Actual source code: mpiaijcusp.cu
petsc-3.5.4 2015-05-23
1: #include <petscconf.h>
2: PETSC_CUDA_EXTERN_C_BEGIN
3: #include <../src/mat/impls/aij/mpi/mpiaij.h> /*I "petscmat.h" I*/
4: PETSC_CUDA_EXTERN_C_END
5: #include <../src/mat/impls/aij/mpi/mpicusp/mpicuspmatimpl.h>
9: PetscErrorCode MatMPIAIJSetPreallocation_MPIAIJCUSP(Mat B,PetscInt d_nz,const PetscInt d_nnz[],PetscInt o_nz,const PetscInt o_nnz[])
10: {
11: Mat_MPIAIJ *b = (Mat_MPIAIJ*)B->data;
12: Mat_MPIAIJCUSP * cuspStruct = (Mat_MPIAIJCUSP*)b->spptr;
14: PetscInt i;
17: PetscLayoutSetUp(B->rmap);
18: PetscLayoutSetUp(B->cmap);
19: if (d_nnz) {
20: for (i=0; i<B->rmap->n; i++) {
21: 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]);
22: }
23: }
24: if (o_nnz) {
25: for (i=0; i<B->rmap->n; i++) {
26: 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]);
27: }
28: }
29: if (!B->preallocated) {
30: /* Explicitly create 2 MATSEQAIJCUSP matrices. */
31: MatCreate(PETSC_COMM_SELF,&b->A);
32: MatSetSizes(b->A,B->rmap->n,B->cmap->n,B->rmap->n,B->cmap->n);
33: MatSetType(b->A,MATSEQAIJCUSP);
34: PetscLogObjectParent((PetscObject)B,(PetscObject)b->A);
35: MatCreate(PETSC_COMM_SELF,&b->B);
36: MatSetSizes(b->B,B->rmap->n,B->cmap->N,B->rmap->n,B->cmap->N);
37: MatSetType(b->B,MATSEQAIJCUSP);
38: PetscLogObjectParent((PetscObject)B,(PetscObject)b->B);
39: }
40: MatSeqAIJSetPreallocation(b->A,d_nz,d_nnz);
41: MatSeqAIJSetPreallocation(b->B,o_nz,o_nnz);
42: MatCUSPSetFormat(b->A,MAT_CUSP_MULT,cuspStruct->diagGPUMatFormat);
43: MatCUSPSetFormat(b->B,MAT_CUSP_MULT,cuspStruct->offdiagGPUMatFormat);
44: MatCUSPSetStream(b->A,cuspStruct->stream);
45: MatCUSPSetStream(b->B,cuspStruct->stream);
46: B->preallocated = PETSC_TRUE;
47: return(0);
48: }
52: PetscErrorCode MatGetVecs_MPIAIJCUSP(Mat mat,Vec *right,Vec *left)
53: {
55: PetscInt rbs,cbs;
58: MatGetBlockSizes(mat,&rbs,&cbs);
59: if (right) {
60: VecCreate(PetscObjectComm((PetscObject)mat),right);
61: VecSetSizes(*right,mat->cmap->n,PETSC_DETERMINE);
62: VecSetBlockSize(*right,cbs);
63: VecSetType(*right,VECCUSP);
64: VecSetLayout(*right,mat->cmap);
65: }
66: if (left) {
67: VecCreate(PetscObjectComm((PetscObject)mat),left);
68: VecSetSizes(*left,mat->rmap->n,PETSC_DETERMINE);
69: VecSetBlockSize(*left,rbs);
70: VecSetType(*left,VECCUSP);
71: VecSetLayout(*left,mat->rmap);
72: }
73: return(0);
74: }
79: PetscErrorCode MatMult_MPIAIJCUSP(Mat A,Vec xx,Vec yy)
80: {
81: /* This multiplication sequence is different sequence
82: than the CPU version. In particular, the diagonal block
83: multiplication kernel is launched in one stream. Then,
84: in a separate stream, the data transfers from DeviceToHost
85: (with MPI messaging in between), then HostToDevice are
86: launched. Once the data transfer stream is synchronized,
87: to ensure messaging is complete, the MatMultAdd kernel
88: is launched in the original (MatMult) stream to protect
89: against race conditions.
91: This sequence should only be called for GPU computation. */
92: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
94: PetscInt nt;
97: VecGetLocalSize(xx,&nt);
98: 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);
99: VecScatterInitializeForGPU(a->Mvctx,xx,SCATTER_FORWARD);
100: (*a->A->ops->mult)(a->A,xx,yy);
101: VecScatterBegin(a->Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);
102: VecScatterEnd(a->Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);
103: (*a->B->ops->multadd)(a->B,a->lvec,yy,yy);
104: VecScatterFinalizeForGPU(a->Mvctx);
105: return(0);
106: }
108: PetscErrorCode MatSetValuesBatch_MPIAIJCUSP(Mat J, PetscInt Ne, PetscInt Nl, PetscInt *elemRows, const PetscScalar *elemMats);
112: PetscErrorCode MatCUSPSetFormat_MPIAIJCUSP(Mat A,MatCUSPFormatOperation op,MatCUSPStorageFormat format)
113: {
114: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
115: Mat_MPIAIJCUSP * cuspStruct = (Mat_MPIAIJCUSP*)a->spptr;
118: switch (op) {
119: case MAT_CUSP_MULT_DIAG:
120: cuspStruct->diagGPUMatFormat = format;
121: break;
122: case MAT_CUSP_MULT_OFFDIAG:
123: cuspStruct->offdiagGPUMatFormat = format;
124: break;
125: case MAT_CUSP_ALL:
126: cuspStruct->diagGPUMatFormat = format;
127: cuspStruct->offdiagGPUMatFormat = format;
128: break;
129: default:
130: SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"unsupported operation %d for MatCUSPFormatOperation. Only MAT_CUSP_MULT_DIAG, MAT_CUSP_MULT_DIAG, and MAT_CUSP_MULT_ALL are currently supported.",op);
131: }
132: return(0);
133: }
137: PetscErrorCode MatSetFromOptions_MPIAIJCUSP(Mat A)
138: {
139: MatCUSPStorageFormat format;
140: PetscErrorCode ierr;
141: PetscBool flg;
144: PetscOptionsHead("MPIAIJCUSP options");
145: PetscObjectOptionsBegin((PetscObject)A);
146: if (A->factortype==MAT_FACTOR_NONE) {
147: PetscOptionsEnum("-mat_cusp_mult_diag_storage_format","sets storage format of the diagonal blocks of (mpi)aijcusp gpu matrices for SpMV",
148: "MatCUSPSetFormat",MatCUSPStorageFormats,(PetscEnum)MAT_CUSP_CSR,(PetscEnum*)&format,&flg);
149: if (flg) {
150: MatCUSPSetFormat(A,MAT_CUSP_MULT_DIAG,format);
151: }
152: PetscOptionsEnum("-mat_cusp_mult_offdiag_storage_format","sets storage format of the off-diagonal blocks (mpi)aijcusp gpu matrices for SpMV",
153: "MatCUSPSetFormat",MatCUSPStorageFormats,(PetscEnum)MAT_CUSP_CSR,(PetscEnum*)&format,&flg);
154: if (flg) {
155: MatCUSPSetFormat(A,MAT_CUSP_MULT_OFFDIAG,format);
156: }
157: PetscOptionsEnum("-mat_cusp_storage_format","sets storage format of the diagonal and off-diagonal blocks (mpi)aijcusp gpu matrices for SpMV",
158: "MatCUSPSetFormat",MatCUSPStorageFormats,(PetscEnum)MAT_CUSP_CSR,(PetscEnum*)&format,&flg);
159: if (flg) {
160: MatCUSPSetFormat(A,MAT_CUSP_ALL,format);
161: }
162: }
163: PetscOptionsEnd();
164: return(0);
165: }
169: PetscErrorCode MatDestroy_MPIAIJCUSP(Mat A)
170: {
172: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
173: Mat_MPIAIJCUSP *cuspStruct = (Mat_MPIAIJCUSP*)a->spptr;
174: cudaError_t err=cudaSuccess;
177: try {
178: err = cudaStreamDestroy(cuspStruct->stream);
179: if (err!=cudaSuccess)
180: SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Mat_MPIAIJCUSP error: %s", cudaGetErrorString(err));
181: delete cuspStruct;
182: } catch(char *ex) {
183: SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Mat_MPIAIJCUSP error: %s", ex);
184: }
185: cuspStruct = 0;
186: MatDestroy_MPIAIJ(A);
187: return(0);
188: }
192: PETSC_EXTERN PetscErrorCode MatCreate_MPIAIJCUSP(Mat A)
193: {
195: Mat_MPIAIJ *a;
196: Mat_MPIAIJCUSP * cuspStruct;
197: cudaError_t err=cudaSuccess;
200: MatCreate_MPIAIJ(A);
201: PetscObjectComposeFunction((PetscObject)A,"MatMPIAIJSetPreallocation_C",MatMPIAIJSetPreallocation_MPIAIJCUSP);
202: A->ops->getvecs = MatGetVecs_MPIAIJCUSP;
203: A->ops->setvaluesbatch = MatSetValuesBatch_MPIAIJCUSP;
205: a = (Mat_MPIAIJ*)A->data;
206: a->spptr = new Mat_MPIAIJCUSP;
207: cuspStruct = (Mat_MPIAIJCUSP*)a->spptr;
209: cuspStruct->diagGPUMatFormat = MAT_CUSP_CSR;
210: cuspStruct->offdiagGPUMatFormat = MAT_CUSP_CSR;
211: err = cudaStreamCreate(&(cuspStruct->stream));
212: if (err!=cudaSuccess)
213: SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Mat_MPIAIJCUSP error: %s", cudaGetErrorString(err));
215: A->ops->mult = MatMult_MPIAIJCUSP;
216: A->ops->setfromoptions = MatSetFromOptions_MPIAIJCUSP;
217: A->ops->destroy = MatDestroy_MPIAIJCUSP;
219: PetscObjectComposeFunction((PetscObject)A,"MatCUSPSetFormat_C", MatCUSPSetFormat_MPIAIJCUSP);
220: PetscObjectChangeTypeName((PetscObject)A,MATMPIAIJCUSP);
221: return(0);
222: }
225: /*@
226: MatCreateAIJCUSP - Creates a sparse matrix in AIJ (compressed row) format
227: (the default parallel PETSc format). This matrix will ultimately pushed down
228: to NVidia GPUs and use the CUSP library for calculations. For good matrix
229: assembly performance the user should preallocate the matrix storage by setting
230: the parameter nz (or the array nnz). By setting these parameters accurately,
231: performance during matrix assembly can be increased by more than a factor of 50.
234: Collective on MPI_Comm
236: Input Parameters:
237: + comm - MPI communicator, set to PETSC_COMM_SELF
238: . m - number of rows
239: . n - number of columns
240: . nz - number of nonzeros per row (same for all rows)
241: - nnz - array containing the number of nonzeros in the various rows
242: (possibly different for each row) or NULL
244: Output Parameter:
245: . A - the matrix
247: It is recommended that one use the MatCreate(), MatSetType() and/or MatSetFromOptions(),
248: MatXXXXSetPreallocation() paradigm instead of this routine directly.
249: [MatXXXXSetPreallocation() is, for example, MatSeqAIJSetPreallocation]
251: Notes:
252: If nnz is given then nz is ignored
254: The AIJ format (also called the Yale sparse matrix format or
255: compressed row storage), is fully compatible with standard Fortran 77
256: storage. That is, the stored row and column indices can begin at
257: either one (as in Fortran) or zero. See the users' manual for details.
259: Specify the preallocated storage with either nz or nnz (not both).
260: Set nz=PETSC_DEFAULT and nnz=NULL for PETSc to control dynamic memory
261: allocation. For large problems you MUST preallocate memory or you
262: will get TERRIBLE performance, see the users' manual chapter on matrices.
264: By default, this format uses inodes (identical nodes) when possible, to
265: improve numerical efficiency of matrix-vector products and solves. We
266: search for consecutive rows with the same nonzero structure, thereby
267: reusing matrix information to achieve increased efficiency.
269: Level: intermediate
271: .seealso: MatCreate(), MatCreateAIJ(), MatSetValues(), MatSeqAIJSetColumnIndices(), MatCreateSeqAIJWithArrays(), MatCreateAIJ(), MATMPIAIJCUSP, MATAIJCUSP
272: @*/
275: PetscErrorCode MatCreateAIJCUSP(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)
276: {
278: PetscMPIInt size;
281: MatCreate(comm,A);
282: MatSetSizes(*A,m,n,M,N);
283: MPI_Comm_size(comm,&size);
284: if (size > 1) {
285: MatSetType(*A,MATMPIAIJCUSP);
286: MatMPIAIJSetPreallocation(*A,d_nz,d_nnz,o_nz,o_nnz);
287: } else {
288: MatSetType(*A,MATSEQAIJCUSP);
289: MatSeqAIJSetPreallocation(*A,d_nz,d_nnz);
290: }
291: return(0);
292: }
294: /*M
295: MATAIJCUSP - MATMPIAIJCUSP= "aijcusp" = "mpiaijcusp" - A matrix type to be used for sparse matrices.
297: A matrix type type whose data resides on Nvidia GPUs. These matrices can be CSR format.
298: All matrix calculations are performed using the CUSP library. DIA and ELL
299: formats are also available
301: This matrix type is identical to MATSEQAIJCUSP when constructed with a single process communicator,
302: and MATMPIAIJCUSP otherwise. As a result, for single process communicators,
303: MatSeqAIJSetPreallocation is supported, and similarly MatMPIAIJSetPreallocation is supported
304: for communicators controlling multiple processes. It is recommended that you call both of
305: the above preallocation routines for simplicity.
307: Options Database Keys:
308: + -mat_type mpiaijcusp - sets the matrix type to "mpiaijcusp" during a call to MatSetFromOptions()
309: . -mat_cusp_storage_format csr - sets the storage format of diagonal and off-diagonal matrices during a call to MatSetFromOptions(). Other storage formats include dia (diagonal) or ell (ellpack).
310: . -mat_cusp_mult_diag_storage_format csr - sets the storage format of diagonal matrix during a call to MatSetFromOptions(). Other storage formats include dia (diagonal) or ell (ellpack).
311: - -mat_cusp_mult_offdiag_storage_format csr - sets the storage format of off-diagonal matrix during a call to MatSetFromOptions(). Other storage formats include dia (diagonal) or ell (ellpack).
313: Level: beginner
315: .seealso: MatCreateAIJCUSP(), MATSEQAIJCUSP, MatCreateSeqAIJCUSP(), MatCUSPSetFormat(), MatCUSPStorageFormat, MatCUSPFormatOperation
316: M*/