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 +1,1260 @@
1
+ /* Copyright (c) 2019 Evan Nemerson <evan@nemerson.com>
2
+ *
3
+ * Permission is hereby granted, free of charge, to any person
4
+ * obtaining a copy of this software and associated documentation
5
+ * files (the "Software"), to deal in the Software without
6
+ * restriction, including without limitation the rights to use, copy,
7
+ * modify, merge, publish, distribute, sublicense, and/or sell copies
8
+ * of the Software, and to permit persons to whom the Software is
9
+ * furnished to do so, subject to the following conditions:
10
+ *
11
+ * The above copyright notice and this permission notice shall be
12
+ * included in all copies or substantial portions of the Software.
13
+ *
14
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
15
+ * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
16
+ * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
17
+ * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
18
+ * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
19
+ * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
20
+ * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
21
+ * SOFTWARE.
22
+ */
23
+
24
+ #define SIMDE_TESTS_CURRENT_NEON_OP sub
25
+ #include <test/arm/neon/test-neon-internal.h>
26
+ #include <simde/arm/neon.h>
27
+
28
+ #if defined(SIMDE_NEON_NATIVE) || defined(SIMDE_NO_NATIVE) || defined(SIMDE_ALWAYS_BUILD_NATIVE_TESTS)
29
+
30
+ static MunitResult
31
+ test_simde_vsub_s8(const MunitParameter params[], void* data) {
32
+ (void) params;
33
+ (void) data;
34
+
35
+ const struct {
36
+ simde_int8x8_t a;
37
+ simde_int8x8_t b;
38
+ simde_int8x8_t r;
39
+ } test_vec[8] = {
40
+ { simde_x_vload_s8(INT8_C( -37), INT8_C( -36), INT8_C( 2), INT8_C( -96),
41
+ INT8_C( -72), INT8_C( -93), INT8_C( 9), INT8_C( -71)),
42
+ simde_x_vload_s8(INT8_C( 96), INT8_C( 106), INT8_C( 75), INT8_C( -96),
43
+ INT8_C( -4), INT8_C( 44), INT8_C( 38), INT8_C( -45)),
44
+ simde_x_vload_s8(INT8_C( 123), INT8_C( 114), INT8_C( -73), INT8_C( 0),
45
+ INT8_C( -68), INT8_C( 119), INT8_C( -29), INT8_C( -26)) },
46
+ { simde_x_vload_s8(INT8_C( -5), INT8_C( -33), INT8_C( -73), INT8_C( 22),
47
+ INT8_C( -99), INT8_C( 42), INT8_C( -15), INT8_C( -34)),
48
+ simde_x_vload_s8(INT8_C( -51), INT8_C( 115), INT8_C( -77), INT8_C( 61),
49
+ INT8_C( -66), INT8_C( -1), INT8_C( 14), INT8_C( 29)),
50
+ simde_x_vload_s8(INT8_C( 46), INT8_C( 108), INT8_C( 4), INT8_C( -39),
51
+ INT8_C( -33), INT8_C( 43), INT8_C( -29), INT8_C( -63)) },
52
+ { simde_x_vload_s8(INT8_C(-128), INT8_C( -74), INT8_C( 5), INT8_C( 28),
53
+ INT8_C( 66), INT8_C( 15), INT8_C( 84), INT8_C( 57)),
54
+ simde_x_vload_s8(INT8_C( 85), INT8_C( -7), INT8_C(-111), INT8_C( 127),
55
+ INT8_C( 29), INT8_C( -36), INT8_C( -49), INT8_C( -64)),
56
+ simde_x_vload_s8(INT8_C( 43), INT8_C( -67), INT8_C( 116), INT8_C( -99),
57
+ INT8_C( 37), INT8_C( 51), INT8_C(-123), INT8_C( 121)) },
58
+ { simde_x_vload_s8(INT8_C( 6), INT8_C( 80), INT8_C( 98), INT8_C( -1),
59
+ INT8_C( -72), INT8_C( 127), INT8_C( -21), INT8_C( 102)),
60
+ simde_x_vload_s8(INT8_C( -66), INT8_C( 79), INT8_C( -10), INT8_C( 116),
61
+ INT8_C(-109), INT8_C( -89), INT8_C( 86), INT8_C( -70)),
62
+ simde_x_vload_s8(INT8_C( 72), INT8_C( 1), INT8_C( 108), INT8_C(-117),
63
+ INT8_C( 37), INT8_C( -40), INT8_C(-107), INT8_C( -84)) },
64
+ { simde_x_vload_s8(INT8_C( 44), INT8_C( 7), INT8_C( 4), INT8_C( -21),
65
+ INT8_C( 60), INT8_C( 93), INT8_C( 58), INT8_C( 62)),
66
+ simde_x_vload_s8(INT8_C( -23), INT8_C( 75), INT8_C( -67), INT8_C( 92),
67
+ INT8_C( 17), INT8_C( 100), INT8_C( 61), INT8_C( 28)),
68
+ simde_x_vload_s8(INT8_C( 67), INT8_C( -68), INT8_C( 71), INT8_C(-113),
69
+ INT8_C( 43), INT8_C( -7), INT8_C( -3), INT8_C( 34)) },
70
+ { simde_x_vload_s8(INT8_C( -94), INT8_C( -96), INT8_C( -5), INT8_C( -70),
71
+ INT8_C( 60), INT8_C( 114), INT8_C( -44), INT8_C( -9)),
72
+ simde_x_vload_s8(INT8_C(-122), INT8_C( 42), INT8_C( 92), INT8_C( 96),
73
+ INT8_C( 37), INT8_C( 121), INT8_C( 9), INT8_C( -91)),
74
+ simde_x_vload_s8(INT8_C( 28), INT8_C( 118), INT8_C( -97), INT8_C( 90),
75
+ INT8_C( 23), INT8_C( -7), INT8_C( -53), INT8_C( 82)) },
76
+ { simde_x_vload_s8(INT8_C( 28), INT8_C( -81), INT8_C( -40), INT8_C(-110),
77
+ INT8_C( 63), INT8_C( 82), INT8_C( -67), INT8_C( 7)),
78
+ simde_x_vload_s8(INT8_C( 119), INT8_C( 6), INT8_C( 56), INT8_C( 11),
79
+ INT8_C( -13), INT8_C( 100), INT8_C(-108), INT8_C(-115)),
80
+ simde_x_vload_s8(INT8_C( -91), INT8_C( -87), INT8_C( -96), INT8_C(-121),
81
+ INT8_C( 76), INT8_C( -18), INT8_C( 41), INT8_C( 122)) },
82
+ { simde_x_vload_s8(INT8_C( 106), INT8_C( -74), INT8_C( 24), INT8_C( -79),
83
+ INT8_C( -4), INT8_C( -55), INT8_C( 90), INT8_C( 10)),
84
+ simde_x_vload_s8(INT8_C( 80), INT8_C( 78), INT8_C( 34), INT8_C( -48),
85
+ INT8_C( 99), INT8_C(-115), INT8_C( 91), INT8_C( 66)),
86
+ simde_x_vload_s8(INT8_C( 26), INT8_C( 104), INT8_C( -10), INT8_C( -31),
87
+ INT8_C(-103), INT8_C( 60), INT8_C( -1), INT8_C( -56)) }
88
+ };
89
+
90
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
91
+ simde_int8x8_t r = simde_vsub_s8(test_vec[i].a, test_vec[i].b);
92
+ simde_neon_assert_int8x8(r, ==, test_vec[i].r);
93
+ }
94
+
95
+ return MUNIT_OK;
96
+ }
97
+
98
+ static MunitResult
99
+ test_simde_vsub_s16(const MunitParameter params[], void* data) {
100
+ (void) params;
101
+ (void) data;
102
+
103
+ const struct {
104
+ simde_int16x4_t a;
105
+ simde_int16x4_t b;
106
+ simde_int16x4_t r;
107
+ } test_vec[8] = {
108
+ { simde_x_vload_s16(INT16_C(-14111), INT16_C( 18824), INT16_C(-21849), INT16_C( 4616)),
109
+ simde_x_vload_s16(INT16_C( 5795), INT16_C( 7143), INT16_C( 5812), INT16_C( 196)),
110
+ simde_x_vload_s16(INT16_C(-19906), INT16_C( 11681), INT16_C(-27661), INT16_C( 4420)) },
111
+ { simde_x_vload_s16(INT16_C( 22250), INT16_C( -1040), INT16_C( 15496), INT16_C( -2604)),
112
+ simde_x_vload_s16(INT16_C( 27298), INT16_C( 16152), INT16_C( -1313), INT16_C( 31117)),
113
+ simde_x_vload_s16(INT16_C( -5048), INT16_C(-17192), INT16_C( 16809), INT16_C( 31815)) },
114
+ { simde_x_vload_s16(INT16_C( 16667), INT16_C( -6715), INT16_C( -3312), INT16_C( 14904)),
115
+ simde_x_vload_s16(INT16_C(-13864), INT16_C( 18124), INT16_C( 31980), INT16_C(-16639)),
116
+ simde_x_vload_s16(INT16_C( 30531), INT16_C(-24839), INT16_C( 30244), INT16_C( 31543)) },
117
+ { simde_x_vload_s16(INT16_C(-13624), INT16_C( 15318), INT16_C(-16753), INT16_C( 5361)),
118
+ simde_x_vload_s16(INT16_C( -997), INT16_C( 7636), INT16_C( 23352), INT16_C( -4061)),
119
+ simde_x_vload_s16(INT16_C(-12627), INT16_C( 7682), INT16_C( 25431), INT16_C( 9422)) },
120
+ { simde_x_vload_s16(INT16_C( 17298), INT16_C( 20617), INT16_C(-15704), INT16_C( 20998)),
121
+ simde_x_vload_s16(INT16_C( -5243), INT16_C(-23055), INT16_C( 11286), INT16_C( 22862)),
122
+ simde_x_vload_s16(INT16_C( 22541), INT16_C(-21864), INT16_C(-26990), INT16_C( -1864)) },
123
+ { simde_x_vload_s16(INT16_C( -8125), INT16_C( 17253), INT16_C(-22390), INT16_C( -1004)),
124
+ simde_x_vload_s16(INT16_C( 20202), INT16_C( 22564), INT16_C( 26065), INT16_C( 31272)),
125
+ simde_x_vload_s16(INT16_C(-28327), INT16_C( -5311), INT16_C( 17081), INT16_C(-32276)) },
126
+ { simde_x_vload_s16(INT16_C( 19932), INT16_C( 743), INT16_C( 22659), INT16_C( 15104)),
127
+ simde_x_vload_s16(INT16_C( 11972), INT16_C( 6408), INT16_C(-27074), INT16_C( 4793)),
128
+ simde_x_vload_s16(INT16_C( 7960), INT16_C( -5665), INT16_C(-15803), INT16_C( 10311)) },
129
+ { simde_x_vload_s16(INT16_C( 13977), INT16_C( 18898), INT16_C(-13590), INT16_C(-20807)),
130
+ simde_x_vload_s16(INT16_C(-25522), INT16_C( -7382), INT16_C(-22841), INT16_C( -5401)),
131
+ simde_x_vload_s16(INT16_C(-26037), INT16_C( 26280), INT16_C( 9251), INT16_C(-15406)) }
132
+ };
133
+
134
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
135
+ simde_int16x4_t r = simde_vsub_s16(test_vec[i].a, test_vec[i].b);
136
+ simde_neon_assert_int16x4(r, ==, test_vec[i].r);
137
+ }
138
+
139
+ return MUNIT_OK;
140
+ }
141
+
142
+ static MunitResult
143
+ test_simde_vsub_s32(const MunitParameter params[], void* data) {
144
+ (void) params;
145
+ (void) data;
146
+
147
+ const struct {
148
+ simde_int32x2_t a;
149
+ simde_int32x2_t b;
150
+ simde_int32x2_t r;
151
+ } test_vec[8] = {
152
+ { simde_x_vload_s32(INT32_C( 1984799789), INT32_C( 1614293491)),
153
+ simde_x_vload_s32(INT32_C( 1366009077), INT32_C( 624537206)),
154
+ simde_x_vload_s32(INT32_C( 618790712), INT32_C( 989756285)) },
155
+ { simde_x_vload_s32(INT32_C(-1594895836), INT32_C( 2102087776)),
156
+ simde_x_vload_s32(INT32_C( 1894603626), INT32_C(-1774211407)),
157
+ simde_x_vload_s32(INT32_C( 805467834), INT32_C( -418668113)) },
158
+ { simde_x_vload_s32(INT32_C(-1108372257), INT32_C( 578916545)),
159
+ simde_x_vload_s32(INT32_C( 717783548), INT32_C( 747324151)),
160
+ simde_x_vload_s32(INT32_C(-1826155805), INT32_C( -168407606)) },
161
+ { simde_x_vload_s32(INT32_C( -494477761), INT32_C( 45737703)),
162
+ simde_x_vload_s32(INT32_C( 1384114837), INT32_C( 1768395276)),
163
+ simde_x_vload_s32(INT32_C(-1878592598), INT32_C(-1722657573)) },
164
+ { simde_x_vload_s32(INT32_C(-1928892032), INT32_C( -625136713)),
165
+ simde_x_vload_s32(INT32_C( -608697091), INT32_C(-1499675239)),
166
+ simde_x_vload_s32(INT32_C(-1320194941), INT32_C( 874538526)) },
167
+ { simde_x_vload_s32(INT32_C( -818130826), INT32_C( 1761083523)),
168
+ simde_x_vload_s32(INT32_C( 747143995), INT32_C( 982923146)),
169
+ simde_x_vload_s32(INT32_C(-1565274821), INT32_C( 778160377)) },
170
+ { simde_x_vload_s32(INT32_C( 727287038), INT32_C( 1476178736)),
171
+ simde_x_vload_s32(INT32_C( 869955011), INT32_C( 1649209025)),
172
+ simde_x_vload_s32(INT32_C( -142667973), INT32_C( -173030289)) },
173
+ { simde_x_vload_s32(INT32_C( 340781678), INT32_C(-1368501231)),
174
+ simde_x_vload_s32(INT32_C( 1930912879), INT32_C( 1045548067)),
175
+ simde_x_vload_s32(INT32_C(-1590131201), INT32_C( 1880917998)) }
176
+ };
177
+
178
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
179
+ simde_int32x2_t r = simde_vsub_s32(test_vec[i].a, test_vec[i].b);
180
+ simde_neon_assert_int32x2(r, ==, test_vec[i].r);
181
+ }
182
+
183
+ return MUNIT_OK;
184
+ }
185
+
186
+ static MunitResult
187
+ test_simde_vsub_s64(const MunitParameter params[], void* data) {
188
+ (void) params;
189
+ (void) data;
190
+
191
+ const struct {
192
+ simde_int64x1_t a;
193
+ simde_int64x1_t b;
194
+ simde_int64x1_t r;
195
+ } test_vec[8] = {
196
+ { simde_x_vload_s64(INT64_C(-4943373233697329301)),
197
+ simde_x_vload_s64(INT64_C( 3512059779770080645)),
198
+ simde_x_vload_s64(INT64_C(-8455433013467409946)) },
199
+ { simde_x_vload_s64(INT64_C(-8185513873329872926)),
200
+ simde_x_vload_s64(INT64_C( 7795236760724609678)),
201
+ simde_x_vload_s64(INT64_C( 2465993439655069012)) },
202
+ { simde_x_vload_s64(INT64_C(-1794875855984181891)),
203
+ simde_x_vload_s64(INT64_C(-5326491736934413717)),
204
+ simde_x_vload_s64(INT64_C( 3531615880950231826)) },
205
+ { simde_x_vload_s64(INT64_C( -136055898301259445)),
206
+ simde_x_vload_s64(INT64_C( 970507860556016427)),
207
+ simde_x_vload_s64(INT64_C(-1106563758857275872)) },
208
+ { simde_x_vload_s64(INT64_C( 2444272399882168566)),
209
+ simde_x_vload_s64(INT64_C(-5872539651416703481)),
210
+ simde_x_vload_s64(INT64_C( 8316812051298872047)) },
211
+ { simde_x_vload_s64(INT64_C( 4263823998790759071)),
212
+ simde_x_vload_s64(INT64_C( 7412649410322713882)),
213
+ simde_x_vload_s64(INT64_C(-3148825411531954811)) },
214
+ { simde_x_vload_s64(INT64_C( 8294241550524916467)),
215
+ simde_x_vload_s64(INT64_C( 3520541369967028956)),
216
+ simde_x_vload_s64(INT64_C( 4773700180557887511)) },
217
+ { simde_x_vload_s64(INT64_C(-7306887266544301894)),
218
+ simde_x_vload_s64(INT64_C(-1889250389483863155)),
219
+ simde_x_vload_s64(INT64_C(-5417636877060438739)) }
220
+ };
221
+
222
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
223
+ simde_int64x1_t r = simde_vsub_s64(test_vec[i].a, test_vec[i].b);
224
+ simde_neon_assert_int64x1(r, ==, test_vec[i].r);
225
+ }
226
+
227
+ return MUNIT_OK;
228
+ }
229
+
230
+ static MunitResult
231
+ test_simde_vsub_u8(const MunitParameter params[], void* data) {
232
+ (void) params;
233
+ (void) data;
234
+
235
+ const struct {
236
+ simde_uint8x8_t a;
237
+ simde_uint8x8_t b;
238
+ simde_uint8x8_t r;
239
+ } test_vec[8] = {
240
+ { simde_x_vload_u8(UINT8_C(191), UINT8_C( 81), UINT8_C(127), UINT8_C(142),
241
+ UINT8_C(249), UINT8_C( 25), UINT8_C(123), UINT8_C( 26)),
242
+ simde_x_vload_u8(UINT8_C( 57), UINT8_C(151), UINT8_C( 54), UINT8_C( 76),
243
+ UINT8_C(173), UINT8_C(145), UINT8_C(101), UINT8_C(119)),
244
+ simde_x_vload_u8(UINT8_C(134), UINT8_C(186), UINT8_C( 73), UINT8_C( 66),
245
+ UINT8_C( 76), UINT8_C(136), UINT8_C( 22), UINT8_C(163)) },
246
+ { simde_x_vload_u8(UINT8_C( 25), UINT8_C( 14), UINT8_C(148), UINT8_C(196),
247
+ UINT8_C(146), UINT8_C(203), UINT8_C( 88), UINT8_C( 77)),
248
+ simde_x_vload_u8(UINT8_C( 81), UINT8_C(233), UINT8_C(126), UINT8_C( 6),
249
+ UINT8_C(206), UINT8_C( 23), UINT8_C( 44), UINT8_C(181)),
250
+ simde_x_vload_u8(UINT8_C(200), UINT8_C( 37), UINT8_C( 22), UINT8_C(190),
251
+ UINT8_C(196), UINT8_C(180), UINT8_C( 44), UINT8_C(152)) },
252
+ { simde_x_vload_u8(UINT8_C(233), UINT8_C( 64), UINT8_C( 17), UINT8_C(225),
253
+ UINT8_C(133), UINT8_C(156), UINT8_C( 24), UINT8_C( 95)),
254
+ simde_x_vload_u8(UINT8_C(198), UINT8_C(250), UINT8_C(135), UINT8_C( 74),
255
+ UINT8_C(254), UINT8_C(137), UINT8_C(162), UINT8_C(168)),
256
+ simde_x_vload_u8(UINT8_C( 35), UINT8_C( 70), UINT8_C(138), UINT8_C(151),
257
+ UINT8_C(135), UINT8_C( 19), UINT8_C(118), UINT8_C(183)) },
258
+ { simde_x_vload_u8(UINT8_C(107), UINT8_C(212), UINT8_C(123), UINT8_C( 50),
259
+ UINT8_C(138), UINT8_C( 37), UINT8_C( 83), UINT8_C(189)),
260
+ simde_x_vload_u8(UINT8_C(221), UINT8_C( 69), UINT8_C(171), UINT8_C(113),
261
+ UINT8_C(118), UINT8_C(148), UINT8_C(175), UINT8_C( 83)),
262
+ simde_x_vload_u8(UINT8_C(142), UINT8_C(143), UINT8_C(208), UINT8_C(193),
263
+ UINT8_C( 20), UINT8_C(145), UINT8_C(164), UINT8_C(106)) },
264
+ { simde_x_vload_u8(UINT8_C(147), UINT8_C( 46), UINT8_C( 36), UINT8_C(247),
265
+ UINT8_C( 68), UINT8_C( 59), UINT8_C( 79), UINT8_C(177)),
266
+ simde_x_vload_u8(UINT8_C(146), UINT8_C(254), UINT8_C(116), UINT8_C(172),
267
+ UINT8_C( 23), UINT8_C(132), UINT8_C( 55), UINT8_C( 35)),
268
+ simde_x_vload_u8(UINT8_C( 1), UINT8_C( 48), UINT8_C(176), UINT8_C( 75),
269
+ UINT8_C( 45), UINT8_C(183), UINT8_C( 24), UINT8_C(142)) },
270
+ { simde_x_vload_u8(UINT8_C( 73), UINT8_C(145), UINT8_C(242), UINT8_C(203),
271
+ UINT8_C(228), UINT8_C( 74), UINT8_C( 0), UINT8_C(236)),
272
+ simde_x_vload_u8(UINT8_C(157), UINT8_C(243), UINT8_C(159), UINT8_C( 43),
273
+ UINT8_C(247), UINT8_C( 45), UINT8_C(197), UINT8_C(248)),
274
+ simde_x_vload_u8(UINT8_C(172), UINT8_C(158), UINT8_C( 83), UINT8_C(160),
275
+ UINT8_C(237), UINT8_C( 29), UINT8_C( 59), UINT8_C(244)) },
276
+ { simde_x_vload_u8(UINT8_C( 55), UINT8_C( 2), UINT8_C( 33), UINT8_C(112),
277
+ UINT8_C(107), UINT8_C( 62), UINT8_C( 81), UINT8_C( 25)),
278
+ simde_x_vload_u8(UINT8_C( 5), UINT8_C( 46), UINT8_C( 9), UINT8_C( 55),
279
+ UINT8_C(137), UINT8_C(243), UINT8_C( 41), UINT8_C( 14)),
280
+ simde_x_vload_u8(UINT8_C( 50), UINT8_C(212), UINT8_C( 24), UINT8_C( 57),
281
+ UINT8_C(226), UINT8_C( 75), UINT8_C( 40), UINT8_C( 11)) },
282
+ { simde_x_vload_u8(UINT8_C(214), UINT8_C( 35), UINT8_C( 24), UINT8_C(152),
283
+ UINT8_C(249), UINT8_C(211), UINT8_C( 34), UINT8_C(148)),
284
+ simde_x_vload_u8(UINT8_C(132), UINT8_C(174), UINT8_C(198), UINT8_C(119),
285
+ UINT8_C( 15), UINT8_C( 53), UINT8_C(142), UINT8_C(101)),
286
+ simde_x_vload_u8(UINT8_C( 82), UINT8_C(117), UINT8_C( 82), UINT8_C( 33),
287
+ UINT8_C(234), UINT8_C(158), UINT8_C(148), UINT8_C( 47)) }
288
+ };
289
+
290
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
291
+ simde_uint8x8_t r = simde_vsub_u8(test_vec[i].a, test_vec[i].b);
292
+ simde_neon_assert_uint8x8(r, ==, test_vec[i].r);
293
+ }
294
+
295
+ return MUNIT_OK;
296
+ }
297
+
298
+ static MunitResult
299
+ test_simde_vsub_u16(const MunitParameter params[], void* data) {
300
+ (void) params;
301
+ (void) data;
302
+
303
+ const struct {
304
+ simde_uint16x4_t a;
305
+ simde_uint16x4_t b;
306
+ simde_uint16x4_t r;
307
+ } test_vec[8] = {
308
+ { simde_x_vload_u16(UINT16_C(48223), UINT16_C(50867), UINT16_C(48082), UINT16_C( 2665)),
309
+ simde_x_vload_u16(UINT16_C(59993), UINT16_C(42266), UINT16_C(10432), UINT16_C(44082)),
310
+ simde_x_vload_u16(UINT16_C(53766), UINT16_C( 8601), UINT16_C(37650), UINT16_C(24119)) },
311
+ { simde_x_vload_u16(UINT16_C(24941), UINT16_C(47948), UINT16_C(11907), UINT16_C(36303)),
312
+ simde_x_vload_u16(UINT16_C( 7815), UINT16_C(10383), UINT16_C(29405), UINT16_C(61652)),
313
+ simde_x_vload_u16(UINT16_C(17126), UINT16_C(37565), UINT16_C(48038), UINT16_C(40187)) },
314
+ { simde_x_vload_u16(UINT16_C(28462), UINT16_C( 9248), UINT16_C( 6523), UINT16_C(25838)),
315
+ simde_x_vload_u16(UINT16_C(30404), UINT16_C(21387), UINT16_C(51331), UINT16_C(27661)),
316
+ simde_x_vload_u16(UINT16_C(63594), UINT16_C(53397), UINT16_C(20728), UINT16_C(63713)) },
317
+ { simde_x_vload_u16(UINT16_C( 9599), UINT16_C( 6357), UINT16_C(47695), UINT16_C(27046)),
318
+ simde_x_vload_u16(UINT16_C(38129), UINT16_C(64014), UINT16_C(48035), UINT16_C(13887)),
319
+ simde_x_vload_u16(UINT16_C(37006), UINT16_C( 7879), UINT16_C(65196), UINT16_C(13159)) },
320
+ { simde_x_vload_u16(UINT16_C(44687), UINT16_C(17981), UINT16_C(31332), UINT16_C(60426)),
321
+ simde_x_vload_u16(UINT16_C(52812), UINT16_C(44247), UINT16_C(38502), UINT16_C(30153)),
322
+ simde_x_vload_u16(UINT16_C(57411), UINT16_C(39270), UINT16_C(58366), UINT16_C(30273)) },
323
+ { simde_x_vload_u16(UINT16_C(42610), UINT16_C(21966), UINT16_C(41660), UINT16_C( 1960)),
324
+ simde_x_vload_u16(UINT16_C(17066), UINT16_C(26127), UINT16_C( 1932), UINT16_C(39429)),
325
+ simde_x_vload_u16(UINT16_C(25544), UINT16_C(61375), UINT16_C(39728), UINT16_C(28067)) },
326
+ { simde_x_vload_u16(UINT16_C(15534), UINT16_C( 4666), UINT16_C(23694), UINT16_C( 9679)),
327
+ simde_x_vload_u16(UINT16_C(53422), UINT16_C(13900), UINT16_C(32150), UINT16_C(31322)),
328
+ simde_x_vload_u16(UINT16_C(27648), UINT16_C(56302), UINT16_C(57080), UINT16_C(43893)) },
329
+ { simde_x_vload_u16(UINT16_C(15202), UINT16_C(61259), UINT16_C(46371), UINT16_C(24403)),
330
+ simde_x_vload_u16(UINT16_C( 2026), UINT16_C(31299), UINT16_C(13786), UINT16_C(26993)),
331
+ simde_x_vload_u16(UINT16_C(13176), UINT16_C(29960), UINT16_C(32585), UINT16_C(62946)) }
332
+ };
333
+
334
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
335
+ simde_uint16x4_t r = simde_vsub_u16(test_vec[i].a, test_vec[i].b);
336
+ simde_neon_assert_uint16x4(r, ==, test_vec[i].r);
337
+ }
338
+
339
+ return MUNIT_OK;
340
+ }
341
+
342
+ static MunitResult
343
+ test_simde_vsub_u32(const MunitParameter params[], void* data) {
344
+ (void) params;
345
+ (void) data;
346
+
347
+ const struct {
348
+ simde_uint32x2_t a;
349
+ simde_uint32x2_t b;
350
+ simde_uint32x2_t r;
351
+ } test_vec[8] = {
352
+ { simde_x_vload_u32(UINT32_C(2975214003), UINT32_C( 741963422)),
353
+ simde_x_vload_u32(UINT32_C(1106378674), UINT32_C(2001202843)),
354
+ simde_x_vload_u32(UINT32_C(1868835329), UINT32_C(3035727875)) },
355
+ { simde_x_vload_u32(UINT32_C( 631537799), UINT32_C( 133948393)),
356
+ simde_x_vload_u32(UINT32_C(2673177213), UINT32_C(2434411774)),
357
+ simde_x_vload_u32(UINT32_C(2253327882), UINT32_C(1994503915)) },
358
+ { simde_x_vload_u32(UINT32_C(2657161563), UINT32_C( 768269797)),
359
+ simde_x_vload_u32(UINT32_C(3116181497), UINT32_C(1425632443)),
360
+ simde_x_vload_u32(UINT32_C(3835947362), UINT32_C(3637604650)) },
361
+ { simde_x_vload_u32(UINT32_C( 294243551), UINT32_C( 265814871)),
362
+ simde_x_vload_u32(UINT32_C(2364431343), UINT32_C( 951230457)),
363
+ simde_x_vload_u32(UINT32_C(2224779504), UINT32_C(3609551710)) },
364
+ { simde_x_vload_u32(UINT32_C( 131983137), UINT32_C(2534419200)),
365
+ simde_x_vload_u32(UINT32_C(3297552523), UINT32_C(1710645813)),
366
+ simde_x_vload_u32(UINT32_C(1129397910), UINT32_C( 823773387)) },
367
+ { simde_x_vload_u32(UINT32_C(4165452150), UINT32_C(1232131517)),
368
+ simde_x_vload_u32(UINT32_C(1456354963), UINT32_C(3902805375)),
369
+ simde_x_vload_u32(UINT32_C(2709097187), UINT32_C(1624293438)) },
370
+ { simde_x_vload_u32(UINT32_C(1447604069), UINT32_C(3135348846)),
371
+ simde_x_vload_u32(UINT32_C(4051253168), UINT32_C(1516205751)),
372
+ simde_x_vload_u32(UINT32_C(1691318197), UINT32_C(1619143095)) },
373
+ { simde_x_vload_u32(UINT32_C(1514440368), UINT32_C(2479035141)),
374
+ simde_x_vload_u32(UINT32_C( 140215595), UINT32_C(2450662976)),
375
+ simde_x_vload_u32(UINT32_C(1374224773), UINT32_C( 28372165)) }
376
+ };
377
+
378
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
379
+ simde_uint32x2_t r = simde_vsub_u32(test_vec[i].a, test_vec[i].b);
380
+ simde_neon_assert_uint32x2(r, ==, test_vec[i].r);
381
+ }
382
+
383
+ return MUNIT_OK;
384
+ }
385
+
386
+ static MunitResult
387
+ test_simde_vsub_u64(const MunitParameter params[], void* data) {
388
+ (void) params;
389
+ (void) data;
390
+
391
+ const struct {
392
+ simde_uint64x1_t a;
393
+ simde_uint64x1_t b;
394
+ simde_uint64x1_t r;
395
+ } test_vec[8] = {
396
+ { simde_x_vload_u64(UINT64_C(1936986223633213948)),
397
+ simde_x_vload_u64(UINT64_C(8847896019390123612)),
398
+ simde_x_vload_u64(UINT64_C(11535834277952641952)) },
399
+ { simde_x_vload_u64(UINT64_C(10718644368272015670)),
400
+ simde_x_vload_u64(UINT64_C(15876979127427977051)),
401
+ simde_x_vload_u64(UINT64_C(13288409314553590235)) },
402
+ { simde_x_vload_u64(UINT64_C(7920926281338567907)),
403
+ simde_x_vload_u64(UINT64_C(5876662223356338855)),
404
+ simde_x_vload_u64(UINT64_C(2044264057982229052)) },
405
+ { simde_x_vload_u64(UINT64_C(2152545993092625936)),
406
+ simde_x_vload_u64(UINT64_C(5968089876997405873)),
407
+ simde_x_vload_u64(UINT64_C(14631200189804771679)) },
408
+ { simde_x_vload_u64(UINT64_C(5384786867744097014)),
409
+ simde_x_vload_u64(UINT64_C(5558818678212911970)),
410
+ simde_x_vload_u64(UINT64_C(18272712263240736660)) },
411
+ { simde_x_vload_u64(UINT64_C(6837335513581327967)),
412
+ simde_x_vload_u64(UINT64_C( 213032869806481304)),
413
+ simde_x_vload_u64(UINT64_C(6624302643774846663)) },
414
+ { simde_x_vload_u64(UINT64_C(12407096746524455118)),
415
+ simde_x_vload_u64(UINT64_C(9728709649850370960)),
416
+ simde_x_vload_u64(UINT64_C(2678387096674084158)) },
417
+ { simde_x_vload_u64(UINT64_C(16295407813227308533)),
418
+ simde_x_vload_u64(UINT64_C(1567153282166012823)),
419
+ simde_x_vload_u64(UINT64_C(14728254531061295710)) }
420
+ };
421
+
422
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
423
+ simde_uint64x1_t r = simde_vsub_u64(test_vec[i].a, test_vec[i].b);
424
+ simde_neon_assert_uint64x1(r, ==, test_vec[i].r);
425
+ }
426
+
427
+ return MUNIT_OK;
428
+ }
429
+
430
+ static MunitResult
431
+ test_simde_vsub_f32(const MunitParameter params[], void* data) {
432
+ (void) params;
433
+ (void) data;
434
+
435
+ const struct {
436
+ simde_float32x2_t a;
437
+ simde_float32x2_t b;
438
+ simde_float32x2_t r;
439
+ } test_vec[8] = {
440
+ { simde_x_vload_f32(SIMDE_FLOAT32_C( -679.55), SIMDE_FLOAT32_C( -380.07)),
441
+ simde_x_vload_f32(SIMDE_FLOAT32_C( -76.24), SIMDE_FLOAT32_C( 210.53)),
442
+ simde_x_vload_f32(SIMDE_FLOAT32_C( -603.31), SIMDE_FLOAT32_C( -590.60)) },
443
+ { simde_x_vload_f32(SIMDE_FLOAT32_C( 251.95), SIMDE_FLOAT32_C( 117.44)),
444
+ simde_x_vload_f32(SIMDE_FLOAT32_C( 178.33), SIMDE_FLOAT32_C( -943.22)),
445
+ simde_x_vload_f32(SIMDE_FLOAT32_C( 73.62), SIMDE_FLOAT32_C( 1060.66)) },
446
+ { simde_x_vload_f32(SIMDE_FLOAT32_C( -913.61), SIMDE_FLOAT32_C( 734.07)),
447
+ simde_x_vload_f32(SIMDE_FLOAT32_C( 239.65), SIMDE_FLOAT32_C( -254.71)),
448
+ simde_x_vload_f32(SIMDE_FLOAT32_C(-1153.26), SIMDE_FLOAT32_C( 988.78)) },
449
+ { simde_x_vload_f32(SIMDE_FLOAT32_C( -492.68), SIMDE_FLOAT32_C( 226.18)),
450
+ simde_x_vload_f32(SIMDE_FLOAT32_C( -939.53), SIMDE_FLOAT32_C( -555.21)),
451
+ simde_x_vload_f32(SIMDE_FLOAT32_C( 446.85), SIMDE_FLOAT32_C( 781.39)) },
452
+ { simde_x_vload_f32(SIMDE_FLOAT32_C( -223.46), SIMDE_FLOAT32_C( 644.31)),
453
+ simde_x_vload_f32(SIMDE_FLOAT32_C( 497.64), SIMDE_FLOAT32_C( -984.04)),
454
+ simde_x_vload_f32(SIMDE_FLOAT32_C( -721.10), SIMDE_FLOAT32_C( 1628.35)) },
455
+ { simde_x_vload_f32(SIMDE_FLOAT32_C( -731.48), SIMDE_FLOAT32_C( -32.34)),
456
+ simde_x_vload_f32(SIMDE_FLOAT32_C( -429.71), SIMDE_FLOAT32_C( 124.70)),
457
+ simde_x_vload_f32(SIMDE_FLOAT32_C( -301.77), SIMDE_FLOAT32_C( -157.04)) },
458
+ { simde_x_vload_f32(SIMDE_FLOAT32_C( -305.10), SIMDE_FLOAT32_C( 187.85)),
459
+ simde_x_vload_f32(SIMDE_FLOAT32_C( -464.37), SIMDE_FLOAT32_C( -869.02)),
460
+ simde_x_vload_f32(SIMDE_FLOAT32_C( 159.27), SIMDE_FLOAT32_C( 1056.87)) },
461
+ { simde_x_vload_f32(SIMDE_FLOAT32_C( -311.70), SIMDE_FLOAT32_C( -825.92)),
462
+ simde_x_vload_f32(SIMDE_FLOAT32_C( 836.51), SIMDE_FLOAT32_C( 663.00)),
463
+ simde_x_vload_f32(SIMDE_FLOAT32_C(-1148.21), SIMDE_FLOAT32_C(-1488.92)) }
464
+ };
465
+
466
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
467
+ simde_float32x2_t r = simde_vsub_f32(test_vec[i].a, test_vec[i].b);
468
+ simde_neon_assert_float32x2_equal(r, test_vec[i].r, 1);
469
+ }
470
+
471
+ return MUNIT_OK;
472
+ }
473
+
474
+ static MunitResult
475
+ test_simde_vsub_f64(const MunitParameter params[], void* data) {
476
+ (void) params;
477
+ (void) data;
478
+
479
+ const struct {
480
+ simde_float64x1_t a;
481
+ simde_float64x1_t b;
482
+ simde_float64x1_t r;
483
+ } test_vec[8] = {
484
+ { simde_x_vload_f64(SIMDE_FLOAT64_C( 184.84)),
485
+ simde_x_vload_f64(SIMDE_FLOAT64_C( 696.66)),
486
+ simde_x_vload_f64(SIMDE_FLOAT64_C( -511.82)) },
487
+ { simde_x_vload_f64(SIMDE_FLOAT64_C( 463.00)),
488
+ simde_x_vload_f64(SIMDE_FLOAT64_C( 280.90)),
489
+ simde_x_vload_f64(SIMDE_FLOAT64_C( 182.10)) },
490
+ { simde_x_vload_f64(SIMDE_FLOAT64_C( 443.58)),
491
+ simde_x_vload_f64(SIMDE_FLOAT64_C( -388.57)),
492
+ simde_x_vload_f64(SIMDE_FLOAT64_C( 832.15)) },
493
+ { simde_x_vload_f64(SIMDE_FLOAT64_C( -74.54)),
494
+ simde_x_vload_f64(SIMDE_FLOAT64_C( -208.53)),
495
+ simde_x_vload_f64(SIMDE_FLOAT64_C( 133.99)) },
496
+ { simde_x_vload_f64(SIMDE_FLOAT64_C( -834.57)),
497
+ simde_x_vload_f64(SIMDE_FLOAT64_C( -2.20)),
498
+ simde_x_vload_f64(SIMDE_FLOAT64_C( -832.37)) },
499
+ { simde_x_vload_f64(SIMDE_FLOAT64_C( 48.05)),
500
+ simde_x_vload_f64(SIMDE_FLOAT64_C( 657.05)),
501
+ simde_x_vload_f64(SIMDE_FLOAT64_C( -609.00)) },
502
+ { simde_x_vload_f64(SIMDE_FLOAT64_C( -559.45)),
503
+ simde_x_vload_f64(SIMDE_FLOAT64_C( -465.28)),
504
+ simde_x_vload_f64(SIMDE_FLOAT64_C( -94.17)) },
505
+ { simde_x_vload_f64(SIMDE_FLOAT64_C( -302.94)),
506
+ simde_x_vload_f64(SIMDE_FLOAT64_C( -629.36)),
507
+ simde_x_vload_f64(SIMDE_FLOAT64_C( 326.42)) }
508
+ };
509
+
510
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
511
+ simde_float64x1_t r = simde_vsub_f64(test_vec[i].a, test_vec[i].b);
512
+ simde_neon_assert_float64x1_equal(r, test_vec[i].r, 1);
513
+ }
514
+
515
+ return MUNIT_OK;
516
+ }
517
+
518
+ static MunitResult
519
+ test_simde_vsubq_s8(const MunitParameter params[], void* data) {
520
+ (void) params;
521
+ (void) data;
522
+
523
+ const struct {
524
+ simde_int8x16_t a;
525
+ simde_int8x16_t b;
526
+ simde_int8x16_t r;
527
+ } test_vec[8] = {
528
+ { simde_x_vloadq_s8(INT8_C(-125), INT8_C( 49), INT8_C( -91), INT8_C( 90),
529
+ INT8_C( 28), INT8_C( -43), INT8_C( -55), INT8_C( 78),
530
+ INT8_C(-125), INT8_C(-127), INT8_C( -32), INT8_C( -71),
531
+ INT8_C( -73), INT8_C( -28), INT8_C( -70), INT8_C(-125)),
532
+ simde_x_vloadq_s8(INT8_C( -94), INT8_C( 124), INT8_C( -70), INT8_C( -91),
533
+ INT8_C( -88), INT8_C( -40), INT8_C( 13), INT8_C( 28),
534
+ INT8_C( -67), INT8_C(-127), INT8_C( 83), INT8_C(-111),
535
+ INT8_C( 104), INT8_C( 91), INT8_C( 97), INT8_C( -60)),
536
+ simde_x_vloadq_s8(INT8_C( -31), INT8_C( -75), INT8_C( -21), INT8_C( -75),
537
+ INT8_C( 116), INT8_C( -3), INT8_C( -68), INT8_C( 50),
538
+ INT8_C( -58), INT8_C( 0), INT8_C(-115), INT8_C( 40),
539
+ INT8_C( 79), INT8_C(-119), INT8_C( 89), INT8_C( -65)) },
540
+ { simde_x_vloadq_s8(INT8_C( 80), INT8_C( 41), INT8_C( 54), INT8_C( -74),
541
+ INT8_C( 49), INT8_C( 1), INT8_C(-119), INT8_C( 27),
542
+ INT8_C( -64), INT8_C( -86), INT8_C( -55), INT8_C( -12),
543
+ INT8_C( -77), INT8_C( 93), INT8_C( -63), INT8_C(-116)),
544
+ simde_x_vloadq_s8(INT8_C(-105), INT8_C( -2), INT8_C( 58), INT8_C( -15),
545
+ INT8_C( 111), INT8_C( 88), INT8_C( -97), INT8_C( -53),
546
+ INT8_C( -24), INT8_C( -59), INT8_C( -69), INT8_C( -56),
547
+ INT8_C( 60), INT8_C( 125), INT8_C( -6), INT8_C( 113)),
548
+ simde_x_vloadq_s8(INT8_C( -71), INT8_C( 43), INT8_C( -4), INT8_C( -59),
549
+ INT8_C( -62), INT8_C( -87), INT8_C( -22), INT8_C( 80),
550
+ INT8_C( -40), INT8_C( -27), INT8_C( 14), INT8_C( 44),
551
+ INT8_C( 119), INT8_C( -32), INT8_C( -57), INT8_C( 27)) },
552
+ { simde_x_vloadq_s8(INT8_C( 84), INT8_C( -45), INT8_C( 18), INT8_C( 47),
553
+ INT8_C( -79), INT8_C( 30), INT8_C( -93), INT8_C( 53),
554
+ INT8_C(-115), INT8_C( -90), INT8_C( -5), INT8_C( 105),
555
+ INT8_C( 85), INT8_C( -95), INT8_C( -57), INT8_C( 54)),
556
+ simde_x_vloadq_s8(INT8_C( 0), INT8_C( 17), INT8_C( -39), INT8_C( 102),
557
+ INT8_C( -2), INT8_C( -94), INT8_C( 28), INT8_C( 10),
558
+ INT8_C( 110), INT8_C( 36), INT8_C( 54), INT8_C( -17),
559
+ INT8_C( 110), INT8_C( -77), INT8_C( -15), INT8_C( -82)),
560
+ simde_x_vloadq_s8(INT8_C( 84), INT8_C( -62), INT8_C( 57), INT8_C( -55),
561
+ INT8_C( -77), INT8_C( 124), INT8_C(-121), INT8_C( 43),
562
+ INT8_C( 31), INT8_C(-126), INT8_C( -59), INT8_C( 122),
563
+ INT8_C( -25), INT8_C( -18), INT8_C( -42), INT8_C(-120)) },
564
+ { simde_x_vloadq_s8(INT8_C( -95), INT8_C( -42), INT8_C( 35), INT8_C( 1),
565
+ INT8_C( 32), INT8_C( 18), INT8_C( 75), INT8_C( 123),
566
+ INT8_C( 82), INT8_C( 23), INT8_C( 27), INT8_C( -61),
567
+ INT8_C( 31), INT8_C( 102), INT8_C( -52), INT8_C( 105)),
568
+ simde_x_vloadq_s8(INT8_C( -39), INT8_C( 114), INT8_C( -84), INT8_C( 102),
569
+ INT8_C( 49), INT8_C(-119), INT8_C( 3), INT8_C(-104),
570
+ INT8_C( -37), INT8_C( -97), INT8_C( -34), INT8_C( 37),
571
+ INT8_C( 23), INT8_C( 11), INT8_C( -54), INT8_C( 29)),
572
+ simde_x_vloadq_s8(INT8_C( -56), INT8_C( 100), INT8_C( 119), INT8_C(-101),
573
+ INT8_C( -17), INT8_C(-119), INT8_C( 72), INT8_C( -29),
574
+ INT8_C( 119), INT8_C( 120), INT8_C( 61), INT8_C( -98),
575
+ INT8_C( 8), INT8_C( 91), INT8_C( 2), INT8_C( 76)) },
576
+ { simde_x_vloadq_s8(INT8_C( -93), INT8_C( 12), INT8_C( -11), INT8_C( -43),
577
+ INT8_C( 52), INT8_C(-115), INT8_C( -79), INT8_C( -77),
578
+ INT8_C( 9), INT8_C( 42), INT8_C( -88), INT8_C( -16),
579
+ INT8_C( 125), INT8_C( -73), INT8_C( 104), INT8_C( -34)),
580
+ simde_x_vloadq_s8(INT8_C( 11), INT8_C( -5), INT8_C( 62), INT8_C( -46),
581
+ INT8_C(-127), INT8_C( 81), INT8_C( 54), INT8_C( -56),
582
+ INT8_C( -83), INT8_C( 97), INT8_C( 81), INT8_C( -3),
583
+ INT8_C( -10), INT8_C( -56), INT8_C( 1), INT8_C( -80)),
584
+ simde_x_vloadq_s8(INT8_C(-104), INT8_C( 17), INT8_C( -73), INT8_C( 3),
585
+ INT8_C( -77), INT8_C( 60), INT8_C( 123), INT8_C( -21),
586
+ INT8_C( 92), INT8_C( -55), INT8_C( 87), INT8_C( -13),
587
+ INT8_C(-121), INT8_C( -17), INT8_C( 103), INT8_C( 46)) },
588
+ { simde_x_vloadq_s8(INT8_C( 127), INT8_C( 102), INT8_C( -72), INT8_C( -98),
589
+ INT8_C( -83), INT8_C( 33), INT8_C( -54), INT8_C( -48),
590
+ INT8_C( -52), INT8_C( -80), INT8_C(-116), INT8_C( 59),
591
+ INT8_C( 83), INT8_C( 20), INT8_C( 70), INT8_C( 112)),
592
+ simde_x_vloadq_s8(INT8_C( -13), INT8_C( 48), INT8_C( 120), INT8_C( -86),
593
+ INT8_C(-109), INT8_C( 27), INT8_C( -25), INT8_C( -55),
594
+ INT8_C( 76), INT8_C( -35), INT8_C( 41), INT8_C( 44),
595
+ INT8_C( 122), INT8_C( 110), INT8_C( 81), INT8_C( 49)),
596
+ simde_x_vloadq_s8(INT8_C(-116), INT8_C( 54), INT8_C( 64), INT8_C( -12),
597
+ INT8_C( 26), INT8_C( 6), INT8_C( -29), INT8_C( 7),
598
+ INT8_C(-128), INT8_C( -45), INT8_C( 99), INT8_C( 15),
599
+ INT8_C( -39), INT8_C( -90), INT8_C( -11), INT8_C( 63)) },
600
+ { simde_x_vloadq_s8(INT8_C( 59), INT8_C( 9), INT8_C( 55), INT8_C( 73),
601
+ INT8_C( 81), INT8_C( 82), INT8_C( 74), INT8_C( -10),
602
+ INT8_C( 73), INT8_C(-112), INT8_C( -42), INT8_C( -39),
603
+ INT8_C( 61), INT8_C( 23), INT8_C( 31), INT8_C( 113)),
604
+ simde_x_vloadq_s8(INT8_C( -38), INT8_C( 82), INT8_C( -46), INT8_C( -94),
605
+ INT8_C( 104), INT8_C( -15), INT8_C( 25), INT8_C( 76),
606
+ INT8_C( 66), INT8_C( -71), INT8_C( 35), INT8_C( -56),
607
+ INT8_C( 85), INT8_C( -52), INT8_C( 101), INT8_C( 124)),
608
+ simde_x_vloadq_s8(INT8_C( 97), INT8_C( -73), INT8_C( 101), INT8_C( -89),
609
+ INT8_C( -23), INT8_C( 97), INT8_C( 49), INT8_C( -86),
610
+ INT8_C( 7), INT8_C( -41), INT8_C( -77), INT8_C( 17),
611
+ INT8_C( -24), INT8_C( 75), INT8_C( -70), INT8_C( -11)) },
612
+ { simde_x_vloadq_s8(INT8_C( 36), INT8_C( -67), INT8_C( -53), INT8_C( 102),
613
+ INT8_C(-101), INT8_C(-120), INT8_C( -51), INT8_C( 96),
614
+ INT8_C( 51), INT8_C( -92), INT8_C( 32), INT8_C( -14),
615
+ INT8_C( 37), INT8_C( -19), INT8_C( 1), INT8_C( -70)),
616
+ simde_x_vloadq_s8(INT8_C(-120), INT8_C( 2), INT8_C( 118), INT8_C( 16),
617
+ INT8_C(-123), INT8_C(-117), INT8_C(-105), INT8_C( 56),
618
+ INT8_C( 96), INT8_C( -12), INT8_C( 51), INT8_C( 97),
619
+ INT8_C( -87), INT8_C( -16), INT8_C( -41), INT8_C( 60)),
620
+ simde_x_vloadq_s8(INT8_C(-100), INT8_C( -69), INT8_C( 85), INT8_C( 86),
621
+ INT8_C( 22), INT8_C( -3), INT8_C( 54), INT8_C( 40),
622
+ INT8_C( -45), INT8_C( -80), INT8_C( -19), INT8_C(-111),
623
+ INT8_C( 124), INT8_C( -3), INT8_C( 42), INT8_C( 126)) }
624
+ };
625
+
626
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
627
+ simde_int8x16_t r = simde_vsubq_s8(test_vec[i].a, test_vec[i].b);
628
+ simde_neon_assert_int8x16(r, ==, test_vec[i].r);
629
+ }
630
+
631
+ return MUNIT_OK;
632
+ }
633
+
634
+ static MunitResult
635
+ test_simde_vsubq_s16(const MunitParameter params[], void* data) {
636
+ (void) params;
637
+ (void) data;
638
+
639
+ const struct {
640
+ simde_int16x8_t a;
641
+ simde_int16x8_t b;
642
+ simde_int16x8_t r;
643
+ } test_vec[8] = {
644
+ { simde_x_vloadq_s16(INT16_C( -3064), INT16_C( -6478), INT16_C(-19169), INT16_C( -8323),
645
+ INT16_C( 10190), INT16_C(-25337), INT16_C( 5190), INT16_C(-23585)),
646
+ simde_x_vloadq_s16(INT16_C(-31375), INT16_C( -4312), INT16_C(-20814), INT16_C( -8615),
647
+ INT16_C( 20734), INT16_C( 20825), INT16_C( -2208), INT16_C(-12727)),
648
+ simde_x_vloadq_s16(INT16_C( 28311), INT16_C( -2166), INT16_C( 1645), INT16_C( 292),
649
+ INT16_C(-10544), INT16_C( 19374), INT16_C( 7398), INT16_C(-10858)) },
650
+ { simde_x_vloadq_s16(INT16_C( 31361), INT16_C( 19716), INT16_C(-25427), INT16_C(-21999),
651
+ INT16_C(-31976), INT16_C( 25709), INT16_C(-31293), INT16_C( 12981)),
652
+ simde_x_vloadq_s16(INT16_C(-24370), INT16_C(-10981), INT16_C( 24886), INT16_C( 29713),
653
+ INT16_C(-28205), INT16_C( 26287), INT16_C( 32689), INT16_C( -6763)),
654
+ simde_x_vloadq_s16(INT16_C( -9805), INT16_C( 30697), INT16_C( 15223), INT16_C( 13824),
655
+ INT16_C( -3771), INT16_C( -578), INT16_C( 1554), INT16_C( 19744)) },
656
+ { simde_x_vloadq_s16(INT16_C( -6586), INT16_C( 18756), INT16_C(-19800), INT16_C( -6814),
657
+ INT16_C( 1525), INT16_C(-17635), INT16_C(-21122), INT16_C( 29494)),
658
+ simde_x_vloadq_s16(INT16_C(-17885), INT16_C( 13224), INT16_C( 24925), INT16_C(-20719),
659
+ INT16_C( -2570), INT16_C(-11542), INT16_C( 17364), INT16_C(-32054)),
660
+ simde_x_vloadq_s16(INT16_C( 11299), INT16_C( 5532), INT16_C( 20811), INT16_C( 13905),
661
+ INT16_C( 4095), INT16_C( -6093), INT16_C( 27050), INT16_C( -3988)) },
662
+ { simde_x_vloadq_s16(INT16_C( 23412), INT16_C( 2384), INT16_C(-11534), INT16_C(-20699),
663
+ INT16_C( -8361), INT16_C( 26604), INT16_C( -8831), INT16_C( 4802)),
664
+ simde_x_vloadq_s16(INT16_C( 21240), INT16_C(-30954), INT16_C( -9256), INT16_C( 14852),
665
+ INT16_C(-31416), INT16_C(-32724), INT16_C(-10577), INT16_C(-23423)),
666
+ simde_x_vloadq_s16(INT16_C( 2172), INT16_C(-32198), INT16_C( -2278), INT16_C( 29985),
667
+ INT16_C( 23055), INT16_C( -6208), INT16_C( 1746), INT16_C( 28225)) },
668
+ { simde_x_vloadq_s16(INT16_C(-26829), INT16_C(-31551), INT16_C( 17339), INT16_C( 26030),
669
+ INT16_C( 18116), INT16_C( 23813), INT16_C(-29427), INT16_C(-31129)),
670
+ simde_x_vloadq_s16(INT16_C( 11744), INT16_C(-26134), INT16_C(-32172), INT16_C( 22772),
671
+ INT16_C( 13692), INT16_C(-12598), INT16_C( 18660), INT16_C( 8856)),
672
+ simde_x_vloadq_s16(INT16_C( 26963), INT16_C( -5417), INT16_C(-16025), INT16_C( 3258),
673
+ INT16_C( 4424), INT16_C(-29125), INT16_C( 17449), INT16_C( 25551)) },
674
+ { simde_x_vloadq_s16(INT16_C( 27534), INT16_C(-12880), INT16_C( 1587), INT16_C( 13556),
675
+ INT16_C( 17408), INT16_C(-27424), INT16_C( 14128), INT16_C(-18368)),
676
+ simde_x_vloadq_s16(INT16_C( 13780), INT16_C( 25990), INT16_C( 13539), INT16_C( -821),
677
+ INT16_C( -7574), INT16_C(-19266), INT16_C( 17436), INT16_C( 13085)),
678
+ simde_x_vloadq_s16(INT16_C( 13754), INT16_C( 26666), INT16_C(-11952), INT16_C( 14377),
679
+ INT16_C( 24982), INT16_C( -8158), INT16_C( -3308), INT16_C(-31453)) },
680
+ { simde_x_vloadq_s16(INT16_C( 27183), INT16_C( 5881), INT16_C(-18942), INT16_C( 15986),
681
+ INT16_C( 9009), INT16_C( -2829), INT16_C( 2978), INT16_C(-16401)),
682
+ simde_x_vloadq_s16(INT16_C(-17269), INT16_C( 32472), INT16_C(-18651), INT16_C( 6285),
683
+ INT16_C(-28747), INT16_C( 22516), INT16_C( 12510), INT16_C(-12894)),
684
+ simde_x_vloadq_s16(INT16_C(-21084), INT16_C(-26591), INT16_C( -291), INT16_C( 9701),
685
+ INT16_C(-27780), INT16_C(-25345), INT16_C( -9532), INT16_C( -3507)) },
686
+ { simde_x_vloadq_s16(INT16_C( 12730), INT16_C( 147), INT16_C( 29151), INT16_C( 5069),
687
+ INT16_C(-32293), INT16_C( 21273), INT16_C( -9141), INT16_C(-30821)),
688
+ simde_x_vloadq_s16(INT16_C( 19519), INT16_C( 7184), INT16_C( 29373), INT16_C( 17347),
689
+ INT16_C( 31501), INT16_C(-17643), INT16_C( 24181), INT16_C( 68)),
690
+ simde_x_vloadq_s16(INT16_C( -6789), INT16_C( -7037), INT16_C( -222), INT16_C(-12278),
691
+ INT16_C( 1742), INT16_C(-26620), INT16_C( 32214), INT16_C(-30889)) }
692
+ };
693
+
694
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
695
+ simde_int16x8_t r = simde_vsubq_s16(test_vec[i].a, test_vec[i].b);
696
+ simde_neon_assert_int16x8(r, ==, test_vec[i].r);
697
+ }
698
+
699
+ return MUNIT_OK;
700
+ }
701
+
702
+ static MunitResult
703
+ test_simde_vsubq_s32(const MunitParameter params[], void* data) {
704
+ (void) params;
705
+ (void) data;
706
+
707
+ const struct {
708
+ simde_int32x4_t a;
709
+ simde_int32x4_t b;
710
+ simde_int32x4_t r;
711
+ } test_vec[8] = {
712
+ { simde_x_vloadq_s32(INT32_C(-1101685234), INT32_C( 1180920411),
713
+ INT32_C( -564596575), INT32_C( 1686309491)),
714
+ simde_x_vloadq_s32(INT32_C( 845411365), INT32_C( 663983193),
715
+ INT32_C(-1054239476), INT32_C( -913377435)),
716
+ simde_x_vloadq_s32(INT32_C(-1947096599), INT32_C( 516937218),
717
+ INT32_C( 489642901), INT32_C(-1695280370)) },
718
+ { simde_x_vloadq_s32(INT32_C(-1972829951), INT32_C( -648135840),
719
+ INT32_C( -561565135), INT32_C( -140995998)),
720
+ simde_x_vloadq_s32(INT32_C( 947827814), INT32_C( 1431067015),
721
+ INT32_C( 1239445055), INT32_C( 373088742)),
722
+ simde_x_vloadq_s32(INT32_C( 1374309531), INT32_C(-2079202855),
723
+ INT32_C(-1801010190), INT32_C( -514084740)) },
724
+ { simde_x_vloadq_s32(INT32_C(-1708152046), INT32_C( 610550290),
725
+ INT32_C( -275708746), INT32_C( 1930775610)),
726
+ simde_x_vloadq_s32(INT32_C(-1432684320), INT32_C( 1920044618),
727
+ INT32_C( 900010268), INT32_C( -243039165)),
728
+ simde_x_vloadq_s32(INT32_C( -275467726), INT32_C(-1309494328),
729
+ INT32_C(-1175719014), INT32_C(-2121152521)) },
730
+ { simde_x_vloadq_s32(INT32_C(-1493380734), INT32_C( 1539675282),
731
+ INT32_C(-1665826128), INT32_C(-2085077931)),
732
+ simde_x_vloadq_s32(INT32_C(-1899313907), INT32_C(-1453149283),
733
+ INT32_C( 1877171515), INT32_C( 1597838518)),
734
+ simde_x_vloadq_s32(INT32_C( 405933173), INT32_C(-1302142731),
735
+ INT32_C( 751969653), INT32_C( 612050847)) },
736
+ { simde_x_vloadq_s32(INT32_C( -247963587), INT32_C( 674028815),
737
+ INT32_C( 130053867), INT32_C( 1037968629)),
738
+ simde_x_vloadq_s32(INT32_C( -299910818), INT32_C( 327709173),
739
+ INT32_C( 439117134), INT32_C(-1117344472)),
740
+ simde_x_vloadq_s32(INT32_C( 51947231), INT32_C( 346319642),
741
+ INT32_C( -309063267), INT32_C(-2139654195)) },
742
+ { simde_x_vloadq_s32(INT32_C( 1689129365), INT32_C( 1866705563),
743
+ INT32_C( 1871339308), INT32_C( 1664204744)),
744
+ simde_x_vloadq_s32(INT32_C( -298376576), INT32_C( 1102343447),
745
+ INT32_C( 726874084), INT32_C(-1691749647)),
746
+ simde_x_vloadq_s32(INT32_C( 1987505941), INT32_C( 764362116),
747
+ INT32_C( 1144465224), INT32_C( -939012905)) },
748
+ { simde_x_vloadq_s32(INT32_C( -563712427), INT32_C( 969271584),
749
+ INT32_C( -123068166), INT32_C( -364046344)),
750
+ simde_x_vloadq_s32(INT32_C( 1838532984), INT32_C( -154506264),
751
+ INT32_C(-1805936744), INT32_C( 1922367201)),
752
+ simde_x_vloadq_s32(INT32_C( 1892721885), INT32_C( 1123777848),
753
+ INT32_C( 1682868578), INT32_C( 2008553751)) },
754
+ { simde_x_vloadq_s32(INT32_C( 1221273160), INT32_C( 899236763),
755
+ INT32_C(-1589608708), INT32_C( 888351319)),
756
+ simde_x_vloadq_s32(INT32_C( 1872743482), INT32_C( 1264023526),
757
+ INT32_C( 2090293152), INT32_C( 1664402116)),
758
+ simde_x_vloadq_s32(INT32_C( -651470322), INT32_C( -364786763),
759
+ INT32_C( 615065436), INT32_C( -776050797)) }
760
+ };
761
+
762
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
763
+ simde_int32x4_t r = simde_vsubq_s32(test_vec[i].a, test_vec[i].b);
764
+ simde_neon_assert_int32x4(r, ==, test_vec[i].r);
765
+ }
766
+
767
+ return MUNIT_OK;
768
+ }
769
+
770
+ static MunitResult
771
+ test_simde_vsubq_s64(const MunitParameter params[], void* data) {
772
+ (void) params;
773
+ (void) data;
774
+
775
+ const struct {
776
+ simde_int64x2_t a;
777
+ simde_int64x2_t b;
778
+ simde_int64x2_t r;
779
+ } test_vec[8] = {
780
+ { simde_x_vloadq_s64(INT64_C( 274222156456610435), INT64_C(-8012782520979457469)),
781
+ simde_x_vloadq_s64(INT64_C(-1951579396529905934), INT64_C(-8949642768782436772)),
782
+ simde_x_vloadq_s64(INT64_C( 2225801552986516369), INT64_C( 936860247802979303)) },
783
+ { simde_x_vloadq_s64(INT64_C( 6363094968655137069), INT64_C(-6522196357844283659)),
784
+ simde_x_vloadq_s64(INT64_C( 171738256209698764), INT64_C( 4332781025410659193)),
785
+ simde_x_vloadq_s64(INT64_C( 6191356712445438305), INT64_C( 7591766690454608764)) },
786
+ { simde_x_vloadq_s64(INT64_C( 7545908718737944913), INT64_C( 6620037987636810772)),
787
+ simde_x_vloadq_s64(INT64_C(-5998593649828178346), INT64_C( 7689368957132035206)),
788
+ simde_x_vloadq_s64(INT64_C(-4902241705143428357), INT64_C(-1069330969495224434)) },
789
+ { simde_x_vloadq_s64(INT64_C(-6078566214002893154), INT64_C( 4141250803303811646)),
790
+ simde_x_vloadq_s64(INT64_C( 5190927894885744620), INT64_C( 8853395625688132922)),
791
+ simde_x_vloadq_s64(INT64_C( 7177249964820913842), INT64_C(-4712144822384321276)) },
792
+ { simde_x_vloadq_s64(INT64_C( 5509488642277061692), INT64_C( 4838938835417695175)),
793
+ simde_x_vloadq_s64(INT64_C( 6187764650123727738), INT64_C(-5535891583034837807)),
794
+ simde_x_vloadq_s64(INT64_C( -678276007846666046), INT64_C(-8071913655257018634)) },
795
+ { simde_x_vloadq_s64(INT64_C(-5696555285264059925), INT64_C( 3640797763463904411)),
796
+ simde_x_vloadq_s64(INT64_C( 1005953358039190781), INT64_C( 5966532970448653645)),
797
+ simde_x_vloadq_s64(INT64_C(-6702508643303250706), INT64_C(-2325735206984749234)) },
798
+ { simde_x_vloadq_s64(INT64_C(-2907053929439635722), INT64_C( 3182714029933510774)),
799
+ simde_x_vloadq_s64(INT64_C(-3401187050751067841), INT64_C( 2862565250883486067)),
800
+ simde_x_vloadq_s64(INT64_C( 494133121311432119), INT64_C( 320148779050024707)) },
801
+ { simde_x_vloadq_s64(INT64_C( 6765691477153603441), INT64_C( 7918281142011594378)),
802
+ simde_x_vloadq_s64(INT64_C( 3538463942809388704), INT64_C( 6955926674885320041)),
803
+ simde_x_vloadq_s64(INT64_C( 3227227534344214737), INT64_C( 962354467126274337)) }
804
+ };
805
+
806
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
807
+ simde_int64x2_t r = simde_vsubq_s64(test_vec[i].a, test_vec[i].b);
808
+ simde_neon_assert_int64x2(r, ==, test_vec[i].r);
809
+ }
810
+
811
+ return MUNIT_OK;
812
+ }
813
+
814
+ static MunitResult
815
+ test_simde_vsubq_u8(const MunitParameter params[], void* data) {
816
+ (void) params;
817
+ (void) data;
818
+
819
+ const struct {
820
+ simde_uint8x16_t a;
821
+ simde_uint8x16_t b;
822
+ simde_uint8x16_t r;
823
+ } test_vec[8] = {
824
+ { simde_x_vloadq_u8(UINT8_C( 87), UINT8_C(166), UINT8_C(133), UINT8_C(200),
825
+ UINT8_C(211), UINT8_C( 23), UINT8_C( 52), UINT8_C(201),
826
+ UINT8_C(140), UINT8_C( 28), UINT8_C(164), UINT8_C( 87),
827
+ UINT8_C(223), UINT8_C(140), UINT8_C( 81), UINT8_C(145) ),
828
+ simde_x_vloadq_u8(UINT8_C(152), UINT8_C(232), UINT8_C( 56), UINT8_C(163),
829
+ UINT8_C(241), UINT8_C(221), UINT8_C(205), UINT8_C(167),
830
+ UINT8_C(140), UINT8_C(111), UINT8_C(223), UINT8_C( 99),
831
+ UINT8_C(240), UINT8_C(161), UINT8_C( 11), UINT8_C(227) ),
832
+ simde_x_vloadq_u8(UINT8_C(191), UINT8_C(190), UINT8_C( 77), UINT8_C( 37),
833
+ UINT8_C(226), UINT8_C( 58), UINT8_C(103), UINT8_C( 34),
834
+ UINT8_C( 0), UINT8_C(173), UINT8_C(197), UINT8_C(244),
835
+ UINT8_C(239), UINT8_C(235), UINT8_C( 70), UINT8_C(174) ) },
836
+ { simde_x_vloadq_u8(UINT8_C( 78), UINT8_C(248), UINT8_C(127), UINT8_C( 50),
837
+ UINT8_C(109), UINT8_C( 16), UINT8_C( 48), UINT8_C(232),
838
+ UINT8_C(142), UINT8_C( 47), UINT8_C(147), UINT8_C(229),
839
+ UINT8_C( 49), UINT8_C(105), UINT8_C( 55), UINT8_C(126) ),
840
+ simde_x_vloadq_u8(UINT8_C(100), UINT8_C( 45), UINT8_C( 97), UINT8_C(206),
841
+ UINT8_C(115), UINT8_C( 34), UINT8_C(225), UINT8_C(163),
842
+ UINT8_C(112), UINT8_C(233), UINT8_C( 18), UINT8_C(183),
843
+ UINT8_C( 19), UINT8_C( 93), UINT8_C(192), UINT8_C(250) ),
844
+ simde_x_vloadq_u8(UINT8_C(234), UINT8_C(203), UINT8_C( 30), UINT8_C(100),
845
+ UINT8_C(250), UINT8_C(238), UINT8_C( 79), UINT8_C( 69),
846
+ UINT8_C( 30), UINT8_C( 70), UINT8_C(129), UINT8_C( 46),
847
+ UINT8_C( 30), UINT8_C( 12), UINT8_C(119), UINT8_C(132) ) },
848
+ { simde_x_vloadq_u8(UINT8_C(142), UINT8_C(126), UINT8_C(125), UINT8_C(192),
849
+ UINT8_C( 49), UINT8_C(244), UINT8_C( 25), UINT8_C(167),
850
+ UINT8_C( 93), UINT8_C(226), UINT8_C(205), UINT8_C(169),
851
+ UINT8_C( 52), UINT8_C( 68), UINT8_C(146), UINT8_C( 69) ),
852
+ simde_x_vloadq_u8(UINT8_C(150), UINT8_C(111), UINT8_C( 21), UINT8_C(234),
853
+ UINT8_C(182), UINT8_C(188), UINT8_C(152), UINT8_C( 59),
854
+ UINT8_C(165), UINT8_C(143), UINT8_C( 0), UINT8_C(157),
855
+ UINT8_C(255), UINT8_C(215), UINT8_C( 72), UINT8_C(154) ),
856
+ simde_x_vloadq_u8(UINT8_C(248), UINT8_C( 15), UINT8_C(104), UINT8_C(214),
857
+ UINT8_C(123), UINT8_C( 56), UINT8_C(129), UINT8_C(108),
858
+ UINT8_C(184), UINT8_C( 83), UINT8_C(205), UINT8_C( 12),
859
+ UINT8_C( 53), UINT8_C(109), UINT8_C( 74), UINT8_C(171) ) },
860
+ { simde_x_vloadq_u8(UINT8_C(233), UINT8_C(157), UINT8_C(171), UINT8_C( 93),
861
+ UINT8_C( 59), UINT8_C(112), UINT8_C(155), UINT8_C(145),
862
+ UINT8_C( 14), UINT8_C(206), UINT8_C(123), UINT8_C( 33),
863
+ UINT8_C(134), UINT8_C(124), UINT8_C(173), UINT8_C( 21) ),
864
+ simde_x_vloadq_u8(UINT8_C(178), UINT8_C(228), UINT8_C(119), UINT8_C(103),
865
+ UINT8_C( 9), UINT8_C( 45), UINT8_C( 47), UINT8_C(186),
866
+ UINT8_C( 59), UINT8_C(218), UINT8_C( 90), UINT8_C(122),
867
+ UINT8_C(239), UINT8_C(194), UINT8_C( 89), UINT8_C(141) ),
868
+ simde_x_vloadq_u8(UINT8_C( 55), UINT8_C(185), UINT8_C( 52), UINT8_C(246),
869
+ UINT8_C( 50), UINT8_C( 67), UINT8_C(108), UINT8_C(215),
870
+ UINT8_C(211), UINT8_C(244), UINT8_C( 33), UINT8_C(167),
871
+ UINT8_C(151), UINT8_C(186), UINT8_C( 84), UINT8_C(136) ) },
872
+ { simde_x_vloadq_u8(UINT8_C( 7), UINT8_C(184), UINT8_C( 89), UINT8_C(229),
873
+ UINT8_C( 26), UINT8_C(225), UINT8_C(171), UINT8_C( 84),
874
+ UINT8_C(163), UINT8_C(191), UINT8_C( 98), UINT8_C(129),
875
+ UINT8_C(143), UINT8_C(170), UINT8_C(148), UINT8_C( 70) ),
876
+ simde_x_vloadq_u8(UINT8_C( 5), UINT8_C(254), UINT8_C(214), UINT8_C(207),
877
+ UINT8_C(152), UINT8_C(188), UINT8_C(231), UINT8_C(116),
878
+ UINT8_C(199), UINT8_C(147), UINT8_C( 90), UINT8_C(229),
879
+ UINT8_C(191), UINT8_C( 52), UINT8_C(224), UINT8_C(113) ),
880
+ simde_x_vloadq_u8(UINT8_C( 2), UINT8_C(186), UINT8_C(131), UINT8_C( 22),
881
+ UINT8_C(130), UINT8_C( 37), UINT8_C(196), UINT8_C(224),
882
+ UINT8_C(220), UINT8_C( 44), UINT8_C( 8), UINT8_C(156),
883
+ UINT8_C(208), UINT8_C(118), UINT8_C(180), UINT8_C(213) ) },
884
+ { simde_x_vloadq_u8(UINT8_C( 64), UINT8_C( 6), UINT8_C( 42), UINT8_C(193),
885
+ UINT8_C(142), UINT8_C( 7), UINT8_C(114), UINT8_C( 83),
886
+ UINT8_C( 65), UINT8_C(177), UINT8_C(114), UINT8_C(206),
887
+ UINT8_C(218), UINT8_C(169), UINT8_C( 31), UINT8_C( 17) ),
888
+ simde_x_vloadq_u8(UINT8_C( 72), UINT8_C(215), UINT8_C( 74), UINT8_C( 77),
889
+ UINT8_C( 51), UINT8_C(124), UINT8_C(148), UINT8_C(247),
890
+ UINT8_C(227), UINT8_C(145), UINT8_C( 46), UINT8_C( 93),
891
+ UINT8_C( 71), UINT8_C(189), UINT8_C( 75), UINT8_C( 95) ),
892
+ simde_x_vloadq_u8(UINT8_C(248), UINT8_C( 47), UINT8_C(224), UINT8_C(116),
893
+ UINT8_C( 91), UINT8_C(139), UINT8_C(222), UINT8_C( 92),
894
+ UINT8_C( 94), UINT8_C( 32), UINT8_C( 68), UINT8_C(113),
895
+ UINT8_C(147), UINT8_C(236), UINT8_C(212), UINT8_C(178) ) },
896
+ { simde_x_vloadq_u8(UINT8_C(229), UINT8_C( 36), UINT8_C( 55), UINT8_C( 75),
897
+ UINT8_C( 32), UINT8_C( 62), UINT8_C(175), UINT8_C(217),
898
+ UINT8_C( 6), UINT8_C( 76), UINT8_C( 16), UINT8_C(228),
899
+ UINT8_C( 98), UINT8_C( 43), UINT8_C(172), UINT8_C( 10) ),
900
+ simde_x_vloadq_u8(UINT8_C(221), UINT8_C( 62), UINT8_C( 60), UINT8_C(117),
901
+ UINT8_C(127), UINT8_C(225), UINT8_C(161), UINT8_C(111),
902
+ UINT8_C( 71), UINT8_C( 10), UINT8_C( 89), UINT8_C(137),
903
+ UINT8_C( 24), UINT8_C( 13), UINT8_C(189), UINT8_C(193) ),
904
+ simde_x_vloadq_u8(UINT8_C( 8), UINT8_C(230), UINT8_C(251), UINT8_C(214),
905
+ UINT8_C(161), UINT8_C( 93), UINT8_C( 14), UINT8_C(106),
906
+ UINT8_C(191), UINT8_C( 66), UINT8_C(183), UINT8_C( 91),
907
+ UINT8_C( 74), UINT8_C( 30), UINT8_C(239), UINT8_C( 73) ) },
908
+ { simde_x_vloadq_u8(UINT8_C(216), UINT8_C( 46), UINT8_C(120), UINT8_C(228),
909
+ UINT8_C(105), UINT8_C(220), UINT8_C( 54), UINT8_C(187),
910
+ UINT8_C(106), UINT8_C( 67), UINT8_C(134), UINT8_C(119),
911
+ UINT8_C(202), UINT8_C( 11), UINT8_C(190), UINT8_C( 21) ),
912
+ simde_x_vloadq_u8(UINT8_C(111), UINT8_C(153), UINT8_C( 58), UINT8_C( 46),
913
+ UINT8_C(233), UINT8_C(135), UINT8_C( 56), UINT8_C( 71),
914
+ UINT8_C(213), UINT8_C(201), UINT8_C( 10), UINT8_C(126),
915
+ UINT8_C(252), UINT8_C( 75), UINT8_C( 50), UINT8_C( 2) ),
916
+ simde_x_vloadq_u8(UINT8_C(105), UINT8_C(149), UINT8_C( 62), UINT8_C(182),
917
+ UINT8_C(128), UINT8_C( 85), UINT8_C(254), UINT8_C(116),
918
+ UINT8_C(149), UINT8_C(122), UINT8_C(124), UINT8_C(249),
919
+ UINT8_C(206), UINT8_C(192), UINT8_C(140), UINT8_C( 19) ) }
920
+ };
921
+
922
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
923
+ simde_uint8x16_t r = simde_vsubq_u8(test_vec[i].a, test_vec[i].b);
924
+ simde_neon_assert_uint8x16(r, ==, test_vec[i].r);
925
+ }
926
+
927
+ return MUNIT_OK;
928
+ }
929
+
930
+ static MunitResult
931
+ test_simde_vsubq_u16(const MunitParameter params[], void* data) {
932
+ (void) params;
933
+ (void) data;
934
+
935
+ const struct {
936
+ simde_uint16x8_t a;
937
+ simde_uint16x8_t b;
938
+ simde_uint16x8_t r;
939
+ } test_vec[8] = {
940
+ { simde_x_vloadq_u16(UINT16_C(32058), UINT16_C(26873), UINT16_C(51051), UINT16_C(31536),
941
+ UINT16_C(10411), UINT16_C(39995), UINT16_C(21610), UINT16_C( 9384)),
942
+ simde_x_vloadq_u16(UINT16_C(59598), UINT16_C(35382), UINT16_C(33176), UINT16_C(25970),
943
+ UINT16_C(41560), UINT16_C(47620), UINT16_C( 1773), UINT16_C(42681)),
944
+ simde_x_vloadq_u16(UINT16_C(37996), UINT16_C(57027), UINT16_C(17875), UINT16_C( 5566),
945
+ UINT16_C(34387), UINT16_C(57911), UINT16_C(19837), UINT16_C(32239)) },
946
+ { simde_x_vloadq_u16(UINT16_C(59399), UINT16_C(53992), UINT16_C(46317), UINT16_C(28372),
947
+ UINT16_C(25727), UINT16_C(30337), UINT16_C(55706), UINT16_C(63759)),
948
+ simde_x_vloadq_u16(UINT16_C( 5797), UINT16_C(19305), UINT16_C(37560), UINT16_C(46549),
949
+ UINT16_C(33738), UINT16_C(15792), UINT16_C(10742), UINT16_C(32928)),
950
+ simde_x_vloadq_u16(UINT16_C(53602), UINT16_C(34687), UINT16_C( 8757), UINT16_C(47359),
951
+ UINT16_C(57525), UINT16_C(14545), UINT16_C(44964), UINT16_C(30831)) },
952
+ { simde_x_vloadq_u16(UINT16_C(53681), UINT16_C(22004), UINT16_C(53966), UINT16_C(35356),
953
+ UINT16_C(62005), UINT16_C(63604), UINT16_C(59737), UINT16_C(55622)),
954
+ simde_x_vloadq_u16(UINT16_C(34367), UINT16_C(61611), UINT16_C(14128), UINT16_C(16396),
955
+ UINT16_C(52618), UINT16_C(49102), UINT16_C( 2031), UINT16_C(63431)),
956
+ simde_x_vloadq_u16(UINT16_C(19314), UINT16_C(25929), UINT16_C(39838), UINT16_C(18960),
957
+ UINT16_C( 9387), UINT16_C(14502), UINT16_C(57706), UINT16_C(57727)) },
958
+ { simde_x_vloadq_u16(UINT16_C(39646), UINT16_C(45587), UINT16_C(23807), UINT16_C(48676),
959
+ UINT16_C( 8603), UINT16_C( 9561), UINT16_C(50591), UINT16_C(45984)),
960
+ simde_x_vloadq_u16(UINT16_C(58421), UINT16_C(30266), UINT16_C(17160), UINT16_C(14127),
961
+ UINT16_C(59428), UINT16_C(25189), UINT16_C(40047), UINT16_C(49620)),
962
+ simde_x_vloadq_u16(UINT16_C(46761), UINT16_C(15321), UINT16_C( 6647), UINT16_C(34549),
963
+ UINT16_C(14711), UINT16_C(49908), UINT16_C(10544), UINT16_C(61900)) },
964
+ { simde_x_vloadq_u16(UINT16_C( 6007), UINT16_C(32558), UINT16_C(60603), UINT16_C( 8094),
965
+ UINT16_C(46041), UINT16_C(14496), UINT16_C( 9960), UINT16_C(64452)),
966
+ simde_x_vloadq_u16(UINT16_C(55919), UINT16_C(48705), UINT16_C( 278), UINT16_C(40315),
967
+ UINT16_C(43583), UINT16_C(32754), UINT16_C(54969), UINT16_C(32537)),
968
+ simde_x_vloadq_u16(UINT16_C(15624), UINT16_C(49389), UINT16_C(60325), UINT16_C(33315),
969
+ UINT16_C( 2458), UINT16_C(47278), UINT16_C(20527), UINT16_C(31915)) },
970
+ { simde_x_vloadq_u16(UINT16_C(29699), UINT16_C( 6878), UINT16_C(61985), UINT16_C(28537),
971
+ UINT16_C(11093), UINT16_C(59294), UINT16_C(49467), UINT16_C(14552)),
972
+ simde_x_vloadq_u16(UINT16_C(28129), UINT16_C(16604), UINT16_C( 2705), UINT16_C(46832),
973
+ UINT16_C(52122), UINT16_C(33620), UINT16_C(45368), UINT16_C(46933)),
974
+ simde_x_vloadq_u16(UINT16_C( 1570), UINT16_C(55810), UINT16_C(59280), UINT16_C(47241),
975
+ UINT16_C(24507), UINT16_C(25674), UINT16_C( 4099), UINT16_C(33155)) },
976
+ { simde_x_vloadq_u16(UINT16_C(10975), UINT16_C( 9937), UINT16_C(22680), UINT16_C( 9723),
977
+ UINT16_C(21167), UINT16_C(13605), UINT16_C(22850), UINT16_C(50938)),
978
+ simde_x_vloadq_u16(UINT16_C( 791), UINT16_C(29997), UINT16_C( 6816), UINT16_C(19884),
979
+ UINT16_C(27158), UINT16_C( 195), UINT16_C(51377), UINT16_C(51054)),
980
+ simde_x_vloadq_u16(UINT16_C(10184), UINT16_C(45476), UINT16_C(15864), UINT16_C(55375),
981
+ UINT16_C(59545), UINT16_C(13410), UINT16_C(37009), UINT16_C(65420)) },
982
+ { simde_x_vloadq_u16(UINT16_C(35815), UINT16_C(48475), UINT16_C(62387), UINT16_C(39859),
983
+ UINT16_C( 8385), UINT16_C( 1980), UINT16_C(63086), UINT16_C(23583)),
984
+ simde_x_vloadq_u16(UINT16_C(32032), UINT16_C(34559), UINT16_C(38712), UINT16_C(51264),
985
+ UINT16_C( 1517), UINT16_C( 1962), UINT16_C( 5545), UINT16_C(52626)),
986
+ simde_x_vloadq_u16(UINT16_C( 3783), UINT16_C(13916), UINT16_C(23675), UINT16_C(54131),
987
+ UINT16_C( 6868), UINT16_C( 18), UINT16_C(57541), UINT16_C(36493)) }
988
+ };
989
+
990
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
991
+ simde_uint16x8_t r = simde_vsubq_u16(test_vec[i].a, test_vec[i].b);
992
+ simde_neon_assert_uint16x8(r, ==, test_vec[i].r);
993
+ }
994
+
995
+ return MUNIT_OK;
996
+ }
997
+
998
+ static MunitResult
999
+ test_simde_vsubq_u32(const MunitParameter params[], void* data) {
1000
+ (void) params;
1001
+ (void) data;
1002
+
1003
+ const struct {
1004
+ simde_uint32x4_t a;
1005
+ simde_uint32x4_t b;
1006
+ simde_uint32x4_t r;
1007
+ } test_vec[8] = {
1008
+ { simde_x_vloadq_u32(UINT32_C(3895408461), UINT32_C(3640128288),
1009
+ UINT32_C(1854574044), UINT32_C(1659558118)),
1010
+ simde_x_vloadq_u32(UINT32_C(1884519680), UINT32_C( 213141937),
1011
+ UINT32_C(2264515241), UINT32_C( 599802617)),
1012
+ simde_x_vloadq_u32(UINT32_C(2010888781), UINT32_C(3426986351),
1013
+ UINT32_C(3885026099), UINT32_C(1059755501)) },
1014
+ { simde_x_vloadq_u32(UINT32_C(3891897214), UINT32_C(1797204056),
1015
+ UINT32_C(3471191210), UINT32_C(3208034210)),
1016
+ simde_x_vloadq_u32(UINT32_C(4270742654), UINT32_C(1803662066),
1017
+ UINT32_C( 905537885), UINT32_C(3774829890)),
1018
+ simde_x_vloadq_u32(UINT32_C(3916121856), UINT32_C(4288509286),
1019
+ UINT32_C(2565653325), UINT32_C(3728171616)) },
1020
+ { simde_x_vloadq_u32(UINT32_C(1716681133), UINT32_C(1373763584),
1021
+ UINT32_C(2129047472), UINT32_C( 660748118)),
1022
+ simde_x_vloadq_u32(UINT32_C(1322003017), UINT32_C(2735955115),
1023
+ UINT32_C(1208054480), UINT32_C(3915748514)),
1024
+ simde_x_vloadq_u32(UINT32_C( 394678116), UINT32_C(2932775765),
1025
+ UINT32_C( 920992992), UINT32_C(1039966900)) },
1026
+ { simde_x_vloadq_u32(UINT32_C(1227832463), UINT32_C( 881560198),
1027
+ UINT32_C(3487996398), UINT32_C(3949354035)),
1028
+ simde_x_vloadq_u32(UINT32_C(2421331614), UINT32_C(2819230368),
1029
+ UINT32_C(2423324594), UINT32_C( 998713619)),
1030
+ simde_x_vloadq_u32(UINT32_C(3101468145), UINT32_C(2357297126),
1031
+ UINT32_C(1064671804), UINT32_C(2950640416)) },
1032
+ { simde_x_vloadq_u32(UINT32_C(3212444536), UINT32_C(3255121650),
1033
+ UINT32_C( 280297817), UINT32_C(2780019175)),
1034
+ simde_x_vloadq_u32(UINT32_C(1634711183), UINT32_C( 74404227),
1035
+ UINT32_C(4155314604), UINT32_C(2276824421)),
1036
+ simde_x_vloadq_u32(UINT32_C(1577733353), UINT32_C(3180717423),
1037
+ UINT32_C( 419950509), UINT32_C( 503194754)) },
1038
+ { simde_x_vloadq_u32(UINT32_C(2531614680), UINT32_C(3065185794),
1039
+ UINT32_C(1559662141), UINT32_C(4245192357)),
1040
+ simde_x_vloadq_u32(UINT32_C( 629083668), UINT32_C( 165904817),
1041
+ UINT32_C(1258087237), UINT32_C( 869898369)),
1042
+ simde_x_vloadq_u32(UINT32_C(1902531012), UINT32_C(2899280977),
1043
+ UINT32_C( 301574904), UINT32_C(3375293988)) },
1044
+ { simde_x_vloadq_u32(UINT32_C(3370588623), UINT32_C( 378752671),
1045
+ UINT32_C( 654547258), UINT32_C(3831311079)),
1046
+ simde_x_vloadq_u32(UINT32_C(1157532660), UINT32_C(1026394530),
1047
+ UINT32_C(1443726588), UINT32_C(1444660531)),
1048
+ simde_x_vloadq_u32(UINT32_C(2213055963), UINT32_C(3647325437),
1049
+ UINT32_C(3505787966), UINT32_C(2386650548)) },
1050
+ { simde_x_vloadq_u32(UINT32_C( 388092756), UINT32_C(2451063360),
1051
+ UINT32_C(2692706503), UINT32_C(3353389405)),
1052
+ simde_x_vloadq_u32(UINT32_C( 598271220), UINT32_C(3357759735),
1053
+ UINT32_C(2490786438), UINT32_C(2782412551)),
1054
+ simde_x_vloadq_u32(UINT32_C(4084788832), UINT32_C(3388270921),
1055
+ UINT32_C( 201920065), UINT32_C( 570976854)) }
1056
+ };
1057
+
1058
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
1059
+ simde_uint32x4_t r = simde_vsubq_u32(test_vec[i].a, test_vec[i].b);
1060
+ simde_neon_assert_uint32x4(r, ==, test_vec[i].r);
1061
+ }
1062
+
1063
+ return MUNIT_OK;
1064
+ }
1065
+
1066
+ static MunitResult
1067
+ test_simde_vsubq_u64(const MunitParameter params[], void* data) {
1068
+ (void) params;
1069
+ (void) data;
1070
+
1071
+ const struct {
1072
+ simde_uint64x2_t a;
1073
+ simde_uint64x2_t b;
1074
+ simde_uint64x2_t r;
1075
+ } test_vec[8] = {
1076
+ { simde_x_vloadq_u64(UINT64_C(18062352261781936342), UINT64_C(7388857429341683123)),
1077
+ simde_x_vloadq_u64(UINT64_C(1553436335104468014), UINT64_C(16839974875395556662)),
1078
+ simde_x_vloadq_u64(UINT64_C(16508915926677468328), UINT64_C(8995626627655678077)) },
1079
+ { simde_x_vloadq_u64(UINT64_C(9781783238890835896), UINT64_C(17293201827280780780)),
1080
+ simde_x_vloadq_u64(UINT64_C(7134357670424035822), UINT64_C(4404198093893325283)),
1081
+ simde_x_vloadq_u64(UINT64_C(2647425568466800074), UINT64_C(12889003733387455497)) },
1082
+ { simde_x_vloadq_u64(UINT64_C(4488961572228535140), UINT64_C(1346563325554989949)),
1083
+ simde_x_vloadq_u64(UINT64_C(2634888985021256352), UINT64_C( 666451602867372356)),
1084
+ simde_x_vloadq_u64(UINT64_C(1854072587207278788), UINT64_C( 680111722687617593)) },
1085
+ { simde_x_vloadq_u64(UINT64_C(1318235667828533745), UINT64_C(14815956286012371914)),
1086
+ simde_x_vloadq_u64(UINT64_C(11432878494825478518), UINT64_C(17685580310423747579)),
1087
+ simde_x_vloadq_u64(UINT64_C(8332101246712606843), UINT64_C(15577120049298175951)) },
1088
+ { simde_x_vloadq_u64(UINT64_C(17790576723876784299), UINT64_C(11663761306291148559)),
1089
+ simde_x_vloadq_u64(UINT64_C(15385151009115536057), UINT64_C(13794328263855967094)),
1090
+ simde_x_vloadq_u64(UINT64_C(2405425714761248242), UINT64_C(16316177116144733081)) },
1091
+ { simde_x_vloadq_u64(UINT64_C( 350914325027652600), UINT64_C(14797590881591740981)),
1092
+ simde_x_vloadq_u64(UINT64_C(12036897114441261299), UINT64_C(8202964999985589665)),
1093
+ simde_x_vloadq_u64(UINT64_C(6760761284295942917), UINT64_C(6594625881606151316)) },
1094
+ { simde_x_vloadq_u64(UINT64_C(8518920335610308466), UINT64_C(16046307191155630549)),
1095
+ simde_x_vloadq_u64(UINT64_C(6106112106205787470), UINT64_C(16756353123063307916)),
1096
+ simde_x_vloadq_u64(UINT64_C(2412808229404520996), UINT64_C(17736698141801874249)) },
1097
+ { simde_x_vloadq_u64(UINT64_C(5938134708777591391), UINT64_C(5692436156143036134)),
1098
+ simde_x_vloadq_u64(UINT64_C(1370976571276529480), UINT64_C(2772233378346164097)),
1099
+ simde_x_vloadq_u64(UINT64_C(4567158137501061911), UINT64_C(2920202777796872037)) }
1100
+ };
1101
+
1102
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
1103
+ simde_uint64x2_t r = simde_vsubq_u64(test_vec[i].a, test_vec[i].b);
1104
+ simde_neon_assert_uint64x2(r, ==, test_vec[i].r);
1105
+ }
1106
+
1107
+ return MUNIT_OK;
1108
+ }
1109
+
1110
+ static MunitResult
1111
+ test_simde_vsubq_f32(const MunitParameter params[], void* data) {
1112
+ (void) params;
1113
+ (void) data;
1114
+
1115
+ const struct {
1116
+ simde_float32x4_t a;
1117
+ simde_float32x4_t b;
1118
+ simde_float32x4_t r;
1119
+ } test_vec[8] = {
1120
+ { simde_x_vloadq_f32(SIMDE_FLOAT32_C( 7.69), SIMDE_FLOAT32_C( 907.62),
1121
+ SIMDE_FLOAT32_C( 434.05), SIMDE_FLOAT32_C( 467.28)),
1122
+ simde_x_vloadq_f32(SIMDE_FLOAT32_C( -933.64), SIMDE_FLOAT32_C( -442.18),
1123
+ SIMDE_FLOAT32_C( 903.29), SIMDE_FLOAT32_C( 399.69)),
1124
+ simde_x_vloadq_f32(SIMDE_FLOAT32_C( 941.33), SIMDE_FLOAT32_C( 1349.80),
1125
+ SIMDE_FLOAT32_C( -469.24), SIMDE_FLOAT32_C( 67.59)) },
1126
+ { simde_x_vloadq_f32(SIMDE_FLOAT32_C( 730.21), SIMDE_FLOAT32_C( 541.58),
1127
+ SIMDE_FLOAT32_C( -112.19), SIMDE_FLOAT32_C( -183.13)),
1128
+ simde_x_vloadq_f32(SIMDE_FLOAT32_C( -894.87), SIMDE_FLOAT32_C( -998.92),
1129
+ SIMDE_FLOAT32_C( -118.73), SIMDE_FLOAT32_C( 578.99)),
1130
+ simde_x_vloadq_f32(SIMDE_FLOAT32_C( 1625.08), SIMDE_FLOAT32_C( 1540.50),
1131
+ SIMDE_FLOAT32_C( 6.54), SIMDE_FLOAT32_C( -762.12)) },
1132
+ { simde_x_vloadq_f32(SIMDE_FLOAT32_C( 39.67), SIMDE_FLOAT32_C( 381.67),
1133
+ SIMDE_FLOAT32_C( 729.76), SIMDE_FLOAT32_C( -88.71)),
1134
+ simde_x_vloadq_f32(SIMDE_FLOAT32_C( 432.56), SIMDE_FLOAT32_C( 323.88),
1135
+ SIMDE_FLOAT32_C( -791.56), SIMDE_FLOAT32_C( 976.13)),
1136
+ simde_x_vloadq_f32(SIMDE_FLOAT32_C( -392.89), SIMDE_FLOAT32_C( 57.79),
1137
+ SIMDE_FLOAT32_C( 1521.32), SIMDE_FLOAT32_C(-1064.84)) },
1138
+ { simde_x_vloadq_f32(SIMDE_FLOAT32_C( -983.53), SIMDE_FLOAT32_C( 399.10),
1139
+ SIMDE_FLOAT32_C( 580.81), SIMDE_FLOAT32_C( 664.21)),
1140
+ simde_x_vloadq_f32(SIMDE_FLOAT32_C( -569.36), SIMDE_FLOAT32_C( 667.71),
1141
+ SIMDE_FLOAT32_C( 933.23), SIMDE_FLOAT32_C( -924.97)),
1142
+ simde_x_vloadq_f32(SIMDE_FLOAT32_C( -414.17), SIMDE_FLOAT32_C( -268.61),
1143
+ SIMDE_FLOAT32_C( -352.42), SIMDE_FLOAT32_C( 1589.18)) },
1144
+ { simde_x_vloadq_f32(SIMDE_FLOAT32_C( 921.09), SIMDE_FLOAT32_C( -843.40),
1145
+ SIMDE_FLOAT32_C( 221.06), SIMDE_FLOAT32_C( -617.96)),
1146
+ simde_x_vloadq_f32(SIMDE_FLOAT32_C( -739.66), SIMDE_FLOAT32_C( 351.29),
1147
+ SIMDE_FLOAT32_C( 257.33), SIMDE_FLOAT32_C( -723.30)),
1148
+ simde_x_vloadq_f32(SIMDE_FLOAT32_C( 1660.75), SIMDE_FLOAT32_C(-1194.69),
1149
+ SIMDE_FLOAT32_C( -36.27), SIMDE_FLOAT32_C( 105.34)) },
1150
+ { simde_x_vloadq_f32(SIMDE_FLOAT32_C( 598.06), SIMDE_FLOAT32_C( -849.67),
1151
+ SIMDE_FLOAT32_C( -665.04), SIMDE_FLOAT32_C( 262.46)),
1152
+ simde_x_vloadq_f32(SIMDE_FLOAT32_C( 403.08), SIMDE_FLOAT32_C( -450.69),
1153
+ SIMDE_FLOAT32_C( 315.14), SIMDE_FLOAT32_C( -172.43)),
1154
+ simde_x_vloadq_f32(SIMDE_FLOAT32_C( 194.98), SIMDE_FLOAT32_C( -398.98),
1155
+ SIMDE_FLOAT32_C( -980.18), SIMDE_FLOAT32_C( 434.89)) },
1156
+ { simde_x_vloadq_f32(SIMDE_FLOAT32_C( -734.13), SIMDE_FLOAT32_C( 313.84),
1157
+ SIMDE_FLOAT32_C( 126.41), SIMDE_FLOAT32_C( 460.30)),
1158
+ simde_x_vloadq_f32(SIMDE_FLOAT32_C( 981.68), SIMDE_FLOAT32_C( -396.47),
1159
+ SIMDE_FLOAT32_C( 734.12), SIMDE_FLOAT32_C( -993.96)),
1160
+ simde_x_vloadq_f32(SIMDE_FLOAT32_C(-1715.81), SIMDE_FLOAT32_C( 710.31),
1161
+ SIMDE_FLOAT32_C( -607.71), SIMDE_FLOAT32_C( 1454.26)) },
1162
+ { simde_x_vloadq_f32(SIMDE_FLOAT32_C( -43.70), SIMDE_FLOAT32_C( 865.01),
1163
+ SIMDE_FLOAT32_C( 778.14), SIMDE_FLOAT32_C( 818.60)),
1164
+ simde_x_vloadq_f32(SIMDE_FLOAT32_C( 597.32), SIMDE_FLOAT32_C( -982.41),
1165
+ SIMDE_FLOAT32_C( 529.34), SIMDE_FLOAT32_C( 891.14)),
1166
+ simde_x_vloadq_f32(SIMDE_FLOAT32_C( -641.02), SIMDE_FLOAT32_C( 1847.42),
1167
+ SIMDE_FLOAT32_C( 248.80), SIMDE_FLOAT32_C( -72.54)) }
1168
+ };
1169
+
1170
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
1171
+ simde_float32x4_t r = simde_vsubq_f32(test_vec[i].a, test_vec[i].b);
1172
+ simde_neon_assert_float32x4_equal(r, test_vec[i].r, 1);
1173
+ }
1174
+
1175
+ return MUNIT_OK;
1176
+ }
1177
+
1178
+ static MunitResult
1179
+ test_simde_vsubq_f64(const MunitParameter params[], void* data) {
1180
+ (void) params;
1181
+ (void) data;
1182
+
1183
+ const struct {
1184
+ simde_float64x2_t a;
1185
+ simde_float64x2_t b;
1186
+ simde_float64x2_t r;
1187
+ } test_vec[8] = {
1188
+ { simde_x_vloadq_f64(SIMDE_FLOAT64_C( 825.97), SIMDE_FLOAT64_C( 668.02)),
1189
+ simde_x_vloadq_f64(SIMDE_FLOAT64_C( -306.36), SIMDE_FLOAT64_C( -166.09)),
1190
+ simde_x_vloadq_f64(SIMDE_FLOAT64_C( 1132.33), SIMDE_FLOAT64_C( 834.11)) },
1191
+ { simde_x_vloadq_f64(SIMDE_FLOAT64_C( -347.97), SIMDE_FLOAT64_C( 878.12)),
1192
+ simde_x_vloadq_f64(SIMDE_FLOAT64_C( -577.04), SIMDE_FLOAT64_C( 589.16)),
1193
+ simde_x_vloadq_f64(SIMDE_FLOAT64_C( 229.07), SIMDE_FLOAT64_C( 288.96)) },
1194
+ { simde_x_vloadq_f64(SIMDE_FLOAT64_C( 977.90), SIMDE_FLOAT64_C( 617.14)),
1195
+ simde_x_vloadq_f64(SIMDE_FLOAT64_C( 203.07), SIMDE_FLOAT64_C( 438.54)),
1196
+ simde_x_vloadq_f64(SIMDE_FLOAT64_C( 774.83), SIMDE_FLOAT64_C( 178.60)) },
1197
+ { simde_x_vloadq_f64(SIMDE_FLOAT64_C( -37.98), SIMDE_FLOAT64_C( -671.53)),
1198
+ simde_x_vloadq_f64(SIMDE_FLOAT64_C( -96.79), SIMDE_FLOAT64_C( 106.17)),
1199
+ simde_x_vloadq_f64(SIMDE_FLOAT64_C( 58.81), SIMDE_FLOAT64_C( -777.70)) },
1200
+ { simde_x_vloadq_f64(SIMDE_FLOAT64_C( -617.94), SIMDE_FLOAT64_C( 127.13)),
1201
+ simde_x_vloadq_f64(SIMDE_FLOAT64_C( 546.61), SIMDE_FLOAT64_C( 958.62)),
1202
+ simde_x_vloadq_f64(SIMDE_FLOAT64_C(-1164.55), SIMDE_FLOAT64_C( -831.49)) },
1203
+ { simde_x_vloadq_f64(SIMDE_FLOAT64_C( -851.38), SIMDE_FLOAT64_C( -117.61)),
1204
+ simde_x_vloadq_f64(SIMDE_FLOAT64_C( -556.18), SIMDE_FLOAT64_C( 808.85)),
1205
+ simde_x_vloadq_f64(SIMDE_FLOAT64_C( -295.20), SIMDE_FLOAT64_C( -926.46)) },
1206
+ { simde_x_vloadq_f64(SIMDE_FLOAT64_C( 810.93), SIMDE_FLOAT64_C( 200.78)),
1207
+ simde_x_vloadq_f64(SIMDE_FLOAT64_C( 375.25), SIMDE_FLOAT64_C( -38.10)),
1208
+ simde_x_vloadq_f64(SIMDE_FLOAT64_C( 435.68), SIMDE_FLOAT64_C( 238.88)) },
1209
+ { simde_x_vloadq_f64(SIMDE_FLOAT64_C( -395.68), SIMDE_FLOAT64_C( 663.06)),
1210
+ simde_x_vloadq_f64(SIMDE_FLOAT64_C( -565.25), SIMDE_FLOAT64_C( 522.03)),
1211
+ simde_x_vloadq_f64(SIMDE_FLOAT64_C( 169.57), SIMDE_FLOAT64_C( 141.03)) }
1212
+ };
1213
+
1214
+ for (size_t i = 0 ; i < (sizeof(test_vec) / sizeof(test_vec[0])); i++) {
1215
+ simde_float64x2_t r = simde_vsubq_f64(test_vec[i].a, test_vec[i].b);
1216
+ simde_neon_assert_float64x2_equal(r, test_vec[i].r, 1);
1217
+ }
1218
+
1219
+ return MUNIT_OK;
1220
+ }
1221
+
1222
+ #endif /* defined(SIMDE_NEON_NATIVE) || defined(SIMDE_NO_NATIVE) || defined(SIMDE_ALWAYS_BUILD_NATIVE_TESTS) */
1223
+
1224
+ HEDLEY_DIAGNOSTIC_PUSH
1225
+ HEDLEY_DIAGNOSTIC_DISABLE_CAST_QUAL
1226
+
1227
+ static MunitTest test_suite_tests[] = {
1228
+ #if defined(SIMDE_NEON_NATIVE) || defined(SIMDE_NO_NATIVE) || defined(SIMDE_ALWAYS_BUILD_NATIVE_TESTS)
1229
+ SIMDE_TESTS_NEON_DEFINE_TEST(s8),
1230
+ SIMDE_TESTS_NEON_DEFINE_TEST(s16),
1231
+ SIMDE_TESTS_NEON_DEFINE_TEST(s32),
1232
+ SIMDE_TESTS_NEON_DEFINE_TEST(s64),
1233
+ SIMDE_TESTS_NEON_DEFINE_TEST(u8),
1234
+ SIMDE_TESTS_NEON_DEFINE_TEST(u16),
1235
+ SIMDE_TESTS_NEON_DEFINE_TEST(u32),
1236
+ SIMDE_TESTS_NEON_DEFINE_TEST(u64),
1237
+ SIMDE_TESTS_NEON_DEFINE_TEST(f32),
1238
+ SIMDE_TESTS_NEON_DEFINE_TEST(f64),
1239
+ SIMDE_TESTS_NEON_DEFINE_TEST_FULL(q, s8),
1240
+ SIMDE_TESTS_NEON_DEFINE_TEST_FULL(q, s16),
1241
+ SIMDE_TESTS_NEON_DEFINE_TEST_FULL(q, s32),
1242
+ SIMDE_TESTS_NEON_DEFINE_TEST_FULL(q, s64),
1243
+ SIMDE_TESTS_NEON_DEFINE_TEST_FULL(q, u8),
1244
+ SIMDE_TESTS_NEON_DEFINE_TEST_FULL(q, u16),
1245
+ SIMDE_TESTS_NEON_DEFINE_TEST_FULL(q, u32),
1246
+ SIMDE_TESTS_NEON_DEFINE_TEST_FULL(q, u64),
1247
+ SIMDE_TESTS_NEON_DEFINE_TEST_FULL(q, f32),
1248
+ SIMDE_TESTS_NEON_DEFINE_TEST_FULL(q, f64),
1249
+ #endif /* defined(SIMDE_NEON_NATIVE) || defined(SIMDE_NO_NATIVE) || defined(SIMDE_ALWAYS_BUILD_NATIVE_TESTS) */
1250
+
1251
+ { NULL, NULL, NULL, NULL, MUNIT_TEST_OPTION_NONE, NULL }
1252
+ };
1253
+
1254
+ HEDLEY_C_DECL MunitSuite* SIMDE_TESTS_GENERATE_SYMBOL(SIMDE_TESTS_CURRENT_NEON_OP)(void) {
1255
+ static MunitSuite suite = { (char*) "/v" HEDLEY_STRINGIFY(SIMDE_TESTS_CURRENT_NEON_OP), test_suite_tests, NULL, 1, MUNIT_SUITE_OPTION_NONE };
1256
+
1257
+ return &suite;
1258
+ }
1259
+
1260
+ HEDLEY_DIAGNOSTIC_POP