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 CHKERRCUDA(cerr) do { \ 29 const cudaError_t _p_cuda_err__ = cerr; \ 30 if (PetscUnlikely(_p_cuda_err__ != cudaSuccess)) { \ 31 const char *name = cudaGetErrorName(_p_cuda_err__); \ 32 const char *descr = cudaGetErrorString(_p_cuda_err__); \ 33 SETERRQ3(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 CHKERRCUDA(cerr) do { \ 39 const cudaError_t _p_cuda_err__ = cerr; \ 40 if (PetscUnlikely(_p_cuda_err__ != cudaSuccess)) { \ 41 SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_GPU,"cuda error %d", \ 42 (PetscErrorCode)_p_cuda_err__); \ 43 } \ 44 } while (0) 45 #endif /* PETSC_PKG_CUDA_VERSION_GE(8,0,0) */ 46 47 #define CHKERRCUBLAS(stat) do { \ 48 const cublasStatus_t _p_cublas_stat__ = stat; \ 49 if (PetscUnlikely(_p_cublas_stat__ != CUBLAS_STATUS_SUCCESS)) { \ 50 const char *name = PetscCUBLASGetErrorName(_p_cublas_stat__); \ 51 if (((_p_cublas_stat__ == CUBLAS_STATUS_NOT_INITIALIZED) || \ 52 (_p_cublas_stat__ == CUBLAS_STATUS_ALLOC_FAILED)) && \ 53 PetscDeviceInitialized(PETSC_DEVICE_CUDA)) { \ 54 SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_GPU_RESOURCE, \ 55 "cuBLAS error %d (%s). " \ 56 "Reports not initialized or alloc failed; " \ 57 "this indicates the GPU may have run out resources", \ 58 (PetscErrorCode)_p_cublas_stat__,name); \ 59 } else { \ 60 SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_GPU,"cuBLAS error %d (%s)", \ 61 (PetscErrorCode)_p_cublas_stat__,name); \ 62 } \ 63 } \ 64 } while (0) 65 66 #if (CUSPARSE_VER_MAJOR > 10 || CUSPARSE_VER_MAJOR == 10 && CUSPARSE_VER_MINOR >= 2) /* According to cuda/10.1.168 on OLCF Summit */ 67 #define CHKERRCUSPARSE(stat)\ 68 do {\ 69 if (PetscUnlikely(stat)) {\ 70 const char *name = cusparseGetErrorName(stat);\ 71 const char *descr = cusparseGetErrorString(stat);\ 72 if ((stat == CUSPARSE_STATUS_NOT_INITIALIZED) || (stat == CUSPARSE_STATUS_ALLOC_FAILED)) SETERRQ3(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)stat,name,descr); \ 73 else SETERRQ3(PETSC_COMM_SELF,PETSC_ERR_GPU,"cuSPARSE errorcode %d (%s) : %s",(int)stat,name,descr);\ 74 }\ 75 } while (0) 76 #else /* (CUSPARSE_VER_MAJOR > 10 || CUSPARSE_VER_MAJOR == 10 && CUSPARSE_VER_MINOR >= 2) */ 77 #define CHKERRCUSPARSE(stat) do {if (PetscUnlikely(stat)) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_GPU,"cuSPARSE errorcode %d",(int)stat);} while (0) 78 #endif /* (CUSPARSE_VER_MAJOR > 10 || CUSPARSE_VER_MAJOR == 10 && CUSPARSE_VER_MINOR >= 2) */ 79 80 #define CHKERRCUSOLVER(stat) do { \ 81 const cusolverStatus_t _p_cusolver_stat__ = stat; \ 82 if (PetscUnlikely(_p_cusolver_stat__ != CUSOLVER_STATUS_SUCCESS)) { \ 83 const char *name = PetscCUSolverGetErrorName(_p_cusolver_stat__); \ 84 if (((_p_cusolver_stat__ == CUSOLVER_STATUS_NOT_INITIALIZED) || \ 85 (_p_cusolver_stat__ == CUSOLVER_STATUS_ALLOC_FAILED) || \ 86 (_p_cusolver_stat__ == CUSOLVER_STATUS_INTERNAL_ERROR)) && \ 87 PetscDeviceInitialized(PETSC_DEVICE_CUDA)) { \ 88 SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_GPU_RESOURCE, \ 89 "cuSolver error %d (%s). " \ 90 "This indicates the GPU may have run out resources", \ 91 (PetscErrorCode)_p_cusolver_stat__,name); \ 92 } else { \ 93 SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_GPU, \ 94 "cuSolver error %d (%s)", \ 95 (PetscErrorCode)_p_cusolver_stat__,name); \ 96 } \ 97 } \ 98 } while (0) 99 100 #define CHKERRCUFFT(res) do { \ 101 const cufftResult_t _p_cufft_stat__ = res; \ 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 SETERRQ2(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 { \ 113 SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_GPU, \ 114 "cuFFT error %d (%s)", \ 115 (PetscErrorCode)_p_cufft_stat__,name); \ 116 } \ 117 } \ 118 } while (0) 119 120 #define CHKERRCURAND(stat) do { \ 121 const curandStatus_t _p_curand_stat__ = stat; \ 122 if (PetscUnlikely(_p_curand_stat__ != CURAND_STATUS_SUCCESS)) { \ 123 if (((_p_curand_stat__ == CURAND_STATUS_INITIALIZATION_FAILED) || \ 124 (_p_curand_stat__ == CURAND_STATUS_ALLOCATION_FAILED)) && \ 125 PetscDeviceInitialized(PETSC_DEVICE_CUDA)) { \ 126 SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_GPU_RESOURCE, \ 127 "cuRAND error %d. " \ 128 "Reports not initialized or alloc failed; " \ 129 "this indicates the GPU has run out resources", \ 130 (PetscErrorCode)_p_curand_stat__); \ 131 } else { \ 132 SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_GPU, \ 133 "cuRand error %d",(PetscErrorCode)_p_curand_stat__); \ 134 } \ 135 } \ 136 } while (0) 137 138 PETSC_EXTERN cudaStream_t PetscDefaultCudaStream; /* The default stream used by PETSc */ 139 140 PETSC_EXTERN PetscErrorCode PetscCUBLASGetHandle(cublasHandle_t*); 141 PETSC_EXTERN PetscErrorCode PetscCUSOLVERDnGetHandle(cusolverDnHandle_t*); 142 #endif /* PETSC_HAVE_CUDA */ 143 144 #if defined(PETSC_HAVE_HIP) 145 #include <hip/hip_runtime.h> 146 #include <hipblas.h> 147 #if defined(__HIP_PLATFORM_NVCC__) 148 #include <cusolverDn.h> 149 #else /* __HIP_PLATFORM_HCC__ */ 150 #include <rocsolver.h> 151 #endif /* __HIP_PLATFORM_NVCC__ */ 152 153 /* REMOVE ME */ 154 #define WaitForHIP() hipDeviceSynchronize() 155 156 /* hipBLAS does not have hipblasGetErrorName(). We create one on our own. */ 157 PETSC_EXTERN const char* PetscHIPBLASGetErrorName(hipblasStatus_t); /* PETSC_EXTERN since it is exposed by the CHKERRHIPBLAS macro */ 158 159 #define CHKERRHIP(cerr) do { \ 160 const hipError_t _p_hip_err__ = cerr; \ 161 if (PetscUnlikely(_p_hip_err__ != hipSuccess)) { \ 162 const char *name = hipGetErrorName(_p_hip_err__); \ 163 const char *descr = hipGetErrorString(_p_hip_err__); \ 164 SETERRQ3(PETSC_COMM_SELF,PETSC_ERR_GPU,"hip error %d (%s) : %s", \ 165 (PetscErrorCode)_p_hip_err__,name,descr); \ 166 } \ 167 } while (0) 168 169 #define CHKERRHIPBLAS(stat) do { \ 170 const hipblasStatus_t _p_hipblas_stat__ = stat; \ 171 if (PetscUnlikely(_p_hipblas_stat__ != HIPBLAS_STATUS_SUCCESS)) { \ 172 const char *name = PetscHIPBLASGetErrorName(_p_hipblas_stat__); \ 173 SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_GPU,"hipBLAS error %d (%s)", \ 174 (PetscErrorCode)_p_hipblas_stat__,name); \ 175 } \ 176 } while (0) 177 178 /* TODO: SEK: Need to figure out the hipsolver issues */ 179 #define CHKERRHIPSOLVER(stat) do { \ 180 const hipsolverStatus_t _p_hipsolver_stat__ = stat; \ 181 if (PetscUnlikely(_p_hipsolver_stat__ /* != HIPSOLVER_STATUS_SUCCESS */)) { \ 182 SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_GPU,"HIPSOLVER error %d", \ 183 (PetscErrorCode)_p_hipsolver_stat__); \ 184 } \ 185 } while (0) 186 187 /* hipSolver does not exist yet so we work around it 188 rocSOLVER users rocBLAS for the handle 189 * */ 190 #if defined(__HIP_PLATFORM_NVCC__) 191 typedef cusolverDnHandle_t hipsolverHandle_t; 192 typedef cusolverStatus_t hipsolverStatus_t; 193 194 /* Alias hipsolverDestroy to cusolverDnDestroy */ 195 PETSC_STATIC_INLINE hipsolverStatus_t hipsolverDestroy(hipsolverHandle_t *hipsolverhandle) 196 { 197 return cusolverDnDestroy(hipsolverhandle) 198 } 199 200 /* Alias hipsolverCreate to cusolverDnCreate */ 201 PETSC_STATIC_INLINE hipsolverStatus_t hipsolverCreate(hipsolverHandle_t *hipsolverhandle) 202 { 203 return cusolverDnCreate(hipsolverhandle) 204 } 205 206 /* Alias hipsolverGetStream to cusolverDnGetStream */ 207 PETSC_STATIC_INLINE hipsolverStatus_t hipsolverGetStream(hipsolverHandle_t handle, hipStream_t *stream) 208 { 209 return cusolverDnGetStream(handle,stream); 210 } 211 212 /* Alias hipsolverSetStream to cusolverDnSetStream */ 213 PETSC_STATIC_INLINE hipsolverStatus_t hipsolverSetStream(hipsolverHandle_t handle, hipStream_t stream) 214 { 215 return cusolveDnSetStream(handle,stream); 216 } 217 #else /* __HIP_PLATFORM_HCC__ */ 218 typedef rocblas_handle hipsolverHandle_t; 219 typedef rocblas_status hipsolverStatus_t; 220 221 /* Alias hipsolverDestroy to rocblas_destroy_handle */ 222 PETSC_STATIC_INLINE hipsolverStatus_t hipsolverDestroy(hipsolverHandle_t hipsolverhandle) 223 { 224 return rocblas_destroy_handle(hipsolverhandle); 225 } 226 227 /* Alias hipsolverCreate to rocblas_destroy_handle */ 228 PETSC_STATIC_INLINE hipsolverStatus_t hipsolverCreate(hipsolverHandle_t *hipsolverhandle) 229 { 230 return rocblas_create_handle(hipsolverhandle); 231 } 232 233 /* Alias hipsolverGetStream to rocblas_get_stream */ 234 PETSC_STATIC_INLINE hipsolverStatus_t hipsolverGetStream(hipsolverHandle_t handle, hipStream_t *stream) 235 { 236 return rocblas_get_stream(handle,stream); 237 } 238 239 /* Alias hipsolverSetStream to rocblas_set_stream */ 240 PETSC_STATIC_INLINE hipsolverStatus_t hipsolverSetStream(hipsolverHandle_t handle, hipStream_t stream) 241 { 242 return rocblas_set_stream(handle,stream); 243 } 244 #endif /* __HIP_PLATFORM_NVCC__ */ 245 PETSC_EXTERN hipStream_t PetscDefaultHipStream; /* The default stream used by PETSc */ 246 247 PETSC_EXTERN PetscErrorCode PetscHIPBLASGetHandle(hipblasHandle_t*); 248 PETSC_EXTERN PetscErrorCode PetscHIPSOLVERGetHandle(hipsolverHandle_t*); 249 #endif /* PETSC_HAVE_HIP */ 250 251 /* Cannot use the device context api without C++11 */ 252 #if defined(PETSC_HAVE_CXX_DIALECT_CXX11) 253 PETSC_EXTERN PetscErrorCode PetscDeviceInitializePackage(void); 254 PETSC_EXTERN PetscErrorCode PetscDeviceFinalizePackage(void); 255 256 /* PetscDevice */ 257 PETSC_EXTERN PetscErrorCode PetscDeviceInitialize(PetscDeviceType); 258 PETSC_EXTERN PetscBool PetscDeviceInitialized(PetscDeviceType); 259 PETSC_EXTERN PetscErrorCode PetscDeviceCreate(PetscDeviceType,PetscInt,PetscDevice*); 260 PETSC_EXTERN PetscErrorCode PetscDeviceConfigure(PetscDevice); 261 PETSC_EXTERN PetscErrorCode PetscDeviceView(PetscDevice,PetscViewer); 262 PETSC_EXTERN PetscErrorCode PetscDeviceDestroy(PetscDevice*); 263 264 /* PetscDeviceContext */ 265 PETSC_EXTERN PetscErrorCode PetscDeviceContextCreate(PetscDeviceContext*); 266 PETSC_EXTERN PetscErrorCode PetscDeviceContextDestroy(PetscDeviceContext*); 267 PETSC_EXTERN PetscErrorCode PetscDeviceContextSetDevice(PetscDeviceContext,PetscDevice); 268 PETSC_EXTERN PetscErrorCode PetscDeviceContextGetDevice(PetscDeviceContext,PetscDevice*); 269 PETSC_EXTERN PetscErrorCode PetscDeviceContextSetStreamType(PetscDeviceContext,PetscStreamType); 270 PETSC_EXTERN PetscErrorCode PetscDeviceContextGetStreamType(PetscDeviceContext,PetscStreamType*); 271 PETSC_EXTERN PetscErrorCode PetscDeviceContextSetUp(PetscDeviceContext); 272 PETSC_EXTERN PetscErrorCode PetscDeviceContextDuplicate(PetscDeviceContext,PetscDeviceContext*); 273 PETSC_EXTERN PetscErrorCode PetscDeviceContextQueryIdle(PetscDeviceContext,PetscBool*); 274 PETSC_EXTERN PetscErrorCode PetscDeviceContextWaitForContext(PetscDeviceContext,PetscDeviceContext); 275 PETSC_EXTERN PetscErrorCode PetscDeviceContextFork(PetscDeviceContext,PetscInt,PetscDeviceContext**); 276 PETSC_EXTERN PetscErrorCode PetscDeviceContextJoin(PetscDeviceContext,PetscInt,PetscDeviceContextJoinMode,PetscDeviceContext**); 277 PETSC_EXTERN PetscErrorCode PetscDeviceContextSynchronize(PetscDeviceContext); 278 PETSC_EXTERN PetscErrorCode PetscDeviceContextGetCurrentContext(PetscDeviceContext*); 279 PETSC_EXTERN PetscErrorCode PetscDeviceContextSetCurrentContext(PetscDeviceContext); 280 PETSC_EXTERN PetscErrorCode PetscDeviceContextSetFromOptions(MPI_Comm,const char[],PetscDeviceContext); 281 #endif /* PETSC_HAVE_CXX_DIALECT_CXX11 */ 282 #endif /* PETSCDEVICE_H */ 283