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 +1,1084 @@
1
+ /* Copyright (c) 2019 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_NEON_OP mul
25
+ #include <test/arm/neon/test-neon-internal.h>
26
+ #include <simde/arm/neon.h>
27
+
28
+ #if defined(SIMDE_NEON_NATIVE) || defined(SIMDE_NO_NATIVE) || defined(SIMDE_ALWAYS_BUILD_NATIVE_TESTS)
29
+
30
+ static MunitResult
31
+ test_simde_vmul_s8(const MunitParameter params[], void* data) {
32
+ (void) params;
33
+ (void) data;
34
+
35
+ const struct {
36
+ simde_int8x8_t a;
37
+ simde_int8x8_t b;
38
+ simde_int8x8_t r;
39
+ } test_vec[8] = {
40
+ { simde_x_vload_s8(INT8_C( 102), INT8_C( -11), INT8_C( -66), INT8_C(-119),
41
+ INT8_C( 43), INT8_C( 110), INT8_C( 64), INT8_C( -56)),
42
+ simde_x_vload_s8(INT8_C( 21), INT8_C(-117), INT8_C( 108), INT8_C(-116),
43
+ INT8_C( -14), INT8_C( 64), INT8_C( 89), INT8_C( 83)),
44
+ simde_x_vload_s8(INT8_C( 94), INT8_C( 7), INT8_C( 40), INT8_C( -20),
45
+ INT8_C( -90), INT8_C(-128), INT8_C( 64), INT8_C( -40)) },
46
+ { simde_x_vload_s8(INT8_C( 117), INT8_C( -87), INT8_C( -31), INT8_C( 110),
47
+ INT8_C( 28), INT8_C( 19), INT8_C( 18), INT8_C( 64)),
48
+ simde_x_vload_s8(INT8_C( 117), INT8_C(-106), INT8_C( 24), INT8_C( -75),
49
+ INT8_C( 22), INT8_C( -42), INT8_C( -80), INT8_C( 85)),
50
+ simde_x_vload_s8(INT8_C( 121), INT8_C( 6), INT8_C( 24), INT8_C( -58),
51
+ INT8_C( 104), INT8_C( -30), INT8_C( 96), INT8_C( 64)) },
52
+ { simde_x_vload_s8(INT8_C( -67), INT8_C( 22), INT8_C( 58), INT8_C(-112),
53
+ INT8_C( 113), INT8_C( 58), INT8_C(-121), INT8_C(-109)),
54
+ simde_x_vload_s8(INT8_C( -53), INT8_C(-109), INT8_C( 9), INT8_C( -8),
55
+ INT8_C( 45), INT8_C( 17), INT8_C( -71), INT8_C( -33)),
56
+ simde_x_vload_s8(INT8_C( -33), INT8_C( -94), INT8_C( 10), INT8_C(-128),
57
+ INT8_C( -35), INT8_C( -38), INT8_C(-113), INT8_C( 13)) },
58
+ { simde_x_vload_s8(INT8_C( -52), INT8_C(-112), INT8_C( -62), INT8_C( 74),
59
+ INT8_C( 1), INT8_C( -87), INT8_C( -43), INT8_C( 20)),
60
+ simde_x_vload_s8(INT8_C( 79), INT8_C( -70), INT8_C( 124), INT8_C(-110),
61
+ INT8_C(-103), INT8_C( 92), INT8_C( 62), INT8_C( -96)),
62
+ simde_x_vload_s8(INT8_C( -12), INT8_C( -96), INT8_C( -8), INT8_C( 52),
63
+ INT8_C(-103), INT8_C( -68), INT8_C(-106), INT8_C(-128)) },
64
+ { simde_x_vload_s8(INT8_C( 106), INT8_C( 23), INT8_C( -98), INT8_C( 21),
65
+ INT8_C(-114), INT8_C( 8), INT8_C( -78), INT8_C( -32)),
66
+ simde_x_vload_s8(INT8_C(-100), INT8_C( 22), INT8_C(-126), INT8_C( -35),
67
+ INT8_C( -1), INT8_C( -46), INT8_C( -93), INT8_C( -32)),
68
+ simde_x_vload_s8(INT8_C(-104), INT8_C( -6), INT8_C( 60), INT8_C( 33),
69
+ INT8_C( 114), INT8_C(-112), INT8_C( 86), INT8_C( 0)) },
70
+ { simde_x_vload_s8(INT8_C( 118), INT8_C( -95), INT8_C( -87), INT8_C( 25),
71
+ INT8_C( -22), INT8_C( 92), INT8_C( 48), INT8_C( 41)),
72
+ simde_x_vload_s8(INT8_C( -2), INT8_C( 61), INT8_C(-127), INT8_C( 120),
73
+ INT8_C( 32), INT8_C(-125), INT8_C( -64), INT8_C( 119)),
74
+ simde_x_vload_s8(INT8_C( 20), INT8_C( 93), INT8_C( 41), INT8_C( -72),
75
+ INT8_C( 64), INT8_C( 20), INT8_C( 0), INT8_C( 15)) },
76
+ { simde_x_vload_s8(INT8_C( -71), INT8_C( 100), INT8_C( 2), INT8_C( -47),
77
+ INT8_C(-105), INT8_C( 83), INT8_C( 36), INT8_C( -48)),
78
+ simde_x_vload_s8(INT8_C( 37), INT8_C( -86), INT8_C( -62), INT8_C(-128),
79
+ INT8_C( 111), INT8_C( 42), INT8_C( -35), INT8_C( 99)),
80
+ simde_x_vload_s8(INT8_C( -67), INT8_C( 104), INT8_C(-124), INT8_C(-128),
81
+ INT8_C( 121), INT8_C( -98), INT8_C( 20), INT8_C( 112)) },
82
+ { simde_x_vload_s8(INT8_C( 31), INT8_C(-109), INT8_C( 79), INT8_C( 57),
83
+ INT8_C( -62), INT8_C( 16), INT8_C( 2), INT8_C( 111)),
84
+ simde_x_vload_s8(INT8_C( -78), INT8_C( -76), INT8_C( -69), INT8_C( -53),
85
+ INT8_C( 0), INT8_C( -37), INT8_C( 19), INT8_C( 18)),
86
+ simde_x_vload_s8(INT8_C(-114), INT8_C( 92), INT8_C( -75), INT8_C( 51),
87
+ INT8_C( 0), INT8_C( -80), INT8_C( 38), INT8_C( -50)) }
88
+ };
89
+
90
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
91
+ simde_int8x8_t r = simde_vmul_s8(test_vec[i].a, test_vec[i].b);
92
+ simde_neon_assert_int8x8(r, ==, test_vec[i].r);
93
+ }
94
+
95
+ return MUNIT_OK;
96
+ }
97
+
98
+ static MunitResult
99
+ test_simde_vmul_s16(const MunitParameter params[], void* data) {
100
+ (void) params;
101
+ (void) data;
102
+
103
+ const struct {
104
+ simde_int16x4_t a;
105
+ simde_int16x4_t b;
106
+ simde_int16x4_t r;
107
+ } test_vec[8] = {
108
+ { simde_x_vload_s16(INT16_C(-31826), INT16_C( -7391), INT16_C( -462), INT16_C(-10931)),
109
+ simde_x_vload_s16(INT16_C(-29948), INT16_C(-20769), INT16_C( 17251), INT16_C( 13600)),
110
+ simde_x_vload_s16(INT16_C(-30536), INT16_C( 18367), INT16_C( 25430), INT16_C(-25952)) },
111
+ { simde_x_vload_s16(INT16_C(-12602), INT16_C( 26994), INT16_C( 15946), INT16_C(-17611)),
112
+ simde_x_vload_s16(INT16_C(-27918), INT16_C( 31417), INT16_C(-31752), INT16_C( 12675)),
113
+ simde_x_vload_s16(INT16_C( 25388), INT16_C(-30878), INT16_C( 13744), INT16_C( -3809)) },
114
+ { simde_x_vload_s16(INT16_C( 25025), INT16_C( 10589), INT16_C( 6804), INT16_C( 7776)),
115
+ simde_x_vload_s16(INT16_C(-25648), INT16_C(-28276), INT16_C( -793), INT16_C( 31106)),
116
+ simde_x_vload_s16(INT16_C( 18384), INT16_C( 19420), INT16_C(-21620), INT16_C(-13120)) },
117
+ { simde_x_vload_s16(INT16_C( 196), INT16_C( 1304), INT16_C(-31290), INT16_C(-32121)),
118
+ simde_x_vload_s16(INT16_C( 21652), INT16_C( -9418), INT16_C( 29473), INT16_C( 15166)),
119
+ simde_x_vload_s16(INT16_C(-16048), INT16_C(-25840), INT16_C( 12422), INT16_C(-17998)) },
120
+ { simde_x_vload_s16(INT16_C(-29855), INT16_C( -248), INT16_C(-21274), INT16_C(-28669)),
121
+ simde_x_vload_s16(INT16_C(-15098), INT16_C( -8131), INT16_C( 30935), INT16_C( -2866)),
122
+ simde_x_vload_s16(INT16_C( -5818), INT16_C(-15128), INT16_C( 1322), INT16_C(-16790)) },
123
+ { simde_x_vload_s16(INT16_C(-12706), INT16_C( 1479), INT16_C(-16812), INT16_C( 7789)),
124
+ simde_x_vload_s16(INT16_C( 25541), INT16_C( 26457), INT16_C( 25772), INT16_C( 31684)),
125
+ simde_x_vload_s16(INT16_C( 10326), INT16_C( 4911), INT16_C(-20368), INT16_C(-21900)) },
126
+ { simde_x_vload_s16(INT16_C( 6250), INT16_C(-24625), INT16_C( 1194), INT16_C( -6051)),
127
+ simde_x_vload_s16(INT16_C( 29420), INT16_C( 25728), INT16_C( 23202), INT16_C( 22987)),
128
+ simde_x_vload_s16(INT16_C(-19016), INT16_C(-15488), INT16_C(-18540), INT16_C(-26945)) },
129
+ { simde_x_vload_s16(INT16_C(-30139), INT16_C( 13263), INT16_C( 10350), INT16_C( 31500)),
130
+ simde_x_vload_s16(INT16_C( -4536), INT16_C( 10209), INT16_C(-32488), INT16_C( 9174)),
131
+ simde_x_vload_s16(INT16_C( 2408), INT16_C( 4591), INT16_C( 14416), INT16_C(-32760)) }
132
+ };
133
+
134
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
135
+ simde_int16x4_t r = simde_vmul_s16(test_vec[i].a, test_vec[i].b);
136
+ simde_neon_assert_int16x4(r, ==, test_vec[i].r);
137
+ }
138
+
139
+ return MUNIT_OK;
140
+ }
141
+
142
+ static MunitResult
143
+ test_simde_vmul_s32(const MunitParameter params[], void* data) {
144
+ (void) params;
145
+ (void) data;
146
+
147
+ const struct {
148
+ simde_int32x2_t a;
149
+ simde_int32x2_t b;
150
+ simde_int32x2_t r;
151
+ } test_vec[8] = {
152
+ { simde_x_vload_s32(INT32_C(-1462480123), INT32_C(-1568774293)),
153
+ simde_x_vload_s32(INT32_C( 855416238), INT32_C( 670298924)),
154
+ simde_x_vload_s32(INT32_C(-1440317850), INT32_C( 1635080036)) },
155
+ { simde_x_vload_s32(INT32_C( 1950510690), INT32_C(-1312310854)),
156
+ simde_x_vload_s32(INT32_C( 1186234862), INT32_C(-1064675530)),
157
+ simde_x_vload_s32(INT32_C( -129367780), INT32_C( -24943812)) },
158
+ { simde_x_vload_s32(INT32_C(-1982046851), INT32_C(-1152099595)),
159
+ simde_x_vload_s32(INT32_C(-1466083134), INT32_C( -479470034)),
160
+ simde_x_vload_s32(INT32_C(-1765093190), INT32_C( 326094342)) },
161
+ { simde_x_vload_s32(INT32_C(-1099740866), INT32_C(-2031221022)),
162
+ simde_x_vload_s32(INT32_C( -272182105), INT32_C( 1934242918)),
163
+ simde_x_vload_s32(INT32_C( 1037049714), INT32_C(-1617910260)) },
164
+ { simde_x_vload_s32(INT32_C( 846494307), INT32_C( 992291677)),
165
+ simde_x_vload_s32(INT32_C(-1062975990), INT32_C( -728231586)),
166
+ simde_x_vload_s32(INT32_C( 2037373406), INT32_C( -222357210)) },
167
+ { simde_x_vload_s32(INT32_C( 1233604686), INT32_C( 913582180)),
168
+ simde_x_vload_s32(INT32_C(-2062706547), INT32_C( 1757955985)),
169
+ simde_x_vload_s32(INT32_C( 1898679030), INT32_C( 1025416356)) },
170
+ { simde_x_vload_s32(INT32_C( 504229488), INT32_C(-2135950055)),
171
+ simde_x_vload_s32(INT32_C( 57241002), INT32_C(-2085005546)),
172
+ simde_x_vload_s32(INT32_C( -4848032), INT32_C(-1875624154)) },
173
+ { simde_x_vload_s32(INT32_C(-2063505862), INT32_C( 1539186741)),
174
+ simde_x_vload_s32(INT32_C( -513744611), INT32_C( 522045179)),
175
+ simde_x_vload_s32(INT32_C(-1681813870), INT32_C( 290689527)) }
176
+ };
177
+
178
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
179
+ simde_int32x2_t r = simde_vmul_s32(test_vec[i].a, test_vec[i].b);
180
+ simde_neon_assert_int32x2(r, ==, test_vec[i].r);
181
+ }
182
+
183
+ return MUNIT_OK;
184
+ }
185
+
186
+ static MunitResult
187
+ test_simde_vmul_u8(const MunitParameter params[], void* data) {
188
+ (void) params;
189
+ (void) data;
190
+
191
+ const struct {
192
+ simde_uint8x8_t a;
193
+ simde_uint8x8_t b;
194
+ simde_uint8x8_t r;
195
+ } test_vec[8] = {
196
+ { simde_x_vload_u8(UINT8_C(122), UINT8_C(173), UINT8_C(147), UINT8_C( 28),
197
+ UINT8_C(219), UINT8_C( 56), UINT8_C(136), UINT8_C(125)),
198
+ simde_x_vload_u8(UINT8_C( 44), UINT8_C(186), UINT8_C(141), UINT8_C(226),
199
+ UINT8_C( 55), UINT8_C(190), UINT8_C(143), UINT8_C(187)),
200
+ simde_x_vload_u8(UINT8_C(248), UINT8_C(178), UINT8_C(247), UINT8_C(184),
201
+ UINT8_C( 13), UINT8_C(144), UINT8_C(248), UINT8_C( 79)) },
202
+ { simde_x_vload_u8(UINT8_C(102), UINT8_C(120), UINT8_C( 9), UINT8_C( 82),
203
+ UINT8_C( 26), UINT8_C(142), UINT8_C(247), UINT8_C( 47)),
204
+ simde_x_vload_u8(UINT8_C(150), UINT8_C(107), UINT8_C(122), UINT8_C(192),
205
+ UINT8_C(153), UINT8_C( 74), UINT8_C(207), UINT8_C( 1)),
206
+ simde_x_vload_u8(UINT8_C(196), UINT8_C( 40), UINT8_C( 74), UINT8_C(128),
207
+ UINT8_C(138), UINT8_C( 12), UINT8_C(185), UINT8_C( 47)) },
208
+ { simde_x_vload_u8(UINT8_C(213), UINT8_C(100), UINT8_C( 93), UINT8_C(149),
209
+ UINT8_C( 25), UINT8_C( 65), UINT8_C(188), UINT8_C(157)),
210
+ simde_x_vload_u8(UINT8_C( 32), UINT8_C(165), UINT8_C( 78), UINT8_C( 59),
211
+ UINT8_C(236), UINT8_C(154), UINT8_C( 33), UINT8_C(249)),
212
+ simde_x_vload_u8(UINT8_C(160), UINT8_C(116), UINT8_C( 86), UINT8_C( 87),
213
+ UINT8_C( 12), UINT8_C( 26), UINT8_C( 60), UINT8_C(181)) },
214
+ { simde_x_vload_u8(UINT8_C( 39), UINT8_C(171), UINT8_C(211), UINT8_C( 66),
215
+ UINT8_C( 72), UINT8_C(124), UINT8_C( 28), UINT8_C( 70)),
216
+ simde_x_vload_u8(UINT8_C(153), UINT8_C( 6), UINT8_C( 71), UINT8_C(240),
217
+ UINT8_C(187), UINT8_C(114), UINT8_C(108), UINT8_C(190)),
218
+ simde_x_vload_u8(UINT8_C( 79), UINT8_C( 2), UINT8_C(133), UINT8_C(224),
219
+ UINT8_C(152), UINT8_C( 56), UINT8_C(208), UINT8_C(244)) },
220
+ { simde_x_vload_u8(UINT8_C(144), UINT8_C(250), UINT8_C( 70), UINT8_C(245),
221
+ UINT8_C(242), UINT8_C( 79), UINT8_C(119), UINT8_C(191)),
222
+ simde_x_vload_u8(UINT8_C( 7), UINT8_C(166), UINT8_C( 88), UINT8_C( 47),
223
+ UINT8_C( 44), UINT8_C( 81), UINT8_C(234), UINT8_C(140)),
224
+ simde_x_vload_u8(UINT8_C(240), UINT8_C( 28), UINT8_C( 16), UINT8_C(251),
225
+ UINT8_C(152), UINT8_C(255), UINT8_C(198), UINT8_C(116)) },
226
+ { simde_x_vload_u8(UINT8_C(144), UINT8_C( 31), UINT8_C(104), UINT8_C( 79),
227
+ UINT8_C( 59), UINT8_C(248), UINT8_C( 83), UINT8_C( 47)),
228
+ simde_x_vload_u8(UINT8_C(213), UINT8_C(219), UINT8_C(235), UINT8_C( 56),
229
+ UINT8_C(101), UINT8_C( 11), UINT8_C(121), UINT8_C(224)),
230
+ simde_x_vload_u8(UINT8_C(208), UINT8_C(133), UINT8_C(120), UINT8_C( 72),
231
+ UINT8_C( 71), UINT8_C(168), UINT8_C( 59), UINT8_C( 32)) },
232
+ { simde_x_vload_u8(UINT8_C( 60), UINT8_C( 93), UINT8_C(193), UINT8_C(141),
233
+ UINT8_C( 94), UINT8_C( 73), UINT8_C( 11), UINT8_C(237)),
234
+ simde_x_vload_u8(UINT8_C(212), UINT8_C(250), UINT8_C(105), UINT8_C( 20),
235
+ UINT8_C( 97), UINT8_C(145), UINT8_C( 38), UINT8_C(145)),
236
+ simde_x_vload_u8(UINT8_C(176), UINT8_C(210), UINT8_C( 41), UINT8_C( 4),
237
+ UINT8_C(158), UINT8_C( 89), UINT8_C(162), UINT8_C( 61)) },
238
+ { simde_x_vload_u8(UINT8_C( 0), UINT8_C( 70), UINT8_C(114), UINT8_C(145),
239
+ UINT8_C( 45), UINT8_C(152), UINT8_C(129), UINT8_C( 87)),
240
+ simde_x_vload_u8(UINT8_C( 26), UINT8_C(240), UINT8_C( 84), UINT8_C(165),
241
+ UINT8_C(123), UINT8_C( 30), UINT8_C( 67), UINT8_C(102)),
242
+ simde_x_vload_u8(UINT8_C( 0), UINT8_C(160), UINT8_C(104), UINT8_C(117),
243
+ UINT8_C(159), UINT8_C(208), UINT8_C(195), UINT8_C(170)) }
244
+ };
245
+
246
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
247
+ simde_uint8x8_t r = simde_vmul_u8(test_vec[i].a, test_vec[i].b);
248
+ simde_neon_assert_uint8x8(r, ==, test_vec[i].r);
249
+ }
250
+
251
+ return MUNIT_OK;
252
+ }
253
+
254
+ static MunitResult
255
+ test_simde_vmul_u16(const MunitParameter params[], void* data) {
256
+ (void) params;
257
+ (void) data;
258
+
259
+ const struct {
260
+ simde_uint16x4_t a;
261
+ simde_uint16x4_t b;
262
+ simde_uint16x4_t r;
263
+ } test_vec[8] = {
264
+ { simde_x_vload_u16(UINT16_C( 3243), UINT16_C(48979), UINT16_C(13740), UINT16_C( 548)),
265
+ simde_x_vload_u16(UINT16_C(33299), UINT16_C(22598), UINT16_C(38440), UINT16_C(37700)),
266
+ simde_x_vload_u16(UINT16_C(50865), UINT16_C(55474), UINT16_C(10976), UINT16_C(15760)) },
267
+ { simde_x_vload_u16(UINT16_C(33812), UINT16_C(24559), UINT16_C(20280), UINT16_C(29924)),
268
+ simde_x_vload_u16(UINT16_C( 1544), UINT16_C(35519), UINT16_C(26256), UINT16_C(45486)),
269
+ simde_x_vload_u16(UINT16_C(39072), UINT16_C(26961), UINT16_C(57216), UINT16_C( 5880)) },
270
+ { simde_x_vload_u16(UINT16_C(42628), UINT16_C(33742), UINT16_C(64614), UINT16_C(12076)),
271
+ simde_x_vload_u16(UINT16_C(34782), UINT16_C(62945), UINT16_C(57802), UINT16_C( 3464)),
272
+ simde_x_vload_u16(UINT16_C( 632), UINT16_C(65038), UINT16_C(52860), UINT16_C(19296)) },
273
+ { simde_x_vload_u16(UINT16_C(30310), UINT16_C(13289), UINT16_C(24252), UINT16_C(44497)),
274
+ simde_x_vload_u16(UINT16_C(60569), UINT16_C(42651), UINT16_C(61697), UINT16_C( 5994)),
275
+ simde_x_vload_u16(UINT16_C(51958), UINT16_C(33811), UINT16_C(23228), UINT16_C(49034)) },
276
+ { simde_x_vload_u16(UINT16_C(13840), UINT16_C(32818), UINT16_C(16327), UINT16_C(45233)),
277
+ simde_x_vload_u16(UINT16_C(53905), UINT16_C(53691), UINT16_C(49691), UINT16_C(13942)),
278
+ simde_x_vload_u16(UINT16_C(48912), UINT16_C(30342), UINT16_C(34813), UINT16_C(51094)) },
279
+ { simde_x_vload_u16(UINT16_C(22976), UINT16_C(40255), UINT16_C(55574), UINT16_C(29449)),
280
+ simde_x_vload_u16(UINT16_C(56709), UINT16_C(22563), UINT16_C(29259), UINT16_C(28398)),
281
+ simde_x_vload_u16(UINT16_C(24768), UINT16_C(10141), UINT16_C(25970), UINT16_C(53342)) },
282
+ { simde_x_vload_u16(UINT16_C(60271), UINT16_C(29991), UINT16_C( 1809), UINT16_C(12217)),
283
+ simde_x_vload_u16(UINT16_C(38755), UINT16_C(35233), UINT16_C(42873), UINT16_C(28048)),
284
+ simde_x_vload_u16(UINT16_C(34029), UINT16_C(35975), UINT16_C(28169), UINT16_C(40208)) },
285
+ { simde_x_vload_u16(UINT16_C(14808), UINT16_C(46271), UINT16_C( 898), UINT16_C(34464)),
286
+ simde_x_vload_u16(UINT16_C(11637), UINT16_C(17181), UINT16_C(45377), UINT16_C(21995)),
287
+ simde_x_vload_u16(UINT16_C(26552), UINT16_C(30371), UINT16_C(50690), UINT16_C(46304)) }
288
+ };
289
+
290
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
291
+ simde_uint16x4_t r = simde_vmul_u16(test_vec[i].a, test_vec[i].b);
292
+ simde_neon_assert_uint16x4(r, ==, test_vec[i].r);
293
+ }
294
+
295
+ return MUNIT_OK;
296
+ }
297
+
298
+ static MunitResult
299
+ test_simde_vmul_u32(const MunitParameter params[], void* data) {
300
+ (void) params;
301
+ (void) data;
302
+
303
+ const struct {
304
+ simde_uint32x2_t a;
305
+ simde_uint32x2_t b;
306
+ simde_uint32x2_t r;
307
+ } test_vec[8] = {
308
+ { simde_x_vload_u32(UINT32_C(3269571689), UINT32_C( 733666179)),
309
+ simde_x_vload_u32(UINT32_C(3150126302), UINT32_C(3039826473)),
310
+ simde_x_vload_u32(UINT32_C(3463507726), UINT32_C(3958136315)) },
311
+ { simde_x_vload_u32(UINT32_C(3052965843), UINT32_C(2222125336)),
312
+ simde_x_vload_u32(UINT32_C( 191761853), UINT32_C(3420348928)),
313
+ simde_x_vload_u32(UINT32_C(3581535687), UINT32_C( 454766592)) },
314
+ { simde_x_vload_u32(UINT32_C( 25380888), UINT32_C(2761087377)),
315
+ simde_x_vload_u32(UINT32_C(2170081949), UINT32_C(2408729295)),
316
+ simde_x_vload_u32(UINT32_C(2129061560), UINT32_C(2572408383)) },
317
+ { simde_x_vload_u32(UINT32_C(1772322605), UINT32_C(2834267239)),
318
+ simde_x_vload_u32(UINT32_C(1803133576), UINT32_C(1820444802)),
319
+ simde_x_vload_u32(UINT32_C(2157644264), UINT32_C( 112501838)) },
320
+ { simde_x_vload_u32(UINT32_C(1769263504), UINT32_C(1439650640)),
321
+ simde_x_vload_u32(UINT32_C(3130677651), UINT32_C(2329460577)),
322
+ simde_x_vload_u32(UINT32_C(3683607984), UINT32_C(2654421328)) },
323
+ { simde_x_vload_u32(UINT32_C(3689635154), UINT32_C( 675235401)),
324
+ simde_x_vload_u32(UINT32_C(2394861819), UINT32_C( 13551513)),
325
+ simde_x_vload_u32(UINT32_C( 784916326), UINT32_C(3425812641)) },
326
+ { simde_x_vload_u32(UINT32_C( 32265746), UINT32_C(1506124520)),
327
+ simde_x_vload_u32(UINT32_C( 586841937), UINT32_C(2995843402)),
328
+ simde_x_vload_u32(UINT32_C(1275996594), UINT32_C(4093476624)) },
329
+ { simde_x_vload_u32(UINT32_C( 493330379), UINT32_C(3333627124)),
330
+ simde_x_vload_u32(UINT32_C(4056152600), UINT32_C( 122221865)),
331
+ simde_x_vload_u32(UINT32_C(3328456968), UINT32_C( 235218708)) }
332
+ };
333
+
334
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
335
+ simde_uint32x2_t r = simde_vmul_u32(test_vec[i].a, test_vec[i].b);
336
+ simde_neon_assert_uint32x2(r, ==, test_vec[i].r);
337
+ }
338
+
339
+ return MUNIT_OK;
340
+ }
341
+
342
+ static MunitResult
343
+ test_simde_vmul_f32(const MunitParameter params[], void* data) {
344
+ (void) params;
345
+ (void) data;
346
+
347
+ const struct {
348
+ simde_float32x2_t a;
349
+ simde_float32x2_t b;
350
+ simde_float32x2_t r;
351
+ } test_vec[8] = {
352
+ { simde_x_vload_f32(SIMDE_FLOAT32_C( 709.32), SIMDE_FLOAT32_C( 328.06)),
353
+ simde_x_vload_f32(SIMDE_FLOAT32_C( -587.68), SIMDE_FLOAT32_C( 662.19)),
354
+ simde_x_vload_f32(SIMDE_FLOAT32_C(-416853.19), SIMDE_FLOAT32_C(217238.05)) },
355
+ { simde_x_vload_f32(SIMDE_FLOAT32_C( 272.36), SIMDE_FLOAT32_C( -510.45)),
356
+ simde_x_vload_f32(SIMDE_FLOAT32_C( 160.08), SIMDE_FLOAT32_C( -874.02)),
357
+ simde_x_vload_f32(SIMDE_FLOAT32_C(43599.39), SIMDE_FLOAT32_C(446143.53)) },
358
+ { simde_x_vload_f32(SIMDE_FLOAT32_C( -793.25), SIMDE_FLOAT32_C( 239.08)),
359
+ simde_x_vload_f32(SIMDE_FLOAT32_C( 670.43), SIMDE_FLOAT32_C( 908.32)),
360
+ simde_x_vload_f32(SIMDE_FLOAT32_C(-531818.56), SIMDE_FLOAT32_C(217161.16)) },
361
+ { simde_x_vload_f32(SIMDE_FLOAT32_C( 536.53), SIMDE_FLOAT32_C( 128.77)),
362
+ simde_x_vload_f32(SIMDE_FLOAT32_C( 717.28), SIMDE_FLOAT32_C( 92.51)),
363
+ simde_x_vload_f32(SIMDE_FLOAT32_C(384842.28), SIMDE_FLOAT32_C(11912.51)) },
364
+ { simde_x_vload_f32(SIMDE_FLOAT32_C( 477.63), SIMDE_FLOAT32_C( 569.89)),
365
+ simde_x_vload_f32(SIMDE_FLOAT32_C( -104.64), SIMDE_FLOAT32_C( -250.44)),
366
+ simde_x_vload_f32(SIMDE_FLOAT32_C(-49979.20), SIMDE_FLOAT32_C(-142723.25)) },
367
+ { simde_x_vload_f32(SIMDE_FLOAT32_C( -148.80), SIMDE_FLOAT32_C( 904.67)),
368
+ simde_x_vload_f32(SIMDE_FLOAT32_C( -644.39), SIMDE_FLOAT32_C( -98.71)),
369
+ simde_x_vload_f32(SIMDE_FLOAT32_C(95885.23), SIMDE_FLOAT32_C(-89299.98)) },
370
+ { simde_x_vload_f32(SIMDE_FLOAT32_C( -491.91), SIMDE_FLOAT32_C( -357.57)),
371
+ simde_x_vload_f32(SIMDE_FLOAT32_C( 647.10), SIMDE_FLOAT32_C( -816.96)),
372
+ simde_x_vload_f32(SIMDE_FLOAT32_C(-318314.94), SIMDE_FLOAT32_C(292120.41)) },
373
+ { simde_x_vload_f32(SIMDE_FLOAT32_C( 942.18), SIMDE_FLOAT32_C( -728.13)),
374
+ simde_x_vload_f32(SIMDE_FLOAT32_C( 796.63), SIMDE_FLOAT32_C( -402.16)),
375
+ simde_x_vload_f32(SIMDE_FLOAT32_C(750568.88), SIMDE_FLOAT32_C(292824.75)) }
376
+ };
377
+
378
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
379
+ simde_float32x2_t r = simde_vmul_f32(test_vec[i].a, test_vec[i].b);
380
+ simde_neon_assert_float32x2_equal(r, test_vec[i].r, 1);
381
+ }
382
+
383
+ return MUNIT_OK;
384
+ }
385
+
386
+ static MunitResult
387
+ test_simde_vmul_f64(const MunitParameter params[], void* data) {
388
+ (void) params;
389
+ (void) data;
390
+
391
+ const struct {
392
+ simde_float64x1_t a;
393
+ simde_float64x1_t b;
394
+ simde_float64x1_t r;
395
+ } test_vec[8] = {
396
+ { simde_x_vload_f64(SIMDE_FLOAT64_C( 550.79)),
397
+ simde_x_vload_f64(SIMDE_FLOAT64_C( 249.76)),
398
+ simde_x_vload_f64(SIMDE_FLOAT64_C(137565.31)) },
399
+ { simde_x_vload_f64(SIMDE_FLOAT64_C( -302.70)),
400
+ simde_x_vload_f64(SIMDE_FLOAT64_C( -373.11)),
401
+ simde_x_vload_f64(SIMDE_FLOAT64_C(112940.40)) },
402
+ { simde_x_vload_f64(SIMDE_FLOAT64_C( 541.09)),
403
+ simde_x_vload_f64(SIMDE_FLOAT64_C( 765.69)),
404
+ simde_x_vload_f64(SIMDE_FLOAT64_C(414307.20)) },
405
+ { simde_x_vload_f64(SIMDE_FLOAT64_C( 829.25)),
406
+ simde_x_vload_f64(SIMDE_FLOAT64_C( 607.20)),
407
+ simde_x_vload_f64(SIMDE_FLOAT64_C(503520.60)) },
408
+ { simde_x_vload_f64(SIMDE_FLOAT64_C( -860.60)),
409
+ simde_x_vload_f64(SIMDE_FLOAT64_C( -496.29)),
410
+ simde_x_vload_f64(SIMDE_FLOAT64_C(427107.17)) },
411
+ { simde_x_vload_f64(SIMDE_FLOAT64_C( 398.42)),
412
+ simde_x_vload_f64(SIMDE_FLOAT64_C( 348.46)),
413
+ simde_x_vload_f64(SIMDE_FLOAT64_C(138833.43)) },
414
+ { simde_x_vload_f64(SIMDE_FLOAT64_C( -13.38)),
415
+ simde_x_vload_f64(SIMDE_FLOAT64_C( 905.20)),
416
+ simde_x_vload_f64(SIMDE_FLOAT64_C(-12111.58)) },
417
+ { simde_x_vload_f64(SIMDE_FLOAT64_C( -883.24)),
418
+ simde_x_vload_f64(SIMDE_FLOAT64_C( 588.39)),
419
+ simde_x_vload_f64(SIMDE_FLOAT64_C(-519689.58)) }
420
+ };
421
+
422
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
423
+ simde_float64x1_t r = simde_vmul_f64(test_vec[i].a, test_vec[i].b);
424
+ simde_neon_assert_float64x1_equal(r, test_vec[i].r, 1);
425
+ }
426
+
427
+ return MUNIT_OK;
428
+ }
429
+
430
+ static MunitResult
431
+ test_simde_vmulq_s8(const MunitParameter params[], void* data) {
432
+ (void) params;
433
+ (void) data;
434
+
435
+ const struct {
436
+ simde_int8x16_t a;
437
+ simde_int8x16_t b;
438
+ simde_int8x16_t r;
439
+ } test_vec[8] = {
440
+ { simde_x_vloadq_s8(INT8_C(-127), INT8_C( 30), INT8_C( -94), INT8_C( 45),
441
+ INT8_C( 22), INT8_C( -3), INT8_C( 39), INT8_C( 115),
442
+ INT8_C( 112), INT8_C( -51), INT8_C( 121), INT8_C( 125),
443
+ INT8_C( -78), INT8_C( 69), INT8_C( -42), INT8_C( 74)),
444
+ simde_x_vloadq_s8(INT8_C( 35), INT8_C( -23), INT8_C( -17), INT8_C(-126),
445
+ INT8_C( -31), INT8_C(-106), INT8_C( -83), INT8_C(-118),
446
+ INT8_C( 40), INT8_C(-126), INT8_C( -6), INT8_C( 38),
447
+ INT8_C( -47), INT8_C( 106), INT8_C( -22), INT8_C( -69)),
448
+ simde_x_vloadq_s8(INT8_C( -93), INT8_C( 78), INT8_C( 62), INT8_C( -38),
449
+ INT8_C( 86), INT8_C( 62), INT8_C( 91), INT8_C( -2),
450
+ INT8_C(-128), INT8_C( 26), INT8_C( 42), INT8_C(-114),
451
+ INT8_C( 82), INT8_C(-110), INT8_C(-100), INT8_C( 14)) },
452
+ { simde_x_vloadq_s8(INT8_C( -46), INT8_C( 10), INT8_C( -18), INT8_C( -86),
453
+ INT8_C( 84), INT8_C( -75), INT8_C( -6), INT8_C( -52),
454
+ INT8_C( -15), INT8_C( 62), INT8_C( 113), INT8_C( 52),
455
+ INT8_C( 118), INT8_C( 40), INT8_C(-114), INT8_C( -53)),
456
+ simde_x_vloadq_s8(INT8_C( 118), INT8_C( -3), INT8_C( 61), INT8_C( 26),
457
+ INT8_C(-111), INT8_C( 7), INT8_C( -62), INT8_C(-109),
458
+ INT8_C( 75), INT8_C( 107), INT8_C( 24), INT8_C( -50),
459
+ INT8_C( -58), INT8_C( -11), INT8_C(-125), INT8_C( 65)),
460
+ simde_x_vloadq_s8(INT8_C( -52), INT8_C( -30), INT8_C( -74), INT8_C( 68),
461
+ INT8_C(-108), INT8_C( -13), INT8_C( 116), INT8_C( 36),
462
+ INT8_C(-101), INT8_C( -22), INT8_C(-104), INT8_C( -40),
463
+ INT8_C( 68), INT8_C( 72), INT8_C( -86), INT8_C(-117)) },
464
+ { simde_x_vloadq_s8(INT8_C( 24), INT8_C( -71), INT8_C( 80), INT8_C( -24),
465
+ INT8_C( 97), INT8_C( 8), INT8_C( 91), INT8_C( -11),
466
+ INT8_C( -1), INT8_C( 82), INT8_C( 52), INT8_C( 98),
467
+ INT8_C( -56), INT8_C( 61), INT8_C( 103), INT8_C( 17)),
468
+ simde_x_vloadq_s8(INT8_C( 108), INT8_C(-120), INT8_C( 84), INT8_C( 59),
469
+ INT8_C( -97), INT8_C( -12), INT8_C( 19), INT8_C(-111),
470
+ INT8_C( -18), INT8_C( 16), INT8_C(-109), INT8_C( 92),
471
+ INT8_C( 19), INT8_C( -30), INT8_C( 92), INT8_C( 34)),
472
+ simde_x_vloadq_s8(INT8_C( 32), INT8_C( 72), INT8_C( 64), INT8_C( 120),
473
+ INT8_C( 63), INT8_C( -96), INT8_C( -63), INT8_C( -59),
474
+ INT8_C( 18), INT8_C( 32), INT8_C( -36), INT8_C( 56),
475
+ INT8_C( -40), INT8_C( -38), INT8_C( 4), INT8_C( 66)) },
476
+ { simde_x_vloadq_s8(INT8_C( 119), INT8_C( 38), INT8_C(-126), INT8_C(-104),
477
+ INT8_C( -9), INT8_C( -61), INT8_C( -53), INT8_C(-102),
478
+ INT8_C( 34), INT8_C(-123), INT8_C( 94), INT8_C( 102),
479
+ INT8_C( 126), INT8_C( -98), INT8_C( -82), INT8_C( 70)),
480
+ simde_x_vloadq_s8(INT8_C( -36), INT8_C( 31), INT8_C( -43), INT8_C( -87),
481
+ INT8_C(-101), INT8_C( 37), INT8_C( 51), INT8_C( 62),
482
+ INT8_C( 55), INT8_C( 107), INT8_C( 19), INT8_C( -79),
483
+ INT8_C( -67), INT8_C(-104), INT8_C( -66), INT8_C( -28)),
484
+ simde_x_vloadq_s8(INT8_C( 68), INT8_C(-102), INT8_C( 42), INT8_C( 88),
485
+ INT8_C(-115), INT8_C( 47), INT8_C( 113), INT8_C( 76),
486
+ INT8_C( 78), INT8_C(-105), INT8_C( -6), INT8_C(-122),
487
+ INT8_C( 6), INT8_C( -48), INT8_C( 36), INT8_C( 88)) },
488
+ { simde_x_vloadq_s8(INT8_C( 34), INT8_C( -79), INT8_C( -76), INT8_C( 98),
489
+ INT8_C( 2), INT8_C( -94), INT8_C( 51), INT8_C( 68),
490
+ INT8_C(-107), INT8_C( 82), INT8_C( -6), INT8_C( -35),
491
+ INT8_C( 7), INT8_C( -99), INT8_C( 36), INT8_C( 3)),
492
+ simde_x_vloadq_s8(INT8_C( 88), INT8_C( 9), INT8_C( -47), INT8_C( -9),
493
+ INT8_C( 52), INT8_C( 114), INT8_C( 79), INT8_C( 52),
494
+ INT8_C( 3), INT8_C( -2), INT8_C(-115), INT8_C( -50),
495
+ INT8_C( 58), INT8_C( 37), INT8_C(-126), INT8_C( 9)),
496
+ simde_x_vloadq_s8(INT8_C( -80), INT8_C( 57), INT8_C( -12), INT8_C(-114),
497
+ INT8_C( 104), INT8_C( 36), INT8_C( -67), INT8_C( -48),
498
+ INT8_C( -65), INT8_C( 92), INT8_C( -78), INT8_C( -42),
499
+ INT8_C(-106), INT8_C( -79), INT8_C( 72), INT8_C( 27)) },
500
+ { simde_x_vloadq_s8(INT8_C( 10), INT8_C( 124), INT8_C( 81), INT8_C( 42),
501
+ INT8_C( 82), INT8_C( -37), INT8_C( 125), INT8_C( 86),
502
+ INT8_C( -51), INT8_C(-126), INT8_C( -82), INT8_C( -29),
503
+ INT8_C( 74), INT8_C( 115), INT8_C( -92), INT8_C(-111)),
504
+ simde_x_vloadq_s8(INT8_C( -46), INT8_C( 126), INT8_C( -67), INT8_C( 122),
505
+ INT8_C( 17), INT8_C( -26), INT8_C( 92), INT8_C( 45),
506
+ INT8_C( -76), INT8_C( 13), INT8_C( 59), INT8_C( 23),
507
+ INT8_C( -74), INT8_C( -28), INT8_C(-126), INT8_C( -57)),
508
+ simde_x_vloadq_s8(INT8_C( 52), INT8_C( 8), INT8_C( -51), INT8_C( 4),
509
+ INT8_C( 114), INT8_C( -62), INT8_C( -20), INT8_C( 30),
510
+ INT8_C( 36), INT8_C(-102), INT8_C( 26), INT8_C( 101),
511
+ INT8_C(-100), INT8_C( 108), INT8_C( 72), INT8_C( -73)) },
512
+ { simde_x_vloadq_s8(INT8_C(-110), INT8_C( 103), INT8_C( -90), INT8_C( -37),
513
+ INT8_C( 10), INT8_C( -77), INT8_C( -99), INT8_C( -84),
514
+ INT8_C( 39), INT8_C( 66), INT8_C( -48), INT8_C( -53),
515
+ INT8_C( 76), INT8_C( 34), INT8_C( -75), INT8_C( 106)),
516
+ simde_x_vloadq_s8(INT8_C( -23), INT8_C( 95), INT8_C( -41), INT8_C( -9),
517
+ INT8_C(-122), INT8_C( 76), INT8_C(-125), INT8_C( -69),
518
+ INT8_C( 22), INT8_C( -84), INT8_C( 36), INT8_C( 108),
519
+ INT8_C(-116), INT8_C(-127), INT8_C( 2), INT8_C( 124)),
520
+ simde_x_vloadq_s8(INT8_C( -30), INT8_C( 57), INT8_C( 106), INT8_C( 77),
521
+ INT8_C( 60), INT8_C( 36), INT8_C( 87), INT8_C( -92),
522
+ INT8_C( 90), INT8_C( 88), INT8_C( 64), INT8_C( -92),
523
+ INT8_C(-112), INT8_C( 34), INT8_C( 106), INT8_C( 88)) },
524
+ { simde_x_vloadq_s8(INT8_C( -61), INT8_C( -14), INT8_C( 90), INT8_C( 34),
525
+ INT8_C( -60), INT8_C( 47), INT8_C( 73), INT8_C( 4),
526
+ INT8_C( 101), INT8_C( -83), INT8_C( 12), INT8_C( -87),
527
+ INT8_C( -41), INT8_C( 33), INT8_C( 95), INT8_C( 47)),
528
+ simde_x_vloadq_s8(INT8_C( 43), INT8_C( -60), INT8_C( 103), INT8_C( 61),
529
+ INT8_C( 87), INT8_C( -37), INT8_C( 8), INT8_C( 89),
530
+ INT8_C( 78), INT8_C( -98), INT8_C(-109), INT8_C( 95),
531
+ INT8_C( 3), INT8_C( -59), INT8_C( 114), INT8_C( -99)),
532
+ simde_x_vloadq_s8(INT8_C( -63), INT8_C( 72), INT8_C( 54), INT8_C( 26),
533
+ INT8_C(-100), INT8_C( 53), INT8_C( 72), INT8_C( 100),
534
+ INT8_C( -58), INT8_C( -58), INT8_C( -28), INT8_C( -73),
535
+ INT8_C(-123), INT8_C( 101), INT8_C( 78), INT8_C( -45)) }
536
+ };
537
+
538
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
539
+ simde_int8x16_t r = simde_vmulq_s8(test_vec[i].a, test_vec[i].b);
540
+ simde_neon_assert_int8x16(r, ==, test_vec[i].r);
541
+ }
542
+
543
+ return MUNIT_OK;
544
+ }
545
+
546
+ static MunitResult
547
+ test_simde_vmulq_s16(const MunitParameter params[], void* data) {
548
+ (void) params;
549
+ (void) data;
550
+
551
+ const struct {
552
+ simde_int16x8_t a;
553
+ simde_int16x8_t b;
554
+ simde_int16x8_t r;
555
+ } test_vec[8] = {
556
+ { simde_x_vloadq_s16(INT16_C( 3606), INT16_C(-26146), INT16_C( 29578), INT16_C(-21917),
557
+ INT16_C( 16291), INT16_C(-12460), INT16_C( 16844), INT16_C( -7447)),
558
+ simde_x_vloadq_s16(INT16_C(-14567), INT16_C( -9750), INT16_C(-13348), INT16_C(-21787),
559
+ INT16_C(-10512), INT16_C(-12251), INT16_C(-12570), INT16_C( 1106)),
560
+ simde_x_vloadq_s16(INT16_C( 31270), INT16_C(-11540), INT16_C(-18280), INT16_C( 10383),
561
+ INT16_C( -5424), INT16_C( 14116), INT16_C( 17736), INT16_C( 21154)) },
562
+ { simde_x_vloadq_s16(INT16_C( -110), INT16_C( 17978), INT16_C(-18156), INT16_C( 14375),
563
+ INT16_C(-18791), INT16_C( 12877), INT16_C( 31925), INT16_C(-31559)),
564
+ simde_x_vloadq_s16(INT16_C(-30054), INT16_C(-28348), INT16_C( 11497), INT16_C(-23041),
565
+ INT16_C(-12411), INT16_C( 8254), INT16_C(-11572), INT16_C( 28407)),
566
+ simde_x_vloadq_s16(INT16_C( 29140), INT16_C(-32408), INT16_C( -7372), INT16_C( 4569),
567
+ INT16_C(-27523), INT16_C(-12634), INT16_C( -9668), INT16_C(-29569)) },
568
+ { simde_x_vloadq_s16(INT16_C( -5983), INT16_C( 29026), INT16_C( 18029), INT16_C( 5032),
569
+ INT16_C(-21014), INT16_C(-14771), INT16_C( 32443), INT16_C(-21683)),
570
+ simde_x_vloadq_s16(INT16_C(-17878), INT16_C( -2583), INT16_C( 19128), INT16_C( -6228),
571
+ INT16_C( 28126), INT16_C(-22741), INT16_C(-24337), INT16_C(-26084)),
572
+ simde_x_vloadq_s16(INT16_C( 9322), INT16_C( -974), INT16_C( 8280), INT16_C(-13088),
573
+ INT16_C( 29420), INT16_C(-30225), INT16_C( 12437), INT16_C( 3692)) },
574
+ { simde_x_vloadq_s16(INT16_C( -2929), INT16_C( -6488), INT16_C(-13187), INT16_C( -1982),
575
+ INT16_C( 17773), INT16_C( 15022), INT16_C( 13905), INT16_C( 10993)),
576
+ simde_x_vloadq_s16(INT16_C( 16456), INT16_C(-22704), INT16_C(-19837), INT16_C( 30752),
577
+ INT16_C(-10591), INT16_C(-26882), INT16_C( 2522), INT16_C( 9207)),
578
+ simde_x_vloadq_s16(INT16_C(-30664), INT16_C(-21376), INT16_C(-29193), INT16_C( -1984),
579
+ INT16_C(-14451), INT16_C( 11428), INT16_C( 6650), INT16_C( 24967)) },
580
+ { simde_x_vloadq_s16(INT16_C( -8487), INT16_C( 18864), INT16_C(-23663), INT16_C( -2393),
581
+ INT16_C(-30875), INT16_C( 25925), INT16_C( 115), INT16_C(-31502)),
582
+ simde_x_vloadq_s16(INT16_C( 19805), INT16_C(-22785), INT16_C( 13361), INT16_C( 31862),
583
+ INT16_C(-27097), INT16_C(-13524), INT16_C( 23305), INT16_C( 4600)),
584
+ simde_x_vloadq_s16(INT16_C( 14805), INT16_C(-31152), INT16_C(-15679), INT16_C(-27398),
585
+ INT16_C(-12701), INT16_C( 7900), INT16_C( -6901), INT16_C( -9104)) },
586
+ { simde_x_vloadq_s16(INT16_C( 3051), INT16_C( 10349), INT16_C( 12948), INT16_C( 9215),
587
+ INT16_C( -2354), INT16_C(-30618), INT16_C(-28405), INT16_C( 9209)),
588
+ simde_x_vloadq_s16(INT16_C(-12161), INT16_C(-14552), INT16_C(-19362), INT16_C( 27494),
589
+ INT16_C( -8686), INT16_C( 1070), INT16_C(-30539), INT16_C( 25944)),
590
+ simde_x_vloadq_s16(INT16_C( -9835), INT16_C( 3080), INT16_C(-23976), INT16_C( -4966),
591
+ INT16_C( -388), INT16_C( 6740), INT16_C( 25799), INT16_C(-25960)) },
592
+ { simde_x_vloadq_s16(INT16_C(-12070), INT16_C( 24160), INT16_C(-23416), INT16_C(-30079),
593
+ INT16_C( 5294), INT16_C( 11657), INT16_C(-31940), INT16_C( 7362)),
594
+ simde_x_vloadq_s16(INT16_C(-14737), INT16_C( 20723), INT16_C( 9772), INT16_C(-12264),
595
+ INT16_C( 24972), INT16_C(-24660), INT16_C( -4901), INT16_C( 23959)),
596
+ simde_x_vloadq_s16(INT16_C( 10886), INT16_C(-27360), INT16_C( 30560), INT16_C(-13288),
597
+ INT16_C( 15656), INT16_C(-20724), INT16_C(-27564), INT16_C( 28782)) },
598
+ { simde_x_vloadq_s16(INT16_C(-19244), INT16_C(-24194), INT16_C(-25952), INT16_C(-16694),
599
+ INT16_C( 20006), INT16_C(-12804), INT16_C( 17167), INT16_C( -934)),
600
+ simde_x_vloadq_s16(INT16_C( -5526), INT16_C( 27910), INT16_C(-12649), INT16_C( -8536),
601
+ INT16_C( 32495), INT16_C( 12909), INT16_C( 11959), INT16_C( 8750)),
602
+ simde_x_vloadq_s16(INT16_C(-22584), INT16_C( 28404), INT16_C( -2976), INT16_C( 24720),
603
+ INT16_C(-22150), INT16_C( -5044), INT16_C(-24135), INT16_C( 19500)) }
604
+ };
605
+
606
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
607
+ simde_int16x8_t r = simde_vmulq_s16(test_vec[i].a, test_vec[i].b);
608
+ simde_neon_assert_int16x8(r, ==, test_vec[i].r);
609
+ }
610
+
611
+ return MUNIT_OK;
612
+ }
613
+
614
+ static MunitResult
615
+ test_simde_vmulq_s32(const MunitParameter params[], void* data) {
616
+ (void) params;
617
+ (void) data;
618
+
619
+ const struct {
620
+ simde_int32x4_t a;
621
+ simde_int32x4_t b;
622
+ simde_int32x4_t r;
623
+ } test_vec[8] = {
624
+ { simde_x_vloadq_s32(INT32_C( 1123170826), INT32_C( 1724201296),
625
+ INT32_C(-1071829400), INT32_C( -605407000)),
626
+ simde_x_vloadq_s32(INT32_C(-1297699783), INT32_C( 827772923),
627
+ INT32_C( 918675174), INT32_C( -817623083)),
628
+ simde_x_vloadq_s32(INT32_C( 1030333498), INT32_C( -845749904),
629
+ INT32_C( -723656336), INT32_C( 1758916872)) },
630
+ { simde_x_vloadq_s32(INT32_C( 1918919185), INT32_C( 262393006),
631
+ INT32_C( 1657094748), INT32_C( 837021098)),
632
+ simde_x_vloadq_s32(INT32_C( 1949462567), INT32_C(-1250655484),
633
+ INT32_C( 1421638944), INT32_C(-1778492349)),
634
+ simde_x_vloadq_s32(INT32_C( 1927831703), INT32_C( -719045448),
635
+ INT32_C(-1784727680), INT32_C(-1195619458)) },
636
+ { simde_x_vloadq_s32(INT32_C(-1515943730), INT32_C( 2003246422),
637
+ INT32_C( 1694426031), INT32_C(-1948948754)),
638
+ simde_x_vloadq_s32(INT32_C( -398862630), INT32_C( 2132045133),
639
+ INT32_C(-1807602235), INT32_C( 1783413341)),
640
+ simde_x_vloadq_s32(INT32_C(-1101614228), INT32_C( -969314082),
641
+ INT32_C( 1699281067), INT32_C( -201895818)) },
642
+ { simde_x_vloadq_s32(INT32_C( -764500520), INT32_C(-1890077122),
643
+ INT32_C( -441183161), INT32_C( 1043231808)),
644
+ simde_x_vloadq_s32(INT32_C( 1295880986), INT32_C( 533621567),
645
+ INT32_C( -165031911), INT32_C( -725208039)),
646
+ simde_x_vloadq_s32(INT32_C( 1693442032), INT32_C( 396739394),
647
+ INT32_C( 946776815), INT32_C( 1806064192)) },
648
+ { simde_x_vloadq_s32(INT32_C(-1009530908), INT32_C(-1685480510),
649
+ INT32_C( -100061526), INT32_C( 486038522)),
650
+ simde_x_vloadq_s32(INT32_C(-2078046313), INT32_C( -449627785),
651
+ INT32_C(-1737499256), INT32_C(-1480286984)),
652
+ simde_x_vloadq_s32(INT32_C( 343942012), INT32_C(-2108974802),
653
+ INT32_C( -304116656), INT32_C( -368217552)) },
654
+ { simde_x_vloadq_s32(INT32_C(-1961990836), INT32_C( 1719683566),
655
+ INT32_C( 1283520022), INT32_C( 427779620)),
656
+ simde_x_vloadq_s32(INT32_C(-1041353110), INT32_C( 2047817713),
657
+ INT32_C( -453743286), INT32_C( -190808905)),
658
+ simde_x_vloadq_s32(INT32_C( 368834936), INT32_C( -812008690),
659
+ INT32_C( 721846364), INT32_C( 369849276)) },
660
+ { simde_x_vloadq_s32(INT32_C( -205929908), INT32_C(-1288775358),
661
+ INT32_C( 1260732697), INT32_C(-1728995848)),
662
+ simde_x_vloadq_s32(INT32_C( -474735779), INT32_C(-1917133953),
663
+ INT32_C( 269438727), INT32_C( 588750565)),
664
+ simde_x_vloadq_s32(INT32_C( 594123164), INT32_C( 739065278),
665
+ INT32_C( -309166417), INT32_C( 149511896)) },
666
+ { simde_x_vloadq_s32(INT32_C( 1285818920), INT32_C( 251500733),
667
+ INT32_C( -957328722), INT32_C( 376096822)),
668
+ simde_x_vloadq_s32(INT32_C(-1080123710), INT32_C( -43912310),
669
+ INT32_C( -764093305), INT32_C( 1166743576)),
670
+ simde_x_vloadq_s32(INT32_C( 1866752592), INT32_C( 1967963362),
671
+ INT32_C(-1566498366), INT32_C( 1612305680)) }
672
+ };
673
+
674
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
675
+ simde_int32x4_t r = simde_vmulq_s32(test_vec[i].a, test_vec[i].b);
676
+ simde_neon_assert_int32x4(r, ==, test_vec[i].r);
677
+ }
678
+
679
+ return MUNIT_OK;
680
+ }
681
+
682
+ static MunitResult
683
+ test_simde_vmulq_u8(const MunitParameter params[], void* data) {
684
+ (void) params;
685
+ (void) data;
686
+
687
+ const struct {
688
+ simde_uint8x16_t a;
689
+ simde_uint8x16_t b;
690
+ simde_uint8x16_t r;
691
+ } test_vec[8] = {
692
+ { simde_x_vloadq_u8(UINT8_C(126), UINT8_C( 74), UINT8_C( 80), UINT8_C(137),
693
+ UINT8_C( 99), UINT8_C(157), UINT8_C(205), UINT8_C(210),
694
+ UINT8_C( 39), UINT8_C( 80), UINT8_C(198), UINT8_C( 44),
695
+ UINT8_C(160), UINT8_C(209), UINT8_C(123), UINT8_C(223) ),
696
+ simde_x_vloadq_u8(UINT8_C(164), UINT8_C( 57), UINT8_C(116), UINT8_C(244),
697
+ UINT8_C( 62), UINT8_C( 78), UINT8_C(109), UINT8_C(114),
698
+ UINT8_C( 12), UINT8_C(108), UINT8_C( 47), UINT8_C(217),
699
+ UINT8_C( 64), UINT8_C(230), UINT8_C( 89), UINT8_C(254) ),
700
+ simde_x_vloadq_u8(UINT8_C(184), UINT8_C(122), UINT8_C( 64), UINT8_C(148),
701
+ UINT8_C(250), UINT8_C(214), UINT8_C( 73), UINT8_C(132),
702
+ UINT8_C(212), UINT8_C(192), UINT8_C( 90), UINT8_C( 76),
703
+ UINT8_C( 0), UINT8_C(198), UINT8_C(195), UINT8_C( 66) ) },
704
+ { simde_x_vloadq_u8(UINT8_C(130), UINT8_C(192), UINT8_C(246), UINT8_C(206),
705
+ UINT8_C(120), UINT8_C(140), UINT8_C(116), UINT8_C(157),
706
+ UINT8_C(128), UINT8_C( 75), UINT8_C(144), UINT8_C(138),
707
+ UINT8_C(244), UINT8_C( 96), UINT8_C(118), UINT8_C(213) ),
708
+ simde_x_vloadq_u8(UINT8_C(132), UINT8_C(156), UINT8_C( 48), UINT8_C( 81),
709
+ UINT8_C( 87), UINT8_C( 60), UINT8_C(167), UINT8_C( 88),
710
+ UINT8_C(202), UINT8_C(206), UINT8_C(182), UINT8_C(144),
711
+ UINT8_C(125), UINT8_C(103), UINT8_C( 27), UINT8_C(109) ),
712
+ simde_x_vloadq_u8(UINT8_C( 8), UINT8_C( 0), UINT8_C( 32), UINT8_C( 46),
713
+ UINT8_C(200), UINT8_C(208), UINT8_C(172), UINT8_C(248),
714
+ UINT8_C( 0), UINT8_C( 90), UINT8_C( 96), UINT8_C(160),
715
+ UINT8_C( 36), UINT8_C(160), UINT8_C(114), UINT8_C(177) ) },
716
+ { simde_x_vloadq_u8(UINT8_C( 57), UINT8_C( 69), UINT8_C( 12), UINT8_C(222),
717
+ UINT8_C( 62), UINT8_C( 70), UINT8_C(174), UINT8_C( 65),
718
+ UINT8_C(234), UINT8_C(200), UINT8_C(157), UINT8_C( 2),
719
+ UINT8_C(123), UINT8_C(247), UINT8_C( 90), UINT8_C(161) ),
720
+ simde_x_vloadq_u8(UINT8_C(243), UINT8_C( 52), UINT8_C(110), UINT8_C(121),
721
+ UINT8_C( 94), UINT8_C(106), UINT8_C(150), UINT8_C(135),
722
+ UINT8_C( 38), UINT8_C(108), UINT8_C( 84), UINT8_C(152),
723
+ UINT8_C( 79), UINT8_C( 50), UINT8_C(197), UINT8_C(194) ),
724
+ simde_x_vloadq_u8(UINT8_C( 27), UINT8_C( 4), UINT8_C( 40), UINT8_C(238),
725
+ UINT8_C(196), UINT8_C(252), UINT8_C(244), UINT8_C( 71),
726
+ UINT8_C(188), UINT8_C( 96), UINT8_C(132), UINT8_C( 48),
727
+ UINT8_C(245), UINT8_C( 62), UINT8_C( 66), UINT8_C( 2) ) },
728
+ { simde_x_vloadq_u8(UINT8_C( 31), UINT8_C( 18), UINT8_C( 4), UINT8_C( 19),
729
+ UINT8_C(118), UINT8_C(179), UINT8_C(138), UINT8_C( 53),
730
+ UINT8_C(106), UINT8_C(228), UINT8_C( 75), UINT8_C(106),
731
+ UINT8_C(142), UINT8_C(177), UINT8_C(130), UINT8_C(187) ),
732
+ simde_x_vloadq_u8(UINT8_C( 81), UINT8_C(104), UINT8_C( 98), UINT8_C(136),
733
+ UINT8_C(147), UINT8_C( 62), UINT8_C(226), UINT8_C(107),
734
+ UINT8_C( 61), UINT8_C(139), UINT8_C(232), UINT8_C( 20),
735
+ UINT8_C(198), UINT8_C(120), UINT8_C( 39), UINT8_C( 95) ),
736
+ simde_x_vloadq_u8(UINT8_C(207), UINT8_C( 80), UINT8_C(136), UINT8_C( 24),
737
+ UINT8_C(194), UINT8_C( 90), UINT8_C(212), UINT8_C( 39),
738
+ UINT8_C( 66), UINT8_C(204), UINT8_C(248), UINT8_C( 72),
739
+ UINT8_C(212), UINT8_C(248), UINT8_C(206), UINT8_C(101) ) },
740
+ { simde_x_vloadq_u8(UINT8_C(115), UINT8_C( 58), UINT8_C(178), UINT8_C(232),
741
+ UINT8_C(203), UINT8_C( 40), UINT8_C( 37), UINT8_C(168),
742
+ UINT8_C(178), UINT8_C(135), UINT8_C(125), UINT8_C(214),
743
+ UINT8_C(244), UINT8_C( 13), UINT8_C(201), UINT8_C(183) ),
744
+ simde_x_vloadq_u8(UINT8_C(182), UINT8_C( 21), UINT8_C(101), UINT8_C( 9),
745
+ UINT8_C(209), UINT8_C( 43), UINT8_C(159), UINT8_C(143),
746
+ UINT8_C( 21), UINT8_C(212), UINT8_C( 67), UINT8_C( 34),
747
+ UINT8_C( 83), UINT8_C( 77), UINT8_C(196), UINT8_C(245) ),
748
+ simde_x_vloadq_u8(UINT8_C(194), UINT8_C(194), UINT8_C( 58), UINT8_C( 40),
749
+ UINT8_C(187), UINT8_C(184), UINT8_C(251), UINT8_C(216),
750
+ UINT8_C(154), UINT8_C(204), UINT8_C(183), UINT8_C(108),
751
+ UINT8_C( 28), UINT8_C(233), UINT8_C(228), UINT8_C( 35) ) },
752
+ { simde_x_vloadq_u8(UINT8_C( 33), UINT8_C( 64), UINT8_C(252), UINT8_C(104),
753
+ UINT8_C( 64), UINT8_C(132), UINT8_C(155), UINT8_C( 5),
754
+ UINT8_C(192), UINT8_C(140), UINT8_C( 9), UINT8_C(119),
755
+ UINT8_C(111), UINT8_C(109), UINT8_C(122), UINT8_C(221) ),
756
+ simde_x_vloadq_u8(UINT8_C(233), UINT8_C( 26), UINT8_C(123), UINT8_C(153),
757
+ UINT8_C( 95), UINT8_C(120), UINT8_C(130), UINT8_C(168),
758
+ UINT8_C( 34), UINT8_C( 37), UINT8_C(150), UINT8_C(150),
759
+ UINT8_C(207), UINT8_C(153), UINT8_C(251), UINT8_C(245) ),
760
+ simde_x_vloadq_u8(UINT8_C( 9), UINT8_C(128), UINT8_C( 20), UINT8_C( 40),
761
+ UINT8_C(192), UINT8_C(224), UINT8_C(182), UINT8_C( 72),
762
+ UINT8_C(128), UINT8_C( 60), UINT8_C( 70), UINT8_C(186),
763
+ UINT8_C(193), UINT8_C( 37), UINT8_C(158), UINT8_C(129) ) },
764
+ { simde_x_vloadq_u8(UINT8_C( 20), UINT8_C(223), UINT8_C(245), UINT8_C(209),
765
+ UINT8_C(145), UINT8_C(183), UINT8_C(224), UINT8_C( 55),
766
+ UINT8_C(113), UINT8_C(226), UINT8_C(146), UINT8_C(229),
767
+ UINT8_C(255), UINT8_C(156), UINT8_C( 39), UINT8_C(101) ),
768
+ simde_x_vloadq_u8(UINT8_C(119), UINT8_C(140), UINT8_C( 6), UINT8_C(233),
769
+ UINT8_C( 68), UINT8_C( 82), UINT8_C( 80), UINT8_C( 61),
770
+ UINT8_C(229), UINT8_C(139), UINT8_C(236), UINT8_C(217),
771
+ UINT8_C( 55), UINT8_C(231), UINT8_C( 17), UINT8_C( 27) ),
772
+ simde_x_vloadq_u8(UINT8_C( 76), UINT8_C(244), UINT8_C(190), UINT8_C( 57),
773
+ UINT8_C(132), UINT8_C(158), UINT8_C( 0), UINT8_C( 27),
774
+ UINT8_C( 21), UINT8_C(182), UINT8_C(152), UINT8_C( 29),
775
+ UINT8_C(201), UINT8_C(196), UINT8_C(151), UINT8_C(167) ) },
776
+ { simde_x_vloadq_u8(UINT8_C(185), UINT8_C(200), UINT8_C(242), UINT8_C(242),
777
+ UINT8_C(137), UINT8_C(106), UINT8_C(236), UINT8_C( 70),
778
+ UINT8_C(127), UINT8_C(191), UINT8_C(140), UINT8_C( 2),
779
+ UINT8_C(230), UINT8_C(137), UINT8_C(180), UINT8_C( 63) ),
780
+ simde_x_vloadq_u8(UINT8_C(243), UINT8_C( 61), UINT8_C( 95), UINT8_C( 53),
781
+ UINT8_C(149), UINT8_C( 43), UINT8_C( 36), UINT8_C(164),
782
+ UINT8_C( 19), UINT8_C(160), UINT8_C(138), UINT8_C( 78),
783
+ UINT8_C(241), UINT8_C(165), UINT8_C(145), UINT8_C(172) ),
784
+ simde_x_vloadq_u8(UINT8_C(155), UINT8_C(168), UINT8_C(206), UINT8_C( 26),
785
+ UINT8_C(189), UINT8_C(206), UINT8_C( 48), UINT8_C(216),
786
+ UINT8_C(109), UINT8_C( 96), UINT8_C(120), UINT8_C(156),
787
+ UINT8_C(134), UINT8_C( 77), UINT8_C(244), UINT8_C( 84) ) }
788
+ };
789
+
790
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
791
+ simde_uint8x16_t r = simde_vmulq_u8(test_vec[i].a, test_vec[i].b);
792
+ simde_neon_assert_uint8x16(r, ==, test_vec[i].r);
793
+ }
794
+
795
+ return MUNIT_OK;
796
+ }
797
+
798
+ static MunitResult
799
+ test_simde_vmulq_u16(const MunitParameter params[], void* data) {
800
+ (void) params;
801
+ (void) data;
802
+
803
+ const struct {
804
+ simde_uint16x8_t a;
805
+ simde_uint16x8_t b;
806
+ simde_uint16x8_t r;
807
+ } test_vec[8] = {
808
+ { simde_x_vloadq_u16(UINT16_C(57954), UINT16_C(44854), UINT16_C(11016), UINT16_C( 7879),
809
+ UINT16_C(53913), UINT16_C(49380), UINT16_C( 4238), UINT16_C(16363)),
810
+ simde_x_vloadq_u16(UINT16_C(23701), UINT16_C(36025), UINT16_C(20163), UINT16_C(52240),
811
+ UINT16_C(58420), UINT16_C(35739), UINT16_C(47224), UINT16_C(50522)),
812
+ simde_x_vloadq_u16(UINT16_C(64266), UINT16_C( 9734), UINT16_C(14104), UINT16_C(32880),
813
+ UINT16_C( 2836), UINT16_C(38412), UINT16_C(53904), UINT16_C(20382)) },
814
+ { simde_x_vloadq_u16(UINT16_C(21507), UINT16_C(16851), UINT16_C(37450), UINT16_C(18168),
815
+ UINT16_C( 2860), UINT16_C(59306), UINT16_C(60364), UINT16_C(13145)),
816
+ simde_x_vloadq_u16(UINT16_C(55394), UINT16_C(37714), UINT16_C(65217), UINT16_C( 7971),
817
+ UINT16_C(10952), UINT16_C(29506), UINT16_C(37052), UINT16_C( 2147)),
818
+ simde_x_vloadq_u16(UINT16_C(45350), UINT16_C(16022), UINT16_C(46538), UINT16_C(48104),
819
+ UINT16_C(62048), UINT16_C( 6100), UINT16_C(59856), UINT16_C(41835)) },
820
+ { simde_x_vloadq_u16(UINT16_C(48292), UINT16_C(62594), UINT16_C( 3256), UINT16_C( 8059),
821
+ UINT16_C(36503), UINT16_C(23744), UINT16_C( 5869), UINT16_C( 8493)),
822
+ simde_x_vloadq_u16(UINT16_C(44723), UINT16_C(64161), UINT16_C(57320), UINT16_C(37355),
823
+ UINT16_C(50612), UINT16_C( 6289), UINT16_C(22199), UINT16_C(32637)),
824
+ simde_x_vloadq_u16(UINT16_C(24236), UINT16_C(47554), UINT16_C(52928), UINT16_C(37097),
825
+ UINT16_C(29996), UINT16_C(35008), UINT16_C( 363), UINT16_C(34297)) },
826
+ { simde_x_vloadq_u16(UINT16_C(13882), UINT16_C(48253), UINT16_C(61217), UINT16_C(33326),
827
+ UINT16_C( 3325), UINT16_C(37911), UINT16_C(32474), UINT16_C( 30)),
828
+ simde_x_vloadq_u16(UINT16_C(59817), UINT16_C(13531), UINT16_C(58022), UINT16_C(27945),
829
+ UINT16_C(18424), UINT16_C(18686), UINT16_C(57158), UINT16_C(37766)),
830
+ simde_x_vloadq_u16(UINT16_C(38474), UINT16_C(41711), UINT16_C(12646), UINT16_C(28510),
831
+ UINT16_C(49176), UINT16_C(26322), UINT16_C(38300), UINT16_C(18868)) },
832
+ { simde_x_vloadq_u16(UINT16_C(29922), UINT16_C(21770), UINT16_C(13743), UINT16_C(64131),
833
+ UINT16_C(44236), UINT16_C( 6320), UINT16_C( 1235), UINT16_C(12856)),
834
+ simde_x_vloadq_u16(UINT16_C(54137), UINT16_C( 9316), UINT16_C( 7539), UINT16_C( 9579),
835
+ UINT16_C( 9529), UINT16_C(46971), UINT16_C(14527), UINT16_C(13778)),
836
+ simde_x_vloadq_u16(UINT16_C(34002), UINT16_C(40936), UINT16_C(61597), UINT16_C(41921),
837
+ UINT16_C(62828), UINT16_C(44176), UINT16_C(49517), UINT16_C(51696)) },
838
+ { simde_x_vloadq_u16(UINT16_C(22341), UINT16_C(29452), UINT16_C(20004), UINT16_C(57521),
839
+ UINT16_C(15740), UINT16_C(16551), UINT16_C(49777), UINT16_C(63636)),
840
+ simde_x_vloadq_u16(UINT16_C(40908), UINT16_C(34280), UINT16_C(49053), UINT16_C(31260),
841
+ UINT16_C(58817), UINT16_C( 2677), UINT16_C(12774), UINT16_C(43621)),
842
+ simde_x_vloadq_u16(UINT16_C(26108), UINT16_C(32480), UINT16_C(51220), UINT16_C(60764),
843
+ UINT16_C(18044), UINT16_C( 4691), UINT16_C(21126), UINT16_C(23140)) },
844
+ { simde_x_vloadq_u16(UINT16_C(38925), UINT16_C(29195), UINT16_C(36350), UINT16_C( 1990),
845
+ UINT16_C( 2672), UINT16_C( 4756), UINT16_C(59542), UINT16_C(32965)),
846
+ simde_x_vloadq_u16(UINT16_C( 7599), UINT16_C(48968), UINT16_C( 7710), UINT16_C(33995),
847
+ UINT16_C(56860), UINT16_C(36311), UINT16_C(10145), UINT16_C(14871)),
848
+ simde_x_vloadq_u16(UINT16_C(27107), UINT16_C(18456), UINT16_C(26564), UINT16_C(16898),
849
+ UINT16_C(17472), UINT16_C( 7756), UINT16_C( 8278), UINT16_C(13235)) },
850
+ { simde_x_vloadq_u16(UINT16_C(50306), UINT16_C(14229), UINT16_C(19051), UINT16_C(24713),
851
+ UINT16_C(37797), UINT16_C(32691), UINT16_C(32387), UINT16_C(53422)),
852
+ simde_x_vloadq_u16(UINT16_C(52105), UINT16_C(57603), UINT16_C(35832), UINT16_C(27715),
853
+ UINT16_C(61559), UINT16_C( 2260), UINT16_C(37277), UINT16_C(38380)),
854
+ simde_x_vloadq_u16(UINT16_C(16274), UINT16_C(39871), UINT16_C(12456), UINT16_C( 4059),
855
+ UINT16_C(20915), UINT16_C(22588), UINT16_C(51543), UINT16_C(42600)) }
856
+ };
857
+
858
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
859
+ simde_uint16x8_t r = simde_vmulq_u16(test_vec[i].a, test_vec[i].b);
860
+ simde_neon_assert_uint16x8(r, ==, test_vec[i].r);
861
+ }
862
+
863
+ return MUNIT_OK;
864
+ }
865
+
866
+ static MunitResult
867
+ test_simde_vmulq_u32(const MunitParameter params[], void* data) {
868
+ (void) params;
869
+ (void) data;
870
+
871
+ const struct {
872
+ simde_uint32x4_t a;
873
+ simde_uint32x4_t b;
874
+ simde_uint32x4_t r;
875
+ } test_vec[8] = {
876
+ { simde_x_vloadq_u32(UINT32_C(3958288120), UINT32_C( 956872615),
877
+ UINT32_C(2294683009), UINT32_C(3805713255)),
878
+ simde_x_vloadq_u32(UINT32_C(1250198096), UINT32_C(3188317560),
879
+ UINT32_C(2975088804), UINT32_C(1157674059)),
880
+ simde_x_vloadq_u32(UINT32_C(3612384640), UINT32_C(2843754824),
881
+ UINT32_C(1671038628), UINT32_C( 394369837)) },
882
+ { simde_x_vloadq_u32(UINT32_C(3253804944), UINT32_C(2275803417),
883
+ UINT32_C(3217780119), UINT32_C(2805772733)),
884
+ simde_x_vloadq_u32(UINT32_C(2303755275), UINT32_C( 739359423),
885
+ UINT32_C( 509143968), UINT32_C(2387159198)),
886
+ simde_x_vloadq_u32(UINT32_C(2527238960), UINT32_C( 810548135),
887
+ UINT32_C(4205189984), UINT32_C( 941636262)) },
888
+ { simde_x_vloadq_u32(UINT32_C(3316533415), UINT32_C(2372288697),
889
+ UINT32_C(2634527217), UINT32_C(3578420369)),
890
+ simde_x_vloadq_u32(UINT32_C(2318049332), UINT32_C(1737870643),
891
+ UINT32_C( 316097046), UINT32_C(1988165889)),
892
+ simde_x_vloadq_u32(UINT32_C( 104145388), UINT32_C( 668994011),
893
+ UINT32_C(3723521206), UINT32_C(2595745169)) },
894
+ { simde_x_vloadq_u32(UINT32_C(2183642275), UINT32_C(2838944061),
895
+ UINT32_C( 961498069), UINT32_C(4012733225)),
896
+ simde_x_vloadq_u32(UINT32_C(2455707487), UINT32_C(1102429086),
897
+ UINT32_C(3614870952), UINT32_C(2124555621)),
898
+ simde_x_vloadq_u32(UINT32_C( 320060797), UINT32_C(2510987942),
899
+ UINT32_C( 536614088), UINT32_C(4236598317)) },
900
+ { simde_x_vloadq_u32(UINT32_C(1746629002), UINT32_C(3420579250),
901
+ UINT32_C(2351405457), UINT32_C( 5315732)),
902
+ simde_x_vloadq_u32(UINT32_C(1374835167), UINT32_C(3729888048),
903
+ UINT32_C(1636607397), UINT32_C(1226229338)),
904
+ simde_x_vloadq_u32(UINT32_C(1565036854), UINT32_C(4203890528),
905
+ UINT32_C(3352063861), UINT32_C(2169930760)) },
906
+ { simde_x_vloadq_u32(UINT32_C( 549320667), UINT32_C(2322323920),
907
+ UINT32_C( 151754473), UINT32_C(4047667140)),
908
+ simde_x_vloadq_u32(UINT32_C(1818929917), UINT32_C(3103405141),
909
+ UINT32_C(1952449955), UINT32_C( 82542915)),
910
+ simde_x_vloadq_u32(UINT32_C(2265145711), UINT32_C( 321753616),
911
+ UINT32_C( 666084187), UINT32_C(4290065484)) },
912
+ { simde_x_vloadq_u32(UINT32_C(2077124603), UINT32_C(2666703247),
913
+ UINT32_C(2161899434), UINT32_C( 575602451)),
914
+ simde_x_vloadq_u32(UINT32_C( 251474702), UINT32_C(1393348005),
915
+ UINT32_C(3757489762), UINT32_C(4131780358)),
916
+ simde_x_vloadq_u32(UINT32_C(1628418234), UINT32_C( 333630507),
917
+ UINT32_C(2841705236), UINT32_C( 135548786)) },
918
+ { simde_x_vloadq_u32(UINT32_C(1086457379), UINT32_C( 14743562),
919
+ UINT32_C(1761520088), UINT32_C(2057869708)),
920
+ simde_x_vloadq_u32(UINT32_C(2746909344), UINT32_C(1754166210),
921
+ UINT32_C(4038189003), UINT32_C(2974027701)),
922
+ simde_x_vloadq_u32(UINT32_C(3201276896), UINT32_C(1601467796),
923
+ UINT32_C( 104584776), UINT32_C( 795067388)) }
924
+ };
925
+
926
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
927
+ simde_uint32x4_t r = simde_vmulq_u32(test_vec[i].a, test_vec[i].b);
928
+ simde_neon_assert_uint32x4(r, ==, test_vec[i].r);
929
+ }
930
+
931
+ return MUNIT_OK;
932
+ }
933
+
934
+ static MunitResult
935
+ test_simde_vmulq_f32(const MunitParameter params[], void* data) {
936
+ (void) params;
937
+ (void) data;
938
+
939
+ const struct {
940
+ simde_float32x4_t a;
941
+ simde_float32x4_t b;
942
+ simde_float32x4_t r;
943
+ } test_vec[8] = {
944
+ { simde_x_vloadq_f32(SIMDE_FLOAT32_C( 764.73), SIMDE_FLOAT32_C( -904.55),
945
+ SIMDE_FLOAT32_C( 692.13), SIMDE_FLOAT32_C( -510.35)),
946
+ simde_x_vloadq_f32(SIMDE_FLOAT32_C( 32.75), SIMDE_FLOAT32_C( 51.72),
947
+ SIMDE_FLOAT32_C( -142.44), SIMDE_FLOAT32_C( 194.67)),
948
+ simde_x_vloadq_f32(SIMDE_FLOAT32_C(25044.91), SIMDE_FLOAT32_C(-46783.33),
949
+ SIMDE_FLOAT32_C(-98587.00), SIMDE_FLOAT32_C(-99349.84)) },
950
+ { simde_x_vloadq_f32(SIMDE_FLOAT32_C( -904.10), SIMDE_FLOAT32_C( -331.30),
951
+ SIMDE_FLOAT32_C( 798.10), SIMDE_FLOAT32_C( -717.75)),
952
+ simde_x_vloadq_f32(SIMDE_FLOAT32_C( 673.61), SIMDE_FLOAT32_C( 902.28),
953
+ SIMDE_FLOAT32_C( 17.82), SIMDE_FLOAT32_C( 262.95)),
954
+ simde_x_vloadq_f32(SIMDE_FLOAT32_C(-609010.75), SIMDE_FLOAT32_C(-298925.38),
955
+ SIMDE_FLOAT32_C(14222.14), SIMDE_FLOAT32_C(-188732.38)) },
956
+ { simde_x_vloadq_f32(SIMDE_FLOAT32_C( -782.37), SIMDE_FLOAT32_C( 839.26),
957
+ SIMDE_FLOAT32_C( 988.92), SIMDE_FLOAT32_C( 194.85)),
958
+ simde_x_vloadq_f32(SIMDE_FLOAT32_C( -438.71), SIMDE_FLOAT32_C( 397.59),
959
+ SIMDE_FLOAT32_C( -906.02), SIMDE_FLOAT32_C( -645.00)),
960
+ simde_x_vloadq_f32(SIMDE_FLOAT32_C(343233.53), SIMDE_FLOAT32_C(333681.38),
961
+ SIMDE_FLOAT32_C(-895981.31), SIMDE_FLOAT32_C(-125678.26)) },
962
+ { simde_x_vloadq_f32(SIMDE_FLOAT32_C( 514.53), SIMDE_FLOAT32_C( 516.76),
963
+ SIMDE_FLOAT32_C( 141.10), SIMDE_FLOAT32_C( 80.68)),
964
+ simde_x_vloadq_f32(SIMDE_FLOAT32_C( -996.77), SIMDE_FLOAT32_C( 465.18),
965
+ SIMDE_FLOAT32_C( 267.79), SIMDE_FLOAT32_C( -96.82)),
966
+ simde_x_vloadq_f32(SIMDE_FLOAT32_C(-512868.09), SIMDE_FLOAT32_C(240386.42),
967
+ SIMDE_FLOAT32_C(37785.17), SIMDE_FLOAT32_C(-7811.44)) },
968
+ { simde_x_vloadq_f32(SIMDE_FLOAT32_C( 17.49), SIMDE_FLOAT32_C( -972.94),
969
+ SIMDE_FLOAT32_C( -9.96), SIMDE_FLOAT32_C( 557.04)),
970
+ simde_x_vloadq_f32(SIMDE_FLOAT32_C( 658.33), SIMDE_FLOAT32_C( 152.31),
971
+ SIMDE_FLOAT32_C( -894.23), SIMDE_FLOAT32_C( -290.74)),
972
+ simde_x_vloadq_f32(SIMDE_FLOAT32_C(11514.19), SIMDE_FLOAT32_C(-148188.48),
973
+ SIMDE_FLOAT32_C( 8906.53), SIMDE_FLOAT32_C(-161953.80)) },
974
+ { simde_x_vloadq_f32(SIMDE_FLOAT32_C( -221.71), SIMDE_FLOAT32_C( 835.86),
975
+ SIMDE_FLOAT32_C( 400.30), SIMDE_FLOAT32_C( -158.20)),
976
+ simde_x_vloadq_f32(SIMDE_FLOAT32_C( 216.50), SIMDE_FLOAT32_C( 439.51),
977
+ SIMDE_FLOAT32_C( -451.11), SIMDE_FLOAT32_C( 42.20)),
978
+ simde_x_vloadq_f32(SIMDE_FLOAT32_C(-48000.21), SIMDE_FLOAT32_C(367368.84),
979
+ SIMDE_FLOAT32_C(-180579.33), SIMDE_FLOAT32_C(-6676.04)) },
980
+ { simde_x_vloadq_f32(SIMDE_FLOAT32_C( -279.50), SIMDE_FLOAT32_C( -471.25),
981
+ SIMDE_FLOAT32_C( 976.16), SIMDE_FLOAT32_C( -178.14)),
982
+ simde_x_vloadq_f32(SIMDE_FLOAT32_C( -66.41), SIMDE_FLOAT32_C( -721.09),
983
+ SIMDE_FLOAT32_C( 642.38), SIMDE_FLOAT32_C( 930.30)),
984
+ simde_x_vloadq_f32(SIMDE_FLOAT32_C(18561.60), SIMDE_FLOAT32_C(339813.69),
985
+ SIMDE_FLOAT32_C(627065.62), SIMDE_FLOAT32_C(-165723.64)) },
986
+ { simde_x_vloadq_f32(SIMDE_FLOAT32_C( -593.59), SIMDE_FLOAT32_C( 407.91),
987
+ SIMDE_FLOAT32_C( -402.54), SIMDE_FLOAT32_C( 522.63)),
988
+ simde_x_vloadq_f32(SIMDE_FLOAT32_C( -180.83), SIMDE_FLOAT32_C( -430.34),
989
+ SIMDE_FLOAT32_C( -204.65), SIMDE_FLOAT32_C( -888.67)),
990
+ simde_x_vloadq_f32(SIMDE_FLOAT32_C(107338.88), SIMDE_FLOAT32_C(-175539.98),
991
+ SIMDE_FLOAT32_C(82379.81), SIMDE_FLOAT32_C(-464445.59)) }
992
+ };
993
+
994
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
995
+ simde_float32x4_t r = simde_vmulq_f32(test_vec[i].a, test_vec[i].b);
996
+ simde_neon_assert_float32x4_equal(r, test_vec[i].r, 1);
997
+ }
998
+
999
+ return MUNIT_OK;
1000
+ }
1001
+
1002
+ static MunitResult
1003
+ test_simde_vmulq_f64(const MunitParameter params[], void* data) {
1004
+ (void) params;
1005
+ (void) data;
1006
+
1007
+ const struct {
1008
+ simde_float64x2_t a;
1009
+ simde_float64x2_t b;
1010
+ simde_float64x2_t r;
1011
+ } test_vec[8] = {
1012
+ { simde_x_vloadq_f64(SIMDE_FLOAT64_C( 261.94), SIMDE_FLOAT64_C( 389.36)),
1013
+ simde_x_vloadq_f64(SIMDE_FLOAT64_C( -604.87), SIMDE_FLOAT64_C( -117.53)),
1014
+ simde_x_vloadq_f64(SIMDE_FLOAT64_C(-158439.65), SIMDE_FLOAT64_C(-45761.48)) },
1015
+ { simde_x_vloadq_f64(SIMDE_FLOAT64_C( 516.97), SIMDE_FLOAT64_C( 458.14)),
1016
+ simde_x_vloadq_f64(SIMDE_FLOAT64_C( 808.10), SIMDE_FLOAT64_C( -730.44)),
1017
+ simde_x_vloadq_f64(SIMDE_FLOAT64_C(417763.46), SIMDE_FLOAT64_C(-334643.78)) },
1018
+ { simde_x_vloadq_f64(SIMDE_FLOAT64_C( 956.22), SIMDE_FLOAT64_C( 84.60)),
1019
+ simde_x_vloadq_f64(SIMDE_FLOAT64_C( 683.61), SIMDE_FLOAT64_C( -964.54)),
1020
+ simde_x_vloadq_f64(SIMDE_FLOAT64_C(653681.55), SIMDE_FLOAT64_C(-81600.08)) },
1021
+ { simde_x_vloadq_f64(SIMDE_FLOAT64_C( -723.12), SIMDE_FLOAT64_C( 136.75)),
1022
+ simde_x_vloadq_f64(SIMDE_FLOAT64_C( -859.96), SIMDE_FLOAT64_C( 319.65)),
1023
+ simde_x_vloadq_f64(SIMDE_FLOAT64_C(621854.28), SIMDE_FLOAT64_C(43712.14)) },
1024
+ { simde_x_vloadq_f64(SIMDE_FLOAT64_C( -937.43), SIMDE_FLOAT64_C( -648.19)),
1025
+ simde_x_vloadq_f64(SIMDE_FLOAT64_C( 344.36), SIMDE_FLOAT64_C( -506.13)),
1026
+ simde_x_vloadq_f64(SIMDE_FLOAT64_C(-322813.39), SIMDE_FLOAT64_C(328068.40)) },
1027
+ { simde_x_vloadq_f64(SIMDE_FLOAT64_C( -967.06), SIMDE_FLOAT64_C( -492.64)),
1028
+ simde_x_vloadq_f64(SIMDE_FLOAT64_C( -564.55), SIMDE_FLOAT64_C( 904.04)),
1029
+ simde_x_vloadq_f64(SIMDE_FLOAT64_C(545953.72), SIMDE_FLOAT64_C(-445366.27)) },
1030
+ { simde_x_vloadq_f64(SIMDE_FLOAT64_C( 86.34), SIMDE_FLOAT64_C( 355.85)),
1031
+ simde_x_vloadq_f64(SIMDE_FLOAT64_C( -892.48), SIMDE_FLOAT64_C( -794.20)),
1032
+ simde_x_vloadq_f64(SIMDE_FLOAT64_C(-77056.72), SIMDE_FLOAT64_C(-282616.07)) },
1033
+ { simde_x_vloadq_f64(SIMDE_FLOAT64_C( 554.46), SIMDE_FLOAT64_C( 210.82)),
1034
+ simde_x_vloadq_f64(SIMDE_FLOAT64_C( 71.90), SIMDE_FLOAT64_C( -553.26)),
1035
+ simde_x_vloadq_f64(SIMDE_FLOAT64_C(39865.67), SIMDE_FLOAT64_C(-116638.27)) }
1036
+ };
1037
+
1038
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
1039
+ simde_float64x2_t r = simde_vmulq_f64(test_vec[i].a, test_vec[i].b);
1040
+ simde_neon_assert_float64x2_equal(r, test_vec[i].r, 1);
1041
+ }
1042
+
1043
+ return MUNIT_OK;
1044
+ }
1045
+
1046
+ #endif /* defined(SIMDE_NEON_NATIVE) || defined(SIMDE_NO_NATIVE) || defined(SIMDE_ALWAYS_BUILD_NATIVE_TESTS) */
1047
+
1048
+ HEDLEY_DIAGNOSTIC_PUSH
1049
+ HEDLEY_DIAGNOSTIC_DISABLE_CAST_QUAL
1050
+
1051
+ static MunitTest test_suite_tests[] = {
1052
+ #if defined(SIMDE_NEON_NATIVE) || defined(SIMDE_NO_NATIVE) || defined(SIMDE_ALWAYS_BUILD_NATIVE_TESTS)
1053
+ SIMDE_TESTS_NEON_DEFINE_TEST(s8),
1054
+ SIMDE_TESTS_NEON_DEFINE_TEST(s16),
1055
+ SIMDE_TESTS_NEON_DEFINE_TEST(s32),
1056
+ // SIMDE_TESTS_NEON_DEFINE_TEST(s64),
1057
+ SIMDE_TESTS_NEON_DEFINE_TEST(u8),
1058
+ SIMDE_TESTS_NEON_DEFINE_TEST(u16),
1059
+ SIMDE_TESTS_NEON_DEFINE_TEST(u32),
1060
+ // SIMDE_TESTS_NEON_DEFINE_TEST(u64),
1061
+ SIMDE_TESTS_NEON_DEFINE_TEST(f32),
1062
+ SIMDE_TESTS_NEON_DEFINE_TEST(f64),
1063
+ SIMDE_TESTS_NEON_DEFINE_TEST_FULL(q, s8),
1064
+ SIMDE_TESTS_NEON_DEFINE_TEST_FULL(q, s16),
1065
+ SIMDE_TESTS_NEON_DEFINE_TEST_FULL(q, s32),
1066
+ // SIMDE_TESTS_NEON_DEFINE_TEST_FULL(q, s64),
1067
+ SIMDE_TESTS_NEON_DEFINE_TEST_FULL(q, u8),
1068
+ SIMDE_TESTS_NEON_DEFINE_TEST_FULL(q, u16),
1069
+ SIMDE_TESTS_NEON_DEFINE_TEST_FULL(q, u32),
1070
+ // SIMDE_TESTS_NEON_DEFINE_TEST_FULL(q, u64),
1071
+ SIMDE_TESTS_NEON_DEFINE_TEST_FULL(q, f32),
1072
+ SIMDE_TESTS_NEON_DEFINE_TEST_FULL(q, f64),
1073
+ #endif /* defined(SIMDE_NEON_NATIVE) || defined(SIMDE_NO_NATIVE) || defined(SIMDE_ALWAYS_BUILD_NATIVE_TESTS) */
1074
+
1075
+ { NULL, NULL, NULL, NULL, MUNIT_TEST_OPTION_NONE, NULL }
1076
+ };
1077
+
1078
+ HEDLEY_C_DECL MunitSuite* SIMDE_TESTS_GENERATE_SYMBOL(SIMDE_TESTS_CURRENT_NEON_OP)(void) {
1079
+ static MunitSuite suite = { (char*) "/v" HEDLEY_STRINGIFY(SIMDE_TESTS_CURRENT_NEON_OP), test_suite_tests, NULL, 1, MUNIT_SUITE_OPTION_NONE };
1080
+
1081
+ return &suite;
1082
+ }
1083
+
1084
+ HEDLEY_DIAGNOSTIC_POP