Actual source code: mpiaijcusparse.cu
petsc-3.11.4 2019-09-28
1: #define PETSC_SKIP_SPINLOCK
3: #include <petscconf.h>
4: #include <../src/mat/impls/aij/mpi/mpiaij.h>
5: #include <../src/mat/impls/aij/mpi/mpicusparse/mpicusparsematimpl.h>
7: PetscErrorCode MatMPIAIJSetPreallocation_MPIAIJCUSPARSE(Mat B,PetscInt d_nz,const PetscInt d_nnz[],PetscInt o_nz,const PetscInt o_nnz[])
8: {
9: Mat_MPIAIJ *b = (Mat_MPIAIJ*)B->data;
10: Mat_MPIAIJCUSPARSE * cusparseStruct = (Mat_MPIAIJCUSPARSE*)b->spptr;
11: PetscErrorCode ierr;
12: PetscInt i;
15: PetscLayoutSetUp(B->rmap);
16: PetscLayoutSetUp(B->cmap);
17: if (d_nnz) {
18: for (i=0; i<B->rmap->n; i++) {
19: if (d_nnz[i] < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"d_nnz cannot be less than 0: local row %D value %D",i,d_nnz[i]);
20: }
21: }
22: if (o_nnz) {
23: for (i=0; i<B->rmap->n; i++) {
24: if (o_nnz[i] < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"o_nnz cannot be less than 0: local row %D value %D",i,o_nnz[i]);
25: }
26: }
27: if (!B->preallocated) {
28: /* Explicitly create 2 MATSEQAIJCUSPARSE matrices. */
29: MatCreate(PETSC_COMM_SELF,&b->A);
30: MatSetSizes(b->A,B->rmap->n,B->cmap->n,B->rmap->n,B->cmap->n);
31: MatSetType(b->A,MATSEQAIJCUSPARSE);
32: PetscLogObjectParent((PetscObject)B,(PetscObject)b->A);
33: MatCreate(PETSC_COMM_SELF,&b->B);
34: MatSetSizes(b->B,B->rmap->n,B->cmap->N,B->rmap->n,B->cmap->N);
35: MatSetType(b->B,MATSEQAIJCUSPARSE);
36: PetscLogObjectParent((PetscObject)B,(PetscObject)b->B);
37: }
38: MatSeqAIJSetPreallocation(b->A,d_nz,d_nnz);
39: MatSeqAIJSetPreallocation(b->B,o_nz,o_nnz);
40: MatCUSPARSESetFormat(b->A,MAT_CUSPARSE_MULT,cusparseStruct->diagGPUMatFormat);
41: MatCUSPARSESetFormat(b->B,MAT_CUSPARSE_MULT,cusparseStruct->offdiagGPUMatFormat);
42: MatCUSPARSESetHandle(b->A,cusparseStruct->handle);
43: MatCUSPARSESetHandle(b->B,cusparseStruct->handle);
44: MatCUSPARSESetStream(b->A,cusparseStruct->stream);
45: MatCUSPARSESetStream(b->B,cusparseStruct->stream);
47: B->preallocated = PETSC_TRUE;
48: return(0);
49: }
51: PetscErrorCode MatMult_MPIAIJCUSPARSE(Mat A,Vec xx,Vec yy)
52: {
53: /* This multiplication sequence is different sequence
54: than the CPU version. In particular, the diagonal block
55: multiplication kernel is launched in one stream. Then,
56: in a separate stream, the data transfers from DeviceToHost
57: (with MPI messaging in between), then HostToDevice are
58: launched. Once the data transfer stream is synchronized,
59: to ensure messaging is complete, the MatMultAdd kernel
60: is launched in the original (MatMult) stream to protect
61: against race conditions.
63: This sequence should only be called for GPU computation. */
64: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
66: PetscInt nt;
69: VecGetLocalSize(xx,&nt);
70: if (nt != A->cmap->n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Incompatible partition of A (%D) and xx (%D)",A->cmap->n,nt);
71: VecScatterInitializeForGPU(a->Mvctx,xx);
72: (*a->A->ops->mult)(a->A,xx,yy);
73: VecScatterBegin(a->Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);
74: VecScatterEnd(a->Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);
75: (*a->B->ops->multadd)(a->B,a->lvec,yy,yy);
76: VecScatterFinalizeForGPU(a->Mvctx);
77: return(0);
78: }
80: PetscErrorCode MatMultTranspose_MPIAIJCUSPARSE(Mat A,Vec xx,Vec yy)
81: {
82: /* This multiplication sequence is different sequence
83: than the CPU version. In particular, the diagonal block
84: multiplication kernel is launched in one stream. Then,
85: in a separate stream, the data transfers from DeviceToHost
86: (with MPI messaging in between), then HostToDevice are
87: launched. Once the data transfer stream is synchronized,
88: to ensure messaging is complete, the MatMultAdd kernel
89: is launched in the original (MatMult) stream to protect
90: against race conditions.
92: This sequence should only be called for GPU computation. */
93: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
95: PetscInt nt;
98: VecGetLocalSize(xx,&nt);
99: if (nt != A->rmap->n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Incompatible partition of A (%D) and xx (%D)",A->rmap->n,nt);
100: VecScatterInitializeForGPU(a->Mvctx,xx);
101: (*a->B->ops->multtranspose)(a->B,xx,a->lvec);
102: (*a->A->ops->multtranspose)(a->A,xx,yy);
103: VecScatterBegin(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);
104: VecScatterEnd(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);
105: VecScatterFinalizeForGPU(a->Mvctx);
106: return(0);
107: }
109: PetscErrorCode MatCUSPARSESetFormat_MPIAIJCUSPARSE(Mat A,MatCUSPARSEFormatOperation op,MatCUSPARSEStorageFormat format)
110: {
111: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
112: Mat_MPIAIJCUSPARSE * cusparseStruct = (Mat_MPIAIJCUSPARSE*)a->spptr;
115: switch (op) {
116: case MAT_CUSPARSE_MULT_DIAG:
117: cusparseStruct->diagGPUMatFormat = format;
118: break;
119: case MAT_CUSPARSE_MULT_OFFDIAG:
120: cusparseStruct->offdiagGPUMatFormat = format;
121: break;
122: case MAT_CUSPARSE_ALL:
123: cusparseStruct->diagGPUMatFormat = format;
124: cusparseStruct->offdiagGPUMatFormat = format;
125: break;
126: default:
127: SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"unsupported operation %d for MatCUSPARSEFormatOperation. Only MAT_CUSPARSE_MULT_DIAG, MAT_CUSPARSE_MULT_DIAG, and MAT_CUSPARSE_MULT_ALL are currently supported.",op);
128: }
129: return(0);
130: }
132: PetscErrorCode MatSetFromOptions_MPIAIJCUSPARSE(PetscOptionItems *PetscOptionsObject,Mat A)
133: {
134: MatCUSPARSEStorageFormat format;
135: PetscErrorCode ierr;
136: PetscBool flg;
137: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
138: Mat_MPIAIJCUSPARSE *cusparseStruct = (Mat_MPIAIJCUSPARSE*)a->spptr;
141: PetscOptionsHead(PetscOptionsObject,"MPIAIJCUSPARSE options");
142: if (A->factortype==MAT_FACTOR_NONE) {
143: PetscOptionsEnum("-mat_cusparse_mult_diag_storage_format","sets storage format of the diagonal blocks of (mpi)aijcusparse gpu matrices for SpMV",
144: "MatCUSPARSESetFormat",MatCUSPARSEStorageFormats,(PetscEnum)cusparseStruct->diagGPUMatFormat,(PetscEnum*)&format,&flg);
145: if (flg) {
146: MatCUSPARSESetFormat(A,MAT_CUSPARSE_MULT_DIAG,format);
147: }
148: PetscOptionsEnum("-mat_cusparse_mult_offdiag_storage_format","sets storage format of the off-diagonal blocks (mpi)aijcusparse gpu matrices for SpMV",
149: "MatCUSPARSESetFormat",MatCUSPARSEStorageFormats,(PetscEnum)cusparseStruct->offdiagGPUMatFormat,(PetscEnum*)&format,&flg);
150: if (flg) {
151: MatCUSPARSESetFormat(A,MAT_CUSPARSE_MULT_OFFDIAG,format);
152: }
153: PetscOptionsEnum("-mat_cusparse_storage_format","sets storage format of the diagonal and off-diagonal blocks (mpi)aijcusparse gpu matrices for SpMV",
154: "MatCUSPARSESetFormat",MatCUSPARSEStorageFormats,(PetscEnum)cusparseStruct->diagGPUMatFormat,(PetscEnum*)&format,&flg);
155: if (flg) {
156: MatCUSPARSESetFormat(A,MAT_CUSPARSE_ALL,format);
157: }
158: }
159: PetscOptionsTail();
160: return(0);
161: }
163: PetscErrorCode MatAssemblyEnd_MPIAIJCUSPARSE(Mat A,MatAssemblyType mode)
164: {
166: Mat_MPIAIJ *mpiaij;
169: mpiaij = (Mat_MPIAIJ*)A->data;
170: MatAssemblyEnd_MPIAIJ(A,mode);
171: if (!A->was_assembled && mode == MAT_FINAL_ASSEMBLY) {
172: VecSetType(mpiaij->lvec,VECSEQCUDA);
173: }
174: return(0);
175: }
177: PetscErrorCode MatDestroy_MPIAIJCUSPARSE(Mat A)
178: {
179: PetscErrorCode ierr;
180: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
181: Mat_MPIAIJCUSPARSE *cusparseStruct = (Mat_MPIAIJCUSPARSE*)a->spptr;
182: cudaError_t err;
183: cusparseStatus_t stat;
186: try {
187: MatCUSPARSEClearHandle(a->A);
188: MatCUSPARSEClearHandle(a->B);
189: stat = cusparseDestroy(cusparseStruct->handle);CHKERRCUDA(stat);
190: err = cudaStreamDestroy(cusparseStruct->stream);CHKERRCUDA(err);
191: delete cusparseStruct;
192: } catch(char *ex) {
193: SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Mat_MPIAIJCUSPARSE error: %s", ex);
194: }
195: cusparseStruct = 0;
197: MatDestroy_MPIAIJ(A);
198: return(0);
199: }
201: PETSC_EXTERN PetscErrorCode MatCreate_MPIAIJCUSPARSE(Mat A)
202: {
203: PetscErrorCode ierr;
204: Mat_MPIAIJ *a;
205: Mat_MPIAIJCUSPARSE * cusparseStruct;
206: cudaError_t err;
207: cusparseStatus_t stat;
210: MatCreate_MPIAIJ(A);
211: PetscObjectComposeFunction((PetscObject)A,"MatMPIAIJSetPreallocation_C",MatMPIAIJSetPreallocation_MPIAIJCUSPARSE);
212: PetscFree(A->defaultvectype);
213: PetscStrallocpy(VECCUDA,&A->defaultvectype);
215: a = (Mat_MPIAIJ*)A->data;
216: a->spptr = new Mat_MPIAIJCUSPARSE;
218: cusparseStruct = (Mat_MPIAIJCUSPARSE*)a->spptr;
219: cusparseStruct->diagGPUMatFormat = MAT_CUSPARSE_CSR;
220: cusparseStruct->offdiagGPUMatFormat = MAT_CUSPARSE_CSR;
221: stat = cusparseCreate(&(cusparseStruct->handle));CHKERRCUDA(stat);
222: err = cudaStreamCreate(&(cusparseStruct->stream));CHKERRCUDA(err);
224: A->ops->assemblyend = MatAssemblyEnd_MPIAIJCUSPARSE;
225: A->ops->mult = MatMult_MPIAIJCUSPARSE;
226: A->ops->multtranspose = MatMultTranspose_MPIAIJCUSPARSE;
227: A->ops->setfromoptions = MatSetFromOptions_MPIAIJCUSPARSE;
228: A->ops->destroy = MatDestroy_MPIAIJCUSPARSE;
230: PetscObjectChangeTypeName((PetscObject)A,MATMPIAIJCUSPARSE);
231: PetscObjectComposeFunction((PetscObject)A,"MatCUSPARSESetFormat_C", MatCUSPARSESetFormat_MPIAIJCUSPARSE);
232: return(0);
233: }
235: /*@
236: MatCreateAIJCUSPARSE - Creates a sparse matrix in AIJ (compressed row) format
237: (the default parallel PETSc format). This matrix will ultimately pushed down
238: to NVidia GPUs and use the CUSPARSE library for calculations. For good matrix
239: assembly performance the user should preallocate the matrix storage by setting
240: the parameter nz (or the array nnz). By setting these parameters accurately,
241: performance during matrix assembly can be increased by more than a factor of 50.
243: Collective on MPI_Comm
245: Input Parameters:
246: + comm - MPI communicator, set to PETSC_COMM_SELF
247: . m - number of rows
248: . n - number of columns
249: . nz - number of nonzeros per row (same for all rows)
250: - nnz - array containing the number of nonzeros in the various rows
251: (possibly different for each row) or NULL
253: Output Parameter:
254: . A - the matrix
256: It is recommended that one use the MatCreate(), MatSetType() and/or MatSetFromOptions(),
257: MatXXXXSetPreallocation() paradigm instead of this routine directly.
258: [MatXXXXSetPreallocation() is, for example, MatSeqAIJSetPreallocation]
260: Notes:
261: If nnz is given then nz is ignored
263: The AIJ format (also called the Yale sparse matrix format or
264: compressed row storage), is fully compatible with standard Fortran 77
265: storage. That is, the stored row and column indices can begin at
266: either one (as in Fortran) or zero. See the users' manual for details.
268: Specify the preallocated storage with either nz or nnz (not both).
269: Set nz=PETSC_DEFAULT and nnz=NULL for PETSc to control dynamic memory
270: allocation. For large problems you MUST preallocate memory or you
271: will get TERRIBLE performance, see the users' manual chapter on matrices.
273: By default, this format uses inodes (identical nodes) when possible, to
274: improve numerical efficiency of matrix-vector products and solves. We
275: search for consecutive rows with the same nonzero structure, thereby
276: reusing matrix information to achieve increased efficiency.
278: Level: intermediate
280: .seealso: MatCreate(), MatCreateAIJ(), MatSetValues(), MatSeqAIJSetColumnIndices(), MatCreateSeqAIJWithArrays(), MatCreateAIJ(), MATMPIAIJCUSPARSE, MATAIJCUSPARSE
281: @*/
282: PetscErrorCode MatCreateAIJCUSPARSE(MPI_Comm comm,PetscInt m,PetscInt n,PetscInt M,PetscInt N,PetscInt d_nz,const PetscInt d_nnz[],PetscInt o_nz,const PetscInt o_nnz[],Mat *A)
283: {
285: PetscMPIInt size;
288: MatCreate(comm,A);
289: MatSetSizes(*A,m,n,M,N);
290: MPI_Comm_size(comm,&size);
291: if (size > 1) {
292: MatSetType(*A,MATMPIAIJCUSPARSE);
293: MatMPIAIJSetPreallocation(*A,d_nz,d_nnz,o_nz,o_nnz);
294: } else {
295: MatSetType(*A,MATSEQAIJCUSPARSE);
296: MatSeqAIJSetPreallocation(*A,d_nz,d_nnz);
297: }
298: return(0);
299: }
301: /*MC
302: MATAIJCUSPARSE - MATMPIAIJCUSPARSE = "aijcusparse" = "mpiaijcusparse" - A matrix type to be used for sparse matrices.
304: A matrix type type whose data resides on Nvidia GPUs. These matrices can be in either
305: CSR, ELL, or Hybrid format. The ELL and HYB formats require CUDA 4.2 or later.
306: All matrix calculations are performed on Nvidia GPUs using the CUSPARSE library.
308: This matrix type is identical to MATSEQAIJCUSPARSE when constructed with a single process communicator,
309: and MATMPIAIJCUSPARSE otherwise. As a result, for single process communicators,
310: MatSeqAIJSetPreallocation is supported, and similarly MatMPIAIJSetPreallocation is supported
311: for communicators controlling multiple processes. It is recommended that you call both of
312: the above preallocation routines for simplicity.
314: Options Database Keys:
315: + -mat_type mpiaijcusparse - sets the matrix type to "mpiaijcusparse" during a call to MatSetFromOptions()
316: . -mat_cusparse_storage_format csr - sets the storage format of diagonal and off-diagonal matrices during a call to MatSetFromOptions(). Other options include ell (ellpack) or hyb (hybrid).
317: . -mat_cusparse_mult_diag_storage_format csr - sets the storage format of diagonal matrix during a call to MatSetFromOptions(). Other options include ell (ellpack) or hyb (hybrid).
318: - -mat_cusparse_mult_offdiag_storage_format csr - sets the storage format of off-diagonal matrix during a call to MatSetFromOptions(). Other options include ell (ellpack) or hyb (hybrid).
320: Level: beginner
322: .seealso: MatCreateAIJCUSPARSE(), MATSEQAIJCUSPARSE, MatCreateSeqAIJCUSPARSE(), MatCUSPARSESetFormat(), MatCUSPARSEStorageFormat, MatCUSPARSEFormatOperation
323: M
324: M*/