Actual source code: mpiaijcusp.cu

petsc-3.6.1 2015-08-06
Report Typos and Errors
  1: #define PETSC_SKIP_COMPLEX

  3: #include <petscconf.h>
  4: PETSC_CUDA_EXTERN_C_BEGIN
  5: #include <../src/mat/impls/aij/mpi/mpiaij.h>   /*I "petscmat.h" I*/
  6: PETSC_CUDA_EXTERN_C_END
  7: #include <../src/mat/impls/aij/mpi/mpicusp/mpicuspmatimpl.h>

 11: PetscErrorCode  MatMPIAIJSetPreallocation_MPIAIJCUSP(Mat B,PetscInt d_nz,const PetscInt d_nnz[],PetscInt o_nz,const PetscInt o_nnz[])
 12: {
 13:   Mat_MPIAIJ *b = (Mat_MPIAIJ*)B->data;
 14:   Mat_MPIAIJCUSP * cuspStruct = (Mat_MPIAIJCUSP*)b->spptr;
 16:   PetscInt       i;

 19:   PetscLayoutSetUp(B->rmap);
 20:   PetscLayoutSetUp(B->cmap);
 21:   if (d_nnz) {
 22:     for (i=0; i<B->rmap->n; i++) {
 23:       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]);
 24:     }
 25:   }
 26:   if (o_nnz) {
 27:     for (i=0; i<B->rmap->n; i++) {
 28:       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]);
 29:     }
 30:   }
 31:   if (!B->preallocated) {
 32:     /* Explicitly create 2 MATSEQAIJCUSP matrices. */
 33:     MatCreate(PETSC_COMM_SELF,&b->A);
 34:     MatSetSizes(b->A,B->rmap->n,B->cmap->n,B->rmap->n,B->cmap->n);
 35:     MatSetType(b->A,MATSEQAIJCUSP);
 36:     PetscLogObjectParent((PetscObject)B,(PetscObject)b->A);
 37:     MatCreate(PETSC_COMM_SELF,&b->B);
 38:     MatSetSizes(b->B,B->rmap->n,B->cmap->N,B->rmap->n,B->cmap->N);
 39:     MatSetType(b->B,MATSEQAIJCUSP);
 40:     PetscLogObjectParent((PetscObject)B,(PetscObject)b->B);
 41:   }
 42:   MatSeqAIJSetPreallocation(b->A,d_nz,d_nnz);
 43:   MatSeqAIJSetPreallocation(b->B,o_nz,o_nnz);
 44:   MatCUSPSetFormat(b->A,MAT_CUSP_MULT,cuspStruct->diagGPUMatFormat);
 45:   MatCUSPSetFormat(b->B,MAT_CUSP_MULT,cuspStruct->offdiagGPUMatFormat);
 46:   MatCUSPSetStream(b->A,cuspStruct->stream);
 47:   MatCUSPSetStream(b->B,cuspStruct->stream);
 48:   B->preallocated = PETSC_TRUE;
 49:   return(0);
 50: }

 54: PetscErrorCode  MatCreateVecs_MPIAIJCUSP(Mat mat,Vec *right,Vec *left)
 55: {
 57:   PetscInt rbs,cbs;

 60:   MatGetBlockSizes(mat,&rbs,&cbs);
 61:   if (right) {
 62:     VecCreate(PetscObjectComm((PetscObject)mat),right);
 63:     VecSetSizes(*right,mat->cmap->n,PETSC_DETERMINE);
 64:     VecSetBlockSize(*right,cbs);
 65:     VecSetType(*right,VECCUSP);
 66:     VecSetLayout(*right,mat->cmap);
 67:   }
 68:   if (left) {
 69:     VecCreate(PetscObjectComm((PetscObject)mat),left);
 70:     VecSetSizes(*left,mat->rmap->n,PETSC_DETERMINE);
 71:     VecSetBlockSize(*left,rbs);
 72:     VecSetType(*left,VECCUSP);
 73:     VecSetLayout(*left,mat->rmap);
 74:   }
 75:   return(0);
 76: }


 81: PetscErrorCode MatMult_MPIAIJCUSP(Mat A,Vec xx,Vec yy)
 82: {
 83:   /* This multiplication sequence is different sequence
 84:      than the CPU version. In particular, the diagonal block
 85:      multiplication kernel is launched in one stream. Then,
 86:      in a separate stream, the data transfers from DeviceToHost
 87:      (with MPI messaging in between), then HostToDevice are
 88:      launched. Once the data transfer stream is synchronized,
 89:      to ensure messaging is complete, the MatMultAdd kernel
 90:      is launched in the original (MatMult) stream to protect
 91:      against race conditions.

 93:      This sequence should only be called for GPU computation. */
 94:   Mat_MPIAIJ     *a = (Mat_MPIAIJ*)A->data;
 96:   PetscInt       nt;

 99:   VecGetLocalSize(xx,&nt);
100:   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);
101:   VecScatterInitializeForGPU(a->Mvctx,xx,SCATTER_FORWARD);
102:   (*a->A->ops->mult)(a->A,xx,yy);
103:   VecScatterBegin(a->Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);
104:   VecScatterEnd(a->Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);
105:   (*a->B->ops->multadd)(a->B,a->lvec,yy,yy);
106:   VecScatterFinalizeForGPU(a->Mvctx);
107:   return(0);
108: }

110: PetscErrorCode MatSetValuesBatch_MPIAIJCUSP(Mat J, PetscInt Ne, PetscInt Nl, PetscInt *elemRows, const PetscScalar *elemMats);

114: PetscErrorCode MatCUSPSetFormat_MPIAIJCUSP(Mat A,MatCUSPFormatOperation op,MatCUSPStorageFormat format)
115: {
116:   Mat_MPIAIJ     *a           = (Mat_MPIAIJ*)A->data;
117:   Mat_MPIAIJCUSP * cuspStruct = (Mat_MPIAIJCUSP*)a->spptr;

120:   switch (op) {
121:   case MAT_CUSP_MULT_DIAG:
122:     cuspStruct->diagGPUMatFormat = format;
123:     break;
124:   case MAT_CUSP_MULT_OFFDIAG:
125:     cuspStruct->offdiagGPUMatFormat = format;
126:     break;
127:   case MAT_CUSP_ALL:
128:     cuspStruct->diagGPUMatFormat    = format;
129:     cuspStruct->offdiagGPUMatFormat = format;
130:     break;
131:   default:
132:     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);
133:   }
134:   return(0);
135: }

139: PetscErrorCode MatSetFromOptions_MPIAIJCUSP(PetscOptions *PetscOptionsObject,Mat A)
140: {
141:   MatCUSPStorageFormat format;
142:   PetscErrorCode       ierr;
143:   PetscBool            flg;
144:   Mat_MPIAIJ           *a = (Mat_MPIAIJ*)A->data;
145:   Mat_MPIAIJCUSP       *cuspStruct = (Mat_MPIAIJCUSP*)a->spptr;

148:   PetscOptionsHead(PetscOptionsObject,"MPIAIJCUSP options");
149:   PetscObjectOptionsBegin((PetscObject)A);
150:   if (A->factortype==MAT_FACTOR_NONE) {
151:     PetscOptionsEnum("-mat_cusp_mult_diag_storage_format","sets storage format of the diagonal blocks of (mpi)aijcusp gpu matrices for SpMV",
152:                             "MatCUSPSetFormat",MatCUSPStorageFormats,(PetscEnum)cuspStruct->diagGPUMatFormat,(PetscEnum*)&format,&flg);
153:     if (flg) {
154:       MatCUSPSetFormat(A,MAT_CUSP_MULT_DIAG,format);
155:     }
156:     PetscOptionsEnum("-mat_cusp_mult_offdiag_storage_format","sets storage format of the off-diagonal blocks (mpi)aijcusp gpu matrices for SpMV",
157:                             "MatCUSPSetFormat",MatCUSPStorageFormats,(PetscEnum)cuspStruct->offdiagGPUMatFormat,(PetscEnum*)&format,&flg);
158:     if (flg) {
159:       MatCUSPSetFormat(A,MAT_CUSP_MULT_OFFDIAG,format);
160:     }
161:     PetscOptionsEnum("-mat_cusp_storage_format","sets storage format of the diagonal and off-diagonal blocks (mpi)aijcusp gpu matrices for SpMV",
162:                             "MatCUSPSetFormat",MatCUSPStorageFormats,(PetscEnum)cuspStruct->diagGPUMatFormat,(PetscEnum*)&format,&flg);
163:     if (flg) {
164:       MatCUSPSetFormat(A,MAT_CUSP_ALL,format);
165:     }
166:   }
167:   PetscOptionsEnd();
168:   return(0);
169: }

173: PetscErrorCode MatDestroy_MPIAIJCUSP(Mat A)
174: {
176:   Mat_MPIAIJ     *a           = (Mat_MPIAIJ*)A->data;
177:   Mat_MPIAIJCUSP *cuspStruct = (Mat_MPIAIJCUSP*)a->spptr;
178:   cudaError_t    err=cudaSuccess;

181:   try {
182:     err = cudaStreamDestroy(cuspStruct->stream);
183:     if (err!=cudaSuccess)
184:       SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Mat_MPIAIJCUSP error: %s", cudaGetErrorString(err));
185:     delete cuspStruct;
186:   } catch(char *ex) {
187:     SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Mat_MPIAIJCUSP error: %s", ex);
188:   }
189:   cuspStruct = 0;
190:   MatDestroy_MPIAIJ(A);
191:   return(0);
192: }

196: PETSC_EXTERN PetscErrorCode MatCreate_MPIAIJCUSP(Mat A)
197: {
199:   Mat_MPIAIJ     *a;
200:   Mat_MPIAIJCUSP * cuspStruct;
201:   cudaError_t    err=cudaSuccess;

204:   MatCreate_MPIAIJ(A);
205:   PetscObjectComposeFunction((PetscObject)A,"MatMPIAIJSetPreallocation_C",MatMPIAIJSetPreallocation_MPIAIJCUSP);
206:   A->ops->getvecs        = MatCreateVecs_MPIAIJCUSP;
207:   A->ops->setvaluesbatch = MatSetValuesBatch_MPIAIJCUSP;

209:   a          = (Mat_MPIAIJ*)A->data;
210:   a->spptr   = new Mat_MPIAIJCUSP;
211:   cuspStruct = (Mat_MPIAIJCUSP*)a->spptr;

213:   cuspStruct->diagGPUMatFormat    = MAT_CUSP_CSR;
214:   cuspStruct->offdiagGPUMatFormat = MAT_CUSP_CSR;
215:   err = cudaStreamCreate(&(cuspStruct->stream));
216:   if (err!=cudaSuccess)
217:     SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_LIB,"Mat_MPIAIJCUSP error: %s", cudaGetErrorString(err));

219:   A->ops->mult           = MatMult_MPIAIJCUSP;
220:   A->ops->setfromoptions = MatSetFromOptions_MPIAIJCUSP;
221:   A->ops->destroy        = MatDestroy_MPIAIJCUSP;

223:   PetscObjectComposeFunction((PetscObject)A,"MatCUSPSetFormat_C", MatCUSPSetFormat_MPIAIJCUSP);
224:   PetscObjectChangeTypeName((PetscObject)A,MATMPIAIJCUSP);
225:   return(0);
226: }


229: /*@
230:    MatCreateAIJCUSP - Creates a sparse matrix in AIJ (compressed row) format
231:    (the default parallel PETSc format).  This matrix will ultimately pushed down
232:    to NVidia GPUs and use the CUSP library for calculations. For good matrix
233:    assembly performance the user should preallocate the matrix storage by setting
234:    the parameter nz (or the array nnz).  By setting these parameters accurately,
235:    performance during matrix assembly can be increased by more than a factor of 50.


238:    Collective on MPI_Comm

240:    Input Parameters:
241: +  comm - MPI communicator, set to PETSC_COMM_SELF
242: .  m - number of rows
243: .  n - number of columns
244: .  nz - number of nonzeros per row (same for all rows)
245: -  nnz - array containing the number of nonzeros in the various rows
246:          (possibly different for each row) or NULL

248:    Output Parameter:
249: .  A - the matrix

251:    It is recommended that one use the MatCreate(), MatSetType() and/or MatSetFromOptions(),
252:    MatXXXXSetPreallocation() paradigm instead of this routine directly.
253:    [MatXXXXSetPreallocation() is, for example, MatSeqAIJSetPreallocation]

255:    Notes:
256:    If nnz is given then nz is ignored

258:    The AIJ format (also called the Yale sparse matrix format or
259:    compressed row storage), is fully compatible with standard Fortran 77
260:    storage.  That is, the stored row and column indices can begin at
261:    either one (as in Fortran) or zero.  See the users' manual for details.

263:    Specify the preallocated storage with either nz or nnz (not both).
264:    Set nz=PETSC_DEFAULT and nnz=NULL for PETSc to control dynamic memory
265:    allocation.  For large problems you MUST preallocate memory or you
266:    will get TERRIBLE performance, see the users' manual chapter on matrices.

268:    By default, this format uses inodes (identical nodes) when possible, to
269:    improve numerical efficiency of matrix-vector products and solves. We
270:    search for consecutive rows with the same nonzero structure, thereby
271:    reusing matrix information to achieve increased efficiency.

273:    Level: intermediate

275: .seealso: MatCreate(), MatCreateAIJ(), MatSetValues(), MatSeqAIJSetColumnIndices(), MatCreateSeqAIJWithArrays(), MatCreateAIJ(), MATMPIAIJCUSP, MATAIJCUSP
276: @*/
279: 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)
280: {
282:   PetscMPIInt    size;

285:   MatCreate(comm,A);
286:   MatSetSizes(*A,m,n,M,N);
287:   MPI_Comm_size(comm,&size);
288:   if (size > 1) {
289:     MatSetType(*A,MATMPIAIJCUSP);
290:     MatMPIAIJSetPreallocation(*A,d_nz,d_nnz,o_nz,o_nnz);
291:   } else {
292:     MatSetType(*A,MATSEQAIJCUSP);
293:     MatSeqAIJSetPreallocation(*A,d_nz,d_nnz);
294:   }
295:   return(0);
296: }

298: /*M
299:    MATAIJCUSP - MATMPIAIJCUSP= "aijcusp" = "mpiaijcusp" - A matrix type to be used for sparse matrices.

301:    A matrix type type whose data resides on Nvidia GPUs. These matrices can be CSR format.
302:    All matrix calculations are performed using the CUSP library. DIA and ELL
303:    formats are also available

305:    This matrix type is identical to MATSEQAIJCUSP when constructed with a single process communicator,
306:    and MATMPIAIJCUSP otherwise.  As a result, for single process communicators,
307:    MatSeqAIJSetPreallocation is supported, and similarly MatMPIAIJSetPreallocation is supported
308:    for communicators controlling multiple processes.  It is recommended that you call both of
309:    the above preallocation routines for simplicity.

311:    Options Database Keys:
312: +  -mat_type mpiaijcusp - sets the matrix type to "mpiaijcusp" during a call to MatSetFromOptions()
313: .  -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).
314: .  -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).
315: -  -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).

317:   Level: beginner

319:  .seealso: MatCreateAIJCUSP(), MATSEQAIJCUSP, MatCreateSeqAIJCUSP(), MatCUSPSetFormat(), MatCUSPStorageFormat, MatCUSPFormatOperation
320: M*/