xref: /petsc/include/petscdevice.h (revision 24ded41b4e3afbef0dd5eaa1b3d8dd0172f6dba2)
1 #if !defined(PETSCDEVICE_H)
2 #define PETSCDEVICE_H
3 
4 #include <petscsys.h>
5 #include <petscdevicetypes.h>
6 #include <petscpkg_version.h>
7 
8 #if defined(PETSC_HAVE_CUDA)
9 #include <cuda.h>
10 #include <cuda_runtime.h>
11 #include <cublas_v2.h>
12 #include <cusolverDn.h>
13 #include <cusolverSp.h>
14 #include <cufft.h>
15 
16 /* cuBLAS does not have cublasGetErrorName(). We create one on our own. */
17 PETSC_EXTERN const char* PetscCUBLASGetErrorName(cublasStatus_t); /* PETSC_EXTERN since it is exposed by the CHKERRCUBLAS macro */
18 PETSC_EXTERN const char* PetscCUSolverGetErrorName(cusolverStatus_t);
19 PETSC_EXTERN const char* PetscCUFFTGetErrorName(cufftResult);
20 
21 /* REMOVE ME */
22 #define WaitForCUDA() cudaDeviceSynchronize()
23 
24 /* CUDART_VERSION = 1000 x major + 10 x minor version */
25 
26 /* Could not find exactly which CUDART_VERSION introduced cudaGetErrorName. At least it was in CUDA 8.0 (Sep. 2016) */
27 #if PETSC_PKG_CUDA_VERSION_GE(8,0,0)
28 #define PetscCallCUDA(...) do {                                         \
29     const cudaError_t _p_cuda_err__ = __VA_ARGS__;                      \
30     if (PetscUnlikely(_p_cuda_err__ != cudaSuccess)) {                  \
31       const char *name  = cudaGetErrorName(_p_cuda_err__);              \
32       const char *descr = cudaGetErrorString(_p_cuda_err__);            \
33       SETERRQ(PETSC_COMM_SELF,PETSC_ERR_GPU,"cuda error %d (%s) : %s",  \
34               (PetscErrorCode)_p_cuda_err__,name,descr);                \
35     }                                                                   \
36   } while (0)
37 #else /* PETSC_PKG_CUDA_VERSION_GE(8,0,0) */
38 #define PetscCallCUDA(...) do {                                                                \
39   const cudaError_t _p_cuda_err__ = __VA_ARGS__;                                               \
40   PetscCheck(_p_cuda_err__ == cudaSuccess,PETSC_COMM_SELF,PETSC_ERR_GPU,"cuda error %d",(PetscErrorCode)_p_cuda_err__); \
41 } while (0)
42 #endif /* PETSC_PKG_CUDA_VERSION_GE(8,0,0) */
43 #define CHKERRCUDA(...) PetscCallCUDA(__VA_ARGS__)
44 
45 #define PetscCallCUBLAS(...) do {                                       \
46     const cublasStatus_t _p_cublas_stat__ = __VA_ARGS__;                \
47     if (PetscUnlikely(_p_cublas_stat__ != CUBLAS_STATUS_SUCCESS)) {     \
48       const char *name = PetscCUBLASGetErrorName(_p_cublas_stat__);     \
49       if (((_p_cublas_stat__ == CUBLAS_STATUS_NOT_INITIALIZED) ||       \
50            (_p_cublas_stat__ == CUBLAS_STATUS_ALLOC_FAILED))   &&       \
51           PetscDeviceInitialized(PETSC_DEVICE_CUDA)) {                  \
52         SETERRQ(PETSC_COMM_SELF,PETSC_ERR_GPU_RESOURCE,                 \
53                 "cuBLAS error %d (%s). "                                \
54                 "Reports not initialized or alloc failed; "             \
55                 "this indicates the GPU may have run out resources",    \
56                 (PetscErrorCode)_p_cublas_stat__,name);                 \
57       } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_GPU,"cuBLAS error %d (%s)",   \
58                 (PetscErrorCode)_p_cublas_stat__,name);                 \
59     }                                                                   \
60   } while (0)
61 #define CHKERRCUBLAS(...) PetscCallCUBLAS(__VA_ARGS__)
62 
63 #if (CUSPARSE_VER_MAJOR > 10 || CUSPARSE_VER_MAJOR == 10 && CUSPARSE_VER_MINOR >= 2) /* According to cuda/10.1.168 on OLCF Summit */
64 #define PetscCallCUSPARSE(...)\
65 do {\
66   const cusparseStatus_t _p_cusparse_stat__ = __VA_ARGS__;\
67   if (PetscUnlikely(_p_cusparse_stat__)) {\
68     const char *name  = cusparseGetErrorName(_p_cusparse_stat__);\
69     const char *descr = cusparseGetErrorString(_p_cusparse_stat__);\
70     PetscCheck((_p_cusparse_stat__ != CUSPARSE_STATUS_NOT_INITIALIZED) && (_p_cusparse_stat__ != CUSPARSE_STATUS_ALLOC_FAILED),PETSC_COMM_SELF,PETSC_ERR_GPU_RESOURCE,"cuSPARSE errorcode %d (%s) : %s. Reports not initialized or alloc failed; this indicates the GPU has run out resources",(int)_p_cusparse_stat__,name,descr); \
71     SETERRQ(PETSC_COMM_SELF,PETSC_ERR_GPU,"cuSPARSE errorcode %d (%s) : %s",(int)_p_cusparse_stat__,name,descr);\
72   }\
73 } while (0)
74 #else  /* (CUSPARSE_VER_MAJOR > 10 || CUSPARSE_VER_MAJOR == 10 && CUSPARSE_VER_MINOR >= 2) */
75 #define PetscCallCUSPARSE(...) do { \
76   const cusparseStatus_t _p_cusparse_stat__ = __VA_ARGS__; \
77   PetscCheck(_p_cusparse_stat__ == CUSPARSE_STATUS_SUCCESS,PETSC_COMM_SELF,PETSC_ERR_GPU,"cuSPARSE errorcode %d",(PetscErrorCode)_p_cusparse_stat__); \
78   } while (0)
79 #endif /* (CUSPARSE_VER_MAJOR > 10 || CUSPARSE_VER_MAJOR == 10 && CUSPARSE_VER_MINOR >= 2) */
80 #define CHKERRCUSPARSE(...) PetscCallCUSPARSE(__VA_ARGS__)
81 
82 #define PetscCallCUSOLVER(...) do {                                     \
83     const cusolverStatus_t _p_cusolver_stat__ = __VA_ARGS__;            \
84     if (PetscUnlikely(_p_cusolver_stat__ != CUSOLVER_STATUS_SUCCESS)) { \
85       const char *name = PetscCUSolverGetErrorName(_p_cusolver_stat__); \
86       if (((_p_cusolver_stat__ == CUSOLVER_STATUS_NOT_INITIALIZED) ||   \
87            (_p_cusolver_stat__ == CUSOLVER_STATUS_ALLOC_FAILED)    ||   \
88            (_p_cusolver_stat__ == CUSOLVER_STATUS_INTERNAL_ERROR)) &&   \
89           PetscDeviceInitialized(PETSC_DEVICE_CUDA)) {                  \
90         SETERRQ(PETSC_COMM_SELF,PETSC_ERR_GPU_RESOURCE,                 \
91                 "cuSolver error %d (%s). "                              \
92                 "This indicates the GPU may have run out resources",    \
93                 (PetscErrorCode)_p_cusolver_stat__,name);               \
94       } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_GPU,"cuSolver error %d (%s)", \
95                 (PetscErrorCode)_p_cusolver_stat__,name);               \
96     }                                                                   \
97   } while (0)
98 #define CHKERRCUSOLVER(...) PetscCallCUSOLVER(__VA_ARGS__)
99 
100 #define PetscCallCUFFT(...)   do {                                      \
101     const cufftResult_t _p_cufft_stat__ = __VA_ARGS__;                  \
102     if (PetscUnlikely(_p_cufft_stat__ != CUFFT_SUCCESS)) {              \
103       const char *name = PetscCUFFTGetErrorName(_p_cufft_stat__);       \
104       if (((_p_cufft_stat__ == CUFFT_SETUP_FAILED)  ||                  \
105            (_p_cufft_stat__ == CUFFT_ALLOC_FAILED)) &&                  \
106           PetscDeviceInitialized(PETSC_DEVICE_CUDA)) {                  \
107         SETERRQ(PETSC_COMM_SELF,PETSC_ERR_GPU_RESOURCE,                 \
108                 "cuFFT error %d (%s). "                                 \
109                 "Reports not initialized or alloc failed; "             \
110                 "this indicates the GPU has run out resources",         \
111                 (PetscErrorCode)_p_cufft_stat__,name);                  \
112       } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_GPU,"cuFFT error %d (%s)",    \
113                 (PetscErrorCode)_p_cufft_stat__,name);                  \
114     }                                                                   \
115   } while (0)
116 #define CHKERRCUFFT(...) PetscCallCUFFT(__VA_ARGS__)
117 
118 #define PetscCallCURAND(...)  do {                                      \
119     const curandStatus_t _p_curand_stat__ = __VA_ARGS__;                \
120     if (PetscUnlikely(_p_curand_stat__ != CURAND_STATUS_SUCCESS)) {     \
121       if (((_p_curand_stat__ == CURAND_STATUS_INITIALIZATION_FAILED) || \
122            (_p_curand_stat__ == CURAND_STATUS_ALLOCATION_FAILED))    && \
123           PetscDeviceInitialized(PETSC_DEVICE_CUDA)) {                  \
124         SETERRQ(PETSC_COMM_SELF,PETSC_ERR_GPU_RESOURCE,                 \
125                 "cuRAND error %d. "                                     \
126                 "Reports not initialized or alloc failed; "             \
127                 "this indicates the GPU has run out resources",         \
128                 (PetscErrorCode)_p_curand_stat__);                      \
129       } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_GPU,                          \
130                 "cuRand error %d",(PetscErrorCode)_p_curand_stat__);    \
131     }                                                                   \
132   } while (0)
133 #define CHKERRCURAND(...) PetscCallCURAND(__VA_ARGS__)
134 
135 PETSC_EXTERN cudaStream_t   PetscDefaultCudaStream; /* The default stream used by PETSc */
136 
137 PETSC_EXTERN PetscErrorCode PetscCUBLASGetHandle(cublasHandle_t*);
138 PETSC_EXTERN PetscErrorCode PetscCUSOLVERDnGetHandle(cusolverDnHandle_t*);
139 #endif /* PETSC_HAVE_CUDA */
140 
141 #if defined(PETSC_HAVE_HIP)
142 #include <hip/hip_runtime.h>
143 #include <hipblas.h>
144 #if defined(__HIP_PLATFORM_NVCC__)
145 #include <cusolverDn.h>
146 #else /* __HIP_PLATFORM_HCC__ */
147 #include <rocsolver.h>
148 #endif /* __HIP_PLATFORM_NVCC__ */
149 
150 /* REMOVE ME */
151 #define WaitForHIP() hipDeviceSynchronize()
152 
153 /* hipBLAS does not have hipblasGetErrorName(). We create one on our own. */
154 PETSC_EXTERN const char* PetscHIPBLASGetErrorName(hipblasStatus_t); /* PETSC_EXTERN since it is exposed by the CHKERRHIPBLAS macro */
155 
156 #define PetscCallHIP(...)     do {                                      \
157     const hipError_t _p_hip_err__ = __VA_ARGS__;                        \
158     if (PetscUnlikely(_p_hip_err__ != hipSuccess)) {                    \
159       const char *name  = hipGetErrorName(_p_hip_err__);                \
160       const char *descr = hipGetErrorString(_p_hip_err__);              \
161       SETERRQ(PETSC_COMM_SELF,PETSC_ERR_GPU,"hip error %d (%s) : %s",   \
162               (PetscErrorCode)_p_hip_err__,name,descr);                 \
163     }                                                                   \
164   } while (0)
165 #define CHKERRHIP(...) PetscCallHIP(__VA_ARGS__)
166 
167 #define PetscCallHIPBLAS(...) do {                                      \
168     const hipblasStatus_t _p_hipblas_stat__ = __VA_ARGS__;              \
169     if (PetscUnlikely(_p_hipblas_stat__ != HIPBLAS_STATUS_SUCCESS)) {   \
170       const char *name = PetscHIPBLASGetErrorName(_p_hipblas_stat__);   \
171       SETERRQ(PETSC_COMM_SELF,PETSC_ERR_GPU,"hipBLAS error %d (%s)",    \
172               (PetscErrorCode)_p_hipblas_stat__,name);                  \
173     }                                                                   \
174   } while (0)
175 #define CHKERRHIPBLAS(...) PetscCallHIPBLAS(__VA_ARGS__)
176 
177 /* TODO: SEK:  Need to figure out the hipsolver issues */
178 #define PetscCallHIPSOLVER(...) do { \
179     const hipsolverStatus_t _p_hipsolver_stat__ = __VA_ARGS__; \
180     PetscCheck(!_p_hipsolver_stat__,PETSC_COMM_SELF,PETSC_ERR_GPU,"HIPSOLVER error %d",(PetscErrorCode)_p_hipsolver_stat__); \
181   } while (0)
182 #define CHKERRHIPSOLVER(...) PetscCallHIPSOLVER(__VA_ARGS__)
183 
184 /* hipSolver does not exist yet so we work around it
185  rocSOLVER users rocBLAS for the handle
186  * */
187 #if defined(__HIP_PLATFORM_NVCC__)
188 typedef cusolverDnHandle_t hipsolverHandle_t;
189 typedef cusolverStatus_t   hipsolverStatus_t;
190 
191 /* Alias hipsolverDestroy to cusolverDnDestroy */
192 static inline hipsolverStatus_t hipsolverDestroy(hipsolverHandle_t *hipsolverhandle)
193 {
194   return cusolverDnDestroy(hipsolverhandle);
195 }
196 
197 /* Alias hipsolverCreate to cusolverDnCreate */
198 static inline hipsolverStatus_t hipsolverCreate(hipsolverHandle_t *hipsolverhandle)
199 {
200   return cusolverDnCreate(hipsolverhandle);
201 }
202 
203 /* Alias hipsolverGetStream to cusolverDnGetStream */
204 static inline hipsolverStatus_t hipsolverGetStream(hipsolverHandle_t handle, hipStream_t *stream)
205 {
206   return cusolverDnGetStream(handle,stream);
207 }
208 
209 /* Alias hipsolverSetStream to cusolverDnSetStream */
210 static inline hipsolverStatus_t hipsolverSetStream(hipsolverHandle_t handle, hipStream_t stream)
211 {
212   return cusolveDnSetStream(handle,stream);
213 }
214 #else /* __HIP_PLATFORM_HCC__ */
215 typedef rocblas_handle hipsolverHandle_t;
216 typedef rocblas_status hipsolverStatus_t;
217 
218 /* Alias hipsolverDestroy to rocblas_destroy_handle */
219 static inline hipsolverStatus_t hipsolverDestroy(hipsolverHandle_t  hipsolverhandle)
220 {
221   return rocblas_destroy_handle(hipsolverhandle);
222 }
223 
224 /* Alias hipsolverCreate to rocblas_destroy_handle */
225 static inline hipsolverStatus_t hipsolverCreate(hipsolverHandle_t *hipsolverhandle)
226 {
227   return rocblas_create_handle(hipsolverhandle);
228 }
229 
230 /* Alias hipsolverGetStream to rocblas_get_stream */
231 static inline hipsolverStatus_t hipsolverGetStream(hipsolverHandle_t handle, hipStream_t *stream)
232 {
233   return rocblas_get_stream(handle,stream);
234 }
235 
236 /* Alias hipsolverSetStream to rocblas_set_stream */
237 static inline hipsolverStatus_t hipsolverSetStream(hipsolverHandle_t handle, hipStream_t stream)
238 {
239   return rocblas_set_stream(handle,stream);
240 }
241 #endif /* __HIP_PLATFORM_NVCC__ */
242 PETSC_EXTERN hipStream_t    PetscDefaultHipStream; /* The default stream used by PETSc */
243 
244 PETSC_EXTERN PetscErrorCode PetscHIPBLASGetHandle(hipblasHandle_t*);
245 PETSC_EXTERN PetscErrorCode PetscHIPSOLVERGetHandle(hipsolverHandle_t*);
246 #endif /* PETSC_HAVE_HIP */
247 
248 /* Cannot use the device context api without C++ */
249 #if defined(PETSC_HAVE_CXX)
250 PETSC_EXTERN PetscErrorCode PetscDeviceInitializePackage(void);
251 PETSC_EXTERN PetscErrorCode PetscDeviceFinalizePackage(void);
252 
253 /* PetscDevice */
254 PETSC_EXTERN PetscErrorCode PetscDeviceInitialize(PetscDeviceType);
255 PETSC_EXTERN PetscBool      PetscDeviceInitialized(PetscDeviceType);
256 PETSC_EXTERN PetscErrorCode PetscDeviceCreate(PetscDeviceType,PetscInt,PetscDevice*);
257 PETSC_EXTERN PetscErrorCode PetscDeviceConfigure(PetscDevice);
258 PETSC_EXTERN PetscErrorCode PetscDeviceView(PetscDevice,PetscViewer);
259 PETSC_EXTERN PetscErrorCode PetscDeviceDestroy(PetscDevice*);
260 PETSC_EXTERN PetscErrorCode PetscDeviceGetDeviceId(PetscDevice,PetscInt*);
261 
262 /* PetscDeviceContext */
263 PETSC_EXTERN PetscErrorCode PetscDeviceContextCreate(PetscDeviceContext*);
264 PETSC_EXTERN PetscErrorCode PetscDeviceContextDestroy(PetscDeviceContext*);
265 PETSC_EXTERN PetscErrorCode PetscDeviceContextSetDevice(PetscDeviceContext,PetscDevice);
266 PETSC_EXTERN PetscErrorCode PetscDeviceContextGetDevice(PetscDeviceContext,PetscDevice*);
267 PETSC_EXTERN PetscErrorCode PetscDeviceContextSetStreamType(PetscDeviceContext,PetscStreamType);
268 PETSC_EXTERN PetscErrorCode PetscDeviceContextGetStreamType(PetscDeviceContext,PetscStreamType*);
269 PETSC_EXTERN PetscErrorCode PetscDeviceContextSetUp(PetscDeviceContext);
270 PETSC_EXTERN PetscErrorCode PetscDeviceContextDuplicate(PetscDeviceContext,PetscDeviceContext*);
271 PETSC_EXTERN PetscErrorCode PetscDeviceContextQueryIdle(PetscDeviceContext,PetscBool*);
272 PETSC_EXTERN PetscErrorCode PetscDeviceContextWaitForContext(PetscDeviceContext,PetscDeviceContext);
273 PETSC_EXTERN PetscErrorCode PetscDeviceContextFork(PetscDeviceContext,PetscInt,PetscDeviceContext**);
274 PETSC_EXTERN PetscErrorCode PetscDeviceContextJoin(PetscDeviceContext,PetscInt,PetscDeviceContextJoinMode,PetscDeviceContext**);
275 PETSC_EXTERN PetscErrorCode PetscDeviceContextSynchronize(PetscDeviceContext);
276 PETSC_EXTERN PetscErrorCode PetscDeviceContextGetCurrentContext(PetscDeviceContext*);
277 PETSC_EXTERN PetscErrorCode PetscDeviceContextSetCurrentContext(PetscDeviceContext);
278 PETSC_EXTERN PetscErrorCode PetscDeviceContextSetFromOptions(MPI_Comm,const char[],PetscDeviceContext);
279 #else
280 #define PetscDeviceInitialize(...)  0
281 #define PetscDeviceInitialized(...) PETSC_FALSE
282 #endif /* PETSC_HAVE_CXX */
283 
284 #endif /* PETSCDEVICE_H */
285