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 #if PETSC_PKG_CUDA_VERSION_GE(8,0,0) 46 #define PetscCallCUDAVoid(...) do { \ 47 const cudaError_t _p_cuda_err__ = __VA_ARGS__; \ 48 PetscCheckAbort(_p_cuda_err__ == cudaSuccess,PETSC_COMM_SELF,PETSC_ERR_GPU,"cuda error %d (%s) : %s",(PetscErrorCode)_p_cuda_err__,cudaGetErrorName(_p_cuda_err__),cudaGetErrorString(_p_cuda_err__)); \ 49 } while (0) 50 #else /* PETSC_PKG_CUDA_VERSION_GE(8,0,0) */ 51 #define PetscCallCUDAVoid(...) do { \ 52 const cudaError_t _p_cuda_err__ = __VA_ARGS__; \ 53 PetscCheckAbort(_p_cuda_err__ == cudaSuccess,PETSC_COMM_SELF,PETSC_ERR_GPU,"cuda error %d",(PetscErrorCode)_p_cuda_err__); \ 54 } while (0) 55 #endif /* PETSC_PKG_CUDA_VERSION_GE(8,0,0) */ 56 57 #define PetscCallCUBLAS(...) do { \ 58 const cublasStatus_t _p_cublas_stat__ = __VA_ARGS__; \ 59 if (PetscUnlikely(_p_cublas_stat__ != CUBLAS_STATUS_SUCCESS)) { \ 60 const char *name = PetscCUBLASGetErrorName(_p_cublas_stat__); \ 61 if (((_p_cublas_stat__ == CUBLAS_STATUS_NOT_INITIALIZED) || \ 62 (_p_cublas_stat__ == CUBLAS_STATUS_ALLOC_FAILED)) && \ 63 PetscDeviceInitialized(PETSC_DEVICE_CUDA)) { \ 64 SETERRQ(PETSC_COMM_SELF,PETSC_ERR_GPU_RESOURCE, \ 65 "cuBLAS error %d (%s). " \ 66 "Reports not initialized or alloc failed; " \ 67 "this indicates the GPU may have run out resources", \ 68 (PetscErrorCode)_p_cublas_stat__,name); \ 69 } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_GPU,"cuBLAS error %d (%s)", \ 70 (PetscErrorCode)_p_cublas_stat__,name); \ 71 } \ 72 } while (0) 73 #define CHKERRCUBLAS(...) PetscCallCUBLAS(__VA_ARGS__) 74 75 #if (CUSPARSE_VER_MAJOR > 10 || CUSPARSE_VER_MAJOR == 10 && CUSPARSE_VER_MINOR >= 2) /* According to cuda/10.1.168 on OLCF Summit */ 76 #define PetscCallCUSPARSE(...)\ 77 do {\ 78 const cusparseStatus_t _p_cusparse_stat__ = __VA_ARGS__;\ 79 if (PetscUnlikely(_p_cusparse_stat__)) {\ 80 const char *name = cusparseGetErrorName(_p_cusparse_stat__);\ 81 const char *descr = cusparseGetErrorString(_p_cusparse_stat__);\ 82 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); \ 83 SETERRQ(PETSC_COMM_SELF,PETSC_ERR_GPU,"cuSPARSE errorcode %d (%s) : %s",(int)_p_cusparse_stat__,name,descr);\ 84 }\ 85 } while (0) 86 #else /* (CUSPARSE_VER_MAJOR > 10 || CUSPARSE_VER_MAJOR == 10 && CUSPARSE_VER_MINOR >= 2) */ 87 #define PetscCallCUSPARSE(...) do { \ 88 const cusparseStatus_t _p_cusparse_stat__ = __VA_ARGS__; \ 89 PetscCheck(_p_cusparse_stat__ == CUSPARSE_STATUS_SUCCESS,PETSC_COMM_SELF,PETSC_ERR_GPU,"cuSPARSE errorcode %d",(PetscErrorCode)_p_cusparse_stat__); \ 90 } while (0) 91 #endif /* (CUSPARSE_VER_MAJOR > 10 || CUSPARSE_VER_MAJOR == 10 && CUSPARSE_VER_MINOR >= 2) */ 92 #define CHKERRCUSPARSE(...) PetscCallCUSPARSE(__VA_ARGS__) 93 94 #define PetscCallCUSOLVER(...) do { \ 95 const cusolverStatus_t _p_cusolver_stat__ = __VA_ARGS__; \ 96 if (PetscUnlikely(_p_cusolver_stat__ != CUSOLVER_STATUS_SUCCESS)) { \ 97 const char *name = PetscCUSolverGetErrorName(_p_cusolver_stat__); \ 98 if (((_p_cusolver_stat__ == CUSOLVER_STATUS_NOT_INITIALIZED) || \ 99 (_p_cusolver_stat__ == CUSOLVER_STATUS_ALLOC_FAILED) || \ 100 (_p_cusolver_stat__ == CUSOLVER_STATUS_INTERNAL_ERROR)) && \ 101 PetscDeviceInitialized(PETSC_DEVICE_CUDA)) { \ 102 SETERRQ(PETSC_COMM_SELF,PETSC_ERR_GPU_RESOURCE, \ 103 "cuSolver error %d (%s). " \ 104 "This indicates the GPU may have run out resources", \ 105 (PetscErrorCode)_p_cusolver_stat__,name); \ 106 } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_GPU,"cuSolver error %d (%s)", \ 107 (PetscErrorCode)_p_cusolver_stat__,name); \ 108 } \ 109 } while (0) 110 #define CHKERRCUSOLVER(...) PetscCallCUSOLVER(__VA_ARGS__) 111 112 #define PetscCallCUFFT(...) do { \ 113 const cufftResult_t _p_cufft_stat__ = __VA_ARGS__; \ 114 if (PetscUnlikely(_p_cufft_stat__ != CUFFT_SUCCESS)) { \ 115 const char *name = PetscCUFFTGetErrorName(_p_cufft_stat__); \ 116 if (((_p_cufft_stat__ == CUFFT_SETUP_FAILED) || \ 117 (_p_cufft_stat__ == CUFFT_ALLOC_FAILED)) && \ 118 PetscDeviceInitialized(PETSC_DEVICE_CUDA)) { \ 119 SETERRQ(PETSC_COMM_SELF,PETSC_ERR_GPU_RESOURCE, \ 120 "cuFFT error %d (%s). " \ 121 "Reports not initialized or alloc failed; " \ 122 "this indicates the GPU has run out resources", \ 123 (PetscErrorCode)_p_cufft_stat__,name); \ 124 } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_GPU,"cuFFT error %d (%s)", \ 125 (PetscErrorCode)_p_cufft_stat__,name); \ 126 } \ 127 } while (0) 128 #define CHKERRCUFFT(...) PetscCallCUFFT(__VA_ARGS__) 129 130 #define PetscCallCURAND(...) do { \ 131 const curandStatus_t _p_curand_stat__ = __VA_ARGS__; \ 132 if (PetscUnlikely(_p_curand_stat__ != CURAND_STATUS_SUCCESS)) { \ 133 if (((_p_curand_stat__ == CURAND_STATUS_INITIALIZATION_FAILED) || \ 134 (_p_curand_stat__ == CURAND_STATUS_ALLOCATION_FAILED)) && \ 135 PetscDeviceInitialized(PETSC_DEVICE_CUDA)) { \ 136 SETERRQ(PETSC_COMM_SELF,PETSC_ERR_GPU_RESOURCE, \ 137 "cuRAND error %d. " \ 138 "Reports not initialized or alloc failed; " \ 139 "this indicates the GPU has run out resources", \ 140 (PetscErrorCode)_p_curand_stat__); \ 141 } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_GPU, \ 142 "cuRand error %d",(PetscErrorCode)_p_curand_stat__); \ 143 } \ 144 } while (0) 145 #define CHKERRCURAND(...) PetscCallCURAND(__VA_ARGS__) 146 147 PETSC_EXTERN cudaStream_t PetscDefaultCudaStream; /* The default stream used by PETSc */ 148 149 PETSC_EXTERN PetscErrorCode PetscCUBLASGetHandle(cublasHandle_t*); 150 PETSC_EXTERN PetscErrorCode PetscCUSOLVERDnGetHandle(cusolverDnHandle_t*); 151 #endif /* PETSC_HAVE_CUDA */ 152 153 #if defined(PETSC_HAVE_HIP) 154 #include <hip/hip_runtime.h> 155 #include <hipblas.h> 156 #if defined(__HIP_PLATFORM_NVCC__) 157 #include <cusolverDn.h> 158 #else /* __HIP_PLATFORM_HCC__ */ 159 #include <rocsolver.h> 160 #endif /* __HIP_PLATFORM_NVCC__ */ 161 162 /* REMOVE ME */ 163 #define WaitForHIP() hipDeviceSynchronize() 164 165 /* hipBLAS does not have hipblasGetErrorName(). We create one on our own. */ 166 PETSC_EXTERN const char* PetscHIPBLASGetErrorName(hipblasStatus_t); /* PETSC_EXTERN since it is exposed by the CHKERRHIPBLAS macro */ 167 168 #define PetscCallHIP(...) do { \ 169 const hipError_t _p_hip_err__ = __VA_ARGS__; \ 170 if (PetscUnlikely(_p_hip_err__ != hipSuccess)) { \ 171 const char *name = hipGetErrorName(_p_hip_err__); \ 172 const char *descr = hipGetErrorString(_p_hip_err__); \ 173 SETERRQ(PETSC_COMM_SELF,PETSC_ERR_GPU,"hip error %d (%s) : %s", \ 174 (PetscErrorCode)_p_hip_err__,name,descr); \ 175 } \ 176 } while (0) 177 #define CHKERRHIP(...) PetscCallHIP(__VA_ARGS__) 178 179 #define PetscCallHIPBLAS(...) do { \ 180 const hipblasStatus_t _p_hipblas_stat__ = __VA_ARGS__; \ 181 if (PetscUnlikely(_p_hipblas_stat__ != HIPBLAS_STATUS_SUCCESS)) { \ 182 const char *name = PetscHIPBLASGetErrorName(_p_hipblas_stat__); \ 183 SETERRQ(PETSC_COMM_SELF,PETSC_ERR_GPU,"hipBLAS error %d (%s)", \ 184 (PetscErrorCode)_p_hipblas_stat__,name); \ 185 } \ 186 } while (0) 187 #define CHKERRHIPBLAS(...) PetscCallHIPBLAS(__VA_ARGS__) 188 189 /* TODO: SEK: Need to figure out the hipsolver issues */ 190 #define PetscCallHIPSOLVER(...) do { \ 191 const hipsolverStatus_t _p_hipsolver_stat__ = __VA_ARGS__; \ 192 PetscCheck(!_p_hipsolver_stat__,PETSC_COMM_SELF,PETSC_ERR_GPU,"HIPSOLVER error %d",(PetscErrorCode)_p_hipsolver_stat__); \ 193 } while (0) 194 #define CHKERRHIPSOLVER(...) PetscCallHIPSOLVER(__VA_ARGS__) 195 196 /* hipSolver does not exist yet so we work around it 197 rocSOLVER users rocBLAS for the handle 198 * */ 199 #if defined(__HIP_PLATFORM_NVCC__) 200 typedef cusolverDnHandle_t hipsolverHandle_t; 201 typedef cusolverStatus_t hipsolverStatus_t; 202 203 /* Alias hipsolverDestroy to cusolverDnDestroy */ 204 static inline hipsolverStatus_t hipsolverDestroy(hipsolverHandle_t *hipsolverhandle) 205 { 206 return cusolverDnDestroy(hipsolverhandle); 207 } 208 209 /* Alias hipsolverCreate to cusolverDnCreate */ 210 static inline hipsolverStatus_t hipsolverCreate(hipsolverHandle_t *hipsolverhandle) 211 { 212 return cusolverDnCreate(hipsolverhandle); 213 } 214 215 /* Alias hipsolverGetStream to cusolverDnGetStream */ 216 static inline hipsolverStatus_t hipsolverGetStream(hipsolverHandle_t handle, hipStream_t *stream) 217 { 218 return cusolverDnGetStream(handle,stream); 219 } 220 221 /* Alias hipsolverSetStream to cusolverDnSetStream */ 222 static inline hipsolverStatus_t hipsolverSetStream(hipsolverHandle_t handle, hipStream_t stream) 223 { 224 return cusolveDnSetStream(handle,stream); 225 } 226 #else /* __HIP_PLATFORM_HCC__ */ 227 typedef rocblas_handle hipsolverHandle_t; 228 typedef rocblas_status hipsolverStatus_t; 229 230 /* Alias hipsolverDestroy to rocblas_destroy_handle */ 231 static inline hipsolverStatus_t hipsolverDestroy(hipsolverHandle_t hipsolverhandle) 232 { 233 return rocblas_destroy_handle(hipsolverhandle); 234 } 235 236 /* Alias hipsolverCreate to rocblas_destroy_handle */ 237 static inline hipsolverStatus_t hipsolverCreate(hipsolverHandle_t *hipsolverhandle) 238 { 239 return rocblas_create_handle(hipsolverhandle); 240 } 241 242 /* Alias hipsolverGetStream to rocblas_get_stream */ 243 static inline hipsolverStatus_t hipsolverGetStream(hipsolverHandle_t handle, hipStream_t *stream) 244 { 245 return rocblas_get_stream(handle,stream); 246 } 247 248 /* Alias hipsolverSetStream to rocblas_set_stream */ 249 static inline hipsolverStatus_t hipsolverSetStream(hipsolverHandle_t handle, hipStream_t stream) 250 { 251 return rocblas_set_stream(handle,stream); 252 } 253 #endif /* __HIP_PLATFORM_NVCC__ */ 254 PETSC_EXTERN hipStream_t PetscDefaultHipStream; /* The default stream used by PETSc */ 255 256 PETSC_EXTERN PetscErrorCode PetscHIPBLASGetHandle(hipblasHandle_t*); 257 PETSC_EXTERN PetscErrorCode PetscHIPSOLVERGetHandle(hipsolverHandle_t*); 258 #endif /* PETSC_HAVE_HIP */ 259 260 /* Cannot use the device context api without C++ */ 261 #if defined(PETSC_HAVE_CXX) 262 PETSC_EXTERN PetscErrorCode PetscDeviceInitializePackage(void); 263 PETSC_EXTERN PetscErrorCode PetscDeviceFinalizePackage(void); 264 265 /* PetscDevice */ 266 PETSC_EXTERN PetscErrorCode PetscDeviceInitialize(PetscDeviceType); 267 PETSC_EXTERN PetscBool PetscDeviceInitialized(PetscDeviceType); 268 PETSC_EXTERN PetscErrorCode PetscDeviceCreate(PetscDeviceType,PetscInt,PetscDevice*); 269 PETSC_EXTERN PetscErrorCode PetscDeviceConfigure(PetscDevice); 270 PETSC_EXTERN PetscErrorCode PetscDeviceView(PetscDevice,PetscViewer); 271 PETSC_EXTERN PetscErrorCode PetscDeviceDestroy(PetscDevice*); 272 PETSC_EXTERN PetscErrorCode PetscDeviceGetDeviceId(PetscDevice,PetscInt*); 273 274 /* PetscDeviceContext */ 275 PETSC_EXTERN PetscErrorCode PetscDeviceContextCreate(PetscDeviceContext*); 276 PETSC_EXTERN PetscErrorCode PetscDeviceContextDestroy(PetscDeviceContext*); 277 PETSC_EXTERN PetscErrorCode PetscDeviceContextSetDevice(PetscDeviceContext,PetscDevice); 278 PETSC_EXTERN PetscErrorCode PetscDeviceContextGetDevice(PetscDeviceContext,PetscDevice*); 279 PETSC_EXTERN PetscErrorCode PetscDeviceContextSetStreamType(PetscDeviceContext,PetscStreamType); 280 PETSC_EXTERN PetscErrorCode PetscDeviceContextGetStreamType(PetscDeviceContext,PetscStreamType*); 281 PETSC_EXTERN PetscErrorCode PetscDeviceContextSetUp(PetscDeviceContext); 282 PETSC_EXTERN PetscErrorCode PetscDeviceContextDuplicate(PetscDeviceContext,PetscDeviceContext*); 283 PETSC_EXTERN PetscErrorCode PetscDeviceContextQueryIdle(PetscDeviceContext,PetscBool*); 284 PETSC_EXTERN PetscErrorCode PetscDeviceContextWaitForContext(PetscDeviceContext,PetscDeviceContext); 285 PETSC_EXTERN PetscErrorCode PetscDeviceContextFork(PetscDeviceContext,PetscInt,PetscDeviceContext**); 286 PETSC_EXTERN PetscErrorCode PetscDeviceContextJoin(PetscDeviceContext,PetscInt,PetscDeviceContextJoinMode,PetscDeviceContext**); 287 PETSC_EXTERN PetscErrorCode PetscDeviceContextSynchronize(PetscDeviceContext); 288 PETSC_EXTERN PetscErrorCode PetscDeviceContextGetCurrentContext(PetscDeviceContext*); 289 PETSC_EXTERN PetscErrorCode PetscDeviceContextSetCurrentContext(PetscDeviceContext); 290 PETSC_EXTERN PetscErrorCode PetscDeviceContextSetFromOptions(MPI_Comm,const char[],PetscDeviceContext); 291 #else 292 #define PetscDeviceInitialize(...) 0 293 #define PetscDeviceInitialized(...) PETSC_FALSE 294 #endif /* PETSC_HAVE_CXX */ 295 296 #endif /* PETSCDEVICE_H */ 297