xref: /aosp_15_r20/external/pytorch/aten/src/ATen/cuda/CUDAContextLight.h (revision da0073e96a02ea20f0ac840b70461e3646d07c45)
1 #pragma once
2 // Light-weight version of CUDAContext.h with fewer transitive includes
3 
4 #include <cstdint>
5 
6 #include <cuda_runtime_api.h>
7 #include <cusparse.h>
8 #include <cublas_v2.h>
9 
10 // cublasLT was introduced in CUDA 10.1 but we enable only for 11.1 that also
11 // added bf16 support
12 #include <cublasLt.h>
13 
14 #ifdef CUDART_VERSION
15 #include <cusolverDn.h>
16 #endif
17 
18 #if defined(USE_CUDSS)
19 #include <cudss.h>
20 #endif
21 
22 #if defined(USE_ROCM)
23 #include <hipsolver/hipsolver.h>
24 #endif
25 
26 #include <c10/core/Allocator.h>
27 #include <c10/cuda/CUDAFunctions.h>
28 
29 namespace c10 {
30 struct Allocator;
31 }
32 
33 namespace at::cuda {
34 
35 /*
36 A common CUDA interface for ATen.
37 
38 This interface is distinct from CUDAHooks, which defines an interface that links
39 to both CPU-only and CUDA builds. That interface is intended for runtime
40 dispatch and should be used from files that are included in both CPU-only and
41 CUDA builds.
42 
43 CUDAContext, on the other hand, should be preferred by files only included in
44 CUDA builds. It is intended to expose CUDA functionality in a consistent
45 manner.
46 
47 This means there is some overlap between the CUDAContext and CUDAHooks, but
48 the choice of which to use is simple: use CUDAContext when in a CUDA-only file,
49 use CUDAHooks otherwise.
50 
51 Note that CUDAContext simply defines an interface with no associated class.
52 It is expected that the modules whose functions compose this interface will
53 manage their own state. There is only a single CUDA context/state.
54 */
55 
56 /**
57  * DEPRECATED: use device_count() instead
58  */
getNumGPUs()59 inline int64_t getNumGPUs() {
60     return c10::cuda::device_count();
61 }
62 
63 /**
64  * CUDA is available if we compiled with CUDA, and there are one or more
65  * devices.  If we compiled with CUDA but there is a driver problem, etc.,
66  * this function will report CUDA is not available (rather than raise an error.)
67  */
is_available()68 inline bool is_available() {
69     return c10::cuda::device_count() > 0;
70 }
71 
72 TORCH_CUDA_CPP_API cudaDeviceProp* getCurrentDeviceProperties();
73 
74 TORCH_CUDA_CPP_API int warp_size();
75 
76 TORCH_CUDA_CPP_API cudaDeviceProp* getDeviceProperties(c10::DeviceIndex device);
77 
78 TORCH_CUDA_CPP_API bool canDeviceAccessPeer(
79     c10::DeviceIndex device,
80     c10::DeviceIndex peer_device);
81 
82 TORCH_CUDA_CPP_API c10::Allocator* getCUDADeviceAllocator();
83 
84 /* Handles */
85 TORCH_CUDA_CPP_API cusparseHandle_t getCurrentCUDASparseHandle();
86 TORCH_CUDA_CPP_API cublasHandle_t getCurrentCUDABlasHandle();
87 TORCH_CUDA_CPP_API cublasLtHandle_t getCurrentCUDABlasLtHandle();
88 
89 TORCH_CUDA_CPP_API void clearCublasWorkspaces();
90 
91 #if defined(CUDART_VERSION) || defined(USE_ROCM)
92 TORCH_CUDA_CPP_API cusolverDnHandle_t getCurrentCUDASolverDnHandle();
93 #endif
94 
95 #if defined(USE_CUDSS)
96 TORCH_CUDA_CPP_API cudssHandle_t getCurrentCudssHandle();
97 #endif
98 
99 } // namespace at::cuda
100