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