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 +1,2557 @@
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_ISAX fma
25
+ #include <simde/x86/fma.h>
26
+ #include <test/x86/test-avx.h>
27
+
28
+ #if defined(SIMDE_FMA_NATIVE) || defined(SIMDE_NO_NATIVE) || defined(SIMDE_ALWAYS_BUILD_NATIVE_TESTS)
29
+
30
+ static MunitResult
31
+ test_simde_mm_fmadd_pd(const MunitParameter params[], void* data) {
32
+ (void) params;
33
+ (void) data;
34
+
35
+ const struct {
36
+ simde__m128d a;
37
+ simde__m128d b;
38
+ simde__m128d c;
39
+ simde__m128d r;
40
+ } test_vec[8] = {
41
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( -364.57), SIMDE_FLOAT64_C( -702.81)),
42
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -890.80), SIMDE_FLOAT64_C( -433.89)),
43
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 486.26), SIMDE_FLOAT64_C( -304.02)),
44
+ simde_mm_set_pd(SIMDE_FLOAT64_C(325245.22), SIMDE_FLOAT64_C(304638.21)) },
45
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( 278.35), SIMDE_FLOAT64_C( -601.69)),
46
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -850.98), SIMDE_FLOAT64_C( 10.48)),
47
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -876.47), SIMDE_FLOAT64_C( -253.46)),
48
+ simde_mm_set_pd(SIMDE_FLOAT64_C(-237746.75), SIMDE_FLOAT64_C(-6559.17)) },
49
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( -495.32), SIMDE_FLOAT64_C( 626.54)),
50
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 243.15), SIMDE_FLOAT64_C( -595.67)),
51
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 652.27), SIMDE_FLOAT64_C( 684.47)),
52
+ simde_mm_set_pd(SIMDE_FLOAT64_C(-119784.79), SIMDE_FLOAT64_C(-372526.61)) },
53
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( -675.35), SIMDE_FLOAT64_C( -855.85)),
54
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 863.96), SIMDE_FLOAT64_C( -244.88)),
55
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 266.15), SIMDE_FLOAT64_C( -217.90)),
56
+ simde_mm_set_pd(SIMDE_FLOAT64_C(-583209.24), SIMDE_FLOAT64_C(209362.65)) },
57
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( -945.02), SIMDE_FLOAT64_C( -266.12)),
58
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 248.34), SIMDE_FLOAT64_C( -754.68)),
59
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 686.19), SIMDE_FLOAT64_C( 201.29)),
60
+ simde_mm_set_pd(SIMDE_FLOAT64_C(-234000.08), SIMDE_FLOAT64_C(201036.73)) },
61
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( -923.40), SIMDE_FLOAT64_C( 347.92)),
62
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -299.09), SIMDE_FLOAT64_C( -322.35)),
63
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -172.16), SIMDE_FLOAT64_C( 792.83)),
64
+ simde_mm_set_pd(SIMDE_FLOAT64_C(276007.55), SIMDE_FLOAT64_C(-111359.18)) },
65
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( -997.96), SIMDE_FLOAT64_C( -774.36)),
66
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 336.57), SIMDE_FLOAT64_C( -666.28)),
67
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 18.66), SIMDE_FLOAT64_C( 857.72)),
68
+ simde_mm_set_pd(SIMDE_FLOAT64_C(-335864.74), SIMDE_FLOAT64_C(516798.30)) },
69
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( 425.15), SIMDE_FLOAT64_C( -554.19)),
70
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -602.50), SIMDE_FLOAT64_C( -329.67)),
71
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -208.43), SIMDE_FLOAT64_C( 819.37)),
72
+ simde_mm_set_pd(SIMDE_FLOAT64_C(-256361.30), SIMDE_FLOAT64_C(183519.19)) }
73
+ };
74
+
75
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
76
+ simde__m128d r = simde_mm_fmadd_pd(test_vec[i].a, test_vec[i].b, test_vec[i].c);
77
+ simde_assert_m128d_close(r, test_vec[i].r, 1);
78
+ }
79
+
80
+ return MUNIT_OK;
81
+ }
82
+
83
+ static MunitResult
84
+ test_simde_mm256_fmadd_pd(const MunitParameter params[], void* data) {
85
+ (void) params;
86
+ (void) data;
87
+
88
+ const struct {
89
+ simde__m256d a;
90
+ simde__m256d b;
91
+ simde__m256d c;
92
+ simde__m256d r;
93
+ } test_vec[8] = {
94
+ { simde_mm256_set_pd(SIMDE_FLOAT64_C( 463.71), SIMDE_FLOAT64_C( -551.83),
95
+ SIMDE_FLOAT64_C( 568.05), SIMDE_FLOAT64_C( -826.17)),
96
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( 440.29), SIMDE_FLOAT64_C( 762.39),
97
+ SIMDE_FLOAT64_C( -806.23), SIMDE_FLOAT64_C( -848.48)),
98
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( 222.93), SIMDE_FLOAT64_C( -604.06),
99
+ SIMDE_FLOAT64_C( -844.49), SIMDE_FLOAT64_C( 221.50)),
100
+ simde_mm256_set_pd(SIMDE_FLOAT64_C(204389.81), SIMDE_FLOAT64_C(-421313.73),
101
+ SIMDE_FLOAT64_C(-458823.44), SIMDE_FLOAT64_C(701210.22)) },
102
+ { simde_mm256_set_pd(SIMDE_FLOAT64_C( 418.00), SIMDE_FLOAT64_C( -725.82),
103
+ SIMDE_FLOAT64_C( -54.90), SIMDE_FLOAT64_C( -342.22)),
104
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( 280.70), SIMDE_FLOAT64_C( 983.58),
105
+ SIMDE_FLOAT64_C( -289.88), SIMDE_FLOAT64_C( 305.30)),
106
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( -404.92), SIMDE_FLOAT64_C( -664.17),
107
+ SIMDE_FLOAT64_C( 164.15), SIMDE_FLOAT64_C( -785.83)),
108
+ simde_mm256_set_pd(SIMDE_FLOAT64_C(116927.68), SIMDE_FLOAT64_C(-714566.21),
109
+ SIMDE_FLOAT64_C( 16078.56), SIMDE_FLOAT64_C(-105265.60)) },
110
+ { simde_mm256_set_pd(SIMDE_FLOAT64_C( 267.84), SIMDE_FLOAT64_C( 153.22),
111
+ SIMDE_FLOAT64_C( 565.53), SIMDE_FLOAT64_C( 45.62)),
112
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( 253.29), SIMDE_FLOAT64_C( -448.85),
113
+ SIMDE_FLOAT64_C( -379.10), SIMDE_FLOAT64_C( 896.99)),
114
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( 305.51), SIMDE_FLOAT64_C( -18.42),
115
+ SIMDE_FLOAT64_C( 560.02), SIMDE_FLOAT64_C( -441.54)),
116
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( 68146.70), SIMDE_FLOAT64_C(-68791.22),
117
+ SIMDE_FLOAT64_C(-213832.40), SIMDE_FLOAT64_C( 40479.14)) },
118
+ { simde_mm256_set_pd(SIMDE_FLOAT64_C( 698.02), SIMDE_FLOAT64_C( -282.65),
119
+ SIMDE_FLOAT64_C( -531.77), SIMDE_FLOAT64_C( -673.05)),
120
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( -227.36), SIMDE_FLOAT64_C( 165.86),
121
+ SIMDE_FLOAT64_C( -853.86), SIMDE_FLOAT64_C( 210.39)),
122
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( -443.02), SIMDE_FLOAT64_C( -362.32),
123
+ SIMDE_FLOAT64_C( 833.55), SIMDE_FLOAT64_C( 692.62)),
124
+ simde_mm256_set_pd(SIMDE_FLOAT64_C(-159144.85), SIMDE_FLOAT64_C(-47242.65),
125
+ SIMDE_FLOAT64_C(454890.68), SIMDE_FLOAT64_C(-140910.37)) },
126
+ { simde_mm256_set_pd(SIMDE_FLOAT64_C( -601.68), SIMDE_FLOAT64_C( 654.88),
127
+ SIMDE_FLOAT64_C( 957.42), SIMDE_FLOAT64_C( 563.37)),
128
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( -958.65), SIMDE_FLOAT64_C( 523.00),
129
+ SIMDE_FLOAT64_C( -211.18), SIMDE_FLOAT64_C( -889.28)),
130
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( 180.36), SIMDE_FLOAT64_C( 481.63),
131
+ SIMDE_FLOAT64_C( -222.77), SIMDE_FLOAT64_C( -51.21)),
132
+ simde_mm256_set_pd(SIMDE_FLOAT64_C(576980.89), SIMDE_FLOAT64_C(342983.87),
133
+ SIMDE_FLOAT64_C(-202410.73), SIMDE_FLOAT64_C(-501044.88)) },
134
+ { simde_mm256_set_pd(SIMDE_FLOAT64_C( 731.42), SIMDE_FLOAT64_C( -631.15),
135
+ SIMDE_FLOAT64_C( -982.89), SIMDE_FLOAT64_C( -397.65)),
136
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( 69.37), SIMDE_FLOAT64_C( -394.43),
137
+ SIMDE_FLOAT64_C( -18.09), SIMDE_FLOAT64_C( 272.24)),
138
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( 920.14), SIMDE_FLOAT64_C( -196.58),
139
+ SIMDE_FLOAT64_C( 324.68), SIMDE_FLOAT64_C( -193.62)),
140
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( 51658.75), SIMDE_FLOAT64_C(248747.91),
141
+ SIMDE_FLOAT64_C( 18105.16), SIMDE_FLOAT64_C(-108449.86)) },
142
+ { simde_mm256_set_pd(SIMDE_FLOAT64_C( -561.20), SIMDE_FLOAT64_C( -459.54),
143
+ SIMDE_FLOAT64_C( -681.08), SIMDE_FLOAT64_C( -49.66)),
144
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( 197.69), SIMDE_FLOAT64_C( -813.71),
145
+ SIMDE_FLOAT64_C( -990.48), SIMDE_FLOAT64_C( -180.87)),
146
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( -453.27), SIMDE_FLOAT64_C( -557.45),
147
+ SIMDE_FLOAT64_C( -780.15), SIMDE_FLOAT64_C( 693.73)),
148
+ simde_mm256_set_pd(SIMDE_FLOAT64_C(-111396.90), SIMDE_FLOAT64_C(373374.84),
149
+ SIMDE_FLOAT64_C(673815.97), SIMDE_FLOAT64_C( 9675.73)) },
150
+ { simde_mm256_set_pd(SIMDE_FLOAT64_C( 297.04), SIMDE_FLOAT64_C( 950.40),
151
+ SIMDE_FLOAT64_C( -454.41), SIMDE_FLOAT64_C( 419.22)),
152
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( 303.25), SIMDE_FLOAT64_C( -917.33),
153
+ SIMDE_FLOAT64_C( 128.78), SIMDE_FLOAT64_C( 208.96)),
154
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( 735.53), SIMDE_FLOAT64_C( 976.90),
155
+ SIMDE_FLOAT64_C( 803.26), SIMDE_FLOAT64_C( 610.54)),
156
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( 90812.91), SIMDE_FLOAT64_C(-870853.53),
157
+ SIMDE_FLOAT64_C(-57715.66), SIMDE_FLOAT64_C( 88210.75)) }
158
+ };
159
+
160
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
161
+ simde__m256d r = simde_mm256_fmadd_pd(test_vec[i].a, test_vec[i].b, test_vec[i].c);
162
+ simde_assert_m256d_close(r, test_vec[i].r, 1);
163
+ }
164
+
165
+ return MUNIT_OK;
166
+ }
167
+
168
+ static MunitResult
169
+ test_simde_mm_fmadd_ps(const MunitParameter params[], void* data) {
170
+ (void) params;
171
+ (void) data;
172
+
173
+ const struct {
174
+ simde__m128 a;
175
+ simde__m128 b;
176
+ simde__m128 c;
177
+ simde__m128 r;
178
+ } test_vec[8] = {
179
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( 68.47), SIMDE_FLOAT32_C( -20.99), SIMDE_FLOAT32_C( -768.39), SIMDE_FLOAT32_C( 464.52)),
180
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 297.92), SIMDE_FLOAT32_C( 902.90), SIMDE_FLOAT32_C( 496.10), SIMDE_FLOAT32_C( -932.73)),
181
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -341.38), SIMDE_FLOAT32_C( -852.40), SIMDE_FLOAT32_C( 426.68), SIMDE_FLOAT32_C( 755.10)),
182
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 20057.20), SIMDE_FLOAT32_C(-19804.27), SIMDE_FLOAT32_C(-380771.59), SIMDE_FLOAT32_C(-432516.62)) },
183
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( 969.18), SIMDE_FLOAT32_C( 318.32), SIMDE_FLOAT32_C( -273.65), SIMDE_FLOAT32_C( 39.39)),
184
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 995.10), SIMDE_FLOAT32_C( 620.67), SIMDE_FLOAT32_C( 664.82), SIMDE_FLOAT32_C( 711.85)),
185
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 473.88), SIMDE_FLOAT32_C( 360.15), SIMDE_FLOAT32_C( -250.82), SIMDE_FLOAT32_C( -88.76)),
186
+ simde_mm_set_ps(SIMDE_FLOAT32_C(964904.88), SIMDE_FLOAT32_C(197931.83), SIMDE_FLOAT32_C(-182178.80), SIMDE_FLOAT32_C( 27951.01)) },
187
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( 164.14), SIMDE_FLOAT32_C( -848.02), SIMDE_FLOAT32_C( 235.35), SIMDE_FLOAT32_C( -999.97)),
188
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 466.54), SIMDE_FLOAT32_C( 41.59), SIMDE_FLOAT32_C( -619.09), SIMDE_FLOAT32_C( 332.19)),
189
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -720.05), SIMDE_FLOAT32_C( 91.37), SIMDE_FLOAT32_C( 3.41), SIMDE_FLOAT32_C( -151.75)),
190
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 75857.83), SIMDE_FLOAT32_C(-35177.78), SIMDE_FLOAT32_C(-145699.44), SIMDE_FLOAT32_C(-332331.78)) },
191
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( -664.98), SIMDE_FLOAT32_C( -765.11), SIMDE_FLOAT32_C( -950.95), SIMDE_FLOAT32_C( 967.68)),
192
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 463.42), SIMDE_FLOAT32_C( 310.01), SIMDE_FLOAT32_C( -859.78), SIMDE_FLOAT32_C( -247.59)),
193
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 372.25), SIMDE_FLOAT32_C( -546.43), SIMDE_FLOAT32_C( -18.65), SIMDE_FLOAT32_C( -608.78)),
194
+ simde_mm_set_ps(SIMDE_FLOAT32_C(-307792.78), SIMDE_FLOAT32_C(-237738.19), SIMDE_FLOAT32_C(817589.19), SIMDE_FLOAT32_C(-240196.67)) },
195
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( 739.81), SIMDE_FLOAT32_C( -275.42), SIMDE_FLOAT32_C( -462.27), SIMDE_FLOAT32_C( -299.55)),
196
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -308.92), SIMDE_FLOAT32_C( 948.18), SIMDE_FLOAT32_C( -344.73), SIMDE_FLOAT32_C( -942.49)),
197
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 780.24), SIMDE_FLOAT32_C( 819.52), SIMDE_FLOAT32_C( -913.65), SIMDE_FLOAT32_C( 715.95)),
198
+ simde_mm_set_ps(SIMDE_FLOAT32_C(-227761.88), SIMDE_FLOAT32_C(-260328.23), SIMDE_FLOAT32_C(158444.69), SIMDE_FLOAT32_C(283038.81)) },
199
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( -766.06), SIMDE_FLOAT32_C( -563.42), SIMDE_FLOAT32_C( -122.27), SIMDE_FLOAT32_C( -338.00)),
200
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 496.12), SIMDE_FLOAT32_C( -751.97), SIMDE_FLOAT32_C( 655.86), SIMDE_FLOAT32_C( 174.40)),
201
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -985.27), SIMDE_FLOAT32_C( 574.75), SIMDE_FLOAT32_C( 212.10), SIMDE_FLOAT32_C( -683.32)),
202
+ simde_mm_set_ps(SIMDE_FLOAT32_C(-381042.97), SIMDE_FLOAT32_C(424249.66), SIMDE_FLOAT32_C(-79979.90), SIMDE_FLOAT32_C(-59630.52)) },
203
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( -663.22), SIMDE_FLOAT32_C( 549.14), SIMDE_FLOAT32_C( 733.90), SIMDE_FLOAT32_C( 785.76)),
204
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -149.60), SIMDE_FLOAT32_C( -221.89), SIMDE_FLOAT32_C( -452.29), SIMDE_FLOAT32_C( -18.14)),
205
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 979.50), SIMDE_FLOAT32_C( -484.31), SIMDE_FLOAT32_C( -965.78), SIMDE_FLOAT32_C( -291.28)),
206
+ simde_mm_set_ps(SIMDE_FLOAT32_C(100197.21), SIMDE_FLOAT32_C(-122332.99), SIMDE_FLOAT32_C(-332901.44), SIMDE_FLOAT32_C(-14544.97)) },
207
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( 82.89), SIMDE_FLOAT32_C( -639.53), SIMDE_FLOAT32_C( 680.97), SIMDE_FLOAT32_C( -745.76)),
208
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 377.68), SIMDE_FLOAT32_C( -229.15), SIMDE_FLOAT32_C( 986.42), SIMDE_FLOAT32_C( -430.87)),
209
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 506.24), SIMDE_FLOAT32_C( -791.48), SIMDE_FLOAT32_C( -896.55), SIMDE_FLOAT32_C( -775.82)),
210
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 31812.13), SIMDE_FLOAT32_C(145756.81), SIMDE_FLOAT32_C(670825.81), SIMDE_FLOAT32_C(320549.81)) }
211
+ };
212
+
213
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
214
+ simde__m128 r = simde_mm_fmadd_ps(test_vec[i].a, test_vec[i].b, test_vec[i].c);
215
+ simde_assert_m128_close(r, test_vec[i].r, 1);
216
+ }
217
+
218
+ return MUNIT_OK;
219
+ }
220
+
221
+ static MunitResult
222
+ test_simde_mm256_fmadd_ps(const MunitParameter params[], void* data) {
223
+ (void) params;
224
+ (void) data;
225
+
226
+ const struct {
227
+ simde__m256 a;
228
+ simde__m256 b;
229
+ simde__m256 c;
230
+ simde__m256 r;
231
+ } test_vec[8] = {
232
+ { simde_mm256_set_ps(SIMDE_FLOAT32_C( 39.90), SIMDE_FLOAT32_C( 46.80),
233
+ SIMDE_FLOAT32_C( -90.30), SIMDE_FLOAT32_C( -57.20),
234
+ SIMDE_FLOAT32_C( 71.50), SIMDE_FLOAT32_C( 75.00),
235
+ SIMDE_FLOAT32_C( -0.30), SIMDE_FLOAT32_C( 14.60)),
236
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( -90.60), SIMDE_FLOAT32_C( -15.30),
237
+ SIMDE_FLOAT32_C( -46.70), SIMDE_FLOAT32_C( 73.50),
238
+ SIMDE_FLOAT32_C( -27.40), SIMDE_FLOAT32_C( -79.00),
239
+ SIMDE_FLOAT32_C( -14.10), SIMDE_FLOAT32_C( 22.30)),
240
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( -19.50), SIMDE_FLOAT32_C( 61.50),
241
+ SIMDE_FLOAT32_C( -38.80), SIMDE_FLOAT32_C( -19.20),
242
+ SIMDE_FLOAT32_C( 54.40), SIMDE_FLOAT32_C( -71.00),
243
+ SIMDE_FLOAT32_C( -11.30), SIMDE_FLOAT32_C( -2.70)),
244
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( -3634.44), SIMDE_FLOAT32_C( -654.54),
245
+ SIMDE_FLOAT32_C( 4178.21), SIMDE_FLOAT32_C( -4223.40),
246
+ SIMDE_FLOAT32_C( -1904.70), SIMDE_FLOAT32_C( -5996.00),
247
+ SIMDE_FLOAT32_C( -7.07), SIMDE_FLOAT32_C( 322.88)) },
248
+ { simde_mm256_set_ps(SIMDE_FLOAT32_C( 56.00), SIMDE_FLOAT32_C( -61.10),
249
+ SIMDE_FLOAT32_C( -84.20), SIMDE_FLOAT32_C( -8.30),
250
+ SIMDE_FLOAT32_C( 96.60), SIMDE_FLOAT32_C( 92.70),
251
+ SIMDE_FLOAT32_C( -19.40), SIMDE_FLOAT32_C( -41.30)),
252
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( -20.80), SIMDE_FLOAT32_C( -77.90),
253
+ SIMDE_FLOAT32_C( 22.80), SIMDE_FLOAT32_C( -62.40),
254
+ SIMDE_FLOAT32_C( 47.20), SIMDE_FLOAT32_C( 23.30),
255
+ SIMDE_FLOAT32_C( -14.70), SIMDE_FLOAT32_C( 1.80)),
256
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( -2.50), SIMDE_FLOAT32_C( -40.20),
257
+ SIMDE_FLOAT32_C( -64.40), SIMDE_FLOAT32_C( 46.00),
258
+ SIMDE_FLOAT32_C( 19.60), SIMDE_FLOAT32_C( 30.00),
259
+ SIMDE_FLOAT32_C( 23.60), SIMDE_FLOAT32_C( 20.60)),
260
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( -1167.30), SIMDE_FLOAT32_C( 4719.49),
261
+ SIMDE_FLOAT32_C( -1984.16), SIMDE_FLOAT32_C( 563.92),
262
+ SIMDE_FLOAT32_C( 4579.12), SIMDE_FLOAT32_C( 2189.91),
263
+ SIMDE_FLOAT32_C( 308.78), SIMDE_FLOAT32_C( -53.74)) },
264
+ { simde_mm256_set_ps(SIMDE_FLOAT32_C( -73.60), SIMDE_FLOAT32_C( -63.70),
265
+ SIMDE_FLOAT32_C( -7.10), SIMDE_FLOAT32_C( -70.90),
266
+ SIMDE_FLOAT32_C( 23.30), SIMDE_FLOAT32_C( 22.20),
267
+ SIMDE_FLOAT32_C( 4.90), SIMDE_FLOAT32_C( -85.30)),
268
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( 75.60), SIMDE_FLOAT32_C( -6.90),
269
+ SIMDE_FLOAT32_C( 73.70), SIMDE_FLOAT32_C( -85.70),
270
+ SIMDE_FLOAT32_C( -25.90), SIMDE_FLOAT32_C( -59.90),
271
+ SIMDE_FLOAT32_C( -56.20), SIMDE_FLOAT32_C( -30.70)),
272
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( 54.00), SIMDE_FLOAT32_C( -79.70),
273
+ SIMDE_FLOAT32_C( 71.20), SIMDE_FLOAT32_C( -74.20),
274
+ SIMDE_FLOAT32_C( -48.90), SIMDE_FLOAT32_C( -7.20),
275
+ SIMDE_FLOAT32_C( -59.10), SIMDE_FLOAT32_C( -84.70)),
276
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( -5510.16), SIMDE_FLOAT32_C( 359.83),
277
+ SIMDE_FLOAT32_C( -452.07), SIMDE_FLOAT32_C( 6001.93),
278
+ SIMDE_FLOAT32_C( -652.37), SIMDE_FLOAT32_C( -1336.98),
279
+ SIMDE_FLOAT32_C( -334.48), SIMDE_FLOAT32_C( 2534.01)) },
280
+ { simde_mm256_set_ps(SIMDE_FLOAT32_C( 57.50), SIMDE_FLOAT32_C( 93.40),
281
+ SIMDE_FLOAT32_C( -2.20), SIMDE_FLOAT32_C( 77.20),
282
+ SIMDE_FLOAT32_C( 79.40), SIMDE_FLOAT32_C( -81.10),
283
+ SIMDE_FLOAT32_C( 25.80), SIMDE_FLOAT32_C( -5.40)),
284
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( -36.80), SIMDE_FLOAT32_C( -46.10),
285
+ SIMDE_FLOAT32_C( 57.50), SIMDE_FLOAT32_C( 47.70),
286
+ SIMDE_FLOAT32_C( 38.00), SIMDE_FLOAT32_C( 48.30),
287
+ SIMDE_FLOAT32_C( 86.60), SIMDE_FLOAT32_C( 85.60)),
288
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( 92.60), SIMDE_FLOAT32_C( 68.60),
289
+ SIMDE_FLOAT32_C( -48.10), SIMDE_FLOAT32_C( -53.80),
290
+ SIMDE_FLOAT32_C( -45.80), SIMDE_FLOAT32_C( 33.60),
291
+ SIMDE_FLOAT32_C( 47.80), SIMDE_FLOAT32_C( 61.30)),
292
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( -2023.40), SIMDE_FLOAT32_C( -4237.14),
293
+ SIMDE_FLOAT32_C( -174.60), SIMDE_FLOAT32_C( 3628.64),
294
+ SIMDE_FLOAT32_C( 2971.40), SIMDE_FLOAT32_C( -3883.53),
295
+ SIMDE_FLOAT32_C( 2282.08), SIMDE_FLOAT32_C( -400.94)) },
296
+ { simde_mm256_set_ps(SIMDE_FLOAT32_C( 39.30), SIMDE_FLOAT32_C( 47.70),
297
+ SIMDE_FLOAT32_C( -46.40), SIMDE_FLOAT32_C( 22.40),
298
+ SIMDE_FLOAT32_C( -47.70), SIMDE_FLOAT32_C( -87.50),
299
+ SIMDE_FLOAT32_C( 56.70), SIMDE_FLOAT32_C( -98.30)),
300
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( 47.80), SIMDE_FLOAT32_C( 25.10),
301
+ SIMDE_FLOAT32_C( 86.40), SIMDE_FLOAT32_C( 20.80),
302
+ SIMDE_FLOAT32_C( -68.30), SIMDE_FLOAT32_C( -7.70),
303
+ SIMDE_FLOAT32_C( 87.10), SIMDE_FLOAT32_C( 24.00)),
304
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( 30.50), SIMDE_FLOAT32_C( 80.40),
305
+ SIMDE_FLOAT32_C( -81.20), SIMDE_FLOAT32_C( -60.10),
306
+ SIMDE_FLOAT32_C( -62.20), SIMDE_FLOAT32_C( 51.30),
307
+ SIMDE_FLOAT32_C( -56.00), SIMDE_FLOAT32_C( -52.90)),
308
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( 1909.04), SIMDE_FLOAT32_C( 1277.67),
309
+ SIMDE_FLOAT32_C( -4090.16), SIMDE_FLOAT32_C( 405.82),
310
+ SIMDE_FLOAT32_C( 3195.71), SIMDE_FLOAT32_C( 725.05),
311
+ SIMDE_FLOAT32_C( 4882.57), SIMDE_FLOAT32_C( -2412.10)) },
312
+ { simde_mm256_set_ps(SIMDE_FLOAT32_C( 35.30), SIMDE_FLOAT32_C( -51.40),
313
+ SIMDE_FLOAT32_C( -71.80), SIMDE_FLOAT32_C( 28.30),
314
+ SIMDE_FLOAT32_C( 41.70), SIMDE_FLOAT32_C( -29.90),
315
+ SIMDE_FLOAT32_C( 47.10), SIMDE_FLOAT32_C( -23.50)),
316
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( -72.20), SIMDE_FLOAT32_C( 5.10),
317
+ SIMDE_FLOAT32_C( 50.30), SIMDE_FLOAT32_C( 8.80),
318
+ SIMDE_FLOAT32_C( 10.30), SIMDE_FLOAT32_C( 88.00),
319
+ SIMDE_FLOAT32_C( -32.10), SIMDE_FLOAT32_C( -71.80)),
320
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( 92.50), SIMDE_FLOAT32_C( -22.70),
321
+ SIMDE_FLOAT32_C( -32.50), SIMDE_FLOAT32_C( -64.00),
322
+ SIMDE_FLOAT32_C( 53.40), SIMDE_FLOAT32_C( 57.00),
323
+ SIMDE_FLOAT32_C( 85.20), SIMDE_FLOAT32_C( 51.90)),
324
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( -2456.16), SIMDE_FLOAT32_C( -284.84),
325
+ SIMDE_FLOAT32_C( -3644.04), SIMDE_FLOAT32_C( 185.04),
326
+ SIMDE_FLOAT32_C( 482.91), SIMDE_FLOAT32_C( -2574.20),
327
+ SIMDE_FLOAT32_C( -1426.71), SIMDE_FLOAT32_C( 1739.20)) },
328
+ { simde_mm256_set_ps(SIMDE_FLOAT32_C( 62.00), SIMDE_FLOAT32_C( -58.50),
329
+ SIMDE_FLOAT32_C( -89.10), SIMDE_FLOAT32_C( 51.50),
330
+ SIMDE_FLOAT32_C( 2.30), SIMDE_FLOAT32_C( -87.50),
331
+ SIMDE_FLOAT32_C( -72.60), SIMDE_FLOAT32_C( 96.30)),
332
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( -25.70), SIMDE_FLOAT32_C( 80.90),
333
+ SIMDE_FLOAT32_C( -77.80), SIMDE_FLOAT32_C( 4.90),
334
+ SIMDE_FLOAT32_C( 70.20), SIMDE_FLOAT32_C( 32.70),
335
+ SIMDE_FLOAT32_C( -60.70), SIMDE_FLOAT32_C( 68.00)),
336
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( -99.00), SIMDE_FLOAT32_C( -12.20),
337
+ SIMDE_FLOAT32_C( 41.70), SIMDE_FLOAT32_C( 9.80),
338
+ SIMDE_FLOAT32_C( -34.40), SIMDE_FLOAT32_C( -50.10),
339
+ SIMDE_FLOAT32_C( 35.40), SIMDE_FLOAT32_C( 62.60)),
340
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( -1692.40), SIMDE_FLOAT32_C( -4744.85),
341
+ SIMDE_FLOAT32_C( 6973.68), SIMDE_FLOAT32_C( 262.15),
342
+ SIMDE_FLOAT32_C( 127.06), SIMDE_FLOAT32_C( -2911.35),
343
+ SIMDE_FLOAT32_C( 4442.22), SIMDE_FLOAT32_C( 6611.00)) },
344
+ { simde_mm256_set_ps(SIMDE_FLOAT32_C( -40.00), SIMDE_FLOAT32_C( 62.80),
345
+ SIMDE_FLOAT32_C( -40.00), SIMDE_FLOAT32_C( 16.60),
346
+ SIMDE_FLOAT32_C( 60.10), SIMDE_FLOAT32_C( 22.60),
347
+ SIMDE_FLOAT32_C( -12.40), SIMDE_FLOAT32_C( 91.30)),
348
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( -98.70), SIMDE_FLOAT32_C( 17.00),
349
+ SIMDE_FLOAT32_C( -23.90), SIMDE_FLOAT32_C( 29.60),
350
+ SIMDE_FLOAT32_C( -52.60), SIMDE_FLOAT32_C( -30.60),
351
+ SIMDE_FLOAT32_C( 43.40), SIMDE_FLOAT32_C( 76.50)),
352
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( 61.00), SIMDE_FLOAT32_C( -10.10),
353
+ SIMDE_FLOAT32_C( 48.20), SIMDE_FLOAT32_C( 50.20),
354
+ SIMDE_FLOAT32_C( 12.20), SIMDE_FLOAT32_C( 64.80),
355
+ SIMDE_FLOAT32_C( -68.90), SIMDE_FLOAT32_C( -86.00)),
356
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( 4009.00), SIMDE_FLOAT32_C( 1057.50),
357
+ SIMDE_FLOAT32_C( 1004.20), SIMDE_FLOAT32_C( 541.56),
358
+ SIMDE_FLOAT32_C( -3149.06), SIMDE_FLOAT32_C( -626.76),
359
+ SIMDE_FLOAT32_C( -607.06), SIMDE_FLOAT32_C( 6898.45)) }
360
+ };
361
+
362
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
363
+ simde__m256 r = simde_mm256_fmadd_ps(test_vec[i].a, test_vec[i].b, test_vec[i].c);
364
+ simde_assert_m256_close(r, test_vec[i].r, 1);
365
+ }
366
+
367
+ return MUNIT_OK;
368
+ }
369
+
370
+ static MunitResult
371
+ test_simde_mm_fmadd_sd(const MunitParameter params[], void* data) {
372
+ (void) params;
373
+ (void) data;
374
+
375
+ const struct {
376
+ simde__m128d a;
377
+ simde__m128d b;
378
+ simde__m128d c;
379
+ simde__m128d r;
380
+ } test_vec[8] = {
381
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( 96.50), SIMDE_FLOAT64_C( -99.50)),
382
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 16.90), SIMDE_FLOAT64_C( -76.80)),
383
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 53.90), SIMDE_FLOAT64_C( 6.40)),
384
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 96.50), SIMDE_FLOAT64_C( 7648.00)) },
385
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( 53.00), SIMDE_FLOAT64_C( -2.10)),
386
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -68.70), SIMDE_FLOAT64_C( -11.70)),
387
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -76.30), SIMDE_FLOAT64_C( 62.90)),
388
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 53.00), SIMDE_FLOAT64_C( 87.47)) },
389
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( 34.00), SIMDE_FLOAT64_C( 30.50)),
390
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -42.50), SIMDE_FLOAT64_C( 32.70)),
391
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -36.20), SIMDE_FLOAT64_C( 36.10)),
392
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 34.00), SIMDE_FLOAT64_C( 1033.45)) },
393
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( -75.50), SIMDE_FLOAT64_C( -58.00)),
394
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 29.70), SIMDE_FLOAT64_C( -42.30)),
395
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 95.20), SIMDE_FLOAT64_C( 92.00)),
396
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -75.50), SIMDE_FLOAT64_C( 2545.40)) },
397
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( -59.60), SIMDE_FLOAT64_C( 12.90)),
398
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -84.80), SIMDE_FLOAT64_C( 50.40)),
399
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 42.20), SIMDE_FLOAT64_C( -77.30)),
400
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -59.60), SIMDE_FLOAT64_C( 572.86)) },
401
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( 78.40), SIMDE_FLOAT64_C( -77.00)),
402
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -76.10), SIMDE_FLOAT64_C( -11.00)),
403
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 5.80), SIMDE_FLOAT64_C( -75.90)),
404
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 78.40), SIMDE_FLOAT64_C( 771.10)) },
405
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( 6.70), SIMDE_FLOAT64_C( 47.60)),
406
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 24.30), SIMDE_FLOAT64_C( 93.60)),
407
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 93.40), SIMDE_FLOAT64_C( -50.10)),
408
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 6.70), SIMDE_FLOAT64_C( 4405.26)) },
409
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( -59.60), SIMDE_FLOAT64_C( -73.40)),
410
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 1.20), SIMDE_FLOAT64_C( 10.90)),
411
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 45.00), SIMDE_FLOAT64_C( -86.50)),
412
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -59.60), SIMDE_FLOAT64_C( -886.56)) }
413
+ };
414
+
415
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
416
+ simde__m128d r = simde_mm_fmadd_sd(test_vec[i].a, test_vec[i].b, test_vec[i].c);
417
+ simde_assert_m128d_close(r, test_vec[i].r, 1);
418
+ }
419
+
420
+ return MUNIT_OK;
421
+ }
422
+
423
+ static MunitResult
424
+ test_simde_mm_fmadd_ss(const MunitParameter params[], void* data) {
425
+ (void) params;
426
+ (void) data;
427
+
428
+ const struct {
429
+ simde__m128 a;
430
+ simde__m128 b;
431
+ simde__m128 c;
432
+ simde__m128 r;
433
+ } test_vec[8] = {
434
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( 29.70), SIMDE_FLOAT32_C( -13.10), SIMDE_FLOAT32_C( -92.70), SIMDE_FLOAT32_C( 44.80)),
435
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 25.90), SIMDE_FLOAT32_C( 67.70), SIMDE_FLOAT32_C( -12.20), SIMDE_FLOAT32_C( 72.20)),
436
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 45.60), SIMDE_FLOAT32_C( 36.90), SIMDE_FLOAT32_C( -98.40), SIMDE_FLOAT32_C( -64.80)),
437
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 29.70), SIMDE_FLOAT32_C( -13.10), SIMDE_FLOAT32_C( -92.70), SIMDE_FLOAT32_C( 3169.76)) },
438
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( 16.70), SIMDE_FLOAT32_C( 85.50), SIMDE_FLOAT32_C( 89.70), SIMDE_FLOAT32_C( -23.80)),
439
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 70.80), SIMDE_FLOAT32_C( 99.80), SIMDE_FLOAT32_C( -87.00), SIMDE_FLOAT32_C( 9.00)),
440
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 42.40), SIMDE_FLOAT32_C( 38.10), SIMDE_FLOAT32_C( -58.60), SIMDE_FLOAT32_C( -71.70)),
441
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 16.70), SIMDE_FLOAT32_C( 85.50), SIMDE_FLOAT32_C( 89.70), SIMDE_FLOAT32_C( -285.90)) },
442
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( -85.70), SIMDE_FLOAT32_C( 66.60), SIMDE_FLOAT32_C( -84.60), SIMDE_FLOAT32_C( -90.50)),
443
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 61.30), SIMDE_FLOAT32_C( -91.00), SIMDE_FLOAT32_C( -35.60), SIMDE_FLOAT32_C( -66.30)),
444
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 76.30), SIMDE_FLOAT32_C( -46.00), SIMDE_FLOAT32_C( 54.10), SIMDE_FLOAT32_C( 17.60)),
445
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -85.70), SIMDE_FLOAT32_C( 66.60), SIMDE_FLOAT32_C( -84.60), SIMDE_FLOAT32_C( 6017.75)) },
446
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( 14.40), SIMDE_FLOAT32_C( -25.60), SIMDE_FLOAT32_C( -65.60), SIMDE_FLOAT32_C( -71.00)),
447
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 65.40), SIMDE_FLOAT32_C( 95.90), SIMDE_FLOAT32_C( 51.70), SIMDE_FLOAT32_C( -84.50)),
448
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -47.60), SIMDE_FLOAT32_C( -50.00), SIMDE_FLOAT32_C( 88.40), SIMDE_FLOAT32_C( -28.80)),
449
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 14.40), SIMDE_FLOAT32_C( -25.60), SIMDE_FLOAT32_C( -65.60), SIMDE_FLOAT32_C( 5970.70)) },
450
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( 29.50), SIMDE_FLOAT32_C( -26.70), SIMDE_FLOAT32_C( 8.30), SIMDE_FLOAT32_C( -34.70)),
451
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 76.70), SIMDE_FLOAT32_C( -34.90), SIMDE_FLOAT32_C( -78.80), SIMDE_FLOAT32_C( 84.50)),
452
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 45.30), SIMDE_FLOAT32_C( -18.40), SIMDE_FLOAT32_C( -36.50), SIMDE_FLOAT32_C( -89.00)),
453
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 29.50), SIMDE_FLOAT32_C( -26.70), SIMDE_FLOAT32_C( 8.30), SIMDE_FLOAT32_C(-3021.15)) },
454
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( 64.00), SIMDE_FLOAT32_C( 46.60), SIMDE_FLOAT32_C( -17.50), SIMDE_FLOAT32_C( 24.10)),
455
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -67.40), SIMDE_FLOAT32_C( -16.40), SIMDE_FLOAT32_C( 38.30), SIMDE_FLOAT32_C( -92.30)),
456
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -66.80), SIMDE_FLOAT32_C( 10.60), SIMDE_FLOAT32_C( -6.70), SIMDE_FLOAT32_C( -49.00)),
457
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 64.00), SIMDE_FLOAT32_C( 46.60), SIMDE_FLOAT32_C( -17.50), SIMDE_FLOAT32_C(-2273.43)) },
458
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( -98.50), SIMDE_FLOAT32_C( 15.30), SIMDE_FLOAT32_C( -33.40), SIMDE_FLOAT32_C( 4.80)),
459
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 97.00), SIMDE_FLOAT32_C( -35.60), SIMDE_FLOAT32_C( 63.50), SIMDE_FLOAT32_C( -94.30)),
460
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -9.90), SIMDE_FLOAT32_C( -97.20), SIMDE_FLOAT32_C( -13.80), SIMDE_FLOAT32_C( 11.60)),
461
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -98.50), SIMDE_FLOAT32_C( 15.30), SIMDE_FLOAT32_C( -33.40), SIMDE_FLOAT32_C( -441.04)) },
462
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( -11.00), SIMDE_FLOAT32_C( -65.00), SIMDE_FLOAT32_C( -76.20), SIMDE_FLOAT32_C( 54.80)),
463
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 37.10), SIMDE_FLOAT32_C( -97.90), SIMDE_FLOAT32_C( -36.50), SIMDE_FLOAT32_C( 50.70)),
464
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -26.80), SIMDE_FLOAT32_C( -74.90), SIMDE_FLOAT32_C( -84.40), SIMDE_FLOAT32_C( 35.70)),
465
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -11.00), SIMDE_FLOAT32_C( -65.00), SIMDE_FLOAT32_C( -76.20), SIMDE_FLOAT32_C( 2814.06)) }
466
+ };
467
+
468
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
469
+ simde__m128 r = simde_mm_fmadd_ss(test_vec[i].a, test_vec[i].b, test_vec[i].c);
470
+ simde_assert_m128_close(r, test_vec[i].r, 1);
471
+ }
472
+
473
+ return MUNIT_OK;
474
+ }
475
+
476
+ static MunitResult
477
+ test_simde_mm_fmaddsub_pd(const MunitParameter params[], void* data) {
478
+ (void) params;
479
+ (void) data;
480
+
481
+ const struct {
482
+ simde__m128d a;
483
+ simde__m128d b;
484
+ simde__m128d c;
485
+ simde__m128d r;
486
+ } test_vec[8] = {
487
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( -72.20), SIMDE_FLOAT64_C( 74.40)),
488
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 41.60), SIMDE_FLOAT64_C( -13.50)),
489
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 18.90), SIMDE_FLOAT64_C( 65.30)),
490
+ simde_mm_set_pd(SIMDE_FLOAT64_C(-2984.62), SIMDE_FLOAT64_C(-1069.70)) },
491
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( 14.70), SIMDE_FLOAT64_C( 97.50)),
492
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 47.70), SIMDE_FLOAT64_C( 86.80)),
493
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 75.80), SIMDE_FLOAT64_C( 19.90)),
494
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 776.99), SIMDE_FLOAT64_C( 8443.10)) },
495
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( -24.00), SIMDE_FLOAT64_C( 39.90)),
496
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -0.00), SIMDE_FLOAT64_C( 42.00)),
497
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -66.10), SIMDE_FLOAT64_C( -55.20)),
498
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -66.10), SIMDE_FLOAT64_C( 1731.00)) },
499
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( -45.20), SIMDE_FLOAT64_C( 65.10)),
500
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -64.40), SIMDE_FLOAT64_C( 58.00)),
501
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 79.80), SIMDE_FLOAT64_C( 19.30)),
502
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 2990.68), SIMDE_FLOAT64_C( 3756.50)) },
503
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( 14.50), SIMDE_FLOAT64_C( -64.90)),
504
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 92.20), SIMDE_FLOAT64_C( -68.80)),
505
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -2.50), SIMDE_FLOAT64_C( -96.40)),
506
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 1334.40), SIMDE_FLOAT64_C( 4561.52)) },
507
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( 61.50), SIMDE_FLOAT64_C( 42.40)),
508
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 77.20), SIMDE_FLOAT64_C( 23.30)),
509
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 32.20), SIMDE_FLOAT64_C( 12.10)),
510
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 4780.00), SIMDE_FLOAT64_C( 975.82)) },
511
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( -38.20), SIMDE_FLOAT64_C( 8.10)),
512
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -10.10), SIMDE_FLOAT64_C( 98.60)),
513
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -14.20), SIMDE_FLOAT64_C( 22.60)),
514
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 371.62), SIMDE_FLOAT64_C( 776.06)) },
515
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( 8.40), SIMDE_FLOAT64_C( -30.70)),
516
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -39.30), SIMDE_FLOAT64_C( 73.40)),
517
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 25.30), SIMDE_FLOAT64_C( 2.90)),
518
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -304.82), SIMDE_FLOAT64_C(-2256.28)) }
519
+ };
520
+
521
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
522
+ simde__m128d r = simde_mm_fmaddsub_pd(test_vec[i].a, test_vec[i].b, test_vec[i].c);
523
+ simde_assert_m128d_close(r, test_vec[i].r, 1);
524
+ }
525
+
526
+ return MUNIT_OK;
527
+ }
528
+
529
+ static MunitResult
530
+ test_simde_mm256_fmaddsub_pd(const MunitParameter params[], void* data) {
531
+ (void) params;
532
+ (void) data;
533
+
534
+ const struct {
535
+ simde__m256d a;
536
+ simde__m256d b;
537
+ simde__m256d c;
538
+ simde__m256d r;
539
+ } test_vec[8] = {
540
+ { simde_mm256_set_pd(SIMDE_FLOAT64_C( -52.10), SIMDE_FLOAT64_C( -92.00),
541
+ SIMDE_FLOAT64_C( -82.90), SIMDE_FLOAT64_C( -49.00)),
542
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( -49.30), SIMDE_FLOAT64_C( -97.40),
543
+ SIMDE_FLOAT64_C( 58.80), SIMDE_FLOAT64_C( 67.60)),
544
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( 23.30), SIMDE_FLOAT64_C( 87.10),
545
+ SIMDE_FLOAT64_C( 71.70), SIMDE_FLOAT64_C( 97.10)),
546
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( 2591.83), SIMDE_FLOAT64_C( 8873.70),
547
+ SIMDE_FLOAT64_C(-4802.82), SIMDE_FLOAT64_C(-3409.50)) },
548
+ { simde_mm256_set_pd(SIMDE_FLOAT64_C( -83.80), SIMDE_FLOAT64_C( 50.40),
549
+ SIMDE_FLOAT64_C( -94.80), SIMDE_FLOAT64_C( -86.80)),
550
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( 3.10), SIMDE_FLOAT64_C( -46.80),
551
+ SIMDE_FLOAT64_C( -3.10), SIMDE_FLOAT64_C( 83.00)),
552
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( -11.70), SIMDE_FLOAT64_C( 76.10),
553
+ SIMDE_FLOAT64_C( 44.50), SIMDE_FLOAT64_C( 28.00)),
554
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( -271.48), SIMDE_FLOAT64_C(-2434.82),
555
+ SIMDE_FLOAT64_C( 338.38), SIMDE_FLOAT64_C(-7232.40)) },
556
+ { simde_mm256_set_pd(SIMDE_FLOAT64_C( -64.40), SIMDE_FLOAT64_C( 40.90),
557
+ SIMDE_FLOAT64_C( 36.80), SIMDE_FLOAT64_C( -1.00)),
558
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( -57.50), SIMDE_FLOAT64_C( 5.00),
559
+ SIMDE_FLOAT64_C( -21.50), SIMDE_FLOAT64_C( -1.70)),
560
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( 21.60), SIMDE_FLOAT64_C( -36.20),
561
+ SIMDE_FLOAT64_C( -67.50), SIMDE_FLOAT64_C( -19.30)),
562
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( 3724.60), SIMDE_FLOAT64_C( 240.70),
563
+ SIMDE_FLOAT64_C( -858.70), SIMDE_FLOAT64_C( 21.00)) },
564
+ { simde_mm256_set_pd(SIMDE_FLOAT64_C( -37.10), SIMDE_FLOAT64_C( 2.20),
565
+ SIMDE_FLOAT64_C( -99.10), SIMDE_FLOAT64_C( 78.20)),
566
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( -60.20), SIMDE_FLOAT64_C( 29.30),
567
+ SIMDE_FLOAT64_C( 2.50), SIMDE_FLOAT64_C( -40.10)),
568
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( 59.60), SIMDE_FLOAT64_C( -28.40),
569
+ SIMDE_FLOAT64_C( 58.10), SIMDE_FLOAT64_C( 96.90)),
570
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( 2293.02), SIMDE_FLOAT64_C( 92.86),
571
+ SIMDE_FLOAT64_C( -189.65), SIMDE_FLOAT64_C(-3232.72)) },
572
+ { simde_mm256_set_pd(SIMDE_FLOAT64_C( -88.70), SIMDE_FLOAT64_C( -20.50),
573
+ SIMDE_FLOAT64_C( 28.00), SIMDE_FLOAT64_C( -13.70)),
574
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( -49.60), SIMDE_FLOAT64_C( -13.90),
575
+ SIMDE_FLOAT64_C( 71.80), SIMDE_FLOAT64_C( -29.40)),
576
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( -95.70), SIMDE_FLOAT64_C( 48.30),
577
+ SIMDE_FLOAT64_C( 78.20), SIMDE_FLOAT64_C( -6.00)),
578
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( 4303.82), SIMDE_FLOAT64_C( 236.65),
579
+ SIMDE_FLOAT64_C( 2088.60), SIMDE_FLOAT64_C( 408.78)) },
580
+ { simde_mm256_set_pd(SIMDE_FLOAT64_C( 74.60), SIMDE_FLOAT64_C( 40.20),
581
+ SIMDE_FLOAT64_C( -4.40), SIMDE_FLOAT64_C( 51.30)),
582
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( 21.60), SIMDE_FLOAT64_C( -83.50),
583
+ SIMDE_FLOAT64_C( -2.00), SIMDE_FLOAT64_C( -6.60)),
584
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( 13.70), SIMDE_FLOAT64_C( 39.10),
585
+ SIMDE_FLOAT64_C( 92.60), SIMDE_FLOAT64_C( -41.90)),
586
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( 1625.06), SIMDE_FLOAT64_C(-3395.80),
587
+ SIMDE_FLOAT64_C( 101.40), SIMDE_FLOAT64_C( -296.68)) },
588
+ { simde_mm256_set_pd(SIMDE_FLOAT64_C( -14.60), SIMDE_FLOAT64_C( -32.40),
589
+ SIMDE_FLOAT64_C( 94.80), SIMDE_FLOAT64_C( -5.20)),
590
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( 15.40), SIMDE_FLOAT64_C( -34.50),
591
+ SIMDE_FLOAT64_C( 91.60), SIMDE_FLOAT64_C( -58.60)),
592
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( -70.70), SIMDE_FLOAT64_C( -91.10),
593
+ SIMDE_FLOAT64_C( -42.30), SIMDE_FLOAT64_C( 64.70)),
594
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( -295.54), SIMDE_FLOAT64_C( 1208.90),
595
+ SIMDE_FLOAT64_C( 8641.38), SIMDE_FLOAT64_C( 240.02)) },
596
+ { simde_mm256_set_pd(SIMDE_FLOAT64_C( 68.20), SIMDE_FLOAT64_C( -45.40),
597
+ SIMDE_FLOAT64_C( 33.10), SIMDE_FLOAT64_C( 17.10)),
598
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( 52.00), SIMDE_FLOAT64_C( 24.80),
599
+ SIMDE_FLOAT64_C( 6.10), SIMDE_FLOAT64_C( 68.80)),
600
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( 30.10), SIMDE_FLOAT64_C( 11.20),
601
+ SIMDE_FLOAT64_C( -78.00), SIMDE_FLOAT64_C( -47.50)),
602
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( 3576.50), SIMDE_FLOAT64_C(-1137.12),
603
+ SIMDE_FLOAT64_C( 123.91), SIMDE_FLOAT64_C( 1223.98)) }
604
+ };
605
+
606
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
607
+ simde__m256d r = simde_mm256_fmaddsub_pd(test_vec[i].a, test_vec[i].b, test_vec[i].c);
608
+ simde_assert_m256d_close(r, test_vec[i].r, 1);
609
+ }
610
+
611
+ return MUNIT_OK;
612
+ }
613
+
614
+ static MunitResult
615
+ test_simde_mm_fmaddsub_ps(const MunitParameter params[], void* data) {
616
+ (void) params;
617
+ (void) data;
618
+
619
+ const struct {
620
+ simde__m128 a;
621
+ simde__m128 b;
622
+ simde__m128 c;
623
+ simde__m128 r;
624
+ } test_vec[8] = {
625
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( -75.30), SIMDE_FLOAT32_C( 37.60), SIMDE_FLOAT32_C( 76.00), SIMDE_FLOAT32_C( -4.20)),
626
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -15.80), SIMDE_FLOAT32_C( 64.20), SIMDE_FLOAT32_C( 50.90), SIMDE_FLOAT32_C( 26.10)),
627
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 64.80), SIMDE_FLOAT32_C( -10.00), SIMDE_FLOAT32_C( -97.40), SIMDE_FLOAT32_C( -90.30)),
628
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 1254.54), SIMDE_FLOAT32_C( 2423.92), SIMDE_FLOAT32_C( 3771.00), SIMDE_FLOAT32_C( -19.32)) },
629
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( -48.70), SIMDE_FLOAT32_C( 50.40), SIMDE_FLOAT32_C( -22.00), SIMDE_FLOAT32_C( 76.40)),
630
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -80.30), SIMDE_FLOAT32_C( -99.30), SIMDE_FLOAT32_C( -86.10), SIMDE_FLOAT32_C( 30.10)),
631
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -41.10), SIMDE_FLOAT32_C( 57.20), SIMDE_FLOAT32_C( -41.90), SIMDE_FLOAT32_C( -88.50)),
632
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 3869.51), SIMDE_FLOAT32_C(-5061.92), SIMDE_FLOAT32_C( 1852.30), SIMDE_FLOAT32_C( 2388.14)) },
633
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( 99.70), SIMDE_FLOAT32_C( 2.10), SIMDE_FLOAT32_C( 41.80), SIMDE_FLOAT32_C( -15.20)),
634
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 16.20), SIMDE_FLOAT32_C( -74.30), SIMDE_FLOAT32_C( -71.40), SIMDE_FLOAT32_C( 51.70)),
635
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 71.10), SIMDE_FLOAT32_C( -90.60), SIMDE_FLOAT32_C( -33.50), SIMDE_FLOAT32_C( -68.40)),
636
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 1686.24), SIMDE_FLOAT32_C( -65.43), SIMDE_FLOAT32_C(-3018.02), SIMDE_FLOAT32_C( -717.44)) },
637
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( 39.90), SIMDE_FLOAT32_C( 12.10), SIMDE_FLOAT32_C( -93.10), SIMDE_FLOAT32_C( -73.80)),
638
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -72.70), SIMDE_FLOAT32_C( -61.90), SIMDE_FLOAT32_C( 1.90), SIMDE_FLOAT32_C( 89.00)),
639
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -63.40), SIMDE_FLOAT32_C( -46.10), SIMDE_FLOAT32_C( 50.20), SIMDE_FLOAT32_C( -74.10)),
640
+ simde_mm_set_ps(SIMDE_FLOAT32_C(-2964.13), SIMDE_FLOAT32_C( -702.89), SIMDE_FLOAT32_C( -126.69), SIMDE_FLOAT32_C(-6494.10)) },
641
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( -37.90), SIMDE_FLOAT32_C( 16.10), SIMDE_FLOAT32_C( 65.80), SIMDE_FLOAT32_C( 65.20)),
642
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -95.90), SIMDE_FLOAT32_C( 9.30), SIMDE_FLOAT32_C( 33.70), SIMDE_FLOAT32_C( -30.80)),
643
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 4.30), SIMDE_FLOAT32_C( -27.90), SIMDE_FLOAT32_C( -62.30), SIMDE_FLOAT32_C( -71.40)),
644
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 3638.91), SIMDE_FLOAT32_C( 177.63), SIMDE_FLOAT32_C( 2155.16), SIMDE_FLOAT32_C(-1936.76)) },
645
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( 40.20), SIMDE_FLOAT32_C( -28.10), SIMDE_FLOAT32_C( -39.20), SIMDE_FLOAT32_C( 15.00)),
646
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -33.70), SIMDE_FLOAT32_C( -55.90), SIMDE_FLOAT32_C( -9.80), SIMDE_FLOAT32_C( -88.20)),
647
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 3.20), SIMDE_FLOAT32_C( -50.90), SIMDE_FLOAT32_C( 35.30), SIMDE_FLOAT32_C( -45.30)),
648
+ simde_mm_set_ps(SIMDE_FLOAT32_C(-1351.54), SIMDE_FLOAT32_C( 1621.69), SIMDE_FLOAT32_C( 419.46), SIMDE_FLOAT32_C(-1277.70)) },
649
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( 2.60), SIMDE_FLOAT32_C( 70.50), SIMDE_FLOAT32_C( 56.20), SIMDE_FLOAT32_C( 5.90)),
650
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -66.40), SIMDE_FLOAT32_C( 95.00), SIMDE_FLOAT32_C( 95.50), SIMDE_FLOAT32_C( -15.90)),
651
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -60.10), SIMDE_FLOAT32_C( -25.30), SIMDE_FLOAT32_C( -69.10), SIMDE_FLOAT32_C( -77.70)),
652
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -232.74), SIMDE_FLOAT32_C( 6722.80), SIMDE_FLOAT32_C( 5298.00), SIMDE_FLOAT32_C( -16.11)) },
653
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( -91.20), SIMDE_FLOAT32_C( 32.90), SIMDE_FLOAT32_C( -8.90), SIMDE_FLOAT32_C( -97.30)),
654
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -78.50), SIMDE_FLOAT32_C( 49.50), SIMDE_FLOAT32_C( 63.70), SIMDE_FLOAT32_C( -83.00)),
655
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 10.30), SIMDE_FLOAT32_C( 73.30), SIMDE_FLOAT32_C( -68.20), SIMDE_FLOAT32_C( 60.90)),
656
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 7169.50), SIMDE_FLOAT32_C( 1555.25), SIMDE_FLOAT32_C( -635.13), SIMDE_FLOAT32_C( 8015.00)) }
657
+ };
658
+
659
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
660
+ simde__m128 r = simde_mm_fmaddsub_ps(test_vec[i].a, test_vec[i].b, test_vec[i].c);
661
+ simde_assert_m128_close(r, test_vec[i].r, 1);
662
+ }
663
+
664
+ return MUNIT_OK;
665
+ }
666
+
667
+ static MunitResult
668
+ test_simde_mm256_fmaddsub_ps(const MunitParameter params[], void* data) {
669
+ (void) params;
670
+ (void) data;
671
+
672
+ const struct {
673
+ simde__m256 a;
674
+ simde__m256 b;
675
+ simde__m256 c;
676
+ simde__m256 r;
677
+ } test_vec[8] = {
678
+ { simde_mm256_set_ps(SIMDE_FLOAT32_C( -61.10), SIMDE_FLOAT32_C( -95.60),
679
+ SIMDE_FLOAT32_C( 56.00), SIMDE_FLOAT32_C( 46.30),
680
+ SIMDE_FLOAT32_C( -62.80), SIMDE_FLOAT32_C( 38.90),
681
+ SIMDE_FLOAT32_C( -92.60), SIMDE_FLOAT32_C( 65.40)),
682
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( -38.30), SIMDE_FLOAT32_C( -1.90),
683
+ SIMDE_FLOAT32_C( -28.00), SIMDE_FLOAT32_C( -43.20),
684
+ SIMDE_FLOAT32_C( -19.40), SIMDE_FLOAT32_C( 57.60),
685
+ SIMDE_FLOAT32_C( -97.20), SIMDE_FLOAT32_C( 81.20)),
686
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( 93.20), SIMDE_FLOAT32_C( -43.00),
687
+ SIMDE_FLOAT32_C( -47.40), SIMDE_FLOAT32_C( -77.00),
688
+ SIMDE_FLOAT32_C( -59.90), SIMDE_FLOAT32_C( 17.90),
689
+ SIMDE_FLOAT32_C( -9.60), SIMDE_FLOAT32_C( -61.30)),
690
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( 2433.33), SIMDE_FLOAT32_C( 224.64),
691
+ SIMDE_FLOAT32_C(-1615.40), SIMDE_FLOAT32_C(-1923.16),
692
+ SIMDE_FLOAT32_C( 1158.42), SIMDE_FLOAT32_C( 2222.74),
693
+ SIMDE_FLOAT32_C( 8991.12), SIMDE_FLOAT32_C( 5371.78)) },
694
+ { simde_mm256_set_ps(SIMDE_FLOAT32_C( -23.40), SIMDE_FLOAT32_C( -24.60),
695
+ SIMDE_FLOAT32_C( 35.70), SIMDE_FLOAT32_C( 59.90),
696
+ SIMDE_FLOAT32_C( -91.00), SIMDE_FLOAT32_C( -25.40),
697
+ SIMDE_FLOAT32_C( -88.30), SIMDE_FLOAT32_C( -99.80)),
698
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( -30.70), SIMDE_FLOAT32_C( 97.10),
699
+ SIMDE_FLOAT32_C( 86.90), SIMDE_FLOAT32_C( -81.10),
700
+ SIMDE_FLOAT32_C( -71.30), SIMDE_FLOAT32_C( -61.20),
701
+ SIMDE_FLOAT32_C( -26.10), SIMDE_FLOAT32_C( 31.60)),
702
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( -74.30), SIMDE_FLOAT32_C( -19.40),
703
+ SIMDE_FLOAT32_C( -70.80), SIMDE_FLOAT32_C( -13.00),
704
+ SIMDE_FLOAT32_C( 82.90), SIMDE_FLOAT32_C( -75.70),
705
+ SIMDE_FLOAT32_C( -31.50), SIMDE_FLOAT32_C( 73.50)),
706
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( 644.08), SIMDE_FLOAT32_C(-2369.26),
707
+ SIMDE_FLOAT32_C( 3031.53), SIMDE_FLOAT32_C(-4844.89),
708
+ SIMDE_FLOAT32_C( 6571.20), SIMDE_FLOAT32_C( 1630.18),
709
+ SIMDE_FLOAT32_C( 2273.13), SIMDE_FLOAT32_C(-3227.18)) },
710
+ { simde_mm256_set_ps(SIMDE_FLOAT32_C( 70.20), SIMDE_FLOAT32_C( -20.40),
711
+ SIMDE_FLOAT32_C( -51.50), SIMDE_FLOAT32_C( 82.30),
712
+ SIMDE_FLOAT32_C( 31.30), SIMDE_FLOAT32_C( 17.80),
713
+ SIMDE_FLOAT32_C( -39.60), SIMDE_FLOAT32_C( 66.80)),
714
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( -72.50), SIMDE_FLOAT32_C( 52.00),
715
+ SIMDE_FLOAT32_C( -54.80), SIMDE_FLOAT32_C( 14.00),
716
+ SIMDE_FLOAT32_C( 91.80), SIMDE_FLOAT32_C( -80.70),
717
+ SIMDE_FLOAT32_C( -97.90), SIMDE_FLOAT32_C( -99.00)),
718
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( -76.10), SIMDE_FLOAT32_C( 26.90),
719
+ SIMDE_FLOAT32_C( 24.90), SIMDE_FLOAT32_C( -50.60),
720
+ SIMDE_FLOAT32_C( 66.90), SIMDE_FLOAT32_C( 82.40),
721
+ SIMDE_FLOAT32_C( 98.50), SIMDE_FLOAT32_C( 9.60)),
722
+ simde_mm256_set_ps(SIMDE_FLOAT32_C(-5165.60), SIMDE_FLOAT32_C(-1087.70),
723
+ SIMDE_FLOAT32_C( 2847.10), SIMDE_FLOAT32_C( 1202.80),
724
+ SIMDE_FLOAT32_C( 2940.24), SIMDE_FLOAT32_C(-1518.86),
725
+ SIMDE_FLOAT32_C( 3975.34), SIMDE_FLOAT32_C(-6622.80)) },
726
+ { simde_mm256_set_ps(SIMDE_FLOAT32_C( -28.40), SIMDE_FLOAT32_C( 22.80),
727
+ SIMDE_FLOAT32_C( 16.40), SIMDE_FLOAT32_C( 80.20),
728
+ SIMDE_FLOAT32_C( -24.10), SIMDE_FLOAT32_C( -83.00),
729
+ SIMDE_FLOAT32_C( -74.10), SIMDE_FLOAT32_C( -49.60)),
730
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( -73.70), SIMDE_FLOAT32_C( 59.00),
731
+ SIMDE_FLOAT32_C( 36.90), SIMDE_FLOAT32_C( 7.50),
732
+ SIMDE_FLOAT32_C( -74.80), SIMDE_FLOAT32_C( -84.40),
733
+ SIMDE_FLOAT32_C( 79.60), SIMDE_FLOAT32_C( -90.70)),
734
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( -76.40), SIMDE_FLOAT32_C( 28.10),
735
+ SIMDE_FLOAT32_C( -13.60), SIMDE_FLOAT32_C( -71.50),
736
+ SIMDE_FLOAT32_C( -52.20), SIMDE_FLOAT32_C( -30.20),
737
+ SIMDE_FLOAT32_C( -62.60), SIMDE_FLOAT32_C( 2.30)),
738
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( 2016.68), SIMDE_FLOAT32_C( 1317.10),
739
+ SIMDE_FLOAT32_C( 591.56), SIMDE_FLOAT32_C( 673.00),
740
+ SIMDE_FLOAT32_C( 1750.48), SIMDE_FLOAT32_C( 7035.40),
741
+ SIMDE_FLOAT32_C(-5960.96), SIMDE_FLOAT32_C( 4496.42)) },
742
+ { simde_mm256_set_ps(SIMDE_FLOAT32_C( 74.80), SIMDE_FLOAT32_C( 79.60),
743
+ SIMDE_FLOAT32_C( -91.30), SIMDE_FLOAT32_C( 86.60),
744
+ SIMDE_FLOAT32_C( 41.70), SIMDE_FLOAT32_C( -74.30),
745
+ SIMDE_FLOAT32_C( -75.60), SIMDE_FLOAT32_C( 28.50)),
746
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( 96.00), SIMDE_FLOAT32_C( 10.10),
747
+ SIMDE_FLOAT32_C( -63.40), SIMDE_FLOAT32_C( 96.90),
748
+ SIMDE_FLOAT32_C( 66.20), SIMDE_FLOAT32_C( -75.30),
749
+ SIMDE_FLOAT32_C( -11.80), SIMDE_FLOAT32_C( 30.70)),
750
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( 31.70), SIMDE_FLOAT32_C( -47.90),
751
+ SIMDE_FLOAT32_C( 27.70), SIMDE_FLOAT32_C( 40.70),
752
+ SIMDE_FLOAT32_C( -22.80), SIMDE_FLOAT32_C( 35.80),
753
+ SIMDE_FLOAT32_C( -30.10), SIMDE_FLOAT32_C( 88.00)),
754
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( 7212.50), SIMDE_FLOAT32_C( 851.86),
755
+ SIMDE_FLOAT32_C( 5816.12), SIMDE_FLOAT32_C( 8350.84),
756
+ SIMDE_FLOAT32_C( 2737.74), SIMDE_FLOAT32_C( 5558.99),
757
+ SIMDE_FLOAT32_C( 861.98), SIMDE_FLOAT32_C( 786.95)) },
758
+ { simde_mm256_set_ps(SIMDE_FLOAT32_C( 91.80), SIMDE_FLOAT32_C( -99.10),
759
+ SIMDE_FLOAT32_C( -91.30), SIMDE_FLOAT32_C( 69.40),
760
+ SIMDE_FLOAT32_C( 38.40), SIMDE_FLOAT32_C( -90.40),
761
+ SIMDE_FLOAT32_C( 62.20), SIMDE_FLOAT32_C( -62.50)),
762
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( 82.70), SIMDE_FLOAT32_C( -63.90),
763
+ SIMDE_FLOAT32_C( 57.00), SIMDE_FLOAT32_C( -53.70),
764
+ SIMDE_FLOAT32_C( -62.00), SIMDE_FLOAT32_C( 87.90),
765
+ SIMDE_FLOAT32_C( -60.70), SIMDE_FLOAT32_C( -94.50)),
766
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( 65.30), SIMDE_FLOAT32_C( 61.10),
767
+ SIMDE_FLOAT32_C( -35.30), SIMDE_FLOAT32_C( -37.60),
768
+ SIMDE_FLOAT32_C( 3.40), SIMDE_FLOAT32_C( 10.20),
769
+ SIMDE_FLOAT32_C( 25.70), SIMDE_FLOAT32_C( 31.10)),
770
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( 7657.16), SIMDE_FLOAT32_C( 6271.39),
771
+ SIMDE_FLOAT32_C(-5239.40), SIMDE_FLOAT32_C(-3689.18),
772
+ SIMDE_FLOAT32_C(-2377.40), SIMDE_FLOAT32_C(-7956.36),
773
+ SIMDE_FLOAT32_C(-3749.84), SIMDE_FLOAT32_C( 5875.15)) },
774
+ { simde_mm256_set_ps(SIMDE_FLOAT32_C( -9.70), SIMDE_FLOAT32_C( 54.70),
775
+ SIMDE_FLOAT32_C( -66.40), SIMDE_FLOAT32_C( -34.70),
776
+ SIMDE_FLOAT32_C( -27.90), SIMDE_FLOAT32_C( 92.40),
777
+ SIMDE_FLOAT32_C( -11.40), SIMDE_FLOAT32_C( 14.60)),
778
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( -0.90), SIMDE_FLOAT32_C( -71.50),
779
+ SIMDE_FLOAT32_C( 67.00), SIMDE_FLOAT32_C( -56.30),
780
+ SIMDE_FLOAT32_C( 74.40), SIMDE_FLOAT32_C( 9.80),
781
+ SIMDE_FLOAT32_C( -5.30), SIMDE_FLOAT32_C( 63.80)),
782
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( -47.10), SIMDE_FLOAT32_C( 81.20),
783
+ SIMDE_FLOAT32_C( 31.00), SIMDE_FLOAT32_C( 11.50),
784
+ SIMDE_FLOAT32_C( 67.80), SIMDE_FLOAT32_C( -14.20),
785
+ SIMDE_FLOAT32_C( -62.80), SIMDE_FLOAT32_C( 84.70)),
786
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( -38.37), SIMDE_FLOAT32_C(-3992.25),
787
+ SIMDE_FLOAT32_C(-4417.80), SIMDE_FLOAT32_C( 1942.11),
788
+ SIMDE_FLOAT32_C(-2007.96), SIMDE_FLOAT32_C( 919.72),
789
+ SIMDE_FLOAT32_C( -2.38), SIMDE_FLOAT32_C( 846.78)) },
790
+ { simde_mm256_set_ps(SIMDE_FLOAT32_C( -14.10), SIMDE_FLOAT32_C( -90.60),
791
+ SIMDE_FLOAT32_C( 37.70), SIMDE_FLOAT32_C( 63.50),
792
+ SIMDE_FLOAT32_C( -67.90), SIMDE_FLOAT32_C( -75.70),
793
+ SIMDE_FLOAT32_C( 48.30), SIMDE_FLOAT32_C( 69.80)),
794
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( 21.20), SIMDE_FLOAT32_C( -56.80),
795
+ SIMDE_FLOAT32_C( -51.20), SIMDE_FLOAT32_C( -55.60),
796
+ SIMDE_FLOAT32_C( 65.10), SIMDE_FLOAT32_C( 21.30),
797
+ SIMDE_FLOAT32_C( -29.20), SIMDE_FLOAT32_C( -61.60)),
798
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( 55.80), SIMDE_FLOAT32_C( -16.50),
799
+ SIMDE_FLOAT32_C( 90.30), SIMDE_FLOAT32_C( 10.50),
800
+ SIMDE_FLOAT32_C( -35.10), SIMDE_FLOAT32_C( 8.40),
801
+ SIMDE_FLOAT32_C( -35.70), SIMDE_FLOAT32_C( 70.80)),
802
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( -243.12), SIMDE_FLOAT32_C( 5162.58),
803
+ SIMDE_FLOAT32_C(-1839.94), SIMDE_FLOAT32_C(-3541.10),
804
+ SIMDE_FLOAT32_C(-4455.39), SIMDE_FLOAT32_C(-1620.81),
805
+ SIMDE_FLOAT32_C(-1446.06), SIMDE_FLOAT32_C(-4370.48)) }
806
+ };
807
+
808
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
809
+ simde__m256 r = simde_mm256_fmaddsub_ps(test_vec[i].a, test_vec[i].b, test_vec[i].c);
810
+ simde_assert_m256_close(r, test_vec[i].r, 1);
811
+ }
812
+
813
+ return MUNIT_OK;
814
+ }
815
+
816
+ static MunitResult
817
+ test_simde_mm_fmsub_pd(const MunitParameter params[], void* data) {
818
+ (void) params;
819
+ (void) data;
820
+
821
+ const struct {
822
+ simde__m128d a;
823
+ simde__m128d b;
824
+ simde__m128d c;
825
+ simde__m128d r;
826
+ } test_vec[8] = {
827
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( -86.00), SIMDE_FLOAT64_C( -88.70)),
828
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 79.20), SIMDE_FLOAT64_C( -72.70)),
829
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 69.30), SIMDE_FLOAT64_C( -94.80)),
830
+ simde_mm_set_pd(SIMDE_FLOAT64_C(-6880.50), SIMDE_FLOAT64_C( 6543.29)) },
831
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( 29.00), SIMDE_FLOAT64_C( -23.00)),
832
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 88.90), SIMDE_FLOAT64_C( 98.30)),
833
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -27.80), SIMDE_FLOAT64_C( -64.50)),
834
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 2605.90), SIMDE_FLOAT64_C(-2196.40)) },
835
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( -56.40), SIMDE_FLOAT64_C( 49.90)),
836
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 79.10), SIMDE_FLOAT64_C( -51.70)),
837
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -66.70), SIMDE_FLOAT64_C( 16.60)),
838
+ simde_mm_set_pd(SIMDE_FLOAT64_C(-4394.54), SIMDE_FLOAT64_C(-2596.43)) },
839
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( -73.50), SIMDE_FLOAT64_C( 25.30)),
840
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 95.60), SIMDE_FLOAT64_C( 38.70)),
841
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 96.40), SIMDE_FLOAT64_C( 40.10)),
842
+ simde_mm_set_pd(SIMDE_FLOAT64_C(-7123.00), SIMDE_FLOAT64_C( 939.01)) },
843
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( 96.70), SIMDE_FLOAT64_C( -25.90)),
844
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -31.20), SIMDE_FLOAT64_C( -59.90)),
845
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -18.60), SIMDE_FLOAT64_C( -15.90)),
846
+ simde_mm_set_pd(SIMDE_FLOAT64_C(-2998.44), SIMDE_FLOAT64_C( 1567.31)) },
847
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( -63.20), SIMDE_FLOAT64_C( -69.90)),
848
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -27.30), SIMDE_FLOAT64_C( 57.70)),
849
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 17.60), SIMDE_FLOAT64_C( 32.80)),
850
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 1707.76), SIMDE_FLOAT64_C(-4066.03)) },
851
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( -33.50), SIMDE_FLOAT64_C( 64.70)),
852
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -68.40), SIMDE_FLOAT64_C( -49.20)),
853
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 44.70), SIMDE_FLOAT64_C( 88.70)),
854
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 2246.70), SIMDE_FLOAT64_C(-3271.94)) },
855
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( -79.60), SIMDE_FLOAT64_C( -61.60)),
856
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -29.20), SIMDE_FLOAT64_C( -21.10)),
857
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -94.70), SIMDE_FLOAT64_C( -26.20)),
858
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 2419.02), SIMDE_FLOAT64_C( 1325.96)) }
859
+ };
860
+
861
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
862
+ simde__m128d r = simde_mm_fmsub_pd(test_vec[i].a, test_vec[i].b, test_vec[i].c);
863
+ simde_assert_m128d_close(r, test_vec[i].r, 1);
864
+ }
865
+
866
+ return MUNIT_OK;
867
+ }
868
+
869
+ static MunitResult
870
+ test_simde_mm256_fmsub_pd(const MunitParameter params[], void* data) {
871
+ (void) params;
872
+ (void) data;
873
+
874
+ const struct {
875
+ simde__m256d a;
876
+ simde__m256d b;
877
+ simde__m256d c;
878
+ simde__m256d r;
879
+ } test_vec[8] = {
880
+ { simde_mm256_set_pd(SIMDE_FLOAT64_C( 34.80), SIMDE_FLOAT64_C( 57.60),
881
+ SIMDE_FLOAT64_C( 21.20), SIMDE_FLOAT64_C( 58.70)),
882
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( -15.50), SIMDE_FLOAT64_C( -85.90),
883
+ SIMDE_FLOAT64_C( 76.40), SIMDE_FLOAT64_C( 37.40)),
884
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( -67.00), SIMDE_FLOAT64_C( -15.40),
885
+ SIMDE_FLOAT64_C( 94.00), SIMDE_FLOAT64_C( -95.50)),
886
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( -472.40), SIMDE_FLOAT64_C(-4932.44),
887
+ SIMDE_FLOAT64_C( 1525.68), SIMDE_FLOAT64_C( 2290.88)) },
888
+ { simde_mm256_set_pd(SIMDE_FLOAT64_C( 79.40), SIMDE_FLOAT64_C( -18.40),
889
+ SIMDE_FLOAT64_C( -87.30), SIMDE_FLOAT64_C( -43.70)),
890
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( 46.70), SIMDE_FLOAT64_C( -61.00),
891
+ SIMDE_FLOAT64_C( 22.50), SIMDE_FLOAT64_C( -19.30)),
892
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( -92.50), SIMDE_FLOAT64_C( 24.60),
893
+ SIMDE_FLOAT64_C( 48.50), SIMDE_FLOAT64_C( 81.10)),
894
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( 3800.48), SIMDE_FLOAT64_C( 1097.80),
895
+ SIMDE_FLOAT64_C(-2012.75), SIMDE_FLOAT64_C( 762.31)) },
896
+ { simde_mm256_set_pd(SIMDE_FLOAT64_C( -53.30), SIMDE_FLOAT64_C( 37.50),
897
+ SIMDE_FLOAT64_C( -12.20), SIMDE_FLOAT64_C( 77.20)),
898
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( 44.30), SIMDE_FLOAT64_C( 68.70),
899
+ SIMDE_FLOAT64_C( 45.00), SIMDE_FLOAT64_C( -94.90)),
900
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( 10.90), SIMDE_FLOAT64_C( -78.60),
901
+ SIMDE_FLOAT64_C( 59.40), SIMDE_FLOAT64_C( 54.70)),
902
+ simde_mm256_set_pd(SIMDE_FLOAT64_C(-2372.09), SIMDE_FLOAT64_C( 2654.85),
903
+ SIMDE_FLOAT64_C( -608.40), SIMDE_FLOAT64_C(-7380.98)) },
904
+ { simde_mm256_set_pd(SIMDE_FLOAT64_C( -45.20), SIMDE_FLOAT64_C( -98.30),
905
+ SIMDE_FLOAT64_C( 6.30), SIMDE_FLOAT64_C( -64.20)),
906
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( 47.00), SIMDE_FLOAT64_C( -17.30),
907
+ SIMDE_FLOAT64_C( 90.50), SIMDE_FLOAT64_C( 33.20)),
908
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( -47.40), SIMDE_FLOAT64_C( -48.00),
909
+ SIMDE_FLOAT64_C( 92.50), SIMDE_FLOAT64_C( -62.30)),
910
+ simde_mm256_set_pd(SIMDE_FLOAT64_C(-2077.00), SIMDE_FLOAT64_C( 1748.59),
911
+ SIMDE_FLOAT64_C( 477.65), SIMDE_FLOAT64_C(-2069.14)) },
912
+ { simde_mm256_set_pd(SIMDE_FLOAT64_C( -66.50), SIMDE_FLOAT64_C( 50.50),
913
+ SIMDE_FLOAT64_C( -60.50), SIMDE_FLOAT64_C( 97.50)),
914
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( -77.70), SIMDE_FLOAT64_C( -31.10),
915
+ SIMDE_FLOAT64_C( 56.50), SIMDE_FLOAT64_C( -49.90)),
916
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( -78.10), SIMDE_FLOAT64_C( -33.20),
917
+ SIMDE_FLOAT64_C( 60.50), SIMDE_FLOAT64_C( 91.20)),
918
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( 5245.15), SIMDE_FLOAT64_C(-1537.35),
919
+ SIMDE_FLOAT64_C(-3478.75), SIMDE_FLOAT64_C(-4956.45)) },
920
+ { simde_mm256_set_pd(SIMDE_FLOAT64_C( 97.80), SIMDE_FLOAT64_C( 3.10),
921
+ SIMDE_FLOAT64_C( -8.70), SIMDE_FLOAT64_C( 56.90)),
922
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( 88.60), SIMDE_FLOAT64_C( -73.80),
923
+ SIMDE_FLOAT64_C( 92.30), SIMDE_FLOAT64_C( 21.50)),
924
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( -81.80), SIMDE_FLOAT64_C( -53.80),
925
+ SIMDE_FLOAT64_C( -76.80), SIMDE_FLOAT64_C( -90.20)),
926
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( 8746.88), SIMDE_FLOAT64_C( -174.98),
927
+ SIMDE_FLOAT64_C( -726.21), SIMDE_FLOAT64_C( 1313.55)) },
928
+ { simde_mm256_set_pd(SIMDE_FLOAT64_C( -20.60), SIMDE_FLOAT64_C( -46.30),
929
+ SIMDE_FLOAT64_C( 51.00), SIMDE_FLOAT64_C( 60.50)),
930
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( -26.30), SIMDE_FLOAT64_C( -65.50),
931
+ SIMDE_FLOAT64_C( -31.40), SIMDE_FLOAT64_C( -0.20)),
932
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( -79.80), SIMDE_FLOAT64_C( 98.80),
933
+ SIMDE_FLOAT64_C( 31.60), SIMDE_FLOAT64_C( -29.00)),
934
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( 621.58), SIMDE_FLOAT64_C( 2933.85),
935
+ SIMDE_FLOAT64_C(-1633.00), SIMDE_FLOAT64_C( 16.90)) },
936
+ { simde_mm256_set_pd(SIMDE_FLOAT64_C( 73.70), SIMDE_FLOAT64_C( -28.30),
937
+ SIMDE_FLOAT64_C( -1.90), SIMDE_FLOAT64_C( -61.50)),
938
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( -19.60), SIMDE_FLOAT64_C( -92.40),
939
+ SIMDE_FLOAT64_C( -22.30), SIMDE_FLOAT64_C( -53.90)),
940
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( -75.90), SIMDE_FLOAT64_C( 72.50),
941
+ SIMDE_FLOAT64_C( -50.10), SIMDE_FLOAT64_C( 18.00)),
942
+ simde_mm256_set_pd(SIMDE_FLOAT64_C(-1368.62), SIMDE_FLOAT64_C( 2542.42),
943
+ SIMDE_FLOAT64_C( 92.47), SIMDE_FLOAT64_C( 3296.85)) }
944
+ };
945
+
946
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
947
+ simde__m256d r = simde_mm256_fmsub_pd(test_vec[i].a, test_vec[i].b, test_vec[i].c);
948
+ simde_assert_m256d_close(r, test_vec[i].r, 1);
949
+ }
950
+
951
+ return MUNIT_OK;
952
+ }
953
+
954
+ static MunitResult
955
+ test_simde_mm_fmsub_ps(const MunitParameter params[], void* data) {
956
+ (void) params;
957
+ (void) data;
958
+
959
+ const struct {
960
+ simde__m128 a;
961
+ simde__m128 b;
962
+ simde__m128 c;
963
+ simde__m128 r;
964
+ } test_vec[8] = {
965
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( -36.50), SIMDE_FLOAT32_C( 13.70), SIMDE_FLOAT32_C( -3.10), SIMDE_FLOAT32_C( 21.40)),
966
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 80.60), SIMDE_FLOAT32_C( 11.30), SIMDE_FLOAT32_C( 96.80), SIMDE_FLOAT32_C( -38.20)),
967
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -8.50), SIMDE_FLOAT32_C( -28.20), SIMDE_FLOAT32_C( -26.80), SIMDE_FLOAT32_C( -95.00)),
968
+ simde_mm_set_ps(SIMDE_FLOAT32_C(-2933.40), SIMDE_FLOAT32_C( 183.01), SIMDE_FLOAT32_C( -273.28), SIMDE_FLOAT32_C( -722.48)) },
969
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( 59.20), SIMDE_FLOAT32_C( -6.20), SIMDE_FLOAT32_C( -52.90), SIMDE_FLOAT32_C( -75.70)),
970
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 61.80), SIMDE_FLOAT32_C( -76.10), SIMDE_FLOAT32_C( -87.70), SIMDE_FLOAT32_C( -40.50)),
971
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 34.20), SIMDE_FLOAT32_C( 37.10), SIMDE_FLOAT32_C( 7.30), SIMDE_FLOAT32_C( 67.80)),
972
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 3624.36), SIMDE_FLOAT32_C( 434.72), SIMDE_FLOAT32_C( 4632.03), SIMDE_FLOAT32_C( 2998.05)) },
973
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( -85.40), SIMDE_FLOAT32_C( 36.60), SIMDE_FLOAT32_C( -55.80), SIMDE_FLOAT32_C( 5.90)),
974
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -37.10), SIMDE_FLOAT32_C( 37.80), SIMDE_FLOAT32_C( -6.30), SIMDE_FLOAT32_C( 90.40)),
975
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 16.90), SIMDE_FLOAT32_C( -83.90), SIMDE_FLOAT32_C( 82.90), SIMDE_FLOAT32_C( 23.00)),
976
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 3151.44), SIMDE_FLOAT32_C( 1467.38), SIMDE_FLOAT32_C( 268.64), SIMDE_FLOAT32_C( 510.36)) },
977
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( 59.30), SIMDE_FLOAT32_C( 97.10), SIMDE_FLOAT32_C( -5.30), SIMDE_FLOAT32_C( -37.40)),
978
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -7.50), SIMDE_FLOAT32_C( 42.80), SIMDE_FLOAT32_C( -32.50), SIMDE_FLOAT32_C( -34.60)),
979
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 25.50), SIMDE_FLOAT32_C( 87.80), SIMDE_FLOAT32_C( 95.90), SIMDE_FLOAT32_C( -68.30)),
980
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -470.25), SIMDE_FLOAT32_C( 4068.08), SIMDE_FLOAT32_C( 76.35), SIMDE_FLOAT32_C( 1362.34)) },
981
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( -87.90), SIMDE_FLOAT32_C( -35.50), SIMDE_FLOAT32_C( -15.00), SIMDE_FLOAT32_C( 72.80)),
982
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 78.40), SIMDE_FLOAT32_C( 83.00), SIMDE_FLOAT32_C( 34.70), SIMDE_FLOAT32_C( -8.50)),
983
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 99.60), SIMDE_FLOAT32_C( 96.00), SIMDE_FLOAT32_C( 45.40), SIMDE_FLOAT32_C( -79.90)),
984
+ simde_mm_set_ps(SIMDE_FLOAT32_C(-6990.96), SIMDE_FLOAT32_C(-3042.50), SIMDE_FLOAT32_C( -565.90), SIMDE_FLOAT32_C( -538.90)) },
985
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( 26.30), SIMDE_FLOAT32_C( 69.80), SIMDE_FLOAT32_C( -48.50), SIMDE_FLOAT32_C( -58.50)),
986
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 85.00), SIMDE_FLOAT32_C( -97.40), SIMDE_FLOAT32_C( 16.90), SIMDE_FLOAT32_C( -37.30)),
987
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -22.30), SIMDE_FLOAT32_C( 21.90), SIMDE_FLOAT32_C( -79.20), SIMDE_FLOAT32_C( -99.20)),
988
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 2257.80), SIMDE_FLOAT32_C(-6820.42), SIMDE_FLOAT32_C( -740.45), SIMDE_FLOAT32_C( 2281.25)) },
989
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( 24.00), SIMDE_FLOAT32_C( 51.40), SIMDE_FLOAT32_C( -24.70), SIMDE_FLOAT32_C( -32.50)),
990
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 34.10), SIMDE_FLOAT32_C( 90.10), SIMDE_FLOAT32_C( 39.10), SIMDE_FLOAT32_C( -33.10)),
991
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 63.90), SIMDE_FLOAT32_C( -54.20), SIMDE_FLOAT32_C( -27.60), SIMDE_FLOAT32_C( 31.70)),
992
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 754.50), SIMDE_FLOAT32_C( 4685.34), SIMDE_FLOAT32_C( -938.17), SIMDE_FLOAT32_C( 1044.05)) },
993
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( 45.70), SIMDE_FLOAT32_C( -95.60), SIMDE_FLOAT32_C( 14.60), SIMDE_FLOAT32_C( -3.40)),
994
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -90.50), SIMDE_FLOAT32_C( -20.20), SIMDE_FLOAT32_C( 91.40), SIMDE_FLOAT32_C( 25.10)),
995
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -26.90), SIMDE_FLOAT32_C( 29.30), SIMDE_FLOAT32_C( 77.50), SIMDE_FLOAT32_C( -80.00)),
996
+ simde_mm_set_ps(SIMDE_FLOAT32_C(-4108.95), SIMDE_FLOAT32_C( 1901.82), SIMDE_FLOAT32_C( 1256.94), SIMDE_FLOAT32_C( -5.34)) }
997
+ };
998
+
999
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
1000
+ simde__m128 r = simde_mm_fmsub_ps(test_vec[i].a, test_vec[i].b, test_vec[i].c);
1001
+ simde_assert_m128_close(r, test_vec[i].r, 1);
1002
+ }
1003
+
1004
+ return MUNIT_OK;
1005
+ }
1006
+
1007
+ static MunitResult
1008
+ test_simde_mm256_fmsub_ps(const MunitParameter params[], void* data) {
1009
+ (void) params;
1010
+ (void) data;
1011
+
1012
+ const struct {
1013
+ simde__m256 a;
1014
+ simde__m256 b;
1015
+ simde__m256 c;
1016
+ simde__m256 r;
1017
+ } test_vec[8] = {
1018
+ { simde_mm256_set_ps(SIMDE_FLOAT32_C( 71.60), SIMDE_FLOAT32_C( 70.70),
1019
+ SIMDE_FLOAT32_C( 40.60), SIMDE_FLOAT32_C( -9.30),
1020
+ SIMDE_FLOAT32_C( -79.10), SIMDE_FLOAT32_C( 52.30),
1021
+ SIMDE_FLOAT32_C( -67.90), SIMDE_FLOAT32_C( 25.70)),
1022
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( -51.50), SIMDE_FLOAT32_C( 43.80),
1023
+ SIMDE_FLOAT32_C( 41.70), SIMDE_FLOAT32_C( -77.20),
1024
+ SIMDE_FLOAT32_C( -5.00), SIMDE_FLOAT32_C( 96.70),
1025
+ SIMDE_FLOAT32_C( -13.50), SIMDE_FLOAT32_C( -2.70)),
1026
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( -28.70), SIMDE_FLOAT32_C( -28.30),
1027
+ SIMDE_FLOAT32_C( 1.80), SIMDE_FLOAT32_C( -81.10),
1028
+ SIMDE_FLOAT32_C( -82.10), SIMDE_FLOAT32_C( -69.80),
1029
+ SIMDE_FLOAT32_C( 42.10), SIMDE_FLOAT32_C( 74.70)),
1030
+ simde_mm256_set_ps(SIMDE_FLOAT32_C(-3658.70), SIMDE_FLOAT32_C( 3124.96),
1031
+ SIMDE_FLOAT32_C( 1691.22), SIMDE_FLOAT32_C( 799.06),
1032
+ SIMDE_FLOAT32_C( 477.60), SIMDE_FLOAT32_C( 5127.21),
1033
+ SIMDE_FLOAT32_C( 874.55), SIMDE_FLOAT32_C( -144.09)) },
1034
+ { simde_mm256_set_ps(SIMDE_FLOAT32_C( -49.00), SIMDE_FLOAT32_C( -78.70),
1035
+ SIMDE_FLOAT32_C( -72.10), SIMDE_FLOAT32_C( 26.10),
1036
+ SIMDE_FLOAT32_C( -91.90), SIMDE_FLOAT32_C( 1.40),
1037
+ SIMDE_FLOAT32_C( 89.80), SIMDE_FLOAT32_C( 94.20)),
1038
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( -13.70), SIMDE_FLOAT32_C( 0.70),
1039
+ SIMDE_FLOAT32_C( 57.80), SIMDE_FLOAT32_C( 33.00),
1040
+ SIMDE_FLOAT32_C( -83.50), SIMDE_FLOAT32_C( -8.10),
1041
+ SIMDE_FLOAT32_C( 91.30), SIMDE_FLOAT32_C( 65.20)),
1042
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( -67.90), SIMDE_FLOAT32_C( -56.40),
1043
+ SIMDE_FLOAT32_C( 5.90), SIMDE_FLOAT32_C( 2.40),
1044
+ SIMDE_FLOAT32_C( 91.80), SIMDE_FLOAT32_C( 50.80),
1045
+ SIMDE_FLOAT32_C( 64.70), SIMDE_FLOAT32_C( -56.10)),
1046
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( 739.20), SIMDE_FLOAT32_C( 1.31),
1047
+ SIMDE_FLOAT32_C(-4173.28), SIMDE_FLOAT32_C( 858.90),
1048
+ SIMDE_FLOAT32_C( 7581.85), SIMDE_FLOAT32_C( -62.14),
1049
+ SIMDE_FLOAT32_C( 8134.04), SIMDE_FLOAT32_C( 6197.94)) },
1050
+ { simde_mm256_set_ps(SIMDE_FLOAT32_C( 40.00), SIMDE_FLOAT32_C( -5.30),
1051
+ SIMDE_FLOAT32_C( 85.00), SIMDE_FLOAT32_C( 83.70),
1052
+ SIMDE_FLOAT32_C( 96.80), SIMDE_FLOAT32_C( -59.70),
1053
+ SIMDE_FLOAT32_C( -72.50), SIMDE_FLOAT32_C( -8.10)),
1054
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( 77.50), SIMDE_FLOAT32_C( 50.00),
1055
+ SIMDE_FLOAT32_C( 72.40), SIMDE_FLOAT32_C( 98.40),
1056
+ SIMDE_FLOAT32_C( 69.10), SIMDE_FLOAT32_C( 35.80),
1057
+ SIMDE_FLOAT32_C( -92.90), SIMDE_FLOAT32_C( 63.70)),
1058
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( 73.80), SIMDE_FLOAT32_C( -94.30),
1059
+ SIMDE_FLOAT32_C( -79.50), SIMDE_FLOAT32_C( 64.60),
1060
+ SIMDE_FLOAT32_C( 63.40), SIMDE_FLOAT32_C( -65.00),
1061
+ SIMDE_FLOAT32_C( 75.20), SIMDE_FLOAT32_C( 48.70)),
1062
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( 3026.20), SIMDE_FLOAT32_C( -170.70),
1063
+ SIMDE_FLOAT32_C( 6233.50), SIMDE_FLOAT32_C( 8171.48),
1064
+ SIMDE_FLOAT32_C( 6625.48), SIMDE_FLOAT32_C(-2072.26),
1065
+ SIMDE_FLOAT32_C( 6660.05), SIMDE_FLOAT32_C( -564.67)) },
1066
+ { simde_mm256_set_ps(SIMDE_FLOAT32_C( -82.20), SIMDE_FLOAT32_C( 93.30),
1067
+ SIMDE_FLOAT32_C( 9.70), SIMDE_FLOAT32_C( -2.70),
1068
+ SIMDE_FLOAT32_C( 86.00), SIMDE_FLOAT32_C( -20.80),
1069
+ SIMDE_FLOAT32_C( 67.70), SIMDE_FLOAT32_C( -47.20)),
1070
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( -83.70), SIMDE_FLOAT32_C( 12.00),
1071
+ SIMDE_FLOAT32_C( 23.10), SIMDE_FLOAT32_C( -42.00),
1072
+ SIMDE_FLOAT32_C( 46.30), SIMDE_FLOAT32_C( 48.20),
1073
+ SIMDE_FLOAT32_C( 86.90), SIMDE_FLOAT32_C( -91.50)),
1074
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( -48.60), SIMDE_FLOAT32_C( 92.30),
1075
+ SIMDE_FLOAT32_C( -12.70), SIMDE_FLOAT32_C( -48.20),
1076
+ SIMDE_FLOAT32_C( 60.90), SIMDE_FLOAT32_C( 43.20),
1077
+ SIMDE_FLOAT32_C( -71.30), SIMDE_FLOAT32_C( -56.60)),
1078
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( 6928.74), SIMDE_FLOAT32_C( 1027.30),
1079
+ SIMDE_FLOAT32_C( 236.77), SIMDE_FLOAT32_C( 161.60),
1080
+ SIMDE_FLOAT32_C( 3920.90), SIMDE_FLOAT32_C(-1045.76),
1081
+ SIMDE_FLOAT32_C( 5954.43), SIMDE_FLOAT32_C( 4375.40)) },
1082
+ { simde_mm256_set_ps(SIMDE_FLOAT32_C( 32.60), SIMDE_FLOAT32_C( 90.30),
1083
+ SIMDE_FLOAT32_C( -31.90), SIMDE_FLOAT32_C( 33.60),
1084
+ SIMDE_FLOAT32_C( 47.40), SIMDE_FLOAT32_C( 49.30),
1085
+ SIMDE_FLOAT32_C( -73.00), SIMDE_FLOAT32_C( 55.80)),
1086
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( -72.60), SIMDE_FLOAT32_C( 34.70),
1087
+ SIMDE_FLOAT32_C( -8.30), SIMDE_FLOAT32_C( -47.40),
1088
+ SIMDE_FLOAT32_C( -91.00), SIMDE_FLOAT32_C( -99.10),
1089
+ SIMDE_FLOAT32_C( -84.60), SIMDE_FLOAT32_C( -13.50)),
1090
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( -84.20), SIMDE_FLOAT32_C( 35.00),
1091
+ SIMDE_FLOAT32_C( -58.10), SIMDE_FLOAT32_C( 81.70),
1092
+ SIMDE_FLOAT32_C( 1.20), SIMDE_FLOAT32_C( -33.20),
1093
+ SIMDE_FLOAT32_C( 36.00), SIMDE_FLOAT32_C( -80.90)),
1094
+ simde_mm256_set_ps(SIMDE_FLOAT32_C(-2282.56), SIMDE_FLOAT32_C( 3098.41),
1095
+ SIMDE_FLOAT32_C( 322.87), SIMDE_FLOAT32_C(-1674.34),
1096
+ SIMDE_FLOAT32_C(-4314.60), SIMDE_FLOAT32_C(-4852.43),
1097
+ SIMDE_FLOAT32_C( 6139.80), SIMDE_FLOAT32_C( -672.40)) },
1098
+ { simde_mm256_set_ps(SIMDE_FLOAT32_C( -58.90), SIMDE_FLOAT32_C( 53.10),
1099
+ SIMDE_FLOAT32_C( -76.60), SIMDE_FLOAT32_C( 83.00),
1100
+ SIMDE_FLOAT32_C( 91.20), SIMDE_FLOAT32_C( -33.50),
1101
+ SIMDE_FLOAT32_C( -65.20), SIMDE_FLOAT32_C( -55.00)),
1102
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( -47.40), SIMDE_FLOAT32_C( -20.10),
1103
+ SIMDE_FLOAT32_C( -89.40), SIMDE_FLOAT32_C( 87.90),
1104
+ SIMDE_FLOAT32_C( -65.50), SIMDE_FLOAT32_C( -20.70),
1105
+ SIMDE_FLOAT32_C( 88.30), SIMDE_FLOAT32_C( 20.40)),
1106
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( 50.40), SIMDE_FLOAT32_C( 75.40),
1107
+ SIMDE_FLOAT32_C( 79.80), SIMDE_FLOAT32_C( 5.10),
1108
+ SIMDE_FLOAT32_C( -6.50), SIMDE_FLOAT32_C( -47.90),
1109
+ SIMDE_FLOAT32_C( 48.50), SIMDE_FLOAT32_C( -69.90)),
1110
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( 2741.46), SIMDE_FLOAT32_C(-1142.71),
1111
+ SIMDE_FLOAT32_C( 6768.24), SIMDE_FLOAT32_C( 7290.60),
1112
+ SIMDE_FLOAT32_C(-5967.10), SIMDE_FLOAT32_C( 741.35),
1113
+ SIMDE_FLOAT32_C(-5805.66), SIMDE_FLOAT32_C(-1052.10)) },
1114
+ { simde_mm256_set_ps(SIMDE_FLOAT32_C( 8.30), SIMDE_FLOAT32_C( 22.80),
1115
+ SIMDE_FLOAT32_C( -55.20), SIMDE_FLOAT32_C( -62.40),
1116
+ SIMDE_FLOAT32_C( -29.10), SIMDE_FLOAT32_C( 56.20),
1117
+ SIMDE_FLOAT32_C( 96.20), SIMDE_FLOAT32_C( 45.90)),
1118
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( 9.40), SIMDE_FLOAT32_C( -58.60),
1119
+ SIMDE_FLOAT32_C( -71.50), SIMDE_FLOAT32_C( 52.70),
1120
+ SIMDE_FLOAT32_C( -96.40), SIMDE_FLOAT32_C( 75.70),
1121
+ SIMDE_FLOAT32_C( -3.70), SIMDE_FLOAT32_C( 35.60)),
1122
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( -87.20), SIMDE_FLOAT32_C( -73.80),
1123
+ SIMDE_FLOAT32_C( -51.80), SIMDE_FLOAT32_C( 49.30),
1124
+ SIMDE_FLOAT32_C( 9.90), SIMDE_FLOAT32_C( 32.40),
1125
+ SIMDE_FLOAT32_C( -44.20), SIMDE_FLOAT32_C( 88.50)),
1126
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( 165.22), SIMDE_FLOAT32_C(-1262.28),
1127
+ SIMDE_FLOAT32_C( 3998.60), SIMDE_FLOAT32_C(-3337.78),
1128
+ SIMDE_FLOAT32_C( 2795.34), SIMDE_FLOAT32_C( 4221.94),
1129
+ SIMDE_FLOAT32_C( -311.74), SIMDE_FLOAT32_C( 1545.54)) },
1130
+ { simde_mm256_set_ps(SIMDE_FLOAT32_C( -71.10), SIMDE_FLOAT32_C( -36.70),
1131
+ SIMDE_FLOAT32_C( 2.00), SIMDE_FLOAT32_C( -19.80),
1132
+ SIMDE_FLOAT32_C( -33.20), SIMDE_FLOAT32_C( 94.30),
1133
+ SIMDE_FLOAT32_C( 1.20), SIMDE_FLOAT32_C( 43.50)),
1134
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( 91.40), SIMDE_FLOAT32_C( 40.00),
1135
+ SIMDE_FLOAT32_C( 26.00), SIMDE_FLOAT32_C( 80.90),
1136
+ SIMDE_FLOAT32_C( -92.20), SIMDE_FLOAT32_C( -86.10),
1137
+ SIMDE_FLOAT32_C( 71.10), SIMDE_FLOAT32_C( 10.10)),
1138
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( 29.80), SIMDE_FLOAT32_C( -33.80),
1139
+ SIMDE_FLOAT32_C( -52.50), SIMDE_FLOAT32_C( 52.00),
1140
+ SIMDE_FLOAT32_C( -20.10), SIMDE_FLOAT32_C( -49.80),
1141
+ SIMDE_FLOAT32_C( 36.10), SIMDE_FLOAT32_C( 37.00)),
1142
+ simde_mm256_set_ps(SIMDE_FLOAT32_C(-6528.34), SIMDE_FLOAT32_C(-1434.20),
1143
+ SIMDE_FLOAT32_C( 104.50), SIMDE_FLOAT32_C(-1653.82),
1144
+ SIMDE_FLOAT32_C( 3081.14), SIMDE_FLOAT32_C(-8069.43),
1145
+ SIMDE_FLOAT32_C( 49.22), SIMDE_FLOAT32_C( 402.35)) }
1146
+ };
1147
+
1148
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
1149
+ simde__m256 r = simde_mm256_fmsub_ps(test_vec[i].a, test_vec[i].b, test_vec[i].c);
1150
+ simde_assert_m256_close(r, test_vec[i].r, 1);
1151
+ }
1152
+
1153
+ return MUNIT_OK;
1154
+ }
1155
+
1156
+ static MunitResult
1157
+ test_simde_mm_fmsub_sd(const MunitParameter params[], void* data) {
1158
+ (void) params;
1159
+ (void) data;
1160
+
1161
+ const struct {
1162
+ simde__m128d a;
1163
+ simde__m128d b;
1164
+ simde__m128d c;
1165
+ simde__m128d r;
1166
+ } test_vec[8] = {
1167
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( 43.20), SIMDE_FLOAT64_C( -60.20)),
1168
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -12.80), SIMDE_FLOAT64_C( 56.50)),
1169
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -27.10), SIMDE_FLOAT64_C( 60.10)),
1170
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 43.20), SIMDE_FLOAT64_C(-3461.40)) },
1171
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( -79.30), SIMDE_FLOAT64_C( 88.90)),
1172
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 44.30), SIMDE_FLOAT64_C( 37.40)),
1173
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 77.70), SIMDE_FLOAT64_C( 22.00)),
1174
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -79.30), SIMDE_FLOAT64_C( 3302.86)) },
1175
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( -1.70), SIMDE_FLOAT64_C( -49.80)),
1176
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 87.10), SIMDE_FLOAT64_C( -41.00)),
1177
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -71.70), SIMDE_FLOAT64_C( 16.20)),
1178
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -1.70), SIMDE_FLOAT64_C( 2025.60)) },
1179
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( -81.20), SIMDE_FLOAT64_C( 22.00)),
1180
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -90.90), SIMDE_FLOAT64_C( 95.10)),
1181
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -81.00), SIMDE_FLOAT64_C( -21.00)),
1182
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -81.20), SIMDE_FLOAT64_C( 2113.20)) },
1183
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( 86.00), SIMDE_FLOAT64_C( 69.40)),
1184
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -68.40), SIMDE_FLOAT64_C( -83.70)),
1185
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 95.80), SIMDE_FLOAT64_C( 94.30)),
1186
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 86.00), SIMDE_FLOAT64_C(-5903.08)) },
1187
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( -19.30), SIMDE_FLOAT64_C( -49.30)),
1188
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -62.80), SIMDE_FLOAT64_C( 42.00)),
1189
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 30.00), SIMDE_FLOAT64_C( -69.00)),
1190
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -19.30), SIMDE_FLOAT64_C(-2001.60)) },
1191
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( 23.40), SIMDE_FLOAT64_C( -19.50)),
1192
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 85.50), SIMDE_FLOAT64_C( 56.90)),
1193
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 27.00), SIMDE_FLOAT64_C( -47.20)),
1194
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 23.40), SIMDE_FLOAT64_C(-1062.35)) },
1195
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( -59.40), SIMDE_FLOAT64_C( 23.00)),
1196
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -67.50), SIMDE_FLOAT64_C( 79.20)),
1197
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 98.40), SIMDE_FLOAT64_C( -48.80)),
1198
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -59.40), SIMDE_FLOAT64_C( 1870.40)) }
1199
+ };
1200
+
1201
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
1202
+ simde__m128d r = simde_mm_fmsub_sd(test_vec[i].a, test_vec[i].b, test_vec[i].c);
1203
+ simde_assert_m128d_close(r, test_vec[i].r, 1);
1204
+ }
1205
+
1206
+ return MUNIT_OK;
1207
+ }
1208
+
1209
+ static MunitResult
1210
+ test_simde_mm_fmsub_ss(const MunitParameter params[], void* data) {
1211
+ (void) params;
1212
+ (void) data;
1213
+
1214
+ const struct {
1215
+ simde__m128 a;
1216
+ simde__m128 b;
1217
+ simde__m128 c;
1218
+ simde__m128 r;
1219
+ } test_vec[8] = {
1220
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( 81.20), SIMDE_FLOAT32_C( 26.30), SIMDE_FLOAT32_C( 21.90), SIMDE_FLOAT32_C( 41.80)),
1221
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 37.70), SIMDE_FLOAT32_C( 61.40), SIMDE_FLOAT32_C( 87.60), SIMDE_FLOAT32_C( -37.20)),
1222
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 85.80), SIMDE_FLOAT32_C( -48.50), SIMDE_FLOAT32_C( 52.10), SIMDE_FLOAT32_C( 67.70)),
1223
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 81.20), SIMDE_FLOAT32_C( 26.30), SIMDE_FLOAT32_C( 21.90), SIMDE_FLOAT32_C(-1622.66)) },
1224
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( -23.60), SIMDE_FLOAT32_C( -82.80), SIMDE_FLOAT32_C( 55.80), SIMDE_FLOAT32_C( -90.30)),
1225
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 53.10), SIMDE_FLOAT32_C( -75.20), SIMDE_FLOAT32_C( -26.00), SIMDE_FLOAT32_C( 93.80)),
1226
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 47.50), SIMDE_FLOAT32_C( 39.90), SIMDE_FLOAT32_C( -49.20), SIMDE_FLOAT32_C( -86.90)),
1227
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -23.60), SIMDE_FLOAT32_C( -82.80), SIMDE_FLOAT32_C( 55.80), SIMDE_FLOAT32_C(-8383.24)) },
1228
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( 0.90), SIMDE_FLOAT32_C( -99.10), SIMDE_FLOAT32_C( 26.00), SIMDE_FLOAT32_C( 32.50)),
1229
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 29.60), SIMDE_FLOAT32_C( -93.20), SIMDE_FLOAT32_C( -96.10), SIMDE_FLOAT32_C( 87.50)),
1230
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 76.20), SIMDE_FLOAT32_C( -98.50), SIMDE_FLOAT32_C( 4.10), SIMDE_FLOAT32_C( -66.50)),
1231
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 0.90), SIMDE_FLOAT32_C( -99.10), SIMDE_FLOAT32_C( 26.00), SIMDE_FLOAT32_C( 2910.25)) },
1232
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( -58.80), SIMDE_FLOAT32_C( 0.70), SIMDE_FLOAT32_C( -50.10), SIMDE_FLOAT32_C( -58.60)),
1233
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 53.70), SIMDE_FLOAT32_C( -83.00), SIMDE_FLOAT32_C( -66.70), SIMDE_FLOAT32_C( 96.60)),
1234
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -97.40), SIMDE_FLOAT32_C( 97.80), SIMDE_FLOAT32_C( 93.40), SIMDE_FLOAT32_C( -82.50)),
1235
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -58.80), SIMDE_FLOAT32_C( 0.70), SIMDE_FLOAT32_C( -50.10), SIMDE_FLOAT32_C(-5578.26)) },
1236
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( 26.20), SIMDE_FLOAT32_C( 0.50), SIMDE_FLOAT32_C( 53.40), SIMDE_FLOAT32_C( 40.40)),
1237
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -60.30), SIMDE_FLOAT32_C( -94.00), SIMDE_FLOAT32_C( 14.10), SIMDE_FLOAT32_C( -94.50)),
1238
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 75.80), SIMDE_FLOAT32_C( 16.70), SIMDE_FLOAT32_C( -3.80), SIMDE_FLOAT32_C( -98.50)),
1239
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 26.20), SIMDE_FLOAT32_C( 0.50), SIMDE_FLOAT32_C( 53.40), SIMDE_FLOAT32_C(-3719.30)) },
1240
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( 6.90), SIMDE_FLOAT32_C( 37.30), SIMDE_FLOAT32_C( 95.60), SIMDE_FLOAT32_C( 26.20)),
1241
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 6.20), SIMDE_FLOAT32_C( 51.70), SIMDE_FLOAT32_C( -27.80), SIMDE_FLOAT32_C( 35.60)),
1242
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 96.60), SIMDE_FLOAT32_C( 16.30), SIMDE_FLOAT32_C( -87.40), SIMDE_FLOAT32_C( 51.00)),
1243
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 6.90), SIMDE_FLOAT32_C( 37.30), SIMDE_FLOAT32_C( 95.60), SIMDE_FLOAT32_C( 881.72)) },
1244
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( -41.80), SIMDE_FLOAT32_C( -50.90), SIMDE_FLOAT32_C( 94.30), SIMDE_FLOAT32_C( 92.50)),
1245
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -51.70), SIMDE_FLOAT32_C( 66.70), SIMDE_FLOAT32_C( 35.70), SIMDE_FLOAT32_C( 84.90)),
1246
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -89.60), SIMDE_FLOAT32_C( -35.50), SIMDE_FLOAT32_C( -45.20), SIMDE_FLOAT32_C( -87.60)),
1247
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -41.80), SIMDE_FLOAT32_C( -50.90), SIMDE_FLOAT32_C( 94.30), SIMDE_FLOAT32_C( 7940.85)) },
1248
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( 60.00), SIMDE_FLOAT32_C( 45.70), SIMDE_FLOAT32_C( 16.60), SIMDE_FLOAT32_C( 40.70)),
1249
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -13.60), SIMDE_FLOAT32_C( -11.50), SIMDE_FLOAT32_C( -61.10), SIMDE_FLOAT32_C( -64.20)),
1250
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 96.70), SIMDE_FLOAT32_C( -80.10), SIMDE_FLOAT32_C( 37.00), SIMDE_FLOAT32_C( 74.70)),
1251
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 60.00), SIMDE_FLOAT32_C( 45.70), SIMDE_FLOAT32_C( 16.60), SIMDE_FLOAT32_C(-2687.64)) }
1252
+ };
1253
+
1254
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
1255
+ simde__m128 r = simde_mm_fmsub_ss(test_vec[i].a, test_vec[i].b, test_vec[i].c);
1256
+ simde_assert_m128_close(r, test_vec[i].r, 1);
1257
+ }
1258
+
1259
+ return MUNIT_OK;
1260
+ }
1261
+
1262
+ static MunitResult
1263
+ test_simde_mm_fmsubadd_pd(const MunitParameter params[], void* data) {
1264
+ (void) params;
1265
+ (void) data;
1266
+
1267
+ const struct {
1268
+ simde__m128d a;
1269
+ simde__m128d b;
1270
+ simde__m128d c;
1271
+ simde__m128d r;
1272
+ } test_vec[8] = {
1273
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( -50.60), SIMDE_FLOAT64_C( -67.20)),
1274
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -79.80), SIMDE_FLOAT64_C( -83.00)),
1275
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 34.70), SIMDE_FLOAT64_C( -10.00)),
1276
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 4003.18), SIMDE_FLOAT64_C( 5567.60)) },
1277
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( -31.80), SIMDE_FLOAT64_C( -73.70)),
1278
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -80.50), SIMDE_FLOAT64_C( 26.40)),
1279
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -10.90), SIMDE_FLOAT64_C( -36.90)),
1280
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 2570.80), SIMDE_FLOAT64_C(-1982.58)) },
1281
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( 7.90), SIMDE_FLOAT64_C( -20.50)),
1282
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 91.90), SIMDE_FLOAT64_C( -31.80)),
1283
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -17.90), SIMDE_FLOAT64_C( -72.60)),
1284
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 743.91), SIMDE_FLOAT64_C( 579.30)) },
1285
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( 0.90), SIMDE_FLOAT64_C( 20.80)),
1286
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 75.20), SIMDE_FLOAT64_C( -63.30)),
1287
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 33.00), SIMDE_FLOAT64_C( 76.30)),
1288
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 34.68), SIMDE_FLOAT64_C(-1240.34)) },
1289
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( 45.60), SIMDE_FLOAT64_C( -62.50)),
1290
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -70.50), SIMDE_FLOAT64_C( 21.00)),
1291
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -20.00), SIMDE_FLOAT64_C( 73.20)),
1292
+ simde_mm_set_pd(SIMDE_FLOAT64_C(-3194.80), SIMDE_FLOAT64_C(-1239.30)) },
1293
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( -10.30), SIMDE_FLOAT64_C( -71.50)),
1294
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 63.70), SIMDE_FLOAT64_C( -56.70)),
1295
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -15.40), SIMDE_FLOAT64_C( 29.10)),
1296
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -640.71), SIMDE_FLOAT64_C( 4083.15)) },
1297
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( -1.00), SIMDE_FLOAT64_C( -6.90)),
1298
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -84.60), SIMDE_FLOAT64_C( -37.70)),
1299
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 53.20), SIMDE_FLOAT64_C( -28.50)),
1300
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 31.40), SIMDE_FLOAT64_C( 231.63)) },
1301
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( -13.10), SIMDE_FLOAT64_C( -9.00)),
1302
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -34.00), SIMDE_FLOAT64_C( -63.90)),
1303
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 31.60), SIMDE_FLOAT64_C( -13.80)),
1304
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 413.80), SIMDE_FLOAT64_C( 561.30)) }
1305
+ };
1306
+
1307
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
1308
+ simde__m128d r = simde_mm_fmsubadd_pd(test_vec[i].a, test_vec[i].b, test_vec[i].c);
1309
+ simde_assert_m128d_close(r, test_vec[i].r, 1);
1310
+ }
1311
+
1312
+ return MUNIT_OK;
1313
+ }
1314
+
1315
+ static MunitResult
1316
+ test_simde_mm256_fmsubadd_pd(const MunitParameter params[], void* data) {
1317
+ (void) params;
1318
+ (void) data;
1319
+
1320
+ const struct {
1321
+ simde__m256d a;
1322
+ simde__m256d b;
1323
+ simde__m256d c;
1324
+ simde__m256d r;
1325
+ } test_vec[8] = {
1326
+ { simde_mm256_set_pd(SIMDE_FLOAT64_C( -49.40), SIMDE_FLOAT64_C( -57.60),
1327
+ SIMDE_FLOAT64_C( -73.20), SIMDE_FLOAT64_C( -70.10)),
1328
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( -0.10), SIMDE_FLOAT64_C( 46.20),
1329
+ SIMDE_FLOAT64_C( -46.70), SIMDE_FLOAT64_C( -70.80)),
1330
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( 79.00), SIMDE_FLOAT64_C( -79.60),
1331
+ SIMDE_FLOAT64_C( 19.80), SIMDE_FLOAT64_C( -16.00)),
1332
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( -74.06), SIMDE_FLOAT64_C(-2740.72),
1333
+ SIMDE_FLOAT64_C( 3398.64), SIMDE_FLOAT64_C( 4947.08)) },
1334
+ { simde_mm256_set_pd(SIMDE_FLOAT64_C( -12.30), SIMDE_FLOAT64_C( 53.50),
1335
+ SIMDE_FLOAT64_C( -97.80), SIMDE_FLOAT64_C( -85.20)),
1336
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( 67.10), SIMDE_FLOAT64_C( -30.10),
1337
+ SIMDE_FLOAT64_C( -0.30), SIMDE_FLOAT64_C( -23.80)),
1338
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( 49.80), SIMDE_FLOAT64_C( 87.30),
1339
+ SIMDE_FLOAT64_C( -23.10), SIMDE_FLOAT64_C( 15.90)),
1340
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( -875.13), SIMDE_FLOAT64_C(-1523.05),
1341
+ SIMDE_FLOAT64_C( 52.44), SIMDE_FLOAT64_C( 2043.66)) },
1342
+ { simde_mm256_set_pd(SIMDE_FLOAT64_C( -27.20), SIMDE_FLOAT64_C( -72.40),
1343
+ SIMDE_FLOAT64_C( 53.20), SIMDE_FLOAT64_C( -9.50)),
1344
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( -71.40), SIMDE_FLOAT64_C( 0.20),
1345
+ SIMDE_FLOAT64_C( -61.10), SIMDE_FLOAT64_C( -97.30)),
1346
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( 85.60), SIMDE_FLOAT64_C( 27.60),
1347
+ SIMDE_FLOAT64_C( 19.30), SIMDE_FLOAT64_C( 46.60)),
1348
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( 1856.48), SIMDE_FLOAT64_C( 13.12),
1349
+ SIMDE_FLOAT64_C(-3269.82), SIMDE_FLOAT64_C( 970.95)) },
1350
+ { simde_mm256_set_pd(SIMDE_FLOAT64_C( 42.40), SIMDE_FLOAT64_C( -47.00),
1351
+ SIMDE_FLOAT64_C( 57.40), SIMDE_FLOAT64_C( -79.50)),
1352
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( -85.60), SIMDE_FLOAT64_C( -55.10),
1353
+ SIMDE_FLOAT64_C( 8.90), SIMDE_FLOAT64_C( -9.70)),
1354
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( 78.80), SIMDE_FLOAT64_C( 18.80),
1355
+ SIMDE_FLOAT64_C( -90.80), SIMDE_FLOAT64_C( 46.00)),
1356
+ simde_mm256_set_pd(SIMDE_FLOAT64_C(-3708.24), SIMDE_FLOAT64_C( 2608.50),
1357
+ SIMDE_FLOAT64_C( 601.66), SIMDE_FLOAT64_C( 817.15)) },
1358
+ { simde_mm256_set_pd(SIMDE_FLOAT64_C( 34.00), SIMDE_FLOAT64_C( 57.40),
1359
+ SIMDE_FLOAT64_C( 76.30), SIMDE_FLOAT64_C( 99.20)),
1360
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( 74.30), SIMDE_FLOAT64_C( 64.30),
1361
+ SIMDE_FLOAT64_C( -88.20), SIMDE_FLOAT64_C( -42.40)),
1362
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( -20.50), SIMDE_FLOAT64_C( 98.80),
1363
+ SIMDE_FLOAT64_C( -81.30), SIMDE_FLOAT64_C( 9.70)),
1364
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( 2546.70), SIMDE_FLOAT64_C( 3789.62),
1365
+ SIMDE_FLOAT64_C(-6648.36), SIMDE_FLOAT64_C(-4196.38)) },
1366
+ { simde_mm256_set_pd(SIMDE_FLOAT64_C( 26.30), SIMDE_FLOAT64_C( -10.40),
1367
+ SIMDE_FLOAT64_C( -16.90), SIMDE_FLOAT64_C( -91.70)),
1368
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( -5.60), SIMDE_FLOAT64_C( -40.40),
1369
+ SIMDE_FLOAT64_C( 57.90), SIMDE_FLOAT64_C( 93.60)),
1370
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( -66.10), SIMDE_FLOAT64_C( -60.00),
1371
+ SIMDE_FLOAT64_C( -42.50), SIMDE_FLOAT64_C( -45.60)),
1372
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( -81.18), SIMDE_FLOAT64_C( 360.16),
1373
+ SIMDE_FLOAT64_C( -936.01), SIMDE_FLOAT64_C(-8628.72)) },
1374
+ { simde_mm256_set_pd(SIMDE_FLOAT64_C( -77.70), SIMDE_FLOAT64_C( 79.90),
1375
+ SIMDE_FLOAT64_C( 16.20), SIMDE_FLOAT64_C( -77.30)),
1376
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( 18.40), SIMDE_FLOAT64_C( 71.60),
1377
+ SIMDE_FLOAT64_C( -95.70), SIMDE_FLOAT64_C( -21.80)),
1378
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( -83.90), SIMDE_FLOAT64_C( 14.30),
1379
+ SIMDE_FLOAT64_C( -44.90), SIMDE_FLOAT64_C( 72.40)),
1380
+ simde_mm256_set_pd(SIMDE_FLOAT64_C(-1345.78), SIMDE_FLOAT64_C( 5735.14),
1381
+ SIMDE_FLOAT64_C(-1505.44), SIMDE_FLOAT64_C( 1757.54)) },
1382
+ { simde_mm256_set_pd(SIMDE_FLOAT64_C( 68.20), SIMDE_FLOAT64_C( 18.60),
1383
+ SIMDE_FLOAT64_C( 38.50), SIMDE_FLOAT64_C( 98.50)),
1384
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( -61.10), SIMDE_FLOAT64_C( -31.60),
1385
+ SIMDE_FLOAT64_C( 70.50), SIMDE_FLOAT64_C( 85.20)),
1386
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( 94.30), SIMDE_FLOAT64_C( 91.40),
1387
+ SIMDE_FLOAT64_C( -28.70), SIMDE_FLOAT64_C( 64.60)),
1388
+ simde_mm256_set_pd(SIMDE_FLOAT64_C(-4261.32), SIMDE_FLOAT64_C( -496.36),
1389
+ SIMDE_FLOAT64_C( 2742.95), SIMDE_FLOAT64_C( 8456.80)) }
1390
+ };
1391
+
1392
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
1393
+ simde__m256d r = simde_mm256_fmsubadd_pd(test_vec[i].a, test_vec[i].b, test_vec[i].c);
1394
+ simde_assert_m256d_close(r, test_vec[i].r, 1);
1395
+ }
1396
+
1397
+ return MUNIT_OK;
1398
+ }
1399
+
1400
+ static MunitResult
1401
+ test_simde_mm_fmsubadd_ps(const MunitParameter params[], void* data) {
1402
+ (void) params;
1403
+ (void) data;
1404
+
1405
+ const struct {
1406
+ simde__m128 a;
1407
+ simde__m128 b;
1408
+ simde__m128 c;
1409
+ simde__m128 r;
1410
+ } test_vec[8] = {
1411
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( -29.90), SIMDE_FLOAT32_C( -50.10), SIMDE_FLOAT32_C( 13.10), SIMDE_FLOAT32_C( 52.30)),
1412
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -63.90), SIMDE_FLOAT32_C( -96.40), SIMDE_FLOAT32_C( 84.20), SIMDE_FLOAT32_C( -48.20)),
1413
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 53.80), SIMDE_FLOAT32_C( -3.40), SIMDE_FLOAT32_C( 13.90), SIMDE_FLOAT32_C( -46.00)),
1414
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 1856.81), SIMDE_FLOAT32_C( 4826.24), SIMDE_FLOAT32_C( 1089.12), SIMDE_FLOAT32_C(-2566.86)) },
1415
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( 35.60), SIMDE_FLOAT32_C( 8.10), SIMDE_FLOAT32_C( -35.10), SIMDE_FLOAT32_C( 22.30)),
1416
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 77.40), SIMDE_FLOAT32_C( -43.50), SIMDE_FLOAT32_C( -53.00), SIMDE_FLOAT32_C( 60.80)),
1417
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 15.60), SIMDE_FLOAT32_C( -4.70), SIMDE_FLOAT32_C( 24.20), SIMDE_FLOAT32_C( -46.70)),
1418
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 2739.84), SIMDE_FLOAT32_C( -357.05), SIMDE_FLOAT32_C( 1836.10), SIMDE_FLOAT32_C( 1309.14)) },
1419
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( -92.90), SIMDE_FLOAT32_C( 31.90), SIMDE_FLOAT32_C( -29.90), SIMDE_FLOAT32_C( -95.30)),
1420
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 46.90), SIMDE_FLOAT32_C( -89.80), SIMDE_FLOAT32_C( 18.10), SIMDE_FLOAT32_C( -72.70)),
1421
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 79.60), SIMDE_FLOAT32_C( -32.40), SIMDE_FLOAT32_C( -3.60), SIMDE_FLOAT32_C( -57.10)),
1422
+ simde_mm_set_ps(SIMDE_FLOAT32_C(-4436.61), SIMDE_FLOAT32_C(-2897.02), SIMDE_FLOAT32_C( -537.59), SIMDE_FLOAT32_C( 6871.21)) },
1423
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( 68.50), SIMDE_FLOAT32_C( 48.90), SIMDE_FLOAT32_C( 86.30), SIMDE_FLOAT32_C( 72.10)),
1424
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -80.00), SIMDE_FLOAT32_C( -44.60), SIMDE_FLOAT32_C( -3.60), SIMDE_FLOAT32_C( -91.10)),
1425
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -57.30), SIMDE_FLOAT32_C( 2.10), SIMDE_FLOAT32_C( -33.70), SIMDE_FLOAT32_C( -13.60)),
1426
+ simde_mm_set_ps(SIMDE_FLOAT32_C(-5422.70), SIMDE_FLOAT32_C(-2178.84), SIMDE_FLOAT32_C( -276.98), SIMDE_FLOAT32_C(-6581.91)) },
1427
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( -38.10), SIMDE_FLOAT32_C( -61.30), SIMDE_FLOAT32_C( 38.90), SIMDE_FLOAT32_C( -79.30)),
1428
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 64.20), SIMDE_FLOAT32_C( 71.60), SIMDE_FLOAT32_C( -99.30), SIMDE_FLOAT32_C( -87.20)),
1429
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -46.40), SIMDE_FLOAT32_C( 45.20), SIMDE_FLOAT32_C( -56.00), SIMDE_FLOAT32_C( 0.40)),
1430
+ simde_mm_set_ps(SIMDE_FLOAT32_C(-2399.62), SIMDE_FLOAT32_C(-4343.88), SIMDE_FLOAT32_C(-3806.77), SIMDE_FLOAT32_C( 6915.36)) },
1431
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( -88.30), SIMDE_FLOAT32_C( -23.50), SIMDE_FLOAT32_C( 48.80), SIMDE_FLOAT32_C( -55.10)),
1432
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -31.80), SIMDE_FLOAT32_C( 50.50), SIMDE_FLOAT32_C( -24.10), SIMDE_FLOAT32_C( -80.20)),
1433
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -18.70), SIMDE_FLOAT32_C( -24.70), SIMDE_FLOAT32_C( -56.50), SIMDE_FLOAT32_C( 57.70)),
1434
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 2826.64), SIMDE_FLOAT32_C(-1211.45), SIMDE_FLOAT32_C(-1119.58), SIMDE_FLOAT32_C( 4476.72)) },
1435
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( 48.60), SIMDE_FLOAT32_C( 33.60), SIMDE_FLOAT32_C( 8.60), SIMDE_FLOAT32_C( 57.90)),
1436
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 52.40), SIMDE_FLOAT32_C( 2.70), SIMDE_FLOAT32_C( 57.50), SIMDE_FLOAT32_C( -10.40)),
1437
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -26.60), SIMDE_FLOAT32_C( -67.20), SIMDE_FLOAT32_C( 5.80), SIMDE_FLOAT32_C( 75.00)),
1438
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 2573.24), SIMDE_FLOAT32_C( 23.52), SIMDE_FLOAT32_C( 488.70), SIMDE_FLOAT32_C( -527.16)) },
1439
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( -25.60), SIMDE_FLOAT32_C( 57.60), SIMDE_FLOAT32_C( -91.00), SIMDE_FLOAT32_C( 53.30)),
1440
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -92.00), SIMDE_FLOAT32_C( 35.10), SIMDE_FLOAT32_C( 8.60), SIMDE_FLOAT32_C( 0.00)),
1441
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -41.20), SIMDE_FLOAT32_C( -81.00), SIMDE_FLOAT32_C( -21.80), SIMDE_FLOAT32_C( -49.00)),
1442
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 2396.40), SIMDE_FLOAT32_C( 1940.76), SIMDE_FLOAT32_C( -760.80), SIMDE_FLOAT32_C( -49.00)) }
1443
+ };
1444
+
1445
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
1446
+ simde__m128 r = simde_mm_fmsubadd_ps(test_vec[i].a, test_vec[i].b, test_vec[i].c);
1447
+ simde_assert_m128_close(r, test_vec[i].r, 1);
1448
+ }
1449
+
1450
+ return MUNIT_OK;
1451
+ }
1452
+
1453
+ static MunitResult
1454
+ test_simde_mm256_fmsubadd_ps(const MunitParameter params[], void* data) {
1455
+ (void) params;
1456
+ (void) data;
1457
+
1458
+ const struct {
1459
+ simde__m256 a;
1460
+ simde__m256 b;
1461
+ simde__m256 c;
1462
+ simde__m256 r;
1463
+ } test_vec[8] = {
1464
+ { simde_mm256_set_ps(SIMDE_FLOAT32_C( 80.60), SIMDE_FLOAT32_C( -80.20),
1465
+ SIMDE_FLOAT32_C( 25.10), SIMDE_FLOAT32_C( 54.40),
1466
+ SIMDE_FLOAT32_C( -94.50), SIMDE_FLOAT32_C( -99.70),
1467
+ SIMDE_FLOAT32_C( 67.30), SIMDE_FLOAT32_C( -5.80)),
1468
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( -98.10), SIMDE_FLOAT32_C( -47.30),
1469
+ SIMDE_FLOAT32_C( -82.80), SIMDE_FLOAT32_C( -26.80),
1470
+ SIMDE_FLOAT32_C( 87.80), SIMDE_FLOAT32_C( 71.10),
1471
+ SIMDE_FLOAT32_C( 92.80), SIMDE_FLOAT32_C( -97.90)),
1472
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( -36.00), SIMDE_FLOAT32_C( -59.40),
1473
+ SIMDE_FLOAT32_C( -69.40), SIMDE_FLOAT32_C( 50.50),
1474
+ SIMDE_FLOAT32_C( 70.50), SIMDE_FLOAT32_C( 26.60),
1475
+ SIMDE_FLOAT32_C( 29.70), SIMDE_FLOAT32_C( -14.80)),
1476
+ simde_mm256_set_ps(SIMDE_FLOAT32_C(-7870.86), SIMDE_FLOAT32_C( 3734.06),
1477
+ SIMDE_FLOAT32_C(-2008.88), SIMDE_FLOAT32_C(-1407.42),
1478
+ SIMDE_FLOAT32_C(-8367.60), SIMDE_FLOAT32_C(-7062.07),
1479
+ SIMDE_FLOAT32_C( 6215.74), SIMDE_FLOAT32_C( 553.02)) },
1480
+ { simde_mm256_set_ps(SIMDE_FLOAT32_C( -68.50), SIMDE_FLOAT32_C( -56.10),
1481
+ SIMDE_FLOAT32_C( 89.00), SIMDE_FLOAT32_C( -96.30),
1482
+ SIMDE_FLOAT32_C( 41.10), SIMDE_FLOAT32_C( -67.50),
1483
+ SIMDE_FLOAT32_C( 59.30), SIMDE_FLOAT32_C( -62.90)),
1484
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( -50.40), SIMDE_FLOAT32_C( -79.00),
1485
+ SIMDE_FLOAT32_C( 93.10), SIMDE_FLOAT32_C( -46.20),
1486
+ SIMDE_FLOAT32_C( -86.10), SIMDE_FLOAT32_C( 19.30),
1487
+ SIMDE_FLOAT32_C( -62.90), SIMDE_FLOAT32_C( -49.30)),
1488
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( -28.70), SIMDE_FLOAT32_C( -24.80),
1489
+ SIMDE_FLOAT32_C( 30.30), SIMDE_FLOAT32_C( -97.00),
1490
+ SIMDE_FLOAT32_C( -57.70), SIMDE_FLOAT32_C( -32.40),
1491
+ SIMDE_FLOAT32_C( -8.20), SIMDE_FLOAT32_C( 75.20)),
1492
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( 3481.10), SIMDE_FLOAT32_C( 4407.10),
1493
+ SIMDE_FLOAT32_C( 8255.60), SIMDE_FLOAT32_C( 4352.06),
1494
+ SIMDE_FLOAT32_C(-3481.01), SIMDE_FLOAT32_C(-1335.15),
1495
+ SIMDE_FLOAT32_C(-3721.77), SIMDE_FLOAT32_C( 3176.17)) },
1496
+ { simde_mm256_set_ps(SIMDE_FLOAT32_C( -40.40), SIMDE_FLOAT32_C( 50.40),
1497
+ SIMDE_FLOAT32_C( 3.90), SIMDE_FLOAT32_C( -96.60),
1498
+ SIMDE_FLOAT32_C( 84.00), SIMDE_FLOAT32_C( 63.30),
1499
+ SIMDE_FLOAT32_C( 71.70), SIMDE_FLOAT32_C( -5.40)),
1500
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( 88.10), SIMDE_FLOAT32_C( 75.30),
1501
+ SIMDE_FLOAT32_C( -17.10), SIMDE_FLOAT32_C( -27.60),
1502
+ SIMDE_FLOAT32_C( 47.20), SIMDE_FLOAT32_C( -72.70),
1503
+ SIMDE_FLOAT32_C( -49.20), SIMDE_FLOAT32_C( -33.10)),
1504
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( 96.20), SIMDE_FLOAT32_C( 64.10),
1505
+ SIMDE_FLOAT32_C( 96.10), SIMDE_FLOAT32_C( -18.70),
1506
+ SIMDE_FLOAT32_C( -31.60), SIMDE_FLOAT32_C( 43.60),
1507
+ SIMDE_FLOAT32_C( -90.90), SIMDE_FLOAT32_C( -27.30)),
1508
+ simde_mm256_set_ps(SIMDE_FLOAT32_C(-3655.44), SIMDE_FLOAT32_C( 3859.22),
1509
+ SIMDE_FLOAT32_C( -162.79), SIMDE_FLOAT32_C( 2647.46),
1510
+ SIMDE_FLOAT32_C( 3996.40), SIMDE_FLOAT32_C(-4558.31),
1511
+ SIMDE_FLOAT32_C(-3436.74), SIMDE_FLOAT32_C( 151.44)) },
1512
+ { simde_mm256_set_ps(SIMDE_FLOAT32_C( -59.60), SIMDE_FLOAT32_C( -83.90),
1513
+ SIMDE_FLOAT32_C( 58.10), SIMDE_FLOAT32_C( -6.90),
1514
+ SIMDE_FLOAT32_C( 99.80), SIMDE_FLOAT32_C( -64.30),
1515
+ SIMDE_FLOAT32_C( 87.70), SIMDE_FLOAT32_C( 55.40)),
1516
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( 53.60), SIMDE_FLOAT32_C( 89.30),
1517
+ SIMDE_FLOAT32_C( -70.60), SIMDE_FLOAT32_C( 99.70),
1518
+ SIMDE_FLOAT32_C( -5.30), SIMDE_FLOAT32_C( 5.60),
1519
+ SIMDE_FLOAT32_C( 86.80), SIMDE_FLOAT32_C( -0.20)),
1520
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( 15.90), SIMDE_FLOAT32_C( -12.20),
1521
+ SIMDE_FLOAT32_C( 93.70), SIMDE_FLOAT32_C( -91.90),
1522
+ SIMDE_FLOAT32_C( 34.20), SIMDE_FLOAT32_C( -64.50),
1523
+ SIMDE_FLOAT32_C( 97.10), SIMDE_FLOAT32_C( -8.00)),
1524
+ simde_mm256_set_ps(SIMDE_FLOAT32_C(-3210.46), SIMDE_FLOAT32_C(-7504.47),
1525
+ SIMDE_FLOAT32_C(-4195.56), SIMDE_FLOAT32_C( -779.83),
1526
+ SIMDE_FLOAT32_C( -563.14), SIMDE_FLOAT32_C( -424.58),
1527
+ SIMDE_FLOAT32_C( 7515.26), SIMDE_FLOAT32_C( -19.08)) },
1528
+ { simde_mm256_set_ps(SIMDE_FLOAT32_C( 49.20), SIMDE_FLOAT32_C( -59.10),
1529
+ SIMDE_FLOAT32_C( -10.90), SIMDE_FLOAT32_C( -67.30),
1530
+ SIMDE_FLOAT32_C( 52.90), SIMDE_FLOAT32_C( -9.10),
1531
+ SIMDE_FLOAT32_C( -30.60), SIMDE_FLOAT32_C( -79.10)),
1532
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( -82.90), SIMDE_FLOAT32_C( 24.10),
1533
+ SIMDE_FLOAT32_C( 5.20), SIMDE_FLOAT32_C( -4.60),
1534
+ SIMDE_FLOAT32_C( -64.40), SIMDE_FLOAT32_C( -6.30),
1535
+ SIMDE_FLOAT32_C( 88.20), SIMDE_FLOAT32_C( 59.20)),
1536
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( 59.30), SIMDE_FLOAT32_C( -23.80),
1537
+ SIMDE_FLOAT32_C( 86.10), SIMDE_FLOAT32_C( 45.80),
1538
+ SIMDE_FLOAT32_C( -77.20), SIMDE_FLOAT32_C( 3.40),
1539
+ SIMDE_FLOAT32_C( 70.60), SIMDE_FLOAT32_C( -87.00)),
1540
+ simde_mm256_set_ps(SIMDE_FLOAT32_C(-4137.98), SIMDE_FLOAT32_C(-1448.11),
1541
+ SIMDE_FLOAT32_C( -142.78), SIMDE_FLOAT32_C( 355.38),
1542
+ SIMDE_FLOAT32_C(-3329.56), SIMDE_FLOAT32_C( 60.73),
1543
+ SIMDE_FLOAT32_C(-2769.52), SIMDE_FLOAT32_C(-4769.72)) },
1544
+ { simde_mm256_set_ps(SIMDE_FLOAT32_C( -41.90), SIMDE_FLOAT32_C( 60.40),
1545
+ SIMDE_FLOAT32_C( -79.60), SIMDE_FLOAT32_C( 95.50),
1546
+ SIMDE_FLOAT32_C( 31.30), SIMDE_FLOAT32_C( -95.40),
1547
+ SIMDE_FLOAT32_C( 27.30), SIMDE_FLOAT32_C( 96.90)),
1548
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( 46.90), SIMDE_FLOAT32_C( -42.30),
1549
+ SIMDE_FLOAT32_C( 95.50), SIMDE_FLOAT32_C( -75.00),
1550
+ SIMDE_FLOAT32_C( 48.70), SIMDE_FLOAT32_C( 76.90),
1551
+ SIMDE_FLOAT32_C( 81.90), SIMDE_FLOAT32_C( 70.10)),
1552
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( -69.40), SIMDE_FLOAT32_C( 89.00),
1553
+ SIMDE_FLOAT32_C( -88.50), SIMDE_FLOAT32_C( 76.60),
1554
+ SIMDE_FLOAT32_C( -55.90), SIMDE_FLOAT32_C( -98.10),
1555
+ SIMDE_FLOAT32_C( -24.00), SIMDE_FLOAT32_C( -35.80)),
1556
+ simde_mm256_set_ps(SIMDE_FLOAT32_C(-1895.71), SIMDE_FLOAT32_C(-2465.92),
1557
+ SIMDE_FLOAT32_C(-7513.30), SIMDE_FLOAT32_C(-7085.90),
1558
+ SIMDE_FLOAT32_C( 1580.21), SIMDE_FLOAT32_C(-7434.36),
1559
+ SIMDE_FLOAT32_C( 2259.87), SIMDE_FLOAT32_C( 6756.89)) },
1560
+ { simde_mm256_set_ps(SIMDE_FLOAT32_C( 12.90), SIMDE_FLOAT32_C( 65.20),
1561
+ SIMDE_FLOAT32_C( 56.70), SIMDE_FLOAT32_C( 39.40),
1562
+ SIMDE_FLOAT32_C( -25.60), SIMDE_FLOAT32_C( -1.40),
1563
+ SIMDE_FLOAT32_C( 44.70), SIMDE_FLOAT32_C( -72.00)),
1564
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( -49.00), SIMDE_FLOAT32_C( 97.00),
1565
+ SIMDE_FLOAT32_C( -63.50), SIMDE_FLOAT32_C( -40.00),
1566
+ SIMDE_FLOAT32_C( 48.40), SIMDE_FLOAT32_C( 30.20),
1567
+ SIMDE_FLOAT32_C( -73.80), SIMDE_FLOAT32_C( -79.50)),
1568
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( 5.30), SIMDE_FLOAT32_C( 28.90),
1569
+ SIMDE_FLOAT32_C( 61.30), SIMDE_FLOAT32_C( -5.70),
1570
+ SIMDE_FLOAT32_C( 39.10), SIMDE_FLOAT32_C( -88.70),
1571
+ SIMDE_FLOAT32_C( 17.20), SIMDE_FLOAT32_C( 0.40)),
1572
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( -637.40), SIMDE_FLOAT32_C( 6353.30),
1573
+ SIMDE_FLOAT32_C(-3661.75), SIMDE_FLOAT32_C(-1581.70),
1574
+ SIMDE_FLOAT32_C(-1278.14), SIMDE_FLOAT32_C( -130.98),
1575
+ SIMDE_FLOAT32_C(-3316.06), SIMDE_FLOAT32_C( 5724.40)) },
1576
+ { simde_mm256_set_ps(SIMDE_FLOAT32_C( 1.50), SIMDE_FLOAT32_C( -69.70),
1577
+ SIMDE_FLOAT32_C( -80.70), SIMDE_FLOAT32_C( 7.80),
1578
+ SIMDE_FLOAT32_C( -92.30), SIMDE_FLOAT32_C( 11.90),
1579
+ SIMDE_FLOAT32_C( 59.30), SIMDE_FLOAT32_C( -21.40)),
1580
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( -78.70), SIMDE_FLOAT32_C( -69.80),
1581
+ SIMDE_FLOAT32_C( 38.10), SIMDE_FLOAT32_C( 22.10),
1582
+ SIMDE_FLOAT32_C( -96.20), SIMDE_FLOAT32_C( 60.20),
1583
+ SIMDE_FLOAT32_C( 49.80), SIMDE_FLOAT32_C( -68.50)),
1584
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( 52.80), SIMDE_FLOAT32_C( 2.20),
1585
+ SIMDE_FLOAT32_C( -17.20), SIMDE_FLOAT32_C( 60.50),
1586
+ SIMDE_FLOAT32_C( -86.40), SIMDE_FLOAT32_C( -89.40),
1587
+ SIMDE_FLOAT32_C( -67.80), SIMDE_FLOAT32_C( 4.40)),
1588
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( -170.85), SIMDE_FLOAT32_C( 4867.26),
1589
+ SIMDE_FLOAT32_C(-3057.47), SIMDE_FLOAT32_C( 232.88),
1590
+ SIMDE_FLOAT32_C( 8965.66), SIMDE_FLOAT32_C( 626.98),
1591
+ SIMDE_FLOAT32_C( 3020.94), SIMDE_FLOAT32_C( 1470.30)) }
1592
+ };
1593
+
1594
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
1595
+ simde__m256 r = simde_mm256_fmsubadd_ps(test_vec[i].a, test_vec[i].b, test_vec[i].c);
1596
+ simde_assert_m256_close(r, test_vec[i].r, 1);
1597
+ }
1598
+
1599
+ return MUNIT_OK;
1600
+ }
1601
+
1602
+ static MunitResult
1603
+ test_simde_mm_fnmadd_pd(const MunitParameter params[], void* data) {
1604
+ (void) params;
1605
+ (void) data;
1606
+
1607
+ const struct {
1608
+ simde__m128d a;
1609
+ simde__m128d b;
1610
+ simde__m128d c;
1611
+ simde__m128d r;
1612
+ } test_vec[8] = {
1613
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( -93.70), SIMDE_FLOAT64_C( 14.40)),
1614
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 8.90), SIMDE_FLOAT64_C( -15.90)),
1615
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -87.90), SIMDE_FLOAT64_C( -34.80)),
1616
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 746.03), SIMDE_FLOAT64_C( 194.16)) },
1617
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( 38.10), SIMDE_FLOAT64_C( -13.20)),
1618
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 63.40), SIMDE_FLOAT64_C( -68.80)),
1619
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 13.30), SIMDE_FLOAT64_C( -61.60)),
1620
+ simde_mm_set_pd(SIMDE_FLOAT64_C(-2402.24), SIMDE_FLOAT64_C( -969.76)) },
1621
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( 94.40), SIMDE_FLOAT64_C( 89.90)),
1622
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -60.60), SIMDE_FLOAT64_C( -24.70)),
1623
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -87.30), SIMDE_FLOAT64_C( 84.80)),
1624
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 5633.34), SIMDE_FLOAT64_C( 2305.33)) },
1625
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( -6.90), SIMDE_FLOAT64_C( 88.80)),
1626
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 6.60), SIMDE_FLOAT64_C( -57.80)),
1627
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 35.50), SIMDE_FLOAT64_C( 30.80)),
1628
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 81.04), SIMDE_FLOAT64_C( 5163.44)) },
1629
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( 51.80), SIMDE_FLOAT64_C( 95.50)),
1630
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -57.60), SIMDE_FLOAT64_C( -59.80)),
1631
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -97.40), SIMDE_FLOAT64_C( -60.30)),
1632
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 2886.28), SIMDE_FLOAT64_C( 5650.60)) },
1633
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( 71.70), SIMDE_FLOAT64_C( -99.40)),
1634
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 27.40), SIMDE_FLOAT64_C( 37.90)),
1635
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 62.10), SIMDE_FLOAT64_C( 17.90)),
1636
+ simde_mm_set_pd(SIMDE_FLOAT64_C(-1902.48), SIMDE_FLOAT64_C( 3785.16)) },
1637
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( 43.60), SIMDE_FLOAT64_C( 78.80)),
1638
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -37.30), SIMDE_FLOAT64_C( -4.80)),
1639
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -23.90), SIMDE_FLOAT64_C( -9.00)),
1640
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 1602.38), SIMDE_FLOAT64_C( 369.24)) },
1641
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( 26.20), SIMDE_FLOAT64_C( -96.60)),
1642
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 57.90), SIMDE_FLOAT64_C( 91.50)),
1643
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 97.90), SIMDE_FLOAT64_C( 18.30)),
1644
+ simde_mm_set_pd(SIMDE_FLOAT64_C(-1419.08), SIMDE_FLOAT64_C( 8857.20)) }
1645
+ };
1646
+
1647
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
1648
+ simde__m128d r = simde_mm_fnmadd_pd(test_vec[i].a, test_vec[i].b, test_vec[i].c);
1649
+ simde_assert_m128d_close(r, test_vec[i].r, 1);
1650
+ }
1651
+
1652
+ return MUNIT_OK;
1653
+ }
1654
+
1655
+ static MunitResult
1656
+ test_simde_mm256_fnmadd_pd(const MunitParameter params[], void* data) {
1657
+ (void) params;
1658
+ (void) data;
1659
+
1660
+ const struct {
1661
+ simde__m256d a;
1662
+ simde__m256d b;
1663
+ simde__m256d c;
1664
+ simde__m256d r;
1665
+ } test_vec[8] = {
1666
+ { simde_mm256_set_pd(SIMDE_FLOAT64_C( 17.60), SIMDE_FLOAT64_C( -99.20),
1667
+ SIMDE_FLOAT64_C( 64.80), SIMDE_FLOAT64_C( -66.40)),
1668
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( -84.50), SIMDE_FLOAT64_C( 62.70),
1669
+ SIMDE_FLOAT64_C( -1.00), SIMDE_FLOAT64_C( 62.00)),
1670
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( 6.20), SIMDE_FLOAT64_C( -52.40),
1671
+ SIMDE_FLOAT64_C( -54.70), SIMDE_FLOAT64_C( 93.30)),
1672
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( 1493.40), SIMDE_FLOAT64_C( 6167.44),
1673
+ SIMDE_FLOAT64_C( 10.10), SIMDE_FLOAT64_C( 4210.10)) },
1674
+ { simde_mm256_set_pd(SIMDE_FLOAT64_C( -70.30), SIMDE_FLOAT64_C( 67.00),
1675
+ SIMDE_FLOAT64_C( 26.40), SIMDE_FLOAT64_C( 52.00)),
1676
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( 8.30), SIMDE_FLOAT64_C( -6.70),
1677
+ SIMDE_FLOAT64_C( -38.30), SIMDE_FLOAT64_C( -42.20)),
1678
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( 77.70), SIMDE_FLOAT64_C( 26.30),
1679
+ SIMDE_FLOAT64_C( 10.50), SIMDE_FLOAT64_C( 36.60)),
1680
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( 661.19), SIMDE_FLOAT64_C( 475.20),
1681
+ SIMDE_FLOAT64_C( 1021.62), SIMDE_FLOAT64_C( 2231.00)) },
1682
+ { simde_mm256_set_pd(SIMDE_FLOAT64_C( -57.10), SIMDE_FLOAT64_C( 58.80),
1683
+ SIMDE_FLOAT64_C( 93.20), SIMDE_FLOAT64_C( -86.80)),
1684
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( 59.10), SIMDE_FLOAT64_C( 76.50),
1685
+ SIMDE_FLOAT64_C( 45.10), SIMDE_FLOAT64_C( 67.70)),
1686
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( 60.20), SIMDE_FLOAT64_C( 65.10),
1687
+ SIMDE_FLOAT64_C( -17.00), SIMDE_FLOAT64_C( -84.40)),
1688
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( 3434.81), SIMDE_FLOAT64_C(-4433.10),
1689
+ SIMDE_FLOAT64_C(-4220.32), SIMDE_FLOAT64_C( 5791.96)) },
1690
+ { simde_mm256_set_pd(SIMDE_FLOAT64_C( 34.60), SIMDE_FLOAT64_C( -5.80),
1691
+ SIMDE_FLOAT64_C( 89.80), SIMDE_FLOAT64_C( -83.20)),
1692
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( 43.00), SIMDE_FLOAT64_C( 3.10),
1693
+ SIMDE_FLOAT64_C( -37.70), SIMDE_FLOAT64_C( -40.60)),
1694
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( 26.40), SIMDE_FLOAT64_C( -59.60),
1695
+ SIMDE_FLOAT64_C( -71.50), SIMDE_FLOAT64_C( 60.40)),
1696
+ simde_mm256_set_pd(SIMDE_FLOAT64_C(-1461.40), SIMDE_FLOAT64_C( -41.62),
1697
+ SIMDE_FLOAT64_C( 3313.96), SIMDE_FLOAT64_C(-3317.52)) },
1698
+ { simde_mm256_set_pd(SIMDE_FLOAT64_C( 40.20), SIMDE_FLOAT64_C( -24.50),
1699
+ SIMDE_FLOAT64_C( -31.60), SIMDE_FLOAT64_C( 3.30)),
1700
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( 39.80), SIMDE_FLOAT64_C( 32.70),
1701
+ SIMDE_FLOAT64_C( 20.30), SIMDE_FLOAT64_C( 49.70)),
1702
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( 69.00), SIMDE_FLOAT64_C( 7.80),
1703
+ SIMDE_FLOAT64_C( 99.70), SIMDE_FLOAT64_C( 49.20)),
1704
+ simde_mm256_set_pd(SIMDE_FLOAT64_C(-1530.96), SIMDE_FLOAT64_C( 808.95),
1705
+ SIMDE_FLOAT64_C( 741.18), SIMDE_FLOAT64_C( -114.81)) },
1706
+ { simde_mm256_set_pd(SIMDE_FLOAT64_C( -82.30), SIMDE_FLOAT64_C( -8.50),
1707
+ SIMDE_FLOAT64_C( -80.50), SIMDE_FLOAT64_C( 9.10)),
1708
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( 52.10), SIMDE_FLOAT64_C( -96.40),
1709
+ SIMDE_FLOAT64_C( 3.00), SIMDE_FLOAT64_C( -86.00)),
1710
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( -93.70), SIMDE_FLOAT64_C( 8.90),
1711
+ SIMDE_FLOAT64_C( 46.10), SIMDE_FLOAT64_C( -50.90)),
1712
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( 4194.13), SIMDE_FLOAT64_C( -810.50),
1713
+ SIMDE_FLOAT64_C( 287.60), SIMDE_FLOAT64_C( 731.70)) },
1714
+ { simde_mm256_set_pd(SIMDE_FLOAT64_C( 72.30), SIMDE_FLOAT64_C( 96.70),
1715
+ SIMDE_FLOAT64_C( -51.00), SIMDE_FLOAT64_C( -38.00)),
1716
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( 37.50), SIMDE_FLOAT64_C( 93.30),
1717
+ SIMDE_FLOAT64_C( 79.70), SIMDE_FLOAT64_C( 71.40)),
1718
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( 54.00), SIMDE_FLOAT64_C( 6.80),
1719
+ SIMDE_FLOAT64_C( -77.40), SIMDE_FLOAT64_C( -48.10)),
1720
+ simde_mm256_set_pd(SIMDE_FLOAT64_C(-2657.25), SIMDE_FLOAT64_C(-9015.31),
1721
+ SIMDE_FLOAT64_C( 3987.30), SIMDE_FLOAT64_C( 2665.10)) },
1722
+ { simde_mm256_set_pd(SIMDE_FLOAT64_C( -52.40), SIMDE_FLOAT64_C( -75.40),
1723
+ SIMDE_FLOAT64_C( -96.00), SIMDE_FLOAT64_C( -23.10)),
1724
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( -21.90), SIMDE_FLOAT64_C( -53.30),
1725
+ SIMDE_FLOAT64_C( -90.50), SIMDE_FLOAT64_C( -18.30)),
1726
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( -63.30), SIMDE_FLOAT64_C( -23.10),
1727
+ SIMDE_FLOAT64_C( -88.90), SIMDE_FLOAT64_C( 67.10)),
1728
+ simde_mm256_set_pd(SIMDE_FLOAT64_C(-1210.86), SIMDE_FLOAT64_C(-4041.92),
1729
+ SIMDE_FLOAT64_C(-8776.90), SIMDE_FLOAT64_C( -355.63)) }
1730
+ };
1731
+
1732
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
1733
+ simde__m256d r = simde_mm256_fnmadd_pd(test_vec[i].a, test_vec[i].b, test_vec[i].c);
1734
+ simde_assert_m256d_close(r, test_vec[i].r, 1);
1735
+ }
1736
+
1737
+ return MUNIT_OK;
1738
+ }
1739
+
1740
+ static MunitResult
1741
+ test_simde_mm_fnmadd_ps(const MunitParameter params[], void* data) {
1742
+ (void) params;
1743
+ (void) data;
1744
+
1745
+ const struct {
1746
+ simde__m128 a;
1747
+ simde__m128 b;
1748
+ simde__m128 c;
1749
+ simde__m128 r;
1750
+ } test_vec[8] = {
1751
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( 5.20), SIMDE_FLOAT32_C( 59.60), SIMDE_FLOAT32_C( 87.70), SIMDE_FLOAT32_C( 47.10)),
1752
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -48.20), SIMDE_FLOAT32_C( -88.00), SIMDE_FLOAT32_C( 90.80), SIMDE_FLOAT32_C( -22.90)),
1753
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -19.00), SIMDE_FLOAT32_C( 40.90), SIMDE_FLOAT32_C( 74.00), SIMDE_FLOAT32_C( 71.80)),
1754
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 231.64), SIMDE_FLOAT32_C( 5285.70), SIMDE_FLOAT32_C(-7889.16), SIMDE_FLOAT32_C( 1150.39)) },
1755
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( 23.70), SIMDE_FLOAT32_C( 46.10), SIMDE_FLOAT32_C( -5.90), SIMDE_FLOAT32_C( 49.30)),
1756
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 6.50), SIMDE_FLOAT32_C( 83.40), SIMDE_FLOAT32_C( -86.10), SIMDE_FLOAT32_C( 15.60)),
1757
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -10.40), SIMDE_FLOAT32_C( -37.00), SIMDE_FLOAT32_C( -97.90), SIMDE_FLOAT32_C( 43.80)),
1758
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -164.45), SIMDE_FLOAT32_C(-3881.74), SIMDE_FLOAT32_C( -605.89), SIMDE_FLOAT32_C( -725.28)) },
1759
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( -41.10), SIMDE_FLOAT32_C( 98.60), SIMDE_FLOAT32_C( -66.40), SIMDE_FLOAT32_C( 31.30)),
1760
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 80.90), SIMDE_FLOAT32_C( -40.10), SIMDE_FLOAT32_C( -24.70), SIMDE_FLOAT32_C( 7.90)),
1761
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 99.20), SIMDE_FLOAT32_C( -40.90), SIMDE_FLOAT32_C( -69.50), SIMDE_FLOAT32_C( 9.90)),
1762
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 3424.19), SIMDE_FLOAT32_C( 3912.96), SIMDE_FLOAT32_C(-1709.58), SIMDE_FLOAT32_C( -237.37)) },
1763
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( 0.30), SIMDE_FLOAT32_C( 18.10), SIMDE_FLOAT32_C( -38.40), SIMDE_FLOAT32_C( -54.30)),
1764
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -5.80), SIMDE_FLOAT32_C( 84.90), SIMDE_FLOAT32_C( -77.80), SIMDE_FLOAT32_C( -32.70)),
1765
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -23.10), SIMDE_FLOAT32_C( 3.00), SIMDE_FLOAT32_C( 5.40), SIMDE_FLOAT32_C( 61.50)),
1766
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -21.36), SIMDE_FLOAT32_C(-1533.69), SIMDE_FLOAT32_C(-2982.12), SIMDE_FLOAT32_C(-1714.11)) },
1767
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( -95.30), SIMDE_FLOAT32_C( -61.60), SIMDE_FLOAT32_C( -95.50), SIMDE_FLOAT32_C( -55.10)),
1768
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -67.20), SIMDE_FLOAT32_C( 95.00), SIMDE_FLOAT32_C( 94.10), SIMDE_FLOAT32_C( 87.40)),
1769
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 45.20), SIMDE_FLOAT32_C( -12.10), SIMDE_FLOAT32_C( -17.00), SIMDE_FLOAT32_C( -48.70)),
1770
+ simde_mm_set_ps(SIMDE_FLOAT32_C(-6358.96), SIMDE_FLOAT32_C( 5839.90), SIMDE_FLOAT32_C( 8969.55), SIMDE_FLOAT32_C( 4767.04)) },
1771
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( -85.20), SIMDE_FLOAT32_C( -17.40), SIMDE_FLOAT32_C( 5.50), SIMDE_FLOAT32_C( 51.40)),
1772
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 84.50), SIMDE_FLOAT32_C( 0.60), SIMDE_FLOAT32_C( 61.30), SIMDE_FLOAT32_C( -9.50)),
1773
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -44.50), SIMDE_FLOAT32_C( -83.00), SIMDE_FLOAT32_C( -17.60), SIMDE_FLOAT32_C( -95.10)),
1774
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 7154.90), SIMDE_FLOAT32_C( -72.56), SIMDE_FLOAT32_C( -354.75), SIMDE_FLOAT32_C( 393.20)) },
1775
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( 40.00), SIMDE_FLOAT32_C( -99.20), SIMDE_FLOAT32_C( -45.30), SIMDE_FLOAT32_C( 65.20)),
1776
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 3.60), SIMDE_FLOAT32_C( -27.50), SIMDE_FLOAT32_C( 92.40), SIMDE_FLOAT32_C( -74.00)),
1777
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 91.90), SIMDE_FLOAT32_C( 62.60), SIMDE_FLOAT32_C( 33.10), SIMDE_FLOAT32_C( 17.30)),
1778
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -52.10), SIMDE_FLOAT32_C(-2665.40), SIMDE_FLOAT32_C( 4218.82), SIMDE_FLOAT32_C( 4842.10)) },
1779
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( 46.00), SIMDE_FLOAT32_C( 7.80), SIMDE_FLOAT32_C( 62.40), SIMDE_FLOAT32_C( 98.20)),
1780
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 60.60), SIMDE_FLOAT32_C( -96.70), SIMDE_FLOAT32_C( 86.60), SIMDE_FLOAT32_C( 94.40)),
1781
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 57.50), SIMDE_FLOAT32_C( -34.30), SIMDE_FLOAT32_C( -42.40), SIMDE_FLOAT32_C( -32.30)),
1782
+ simde_mm_set_ps(SIMDE_FLOAT32_C(-2730.10), SIMDE_FLOAT32_C( 719.96), SIMDE_FLOAT32_C(-5446.24), SIMDE_FLOAT32_C(-9302.38)) }
1783
+ };
1784
+
1785
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
1786
+ simde__m128 r = simde_mm_fnmadd_ps(test_vec[i].a, test_vec[i].b, test_vec[i].c);
1787
+ simde_assert_m128_close(r, test_vec[i].r, 1);
1788
+ }
1789
+
1790
+ return MUNIT_OK;
1791
+ }
1792
+
1793
+ static MunitResult
1794
+ test_simde_mm256_fnmadd_ps(const MunitParameter params[], void* data) {
1795
+ (void) params;
1796
+ (void) data;
1797
+
1798
+ const struct {
1799
+ simde__m256 a;
1800
+ simde__m256 b;
1801
+ simde__m256 c;
1802
+ simde__m256 r;
1803
+ } test_vec[8] = {
1804
+ { simde_mm256_set_ps(SIMDE_FLOAT32_C( -74.50), SIMDE_FLOAT32_C( 76.00),
1805
+ SIMDE_FLOAT32_C( -65.60), SIMDE_FLOAT32_C( -57.80),
1806
+ SIMDE_FLOAT32_C( 48.90), SIMDE_FLOAT32_C( 17.90),
1807
+ SIMDE_FLOAT32_C( 92.90), SIMDE_FLOAT32_C( 17.80)),
1808
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( -79.20), SIMDE_FLOAT32_C( 67.50),
1809
+ SIMDE_FLOAT32_C( -50.60), SIMDE_FLOAT32_C( 96.50),
1810
+ SIMDE_FLOAT32_C( -92.70), SIMDE_FLOAT32_C( 12.20),
1811
+ SIMDE_FLOAT32_C( -41.10), SIMDE_FLOAT32_C( -24.10)),
1812
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( -20.70), SIMDE_FLOAT32_C( 76.20),
1813
+ SIMDE_FLOAT32_C( -47.10), SIMDE_FLOAT32_C( -61.40),
1814
+ SIMDE_FLOAT32_C( 55.90), SIMDE_FLOAT32_C( 79.30),
1815
+ SIMDE_FLOAT32_C( -95.40), SIMDE_FLOAT32_C( 98.20)),
1816
+ simde_mm256_set_ps(SIMDE_FLOAT32_C(-5921.10), SIMDE_FLOAT32_C(-5053.80),
1817
+ SIMDE_FLOAT32_C(-3366.46), SIMDE_FLOAT32_C( 5516.30),
1818
+ SIMDE_FLOAT32_C( 4588.93), SIMDE_FLOAT32_C( -139.08),
1819
+ SIMDE_FLOAT32_C( 3722.79), SIMDE_FLOAT32_C( 527.18)) },
1820
+ { simde_mm256_set_ps(SIMDE_FLOAT32_C( 34.30), SIMDE_FLOAT32_C( 90.50),
1821
+ SIMDE_FLOAT32_C( -43.40), SIMDE_FLOAT32_C( -95.00),
1822
+ SIMDE_FLOAT32_C( -62.70), SIMDE_FLOAT32_C( -17.10),
1823
+ SIMDE_FLOAT32_C( 30.50), SIMDE_FLOAT32_C( 1.00)),
1824
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( -23.20), SIMDE_FLOAT32_C( 28.90),
1825
+ SIMDE_FLOAT32_C( 78.70), SIMDE_FLOAT32_C( 6.50),
1826
+ SIMDE_FLOAT32_C( -13.60), SIMDE_FLOAT32_C( 7.60),
1827
+ SIMDE_FLOAT32_C( -56.70), SIMDE_FLOAT32_C( 52.80)),
1828
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( -24.30), SIMDE_FLOAT32_C( 65.20),
1829
+ SIMDE_FLOAT32_C( 27.90), SIMDE_FLOAT32_C( -88.40),
1830
+ SIMDE_FLOAT32_C( -43.70), SIMDE_FLOAT32_C( 61.70),
1831
+ SIMDE_FLOAT32_C( -22.10), SIMDE_FLOAT32_C( -51.30)),
1832
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( 771.46), SIMDE_FLOAT32_C(-2550.25),
1833
+ SIMDE_FLOAT32_C( 3443.48), SIMDE_FLOAT32_C( 529.10),
1834
+ SIMDE_FLOAT32_C( -896.42), SIMDE_FLOAT32_C( 191.66),
1835
+ SIMDE_FLOAT32_C( 1707.25), SIMDE_FLOAT32_C( -104.10)) },
1836
+ { simde_mm256_set_ps(SIMDE_FLOAT32_C( 95.30), SIMDE_FLOAT32_C( -81.70),
1837
+ SIMDE_FLOAT32_C( 51.00), SIMDE_FLOAT32_C( 6.50),
1838
+ SIMDE_FLOAT32_C( 46.00), SIMDE_FLOAT32_C( 76.10),
1839
+ SIMDE_FLOAT32_C( -72.70), SIMDE_FLOAT32_C( 10.10)),
1840
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( 67.00), SIMDE_FLOAT32_C( -43.80),
1841
+ SIMDE_FLOAT32_C( 5.70), SIMDE_FLOAT32_C( 9.00),
1842
+ SIMDE_FLOAT32_C( 39.70), SIMDE_FLOAT32_C( -47.40),
1843
+ SIMDE_FLOAT32_C( -89.40), SIMDE_FLOAT32_C( -69.10)),
1844
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( 5.10), SIMDE_FLOAT32_C( -27.10),
1845
+ SIMDE_FLOAT32_C( 24.30), SIMDE_FLOAT32_C( -90.10),
1846
+ SIMDE_FLOAT32_C( 48.70), SIMDE_FLOAT32_C( 91.00),
1847
+ SIMDE_FLOAT32_C( 80.80), SIMDE_FLOAT32_C( -24.60)),
1848
+ simde_mm256_set_ps(SIMDE_FLOAT32_C(-6380.00), SIMDE_FLOAT32_C(-3605.56),
1849
+ SIMDE_FLOAT32_C( -266.40), SIMDE_FLOAT32_C( -148.60),
1850
+ SIMDE_FLOAT32_C(-1777.50), SIMDE_FLOAT32_C( 3698.14),
1851
+ SIMDE_FLOAT32_C(-6418.58), SIMDE_FLOAT32_C( 673.31)) },
1852
+ { simde_mm256_set_ps(SIMDE_FLOAT32_C( 5.80), SIMDE_FLOAT32_C( 2.80),
1853
+ SIMDE_FLOAT32_C( 37.80), SIMDE_FLOAT32_C( -55.50),
1854
+ SIMDE_FLOAT32_C( 60.80), SIMDE_FLOAT32_C( -46.40),
1855
+ SIMDE_FLOAT32_C( -53.70), SIMDE_FLOAT32_C( -55.70)),
1856
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( 76.00), SIMDE_FLOAT32_C( 65.10),
1857
+ SIMDE_FLOAT32_C( 67.70), SIMDE_FLOAT32_C( -84.20),
1858
+ SIMDE_FLOAT32_C( 63.00), SIMDE_FLOAT32_C( -82.10),
1859
+ SIMDE_FLOAT32_C( -55.20), SIMDE_FLOAT32_C( 20.10)),
1860
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( 60.20), SIMDE_FLOAT32_C( -85.50),
1861
+ SIMDE_FLOAT32_C( 58.00), SIMDE_FLOAT32_C( 40.40),
1862
+ SIMDE_FLOAT32_C( 31.70), SIMDE_FLOAT32_C( -6.20),
1863
+ SIMDE_FLOAT32_C( 83.70), SIMDE_FLOAT32_C( -68.50)),
1864
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( -380.60), SIMDE_FLOAT32_C( -267.78),
1865
+ SIMDE_FLOAT32_C(-2501.06), SIMDE_FLOAT32_C(-4632.70),
1866
+ SIMDE_FLOAT32_C(-3798.70), SIMDE_FLOAT32_C(-3815.64),
1867
+ SIMDE_FLOAT32_C(-2880.54), SIMDE_FLOAT32_C( 1051.07)) },
1868
+ { simde_mm256_set_ps(SIMDE_FLOAT32_C( -30.40), SIMDE_FLOAT32_C( 81.70),
1869
+ SIMDE_FLOAT32_C( -68.60), SIMDE_FLOAT32_C( 46.50),
1870
+ SIMDE_FLOAT32_C( 53.40), SIMDE_FLOAT32_C( -1.10),
1871
+ SIMDE_FLOAT32_C( -70.80), SIMDE_FLOAT32_C( 10.20)),
1872
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( 99.90), SIMDE_FLOAT32_C( -78.30),
1873
+ SIMDE_FLOAT32_C( -52.60), SIMDE_FLOAT32_C( 28.60),
1874
+ SIMDE_FLOAT32_C( 62.90), SIMDE_FLOAT32_C( -65.50),
1875
+ SIMDE_FLOAT32_C( -51.00), SIMDE_FLOAT32_C( -0.20)),
1876
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( 25.50), SIMDE_FLOAT32_C( 2.70),
1877
+ SIMDE_FLOAT32_C( 99.80), SIMDE_FLOAT32_C( -76.10),
1878
+ SIMDE_FLOAT32_C( -4.50), SIMDE_FLOAT32_C( 7.40),
1879
+ SIMDE_FLOAT32_C( 81.50), SIMDE_FLOAT32_C( 1.50)),
1880
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( 3062.46), SIMDE_FLOAT32_C( 6399.81),
1881
+ SIMDE_FLOAT32_C(-3508.56), SIMDE_FLOAT32_C(-1406.00),
1882
+ SIMDE_FLOAT32_C(-3363.36), SIMDE_FLOAT32_C( -64.65),
1883
+ SIMDE_FLOAT32_C(-3529.30), SIMDE_FLOAT32_C( 3.54)) },
1884
+ { simde_mm256_set_ps(SIMDE_FLOAT32_C( 78.00), SIMDE_FLOAT32_C( 45.70),
1885
+ SIMDE_FLOAT32_C( 59.30), SIMDE_FLOAT32_C( 35.50),
1886
+ SIMDE_FLOAT32_C( 91.10), SIMDE_FLOAT32_C( -96.00),
1887
+ SIMDE_FLOAT32_C( -93.90), SIMDE_FLOAT32_C( -0.10)),
1888
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( -80.10), SIMDE_FLOAT32_C( 37.30),
1889
+ SIMDE_FLOAT32_C( 94.60), SIMDE_FLOAT32_C( -45.10),
1890
+ SIMDE_FLOAT32_C( -34.70), SIMDE_FLOAT32_C( -33.50),
1891
+ SIMDE_FLOAT32_C( -17.00), SIMDE_FLOAT32_C( -46.50)),
1892
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( -72.60), SIMDE_FLOAT32_C( 47.00),
1893
+ SIMDE_FLOAT32_C( -93.40), SIMDE_FLOAT32_C( -34.90),
1894
+ SIMDE_FLOAT32_C( 77.40), SIMDE_FLOAT32_C( -96.40),
1895
+ SIMDE_FLOAT32_C( 74.90), SIMDE_FLOAT32_C( 16.90)),
1896
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( 6175.20), SIMDE_FLOAT32_C(-1657.61),
1897
+ SIMDE_FLOAT32_C(-5703.18), SIMDE_FLOAT32_C( 1566.15),
1898
+ SIMDE_FLOAT32_C( 3238.57), SIMDE_FLOAT32_C(-3312.40),
1899
+ SIMDE_FLOAT32_C(-1521.40), SIMDE_FLOAT32_C( 12.25)) },
1900
+ { simde_mm256_set_ps(SIMDE_FLOAT32_C( -42.00), SIMDE_FLOAT32_C( 100.00),
1901
+ SIMDE_FLOAT32_C( -84.50), SIMDE_FLOAT32_C( 27.60),
1902
+ SIMDE_FLOAT32_C( 27.10), SIMDE_FLOAT32_C( -76.60),
1903
+ SIMDE_FLOAT32_C( -36.20), SIMDE_FLOAT32_C( 16.50)),
1904
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( 98.50), SIMDE_FLOAT32_C( -46.90),
1905
+ SIMDE_FLOAT32_C( -21.70), SIMDE_FLOAT32_C( 90.80),
1906
+ SIMDE_FLOAT32_C( 42.70), SIMDE_FLOAT32_C( 48.80),
1907
+ SIMDE_FLOAT32_C( 91.30), SIMDE_FLOAT32_C( 90.10)),
1908
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( 74.40), SIMDE_FLOAT32_C( -15.10),
1909
+ SIMDE_FLOAT32_C( 42.70), SIMDE_FLOAT32_C( -90.90),
1910
+ SIMDE_FLOAT32_C( -30.80), SIMDE_FLOAT32_C( 48.00),
1911
+ SIMDE_FLOAT32_C( 12.60), SIMDE_FLOAT32_C( 59.70)),
1912
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( 4211.40), SIMDE_FLOAT32_C( 4674.90),
1913
+ SIMDE_FLOAT32_C(-1790.95), SIMDE_FLOAT32_C(-2596.98),
1914
+ SIMDE_FLOAT32_C(-1187.97), SIMDE_FLOAT32_C( 3786.08),
1915
+ SIMDE_FLOAT32_C( 3317.66), SIMDE_FLOAT32_C(-1426.95)) },
1916
+ { simde_mm256_set_ps(SIMDE_FLOAT32_C( -49.00), SIMDE_FLOAT32_C( 74.90),
1917
+ SIMDE_FLOAT32_C( -48.00), SIMDE_FLOAT32_C( 46.70),
1918
+ SIMDE_FLOAT32_C( 4.40), SIMDE_FLOAT32_C( 44.70),
1919
+ SIMDE_FLOAT32_C( -68.40), SIMDE_FLOAT32_C( 74.50)),
1920
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( 8.00), SIMDE_FLOAT32_C( -94.30),
1921
+ SIMDE_FLOAT32_C( -6.20), SIMDE_FLOAT32_C( -21.50),
1922
+ SIMDE_FLOAT32_C( 61.90), SIMDE_FLOAT32_C( 14.50),
1923
+ SIMDE_FLOAT32_C( -69.00), SIMDE_FLOAT32_C( -34.50)),
1924
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( 11.20), SIMDE_FLOAT32_C( 22.30),
1925
+ SIMDE_FLOAT32_C( -35.00), SIMDE_FLOAT32_C( 30.60),
1926
+ SIMDE_FLOAT32_C( 72.90), SIMDE_FLOAT32_C( 97.50),
1927
+ SIMDE_FLOAT32_C( 2.70), SIMDE_FLOAT32_C( 72.80)),
1928
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( 403.20), SIMDE_FLOAT32_C( 7085.37),
1929
+ SIMDE_FLOAT32_C( -332.60), SIMDE_FLOAT32_C( 1034.65),
1930
+ SIMDE_FLOAT32_C( -199.46), SIMDE_FLOAT32_C( -550.65),
1931
+ SIMDE_FLOAT32_C(-4716.90), SIMDE_FLOAT32_C( 2643.05)) }
1932
+ };
1933
+
1934
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
1935
+ simde__m256 r = simde_mm256_fnmadd_ps(test_vec[i].a, test_vec[i].b, test_vec[i].c);
1936
+ simde_assert_m256_close(r, test_vec[i].r, 1);
1937
+ }
1938
+
1939
+ return MUNIT_OK;
1940
+ }
1941
+
1942
+ static MunitResult
1943
+ test_simde_mm_fnmadd_sd(const MunitParameter params[], void* data) {
1944
+ (void) params;
1945
+ (void) data;
1946
+
1947
+ const struct {
1948
+ simde__m128d a;
1949
+ simde__m128d b;
1950
+ simde__m128d c;
1951
+ simde__m128d r;
1952
+ } test_vec[8] = {
1953
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( 47.80), SIMDE_FLOAT64_C( -80.70)),
1954
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 55.10), SIMDE_FLOAT64_C( 17.30)),
1955
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -52.00), SIMDE_FLOAT64_C( -7.60)),
1956
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 47.80), SIMDE_FLOAT64_C( 1388.51)) },
1957
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( -16.50), SIMDE_FLOAT64_C( 77.80)),
1958
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -98.70), SIMDE_FLOAT64_C( -77.50)),
1959
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -99.70), SIMDE_FLOAT64_C( 69.50)),
1960
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -16.50), SIMDE_FLOAT64_C( 6099.00)) },
1961
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( 40.70), SIMDE_FLOAT64_C( -56.00)),
1962
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -41.00), SIMDE_FLOAT64_C( -43.20)),
1963
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -9.90), SIMDE_FLOAT64_C( 48.90)),
1964
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 40.70), SIMDE_FLOAT64_C(-2370.30)) },
1965
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( -21.60), SIMDE_FLOAT64_C( -51.90)),
1966
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 91.50), SIMDE_FLOAT64_C( 24.80)),
1967
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 83.10), SIMDE_FLOAT64_C( 15.90)),
1968
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -21.60), SIMDE_FLOAT64_C( 1303.02)) },
1969
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( 36.60), SIMDE_FLOAT64_C( 90.60)),
1970
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 25.20), SIMDE_FLOAT64_C( -17.40)),
1971
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 21.60), SIMDE_FLOAT64_C( 29.60)),
1972
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 36.60), SIMDE_FLOAT64_C( 1606.04)) },
1973
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( 28.70), SIMDE_FLOAT64_C( -13.60)),
1974
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -49.50), SIMDE_FLOAT64_C( 1.80)),
1975
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 78.40), SIMDE_FLOAT64_C( 70.60)),
1976
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 28.70), SIMDE_FLOAT64_C( 95.08)) },
1977
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( -83.30), SIMDE_FLOAT64_C( -83.00)),
1978
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 89.10), SIMDE_FLOAT64_C( 5.70)),
1979
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -0.10), SIMDE_FLOAT64_C( 56.50)),
1980
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -83.30), SIMDE_FLOAT64_C( 529.60)) },
1981
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( -8.80), SIMDE_FLOAT64_C( 91.50)),
1982
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 82.10), SIMDE_FLOAT64_C( -69.30)),
1983
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 92.70), SIMDE_FLOAT64_C( -85.50)),
1984
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -8.80), SIMDE_FLOAT64_C( 6255.45)) }
1985
+ };
1986
+
1987
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
1988
+ simde__m128d r = simde_mm_fnmadd_sd(test_vec[i].a, test_vec[i].b, test_vec[i].c);
1989
+ simde_assert_m128d_close(r, test_vec[i].r, 1);
1990
+ }
1991
+
1992
+ return MUNIT_OK;
1993
+ }
1994
+
1995
+ static MunitResult
1996
+ test_simde_mm_fnmadd_ss(const MunitParameter params[], void* data) {
1997
+ (void) params;
1998
+ (void) data;
1999
+
2000
+ const struct {
2001
+ simde__m128 a;
2002
+ simde__m128 b;
2003
+ simde__m128 c;
2004
+ simde__m128 r;
2005
+ } test_vec[8] = {
2006
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( -92.60), SIMDE_FLOAT32_C( -98.70), SIMDE_FLOAT32_C( 10.90), SIMDE_FLOAT32_C( -61.00)),
2007
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 48.20), SIMDE_FLOAT32_C( 4.60), SIMDE_FLOAT32_C( -98.40), SIMDE_FLOAT32_C( 56.90)),
2008
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -83.90), SIMDE_FLOAT32_C( 54.30), SIMDE_FLOAT32_C( 54.70), SIMDE_FLOAT32_C( 20.70)),
2009
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -92.60), SIMDE_FLOAT32_C( -98.70), SIMDE_FLOAT32_C( 10.90), SIMDE_FLOAT32_C( 3491.60)) },
2010
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( -61.10), SIMDE_FLOAT32_C( -33.00), SIMDE_FLOAT32_C( -47.10), SIMDE_FLOAT32_C( 31.40)),
2011
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 47.10), SIMDE_FLOAT32_C( -73.50), SIMDE_FLOAT32_C( -40.70), SIMDE_FLOAT32_C( -95.70)),
2012
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -73.00), SIMDE_FLOAT32_C( -68.20), SIMDE_FLOAT32_C( 35.20), SIMDE_FLOAT32_C( 48.50)),
2013
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -61.10), SIMDE_FLOAT32_C( -33.00), SIMDE_FLOAT32_C( -47.10), SIMDE_FLOAT32_C( 3053.48)) },
2014
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( -17.60), SIMDE_FLOAT32_C( -75.20), SIMDE_FLOAT32_C( -94.50), SIMDE_FLOAT32_C( 95.50)),
2015
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 56.20), SIMDE_FLOAT32_C( -24.90), SIMDE_FLOAT32_C( 6.00), SIMDE_FLOAT32_C( -33.20)),
2016
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -8.10), SIMDE_FLOAT32_C( 95.10), SIMDE_FLOAT32_C( -66.20), SIMDE_FLOAT32_C( 51.50)),
2017
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -17.60), SIMDE_FLOAT32_C( -75.20), SIMDE_FLOAT32_C( -94.50), SIMDE_FLOAT32_C( 3222.10)) },
2018
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( 15.40), SIMDE_FLOAT32_C( -42.20), SIMDE_FLOAT32_C( -38.90), SIMDE_FLOAT32_C( -40.10)),
2019
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -84.90), SIMDE_FLOAT32_C( -51.00), SIMDE_FLOAT32_C( -45.70), SIMDE_FLOAT32_C( 14.40)),
2020
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -82.20), SIMDE_FLOAT32_C( 60.00), SIMDE_FLOAT32_C( -19.40), SIMDE_FLOAT32_C( 90.10)),
2021
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 15.40), SIMDE_FLOAT32_C( -42.20), SIMDE_FLOAT32_C( -38.90), SIMDE_FLOAT32_C( 667.54)) },
2022
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( -29.00), SIMDE_FLOAT32_C( 45.90), SIMDE_FLOAT32_C( -65.60), SIMDE_FLOAT32_C( -2.20)),
2023
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -46.90), SIMDE_FLOAT32_C( 6.70), SIMDE_FLOAT32_C( -97.90), SIMDE_FLOAT32_C( -72.50)),
2024
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -7.70), SIMDE_FLOAT32_C( -29.90), SIMDE_FLOAT32_C( 69.80), SIMDE_FLOAT32_C( -66.50)),
2025
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -29.00), SIMDE_FLOAT32_C( 45.90), SIMDE_FLOAT32_C( -65.60), SIMDE_FLOAT32_C( -226.00)) },
2026
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( -58.10), SIMDE_FLOAT32_C( -47.10), SIMDE_FLOAT32_C( 68.70), SIMDE_FLOAT32_C( 33.80)),
2027
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 40.90), SIMDE_FLOAT32_C( -18.60), SIMDE_FLOAT32_C( -92.90), SIMDE_FLOAT32_C( 19.60)),
2028
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -14.90), SIMDE_FLOAT32_C( 50.40), SIMDE_FLOAT32_C( -64.40), SIMDE_FLOAT32_C( -4.60)),
2029
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -58.10), SIMDE_FLOAT32_C( -47.10), SIMDE_FLOAT32_C( 68.70), SIMDE_FLOAT32_C( -667.08)) },
2030
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( -90.60), SIMDE_FLOAT32_C( -45.50), SIMDE_FLOAT32_C( -20.60), SIMDE_FLOAT32_C( -95.50)),
2031
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -94.80), SIMDE_FLOAT32_C( 21.50), SIMDE_FLOAT32_C( 77.40), SIMDE_FLOAT32_C( -58.90)),
2032
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -8.20), SIMDE_FLOAT32_C( 56.80), SIMDE_FLOAT32_C( 16.40), SIMDE_FLOAT32_C( -52.50)),
2033
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -90.60), SIMDE_FLOAT32_C( -45.50), SIMDE_FLOAT32_C( -20.60), SIMDE_FLOAT32_C(-5677.45)) },
2034
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( 16.00), SIMDE_FLOAT32_C( -22.20), SIMDE_FLOAT32_C( -70.50), SIMDE_FLOAT32_C( -57.60)),
2035
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 86.70), SIMDE_FLOAT32_C( 31.60), SIMDE_FLOAT32_C( -15.30), SIMDE_FLOAT32_C( -77.20)),
2036
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -72.10), SIMDE_FLOAT32_C( 13.20), SIMDE_FLOAT32_C( 17.70), SIMDE_FLOAT32_C( -65.30)),
2037
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 16.00), SIMDE_FLOAT32_C( -22.20), SIMDE_FLOAT32_C( -70.50), SIMDE_FLOAT32_C(-4512.02)) }
2038
+ };
2039
+
2040
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
2041
+ simde__m128 r = simde_mm_fnmadd_ss(test_vec[i].a, test_vec[i].b, test_vec[i].c);
2042
+ simde_assert_m128_close(r, test_vec[i].r, 1);
2043
+ }
2044
+
2045
+ return MUNIT_OK;
2046
+ }
2047
+
2048
+ static MunitResult
2049
+ test_simde_mm_fnmsub_pd(const MunitParameter params[], void* data) {
2050
+ (void) params;
2051
+ (void) data;
2052
+
2053
+ const struct {
2054
+ simde__m128d a;
2055
+ simde__m128d b;
2056
+ simde__m128d c;
2057
+ simde__m128d r;
2058
+ } test_vec[8] = {
2059
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( -85.20), SIMDE_FLOAT64_C( -77.60)),
2060
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 22.30), SIMDE_FLOAT64_C( 10.00)),
2061
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 40.90), SIMDE_FLOAT64_C( 66.70)),
2062
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 1859.06), SIMDE_FLOAT64_C( 709.30)) },
2063
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( 12.10), SIMDE_FLOAT64_C( -42.90)),
2064
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -71.60), SIMDE_FLOAT64_C( -43.70)),
2065
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 73.80), SIMDE_FLOAT64_C( -65.70)),
2066
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 792.56), SIMDE_FLOAT64_C(-1809.03)) },
2067
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( 47.90), SIMDE_FLOAT64_C( 8.70)),
2068
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 4.00), SIMDE_FLOAT64_C( -70.80)),
2069
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -58.60), SIMDE_FLOAT64_C( -21.30)),
2070
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -133.00), SIMDE_FLOAT64_C( 637.26)) },
2071
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( 70.80), SIMDE_FLOAT64_C( -62.90)),
2072
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 22.60), SIMDE_FLOAT64_C( -27.80)),
2073
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 58.00), SIMDE_FLOAT64_C( 35.60)),
2074
+ simde_mm_set_pd(SIMDE_FLOAT64_C(-1658.08), SIMDE_FLOAT64_C(-1784.22)) },
2075
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( -59.60), SIMDE_FLOAT64_C( -26.20)),
2076
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 10.40), SIMDE_FLOAT64_C( -3.80)),
2077
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -36.80), SIMDE_FLOAT64_C( -20.60)),
2078
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 656.64), SIMDE_FLOAT64_C( -78.96)) },
2079
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( 75.80), SIMDE_FLOAT64_C( -40.50)),
2080
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 82.60), SIMDE_FLOAT64_C( 14.90)),
2081
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -69.00), SIMDE_FLOAT64_C( 52.40)),
2082
+ simde_mm_set_pd(SIMDE_FLOAT64_C(-6192.08), SIMDE_FLOAT64_C( 551.05)) },
2083
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( -54.60), SIMDE_FLOAT64_C( 2.00)),
2084
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 17.90), SIMDE_FLOAT64_C( 72.80)),
2085
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 95.70), SIMDE_FLOAT64_C( 56.80)),
2086
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 881.64), SIMDE_FLOAT64_C( -202.40)) },
2087
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( 21.40), SIMDE_FLOAT64_C( 40.20)),
2088
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 42.50), SIMDE_FLOAT64_C( 29.10)),
2089
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 24.40), SIMDE_FLOAT64_C( -57.60)),
2090
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -933.90), SIMDE_FLOAT64_C(-1112.22)) }
2091
+ };
2092
+
2093
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
2094
+ simde__m128d r = simde_mm_fnmsub_pd(test_vec[i].a, test_vec[i].b, test_vec[i].c);
2095
+ simde_assert_m128d_close(r, test_vec[i].r, 1);
2096
+ }
2097
+
2098
+ return MUNIT_OK;
2099
+ }
2100
+
2101
+ static MunitResult
2102
+ test_simde_mm256_fnmsub_pd(const MunitParameter params[], void* data) {
2103
+ (void) params;
2104
+ (void) data;
2105
+
2106
+ const struct {
2107
+ simde__m256d a;
2108
+ simde__m256d b;
2109
+ simde__m256d c;
2110
+ simde__m256d r;
2111
+ } test_vec[8] = {
2112
+ { simde_mm256_set_pd(SIMDE_FLOAT64_C( -97.30), SIMDE_FLOAT64_C( 40.60),
2113
+ SIMDE_FLOAT64_C( -78.70), SIMDE_FLOAT64_C( 0.60)),
2114
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( 43.40), SIMDE_FLOAT64_C( -67.40),
2115
+ SIMDE_FLOAT64_C( 62.50), SIMDE_FLOAT64_C( -5.70)),
2116
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( 51.10), SIMDE_FLOAT64_C( 66.40),
2117
+ SIMDE_FLOAT64_C( 79.40), SIMDE_FLOAT64_C( 4.20)),
2118
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( 4171.72), SIMDE_FLOAT64_C( 2670.04),
2119
+ SIMDE_FLOAT64_C( 4839.35), SIMDE_FLOAT64_C( -0.78)) },
2120
+ { simde_mm256_set_pd(SIMDE_FLOAT64_C( -37.90), SIMDE_FLOAT64_C( -91.10),
2121
+ SIMDE_FLOAT64_C( 99.40), SIMDE_FLOAT64_C( -64.60)),
2122
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( 55.70), SIMDE_FLOAT64_C( -31.70),
2123
+ SIMDE_FLOAT64_C( 33.10), SIMDE_FLOAT64_C( 94.30)),
2124
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( 18.80), SIMDE_FLOAT64_C( 56.10),
2125
+ SIMDE_FLOAT64_C( -19.80), SIMDE_FLOAT64_C( -98.50)),
2126
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( 2092.23), SIMDE_FLOAT64_C(-2943.97),
2127
+ SIMDE_FLOAT64_C(-3270.34), SIMDE_FLOAT64_C( 6190.28)) },
2128
+ { simde_mm256_set_pd(SIMDE_FLOAT64_C( 44.20), SIMDE_FLOAT64_C( 98.00),
2129
+ SIMDE_FLOAT64_C( -20.60), SIMDE_FLOAT64_C( 99.20)),
2130
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( -36.50), SIMDE_FLOAT64_C( 37.70),
2131
+ SIMDE_FLOAT64_C( 27.10), SIMDE_FLOAT64_C( -85.00)),
2132
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( -54.80), SIMDE_FLOAT64_C( 46.70),
2133
+ SIMDE_FLOAT64_C( -59.70), SIMDE_FLOAT64_C( -80.00)),
2134
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( 1668.10), SIMDE_FLOAT64_C(-3741.30),
2135
+ SIMDE_FLOAT64_C( 617.96), SIMDE_FLOAT64_C( 8512.00)) },
2136
+ { simde_mm256_set_pd(SIMDE_FLOAT64_C( -13.40), SIMDE_FLOAT64_C( 16.00),
2137
+ SIMDE_FLOAT64_C( -82.10), SIMDE_FLOAT64_C( 27.40)),
2138
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( -41.30), SIMDE_FLOAT64_C( 84.40),
2139
+ SIMDE_FLOAT64_C( -52.10), SIMDE_FLOAT64_C( 16.60)),
2140
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( 7.30), SIMDE_FLOAT64_C( -49.40),
2141
+ SIMDE_FLOAT64_C( -31.90), SIMDE_FLOAT64_C( 69.30)),
2142
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( -560.72), SIMDE_FLOAT64_C(-1301.00),
2143
+ SIMDE_FLOAT64_C(-4245.51), SIMDE_FLOAT64_C( -524.14)) },
2144
+ { simde_mm256_set_pd(SIMDE_FLOAT64_C( -16.80), SIMDE_FLOAT64_C( -78.00),
2145
+ SIMDE_FLOAT64_C( -43.90), SIMDE_FLOAT64_C( -53.60)),
2146
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( -11.30), SIMDE_FLOAT64_C( -83.60),
2147
+ SIMDE_FLOAT64_C( -78.30), SIMDE_FLOAT64_C( -1.10)),
2148
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( -94.20), SIMDE_FLOAT64_C( 36.20),
2149
+ SIMDE_FLOAT64_C( 66.40), SIMDE_FLOAT64_C( 12.70)),
2150
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( -95.64), SIMDE_FLOAT64_C(-6557.00),
2151
+ SIMDE_FLOAT64_C(-3503.77), SIMDE_FLOAT64_C( -71.66)) },
2152
+ { simde_mm256_set_pd(SIMDE_FLOAT64_C( 94.80), SIMDE_FLOAT64_C( 27.60),
2153
+ SIMDE_FLOAT64_C( 5.70), SIMDE_FLOAT64_C( -73.90)),
2154
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( 11.80), SIMDE_FLOAT64_C( -83.40),
2155
+ SIMDE_FLOAT64_C( 89.00), SIMDE_FLOAT64_C( 39.10)),
2156
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( 61.70), SIMDE_FLOAT64_C( 98.90),
2157
+ SIMDE_FLOAT64_C( -6.00), SIMDE_FLOAT64_C( -89.20)),
2158
+ simde_mm256_set_pd(SIMDE_FLOAT64_C(-1180.34), SIMDE_FLOAT64_C( 2202.94),
2159
+ SIMDE_FLOAT64_C( -501.30), SIMDE_FLOAT64_C( 2978.69)) },
2160
+ { simde_mm256_set_pd(SIMDE_FLOAT64_C( -66.00), SIMDE_FLOAT64_C( -99.10),
2161
+ SIMDE_FLOAT64_C( -51.20), SIMDE_FLOAT64_C( 98.20)),
2162
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( 58.10), SIMDE_FLOAT64_C( -66.70),
2163
+ SIMDE_FLOAT64_C( -86.20), SIMDE_FLOAT64_C( 25.30)),
2164
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( -71.40), SIMDE_FLOAT64_C( 40.80),
2165
+ SIMDE_FLOAT64_C( -71.40), SIMDE_FLOAT64_C( 8.90)),
2166
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( 3906.00), SIMDE_FLOAT64_C(-6650.77),
2167
+ SIMDE_FLOAT64_C(-4342.04), SIMDE_FLOAT64_C(-2493.36)) },
2168
+ { simde_mm256_set_pd(SIMDE_FLOAT64_C( 63.90), SIMDE_FLOAT64_C( 7.50),
2169
+ SIMDE_FLOAT64_C( -0.00), SIMDE_FLOAT64_C( -97.90)),
2170
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( -39.10), SIMDE_FLOAT64_C( -73.10),
2171
+ SIMDE_FLOAT64_C( -53.20), SIMDE_FLOAT64_C( 81.20)),
2172
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( -32.20), SIMDE_FLOAT64_C( 71.70),
2173
+ SIMDE_FLOAT64_C( 39.30), SIMDE_FLOAT64_C( -11.60)),
2174
+ simde_mm256_set_pd(SIMDE_FLOAT64_C( 2530.69), SIMDE_FLOAT64_C( 476.55),
2175
+ SIMDE_FLOAT64_C( -39.30), SIMDE_FLOAT64_C( 7961.08)) }
2176
+ };
2177
+
2178
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
2179
+ simde__m256d r = simde_mm256_fnmsub_pd(test_vec[i].a, test_vec[i].b, test_vec[i].c);
2180
+ simde_assert_m256d_close(r, test_vec[i].r, 1);
2181
+ }
2182
+
2183
+ return MUNIT_OK;
2184
+ }
2185
+
2186
+ static MunitResult
2187
+ test_simde_mm_fnmsub_ps(const MunitParameter params[], void* data) {
2188
+ (void) params;
2189
+ (void) data;
2190
+
2191
+ const struct {
2192
+ simde__m128 a;
2193
+ simde__m128 b;
2194
+ simde__m128 c;
2195
+ simde__m128 r;
2196
+ } test_vec[8] = {
2197
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( 18.60), SIMDE_FLOAT32_C( -96.60), SIMDE_FLOAT32_C( -17.10), SIMDE_FLOAT32_C( -50.50)),
2198
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 3.20), SIMDE_FLOAT32_C( -15.90), SIMDE_FLOAT32_C( 83.80), SIMDE_FLOAT32_C( -57.70)),
2199
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -50.90), SIMDE_FLOAT32_C( -53.70), SIMDE_FLOAT32_C( 66.30), SIMDE_FLOAT32_C( 53.50)),
2200
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -8.62), SIMDE_FLOAT32_C(-1482.24), SIMDE_FLOAT32_C( 1366.68), SIMDE_FLOAT32_C(-2967.35)) },
2201
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( 19.40), SIMDE_FLOAT32_C( 81.00), SIMDE_FLOAT32_C( 0.70), SIMDE_FLOAT32_C( 26.20)),
2202
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -29.00), SIMDE_FLOAT32_C( 15.30), SIMDE_FLOAT32_C( -89.70), SIMDE_FLOAT32_C( -71.80)),
2203
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 61.60), SIMDE_FLOAT32_C( -84.00), SIMDE_FLOAT32_C( -77.60), SIMDE_FLOAT32_C( 49.30)),
2204
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 501.00), SIMDE_FLOAT32_C(-1155.30), SIMDE_FLOAT32_C( 140.39), SIMDE_FLOAT32_C( 1831.86)) },
2205
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( 10.90), SIMDE_FLOAT32_C( -73.80), SIMDE_FLOAT32_C( -37.10), SIMDE_FLOAT32_C( 92.30)),
2206
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 27.80), SIMDE_FLOAT32_C( 31.90), SIMDE_FLOAT32_C( -77.70), SIMDE_FLOAT32_C( -29.80)),
2207
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 72.70), SIMDE_FLOAT32_C( 50.20), SIMDE_FLOAT32_C( -64.40), SIMDE_FLOAT32_C( 81.90)),
2208
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -375.72), SIMDE_FLOAT32_C( 2304.02), SIMDE_FLOAT32_C(-2818.27), SIMDE_FLOAT32_C( 2668.64)) },
2209
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( -2.50), SIMDE_FLOAT32_C( -77.00), SIMDE_FLOAT32_C( -97.10), SIMDE_FLOAT32_C( -6.20)),
2210
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 1.40), SIMDE_FLOAT32_C( 38.10), SIMDE_FLOAT32_C( 96.80), SIMDE_FLOAT32_C( -90.40)),
2211
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -31.70), SIMDE_FLOAT32_C( -86.40), SIMDE_FLOAT32_C( -62.20), SIMDE_FLOAT32_C( -64.70)),
2212
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 35.20), SIMDE_FLOAT32_C( 3020.10), SIMDE_FLOAT32_C( 9461.48), SIMDE_FLOAT32_C( -495.78)) },
2213
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( -29.50), SIMDE_FLOAT32_C( -45.60), SIMDE_FLOAT32_C( -87.90), SIMDE_FLOAT32_C( -82.00)),
2214
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -16.40), SIMDE_FLOAT32_C( -50.10), SIMDE_FLOAT32_C( -30.70), SIMDE_FLOAT32_C( -73.60)),
2215
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 49.20), SIMDE_FLOAT32_C( 55.00), SIMDE_FLOAT32_C( 57.30), SIMDE_FLOAT32_C( -33.30)),
2216
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -533.00), SIMDE_FLOAT32_C(-2339.56), SIMDE_FLOAT32_C(-2755.83), SIMDE_FLOAT32_C(-6001.90)) },
2217
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( -90.10), SIMDE_FLOAT32_C( 83.90), SIMDE_FLOAT32_C( -87.40), SIMDE_FLOAT32_C( -87.30)),
2218
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -1.40), SIMDE_FLOAT32_C( -10.10), SIMDE_FLOAT32_C( 29.30), SIMDE_FLOAT32_C( -74.80)),
2219
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -21.90), SIMDE_FLOAT32_C( 46.80), SIMDE_FLOAT32_C( -76.50), SIMDE_FLOAT32_C( -94.40)),
2220
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -104.24), SIMDE_FLOAT32_C( 800.59), SIMDE_FLOAT32_C( 2637.32), SIMDE_FLOAT32_C(-6435.64)) },
2221
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( -29.30), SIMDE_FLOAT32_C( -94.30), SIMDE_FLOAT32_C( -8.20), SIMDE_FLOAT32_C( -67.90)),
2222
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -39.50), SIMDE_FLOAT32_C( 47.60), SIMDE_FLOAT32_C( 50.70), SIMDE_FLOAT32_C( 19.40)),
2223
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 2.50), SIMDE_FLOAT32_C( 40.50), SIMDE_FLOAT32_C( -73.30), SIMDE_FLOAT32_C( 7.40)),
2224
+ simde_mm_set_ps(SIMDE_FLOAT32_C(-1159.85), SIMDE_FLOAT32_C( 4448.18), SIMDE_FLOAT32_C( 489.04), SIMDE_FLOAT32_C( 1309.86)) },
2225
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( 15.30), SIMDE_FLOAT32_C( 14.60), SIMDE_FLOAT32_C( -68.80), SIMDE_FLOAT32_C( 92.80)),
2226
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -34.50), SIMDE_FLOAT32_C( 77.40), SIMDE_FLOAT32_C( 73.70), SIMDE_FLOAT32_C( -25.10)),
2227
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -70.20), SIMDE_FLOAT32_C( -4.40), SIMDE_FLOAT32_C( -93.70), SIMDE_FLOAT32_C( 16.30)),
2228
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 598.05), SIMDE_FLOAT32_C(-1125.64), SIMDE_FLOAT32_C( 5164.26), SIMDE_FLOAT32_C( 2312.98)) }
2229
+ };
2230
+
2231
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
2232
+ simde__m128 r = simde_mm_fnmsub_ps(test_vec[i].a, test_vec[i].b, test_vec[i].c);
2233
+ simde_assert_m128_close(r, test_vec[i].r, 1);
2234
+ }
2235
+
2236
+ return MUNIT_OK;
2237
+ }
2238
+
2239
+ static MunitResult
2240
+ test_simde_mm256_fnmsub_ps(const MunitParameter params[], void* data) {
2241
+ (void) params;
2242
+ (void) data;
2243
+
2244
+ const struct {
2245
+ simde__m256 a;
2246
+ simde__m256 b;
2247
+ simde__m256 c;
2248
+ simde__m256 r;
2249
+ } test_vec[8] = {
2250
+ { simde_mm256_set_ps(SIMDE_FLOAT32_C( -91.80), SIMDE_FLOAT32_C( -53.10),
2251
+ SIMDE_FLOAT32_C( -79.10), SIMDE_FLOAT32_C( 50.50),
2252
+ SIMDE_FLOAT32_C( -81.20), SIMDE_FLOAT32_C( -11.90),
2253
+ SIMDE_FLOAT32_C( -72.60), SIMDE_FLOAT32_C( 13.60)),
2254
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( 11.90), SIMDE_FLOAT32_C( -8.00),
2255
+ SIMDE_FLOAT32_C( 73.10), SIMDE_FLOAT32_C( 73.00),
2256
+ SIMDE_FLOAT32_C( -15.70), SIMDE_FLOAT32_C( 33.70),
2257
+ SIMDE_FLOAT32_C( -36.30), SIMDE_FLOAT32_C( -25.80)),
2258
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( -90.00), SIMDE_FLOAT32_C( -13.00),
2259
+ SIMDE_FLOAT32_C( -28.10), SIMDE_FLOAT32_C( -49.40),
2260
+ SIMDE_FLOAT32_C( -74.60), SIMDE_FLOAT32_C( -32.00),
2261
+ SIMDE_FLOAT32_C( -63.50), SIMDE_FLOAT32_C( -18.90)),
2262
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( 1182.42), SIMDE_FLOAT32_C( -411.80),
2263
+ SIMDE_FLOAT32_C( 5810.31), SIMDE_FLOAT32_C(-3637.10),
2264
+ SIMDE_FLOAT32_C(-1200.24), SIMDE_FLOAT32_C( 433.03),
2265
+ SIMDE_FLOAT32_C(-2571.88), SIMDE_FLOAT32_C( 369.78)) },
2266
+ { simde_mm256_set_ps(SIMDE_FLOAT32_C( -57.30), SIMDE_FLOAT32_C( 71.50),
2267
+ SIMDE_FLOAT32_C( 39.90), SIMDE_FLOAT32_C( -77.10),
2268
+ SIMDE_FLOAT32_C( -9.90), SIMDE_FLOAT32_C( -16.00),
2269
+ SIMDE_FLOAT32_C( 74.80), SIMDE_FLOAT32_C( 77.40)),
2270
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( -53.60), SIMDE_FLOAT32_C( 49.60),
2271
+ SIMDE_FLOAT32_C( 94.20), SIMDE_FLOAT32_C( 1.20),
2272
+ SIMDE_FLOAT32_C( -56.30), SIMDE_FLOAT32_C( 26.10),
2273
+ SIMDE_FLOAT32_C( -23.40), SIMDE_FLOAT32_C( -47.00)),
2274
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( 82.40), SIMDE_FLOAT32_C( -13.50),
2275
+ SIMDE_FLOAT32_C( -97.40), SIMDE_FLOAT32_C( 84.50),
2276
+ SIMDE_FLOAT32_C( -48.30), SIMDE_FLOAT32_C( 98.50),
2277
+ SIMDE_FLOAT32_C( -91.50), SIMDE_FLOAT32_C( 24.60)),
2278
+ simde_mm256_set_ps(SIMDE_FLOAT32_C(-3153.68), SIMDE_FLOAT32_C(-3532.90),
2279
+ SIMDE_FLOAT32_C(-3661.18), SIMDE_FLOAT32_C( 8.02),
2280
+ SIMDE_FLOAT32_C( -509.07), SIMDE_FLOAT32_C( 319.10),
2281
+ SIMDE_FLOAT32_C( 1841.82), SIMDE_FLOAT32_C( 3613.20)) },
2282
+ { simde_mm256_set_ps(SIMDE_FLOAT32_C( 11.60), SIMDE_FLOAT32_C( 59.90),
2283
+ SIMDE_FLOAT32_C( -3.20), SIMDE_FLOAT32_C( 4.40),
2284
+ SIMDE_FLOAT32_C( -98.80), SIMDE_FLOAT32_C( 29.00),
2285
+ SIMDE_FLOAT32_C( -86.20), SIMDE_FLOAT32_C( 19.50)),
2286
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( 63.60), SIMDE_FLOAT32_C( -94.60),
2287
+ SIMDE_FLOAT32_C( -81.40), SIMDE_FLOAT32_C( 9.90),
2288
+ SIMDE_FLOAT32_C( -69.00), SIMDE_FLOAT32_C( -83.90),
2289
+ SIMDE_FLOAT32_C( 22.00), SIMDE_FLOAT32_C( -56.30)),
2290
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( 80.90), SIMDE_FLOAT32_C( -7.90),
2291
+ SIMDE_FLOAT32_C( -92.10), SIMDE_FLOAT32_C( 65.40),
2292
+ SIMDE_FLOAT32_C( -26.30), SIMDE_FLOAT32_C( -26.90),
2293
+ SIMDE_FLOAT32_C( -44.20), SIMDE_FLOAT32_C( -39.60)),
2294
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( -818.66), SIMDE_FLOAT32_C( 5674.44),
2295
+ SIMDE_FLOAT32_C( -168.38), SIMDE_FLOAT32_C( -108.96),
2296
+ SIMDE_FLOAT32_C(-6790.90), SIMDE_FLOAT32_C( 2460.00),
2297
+ SIMDE_FLOAT32_C( 1940.60), SIMDE_FLOAT32_C( 1137.45)) },
2298
+ { simde_mm256_set_ps(SIMDE_FLOAT32_C( -73.70), SIMDE_FLOAT32_C( -39.20),
2299
+ SIMDE_FLOAT32_C( 40.90), SIMDE_FLOAT32_C( 0.60),
2300
+ SIMDE_FLOAT32_C( -64.50), SIMDE_FLOAT32_C( 35.70),
2301
+ SIMDE_FLOAT32_C( -58.10), SIMDE_FLOAT32_C( -23.80)),
2302
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( 98.30), SIMDE_FLOAT32_C( 77.60),
2303
+ SIMDE_FLOAT32_C( 33.80), SIMDE_FLOAT32_C( 94.20),
2304
+ SIMDE_FLOAT32_C( 8.60), SIMDE_FLOAT32_C( -96.70),
2305
+ SIMDE_FLOAT32_C( -22.70), SIMDE_FLOAT32_C( -38.00)),
2306
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( -52.20), SIMDE_FLOAT32_C( -35.80),
2307
+ SIMDE_FLOAT32_C( 76.20), SIMDE_FLOAT32_C( -32.30),
2308
+ SIMDE_FLOAT32_C( -84.60), SIMDE_FLOAT32_C( 76.00),
2309
+ SIMDE_FLOAT32_C( -84.30), SIMDE_FLOAT32_C( 87.60)),
2310
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( 7296.91), SIMDE_FLOAT32_C( 3077.72),
2311
+ SIMDE_FLOAT32_C(-1458.62), SIMDE_FLOAT32_C( -24.22),
2312
+ SIMDE_FLOAT32_C( 639.30), SIMDE_FLOAT32_C( 3376.19),
2313
+ SIMDE_FLOAT32_C(-1234.57), SIMDE_FLOAT32_C( -992.00)) },
2314
+ { simde_mm256_set_ps(SIMDE_FLOAT32_C( -10.70), SIMDE_FLOAT32_C( 52.50),
2315
+ SIMDE_FLOAT32_C( 95.50), SIMDE_FLOAT32_C( -35.90),
2316
+ SIMDE_FLOAT32_C( -55.60), SIMDE_FLOAT32_C( 1.10),
2317
+ SIMDE_FLOAT32_C( -20.80), SIMDE_FLOAT32_C( -55.50)),
2318
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( -20.50), SIMDE_FLOAT32_C( 25.80),
2319
+ SIMDE_FLOAT32_C( 85.10), SIMDE_FLOAT32_C( -30.10),
2320
+ SIMDE_FLOAT32_C( 98.50), SIMDE_FLOAT32_C( -42.90),
2321
+ SIMDE_FLOAT32_C( 14.30), SIMDE_FLOAT32_C( 52.60)),
2322
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( 81.70), SIMDE_FLOAT32_C( 21.40),
2323
+ SIMDE_FLOAT32_C( 41.10), SIMDE_FLOAT32_C( 65.30),
2324
+ SIMDE_FLOAT32_C( -66.60), SIMDE_FLOAT32_C( 6.20),
2325
+ SIMDE_FLOAT32_C( 29.60), SIMDE_FLOAT32_C( 47.20)),
2326
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( -301.05), SIMDE_FLOAT32_C(-1375.90),
2327
+ SIMDE_FLOAT32_C(-8168.15), SIMDE_FLOAT32_C(-1145.89),
2328
+ SIMDE_FLOAT32_C( 5543.20), SIMDE_FLOAT32_C( 40.99),
2329
+ SIMDE_FLOAT32_C( 267.84), SIMDE_FLOAT32_C( 2872.10)) },
2330
+ { simde_mm256_set_ps(SIMDE_FLOAT32_C( -94.70), SIMDE_FLOAT32_C( 14.30),
2331
+ SIMDE_FLOAT32_C( 36.30), SIMDE_FLOAT32_C( -95.40),
2332
+ SIMDE_FLOAT32_C( -85.70), SIMDE_FLOAT32_C( 15.60),
2333
+ SIMDE_FLOAT32_C( -45.20), SIMDE_FLOAT32_C( -87.50)),
2334
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( -75.50), SIMDE_FLOAT32_C( 52.00),
2335
+ SIMDE_FLOAT32_C( 88.60), SIMDE_FLOAT32_C( -12.10),
2336
+ SIMDE_FLOAT32_C( -27.40), SIMDE_FLOAT32_C( 41.00),
2337
+ SIMDE_FLOAT32_C( -70.80), SIMDE_FLOAT32_C( 22.60)),
2338
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( -14.10), SIMDE_FLOAT32_C( -90.60),
2339
+ SIMDE_FLOAT32_C( 84.80), SIMDE_FLOAT32_C( -47.50),
2340
+ SIMDE_FLOAT32_C( -49.90), SIMDE_FLOAT32_C( 72.50),
2341
+ SIMDE_FLOAT32_C( 90.90), SIMDE_FLOAT32_C( -74.60)),
2342
+ simde_mm256_set_ps(SIMDE_FLOAT32_C(-7135.75), SIMDE_FLOAT32_C( -653.00),
2343
+ SIMDE_FLOAT32_C(-3300.98), SIMDE_FLOAT32_C(-1106.84),
2344
+ SIMDE_FLOAT32_C(-2298.28), SIMDE_FLOAT32_C( -712.10),
2345
+ SIMDE_FLOAT32_C(-3291.06), SIMDE_FLOAT32_C( 2052.10)) },
2346
+ { simde_mm256_set_ps(SIMDE_FLOAT32_C( -59.20), SIMDE_FLOAT32_C( -79.60),
2347
+ SIMDE_FLOAT32_C( 47.00), SIMDE_FLOAT32_C( -96.90),
2348
+ SIMDE_FLOAT32_C( -44.60), SIMDE_FLOAT32_C( 50.20),
2349
+ SIMDE_FLOAT32_C( 10.60), SIMDE_FLOAT32_C( -70.80)),
2350
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( -68.90), SIMDE_FLOAT32_C( 37.70),
2351
+ SIMDE_FLOAT32_C( 58.60), SIMDE_FLOAT32_C( -25.80),
2352
+ SIMDE_FLOAT32_C( 57.80), SIMDE_FLOAT32_C( -89.20),
2353
+ SIMDE_FLOAT32_C( 27.50), SIMDE_FLOAT32_C( 46.60)),
2354
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( 61.30), SIMDE_FLOAT32_C( -66.60),
2355
+ SIMDE_FLOAT32_C( 75.60), SIMDE_FLOAT32_C( -6.00),
2356
+ SIMDE_FLOAT32_C( -95.90), SIMDE_FLOAT32_C( 11.80),
2357
+ SIMDE_FLOAT32_C( 59.10), SIMDE_FLOAT32_C( 34.90)),
2358
+ simde_mm256_set_ps(SIMDE_FLOAT32_C(-4140.18), SIMDE_FLOAT32_C( 3067.52),
2359
+ SIMDE_FLOAT32_C(-2829.80), SIMDE_FLOAT32_C(-2494.02),
2360
+ SIMDE_FLOAT32_C( 2673.78), SIMDE_FLOAT32_C( 4466.04),
2361
+ SIMDE_FLOAT32_C( -350.60), SIMDE_FLOAT32_C( 3264.38)) },
2362
+ { simde_mm256_set_ps(SIMDE_FLOAT32_C( -53.00), SIMDE_FLOAT32_C( -46.10),
2363
+ SIMDE_FLOAT32_C( 53.90), SIMDE_FLOAT32_C( 19.20),
2364
+ SIMDE_FLOAT32_C( -73.10), SIMDE_FLOAT32_C( 23.40),
2365
+ SIMDE_FLOAT32_C( 67.90), SIMDE_FLOAT32_C( -74.50)),
2366
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( -80.10), SIMDE_FLOAT32_C( 56.30),
2367
+ SIMDE_FLOAT32_C( -45.20), SIMDE_FLOAT32_C( 32.20),
2368
+ SIMDE_FLOAT32_C( -17.90), SIMDE_FLOAT32_C( -44.50),
2369
+ SIMDE_FLOAT32_C( 62.10), SIMDE_FLOAT32_C( 8.00)),
2370
+ simde_mm256_set_ps(SIMDE_FLOAT32_C( 30.70), SIMDE_FLOAT32_C( 11.70),
2371
+ SIMDE_FLOAT32_C( -61.10), SIMDE_FLOAT32_C( 76.30),
2372
+ SIMDE_FLOAT32_C( 57.20), SIMDE_FLOAT32_C( 36.40),
2373
+ SIMDE_FLOAT32_C( 67.50), SIMDE_FLOAT32_C( 2.90)),
2374
+ simde_mm256_set_ps(SIMDE_FLOAT32_C(-4276.00), SIMDE_FLOAT32_C( 2583.73),
2375
+ SIMDE_FLOAT32_C( 2497.38), SIMDE_FLOAT32_C( -694.54),
2376
+ SIMDE_FLOAT32_C(-1365.69), SIMDE_FLOAT32_C( 1004.90),
2377
+ SIMDE_FLOAT32_C(-4284.09), SIMDE_FLOAT32_C( 593.10)) }
2378
+ };
2379
+
2380
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
2381
+ simde__m256 r = simde_mm256_fnmsub_ps(test_vec[i].a, test_vec[i].b, test_vec[i].c);
2382
+ simde_assert_m256_close(r, test_vec[i].r, 1);
2383
+ }
2384
+
2385
+ return MUNIT_OK;
2386
+ }
2387
+
2388
+ static MunitResult
2389
+ test_simde_mm_fnmsub_sd(const MunitParameter params[], void* data) {
2390
+ (void) params;
2391
+ (void) data;
2392
+
2393
+ const struct {
2394
+ simde__m128d a;
2395
+ simde__m128d b;
2396
+ simde__m128d c;
2397
+ simde__m128d r;
2398
+ } test_vec[8] = {
2399
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( -60.10), SIMDE_FLOAT64_C( -84.00)),
2400
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -22.80), SIMDE_FLOAT64_C( 63.20)),
2401
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 81.10), SIMDE_FLOAT64_C( -77.20)),
2402
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -60.10), SIMDE_FLOAT64_C( 5386.00)) },
2403
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( 37.60), SIMDE_FLOAT64_C( 78.90)),
2404
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -10.00), SIMDE_FLOAT64_C( 53.60)),
2405
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -88.70), SIMDE_FLOAT64_C( 54.90)),
2406
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 37.60), SIMDE_FLOAT64_C(-4283.94)) },
2407
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( -38.20), SIMDE_FLOAT64_C( 72.10)),
2408
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -86.10), SIMDE_FLOAT64_C( 25.50)),
2409
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -52.90), SIMDE_FLOAT64_C( -86.30)),
2410
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -38.20), SIMDE_FLOAT64_C(-1752.25)) },
2411
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( 7.50), SIMDE_FLOAT64_C( 35.00)),
2412
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 5.30), SIMDE_FLOAT64_C( 97.00)),
2413
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -49.40), SIMDE_FLOAT64_C( -58.60)),
2414
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 7.50), SIMDE_FLOAT64_C(-3336.40)) },
2415
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( 41.40), SIMDE_FLOAT64_C( -46.90)),
2416
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -70.70), SIMDE_FLOAT64_C( -78.10)),
2417
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 77.70), SIMDE_FLOAT64_C( -33.20)),
2418
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 41.40), SIMDE_FLOAT64_C(-3629.69)) },
2419
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( -63.30), SIMDE_FLOAT64_C( -78.40)),
2420
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 58.80), SIMDE_FLOAT64_C( 11.50)),
2421
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -18.00), SIMDE_FLOAT64_C( -49.30)),
2422
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -63.30), SIMDE_FLOAT64_C( 950.90)) },
2423
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( -80.50), SIMDE_FLOAT64_C( 28.50)),
2424
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -48.00), SIMDE_FLOAT64_C( 38.10)),
2425
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -73.40), SIMDE_FLOAT64_C( -29.00)),
2426
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -80.50), SIMDE_FLOAT64_C(-1056.85)) },
2427
+ { simde_mm_set_pd(SIMDE_FLOAT64_C( 42.80), SIMDE_FLOAT64_C( -10.90)),
2428
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 8.60), SIMDE_FLOAT64_C( -39.60)),
2429
+ simde_mm_set_pd(SIMDE_FLOAT64_C( -25.50), SIMDE_FLOAT64_C( 42.10)),
2430
+ simde_mm_set_pd(SIMDE_FLOAT64_C( 42.80), SIMDE_FLOAT64_C( -473.74)) }
2431
+ };
2432
+
2433
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
2434
+ simde__m128d r = simde_mm_fnmsub_sd(test_vec[i].a, test_vec[i].b, test_vec[i].c);
2435
+ simde_assert_m128d_close(r, test_vec[i].r, 1);
2436
+ }
2437
+
2438
+ return MUNIT_OK;
2439
+ }
2440
+
2441
+ static MunitResult
2442
+ test_simde_mm_fnmsub_ss(const MunitParameter params[], void* data) {
2443
+ (void) params;
2444
+ (void) data;
2445
+
2446
+ const struct {
2447
+ simde__m128 a;
2448
+ simde__m128 b;
2449
+ simde__m128 c;
2450
+ simde__m128 r;
2451
+ } test_vec[8] = {
2452
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( 17.30), SIMDE_FLOAT32_C( 17.40), SIMDE_FLOAT32_C( 41.70), SIMDE_FLOAT32_C( 37.00)),
2453
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 61.00), SIMDE_FLOAT32_C( 15.10), SIMDE_FLOAT32_C( 2.00), SIMDE_FLOAT32_C( 43.30)),
2454
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 28.00), SIMDE_FLOAT32_C( 83.70), SIMDE_FLOAT32_C( 43.30), SIMDE_FLOAT32_C( -38.20)),
2455
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 17.30), SIMDE_FLOAT32_C( 17.40), SIMDE_FLOAT32_C( 41.70), SIMDE_FLOAT32_C(-1563.90)) },
2456
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( -37.00), SIMDE_FLOAT32_C( -28.20), SIMDE_FLOAT32_C( 12.60), SIMDE_FLOAT32_C( -73.50)),
2457
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 5.90), SIMDE_FLOAT32_C( 68.10), SIMDE_FLOAT32_C( 57.10), SIMDE_FLOAT32_C( 23.80)),
2458
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 17.40), SIMDE_FLOAT32_C( 89.40), SIMDE_FLOAT32_C( 38.60), SIMDE_FLOAT32_C( -36.50)),
2459
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -37.00), SIMDE_FLOAT32_C( -28.20), SIMDE_FLOAT32_C( 12.60), SIMDE_FLOAT32_C( 1785.80)) },
2460
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( 35.60), SIMDE_FLOAT32_C( -64.00), SIMDE_FLOAT32_C( 95.10), SIMDE_FLOAT32_C( -83.40)),
2461
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 90.40), SIMDE_FLOAT32_C( 58.10), SIMDE_FLOAT32_C( -8.40), SIMDE_FLOAT32_C( -87.90)),
2462
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 26.50), SIMDE_FLOAT32_C( -91.50), SIMDE_FLOAT32_C( 38.20), SIMDE_FLOAT32_C( 39.20)),
2463
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 35.60), SIMDE_FLOAT32_C( -64.00), SIMDE_FLOAT32_C( 95.10), SIMDE_FLOAT32_C(-7370.06)) },
2464
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( 26.00), SIMDE_FLOAT32_C( 35.10), SIMDE_FLOAT32_C( 90.70), SIMDE_FLOAT32_C( -77.20)),
2465
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -18.70), SIMDE_FLOAT32_C( 97.20), SIMDE_FLOAT32_C( -13.90), SIMDE_FLOAT32_C( 3.60)),
2466
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -66.00), SIMDE_FLOAT32_C( -38.90), SIMDE_FLOAT32_C( 92.90), SIMDE_FLOAT32_C( 44.20)),
2467
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 26.00), SIMDE_FLOAT32_C( 35.10), SIMDE_FLOAT32_C( 90.70), SIMDE_FLOAT32_C( 233.72)) },
2468
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( 63.90), SIMDE_FLOAT32_C( -84.10), SIMDE_FLOAT32_C( 20.70), SIMDE_FLOAT32_C( -87.00)),
2469
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 2.30), SIMDE_FLOAT32_C( -39.50), SIMDE_FLOAT32_C( -17.30), SIMDE_FLOAT32_C( -98.60)),
2470
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -59.10), SIMDE_FLOAT32_C( -12.50), SIMDE_FLOAT32_C( 12.60), SIMDE_FLOAT32_C( 34.90)),
2471
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 63.90), SIMDE_FLOAT32_C( -84.10), SIMDE_FLOAT32_C( 20.70), SIMDE_FLOAT32_C(-8613.10)) },
2472
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( -37.30), SIMDE_FLOAT32_C( -17.50), SIMDE_FLOAT32_C( -37.30), SIMDE_FLOAT32_C( -7.90)),
2473
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 85.10), SIMDE_FLOAT32_C( -93.00), SIMDE_FLOAT32_C( -6.70), SIMDE_FLOAT32_C( 16.00)),
2474
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 17.50), SIMDE_FLOAT32_C( -83.60), SIMDE_FLOAT32_C( 98.60), SIMDE_FLOAT32_C( -20.50)),
2475
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -37.30), SIMDE_FLOAT32_C( -17.50), SIMDE_FLOAT32_C( -37.30), SIMDE_FLOAT32_C( 146.90)) },
2476
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( -98.40), SIMDE_FLOAT32_C( 46.60), SIMDE_FLOAT32_C( -57.20), SIMDE_FLOAT32_C( -62.70)),
2477
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 23.90), SIMDE_FLOAT32_C( 59.10), SIMDE_FLOAT32_C( 62.20), SIMDE_FLOAT32_C( 48.60)),
2478
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -80.80), SIMDE_FLOAT32_C( -51.00), SIMDE_FLOAT32_C( 63.40), SIMDE_FLOAT32_C( 30.10)),
2479
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -98.40), SIMDE_FLOAT32_C( 46.60), SIMDE_FLOAT32_C( -57.20), SIMDE_FLOAT32_C( 3017.12)) },
2480
+ { simde_mm_set_ps(SIMDE_FLOAT32_C( -52.30), SIMDE_FLOAT32_C( 90.80), SIMDE_FLOAT32_C( 10.20), SIMDE_FLOAT32_C( 40.80)),
2481
+ simde_mm_set_ps(SIMDE_FLOAT32_C( 25.10), SIMDE_FLOAT32_C( -1.00), SIMDE_FLOAT32_C( 38.80), SIMDE_FLOAT32_C( 1.50)),
2482
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -21.30), SIMDE_FLOAT32_C( -30.30), SIMDE_FLOAT32_C( 80.90), SIMDE_FLOAT32_C( -98.10)),
2483
+ simde_mm_set_ps(SIMDE_FLOAT32_C( -52.30), SIMDE_FLOAT32_C( 90.80), SIMDE_FLOAT32_C( 10.20), SIMDE_FLOAT32_C( 36.90)) }
2484
+ };
2485
+
2486
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
2487
+ simde__m128 r = simde_mm_fnmsub_ss(test_vec[i].a, test_vec[i].b, test_vec[i].c);
2488
+ simde_assert_m128_close(r, test_vec[i].r, 1);
2489
+ }
2490
+
2491
+ return MUNIT_OK;
2492
+ }
2493
+
2494
+ #endif /* defined(SIMDE_FMA_NATIVE) || defined(SIMDE_NO_NATIVE) || defined(SIMDE_ALWAYS_BUILD_NATIVE_TESTS) */
2495
+
2496
+ HEDLEY_DIAGNOSTIC_PUSH
2497
+ HEDLEY_DIAGNOSTIC_DISABLE_CAST_QUAL
2498
+
2499
+ static MunitTest test_suite_tests[] = {
2500
+ #if defined(SIMDE_FMA_NATIVE) || defined(SIMDE_NO_NATIVE) || defined(SIMDE_ALWAYS_BUILD_NATIVE_TESTS)
2501
+ SIMDE_TESTS_DEFINE_TEST(mm_fmadd_pd),
2502
+ SIMDE_TESTS_DEFINE_TEST(mm256_fmadd_pd),
2503
+ SIMDE_TESTS_DEFINE_TEST(mm_fmadd_ps),
2504
+ SIMDE_TESTS_DEFINE_TEST(mm256_fmadd_ps),
2505
+ SIMDE_TESTS_DEFINE_TEST(mm_fmadd_sd),
2506
+ SIMDE_TESTS_DEFINE_TEST(mm_fmadd_ss),
2507
+
2508
+ SIMDE_TESTS_DEFINE_TEST(mm_fmaddsub_pd),
2509
+ SIMDE_TESTS_DEFINE_TEST(mm256_fmaddsub_pd),
2510
+ SIMDE_TESTS_DEFINE_TEST(mm_fmaddsub_ps),
2511
+ SIMDE_TESTS_DEFINE_TEST(mm256_fmaddsub_ps),
2512
+
2513
+ SIMDE_TESTS_DEFINE_TEST(mm_fmsub_pd),
2514
+ SIMDE_TESTS_DEFINE_TEST(mm256_fmsub_pd),
2515
+ SIMDE_TESTS_DEFINE_TEST(mm_fmsub_ps),
2516
+ SIMDE_TESTS_DEFINE_TEST(mm256_fmsub_ps),
2517
+ SIMDE_TESTS_DEFINE_TEST(mm_fmsub_sd),
2518
+ SIMDE_TESTS_DEFINE_TEST(mm_fmsub_ss),
2519
+
2520
+ SIMDE_TESTS_DEFINE_TEST(mm_fmsubadd_pd),
2521
+ SIMDE_TESTS_DEFINE_TEST(mm256_fmsubadd_pd),
2522
+ SIMDE_TESTS_DEFINE_TEST(mm_fmsubadd_ps),
2523
+ SIMDE_TESTS_DEFINE_TEST(mm256_fmsubadd_ps),
2524
+
2525
+ SIMDE_TESTS_DEFINE_TEST(mm_fnmadd_pd),
2526
+ SIMDE_TESTS_DEFINE_TEST(mm256_fnmadd_pd),
2527
+ SIMDE_TESTS_DEFINE_TEST(mm_fnmadd_ps),
2528
+ SIMDE_TESTS_DEFINE_TEST(mm256_fnmadd_ps),
2529
+ SIMDE_TESTS_DEFINE_TEST(mm_fnmadd_sd),
2530
+ SIMDE_TESTS_DEFINE_TEST(mm_fnmadd_ss),
2531
+
2532
+ SIMDE_TESTS_DEFINE_TEST(mm_fnmsub_pd),
2533
+ SIMDE_TESTS_DEFINE_TEST(mm256_fnmsub_pd),
2534
+ SIMDE_TESTS_DEFINE_TEST(mm_fnmsub_ps),
2535
+ SIMDE_TESTS_DEFINE_TEST(mm256_fnmsub_ps),
2536
+ SIMDE_TESTS_DEFINE_TEST(mm_fnmsub_sd),
2537
+ SIMDE_TESTS_DEFINE_TEST(mm_fnmsub_ss),
2538
+ #endif /* defined(SIMDE_FMA_NATIVE) || defined(SIMDE_NO_NATIVE) || defined(SIMDE_ALWAYS_BUILD_NATIVE_TESTS) */
2539
+
2540
+ { NULL, NULL, NULL, NULL, MUNIT_TEST_OPTION_NONE, NULL }
2541
+ };
2542
+
2543
+ HEDLEY_C_DECL MunitSuite* SIMDE_TESTS_GENERATE_SYMBOL(suite)(void) {
2544
+ static MunitSuite suite = { (char*) "/" HEDLEY_STRINGIFY(SIMDE_TESTS_CURRENT_ISAX), test_suite_tests, NULL, 1, MUNIT_SUITE_OPTION_NONE };
2545
+
2546
+ return &suite;
2547
+ }
2548
+
2549
+ #if defined(SIMDE_TESTS_SINGLE_ISAX)
2550
+ int main(int argc, char* argv[HEDLEY_ARRAY_PARAM(argc + 1)]) {
2551
+ static MunitSuite suite = { "", test_suite_tests, NULL, 1, MUNIT_SUITE_OPTION_NONE };
2552
+
2553
+ return munit_suite_main(&suite, NULL, argc, argv);
2554
+ }
2555
+ #endif /* defined(SIMDE_TESTS_SINGLE_ISAX) */
2556
+
2557
+ HEDLEY_DIAGNOSTIC_POP