usearch 2.9.2 → 2.10.1

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (172) hide show
  1. package/README.md +39 -38
  2. package/javascript/{types → dist/cjs}/usearch.d.ts +90 -115
  3. package/javascript/dist/cjs/usearch.js +531 -0
  4. package/javascript/dist/esm/usearch.d.ts +212 -0
  5. package/javascript/{usearch.js → dist/esm/usearch.js} +199 -153
  6. package/package.json +25 -23
  7. package/prebuilds/darwin-arm64+x64/usearch.node +0 -0
  8. package/prebuilds/linux-arm64/usearch.node +0 -0
  9. package/prebuilds/linux-x64/usearch.node +0 -0
  10. package/prebuilds/win32-ia32/usearch.node +0 -0
  11. package/prebuilds/win32-x64/usearch.node +0 -0
  12. package/.clang-format +0 -10
  13. package/.clang-tidy +0 -60
  14. package/.cmake-format.py +0 -19
  15. package/.dockerignore +0 -93
  16. package/.gitattributes +0 -9
  17. package/.github/ISSUE_TEMPLATE/bug_report.yml +0 -90
  18. package/.github/ISSUE_TEMPLATE/config.yml +0 -5
  19. package/.github/ISSUE_TEMPLATE/feature_request.yml +0 -64
  20. package/.github/workflows/Dockerfile.libs +0 -36
  21. package/.github/workflows/prerelease.yml +0 -331
  22. package/.github/workflows/release.yml +0 -722
  23. package/.github/workflows/update_version.sh +0 -15
  24. package/.gitmodules +0 -9
  25. package/.releaserc +0 -90
  26. package/CITATION.cff +0 -11
  27. package/CMakeLists.txt +0 -324
  28. package/CODE_OF_CONDUCT.md +0 -128
  29. package/CONTRIBUTING.md +0 -361
  30. package/Cargo.toml +0 -19
  31. package/Dockerfile +0 -16
  32. package/VERSION +0 -1
  33. package/assets/unum.png +0 -0
  34. package/assets/usearch-approaches-transparent.png +0 -0
  35. package/assets/usearch-approaches-white.png +0 -0
  36. package/assets/usearch-neighbor-types.png +0 -0
  37. package/binding.gyp +0 -87
  38. package/build.gradle +0 -204
  39. package/build.rs +0 -18
  40. package/c/CMakeLists.txt +0 -25
  41. package/c/README.md +0 -34
  42. package/c/lib.cpp +0 -358
  43. package/c/test.c +0 -382
  44. package/c/usearch.h +0 -311
  45. package/cmake/clipp.cmake +0 -7
  46. package/cmake/config.cmake.in +0 -15
  47. package/cmake/emscripten.cmake +0 -8
  48. package/cmake/jemalloc.cmake +0 -62
  49. package/cmake/pkg-config.pc.in +0 -4
  50. package/cmake/usearchConfigVersion.cmake.in +0 -20
  51. package/conanfile.py +0 -26
  52. package/cpp/CMakeLists.txt +0 -14
  53. package/cpp/README.md +0 -161
  54. package/cpp/bench.cpp +0 -618
  55. package/cpp/test.cpp +0 -323
  56. package/fp16/CMakeLists.txt +0 -176
  57. package/fp16/LICENSE +0 -11
  58. package/fp16/README.md +0 -20
  59. package/fp16/bench/alt-element.cc +0 -89
  60. package/fp16/bench/from-alt-array.cc +0 -175
  61. package/fp16/bench/from-ieee-array.cc +0 -399
  62. package/fp16/bench/ieee-element.cc +0 -272
  63. package/fp16/bench/to-alt-array.cc +0 -91
  64. package/fp16/bench/to-ieee-array.cc +0 -303
  65. package/fp16/cmake/DownloadGoogleBenchmark.cmake +0 -15
  66. package/fp16/cmake/DownloadGoogleTest.cmake +0 -15
  67. package/fp16/cmake/DownloadPSimd.cmake +0 -15
  68. package/fp16/configure.py +0 -65
  69. package/fp16/confu.yaml +0 -8
  70. package/fp16/include/fp16/__init__.py +0 -0
  71. package/fp16/include/fp16/avx.py +0 -50
  72. package/fp16/include/fp16/avx2.py +0 -53
  73. package/fp16/include/fp16/bitcasts.h +0 -92
  74. package/fp16/include/fp16/fp16.h +0 -451
  75. package/fp16/include/fp16/psimd.h +0 -131
  76. package/fp16/include/fp16.h +0 -11
  77. package/fp16/jni/Android.mk +0 -6
  78. package/fp16/jni/Application.mk +0 -4
  79. package/fp16/test/alt-from-fp32-value.cc +0 -498
  80. package/fp16/test/alt-to-fp32-bits.cc +0 -263
  81. package/fp16/test/alt-to-fp32-psimd.cc +0 -145
  82. package/fp16/test/alt-to-fp32-value.cc +0 -348
  83. package/fp16/test/alt-to-fp32x2-psimd.cc +0 -245
  84. package/fp16/test/bitcasts.cc +0 -56
  85. package/fp16/test/ieee-from-fp32-value.cc +0 -523
  86. package/fp16/test/ieee-to-fp32-bits.cc +0 -333
  87. package/fp16/test/ieee-to-fp32-psimd.cc +0 -312
  88. package/fp16/test/ieee-to-fp32-value.cc +0 -420
  89. package/fp16/test/ieee-to-fp32x2-psimd.cc +0 -541
  90. package/fp16/test/peachpy/alt-xmm-to-fp32-xmm-avx.cc +0 -96
  91. package/fp16/test/peachpy/alt-xmm-to-fp32-ymm-avx2.cc +0 -96
  92. package/fp16/test/peachpy/stubs.py +0 -38
  93. package/fp16/test/tables.cc +0 -530
  94. package/fp16/test/tables.h +0 -13
  95. package/fp16/third-party/THHalf.h +0 -92
  96. package/fp16/third-party/eigen-half.h +0 -170
  97. package/fp16/third-party/float16-compressor.h +0 -80
  98. package/fp16/third-party/half.hpp +0 -2916
  99. package/fp16/third-party/npy-halffloat.h +0 -328
  100. package/include/usearch/index.hpp +0 -3836
  101. package/include/usearch/index_dense.hpp +0 -2004
  102. package/include/usearch/index_plugins.hpp +0 -2213
  103. package/javascript/lib.cpp +0 -341
  104. package/javascript/types/usearch.d.ts.map +0 -1
  105. package/javascript/usearch.test.js +0 -74
  106. package/package-ci.json +0 -10
  107. package/pyproject.toml +0 -51
  108. package/setup.py +0 -201
  109. package/simsimd/.releaserc +0 -86
  110. package/simsimd/CONTRIBUTING.md +0 -88
  111. package/simsimd/Cargo.lock +0 -709
  112. package/simsimd/Cargo.toml +0 -44
  113. package/simsimd/LICENSE +0 -201
  114. package/simsimd/README.md +0 -317
  115. package/simsimd/binding.gyp +0 -36
  116. package/simsimd/build.rs +0 -20
  117. package/simsimd/include/simsimd/binary.h +0 -245
  118. package/simsimd/include/simsimd/complex.h +0 -174
  119. package/simsimd/include/simsimd/probability.h +0 -548
  120. package/simsimd/include/simsimd/simsimd.h +0 -465
  121. package/simsimd/include/simsimd/spatial.h +0 -1267
  122. package/simsimd/include/simsimd/types.h +0 -208
  123. package/simsimd/javascript/bench.js +0 -127
  124. package/simsimd/javascript/fallback.ts +0 -193
  125. package/simsimd/javascript/lib.c +0 -102
  126. package/simsimd/javascript/node-gyp-build.d.ts +0 -1
  127. package/simsimd/javascript/simsimd.ts +0 -124
  128. package/simsimd/javascript/test.mjs +0 -153
  129. package/simsimd/javascript/tsconfig-base.json +0 -25
  130. package/simsimd/javascript/tsconfig-cjs.json +0 -8
  131. package/simsimd/javascript/tsconfig-esm.json +0 -8
  132. package/simsimd/package-lock.json +0 -1688
  133. package/simsimd/package.json +0 -61
  134. package/simsimd/rust/benches/cosine.rs +0 -27
  135. package/simsimd/rust/benches/native.rs +0 -30
  136. package/simsimd/rust/benches/sqeuclidean.rs +0 -27
  137. package/simsimd/rust/lib.c +0 -45
  138. package/simsimd/rust/lib.rs +0 -550
  139. package/sqlite/README.md +0 -178
  140. package/sqlite/lib.cpp +0 -325
  141. package/stringzilla/CONTRIBUTING.md +0 -446
  142. package/stringzilla/Cargo.lock +0 -25
  143. package/stringzilla/Cargo.toml +0 -26
  144. package/stringzilla/LICENSE +0 -201
  145. package/stringzilla/README.md +0 -1422
  146. package/stringzilla/VERSION +0 -1
  147. package/stringzilla/assets/cover-strinzilla.jpeg +0 -0
  148. package/stringzilla/assets/meme-stringzilla-v2.jpeg +0 -0
  149. package/stringzilla/assets/meme-stringzilla-v3.jpeg +0 -0
  150. package/stringzilla/binding.gyp +0 -16
  151. package/stringzilla/build.rs +0 -87
  152. package/stringzilla/c/lib.c +0 -337
  153. package/stringzilla/cli/README.md +0 -65
  154. package/stringzilla/cli/__init__.py +0 -0
  155. package/stringzilla/cli/split.py +0 -105
  156. package/stringzilla/cli/wc.py +0 -117
  157. package/stringzilla/include/stringzilla/experimental.h +0 -845
  158. package/stringzilla/include/stringzilla/module.modulemap +0 -9
  159. package/stringzilla/include/stringzilla/stringzilla.h +0 -5370
  160. package/stringzilla/include/stringzilla/stringzilla.hpp +0 -3838
  161. package/stringzilla/javascript/lib.c +0 -125
  162. package/stringzilla/javascript/stringzilla.js +0 -22
  163. package/stringzilla/package-ci.json +0 -10
  164. package/stringzilla/package.json +0 -32
  165. package/stringzilla/rust/lib.rs +0 -873
  166. package/stringzilla/swift/StringProtocol+StringZilla.swift +0 -280
  167. package/stringzilla/swift/Test.swift +0 -60
  168. package/tsconfig.json +0 -21
  169. package/wasmer.toml +0 -12
  170. package/winlibinstaller.bat +0 -42
  171. /package/{simsimd/javascript/dist-package-cjs.json → javascript/dist/cjs/package.json} +0 -0
  172. /package/{simsimd/javascript/dist-package-esm.json → javascript/dist/esm/package.json} +0 -0
package/fp16/test/ieee-to-fp32x2-psimd.cc DELETED
@@ -1,541 +0,0 @@
1
- #include <gtest/gtest.h>
2
-
3
- #include <cstdint>
4
-
5
- #include <fp16.h>
6
- #include <fp16/psimd.h>
7
-
8
-
9
- TEST(FP16_IEEE_TO_FP32x2_PSIMD, infinity) {
10
- const uint16_t positive_infinity_f16 = UINT16_C(0x7C00);
11
- const uint16_t negative_infinity_f16 = UINT16_C(0xFC00);
12
-
13
- const uint32_t positive_infinity_f32 = UINT32_C(0x7F800000);
14
- const uint32_t negative_infinity_f32 = UINT32_C(0xFF800000);
15
-
16
- const psimd_u16 fp16 = {
17
- positive_infinity_f16, negative_infinity_f16,
18
- negative_infinity_f16, positive_infinity_f16,
19
- positive_infinity_f16, positive_infinity_f16,
20
- negative_infinity_f16, negative_infinity_f16
21
- };
22
- const psimd_u32x2 fp32 =
23
- psimd_cast_f32x2_u32x2(fp16_ieee_to_fp32x2_psimd(fp16));
24
-
25
- EXPECT_EQ(positive_infinity_f32, fp32.lo[0]) <<
26
- std::hex << std::uppercase << std::setfill('0') <<
27
- "F16 = 0x" << std::setw(4) << fp16[0] << ", " <<
28
- "F32(F16) = 0x" << std::setw(8) << fp32.lo[0] << ", " <<
29
- "F32 = 0x" << std::setw(8) << positive_infinity_f32;
30
- EXPECT_EQ(negative_infinity_f32, fp32.lo[1]) <<
31
- std::hex << std::uppercase << std::setfill('0') <<
32
- "F16 = 0x" << std::setw(4) << fp16[1] << ", " <<
33
- "F32(F16) = 0x" << std::setw(8) << fp32.lo[1] << ", " <<
34
- "F32 = 0x" << std::setw(8) << negative_infinity_f32;
35
- EXPECT_EQ(negative_infinity_f32, fp32.lo[2]) <<
36
- std::hex << std::uppercase << std::setfill('0') <<
37
- "F16 = 0x" << std::setw(4) << fp16[2] << ", " <<
38
- "F32(F16) = 0x" << std::setw(8) << fp32.lo[2] << ", " <<
39
- "F32 = 0x" << std::setw(8) << negative_infinity_f32;
40
- EXPECT_EQ(positive_infinity_f32, fp32.lo[3]) <<
41
- std::hex << std::uppercase << std::setfill('0') <<
42
- "F16 = 0x" << std::setw(4) << fp16[3] << ", " <<
43
- "F32(F16) = 0x" << std::setw(8) << fp32.lo[3] << ", " <<
44
- "F32 = 0x" << std::setw(8) << positive_infinity_f32;
45
- EXPECT_EQ(positive_infinity_f32, fp32.hi[0]) <<
46
- std::hex << std::uppercase << std::setfill('0') <<
47
- "F16 = 0x" << std::setw(4) << fp16[4] << ", " <<
48
- "F32(F16) = 0x" << std::setw(8) << fp32.hi[0] << ", " <<
49
- "F32 = 0x" << std::setw(8) << positive_infinity_f32;
50
- EXPECT_EQ(positive_infinity_f32, fp32.hi[1]) <<
51
- std::hex << std::uppercase << std::setfill('0') <<
52
- "F16 = 0x" << std::setw(4) << fp16[5] << ", " <<
53
- "F32(F16) = 0x" << std::setw(8) << fp32.hi[1] << ", " <<
54
- "F32 = 0x" << std::setw(8) << positive_infinity_f32;
55
- EXPECT_EQ(negative_infinity_f32, fp32.hi[2]) <<
56
- std::hex << std::uppercase << std::setfill('0') <<
57
- "F16 = 0x" << std::setw(4) << fp16[6] << ", " <<
58
- "F32(F16) = 0x" << std::setw(8) << fp32.hi[2] << ", " <<
59
- "F32 = 0x" << std::setw(8) << negative_infinity_f32;
60
- EXPECT_EQ(negative_infinity_f32, fp32.hi[3]) <<
61
- std::hex << std::uppercase << std::setfill('0') <<
62
- "F16 = 0x" << std::setw(4) << fp16[7] << ", " <<
63
- "F32(F16) = 0x" << std::setw(8) << fp32.hi[3] << ", " <<
64
- "F32 = 0x" << std::setw(8) << negative_infinity_f32;
65
- }
66
-
67
- TEST(FP16_IEEE_TO_FP32x2_PSIMD, positive_nan) {
68
- for (uint16_t h = 0; h < 0x0400; h += 8) {
69
- const psimd_u16 fp16 = {
70
- (uint16_t) (h + 0x7C00 + (h == 0)) /* Avoid infinity */,
71
- (uint16_t) (h + 0x7C01),
72
- (uint16_t) (h + 0x7C02),
73
- (uint16_t) (h + 0x7C03),
74
- (uint16_t) (h + 0x7C04),
75
- (uint16_t) (h + 0x7C05),
76
- (uint16_t) (h + 0x7C06),
77
- (uint16_t) (h + 0x7C07),
78
- };
79
- const psimd_u32x2 fp32 =
80
- psimd_cast_f32x2_u32x2(fp16_ieee_to_fp32x2_psimd(fp16));
81
-
82
- /* Check sign */
83
- EXPECT_EQ(fp32.lo[0] & UINT32_C(0x80000000), 0) <<
84
- std::hex << std::uppercase << std::setfill('0') <<
85
- "F16 = 0x" << std::setw(4) << fp16[0] << ", " <<
86
- "F32(F16) = 0x" << std::setw(8) << fp32.lo[0];
87
- EXPECT_EQ(fp32.lo[1] & UINT32_C(0x80000000), 0) <<
88
- std::hex << std::uppercase << std::setfill('0') <<
89
- "F16 = 0x" << std::setw(4) << fp16[1] << ", " <<
90
- "F32(F16) = 0x" << std::setw(8) << fp32.lo[1];
91
- EXPECT_EQ(fp32.lo[2] & UINT32_C(0x80000000), 0) <<
92
- std::hex << std::uppercase << std::setfill('0') <<
93
- "F16 = 0x" << std::setw(4) << fp16[2] << ", " <<
94
- "F32(F16) = 0x" << std::setw(8) << fp32.lo[2];
95
- EXPECT_EQ(fp32.lo[3] & UINT32_C(0x80000000), 0) <<
96
- std::hex << std::uppercase << std::setfill('0') <<
97
- "F16 = 0x" << std::setw(4) << fp16[3] << ", " <<
98
- "F32(F16) = 0x" << std::setw(8) << fp32.lo[3];
99
- EXPECT_EQ(fp32.hi[0] & UINT32_C(0x80000000), 0) <<
100
- std::hex << std::uppercase << std::setfill('0') <<
101
- "F16 = 0x" << std::setw(4) << fp16[4] << ", " <<
102
- "F32(F16) = 0x" << std::setw(8) << fp32.hi[0];
103
- EXPECT_EQ(fp32.hi[1] & UINT32_C(0x80000000), 0) <<
104
- std::hex << std::uppercase << std::setfill('0') <<
105
- "F16 = 0x" << std::setw(4) << fp16[5] << ", " <<
106
- "F32(F16) = 0x" << std::setw(8) << fp32.hi[1];
107
- EXPECT_EQ(fp32.hi[2] & UINT32_C(0x80000000), 0) <<
108
- std::hex << std::uppercase << std::setfill('0') <<
109
- "F16 = 0x" << std::setw(4) << fp16[6] << ", " <<
110
- "F32(F16) = 0x" << std::setw(8) << fp32.hi[2];
111
- EXPECT_EQ(fp32.hi[3] & UINT32_C(0x80000000), 0) <<
112
- std::hex << std::uppercase << std::setfill('0') <<
113
- "F16 = 0x" << std::setw(4) << fp16[7] << ", " <<
114
- "F32(F16) = 0x" << std::setw(8) << fp32.hi[3];
115
-
116
- /* Check exponent */
117
- EXPECT_EQ(fp32.lo[0] & UINT32_C(0x7F800000), UINT32_C(0x7F800000)) <<
118
- std::hex << std::uppercase << std::setfill('0') <<
119
- "F16 = 0x" << std::setw(4) << fp16[0] << ", " <<
120
- "F32(F16) = 0x" << std::setw(8) << fp32.lo[0];
121
- EXPECT_EQ(fp32.lo[1] & UINT32_C(0x7F800000), UINT32_C(0x7F800000)) <<
122
- std::hex << std::uppercase << std::setfill('0') <<
123
- "F16 = 0x" << std::setw(4) << fp16[1] << ", " <<
124
- "F32(F16) = 0x" << std::setw(8) << fp32.lo[1];
125
- EXPECT_EQ(fp32.lo[2] & UINT32_C(0x7F800000), UINT32_C(0x7F800000)) <<
126
- std::hex << std::uppercase << std::setfill('0') <<
127
- "F16 = 0x" << std::setw(4) << fp16[2] << ", " <<
128
- "F32(F16) = 0x" << std::setw(8) << fp32.lo[2];
129
- EXPECT_EQ(fp32.lo[3] & UINT32_C(0x7F800000), UINT32_C(0x7F800000)) <<
130
- std::hex << std::uppercase << std::setfill('0') <<
131
- "F16 = 0x" << std::setw(4) << fp16[3] << ", " <<
132
- "F32(F16) = 0x" << std::setw(8) << fp32.lo[3];
133
- EXPECT_EQ(fp32.hi[0] & UINT32_C(0x7F800000), UINT32_C(0x7F800000)) <<
134
- std::hex << std::uppercase << std::setfill('0') <<
135
- "F16 = 0x" << std::setw(4) << fp16[4] << ", " <<
136
- "F32(F16) = 0x" << std::setw(8) << fp32.hi[0];
137
- EXPECT_EQ(fp32.hi[1] & UINT32_C(0x7F800000), UINT32_C(0x7F800000)) <<
138
- std::hex << std::uppercase << std::setfill('0') <<
139
- "F16 = 0x" << std::setw(4) << fp16[5] << ", " <<
140
- "F32(F16) = 0x" << std::setw(8) << fp32.hi[1];
141
- EXPECT_EQ(fp32.hi[2] & UINT32_C(0x7F800000), UINT32_C(0x7F800000)) <<
142
- std::hex << std::uppercase << std::setfill('0') <<
143
- "F16 = 0x" << std::setw(4) << fp16[6] << ", " <<
144
- "F32(F16) = 0x" << std::setw(8) << fp32.hi[2];
145
- EXPECT_EQ(fp32.hi[3] & UINT32_C(0x7F800000), UINT32_C(0x7F800000)) <<
146
- std::hex << std::uppercase << std::setfill('0') <<
147
- "F16 = 0x" << std::setw(4) << fp16[7] << ", " <<
148
- "F32(F16) = 0x" << std::setw(8) << fp32.hi[3];
149
-
150
- /* Check mantissa */
151
- EXPECT_NE(fp32.lo[0] & UINT32_C(0x007FFFFF), UINT32_C(0)) <<
152
- std::hex << std::uppercase << std::setfill('0') <<
153
- "F16 = 0x" << std::setw(4) << fp16[0] << ", " <<
154
- "F32(F16) = 0x" << std::setw(8) << fp32.lo[0];
155
- EXPECT_NE(fp32.lo[1] & UINT32_C(0x007FFFFF), UINT32_C(0)) <<
156
- std::hex << std::uppercase << std::setfill('0') <<
157
- "F16 = 0x" << std::setw(4) << fp16[1] << ", " <<
158
- "F32(F16) = 0x" << std::setw(8) << fp32.lo[1];
159
- EXPECT_NE(fp32.lo[2] & UINT32_C(0x007FFFFF), UINT32_C(0)) <<
160
- std::hex << std::uppercase << std::setfill('0') <<
161
- "F16 = 0x" << std::setw(4) << fp16[2] << ", " <<
162
- "F32(F16) = 0x" << std::setw(8) << fp32.lo[2];
163
- EXPECT_NE(fp32.lo[3] & UINT32_C(0x007FFFFF), UINT32_C(0)) <<
164
- std::hex << std::uppercase << std::setfill('0') <<
165
- "F16 = 0x" << std::setw(4) << fp16[3] << ", " <<
166
- "F32(F16) = 0x" << std::setw(8) << fp32.lo[3];
167
- EXPECT_NE(fp32.hi[0] & UINT32_C(0x007FFFFF), UINT32_C(0)) <<
168
- std::hex << std::uppercase << std::setfill('0') <<
169
- "F16 = 0x" << std::setw(4) << fp16[4] << ", " <<
170
- "F32(F16) = 0x" << std::setw(8) << fp32.hi[0];
171
- EXPECT_NE(fp32.hi[1] & UINT32_C(0x007FFFFF), UINT32_C(0)) <<
172
- std::hex << std::uppercase << std::setfill('0') <<
173
- "F16 = 0x" << std::setw(4) << fp16[5] << ", " <<
174
- "F32(F16) = 0x" << std::setw(8) << fp32.hi[1];
175
- EXPECT_NE(fp32.hi[2] & UINT32_C(0x007FFFFF), UINT32_C(0)) <<
176
- std::hex << std::uppercase << std::setfill('0') <<
177
- "F16 = 0x" << std::setw(4) << fp16[6] << ", " <<
178
- "F32(F16) = 0x" << std::setw(8) << fp32.hi[2];
179
- EXPECT_NE(fp32.hi[3] & UINT32_C(0x007FFFFF), UINT32_C(0)) <<
180
- std::hex << std::uppercase << std::setfill('0') <<
181
- "F16 = 0x" << std::setw(4) << fp16[7] << ", " <<
182
- "F32(F16) = 0x" << std::setw(8) << fp32.hi[3];
183
- }
184
- }
185
-
186
- TEST(FP16_IEEE_TO_FP32x2_PSIMD, negative_nan) {
187
- for (uint16_t h = 0; h < 0x0400; h += 8) {
188
- const psimd_u16 fp16 = {
189
- (uint16_t) (h + 0xFC00 + (h == 0)) /* Avoid infinity */,
190
- (uint16_t) (h + 0xFC01),
191
- (uint16_t) (h + 0xFC02),
192
- (uint16_t) (h + 0xFC03),
193
- (uint16_t) (h + 0xFC04),
194
- (uint16_t) (h + 0xFC05),
195
- (uint16_t) (h + 0xFC06),
196
- (uint16_t) (h + 0xFC07),
197
- };
198
- const psimd_u32x2 fp32 =
199
- psimd_cast_f32x2_u32x2(fp16_ieee_to_fp32x2_psimd(fp16));
200
-
201
- /* Check sign */
202
- EXPECT_EQ(fp32.lo[0] & UINT32_C(0x80000000), UINT32_C(0x80000000)) <<
203
- std::hex << std::uppercase << std::setfill('0') <<
204
- "F16 = 0x" << std::setw(4) << fp16[0] << ", " <<
205
- "F32(F16) = 0x" << std::setw(8) << fp32.lo[0];
206
- EXPECT_EQ(fp32.lo[1] & UINT32_C(0x80000000), UINT32_C(0x80000000)) <<
207
- std::hex << std::uppercase << std::setfill('0') <<
208
- "F16 = 0x" << std::setw(4) << fp16[1] << ", " <<
209
- "F32(F16) = 0x" << std::setw(8) << fp32.lo[1];
210
- EXPECT_EQ(fp32.lo[2] & UINT32_C(0x80000000), UINT32_C(0x80000000)) <<
211
- std::hex << std::uppercase << std::setfill('0') <<
212
- "F16 = 0x" << std::setw(4) << fp16[2] << ", " <<
213
- "F32(F16) = 0x" << std::setw(8) << fp32.lo[2];
214
- EXPECT_EQ(fp32.lo[3] & UINT32_C(0x80000000), UINT32_C(0x80000000)) <<
215
- std::hex << std::uppercase << std::setfill('0') <<
216
- "F16 = 0x" << std::setw(4) << fp16[3] << ", " <<
217
- "F32(F16) = 0x" << std::setw(8) << fp32.lo[3];
218
- EXPECT_EQ(fp32.hi[0] & UINT32_C(0x80000000), UINT32_C(0x80000000)) <<
219
- std::hex << std::uppercase << std::setfill('0') <<
220
- "F16 = 0x" << std::setw(4) << fp16[4] << ", " <<
221
- "F32(F16) = 0x" << std::setw(8) << fp32.hi[0];
222
- EXPECT_EQ(fp32.hi[1] & UINT32_C(0x80000000), UINT32_C(0x80000000)) <<
223
- std::hex << std::uppercase << std::setfill('0') <<
224
- "F16 = 0x" << std::setw(4) << fp16[5] << ", " <<
225
- "F32(F16) = 0x" << std::setw(8) << fp32.hi[1];
226
- EXPECT_EQ(fp32.hi[2] & UINT32_C(0x80000000), UINT32_C(0x80000000)) <<
227
- std::hex << std::uppercase << std::setfill('0') <<
228
- "F16 = 0x" << std::setw(4) << fp16[6] << ", " <<
229
- "F32(F16) = 0x" << std::setw(8) << fp32.hi[2];
230
- EXPECT_EQ(fp32.hi[3] & UINT32_C(0x80000000), UINT32_C(0x80000000)) <<
231
- std::hex << std::uppercase << std::setfill('0') <<
232
- "F16 = 0x" << std::setw(4) << fp16[7] << ", " <<
233
- "F32(F16) = 0x" << std::setw(8) << fp32.hi[3];
234
-
235
- /* Check exponent */
236
- EXPECT_EQ(fp32.lo[0] & UINT32_C(0x7F800000), UINT32_C(0x7F800000)) <<
237
- std::hex << std::uppercase << std::setfill('0') <<
238
- "F16 = 0x" << std::setw(4) << fp16[0] << ", " <<
239
- "F32(F16) = 0x" << std::setw(8) << fp32.lo[0];
240
- EXPECT_EQ(fp32.lo[1] & UINT32_C(0x7F800000), UINT32_C(0x7F800000)) <<
241
- std::hex << std::uppercase << std::setfill('0') <<
242
- "F16 = 0x" << std::setw(4) << fp16[1] << ", " <<
243
- "F32(F16) = 0x" << std::setw(8) << fp32.lo[1];
244
- EXPECT_EQ(fp32.lo[2] & UINT32_C(0x7F800000), UINT32_C(0x7F800000)) <<
245
- std::hex << std::uppercase << std::setfill('0') <<
246
- "F16 = 0x" << std::setw(4) << fp16[2] << ", " <<
247
- "F32(F16) = 0x" << std::setw(8) << fp32.lo[2];
248
- EXPECT_EQ(fp32.lo[3] & UINT32_C(0x7F800000), UINT32_C(0x7F800000)) <<
249
- std::hex << std::uppercase << std::setfill('0') <<
250
- "F16 = 0x" << std::setw(4) << fp16[3] << ", " <<
251
- "F32(F16) = 0x" << std::setw(8) << fp32.lo[3];
252
- EXPECT_EQ(fp32.hi[0] & UINT32_C(0x7F800000), UINT32_C(0x7F800000)) <<
253
- std::hex << std::uppercase << std::setfill('0') <<
254
- "F16 = 0x" << std::setw(4) << fp16[4] << ", " <<
255
- "F32(F16) = 0x" << std::setw(8) << fp32.hi[0];
256
- EXPECT_EQ(fp32.hi[1] & UINT32_C(0x7F800000), UINT32_C(0x7F800000)) <<
257
- std::hex << std::uppercase << std::setfill('0') <<
258
- "F16 = 0x" << std::setw(4) << fp16[5] << ", " <<
259
- "F32(F16) = 0x" << std::setw(8) << fp32.hi[1];
260
- EXPECT_EQ(fp32.hi[2] & UINT32_C(0x7F800000), UINT32_C(0x7F800000)) <<
261
- std::hex << std::uppercase << std::setfill('0') <<
262
- "F16 = 0x" << std::setw(4) << fp16[6] << ", " <<
263
- "F32(F16) = 0x" << std::setw(8) << fp32.hi[2];
264
- EXPECT_EQ(fp32.hi[3] & UINT32_C(0x7F800000), UINT32_C(0x7F800000)) <<
265
- std::hex << std::uppercase << std::setfill('0') <<
266
- "F16 = 0x" << std::setw(4) << fp16[7] << ", " <<
267
- "F32(F16) = 0x" << std::setw(8) << fp32.hi[3];
268
-
269
- /* Check mantissa */
270
- EXPECT_NE(fp32.lo[0] & UINT32_C(0x007FFFFF), UINT32_C(0)) <<
271
- std::hex << std::uppercase << std::setfill('0') <<
272
- "F16 = 0x" << std::setw(4) << fp16[0] << ", " <<
273
- "F32(F16) = 0x" << std::setw(8) << fp32.lo[0];
274
- EXPECT_NE(fp32.lo[1] & UINT32_C(0x007FFFFF), UINT32_C(0)) <<
275
- std::hex << std::uppercase << std::setfill('0') <<
276
- "F16 = 0x" << std::setw(4) << fp16[1] << ", " <<
277
- "F32(F16) = 0x" << std::setw(8) << fp32.lo[1];
278
- EXPECT_NE(fp32.lo[2] & UINT32_C(0x007FFFFF), UINT32_C(0)) <<
279
- std::hex << std::uppercase << std::setfill('0') <<
280
- "F16 = 0x" << std::setw(4) << fp16[2] << ", " <<
281
- "F32(F16) = 0x" << std::setw(8) << fp32.lo[2];
282
- EXPECT_NE(fp32.lo[3] & UINT32_C(0x007FFFFF), UINT32_C(0)) <<
283
- std::hex << std::uppercase << std::setfill('0') <<
284
- "F16 = 0x" << std::setw(4) << fp16[3] << ", " <<
285
- "F32(F16) = 0x" << std::setw(8) << fp32.lo[3];
286
- EXPECT_NE(fp32.hi[0] & UINT32_C(0x007FFFFF), UINT32_C(0)) <<
287
- std::hex << std::uppercase << std::setfill('0') <<
288
- "F16 = 0x" << std::setw(4) << fp16[4] << ", " <<
289
- "F32(F16) = 0x" << std::setw(8) << fp32.hi[0];
290
- EXPECT_NE(fp32.hi[1] & UINT32_C(0x007FFFFF), UINT32_C(0)) <<
291
- std::hex << std::uppercase << std::setfill('0') <<
292
- "F16 = 0x" << std::setw(4) << fp16[5] << ", " <<
293
- "F32(F16) = 0x" << std::setw(8) << fp32.hi[1];
294
- EXPECT_NE(fp32.hi[2] & UINT32_C(0x007FFFFF), UINT32_C(0)) <<
295
- std::hex << std::uppercase << std::setfill('0') <<
296
- "F16 = 0x" << std::setw(4) << fp16[6] << ", " <<
297
- "F32(F16) = 0x" << std::setw(8) << fp32.hi[2];
298
- EXPECT_NE(fp32.hi[3] & UINT32_C(0x007FFFFF), UINT32_C(0)) <<
299
- std::hex << std::uppercase << std::setfill('0') <<
300
- "F16 = 0x" << std::setw(4) << fp16[7] << ", " <<
301
- "F32(F16) = 0x" << std::setw(8) << fp32.hi[3];
302
- }
303
- }
304
-
305
- TEST(FP16_IEEE_TO_FP32x2_PSIMD, positive_normalized_values) {
306
- const uint32_t exponentBias = 15;
307
- for (int32_t e = -14; e <= 15; e++) {
308
- for (uint16_t h = 0; h < 0x0400; h += 8) {
309
- const psimd_u16 fp16 = {
310
- (uint16_t) (h + ((e + exponentBias) << 10) + 0),
311
- (uint16_t) (h + ((e + exponentBias) << 10) + 1),
312
- (uint16_t) (h + ((e + exponentBias) << 10) + 2),
313
- (uint16_t) (h + ((e + exponentBias) << 10) + 3),
314
- (uint16_t) (h + ((e + exponentBias) << 10) + 4),
315
- (uint16_t) (h + ((e + exponentBias) << 10) + 5),
316
- (uint16_t) (h + ((e + exponentBias) << 10) + 6),
317
- (uint16_t) (h + ((e + exponentBias) << 10) + 7)
318
- };
319
- const psimd_u32x2 fp32 =
320
- psimd_cast_f32x2_u32x2(fp16_ieee_to_fp32x2_psimd(fp16));
321
-
322
- EXPECT_EQ(fp16_ieee_to_fp32_bits(fp16[0]), fp32.lo[0]) <<
323
- std::hex << std::uppercase << std::setfill('0') <<
324
- "F16 = 0x" << std::setw(4) << fp16[0] << ", " <<
325
- "F32(F16) = 0x" << std::setw(8) << fp32.lo[0] << ", " <<
326
- "F32 = 0x" << std::setw(8) << fp16_ieee_to_fp32_bits(fp16[0]);
327
- EXPECT_EQ(fp16_ieee_to_fp32_bits(fp16[1]), fp32.lo[1]) <<
328
- std::hex << std::uppercase << std::setfill('0') <<
329
- "F16 = 0x" << std::setw(4) << fp16[1] << ", " <<
330
- "F32(F16) = 0x" << std::setw(8) << fp32.lo[1] << ", " <<
331
- "F32 = 0x" << std::setw(8) << fp16_ieee_to_fp32_bits(fp16[1]);
332
- EXPECT_EQ(fp16_ieee_to_fp32_bits(fp16[2]), fp32.lo[2]) <<
333
- std::hex << std::uppercase << std::setfill('0') <<
334
- "F16 = 0x" << std::setw(4) << fp16[2] << ", " <<
335
- "F32(F16) = 0x" << std::setw(8) << fp32.lo[2] << ", " <<
336
- "F32 = 0x" << std::setw(8) << fp16_ieee_to_fp32_bits(fp16[2]);
337
- EXPECT_EQ(fp16_ieee_to_fp32_bits(fp16[3]), fp32.lo[3]) <<
338
- std::hex << std::uppercase << std::setfill('0') <<
339
- "F16 = 0x" << std::setw(4) << fp16[3] << ", " <<
340
- "F32(F16) = 0x" << std::setw(8) << fp32.lo[3] << ", " <<
341
- "F32 = 0x" << std::setw(8) << fp16_ieee_to_fp32_bits(fp16[3]);
342
- EXPECT_EQ(fp16_ieee_to_fp32_bits(fp16[4]), fp32.hi[0]) <<
343
- std::hex << std::uppercase << std::setfill('0') <<
344
- "F16 = 0x" << std::setw(4) << fp16[4] << ", " <<
345
- "F32(F16) = 0x" << std::setw(8) << fp32.hi[0] << ", " <<
346
- "F32 = 0x" << std::setw(8) << fp16_ieee_to_fp32_bits(fp16[4]);
347
- EXPECT_EQ(fp16_ieee_to_fp32_bits(fp16[5]), fp32.hi[1]) <<
348
- std::hex << std::uppercase << std::setfill('0') <<
349
- "F16 = 0x" << std::setw(4) << fp16[5] << ", " <<
350
- "F32(F16) = 0x" << std::setw(8) << fp32.hi[1] << ", " <<
351
- "F32 = 0x" << std::setw(8) << fp16_ieee_to_fp32_bits(fp16[5]);
352
- EXPECT_EQ(fp16_ieee_to_fp32_bits(fp16[6]), fp32.hi[2]) <<
353
- std::hex << std::uppercase << std::setfill('0') <<
354
- "F16 = 0x" << std::setw(4) << fp16[6] << ", " <<
355
- "F32(F16) = 0x" << std::setw(8) << fp32.hi[2] << ", " <<
356
- "F32 = 0x" << std::setw(8) << fp16_ieee_to_fp32_bits(fp16[6]);
357
- EXPECT_EQ(fp16_ieee_to_fp32_bits(fp16[7]), fp32.hi[3]) <<
358
- std::hex << std::uppercase << std::setfill('0') <<
359
- "F16 = 0x" << std::setw(4) << fp16[7] << ", " <<
360
- "F32(F16) = 0x" << std::setw(8) << fp32.hi[3] << ", " <<
361
- "F32 = 0x" << std::setw(8) << fp16_ieee_to_fp32_bits(fp16[7]);
362
- }
363
- }
364
- }
365
-
366
- TEST(FP16_IEEE_TO_FP32x2_PSIMD, negative_normalized_values) {
367
- const uint32_t exponentBias = 15;
368
- for (int32_t e = -14; e <= 15; e++) {
369
- for (uint16_t h = 0; h < 0x0400; h += 8) {
370
- const psimd_u16 fp16 = {
371
- (uint16_t) (h + ((e + exponentBias) << 10) + 0x8000),
372
- (uint16_t) (h + ((e + exponentBias) << 10) + 0x8001),
373
- (uint16_t) (h + ((e + exponentBias) << 10) + 0x8002),
374
- (uint16_t) (h + ((e + exponentBias) << 10) + 0x8003),
375
- (uint16_t) (h + ((e + exponentBias) << 10) + 0x8004),
376
- (uint16_t) (h + ((e + exponentBias) << 10) + 0x8005),
377
- (uint16_t) (h + ((e + exponentBias) << 10) + 0x8006),
378
- (uint16_t) (h + ((e + exponentBias) << 10) + 0x8007)
379
- };
380
- const psimd_u32x2 fp32 =
381
- psimd_cast_f32x2_u32x2(fp16_ieee_to_fp32x2_psimd(fp16));
382
-
383
- EXPECT_EQ(fp16_ieee_to_fp32_bits(fp16[0]), fp32.lo[0]) <<
384
- std::hex << std::uppercase << std::setfill('0') <<
385
- "F16 = 0x" << std::setw(4) << fp16[0] << ", " <<
386
- "F32(F16) = 0x" << std::setw(8) << fp32.lo[0] << ", " <<
387
- "F32 = 0x" << std::setw(8) << fp16_ieee_to_fp32_bits(fp16[0]);
388
- EXPECT_EQ(fp16_ieee_to_fp32_bits(fp16[1]), fp32.lo[1]) <<
389
- std::hex << std::uppercase << std::setfill('0') <<
390
- "F16 = 0x" << std::setw(4) << fp16[1] << ", " <<
391
- "F32(F16) = 0x" << std::setw(8) << fp32.lo[1] << ", " <<
392
- "F32 = 0x" << std::setw(8) << fp16_ieee_to_fp32_bits(fp16[1]);
393
- EXPECT_EQ(fp16_ieee_to_fp32_bits(fp16[2]), fp32.lo[2]) <<
394
- std::hex << std::uppercase << std::setfill('0') <<
395
- "F16 = 0x" << std::setw(4) << fp16[2] << ", " <<
396
- "F32(F16) = 0x" << std::setw(8) << fp32.lo[2] << ", " <<
397
- "F32 = 0x" << std::setw(8) << fp16_ieee_to_fp32_bits(fp16[2]);
398
- EXPECT_EQ(fp16_ieee_to_fp32_bits(fp16[3]), fp32.lo[3]) <<
399
- std::hex << std::uppercase << std::setfill('0') <<
400
- "F16 = 0x" << std::setw(4) << fp16[3] << ", " <<
401
- "F32(F16) = 0x" << std::setw(8) << fp32.lo[3] << ", " <<
402
- "F32 = 0x" << std::setw(8) << fp16_ieee_to_fp32_bits(fp16[3]);
403
- EXPECT_EQ(fp16_ieee_to_fp32_bits(fp16[4]), fp32.hi[0]) <<
404
- std::hex << std::uppercase << std::setfill('0') <<
405
- "F16 = 0x" << std::setw(4) << fp16[4] << ", " <<
406
- "F32(F16) = 0x" << std::setw(8) << fp32.hi[0] << ", " <<
407
- "F32 = 0x" << std::setw(8) << fp16_ieee_to_fp32_bits(fp16[4]);
408
- EXPECT_EQ(fp16_ieee_to_fp32_bits(fp16[5]), fp32.hi[1]) <<
409
- std::hex << std::uppercase << std::setfill('0') <<
410
- "F16 = 0x" << std::setw(4) << fp16[5] << ", " <<
411
- "F32(F16) = 0x" << std::setw(8) << fp32.hi[1] << ", " <<
412
- "F32 = 0x" << std::setw(8) << fp16_ieee_to_fp32_bits(fp16[5]);
413
- EXPECT_EQ(fp16_ieee_to_fp32_bits(fp16[6]), fp32.hi[2]) <<
414
- std::hex << std::uppercase << std::setfill('0') <<
415
- "F16 = 0x" << std::setw(4) << fp16[6] << ", " <<
416
- "F32(F16) = 0x" << std::setw(8) << fp32.hi[2] << ", " <<
417
- "F32 = 0x" << std::setw(8) << fp16_ieee_to_fp32_bits(fp16[6]);
418
- EXPECT_EQ(fp16_ieee_to_fp32_bits(fp16[7]), fp32.hi[3]) <<
419
- std::hex << std::uppercase << std::setfill('0') <<
420
- "F16 = 0x" << std::setw(4) << fp16[7] << ", " <<
421
- "F32(F16) = 0x" << std::setw(8) << fp32.hi[3] << ", " <<
422
- "F32 = 0x" << std::setw(8) << fp16_ieee_to_fp32_bits(fp16[7]);
423
- }
424
- }
425
- }
426
-
427
- TEST(FP16_IEEE_TO_FP32x2_PSIMD, positive_denormalized_values) {
428
- for (uint16_t h = 0; h < 0x0400; h += 8) {
429
- const psimd_u16 fp16 = {
430
- (uint16_t) (h + 0),
431
- (uint16_t) (h + 1),
432
- (uint16_t) (h + 2),
433
- (uint16_t) (h + 3),
434
- (uint16_t) (h + 4),
435
- (uint16_t) (h + 5),
436
- (uint16_t) (h + 6),
437
- (uint16_t) (h + 7)
438
- };
439
- const psimd_u32x2 fp32 =
440
- psimd_cast_f32x2_u32x2(fp16_ieee_to_fp32x2_psimd(fp16));
441
-
442
- EXPECT_EQ(fp16_ieee_to_fp32_bits(fp16[0]), fp32.lo[0]) <<
443
- std::hex << std::uppercase << std::setfill('0') <<
444
- "F16 = 0x" << std::setw(4) << fp16[0] << ", " <<
445
- "F32(F16) = 0x" << std::setw(8) << fp32.lo[0] << ", " <<
446
- "F32 = 0x" << std::setw(8) << fp16_ieee_to_fp32_bits(fp16[0]);
447
- EXPECT_EQ(fp16_ieee_to_fp32_bits(fp16[1]), fp32.lo[1]) <<
448
- std::hex << std::uppercase << std::setfill('0') <<
449
- "F16 = 0x" << std::setw(4) << fp16[1] << ", " <<
450
- "F32(F16) = 0x" << std::setw(8) << fp32.lo[1] << ", " <<
451
- "F32 = 0x" << std::setw(8) << fp16_ieee_to_fp32_bits(fp16[1]);
452
- EXPECT_EQ(fp16_ieee_to_fp32_bits(fp16[2]), fp32.lo[2]) <<
453
- std::hex << std::uppercase << std::setfill('0') <<
454
- "F16 = 0x" << std::setw(4) << fp16[2] << ", " <<
455
- "F32(F16) = 0x" << std::setw(8) << fp32.lo[2] << ", " <<
456
- "F32 = 0x" << std::setw(8) << fp16_ieee_to_fp32_bits(fp16[2]);
457
- EXPECT_EQ(fp16_ieee_to_fp32_bits(fp16[3]), fp32.lo[3]) <<
458
- std::hex << std::uppercase << std::setfill('0') <<
459
- "F16 = 0x" << std::setw(4) << fp16[3] << ", " <<
460
- "F32(F16) = 0x" << std::setw(8) << fp32.lo[3] << ", " <<
461
- "F32 = 0x" << std::setw(8) << fp16_ieee_to_fp32_bits(fp16[3]);
462
- EXPECT_EQ(fp16_ieee_to_fp32_bits(fp16[4]), fp32.hi[0]) <<
463
- std::hex << std::uppercase << std::setfill('0') <<
464
- "F16 = 0x" << std::setw(4) << fp16[4] << ", " <<
465
- "F32(F16) = 0x" << std::setw(8) << fp32.hi[0] << ", " <<
466
- "F32 = 0x" << std::setw(8) << fp16_ieee_to_fp32_bits(fp16[4]);
467
- EXPECT_EQ(fp16_ieee_to_fp32_bits(fp16[5]), fp32.hi[1]) <<
468
- std::hex << std::uppercase << std::setfill('0') <<
469
- "F16 = 0x" << std::setw(4) << fp16[5] << ", " <<
470
- "F32(F16) = 0x" << std::setw(8) << fp32.hi[1] << ", " <<
471
- "F32 = 0x" << std::setw(8) << fp16_ieee_to_fp32_bits(fp16[5]);
472
- EXPECT_EQ(fp16_ieee_to_fp32_bits(fp16[6]), fp32.hi[2]) <<
473
- std::hex << std::uppercase << std::setfill('0') <<
474
- "F16 = 0x" << std::setw(4) << fp16[6] << ", " <<
475
- "F32(F16) = 0x" << std::setw(8) << fp32.hi[2] << ", " <<
476
- "F32 = 0x" << std::setw(8) << fp16_ieee_to_fp32_bits(fp16[6]);
477
- EXPECT_EQ(fp16_ieee_to_fp32_bits(fp16[7]), fp32.hi[3]) <<
478
- std::hex << std::uppercase << std::setfill('0') <<
479
- "F16 = 0x" << std::setw(4) << fp16[7] << ", " <<
480
- "F32(F16) = 0x" << std::setw(8) << fp32.hi[3] << ", " <<
481
- "F32 = 0x" << std::setw(8) << fp16_ieee_to_fp32_bits(fp16[7]);
482
- }
483
- }
484
-
485
- TEST(FP16_IEEE_TO_FP32x2_PSIMD, negative_denormalized_values) {
486
- for (uint16_t h = 0; h < 0x0400; h += 8) {
487
- const psimd_u16 fp16 = {
488
- (uint16_t) (h + 0x8000),
489
- (uint16_t) (h + 0x8001),
490
- (uint16_t) (h + 0x8002),
491
- (uint16_t) (h + 0x8003),
492
- (uint16_t) (h + 0x8004),
493
- (uint16_t) (h + 0x8005),
494
- (uint16_t) (h + 0x8006),
495
- (uint16_t) (h + 0x8007)
496
- };
497
- const psimd_u32x2 fp32 =
498
- psimd_cast_f32x2_u32x2(fp16_ieee_to_fp32x2_psimd(fp16));
499
-
500
- EXPECT_EQ(fp16_ieee_to_fp32_bits(fp16[0]), fp32.lo[0]) <<
501
- std::hex << std::uppercase << std::setfill('0') <<
502
- "F16 = 0x" << std::setw(4) << fp16[0] << ", " <<
503
- "F32(F16) = 0x" << std::setw(8) << fp32.lo[0] << ", " <<
504
- "F32 = 0x" << std::setw(8) << fp16_ieee_to_fp32_bits(fp16[0]);
505
- EXPECT_EQ(fp16_ieee_to_fp32_bits(fp16[1]), fp32.lo[1]) <<
506
- std::hex << std::uppercase << std::setfill('0') <<
507
- "F16 = 0x" << std::setw(4) << fp16[1] << ", " <<
508
- "F32(F16) = 0x" << std::setw(8) << fp32.lo[1] << ", " <<
509
- "F32 = 0x" << std::setw(8) << fp16_ieee_to_fp32_bits(fp16[1]);
510
- EXPECT_EQ(fp16_ieee_to_fp32_bits(fp16[2]), fp32.lo[2]) <<
511
- std::hex << std::uppercase << std::setfill('0') <<
512
- "F16 = 0x" << std::setw(4) << fp16[2] << ", " <<
513
- "F32(F16) = 0x" << std::setw(8) << fp32.lo[2] << ", " <<
514
- "F32 = 0x" << std::setw(8) << fp16_ieee_to_fp32_bits(fp16[2]);
515
- EXPECT_EQ(fp16_ieee_to_fp32_bits(fp16[3]), fp32.lo[3]) <<
516
- std::hex << std::uppercase << std::setfill('0') <<
517
- "F16 = 0x" << std::setw(4) << fp16[3] << ", " <<
518
- "F32(F16) = 0x" << std::setw(8) << fp32.lo[3] << ", " <<
519
- "F32 = 0x" << std::setw(8) << fp16_ieee_to_fp32_bits(fp16[3]);
520
- EXPECT_EQ(fp16_ieee_to_fp32_bits(fp16[4]), fp32.hi[0]) <<
521
- std::hex << std::uppercase << std::setfill('0') <<
522
- "F16 = 0x" << std::setw(4) << fp16[4] << ", " <<
523
- "F32(F16) = 0x" << std::setw(8) << fp32.hi[0] << ", " <<
524
- "F32 = 0x" << std::setw(8) << fp16_ieee_to_fp32_bits(fp16[4]);
525
- EXPECT_EQ(fp16_ieee_to_fp32_bits(fp16[5]), fp32.hi[1]) <<
526
- std::hex << std::uppercase << std::setfill('0') <<
527
- "F16 = 0x" << std::setw(4) << fp16[5] << ", " <<
528
- "F32(F16) = 0x" << std::setw(8) << fp32.hi[1] << ", " <<
529
- "F32 = 0x" << std::setw(8) << fp16_ieee_to_fp32_bits(fp16[5]);
530
- EXPECT_EQ(fp16_ieee_to_fp32_bits(fp16[6]), fp32.hi[2]) <<
531
- std::hex << std::uppercase << std::setfill('0') <<
532
- "F16 = 0x" << std::setw(4) << fp16[6] << ", " <<
533
- "F32(F16) = 0x" << std::setw(8) << fp32.hi[2] << ", " <<
534
- "F32 = 0x" << std::setw(8) << fp16_ieee_to_fp32_bits(fp16[6]);
535
- EXPECT_EQ(fp16_ieee_to_fp32_bits(fp16[7]), fp32.hi[3]) <<
536
- std::hex << std::uppercase << std::setfill('0') <<
537
- "F16 = 0x" << std::setw(4) << fp16[7] << ", " <<
538
- "F32(F16) = 0x" << std::setw(8) << fp32.hi[3] << ", " <<
539
- "F32 = 0x" << std::setw(8) << fp16_ieee_to_fp32_bits(fp16[7]);
540
- }
541
- }
package/fp16/test/peachpy/alt-xmm-to-fp32-xmm-avx.cc DELETED
@@ -1,96 +0,0 @@
1
- #include <gtest/gtest.h>
2
-
3
- #include <cstdint>
4
-
5
- #include <fp16.h>
6
-
7
-
8
- extern "C" void fp16_alt_xmm_to_fp32_xmm_peachpy__avx(const uint16_t* fp16, uint32_t* fp32);
9
- const size_t vector_elements = 4;
10
-
11
-
12
- TEST(FP16_ALT_XMM_TO_FP32_XMM, positive_normalized_values) {
13
- const uint32_t exponentBias = 15;
14
- for (int32_t e = -14; e <= 16; e++) {
15
- for (uint16_t h = 0; h < 0x0400; h += vector_elements) {
16
- uint16_t fp16[vector_elements];
17
- for (size_t i = 0; i < vector_elements; i++) {
18
- fp16[i] = h + ((e + exponentBias) << 10) + i;
19
- }
20
- uint32_t fp32[vector_elements];
21
- fp16_alt_xmm_to_fp32_xmm_peachpy__avx(fp16, fp32);
22
-
23
- for (size_t i = 0; i < vector_elements; i++) {
24
- EXPECT_EQ(fp16_alt_to_fp32_bits(fp16[i]), fp32[i]) <<
25
- std::hex << std::uppercase << std::setfill('0') <<
26
- "F16 = 0x" << std::setw(4) << fp16[i] << ", " <<
27
- "F32(F16) = 0x" << std::setw(8) << fp32[i] << ", " <<
28
- "F32 = 0x" << std::setw(8) << fp16_alt_to_fp32_bits(fp16[i]) <<
29
- ", lane " << i << "/" << vector_elements;
30
- }
31
- }
32
- }
33
- }
34
-
35
- TEST(FP16_ALT_XMM_TO_FP32_XMM, negative_normalized_values) {
36
- const uint32_t exponentBias = 15;
37
- for (int32_t e = -14; e <= 16; e++) {
38
- for (uint16_t h = 0; h < 0x0400; h += vector_elements) {
39
- uint16_t fp16[vector_elements];
40
- for (size_t i = 0; i < vector_elements; i++) {
41
- fp16[i] = 0x8000 + h + ((e + exponentBias) << 10) + i;
42
- }
43
- uint32_t fp32[vector_elements];
44
- fp16_alt_xmm_to_fp32_xmm_peachpy__avx(fp16, fp32);
45
-
46
- for (size_t i = 0; i < vector_elements; i++) {
47
- EXPECT_EQ(fp16_alt_to_fp32_bits(fp16[i]), fp32[i]) <<
48
- std::hex << std::uppercase << std::setfill('0') <<
49
- "F16 = 0x" << std::setw(4) << fp16[i] << ", " <<
50
- "F32(F16) = 0x" << std::setw(8) << fp32[i] << ", " <<
51
- "F32 = 0x" << std::setw(8) << fp16_alt_to_fp32_bits(fp16[i]) <<
52
- ", lane " << i << "/" << vector_elements;
53
- }
54
- }
55
- }
56
- }
57
-
58
- TEST(FP16_ALT_XMM_TO_FP32_XMM, positive_denormalized_values) {
59
- for (uint16_t h = 0; h < 0x0400; h += vector_elements) {
60
- uint16_t fp16[vector_elements];
61
- for (size_t i = 0; i < vector_elements; i++) {
62
- fp16[i] = h + i;
63
- }
64
- uint32_t fp32[vector_elements];
65
- fp16_alt_xmm_to_fp32_xmm_peachpy__avx(fp16, fp32);
66
-
67
- for (size_t i = 0; i < vector_elements; i++) {
68
- EXPECT_EQ(fp16_alt_to_fp32_bits(fp16[i]), fp32[i]) <<
69
- std::hex << std::uppercase << std::setfill('0') <<
70
- "F16 = 0x" << std::setw(4) << fp16[i] << ", " <<
71
- "F32(F16) = 0x" << std::setw(8) << fp32[i] << ", " <<
72
- "F32 = 0x" << std::setw(8) << fp16_alt_to_fp32_bits(fp16[i]) <<
73
- ", lane " << i << "/" << vector_elements;
74
- }
75
- }
76
- }
77
-
78
- TEST(FP16_ALT_XMM_TO_FP32_XMM, negative_denormalized_values) {
79
- for (uint16_t h = 0; h < 0x0400; h += vector_elements) {
80
- uint16_t fp16[vector_elements];
81
- for (size_t i = 0; i < vector_elements; i++) {
82
- fp16[i] = 0x8000 + h + i;
83
- }
84
- uint32_t fp32[vector_elements];
85
- fp16_alt_xmm_to_fp32_xmm_peachpy__avx(fp16, fp32);
86
-
87
- for (size_t i = 0; i < vector_elements; i++) {
88
- EXPECT_EQ(fp16_alt_to_fp32_bits(fp16[i]), fp32[i]) <<
89
- std::hex << std::uppercase << std::setfill('0') <<
90
- "F16 = 0x" << std::setw(4) << fp16[i] << ", " <<
91
- "F32(F16) = 0x" << std::setw(8) << fp32[i] << ", " <<
92
- "F32 = 0x" << std::setw(8) << fp16_alt_to_fp32_bits(fp16[i]) <<
93
- ", lane " << i << "/" << vector_elements;
94
- }
95
- }
96
- }