nvmath-python 1.0.0__cp314-cp314t-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 (332) 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 +130 -0
  7. nvmath/_internal/threadsafe.py +106 -0
  8. nvmath/_internal/utils.py +43 -0
  9. nvmath/_internal/workspace.py +490 -0
  10. nvmath/_utils.py +147 -0
  11. nvmath/bindings/__init__.py +60 -0
  12. nvmath/bindings/_internal/__init__.pxd +0 -0
  13. nvmath/bindings/_internal/__init__.py +0 -0
  14. nvmath/bindings/_internal/common_types.pxd +31 -0
  15. nvmath/bindings/_internal/cublas.cp314t-win_amd64.pyd +0 -0
  16. nvmath/bindings/_internal/cublas.pxd +530 -0
  17. nvmath/bindings/_internal/cublasLt.cp314t-win_amd64.pyd +0 -0
  18. nvmath/bindings/_internal/cublasLt.pxd +59 -0
  19. nvmath/bindings/_internal/cublasMp.pxd +52 -0
  20. nvmath/bindings/_internal/cudss.cp314t-win_amd64.pyd +0 -0
  21. nvmath/bindings/_internal/cudss.pxd +54 -0
  22. nvmath/bindings/_internal/cufft.cp314t-win_amd64.pyd +0 -0
  23. nvmath/bindings/_internal/cufft.pxd +70 -0
  24. nvmath/bindings/_internal/cufftMp.pxd +77 -0
  25. nvmath/bindings/_internal/curand.cp314t-win_amd64.pyd +0 -0
  26. nvmath/bindings/_internal/curand.pxd +42 -0
  27. nvmath/bindings/_internal/cusolver.cp314t-win_amd64.pyd +0 -0
  28. nvmath/bindings/_internal/cusolver.pxd +15 -0
  29. nvmath/bindings/_internal/cusolverDn.cp314t-win_amd64.pyd +0 -0
  30. nvmath/bindings/_internal/cusolverDn.pxd +406 -0
  31. nvmath/bindings/_internal/cusolverMp.pxd +71 -0
  32. nvmath/bindings/_internal/cusolverSp.cp314t-win_amd64.pyd +0 -0
  33. nvmath/bindings/_internal/cusolverSp.pxd +75 -0
  34. nvmath/bindings/_internal/cusparse.cp314t-win_amd64.pyd +0 -0
  35. nvmath/bindings/_internal/cusparse.pxd +471 -0
  36. nvmath/bindings/_internal/cusparseLt.cp314t-win_amd64.pyd +0 -0
  37. nvmath/bindings/_internal/cusparseLt.pxd +48 -0
  38. nvmath/bindings/_internal/cutensor.cp314t-win_amd64.pyd +0 -0
  39. nvmath/bindings/_internal/cutensor.pxd +58 -0
  40. nvmath/bindings/_internal/mathdx.cp314t-win_amd64.pyd +0 -0
  41. nvmath/bindings/_internal/mathdx.pxd +116 -0
  42. nvmath/bindings/_internal/nvshmem.pxd +29 -0
  43. nvmath/bindings/_internal/utils.cp314t-win_amd64.pyd +0 -0
  44. nvmath/bindings/_internal/utils.pxd +174 -0
  45. nvmath/bindings/_internal/utils.pyi +10 -0
  46. nvmath/bindings/cublas.cp314t-win_amd64.pyd +0 -0
  47. nvmath/bindings/cublas.pxd +558 -0
  48. nvmath/bindings/cublas.pyi +812 -0
  49. nvmath/bindings/cublasLt.cp314t-win_amd64.pyd +0 -0
  50. nvmath/bindings/cublasLt.pxd +109 -0
  51. nvmath/bindings/cublasLt.pyi +1461 -0
  52. nvmath/bindings/cublasMp.pxd +85 -0
  53. nvmath/bindings/cublasMp.pyi +267 -0
  54. nvmath/bindings/cudss.cp314t-win_amd64.pyd +0 -0
  55. nvmath/bindings/cudss.pxd +98 -0
  56. nvmath/bindings/cudss.pyi +443 -0
  57. nvmath/bindings/cufft.cp314t-win_amd64.pyd +0 -0
  58. nvmath/bindings/cufft.pxd +118 -0
  59. nvmath/bindings/cufft.pyi +301 -0
  60. nvmath/bindings/cufftMp.pxd +124 -0
  61. nvmath/bindings/cufftMp.pyi +326 -0
  62. nvmath/bindings/curand.cp314t-win_amd64.pyd +0 -0
  63. nvmath/bindings/curand.pxd +71 -0
  64. nvmath/bindings/curand.pyi +189 -0
  65. nvmath/bindings/cusolver.cp314t-win_amd64.pyd +0 -0
  66. nvmath/bindings/cusolver.pxd +62 -0
  67. nvmath/bindings/cusolver.pyi +320 -0
  68. nvmath/bindings/cusolverDn.cp314t-win_amd64.pyd +0 -0
  69. nvmath/bindings/cusolverDn.pxd +430 -0
  70. nvmath/bindings/cusolverDn.pyi +422 -0
  71. nvmath/bindings/cusolverMp.pxd +98 -0
  72. nvmath/bindings/cusolverMp.pyi +114 -0
  73. nvmath/bindings/cusolverSp.cp314t-win_amd64.pyd +0 -0
  74. nvmath/bindings/cusolverSp.pxd +95 -0
  75. nvmath/bindings/cusolverSp.pyi +70 -0
  76. nvmath/bindings/cusparse.cp314t-win_amd64.pyd +0 -0
  77. nvmath/bindings/cusparse.pxd +546 -0
  78. nvmath/bindings/cusparse.pyi +1017 -0
  79. nvmath/bindings/cusparseLt.cp314t-win_amd64.pyd +0 -0
  80. nvmath/bindings/cusparseLt.pxd +99 -0
  81. nvmath/bindings/cusparseLt.pyi +252 -0
  82. nvmath/bindings/cutensor.cp314t-win_amd64.pyd +0 -0
  83. nvmath/bindings/cutensor.pxd +98 -0
  84. nvmath/bindings/cutensor.pyi +324 -0
  85. nvmath/bindings/cycublas.cp314t-win_amd64.pyd +0 -0
  86. nvmath/bindings/cycublas.pxd +664 -0
  87. nvmath/bindings/cycublasLt.cp314t-win_amd64.pyd +0 -0
  88. nvmath/bindings/cycublasLt.pxd +1045 -0
  89. nvmath/bindings/cycublasMp.pxd +171 -0
  90. nvmath/bindings/cycudss.cp314t-win_amd64.pyd +0 -0
  91. nvmath/bindings/cycudss.pxd +277 -0
  92. nvmath/bindings/cycufft.cp314t-win_amd64.pyd +0 -0
  93. nvmath/bindings/cycufft.pxd +333 -0
  94. nvmath/bindings/cycufftMp.pxd +342 -0
  95. nvmath/bindings/cycurand.cp314t-win_amd64.pyd +0 -0
  96. nvmath/bindings/cycurand.pxd +141 -0
  97. nvmath/bindings/cycusolver.cp314t-win_amd64.pyd +0 -0
  98. nvmath/bindings/cycusolver.pxd +137 -0
  99. nvmath/bindings/cycusolverDn.cp314t-win_amd64.pyd +0 -0
  100. nvmath/bindings/cycusolverDn.pxd +443 -0
  101. nvmath/bindings/cycusolverMp.pxd +107 -0
  102. nvmath/bindings/cycusolverSp.cp314t-win_amd64.pyd +0 -0
  103. nvmath/bindings/cycusolverSp.pxd +93 -0
  104. nvmath/bindings/cycusparse.cp314t-win_amd64.pyd +0 -0
  105. nvmath/bindings/cycusparse.pxd +679 -0
  106. nvmath/bindings/cycusparseLt.cp314t-win_amd64.pyd +0 -0
  107. nvmath/bindings/cycusparseLt.pxd +135 -0
  108. nvmath/bindings/cycutensor.cp314t-win_amd64.pyd +0 -0
  109. nvmath/bindings/cycutensor.pxd +189 -0
  110. nvmath/bindings/cymathdx.cp314t-win_amd64.pyd +0 -0
  111. nvmath/bindings/cymathdx.pxd +552 -0
  112. nvmath/bindings/cynvshmem.pxd +118 -0
  113. nvmath/bindings/mathdx.cp314t-win_amd64.pyd +0 -0
  114. nvmath/bindings/mathdx.pxd +182 -0
  115. nvmath/bindings/mathdx.pyi +1562 -0
  116. nvmath/bindings/nvpl/__init__.pxd +0 -0
  117. nvmath/bindings/nvpl/__init__.py +13 -0
  118. nvmath/bindings/nvpl/_internal/__init__.pxd +0 -0
  119. nvmath/bindings/nvpl/_internal/__init__.py +0 -0
  120. nvmath/bindings/nvpl/_internal/blas.cp314t-win_amd64.pyd +0 -0
  121. nvmath/bindings/nvpl/_internal/blas.pxd +237 -0
  122. nvmath/bindings/nvpl/_internal/fft.cp314t-win_amd64.pyd +0 -0
  123. nvmath/bindings/nvpl/_internal/fft.pxd +36 -0
  124. nvmath/bindings/nvpl/blas.cp314t-win_amd64.pyd +0 -0
  125. nvmath/bindings/nvpl/blas.pxd +131 -0
  126. nvmath/bindings/nvpl/blas.pyi +168 -0
  127. nvmath/bindings/nvpl/cyblas.cp314t-win_amd64.pyd +0 -0
  128. nvmath/bindings/nvpl/cyblas.pxd +280 -0
  129. nvmath/bindings/nvpl/cyfft.cp314t-win_amd64.pyd +0 -0
  130. nvmath/bindings/nvpl/cyfft.pxd +93 -0
  131. nvmath/bindings/nvpl/fft.cp314t-win_amd64.pyd +0 -0
  132. nvmath/bindings/nvpl/fft.pxd +100 -0
  133. nvmath/bindings/nvpl/fft.pyi +168 -0
  134. nvmath/bindings/nvshmem.pxd +54 -0
  135. nvmath/bindings/nvshmem.pyi +191 -0
  136. nvmath/device/__init__.py +38 -0
  137. nvmath/device/_deprecated.py +33 -0
  138. nvmath/device/common.py +315 -0
  139. nvmath/device/common_backend.py +131 -0
  140. nvmath/device/common_cuda.py +201 -0
  141. nvmath/device/common_numba.py +300 -0
  142. nvmath/device/common_numba_cuda_mlir.py +202 -0
  143. nvmath/device/common_opaque_tensor.py +201 -0
  144. nvmath/device/cublasdx.py +1606 -0
  145. nvmath/device/cublasdx_backend.py +860 -0
  146. nvmath/device/cublasdx_numba.py +1534 -0
  147. nvmath/device/cublasdx_numba_cuda_mlir.py +208 -0
  148. nvmath/device/cufftdx.py +373 -0
  149. nvmath/device/cufftdx_backend.py +220 -0
  150. nvmath/device/cufftdx_numba.py +140 -0
  151. nvmath/device/cufftdx_numba_cuda_mlir.py +79 -0
  152. nvmath/device/curand_kernel.py +9147 -0
  153. nvmath/device/cusolverdx.py +2708 -0
  154. nvmath/device/cusolverdx_backend.py +440 -0
  155. nvmath/device/cusolverdx_numba.py +567 -0
  156. nvmath/device/cusolverdx_numba_cuda_mlir.py +604 -0
  157. nvmath/device/cusolverdx_overload_backend.py +1029 -0
  158. nvmath/device/llvm_array.py +29 -0
  159. nvmath/device/random.py +441 -0
  160. nvmath/device/random_helpers.py +23 -0
  161. nvmath/device/random_states.py +187 -0
  162. nvmath/device/types.py +138 -0
  163. nvmath/device/vector_types_numba.py +259 -0
  164. nvmath/distributed/__init__.py +200 -0
  165. nvmath/distributed/_internal/__init__.py +0 -0
  166. nvmath/distributed/_internal/nccl.py +86 -0
  167. nvmath/distributed/_internal/nvshmem.py +307 -0
  168. nvmath/distributed/_internal/symmetric_memory.py +35 -0
  169. nvmath/distributed/_internal/tensor_ifc.py +70 -0
  170. nvmath/distributed/_internal/tensor_ifc_cupy.py +68 -0
  171. nvmath/distributed/_internal/tensor_ifc_host_device.py +172 -0
  172. nvmath/distributed/_internal/tensor_ifc_numpy.py +46 -0
  173. nvmath/distributed/_internal/tensor_ifc_torch.py +162 -0
  174. nvmath/distributed/_internal/tensor_wrapper.py +81 -0
  175. nvmath/distributed/_utils.py +167 -0
  176. nvmath/distributed/distribution/__init__.py +30 -0
  177. nvmath/distributed/distribution/_configuration.py +39 -0
  178. nvmath/distributed/distribution/distributions.py +1024 -0
  179. nvmath/distributed/distribution/redistribute.py +1284 -0
  180. nvmath/distributed/fft/__init__.py +7 -0
  181. nvmath/distributed/fft/_configuration.py +82 -0
  182. nvmath/distributed/fft/fft.py +2742 -0
  183. nvmath/distributed/linalg/__init__.py +22 -0
  184. nvmath/distributed/linalg/_internal/__init__.py +3 -0
  185. nvmath/distributed/linalg/_internal/epilog_protocol.py +586 -0
  186. nvmath/distributed/linalg/_internal/matmul_desc_ifc.py +28 -0
  187. nvmath/distributed/linalg/advanced/__init__.py +8 -0
  188. nvmath/distributed/linalg/advanced/_configuration.py +171 -0
  189. nvmath/distributed/linalg/advanced/matmulmod.py +3573 -0
  190. nvmath/distributed/linalg/generic/__init__.py +8 -0
  191. nvmath/distributed/linalg/generic/_caching.py +66 -0
  192. nvmath/distributed/linalg/generic/_configuration.py +61 -0
  193. nvmath/distributed/linalg/generic/_factorization.py +172 -0
  194. nvmath/distributed/linalg/generic/_initialization.py +966 -0
  195. nvmath/distributed/linalg/generic/_problem_spec.py +511 -0
  196. nvmath/distributed/linalg/generic/solvermod.py +1368 -0
  197. nvmath/distributed/process_group.py +408 -0
  198. nvmath/fft/__init__.py +7 -0
  199. nvmath/fft/_configuration.py +189 -0
  200. nvmath/fft/_exec_utils.py +82 -0
  201. nvmath/fft/_helpers.py +237 -0
  202. nvmath/fft/fft.py +3122 -0
  203. nvmath/internal/__init__.pxd +3 -0
  204. nvmath/internal/__init__.py +10 -0
  205. nvmath/internal/_bindings.cp314t-win_amd64.pyd +0 -0
  206. nvmath/internal/_bindings.pxd +18 -0
  207. nvmath/internal/_device_utils.py +45 -0
  208. nvmath/internal/_layout/__init__.pxd +3 -0
  209. nvmath/internal/_layout/__init__.py +7 -0
  210. nvmath/internal/_layout/_layout.cp314t-win_amd64.pyd +0 -0
  211. nvmath/internal/_layout/_layout.pxd +1303 -0
  212. nvmath/internal/_layout/_layout.pyi +1145 -0
  213. nvmath/internal/enum_utils.py +142 -0
  214. nvmath/internal/formatters.py +87 -0
  215. nvmath/internal/mem_limit.py +51 -0
  216. nvmath/internal/memory.cp314t-win_amd64.pyd +0 -0
  217. nvmath/internal/memory.pxd +13 -0
  218. nvmath/internal/memory.pyi +50 -0
  219. nvmath/internal/ndbuffer/__init__.pxd +3 -0
  220. nvmath/internal/ndbuffer/__init__.py +9 -0
  221. nvmath/internal/ndbuffer/_copy_kernel.cp314t-win_amd64.pyd +0 -0
  222. nvmath/internal/ndbuffer/_copy_kernel.pxd +10 -0
  223. nvmath/internal/ndbuffer/_jit.cp314t-win_amd64.pyd +0 -0
  224. nvmath/internal/ndbuffer/_jit.pxd +7 -0
  225. nvmath/internal/ndbuffer/_ndbuffer.cp314t-win_amd64.pyd +0 -0
  226. nvmath/internal/ndbuffer/_ndbuffer.pxd +40 -0
  227. nvmath/internal/ndbuffer/_ndbuffer.pyi +463 -0
  228. nvmath/internal/ndbuffer/copy_kernel/args.h +34 -0
  229. nvmath/internal/ndbuffer/copy_kernel/copy_kernel_impl/array_view.h +52 -0
  230. nvmath/internal/ndbuffer/copy_kernel/copy_kernel_impl/elementwise.h +68 -0
  231. nvmath/internal/ndbuffer/copy_kernel/copy_kernel_impl/grid_indexer.h +69 -0
  232. nvmath/internal/ndbuffer/copy_kernel/copy_kernel_impl/transposed.h +242 -0
  233. nvmath/internal/ndbuffer/copy_kernel/copy_kernel_impl/type_utils.h +39 -0
  234. nvmath/internal/ndbuffer/copy_kernel/copy_kernel_impl/utils.h +132 -0
  235. nvmath/internal/ndbuffer/copy_kernel/copy_kernel_impl/vec.h +159 -0
  236. nvmath/internal/ndbuffer/copy_kernel/elementwise.h +53 -0
  237. nvmath/internal/ndbuffer/copy_kernel/transposed.h +58 -0
  238. nvmath/internal/package_ifc.py +168 -0
  239. nvmath/internal/package_ifc_cuda.py +57 -0
  240. nvmath/internal/package_ifc_cupy.py +67 -0
  241. nvmath/internal/package_ifc_torch.py +69 -0
  242. nvmath/internal/package_wrapper.py +14 -0
  243. nvmath/internal/tensor_ifc.py +179 -0
  244. nvmath/internal/tensor_ifc_cupy.py +234 -0
  245. nvmath/internal/tensor_ifc_ndbuffer.py +147 -0
  246. nvmath/internal/tensor_ifc_numpy.py +184 -0
  247. nvmath/internal/tensor_ifc_torch.py +178 -0
  248. nvmath/internal/tensor_wrapper.py +160 -0
  249. nvmath/internal/typemaps.py +113 -0
  250. nvmath/internal/utils.py +805 -0
  251. nvmath/linalg/__init__.py +56 -0
  252. nvmath/linalg/_internal/__init__.py +3 -0
  253. nvmath/linalg/_internal/algo_cap_ifc.py +82 -0
  254. nvmath/linalg/_internal/algo_config_ifc.py +43 -0
  255. nvmath/linalg/_internal/batch.py +234 -0
  256. nvmath/linalg/_internal/enum_to_tuples.py +64 -0
  257. nvmath/linalg/_internal/epilog_protocol.py +766 -0
  258. nvmath/linalg/_internal/layout.py +624 -0
  259. nvmath/linalg/_internal/matmul_desc_ifc.py +28 -0
  260. nvmath/linalg/_internal/matmul_pref_ifc.py +27 -0
  261. nvmath/linalg/_internal/matrix_layout_ifc.py +26 -0
  262. nvmath/linalg/_internal/solver_utils.py +432 -0
  263. nvmath/linalg/_internal/typemaps.py +144 -0
  264. nvmath/linalg/_internal/utils.py +157 -0
  265. nvmath/linalg/advanced/__init__.py +8 -0
  266. nvmath/linalg/advanced/_algorithmmod.py +170 -0
  267. nvmath/linalg/advanced/_configuration.py +351 -0
  268. nvmath/linalg/advanced/helpers/__init__.py +5 -0
  269. nvmath/linalg/advanced/helpers/matmul.py +1316 -0
  270. nvmath/linalg/advanced/matmulmod.py +3734 -0
  271. nvmath/linalg/generic/__init__.py +53 -0
  272. nvmath/linalg/generic/_configuration/__init__.py +39 -0
  273. nvmath/linalg/generic/_configuration/layout.py +263 -0
  274. nvmath/linalg/generic/_configuration/match.py +734 -0
  275. nvmath/linalg/generic/_configuration/qualifiers.py +493 -0
  276. nvmath/linalg/generic/_configuration/solver_configuration.py +59 -0
  277. nvmath/linalg/generic/_configuration/wrap.py +217 -0
  278. nvmath/linalg/generic/_dtype.py +15 -0
  279. nvmath/linalg/generic/matmulmod.py +2094 -0
  280. nvmath/linalg/generic/solvermod.py +1301 -0
  281. nvmath/memory.py +279 -0
  282. nvmath/sparse/__init__.py +38 -0
  283. nvmath/sparse/_internal/__init__.py +21 -0
  284. nvmath/sparse/_internal/common_utils.py +147 -0
  285. nvmath/sparse/_internal/cudss_config_ifc.py +702 -0
  286. nvmath/sparse/_internal/cudss_data_ifc.py +399 -0
  287. nvmath/sparse/_internal/cudss_utils.py +506 -0
  288. nvmath/sparse/_internal/cusparse_utils.py +382 -0
  289. nvmath/sparse/_internal/sparse_bsc_ifc.py +303 -0
  290. nvmath/sparse/_internal/sparse_bsr_ifc.py +305 -0
  291. nvmath/sparse/_internal/sparse_coo_ifc.py +256 -0
  292. nvmath/sparse/_internal/sparse_csc_ifc.py +268 -0
  293. nvmath/sparse/_internal/sparse_csr_ifc.py +288 -0
  294. nvmath/sparse/_internal/sparse_dia_ifc.py +242 -0
  295. nvmath/sparse/_internal/sparse_format_helpers.py +601 -0
  296. nvmath/sparse/_internal/sparse_tensor_ifc.py +133 -0
  297. nvmath/sparse/_internal/sparse_ust_ifc.py +141 -0
  298. nvmath/sparse/_internal/utils.py +56 -0
  299. nvmath/sparse/advanced/__init__.py +7 -0
  300. nvmath/sparse/advanced/_configuration.py +227 -0
  301. nvmath/sparse/advanced/direct_solver.py +2069 -0
  302. nvmath/sparse/generic/__init__.py +7 -0
  303. nvmath/sparse/generic/_configuration.py +129 -0
  304. nvmath/sparse/generic/_helpers.py +137 -0
  305. nvmath/sparse/generic/_thunks.py +21 -0
  306. nvmath/sparse/generic/matmulmod.py +2353 -0
  307. nvmath/sparse/ust/__init__.py +7 -0
  308. nvmath/sparse/ust/_converters.py +422 -0
  309. nvmath/sparse/ust/_cpp.py +28 -0
  310. nvmath/sparse/ust/_drawer.py +565 -0
  311. nvmath/sparse/ust/_emitter.py +1033 -0
  312. nvmath/sparse/ust/_jit.py +188 -0
  313. nvmath/sparse/ust/_kernel.py +282 -0
  314. nvmath/sparse/ust/_utils.py +149 -0
  315. nvmath/sparse/ust/interfaces/__init__.py +0 -0
  316. nvmath/sparse/ust/interfaces/torch_interface.py +476 -0
  317. nvmath/sparse/ust/tensor.py +1016 -0
  318. nvmath/sparse/ust/tensor_format.py +957 -0
  319. nvmath/tensor/__init__.py +6 -0
  320. nvmath/tensor/_configuration.py +120 -0
  321. nvmath/tensor/_internal/__init__.py +3 -0
  322. nvmath/tensor/_internal/cutensor_config_ifc.py +279 -0
  323. nvmath/tensor/_internal/cutensor_utils.py +230 -0
  324. nvmath/tensor/_internal/data.py +43 -0
  325. nvmath/tensor/_internal/einsum_parser.py +444 -0
  326. nvmath/tensor/_internal/typemaps.py +96 -0
  327. nvmath/tensor/contract.py +1900 -0
  328. nvmath_python-1.0.0.dist-info/METADATA +134 -0
  329. nvmath_python-1.0.0.dist-info/RECORD +332 -0
  330. nvmath_python-1.0.0.dist-info/WHEEL +5 -0
  331. nvmath_python-1.0.0.dist-info/licenses/LICENSE +177 -0
  332. nvmath_python-1.0.0.dist-info/top_level.txt +2 -0
@@ -0,0 +1,1045 @@
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.dev1471+g7122059e9. 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
+ from ._internal.common_types cimport cudaDataType, cudaDataType_t, cudaStream_t, libraryPropertyType, libraryPropertyType_t, cuComplex, cuDoubleComplex
14
+
15
+
16
+ ###############################################################################
17
+ # Types (structs, enums, ...)
18
+ ###############################################################################
19
+
20
+ # enums
21
+ ctypedef enum cublasLtMatmulTile_t "cublasLtMatmulTile_t":
22
+ CUBLASLT_MATMUL_TILE_UNDEFINED "CUBLASLT_MATMUL_TILE_UNDEFINED" = 0
23
+ CUBLASLT_MATMUL_TILE_8x8 "CUBLASLT_MATMUL_TILE_8x8" = 1
24
+ CUBLASLT_MATMUL_TILE_8x16 "CUBLASLT_MATMUL_TILE_8x16" = 2
25
+ CUBLASLT_MATMUL_TILE_16x8 "CUBLASLT_MATMUL_TILE_16x8" = 3
26
+ CUBLASLT_MATMUL_TILE_8x32 "CUBLASLT_MATMUL_TILE_8x32" = 4
27
+ CUBLASLT_MATMUL_TILE_16x16 "CUBLASLT_MATMUL_TILE_16x16" = 5
28
+ CUBLASLT_MATMUL_TILE_32x8 "CUBLASLT_MATMUL_TILE_32x8" = 6
29
+ CUBLASLT_MATMUL_TILE_8x64 "CUBLASLT_MATMUL_TILE_8x64" = 7
30
+ CUBLASLT_MATMUL_TILE_16x32 "CUBLASLT_MATMUL_TILE_16x32" = 8
31
+ CUBLASLT_MATMUL_TILE_32x16 "CUBLASLT_MATMUL_TILE_32x16" = 9
32
+ CUBLASLT_MATMUL_TILE_64x8 "CUBLASLT_MATMUL_TILE_64x8" = 10
33
+ CUBLASLT_MATMUL_TILE_32x32 "CUBLASLT_MATMUL_TILE_32x32" = 11
34
+ CUBLASLT_MATMUL_TILE_32x64 "CUBLASLT_MATMUL_TILE_32x64" = 12
35
+ CUBLASLT_MATMUL_TILE_64x32 "CUBLASLT_MATMUL_TILE_64x32" = 13
36
+ CUBLASLT_MATMUL_TILE_32x128 "CUBLASLT_MATMUL_TILE_32x128" = 14
37
+ CUBLASLT_MATMUL_TILE_64x64 "CUBLASLT_MATMUL_TILE_64x64" = 15
38
+ CUBLASLT_MATMUL_TILE_128x32 "CUBLASLT_MATMUL_TILE_128x32" = 16
39
+ CUBLASLT_MATMUL_TILE_64x128 "CUBLASLT_MATMUL_TILE_64x128" = 17
40
+ CUBLASLT_MATMUL_TILE_128x64 "CUBLASLT_MATMUL_TILE_128x64" = 18
41
+ CUBLASLT_MATMUL_TILE_64x256 "CUBLASLT_MATMUL_TILE_64x256" = 19
42
+ CUBLASLT_MATMUL_TILE_128x128 "CUBLASLT_MATMUL_TILE_128x128" = 20
43
+ CUBLASLT_MATMUL_TILE_256x64 "CUBLASLT_MATMUL_TILE_256x64" = 21
44
+ CUBLASLT_MATMUL_TILE_64x512 "CUBLASLT_MATMUL_TILE_64x512" = 22
45
+ CUBLASLT_MATMUL_TILE_128x256 "CUBLASLT_MATMUL_TILE_128x256" = 23
46
+ CUBLASLT_MATMUL_TILE_256x128 "CUBLASLT_MATMUL_TILE_256x128" = 24
47
+ CUBLASLT_MATMUL_TILE_512x64 "CUBLASLT_MATMUL_TILE_512x64" = 25
48
+ CUBLASLT_MATMUL_TILE_64x96 "CUBLASLT_MATMUL_TILE_64x96" = 26
49
+ CUBLASLT_MATMUL_TILE_96x64 "CUBLASLT_MATMUL_TILE_96x64" = 27
50
+ CUBLASLT_MATMUL_TILE_96x128 "CUBLASLT_MATMUL_TILE_96x128" = 28
51
+ CUBLASLT_MATMUL_TILE_128x160 "CUBLASLT_MATMUL_TILE_128x160" = 29
52
+ CUBLASLT_MATMUL_TILE_160x128 "CUBLASLT_MATMUL_TILE_160x128" = 30
53
+ CUBLASLT_MATMUL_TILE_192x128 "CUBLASLT_MATMUL_TILE_192x128" = 31
54
+ CUBLASLT_MATMUL_TILE_128x192 "CUBLASLT_MATMUL_TILE_128x192" = 32
55
+ CUBLASLT_MATMUL_TILE_128x96 "CUBLASLT_MATMUL_TILE_128x96" = 33
56
+ CUBLASLT_MATMUL_TILE_32x256 "CUBLASLT_MATMUL_TILE_32x256" = 34
57
+ CUBLASLT_MATMUL_TILE_256x32 "CUBLASLT_MATMUL_TILE_256x32" = 35
58
+ CUBLASLT_MATMUL_TILE_8x128 "CUBLASLT_MATMUL_TILE_8x128" = 36
59
+ CUBLASLT_MATMUL_TILE_8x192 "CUBLASLT_MATMUL_TILE_8x192" = 37
60
+ CUBLASLT_MATMUL_TILE_8x256 "CUBLASLT_MATMUL_TILE_8x256" = 38
61
+ CUBLASLT_MATMUL_TILE_8x320 "CUBLASLT_MATMUL_TILE_8x320" = 39
62
+ CUBLASLT_MATMUL_TILE_8x384 "CUBLASLT_MATMUL_TILE_8x384" = 40
63
+ CUBLASLT_MATMUL_TILE_8x448 "CUBLASLT_MATMUL_TILE_8x448" = 41
64
+ CUBLASLT_MATMUL_TILE_8x512 "CUBLASLT_MATMUL_TILE_8x512" = 42
65
+ CUBLASLT_MATMUL_TILE_8x576 "CUBLASLT_MATMUL_TILE_8x576" = 43
66
+ CUBLASLT_MATMUL_TILE_8x640 "CUBLASLT_MATMUL_TILE_8x640" = 44
67
+ CUBLASLT_MATMUL_TILE_8x704 "CUBLASLT_MATMUL_TILE_8x704" = 45
68
+ CUBLASLT_MATMUL_TILE_8x768 "CUBLASLT_MATMUL_TILE_8x768" = 46
69
+ CUBLASLT_MATMUL_TILE_16x64 "CUBLASLT_MATMUL_TILE_16x64" = 47
70
+ CUBLASLT_MATMUL_TILE_16x128 "CUBLASLT_MATMUL_TILE_16x128" = 48
71
+ CUBLASLT_MATMUL_TILE_16x192 "CUBLASLT_MATMUL_TILE_16x192" = 49
72
+ CUBLASLT_MATMUL_TILE_16x256 "CUBLASLT_MATMUL_TILE_16x256" = 50
73
+ CUBLASLT_MATMUL_TILE_16x320 "CUBLASLT_MATMUL_TILE_16x320" = 51
74
+ CUBLASLT_MATMUL_TILE_16x384 "CUBLASLT_MATMUL_TILE_16x384" = 52
75
+ CUBLASLT_MATMUL_TILE_16x448 "CUBLASLT_MATMUL_TILE_16x448" = 53
76
+ CUBLASLT_MATMUL_TILE_16x512 "CUBLASLT_MATMUL_TILE_16x512" = 54
77
+ CUBLASLT_MATMUL_TILE_16x576 "CUBLASLT_MATMUL_TILE_16x576" = 55
78
+ CUBLASLT_MATMUL_TILE_16x640 "CUBLASLT_MATMUL_TILE_16x640" = 56
79
+ CUBLASLT_MATMUL_TILE_16x704 "CUBLASLT_MATMUL_TILE_16x704" = 57
80
+ CUBLASLT_MATMUL_TILE_16x768 "CUBLASLT_MATMUL_TILE_16x768" = 58
81
+ CUBLASLT_MATMUL_TILE_24x64 "CUBLASLT_MATMUL_TILE_24x64" = 59
82
+ CUBLASLT_MATMUL_TILE_24x128 "CUBLASLT_MATMUL_TILE_24x128" = 60
83
+ CUBLASLT_MATMUL_TILE_24x192 "CUBLASLT_MATMUL_TILE_24x192" = 61
84
+ CUBLASLT_MATMUL_TILE_24x256 "CUBLASLT_MATMUL_TILE_24x256" = 62
85
+ CUBLASLT_MATMUL_TILE_24x320 "CUBLASLT_MATMUL_TILE_24x320" = 63
86
+ CUBLASLT_MATMUL_TILE_24x384 "CUBLASLT_MATMUL_TILE_24x384" = 64
87
+ CUBLASLT_MATMUL_TILE_24x448 "CUBLASLT_MATMUL_TILE_24x448" = 65
88
+ CUBLASLT_MATMUL_TILE_24x512 "CUBLASLT_MATMUL_TILE_24x512" = 66
89
+ CUBLASLT_MATMUL_TILE_24x576 "CUBLASLT_MATMUL_TILE_24x576" = 67
90
+ CUBLASLT_MATMUL_TILE_24x640 "CUBLASLT_MATMUL_TILE_24x640" = 68
91
+ CUBLASLT_MATMUL_TILE_24x704 "CUBLASLT_MATMUL_TILE_24x704" = 69
92
+ CUBLASLT_MATMUL_TILE_24x768 "CUBLASLT_MATMUL_TILE_24x768" = 70
93
+ CUBLASLT_MATMUL_TILE_32x192 "CUBLASLT_MATMUL_TILE_32x192" = 71
94
+ CUBLASLT_MATMUL_TILE_32x320 "CUBLASLT_MATMUL_TILE_32x320" = 72
95
+ CUBLASLT_MATMUL_TILE_32x384 "CUBLASLT_MATMUL_TILE_32x384" = 73
96
+ CUBLASLT_MATMUL_TILE_32x448 "CUBLASLT_MATMUL_TILE_32x448" = 74
97
+ CUBLASLT_MATMUL_TILE_32x512 "CUBLASLT_MATMUL_TILE_32x512" = 75
98
+ CUBLASLT_MATMUL_TILE_32x576 "CUBLASLT_MATMUL_TILE_32x576" = 76
99
+ CUBLASLT_MATMUL_TILE_32x640 "CUBLASLT_MATMUL_TILE_32x640" = 77
100
+ CUBLASLT_MATMUL_TILE_32x704 "CUBLASLT_MATMUL_TILE_32x704" = 78
101
+ CUBLASLT_MATMUL_TILE_32x768 "CUBLASLT_MATMUL_TILE_32x768" = 79
102
+ CUBLASLT_MATMUL_TILE_40x64 "CUBLASLT_MATMUL_TILE_40x64" = 80
103
+ CUBLASLT_MATMUL_TILE_40x128 "CUBLASLT_MATMUL_TILE_40x128" = 81
104
+ CUBLASLT_MATMUL_TILE_40x192 "CUBLASLT_MATMUL_TILE_40x192" = 82
105
+ CUBLASLT_MATMUL_TILE_40x256 "CUBLASLT_MATMUL_TILE_40x256" = 83
106
+ CUBLASLT_MATMUL_TILE_40x320 "CUBLASLT_MATMUL_TILE_40x320" = 84
107
+ CUBLASLT_MATMUL_TILE_40x384 "CUBLASLT_MATMUL_TILE_40x384" = 85
108
+ CUBLASLT_MATMUL_TILE_40x448 "CUBLASLT_MATMUL_TILE_40x448" = 86
109
+ CUBLASLT_MATMUL_TILE_40x512 "CUBLASLT_MATMUL_TILE_40x512" = 87
110
+ CUBLASLT_MATMUL_TILE_40x576 "CUBLASLT_MATMUL_TILE_40x576" = 88
111
+ CUBLASLT_MATMUL_TILE_40x640 "CUBLASLT_MATMUL_TILE_40x640" = 89
112
+ CUBLASLT_MATMUL_TILE_40x704 "CUBLASLT_MATMUL_TILE_40x704" = 90
113
+ CUBLASLT_MATMUL_TILE_40x768 "CUBLASLT_MATMUL_TILE_40x768" = 91
114
+ CUBLASLT_MATMUL_TILE_48x64 "CUBLASLT_MATMUL_TILE_48x64" = 92
115
+ CUBLASLT_MATMUL_TILE_48x128 "CUBLASLT_MATMUL_TILE_48x128" = 93
116
+ CUBLASLT_MATMUL_TILE_48x192 "CUBLASLT_MATMUL_TILE_48x192" = 94
117
+ CUBLASLT_MATMUL_TILE_48x256 "CUBLASLT_MATMUL_TILE_48x256" = 95
118
+ CUBLASLT_MATMUL_TILE_48x320 "CUBLASLT_MATMUL_TILE_48x320" = 96
119
+ CUBLASLT_MATMUL_TILE_48x384 "CUBLASLT_MATMUL_TILE_48x384" = 97
120
+ CUBLASLT_MATMUL_TILE_48x448 "CUBLASLT_MATMUL_TILE_48x448" = 98
121
+ CUBLASLT_MATMUL_TILE_48x512 "CUBLASLT_MATMUL_TILE_48x512" = 99
122
+ CUBLASLT_MATMUL_TILE_48x576 "CUBLASLT_MATMUL_TILE_48x576" = 100
123
+ CUBLASLT_MATMUL_TILE_48x640 "CUBLASLT_MATMUL_TILE_48x640" = 101
124
+ CUBLASLT_MATMUL_TILE_48x704 "CUBLASLT_MATMUL_TILE_48x704" = 102
125
+ CUBLASLT_MATMUL_TILE_48x768 "CUBLASLT_MATMUL_TILE_48x768" = 103
126
+ CUBLASLT_MATMUL_TILE_56x64 "CUBLASLT_MATMUL_TILE_56x64" = 104
127
+ CUBLASLT_MATMUL_TILE_56x128 "CUBLASLT_MATMUL_TILE_56x128" = 105
128
+ CUBLASLT_MATMUL_TILE_56x192 "CUBLASLT_MATMUL_TILE_56x192" = 106
129
+ CUBLASLT_MATMUL_TILE_56x256 "CUBLASLT_MATMUL_TILE_56x256" = 107
130
+ CUBLASLT_MATMUL_TILE_56x320 "CUBLASLT_MATMUL_TILE_56x320" = 108
131
+ CUBLASLT_MATMUL_TILE_56x384 "CUBLASLT_MATMUL_TILE_56x384" = 109
132
+ CUBLASLT_MATMUL_TILE_56x448 "CUBLASLT_MATMUL_TILE_56x448" = 110
133
+ CUBLASLT_MATMUL_TILE_56x512 "CUBLASLT_MATMUL_TILE_56x512" = 111
134
+ CUBLASLT_MATMUL_TILE_56x576 "CUBLASLT_MATMUL_TILE_56x576" = 112
135
+ CUBLASLT_MATMUL_TILE_56x640 "CUBLASLT_MATMUL_TILE_56x640" = 113
136
+ CUBLASLT_MATMUL_TILE_56x704 "CUBLASLT_MATMUL_TILE_56x704" = 114
137
+ CUBLASLT_MATMUL_TILE_56x768 "CUBLASLT_MATMUL_TILE_56x768" = 115
138
+ CUBLASLT_MATMUL_TILE_64x192 "CUBLASLT_MATMUL_TILE_64x192" = 116
139
+ CUBLASLT_MATMUL_TILE_64x320 "CUBLASLT_MATMUL_TILE_64x320" = 117
140
+ CUBLASLT_MATMUL_TILE_64x384 "CUBLASLT_MATMUL_TILE_64x384" = 118
141
+ CUBLASLT_MATMUL_TILE_64x448 "CUBLASLT_MATMUL_TILE_64x448" = 119
142
+ CUBLASLT_MATMUL_TILE_64x576 "CUBLASLT_MATMUL_TILE_64x576" = 120
143
+ CUBLASLT_MATMUL_TILE_64x640 "CUBLASLT_MATMUL_TILE_64x640" = 121
144
+ CUBLASLT_MATMUL_TILE_64x704 "CUBLASLT_MATMUL_TILE_64x704" = 122
145
+ CUBLASLT_MATMUL_TILE_64x768 "CUBLASLT_MATMUL_TILE_64x768" = 123
146
+ CUBLASLT_MATMUL_TILE_72x64 "CUBLASLT_MATMUL_TILE_72x64" = 124
147
+ CUBLASLT_MATMUL_TILE_72x128 "CUBLASLT_MATMUL_TILE_72x128" = 125
148
+ CUBLASLT_MATMUL_TILE_72x192 "CUBLASLT_MATMUL_TILE_72x192" = 126
149
+ CUBLASLT_MATMUL_TILE_72x256 "CUBLASLT_MATMUL_TILE_72x256" = 127
150
+ CUBLASLT_MATMUL_TILE_72x320 "CUBLASLT_MATMUL_TILE_72x320" = 128
151
+ CUBLASLT_MATMUL_TILE_72x384 "CUBLASLT_MATMUL_TILE_72x384" = 129
152
+ CUBLASLT_MATMUL_TILE_72x448 "CUBLASLT_MATMUL_TILE_72x448" = 130
153
+ CUBLASLT_MATMUL_TILE_72x512 "CUBLASLT_MATMUL_TILE_72x512" = 131
154
+ CUBLASLT_MATMUL_TILE_72x576 "CUBLASLT_MATMUL_TILE_72x576" = 132
155
+ CUBLASLT_MATMUL_TILE_72x640 "CUBLASLT_MATMUL_TILE_72x640" = 133
156
+ CUBLASLT_MATMUL_TILE_80x64 "CUBLASLT_MATMUL_TILE_80x64" = 134
157
+ CUBLASLT_MATMUL_TILE_80x128 "CUBLASLT_MATMUL_TILE_80x128" = 135
158
+ CUBLASLT_MATMUL_TILE_80x192 "CUBLASLT_MATMUL_TILE_80x192" = 136
159
+ CUBLASLT_MATMUL_TILE_80x256 "CUBLASLT_MATMUL_TILE_80x256" = 137
160
+ CUBLASLT_MATMUL_TILE_80x320 "CUBLASLT_MATMUL_TILE_80x320" = 138
161
+ CUBLASLT_MATMUL_TILE_80x384 "CUBLASLT_MATMUL_TILE_80x384" = 139
162
+ CUBLASLT_MATMUL_TILE_80x448 "CUBLASLT_MATMUL_TILE_80x448" = 140
163
+ CUBLASLT_MATMUL_TILE_80x512 "CUBLASLT_MATMUL_TILE_80x512" = 141
164
+ CUBLASLT_MATMUL_TILE_80x576 "CUBLASLT_MATMUL_TILE_80x576" = 142
165
+ CUBLASLT_MATMUL_TILE_88x64 "CUBLASLT_MATMUL_TILE_88x64" = 143
166
+ CUBLASLT_MATMUL_TILE_88x128 "CUBLASLT_MATMUL_TILE_88x128" = 144
167
+ CUBLASLT_MATMUL_TILE_88x192 "CUBLASLT_MATMUL_TILE_88x192" = 145
168
+ CUBLASLT_MATMUL_TILE_88x256 "CUBLASLT_MATMUL_TILE_88x256" = 146
169
+ CUBLASLT_MATMUL_TILE_88x320 "CUBLASLT_MATMUL_TILE_88x320" = 147
170
+ CUBLASLT_MATMUL_TILE_88x384 "CUBLASLT_MATMUL_TILE_88x384" = 148
171
+ CUBLASLT_MATMUL_TILE_88x448 "CUBLASLT_MATMUL_TILE_88x448" = 149
172
+ CUBLASLT_MATMUL_TILE_88x512 "CUBLASLT_MATMUL_TILE_88x512" = 150
173
+ CUBLASLT_MATMUL_TILE_96x192 "CUBLASLT_MATMUL_TILE_96x192" = 151
174
+ CUBLASLT_MATMUL_TILE_96x256 "CUBLASLT_MATMUL_TILE_96x256" = 152
175
+ CUBLASLT_MATMUL_TILE_96x320 "CUBLASLT_MATMUL_TILE_96x320" = 153
176
+ CUBLASLT_MATMUL_TILE_96x384 "CUBLASLT_MATMUL_TILE_96x384" = 154
177
+ CUBLASLT_MATMUL_TILE_96x448 "CUBLASLT_MATMUL_TILE_96x448" = 155
178
+ CUBLASLT_MATMUL_TILE_96x512 "CUBLASLT_MATMUL_TILE_96x512" = 156
179
+ CUBLASLT_MATMUL_TILE_104x64 "CUBLASLT_MATMUL_TILE_104x64" = 157
180
+ CUBLASLT_MATMUL_TILE_104x128 "CUBLASLT_MATMUL_TILE_104x128" = 158
181
+ CUBLASLT_MATMUL_TILE_104x192 "CUBLASLT_MATMUL_TILE_104x192" = 159
182
+ CUBLASLT_MATMUL_TILE_104x256 "CUBLASLT_MATMUL_TILE_104x256" = 160
183
+ CUBLASLT_MATMUL_TILE_104x320 "CUBLASLT_MATMUL_TILE_104x320" = 161
184
+ CUBLASLT_MATMUL_TILE_104x384 "CUBLASLT_MATMUL_TILE_104x384" = 162
185
+ CUBLASLT_MATMUL_TILE_104x448 "CUBLASLT_MATMUL_TILE_104x448" = 163
186
+ CUBLASLT_MATMUL_TILE_112x64 "CUBLASLT_MATMUL_TILE_112x64" = 164
187
+ CUBLASLT_MATMUL_TILE_112x128 "CUBLASLT_MATMUL_TILE_112x128" = 165
188
+ CUBLASLT_MATMUL_TILE_112x192 "CUBLASLT_MATMUL_TILE_112x192" = 166
189
+ CUBLASLT_MATMUL_TILE_112x256 "CUBLASLT_MATMUL_TILE_112x256" = 167
190
+ CUBLASLT_MATMUL_TILE_112x320 "CUBLASLT_MATMUL_TILE_112x320" = 168
191
+ CUBLASLT_MATMUL_TILE_112x384 "CUBLASLT_MATMUL_TILE_112x384" = 169
192
+ CUBLASLT_MATMUL_TILE_120x64 "CUBLASLT_MATMUL_TILE_120x64" = 170
193
+ CUBLASLT_MATMUL_TILE_120x128 "CUBLASLT_MATMUL_TILE_120x128" = 171
194
+ CUBLASLT_MATMUL_TILE_120x192 "CUBLASLT_MATMUL_TILE_120x192" = 172
195
+ CUBLASLT_MATMUL_TILE_120x256 "CUBLASLT_MATMUL_TILE_120x256" = 173
196
+ CUBLASLT_MATMUL_TILE_120x320 "CUBLASLT_MATMUL_TILE_120x320" = 174
197
+ CUBLASLT_MATMUL_TILE_120x384 "CUBLASLT_MATMUL_TILE_120x384" = 175
198
+ CUBLASLT_MATMUL_TILE_128x320 "CUBLASLT_MATMUL_TILE_128x320" = 176
199
+ CUBLASLT_MATMUL_TILE_128x384 "CUBLASLT_MATMUL_TILE_128x384" = 177
200
+ CUBLASLT_MATMUL_TILE_136x64 "CUBLASLT_MATMUL_TILE_136x64" = 178
201
+ CUBLASLT_MATMUL_TILE_136x128 "CUBLASLT_MATMUL_TILE_136x128" = 179
202
+ CUBLASLT_MATMUL_TILE_136x192 "CUBLASLT_MATMUL_TILE_136x192" = 180
203
+ CUBLASLT_MATMUL_TILE_136x256 "CUBLASLT_MATMUL_TILE_136x256" = 181
204
+ CUBLASLT_MATMUL_TILE_136x320 "CUBLASLT_MATMUL_TILE_136x320" = 182
205
+ CUBLASLT_MATMUL_TILE_144x64 "CUBLASLT_MATMUL_TILE_144x64" = 183
206
+ CUBLASLT_MATMUL_TILE_144x128 "CUBLASLT_MATMUL_TILE_144x128" = 184
207
+ CUBLASLT_MATMUL_TILE_144x192 "CUBLASLT_MATMUL_TILE_144x192" = 185
208
+ CUBLASLT_MATMUL_TILE_144x256 "CUBLASLT_MATMUL_TILE_144x256" = 186
209
+ CUBLASLT_MATMUL_TILE_144x320 "CUBLASLT_MATMUL_TILE_144x320" = 187
210
+ CUBLASLT_MATMUL_TILE_152x64 "CUBLASLT_MATMUL_TILE_152x64" = 188
211
+ CUBLASLT_MATMUL_TILE_152x128 "CUBLASLT_MATMUL_TILE_152x128" = 189
212
+ CUBLASLT_MATMUL_TILE_152x192 "CUBLASLT_MATMUL_TILE_152x192" = 190
213
+ CUBLASLT_MATMUL_TILE_152x256 "CUBLASLT_MATMUL_TILE_152x256" = 191
214
+ CUBLASLT_MATMUL_TILE_152x320 "CUBLASLT_MATMUL_TILE_152x320" = 192
215
+ CUBLASLT_MATMUL_TILE_160x64 "CUBLASLT_MATMUL_TILE_160x64" = 193
216
+ CUBLASLT_MATMUL_TILE_160x192 "CUBLASLT_MATMUL_TILE_160x192" = 194
217
+ CUBLASLT_MATMUL_TILE_160x256 "CUBLASLT_MATMUL_TILE_160x256" = 195
218
+ CUBLASLT_MATMUL_TILE_168x64 "CUBLASLT_MATMUL_TILE_168x64" = 196
219
+ CUBLASLT_MATMUL_TILE_168x128 "CUBLASLT_MATMUL_TILE_168x128" = 197
220
+ CUBLASLT_MATMUL_TILE_168x192 "CUBLASLT_MATMUL_TILE_168x192" = 198
221
+ CUBLASLT_MATMUL_TILE_168x256 "CUBLASLT_MATMUL_TILE_168x256" = 199
222
+ CUBLASLT_MATMUL_TILE_176x64 "CUBLASLT_MATMUL_TILE_176x64" = 200
223
+ CUBLASLT_MATMUL_TILE_176x128 "CUBLASLT_MATMUL_TILE_176x128" = 201
224
+ CUBLASLT_MATMUL_TILE_176x192 "CUBLASLT_MATMUL_TILE_176x192" = 202
225
+ CUBLASLT_MATMUL_TILE_176x256 "CUBLASLT_MATMUL_TILE_176x256" = 203
226
+ CUBLASLT_MATMUL_TILE_184x64 "CUBLASLT_MATMUL_TILE_184x64" = 204
227
+ CUBLASLT_MATMUL_TILE_184x128 "CUBLASLT_MATMUL_TILE_184x128" = 205
228
+ CUBLASLT_MATMUL_TILE_184x192 "CUBLASLT_MATMUL_TILE_184x192" = 206
229
+ CUBLASLT_MATMUL_TILE_184x256 "CUBLASLT_MATMUL_TILE_184x256" = 207
230
+ CUBLASLT_MATMUL_TILE_192x64 "CUBLASLT_MATMUL_TILE_192x64" = 208
231
+ CUBLASLT_MATMUL_TILE_192x192 "CUBLASLT_MATMUL_TILE_192x192" = 209
232
+ CUBLASLT_MATMUL_TILE_192x256 "CUBLASLT_MATMUL_TILE_192x256" = 210
233
+ CUBLASLT_MATMUL_TILE_200x64 "CUBLASLT_MATMUL_TILE_200x64" = 211
234
+ CUBLASLT_MATMUL_TILE_200x128 "CUBLASLT_MATMUL_TILE_200x128" = 212
235
+ CUBLASLT_MATMUL_TILE_200x192 "CUBLASLT_MATMUL_TILE_200x192" = 213
236
+ CUBLASLT_MATMUL_TILE_208x64 "CUBLASLT_MATMUL_TILE_208x64" = 214
237
+ CUBLASLT_MATMUL_TILE_208x128 "CUBLASLT_MATMUL_TILE_208x128" = 215
238
+ CUBLASLT_MATMUL_TILE_208x192 "CUBLASLT_MATMUL_TILE_208x192" = 216
239
+ CUBLASLT_MATMUL_TILE_216x64 "CUBLASLT_MATMUL_TILE_216x64" = 217
240
+ CUBLASLT_MATMUL_TILE_216x128 "CUBLASLT_MATMUL_TILE_216x128" = 218
241
+ CUBLASLT_MATMUL_TILE_216x192 "CUBLASLT_MATMUL_TILE_216x192" = 219
242
+ CUBLASLT_MATMUL_TILE_224x64 "CUBLASLT_MATMUL_TILE_224x64" = 220
243
+ CUBLASLT_MATMUL_TILE_224x128 "CUBLASLT_MATMUL_TILE_224x128" = 221
244
+ CUBLASLT_MATMUL_TILE_224x192 "CUBLASLT_MATMUL_TILE_224x192" = 222
245
+ CUBLASLT_MATMUL_TILE_232x64 "CUBLASLT_MATMUL_TILE_232x64" = 223
246
+ CUBLASLT_MATMUL_TILE_232x128 "CUBLASLT_MATMUL_TILE_232x128" = 224
247
+ CUBLASLT_MATMUL_TILE_232x192 "CUBLASLT_MATMUL_TILE_232x192" = 225
248
+ CUBLASLT_MATMUL_TILE_240x64 "CUBLASLT_MATMUL_TILE_240x64" = 226
249
+ CUBLASLT_MATMUL_TILE_240x128 "CUBLASLT_MATMUL_TILE_240x128" = 227
250
+ CUBLASLT_MATMUL_TILE_240x192 "CUBLASLT_MATMUL_TILE_240x192" = 228
251
+ CUBLASLT_MATMUL_TILE_248x64 "CUBLASLT_MATMUL_TILE_248x64" = 229
252
+ CUBLASLT_MATMUL_TILE_248x128 "CUBLASLT_MATMUL_TILE_248x128" = 230
253
+ CUBLASLT_MATMUL_TILE_248x192 "CUBLASLT_MATMUL_TILE_248x192" = 231
254
+ CUBLASLT_MATMUL_TILE_256x192 "CUBLASLT_MATMUL_TILE_256x192" = 232
255
+ CUBLASLT_MATMUL_TILE_264x64 "CUBLASLT_MATMUL_TILE_264x64" = 233
256
+ CUBLASLT_MATMUL_TILE_264x128 "CUBLASLT_MATMUL_TILE_264x128" = 234
257
+ CUBLASLT_MATMUL_TILE_272x64 "CUBLASLT_MATMUL_TILE_272x64" = 235
258
+ CUBLASLT_MATMUL_TILE_272x128 "CUBLASLT_MATMUL_TILE_272x128" = 236
259
+ CUBLASLT_MATMUL_TILE_280x64 "CUBLASLT_MATMUL_TILE_280x64" = 237
260
+ CUBLASLT_MATMUL_TILE_280x128 "CUBLASLT_MATMUL_TILE_280x128" = 238
261
+ CUBLASLT_MATMUL_TILE_288x64 "CUBLASLT_MATMUL_TILE_288x64" = 239
262
+ CUBLASLT_MATMUL_TILE_288x128 "CUBLASLT_MATMUL_TILE_288x128" = 240
263
+ CUBLASLT_MATMUL_TILE_296x64 "CUBLASLT_MATMUL_TILE_296x64" = 241
264
+ CUBLASLT_MATMUL_TILE_296x128 "CUBLASLT_MATMUL_TILE_296x128" = 242
265
+ CUBLASLT_MATMUL_TILE_304x64 "CUBLASLT_MATMUL_TILE_304x64" = 243
266
+ CUBLASLT_MATMUL_TILE_304x128 "CUBLASLT_MATMUL_TILE_304x128" = 244
267
+ CUBLASLT_MATMUL_TILE_312x64 "CUBLASLT_MATMUL_TILE_312x64" = 245
268
+ CUBLASLT_MATMUL_TILE_312x128 "CUBLASLT_MATMUL_TILE_312x128" = 246
269
+ CUBLASLT_MATMUL_TILE_320x64 "CUBLASLT_MATMUL_TILE_320x64" = 247
270
+ CUBLASLT_MATMUL_TILE_320x128 "CUBLASLT_MATMUL_TILE_320x128" = 248
271
+ CUBLASLT_MATMUL_TILE_328x64 "CUBLASLT_MATMUL_TILE_328x64" = 249
272
+ CUBLASLT_MATMUL_TILE_328x128 "CUBLASLT_MATMUL_TILE_328x128" = 250
273
+ CUBLASLT_MATMUL_TILE_336x64 "CUBLASLT_MATMUL_TILE_336x64" = 251
274
+ CUBLASLT_MATMUL_TILE_336x128 "CUBLASLT_MATMUL_TILE_336x128" = 252
275
+ CUBLASLT_MATMUL_TILE_344x64 "CUBLASLT_MATMUL_TILE_344x64" = 253
276
+ CUBLASLT_MATMUL_TILE_344x128 "CUBLASLT_MATMUL_TILE_344x128" = 254
277
+ CUBLASLT_MATMUL_TILE_352x64 "CUBLASLT_MATMUL_TILE_352x64" = 255
278
+ CUBLASLT_MATMUL_TILE_352x128 "CUBLASLT_MATMUL_TILE_352x128" = 256
279
+ CUBLASLT_MATMUL_TILE_360x64 "CUBLASLT_MATMUL_TILE_360x64" = 257
280
+ CUBLASLT_MATMUL_TILE_360x128 "CUBLASLT_MATMUL_TILE_360x128" = 258
281
+ CUBLASLT_MATMUL_TILE_368x64 "CUBLASLT_MATMUL_TILE_368x64" = 259
282
+ CUBLASLT_MATMUL_TILE_368x128 "CUBLASLT_MATMUL_TILE_368x128" = 260
283
+ CUBLASLT_MATMUL_TILE_376x64 "CUBLASLT_MATMUL_TILE_376x64" = 261
284
+ CUBLASLT_MATMUL_TILE_376x128 "CUBLASLT_MATMUL_TILE_376x128" = 262
285
+ CUBLASLT_MATMUL_TILE_384x64 "CUBLASLT_MATMUL_TILE_384x64" = 263
286
+ CUBLASLT_MATMUL_TILE_384x128 "CUBLASLT_MATMUL_TILE_384x128" = 264
287
+ CUBLASLT_MATMUL_TILE_392x64 "CUBLASLT_MATMUL_TILE_392x64" = 265
288
+ CUBLASLT_MATMUL_TILE_400x64 "CUBLASLT_MATMUL_TILE_400x64" = 266
289
+ CUBLASLT_MATMUL_TILE_408x64 "CUBLASLT_MATMUL_TILE_408x64" = 267
290
+ CUBLASLT_MATMUL_TILE_416x64 "CUBLASLT_MATMUL_TILE_416x64" = 268
291
+ CUBLASLT_MATMUL_TILE_424x64 "CUBLASLT_MATMUL_TILE_424x64" = 269
292
+ CUBLASLT_MATMUL_TILE_432x64 "CUBLASLT_MATMUL_TILE_432x64" = 270
293
+ CUBLASLT_MATMUL_TILE_440x64 "CUBLASLT_MATMUL_TILE_440x64" = 271
294
+ CUBLASLT_MATMUL_TILE_448x64 "CUBLASLT_MATMUL_TILE_448x64" = 272
295
+ CUBLASLT_MATMUL_TILE_456x64 "CUBLASLT_MATMUL_TILE_456x64" = 273
296
+ CUBLASLT_MATMUL_TILE_464x64 "CUBLASLT_MATMUL_TILE_464x64" = 274
297
+ CUBLASLT_MATMUL_TILE_472x64 "CUBLASLT_MATMUL_TILE_472x64" = 275
298
+ CUBLASLT_MATMUL_TILE_480x64 "CUBLASLT_MATMUL_TILE_480x64" = 276
299
+ CUBLASLT_MATMUL_TILE_488x64 "CUBLASLT_MATMUL_TILE_488x64" = 277
300
+ CUBLASLT_MATMUL_TILE_496x64 "CUBLASLT_MATMUL_TILE_496x64" = 278
301
+ CUBLASLT_MATMUL_TILE_504x64 "CUBLASLT_MATMUL_TILE_504x64" = 279
302
+ CUBLASLT_MATMUL_TILE_520x64 "CUBLASLT_MATMUL_TILE_520x64" = 280
303
+ CUBLASLT_MATMUL_TILE_528x64 "CUBLASLT_MATMUL_TILE_528x64" = 281
304
+ CUBLASLT_MATMUL_TILE_536x64 "CUBLASLT_MATMUL_TILE_536x64" = 282
305
+ CUBLASLT_MATMUL_TILE_544x64 "CUBLASLT_MATMUL_TILE_544x64" = 283
306
+ CUBLASLT_MATMUL_TILE_552x64 "CUBLASLT_MATMUL_TILE_552x64" = 284
307
+ CUBLASLT_MATMUL_TILE_560x64 "CUBLASLT_MATMUL_TILE_560x64" = 285
308
+ CUBLASLT_MATMUL_TILE_568x64 "CUBLASLT_MATMUL_TILE_568x64" = 286
309
+ CUBLASLT_MATMUL_TILE_576x64 "CUBLASLT_MATMUL_TILE_576x64" = 287
310
+ CUBLASLT_MATMUL_TILE_584x64 "CUBLASLT_MATMUL_TILE_584x64" = 288
311
+ CUBLASLT_MATMUL_TILE_592x64 "CUBLASLT_MATMUL_TILE_592x64" = 289
312
+ CUBLASLT_MATMUL_TILE_600x64 "CUBLASLT_MATMUL_TILE_600x64" = 290
313
+ CUBLASLT_MATMUL_TILE_608x64 "CUBLASLT_MATMUL_TILE_608x64" = 291
314
+ CUBLASLT_MATMUL_TILE_616x64 "CUBLASLT_MATMUL_TILE_616x64" = 292
315
+ CUBLASLT_MATMUL_TILE_624x64 "CUBLASLT_MATMUL_TILE_624x64" = 293
316
+ CUBLASLT_MATMUL_TILE_632x64 "CUBLASLT_MATMUL_TILE_632x64" = 294
317
+ CUBLASLT_MATMUL_TILE_640x64 "CUBLASLT_MATMUL_TILE_640x64" = 295
318
+ CUBLASLT_MATMUL_TILE_648x64 "CUBLASLT_MATMUL_TILE_648x64" = 296
319
+ CUBLASLT_MATMUL_TILE_656x64 "CUBLASLT_MATMUL_TILE_656x64" = 297
320
+ CUBLASLT_MATMUL_TILE_664x64 "CUBLASLT_MATMUL_TILE_664x64" = 298
321
+ CUBLASLT_MATMUL_TILE_672x64 "CUBLASLT_MATMUL_TILE_672x64" = 299
322
+ CUBLASLT_MATMUL_TILE_680x64 "CUBLASLT_MATMUL_TILE_680x64" = 300
323
+ CUBLASLT_MATMUL_TILE_688x64 "CUBLASLT_MATMUL_TILE_688x64" = 301
324
+ CUBLASLT_MATMUL_TILE_696x64 "CUBLASLT_MATMUL_TILE_696x64" = 302
325
+ CUBLASLT_MATMUL_TILE_704x64 "CUBLASLT_MATMUL_TILE_704x64" = 303
326
+ CUBLASLT_MATMUL_TILE_712x64 "CUBLASLT_MATMUL_TILE_712x64" = 304
327
+ CUBLASLT_MATMUL_TILE_720x64 "CUBLASLT_MATMUL_TILE_720x64" = 305
328
+ CUBLASLT_MATMUL_TILE_728x64 "CUBLASLT_MATMUL_TILE_728x64" = 306
329
+ CUBLASLT_MATMUL_TILE_736x64 "CUBLASLT_MATMUL_TILE_736x64" = 307
330
+ CUBLASLT_MATMUL_TILE_744x64 "CUBLASLT_MATMUL_TILE_744x64" = 308
331
+ CUBLASLT_MATMUL_TILE_752x64 "CUBLASLT_MATMUL_TILE_752x64" = 309
332
+ CUBLASLT_MATMUL_TILE_760x64 "CUBLASLT_MATMUL_TILE_760x64" = 310
333
+ CUBLASLT_MATMUL_TILE_768x64 "CUBLASLT_MATMUL_TILE_768x64" = 311
334
+ CUBLASLT_MATMUL_TILE_64x16 "CUBLASLT_MATMUL_TILE_64x16" = 312
335
+ CUBLASLT_MATMUL_TILE_64x24 "CUBLASLT_MATMUL_TILE_64x24" = 313
336
+ CUBLASLT_MATMUL_TILE_64x40 "CUBLASLT_MATMUL_TILE_64x40" = 314
337
+ CUBLASLT_MATMUL_TILE_64x48 "CUBLASLT_MATMUL_TILE_64x48" = 315
338
+ CUBLASLT_MATMUL_TILE_64x56 "CUBLASLT_MATMUL_TILE_64x56" = 316
339
+ CUBLASLT_MATMUL_TILE_64x72 "CUBLASLT_MATMUL_TILE_64x72" = 317
340
+ CUBLASLT_MATMUL_TILE_64x80 "CUBLASLT_MATMUL_TILE_64x80" = 318
341
+ CUBLASLT_MATMUL_TILE_64x88 "CUBLASLT_MATMUL_TILE_64x88" = 319
342
+ CUBLASLT_MATMUL_TILE_64x104 "CUBLASLT_MATMUL_TILE_64x104" = 320
343
+ CUBLASLT_MATMUL_TILE_64x112 "CUBLASLT_MATMUL_TILE_64x112" = 321
344
+ CUBLASLT_MATMUL_TILE_64x120 "CUBLASLT_MATMUL_TILE_64x120" = 322
345
+ CUBLASLT_MATMUL_TILE_64x136 "CUBLASLT_MATMUL_TILE_64x136" = 323
346
+ CUBLASLT_MATMUL_TILE_64x144 "CUBLASLT_MATMUL_TILE_64x144" = 324
347
+ CUBLASLT_MATMUL_TILE_64x152 "CUBLASLT_MATMUL_TILE_64x152" = 325
348
+ CUBLASLT_MATMUL_TILE_64x160 "CUBLASLT_MATMUL_TILE_64x160" = 326
349
+ CUBLASLT_MATMUL_TILE_64x168 "CUBLASLT_MATMUL_TILE_64x168" = 327
350
+ CUBLASLT_MATMUL_TILE_64x176 "CUBLASLT_MATMUL_TILE_64x176" = 328
351
+ CUBLASLT_MATMUL_TILE_64x184 "CUBLASLT_MATMUL_TILE_64x184" = 329
352
+ CUBLASLT_MATMUL_TILE_64x200 "CUBLASLT_MATMUL_TILE_64x200" = 330
353
+ CUBLASLT_MATMUL_TILE_64x208 "CUBLASLT_MATMUL_TILE_64x208" = 331
354
+ CUBLASLT_MATMUL_TILE_64x216 "CUBLASLT_MATMUL_TILE_64x216" = 332
355
+ CUBLASLT_MATMUL_TILE_64x224 "CUBLASLT_MATMUL_TILE_64x224" = 333
356
+ CUBLASLT_MATMUL_TILE_64x232 "CUBLASLT_MATMUL_TILE_64x232" = 334
357
+ CUBLASLT_MATMUL_TILE_64x240 "CUBLASLT_MATMUL_TILE_64x240" = 335
358
+ CUBLASLT_MATMUL_TILE_64x248 "CUBLASLT_MATMUL_TILE_64x248" = 336
359
+ CUBLASLT_MATMUL_TILE_64x264 "CUBLASLT_MATMUL_TILE_64x264" = 337
360
+ CUBLASLT_MATMUL_TILE_64x272 "CUBLASLT_MATMUL_TILE_64x272" = 338
361
+ CUBLASLT_MATMUL_TILE_64x280 "CUBLASLT_MATMUL_TILE_64x280" = 339
362
+ CUBLASLT_MATMUL_TILE_64x288 "CUBLASLT_MATMUL_TILE_64x288" = 340
363
+ CUBLASLT_MATMUL_TILE_64x296 "CUBLASLT_MATMUL_TILE_64x296" = 341
364
+ CUBLASLT_MATMUL_TILE_64x304 "CUBLASLT_MATMUL_TILE_64x304" = 342
365
+ CUBLASLT_MATMUL_TILE_64x312 "CUBLASLT_MATMUL_TILE_64x312" = 343
366
+ CUBLASLT_MATMUL_TILE_64x328 "CUBLASLT_MATMUL_TILE_64x328" = 344
367
+ CUBLASLT_MATMUL_TILE_64x336 "CUBLASLT_MATMUL_TILE_64x336" = 345
368
+ CUBLASLT_MATMUL_TILE_64x344 "CUBLASLT_MATMUL_TILE_64x344" = 346
369
+ CUBLASLT_MATMUL_TILE_64x352 "CUBLASLT_MATMUL_TILE_64x352" = 347
370
+ CUBLASLT_MATMUL_TILE_64x360 "CUBLASLT_MATMUL_TILE_64x360" = 348
371
+ CUBLASLT_MATMUL_TILE_64x368 "CUBLASLT_MATMUL_TILE_64x368" = 349
372
+ CUBLASLT_MATMUL_TILE_64x376 "CUBLASLT_MATMUL_TILE_64x376" = 350
373
+ CUBLASLT_MATMUL_TILE_64x392 "CUBLASLT_MATMUL_TILE_64x392" = 351
374
+ CUBLASLT_MATMUL_TILE_64x400 "CUBLASLT_MATMUL_TILE_64x400" = 352
375
+ CUBLASLT_MATMUL_TILE_64x408 "CUBLASLT_MATMUL_TILE_64x408" = 353
376
+ CUBLASLT_MATMUL_TILE_64x416 "CUBLASLT_MATMUL_TILE_64x416" = 354
377
+ CUBLASLT_MATMUL_TILE_64x424 "CUBLASLT_MATMUL_TILE_64x424" = 355
378
+ CUBLASLT_MATMUL_TILE_64x432 "CUBLASLT_MATMUL_TILE_64x432" = 356
379
+ CUBLASLT_MATMUL_TILE_64x440 "CUBLASLT_MATMUL_TILE_64x440" = 357
380
+ CUBLASLT_MATMUL_TILE_64x456 "CUBLASLT_MATMUL_TILE_64x456" = 358
381
+ CUBLASLT_MATMUL_TILE_64x464 "CUBLASLT_MATMUL_TILE_64x464" = 359
382
+ CUBLASLT_MATMUL_TILE_64x472 "CUBLASLT_MATMUL_TILE_64x472" = 360
383
+ CUBLASLT_MATMUL_TILE_64x480 "CUBLASLT_MATMUL_TILE_64x480" = 361
384
+ CUBLASLT_MATMUL_TILE_64x488 "CUBLASLT_MATMUL_TILE_64x488" = 362
385
+ CUBLASLT_MATMUL_TILE_64x496 "CUBLASLT_MATMUL_TILE_64x496" = 363
386
+ CUBLASLT_MATMUL_TILE_64x504 "CUBLASLT_MATMUL_TILE_64x504" = 364
387
+ CUBLASLT_MATMUL_TILE_64x520 "CUBLASLT_MATMUL_TILE_64x520" = 365
388
+ CUBLASLT_MATMUL_TILE_64x528 "CUBLASLT_MATMUL_TILE_64x528" = 366
389
+ CUBLASLT_MATMUL_TILE_64x536 "CUBLASLT_MATMUL_TILE_64x536" = 367
390
+ CUBLASLT_MATMUL_TILE_64x544 "CUBLASLT_MATMUL_TILE_64x544" = 368
391
+ CUBLASLT_MATMUL_TILE_64x552 "CUBLASLT_MATMUL_TILE_64x552" = 369
392
+ CUBLASLT_MATMUL_TILE_64x560 "CUBLASLT_MATMUL_TILE_64x560" = 370
393
+ CUBLASLT_MATMUL_TILE_64x568 "CUBLASLT_MATMUL_TILE_64x568" = 371
394
+ CUBLASLT_MATMUL_TILE_64x584 "CUBLASLT_MATMUL_TILE_64x584" = 372
395
+ CUBLASLT_MATMUL_TILE_64x592 "CUBLASLT_MATMUL_TILE_64x592" = 373
396
+ CUBLASLT_MATMUL_TILE_64x600 "CUBLASLT_MATMUL_TILE_64x600" = 374
397
+ CUBLASLT_MATMUL_TILE_64x608 "CUBLASLT_MATMUL_TILE_64x608" = 375
398
+ CUBLASLT_MATMUL_TILE_64x616 "CUBLASLT_MATMUL_TILE_64x616" = 376
399
+ CUBLASLT_MATMUL_TILE_64x624 "CUBLASLT_MATMUL_TILE_64x624" = 377
400
+ CUBLASLT_MATMUL_TILE_64x632 "CUBLASLT_MATMUL_TILE_64x632" = 378
401
+ CUBLASLT_MATMUL_TILE_64x648 "CUBLASLT_MATMUL_TILE_64x648" = 379
402
+ CUBLASLT_MATMUL_TILE_64x656 "CUBLASLT_MATMUL_TILE_64x656" = 380
403
+ CUBLASLT_MATMUL_TILE_64x664 "CUBLASLT_MATMUL_TILE_64x664" = 381
404
+ CUBLASLT_MATMUL_TILE_64x672 "CUBLASLT_MATMUL_TILE_64x672" = 382
405
+ CUBLASLT_MATMUL_TILE_64x680 "CUBLASLT_MATMUL_TILE_64x680" = 383
406
+ CUBLASLT_MATMUL_TILE_64x688 "CUBLASLT_MATMUL_TILE_64x688" = 384
407
+ CUBLASLT_MATMUL_TILE_64x696 "CUBLASLT_MATMUL_TILE_64x696" = 385
408
+ CUBLASLT_MATMUL_TILE_64x712 "CUBLASLT_MATMUL_TILE_64x712" = 386
409
+ CUBLASLT_MATMUL_TILE_64x720 "CUBLASLT_MATMUL_TILE_64x720" = 387
410
+ CUBLASLT_MATMUL_TILE_64x728 "CUBLASLT_MATMUL_TILE_64x728" = 388
411
+ CUBLASLT_MATMUL_TILE_64x736 "CUBLASLT_MATMUL_TILE_64x736" = 389
412
+ CUBLASLT_MATMUL_TILE_64x744 "CUBLASLT_MATMUL_TILE_64x744" = 390
413
+ CUBLASLT_MATMUL_TILE_64x752 "CUBLASLT_MATMUL_TILE_64x752" = 391
414
+ CUBLASLT_MATMUL_TILE_64x760 "CUBLASLT_MATMUL_TILE_64x760" = 392
415
+ CUBLASLT_MATMUL_TILE_128x8 "CUBLASLT_MATMUL_TILE_128x8" = 393
416
+ CUBLASLT_MATMUL_TILE_128x16 "CUBLASLT_MATMUL_TILE_128x16" = 394
417
+ CUBLASLT_MATMUL_TILE_128x24 "CUBLASLT_MATMUL_TILE_128x24" = 395
418
+ CUBLASLT_MATMUL_TILE_128x40 "CUBLASLT_MATMUL_TILE_128x40" = 396
419
+ CUBLASLT_MATMUL_TILE_128x48 "CUBLASLT_MATMUL_TILE_128x48" = 397
420
+ CUBLASLT_MATMUL_TILE_128x56 "CUBLASLT_MATMUL_TILE_128x56" = 398
421
+ CUBLASLT_MATMUL_TILE_128x72 "CUBLASLT_MATMUL_TILE_128x72" = 399
422
+ CUBLASLT_MATMUL_TILE_128x80 "CUBLASLT_MATMUL_TILE_128x80" = 400
423
+ CUBLASLT_MATMUL_TILE_128x88 "CUBLASLT_MATMUL_TILE_128x88" = 401
424
+ CUBLASLT_MATMUL_TILE_128x104 "CUBLASLT_MATMUL_TILE_128x104" = 402
425
+ CUBLASLT_MATMUL_TILE_128x112 "CUBLASLT_MATMUL_TILE_128x112" = 403
426
+ CUBLASLT_MATMUL_TILE_128x120 "CUBLASLT_MATMUL_TILE_128x120" = 404
427
+ CUBLASLT_MATMUL_TILE_128x136 "CUBLASLT_MATMUL_TILE_128x136" = 405
428
+ CUBLASLT_MATMUL_TILE_128x144 "CUBLASLT_MATMUL_TILE_128x144" = 406
429
+ CUBLASLT_MATMUL_TILE_128x152 "CUBLASLT_MATMUL_TILE_128x152" = 407
430
+ CUBLASLT_MATMUL_TILE_128x168 "CUBLASLT_MATMUL_TILE_128x168" = 408
431
+ CUBLASLT_MATMUL_TILE_128x176 "CUBLASLT_MATMUL_TILE_128x176" = 409
432
+ CUBLASLT_MATMUL_TILE_128x184 "CUBLASLT_MATMUL_TILE_128x184" = 410
433
+ CUBLASLT_MATMUL_TILE_128x200 "CUBLASLT_MATMUL_TILE_128x200" = 411
434
+ CUBLASLT_MATMUL_TILE_128x208 "CUBLASLT_MATMUL_TILE_128x208" = 412
435
+ CUBLASLT_MATMUL_TILE_128x216 "CUBLASLT_MATMUL_TILE_128x216" = 413
436
+ CUBLASLT_MATMUL_TILE_128x224 "CUBLASLT_MATMUL_TILE_128x224" = 414
437
+ CUBLASLT_MATMUL_TILE_128x232 "CUBLASLT_MATMUL_TILE_128x232" = 415
438
+ CUBLASLT_MATMUL_TILE_128x240 "CUBLASLT_MATMUL_TILE_128x240" = 416
439
+ CUBLASLT_MATMUL_TILE_128x248 "CUBLASLT_MATMUL_TILE_128x248" = 417
440
+ CUBLASLT_MATMUL_TILE_128x264 "CUBLASLT_MATMUL_TILE_128x264" = 418
441
+ CUBLASLT_MATMUL_TILE_128x272 "CUBLASLT_MATMUL_TILE_128x272" = 419
442
+ CUBLASLT_MATMUL_TILE_128x280 "CUBLASLT_MATMUL_TILE_128x280" = 420
443
+ CUBLASLT_MATMUL_TILE_128x288 "CUBLASLT_MATMUL_TILE_128x288" = 421
444
+ CUBLASLT_MATMUL_TILE_128x296 "CUBLASLT_MATMUL_TILE_128x296" = 422
445
+ CUBLASLT_MATMUL_TILE_128x304 "CUBLASLT_MATMUL_TILE_128x304" = 423
446
+ CUBLASLT_MATMUL_TILE_128x312 "CUBLASLT_MATMUL_TILE_128x312" = 424
447
+ CUBLASLT_MATMUL_TILE_128x328 "CUBLASLT_MATMUL_TILE_128x328" = 425
448
+ CUBLASLT_MATMUL_TILE_128x336 "CUBLASLT_MATMUL_TILE_128x336" = 426
449
+ CUBLASLT_MATMUL_TILE_128x344 "CUBLASLT_MATMUL_TILE_128x344" = 427
450
+ CUBLASLT_MATMUL_TILE_128x352 "CUBLASLT_MATMUL_TILE_128x352" = 428
451
+ CUBLASLT_MATMUL_TILE_128x360 "CUBLASLT_MATMUL_TILE_128x360" = 429
452
+ CUBLASLT_MATMUL_TILE_128x368 "CUBLASLT_MATMUL_TILE_128x368" = 430
453
+ CUBLASLT_MATMUL_TILE_128x376 "CUBLASLT_MATMUL_TILE_128x376" = 431
454
+ CUBLASLT_MATMUL_TILE_128x392 "CUBLASLT_MATMUL_TILE_128x392" = 432
455
+ CUBLASLT_MATMUL_TILE_128x400 "CUBLASLT_MATMUL_TILE_128x400" = 433
456
+ CUBLASLT_MATMUL_TILE_128x408 "CUBLASLT_MATMUL_TILE_128x408" = 434
457
+ CUBLASLT_MATMUL_TILE_128x416 "CUBLASLT_MATMUL_TILE_128x416" = 435
458
+ CUBLASLT_MATMUL_TILE_128x424 "CUBLASLT_MATMUL_TILE_128x424" = 436
459
+ CUBLASLT_MATMUL_TILE_128x432 "CUBLASLT_MATMUL_TILE_128x432" = 437
460
+ CUBLASLT_MATMUL_TILE_128x440 "CUBLASLT_MATMUL_TILE_128x440" = 438
461
+ CUBLASLT_MATMUL_TILE_128x448 "CUBLASLT_MATMUL_TILE_128x448" = 439
462
+ CUBLASLT_MATMUL_TILE_128x456 "CUBLASLT_MATMUL_TILE_128x456" = 440
463
+ CUBLASLT_MATMUL_TILE_128x464 "CUBLASLT_MATMUL_TILE_128x464" = 441
464
+ CUBLASLT_MATMUL_TILE_128x472 "CUBLASLT_MATMUL_TILE_128x472" = 442
465
+ CUBLASLT_MATMUL_TILE_128x480 "CUBLASLT_MATMUL_TILE_128x480" = 443
466
+ CUBLASLT_MATMUL_TILE_128x488 "CUBLASLT_MATMUL_TILE_128x488" = 444
467
+ CUBLASLT_MATMUL_TILE_128x496 "CUBLASLT_MATMUL_TILE_128x496" = 445
468
+ CUBLASLT_MATMUL_TILE_128x504 "CUBLASLT_MATMUL_TILE_128x504" = 446
469
+ CUBLASLT_MATMUL_TILE_128x512 "CUBLASLT_MATMUL_TILE_128x512" = 447
470
+ CUBLASLT_MATMUL_TILE_192x8 "CUBLASLT_MATMUL_TILE_192x8" = 448
471
+ CUBLASLT_MATMUL_TILE_192x16 "CUBLASLT_MATMUL_TILE_192x16" = 449
472
+ CUBLASLT_MATMUL_TILE_192x24 "CUBLASLT_MATMUL_TILE_192x24" = 450
473
+ CUBLASLT_MATMUL_TILE_192x32 "CUBLASLT_MATMUL_TILE_192x32" = 451
474
+ CUBLASLT_MATMUL_TILE_192x40 "CUBLASLT_MATMUL_TILE_192x40" = 452
475
+ CUBLASLT_MATMUL_TILE_192x48 "CUBLASLT_MATMUL_TILE_192x48" = 453
476
+ CUBLASLT_MATMUL_TILE_192x56 "CUBLASLT_MATMUL_TILE_192x56" = 454
477
+ CUBLASLT_MATMUL_TILE_192x72 "CUBLASLT_MATMUL_TILE_192x72" = 455
478
+ CUBLASLT_MATMUL_TILE_192x80 "CUBLASLT_MATMUL_TILE_192x80" = 456
479
+ CUBLASLT_MATMUL_TILE_192x88 "CUBLASLT_MATMUL_TILE_192x88" = 457
480
+ CUBLASLT_MATMUL_TILE_192x96 "CUBLASLT_MATMUL_TILE_192x96" = 458
481
+ CUBLASLT_MATMUL_TILE_192x104 "CUBLASLT_MATMUL_TILE_192x104" = 459
482
+ CUBLASLT_MATMUL_TILE_192x112 "CUBLASLT_MATMUL_TILE_192x112" = 460
483
+ CUBLASLT_MATMUL_TILE_192x120 "CUBLASLT_MATMUL_TILE_192x120" = 461
484
+ CUBLASLT_MATMUL_TILE_192x136 "CUBLASLT_MATMUL_TILE_192x136" = 462
485
+ CUBLASLT_MATMUL_TILE_192x144 "CUBLASLT_MATMUL_TILE_192x144" = 463
486
+ CUBLASLT_MATMUL_TILE_192x152 "CUBLASLT_MATMUL_TILE_192x152" = 464
487
+ CUBLASLT_MATMUL_TILE_192x160 "CUBLASLT_MATMUL_TILE_192x160" = 465
488
+ CUBLASLT_MATMUL_TILE_192x168 "CUBLASLT_MATMUL_TILE_192x168" = 466
489
+ CUBLASLT_MATMUL_TILE_192x176 "CUBLASLT_MATMUL_TILE_192x176" = 467
490
+ CUBLASLT_MATMUL_TILE_192x184 "CUBLASLT_MATMUL_TILE_192x184" = 468
491
+ CUBLASLT_MATMUL_TILE_192x200 "CUBLASLT_MATMUL_TILE_192x200" = 469
492
+ CUBLASLT_MATMUL_TILE_192x208 "CUBLASLT_MATMUL_TILE_192x208" = 470
493
+ CUBLASLT_MATMUL_TILE_192x216 "CUBLASLT_MATMUL_TILE_192x216" = 471
494
+ CUBLASLT_MATMUL_TILE_192x224 "CUBLASLT_MATMUL_TILE_192x224" = 472
495
+ CUBLASLT_MATMUL_TILE_192x232 "CUBLASLT_MATMUL_TILE_192x232" = 473
496
+ CUBLASLT_MATMUL_TILE_192x240 "CUBLASLT_MATMUL_TILE_192x240" = 474
497
+ CUBLASLT_MATMUL_TILE_192x248 "CUBLASLT_MATMUL_TILE_192x248" = 475
498
+ CUBLASLT_MATMUL_TILE_192x264 "CUBLASLT_MATMUL_TILE_192x264" = 476
499
+ CUBLASLT_MATMUL_TILE_192x272 "CUBLASLT_MATMUL_TILE_192x272" = 477
500
+ CUBLASLT_MATMUL_TILE_192x280 "CUBLASLT_MATMUL_TILE_192x280" = 478
501
+ CUBLASLT_MATMUL_TILE_192x288 "CUBLASLT_MATMUL_TILE_192x288" = 479
502
+ CUBLASLT_MATMUL_TILE_192x296 "CUBLASLT_MATMUL_TILE_192x296" = 480
503
+ CUBLASLT_MATMUL_TILE_192x304 "CUBLASLT_MATMUL_TILE_192x304" = 481
504
+ CUBLASLT_MATMUL_TILE_192x312 "CUBLASLT_MATMUL_TILE_192x312" = 482
505
+ CUBLASLT_MATMUL_TILE_192x320 "CUBLASLT_MATMUL_TILE_192x320" = 483
506
+ CUBLASLT_MATMUL_TILE_192x328 "CUBLASLT_MATMUL_TILE_192x328" = 484
507
+ CUBLASLT_MATMUL_TILE_192x336 "CUBLASLT_MATMUL_TILE_192x336" = 485
508
+ CUBLASLT_MATMUL_TILE_256x8 "CUBLASLT_MATMUL_TILE_256x8" = 486
509
+ CUBLASLT_MATMUL_TILE_256x16 "CUBLASLT_MATMUL_TILE_256x16" = 487
510
+ CUBLASLT_MATMUL_TILE_256x24 "CUBLASLT_MATMUL_TILE_256x24" = 488
511
+ CUBLASLT_MATMUL_TILE_256x40 "CUBLASLT_MATMUL_TILE_256x40" = 489
512
+ CUBLASLT_MATMUL_TILE_256x48 "CUBLASLT_MATMUL_TILE_256x48" = 490
513
+ CUBLASLT_MATMUL_TILE_256x56 "CUBLASLT_MATMUL_TILE_256x56" = 491
514
+ CUBLASLT_MATMUL_TILE_256x72 "CUBLASLT_MATMUL_TILE_256x72" = 492
515
+ CUBLASLT_MATMUL_TILE_256x80 "CUBLASLT_MATMUL_TILE_256x80" = 493
516
+ CUBLASLT_MATMUL_TILE_256x88 "CUBLASLT_MATMUL_TILE_256x88" = 494
517
+ CUBLASLT_MATMUL_TILE_256x96 "CUBLASLT_MATMUL_TILE_256x96" = 495
518
+ CUBLASLT_MATMUL_TILE_256x104 "CUBLASLT_MATMUL_TILE_256x104" = 496
519
+ CUBLASLT_MATMUL_TILE_256x112 "CUBLASLT_MATMUL_TILE_256x112" = 497
520
+ CUBLASLT_MATMUL_TILE_256x120 "CUBLASLT_MATMUL_TILE_256x120" = 498
521
+ CUBLASLT_MATMUL_TILE_256x136 "CUBLASLT_MATMUL_TILE_256x136" = 499
522
+ CUBLASLT_MATMUL_TILE_256x144 "CUBLASLT_MATMUL_TILE_256x144" = 500
523
+ CUBLASLT_MATMUL_TILE_256x152 "CUBLASLT_MATMUL_TILE_256x152" = 501
524
+ CUBLASLT_MATMUL_TILE_256x160 "CUBLASLT_MATMUL_TILE_256x160" = 502
525
+ CUBLASLT_MATMUL_TILE_256x168 "CUBLASLT_MATMUL_TILE_256x168" = 503
526
+ CUBLASLT_MATMUL_TILE_256x176 "CUBLASLT_MATMUL_TILE_256x176" = 504
527
+ CUBLASLT_MATMUL_TILE_256x184 "CUBLASLT_MATMUL_TILE_256x184" = 505
528
+ CUBLASLT_MATMUL_TILE_256x200 "CUBLASLT_MATMUL_TILE_256x200" = 506
529
+ CUBLASLT_MATMUL_TILE_256x208 "CUBLASLT_MATMUL_TILE_256x208" = 507
530
+ CUBLASLT_MATMUL_TILE_256x216 "CUBLASLT_MATMUL_TILE_256x216" = 508
531
+ CUBLASLT_MATMUL_TILE_256x224 "CUBLASLT_MATMUL_TILE_256x224" = 509
532
+ CUBLASLT_MATMUL_TILE_256x232 "CUBLASLT_MATMUL_TILE_256x232" = 510
533
+ CUBLASLT_MATMUL_TILE_256x240 "CUBLASLT_MATMUL_TILE_256x240" = 511
534
+ CUBLASLT_MATMUL_TILE_256x248 "CUBLASLT_MATMUL_TILE_256x248" = 512
535
+ CUBLASLT_MATMUL_TILE_256x256 "CUBLASLT_MATMUL_TILE_256x256" = 513
536
+ CUBLASLT_MATMUL_TILE_320x8 "CUBLASLT_MATMUL_TILE_320x8" = 514
537
+ CUBLASLT_MATMUL_TILE_320x16 "CUBLASLT_MATMUL_TILE_320x16" = 515
538
+ CUBLASLT_MATMUL_TILE_320x24 "CUBLASLT_MATMUL_TILE_320x24" = 516
539
+ CUBLASLT_MATMUL_TILE_320x32 "CUBLASLT_MATMUL_TILE_320x32" = 517
540
+ CUBLASLT_MATMUL_TILE_320x40 "CUBLASLT_MATMUL_TILE_320x40" = 518
541
+ CUBLASLT_MATMUL_TILE_320x48 "CUBLASLT_MATMUL_TILE_320x48" = 519
542
+ CUBLASLT_MATMUL_TILE_320x56 "CUBLASLT_MATMUL_TILE_320x56" = 520
543
+ CUBLASLT_MATMUL_TILE_320x72 "CUBLASLT_MATMUL_TILE_320x72" = 521
544
+ CUBLASLT_MATMUL_TILE_320x80 "CUBLASLT_MATMUL_TILE_320x80" = 522
545
+ CUBLASLT_MATMUL_TILE_320x88 "CUBLASLT_MATMUL_TILE_320x88" = 523
546
+ CUBLASLT_MATMUL_TILE_320x96 "CUBLASLT_MATMUL_TILE_320x96" = 524
547
+ CUBLASLT_MATMUL_TILE_320x104 "CUBLASLT_MATMUL_TILE_320x104" = 525
548
+ CUBLASLT_MATMUL_TILE_320x112 "CUBLASLT_MATMUL_TILE_320x112" = 526
549
+ CUBLASLT_MATMUL_TILE_320x120 "CUBLASLT_MATMUL_TILE_320x120" = 527
550
+ CUBLASLT_MATMUL_TILE_320x136 "CUBLASLT_MATMUL_TILE_320x136" = 528
551
+ CUBLASLT_MATMUL_TILE_320x144 "CUBLASLT_MATMUL_TILE_320x144" = 529
552
+ CUBLASLT_MATMUL_TILE_320x152 "CUBLASLT_MATMUL_TILE_320x152" = 530
553
+ CUBLASLT_MATMUL_TILE_320x160 "CUBLASLT_MATMUL_TILE_320x160" = 531
554
+ CUBLASLT_MATMUL_TILE_320x168 "CUBLASLT_MATMUL_TILE_320x168" = 532
555
+ CUBLASLT_MATMUL_TILE_320x176 "CUBLASLT_MATMUL_TILE_320x176" = 533
556
+ CUBLASLT_MATMUL_TILE_320x184 "CUBLASLT_MATMUL_TILE_320x184" = 534
557
+ CUBLASLT_MATMUL_TILE_320x192 "CUBLASLT_MATMUL_TILE_320x192" = 535
558
+ CUBLASLT_MATMUL_TILE_320x200 "CUBLASLT_MATMUL_TILE_320x200" = 536
559
+ CUBLASLT_MATMUL_TILE_384x8 "CUBLASLT_MATMUL_TILE_384x8" = 537
560
+ CUBLASLT_MATMUL_TILE_384x16 "CUBLASLT_MATMUL_TILE_384x16" = 538
561
+ CUBLASLT_MATMUL_TILE_384x24 "CUBLASLT_MATMUL_TILE_384x24" = 539
562
+ CUBLASLT_MATMUL_TILE_384x32 "CUBLASLT_MATMUL_TILE_384x32" = 540
563
+ CUBLASLT_MATMUL_TILE_384x40 "CUBLASLT_MATMUL_TILE_384x40" = 541
564
+ CUBLASLT_MATMUL_TILE_384x48 "CUBLASLT_MATMUL_TILE_384x48" = 542
565
+ CUBLASLT_MATMUL_TILE_384x56 "CUBLASLT_MATMUL_TILE_384x56" = 543
566
+ CUBLASLT_MATMUL_TILE_384x72 "CUBLASLT_MATMUL_TILE_384x72" = 544
567
+ CUBLASLT_MATMUL_TILE_384x80 "CUBLASLT_MATMUL_TILE_384x80" = 545
568
+ CUBLASLT_MATMUL_TILE_384x88 "CUBLASLT_MATMUL_TILE_384x88" = 546
569
+ CUBLASLT_MATMUL_TILE_384x96 "CUBLASLT_MATMUL_TILE_384x96" = 547
570
+ CUBLASLT_MATMUL_TILE_384x104 "CUBLASLT_MATMUL_TILE_384x104" = 548
571
+ CUBLASLT_MATMUL_TILE_384x112 "CUBLASLT_MATMUL_TILE_384x112" = 549
572
+ CUBLASLT_MATMUL_TILE_384x120 "CUBLASLT_MATMUL_TILE_384x120" = 550
573
+ CUBLASLT_MATMUL_TILE_384x136 "CUBLASLT_MATMUL_TILE_384x136" = 551
574
+ CUBLASLT_MATMUL_TILE_384x144 "CUBLASLT_MATMUL_TILE_384x144" = 552
575
+ CUBLASLT_MATMUL_TILE_384x152 "CUBLASLT_MATMUL_TILE_384x152" = 553
576
+ CUBLASLT_MATMUL_TILE_384x160 "CUBLASLT_MATMUL_TILE_384x160" = 554
577
+ CUBLASLT_MATMUL_TILE_384x168 "CUBLASLT_MATMUL_TILE_384x168" = 555
578
+ CUBLASLT_MATMUL_TILE_448x8 "CUBLASLT_MATMUL_TILE_448x8" = 556
579
+ CUBLASLT_MATMUL_TILE_448x16 "CUBLASLT_MATMUL_TILE_448x16" = 557
580
+ CUBLASLT_MATMUL_TILE_448x24 "CUBLASLT_MATMUL_TILE_448x24" = 558
581
+ CUBLASLT_MATMUL_TILE_448x32 "CUBLASLT_MATMUL_TILE_448x32" = 559
582
+ CUBLASLT_MATMUL_TILE_448x40 "CUBLASLT_MATMUL_TILE_448x40" = 560
583
+ CUBLASLT_MATMUL_TILE_448x48 "CUBLASLT_MATMUL_TILE_448x48" = 561
584
+ CUBLASLT_MATMUL_TILE_448x56 "CUBLASLT_MATMUL_TILE_448x56" = 562
585
+ CUBLASLT_MATMUL_TILE_448x72 "CUBLASLT_MATMUL_TILE_448x72" = 563
586
+ CUBLASLT_MATMUL_TILE_448x80 "CUBLASLT_MATMUL_TILE_448x80" = 564
587
+ CUBLASLT_MATMUL_TILE_448x88 "CUBLASLT_MATMUL_TILE_448x88" = 565
588
+ CUBLASLT_MATMUL_TILE_448x96 "CUBLASLT_MATMUL_TILE_448x96" = 566
589
+ CUBLASLT_MATMUL_TILE_448x104 "CUBLASLT_MATMUL_TILE_448x104" = 567
590
+ CUBLASLT_MATMUL_TILE_448x112 "CUBLASLT_MATMUL_TILE_448x112" = 568
591
+ CUBLASLT_MATMUL_TILE_448x120 "CUBLASLT_MATMUL_TILE_448x120" = 569
592
+ CUBLASLT_MATMUL_TILE_448x128 "CUBLASLT_MATMUL_TILE_448x128" = 570
593
+ CUBLASLT_MATMUL_TILE_448x136 "CUBLASLT_MATMUL_TILE_448x136" = 571
594
+ CUBLASLT_MATMUL_TILE_448x144 "CUBLASLT_MATMUL_TILE_448x144" = 572
595
+ CUBLASLT_MATMUL_TILE_512x8 "CUBLASLT_MATMUL_TILE_512x8" = 573
596
+ CUBLASLT_MATMUL_TILE_512x16 "CUBLASLT_MATMUL_TILE_512x16" = 574
597
+ CUBLASLT_MATMUL_TILE_512x24 "CUBLASLT_MATMUL_TILE_512x24" = 575
598
+ CUBLASLT_MATMUL_TILE_512x32 "CUBLASLT_MATMUL_TILE_512x32" = 576
599
+ CUBLASLT_MATMUL_TILE_512x40 "CUBLASLT_MATMUL_TILE_512x40" = 577
600
+ CUBLASLT_MATMUL_TILE_512x48 "CUBLASLT_MATMUL_TILE_512x48" = 578
601
+ CUBLASLT_MATMUL_TILE_512x56 "CUBLASLT_MATMUL_TILE_512x56" = 579
602
+ CUBLASLT_MATMUL_TILE_512x72 "CUBLASLT_MATMUL_TILE_512x72" = 580
603
+ CUBLASLT_MATMUL_TILE_512x80 "CUBLASLT_MATMUL_TILE_512x80" = 581
604
+ CUBLASLT_MATMUL_TILE_512x88 "CUBLASLT_MATMUL_TILE_512x88" = 582
605
+ CUBLASLT_MATMUL_TILE_512x96 "CUBLASLT_MATMUL_TILE_512x96" = 583
606
+ CUBLASLT_MATMUL_TILE_512x104 "CUBLASLT_MATMUL_TILE_512x104" = 584
607
+ CUBLASLT_MATMUL_TILE_512x112 "CUBLASLT_MATMUL_TILE_512x112" = 585
608
+ CUBLASLT_MATMUL_TILE_512x120 "CUBLASLT_MATMUL_TILE_512x120" = 586
609
+ CUBLASLT_MATMUL_TILE_512x128 "CUBLASLT_MATMUL_TILE_512x128" = 587
610
+ CUBLASLT_MATMUL_TILE_576x8 "CUBLASLT_MATMUL_TILE_576x8" = 588
611
+ CUBLASLT_MATMUL_TILE_576x16 "CUBLASLT_MATMUL_TILE_576x16" = 589
612
+ CUBLASLT_MATMUL_TILE_576x24 "CUBLASLT_MATMUL_TILE_576x24" = 590
613
+ CUBLASLT_MATMUL_TILE_576x32 "CUBLASLT_MATMUL_TILE_576x32" = 591
614
+ CUBLASLT_MATMUL_TILE_576x40 "CUBLASLT_MATMUL_TILE_576x40" = 592
615
+ CUBLASLT_MATMUL_TILE_576x48 "CUBLASLT_MATMUL_TILE_576x48" = 593
616
+ CUBLASLT_MATMUL_TILE_576x56 "CUBLASLT_MATMUL_TILE_576x56" = 594
617
+ CUBLASLT_MATMUL_TILE_576x72 "CUBLASLT_MATMUL_TILE_576x72" = 595
618
+ CUBLASLT_MATMUL_TILE_576x80 "CUBLASLT_MATMUL_TILE_576x80" = 596
619
+ CUBLASLT_MATMUL_TILE_576x88 "CUBLASLT_MATMUL_TILE_576x88" = 597
620
+ CUBLASLT_MATMUL_TILE_576x96 "CUBLASLT_MATMUL_TILE_576x96" = 598
621
+ CUBLASLT_MATMUL_TILE_576x104 "CUBLASLT_MATMUL_TILE_576x104" = 599
622
+ CUBLASLT_MATMUL_TILE_576x112 "CUBLASLT_MATMUL_TILE_576x112" = 600
623
+ CUBLASLT_MATMUL_TILE_640x8 "CUBLASLT_MATMUL_TILE_640x8" = 601
624
+ CUBLASLT_MATMUL_TILE_640x16 "CUBLASLT_MATMUL_TILE_640x16" = 602
625
+ CUBLASLT_MATMUL_TILE_640x24 "CUBLASLT_MATMUL_TILE_640x24" = 603
626
+ CUBLASLT_MATMUL_TILE_640x32 "CUBLASLT_MATMUL_TILE_640x32" = 604
627
+ CUBLASLT_MATMUL_TILE_640x40 "CUBLASLT_MATMUL_TILE_640x40" = 605
628
+ CUBLASLT_MATMUL_TILE_640x48 "CUBLASLT_MATMUL_TILE_640x48" = 606
629
+ CUBLASLT_MATMUL_TILE_640x56 "CUBLASLT_MATMUL_TILE_640x56" = 607
630
+ CUBLASLT_MATMUL_TILE_640x72 "CUBLASLT_MATMUL_TILE_640x72" = 608
631
+ CUBLASLT_MATMUL_TILE_640x80 "CUBLASLT_MATMUL_TILE_640x80" = 609
632
+ CUBLASLT_MATMUL_TILE_640x88 "CUBLASLT_MATMUL_TILE_640x88" = 610
633
+ CUBLASLT_MATMUL_TILE_640x96 "CUBLASLT_MATMUL_TILE_640x96" = 611
634
+ CUBLASLT_MATMUL_TILE_704x8 "CUBLASLT_MATMUL_TILE_704x8" = 612
635
+ CUBLASLT_MATMUL_TILE_704x16 "CUBLASLT_MATMUL_TILE_704x16" = 613
636
+ CUBLASLT_MATMUL_TILE_704x24 "CUBLASLT_MATMUL_TILE_704x24" = 614
637
+ CUBLASLT_MATMUL_TILE_704x32 "CUBLASLT_MATMUL_TILE_704x32" = 615
638
+ CUBLASLT_MATMUL_TILE_704x40 "CUBLASLT_MATMUL_TILE_704x40" = 616
639
+ CUBLASLT_MATMUL_TILE_704x48 "CUBLASLT_MATMUL_TILE_704x48" = 617
640
+ CUBLASLT_MATMUL_TILE_704x56 "CUBLASLT_MATMUL_TILE_704x56" = 618
641
+ CUBLASLT_MATMUL_TILE_704x72 "CUBLASLT_MATMUL_TILE_704x72" = 619
642
+ CUBLASLT_MATMUL_TILE_704x80 "CUBLASLT_MATMUL_TILE_704x80" = 620
643
+ CUBLASLT_MATMUL_TILE_704x88 "CUBLASLT_MATMUL_TILE_704x88" = 621
644
+ CUBLASLT_MATMUL_TILE_768x8 "CUBLASLT_MATMUL_TILE_768x8" = 622
645
+ CUBLASLT_MATMUL_TILE_768x16 "CUBLASLT_MATMUL_TILE_768x16" = 623
646
+ CUBLASLT_MATMUL_TILE_768x24 "CUBLASLT_MATMUL_TILE_768x24" = 624
647
+ CUBLASLT_MATMUL_TILE_768x32 "CUBLASLT_MATMUL_TILE_768x32" = 625
648
+ CUBLASLT_MATMUL_TILE_768x40 "CUBLASLT_MATMUL_TILE_768x40" = 626
649
+ CUBLASLT_MATMUL_TILE_768x48 "CUBLASLT_MATMUL_TILE_768x48" = 627
650
+ CUBLASLT_MATMUL_TILE_768x56 "CUBLASLT_MATMUL_TILE_768x56" = 628
651
+ CUBLASLT_MATMUL_TILE_768x72 "CUBLASLT_MATMUL_TILE_768x72" = 629
652
+ CUBLASLT_MATMUL_TILE_768x80 "CUBLASLT_MATMUL_TILE_768x80" = 630
653
+ CUBLASLT_MATMUL_TILE_256x512 "CUBLASLT_MATMUL_TILE_256x512" = 631
654
+ CUBLASLT_MATMUL_TILE_256x1024 "CUBLASLT_MATMUL_TILE_256x1024" = 632
655
+ CUBLASLT_MATMUL_TILE_512x512 "CUBLASLT_MATMUL_TILE_512x512" = 633
656
+ CUBLASLT_MATMUL_TILE_512x1024 "CUBLASLT_MATMUL_TILE_512x1024" = 634
657
+
658
+ ctypedef enum cublasLtMatmulStages_t "cublasLtMatmulStages_t":
659
+ CUBLASLT_MATMUL_STAGES_UNDEFINED "CUBLASLT_MATMUL_STAGES_UNDEFINED" = 0
660
+ CUBLASLT_MATMUL_STAGES_16x1 "CUBLASLT_MATMUL_STAGES_16x1" = 1
661
+ CUBLASLT_MATMUL_STAGES_16x2 "CUBLASLT_MATMUL_STAGES_16x2" = 2
662
+ CUBLASLT_MATMUL_STAGES_16x3 "CUBLASLT_MATMUL_STAGES_16x3" = 3
663
+ CUBLASLT_MATMUL_STAGES_16x4 "CUBLASLT_MATMUL_STAGES_16x4" = 4
664
+ CUBLASLT_MATMUL_STAGES_16x5 "CUBLASLT_MATMUL_STAGES_16x5" = 5
665
+ CUBLASLT_MATMUL_STAGES_16x6 "CUBLASLT_MATMUL_STAGES_16x6" = 6
666
+ CUBLASLT_MATMUL_STAGES_32x1 "CUBLASLT_MATMUL_STAGES_32x1" = 7
667
+ CUBLASLT_MATMUL_STAGES_32x2 "CUBLASLT_MATMUL_STAGES_32x2" = 8
668
+ CUBLASLT_MATMUL_STAGES_32x3 "CUBLASLT_MATMUL_STAGES_32x3" = 9
669
+ CUBLASLT_MATMUL_STAGES_32x4 "CUBLASLT_MATMUL_STAGES_32x4" = 10
670
+ CUBLASLT_MATMUL_STAGES_32x5 "CUBLASLT_MATMUL_STAGES_32x5" = 11
671
+ CUBLASLT_MATMUL_STAGES_32x6 "CUBLASLT_MATMUL_STAGES_32x6" = 12
672
+ CUBLASLT_MATMUL_STAGES_64x1 "CUBLASLT_MATMUL_STAGES_64x1" = 13
673
+ CUBLASLT_MATMUL_STAGES_64x2 "CUBLASLT_MATMUL_STAGES_64x2" = 14
674
+ CUBLASLT_MATMUL_STAGES_64x3 "CUBLASLT_MATMUL_STAGES_64x3" = 15
675
+ CUBLASLT_MATMUL_STAGES_64x4 "CUBLASLT_MATMUL_STAGES_64x4" = 16
676
+ CUBLASLT_MATMUL_STAGES_64x5 "CUBLASLT_MATMUL_STAGES_64x5" = 17
677
+ CUBLASLT_MATMUL_STAGES_64x6 "CUBLASLT_MATMUL_STAGES_64x6" = 18
678
+ CUBLASLT_MATMUL_STAGES_128x1 "CUBLASLT_MATMUL_STAGES_128x1" = 19
679
+ CUBLASLT_MATMUL_STAGES_128x2 "CUBLASLT_MATMUL_STAGES_128x2" = 20
680
+ CUBLASLT_MATMUL_STAGES_128x3 "CUBLASLT_MATMUL_STAGES_128x3" = 21
681
+ CUBLASLT_MATMUL_STAGES_128x4 "CUBLASLT_MATMUL_STAGES_128x4" = 22
682
+ CUBLASLT_MATMUL_STAGES_128x5 "CUBLASLT_MATMUL_STAGES_128x5" = 23
683
+ CUBLASLT_MATMUL_STAGES_128x6 "CUBLASLT_MATMUL_STAGES_128x6" = 24
684
+ CUBLASLT_MATMUL_STAGES_32x10 "CUBLASLT_MATMUL_STAGES_32x10" = 25
685
+ CUBLASLT_MATMUL_STAGES_8x4 "CUBLASLT_MATMUL_STAGES_8x4" = 26
686
+ CUBLASLT_MATMUL_STAGES_16x10 "CUBLASLT_MATMUL_STAGES_16x10" = 27
687
+ CUBLASLT_MATMUL_STAGES_8x5 "CUBLASLT_MATMUL_STAGES_8x5" = 28
688
+ CUBLASLT_MATMUL_STAGES_8x3 "CUBLASLT_MATMUL_STAGES_8x3" = 31
689
+ CUBLASLT_MATMUL_STAGES_8xAUTO "CUBLASLT_MATMUL_STAGES_8xAUTO" = 32
690
+ CUBLASLT_MATMUL_STAGES_16xAUTO "CUBLASLT_MATMUL_STAGES_16xAUTO" = 33
691
+ CUBLASLT_MATMUL_STAGES_32xAUTO "CUBLASLT_MATMUL_STAGES_32xAUTO" = 34
692
+ CUBLASLT_MATMUL_STAGES_64xAUTO "CUBLASLT_MATMUL_STAGES_64xAUTO" = 35
693
+ CUBLASLT_MATMUL_STAGES_128xAUTO "CUBLASLT_MATMUL_STAGES_128xAUTO" = 36
694
+ CUBLASLT_MATMUL_STAGES_256xAUTO "CUBLASLT_MATMUL_STAGES_256xAUTO" = 37
695
+ CUBLASLT_MATMUL_STAGES_768xAUTO "CUBLASLT_MATMUL_STAGES_768xAUTO" = 38
696
+ CUBLASLT_MATMUL_STAGES_16x80 "CUBLASLT_MATMUL_STAGES_16x80" = 29
697
+ CUBLASLT_MATMUL_STAGES_64x80 "CUBLASLT_MATMUL_STAGES_64x80" = 30
698
+
699
+ ctypedef enum cublasLtPointerMode_t "cublasLtPointerMode_t":
700
+ CUBLASLT_POINTER_MODE_HOST "CUBLASLT_POINTER_MODE_HOST" = CUBLAS_POINTER_MODE_HOST
701
+ CUBLASLT_POINTER_MODE_DEVICE "CUBLASLT_POINTER_MODE_DEVICE" = CUBLAS_POINTER_MODE_DEVICE
702
+ CUBLASLT_POINTER_MODE_DEVICE_VECTOR "CUBLASLT_POINTER_MODE_DEVICE_VECTOR" = 2
703
+ CUBLASLT_POINTER_MODE_ALPHA_DEVICE_VECTOR_BETA_ZERO "CUBLASLT_POINTER_MODE_ALPHA_DEVICE_VECTOR_BETA_ZERO" = 3
704
+ CUBLASLT_POINTER_MODE_ALPHA_DEVICE_VECTOR_BETA_HOST "CUBLASLT_POINTER_MODE_ALPHA_DEVICE_VECTOR_BETA_HOST" = 4
705
+
706
+ ctypedef enum cublasLtPointerModeMask_t "cublasLtPointerModeMask_t":
707
+ CUBLASLT_POINTER_MODE_MASK_HOST "CUBLASLT_POINTER_MODE_MASK_HOST" = 1
708
+ CUBLASLT_POINTER_MODE_MASK_DEVICE "CUBLASLT_POINTER_MODE_MASK_DEVICE" = 2
709
+ CUBLASLT_POINTER_MODE_MASK_DEVICE_VECTOR "CUBLASLT_POINTER_MODE_MASK_DEVICE_VECTOR" = 4
710
+ CUBLASLT_POINTER_MODE_MASK_ALPHA_DEVICE_VECTOR_BETA_ZERO "CUBLASLT_POINTER_MODE_MASK_ALPHA_DEVICE_VECTOR_BETA_ZERO" = 8
711
+ CUBLASLT_POINTER_MODE_MASK_ALPHA_DEVICE_VECTOR_BETA_HOST "CUBLASLT_POINTER_MODE_MASK_ALPHA_DEVICE_VECTOR_BETA_HOST" = 16
712
+ CUBLASLT_POINTER_MODE_MASK_NO_FILTERING "CUBLASLT_POINTER_MODE_MASK_NO_FILTERING" = 0
713
+
714
+ ctypedef enum cublasLtOrder_t "cublasLtOrder_t":
715
+ CUBLASLT_ORDER_COL "CUBLASLT_ORDER_COL" = 0
716
+ CUBLASLT_ORDER_ROW "CUBLASLT_ORDER_ROW" = 1
717
+ CUBLASLT_ORDER_COL32 "CUBLASLT_ORDER_COL32" = 2
718
+ CUBLASLT_ORDER_COL4_4R2_8C "CUBLASLT_ORDER_COL4_4R2_8C" = 3
719
+ CUBLASLT_ORDER_COL32_2R_4R4 "CUBLASLT_ORDER_COL32_2R_4R4" = 4
720
+
721
+ ctypedef enum cublasLtMatrixLayoutAttribute_t "cublasLtMatrixLayoutAttribute_t":
722
+ CUBLASLT_MATRIX_LAYOUT_TYPE "CUBLASLT_MATRIX_LAYOUT_TYPE" = 0
723
+ CUBLASLT_MATRIX_LAYOUT_ORDER "CUBLASLT_MATRIX_LAYOUT_ORDER" = 1
724
+ CUBLASLT_MATRIX_LAYOUT_ROWS "CUBLASLT_MATRIX_LAYOUT_ROWS" = 2
725
+ CUBLASLT_MATRIX_LAYOUT_COLS "CUBLASLT_MATRIX_LAYOUT_COLS" = 3
726
+ CUBLASLT_MATRIX_LAYOUT_LD "CUBLASLT_MATRIX_LAYOUT_LD" = 4
727
+ CUBLASLT_MATRIX_LAYOUT_BATCH_COUNT "CUBLASLT_MATRIX_LAYOUT_BATCH_COUNT" = 5
728
+ CUBLASLT_MATRIX_LAYOUT_STRIDED_BATCH_OFFSET "CUBLASLT_MATRIX_LAYOUT_STRIDED_BATCH_OFFSET" = 6
729
+ CUBLASLT_MATRIX_LAYOUT_PLANE_OFFSET "CUBLASLT_MATRIX_LAYOUT_PLANE_OFFSET" = 7
730
+ CUBLASLT_MATRIX_LAYOUT_BATCH_MODE "CUBLASLT_MATRIX_LAYOUT_BATCH_MODE" = 8
731
+ CUBLASLT_GROUPED_MATRIX_LAYOUT_ROWS_ARRAY "CUBLASLT_GROUPED_MATRIX_LAYOUT_ROWS_ARRAY" = 9
732
+ CUBLASLT_GROUPED_MATRIX_LAYOUT_COLS_ARRAY "CUBLASLT_GROUPED_MATRIX_LAYOUT_COLS_ARRAY" = 10
733
+ CUBLASLT_GROUPED_MATRIX_LAYOUT_LD_ARRAY "CUBLASLT_GROUPED_MATRIX_LAYOUT_LD_ARRAY" = 11
734
+ CUBLASLT_GROUPED_MATRIX_LAYOUT_ROWS_COLS_ARRAY_INTEGER_WIDTH "CUBLASLT_GROUPED_MATRIX_LAYOUT_ROWS_COLS_ARRAY_INTEGER_WIDTH" = 12
735
+ CUBLASLT_GROUPED_MATRIX_LAYOUT_LD_ARRAY_INTEGER_WIDTH "CUBLASLT_GROUPED_MATRIX_LAYOUT_LD_ARRAY_INTEGER_WIDTH" = 13
736
+
737
+ ctypedef enum cublasLtMatmulDescAttributes_t "cublasLtMatmulDescAttributes_t":
738
+ CUBLASLT_MATMUL_DESC_COMPUTE_TYPE "CUBLASLT_MATMUL_DESC_COMPUTE_TYPE" = 0
739
+ CUBLASLT_MATMUL_DESC_SCALE_TYPE "CUBLASLT_MATMUL_DESC_SCALE_TYPE" = 1
740
+ CUBLASLT_MATMUL_DESC_POINTER_MODE "CUBLASLT_MATMUL_DESC_POINTER_MODE" = 2
741
+ CUBLASLT_MATMUL_DESC_TRANSA "CUBLASLT_MATMUL_DESC_TRANSA" = 3
742
+ CUBLASLT_MATMUL_DESC_TRANSB "CUBLASLT_MATMUL_DESC_TRANSB" = 4
743
+ CUBLASLT_MATMUL_DESC_TRANSC "CUBLASLT_MATMUL_DESC_TRANSC" = 5
744
+ CUBLASLT_MATMUL_DESC_FILL_MODE "CUBLASLT_MATMUL_DESC_FILL_MODE" = 6
745
+ CUBLASLT_MATMUL_DESC_EPILOGUE "CUBLASLT_MATMUL_DESC_EPILOGUE" = 7
746
+ CUBLASLT_MATMUL_DESC_BIAS_POINTER "CUBLASLT_MATMUL_DESC_BIAS_POINTER" = 8
747
+ CUBLASLT_MATMUL_DESC_BIAS_BATCH_STRIDE "CUBLASLT_MATMUL_DESC_BIAS_BATCH_STRIDE" = 10
748
+ CUBLASLT_MATMUL_DESC_EPILOGUE_AUX_POINTER "CUBLASLT_MATMUL_DESC_EPILOGUE_AUX_POINTER" = 11
749
+ CUBLASLT_MATMUL_DESC_EPILOGUE_AUX_LD "CUBLASLT_MATMUL_DESC_EPILOGUE_AUX_LD" = 12
750
+ CUBLASLT_MATMUL_DESC_EPILOGUE_AUX_BATCH_STRIDE "CUBLASLT_MATMUL_DESC_EPILOGUE_AUX_BATCH_STRIDE" = 13
751
+ CUBLASLT_MATMUL_DESC_ALPHA_VECTOR_BATCH_STRIDE "CUBLASLT_MATMUL_DESC_ALPHA_VECTOR_BATCH_STRIDE" = 14
752
+ CUBLASLT_MATMUL_DESC_SM_COUNT_TARGET "CUBLASLT_MATMUL_DESC_SM_COUNT_TARGET" = 15
753
+ CUBLASLT_MATMUL_DESC_A_SCALE_POINTER "CUBLASLT_MATMUL_DESC_A_SCALE_POINTER" = 17
754
+ CUBLASLT_MATMUL_DESC_B_SCALE_POINTER "CUBLASLT_MATMUL_DESC_B_SCALE_POINTER" = 18
755
+ CUBLASLT_MATMUL_DESC_C_SCALE_POINTER "CUBLASLT_MATMUL_DESC_C_SCALE_POINTER" = 19
756
+ CUBLASLT_MATMUL_DESC_D_SCALE_POINTER "CUBLASLT_MATMUL_DESC_D_SCALE_POINTER" = 20
757
+ CUBLASLT_MATMUL_DESC_AMAX_D_POINTER "CUBLASLT_MATMUL_DESC_AMAX_D_POINTER" = 21
758
+ CUBLASLT_MATMUL_DESC_EPILOGUE_AUX_DATA_TYPE "CUBLASLT_MATMUL_DESC_EPILOGUE_AUX_DATA_TYPE" = 22
759
+ CUBLASLT_MATMUL_DESC_EPILOGUE_AUX_SCALE_POINTER "CUBLASLT_MATMUL_DESC_EPILOGUE_AUX_SCALE_POINTER" = 23
760
+ CUBLASLT_MATMUL_DESC_EPILOGUE_AUX_AMAX_POINTER "CUBLASLT_MATMUL_DESC_EPILOGUE_AUX_AMAX_POINTER" = 24
761
+ CUBLASLT_MATMUL_DESC_FAST_ACCUM "CUBLASLT_MATMUL_DESC_FAST_ACCUM" = 25
762
+ CUBLASLT_MATMUL_DESC_BIAS_DATA_TYPE "CUBLASLT_MATMUL_DESC_BIAS_DATA_TYPE" = 26
763
+ CUBLASLT_MATMUL_DESC_A_SCALE_MODE "CUBLASLT_MATMUL_DESC_A_SCALE_MODE" = 31
764
+ CUBLASLT_MATMUL_DESC_B_SCALE_MODE "CUBLASLT_MATMUL_DESC_B_SCALE_MODE" = 32
765
+ CUBLASLT_MATMUL_DESC_C_SCALE_MODE "CUBLASLT_MATMUL_DESC_C_SCALE_MODE" = 33
766
+ CUBLASLT_MATMUL_DESC_D_SCALE_MODE "CUBLASLT_MATMUL_DESC_D_SCALE_MODE" = 34
767
+ CUBLASLT_MATMUL_DESC_EPILOGUE_AUX_SCALE_MODE "CUBLASLT_MATMUL_DESC_EPILOGUE_AUX_SCALE_MODE" = 35
768
+ CUBLASLT_MATMUL_DESC_D_OUT_SCALE_POINTER "CUBLASLT_MATMUL_DESC_D_OUT_SCALE_POINTER" = 36
769
+ CUBLASLT_MATMUL_DESC_D_OUT_SCALE_MODE "CUBLASLT_MATMUL_DESC_D_OUT_SCALE_MODE" = 37
770
+ CUBLASLT_MATMUL_DESC_EMULATION_DESCRIPTOR "CUBLASLT_MATMUL_DESC_EMULATION_DESCRIPTOR" = 38
771
+ CUBLASLT_MATMUL_DESC_ALPHA_BATCH_STRIDE "CUBLASLT_MATMUL_DESC_ALPHA_BATCH_STRIDE" = 39
772
+ CUBLASLT_MATMUL_DESC_BETA_BATCH_STRIDE "CUBLASLT_MATMUL_DESC_BETA_BATCH_STRIDE" = 40
773
+ CUBLASLT_MATMUL_DESC_ATOMIC_SYNC_NUM_CHUNKS_D_ROWS "CUBLASLT_MATMUL_DESC_ATOMIC_SYNC_NUM_CHUNKS_D_ROWS" = 27
774
+ CUBLASLT_MATMUL_DESC_ATOMIC_SYNC_NUM_CHUNKS_D_COLS "CUBLASLT_MATMUL_DESC_ATOMIC_SYNC_NUM_CHUNKS_D_COLS" = 28
775
+ CUBLASLT_MATMUL_DESC_ATOMIC_SYNC_IN_COUNTERS_POINTER "CUBLASLT_MATMUL_DESC_ATOMIC_SYNC_IN_COUNTERS_POINTER" = 29
776
+ CUBLASLT_MATMUL_DESC_ATOMIC_SYNC_OUT_COUNTERS_POINTER "CUBLASLT_MATMUL_DESC_ATOMIC_SYNC_OUT_COUNTERS_POINTER" = 30
777
+
778
+ ctypedef enum cublasLtMatrixTransformDescAttributes_t "cublasLtMatrixTransformDescAttributes_t":
779
+ CUBLASLT_MATRIX_TRANSFORM_DESC_SCALE_TYPE "CUBLASLT_MATRIX_TRANSFORM_DESC_SCALE_TYPE"
780
+ CUBLASLT_MATRIX_TRANSFORM_DESC_POINTER_MODE "CUBLASLT_MATRIX_TRANSFORM_DESC_POINTER_MODE"
781
+ CUBLASLT_MATRIX_TRANSFORM_DESC_TRANSA "CUBLASLT_MATRIX_TRANSFORM_DESC_TRANSA"
782
+ CUBLASLT_MATRIX_TRANSFORM_DESC_TRANSB "CUBLASLT_MATRIX_TRANSFORM_DESC_TRANSB"
783
+
784
+ ctypedef enum cublasLtReductionScheme_t "cublasLtReductionScheme_t":
785
+ CUBLASLT_REDUCTION_SCHEME_NONE "CUBLASLT_REDUCTION_SCHEME_NONE" = 0
786
+ CUBLASLT_REDUCTION_SCHEME_INPLACE "CUBLASLT_REDUCTION_SCHEME_INPLACE" = 1
787
+ CUBLASLT_REDUCTION_SCHEME_COMPUTE_TYPE "CUBLASLT_REDUCTION_SCHEME_COMPUTE_TYPE" = 2
788
+ CUBLASLT_REDUCTION_SCHEME_OUTPUT_TYPE "CUBLASLT_REDUCTION_SCHEME_OUTPUT_TYPE" = 4
789
+ CUBLASLT_REDUCTION_SCHEME_MASK "CUBLASLT_REDUCTION_SCHEME_MASK" = 0x7
790
+
791
+ ctypedef enum cublasLtEpilogue_t "cublasLtEpilogue_t":
792
+ CUBLASLT_EPILOGUE_DEFAULT "CUBLASLT_EPILOGUE_DEFAULT" = 1
793
+ CUBLASLT_EPILOGUE_RELU "CUBLASLT_EPILOGUE_RELU" = 2
794
+ CUBLASLT_EPILOGUE_RELU_AUX "CUBLASLT_EPILOGUE_RELU_AUX" = (CUBLASLT_EPILOGUE_RELU | 128)
795
+ CUBLASLT_EPILOGUE_BIAS "CUBLASLT_EPILOGUE_BIAS" = 4
796
+ CUBLASLT_EPILOGUE_RELU_BIAS "CUBLASLT_EPILOGUE_RELU_BIAS" = (CUBLASLT_EPILOGUE_RELU | CUBLASLT_EPILOGUE_BIAS)
797
+ CUBLASLT_EPILOGUE_RELU_AUX_BIAS "CUBLASLT_EPILOGUE_RELU_AUX_BIAS" = (CUBLASLT_EPILOGUE_RELU_AUX | CUBLASLT_EPILOGUE_BIAS)
798
+ CUBLASLT_EPILOGUE_DRELU "CUBLASLT_EPILOGUE_DRELU" = (8 | 128)
799
+ CUBLASLT_EPILOGUE_DRELU_BGRAD "CUBLASLT_EPILOGUE_DRELU_BGRAD" = (CUBLASLT_EPILOGUE_DRELU | 16)
800
+ CUBLASLT_EPILOGUE_GELU "CUBLASLT_EPILOGUE_GELU" = 32
801
+ CUBLASLT_EPILOGUE_GELU_AUX "CUBLASLT_EPILOGUE_GELU_AUX" = (CUBLASLT_EPILOGUE_GELU | 128)
802
+ CUBLASLT_EPILOGUE_GELU_BIAS "CUBLASLT_EPILOGUE_GELU_BIAS" = (CUBLASLT_EPILOGUE_GELU | CUBLASLT_EPILOGUE_BIAS)
803
+ CUBLASLT_EPILOGUE_GELU_AUX_BIAS "CUBLASLT_EPILOGUE_GELU_AUX_BIAS" = (CUBLASLT_EPILOGUE_GELU_AUX | CUBLASLT_EPILOGUE_BIAS)
804
+ CUBLASLT_EPILOGUE_DGELU "CUBLASLT_EPILOGUE_DGELU" = (64 | 128)
805
+ CUBLASLT_EPILOGUE_DGELU_BGRAD "CUBLASLT_EPILOGUE_DGELU_BGRAD" = (CUBLASLT_EPILOGUE_DGELU | 16)
806
+ CUBLASLT_EPILOGUE_BGRADA "CUBLASLT_EPILOGUE_BGRADA" = 256
807
+ CUBLASLT_EPILOGUE_BGRADB "CUBLASLT_EPILOGUE_BGRADB" = 512
808
+
809
+ ctypedef enum cublasLtMatmulSearch_t "cublasLtMatmulSearch_t":
810
+ CUBLASLT_SEARCH_BEST_FIT "CUBLASLT_SEARCH_BEST_FIT" = 0
811
+ CUBLASLT_SEARCH_LIMITED_BY_ALGO_ID "CUBLASLT_SEARCH_LIMITED_BY_ALGO_ID" = 1
812
+ CUBLASLT_SEARCH_RESERVED_02 "CUBLASLT_SEARCH_RESERVED_02" = 2
813
+ CUBLASLT_SEARCH_RESERVED_03 "CUBLASLT_SEARCH_RESERVED_03" = 3
814
+ CUBLASLT_SEARCH_RESERVED_04 "CUBLASLT_SEARCH_RESERVED_04" = 4
815
+ CUBLASLT_SEARCH_RESERVED_05 "CUBLASLT_SEARCH_RESERVED_05" = 5
816
+ CUBLASLT_SEARCH_RESERVED_06 "CUBLASLT_SEARCH_RESERVED_06" = 6
817
+ CUBLASLT_SEARCH_RESERVED_07 "CUBLASLT_SEARCH_RESERVED_07" = 7
818
+ CUBLASLT_SEARCH_RESERVED_08 "CUBLASLT_SEARCH_RESERVED_08" = 8
819
+ CUBLASLT_SEARCH_RESERVED_09 "CUBLASLT_SEARCH_RESERVED_09" = 9
820
+
821
+ ctypedef enum cublasLtMatmulPreferenceAttributes_t "cublasLtMatmulPreferenceAttributes_t":
822
+ CUBLASLT_MATMUL_PREF_SEARCH_MODE "CUBLASLT_MATMUL_PREF_SEARCH_MODE" = 0
823
+ CUBLASLT_MATMUL_PREF_MAX_WORKSPACE_BYTES "CUBLASLT_MATMUL_PREF_MAX_WORKSPACE_BYTES" = 1
824
+ CUBLASLT_MATMUL_PREF_REDUCTION_SCHEME_MASK "CUBLASLT_MATMUL_PREF_REDUCTION_SCHEME_MASK" = 3
825
+ CUBLASLT_MATMUL_PREF_MIN_ALIGNMENT_A_BYTES "CUBLASLT_MATMUL_PREF_MIN_ALIGNMENT_A_BYTES" = 5
826
+ CUBLASLT_MATMUL_PREF_MIN_ALIGNMENT_B_BYTES "CUBLASLT_MATMUL_PREF_MIN_ALIGNMENT_B_BYTES" = 6
827
+ CUBLASLT_MATMUL_PREF_MIN_ALIGNMENT_C_BYTES "CUBLASLT_MATMUL_PREF_MIN_ALIGNMENT_C_BYTES" = 7
828
+ CUBLASLT_MATMUL_PREF_MIN_ALIGNMENT_D_BYTES "CUBLASLT_MATMUL_PREF_MIN_ALIGNMENT_D_BYTES" = 8
829
+ CUBLASLT_MATMUL_PREF_MAX_WAVES_COUNT "CUBLASLT_MATMUL_PREF_MAX_WAVES_COUNT" = 9
830
+ CUBLASLT_MATMUL_PREF_IMPL_MASK "CUBLASLT_MATMUL_PREF_IMPL_MASK" = 12
831
+ CUBLASLT_MATMUL_PREF_GROUPED_AVERAGE_REDUCTION_DIM "CUBLASLT_MATMUL_PREF_GROUPED_AVERAGE_REDUCTION_DIM" = 13
832
+ CUBLASLT_MATMUL_PREF_GROUPED_DESC_D_AVERAGE_ROWS "CUBLASLT_MATMUL_PREF_GROUPED_DESC_D_AVERAGE_ROWS" = 14
833
+ CUBLASLT_MATMUL_PREF_GROUPED_DESC_D_AVERAGE_COLS "CUBLASLT_MATMUL_PREF_GROUPED_DESC_D_AVERAGE_COLS" = 15
834
+ CUBLASLT_MATMUL_PREF_MATH_MODE_MASK "CUBLASLT_MATMUL_PREF_MATH_MODE_MASK" = 2
835
+ CUBLASLT_MATMUL_PREF_GAUSSIAN_MODE_MASK "CUBLASLT_MATMUL_PREF_GAUSSIAN_MODE_MASK" = 4
836
+ CUBLASLT_MATMUL_PREF_POINTER_MODE_MASK "CUBLASLT_MATMUL_PREF_POINTER_MODE_MASK" = 10
837
+ CUBLASLT_MATMUL_PREF_EPILOGUE_MASK "CUBLASLT_MATMUL_PREF_EPILOGUE_MASK" = 11
838
+ CUBLASLT_MATMUL_PREF_SM_COUNT_TARGET "CUBLASLT_MATMUL_PREF_SM_COUNT_TARGET" = 13
839
+
840
+ ctypedef enum cublasLtMatmulAlgoCapAttributes_t "cublasLtMatmulAlgoCapAttributes_t":
841
+ CUBLASLT_ALGO_CAP_SPLITK_SUPPORT "CUBLASLT_ALGO_CAP_SPLITK_SUPPORT" = 0
842
+ CUBLASLT_ALGO_CAP_REDUCTION_SCHEME_MASK "CUBLASLT_ALGO_CAP_REDUCTION_SCHEME_MASK" = 1
843
+ CUBLASLT_ALGO_CAP_CTA_SWIZZLING_SUPPORT "CUBLASLT_ALGO_CAP_CTA_SWIZZLING_SUPPORT" = 2
844
+ CUBLASLT_ALGO_CAP_STRIDED_BATCH_SUPPORT "CUBLASLT_ALGO_CAP_STRIDED_BATCH_SUPPORT" = 3
845
+ CUBLASLT_ALGO_CAP_OUT_OF_PLACE_RESULT_SUPPORT "CUBLASLT_ALGO_CAP_OUT_OF_PLACE_RESULT_SUPPORT" = 4
846
+ CUBLASLT_ALGO_CAP_UPLO_SUPPORT "CUBLASLT_ALGO_CAP_UPLO_SUPPORT" = 5
847
+ CUBLASLT_ALGO_CAP_TILE_IDS "CUBLASLT_ALGO_CAP_TILE_IDS" = 6
848
+ CUBLASLT_ALGO_CAP_CUSTOM_OPTION_MAX "CUBLASLT_ALGO_CAP_CUSTOM_OPTION_MAX" = 7
849
+ CUBLASLT_ALGO_CAP_CUSTOM_MEMORY_ORDER "CUBLASLT_ALGO_CAP_CUSTOM_MEMORY_ORDER" = 10
850
+ CUBLASLT_ALGO_CAP_POINTER_MODE_MASK "CUBLASLT_ALGO_CAP_POINTER_MODE_MASK" = 11
851
+ CUBLASLT_ALGO_CAP_EPILOGUE_MASK "CUBLASLT_ALGO_CAP_EPILOGUE_MASK" = 12
852
+ CUBLASLT_ALGO_CAP_STAGES_IDS "CUBLASLT_ALGO_CAP_STAGES_IDS" = 13
853
+ CUBLASLT_ALGO_CAP_LD_NEGATIVE "CUBLASLT_ALGO_CAP_LD_NEGATIVE" = 14
854
+ CUBLASLT_ALGO_CAP_NUMERICAL_IMPL_FLAGS "CUBLASLT_ALGO_CAP_NUMERICAL_IMPL_FLAGS" = 15
855
+ CUBLASLT_ALGO_CAP_MIN_ALIGNMENT_A_BYTES "CUBLASLT_ALGO_CAP_MIN_ALIGNMENT_A_BYTES" = 16
856
+ CUBLASLT_ALGO_CAP_MIN_ALIGNMENT_B_BYTES "CUBLASLT_ALGO_CAP_MIN_ALIGNMENT_B_BYTES" = 17
857
+ CUBLASLT_ALGO_CAP_MIN_ALIGNMENT_C_BYTES "CUBLASLT_ALGO_CAP_MIN_ALIGNMENT_C_BYTES" = 18
858
+ CUBLASLT_ALGO_CAP_MIN_ALIGNMENT_D_BYTES "CUBLASLT_ALGO_CAP_MIN_ALIGNMENT_D_BYTES" = 19
859
+ CUBLASLT_ALGO_CAP_POINTER_ARRAY_BATCH_SUPPORT "CUBLASLT_ALGO_CAP_POINTER_ARRAY_BATCH_SUPPORT" = 21
860
+ CUBLASLT_ALGO_CAP_FLOATING_POINT_EMULATION_SUPPORT "CUBLASLT_ALGO_CAP_FLOATING_POINT_EMULATION_SUPPORT" = 22
861
+ CUBLASLT_ALGO_CAP_POINTER_ARRAY_GROUPED_SUPPORT "CUBLASLT_ALGO_CAP_POINTER_ARRAY_GROUPED_SUPPORT" = 23
862
+ CUBLASLT_ALGO_CAP_ATOMIC_SYNC "CUBLASLT_ALGO_CAP_ATOMIC_SYNC" = 20
863
+ CUBLASLT_ALGO_CAP_MATHMODE_IMPL "CUBLASLT_ALGO_CAP_MATHMODE_IMPL" = 8
864
+ CUBLASLT_ALGO_CAP_GAUSSIAN_IMPL "CUBLASLT_ALGO_CAP_GAUSSIAN_IMPL" = 9
865
+
866
+ ctypedef enum cublasLtMatmulAlgoConfigAttributes_t "cublasLtMatmulAlgoConfigAttributes_t":
867
+ CUBLASLT_ALGO_CONFIG_ID "CUBLASLT_ALGO_CONFIG_ID" = 0
868
+ CUBLASLT_ALGO_CONFIG_TILE_ID "CUBLASLT_ALGO_CONFIG_TILE_ID" = 1
869
+ CUBLASLT_ALGO_CONFIG_SPLITK_NUM "CUBLASLT_ALGO_CONFIG_SPLITK_NUM" = 2
870
+ CUBLASLT_ALGO_CONFIG_REDUCTION_SCHEME "CUBLASLT_ALGO_CONFIG_REDUCTION_SCHEME" = 3
871
+ CUBLASLT_ALGO_CONFIG_CTA_SWIZZLING "CUBLASLT_ALGO_CONFIG_CTA_SWIZZLING" = 4
872
+ CUBLASLT_ALGO_CONFIG_CUSTOM_OPTION "CUBLASLT_ALGO_CONFIG_CUSTOM_OPTION" = 5
873
+ CUBLASLT_ALGO_CONFIG_STAGES_ID "CUBLASLT_ALGO_CONFIG_STAGES_ID" = 6
874
+ CUBLASLT_ALGO_CONFIG_INNER_SHAPE_ID "CUBLASLT_ALGO_CONFIG_INNER_SHAPE_ID" = 7
875
+ CUBLASLT_ALGO_CONFIG_CLUSTER_SHAPE_ID "CUBLASLT_ALGO_CONFIG_CLUSTER_SHAPE_ID" = 8
876
+
877
+ ctypedef enum cublasLtClusterShape_t "cublasLtClusterShape_t":
878
+ CUBLASLT_CLUSTER_SHAPE_AUTO "CUBLASLT_CLUSTER_SHAPE_AUTO" = 0
879
+ CUBLASLT_CLUSTER_SHAPE_1x1x1 "CUBLASLT_CLUSTER_SHAPE_1x1x1" = 2
880
+ CUBLASLT_CLUSTER_SHAPE_2x1x1 "CUBLASLT_CLUSTER_SHAPE_2x1x1" = 3
881
+ CUBLASLT_CLUSTER_SHAPE_4x1x1 "CUBLASLT_CLUSTER_SHAPE_4x1x1" = 4
882
+ CUBLASLT_CLUSTER_SHAPE_1x2x1 "CUBLASLT_CLUSTER_SHAPE_1x2x1" = 5
883
+ CUBLASLT_CLUSTER_SHAPE_2x2x1 "CUBLASLT_CLUSTER_SHAPE_2x2x1" = 6
884
+ CUBLASLT_CLUSTER_SHAPE_4x2x1 "CUBLASLT_CLUSTER_SHAPE_4x2x1" = 7
885
+ CUBLASLT_CLUSTER_SHAPE_1x4x1 "CUBLASLT_CLUSTER_SHAPE_1x4x1" = 8
886
+ CUBLASLT_CLUSTER_SHAPE_2x4x1 "CUBLASLT_CLUSTER_SHAPE_2x4x1" = 9
887
+ CUBLASLT_CLUSTER_SHAPE_4x4x1 "CUBLASLT_CLUSTER_SHAPE_4x4x1" = 10
888
+ CUBLASLT_CLUSTER_SHAPE_8x1x1 "CUBLASLT_CLUSTER_SHAPE_8x1x1" = 11
889
+ CUBLASLT_CLUSTER_SHAPE_1x8x1 "CUBLASLT_CLUSTER_SHAPE_1x8x1" = 12
890
+ CUBLASLT_CLUSTER_SHAPE_8x2x1 "CUBLASLT_CLUSTER_SHAPE_8x2x1" = 13
891
+ CUBLASLT_CLUSTER_SHAPE_2x8x1 "CUBLASLT_CLUSTER_SHAPE_2x8x1" = 14
892
+ CUBLASLT_CLUSTER_SHAPE_16x1x1 "CUBLASLT_CLUSTER_SHAPE_16x1x1" = 15
893
+ CUBLASLT_CLUSTER_SHAPE_1x16x1 "CUBLASLT_CLUSTER_SHAPE_1x16x1" = 16
894
+ CUBLASLT_CLUSTER_SHAPE_3x1x1 "CUBLASLT_CLUSTER_SHAPE_3x1x1" = 17
895
+ CUBLASLT_CLUSTER_SHAPE_5x1x1 "CUBLASLT_CLUSTER_SHAPE_5x1x1" = 18
896
+ CUBLASLT_CLUSTER_SHAPE_6x1x1 "CUBLASLT_CLUSTER_SHAPE_6x1x1" = 19
897
+ CUBLASLT_CLUSTER_SHAPE_7x1x1 "CUBLASLT_CLUSTER_SHAPE_7x1x1" = 20
898
+ CUBLASLT_CLUSTER_SHAPE_9x1x1 "CUBLASLT_CLUSTER_SHAPE_9x1x1" = 21
899
+ CUBLASLT_CLUSTER_SHAPE_10x1x1 "CUBLASLT_CLUSTER_SHAPE_10x1x1" = 22
900
+ CUBLASLT_CLUSTER_SHAPE_11x1x1 "CUBLASLT_CLUSTER_SHAPE_11x1x1" = 23
901
+ CUBLASLT_CLUSTER_SHAPE_12x1x1 "CUBLASLT_CLUSTER_SHAPE_12x1x1" = 24
902
+ CUBLASLT_CLUSTER_SHAPE_13x1x1 "CUBLASLT_CLUSTER_SHAPE_13x1x1" = 25
903
+ CUBLASLT_CLUSTER_SHAPE_14x1x1 "CUBLASLT_CLUSTER_SHAPE_14x1x1" = 26
904
+ CUBLASLT_CLUSTER_SHAPE_15x1x1 "CUBLASLT_CLUSTER_SHAPE_15x1x1" = 27
905
+ CUBLASLT_CLUSTER_SHAPE_3x2x1 "CUBLASLT_CLUSTER_SHAPE_3x2x1" = 28
906
+ CUBLASLT_CLUSTER_SHAPE_5x2x1 "CUBLASLT_CLUSTER_SHAPE_5x2x1" = 29
907
+ CUBLASLT_CLUSTER_SHAPE_6x2x1 "CUBLASLT_CLUSTER_SHAPE_6x2x1" = 30
908
+ CUBLASLT_CLUSTER_SHAPE_7x2x1 "CUBLASLT_CLUSTER_SHAPE_7x2x1" = 31
909
+ CUBLASLT_CLUSTER_SHAPE_1x3x1 "CUBLASLT_CLUSTER_SHAPE_1x3x1" = 32
910
+ CUBLASLT_CLUSTER_SHAPE_2x3x1 "CUBLASLT_CLUSTER_SHAPE_2x3x1" = 33
911
+ CUBLASLT_CLUSTER_SHAPE_3x3x1 "CUBLASLT_CLUSTER_SHAPE_3x3x1" = 34
912
+ CUBLASLT_CLUSTER_SHAPE_4x3x1 "CUBLASLT_CLUSTER_SHAPE_4x3x1" = 35
913
+ CUBLASLT_CLUSTER_SHAPE_5x3x1 "CUBLASLT_CLUSTER_SHAPE_5x3x1" = 36
914
+ CUBLASLT_CLUSTER_SHAPE_3x4x1 "CUBLASLT_CLUSTER_SHAPE_3x4x1" = 37
915
+ CUBLASLT_CLUSTER_SHAPE_1x5x1 "CUBLASLT_CLUSTER_SHAPE_1x5x1" = 38
916
+ CUBLASLT_CLUSTER_SHAPE_2x5x1 "CUBLASLT_CLUSTER_SHAPE_2x5x1" = 39
917
+ CUBLASLT_CLUSTER_SHAPE_3x5x1 "CUBLASLT_CLUSTER_SHAPE_3x5x1" = 40
918
+ CUBLASLT_CLUSTER_SHAPE_1x6x1 "CUBLASLT_CLUSTER_SHAPE_1x6x1" = 41
919
+ CUBLASLT_CLUSTER_SHAPE_2x6x1 "CUBLASLT_CLUSTER_SHAPE_2x6x1" = 42
920
+ CUBLASLT_CLUSTER_SHAPE_1x7x1 "CUBLASLT_CLUSTER_SHAPE_1x7x1" = 43
921
+ CUBLASLT_CLUSTER_SHAPE_2x7x1 "CUBLASLT_CLUSTER_SHAPE_2x7x1" = 44
922
+ CUBLASLT_CLUSTER_SHAPE_1x9x1 "CUBLASLT_CLUSTER_SHAPE_1x9x1" = 45
923
+ CUBLASLT_CLUSTER_SHAPE_1x10x1 "CUBLASLT_CLUSTER_SHAPE_1x10x1" = 46
924
+ CUBLASLT_CLUSTER_SHAPE_1x11x1 "CUBLASLT_CLUSTER_SHAPE_1x11x1" = 47
925
+ CUBLASLT_CLUSTER_SHAPE_1x12x1 "CUBLASLT_CLUSTER_SHAPE_1x12x1" = 48
926
+ CUBLASLT_CLUSTER_SHAPE_1x13x1 "CUBLASLT_CLUSTER_SHAPE_1x13x1" = 49
927
+ CUBLASLT_CLUSTER_SHAPE_1x14x1 "CUBLASLT_CLUSTER_SHAPE_1x14x1" = 50
928
+ CUBLASLT_CLUSTER_SHAPE_1x15x1 "CUBLASLT_CLUSTER_SHAPE_1x15x1" = 51
929
+
930
+ ctypedef enum cublasLtMatmulInnerShape_t "cublasLtMatmulInnerShape_t":
931
+ CUBLASLT_MATMUL_INNER_SHAPE_UNDEFINED "CUBLASLT_MATMUL_INNER_SHAPE_UNDEFINED" = 0
932
+ CUBLASLT_MATMUL_INNER_SHAPE_MMA884 "CUBLASLT_MATMUL_INNER_SHAPE_MMA884" = 1
933
+ CUBLASLT_MATMUL_INNER_SHAPE_MMA1684 "CUBLASLT_MATMUL_INNER_SHAPE_MMA1684" = 2
934
+ CUBLASLT_MATMUL_INNER_SHAPE_MMA1688 "CUBLASLT_MATMUL_INNER_SHAPE_MMA1688" = 3
935
+ CUBLASLT_MATMUL_INNER_SHAPE_MMA16816 "CUBLASLT_MATMUL_INNER_SHAPE_MMA16816" = 4
936
+
937
+ ctypedef enum cublasLtMatmulMatrixScale_t "cublasLtMatmulMatrixScale_t":
938
+ CUBLASLT_MATMUL_MATRIX_SCALE_SCALAR_32F "CUBLASLT_MATMUL_MATRIX_SCALE_SCALAR_32F" = 0
939
+ CUBLASLT_MATMUL_MATRIX_SCALE_VEC16_UE4M3 "CUBLASLT_MATMUL_MATRIX_SCALE_VEC16_UE4M3" = 1
940
+ CUBLASLT_MATMUL_MATRIX_SCALE_VEC32_UE8M0 "CUBLASLT_MATMUL_MATRIX_SCALE_VEC32_UE8M0" = 2
941
+ CUBLASLT_MATMUL_MATRIX_SCALE_OUTER_VEC_32F "CUBLASLT_MATMUL_MATRIX_SCALE_OUTER_VEC_32F" = 3
942
+ CUBLASLT_MATMUL_MATRIX_SCALE_VEC128_32F "CUBLASLT_MATMUL_MATRIX_SCALE_VEC128_32F" = 4
943
+ CUBLASLT_MATMUL_MATRIX_SCALE_BLK128x128_32F "CUBLASLT_MATMUL_MATRIX_SCALE_BLK128x128_32F" = 5
944
+ CUBLASLT_MATMUL_MATRIX_SCALE_PER_BATCH_SCALAR_32F "CUBLASLT_MATMUL_MATRIX_SCALE_PER_BATCH_SCALAR_32F" = 6
945
+
946
+ ctypedef enum cublasLtBatchMode_t "cublasLtBatchMode_t":
947
+ CUBLASLT_BATCH_MODE_STRIDED "CUBLASLT_BATCH_MODE_STRIDED" = 0
948
+ CUBLASLT_BATCH_MODE_POINTER_ARRAY "CUBLASLT_BATCH_MODE_POINTER_ARRAY" = 1
949
+ CUBLASLT_BATCH_MODE_GROUPED "CUBLASLT_BATCH_MODE_GROUPED" = 2
950
+
951
+ ctypedef enum cublasLtIntegerWidth_t "cublasLtIntegerWidth_t":
952
+ CUBLASLT_INTEGER_WIDTH_32 "CUBLASLT_INTEGER_WIDTH_32" = 0
953
+ CUBLASLT_INTEGER_WIDTH_64 "CUBLASLT_INTEGER_WIDTH_64" = 1
954
+
955
+ ctypedef enum cublasLtEmulationDescAttributes_t "cublasLtEmulationDescAttributes_t":
956
+ CUBLASLT_EMULATION_DESC_STRATEGY "CUBLASLT_EMULATION_DESC_STRATEGY" = 0
957
+ CUBLASLT_EMULATION_DESC_SPECIAL_VALUES_SUPPORT "CUBLASLT_EMULATION_DESC_SPECIAL_VALUES_SUPPORT" = 1
958
+ CUBLASLT_EMULATION_DESC_FIXEDPOINT_MANTISSA_CONTROL "CUBLASLT_EMULATION_DESC_FIXEDPOINT_MANTISSA_CONTROL" = 2
959
+ CUBLASLT_EMULATION_DESC_FIXEDPOINT_MAX_MANTISSA_BIT_COUNT "CUBLASLT_EMULATION_DESC_FIXEDPOINT_MAX_MANTISSA_BIT_COUNT" = 3
960
+ CUBLASLT_EMULATION_DESC_FIXEDPOINT_MANTISSA_BIT_OFFSET "CUBLASLT_EMULATION_DESC_FIXEDPOINT_MANTISSA_BIT_OFFSET" = 4
961
+ CUBLASLT_EMULATION_DESC_FIXEDPOINT_MANTISSA_BIT_COUNT_POINTER "CUBLASLT_EMULATION_DESC_FIXEDPOINT_MANTISSA_BIT_COUNT_POINTER" = 5
962
+
963
+
964
+ # types
965
+ ctypedef uint64_t cublasLtNumericalImplFlags_t 'cublasLtNumericalImplFlags_t'
966
+
967
+ ctypedef void* cublasLtHandle_t 'cublasLtHandle_t'
968
+
969
+ ctypedef void* cublasLtMatrixLayout_t 'cublasLtMatrixLayout_t'
970
+
971
+ ctypedef void* cublasLtMatmulDesc_t 'cublasLtMatmulDesc_t'
972
+
973
+ ctypedef void* cublasLtMatrixTransformDesc_t 'cublasLtMatrixTransformDesc_t'
974
+
975
+ ctypedef void* cublasLtMatmulPreference_t 'cublasLtMatmulPreference_t'
976
+
977
+ ctypedef void* cublasLtEmulationDesc_t 'cublasLtEmulationDesc_t'
978
+
979
+ ctypedef struct cublasLtMatmulAlgo_t 'cublasLtMatmulAlgo_t':
980
+ uint64_t data[8]
981
+
982
+ ctypedef void (*cublasLtLoggerCallback_t 'cublasLtLoggerCallback_t')(
983
+ int logLevel,
984
+ const char* functionName,
985
+ const char* message
986
+ )
987
+
988
+ ctypedef struct cublasLtMatmulHeuristicResult_t 'cublasLtMatmulHeuristicResult_t':
989
+ cublasLtMatmulAlgo_t algo
990
+ size_t workspaceSize
991
+ cublasStatus_t state
992
+ float wavesCount
993
+ int reserved[4]
994
+
995
+
996
+ ###############################################################################
997
+ # Functions
998
+ ###############################################################################
999
+
1000
+ cdef cublasStatus_t cublasLtCreate(cublasLtHandle_t* lightHandle) except?_CUBLASSTATUS_T_INTERNAL_LOADING_ERROR nogil
1001
+ cdef cublasStatus_t cublasLtDestroy(cublasLtHandle_t lightHandle) except?_CUBLASSTATUS_T_INTERNAL_LOADING_ERROR nogil
1002
+ cdef size_t cublasLtGetVersion() except?0 nogil
1003
+ cdef size_t cublasLtGetCudartVersion() except?0 nogil
1004
+ cdef cublasStatus_t cublasLtGetProperty(libraryPropertyType type, int* value) except?_CUBLASSTATUS_T_INTERNAL_LOADING_ERROR nogil
1005
+ 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
1006
+ 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
1007
+ 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
1008
+ cdef cublasStatus_t cublasLtMatrixLayoutDestroy(cublasLtMatrixLayout_t matLayout) except?_CUBLASSTATUS_T_INTERNAL_LOADING_ERROR nogil
1009
+ cdef cublasStatus_t cublasLtMatrixLayoutSetAttribute(cublasLtMatrixLayout_t matLayout, cublasLtMatrixLayoutAttribute_t attr, const void* buf, size_t sizeInBytes) except?_CUBLASSTATUS_T_INTERNAL_LOADING_ERROR nogil
1010
+ 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
1011
+ cdef cublasStatus_t cublasLtMatmulDescCreate(cublasLtMatmulDesc_t* matmulDesc, cublasComputeType_t computeType, cudaDataType_t scaleType) except?_CUBLASSTATUS_T_INTERNAL_LOADING_ERROR nogil
1012
+ cdef cublasStatus_t cublasLtMatmulDescDestroy(cublasLtMatmulDesc_t matmulDesc) except?_CUBLASSTATUS_T_INTERNAL_LOADING_ERROR nogil
1013
+ cdef cublasStatus_t cublasLtMatmulDescSetAttribute(cublasLtMatmulDesc_t matmulDesc, cublasLtMatmulDescAttributes_t attr, const void* buf, size_t sizeInBytes) except?_CUBLASSTATUS_T_INTERNAL_LOADING_ERROR nogil
1014
+ 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
1015
+ cdef cublasStatus_t cublasLtMatrixTransformDescCreate(cublasLtMatrixTransformDesc_t* transformDesc, cudaDataType scaleType) except?_CUBLASSTATUS_T_INTERNAL_LOADING_ERROR nogil
1016
+ cdef cublasStatus_t cublasLtMatrixTransformDescDestroy(cublasLtMatrixTransformDesc_t transformDesc) except?_CUBLASSTATUS_T_INTERNAL_LOADING_ERROR nogil
1017
+ cdef cublasStatus_t cublasLtMatrixTransformDescSetAttribute(cublasLtMatrixTransformDesc_t transformDesc, cublasLtMatrixTransformDescAttributes_t attr, const void* buf, size_t sizeInBytes) except?_CUBLASSTATUS_T_INTERNAL_LOADING_ERROR nogil
1018
+ 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
1019
+ cdef cublasStatus_t cublasLtMatmulPreferenceCreate(cublasLtMatmulPreference_t* pref) except?_CUBLASSTATUS_T_INTERNAL_LOADING_ERROR nogil
1020
+ cdef cublasStatus_t cublasLtMatmulPreferenceDestroy(cublasLtMatmulPreference_t pref) except?_CUBLASSTATUS_T_INTERNAL_LOADING_ERROR nogil
1021
+ cdef cublasStatus_t cublasLtMatmulPreferenceSetAttribute(cublasLtMatmulPreference_t pref, cublasLtMatmulPreferenceAttributes_t attr, const void* buf, size_t sizeInBytes) except?_CUBLASSTATUS_T_INTERNAL_LOADING_ERROR nogil
1022
+ 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
1023
+ 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
1024
+ 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
1025
+ 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
1026
+ 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
1027
+ 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
1028
+ cdef cublasStatus_t cublasLtMatmulAlgoConfigSetAttribute(cublasLtMatmulAlgo_t* algo, cublasLtMatmulAlgoConfigAttributes_t attr, const void* buf, size_t sizeInBytes) except?_CUBLASSTATUS_T_INTERNAL_LOADING_ERROR nogil
1029
+ 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
1030
+ cdef cublasStatus_t cublasLtLoggerSetCallback(cublasLtLoggerCallback_t callback) except?_CUBLASSTATUS_T_INTERNAL_LOADING_ERROR nogil
1031
+ cdef cublasStatus_t cublasLtLoggerSetFile(FILE* file) except?_CUBLASSTATUS_T_INTERNAL_LOADING_ERROR nogil
1032
+ cdef cublasStatus_t cublasLtLoggerOpenFile(const char* logFile) except?_CUBLASSTATUS_T_INTERNAL_LOADING_ERROR nogil
1033
+ cdef cublasStatus_t cublasLtLoggerSetLevel(int level) except?_CUBLASSTATUS_T_INTERNAL_LOADING_ERROR nogil
1034
+ cdef cublasStatus_t cublasLtLoggerSetMask(int mask) except?_CUBLASSTATUS_T_INTERNAL_LOADING_ERROR nogil
1035
+ cdef cublasStatus_t cublasLtLoggerForceDisable() except?_CUBLASSTATUS_T_INTERNAL_LOADING_ERROR nogil
1036
+ cdef const char* cublasLtGetStatusName(cublasStatus_t status) except?NULL nogil
1037
+ cdef const char* cublasLtGetStatusString(cublasStatus_t status) except?NULL nogil
1038
+ cdef cublasStatus_t cublasLtHeuristicsCacheGetCapacity(size_t* capacity) except?_CUBLASSTATUS_T_INTERNAL_LOADING_ERROR nogil
1039
+ cdef cublasStatus_t cublasLtHeuristicsCacheSetCapacity(size_t capacity) except?_CUBLASSTATUS_T_INTERNAL_LOADING_ERROR nogil
1040
+ cdef unsigned cublasLtDisableCpuInstructionsSetMask(unsigned mask) except?0 nogil
1041
+ 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
1042
+ cdef cublasStatus_t cublasLtEmulationDescCreate(cublasLtEmulationDesc_t* emulationDesc) except?_CUBLASSTATUS_T_INTERNAL_LOADING_ERROR nogil
1043
+ cdef cublasStatus_t cublasLtEmulationDescDestroy(cublasLtEmulationDesc_t emulationDesc) except?_CUBLASSTATUS_T_INTERNAL_LOADING_ERROR nogil
1044
+ cdef cublasStatus_t cublasLtEmulationDescSetAttribute(cublasLtEmulationDesc_t emulationDesc, cublasLtEmulationDescAttributes_t attr, const void* buf, size_t sizeInBytes) except?_CUBLASSTATUS_T_INTERNAL_LOADING_ERROR nogil
1045
+ 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