nvmath-python 0.9.0__cp314-cp314-win_amd64.whl

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 (309) hide show
  1. nvmath/__init__.pxd +0 -0
  2. nvmath/__init__.py +45 -0
  3. nvmath/_internal/__init__.py +0 -0
  4. nvmath/_internal/attribute_ifc_factory.py +330 -0
  5. nvmath/_internal/layout.py +70 -0
  6. nvmath/_internal/templates.py +383 -0
  7. nvmath/_utils.py +150 -0
  8. nvmath/bindings/__init__.py +51 -0
  9. nvmath/bindings/_internal/__init__.pxd +0 -0
  10. nvmath/bindings/_internal/__init__.py +0 -0
  11. nvmath/bindings/_internal/cublas.cp314-win_amd64.pyd +0 -0
  12. nvmath/bindings/_internal/cublas.pxd +530 -0
  13. nvmath/bindings/_internal/cublasLt.cp314-win_amd64.pyd +0 -0
  14. nvmath/bindings/_internal/cublasLt.pxd +59 -0
  15. nvmath/bindings/_internal/cublasMp.pxd +52 -0
  16. nvmath/bindings/_internal/cudss.cp314-win_amd64.pyd +0 -0
  17. nvmath/bindings/_internal/cudss.pxd +47 -0
  18. nvmath/bindings/_internal/cufft.cp314-win_amd64.pyd +0 -0
  19. nvmath/bindings/_internal/cufft.pxd +70 -0
  20. nvmath/bindings/_internal/cufftMp.pxd +77 -0
  21. nvmath/bindings/_internal/curand.cp314-win_amd64.pyd +0 -0
  22. nvmath/bindings/_internal/curand.pxd +42 -0
  23. nvmath/bindings/_internal/cusolver.cp314-win_amd64.pyd +0 -0
  24. nvmath/bindings/_internal/cusolver.pxd +15 -0
  25. nvmath/bindings/_internal/cusolverDn.cp314-win_amd64.pyd +0 -0
  26. nvmath/bindings/_internal/cusolverDn.pxd +406 -0
  27. nvmath/bindings/_internal/cusparse.cp314-win_amd64.pyd +0 -0
  28. nvmath/bindings/_internal/cusparse.pxd +277 -0
  29. nvmath/bindings/_internal/cusparseLt.cp314-win_amd64.pyd +0 -0
  30. nvmath/bindings/_internal/cusparseLt.pxd +48 -0
  31. nvmath/bindings/_internal/cutensor.cp314-win_amd64.pyd +0 -0
  32. nvmath/bindings/_internal/cutensor.pxd +58 -0
  33. nvmath/bindings/_internal/mathdx.cp314-win_amd64.pyd +0 -0
  34. nvmath/bindings/_internal/mathdx.pxd +116 -0
  35. nvmath/bindings/_internal/nvshmem.pxd +29 -0
  36. nvmath/bindings/_internal/utils.cp314-win_amd64.pyd +0 -0
  37. nvmath/bindings/_internal/utils.pxd +179 -0
  38. nvmath/bindings/_internal/utils.pyi +10 -0
  39. nvmath/bindings/cublas.cp314-win_amd64.pyd +0 -0
  40. nvmath/bindings/cublas.pxd +558 -0
  41. nvmath/bindings/cublas.pyi +746 -0
  42. nvmath/bindings/cublasLt.cp314-win_amd64.pyd +0 -0
  43. nvmath/bindings/cublasLt.pxd +109 -0
  44. nvmath/bindings/cublasLt.pyi +1337 -0
  45. nvmath/bindings/cublasMp.pxd +85 -0
  46. nvmath/bindings/cublasMp.pyi +219 -0
  47. nvmath/bindings/cudss.cp314-win_amd64.pyd +0 -0
  48. nvmath/bindings/cudss.pxd +86 -0
  49. nvmath/bindings/cudss.pyi +282 -0
  50. nvmath/bindings/cufft.cp314-win_amd64.pyd +0 -0
  51. nvmath/bindings/cufft.pxd +118 -0
  52. nvmath/bindings/cufft.pyi +241 -0
  53. nvmath/bindings/cufftMp.pxd +124 -0
  54. nvmath/bindings/cufftMp.pyi +260 -0
  55. nvmath/bindings/curand.cp314-win_amd64.pyd +0 -0
  56. nvmath/bindings/curand.pxd +71 -0
  57. nvmath/bindings/curand.pyi +159 -0
  58. nvmath/bindings/cusolver.cp314-win_amd64.pyd +0 -0
  59. nvmath/bindings/cusolver.pxd +62 -0
  60. nvmath/bindings/cusolver.pyi +242 -0
  61. nvmath/bindings/cusolverDn.cp314-win_amd64.pyd +0 -0
  62. nvmath/bindings/cusolverDn.pxd +430 -0
  63. nvmath/bindings/cusolverDn.pyi +416 -0
  64. nvmath/bindings/cusparse.cp314-win_amd64.pyd +0 -0
  65. nvmath/bindings/cusparse.pxd +338 -0
  66. nvmath/bindings/cusparse.pyi +659 -0
  67. nvmath/bindings/cusparseLt.cp314-win_amd64.pyd +0 -0
  68. nvmath/bindings/cusparseLt.pxd +99 -0
  69. nvmath/bindings/cusparseLt.pyi +198 -0
  70. nvmath/bindings/cutensor.cp314-win_amd64.pyd +0 -0
  71. nvmath/bindings/cutensor.pxd +98 -0
  72. nvmath/bindings/cutensor.pyi +294 -0
  73. nvmath/bindings/cycublas.cp314-win_amd64.pyd +0 -0
  74. nvmath/bindings/cycublas.pxd +684 -0
  75. nvmath/bindings/cycublasLt.cp314-win_amd64.pyd +0 -0
  76. nvmath/bindings/cycublasLt.pxd +1055 -0
  77. nvmath/bindings/cycublasMp.pxd +183 -0
  78. nvmath/bindings/cycudss.cp314-win_amd64.pyd +0 -0
  79. nvmath/bindings/cycudss.pxd +224 -0
  80. nvmath/bindings/cycufft.cp314-win_amd64.pyd +0 -0
  81. nvmath/bindings/cycufft.pxd +326 -0
  82. nvmath/bindings/cycufftMp.pxd +334 -0
  83. nvmath/bindings/cycurand.cp314-win_amd64.pyd +0 -0
  84. nvmath/bindings/cycurand.pxd +146 -0
  85. nvmath/bindings/cycusolver.cp314-win_amd64.pyd +0 -0
  86. nvmath/bindings/cycusolver.pxd +154 -0
  87. nvmath/bindings/cycusolverDn.cp314-win_amd64.pyd +0 -0
  88. nvmath/bindings/cycusolverDn.pxd +446 -0
  89. nvmath/bindings/cycusparse.cp314-win_amd64.pyd +0 -0
  90. nvmath/bindings/cycusparse.pxd +470 -0
  91. nvmath/bindings/cycusparseLt.cp314-win_amd64.pyd +0 -0
  92. nvmath/bindings/cycusparseLt.pxd +150 -0
  93. nvmath/bindings/cycutensor.cp314-win_amd64.pyd +0 -0
  94. nvmath/bindings/cycutensor.pxd +192 -0
  95. nvmath/bindings/cymathdx.cp314-win_amd64.pyd +0 -0
  96. nvmath/bindings/cymathdx.pxd +552 -0
  97. nvmath/bindings/cynvshmem.pxd +126 -0
  98. nvmath/bindings/mathdx.cp314-win_amd64.pyd +0 -0
  99. nvmath/bindings/mathdx.pxd +182 -0
  100. nvmath/bindings/mathdx.pyi +1244 -0
  101. nvmath/bindings/nvpl/__init__.pxd +0 -0
  102. nvmath/bindings/nvpl/__init__.py +13 -0
  103. nvmath/bindings/nvpl/_internal/__init__.pxd +0 -0
  104. nvmath/bindings/nvpl/_internal/__init__.py +0 -0
  105. nvmath/bindings/nvpl/_internal/blas.cp314-win_amd64.pyd +0 -0
  106. nvmath/bindings/nvpl/_internal/blas.pxd +125 -0
  107. nvmath/bindings/nvpl/_internal/fft.cp314-win_amd64.pyd +0 -0
  108. nvmath/bindings/nvpl/_internal/fft.pxd +36 -0
  109. nvmath/bindings/nvpl/blas.cp314-win_amd64.pyd +0 -0
  110. nvmath/bindings/nvpl/blas.pxd +85 -0
  111. nvmath/bindings/nvpl/blas.pyi +122 -0
  112. nvmath/bindings/nvpl/cyblas.cp314-win_amd64.pyd +0 -0
  113. nvmath/bindings/nvpl/cyblas.pxd +166 -0
  114. nvmath/bindings/nvpl/cyfft.cp314-win_amd64.pyd +0 -0
  115. nvmath/bindings/nvpl/cyfft.pxd +92 -0
  116. nvmath/bindings/nvpl/fft.cp314-win_amd64.pyd +0 -0
  117. nvmath/bindings/nvpl/fft.pxd +100 -0
  118. nvmath/bindings/nvpl/fft.pyi +100 -0
  119. nvmath/bindings/nvshmem.pxd +54 -0
  120. nvmath/bindings/nvshmem.pyi +179 -0
  121. nvmath/device/__init__.py +21 -0
  122. nvmath/device/_deprecated.py +33 -0
  123. nvmath/device/common.py +313 -0
  124. nvmath/device/common_backend.py +131 -0
  125. nvmath/device/common_cuda.py +201 -0
  126. nvmath/device/common_numba.py +310 -0
  127. nvmath/device/common_opaque_tensor.py +201 -0
  128. nvmath/device/cublasdx.py +1805 -0
  129. nvmath/device/cublasdx_backend.py +807 -0
  130. nvmath/device/cublasdx_numba.py +1612 -0
  131. nvmath/device/cufftdx.py +510 -0
  132. nvmath/device/cufftdx_backend.py +196 -0
  133. nvmath/device/cufftdx_numba.py +196 -0
  134. nvmath/device/curand_kernel.py +9147 -0
  135. nvmath/device/cusolverdx.py +2690 -0
  136. nvmath/device/cusolverdx_backend.py +440 -0
  137. nvmath/device/cusolverdx_numba.py +1624 -0
  138. nvmath/device/llvm_array.py +29 -0
  139. nvmath/device/random.py +445 -0
  140. nvmath/device/random_helpers.py +23 -0
  141. nvmath/device/random_states.py +187 -0
  142. nvmath/device/types.py +147 -0
  143. nvmath/device/vector_types_numba.py +203 -0
  144. nvmath/distributed/__init__.py +205 -0
  145. nvmath/distributed/_internal/__init__.py +0 -0
  146. nvmath/distributed/_internal/nvshmem.py +302 -0
  147. nvmath/distributed/_internal/tensor_ifc.py +67 -0
  148. nvmath/distributed/_internal/tensor_ifc_cupy.py +62 -0
  149. nvmath/distributed/_internal/tensor_ifc_host_device.py +165 -0
  150. nvmath/distributed/_internal/tensor_ifc_numpy.py +41 -0
  151. nvmath/distributed/_internal/tensor_ifc_torch.py +141 -0
  152. nvmath/distributed/_internal/tensor_wrapper.py +78 -0
  153. nvmath/distributed/_utils.py +167 -0
  154. nvmath/distributed/distribution.py +770 -0
  155. nvmath/distributed/fft/__init__.py +7 -0
  156. nvmath/distributed/fft/_configuration.py +79 -0
  157. nvmath/distributed/fft/fft.py +2801 -0
  158. nvmath/distributed/linalg/__init__.py +12 -0
  159. nvmath/distributed/linalg/_internal/__init__.py +3 -0
  160. nvmath/distributed/linalg/_internal/epilog_protocol.py +496 -0
  161. nvmath/distributed/linalg/_internal/matmul_desc_ifc.py +28 -0
  162. nvmath/distributed/linalg/advanced/__init__.py +8 -0
  163. nvmath/distributed/linalg/advanced/_configuration.py +166 -0
  164. nvmath/distributed/linalg/advanced/matmulmod.py +2908 -0
  165. nvmath/distributed/process_group.py +408 -0
  166. nvmath/distributed/reshape/__init__.py +6 -0
  167. nvmath/distributed/reshape/_configuration.py +39 -0
  168. nvmath/distributed/reshape/reshape.py +1256 -0
  169. nvmath/fft/__init__.py +7 -0
  170. nvmath/fft/_configuration.py +208 -0
  171. nvmath/fft/_exec_utils.py +109 -0
  172. nvmath/fft/_helpers.py +237 -0
  173. nvmath/fft/fft.py +2903 -0
  174. nvmath/internal/__init__.pxd +3 -0
  175. nvmath/internal/__init__.py +10 -0
  176. nvmath/internal/_bindings.cp314-win_amd64.pyd +0 -0
  177. nvmath/internal/_bindings.pxd +12 -0
  178. nvmath/internal/_device_utils.py +54 -0
  179. nvmath/internal/enum_utils.py +142 -0
  180. nvmath/internal/formatters.py +87 -0
  181. nvmath/internal/mem_limit.py +51 -0
  182. nvmath/internal/memory.cp314-win_amd64.pyd +0 -0
  183. nvmath/internal/memory.pxd +10 -0
  184. nvmath/internal/memory.pyi +46 -0
  185. nvmath/internal/ndbuffer/__init__.pxd +3 -0
  186. nvmath/internal/ndbuffer/__init__.py +7 -0
  187. nvmath/internal/ndbuffer/copy_kernel/args.h +34 -0
  188. nvmath/internal/ndbuffer/copy_kernel/copy_kernel_impl/array_view.h +52 -0
  189. nvmath/internal/ndbuffer/copy_kernel/copy_kernel_impl/elementwise.h +68 -0
  190. nvmath/internal/ndbuffer/copy_kernel/copy_kernel_impl/grid_indexer.h +69 -0
  191. nvmath/internal/ndbuffer/copy_kernel/copy_kernel_impl/transposed.h +240 -0
  192. nvmath/internal/ndbuffer/copy_kernel/copy_kernel_impl/type_utils.h +39 -0
  193. nvmath/internal/ndbuffer/copy_kernel/copy_kernel_impl/utils.h +132 -0
  194. nvmath/internal/ndbuffer/copy_kernel/copy_kernel_impl/vec.h +159 -0
  195. nvmath/internal/ndbuffer/copy_kernel/elementwise.h +53 -0
  196. nvmath/internal/ndbuffer/copy_kernel/transposed.h +58 -0
  197. nvmath/internal/ndbuffer/copy_kernel.cp314-win_amd64.pyd +0 -0
  198. nvmath/internal/ndbuffer/copy_kernel.pxd +9 -0
  199. nvmath/internal/ndbuffer/copy_kernel.pyi +9 -0
  200. nvmath/internal/ndbuffer/data_layout.cp314-win_amd64.pyd +0 -0
  201. nvmath/internal/ndbuffer/data_layout.pxd +72 -0
  202. nvmath/internal/ndbuffer/data_layout.pyi +23 -0
  203. nvmath/internal/ndbuffer/jit.cp314-win_amd64.pyd +0 -0
  204. nvmath/internal/ndbuffer/jit.pxd +7 -0
  205. nvmath/internal/ndbuffer/jit.pyi +6 -0
  206. nvmath/internal/ndbuffer/ndbuffer.cp314-win_amd64.pyd +0 -0
  207. nvmath/internal/ndbuffer/ndbuffer.pxd +36 -0
  208. nvmath/internal/ndbuffer/ndbuffer.pyi +41 -0
  209. nvmath/internal/ndbuffer/package_utils.cp314-win_amd64.pyd +0 -0
  210. nvmath/internal/ndbuffer/package_utils.pxd +11 -0
  211. nvmath/internal/ndbuffer/package_utils.pyi +12 -0
  212. nvmath/internal/package_ifc.py +147 -0
  213. nvmath/internal/package_ifc_cuda.py +49 -0
  214. nvmath/internal/package_ifc_cupy.py +49 -0
  215. nvmath/internal/package_ifc_torch.py +31 -0
  216. nvmath/internal/package_wrapper.py +14 -0
  217. nvmath/internal/tensor_ifc.py +194 -0
  218. nvmath/internal/tensor_ifc_cupy.py +241 -0
  219. nvmath/internal/tensor_ifc_ndbuffer.py +153 -0
  220. nvmath/internal/tensor_ifc_numpy.py +183 -0
  221. nvmath/internal/tensor_ifc_torch.py +176 -0
  222. nvmath/internal/tensor_wrapper.py +157 -0
  223. nvmath/internal/typemaps.py +162 -0
  224. nvmath/internal/utils.py +771 -0
  225. nvmath/linalg/__init__.py +46 -0
  226. nvmath/linalg/_internal/__init__.py +3 -0
  227. nvmath/linalg/_internal/algo_cap_ifc.py +82 -0
  228. nvmath/linalg/_internal/algo_config_ifc.py +43 -0
  229. nvmath/linalg/_internal/batch.py +217 -0
  230. nvmath/linalg/_internal/enum_to_tuples.py +64 -0
  231. nvmath/linalg/_internal/epilog_protocol.py +766 -0
  232. nvmath/linalg/_internal/layout.py +624 -0
  233. nvmath/linalg/_internal/matmul_desc_ifc.py +28 -0
  234. nvmath/linalg/_internal/matmul_pref_ifc.py +27 -0
  235. nvmath/linalg/_internal/matrix_layout_ifc.py +26 -0
  236. nvmath/linalg/_internal/typemaps.py +144 -0
  237. nvmath/linalg/_internal/utils.py +134 -0
  238. nvmath/linalg/advanced/__init__.py +8 -0
  239. nvmath/linalg/advanced/_algorithmmod.py +149 -0
  240. nvmath/linalg/advanced/_configuration.py +349 -0
  241. nvmath/linalg/advanced/helpers/__init__.py +5 -0
  242. nvmath/linalg/advanced/helpers/matmul.py +1348 -0
  243. nvmath/linalg/advanced/matmulmod.py +3937 -0
  244. nvmath/linalg/generic/__init__.py +43 -0
  245. nvmath/linalg/generic/_configuration/__init__.py +37 -0
  246. nvmath/linalg/generic/_configuration/layout.py +263 -0
  247. nvmath/linalg/generic/_configuration/match.py +609 -0
  248. nvmath/linalg/generic/_configuration/qualifiers.py +494 -0
  249. nvmath/linalg/generic/_configuration/wrap.py +217 -0
  250. nvmath/linalg/generic/_dtype.py +15 -0
  251. nvmath/linalg/generic/matmulmod.py +1060 -0
  252. nvmath/memory.py +282 -0
  253. nvmath/sparse/__init__.py +38 -0
  254. nvmath/sparse/_internal/__init__.py +21 -0
  255. nvmath/sparse/_internal/common_utils.py +144 -0
  256. nvmath/sparse/_internal/cudss_config_ifc.py +741 -0
  257. nvmath/sparse/_internal/cudss_data_ifc.py +399 -0
  258. nvmath/sparse/_internal/cudss_utils.py +445 -0
  259. nvmath/sparse/_internal/cusparse_utils.py +382 -0
  260. nvmath/sparse/_internal/sparse_bsc_ifc.py +300 -0
  261. nvmath/sparse/_internal/sparse_bsr_ifc.py +303 -0
  262. nvmath/sparse/_internal/sparse_coo_ifc.py +254 -0
  263. nvmath/sparse/_internal/sparse_csc_ifc.py +266 -0
  264. nvmath/sparse/_internal/sparse_csr_ifc.py +266 -0
  265. nvmath/sparse/_internal/sparse_dia_ifc.py +239 -0
  266. nvmath/sparse/_internal/sparse_format_helpers.py +601 -0
  267. nvmath/sparse/_internal/sparse_tensor_ifc.py +121 -0
  268. nvmath/sparse/_internal/sparse_ust_ifc.py +146 -0
  269. nvmath/sparse/_internal/utils.py +56 -0
  270. nvmath/sparse/advanced/__init__.py +7 -0
  271. nvmath/sparse/advanced/_configuration.py +230 -0
  272. nvmath/sparse/advanced/direct_solver.py +1884 -0
  273. nvmath/sparse/generic/__init__.py +7 -0
  274. nvmath/sparse/generic/_configuration.py +129 -0
  275. nvmath/sparse/generic/_helpers.py +137 -0
  276. nvmath/sparse/generic/_thunks.py +21 -0
  277. nvmath/sparse/generic/matmulmod.py +2426 -0
  278. nvmath/sparse/ust/__init__.py +7 -0
  279. nvmath/sparse/ust/_converters.py +378 -0
  280. nvmath/sparse/ust/_drawer.py +552 -0
  281. nvmath/sparse/ust/_emitter.py +553 -0
  282. nvmath/sparse/ust/_jit.py +212 -0
  283. nvmath/sparse/ust/_kernel.py +346 -0
  284. nvmath/sparse/ust/_semiring.py +71 -0
  285. nvmath/sparse/ust/_utils.py +102 -0
  286. nvmath/sparse/ust/interfaces/__init__.py +0 -0
  287. nvmath/sparse/ust/interfaces/torch_interface.py +476 -0
  288. nvmath/sparse/ust/semiring_ops.py +71 -0
  289. nvmath/sparse/ust/tensor.py +1027 -0
  290. nvmath/sparse/ust/tensor_drawer.py +552 -0
  291. nvmath/sparse/ust/tensor_emitter.py +558 -0
  292. nvmath/sparse/ust/tensor_format.py +935 -0
  293. nvmath/sparse/ust/tensor_jit.py +212 -0
  294. nvmath/sparse/ust/tensor_kernel.py +348 -0
  295. nvmath/sparse/ust/tensor_utils.py +380 -0
  296. nvmath/tensor/__init__.py +6 -0
  297. nvmath/tensor/_configuration.py +95 -0
  298. nvmath/tensor/_internal/__init__.py +3 -0
  299. nvmath/tensor/_internal/cutensor_config_ifc.py +247 -0
  300. nvmath/tensor/_internal/cutensor_utils.py +162 -0
  301. nvmath/tensor/_internal/data.py +43 -0
  302. nvmath/tensor/_internal/einsum_parser.py +444 -0
  303. nvmath/tensor/_internal/typemaps.py +96 -0
  304. nvmath/tensor/contract.py +1861 -0
  305. nvmath_python-0.9.0.dist-info/METADATA +127 -0
  306. nvmath_python-0.9.0.dist-info/RECORD +309 -0
  307. nvmath_python-0.9.0.dist-info/WHEEL +5 -0
  308. nvmath_python-0.9.0.dist-info/licenses/LICENSE +177 -0
  309. nvmath_python-0.9.0.dist-info/top_level.txt +2 -0
@@ -0,0 +1,1055 @@
1
+ # Copyright (c) 2024-2026, NVIDIA CORPORATION & AFFILIATES. ALL RIGHTS RESERVED.
2
+ #
3
+ # SPDX-License-Identifier: Apache-2.0
4
+ #
5
+ # This code was automatically generated across versions from 11.0.3 to 13.2.0, generator version 0.3.1.dev1301+g7215ac36e. Do not modify it directly.
6
+ # This layer exposes the C header to Cython as-is.
7
+
8
+ from libc.stdint cimport int64_t, uint64_t
9
+ from libc.stdio cimport FILE
10
+
11
+ from .cycublas cimport cublasStatus_t, cublasComputeType_t, CUBLAS_POINTER_MODE_HOST, CUBLAS_POINTER_MODE_DEVICE
12
+ from .cycublas cimport _CUBLASSTATUS_T_INTERNAL_LOADING_ERROR
13
+
14
+
15
+ ###############################################################################
16
+ # Types (structs, enums, ...)
17
+ ###############################################################################
18
+
19
+ # enums
20
+ ctypedef enum cublasLtMatmulTile_t "cublasLtMatmulTile_t":
21
+ CUBLASLT_MATMUL_TILE_UNDEFINED "CUBLASLT_MATMUL_TILE_UNDEFINED" = 0
22
+ CUBLASLT_MATMUL_TILE_8x8 "CUBLASLT_MATMUL_TILE_8x8" = 1
23
+ CUBLASLT_MATMUL_TILE_8x16 "CUBLASLT_MATMUL_TILE_8x16" = 2
24
+ CUBLASLT_MATMUL_TILE_16x8 "CUBLASLT_MATMUL_TILE_16x8" = 3
25
+ CUBLASLT_MATMUL_TILE_8x32 "CUBLASLT_MATMUL_TILE_8x32" = 4
26
+ CUBLASLT_MATMUL_TILE_16x16 "CUBLASLT_MATMUL_TILE_16x16" = 5
27
+ CUBLASLT_MATMUL_TILE_32x8 "CUBLASLT_MATMUL_TILE_32x8" = 6
28
+ CUBLASLT_MATMUL_TILE_8x64 "CUBLASLT_MATMUL_TILE_8x64" = 7
29
+ CUBLASLT_MATMUL_TILE_16x32 "CUBLASLT_MATMUL_TILE_16x32" = 8
30
+ CUBLASLT_MATMUL_TILE_32x16 "CUBLASLT_MATMUL_TILE_32x16" = 9
31
+ CUBLASLT_MATMUL_TILE_64x8 "CUBLASLT_MATMUL_TILE_64x8" = 10
32
+ CUBLASLT_MATMUL_TILE_32x32 "CUBLASLT_MATMUL_TILE_32x32" = 11
33
+ CUBLASLT_MATMUL_TILE_32x64 "CUBLASLT_MATMUL_TILE_32x64" = 12
34
+ CUBLASLT_MATMUL_TILE_64x32 "CUBLASLT_MATMUL_TILE_64x32" = 13
35
+ CUBLASLT_MATMUL_TILE_32x128 "CUBLASLT_MATMUL_TILE_32x128" = 14
36
+ CUBLASLT_MATMUL_TILE_64x64 "CUBLASLT_MATMUL_TILE_64x64" = 15
37
+ CUBLASLT_MATMUL_TILE_128x32 "CUBLASLT_MATMUL_TILE_128x32" = 16
38
+ CUBLASLT_MATMUL_TILE_64x128 "CUBLASLT_MATMUL_TILE_64x128" = 17
39
+ CUBLASLT_MATMUL_TILE_128x64 "CUBLASLT_MATMUL_TILE_128x64" = 18
40
+ CUBLASLT_MATMUL_TILE_64x256 "CUBLASLT_MATMUL_TILE_64x256" = 19
41
+ CUBLASLT_MATMUL_TILE_128x128 "CUBLASLT_MATMUL_TILE_128x128" = 20
42
+ CUBLASLT_MATMUL_TILE_256x64 "CUBLASLT_MATMUL_TILE_256x64" = 21
43
+ CUBLASLT_MATMUL_TILE_64x512 "CUBLASLT_MATMUL_TILE_64x512" = 22
44
+ CUBLASLT_MATMUL_TILE_128x256 "CUBLASLT_MATMUL_TILE_128x256" = 23
45
+ CUBLASLT_MATMUL_TILE_256x128 "CUBLASLT_MATMUL_TILE_256x128" = 24
46
+ CUBLASLT_MATMUL_TILE_512x64 "CUBLASLT_MATMUL_TILE_512x64" = 25
47
+ CUBLASLT_MATMUL_TILE_64x96 "CUBLASLT_MATMUL_TILE_64x96" = 26
48
+ CUBLASLT_MATMUL_TILE_96x64 "CUBLASLT_MATMUL_TILE_96x64" = 27
49
+ CUBLASLT_MATMUL_TILE_96x128 "CUBLASLT_MATMUL_TILE_96x128" = 28
50
+ CUBLASLT_MATMUL_TILE_128x160 "CUBLASLT_MATMUL_TILE_128x160" = 29
51
+ CUBLASLT_MATMUL_TILE_160x128 "CUBLASLT_MATMUL_TILE_160x128" = 30
52
+ CUBLASLT_MATMUL_TILE_192x128 "CUBLASLT_MATMUL_TILE_192x128" = 31
53
+ CUBLASLT_MATMUL_TILE_128x192 "CUBLASLT_MATMUL_TILE_128x192" = 32
54
+ CUBLASLT_MATMUL_TILE_128x96 "CUBLASLT_MATMUL_TILE_128x96" = 33
55
+ CUBLASLT_MATMUL_TILE_32x256 "CUBLASLT_MATMUL_TILE_32x256" = 34
56
+ CUBLASLT_MATMUL_TILE_256x32 "CUBLASLT_MATMUL_TILE_256x32" = 35
57
+ CUBLASLT_MATMUL_TILE_8x128 "CUBLASLT_MATMUL_TILE_8x128" = 36
58
+ CUBLASLT_MATMUL_TILE_8x192 "CUBLASLT_MATMUL_TILE_8x192" = 37
59
+ CUBLASLT_MATMUL_TILE_8x256 "CUBLASLT_MATMUL_TILE_8x256" = 38
60
+ CUBLASLT_MATMUL_TILE_8x320 "CUBLASLT_MATMUL_TILE_8x320" = 39
61
+ CUBLASLT_MATMUL_TILE_8x384 "CUBLASLT_MATMUL_TILE_8x384" = 40
62
+ CUBLASLT_MATMUL_TILE_8x448 "CUBLASLT_MATMUL_TILE_8x448" = 41
63
+ CUBLASLT_MATMUL_TILE_8x512 "CUBLASLT_MATMUL_TILE_8x512" = 42
64
+ CUBLASLT_MATMUL_TILE_8x576 "CUBLASLT_MATMUL_TILE_8x576" = 43
65
+ CUBLASLT_MATMUL_TILE_8x640 "CUBLASLT_MATMUL_TILE_8x640" = 44
66
+ CUBLASLT_MATMUL_TILE_8x704 "CUBLASLT_MATMUL_TILE_8x704" = 45
67
+ CUBLASLT_MATMUL_TILE_8x768 "CUBLASLT_MATMUL_TILE_8x768" = 46
68
+ CUBLASLT_MATMUL_TILE_16x64 "CUBLASLT_MATMUL_TILE_16x64" = 47
69
+ CUBLASLT_MATMUL_TILE_16x128 "CUBLASLT_MATMUL_TILE_16x128" = 48
70
+ CUBLASLT_MATMUL_TILE_16x192 "CUBLASLT_MATMUL_TILE_16x192" = 49
71
+ CUBLASLT_MATMUL_TILE_16x256 "CUBLASLT_MATMUL_TILE_16x256" = 50
72
+ CUBLASLT_MATMUL_TILE_16x320 "CUBLASLT_MATMUL_TILE_16x320" = 51
73
+ CUBLASLT_MATMUL_TILE_16x384 "CUBLASLT_MATMUL_TILE_16x384" = 52
74
+ CUBLASLT_MATMUL_TILE_16x448 "CUBLASLT_MATMUL_TILE_16x448" = 53
75
+ CUBLASLT_MATMUL_TILE_16x512 "CUBLASLT_MATMUL_TILE_16x512" = 54
76
+ CUBLASLT_MATMUL_TILE_16x576 "CUBLASLT_MATMUL_TILE_16x576" = 55
77
+ CUBLASLT_MATMUL_TILE_16x640 "CUBLASLT_MATMUL_TILE_16x640" = 56
78
+ CUBLASLT_MATMUL_TILE_16x704 "CUBLASLT_MATMUL_TILE_16x704" = 57
79
+ CUBLASLT_MATMUL_TILE_16x768 "CUBLASLT_MATMUL_TILE_16x768" = 58
80
+ CUBLASLT_MATMUL_TILE_24x64 "CUBLASLT_MATMUL_TILE_24x64" = 59
81
+ CUBLASLT_MATMUL_TILE_24x128 "CUBLASLT_MATMUL_TILE_24x128" = 60
82
+ CUBLASLT_MATMUL_TILE_24x192 "CUBLASLT_MATMUL_TILE_24x192" = 61
83
+ CUBLASLT_MATMUL_TILE_24x256 "CUBLASLT_MATMUL_TILE_24x256" = 62
84
+ CUBLASLT_MATMUL_TILE_24x320 "CUBLASLT_MATMUL_TILE_24x320" = 63
85
+ CUBLASLT_MATMUL_TILE_24x384 "CUBLASLT_MATMUL_TILE_24x384" = 64
86
+ CUBLASLT_MATMUL_TILE_24x448 "CUBLASLT_MATMUL_TILE_24x448" = 65
87
+ CUBLASLT_MATMUL_TILE_24x512 "CUBLASLT_MATMUL_TILE_24x512" = 66
88
+ CUBLASLT_MATMUL_TILE_24x576 "CUBLASLT_MATMUL_TILE_24x576" = 67
89
+ CUBLASLT_MATMUL_TILE_24x640 "CUBLASLT_MATMUL_TILE_24x640" = 68
90
+ CUBLASLT_MATMUL_TILE_24x704 "CUBLASLT_MATMUL_TILE_24x704" = 69
91
+ CUBLASLT_MATMUL_TILE_24x768 "CUBLASLT_MATMUL_TILE_24x768" = 70
92
+ CUBLASLT_MATMUL_TILE_32x192 "CUBLASLT_MATMUL_TILE_32x192" = 71
93
+ CUBLASLT_MATMUL_TILE_32x320 "CUBLASLT_MATMUL_TILE_32x320" = 72
94
+ CUBLASLT_MATMUL_TILE_32x384 "CUBLASLT_MATMUL_TILE_32x384" = 73
95
+ CUBLASLT_MATMUL_TILE_32x448 "CUBLASLT_MATMUL_TILE_32x448" = 74
96
+ CUBLASLT_MATMUL_TILE_32x512 "CUBLASLT_MATMUL_TILE_32x512" = 75
97
+ CUBLASLT_MATMUL_TILE_32x576 "CUBLASLT_MATMUL_TILE_32x576" = 76
98
+ CUBLASLT_MATMUL_TILE_32x640 "CUBLASLT_MATMUL_TILE_32x640" = 77
99
+ CUBLASLT_MATMUL_TILE_32x704 "CUBLASLT_MATMUL_TILE_32x704" = 78
100
+ CUBLASLT_MATMUL_TILE_32x768 "CUBLASLT_MATMUL_TILE_32x768" = 79
101
+ CUBLASLT_MATMUL_TILE_40x64 "CUBLASLT_MATMUL_TILE_40x64" = 80
102
+ CUBLASLT_MATMUL_TILE_40x128 "CUBLASLT_MATMUL_TILE_40x128" = 81
103
+ CUBLASLT_MATMUL_TILE_40x192 "CUBLASLT_MATMUL_TILE_40x192" = 82
104
+ CUBLASLT_MATMUL_TILE_40x256 "CUBLASLT_MATMUL_TILE_40x256" = 83
105
+ CUBLASLT_MATMUL_TILE_40x320 "CUBLASLT_MATMUL_TILE_40x320" = 84
106
+ CUBLASLT_MATMUL_TILE_40x384 "CUBLASLT_MATMUL_TILE_40x384" = 85
107
+ CUBLASLT_MATMUL_TILE_40x448 "CUBLASLT_MATMUL_TILE_40x448" = 86
108
+ CUBLASLT_MATMUL_TILE_40x512 "CUBLASLT_MATMUL_TILE_40x512" = 87
109
+ CUBLASLT_MATMUL_TILE_40x576 "CUBLASLT_MATMUL_TILE_40x576" = 88
110
+ CUBLASLT_MATMUL_TILE_40x640 "CUBLASLT_MATMUL_TILE_40x640" = 89
111
+ CUBLASLT_MATMUL_TILE_40x704 "CUBLASLT_MATMUL_TILE_40x704" = 90
112
+ CUBLASLT_MATMUL_TILE_40x768 "CUBLASLT_MATMUL_TILE_40x768" = 91
113
+ CUBLASLT_MATMUL_TILE_48x64 "CUBLASLT_MATMUL_TILE_48x64" = 92
114
+ CUBLASLT_MATMUL_TILE_48x128 "CUBLASLT_MATMUL_TILE_48x128" = 93
115
+ CUBLASLT_MATMUL_TILE_48x192 "CUBLASLT_MATMUL_TILE_48x192" = 94
116
+ CUBLASLT_MATMUL_TILE_48x256 "CUBLASLT_MATMUL_TILE_48x256" = 95
117
+ CUBLASLT_MATMUL_TILE_48x320 "CUBLASLT_MATMUL_TILE_48x320" = 96
118
+ CUBLASLT_MATMUL_TILE_48x384 "CUBLASLT_MATMUL_TILE_48x384" = 97
119
+ CUBLASLT_MATMUL_TILE_48x448 "CUBLASLT_MATMUL_TILE_48x448" = 98
120
+ CUBLASLT_MATMUL_TILE_48x512 "CUBLASLT_MATMUL_TILE_48x512" = 99
121
+ CUBLASLT_MATMUL_TILE_48x576 "CUBLASLT_MATMUL_TILE_48x576" = 100
122
+ CUBLASLT_MATMUL_TILE_48x640 "CUBLASLT_MATMUL_TILE_48x640" = 101
123
+ CUBLASLT_MATMUL_TILE_48x704 "CUBLASLT_MATMUL_TILE_48x704" = 102
124
+ CUBLASLT_MATMUL_TILE_48x768 "CUBLASLT_MATMUL_TILE_48x768" = 103
125
+ CUBLASLT_MATMUL_TILE_56x64 "CUBLASLT_MATMUL_TILE_56x64" = 104
126
+ CUBLASLT_MATMUL_TILE_56x128 "CUBLASLT_MATMUL_TILE_56x128" = 105
127
+ CUBLASLT_MATMUL_TILE_56x192 "CUBLASLT_MATMUL_TILE_56x192" = 106
128
+ CUBLASLT_MATMUL_TILE_56x256 "CUBLASLT_MATMUL_TILE_56x256" = 107
129
+ CUBLASLT_MATMUL_TILE_56x320 "CUBLASLT_MATMUL_TILE_56x320" = 108
130
+ CUBLASLT_MATMUL_TILE_56x384 "CUBLASLT_MATMUL_TILE_56x384" = 109
131
+ CUBLASLT_MATMUL_TILE_56x448 "CUBLASLT_MATMUL_TILE_56x448" = 110
132
+ CUBLASLT_MATMUL_TILE_56x512 "CUBLASLT_MATMUL_TILE_56x512" = 111
133
+ CUBLASLT_MATMUL_TILE_56x576 "CUBLASLT_MATMUL_TILE_56x576" = 112
134
+ CUBLASLT_MATMUL_TILE_56x640 "CUBLASLT_MATMUL_TILE_56x640" = 113
135
+ CUBLASLT_MATMUL_TILE_56x704 "CUBLASLT_MATMUL_TILE_56x704" = 114
136
+ CUBLASLT_MATMUL_TILE_56x768 "CUBLASLT_MATMUL_TILE_56x768" = 115
137
+ CUBLASLT_MATMUL_TILE_64x192 "CUBLASLT_MATMUL_TILE_64x192" = 116
138
+ CUBLASLT_MATMUL_TILE_64x320 "CUBLASLT_MATMUL_TILE_64x320" = 117
139
+ CUBLASLT_MATMUL_TILE_64x384 "CUBLASLT_MATMUL_TILE_64x384" = 118
140
+ CUBLASLT_MATMUL_TILE_64x448 "CUBLASLT_MATMUL_TILE_64x448" = 119
141
+ CUBLASLT_MATMUL_TILE_64x576 "CUBLASLT_MATMUL_TILE_64x576" = 120
142
+ CUBLASLT_MATMUL_TILE_64x640 "CUBLASLT_MATMUL_TILE_64x640" = 121
143
+ CUBLASLT_MATMUL_TILE_64x704 "CUBLASLT_MATMUL_TILE_64x704" = 122
144
+ CUBLASLT_MATMUL_TILE_64x768 "CUBLASLT_MATMUL_TILE_64x768" = 123
145
+ CUBLASLT_MATMUL_TILE_72x64 "CUBLASLT_MATMUL_TILE_72x64" = 124
146
+ CUBLASLT_MATMUL_TILE_72x128 "CUBLASLT_MATMUL_TILE_72x128" = 125
147
+ CUBLASLT_MATMUL_TILE_72x192 "CUBLASLT_MATMUL_TILE_72x192" = 126
148
+ CUBLASLT_MATMUL_TILE_72x256 "CUBLASLT_MATMUL_TILE_72x256" = 127
149
+ CUBLASLT_MATMUL_TILE_72x320 "CUBLASLT_MATMUL_TILE_72x320" = 128
150
+ CUBLASLT_MATMUL_TILE_72x384 "CUBLASLT_MATMUL_TILE_72x384" = 129
151
+ CUBLASLT_MATMUL_TILE_72x448 "CUBLASLT_MATMUL_TILE_72x448" = 130
152
+ CUBLASLT_MATMUL_TILE_72x512 "CUBLASLT_MATMUL_TILE_72x512" = 131
153
+ CUBLASLT_MATMUL_TILE_72x576 "CUBLASLT_MATMUL_TILE_72x576" = 132
154
+ CUBLASLT_MATMUL_TILE_72x640 "CUBLASLT_MATMUL_TILE_72x640" = 133
155
+ CUBLASLT_MATMUL_TILE_80x64 "CUBLASLT_MATMUL_TILE_80x64" = 134
156
+ CUBLASLT_MATMUL_TILE_80x128 "CUBLASLT_MATMUL_TILE_80x128" = 135
157
+ CUBLASLT_MATMUL_TILE_80x192 "CUBLASLT_MATMUL_TILE_80x192" = 136
158
+ CUBLASLT_MATMUL_TILE_80x256 "CUBLASLT_MATMUL_TILE_80x256" = 137
159
+ CUBLASLT_MATMUL_TILE_80x320 "CUBLASLT_MATMUL_TILE_80x320" = 138
160
+ CUBLASLT_MATMUL_TILE_80x384 "CUBLASLT_MATMUL_TILE_80x384" = 139
161
+ CUBLASLT_MATMUL_TILE_80x448 "CUBLASLT_MATMUL_TILE_80x448" = 140
162
+ CUBLASLT_MATMUL_TILE_80x512 "CUBLASLT_MATMUL_TILE_80x512" = 141
163
+ CUBLASLT_MATMUL_TILE_80x576 "CUBLASLT_MATMUL_TILE_80x576" = 142
164
+ CUBLASLT_MATMUL_TILE_88x64 "CUBLASLT_MATMUL_TILE_88x64" = 143
165
+ CUBLASLT_MATMUL_TILE_88x128 "CUBLASLT_MATMUL_TILE_88x128" = 144
166
+ CUBLASLT_MATMUL_TILE_88x192 "CUBLASLT_MATMUL_TILE_88x192" = 145
167
+ CUBLASLT_MATMUL_TILE_88x256 "CUBLASLT_MATMUL_TILE_88x256" = 146
168
+ CUBLASLT_MATMUL_TILE_88x320 "CUBLASLT_MATMUL_TILE_88x320" = 147
169
+ CUBLASLT_MATMUL_TILE_88x384 "CUBLASLT_MATMUL_TILE_88x384" = 148
170
+ CUBLASLT_MATMUL_TILE_88x448 "CUBLASLT_MATMUL_TILE_88x448" = 149
171
+ CUBLASLT_MATMUL_TILE_88x512 "CUBLASLT_MATMUL_TILE_88x512" = 150
172
+ CUBLASLT_MATMUL_TILE_96x192 "CUBLASLT_MATMUL_TILE_96x192" = 151
173
+ CUBLASLT_MATMUL_TILE_96x256 "CUBLASLT_MATMUL_TILE_96x256" = 152
174
+ CUBLASLT_MATMUL_TILE_96x320 "CUBLASLT_MATMUL_TILE_96x320" = 153
175
+ CUBLASLT_MATMUL_TILE_96x384 "CUBLASLT_MATMUL_TILE_96x384" = 154
176
+ CUBLASLT_MATMUL_TILE_96x448 "CUBLASLT_MATMUL_TILE_96x448" = 155
177
+ CUBLASLT_MATMUL_TILE_96x512 "CUBLASLT_MATMUL_TILE_96x512" = 156
178
+ CUBLASLT_MATMUL_TILE_104x64 "CUBLASLT_MATMUL_TILE_104x64" = 157
179
+ CUBLASLT_MATMUL_TILE_104x128 "CUBLASLT_MATMUL_TILE_104x128" = 158
180
+ CUBLASLT_MATMUL_TILE_104x192 "CUBLASLT_MATMUL_TILE_104x192" = 159
181
+ CUBLASLT_MATMUL_TILE_104x256 "CUBLASLT_MATMUL_TILE_104x256" = 160
182
+ CUBLASLT_MATMUL_TILE_104x320 "CUBLASLT_MATMUL_TILE_104x320" = 161
183
+ CUBLASLT_MATMUL_TILE_104x384 "CUBLASLT_MATMUL_TILE_104x384" = 162
184
+ CUBLASLT_MATMUL_TILE_104x448 "CUBLASLT_MATMUL_TILE_104x448" = 163
185
+ CUBLASLT_MATMUL_TILE_112x64 "CUBLASLT_MATMUL_TILE_112x64" = 164
186
+ CUBLASLT_MATMUL_TILE_112x128 "CUBLASLT_MATMUL_TILE_112x128" = 165
187
+ CUBLASLT_MATMUL_TILE_112x192 "CUBLASLT_MATMUL_TILE_112x192" = 166
188
+ CUBLASLT_MATMUL_TILE_112x256 "CUBLASLT_MATMUL_TILE_112x256" = 167
189
+ CUBLASLT_MATMUL_TILE_112x320 "CUBLASLT_MATMUL_TILE_112x320" = 168
190
+ CUBLASLT_MATMUL_TILE_112x384 "CUBLASLT_MATMUL_TILE_112x384" = 169
191
+ CUBLASLT_MATMUL_TILE_120x64 "CUBLASLT_MATMUL_TILE_120x64" = 170
192
+ CUBLASLT_MATMUL_TILE_120x128 "CUBLASLT_MATMUL_TILE_120x128" = 171
193
+ CUBLASLT_MATMUL_TILE_120x192 "CUBLASLT_MATMUL_TILE_120x192" = 172
194
+ CUBLASLT_MATMUL_TILE_120x256 "CUBLASLT_MATMUL_TILE_120x256" = 173
195
+ CUBLASLT_MATMUL_TILE_120x320 "CUBLASLT_MATMUL_TILE_120x320" = 174
196
+ CUBLASLT_MATMUL_TILE_120x384 "CUBLASLT_MATMUL_TILE_120x384" = 175
197
+ CUBLASLT_MATMUL_TILE_128x320 "CUBLASLT_MATMUL_TILE_128x320" = 176
198
+ CUBLASLT_MATMUL_TILE_128x384 "CUBLASLT_MATMUL_TILE_128x384" = 177
199
+ CUBLASLT_MATMUL_TILE_136x64 "CUBLASLT_MATMUL_TILE_136x64" = 178
200
+ CUBLASLT_MATMUL_TILE_136x128 "CUBLASLT_MATMUL_TILE_136x128" = 179
201
+ CUBLASLT_MATMUL_TILE_136x192 "CUBLASLT_MATMUL_TILE_136x192" = 180
202
+ CUBLASLT_MATMUL_TILE_136x256 "CUBLASLT_MATMUL_TILE_136x256" = 181
203
+ CUBLASLT_MATMUL_TILE_136x320 "CUBLASLT_MATMUL_TILE_136x320" = 182
204
+ CUBLASLT_MATMUL_TILE_144x64 "CUBLASLT_MATMUL_TILE_144x64" = 183
205
+ CUBLASLT_MATMUL_TILE_144x128 "CUBLASLT_MATMUL_TILE_144x128" = 184
206
+ CUBLASLT_MATMUL_TILE_144x192 "CUBLASLT_MATMUL_TILE_144x192" = 185
207
+ CUBLASLT_MATMUL_TILE_144x256 "CUBLASLT_MATMUL_TILE_144x256" = 186
208
+ CUBLASLT_MATMUL_TILE_144x320 "CUBLASLT_MATMUL_TILE_144x320" = 187
209
+ CUBLASLT_MATMUL_TILE_152x64 "CUBLASLT_MATMUL_TILE_152x64" = 188
210
+ CUBLASLT_MATMUL_TILE_152x128 "CUBLASLT_MATMUL_TILE_152x128" = 189
211
+ CUBLASLT_MATMUL_TILE_152x192 "CUBLASLT_MATMUL_TILE_152x192" = 190
212
+ CUBLASLT_MATMUL_TILE_152x256 "CUBLASLT_MATMUL_TILE_152x256" = 191
213
+ CUBLASLT_MATMUL_TILE_152x320 "CUBLASLT_MATMUL_TILE_152x320" = 192
214
+ CUBLASLT_MATMUL_TILE_160x64 "CUBLASLT_MATMUL_TILE_160x64" = 193
215
+ CUBLASLT_MATMUL_TILE_160x192 "CUBLASLT_MATMUL_TILE_160x192" = 194
216
+ CUBLASLT_MATMUL_TILE_160x256 "CUBLASLT_MATMUL_TILE_160x256" = 195
217
+ CUBLASLT_MATMUL_TILE_168x64 "CUBLASLT_MATMUL_TILE_168x64" = 196
218
+ CUBLASLT_MATMUL_TILE_168x128 "CUBLASLT_MATMUL_TILE_168x128" = 197
219
+ CUBLASLT_MATMUL_TILE_168x192 "CUBLASLT_MATMUL_TILE_168x192" = 198
220
+ CUBLASLT_MATMUL_TILE_168x256 "CUBLASLT_MATMUL_TILE_168x256" = 199
221
+ CUBLASLT_MATMUL_TILE_176x64 "CUBLASLT_MATMUL_TILE_176x64" = 200
222
+ CUBLASLT_MATMUL_TILE_176x128 "CUBLASLT_MATMUL_TILE_176x128" = 201
223
+ CUBLASLT_MATMUL_TILE_176x192 "CUBLASLT_MATMUL_TILE_176x192" = 202
224
+ CUBLASLT_MATMUL_TILE_176x256 "CUBLASLT_MATMUL_TILE_176x256" = 203
225
+ CUBLASLT_MATMUL_TILE_184x64 "CUBLASLT_MATMUL_TILE_184x64" = 204
226
+ CUBLASLT_MATMUL_TILE_184x128 "CUBLASLT_MATMUL_TILE_184x128" = 205
227
+ CUBLASLT_MATMUL_TILE_184x192 "CUBLASLT_MATMUL_TILE_184x192" = 206
228
+ CUBLASLT_MATMUL_TILE_184x256 "CUBLASLT_MATMUL_TILE_184x256" = 207
229
+ CUBLASLT_MATMUL_TILE_192x64 "CUBLASLT_MATMUL_TILE_192x64" = 208
230
+ CUBLASLT_MATMUL_TILE_192x192 "CUBLASLT_MATMUL_TILE_192x192" = 209
231
+ CUBLASLT_MATMUL_TILE_192x256 "CUBLASLT_MATMUL_TILE_192x256" = 210
232
+ CUBLASLT_MATMUL_TILE_200x64 "CUBLASLT_MATMUL_TILE_200x64" = 211
233
+ CUBLASLT_MATMUL_TILE_200x128 "CUBLASLT_MATMUL_TILE_200x128" = 212
234
+ CUBLASLT_MATMUL_TILE_200x192 "CUBLASLT_MATMUL_TILE_200x192" = 213
235
+ CUBLASLT_MATMUL_TILE_208x64 "CUBLASLT_MATMUL_TILE_208x64" = 214
236
+ CUBLASLT_MATMUL_TILE_208x128 "CUBLASLT_MATMUL_TILE_208x128" = 215
237
+ CUBLASLT_MATMUL_TILE_208x192 "CUBLASLT_MATMUL_TILE_208x192" = 216
238
+ CUBLASLT_MATMUL_TILE_216x64 "CUBLASLT_MATMUL_TILE_216x64" = 217
239
+ CUBLASLT_MATMUL_TILE_216x128 "CUBLASLT_MATMUL_TILE_216x128" = 218
240
+ CUBLASLT_MATMUL_TILE_216x192 "CUBLASLT_MATMUL_TILE_216x192" = 219
241
+ CUBLASLT_MATMUL_TILE_224x64 "CUBLASLT_MATMUL_TILE_224x64" = 220
242
+ CUBLASLT_MATMUL_TILE_224x128 "CUBLASLT_MATMUL_TILE_224x128" = 221
243
+ CUBLASLT_MATMUL_TILE_224x192 "CUBLASLT_MATMUL_TILE_224x192" = 222
244
+ CUBLASLT_MATMUL_TILE_232x64 "CUBLASLT_MATMUL_TILE_232x64" = 223
245
+ CUBLASLT_MATMUL_TILE_232x128 "CUBLASLT_MATMUL_TILE_232x128" = 224
246
+ CUBLASLT_MATMUL_TILE_232x192 "CUBLASLT_MATMUL_TILE_232x192" = 225
247
+ CUBLASLT_MATMUL_TILE_240x64 "CUBLASLT_MATMUL_TILE_240x64" = 226
248
+ CUBLASLT_MATMUL_TILE_240x128 "CUBLASLT_MATMUL_TILE_240x128" = 227
249
+ CUBLASLT_MATMUL_TILE_240x192 "CUBLASLT_MATMUL_TILE_240x192" = 228
250
+ CUBLASLT_MATMUL_TILE_248x64 "CUBLASLT_MATMUL_TILE_248x64" = 229
251
+ CUBLASLT_MATMUL_TILE_248x128 "CUBLASLT_MATMUL_TILE_248x128" = 230
252
+ CUBLASLT_MATMUL_TILE_248x192 "CUBLASLT_MATMUL_TILE_248x192" = 231
253
+ CUBLASLT_MATMUL_TILE_256x192 "CUBLASLT_MATMUL_TILE_256x192" = 232
254
+ CUBLASLT_MATMUL_TILE_264x64 "CUBLASLT_MATMUL_TILE_264x64" = 233
255
+ CUBLASLT_MATMUL_TILE_264x128 "CUBLASLT_MATMUL_TILE_264x128" = 234
256
+ CUBLASLT_MATMUL_TILE_272x64 "CUBLASLT_MATMUL_TILE_272x64" = 235
257
+ CUBLASLT_MATMUL_TILE_272x128 "CUBLASLT_MATMUL_TILE_272x128" = 236
258
+ CUBLASLT_MATMUL_TILE_280x64 "CUBLASLT_MATMUL_TILE_280x64" = 237
259
+ CUBLASLT_MATMUL_TILE_280x128 "CUBLASLT_MATMUL_TILE_280x128" = 238
260
+ CUBLASLT_MATMUL_TILE_288x64 "CUBLASLT_MATMUL_TILE_288x64" = 239
261
+ CUBLASLT_MATMUL_TILE_288x128 "CUBLASLT_MATMUL_TILE_288x128" = 240
262
+ CUBLASLT_MATMUL_TILE_296x64 "CUBLASLT_MATMUL_TILE_296x64" = 241
263
+ CUBLASLT_MATMUL_TILE_296x128 "CUBLASLT_MATMUL_TILE_296x128" = 242
264
+ CUBLASLT_MATMUL_TILE_304x64 "CUBLASLT_MATMUL_TILE_304x64" = 243
265
+ CUBLASLT_MATMUL_TILE_304x128 "CUBLASLT_MATMUL_TILE_304x128" = 244
266
+ CUBLASLT_MATMUL_TILE_312x64 "CUBLASLT_MATMUL_TILE_312x64" = 245
267
+ CUBLASLT_MATMUL_TILE_312x128 "CUBLASLT_MATMUL_TILE_312x128" = 246
268
+ CUBLASLT_MATMUL_TILE_320x64 "CUBLASLT_MATMUL_TILE_320x64" = 247
269
+ CUBLASLT_MATMUL_TILE_320x128 "CUBLASLT_MATMUL_TILE_320x128" = 248
270
+ CUBLASLT_MATMUL_TILE_328x64 "CUBLASLT_MATMUL_TILE_328x64" = 249
271
+ CUBLASLT_MATMUL_TILE_328x128 "CUBLASLT_MATMUL_TILE_328x128" = 250
272
+ CUBLASLT_MATMUL_TILE_336x64 "CUBLASLT_MATMUL_TILE_336x64" = 251
273
+ CUBLASLT_MATMUL_TILE_336x128 "CUBLASLT_MATMUL_TILE_336x128" = 252
274
+ CUBLASLT_MATMUL_TILE_344x64 "CUBLASLT_MATMUL_TILE_344x64" = 253
275
+ CUBLASLT_MATMUL_TILE_344x128 "CUBLASLT_MATMUL_TILE_344x128" = 254
276
+ CUBLASLT_MATMUL_TILE_352x64 "CUBLASLT_MATMUL_TILE_352x64" = 255
277
+ CUBLASLT_MATMUL_TILE_352x128 "CUBLASLT_MATMUL_TILE_352x128" = 256
278
+ CUBLASLT_MATMUL_TILE_360x64 "CUBLASLT_MATMUL_TILE_360x64" = 257
279
+ CUBLASLT_MATMUL_TILE_360x128 "CUBLASLT_MATMUL_TILE_360x128" = 258
280
+ CUBLASLT_MATMUL_TILE_368x64 "CUBLASLT_MATMUL_TILE_368x64" = 259
281
+ CUBLASLT_MATMUL_TILE_368x128 "CUBLASLT_MATMUL_TILE_368x128" = 260
282
+ CUBLASLT_MATMUL_TILE_376x64 "CUBLASLT_MATMUL_TILE_376x64" = 261
283
+ CUBLASLT_MATMUL_TILE_376x128 "CUBLASLT_MATMUL_TILE_376x128" = 262
284
+ CUBLASLT_MATMUL_TILE_384x64 "CUBLASLT_MATMUL_TILE_384x64" = 263
285
+ CUBLASLT_MATMUL_TILE_384x128 "CUBLASLT_MATMUL_TILE_384x128" = 264
286
+ CUBLASLT_MATMUL_TILE_392x64 "CUBLASLT_MATMUL_TILE_392x64" = 265
287
+ CUBLASLT_MATMUL_TILE_400x64 "CUBLASLT_MATMUL_TILE_400x64" = 266
288
+ CUBLASLT_MATMUL_TILE_408x64 "CUBLASLT_MATMUL_TILE_408x64" = 267
289
+ CUBLASLT_MATMUL_TILE_416x64 "CUBLASLT_MATMUL_TILE_416x64" = 268
290
+ CUBLASLT_MATMUL_TILE_424x64 "CUBLASLT_MATMUL_TILE_424x64" = 269
291
+ CUBLASLT_MATMUL_TILE_432x64 "CUBLASLT_MATMUL_TILE_432x64" = 270
292
+ CUBLASLT_MATMUL_TILE_440x64 "CUBLASLT_MATMUL_TILE_440x64" = 271
293
+ CUBLASLT_MATMUL_TILE_448x64 "CUBLASLT_MATMUL_TILE_448x64" = 272
294
+ CUBLASLT_MATMUL_TILE_456x64 "CUBLASLT_MATMUL_TILE_456x64" = 273
295
+ CUBLASLT_MATMUL_TILE_464x64 "CUBLASLT_MATMUL_TILE_464x64" = 274
296
+ CUBLASLT_MATMUL_TILE_472x64 "CUBLASLT_MATMUL_TILE_472x64" = 275
297
+ CUBLASLT_MATMUL_TILE_480x64 "CUBLASLT_MATMUL_TILE_480x64" = 276
298
+ CUBLASLT_MATMUL_TILE_488x64 "CUBLASLT_MATMUL_TILE_488x64" = 277
299
+ CUBLASLT_MATMUL_TILE_496x64 "CUBLASLT_MATMUL_TILE_496x64" = 278
300
+ CUBLASLT_MATMUL_TILE_504x64 "CUBLASLT_MATMUL_TILE_504x64" = 279
301
+ CUBLASLT_MATMUL_TILE_520x64 "CUBLASLT_MATMUL_TILE_520x64" = 280
302
+ CUBLASLT_MATMUL_TILE_528x64 "CUBLASLT_MATMUL_TILE_528x64" = 281
303
+ CUBLASLT_MATMUL_TILE_536x64 "CUBLASLT_MATMUL_TILE_536x64" = 282
304
+ CUBLASLT_MATMUL_TILE_544x64 "CUBLASLT_MATMUL_TILE_544x64" = 283
305
+ CUBLASLT_MATMUL_TILE_552x64 "CUBLASLT_MATMUL_TILE_552x64" = 284
306
+ CUBLASLT_MATMUL_TILE_560x64 "CUBLASLT_MATMUL_TILE_560x64" = 285
307
+ CUBLASLT_MATMUL_TILE_568x64 "CUBLASLT_MATMUL_TILE_568x64" = 286
308
+ CUBLASLT_MATMUL_TILE_576x64 "CUBLASLT_MATMUL_TILE_576x64" = 287
309
+ CUBLASLT_MATMUL_TILE_584x64 "CUBLASLT_MATMUL_TILE_584x64" = 288
310
+ CUBLASLT_MATMUL_TILE_592x64 "CUBLASLT_MATMUL_TILE_592x64" = 289
311
+ CUBLASLT_MATMUL_TILE_600x64 "CUBLASLT_MATMUL_TILE_600x64" = 290
312
+ CUBLASLT_MATMUL_TILE_608x64 "CUBLASLT_MATMUL_TILE_608x64" = 291
313
+ CUBLASLT_MATMUL_TILE_616x64 "CUBLASLT_MATMUL_TILE_616x64" = 292
314
+ CUBLASLT_MATMUL_TILE_624x64 "CUBLASLT_MATMUL_TILE_624x64" = 293
315
+ CUBLASLT_MATMUL_TILE_632x64 "CUBLASLT_MATMUL_TILE_632x64" = 294
316
+ CUBLASLT_MATMUL_TILE_640x64 "CUBLASLT_MATMUL_TILE_640x64" = 295
317
+ CUBLASLT_MATMUL_TILE_648x64 "CUBLASLT_MATMUL_TILE_648x64" = 296
318
+ CUBLASLT_MATMUL_TILE_656x64 "CUBLASLT_MATMUL_TILE_656x64" = 297
319
+ CUBLASLT_MATMUL_TILE_664x64 "CUBLASLT_MATMUL_TILE_664x64" = 298
320
+ CUBLASLT_MATMUL_TILE_672x64 "CUBLASLT_MATMUL_TILE_672x64" = 299
321
+ CUBLASLT_MATMUL_TILE_680x64 "CUBLASLT_MATMUL_TILE_680x64" = 300
322
+ CUBLASLT_MATMUL_TILE_688x64 "CUBLASLT_MATMUL_TILE_688x64" = 301
323
+ CUBLASLT_MATMUL_TILE_696x64 "CUBLASLT_MATMUL_TILE_696x64" = 302
324
+ CUBLASLT_MATMUL_TILE_704x64 "CUBLASLT_MATMUL_TILE_704x64" = 303
325
+ CUBLASLT_MATMUL_TILE_712x64 "CUBLASLT_MATMUL_TILE_712x64" = 304
326
+ CUBLASLT_MATMUL_TILE_720x64 "CUBLASLT_MATMUL_TILE_720x64" = 305
327
+ CUBLASLT_MATMUL_TILE_728x64 "CUBLASLT_MATMUL_TILE_728x64" = 306
328
+ CUBLASLT_MATMUL_TILE_736x64 "CUBLASLT_MATMUL_TILE_736x64" = 307
329
+ CUBLASLT_MATMUL_TILE_744x64 "CUBLASLT_MATMUL_TILE_744x64" = 308
330
+ CUBLASLT_MATMUL_TILE_752x64 "CUBLASLT_MATMUL_TILE_752x64" = 309
331
+ CUBLASLT_MATMUL_TILE_760x64 "CUBLASLT_MATMUL_TILE_760x64" = 310
332
+ CUBLASLT_MATMUL_TILE_768x64 "CUBLASLT_MATMUL_TILE_768x64" = 311
333
+ CUBLASLT_MATMUL_TILE_64x16 "CUBLASLT_MATMUL_TILE_64x16" = 312
334
+ CUBLASLT_MATMUL_TILE_64x24 "CUBLASLT_MATMUL_TILE_64x24" = 313
335
+ CUBLASLT_MATMUL_TILE_64x40 "CUBLASLT_MATMUL_TILE_64x40" = 314
336
+ CUBLASLT_MATMUL_TILE_64x48 "CUBLASLT_MATMUL_TILE_64x48" = 315
337
+ CUBLASLT_MATMUL_TILE_64x56 "CUBLASLT_MATMUL_TILE_64x56" = 316
338
+ CUBLASLT_MATMUL_TILE_64x72 "CUBLASLT_MATMUL_TILE_64x72" = 317
339
+ CUBLASLT_MATMUL_TILE_64x80 "CUBLASLT_MATMUL_TILE_64x80" = 318
340
+ CUBLASLT_MATMUL_TILE_64x88 "CUBLASLT_MATMUL_TILE_64x88" = 319
341
+ CUBLASLT_MATMUL_TILE_64x104 "CUBLASLT_MATMUL_TILE_64x104" = 320
342
+ CUBLASLT_MATMUL_TILE_64x112 "CUBLASLT_MATMUL_TILE_64x112" = 321
343
+ CUBLASLT_MATMUL_TILE_64x120 "CUBLASLT_MATMUL_TILE_64x120" = 322
344
+ CUBLASLT_MATMUL_TILE_64x136 "CUBLASLT_MATMUL_TILE_64x136" = 323
345
+ CUBLASLT_MATMUL_TILE_64x144 "CUBLASLT_MATMUL_TILE_64x144" = 324
346
+ CUBLASLT_MATMUL_TILE_64x152 "CUBLASLT_MATMUL_TILE_64x152" = 325
347
+ CUBLASLT_MATMUL_TILE_64x160 "CUBLASLT_MATMUL_TILE_64x160" = 326
348
+ CUBLASLT_MATMUL_TILE_64x168 "CUBLASLT_MATMUL_TILE_64x168" = 327
349
+ CUBLASLT_MATMUL_TILE_64x176 "CUBLASLT_MATMUL_TILE_64x176" = 328
350
+ CUBLASLT_MATMUL_TILE_64x184 "CUBLASLT_MATMUL_TILE_64x184" = 329
351
+ CUBLASLT_MATMUL_TILE_64x200 "CUBLASLT_MATMUL_TILE_64x200" = 330
352
+ CUBLASLT_MATMUL_TILE_64x208 "CUBLASLT_MATMUL_TILE_64x208" = 331
353
+ CUBLASLT_MATMUL_TILE_64x216 "CUBLASLT_MATMUL_TILE_64x216" = 332
354
+ CUBLASLT_MATMUL_TILE_64x224 "CUBLASLT_MATMUL_TILE_64x224" = 333
355
+ CUBLASLT_MATMUL_TILE_64x232 "CUBLASLT_MATMUL_TILE_64x232" = 334
356
+ CUBLASLT_MATMUL_TILE_64x240 "CUBLASLT_MATMUL_TILE_64x240" = 335
357
+ CUBLASLT_MATMUL_TILE_64x248 "CUBLASLT_MATMUL_TILE_64x248" = 336
358
+ CUBLASLT_MATMUL_TILE_64x264 "CUBLASLT_MATMUL_TILE_64x264" = 337
359
+ CUBLASLT_MATMUL_TILE_64x272 "CUBLASLT_MATMUL_TILE_64x272" = 338
360
+ CUBLASLT_MATMUL_TILE_64x280 "CUBLASLT_MATMUL_TILE_64x280" = 339
361
+ CUBLASLT_MATMUL_TILE_64x288 "CUBLASLT_MATMUL_TILE_64x288" = 340
362
+ CUBLASLT_MATMUL_TILE_64x296 "CUBLASLT_MATMUL_TILE_64x296" = 341
363
+ CUBLASLT_MATMUL_TILE_64x304 "CUBLASLT_MATMUL_TILE_64x304" = 342
364
+ CUBLASLT_MATMUL_TILE_64x312 "CUBLASLT_MATMUL_TILE_64x312" = 343
365
+ CUBLASLT_MATMUL_TILE_64x328 "CUBLASLT_MATMUL_TILE_64x328" = 344
366
+ CUBLASLT_MATMUL_TILE_64x336 "CUBLASLT_MATMUL_TILE_64x336" = 345
367
+ CUBLASLT_MATMUL_TILE_64x344 "CUBLASLT_MATMUL_TILE_64x344" = 346
368
+ CUBLASLT_MATMUL_TILE_64x352 "CUBLASLT_MATMUL_TILE_64x352" = 347
369
+ CUBLASLT_MATMUL_TILE_64x360 "CUBLASLT_MATMUL_TILE_64x360" = 348
370
+ CUBLASLT_MATMUL_TILE_64x368 "CUBLASLT_MATMUL_TILE_64x368" = 349
371
+ CUBLASLT_MATMUL_TILE_64x376 "CUBLASLT_MATMUL_TILE_64x376" = 350
372
+ CUBLASLT_MATMUL_TILE_64x392 "CUBLASLT_MATMUL_TILE_64x392" = 351
373
+ CUBLASLT_MATMUL_TILE_64x400 "CUBLASLT_MATMUL_TILE_64x400" = 352
374
+ CUBLASLT_MATMUL_TILE_64x408 "CUBLASLT_MATMUL_TILE_64x408" = 353
375
+ CUBLASLT_MATMUL_TILE_64x416 "CUBLASLT_MATMUL_TILE_64x416" = 354
376
+ CUBLASLT_MATMUL_TILE_64x424 "CUBLASLT_MATMUL_TILE_64x424" = 355
377
+ CUBLASLT_MATMUL_TILE_64x432 "CUBLASLT_MATMUL_TILE_64x432" = 356
378
+ CUBLASLT_MATMUL_TILE_64x440 "CUBLASLT_MATMUL_TILE_64x440" = 357
379
+ CUBLASLT_MATMUL_TILE_64x456 "CUBLASLT_MATMUL_TILE_64x456" = 358
380
+ CUBLASLT_MATMUL_TILE_64x464 "CUBLASLT_MATMUL_TILE_64x464" = 359
381
+ CUBLASLT_MATMUL_TILE_64x472 "CUBLASLT_MATMUL_TILE_64x472" = 360
382
+ CUBLASLT_MATMUL_TILE_64x480 "CUBLASLT_MATMUL_TILE_64x480" = 361
383
+ CUBLASLT_MATMUL_TILE_64x488 "CUBLASLT_MATMUL_TILE_64x488" = 362
384
+ CUBLASLT_MATMUL_TILE_64x496 "CUBLASLT_MATMUL_TILE_64x496" = 363
385
+ CUBLASLT_MATMUL_TILE_64x504 "CUBLASLT_MATMUL_TILE_64x504" = 364
386
+ CUBLASLT_MATMUL_TILE_64x520 "CUBLASLT_MATMUL_TILE_64x520" = 365
387
+ CUBLASLT_MATMUL_TILE_64x528 "CUBLASLT_MATMUL_TILE_64x528" = 366
388
+ CUBLASLT_MATMUL_TILE_64x536 "CUBLASLT_MATMUL_TILE_64x536" = 367
389
+ CUBLASLT_MATMUL_TILE_64x544 "CUBLASLT_MATMUL_TILE_64x544" = 368
390
+ CUBLASLT_MATMUL_TILE_64x552 "CUBLASLT_MATMUL_TILE_64x552" = 369
391
+ CUBLASLT_MATMUL_TILE_64x560 "CUBLASLT_MATMUL_TILE_64x560" = 370
392
+ CUBLASLT_MATMUL_TILE_64x568 "CUBLASLT_MATMUL_TILE_64x568" = 371
393
+ CUBLASLT_MATMUL_TILE_64x584 "CUBLASLT_MATMUL_TILE_64x584" = 372
394
+ CUBLASLT_MATMUL_TILE_64x592 "CUBLASLT_MATMUL_TILE_64x592" = 373
395
+ CUBLASLT_MATMUL_TILE_64x600 "CUBLASLT_MATMUL_TILE_64x600" = 374
396
+ CUBLASLT_MATMUL_TILE_64x608 "CUBLASLT_MATMUL_TILE_64x608" = 375
397
+ CUBLASLT_MATMUL_TILE_64x616 "CUBLASLT_MATMUL_TILE_64x616" = 376
398
+ CUBLASLT_MATMUL_TILE_64x624 "CUBLASLT_MATMUL_TILE_64x624" = 377
399
+ CUBLASLT_MATMUL_TILE_64x632 "CUBLASLT_MATMUL_TILE_64x632" = 378
400
+ CUBLASLT_MATMUL_TILE_64x648 "CUBLASLT_MATMUL_TILE_64x648" = 379
401
+ CUBLASLT_MATMUL_TILE_64x656 "CUBLASLT_MATMUL_TILE_64x656" = 380
402
+ CUBLASLT_MATMUL_TILE_64x664 "CUBLASLT_MATMUL_TILE_64x664" = 381
403
+ CUBLASLT_MATMUL_TILE_64x672 "CUBLASLT_MATMUL_TILE_64x672" = 382
404
+ CUBLASLT_MATMUL_TILE_64x680 "CUBLASLT_MATMUL_TILE_64x680" = 383
405
+ CUBLASLT_MATMUL_TILE_64x688 "CUBLASLT_MATMUL_TILE_64x688" = 384
406
+ CUBLASLT_MATMUL_TILE_64x696 "CUBLASLT_MATMUL_TILE_64x696" = 385
407
+ CUBLASLT_MATMUL_TILE_64x712 "CUBLASLT_MATMUL_TILE_64x712" = 386
408
+ CUBLASLT_MATMUL_TILE_64x720 "CUBLASLT_MATMUL_TILE_64x720" = 387
409
+ CUBLASLT_MATMUL_TILE_64x728 "CUBLASLT_MATMUL_TILE_64x728" = 388
410
+ CUBLASLT_MATMUL_TILE_64x736 "CUBLASLT_MATMUL_TILE_64x736" = 389
411
+ CUBLASLT_MATMUL_TILE_64x744 "CUBLASLT_MATMUL_TILE_64x744" = 390
412
+ CUBLASLT_MATMUL_TILE_64x752 "CUBLASLT_MATMUL_TILE_64x752" = 391
413
+ CUBLASLT_MATMUL_TILE_64x760 "CUBLASLT_MATMUL_TILE_64x760" = 392
414
+ CUBLASLT_MATMUL_TILE_128x8 "CUBLASLT_MATMUL_TILE_128x8" = 393
415
+ CUBLASLT_MATMUL_TILE_128x16 "CUBLASLT_MATMUL_TILE_128x16" = 394
416
+ CUBLASLT_MATMUL_TILE_128x24 "CUBLASLT_MATMUL_TILE_128x24" = 395
417
+ CUBLASLT_MATMUL_TILE_128x40 "CUBLASLT_MATMUL_TILE_128x40" = 396
418
+ CUBLASLT_MATMUL_TILE_128x48 "CUBLASLT_MATMUL_TILE_128x48" = 397
419
+ CUBLASLT_MATMUL_TILE_128x56 "CUBLASLT_MATMUL_TILE_128x56" = 398
420
+ CUBLASLT_MATMUL_TILE_128x72 "CUBLASLT_MATMUL_TILE_128x72" = 399
421
+ CUBLASLT_MATMUL_TILE_128x80 "CUBLASLT_MATMUL_TILE_128x80" = 400
422
+ CUBLASLT_MATMUL_TILE_128x88 "CUBLASLT_MATMUL_TILE_128x88" = 401
423
+ CUBLASLT_MATMUL_TILE_128x104 "CUBLASLT_MATMUL_TILE_128x104" = 402
424
+ CUBLASLT_MATMUL_TILE_128x112 "CUBLASLT_MATMUL_TILE_128x112" = 403
425
+ CUBLASLT_MATMUL_TILE_128x120 "CUBLASLT_MATMUL_TILE_128x120" = 404
426
+ CUBLASLT_MATMUL_TILE_128x136 "CUBLASLT_MATMUL_TILE_128x136" = 405
427
+ CUBLASLT_MATMUL_TILE_128x144 "CUBLASLT_MATMUL_TILE_128x144" = 406
428
+ CUBLASLT_MATMUL_TILE_128x152 "CUBLASLT_MATMUL_TILE_128x152" = 407
429
+ CUBLASLT_MATMUL_TILE_128x168 "CUBLASLT_MATMUL_TILE_128x168" = 408
430
+ CUBLASLT_MATMUL_TILE_128x176 "CUBLASLT_MATMUL_TILE_128x176" = 409
431
+ CUBLASLT_MATMUL_TILE_128x184 "CUBLASLT_MATMUL_TILE_128x184" = 410
432
+ CUBLASLT_MATMUL_TILE_128x200 "CUBLASLT_MATMUL_TILE_128x200" = 411
433
+ CUBLASLT_MATMUL_TILE_128x208 "CUBLASLT_MATMUL_TILE_128x208" = 412
434
+ CUBLASLT_MATMUL_TILE_128x216 "CUBLASLT_MATMUL_TILE_128x216" = 413
435
+ CUBLASLT_MATMUL_TILE_128x224 "CUBLASLT_MATMUL_TILE_128x224" = 414
436
+ CUBLASLT_MATMUL_TILE_128x232 "CUBLASLT_MATMUL_TILE_128x232" = 415
437
+ CUBLASLT_MATMUL_TILE_128x240 "CUBLASLT_MATMUL_TILE_128x240" = 416
438
+ CUBLASLT_MATMUL_TILE_128x248 "CUBLASLT_MATMUL_TILE_128x248" = 417
439
+ CUBLASLT_MATMUL_TILE_128x264 "CUBLASLT_MATMUL_TILE_128x264" = 418
440
+ CUBLASLT_MATMUL_TILE_128x272 "CUBLASLT_MATMUL_TILE_128x272" = 419
441
+ CUBLASLT_MATMUL_TILE_128x280 "CUBLASLT_MATMUL_TILE_128x280" = 420
442
+ CUBLASLT_MATMUL_TILE_128x288 "CUBLASLT_MATMUL_TILE_128x288" = 421
443
+ CUBLASLT_MATMUL_TILE_128x296 "CUBLASLT_MATMUL_TILE_128x296" = 422
444
+ CUBLASLT_MATMUL_TILE_128x304 "CUBLASLT_MATMUL_TILE_128x304" = 423
445
+ CUBLASLT_MATMUL_TILE_128x312 "CUBLASLT_MATMUL_TILE_128x312" = 424
446
+ CUBLASLT_MATMUL_TILE_128x328 "CUBLASLT_MATMUL_TILE_128x328" = 425
447
+ CUBLASLT_MATMUL_TILE_128x336 "CUBLASLT_MATMUL_TILE_128x336" = 426
448
+ CUBLASLT_MATMUL_TILE_128x344 "CUBLASLT_MATMUL_TILE_128x344" = 427
449
+ CUBLASLT_MATMUL_TILE_128x352 "CUBLASLT_MATMUL_TILE_128x352" = 428
450
+ CUBLASLT_MATMUL_TILE_128x360 "CUBLASLT_MATMUL_TILE_128x360" = 429
451
+ CUBLASLT_MATMUL_TILE_128x368 "CUBLASLT_MATMUL_TILE_128x368" = 430
452
+ CUBLASLT_MATMUL_TILE_128x376 "CUBLASLT_MATMUL_TILE_128x376" = 431
453
+ CUBLASLT_MATMUL_TILE_128x392 "CUBLASLT_MATMUL_TILE_128x392" = 432
454
+ CUBLASLT_MATMUL_TILE_128x400 "CUBLASLT_MATMUL_TILE_128x400" = 433
455
+ CUBLASLT_MATMUL_TILE_128x408 "CUBLASLT_MATMUL_TILE_128x408" = 434
456
+ CUBLASLT_MATMUL_TILE_128x416 "CUBLASLT_MATMUL_TILE_128x416" = 435
457
+ CUBLASLT_MATMUL_TILE_128x424 "CUBLASLT_MATMUL_TILE_128x424" = 436
458
+ CUBLASLT_MATMUL_TILE_128x432 "CUBLASLT_MATMUL_TILE_128x432" = 437
459
+ CUBLASLT_MATMUL_TILE_128x440 "CUBLASLT_MATMUL_TILE_128x440" = 438
460
+ CUBLASLT_MATMUL_TILE_128x448 "CUBLASLT_MATMUL_TILE_128x448" = 439
461
+ CUBLASLT_MATMUL_TILE_128x456 "CUBLASLT_MATMUL_TILE_128x456" = 440
462
+ CUBLASLT_MATMUL_TILE_128x464 "CUBLASLT_MATMUL_TILE_128x464" = 441
463
+ CUBLASLT_MATMUL_TILE_128x472 "CUBLASLT_MATMUL_TILE_128x472" = 442
464
+ CUBLASLT_MATMUL_TILE_128x480 "CUBLASLT_MATMUL_TILE_128x480" = 443
465
+ CUBLASLT_MATMUL_TILE_128x488 "CUBLASLT_MATMUL_TILE_128x488" = 444
466
+ CUBLASLT_MATMUL_TILE_128x496 "CUBLASLT_MATMUL_TILE_128x496" = 445
467
+ CUBLASLT_MATMUL_TILE_128x504 "CUBLASLT_MATMUL_TILE_128x504" = 446
468
+ CUBLASLT_MATMUL_TILE_128x512 "CUBLASLT_MATMUL_TILE_128x512" = 447
469
+ CUBLASLT_MATMUL_TILE_192x8 "CUBLASLT_MATMUL_TILE_192x8" = 448
470
+ CUBLASLT_MATMUL_TILE_192x16 "CUBLASLT_MATMUL_TILE_192x16" = 449
471
+ CUBLASLT_MATMUL_TILE_192x24 "CUBLASLT_MATMUL_TILE_192x24" = 450
472
+ CUBLASLT_MATMUL_TILE_192x32 "CUBLASLT_MATMUL_TILE_192x32" = 451
473
+ CUBLASLT_MATMUL_TILE_192x40 "CUBLASLT_MATMUL_TILE_192x40" = 452
474
+ CUBLASLT_MATMUL_TILE_192x48 "CUBLASLT_MATMUL_TILE_192x48" = 453
475
+ CUBLASLT_MATMUL_TILE_192x56 "CUBLASLT_MATMUL_TILE_192x56" = 454
476
+ CUBLASLT_MATMUL_TILE_192x72 "CUBLASLT_MATMUL_TILE_192x72" = 455
477
+ CUBLASLT_MATMUL_TILE_192x80 "CUBLASLT_MATMUL_TILE_192x80" = 456
478
+ CUBLASLT_MATMUL_TILE_192x88 "CUBLASLT_MATMUL_TILE_192x88" = 457
479
+ CUBLASLT_MATMUL_TILE_192x96 "CUBLASLT_MATMUL_TILE_192x96" = 458
480
+ CUBLASLT_MATMUL_TILE_192x104 "CUBLASLT_MATMUL_TILE_192x104" = 459
481
+ CUBLASLT_MATMUL_TILE_192x112 "CUBLASLT_MATMUL_TILE_192x112" = 460
482
+ CUBLASLT_MATMUL_TILE_192x120 "CUBLASLT_MATMUL_TILE_192x120" = 461
483
+ CUBLASLT_MATMUL_TILE_192x136 "CUBLASLT_MATMUL_TILE_192x136" = 462
484
+ CUBLASLT_MATMUL_TILE_192x144 "CUBLASLT_MATMUL_TILE_192x144" = 463
485
+ CUBLASLT_MATMUL_TILE_192x152 "CUBLASLT_MATMUL_TILE_192x152" = 464
486
+ CUBLASLT_MATMUL_TILE_192x160 "CUBLASLT_MATMUL_TILE_192x160" = 465
487
+ CUBLASLT_MATMUL_TILE_192x168 "CUBLASLT_MATMUL_TILE_192x168" = 466
488
+ CUBLASLT_MATMUL_TILE_192x176 "CUBLASLT_MATMUL_TILE_192x176" = 467
489
+ CUBLASLT_MATMUL_TILE_192x184 "CUBLASLT_MATMUL_TILE_192x184" = 468
490
+ CUBLASLT_MATMUL_TILE_192x200 "CUBLASLT_MATMUL_TILE_192x200" = 469
491
+ CUBLASLT_MATMUL_TILE_192x208 "CUBLASLT_MATMUL_TILE_192x208" = 470
492
+ CUBLASLT_MATMUL_TILE_192x216 "CUBLASLT_MATMUL_TILE_192x216" = 471
493
+ CUBLASLT_MATMUL_TILE_192x224 "CUBLASLT_MATMUL_TILE_192x224" = 472
494
+ CUBLASLT_MATMUL_TILE_192x232 "CUBLASLT_MATMUL_TILE_192x232" = 473
495
+ CUBLASLT_MATMUL_TILE_192x240 "CUBLASLT_MATMUL_TILE_192x240" = 474
496
+ CUBLASLT_MATMUL_TILE_192x248 "CUBLASLT_MATMUL_TILE_192x248" = 475
497
+ CUBLASLT_MATMUL_TILE_192x264 "CUBLASLT_MATMUL_TILE_192x264" = 476
498
+ CUBLASLT_MATMUL_TILE_192x272 "CUBLASLT_MATMUL_TILE_192x272" = 477
499
+ CUBLASLT_MATMUL_TILE_192x280 "CUBLASLT_MATMUL_TILE_192x280" = 478
500
+ CUBLASLT_MATMUL_TILE_192x288 "CUBLASLT_MATMUL_TILE_192x288" = 479
501
+ CUBLASLT_MATMUL_TILE_192x296 "CUBLASLT_MATMUL_TILE_192x296" = 480
502
+ CUBLASLT_MATMUL_TILE_192x304 "CUBLASLT_MATMUL_TILE_192x304" = 481
503
+ CUBLASLT_MATMUL_TILE_192x312 "CUBLASLT_MATMUL_TILE_192x312" = 482
504
+ CUBLASLT_MATMUL_TILE_192x320 "CUBLASLT_MATMUL_TILE_192x320" = 483
505
+ CUBLASLT_MATMUL_TILE_192x328 "CUBLASLT_MATMUL_TILE_192x328" = 484
506
+ CUBLASLT_MATMUL_TILE_192x336 "CUBLASLT_MATMUL_TILE_192x336" = 485
507
+ CUBLASLT_MATMUL_TILE_256x8 "CUBLASLT_MATMUL_TILE_256x8" = 486
508
+ CUBLASLT_MATMUL_TILE_256x16 "CUBLASLT_MATMUL_TILE_256x16" = 487
509
+ CUBLASLT_MATMUL_TILE_256x24 "CUBLASLT_MATMUL_TILE_256x24" = 488
510
+ CUBLASLT_MATMUL_TILE_256x40 "CUBLASLT_MATMUL_TILE_256x40" = 489
511
+ CUBLASLT_MATMUL_TILE_256x48 "CUBLASLT_MATMUL_TILE_256x48" = 490
512
+ CUBLASLT_MATMUL_TILE_256x56 "CUBLASLT_MATMUL_TILE_256x56" = 491
513
+ CUBLASLT_MATMUL_TILE_256x72 "CUBLASLT_MATMUL_TILE_256x72" = 492
514
+ CUBLASLT_MATMUL_TILE_256x80 "CUBLASLT_MATMUL_TILE_256x80" = 493
515
+ CUBLASLT_MATMUL_TILE_256x88 "CUBLASLT_MATMUL_TILE_256x88" = 494
516
+ CUBLASLT_MATMUL_TILE_256x96 "CUBLASLT_MATMUL_TILE_256x96" = 495
517
+ CUBLASLT_MATMUL_TILE_256x104 "CUBLASLT_MATMUL_TILE_256x104" = 496
518
+ CUBLASLT_MATMUL_TILE_256x112 "CUBLASLT_MATMUL_TILE_256x112" = 497
519
+ CUBLASLT_MATMUL_TILE_256x120 "CUBLASLT_MATMUL_TILE_256x120" = 498
520
+ CUBLASLT_MATMUL_TILE_256x136 "CUBLASLT_MATMUL_TILE_256x136" = 499
521
+ CUBLASLT_MATMUL_TILE_256x144 "CUBLASLT_MATMUL_TILE_256x144" = 500
522
+ CUBLASLT_MATMUL_TILE_256x152 "CUBLASLT_MATMUL_TILE_256x152" = 501
523
+ CUBLASLT_MATMUL_TILE_256x160 "CUBLASLT_MATMUL_TILE_256x160" = 502
524
+ CUBLASLT_MATMUL_TILE_256x168 "CUBLASLT_MATMUL_TILE_256x168" = 503
525
+ CUBLASLT_MATMUL_TILE_256x176 "CUBLASLT_MATMUL_TILE_256x176" = 504
526
+ CUBLASLT_MATMUL_TILE_256x184 "CUBLASLT_MATMUL_TILE_256x184" = 505
527
+ CUBLASLT_MATMUL_TILE_256x200 "CUBLASLT_MATMUL_TILE_256x200" = 506
528
+ CUBLASLT_MATMUL_TILE_256x208 "CUBLASLT_MATMUL_TILE_256x208" = 507
529
+ CUBLASLT_MATMUL_TILE_256x216 "CUBLASLT_MATMUL_TILE_256x216" = 508
530
+ CUBLASLT_MATMUL_TILE_256x224 "CUBLASLT_MATMUL_TILE_256x224" = 509
531
+ CUBLASLT_MATMUL_TILE_256x232 "CUBLASLT_MATMUL_TILE_256x232" = 510
532
+ CUBLASLT_MATMUL_TILE_256x240 "CUBLASLT_MATMUL_TILE_256x240" = 511
533
+ CUBLASLT_MATMUL_TILE_256x248 "CUBLASLT_MATMUL_TILE_256x248" = 512
534
+ CUBLASLT_MATMUL_TILE_256x256 "CUBLASLT_MATMUL_TILE_256x256" = 513
535
+ CUBLASLT_MATMUL_TILE_320x8 "CUBLASLT_MATMUL_TILE_320x8" = 514
536
+ CUBLASLT_MATMUL_TILE_320x16 "CUBLASLT_MATMUL_TILE_320x16" = 515
537
+ CUBLASLT_MATMUL_TILE_320x24 "CUBLASLT_MATMUL_TILE_320x24" = 516
538
+ CUBLASLT_MATMUL_TILE_320x32 "CUBLASLT_MATMUL_TILE_320x32" = 517
539
+ CUBLASLT_MATMUL_TILE_320x40 "CUBLASLT_MATMUL_TILE_320x40" = 518
540
+ CUBLASLT_MATMUL_TILE_320x48 "CUBLASLT_MATMUL_TILE_320x48" = 519
541
+ CUBLASLT_MATMUL_TILE_320x56 "CUBLASLT_MATMUL_TILE_320x56" = 520
542
+ CUBLASLT_MATMUL_TILE_320x72 "CUBLASLT_MATMUL_TILE_320x72" = 521
543
+ CUBLASLT_MATMUL_TILE_320x80 "CUBLASLT_MATMUL_TILE_320x80" = 522
544
+ CUBLASLT_MATMUL_TILE_320x88 "CUBLASLT_MATMUL_TILE_320x88" = 523
545
+ CUBLASLT_MATMUL_TILE_320x96 "CUBLASLT_MATMUL_TILE_320x96" = 524
546
+ CUBLASLT_MATMUL_TILE_320x104 "CUBLASLT_MATMUL_TILE_320x104" = 525
547
+ CUBLASLT_MATMUL_TILE_320x112 "CUBLASLT_MATMUL_TILE_320x112" = 526
548
+ CUBLASLT_MATMUL_TILE_320x120 "CUBLASLT_MATMUL_TILE_320x120" = 527
549
+ CUBLASLT_MATMUL_TILE_320x136 "CUBLASLT_MATMUL_TILE_320x136" = 528
550
+ CUBLASLT_MATMUL_TILE_320x144 "CUBLASLT_MATMUL_TILE_320x144" = 529
551
+ CUBLASLT_MATMUL_TILE_320x152 "CUBLASLT_MATMUL_TILE_320x152" = 530
552
+ CUBLASLT_MATMUL_TILE_320x160 "CUBLASLT_MATMUL_TILE_320x160" = 531
553
+ CUBLASLT_MATMUL_TILE_320x168 "CUBLASLT_MATMUL_TILE_320x168" = 532
554
+ CUBLASLT_MATMUL_TILE_320x176 "CUBLASLT_MATMUL_TILE_320x176" = 533
555
+ CUBLASLT_MATMUL_TILE_320x184 "CUBLASLT_MATMUL_TILE_320x184" = 534
556
+ CUBLASLT_MATMUL_TILE_320x192 "CUBLASLT_MATMUL_TILE_320x192" = 535
557
+ CUBLASLT_MATMUL_TILE_320x200 "CUBLASLT_MATMUL_TILE_320x200" = 536
558
+ CUBLASLT_MATMUL_TILE_384x8 "CUBLASLT_MATMUL_TILE_384x8" = 537
559
+ CUBLASLT_MATMUL_TILE_384x16 "CUBLASLT_MATMUL_TILE_384x16" = 538
560
+ CUBLASLT_MATMUL_TILE_384x24 "CUBLASLT_MATMUL_TILE_384x24" = 539
561
+ CUBLASLT_MATMUL_TILE_384x32 "CUBLASLT_MATMUL_TILE_384x32" = 540
562
+ CUBLASLT_MATMUL_TILE_384x40 "CUBLASLT_MATMUL_TILE_384x40" = 541
563
+ CUBLASLT_MATMUL_TILE_384x48 "CUBLASLT_MATMUL_TILE_384x48" = 542
564
+ CUBLASLT_MATMUL_TILE_384x56 "CUBLASLT_MATMUL_TILE_384x56" = 543
565
+ CUBLASLT_MATMUL_TILE_384x72 "CUBLASLT_MATMUL_TILE_384x72" = 544
566
+ CUBLASLT_MATMUL_TILE_384x80 "CUBLASLT_MATMUL_TILE_384x80" = 545
567
+ CUBLASLT_MATMUL_TILE_384x88 "CUBLASLT_MATMUL_TILE_384x88" = 546
568
+ CUBLASLT_MATMUL_TILE_384x96 "CUBLASLT_MATMUL_TILE_384x96" = 547
569
+ CUBLASLT_MATMUL_TILE_384x104 "CUBLASLT_MATMUL_TILE_384x104" = 548
570
+ CUBLASLT_MATMUL_TILE_384x112 "CUBLASLT_MATMUL_TILE_384x112" = 549
571
+ CUBLASLT_MATMUL_TILE_384x120 "CUBLASLT_MATMUL_TILE_384x120" = 550
572
+ CUBLASLT_MATMUL_TILE_384x136 "CUBLASLT_MATMUL_TILE_384x136" = 551
573
+ CUBLASLT_MATMUL_TILE_384x144 "CUBLASLT_MATMUL_TILE_384x144" = 552
574
+ CUBLASLT_MATMUL_TILE_384x152 "CUBLASLT_MATMUL_TILE_384x152" = 553
575
+ CUBLASLT_MATMUL_TILE_384x160 "CUBLASLT_MATMUL_TILE_384x160" = 554
576
+ CUBLASLT_MATMUL_TILE_384x168 "CUBLASLT_MATMUL_TILE_384x168" = 555
577
+ CUBLASLT_MATMUL_TILE_448x8 "CUBLASLT_MATMUL_TILE_448x8" = 556
578
+ CUBLASLT_MATMUL_TILE_448x16 "CUBLASLT_MATMUL_TILE_448x16" = 557
579
+ CUBLASLT_MATMUL_TILE_448x24 "CUBLASLT_MATMUL_TILE_448x24" = 558
580
+ CUBLASLT_MATMUL_TILE_448x32 "CUBLASLT_MATMUL_TILE_448x32" = 559
581
+ CUBLASLT_MATMUL_TILE_448x40 "CUBLASLT_MATMUL_TILE_448x40" = 560
582
+ CUBLASLT_MATMUL_TILE_448x48 "CUBLASLT_MATMUL_TILE_448x48" = 561
583
+ CUBLASLT_MATMUL_TILE_448x56 "CUBLASLT_MATMUL_TILE_448x56" = 562
584
+ CUBLASLT_MATMUL_TILE_448x72 "CUBLASLT_MATMUL_TILE_448x72" = 563
585
+ CUBLASLT_MATMUL_TILE_448x80 "CUBLASLT_MATMUL_TILE_448x80" = 564
586
+ CUBLASLT_MATMUL_TILE_448x88 "CUBLASLT_MATMUL_TILE_448x88" = 565
587
+ CUBLASLT_MATMUL_TILE_448x96 "CUBLASLT_MATMUL_TILE_448x96" = 566
588
+ CUBLASLT_MATMUL_TILE_448x104 "CUBLASLT_MATMUL_TILE_448x104" = 567
589
+ CUBLASLT_MATMUL_TILE_448x112 "CUBLASLT_MATMUL_TILE_448x112" = 568
590
+ CUBLASLT_MATMUL_TILE_448x120 "CUBLASLT_MATMUL_TILE_448x120" = 569
591
+ CUBLASLT_MATMUL_TILE_448x128 "CUBLASLT_MATMUL_TILE_448x128" = 570
592
+ CUBLASLT_MATMUL_TILE_448x136 "CUBLASLT_MATMUL_TILE_448x136" = 571
593
+ CUBLASLT_MATMUL_TILE_448x144 "CUBLASLT_MATMUL_TILE_448x144" = 572
594
+ CUBLASLT_MATMUL_TILE_512x8 "CUBLASLT_MATMUL_TILE_512x8" = 573
595
+ CUBLASLT_MATMUL_TILE_512x16 "CUBLASLT_MATMUL_TILE_512x16" = 574
596
+ CUBLASLT_MATMUL_TILE_512x24 "CUBLASLT_MATMUL_TILE_512x24" = 575
597
+ CUBLASLT_MATMUL_TILE_512x32 "CUBLASLT_MATMUL_TILE_512x32" = 576
598
+ CUBLASLT_MATMUL_TILE_512x40 "CUBLASLT_MATMUL_TILE_512x40" = 577
599
+ CUBLASLT_MATMUL_TILE_512x48 "CUBLASLT_MATMUL_TILE_512x48" = 578
600
+ CUBLASLT_MATMUL_TILE_512x56 "CUBLASLT_MATMUL_TILE_512x56" = 579
601
+ CUBLASLT_MATMUL_TILE_512x72 "CUBLASLT_MATMUL_TILE_512x72" = 580
602
+ CUBLASLT_MATMUL_TILE_512x80 "CUBLASLT_MATMUL_TILE_512x80" = 581
603
+ CUBLASLT_MATMUL_TILE_512x88 "CUBLASLT_MATMUL_TILE_512x88" = 582
604
+ CUBLASLT_MATMUL_TILE_512x96 "CUBLASLT_MATMUL_TILE_512x96" = 583
605
+ CUBLASLT_MATMUL_TILE_512x104 "CUBLASLT_MATMUL_TILE_512x104" = 584
606
+ CUBLASLT_MATMUL_TILE_512x112 "CUBLASLT_MATMUL_TILE_512x112" = 585
607
+ CUBLASLT_MATMUL_TILE_512x120 "CUBLASLT_MATMUL_TILE_512x120" = 586
608
+ CUBLASLT_MATMUL_TILE_512x128 "CUBLASLT_MATMUL_TILE_512x128" = 587
609
+ CUBLASLT_MATMUL_TILE_576x8 "CUBLASLT_MATMUL_TILE_576x8" = 588
610
+ CUBLASLT_MATMUL_TILE_576x16 "CUBLASLT_MATMUL_TILE_576x16" = 589
611
+ CUBLASLT_MATMUL_TILE_576x24 "CUBLASLT_MATMUL_TILE_576x24" = 590
612
+ CUBLASLT_MATMUL_TILE_576x32 "CUBLASLT_MATMUL_TILE_576x32" = 591
613
+ CUBLASLT_MATMUL_TILE_576x40 "CUBLASLT_MATMUL_TILE_576x40" = 592
614
+ CUBLASLT_MATMUL_TILE_576x48 "CUBLASLT_MATMUL_TILE_576x48" = 593
615
+ CUBLASLT_MATMUL_TILE_576x56 "CUBLASLT_MATMUL_TILE_576x56" = 594
616
+ CUBLASLT_MATMUL_TILE_576x72 "CUBLASLT_MATMUL_TILE_576x72" = 595
617
+ CUBLASLT_MATMUL_TILE_576x80 "CUBLASLT_MATMUL_TILE_576x80" = 596
618
+ CUBLASLT_MATMUL_TILE_576x88 "CUBLASLT_MATMUL_TILE_576x88" = 597
619
+ CUBLASLT_MATMUL_TILE_576x96 "CUBLASLT_MATMUL_TILE_576x96" = 598
620
+ CUBLASLT_MATMUL_TILE_576x104 "CUBLASLT_MATMUL_TILE_576x104" = 599
621
+ CUBLASLT_MATMUL_TILE_576x112 "CUBLASLT_MATMUL_TILE_576x112" = 600
622
+ CUBLASLT_MATMUL_TILE_640x8 "CUBLASLT_MATMUL_TILE_640x8" = 601
623
+ CUBLASLT_MATMUL_TILE_640x16 "CUBLASLT_MATMUL_TILE_640x16" = 602
624
+ CUBLASLT_MATMUL_TILE_640x24 "CUBLASLT_MATMUL_TILE_640x24" = 603
625
+ CUBLASLT_MATMUL_TILE_640x32 "CUBLASLT_MATMUL_TILE_640x32" = 604
626
+ CUBLASLT_MATMUL_TILE_640x40 "CUBLASLT_MATMUL_TILE_640x40" = 605
627
+ CUBLASLT_MATMUL_TILE_640x48 "CUBLASLT_MATMUL_TILE_640x48" = 606
628
+ CUBLASLT_MATMUL_TILE_640x56 "CUBLASLT_MATMUL_TILE_640x56" = 607
629
+ CUBLASLT_MATMUL_TILE_640x72 "CUBLASLT_MATMUL_TILE_640x72" = 608
630
+ CUBLASLT_MATMUL_TILE_640x80 "CUBLASLT_MATMUL_TILE_640x80" = 609
631
+ CUBLASLT_MATMUL_TILE_640x88 "CUBLASLT_MATMUL_TILE_640x88" = 610
632
+ CUBLASLT_MATMUL_TILE_640x96 "CUBLASLT_MATMUL_TILE_640x96" = 611
633
+ CUBLASLT_MATMUL_TILE_704x8 "CUBLASLT_MATMUL_TILE_704x8" = 612
634
+ CUBLASLT_MATMUL_TILE_704x16 "CUBLASLT_MATMUL_TILE_704x16" = 613
635
+ CUBLASLT_MATMUL_TILE_704x24 "CUBLASLT_MATMUL_TILE_704x24" = 614
636
+ CUBLASLT_MATMUL_TILE_704x32 "CUBLASLT_MATMUL_TILE_704x32" = 615
637
+ CUBLASLT_MATMUL_TILE_704x40 "CUBLASLT_MATMUL_TILE_704x40" = 616
638
+ CUBLASLT_MATMUL_TILE_704x48 "CUBLASLT_MATMUL_TILE_704x48" = 617
639
+ CUBLASLT_MATMUL_TILE_704x56 "CUBLASLT_MATMUL_TILE_704x56" = 618
640
+ CUBLASLT_MATMUL_TILE_704x72 "CUBLASLT_MATMUL_TILE_704x72" = 619
641
+ CUBLASLT_MATMUL_TILE_704x80 "CUBLASLT_MATMUL_TILE_704x80" = 620
642
+ CUBLASLT_MATMUL_TILE_704x88 "CUBLASLT_MATMUL_TILE_704x88" = 621
643
+ CUBLASLT_MATMUL_TILE_768x8 "CUBLASLT_MATMUL_TILE_768x8" = 622
644
+ CUBLASLT_MATMUL_TILE_768x16 "CUBLASLT_MATMUL_TILE_768x16" = 623
645
+ CUBLASLT_MATMUL_TILE_768x24 "CUBLASLT_MATMUL_TILE_768x24" = 624
646
+ CUBLASLT_MATMUL_TILE_768x32 "CUBLASLT_MATMUL_TILE_768x32" = 625
647
+ CUBLASLT_MATMUL_TILE_768x40 "CUBLASLT_MATMUL_TILE_768x40" = 626
648
+ CUBLASLT_MATMUL_TILE_768x48 "CUBLASLT_MATMUL_TILE_768x48" = 627
649
+ CUBLASLT_MATMUL_TILE_768x56 "CUBLASLT_MATMUL_TILE_768x56" = 628
650
+ CUBLASLT_MATMUL_TILE_768x72 "CUBLASLT_MATMUL_TILE_768x72" = 629
651
+ CUBLASLT_MATMUL_TILE_768x80 "CUBLASLT_MATMUL_TILE_768x80" = 630
652
+ CUBLASLT_MATMUL_TILE_256x512 "CUBLASLT_MATMUL_TILE_256x512" = 631
653
+ CUBLASLT_MATMUL_TILE_256x1024 "CUBLASLT_MATMUL_TILE_256x1024" = 632
654
+ CUBLASLT_MATMUL_TILE_512x512 "CUBLASLT_MATMUL_TILE_512x512" = 633
655
+ CUBLASLT_MATMUL_TILE_512x1024 "CUBLASLT_MATMUL_TILE_512x1024" = 634
656
+
657
+ ctypedef enum cublasLtMatmulStages_t "cublasLtMatmulStages_t":
658
+ CUBLASLT_MATMUL_STAGES_UNDEFINED "CUBLASLT_MATMUL_STAGES_UNDEFINED" = 0
659
+ CUBLASLT_MATMUL_STAGES_16x1 "CUBLASLT_MATMUL_STAGES_16x1" = 1
660
+ CUBLASLT_MATMUL_STAGES_16x2 "CUBLASLT_MATMUL_STAGES_16x2" = 2
661
+ CUBLASLT_MATMUL_STAGES_16x3 "CUBLASLT_MATMUL_STAGES_16x3" = 3
662
+ CUBLASLT_MATMUL_STAGES_16x4 "CUBLASLT_MATMUL_STAGES_16x4" = 4
663
+ CUBLASLT_MATMUL_STAGES_16x5 "CUBLASLT_MATMUL_STAGES_16x5" = 5
664
+ CUBLASLT_MATMUL_STAGES_16x6 "CUBLASLT_MATMUL_STAGES_16x6" = 6
665
+ CUBLASLT_MATMUL_STAGES_32x1 "CUBLASLT_MATMUL_STAGES_32x1" = 7
666
+ CUBLASLT_MATMUL_STAGES_32x2 "CUBLASLT_MATMUL_STAGES_32x2" = 8
667
+ CUBLASLT_MATMUL_STAGES_32x3 "CUBLASLT_MATMUL_STAGES_32x3" = 9
668
+ CUBLASLT_MATMUL_STAGES_32x4 "CUBLASLT_MATMUL_STAGES_32x4" = 10
669
+ CUBLASLT_MATMUL_STAGES_32x5 "CUBLASLT_MATMUL_STAGES_32x5" = 11
670
+ CUBLASLT_MATMUL_STAGES_32x6 "CUBLASLT_MATMUL_STAGES_32x6" = 12
671
+ CUBLASLT_MATMUL_STAGES_64x1 "CUBLASLT_MATMUL_STAGES_64x1" = 13
672
+ CUBLASLT_MATMUL_STAGES_64x2 "CUBLASLT_MATMUL_STAGES_64x2" = 14
673
+ CUBLASLT_MATMUL_STAGES_64x3 "CUBLASLT_MATMUL_STAGES_64x3" = 15
674
+ CUBLASLT_MATMUL_STAGES_64x4 "CUBLASLT_MATMUL_STAGES_64x4" = 16
675
+ CUBLASLT_MATMUL_STAGES_64x5 "CUBLASLT_MATMUL_STAGES_64x5" = 17
676
+ CUBLASLT_MATMUL_STAGES_64x6 "CUBLASLT_MATMUL_STAGES_64x6" = 18
677
+ CUBLASLT_MATMUL_STAGES_128x1 "CUBLASLT_MATMUL_STAGES_128x1" = 19
678
+ CUBLASLT_MATMUL_STAGES_128x2 "CUBLASLT_MATMUL_STAGES_128x2" = 20
679
+ CUBLASLT_MATMUL_STAGES_128x3 "CUBLASLT_MATMUL_STAGES_128x3" = 21
680
+ CUBLASLT_MATMUL_STAGES_128x4 "CUBLASLT_MATMUL_STAGES_128x4" = 22
681
+ CUBLASLT_MATMUL_STAGES_128x5 "CUBLASLT_MATMUL_STAGES_128x5" = 23
682
+ CUBLASLT_MATMUL_STAGES_128x6 "CUBLASLT_MATMUL_STAGES_128x6" = 24
683
+ CUBLASLT_MATMUL_STAGES_32x10 "CUBLASLT_MATMUL_STAGES_32x10" = 25
684
+ CUBLASLT_MATMUL_STAGES_8x4 "CUBLASLT_MATMUL_STAGES_8x4" = 26
685
+ CUBLASLT_MATMUL_STAGES_16x10 "CUBLASLT_MATMUL_STAGES_16x10" = 27
686
+ CUBLASLT_MATMUL_STAGES_8x5 "CUBLASLT_MATMUL_STAGES_8x5" = 28
687
+ CUBLASLT_MATMUL_STAGES_8x3 "CUBLASLT_MATMUL_STAGES_8x3" = 31
688
+ CUBLASLT_MATMUL_STAGES_8xAUTO "CUBLASLT_MATMUL_STAGES_8xAUTO" = 32
689
+ CUBLASLT_MATMUL_STAGES_16xAUTO "CUBLASLT_MATMUL_STAGES_16xAUTO" = 33
690
+ CUBLASLT_MATMUL_STAGES_32xAUTO "CUBLASLT_MATMUL_STAGES_32xAUTO" = 34
691
+ CUBLASLT_MATMUL_STAGES_64xAUTO "CUBLASLT_MATMUL_STAGES_64xAUTO" = 35
692
+ CUBLASLT_MATMUL_STAGES_128xAUTO "CUBLASLT_MATMUL_STAGES_128xAUTO" = 36
693
+ CUBLASLT_MATMUL_STAGES_256xAUTO "CUBLASLT_MATMUL_STAGES_256xAUTO" = 37
694
+ CUBLASLT_MATMUL_STAGES_768xAUTO "CUBLASLT_MATMUL_STAGES_768xAUTO" = 38
695
+ CUBLASLT_MATMUL_STAGES_16x80 "CUBLASLT_MATMUL_STAGES_16x80" = 29
696
+ CUBLASLT_MATMUL_STAGES_64x80 "CUBLASLT_MATMUL_STAGES_64x80" = 30
697
+
698
+ ctypedef enum cublasLtPointerMode_t "cublasLtPointerMode_t":
699
+ CUBLASLT_POINTER_MODE_HOST "CUBLASLT_POINTER_MODE_HOST" = CUBLAS_POINTER_MODE_HOST
700
+ CUBLASLT_POINTER_MODE_DEVICE "CUBLASLT_POINTER_MODE_DEVICE" = CUBLAS_POINTER_MODE_DEVICE
701
+ CUBLASLT_POINTER_MODE_DEVICE_VECTOR "CUBLASLT_POINTER_MODE_DEVICE_VECTOR" = 2
702
+ CUBLASLT_POINTER_MODE_ALPHA_DEVICE_VECTOR_BETA_ZERO "CUBLASLT_POINTER_MODE_ALPHA_DEVICE_VECTOR_BETA_ZERO" = 3
703
+ CUBLASLT_POINTER_MODE_ALPHA_DEVICE_VECTOR_BETA_HOST "CUBLASLT_POINTER_MODE_ALPHA_DEVICE_VECTOR_BETA_HOST" = 4
704
+
705
+ ctypedef enum cublasLtPointerModeMask_t "cublasLtPointerModeMask_t":
706
+ CUBLASLT_POINTER_MODE_MASK_HOST "CUBLASLT_POINTER_MODE_MASK_HOST" = 1
707
+ CUBLASLT_POINTER_MODE_MASK_DEVICE "CUBLASLT_POINTER_MODE_MASK_DEVICE" = 2
708
+ CUBLASLT_POINTER_MODE_MASK_DEVICE_VECTOR "CUBLASLT_POINTER_MODE_MASK_DEVICE_VECTOR" = 4
709
+ CUBLASLT_POINTER_MODE_MASK_ALPHA_DEVICE_VECTOR_BETA_ZERO "CUBLASLT_POINTER_MODE_MASK_ALPHA_DEVICE_VECTOR_BETA_ZERO" = 8
710
+ CUBLASLT_POINTER_MODE_MASK_ALPHA_DEVICE_VECTOR_BETA_HOST "CUBLASLT_POINTER_MODE_MASK_ALPHA_DEVICE_VECTOR_BETA_HOST" = 16
711
+ CUBLASLT_POINTER_MODE_MASK_NO_FILTERING "CUBLASLT_POINTER_MODE_MASK_NO_FILTERING" = 0
712
+
713
+ ctypedef enum cublasLtOrder_t "cublasLtOrder_t":
714
+ CUBLASLT_ORDER_COL "CUBLASLT_ORDER_COL" = 0
715
+ CUBLASLT_ORDER_ROW "CUBLASLT_ORDER_ROW" = 1
716
+ CUBLASLT_ORDER_COL32 "CUBLASLT_ORDER_COL32" = 2
717
+ CUBLASLT_ORDER_COL4_4R2_8C "CUBLASLT_ORDER_COL4_4R2_8C" = 3
718
+ CUBLASLT_ORDER_COL32_2R_4R4 "CUBLASLT_ORDER_COL32_2R_4R4" = 4
719
+
720
+ ctypedef enum cublasLtMatrixLayoutAttribute_t "cublasLtMatrixLayoutAttribute_t":
721
+ CUBLASLT_MATRIX_LAYOUT_TYPE "CUBLASLT_MATRIX_LAYOUT_TYPE" = 0
722
+ CUBLASLT_MATRIX_LAYOUT_ORDER "CUBLASLT_MATRIX_LAYOUT_ORDER" = 1
723
+ CUBLASLT_MATRIX_LAYOUT_ROWS "CUBLASLT_MATRIX_LAYOUT_ROWS" = 2
724
+ CUBLASLT_MATRIX_LAYOUT_COLS "CUBLASLT_MATRIX_LAYOUT_COLS" = 3
725
+ CUBLASLT_MATRIX_LAYOUT_LD "CUBLASLT_MATRIX_LAYOUT_LD" = 4
726
+ CUBLASLT_MATRIX_LAYOUT_BATCH_COUNT "CUBLASLT_MATRIX_LAYOUT_BATCH_COUNT" = 5
727
+ CUBLASLT_MATRIX_LAYOUT_STRIDED_BATCH_OFFSET "CUBLASLT_MATRIX_LAYOUT_STRIDED_BATCH_OFFSET" = 6
728
+ CUBLASLT_MATRIX_LAYOUT_PLANE_OFFSET "CUBLASLT_MATRIX_LAYOUT_PLANE_OFFSET" = 7
729
+ CUBLASLT_MATRIX_LAYOUT_BATCH_MODE "CUBLASLT_MATRIX_LAYOUT_BATCH_MODE" = 8
730
+ CUBLASLT_GROUPED_MATRIX_LAYOUT_ROWS_ARRAY "CUBLASLT_GROUPED_MATRIX_LAYOUT_ROWS_ARRAY" = 9
731
+ CUBLASLT_GROUPED_MATRIX_LAYOUT_COLS_ARRAY "CUBLASLT_GROUPED_MATRIX_LAYOUT_COLS_ARRAY" = 10
732
+ CUBLASLT_GROUPED_MATRIX_LAYOUT_LD_ARRAY "CUBLASLT_GROUPED_MATRIX_LAYOUT_LD_ARRAY" = 11
733
+ CUBLASLT_GROUPED_MATRIX_LAYOUT_ROWS_COLS_ARRAY_INTEGER_WIDTH "CUBLASLT_GROUPED_MATRIX_LAYOUT_ROWS_COLS_ARRAY_INTEGER_WIDTH" = 12
734
+ CUBLASLT_GROUPED_MATRIX_LAYOUT_LD_ARRAY_INTEGER_WIDTH "CUBLASLT_GROUPED_MATRIX_LAYOUT_LD_ARRAY_INTEGER_WIDTH" = 13
735
+
736
+ ctypedef enum cublasLtMatmulDescAttributes_t "cublasLtMatmulDescAttributes_t":
737
+ CUBLASLT_MATMUL_DESC_COMPUTE_TYPE "CUBLASLT_MATMUL_DESC_COMPUTE_TYPE" = 0
738
+ CUBLASLT_MATMUL_DESC_SCALE_TYPE "CUBLASLT_MATMUL_DESC_SCALE_TYPE" = 1
739
+ CUBLASLT_MATMUL_DESC_POINTER_MODE "CUBLASLT_MATMUL_DESC_POINTER_MODE" = 2
740
+ CUBLASLT_MATMUL_DESC_TRANSA "CUBLASLT_MATMUL_DESC_TRANSA" = 3
741
+ CUBLASLT_MATMUL_DESC_TRANSB "CUBLASLT_MATMUL_DESC_TRANSB" = 4
742
+ CUBLASLT_MATMUL_DESC_TRANSC "CUBLASLT_MATMUL_DESC_TRANSC" = 5
743
+ CUBLASLT_MATMUL_DESC_FILL_MODE "CUBLASLT_MATMUL_DESC_FILL_MODE" = 6
744
+ CUBLASLT_MATMUL_DESC_EPILOGUE "CUBLASLT_MATMUL_DESC_EPILOGUE" = 7
745
+ CUBLASLT_MATMUL_DESC_BIAS_POINTER "CUBLASLT_MATMUL_DESC_BIAS_POINTER" = 8
746
+ CUBLASLT_MATMUL_DESC_BIAS_BATCH_STRIDE "CUBLASLT_MATMUL_DESC_BIAS_BATCH_STRIDE" = 10
747
+ CUBLASLT_MATMUL_DESC_EPILOGUE_AUX_POINTER "CUBLASLT_MATMUL_DESC_EPILOGUE_AUX_POINTER" = 11
748
+ CUBLASLT_MATMUL_DESC_EPILOGUE_AUX_LD "CUBLASLT_MATMUL_DESC_EPILOGUE_AUX_LD" = 12
749
+ CUBLASLT_MATMUL_DESC_EPILOGUE_AUX_BATCH_STRIDE "CUBLASLT_MATMUL_DESC_EPILOGUE_AUX_BATCH_STRIDE" = 13
750
+ CUBLASLT_MATMUL_DESC_ALPHA_VECTOR_BATCH_STRIDE "CUBLASLT_MATMUL_DESC_ALPHA_VECTOR_BATCH_STRIDE" = 14
751
+ CUBLASLT_MATMUL_DESC_SM_COUNT_TARGET "CUBLASLT_MATMUL_DESC_SM_COUNT_TARGET" = 15
752
+ CUBLASLT_MATMUL_DESC_A_SCALE_POINTER "CUBLASLT_MATMUL_DESC_A_SCALE_POINTER" = 17
753
+ CUBLASLT_MATMUL_DESC_B_SCALE_POINTER "CUBLASLT_MATMUL_DESC_B_SCALE_POINTER" = 18
754
+ CUBLASLT_MATMUL_DESC_C_SCALE_POINTER "CUBLASLT_MATMUL_DESC_C_SCALE_POINTER" = 19
755
+ CUBLASLT_MATMUL_DESC_D_SCALE_POINTER "CUBLASLT_MATMUL_DESC_D_SCALE_POINTER" = 20
756
+ CUBLASLT_MATMUL_DESC_AMAX_D_POINTER "CUBLASLT_MATMUL_DESC_AMAX_D_POINTER" = 21
757
+ CUBLASLT_MATMUL_DESC_EPILOGUE_AUX_DATA_TYPE "CUBLASLT_MATMUL_DESC_EPILOGUE_AUX_DATA_TYPE" = 22
758
+ CUBLASLT_MATMUL_DESC_EPILOGUE_AUX_SCALE_POINTER "CUBLASLT_MATMUL_DESC_EPILOGUE_AUX_SCALE_POINTER" = 23
759
+ CUBLASLT_MATMUL_DESC_EPILOGUE_AUX_AMAX_POINTER "CUBLASLT_MATMUL_DESC_EPILOGUE_AUX_AMAX_POINTER" = 24
760
+ CUBLASLT_MATMUL_DESC_FAST_ACCUM "CUBLASLT_MATMUL_DESC_FAST_ACCUM" = 25
761
+ CUBLASLT_MATMUL_DESC_BIAS_DATA_TYPE "CUBLASLT_MATMUL_DESC_BIAS_DATA_TYPE" = 26
762
+ CUBLASLT_MATMUL_DESC_A_SCALE_MODE "CUBLASLT_MATMUL_DESC_A_SCALE_MODE" = 31
763
+ CUBLASLT_MATMUL_DESC_B_SCALE_MODE "CUBLASLT_MATMUL_DESC_B_SCALE_MODE" = 32
764
+ CUBLASLT_MATMUL_DESC_C_SCALE_MODE "CUBLASLT_MATMUL_DESC_C_SCALE_MODE" = 33
765
+ CUBLASLT_MATMUL_DESC_D_SCALE_MODE "CUBLASLT_MATMUL_DESC_D_SCALE_MODE" = 34
766
+ CUBLASLT_MATMUL_DESC_EPILOGUE_AUX_SCALE_MODE "CUBLASLT_MATMUL_DESC_EPILOGUE_AUX_SCALE_MODE" = 35
767
+ CUBLASLT_MATMUL_DESC_D_OUT_SCALE_POINTER "CUBLASLT_MATMUL_DESC_D_OUT_SCALE_POINTER" = 36
768
+ CUBLASLT_MATMUL_DESC_D_OUT_SCALE_MODE "CUBLASLT_MATMUL_DESC_D_OUT_SCALE_MODE" = 37
769
+ CUBLASLT_MATMUL_DESC_EMULATION_DESCRIPTOR "CUBLASLT_MATMUL_DESC_EMULATION_DESCRIPTOR" = 38
770
+ CUBLASLT_MATMUL_DESC_ALPHA_BATCH_STRIDE "CUBLASLT_MATMUL_DESC_ALPHA_BATCH_STRIDE" = 39
771
+ CUBLASLT_MATMUL_DESC_BETA_BATCH_STRIDE "CUBLASLT_MATMUL_DESC_BETA_BATCH_STRIDE" = 40
772
+ CUBLASLT_MATMUL_DESC_ATOMIC_SYNC_NUM_CHUNKS_D_ROWS "CUBLASLT_MATMUL_DESC_ATOMIC_SYNC_NUM_CHUNKS_D_ROWS" = 27
773
+ CUBLASLT_MATMUL_DESC_ATOMIC_SYNC_NUM_CHUNKS_D_COLS "CUBLASLT_MATMUL_DESC_ATOMIC_SYNC_NUM_CHUNKS_D_COLS" = 28
774
+ CUBLASLT_MATMUL_DESC_ATOMIC_SYNC_IN_COUNTERS_POINTER "CUBLASLT_MATMUL_DESC_ATOMIC_SYNC_IN_COUNTERS_POINTER" = 29
775
+ CUBLASLT_MATMUL_DESC_ATOMIC_SYNC_OUT_COUNTERS_POINTER "CUBLASLT_MATMUL_DESC_ATOMIC_SYNC_OUT_COUNTERS_POINTER" = 30
776
+
777
+ ctypedef enum cublasLtMatrixTransformDescAttributes_t "cublasLtMatrixTransformDescAttributes_t":
778
+ CUBLASLT_MATRIX_TRANSFORM_DESC_SCALE_TYPE "CUBLASLT_MATRIX_TRANSFORM_DESC_SCALE_TYPE"
779
+ CUBLASLT_MATRIX_TRANSFORM_DESC_POINTER_MODE "CUBLASLT_MATRIX_TRANSFORM_DESC_POINTER_MODE"
780
+ CUBLASLT_MATRIX_TRANSFORM_DESC_TRANSA "CUBLASLT_MATRIX_TRANSFORM_DESC_TRANSA"
781
+ CUBLASLT_MATRIX_TRANSFORM_DESC_TRANSB "CUBLASLT_MATRIX_TRANSFORM_DESC_TRANSB"
782
+
783
+ ctypedef enum cublasLtReductionScheme_t "cublasLtReductionScheme_t":
784
+ CUBLASLT_REDUCTION_SCHEME_NONE "CUBLASLT_REDUCTION_SCHEME_NONE" = 0
785
+ CUBLASLT_REDUCTION_SCHEME_INPLACE "CUBLASLT_REDUCTION_SCHEME_INPLACE" = 1
786
+ CUBLASLT_REDUCTION_SCHEME_COMPUTE_TYPE "CUBLASLT_REDUCTION_SCHEME_COMPUTE_TYPE" = 2
787
+ CUBLASLT_REDUCTION_SCHEME_OUTPUT_TYPE "CUBLASLT_REDUCTION_SCHEME_OUTPUT_TYPE" = 4
788
+ CUBLASLT_REDUCTION_SCHEME_MASK "CUBLASLT_REDUCTION_SCHEME_MASK" = 0x7
789
+
790
+ ctypedef enum cublasLtEpilogue_t "cublasLtEpilogue_t":
791
+ CUBLASLT_EPILOGUE_DEFAULT "CUBLASLT_EPILOGUE_DEFAULT" = 1
792
+ CUBLASLT_EPILOGUE_RELU "CUBLASLT_EPILOGUE_RELU" = 2
793
+ CUBLASLT_EPILOGUE_RELU_AUX "CUBLASLT_EPILOGUE_RELU_AUX" = (CUBLASLT_EPILOGUE_RELU | 128)
794
+ CUBLASLT_EPILOGUE_BIAS "CUBLASLT_EPILOGUE_BIAS" = 4
795
+ CUBLASLT_EPILOGUE_RELU_BIAS "CUBLASLT_EPILOGUE_RELU_BIAS" = (CUBLASLT_EPILOGUE_RELU | CUBLASLT_EPILOGUE_BIAS)
796
+ CUBLASLT_EPILOGUE_RELU_AUX_BIAS "CUBLASLT_EPILOGUE_RELU_AUX_BIAS" = (CUBLASLT_EPILOGUE_RELU_AUX | CUBLASLT_EPILOGUE_BIAS)
797
+ CUBLASLT_EPILOGUE_DRELU "CUBLASLT_EPILOGUE_DRELU" = (8 | 128)
798
+ CUBLASLT_EPILOGUE_DRELU_BGRAD "CUBLASLT_EPILOGUE_DRELU_BGRAD" = (CUBLASLT_EPILOGUE_DRELU | 16)
799
+ CUBLASLT_EPILOGUE_GELU "CUBLASLT_EPILOGUE_GELU" = 32
800
+ CUBLASLT_EPILOGUE_GELU_AUX "CUBLASLT_EPILOGUE_GELU_AUX" = (CUBLASLT_EPILOGUE_GELU | 128)
801
+ CUBLASLT_EPILOGUE_GELU_BIAS "CUBLASLT_EPILOGUE_GELU_BIAS" = (CUBLASLT_EPILOGUE_GELU | CUBLASLT_EPILOGUE_BIAS)
802
+ CUBLASLT_EPILOGUE_GELU_AUX_BIAS "CUBLASLT_EPILOGUE_GELU_AUX_BIAS" = (CUBLASLT_EPILOGUE_GELU_AUX | CUBLASLT_EPILOGUE_BIAS)
803
+ CUBLASLT_EPILOGUE_DGELU "CUBLASLT_EPILOGUE_DGELU" = (64 | 128)
804
+ CUBLASLT_EPILOGUE_DGELU_BGRAD "CUBLASLT_EPILOGUE_DGELU_BGRAD" = (CUBLASLT_EPILOGUE_DGELU | 16)
805
+ CUBLASLT_EPILOGUE_BGRADA "CUBLASLT_EPILOGUE_BGRADA" = 256
806
+ CUBLASLT_EPILOGUE_BGRADB "CUBLASLT_EPILOGUE_BGRADB" = 512
807
+
808
+ ctypedef enum cublasLtMatmulSearch_t "cublasLtMatmulSearch_t":
809
+ CUBLASLT_SEARCH_BEST_FIT "CUBLASLT_SEARCH_BEST_FIT" = 0
810
+ CUBLASLT_SEARCH_LIMITED_BY_ALGO_ID "CUBLASLT_SEARCH_LIMITED_BY_ALGO_ID" = 1
811
+ CUBLASLT_SEARCH_RESERVED_02 "CUBLASLT_SEARCH_RESERVED_02" = 2
812
+ CUBLASLT_SEARCH_RESERVED_03 "CUBLASLT_SEARCH_RESERVED_03" = 3
813
+ CUBLASLT_SEARCH_RESERVED_04 "CUBLASLT_SEARCH_RESERVED_04" = 4
814
+ CUBLASLT_SEARCH_RESERVED_05 "CUBLASLT_SEARCH_RESERVED_05" = 5
815
+ CUBLASLT_SEARCH_RESERVED_06 "CUBLASLT_SEARCH_RESERVED_06" = 6
816
+ CUBLASLT_SEARCH_RESERVED_07 "CUBLASLT_SEARCH_RESERVED_07" = 7
817
+ CUBLASLT_SEARCH_RESERVED_08 "CUBLASLT_SEARCH_RESERVED_08" = 8
818
+ CUBLASLT_SEARCH_RESERVED_09 "CUBLASLT_SEARCH_RESERVED_09" = 9
819
+
820
+ ctypedef enum cublasLtMatmulPreferenceAttributes_t "cublasLtMatmulPreferenceAttributes_t":
821
+ CUBLASLT_MATMUL_PREF_SEARCH_MODE "CUBLASLT_MATMUL_PREF_SEARCH_MODE" = 0
822
+ CUBLASLT_MATMUL_PREF_MAX_WORKSPACE_BYTES "CUBLASLT_MATMUL_PREF_MAX_WORKSPACE_BYTES" = 1
823
+ CUBLASLT_MATMUL_PREF_REDUCTION_SCHEME_MASK "CUBLASLT_MATMUL_PREF_REDUCTION_SCHEME_MASK" = 3
824
+ CUBLASLT_MATMUL_PREF_MIN_ALIGNMENT_A_BYTES "CUBLASLT_MATMUL_PREF_MIN_ALIGNMENT_A_BYTES" = 5
825
+ CUBLASLT_MATMUL_PREF_MIN_ALIGNMENT_B_BYTES "CUBLASLT_MATMUL_PREF_MIN_ALIGNMENT_B_BYTES" = 6
826
+ CUBLASLT_MATMUL_PREF_MIN_ALIGNMENT_C_BYTES "CUBLASLT_MATMUL_PREF_MIN_ALIGNMENT_C_BYTES" = 7
827
+ CUBLASLT_MATMUL_PREF_MIN_ALIGNMENT_D_BYTES "CUBLASLT_MATMUL_PREF_MIN_ALIGNMENT_D_BYTES" = 8
828
+ CUBLASLT_MATMUL_PREF_MAX_WAVES_COUNT "CUBLASLT_MATMUL_PREF_MAX_WAVES_COUNT" = 9
829
+ CUBLASLT_MATMUL_PREF_IMPL_MASK "CUBLASLT_MATMUL_PREF_IMPL_MASK" = 12
830
+ CUBLASLT_MATMUL_PREF_GROUPED_AVERAGE_REDUCTION_DIM "CUBLASLT_MATMUL_PREF_GROUPED_AVERAGE_REDUCTION_DIM" = 13
831
+ CUBLASLT_MATMUL_PREF_GROUPED_DESC_D_AVERAGE_ROWS "CUBLASLT_MATMUL_PREF_GROUPED_DESC_D_AVERAGE_ROWS" = 14
832
+ CUBLASLT_MATMUL_PREF_GROUPED_DESC_D_AVERAGE_COLS "CUBLASLT_MATMUL_PREF_GROUPED_DESC_D_AVERAGE_COLS" = 15
833
+ CUBLASLT_MATMUL_PREF_MATH_MODE_MASK "CUBLASLT_MATMUL_PREF_MATH_MODE_MASK" = 2
834
+ CUBLASLT_MATMUL_PREF_GAUSSIAN_MODE_MASK "CUBLASLT_MATMUL_PREF_GAUSSIAN_MODE_MASK" = 4
835
+ CUBLASLT_MATMUL_PREF_POINTER_MODE_MASK "CUBLASLT_MATMUL_PREF_POINTER_MODE_MASK" = 10
836
+ CUBLASLT_MATMUL_PREF_EPILOGUE_MASK "CUBLASLT_MATMUL_PREF_EPILOGUE_MASK" = 11
837
+ CUBLASLT_MATMUL_PREF_SM_COUNT_TARGET "CUBLASLT_MATMUL_PREF_SM_COUNT_TARGET" = 13
838
+
839
+ ctypedef enum cublasLtMatmulAlgoCapAttributes_t "cublasLtMatmulAlgoCapAttributes_t":
840
+ CUBLASLT_ALGO_CAP_SPLITK_SUPPORT "CUBLASLT_ALGO_CAP_SPLITK_SUPPORT" = 0
841
+ CUBLASLT_ALGO_CAP_REDUCTION_SCHEME_MASK "CUBLASLT_ALGO_CAP_REDUCTION_SCHEME_MASK" = 1
842
+ CUBLASLT_ALGO_CAP_CTA_SWIZZLING_SUPPORT "CUBLASLT_ALGO_CAP_CTA_SWIZZLING_SUPPORT" = 2
843
+ CUBLASLT_ALGO_CAP_STRIDED_BATCH_SUPPORT "CUBLASLT_ALGO_CAP_STRIDED_BATCH_SUPPORT" = 3
844
+ CUBLASLT_ALGO_CAP_OUT_OF_PLACE_RESULT_SUPPORT "CUBLASLT_ALGO_CAP_OUT_OF_PLACE_RESULT_SUPPORT" = 4
845
+ CUBLASLT_ALGO_CAP_UPLO_SUPPORT "CUBLASLT_ALGO_CAP_UPLO_SUPPORT" = 5
846
+ CUBLASLT_ALGO_CAP_TILE_IDS "CUBLASLT_ALGO_CAP_TILE_IDS" = 6
847
+ CUBLASLT_ALGO_CAP_CUSTOM_OPTION_MAX "CUBLASLT_ALGO_CAP_CUSTOM_OPTION_MAX" = 7
848
+ CUBLASLT_ALGO_CAP_CUSTOM_MEMORY_ORDER "CUBLASLT_ALGO_CAP_CUSTOM_MEMORY_ORDER" = 10
849
+ CUBLASLT_ALGO_CAP_POINTER_MODE_MASK "CUBLASLT_ALGO_CAP_POINTER_MODE_MASK" = 11
850
+ CUBLASLT_ALGO_CAP_EPILOGUE_MASK "CUBLASLT_ALGO_CAP_EPILOGUE_MASK" = 12
851
+ CUBLASLT_ALGO_CAP_STAGES_IDS "CUBLASLT_ALGO_CAP_STAGES_IDS" = 13
852
+ CUBLASLT_ALGO_CAP_LD_NEGATIVE "CUBLASLT_ALGO_CAP_LD_NEGATIVE" = 14
853
+ CUBLASLT_ALGO_CAP_NUMERICAL_IMPL_FLAGS "CUBLASLT_ALGO_CAP_NUMERICAL_IMPL_FLAGS" = 15
854
+ CUBLASLT_ALGO_CAP_MIN_ALIGNMENT_A_BYTES "CUBLASLT_ALGO_CAP_MIN_ALIGNMENT_A_BYTES" = 16
855
+ CUBLASLT_ALGO_CAP_MIN_ALIGNMENT_B_BYTES "CUBLASLT_ALGO_CAP_MIN_ALIGNMENT_B_BYTES" = 17
856
+ CUBLASLT_ALGO_CAP_MIN_ALIGNMENT_C_BYTES "CUBLASLT_ALGO_CAP_MIN_ALIGNMENT_C_BYTES" = 18
857
+ CUBLASLT_ALGO_CAP_MIN_ALIGNMENT_D_BYTES "CUBLASLT_ALGO_CAP_MIN_ALIGNMENT_D_BYTES" = 19
858
+ CUBLASLT_ALGO_CAP_POINTER_ARRAY_BATCH_SUPPORT "CUBLASLT_ALGO_CAP_POINTER_ARRAY_BATCH_SUPPORT" = 21
859
+ CUBLASLT_ALGO_CAP_FLOATING_POINT_EMULATION_SUPPORT "CUBLASLT_ALGO_CAP_FLOATING_POINT_EMULATION_SUPPORT" = 22
860
+ CUBLASLT_ALGO_CAP_POINTER_ARRAY_GROUPED_SUPPORT "CUBLASLT_ALGO_CAP_POINTER_ARRAY_GROUPED_SUPPORT" = 23
861
+ CUBLASLT_ALGO_CAP_ATOMIC_SYNC "CUBLASLT_ALGO_CAP_ATOMIC_SYNC" = 20
862
+ CUBLASLT_ALGO_CAP_MATHMODE_IMPL "CUBLASLT_ALGO_CAP_MATHMODE_IMPL" = 8
863
+ CUBLASLT_ALGO_CAP_GAUSSIAN_IMPL "CUBLASLT_ALGO_CAP_GAUSSIAN_IMPL" = 9
864
+
865
+ ctypedef enum cublasLtMatmulAlgoConfigAttributes_t "cublasLtMatmulAlgoConfigAttributes_t":
866
+ CUBLASLT_ALGO_CONFIG_ID "CUBLASLT_ALGO_CONFIG_ID" = 0
867
+ CUBLASLT_ALGO_CONFIG_TILE_ID "CUBLASLT_ALGO_CONFIG_TILE_ID" = 1
868
+ CUBLASLT_ALGO_CONFIG_SPLITK_NUM "CUBLASLT_ALGO_CONFIG_SPLITK_NUM" = 2
869
+ CUBLASLT_ALGO_CONFIG_REDUCTION_SCHEME "CUBLASLT_ALGO_CONFIG_REDUCTION_SCHEME" = 3
870
+ CUBLASLT_ALGO_CONFIG_CTA_SWIZZLING "CUBLASLT_ALGO_CONFIG_CTA_SWIZZLING" = 4
871
+ CUBLASLT_ALGO_CONFIG_CUSTOM_OPTION "CUBLASLT_ALGO_CONFIG_CUSTOM_OPTION" = 5
872
+ CUBLASLT_ALGO_CONFIG_STAGES_ID "CUBLASLT_ALGO_CONFIG_STAGES_ID" = 6
873
+ CUBLASLT_ALGO_CONFIG_INNER_SHAPE_ID "CUBLASLT_ALGO_CONFIG_INNER_SHAPE_ID" = 7
874
+ CUBLASLT_ALGO_CONFIG_CLUSTER_SHAPE_ID "CUBLASLT_ALGO_CONFIG_CLUSTER_SHAPE_ID" = 8
875
+
876
+ ctypedef enum cublasLtClusterShape_t "cublasLtClusterShape_t":
877
+ CUBLASLT_CLUSTER_SHAPE_AUTO "CUBLASLT_CLUSTER_SHAPE_AUTO" = 0
878
+ CUBLASLT_CLUSTER_SHAPE_1x1x1 "CUBLASLT_CLUSTER_SHAPE_1x1x1" = 2
879
+ CUBLASLT_CLUSTER_SHAPE_2x1x1 "CUBLASLT_CLUSTER_SHAPE_2x1x1" = 3
880
+ CUBLASLT_CLUSTER_SHAPE_4x1x1 "CUBLASLT_CLUSTER_SHAPE_4x1x1" = 4
881
+ CUBLASLT_CLUSTER_SHAPE_1x2x1 "CUBLASLT_CLUSTER_SHAPE_1x2x1" = 5
882
+ CUBLASLT_CLUSTER_SHAPE_2x2x1 "CUBLASLT_CLUSTER_SHAPE_2x2x1" = 6
883
+ CUBLASLT_CLUSTER_SHAPE_4x2x1 "CUBLASLT_CLUSTER_SHAPE_4x2x1" = 7
884
+ CUBLASLT_CLUSTER_SHAPE_1x4x1 "CUBLASLT_CLUSTER_SHAPE_1x4x1" = 8
885
+ CUBLASLT_CLUSTER_SHAPE_2x4x1 "CUBLASLT_CLUSTER_SHAPE_2x4x1" = 9
886
+ CUBLASLT_CLUSTER_SHAPE_4x4x1 "CUBLASLT_CLUSTER_SHAPE_4x4x1" = 10
887
+ CUBLASLT_CLUSTER_SHAPE_8x1x1 "CUBLASLT_CLUSTER_SHAPE_8x1x1" = 11
888
+ CUBLASLT_CLUSTER_SHAPE_1x8x1 "CUBLASLT_CLUSTER_SHAPE_1x8x1" = 12
889
+ CUBLASLT_CLUSTER_SHAPE_8x2x1 "CUBLASLT_CLUSTER_SHAPE_8x2x1" = 13
890
+ CUBLASLT_CLUSTER_SHAPE_2x8x1 "CUBLASLT_CLUSTER_SHAPE_2x8x1" = 14
891
+ CUBLASLT_CLUSTER_SHAPE_16x1x1 "CUBLASLT_CLUSTER_SHAPE_16x1x1" = 15
892
+ CUBLASLT_CLUSTER_SHAPE_1x16x1 "CUBLASLT_CLUSTER_SHAPE_1x16x1" = 16
893
+ CUBLASLT_CLUSTER_SHAPE_3x1x1 "CUBLASLT_CLUSTER_SHAPE_3x1x1" = 17
894
+ CUBLASLT_CLUSTER_SHAPE_5x1x1 "CUBLASLT_CLUSTER_SHAPE_5x1x1" = 18
895
+ CUBLASLT_CLUSTER_SHAPE_6x1x1 "CUBLASLT_CLUSTER_SHAPE_6x1x1" = 19
896
+ CUBLASLT_CLUSTER_SHAPE_7x1x1 "CUBLASLT_CLUSTER_SHAPE_7x1x1" = 20
897
+ CUBLASLT_CLUSTER_SHAPE_9x1x1 "CUBLASLT_CLUSTER_SHAPE_9x1x1" = 21
898
+ CUBLASLT_CLUSTER_SHAPE_10x1x1 "CUBLASLT_CLUSTER_SHAPE_10x1x1" = 22
899
+ CUBLASLT_CLUSTER_SHAPE_11x1x1 "CUBLASLT_CLUSTER_SHAPE_11x1x1" = 23
900
+ CUBLASLT_CLUSTER_SHAPE_12x1x1 "CUBLASLT_CLUSTER_SHAPE_12x1x1" = 24
901
+ CUBLASLT_CLUSTER_SHAPE_13x1x1 "CUBLASLT_CLUSTER_SHAPE_13x1x1" = 25
902
+ CUBLASLT_CLUSTER_SHAPE_14x1x1 "CUBLASLT_CLUSTER_SHAPE_14x1x1" = 26
903
+ CUBLASLT_CLUSTER_SHAPE_15x1x1 "CUBLASLT_CLUSTER_SHAPE_15x1x1" = 27
904
+ CUBLASLT_CLUSTER_SHAPE_3x2x1 "CUBLASLT_CLUSTER_SHAPE_3x2x1" = 28
905
+ CUBLASLT_CLUSTER_SHAPE_5x2x1 "CUBLASLT_CLUSTER_SHAPE_5x2x1" = 29
906
+ CUBLASLT_CLUSTER_SHAPE_6x2x1 "CUBLASLT_CLUSTER_SHAPE_6x2x1" = 30
907
+ CUBLASLT_CLUSTER_SHAPE_7x2x1 "CUBLASLT_CLUSTER_SHAPE_7x2x1" = 31
908
+ CUBLASLT_CLUSTER_SHAPE_1x3x1 "CUBLASLT_CLUSTER_SHAPE_1x3x1" = 32
909
+ CUBLASLT_CLUSTER_SHAPE_2x3x1 "CUBLASLT_CLUSTER_SHAPE_2x3x1" = 33
910
+ CUBLASLT_CLUSTER_SHAPE_3x3x1 "CUBLASLT_CLUSTER_SHAPE_3x3x1" = 34
911
+ CUBLASLT_CLUSTER_SHAPE_4x3x1 "CUBLASLT_CLUSTER_SHAPE_4x3x1" = 35
912
+ CUBLASLT_CLUSTER_SHAPE_5x3x1 "CUBLASLT_CLUSTER_SHAPE_5x3x1" = 36
913
+ CUBLASLT_CLUSTER_SHAPE_3x4x1 "CUBLASLT_CLUSTER_SHAPE_3x4x1" = 37
914
+ CUBLASLT_CLUSTER_SHAPE_1x5x1 "CUBLASLT_CLUSTER_SHAPE_1x5x1" = 38
915
+ CUBLASLT_CLUSTER_SHAPE_2x5x1 "CUBLASLT_CLUSTER_SHAPE_2x5x1" = 39
916
+ CUBLASLT_CLUSTER_SHAPE_3x5x1 "CUBLASLT_CLUSTER_SHAPE_3x5x1" = 40
917
+ CUBLASLT_CLUSTER_SHAPE_1x6x1 "CUBLASLT_CLUSTER_SHAPE_1x6x1" = 41
918
+ CUBLASLT_CLUSTER_SHAPE_2x6x1 "CUBLASLT_CLUSTER_SHAPE_2x6x1" = 42
919
+ CUBLASLT_CLUSTER_SHAPE_1x7x1 "CUBLASLT_CLUSTER_SHAPE_1x7x1" = 43
920
+ CUBLASLT_CLUSTER_SHAPE_2x7x1 "CUBLASLT_CLUSTER_SHAPE_2x7x1" = 44
921
+ CUBLASLT_CLUSTER_SHAPE_1x9x1 "CUBLASLT_CLUSTER_SHAPE_1x9x1" = 45
922
+ CUBLASLT_CLUSTER_SHAPE_1x10x1 "CUBLASLT_CLUSTER_SHAPE_1x10x1" = 46
923
+ CUBLASLT_CLUSTER_SHAPE_1x11x1 "CUBLASLT_CLUSTER_SHAPE_1x11x1" = 47
924
+ CUBLASLT_CLUSTER_SHAPE_1x12x1 "CUBLASLT_CLUSTER_SHAPE_1x12x1" = 48
925
+ CUBLASLT_CLUSTER_SHAPE_1x13x1 "CUBLASLT_CLUSTER_SHAPE_1x13x1" = 49
926
+ CUBLASLT_CLUSTER_SHAPE_1x14x1 "CUBLASLT_CLUSTER_SHAPE_1x14x1" = 50
927
+ CUBLASLT_CLUSTER_SHAPE_1x15x1 "CUBLASLT_CLUSTER_SHAPE_1x15x1" = 51
928
+
929
+ ctypedef enum cublasLtMatmulInnerShape_t "cublasLtMatmulInnerShape_t":
930
+ CUBLASLT_MATMUL_INNER_SHAPE_UNDEFINED "CUBLASLT_MATMUL_INNER_SHAPE_UNDEFINED" = 0
931
+ CUBLASLT_MATMUL_INNER_SHAPE_MMA884 "CUBLASLT_MATMUL_INNER_SHAPE_MMA884" = 1
932
+ CUBLASLT_MATMUL_INNER_SHAPE_MMA1684 "CUBLASLT_MATMUL_INNER_SHAPE_MMA1684" = 2
933
+ CUBLASLT_MATMUL_INNER_SHAPE_MMA1688 "CUBLASLT_MATMUL_INNER_SHAPE_MMA1688" = 3
934
+ CUBLASLT_MATMUL_INNER_SHAPE_MMA16816 "CUBLASLT_MATMUL_INNER_SHAPE_MMA16816" = 4
935
+
936
+ ctypedef enum cublasLtMatmulMatrixScale_t "cublasLtMatmulMatrixScale_t":
937
+ CUBLASLT_MATMUL_MATRIX_SCALE_SCALAR_32F "CUBLASLT_MATMUL_MATRIX_SCALE_SCALAR_32F" = 0
938
+ CUBLASLT_MATMUL_MATRIX_SCALE_VEC16_UE4M3 "CUBLASLT_MATMUL_MATRIX_SCALE_VEC16_UE4M3" = 1
939
+ CUBLASLT_MATMUL_MATRIX_SCALE_VEC32_UE8M0 "CUBLASLT_MATMUL_MATRIX_SCALE_VEC32_UE8M0" = 2
940
+ CUBLASLT_MATMUL_MATRIX_SCALE_OUTER_VEC_32F "CUBLASLT_MATMUL_MATRIX_SCALE_OUTER_VEC_32F" = 3
941
+ CUBLASLT_MATMUL_MATRIX_SCALE_VEC128_32F "CUBLASLT_MATMUL_MATRIX_SCALE_VEC128_32F" = 4
942
+ CUBLASLT_MATMUL_MATRIX_SCALE_BLK128x128_32F "CUBLASLT_MATMUL_MATRIX_SCALE_BLK128x128_32F" = 5
943
+ CUBLASLT_MATMUL_MATRIX_SCALE_PER_BATCH_SCALAR_32F "CUBLASLT_MATMUL_MATRIX_SCALE_PER_BATCH_SCALAR_32F" = 6
944
+
945
+ ctypedef enum cublasLtBatchMode_t "cublasLtBatchMode_t":
946
+ CUBLASLT_BATCH_MODE_STRIDED "CUBLASLT_BATCH_MODE_STRIDED" = 0
947
+ CUBLASLT_BATCH_MODE_POINTER_ARRAY "CUBLASLT_BATCH_MODE_POINTER_ARRAY" = 1
948
+ CUBLASLT_BATCH_MODE_GROUPED "CUBLASLT_BATCH_MODE_GROUPED" = 2
949
+
950
+ ctypedef enum cublasLtIntegerWidth_t "cublasLtIntegerWidth_t":
951
+ CUBLASLT_INTEGER_WIDTH_32 "CUBLASLT_INTEGER_WIDTH_32" = 0
952
+ CUBLASLT_INTEGER_WIDTH_64 "CUBLASLT_INTEGER_WIDTH_64" = 1
953
+
954
+ ctypedef enum cublasLtEmulationDescAttributes_t "cublasLtEmulationDescAttributes_t":
955
+ CUBLASLT_EMULATION_DESC_STRATEGY "CUBLASLT_EMULATION_DESC_STRATEGY" = 0
956
+ CUBLASLT_EMULATION_DESC_SPECIAL_VALUES_SUPPORT "CUBLASLT_EMULATION_DESC_SPECIAL_VALUES_SUPPORT" = 1
957
+ CUBLASLT_EMULATION_DESC_FIXEDPOINT_MANTISSA_CONTROL "CUBLASLT_EMULATION_DESC_FIXEDPOINT_MANTISSA_CONTROL" = 2
958
+ CUBLASLT_EMULATION_DESC_FIXEDPOINT_MAX_MANTISSA_BIT_COUNT "CUBLASLT_EMULATION_DESC_FIXEDPOINT_MAX_MANTISSA_BIT_COUNT" = 3
959
+ CUBLASLT_EMULATION_DESC_FIXEDPOINT_MANTISSA_BIT_OFFSET "CUBLASLT_EMULATION_DESC_FIXEDPOINT_MANTISSA_BIT_OFFSET" = 4
960
+ CUBLASLT_EMULATION_DESC_FIXEDPOINT_MANTISSA_BIT_COUNT_POINTER "CUBLASLT_EMULATION_DESC_FIXEDPOINT_MANTISSA_BIT_COUNT_POINTER" = 5
961
+
962
+
963
+ # types
964
+ cdef extern from *:
965
+ """
966
+ #include <driver_types.h>
967
+ #include <library_types.h>
968
+ #include <cuComplex.h>
969
+ """
970
+ ctypedef void* cudaStream_t 'cudaStream_t'
971
+ ctypedef int cudaDataType_t 'cudaDataType_t'
972
+ ctypedef int cudaDataType 'cudaDataType'
973
+ ctypedef int libraryPropertyType_t 'libraryPropertyType_t'
974
+ ctypedef int libraryPropertyType 'libraryPropertyType'
975
+
976
+ ctypedef struct cuComplex:
977
+ pass
978
+ ctypedef struct cuDoubleComplex:
979
+ pass
980
+
981
+
982
+ ctypedef uint64_t cublasLtNumericalImplFlags_t 'cublasLtNumericalImplFlags_t'
983
+ ctypedef void* cublasLtHandle_t 'cublasLtHandle_t'
984
+ ctypedef void* cublasLtMatrixLayout_t 'cublasLtMatrixLayout_t'
985
+ ctypedef void* cublasLtMatmulDesc_t 'cublasLtMatmulDesc_t'
986
+ ctypedef void* cublasLtMatrixTransformDesc_t 'cublasLtMatrixTransformDesc_t'
987
+ ctypedef void* cublasLtMatmulPreference_t 'cublasLtMatmulPreference_t'
988
+ ctypedef void* cublasLtEmulationDesc_t 'cublasLtEmulationDesc_t'
989
+ ctypedef struct cublasLtMatmulAlgo_t 'cublasLtMatmulAlgo_t':
990
+ uint64_t data[8]
991
+
992
+ ctypedef void (*cublasLtLoggerCallback_t 'cublasLtLoggerCallback_t')(
993
+ int logLevel,
994
+ const char* functionName,
995
+ const char* message
996
+ )
997
+ ctypedef struct cublasLtMatmulHeuristicResult_t 'cublasLtMatmulHeuristicResult_t':
998
+ cublasLtMatmulAlgo_t algo
999
+ size_t workspaceSize
1000
+ cublasStatus_t state
1001
+ float wavesCount
1002
+ int reserved[4]
1003
+
1004
+
1005
+
1006
+ ###############################################################################
1007
+ # Functions
1008
+ ###############################################################################
1009
+
1010
+ cdef cublasStatus_t cublasLtCreate(cublasLtHandle_t* lightHandle) except?_CUBLASSTATUS_T_INTERNAL_LOADING_ERROR nogil
1011
+ cdef cublasStatus_t cublasLtDestroy(cublasLtHandle_t lightHandle) except?_CUBLASSTATUS_T_INTERNAL_LOADING_ERROR nogil
1012
+ cdef size_t cublasLtGetVersion() except?0 nogil
1013
+ cdef size_t cublasLtGetCudartVersion() except?0 nogil
1014
+ cdef cublasStatus_t cublasLtGetProperty(libraryPropertyType type, int* value) except?_CUBLASSTATUS_T_INTERNAL_LOADING_ERROR nogil
1015
+ cdef cublasStatus_t cublasLtMatmul(cublasLtHandle_t lightHandle, cublasLtMatmulDesc_t computeDesc, const void* alpha, const void* A, cublasLtMatrixLayout_t Adesc, const void* B, cublasLtMatrixLayout_t Bdesc, const void* beta, const void* C, cublasLtMatrixLayout_t Cdesc, void* D, cublasLtMatrixLayout_t Ddesc, const cublasLtMatmulAlgo_t* algo, void* workspace, size_t workspaceSizeInBytes, cudaStream_t stream) except?_CUBLASSTATUS_T_INTERNAL_LOADING_ERROR nogil
1016
+ cdef cublasStatus_t cublasLtMatrixTransform(cublasLtHandle_t lightHandle, cublasLtMatrixTransformDesc_t transformDesc, const void* alpha, const void* A, cublasLtMatrixLayout_t Adesc, const void* beta, const void* B, cublasLtMatrixLayout_t Bdesc, void* C, cublasLtMatrixLayout_t Cdesc, cudaStream_t stream) except?_CUBLASSTATUS_T_INTERNAL_LOADING_ERROR nogil
1017
+ cdef cublasStatus_t cublasLtMatrixLayoutCreate(cublasLtMatrixLayout_t* matLayout, cudaDataType type, uint64_t rows, uint64_t cols, int64_t ld) except?_CUBLASSTATUS_T_INTERNAL_LOADING_ERROR nogil
1018
+ cdef cublasStatus_t cublasLtMatrixLayoutDestroy(cublasLtMatrixLayout_t matLayout) except?_CUBLASSTATUS_T_INTERNAL_LOADING_ERROR nogil
1019
+ cdef cublasStatus_t cublasLtMatrixLayoutSetAttribute(cublasLtMatrixLayout_t matLayout, cublasLtMatrixLayoutAttribute_t attr, const void* buf, size_t sizeInBytes) except?_CUBLASSTATUS_T_INTERNAL_LOADING_ERROR nogil
1020
+ cdef cublasStatus_t cublasLtMatrixLayoutGetAttribute(cublasLtMatrixLayout_t matLayout, cublasLtMatrixLayoutAttribute_t attr, void* buf, size_t sizeInBytes, size_t* sizeWritten) except?_CUBLASSTATUS_T_INTERNAL_LOADING_ERROR nogil
1021
+ cdef cublasStatus_t cublasLtMatmulDescCreate(cublasLtMatmulDesc_t* matmulDesc, cublasComputeType_t computeType, cudaDataType_t scaleType) except?_CUBLASSTATUS_T_INTERNAL_LOADING_ERROR nogil
1022
+ cdef cublasStatus_t cublasLtMatmulDescDestroy(cublasLtMatmulDesc_t matmulDesc) except?_CUBLASSTATUS_T_INTERNAL_LOADING_ERROR nogil
1023
+ cdef cublasStatus_t cublasLtMatmulDescSetAttribute(cublasLtMatmulDesc_t matmulDesc, cublasLtMatmulDescAttributes_t attr, const void* buf, size_t sizeInBytes) except?_CUBLASSTATUS_T_INTERNAL_LOADING_ERROR nogil
1024
+ cdef cublasStatus_t cublasLtMatmulDescGetAttribute(cublasLtMatmulDesc_t matmulDesc, cublasLtMatmulDescAttributes_t attr, void* buf, size_t sizeInBytes, size_t* sizeWritten) except?_CUBLASSTATUS_T_INTERNAL_LOADING_ERROR nogil
1025
+ cdef cublasStatus_t cublasLtMatrixTransformDescCreate(cublasLtMatrixTransformDesc_t* transformDesc, cudaDataType scaleType) except?_CUBLASSTATUS_T_INTERNAL_LOADING_ERROR nogil
1026
+ cdef cublasStatus_t cublasLtMatrixTransformDescDestroy(cublasLtMatrixTransformDesc_t transformDesc) except?_CUBLASSTATUS_T_INTERNAL_LOADING_ERROR nogil
1027
+ cdef cublasStatus_t cublasLtMatrixTransformDescSetAttribute(cublasLtMatrixTransformDesc_t transformDesc, cublasLtMatrixTransformDescAttributes_t attr, const void* buf, size_t sizeInBytes) except?_CUBLASSTATUS_T_INTERNAL_LOADING_ERROR nogil
1028
+ cdef cublasStatus_t cublasLtMatrixTransformDescGetAttribute(cublasLtMatrixTransformDesc_t transformDesc, cublasLtMatrixTransformDescAttributes_t attr, void* buf, size_t sizeInBytes, size_t* sizeWritten) except?_CUBLASSTATUS_T_INTERNAL_LOADING_ERROR nogil
1029
+ cdef cublasStatus_t cublasLtMatmulPreferenceCreate(cublasLtMatmulPreference_t* pref) except?_CUBLASSTATUS_T_INTERNAL_LOADING_ERROR nogil
1030
+ cdef cublasStatus_t cublasLtMatmulPreferenceDestroy(cublasLtMatmulPreference_t pref) except?_CUBLASSTATUS_T_INTERNAL_LOADING_ERROR nogil
1031
+ cdef cublasStatus_t cublasLtMatmulPreferenceSetAttribute(cublasLtMatmulPreference_t pref, cublasLtMatmulPreferenceAttributes_t attr, const void* buf, size_t sizeInBytes) except?_CUBLASSTATUS_T_INTERNAL_LOADING_ERROR nogil
1032
+ cdef cublasStatus_t cublasLtMatmulPreferenceGetAttribute(cublasLtMatmulPreference_t pref, cublasLtMatmulPreferenceAttributes_t attr, void* buf, size_t sizeInBytes, size_t* sizeWritten) except?_CUBLASSTATUS_T_INTERNAL_LOADING_ERROR nogil
1033
+ cdef cublasStatus_t cublasLtMatmulAlgoGetHeuristic(cublasLtHandle_t lightHandle, cublasLtMatmulDesc_t operationDesc, cublasLtMatrixLayout_t Adesc, cublasLtMatrixLayout_t Bdesc, cublasLtMatrixLayout_t Cdesc, cublasLtMatrixLayout_t Ddesc, cublasLtMatmulPreference_t preference, int requestedAlgoCount, cublasLtMatmulHeuristicResult_t heuristicResultsArray[], int* returnAlgoCount) except?_CUBLASSTATUS_T_INTERNAL_LOADING_ERROR nogil
1034
+ cdef cublasStatus_t cublasLtMatmulAlgoGetIds(cublasLtHandle_t lightHandle, cublasComputeType_t computeType, cudaDataType_t scaleType, cudaDataType_t Atype, cudaDataType_t Btype, cudaDataType_t Ctype, cudaDataType_t Dtype, int requestedAlgoCount, int algoIdsArray[], int* returnAlgoCount) except?_CUBLASSTATUS_T_INTERNAL_LOADING_ERROR nogil
1035
+ cdef cublasStatus_t cublasLtMatmulAlgoInit(cublasLtHandle_t lightHandle, cublasComputeType_t computeType, cudaDataType_t scaleType, cudaDataType_t Atype, cudaDataType_t Btype, cudaDataType_t Ctype, cudaDataType_t Dtype, int algoId, cublasLtMatmulAlgo_t* algo) except?_CUBLASSTATUS_T_INTERNAL_LOADING_ERROR nogil
1036
+ cdef cublasStatus_t cublasLtMatmulAlgoCheck(cublasLtHandle_t lightHandle, cublasLtMatmulDesc_t operationDesc, cublasLtMatrixLayout_t Adesc, cublasLtMatrixLayout_t Bdesc, cublasLtMatrixLayout_t Cdesc, cublasLtMatrixLayout_t Ddesc, const cublasLtMatmulAlgo_t* algo, cublasLtMatmulHeuristicResult_t* result) except?_CUBLASSTATUS_T_INTERNAL_LOADING_ERROR nogil
1037
+ cdef cublasStatus_t cublasLtMatmulAlgoCapGetAttribute(const cublasLtMatmulAlgo_t* algo, cublasLtMatmulAlgoCapAttributes_t attr, void* buf, size_t sizeInBytes, size_t* sizeWritten) except?_CUBLASSTATUS_T_INTERNAL_LOADING_ERROR nogil
1038
+ cdef cublasStatus_t cublasLtMatmulAlgoConfigSetAttribute(cublasLtMatmulAlgo_t* algo, cublasLtMatmulAlgoConfigAttributes_t attr, const void* buf, size_t sizeInBytes) except?_CUBLASSTATUS_T_INTERNAL_LOADING_ERROR nogil
1039
+ cdef cublasStatus_t cublasLtMatmulAlgoConfigGetAttribute(const cublasLtMatmulAlgo_t* algo, cublasLtMatmulAlgoConfigAttributes_t attr, void* buf, size_t sizeInBytes, size_t* sizeWritten) except?_CUBLASSTATUS_T_INTERNAL_LOADING_ERROR nogil
1040
+ cdef cublasStatus_t cublasLtLoggerSetCallback(cublasLtLoggerCallback_t callback) except?_CUBLASSTATUS_T_INTERNAL_LOADING_ERROR nogil
1041
+ cdef cublasStatus_t cublasLtLoggerSetFile(FILE* file) except?_CUBLASSTATUS_T_INTERNAL_LOADING_ERROR nogil
1042
+ cdef cublasStatus_t cublasLtLoggerOpenFile(const char* logFile) except?_CUBLASSTATUS_T_INTERNAL_LOADING_ERROR nogil
1043
+ cdef cublasStatus_t cublasLtLoggerSetLevel(int level) except?_CUBLASSTATUS_T_INTERNAL_LOADING_ERROR nogil
1044
+ cdef cublasStatus_t cublasLtLoggerSetMask(int mask) except?_CUBLASSTATUS_T_INTERNAL_LOADING_ERROR nogil
1045
+ cdef cublasStatus_t cublasLtLoggerForceDisable() except?_CUBLASSTATUS_T_INTERNAL_LOADING_ERROR nogil
1046
+ cdef const char* cublasLtGetStatusName(cublasStatus_t status) except?NULL nogil
1047
+ cdef const char* cublasLtGetStatusString(cublasStatus_t status) except?NULL nogil
1048
+ cdef cublasStatus_t cublasLtHeuristicsCacheGetCapacity(size_t* capacity) except?_CUBLASSTATUS_T_INTERNAL_LOADING_ERROR nogil
1049
+ cdef cublasStatus_t cublasLtHeuristicsCacheSetCapacity(size_t capacity) except?_CUBLASSTATUS_T_INTERNAL_LOADING_ERROR nogil
1050
+ cdef unsigned cublasLtDisableCpuInstructionsSetMask(unsigned mask) except?0 nogil
1051
+ cdef cublasStatus_t cublasLtGroupedMatrixLayoutCreate(cublasLtMatrixLayout_t* matLayout, cudaDataType type, int groupCount, const void* rows_array, const void* cols_array, const void* ld_array) except?_CUBLASSTATUS_T_INTERNAL_LOADING_ERROR nogil
1052
+ cdef cublasStatus_t cublasLtEmulationDescCreate(cublasLtEmulationDesc_t* emulationDesc) except?_CUBLASSTATUS_T_INTERNAL_LOADING_ERROR nogil
1053
+ cdef cublasStatus_t cublasLtEmulationDescDestroy(cublasLtEmulationDesc_t emulationDesc) except?_CUBLASSTATUS_T_INTERNAL_LOADING_ERROR nogil
1054
+ cdef cublasStatus_t cublasLtEmulationDescSetAttribute(cublasLtEmulationDesc_t emulationDesc, cublasLtEmulationDescAttributes_t attr, const void* buf, size_t sizeInBytes) except?_CUBLASSTATUS_T_INTERNAL_LOADING_ERROR nogil
1055
+ cdef cublasStatus_t cublasLtEmulationDescGetAttribute(cublasLtEmulationDesc_t emulationDesc, cublasLtEmulationDescAttributes_t attr, void* buf, size_t sizeInBytes, size_t* sizeWritten) except?_CUBLASSTATUS_T_INTERNAL_LOADING_ERROR nogil