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.
- nvmath/__init__.pxd +0 -0
- nvmath/__init__.py +45 -0
- nvmath/_internal/__init__.py +0 -0
- nvmath/_internal/attribute_ifc_factory.py +330 -0
- nvmath/_internal/layout.py +70 -0
- nvmath/_internal/templates.py +130 -0
- nvmath/_internal/threadsafe.py +106 -0
- nvmath/_internal/utils.py +43 -0
- nvmath/_internal/workspace.py +490 -0
- nvmath/_utils.py +147 -0
- nvmath/bindings/__init__.py +60 -0
- nvmath/bindings/_internal/__init__.pxd +0 -0
- nvmath/bindings/_internal/__init__.py +0 -0
- nvmath/bindings/_internal/common_types.pxd +31 -0
- nvmath/bindings/_internal/cublas.cp314t-win_amd64.pyd +0 -0
- nvmath/bindings/_internal/cublas.pxd +530 -0
- nvmath/bindings/_internal/cublasLt.cp314t-win_amd64.pyd +0 -0
- nvmath/bindings/_internal/cublasLt.pxd +59 -0
- nvmath/bindings/_internal/cublasMp.pxd +52 -0
- nvmath/bindings/_internal/cudss.cp314t-win_amd64.pyd +0 -0
- nvmath/bindings/_internal/cudss.pxd +54 -0
- nvmath/bindings/_internal/cufft.cp314t-win_amd64.pyd +0 -0
- nvmath/bindings/_internal/cufft.pxd +70 -0
- nvmath/bindings/_internal/cufftMp.pxd +77 -0
- nvmath/bindings/_internal/curand.cp314t-win_amd64.pyd +0 -0
- nvmath/bindings/_internal/curand.pxd +42 -0
- nvmath/bindings/_internal/cusolver.cp314t-win_amd64.pyd +0 -0
- nvmath/bindings/_internal/cusolver.pxd +15 -0
- nvmath/bindings/_internal/cusolverDn.cp314t-win_amd64.pyd +0 -0
- nvmath/bindings/_internal/cusolverDn.pxd +406 -0
- nvmath/bindings/_internal/cusolverMp.pxd +71 -0
- nvmath/bindings/_internal/cusolverSp.cp314t-win_amd64.pyd +0 -0
- nvmath/bindings/_internal/cusolverSp.pxd +75 -0
- nvmath/bindings/_internal/cusparse.cp314t-win_amd64.pyd +0 -0
- nvmath/bindings/_internal/cusparse.pxd +471 -0
- nvmath/bindings/_internal/cusparseLt.cp314t-win_amd64.pyd +0 -0
- nvmath/bindings/_internal/cusparseLt.pxd +48 -0
- nvmath/bindings/_internal/cutensor.cp314t-win_amd64.pyd +0 -0
- nvmath/bindings/_internal/cutensor.pxd +58 -0
- nvmath/bindings/_internal/mathdx.cp314t-win_amd64.pyd +0 -0
- nvmath/bindings/_internal/mathdx.pxd +116 -0
- nvmath/bindings/_internal/nvshmem.pxd +29 -0
- nvmath/bindings/_internal/utils.cp314t-win_amd64.pyd +0 -0
- nvmath/bindings/_internal/utils.pxd +174 -0
- nvmath/bindings/_internal/utils.pyi +10 -0
- nvmath/bindings/cublas.cp314t-win_amd64.pyd +0 -0
- nvmath/bindings/cublas.pxd +558 -0
- nvmath/bindings/cublas.pyi +812 -0
- nvmath/bindings/cublasLt.cp314t-win_amd64.pyd +0 -0
- nvmath/bindings/cublasLt.pxd +109 -0
- nvmath/bindings/cublasLt.pyi +1461 -0
- nvmath/bindings/cublasMp.pxd +85 -0
- nvmath/bindings/cublasMp.pyi +267 -0
- nvmath/bindings/cudss.cp314t-win_amd64.pyd +0 -0
- nvmath/bindings/cudss.pxd +98 -0
- nvmath/bindings/cudss.pyi +443 -0
- nvmath/bindings/cufft.cp314t-win_amd64.pyd +0 -0
- nvmath/bindings/cufft.pxd +118 -0
- nvmath/bindings/cufft.pyi +301 -0
- nvmath/bindings/cufftMp.pxd +124 -0
- nvmath/bindings/cufftMp.pyi +326 -0
- nvmath/bindings/curand.cp314t-win_amd64.pyd +0 -0
- nvmath/bindings/curand.pxd +71 -0
- nvmath/bindings/curand.pyi +189 -0
- nvmath/bindings/cusolver.cp314t-win_amd64.pyd +0 -0
- nvmath/bindings/cusolver.pxd +62 -0
- nvmath/bindings/cusolver.pyi +320 -0
- nvmath/bindings/cusolverDn.cp314t-win_amd64.pyd +0 -0
- nvmath/bindings/cusolverDn.pxd +430 -0
- nvmath/bindings/cusolverDn.pyi +422 -0
- nvmath/bindings/cusolverMp.pxd +98 -0
- nvmath/bindings/cusolverMp.pyi +114 -0
- nvmath/bindings/cusolverSp.cp314t-win_amd64.pyd +0 -0
- nvmath/bindings/cusolverSp.pxd +95 -0
- nvmath/bindings/cusolverSp.pyi +70 -0
- nvmath/bindings/cusparse.cp314t-win_amd64.pyd +0 -0
- nvmath/bindings/cusparse.pxd +546 -0
- nvmath/bindings/cusparse.pyi +1017 -0
- nvmath/bindings/cusparseLt.cp314t-win_amd64.pyd +0 -0
- nvmath/bindings/cusparseLt.pxd +99 -0
- nvmath/bindings/cusparseLt.pyi +252 -0
- nvmath/bindings/cutensor.cp314t-win_amd64.pyd +0 -0
- nvmath/bindings/cutensor.pxd +98 -0
- nvmath/bindings/cutensor.pyi +324 -0
- nvmath/bindings/cycublas.cp314t-win_amd64.pyd +0 -0
- nvmath/bindings/cycublas.pxd +664 -0
- nvmath/bindings/cycublasLt.cp314t-win_amd64.pyd +0 -0
- nvmath/bindings/cycublasLt.pxd +1045 -0
- nvmath/bindings/cycublasMp.pxd +171 -0
- nvmath/bindings/cycudss.cp314t-win_amd64.pyd +0 -0
- nvmath/bindings/cycudss.pxd +277 -0
- nvmath/bindings/cycufft.cp314t-win_amd64.pyd +0 -0
- nvmath/bindings/cycufft.pxd +333 -0
- nvmath/bindings/cycufftMp.pxd +342 -0
- nvmath/bindings/cycurand.cp314t-win_amd64.pyd +0 -0
- nvmath/bindings/cycurand.pxd +141 -0
- nvmath/bindings/cycusolver.cp314t-win_amd64.pyd +0 -0
- nvmath/bindings/cycusolver.pxd +137 -0
- nvmath/bindings/cycusolverDn.cp314t-win_amd64.pyd +0 -0
- nvmath/bindings/cycusolverDn.pxd +443 -0
- nvmath/bindings/cycusolverMp.pxd +107 -0
- nvmath/bindings/cycusolverSp.cp314t-win_amd64.pyd +0 -0
- nvmath/bindings/cycusolverSp.pxd +93 -0
- nvmath/bindings/cycusparse.cp314t-win_amd64.pyd +0 -0
- nvmath/bindings/cycusparse.pxd +679 -0
- nvmath/bindings/cycusparseLt.cp314t-win_amd64.pyd +0 -0
- nvmath/bindings/cycusparseLt.pxd +135 -0
- nvmath/bindings/cycutensor.cp314t-win_amd64.pyd +0 -0
- nvmath/bindings/cycutensor.pxd +189 -0
- nvmath/bindings/cymathdx.cp314t-win_amd64.pyd +0 -0
- nvmath/bindings/cymathdx.pxd +552 -0
- nvmath/bindings/cynvshmem.pxd +118 -0
- nvmath/bindings/mathdx.cp314t-win_amd64.pyd +0 -0
- nvmath/bindings/mathdx.pxd +182 -0
- nvmath/bindings/mathdx.pyi +1562 -0
- nvmath/bindings/nvpl/__init__.pxd +0 -0
- nvmath/bindings/nvpl/__init__.py +13 -0
- nvmath/bindings/nvpl/_internal/__init__.pxd +0 -0
- nvmath/bindings/nvpl/_internal/__init__.py +0 -0
- nvmath/bindings/nvpl/_internal/blas.cp314t-win_amd64.pyd +0 -0
- nvmath/bindings/nvpl/_internal/blas.pxd +237 -0
- nvmath/bindings/nvpl/_internal/fft.cp314t-win_amd64.pyd +0 -0
- nvmath/bindings/nvpl/_internal/fft.pxd +36 -0
- nvmath/bindings/nvpl/blas.cp314t-win_amd64.pyd +0 -0
- nvmath/bindings/nvpl/blas.pxd +131 -0
- nvmath/bindings/nvpl/blas.pyi +168 -0
- nvmath/bindings/nvpl/cyblas.cp314t-win_amd64.pyd +0 -0
- nvmath/bindings/nvpl/cyblas.pxd +280 -0
- nvmath/bindings/nvpl/cyfft.cp314t-win_amd64.pyd +0 -0
- nvmath/bindings/nvpl/cyfft.pxd +93 -0
- nvmath/bindings/nvpl/fft.cp314t-win_amd64.pyd +0 -0
- nvmath/bindings/nvpl/fft.pxd +100 -0
- nvmath/bindings/nvpl/fft.pyi +168 -0
- nvmath/bindings/nvshmem.pxd +54 -0
- nvmath/bindings/nvshmem.pyi +191 -0
- nvmath/device/__init__.py +38 -0
- nvmath/device/_deprecated.py +33 -0
- nvmath/device/common.py +315 -0
- nvmath/device/common_backend.py +131 -0
- nvmath/device/common_cuda.py +201 -0
- nvmath/device/common_numba.py +300 -0
- nvmath/device/common_numba_cuda_mlir.py +202 -0
- nvmath/device/common_opaque_tensor.py +201 -0
- nvmath/device/cublasdx.py +1606 -0
- nvmath/device/cublasdx_backend.py +860 -0
- nvmath/device/cublasdx_numba.py +1534 -0
- nvmath/device/cublasdx_numba_cuda_mlir.py +208 -0
- nvmath/device/cufftdx.py +373 -0
- nvmath/device/cufftdx_backend.py +220 -0
- nvmath/device/cufftdx_numba.py +140 -0
- nvmath/device/cufftdx_numba_cuda_mlir.py +79 -0
- nvmath/device/curand_kernel.py +9147 -0
- nvmath/device/cusolverdx.py +2708 -0
- nvmath/device/cusolverdx_backend.py +440 -0
- nvmath/device/cusolverdx_numba.py +567 -0
- nvmath/device/cusolverdx_numba_cuda_mlir.py +604 -0
- nvmath/device/cusolverdx_overload_backend.py +1029 -0
- nvmath/device/llvm_array.py +29 -0
- nvmath/device/random.py +441 -0
- nvmath/device/random_helpers.py +23 -0
- nvmath/device/random_states.py +187 -0
- nvmath/device/types.py +138 -0
- nvmath/device/vector_types_numba.py +259 -0
- nvmath/distributed/__init__.py +200 -0
- nvmath/distributed/_internal/__init__.py +0 -0
- nvmath/distributed/_internal/nccl.py +86 -0
- nvmath/distributed/_internal/nvshmem.py +307 -0
- nvmath/distributed/_internal/symmetric_memory.py +35 -0
- nvmath/distributed/_internal/tensor_ifc.py +70 -0
- nvmath/distributed/_internal/tensor_ifc_cupy.py +68 -0
- nvmath/distributed/_internal/tensor_ifc_host_device.py +172 -0
- nvmath/distributed/_internal/tensor_ifc_numpy.py +46 -0
- nvmath/distributed/_internal/tensor_ifc_torch.py +162 -0
- nvmath/distributed/_internal/tensor_wrapper.py +81 -0
- nvmath/distributed/_utils.py +167 -0
- nvmath/distributed/distribution/__init__.py +30 -0
- nvmath/distributed/distribution/_configuration.py +39 -0
- nvmath/distributed/distribution/distributions.py +1024 -0
- nvmath/distributed/distribution/redistribute.py +1284 -0
- nvmath/distributed/fft/__init__.py +7 -0
- nvmath/distributed/fft/_configuration.py +82 -0
- nvmath/distributed/fft/fft.py +2742 -0
- nvmath/distributed/linalg/__init__.py +22 -0
- nvmath/distributed/linalg/_internal/__init__.py +3 -0
- nvmath/distributed/linalg/_internal/epilog_protocol.py +586 -0
- nvmath/distributed/linalg/_internal/matmul_desc_ifc.py +28 -0
- nvmath/distributed/linalg/advanced/__init__.py +8 -0
- nvmath/distributed/linalg/advanced/_configuration.py +171 -0
- nvmath/distributed/linalg/advanced/matmulmod.py +3573 -0
- nvmath/distributed/linalg/generic/__init__.py +8 -0
- nvmath/distributed/linalg/generic/_caching.py +66 -0
- nvmath/distributed/linalg/generic/_configuration.py +61 -0
- nvmath/distributed/linalg/generic/_factorization.py +172 -0
- nvmath/distributed/linalg/generic/_initialization.py +966 -0
- nvmath/distributed/linalg/generic/_problem_spec.py +511 -0
- nvmath/distributed/linalg/generic/solvermod.py +1368 -0
- nvmath/distributed/process_group.py +408 -0
- nvmath/fft/__init__.py +7 -0
- nvmath/fft/_configuration.py +189 -0
- nvmath/fft/_exec_utils.py +82 -0
- nvmath/fft/_helpers.py +237 -0
- nvmath/fft/fft.py +3122 -0
- nvmath/internal/__init__.pxd +3 -0
- nvmath/internal/__init__.py +10 -0
- nvmath/internal/_bindings.cp314t-win_amd64.pyd +0 -0
- nvmath/internal/_bindings.pxd +18 -0
- nvmath/internal/_device_utils.py +45 -0
- nvmath/internal/_layout/__init__.pxd +3 -0
- nvmath/internal/_layout/__init__.py +7 -0
- nvmath/internal/_layout/_layout.cp314t-win_amd64.pyd +0 -0
- nvmath/internal/_layout/_layout.pxd +1303 -0
- nvmath/internal/_layout/_layout.pyi +1145 -0
- nvmath/internal/enum_utils.py +142 -0
- nvmath/internal/formatters.py +87 -0
- nvmath/internal/mem_limit.py +51 -0
- nvmath/internal/memory.cp314t-win_amd64.pyd +0 -0
- nvmath/internal/memory.pxd +13 -0
- nvmath/internal/memory.pyi +50 -0
- nvmath/internal/ndbuffer/__init__.pxd +3 -0
- nvmath/internal/ndbuffer/__init__.py +9 -0
- nvmath/internal/ndbuffer/_copy_kernel.cp314t-win_amd64.pyd +0 -0
- nvmath/internal/ndbuffer/_copy_kernel.pxd +10 -0
- nvmath/internal/ndbuffer/_jit.cp314t-win_amd64.pyd +0 -0
- nvmath/internal/ndbuffer/_jit.pxd +7 -0
- nvmath/internal/ndbuffer/_ndbuffer.cp314t-win_amd64.pyd +0 -0
- nvmath/internal/ndbuffer/_ndbuffer.pxd +40 -0
- nvmath/internal/ndbuffer/_ndbuffer.pyi +463 -0
- nvmath/internal/ndbuffer/copy_kernel/args.h +34 -0
- nvmath/internal/ndbuffer/copy_kernel/copy_kernel_impl/array_view.h +52 -0
- nvmath/internal/ndbuffer/copy_kernel/copy_kernel_impl/elementwise.h +68 -0
- nvmath/internal/ndbuffer/copy_kernel/copy_kernel_impl/grid_indexer.h +69 -0
- nvmath/internal/ndbuffer/copy_kernel/copy_kernel_impl/transposed.h +242 -0
- nvmath/internal/ndbuffer/copy_kernel/copy_kernel_impl/type_utils.h +39 -0
- nvmath/internal/ndbuffer/copy_kernel/copy_kernel_impl/utils.h +132 -0
- nvmath/internal/ndbuffer/copy_kernel/copy_kernel_impl/vec.h +159 -0
- nvmath/internal/ndbuffer/copy_kernel/elementwise.h +53 -0
- nvmath/internal/ndbuffer/copy_kernel/transposed.h +58 -0
- nvmath/internal/package_ifc.py +168 -0
- nvmath/internal/package_ifc_cuda.py +57 -0
- nvmath/internal/package_ifc_cupy.py +67 -0
- nvmath/internal/package_ifc_torch.py +69 -0
- nvmath/internal/package_wrapper.py +14 -0
- nvmath/internal/tensor_ifc.py +179 -0
- nvmath/internal/tensor_ifc_cupy.py +234 -0
- nvmath/internal/tensor_ifc_ndbuffer.py +147 -0
- nvmath/internal/tensor_ifc_numpy.py +184 -0
- nvmath/internal/tensor_ifc_torch.py +178 -0
- nvmath/internal/tensor_wrapper.py +160 -0
- nvmath/internal/typemaps.py +113 -0
- nvmath/internal/utils.py +805 -0
- nvmath/linalg/__init__.py +56 -0
- nvmath/linalg/_internal/__init__.py +3 -0
- nvmath/linalg/_internal/algo_cap_ifc.py +82 -0
- nvmath/linalg/_internal/algo_config_ifc.py +43 -0
- nvmath/linalg/_internal/batch.py +234 -0
- nvmath/linalg/_internal/enum_to_tuples.py +64 -0
- nvmath/linalg/_internal/epilog_protocol.py +766 -0
- nvmath/linalg/_internal/layout.py +624 -0
- nvmath/linalg/_internal/matmul_desc_ifc.py +28 -0
- nvmath/linalg/_internal/matmul_pref_ifc.py +27 -0
- nvmath/linalg/_internal/matrix_layout_ifc.py +26 -0
- nvmath/linalg/_internal/solver_utils.py +432 -0
- nvmath/linalg/_internal/typemaps.py +144 -0
- nvmath/linalg/_internal/utils.py +157 -0
- nvmath/linalg/advanced/__init__.py +8 -0
- nvmath/linalg/advanced/_algorithmmod.py +170 -0
- nvmath/linalg/advanced/_configuration.py +351 -0
- nvmath/linalg/advanced/helpers/__init__.py +5 -0
- nvmath/linalg/advanced/helpers/matmul.py +1316 -0
- nvmath/linalg/advanced/matmulmod.py +3734 -0
- nvmath/linalg/generic/__init__.py +53 -0
- nvmath/linalg/generic/_configuration/__init__.py +39 -0
- nvmath/linalg/generic/_configuration/layout.py +263 -0
- nvmath/linalg/generic/_configuration/match.py +734 -0
- nvmath/linalg/generic/_configuration/qualifiers.py +493 -0
- nvmath/linalg/generic/_configuration/solver_configuration.py +59 -0
- nvmath/linalg/generic/_configuration/wrap.py +217 -0
- nvmath/linalg/generic/_dtype.py +15 -0
- nvmath/linalg/generic/matmulmod.py +2094 -0
- nvmath/linalg/generic/solvermod.py +1301 -0
- nvmath/memory.py +279 -0
- nvmath/sparse/__init__.py +38 -0
- nvmath/sparse/_internal/__init__.py +21 -0
- nvmath/sparse/_internal/common_utils.py +147 -0
- nvmath/sparse/_internal/cudss_config_ifc.py +702 -0
- nvmath/sparse/_internal/cudss_data_ifc.py +399 -0
- nvmath/sparse/_internal/cudss_utils.py +506 -0
- nvmath/sparse/_internal/cusparse_utils.py +382 -0
- nvmath/sparse/_internal/sparse_bsc_ifc.py +303 -0
- nvmath/sparse/_internal/sparse_bsr_ifc.py +305 -0
- nvmath/sparse/_internal/sparse_coo_ifc.py +256 -0
- nvmath/sparse/_internal/sparse_csc_ifc.py +268 -0
- nvmath/sparse/_internal/sparse_csr_ifc.py +288 -0
- nvmath/sparse/_internal/sparse_dia_ifc.py +242 -0
- nvmath/sparse/_internal/sparse_format_helpers.py +601 -0
- nvmath/sparse/_internal/sparse_tensor_ifc.py +133 -0
- nvmath/sparse/_internal/sparse_ust_ifc.py +141 -0
- nvmath/sparse/_internal/utils.py +56 -0
- nvmath/sparse/advanced/__init__.py +7 -0
- nvmath/sparse/advanced/_configuration.py +227 -0
- nvmath/sparse/advanced/direct_solver.py +2069 -0
- nvmath/sparse/generic/__init__.py +7 -0
- nvmath/sparse/generic/_configuration.py +129 -0
- nvmath/sparse/generic/_helpers.py +137 -0
- nvmath/sparse/generic/_thunks.py +21 -0
- nvmath/sparse/generic/matmulmod.py +2353 -0
- nvmath/sparse/ust/__init__.py +7 -0
- nvmath/sparse/ust/_converters.py +422 -0
- nvmath/sparse/ust/_cpp.py +28 -0
- nvmath/sparse/ust/_drawer.py +565 -0
- nvmath/sparse/ust/_emitter.py +1033 -0
- nvmath/sparse/ust/_jit.py +188 -0
- nvmath/sparse/ust/_kernel.py +282 -0
- nvmath/sparse/ust/_utils.py +149 -0
- nvmath/sparse/ust/interfaces/__init__.py +0 -0
- nvmath/sparse/ust/interfaces/torch_interface.py +476 -0
- nvmath/sparse/ust/tensor.py +1016 -0
- nvmath/sparse/ust/tensor_format.py +957 -0
- nvmath/tensor/__init__.py +6 -0
- nvmath/tensor/_configuration.py +120 -0
- nvmath/tensor/_internal/__init__.py +3 -0
- nvmath/tensor/_internal/cutensor_config_ifc.py +279 -0
- nvmath/tensor/_internal/cutensor_utils.py +230 -0
- nvmath/tensor/_internal/data.py +43 -0
- nvmath/tensor/_internal/einsum_parser.py +444 -0
- nvmath/tensor/_internal/typemaps.py +96 -0
- nvmath/tensor/contract.py +1900 -0
- nvmath_python-1.0.0.dist-info/METADATA +134 -0
- nvmath_python-1.0.0.dist-info/RECORD +332 -0
- nvmath_python-1.0.0.dist-info/WHEEL +5 -0
- nvmath_python-1.0.0.dist-info/licenses/LICENSE +177 -0
- nvmath_python-1.0.0.dist-info/top_level.txt +2 -0
|
Binary file
|
|
@@ -0,0 +1,109 @@
|
|
|
1
|
+
# Copyright (c) 2024-2026, NVIDIA CORPORATION & AFFILIATES. ALL RIGHTS RESERVED.
|
|
2
|
+
#
|
|
3
|
+
# SPDX-License-Identifier: Apache-2.0
|
|
4
|
+
#
|
|
5
|
+
# This code was automatically generated across versions from 11.0.3 to 13.2.0, generator version 0.3.1.dev1301+g7215ac36e. Do not modify it directly.
|
|
6
|
+
|
|
7
|
+
from libc.stdint cimport intptr_t
|
|
8
|
+
|
|
9
|
+
from .cycublasLt cimport *
|
|
10
|
+
|
|
11
|
+
|
|
12
|
+
###############################################################################
|
|
13
|
+
# Types
|
|
14
|
+
###############################################################################
|
|
15
|
+
|
|
16
|
+
ctypedef cublasLtHandle_t Handle
|
|
17
|
+
ctypedef cublasLtMatrixLayout_t MatrixLayout
|
|
18
|
+
ctypedef cublasLtMatmulDesc_t MatmulDesc
|
|
19
|
+
ctypedef cublasLtMatrixTransformDesc_t MatrixTransformDesc
|
|
20
|
+
ctypedef cublasLtMatmulPreference_t MatmulPreference
|
|
21
|
+
ctypedef cublasLtEmulationDesc_t EmulationDesc
|
|
22
|
+
ctypedef cublasLtLoggerCallback_t LoggerCallback
|
|
23
|
+
|
|
24
|
+
ctypedef cudaStream_t Stream
|
|
25
|
+
ctypedef cudaDataType DataType
|
|
26
|
+
ctypedef libraryPropertyType_t LibraryPropertyType
|
|
27
|
+
|
|
28
|
+
|
|
29
|
+
###############################################################################
|
|
30
|
+
# Enum
|
|
31
|
+
###############################################################################
|
|
32
|
+
|
|
33
|
+
ctypedef cublasLtMatmulTile_t _MatmulTile
|
|
34
|
+
ctypedef cublasLtMatmulStages_t _MatmulStages
|
|
35
|
+
ctypedef cublasLtPointerMode_t _PointerMode
|
|
36
|
+
ctypedef cublasLtPointerModeMask_t _PointerModeMask
|
|
37
|
+
ctypedef cublasLtOrder_t _Order
|
|
38
|
+
ctypedef cublasLtMatrixLayoutAttribute_t _MatrixLayoutAttribute
|
|
39
|
+
ctypedef cublasLtMatmulDescAttributes_t _MatmulDescAttribute
|
|
40
|
+
ctypedef cublasLtMatrixTransformDescAttributes_t _MatrixTransformDescAttribute
|
|
41
|
+
ctypedef cublasLtReductionScheme_t _ReductionScheme
|
|
42
|
+
ctypedef cublasLtEpilogue_t _Epilogue
|
|
43
|
+
ctypedef cublasLtMatmulSearch_t _MatmulSearch
|
|
44
|
+
ctypedef cublasLtMatmulPreferenceAttributes_t _MatmulPreferenceAttribute
|
|
45
|
+
ctypedef cublasLtMatmulAlgoCapAttributes_t _MatmulAlgoCapAttribute
|
|
46
|
+
ctypedef cublasLtMatmulAlgoConfigAttributes_t _MatmulAlgoConfigAttribute
|
|
47
|
+
ctypedef cublasLtClusterShape_t _ClusterShape
|
|
48
|
+
ctypedef cublasLtMatmulInnerShape_t _MatmulInnerShape
|
|
49
|
+
ctypedef cublasLtMatmulMatrixScale_t _MatmulMatrixScale
|
|
50
|
+
ctypedef cublasLtBatchMode_t _BatchMode
|
|
51
|
+
ctypedef cublasLtIntegerWidth_t _IntegerWidth
|
|
52
|
+
ctypedef cublasLtEmulationDescAttributes_t _EmulationDescAttribute
|
|
53
|
+
|
|
54
|
+
|
|
55
|
+
###############################################################################
|
|
56
|
+
# Functions
|
|
57
|
+
###############################################################################
|
|
58
|
+
|
|
59
|
+
cpdef intptr_t create() except? 0
|
|
60
|
+
cpdef destroy(intptr_t light_handle)
|
|
61
|
+
cpdef size_t get_version() except? 0
|
|
62
|
+
cpdef size_t get_cudart_version() except? 0
|
|
63
|
+
cpdef int get_property(int type) except? -1
|
|
64
|
+
cpdef matmul(intptr_t light_handle, intptr_t compute_desc, intptr_t alpha, intptr_t a, intptr_t adesc, intptr_t b, intptr_t bdesc, intptr_t beta, intptr_t c, intptr_t cdesc, intptr_t d, intptr_t ddesc, intptr_t algo, intptr_t workspace, size_t workspace_size_in_bytes, intptr_t stream)
|
|
65
|
+
cpdef matrix_transform(intptr_t light_handle, intptr_t transform_desc, intptr_t alpha, intptr_t a, intptr_t adesc, intptr_t beta, intptr_t b, intptr_t bdesc, intptr_t c, intptr_t cdesc, intptr_t stream)
|
|
66
|
+
cpdef intptr_t matrix_layout_create(int type, uint64_t rows, uint64_t cols, int64_t ld) except? 0
|
|
67
|
+
cpdef matrix_layout_destroy(intptr_t mat_layout)
|
|
68
|
+
cpdef get_matrix_layout_attribute_dtype(int attr)
|
|
69
|
+
cpdef matrix_layout_set_attribute(intptr_t mat_layout, int attr, intptr_t buf, size_t size_in_bytes)
|
|
70
|
+
cpdef matrix_layout_get_attribute(intptr_t mat_layout, int attr, intptr_t buf, size_t size_in_bytes, intptr_t size_written)
|
|
71
|
+
cpdef intptr_t matmul_desc_create(int compute_type, int scale_type) except? 0
|
|
72
|
+
cpdef matmul_desc_destroy(intptr_t matmul_desc)
|
|
73
|
+
cpdef get_matmul_desc_attribute_dtype(int attr)
|
|
74
|
+
cpdef matmul_desc_set_attribute(intptr_t matmul_desc, int attr, intptr_t buf, size_t size_in_bytes)
|
|
75
|
+
cpdef matmul_desc_get_attribute(intptr_t matmul_desc, int attr, intptr_t buf, size_t size_in_bytes, intptr_t size_written)
|
|
76
|
+
cpdef intptr_t matrix_transform_desc_create(int scale_type) except? 0
|
|
77
|
+
cpdef matrix_transform_desc_destroy(intptr_t transform_desc)
|
|
78
|
+
cpdef get_matrix_transform_desc_attribute_dtype(int attr)
|
|
79
|
+
cpdef matrix_transform_desc_set_attribute(intptr_t transform_desc, int attr, intptr_t buf, size_t size_in_bytes)
|
|
80
|
+
cpdef matrix_transform_desc_get_attribute(intptr_t transform_desc, int attr, intptr_t buf, size_t size_in_bytes, intptr_t size_written)
|
|
81
|
+
cpdef intptr_t matmul_preference_create() except? 0
|
|
82
|
+
cpdef matmul_preference_destroy(intptr_t pref)
|
|
83
|
+
cpdef get_matmul_preference_attribute_dtype(int attr)
|
|
84
|
+
cpdef matmul_preference_set_attribute(intptr_t pref, int attr, intptr_t buf, size_t size_in_bytes)
|
|
85
|
+
cpdef matmul_preference_get_attribute(intptr_t pref, int attr, intptr_t buf, size_t size_in_bytes, intptr_t size_written)
|
|
86
|
+
cpdef matmul_algo_get_heuristic(intptr_t light_handle, intptr_t operation_desc, intptr_t adesc, intptr_t bdesc, intptr_t cdesc, intptr_t ddesc, intptr_t preference, int requested_algo_count, intptr_t heuristic_results_array, intptr_t return_algo_count)
|
|
87
|
+
cpdef matmul_algo_init(intptr_t light_handle, int compute_type, int scale_type, int atype, int btype, int ctype, int dtype, int algo_id, intptr_t algo)
|
|
88
|
+
cpdef matmul_algo_check(intptr_t light_handle, intptr_t operation_desc, intptr_t adesc, intptr_t bdesc, intptr_t cdesc, intptr_t ddesc, intptr_t algo, intptr_t result)
|
|
89
|
+
cpdef get_matmul_algo_cap_attribute_dtype(int attr)
|
|
90
|
+
cpdef matmul_algo_cap_get_attribute(intptr_t algo, int attr, intptr_t buf, size_t size_in_bytes, intptr_t size_written)
|
|
91
|
+
cpdef get_matmul_algo_config_attribute_dtype(int attr)
|
|
92
|
+
cpdef matmul_algo_config_set_attribute(intptr_t algo, int attr, intptr_t buf, size_t size_in_bytes)
|
|
93
|
+
cpdef matmul_algo_config_get_attribute(intptr_t algo, int attr, intptr_t buf, size_t size_in_bytes, intptr_t size_written)
|
|
94
|
+
cpdef logger_open_file(log_file)
|
|
95
|
+
cpdef logger_set_level(int level)
|
|
96
|
+
cpdef logger_set_mask(int mask)
|
|
97
|
+
cpdef logger_force_disable()
|
|
98
|
+
cpdef str get_status_name(int status)
|
|
99
|
+
cpdef str get_status_string(int status)
|
|
100
|
+
cpdef size_t heuristics_cache_get_capacity() except? 0
|
|
101
|
+
cpdef heuristics_cache_set_capacity(size_t capacity)
|
|
102
|
+
cpdef disable_cpu_instructions_set_mask(unsigned mask)
|
|
103
|
+
cpdef intptr_t grouped_matrix_layout_create(int type, int group_count, intptr_t rows_array, intptr_t cols_array, intptr_t ld_array) except? 0
|
|
104
|
+
cpdef intptr_t emulation_desc_create() except? 0
|
|
105
|
+
cpdef emulation_desc_destroy(intptr_t emulation_desc)
|
|
106
|
+
cpdef get_emulation_desc_attribute_dtype(int attr)
|
|
107
|
+
cpdef emulation_desc_set_attribute(intptr_t emulation_desc, int attr, intptr_t buf, size_t size_in_bytes)
|
|
108
|
+
cpdef emulation_desc_get_attribute(intptr_t emulation_desc, int attr, intptr_t buf, size_t size_in_bytes, intptr_t size_written)
|
|
109
|
+
cpdef tuple matmul_algo_get_ids(intptr_t light_handle, cublasComputeType_t compute_type, size_t scale_type, size_t atype, size_t btype, size_t ctype, size_t dtype, int requested_algo_count)
|