_FLOAT32_C( -558.00), SIMDE_FLOAT32_C( 864.00), SIMDE_FLOAT32_C( -576.00)) },
2895
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( 669.69), SIMDE_FLOAT32_C( 902.50), SIMDE_FLOAT32_C( -479.42), SIMDE_FLOAT32_C( 565.85)),
2896
+ SIMDE_MM_FROUND_TO_ZERO,
2897
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 669.00), SIMDE_FLOAT32_C( 902.00), SIMDE_FLOAT32_C( -479.00), SIMDE_FLOAT32_C( 565.00)) }
2898
+ };
2899
+
2900
+ for (size_t i = 0 ; i < 2 ; i++) {
2901
+ simde__m128 r = simde_mm_round_ps(test_vec[i].a, SIMDE_MM_FROUND_TO_NEAREST_INT);
2902
+ simde_assert_m128_close(r, test_vec[i].r, 1);
2903
+ }
2904
+ for (size_t i = 2 ; i < 4 ; i++) {
2905
+ simde__m128 r = simde_mm_round_ps(test_vec[i].a, SIMDE_MM_FROUND_TO_NEG_INF);
2906
+ simde_assert_m128_close(r, test_vec[i].r, 1);
2907
+ }
2908
+ for (size_t i = 4 ; i < 6 ; i++) {
2909
+ simde__m128 r = simde_mm_round_ps(test_vec[i].a, SIMDE_MM_FROUND_TO_POS_INF);
2910
+ simde_assert_m128_close(r, test_vec[i].r, 1);
2911
+ }
2912
+ for (size_t i = 6 ; i < 8 ; i++) {
2913
+ simde__m128 r = simde_mm_round_ps(test_vec[i].a, SIMDE_MM_FROUND_TO_ZERO);
2914
+ simde_assert_m128_close(r, test_vec[i].r, 1);
2915
+ }
2916
+
2917
+ return MUNIT_OK;
2918
+ }
2919
+
2920
+ static MunitResult
2921
+ test_simde_mm_round_sd(const MunitParameter params[], void* data) {
2922
+ (void) params;
2923
+ (void) data;
2924
+
2925
+ const struct {
2926
+ simde__m128d a;
2927
+ simde__m128d b;
2928
+ int rounding;
2929
+ simde__m128d r;
2930
+ } test_vec[8] = {
2931
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( -57.59), SIMDE_FLOAT64_C( -91.48)),
2932
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -114.01), SIMDE_FLOAT64_C( 129.18)),
2933
+ SIMDE_MM_FROUND_TO_NEAREST_INT,
2934
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -57.59), SIMDE_FLOAT64_C( 129.00)) },
2935
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( 367.02), SIMDE_FLOAT64_C( -228.73)),
2936
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -106.57), SIMDE_FLOAT64_C( -248.85)),
2937
+ SIMDE_MM_FROUND_TO_NEAREST_INT,
2938
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 367.02), SIMDE_FLOAT64_C( -249.00)) },
2939
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( 218.19), SIMDE_FLOAT64_C( -121.02)),
2940
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -367.84), SIMDE_FLOAT64_C( -492.19)),
2941
+ SIMDE_MM_FROUND_TO_NEG_INF,
2942
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 218.19), SIMDE_FLOAT64_C( -493.00)) },
2943
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( 706.26), SIMDE_FLOAT64_C( 710.85)),
2944
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -391.17), SIMDE_FLOAT64_C( -834.44)),
2945
+ SIMDE_MM_FROUND_TO_NEG_INF,
2946
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 706.26), SIMDE_FLOAT64_C( -835.00)) },
2947
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( -784.90), SIMDE_FLOAT64_C( -810.73)),
2948
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -554.44), SIMDE_FLOAT64_C( -463.60)),
2949
+ SIMDE_MM_FROUND_TO_POS_INF,
2950
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -784.90), SIMDE_FLOAT64_C( -463.00)) },
2951
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( -997.83), SIMDE_FLOAT64_C( 6.91)),
2952
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -51.24), SIMDE_FLOAT64_C( -801.09)),
2953
+ SIMDE_MM_FROUND_TO_POS_INF,
2954
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -997.83), SIMDE_FLOAT64_C( -801.00)) },
2955
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( -237.53), SIMDE_FLOAT64_C( 215.82)),
2956
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -338.32), SIMDE_FLOAT64_C( -289.00)),
2957
+ SIMDE_MM_FROUND_TO_ZERO,
2958
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -237.53), SIMDE_FLOAT64_C( -289.00)) },
2959
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( 964.37), SIMDE_FLOAT64_C( -304.03)),
2960
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 237.14), SIMDE_FLOAT64_C( 621.16)),
2961
+ SIMDE_MM_FROUND_TO_ZERO,
2962
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 964.37), SIMDE_FLOAT64_C( 621.00)) }
2963
+ };
2964
+
2965
+ for (size_t i = 0 ; i < 2 ; i++) {
2966
+ simde__m128d r = simde_mm_round_sd(test_vec[i].a, test_vec[i].b, SIMDE_MM_FROUND_TO_NEAREST_INT);
2967
+ simde_assert_m128d_close(r, test_vec[i].r, 1);
2968
+ }
2969
+ for (size_t i = 2 ; i < 4 ; i++) {
2970
+ simde__m128d r = simde_mm_round_sd(test_vec[i].a, test_vec[i].b, SIMDE_MM_FROUND_TO_NEG_INF);
2971
+ simde_assert_m128d_close(r, test_vec[i].r, 1);
2972
+ }
2973
+ for (size_t i = 4 ; i < 6 ; i++) {
2974
+ simde__m128d r = simde_mm_round_sd(test_vec[i].a, test_vec[i].b, SIMDE_MM_FROUND_TO_POS_INF);
2975
+ simde_assert_m128d_close(r, test_vec[i].r, 1);
2976
+ }
2977
+ for (size_t i = 6 ; i < 8 ; i++) {
2978
+ simde__m128d r = simde_mm_round_sd(test_vec[i].a, test_vec[i].b, SIMDE_MM_FROUND_TO_ZERO);
2979
+ simde_assert_m128d_close(r, test_vec[i].r, 1);
2980
+ }
2981
+
2982
+ return MUNIT_OK;
2983
+ }
2984
+
2985
+ static MunitResult
2986
+ test_simde_mm_round_ss(const MunitParameter params[], void* data) {
2987
+ (void) params;
2988
+ (void) data;
2989
+
2990
+ const struct {
2991
+ simde__m128 a;
2992
+ simde__m128 b;
2993
+ int rounding;
2994
+ simde__m128 r;
2995
+ } test_vec[8] = {
2996
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( -963.12), SIMDE_FLOAT32_C( -638.96), SIMDE_FLOAT32_C( -855.75), SIMDE_FLOAT32_C( 789.49)),
2997
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 969.82), SIMDE_FLOAT32_C( 14.36), SIMDE_FLOAT32_C( 60.33), SIMDE_FLOAT32_C( -666.61)),
2998
+ SIMDE_MM_FROUND_TO_NEAREST_INT,
2999
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -963.12), SIMDE_FLOAT32_C( -638.96), SIMDE_FLOAT32_C( -855.75), SIMDE_FLOAT32_C( -667.00)) },
3000
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( -53.45), SIMDE_FLOAT32_C( 397.58), SIMDE_FLOAT32_C( 386.64), SIMDE_FLOAT32_C( 779.00)),
3001
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 497.83), SIMDE_FLOAT32_C( -657.35), SIMDE_FLOAT32_C( -712.13), SIMDE_FLOAT32_C( 600.92)),
3002
+ SIMDE_MM_FROUND_TO_NEAREST_INT,
3003
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -53.45), SIMDE_FLOAT32_C( 397.58), SIMDE_FLOAT32_C( 386.64), SIMDE_FLOAT32_C( 601.00)) },
3004
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( -527.69), SIMDE_FLOAT32_C( -487.47), SIMDE_FLOAT32_C( -52.23), SIMDE_FLOAT32_C( 18.71)),
3005
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -912.96), SIMDE_FLOAT32_C( 897.40), SIMDE_FLOAT32_C( -873.59), SIMDE_FLOAT32_C( 52.18)),
3006
+ SIMDE_MM_FROUND_TO_NEG_INF,
3007
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -527.69), SIMDE_FLOAT32_C( -487.47), SIMDE_FLOAT32_C( -52.23), SIMDE_FLOAT32_C( 52.00)) },
3008
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( 857.48), SIMDE_FLOAT32_C( -849.82), SIMDE_FLOAT32_C( 971.09), SIMDE_FLOAT32_C( -929.13)),
3009
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 871.04), SIMDE_FLOAT32_C( 284.36), SIMDE_FLOAT32_C( 561.82), SIMDE_FLOAT32_C( -146.68)),
3010
+ SIMDE_MM_FROUND_TO_NEG_INF,
3011
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 857.48), SIMDE_FLOAT32_C( -849.82), SIMDE_FLOAT32_C( 971.09), SIMDE_FLOAT32_C( -147.00)) },
3012
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( -859.81), SIMDE_FLOAT32_C( -105.89), SIMDE_FLOAT32_C( -18.60), SIMDE_FLOAT32_C( -214.59)),
3013
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -360.47), SIMDE_FLOAT32_C( 472.18), SIMDE_FLOAT32_C( 960.99), SIMDE_FLOAT32_C( -396.53)),
3014
+ SIMDE_MM_FROUND_TO_POS_INF,
3015
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -859.81), SIMDE_FLOAT32_C( -105.89), SIMDE_FLOAT32_C( -18.60), SIMDE_FLOAT32_C( -396.00)) },
3016
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( -657.82), SIMDE_FLOAT32_C( -922.27), SIMDE_FLOAT32_C( -130.03), SIMDE_FLOAT32_C( 877.75)),
3017
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -413.60), SIMDE_FLOAT32_C( 824.69), SIMDE_FLOAT32_C( 124.72), SIMDE_FLOAT32_C( 312.34)),
3018
+ SIMDE_MM_FROUND_TO_POS_INF,
3019
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -657.82), SIMDE_FLOAT32_C( -922.27), SIMDE_FLOAT32_C( -130.03), SIMDE_FLOAT32_C( 313.00)) },
3020
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( -636.42), SIMDE_FLOAT32_C( -978.69), SIMDE_FLOAT32_C( 787.73), SIMDE_FLOAT32_C( -705.33)),
3021
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -179.55), SIMDE_FLOAT32_C( 391.06), SIMDE_FLOAT32_C( -805.63), SIMDE_FLOAT32_C( 831.10)),
3022
+ SIMDE_MM_FROUND_TO_ZERO,
3023
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -636.42), SIMDE_FLOAT32_C( -978.69), SIMDE_FLOAT32_C( 787.73), SIMDE_FLOAT32_C( 831.00)) },
3024
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( 529.08), SIMDE_FLOAT32_C( -665.00), SIMDE_FLOAT32_C( 786.13), SIMDE_FLOAT32_C( 843.11)),
3025
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -295.93), SIMDE_FLOAT32_C( -148.90), SIMDE_FLOAT32_C( 483.10), SIMDE_FLOAT32_C( -790.07)),
3026
+ SIMDE_MM_FROUND_TO_ZERO,
3027
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 529.08), SIMDE_FLOAT32_C( -665.00), SIMDE_FLOAT32_C( 786.13), SIMDE_FLOAT32_C( -790.00)) }
3028
+ };
3029
+
3030
+ for (size_t i = 0 ; i < 2 ; i++) {
3031
+ simde__m128 r = simde_mm_round_ss(test_vec[i].a, test_vec[i].b, SIMDE_MM_FROUND_TO_NEAREST_INT);
3032
+ simde_assert_m128_close(r, test_vec[i].r, 1);
3033
+ }
3034
+ for (size_t i = 2 ; i < 4 ; i++) {
3035
+ simde__m128 r = simde_mm_round_ss(test_vec[i].a, test_vec[i].b, SIMDE_MM_FROUND_TO_NEG_INF);
3036
+ simde_assert_m128_close(r, test_vec[i].r, 1);
3037
+ }
3038
+ for (size_t i = 4 ; i < 6 ; i++) {
3039
+ simde__m128 r = simde_mm_round_ss(test_vec[i].a, test_vec[i].b, SIMDE_MM_FROUND_TO_POS_INF);
3040
+ simde_assert_m128_close(r, test_vec[i].r, 1);
3041
+ }
3042
+ for (size_t i = 6 ; i < 8 ; i++) {
3043
+ simde__m128 r = simde_mm_round_ss(test_vec[i].a, test_vec[i].b, SIMDE_MM_FROUND_TO_ZERO);
3044
+ simde_assert_m128_close(r, test_vec[i].r, 1);
3045
+ }
3046
+
3047
+ return MUNIT_OK;
3048
+ }
3049
+
3050
+ static MunitResult
3051
+ test_simde_mm_stream_load_si128(const MunitParameter params[], void* data) {
3052
+ (void) params;
3053
+ (void) data;
3054
+
3055
+ const struct {
3056
+ simde__m128i a;
3057
+ simde__m128i r;
3058
+ } test_vec[8] = {
3059
+ { simde_mm_set_epi32(INT32_C( -547904967), INT32_C(-1362990942), INT32_C( 2066137598), INT32_C(-1778922116)),
3060
+ simde_mm_set_epi32(INT32_C( -547904967), INT32_C(-1362990942), INT32_C( 2066137598), INT32_C(-1778922116)) },
3061
+ { simde_mm_set_epi32(INT32_C(-2087163583), INT32_C( 2115433203), INT32_C(-1978815624), INT32_C( 1025228414)),
3062
+ simde_mm_set_epi32(INT32_C(-2087163583), INT32_C( 2115433203), INT32_C(-1978815624), INT32_C( 1025228414)) },
3063
+ { simde_mm_set_epi32(INT32_C(-1244490055), INT32_C( 1943548537), INT32_C( 589236062), INT32_C( -751161899)),
3064
+ simde_mm_set_epi32(INT32_C(-1244490055), INT32_C( 1943548537), INT32_C( 589236062), INT32_C( -751161899)) },
3065
+ { simde_mm_set_epi32(INT32_C(-1274215918), INT32_C(-1858225286), INT32_C( -472727069), INT32_C( 528679049)),
3066
+ simde_mm_set_epi32(INT32_C(-1274215918), INT32_C(-1858225286), INT32_C( -472727069), INT32_C( 528679049)) },
3067
+ { simde_mm_set_epi32(INT32_C(-1568552575), INT32_C( 411221897), INT32_C(-1334506552), INT32_C( 1554165859)),
3068
+ simde_mm_set_epi32(INT32_C(-1568552575), INT32_C( 411221897), INT32_C(-1334506552), INT32_C( 1554165859)) },
3069
+ { simde_mm_set_epi32(INT32_C( 1044144940), INT32_C(-1916851863), INT32_C( -589120926), INT32_C( 651795910)),
3070
+ simde_mm_set_epi32(INT32_C( 1044144940), INT32_C(-1916851863), INT32_C( -589120926), INT32_C( 651795910)) },
3071
+ { simde_mm_set_epi32(INT32_C( -25575503), INT32_C(-1782121708), INT32_C( 751836326), INT32_C( 1005598033)),
3072
+ simde_mm_set_epi32(INT32_C( -25575503), INT32_C(-1782121708), INT32_C( 751836326), INT32_C( 1005598033)) },
3073
+ { simde_mm_set_epi32(INT32_C(-1873195901), INT32_C( -450900536), INT32_C( 904584108), INT32_C( 251416593)),
3074
+ simde_mm_set_epi32(INT32_C(-1873195901), INT32_C( -450900536), INT32_C( 904584108), INT32_C( 251416593)) }
3075
+ };
3076
+
3077
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
3078
+ simde__m128i r = simde_mm_stream_load_si128(&(test_vec[i].a));
3079
+ simde_assert_m128i_i32(r, ==, test_vec[i].r);
3080
+ }
3081
+
3082
+ return MUNIT_OK;
3083
+ }
3084
+
3085
+ static MunitResult
3086
+ test_simde_mm_test_all_ones(const MunitParameter params[], void* data) {
3087
+ (void) params;
3088
+ (void) data;
3089
+
3090
+ const struct {
3091
+ simde__m128i a;
3092
+ int r;
3093
+ } test_vec[8] = {
3094
+ { simde_x_mm_set_epu32(~UINT32_C( 0), UINT32_C(1993838502), UINT32_C(1216384299), UINT32_C(2484321284)),
3095
+ 0 },
3096
+ { simde_x_mm_set_epu32(UINT32_C( 708002161), ~UINT32_C( 0), UINT32_C(1641125312), UINT32_C(1784735378)),
3097
+ 0 },
3098
+ { simde_x_mm_set_epu32(UINT32_C(3153393974), UINT32_C(3299961049), ~UINT32_C( 0), UINT32_C( 256198680)),
3099
+ 0 },
3100
+ { simde_x_mm_set_epu32(UINT32_C(3219291087), UINT32_C(2995285646), UINT32_C(3353269087), ~UINT32_C( 0)),
3101
+ 0 },
3102
+ { simde_x_mm_set_epu32(~UINT32_C( 0), UINT32_C(1590943288), UINT32_C(4263015539), UINT32_C(1245312939)),
3103
+ 0 },
3104
+ { simde_x_mm_set_epu32(~UINT32_C( 0), ~UINT32_C( 0), UINT32_C( 453365074), UINT32_C( 94225678)),
3105
+ 0 },
3106
+ { simde_x_mm_set_epu32(~UINT32_C( 0), ~UINT32_C( 0), ~UINT32_C( 0), UINT32_C(3354153743)),
3107
+ 0 },
3108
+ { simde_x_mm_set_epu32(~UINT32_C( 0), ~UINT32_C( 0), ~UINT32_C( 0), ~UINT32_C( 0)),
3109
+ 1 }
3110
+ };
3111
+
3112
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
3113
+ int r = simde_mm_test_all_ones(test_vec[i].a);
3114
+ munit_assert_int(r, ==, test_vec[i].r);
3115
+ }
3116
+
3117
+ return MUNIT_OK;
3118
+ }
3119
+
3120
+ static MunitResult
3121
+ test_simde_mm_test_all_zeros(const MunitParameter params[], void* data) {
3122
+ (void) params;
3123
+ (void) data;
3124
+
3125
+ const struct {
3126
+ simde__m128i a;
3127
+ simde__m128i b;
3128
+ int r;
3129
+ } test_vec[8] = {
3130
+ { simde_x_mm_set_epu32(UINT32_C(1201894317), UINT32_C(2319231887), UINT32_C(1402528928), UINT32_C(3687328034)),
3131
+ simde_x_mm_set_epu32(UINT32_C(1967137742), UINT32_C(1472851836), UINT32_C(2181608867), UINT32_C(1889365185)),
3132
+ 0 },
3133
+ { simde_x_mm_set_epu32(UINT32_C(2135448036), UINT32_C(1619848109), UINT32_C(2289839916), UINT32_C(3523984647)),
3134
+ simde_x_mm_set_epu32(UINT32_C(2046676780), UINT32_C(1214034345), UINT32_C( 106141409), UINT32_C(4219215983)),
3135
+ 0 },
3136
+ { simde_x_mm_set_epu32(UINT32_C( 657893603), UINT32_C(3270868697), UINT32_C(1314977284), UINT32_C(2544509951)),
3137
+ simde_x_mm_set_epu32(UINT32_C(2424312328), UINT32_C( 687898658), UINT32_C(2425031026), UINT32_C( 672223232)),
3138
+ 1 },
3139
+ { simde_x_mm_set_epu32(UINT32_C(3537640743), UINT32_C( 372152333), UINT32_C(4133234814), UINT32_C( 396143667)),
3140
+ simde_x_mm_set_epu32(UINT32_C( 671105736), UINT32_C(3909093360), UINT32_C( 8553473), UINT32_C( 541283144)),
3141
+ 1 },
3142
+ { simde_x_mm_set_epu32(UINT32_C(2182578541), UINT32_C(4064427053), UINT32_C(2602515508), UINT32_C(4049235221)),
3143
+ simde_x_mm_set_epu32(UINT32_C( 203975314), UINT32_C( 227393856), UINT32_C( 536871690), UINT32_C( 75760138)),
3144
+ 1 },
3145
+ { simde_x_mm_set_epu32(UINT32_C( 152354605), UINT32_C(2176274914), UINT32_C(3011079004), UINT32_C(3847789567)),
3146
+ simde_x_mm_set_epu32(UINT32_C(1562291537), UINT32_C( 354162251), UINT32_C(1134881489), UINT32_C(2956892685)),
3147
+ 0 },
3148
+ { simde_x_mm_set_epu32(UINT32_C(2096532328), UINT32_C(1165454747), UINT32_C(3275710596), UINT32_C(1593733565)),
3149
+ simde_x_mm_set_epu32(UINT32_C(1988174269), UINT32_C(1339821769), UINT32_C(1810810088), UINT32_C( 840755548)),
3150
+ 0 },
3151
+ { simde_x_mm_set_epu32(UINT32_C(3588833624), UINT32_C( 610042197), UINT32_C(1310204993), UINT32_C(1509811379)),
3152
+ simde_x_mm_set_epu32(UINT32_C( 571869223), UINT32_C(1235317762), UINT32_C( 29362474), UINT32_C( 570561100)),
3153
+ 1 }
3154
+ };
3155
+
3156
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
3157
+ int r = simde_mm_test_all_zeros(test_vec[i].a, test_vec[i].b);
3158
+ munit_assert_int(r, ==, test_vec[i].r);
3159
+ }
3160
+
3161
+ return MUNIT_OK;
3162
+ }
3163
+
3164
+ static MunitResult
3165
+ test_simde_mm_test_mix_ones_zeros(const MunitParameter params[], void* data) {
3166
+ (void) params;
3167
+ (void) data;
3168
+
3169
+ const struct {
3170
+ simde__m128i a;
3171
+ simde__m128i b;
3172
+ int r;
3173
+ } test_vec[8] = {
3174
+ { simde_x_mm_set_epu32(UINT32_C( 817555810), UINT32_C(3902491734), UINT32_C(3977937498), UINT32_C(2772762421)),
3175
+ simde_x_mm_set_epu32(UINT32_C(1924103348), UINT32_C(1357836992), UINT32_C(4064253767), UINT32_C(3196943489)),
3176
+ 1 },
3177
+ { simde_x_mm_set_epu32(UINT32_C(1969146297), UINT32_C(3716202658), UINT32_C( 759743671), UINT32_C(2926294225)),
3178
+ simde_x_mm_set_epu32(UINT32_C(2325820998), UINT32_C( 578764637), UINT32_C(3535223624), UINT32_C(1368673070)),
3179
+ 0 },
3180
+ { simde_x_mm_set_epu32(UINT32_C(3794246328), UINT32_C(2295817460), UINT32_C(2963593754), UINT32_C(2249033119)),
3181
+ simde_x_mm_set_epu32(UINT32_C(1810964868), UINT32_C(3773017844), UINT32_C(1790481610), UINT32_C(4104297207)),
3182
+ 1 },
3183
+ { simde_x_mm_set_epu32(UINT32_C(3721359930), UINT32_C(2092677361), UINT32_C(2672746089), UINT32_C(2784499405)),
3184
+ simde_x_mm_set_epu32(UINT32_C( 573607365), UINT32_C(2202289934), UINT32_C(1622221206), UINT32_C(1510467890)),
3185
+ 0 },
3186
+ { simde_x_mm_set_epu32(UINT32_C(1688049548), UINT32_C( 227731129), UINT32_C(2187006143), UINT32_C(2191704962)),
3187
+ simde_x_mm_set_epu32(UINT32_C(3739754335), UINT32_C(1136006144), UINT32_C(2036111421), UINT32_C(3950298731)),
3188
+ 1 },
3189
+ { simde_x_mm_set_epu32(UINT32_C(2420605154), UINT32_C(1034465439), UINT32_C( 270210360), UINT32_C(1245331438)),
3190
+ simde_x_mm_set_epu32(UINT32_C(1874362141), UINT32_C(3260501856), UINT32_C(4024756935), UINT32_C(3049635857)),
3191
+ 0 },
3192
+ { simde_x_mm_set_epu32(UINT32_C(2202152506), UINT32_C( 285372976), UINT32_C(4113851867), UINT32_C(1775650196)),
3193
+ simde_x_mm_set_epu32(UINT32_C(2951443467), UINT32_C(1206063205), UINT32_C(2278810499), UINT32_C(1561806132)),
3194
+ 1 },
3195
+ { simde_x_mm_set_epu32(UINT32_C(2386514979), UINT32_C(2103226597), UINT32_C( 427456035), UINT32_C(2075844151)),
3196
+ simde_x_mm_set_epu32(UINT32_C(2028307641), UINT32_C(3795980391), UINT32_C(1321164258), UINT32_C(1543009672)),
3197
+ 1 }
3198
+ };
3199
+
3200
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
3201
+ int r = simde_mm_test_mix_ones_zeros(test_vec[i].a, test_vec[i].b);
3202
+ munit_assert_int(r, ==, test_vec[i].r);
3203
+ }
3204
+
3205
+ return MUNIT_OK;
3206
+ }
3207
+
3208
+ static MunitResult
3209
+ test_simde_mm_testc_si128(const MunitParameter params[], void* data) {
3210
+ (void) params;
3211
+ (void) data;
3212
+
3213
+ const struct {
3214
+ simde__m128i a;
3215
+ simde__m128i b;
3216
+ int r;
3217
+ } test_vec[8] = {
3218
+ { simde_x_mm_set_epu32(UINT32_C(2830619595), UINT32_C(3721899845), UINT32_C(1506484780), UINT32_C(3754634237)),
3219
+ simde_x_mm_set_epu32(UINT32_C( 135762121), UINT32_C(1485021188), UINT32_C( 66080), UINT32_C(1279789969)),
3220
+ 1 },
3221
+ { simde_x_mm_set_epu32(UINT32_C( 193385482), UINT32_C(2729396552), UINT32_C(3295558360), UINT32_C(1810466520)),
3222
+ simde_x_mm_set_epu32(UINT32_C(2126857072), UINT32_C(2130569152), UINT32_C( 589776977), UINT32_C(3538187411)),
3223
+ 0 },
3224
+ { simde_x_mm_set_epu32(UINT32_C( 786883487), UINT32_C(1935834902), UINT32_C(1176572533), UINT32_C(2648243675)),
3225
+ simde_x_mm_set_epu32(UINT32_C( 782262300), UINT32_C(1617050388), UINT32_C( 33620084), UINT32_C( 294672713)),
3226
+ 1 },
3227
+ { simde_x_mm_set_epu32(UINT32_C(1041307471), UINT32_C( 45744950), UINT32_C(1982345789), UINT32_C(3152632048)),
3228
+ simde_x_mm_set_epu32(UINT32_C(1387031193), UINT32_C( 524221074), UINT32_C(2963552394), UINT32_C( 26167124)),
3229
+ 0 },
3230
+ { simde_x_mm_set_epu32(UINT32_C(4096311220), UINT32_C(1013684968), UINT32_C(3348640547), UINT32_C(2194828263)),
3231
+ simde_x_mm_set_epu32(UINT32_C(3288368404), UINT32_C( 873007168), UINT32_C( 127402785), UINT32_C( 43000416)),
3232
+ 1 },
3233
+ { simde_x_mm_set_epu32(UINT32_C(3620270306), UINT32_C( 102683939), UINT32_C(1921801133), UINT32_C(4285709000)),
3234
+ simde_x_mm_set_epu32(UINT32_C(2218828002), UINT32_C( 67671297), UINT32_C(1384665613), UINT32_C(2030217920)),
3235
+ 1 },
3236
+ { simde_x_mm_set_epu32(UINT32_C( 948424904), UINT32_C(1962159425), UINT32_C(1599818052), UINT32_C( 324039513)),
3237
+ simde_x_mm_set_epu32(UINT32_C( 671486144), UINT32_C(1888759104), UINT32_C(1191182592), UINT32_C( 319836488)),
3238
+ 1 },
3239
+ { simde_x_mm_set_epu32(UINT32_C(2926329185), UINT32_C(3270382597), UINT32_C( 628918622), UINT32_C(3097066353)),
3240
+ simde_x_mm_set_epu32(UINT32_C(2282240352), UINT32_C(3265794053), UINT32_C( 71305480), UINT32_C( 948529457)),
3241
+ 1 }
3242
+ };
3243
+
3244
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
3245
+ int r = simde_mm_testc_si128(test_vec[i].a, test_vec[i].b);
3246
+ munit_assert_int(r, ==, test_vec[i].r);
3247
+ }
3248
+
3249
+ return MUNIT_OK;
3250
+ }
3251
+
3252
+ static MunitResult
3253
+ test_simde_mm_testnzc_si128(const MunitParameter params[], void* data) {
3254
+ (void) params;
3255
+ (void) data;
3256
+
3257
+ const struct {
3258
+ simde__m128i a;
3259
+ simde__m128i b;
3260
+ int r;
3261
+ } test_vec[8] = {
3262
+ { simde_x_mm_set_epu32(UINT32_C(1875425415), UINT32_C(3970564046), UINT32_C(3151364481), UINT32_C(3646821095)),
3263
+ simde_x_mm_set_epu32(UINT32_C( 939391786), UINT32_C(1418582693), UINT32_C(3583579146), UINT32_C(2978542122)),
3264
+ 1 },
3265
+ { simde_x_mm_set_epu32(UINT32_C(1948733162), UINT32_C(3713693944), UINT32_C(2694213253), UINT32_C(1663204587)),
3266
+ simde_x_mm_set_epu32(UINT32_C(1140998794), UINT32_C(1141393520), UINT32_C(2685534720), UINT32_C(1124213826)),
3267
+ 0 },
3268
+ { simde_x_mm_set_epu32(UINT32_C(1383702507), UINT32_C(4170492791), UINT32_C( 81283149), UINT32_C(2572004949)),
3269
+ simde_x_mm_set_epu32(UINT32_C(1383112482), UINT32_C( 1049365), UINT32_C( 67110920), UINT32_C( 33301)),
3270
+ 0 },
3271
+ { simde_x_mm_set_epu32(UINT32_C( 122650173), UINT32_C(3420397620), UINT32_C(2510350310), UINT32_C(1202770589)),
3272
+ simde_x_mm_set_epu32(UINT32_C(3675302012), UINT32_C( 291962269), UINT32_C(1109447557), UINT32_C( 739260950)),
3273
+ 1 },
3274
+ { simde_x_mm_set_epu32(UINT32_C(1579102285), UINT32_C( 270245457), UINT32_C(1078977019), UINT32_C(2085912657)),
3275
+ simde_x_mm_set_epu32(UINT32_C(1242112064), UINT32_C( 1606145), UINT32_C( 688393), UINT32_C( 402685968)),
3276
+ 0 },
3277
+ { simde_x_mm_set_epu32(UINT32_C(3394408789), UINT32_C(2660724232), UINT32_C(2317277326), UINT32_C(3204989336)),
3278
+ simde_x_mm_set_epu32(UINT32_C(1819319362), UINT32_C( 329857933), UINT32_C(3391127410), UINT32_C( 743662239)),
3279
+ 1 },
3280
+ { simde_x_mm_set_epu32(UINT32_C(4125624985), UINT32_C( 2864942), UINT32_C( 875515634), UINT32_C(2091317004)),
3281
+ simde_x_mm_set_epu32(UINT32_C( 617089160), UINT32_C( 112418), UINT32_C( 606079728), UINT32_C(1881276672)),
3282
+ 0 },
3283
+ { simde_x_mm_set_epu32(UINT32_C(2421997128), UINT32_C(3685823722), UINT32_C(2588454762), UINT32_C(2067412407)),
3284
+ simde_x_mm_set_epu32(UINT32_C(3926740437), UINT32_C( 589344338), UINT32_C(2958265721), UINT32_C(1121386306)),
3285
+ 1 }
3286
+ };
3287
+
3288
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
3289
+ int r = simde_mm_testnzc_si128(test_vec[i].a, test_vec[i].b);
3290
+ munit_assert_int(r, ==, test_vec[i].r);
3291
+ }
3292
+
3293
+ return MUNIT_OK;
3294
+ }
3295
+
3296
+ static MunitResult
3297
+ test_simde_mm_testz_si128(const MunitParameter params[], void* data) {
3298
+ (void) params;
3299
+ (void) data;
3300
+
3301
+ const struct {
3302
+ simde__m128i a;
3303
+ simde__m128i b;
3304
+ int r;
3305
+ } test_vec[8] = {
3306
+ { simde_x_mm_set_epu32(UINT32_C(3990889899), UINT32_C(4230789057), UINT32_C(3808461042), UINT32_C(3371427410)),
3307
+ simde_x_mm_set_epu32(UINT32_C( 331792), UINT32_C( 63127558), UINT32_C( 483990789), UINT32_C( 530828)),
3308
+ 1 },
3309
+ { simde_x_mm_set_epu32(UINT32_C(3603350831), UINT32_C(2431793515), UINT32_C(1875486201), UINT32_C(3530533257)),
3310
+ simde_x_mm_set_epu32(UINT32_C( 19933776), UINT32_C(1174733824), UINT32_C(2147880964), UINT32_C( 8393234)),
3311
+ 1 },
3312
+ { simde_x_mm_set_epu32(UINT32_C(3124097223), UINT32_C(4066616999), UINT32_C(2076362673), UINT32_C( 124456258)),
3313
+ simde_x_mm_set_epu32(UINT32_C(1170213928), UINT32_C( 84416520), UINT32_C(2150893634), UINT32_C(3499413545)),
3314
+ 1 },
3315
+ { simde_x_mm_set_epu32(UINT32_C(4156796004), UINT32_C(1994002854), UINT32_C(4034407880), UINT32_C( 600089084)),
3316
+ simde_x_mm_set_epu32(UINT32_C(2971079954), UINT32_C(3997808651), UINT32_C( 222740062), UINT32_C(3989806580)),
3317
+ 0 },
3318
+ { simde_x_mm_set_epu32(UINT32_C(4130552440), UINT32_C( 812777701), UINT32_C(2016424386), UINT32_C( 886379222)),
3319
+ simde_x_mm_set_epu32(UINT32_C( 70177905), UINT32_C(2046022589), UINT32_C( 670901459), UINT32_C(2978865170)),
3320
+ 0 },
3321
+ { simde_x_mm_set_epu32(UINT32_C( 192532312), UINT32_C(1467133872), UINT32_C(1075139299), UINT32_C( 191167596)),
3322
+ simde_x_mm_set_epu32(UINT32_C(1026534956), UINT32_C(4130843248), UINT32_C(3619306010), UINT32_C(3598027302)),
3323
+ 0 },
3324
+ { simde_x_mm_set_epu32(UINT32_C(2410417876), UINT32_C( 74289906), UINT32_C(4031840239), UINT32_C(2837881625)),
3325
+ simde_x_mm_set_epu32(UINT32_C( 542327083), UINT32_C( 680656909), UINT32_C( 184549392), UINT32_C(1421938882)),
3326
+ 1 },
3327
+ { simde_x_mm_set_epu32(UINT32_C(3782150825), UINT32_C(3325635017), UINT32_C(1617333560), UINT32_C(3634437083)),
3328
+ simde_x_mm_set_epu32(UINT32_C( 167825730), UINT32_C( 294047748), UINT32_C( 126906945), UINT32_C( 17884164)),
3329
+ 1 }
3330
+ };
3331
+
3332
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
3333
+ int r = simde_mm_testz_si128(test_vec[i].a, test_vec[i].b);
3334
+ munit_assert_int(r, ==, test_vec[i].r);
3335
+ }
3336
+
3337
+ return MUNIT_OK;
3338
+ }
3339
+
3340
+ #endif /* defined(SIMDE_SSE4_1_NATIVE) || defined(SIMDE_NO_NATIVE) || defined(SIMDE_ALWAYS_BUILD_NATIVE_TESTS) */
3341
+
3342
+ HEDLEY_DIAGNOSTIC_PUSH
3343
+ HEDLEY_DIAGNOSTIC_DISABLE_CAST_QUAL
3344
+
3345
+ static MunitTest test_suite_tests[] = {
3346
+ #if defined(SIMDE_SSE4_1_NATIVE) || defined(SIMDE_NO_NATIVE) || defined(SIMDE_ALWAYS_BUILD_NATIVE_TESTS)
3347
+ SIMDE_TESTS_DEFINE_TEST(mm_blend_epi16),
3348
+ SIMDE_TESTS_DEFINE_TEST(mm_blend_pd),
3349
+ SIMDE_TESTS_DEFINE_TEST(mm_blend_ps),
3350
+
3351
+ SIMDE_TESTS_DEFINE_TEST(mm_blendv_epi8),
3352
+ SIMDE_TESTS_DEFINE_TEST(x_mm_blendv_epi16),
3353
+ SIMDE_TESTS_DEFINE_TEST(x_mm_blendv_epi32),
3354
+ SIMDE_TESTS_DEFINE_TEST(x_mm_blendv_epi64),
3355
+ SIMDE_TESTS_DEFINE_TEST(mm_blendv_pd),
3356
+ SIMDE_TESTS_DEFINE_TEST(mm_blendv_ps),
3357
+
3358
+ SIMDE_TESTS_DEFINE_TEST(mm_ceil_pd),
3359
+ SIMDE_TESTS_DEFINE_TEST(mm_ceil_ps),
3360
+ SIMDE_TESTS_DEFINE_TEST(mm_ceil_sd),
3361
+ SIMDE_TESTS_DEFINE_TEST(mm_ceil_ss),
3362
+
3363
+ SIMDE_TESTS_DEFINE_TEST(mm_cmpeq_epi64),
3364
+
3365
+ SIMDE_TESTS_DEFINE_TEST(mm_cvtepi8_epi16),
3366
+ SIMDE_TESTS_DEFINE_TEST(mm_cvtepi8_epi32),
3367
+ SIMDE_TESTS_DEFINE_TEST(mm_cvtepi8_epi64),
3368
+ SIMDE_TESTS_DEFINE_TEST(mm_cvtepu8_epi16),
3369
+ SIMDE_TESTS_DEFINE_TEST(mm_cvtepu8_epi32),
3370
+ SIMDE_TESTS_DEFINE_TEST(mm_cvtepu8_epi64),
3371
+ SIMDE_TESTS_DEFINE_TEST(mm_cvtepi16_epi32),
3372
+ SIMDE_TESTS_DEFINE_TEST(mm_cvtepi16_epi64),
3373
+ SIMDE_TESTS_DEFINE_TEST(mm_cvtepu16_epi32),
3374
+ SIMDE_TESTS_DEFINE_TEST(mm_cvtepu16_epi64),
3375
+ SIMDE_TESTS_DEFINE_TEST(mm_cvtepi32_epi64),
3376
+ SIMDE_TESTS_DEFINE_TEST(mm_cvtepu32_epi64),
3377
+
3378
+ SIMDE_TESTS_DEFINE_TEST(mm_dp_pd),
3379
+ SIMDE_TESTS_DEFINE_TEST(mm_dp_ps),
3380
+
3381
+ SIMDE_TESTS_DEFINE_TEST(mm_extract_epi8),
3382
+ SIMDE_TESTS_DEFINE_TEST(mm_extract_epi32),
3383
+ SIMDE_TESTS_DEFINE_TEST(mm_extract_epi64),
3384
+
3385
+ SIMDE_TESTS_DEFINE_TEST(mm_floor_pd),
3386
+ SIMDE_TESTS_DEFINE_TEST(mm_floor_ps),
3387
+ SIMDE_TESTS_DEFINE_TEST(mm_floor_sd),
3388
+ SIMDE_TESTS_DEFINE_TEST(mm_floor_ss),
3389
+
3390
+ SIMDE_TESTS_DEFINE_TEST(mm_insert_epi8),
3391
+ SIMDE_TESTS_DEFINE_TEST(mm_insert_epi32),
3392
+ SIMDE_TESTS_DEFINE_TEST(mm_insert_epi64),
3393
+ SIMDE_TESTS_DEFINE_TEST(mm_insert_ps),
3394
+
3395
+ SIMDE_TESTS_DEFINE_TEST(mm_max_epi8),
3396
+ SIMDE_TESTS_DEFINE_TEST(mm_max_epi32),
3397
+ SIMDE_TESTS_DEFINE_TEST(mm_max_epu16),
3398
+ SIMDE_TESTS_DEFINE_TEST(mm_max_epu32),
3399
+
3400
+ SIMDE_TESTS_DEFINE_TEST(mm_min_epi8),
3401
+ SIMDE_TESTS_DEFINE_TEST(mm_min_epi32),
3402
+ SIMDE_TESTS_DEFINE_TEST(mm_min_epu16),
3403
+ SIMDE_TESTS_DEFINE_TEST(mm_min_epu32),
3404
+
3405
+ SIMDE_TESTS_DEFINE_TEST(mm_minpos_epu16),
3406
+
3407
+ SIMDE_TESTS_DEFINE_TEST(mm_mpsadbw_epu8),
3408
+
3409
+ SIMDE_TESTS_DEFINE_TEST(mm_mul_epi32),
3410
+ SIMDE_TESTS_DEFINE_TEST(mm_mullo_epi32),
3411
+
3412
+ SIMDE_TESTS_DEFINE_TEST(mm_packus_epi32),
3413
+
3414
+ SIMDE_TESTS_DEFINE_TEST(mm_round_pd),
3415
+ SIMDE_TESTS_DEFINE_TEST(mm_round_ps),
3416
+ SIMDE_TESTS_DEFINE_TEST(mm_round_sd),
3417
+ SIMDE_TESTS_DEFINE_TEST(mm_round_ss),
3418
+
3419
+ SIMDE_TESTS_DEFINE_TEST(mm_stream_load_si128),
3420
+
3421
+ SIMDE_TESTS_DEFINE_TEST(mm_test_all_ones),
3422
+ SIMDE_TESTS_DEFINE_TEST(mm_test_all_zeros),
3423
+ SIMDE_TESTS_DEFINE_TEST(mm_test_mix_ones_zeros),
3424
+ SIMDE_TESTS_DEFINE_TEST(mm_testc_si128),
3425
+ SIMDE_TESTS_DEFINE_TEST(mm_testnzc_si128),
3426
+ SIMDE_TESTS_DEFINE_TEST(mm_testz_si128),
3427
+ #endif /* defined(SIMDE_SSE4_1_NATIVE) || defined(SIMDE_NO_NATIVE) || defined(SIMDE_ALWAYS_BUILD_NATIVE_TESTS) */
3428
+
3429
+ { NULL, NULL, NULL, NULL, MUNIT_TEST_OPTION_NONE, NULL }
3430
+ };
3431
+
3432
+ HEDLEY_C_DECL MunitSuite* SIMDE_TESTS_GENERATE_SYMBOL(suite)(void) {
3433
+ static MunitSuite suite = { (char*) "/" HEDLEY_STRINGIFY(SIMDE_TESTS_CURRENT_ISAX), test_suite_tests, NULL, 1, MUNIT_SUITE_OPTION_NONE };
3434
+
3435
+ return &suite;
3436
+ }
3437
+
3438
+ #if defined(SIMDE_TESTS_SINGLE_ISAX)
3439
+ int main(int argc, char* argv[HEDLEY_ARRAY_PARAM(argc + 1)]) {
3440
+ static MunitSuite suite = { "", test_suite_tests, NULL, 1, MUNIT_SUITE_OPTION_NONE };
3441
+
3442
+ return munit_suite_main(&suite, NULL, argc, argv);
3443
+ }
3444
+ #endif /* defined(SIMDE_TESTS_SINGLE_ISAX) */
3445
+
3446
+ HEDLEY_DIAGNOSTIC_POP