Actual source code: mpiaijcusparse.cu
1: #define PETSC_SKIP_IMMINTRIN_H_CUDAWORKAROUND 1
3: #include <petscconf.h>
4: #include <../src/mat/impls/aij/mpi/mpiaij.h>
5: #include <../src/mat/impls/aij/seq/seqcusparse/cusparsematimpl.h>
6: #include <../src/mat/impls/aij/mpi/mpicusparse/mpicusparsematimpl.h>
7: #include <thrust/advance.h>
8: #include <thrust/partition.h>
9: #include <thrust/sort.h>
10: #include <thrust/unique.h>
11: #include <petscsf.h>
13: struct VecCUDAEquals {
14: template <typename Tuple>
15: __host__ __device__ void operator()(Tuple t)
16: {
17: thrust::get<1>(t) = thrust::get<0>(t);
18: }
19: };
21: static PetscErrorCode MatCOOStructDestroy_MPIAIJCUSPARSE(void **data)
22: {
23: MatCOOStruct_MPIAIJ *coo = (MatCOOStruct_MPIAIJ *)*data;
25: PetscFunctionBegin;
26: PetscCall(PetscSFDestroy(&coo->sf));
27: PetscCallCUDA(cudaFree(coo->Ajmap1));
28: PetscCallCUDA(cudaFree(coo->Aperm1));
29: PetscCallCUDA(cudaFree(coo->Bjmap1));
30: PetscCallCUDA(cudaFree(coo->Bperm1));
31: PetscCallCUDA(cudaFree(coo->Aimap2));
32: PetscCallCUDA(cudaFree(coo->Ajmap2));
33: PetscCallCUDA(cudaFree(coo->Aperm2));
34: PetscCallCUDA(cudaFree(coo->Bimap2));
35: PetscCallCUDA(cudaFree(coo->Bjmap2));
36: PetscCallCUDA(cudaFree(coo->Bperm2));
37: PetscCallCUDA(cudaFree(coo->Cperm1));
38: PetscCallCUDA(cudaFree(coo->sendbuf));
39: PetscCallCUDA(cudaFree(coo->recvbuf));
40: PetscCall(PetscFree(coo));
41: PetscFunctionReturn(PETSC_SUCCESS);
42: }
44: static PetscErrorCode MatSetPreallocationCOO_MPIAIJCUSPARSE(Mat mat, PetscCount coo_n, PetscInt coo_i[], PetscInt coo_j[])
45: {
46: Mat_MPIAIJ *mpiaij = (Mat_MPIAIJ *)mat->data;
47: PetscBool dev_ij = PETSC_FALSE;
48: PetscMemType mtype = PETSC_MEMTYPE_HOST;
49: PetscInt *i, *j;
50: PetscContainer container_h;
51: MatCOOStruct_MPIAIJ *coo_h, *coo_d;
53: PetscFunctionBegin;
54: PetscCall(PetscFree(mpiaij->garray));
55: PetscCall(VecDestroy(&mpiaij->lvec));
56: #if defined(PETSC_USE_CTABLE)
57: PetscCall(PetscHMapIDestroy(&mpiaij->colmap));
58: #else
59: PetscCall(PetscFree(mpiaij->colmap));
60: #endif
61: PetscCall(VecScatterDestroy(&mpiaij->Mvctx));
62: mat->assembled = PETSC_FALSE;
63: mat->was_assembled = PETSC_FALSE;
64: PetscCall(PetscGetMemType(coo_i, &mtype));
65: if (PetscMemTypeDevice(mtype)) {
66: dev_ij = PETSC_TRUE;
67: PetscCall(PetscMalloc2(coo_n, &i, coo_n, &j));
68: PetscCallCUDA(cudaMemcpy(i, coo_i, coo_n * sizeof(PetscInt), cudaMemcpyDeviceToHost));
69: PetscCallCUDA(cudaMemcpy(j, coo_j, coo_n * sizeof(PetscInt), cudaMemcpyDeviceToHost));
70: } else {
71: i = coo_i;
72: j = coo_j;
73: }
75: PetscCall(MatSetPreallocationCOO_MPIAIJ(mat, coo_n, i, j));
76: if (dev_ij) PetscCall(PetscFree2(i, j));
77: mat->offloadmask = PETSC_OFFLOAD_CPU;
78: // Create the GPU memory
79: PetscCall(MatSeqAIJCUSPARSECopyToGPU(mpiaij->A));
80: PetscCall(MatSeqAIJCUSPARSECopyToGPU(mpiaij->B));
82: // Copy the COO struct to device
83: PetscCall(PetscObjectQuery((PetscObject)mat, "__PETSc_MatCOOStruct_Host", (PetscObject *)&container_h));
84: PetscCall(PetscContainerGetPointer(container_h, (void **)&coo_h));
85: PetscCall(PetscMalloc1(1, &coo_d));
86: *coo_d = *coo_h; // do a shallow copy and then amend fields in coo_d
88: PetscCall(PetscObjectReference((PetscObject)coo_d->sf)); // Since we destroy the sf in both coo_h and coo_d
89: PetscCallCUDA(cudaMalloc((void **)&coo_d->Ajmap1, (coo_h->Annz + 1) * sizeof(PetscCount)));
90: PetscCallCUDA(cudaMalloc((void **)&coo_d->Aperm1, coo_h->Atot1 * sizeof(PetscCount)));
91: PetscCallCUDA(cudaMalloc((void **)&coo_d->Bjmap1, (coo_h->Bnnz + 1) * sizeof(PetscCount)));
92: PetscCallCUDA(cudaMalloc((void **)&coo_d->Bperm1, coo_h->Btot1 * sizeof(PetscCount)));
93: PetscCallCUDA(cudaMalloc((void **)&coo_d->Aimap2, coo_h->Annz2 * sizeof(PetscCount)));
94: PetscCallCUDA(cudaMalloc((void **)&coo_d->Ajmap2, (coo_h->Annz2 + 1) * sizeof(PetscCount)));
95: PetscCallCUDA(cudaMalloc((void **)&coo_d->Aperm2, coo_h->Atot2 * sizeof(PetscCount)));
96: PetscCallCUDA(cudaMalloc((void **)&coo_d->Bimap2, coo_h->Bnnz2 * sizeof(PetscCount)));
97: PetscCallCUDA(cudaMalloc((void **)&coo_d->Bjmap2, (coo_h->Bnnz2 + 1) * sizeof(PetscCount)));
98: PetscCallCUDA(cudaMalloc((void **)&coo_d->Bperm2, coo_h->Btot2 * sizeof(PetscCount)));
99: PetscCallCUDA(cudaMalloc((void **)&coo_d->Cperm1, coo_h->sendlen * sizeof(PetscCount)));
100: PetscCallCUDA(cudaMalloc((void **)&coo_d->sendbuf, coo_h->sendlen * sizeof(PetscScalar)));
101: PetscCallCUDA(cudaMalloc((void **)&coo_d->recvbuf, coo_h->recvlen * sizeof(PetscScalar)));
103: PetscCallCUDA(cudaMemcpy(coo_d->Ajmap1, coo_h->Ajmap1, (coo_h->Annz + 1) * sizeof(PetscCount), cudaMemcpyHostToDevice));
104: PetscCallCUDA(cudaMemcpy(coo_d->Aperm1, coo_h->Aperm1, coo_h->Atot1 * sizeof(PetscCount), cudaMemcpyHostToDevice));
105: PetscCallCUDA(cudaMemcpy(coo_d->Bjmap1, coo_h->Bjmap1, (coo_h->Bnnz + 1) * sizeof(PetscCount), cudaMemcpyHostToDevice));
106: PetscCallCUDA(cudaMemcpy(coo_d->Bperm1, coo_h->Bperm1, coo_h->Btot1 * sizeof(PetscCount), cudaMemcpyHostToDevice));
107: PetscCallCUDA(cudaMemcpy(coo_d->Aimap2, coo_h->Aimap2, coo_h->Annz2 * sizeof(PetscCount), cudaMemcpyHostToDevice));
108: PetscCallCUDA(cudaMemcpy(coo_d->Ajmap2, coo_h->Ajmap2, (coo_h->Annz2 + 1) * sizeof(PetscCount), cudaMemcpyHostToDevice));
109: PetscCallCUDA(cudaMemcpy(coo_d->Aperm2, coo_h->Aperm2, coo_h->Atot2 * sizeof(PetscCount), cudaMemcpyHostToDevice));
110: PetscCallCUDA(cudaMemcpy(coo_d->Bimap2, coo_h->Bimap2, coo_h->Bnnz2 * sizeof(PetscCount), cudaMemcpyHostToDevice));
111: PetscCallCUDA(cudaMemcpy(coo_d->Bjmap2, coo_h->Bjmap2, (coo_h->Bnnz2 + 1) * sizeof(PetscCount), cudaMemcpyHostToDevice));
112: PetscCallCUDA(cudaMemcpy(coo_d->Bperm2, coo_h->Bperm2, coo_h->Btot2 * sizeof(PetscCount), cudaMemcpyHostToDevice));
113: PetscCallCUDA(cudaMemcpy(coo_d->Cperm1, coo_h->Cperm1, coo_h->sendlen * sizeof(PetscCount), cudaMemcpyHostToDevice));
115: // Put the COO struct in a container and then attach that to the matrix
116: PetscCall(PetscObjectContainerCompose((PetscObject)mat, "__PETSc_MatCOOStruct_Device", coo_d, MatCOOStructDestroy_MPIAIJCUSPARSE));
117: PetscFunctionReturn(PETSC_SUCCESS);
118: }
120: __global__ static void MatPackCOOValues(const PetscScalar kv[], PetscCount nnz, const PetscCount perm[], PetscScalar buf[])
121: {
122: PetscCount i = blockIdx.x * blockDim.x + threadIdx.x;
123: const PetscCount grid_size = gridDim.x * blockDim.x;
124: for (; i < nnz; i += grid_size) buf[i] = kv[perm[i]];
125: }
127: __global__ static void MatAddLocalCOOValues(const PetscScalar kv[], InsertMode imode, PetscCount Annz, const PetscCount Ajmap1[], const PetscCount Aperm1[], PetscScalar Aa[], PetscCount Bnnz, const PetscCount Bjmap1[], const PetscCount Bperm1[], PetscScalar Ba[])
128: {
129: PetscCount i = blockIdx.x * blockDim.x + threadIdx.x;
130: const PetscCount grid_size = gridDim.x * blockDim.x;
131: for (; i < Annz + Bnnz; i += grid_size) {
132: PetscScalar sum = 0.0;
133: if (i < Annz) {
134: for (PetscCount k = Ajmap1[i]; k < Ajmap1[i + 1]; k++) sum += kv[Aperm1[k]];
135: Aa[i] = (imode == INSERT_VALUES ? 0.0 : Aa[i]) + sum;
136: } else {
137: i -= Annz;
138: for (PetscCount k = Bjmap1[i]; k < Bjmap1[i + 1]; k++) sum += kv[Bperm1[k]];
139: Ba[i] = (imode == INSERT_VALUES ? 0.0 : Ba[i]) + sum;
140: }
141: }
142: }
144: __global__ static void MatAddRemoteCOOValues(const PetscScalar kv[], PetscCount Annz2, const PetscCount Aimap2[], const PetscCount Ajmap2[], const PetscCount Aperm2[], PetscScalar Aa[], PetscCount Bnnz2, const PetscCount Bimap2[], const PetscCount Bjmap2[], const PetscCount Bperm2[], PetscScalar Ba[])
145: {
146: PetscCount i = blockIdx.x * blockDim.x + threadIdx.x;
147: const PetscCount grid_size = gridDim.x * blockDim.x;
148: for (; i < Annz2 + Bnnz2; i += grid_size) {
149: if (i < Annz2) {
150: for (PetscCount k = Ajmap2[i]; k < Ajmap2[i + 1]; k++) Aa[Aimap2[i]] += kv[Aperm2[k]];
151: } else {
152: i -= Annz2;
153: for (PetscCount k = Bjmap2[i]; k < Bjmap2[i + 1]; k++) Ba[Bimap2[i]] += kv[Bperm2[k]];
154: }
155: }
156: }
158: static PetscErrorCode MatSetValuesCOO_MPIAIJCUSPARSE(Mat mat, const PetscScalar v[], InsertMode imode)
159: {
160: Mat_MPIAIJ *mpiaij = static_cast<Mat_MPIAIJ *>(mat->data);
161: Mat A = mpiaij->A, B = mpiaij->B;
162: PetscScalar *Aa, *Ba;
163: const PetscScalar *v1 = v;
164: PetscMemType memtype;
165: PetscContainer container;
166: MatCOOStruct_MPIAIJ *coo;
168: PetscFunctionBegin;
169: PetscCall(PetscObjectQuery((PetscObject)mat, "__PETSc_MatCOOStruct_Device", (PetscObject *)&container));
170: PetscCheck(container, PetscObjectComm((PetscObject)mat), PETSC_ERR_PLIB, "Not found MatCOOStruct on this matrix");
171: PetscCall(PetscContainerGetPointer(container, (void **)&coo));
173: const auto &Annz = coo->Annz;
174: const auto &Annz2 = coo->Annz2;
175: const auto &Bnnz = coo->Bnnz;
176: const auto &Bnnz2 = coo->Bnnz2;
177: const auto &vsend = coo->sendbuf;
178: const auto &v2 = coo->recvbuf;
179: const auto &Ajmap1 = coo->Ajmap1;
180: const auto &Ajmap2 = coo->Ajmap2;
181: const auto &Aimap2 = coo->Aimap2;
182: const auto &Bjmap1 = coo->Bjmap1;
183: const auto &Bjmap2 = coo->Bjmap2;
184: const auto &Bimap2 = coo->Bimap2;
185: const auto &Aperm1 = coo->Aperm1;
186: const auto &Aperm2 = coo->Aperm2;
187: const auto &Bperm1 = coo->Bperm1;
188: const auto &Bperm2 = coo->Bperm2;
189: const auto &Cperm1 = coo->Cperm1;
191: PetscCall(PetscGetMemType(v, &memtype));
192: if (PetscMemTypeHost(memtype)) { /* If user gave v[] in host, we need to copy it to device */
193: PetscCallCUDA(cudaMalloc((void **)&v1, coo->n * sizeof(PetscScalar)));
194: PetscCallCUDA(cudaMemcpy((void *)v1, v, coo->n * sizeof(PetscScalar), cudaMemcpyHostToDevice));
195: }
197: if (imode == INSERT_VALUES) {
198: PetscCall(MatSeqAIJCUSPARSEGetArrayWrite(A, &Aa)); /* write matrix values */
199: PetscCall(MatSeqAIJCUSPARSEGetArrayWrite(B, &Ba));
200: } else {
201: PetscCall(MatSeqAIJCUSPARSEGetArray(A, &Aa)); /* read & write matrix values */
202: PetscCall(MatSeqAIJCUSPARSEGetArray(B, &Ba));
203: }
205: PetscCall(PetscLogGpuTimeBegin());
206: /* Pack entries to be sent to remote */
207: if (coo->sendlen) {
208: MatPackCOOValues<<<(coo->sendlen + 255) / 256, 256>>>(v1, coo->sendlen, Cperm1, vsend);
209: PetscCallCUDA(cudaPeekAtLastError());
210: }
212: /* Send remote entries to their owner and overlap the communication with local computation */
213: PetscCall(PetscSFReduceWithMemTypeBegin(coo->sf, MPIU_SCALAR, PETSC_MEMTYPE_CUDA, vsend, PETSC_MEMTYPE_CUDA, v2, MPI_REPLACE));
214: /* Add local entries to A and B */
215: if (Annz + Bnnz > 0) {
216: MatAddLocalCOOValues<<<(int)((Annz + Bnnz + 255) / 256), 256>>>(v1, imode, Annz, Ajmap1, Aperm1, Aa, Bnnz, Bjmap1, Bperm1, Ba);
217: PetscCallCUDA(cudaPeekAtLastError());
218: }
219: PetscCall(PetscSFReduceEnd(coo->sf, MPIU_SCALAR, vsend, v2, MPI_REPLACE));
221: /* Add received remote entries to A and B */
222: if (Annz2 + Bnnz2 > 0) {
223: MatAddRemoteCOOValues<<<(int)((Annz2 + Bnnz2 + 255) / 256), 256>>>(v2, Annz2, Aimap2, Ajmap2, Aperm2, Aa, Bnnz2, Bimap2, Bjmap2, Bperm2, Ba);
224: PetscCallCUDA(cudaPeekAtLastError());
225: }
226: PetscCall(PetscLogGpuTimeEnd());
228: if (imode == INSERT_VALUES) {
229: PetscCall(MatSeqAIJCUSPARSERestoreArrayWrite(A, &Aa));
230: PetscCall(MatSeqAIJCUSPARSERestoreArrayWrite(B, &Ba));
231: } else {
232: PetscCall(MatSeqAIJCUSPARSERestoreArray(A, &Aa));
233: PetscCall(MatSeqAIJCUSPARSERestoreArray(B, &Ba));
234: }
235: if (PetscMemTypeHost(memtype)) PetscCallCUDA(cudaFree((void *)v1));
236: mat->offloadmask = PETSC_OFFLOAD_GPU;
237: PetscFunctionReturn(PETSC_SUCCESS);
238: }
240: static PetscErrorCode MatMPIAIJGetLocalMatMerge_MPIAIJCUSPARSE(Mat A, MatReuse scall, IS *glob, Mat *A_loc)
241: {
242: Mat Ad, Ao;
243: const PetscInt *cmap;
245: PetscFunctionBegin;
246: PetscCall(MatMPIAIJGetSeqAIJ(A, &Ad, &Ao, &cmap));
247: PetscCall(MatSeqAIJCUSPARSEMergeMats(Ad, Ao, scall, A_loc));
248: if (glob) {
249: PetscInt cst, i, dn, on, *gidx;
251: PetscCall(MatGetLocalSize(Ad, NULL, &dn));
252: PetscCall(MatGetLocalSize(Ao, NULL, &on));
253: PetscCall(MatGetOwnershipRangeColumn(A, &cst, NULL));
254: PetscCall(PetscMalloc1(dn + on, &gidx));
255: for (i = 0; i < dn; i++) gidx[i] = cst + i;
256: for (i = 0; i < on; i++) gidx[i + dn] = cmap[i];
257: PetscCall(ISCreateGeneral(PetscObjectComm((PetscObject)Ad), dn + on, gidx, PETSC_OWN_POINTER, glob));
258: }
259: PetscFunctionReturn(PETSC_SUCCESS);
260: }
262: static PetscErrorCode MatMPIAIJSetPreallocation_MPIAIJCUSPARSE(Mat B, PetscInt d_nz, const PetscInt d_nnz[], PetscInt o_nz, const PetscInt o_nnz[])
263: {
264: Mat_MPIAIJ *b = (Mat_MPIAIJ *)B->data;
265: Mat_MPIAIJCUSPARSE *cusparseStruct = (Mat_MPIAIJCUSPARSE *)b->spptr;
266: PetscInt i;
268: PetscFunctionBegin;
269: if (B->hash_active) {
270: B->ops[0] = b->cops;
271: B->hash_active = PETSC_FALSE;
272: }
273: PetscCall(PetscLayoutSetUp(B->rmap));
274: PetscCall(PetscLayoutSetUp(B->cmap));
275: if (PetscDefined(USE_DEBUG) && d_nnz) {
276: for (i = 0; i < B->rmap->n; i++) PetscCheck(d_nnz[i] >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "d_nnz cannot be less than 0: local row %" PetscInt_FMT " value %" PetscInt_FMT, i, d_nnz[i]);
277: }
278: if (PetscDefined(USE_DEBUG) && o_nnz) {
279: for (i = 0; i < B->rmap->n; i++) PetscCheck(o_nnz[i] >= 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "o_nnz cannot be less than 0: local row %" PetscInt_FMT " value %" PetscInt_FMT, i, o_nnz[i]);
280: }
281: #if defined(PETSC_USE_CTABLE)
282: PetscCall(PetscHMapIDestroy(&b->colmap));
283: #else
284: PetscCall(PetscFree(b->colmap));
285: #endif
286: PetscCall(PetscFree(b->garray));
287: PetscCall(VecDestroy(&b->lvec));
288: PetscCall(VecScatterDestroy(&b->Mvctx));
289: /* Because the B will have been resized we simply destroy it and create a new one each time */
290: PetscCall(MatDestroy(&b->B));
291: if (!b->A) {
292: PetscCall(MatCreate(PETSC_COMM_SELF, &b->A));
293: PetscCall(MatSetSizes(b->A, B->rmap->n, B->cmap->n, B->rmap->n, B->cmap->n));
294: }
295: if (!b->B) {
296: PetscMPIInt size;
297: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)B), &size));
298: PetscCall(MatCreate(PETSC_COMM_SELF, &b->B));
299: PetscCall(MatSetSizes(b->B, B->rmap->n, size > 1 ? B->cmap->N : 0, B->rmap->n, size > 1 ? B->cmap->N : 0));
300: }
301: PetscCall(MatSetType(b->A, MATSEQAIJCUSPARSE));
302: PetscCall(MatSetType(b->B, MATSEQAIJCUSPARSE));
303: PetscCall(MatBindToCPU(b->A, B->boundtocpu));
304: PetscCall(MatBindToCPU(b->B, B->boundtocpu));
305: PetscCall(MatSeqAIJSetPreallocation(b->A, d_nz, d_nnz));
306: PetscCall(MatSeqAIJSetPreallocation(b->B, o_nz, o_nnz));
307: PetscCall(MatCUSPARSESetFormat(b->A, MAT_CUSPARSE_MULT, cusparseStruct->diagGPUMatFormat));
308: PetscCall(MatCUSPARSESetFormat(b->B, MAT_CUSPARSE_MULT, cusparseStruct->offdiagGPUMatFormat));
309: B->preallocated = PETSC_TRUE;
310: B->was_assembled = PETSC_FALSE;
311: B->assembled = PETSC_FALSE;
312: PetscFunctionReturn(PETSC_SUCCESS);
313: }
315: static PetscErrorCode MatMult_MPIAIJCUSPARSE(Mat A, Vec xx, Vec yy)
316: {
317: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
319: PetscFunctionBegin;
320: PetscCall(VecScatterBegin(a->Mvctx, xx, a->lvec, INSERT_VALUES, SCATTER_FORWARD));
321: PetscCall((*a->A->ops->mult)(a->A, xx, yy));
322: PetscCall(VecScatterEnd(a->Mvctx, xx, a->lvec, INSERT_VALUES, SCATTER_FORWARD));
323: PetscCall((*a->B->ops->multadd)(a->B, a->lvec, yy, yy));
324: PetscFunctionReturn(PETSC_SUCCESS);
325: }
327: static PetscErrorCode MatZeroEntries_MPIAIJCUSPARSE(Mat A)
328: {
329: Mat_MPIAIJ *l = (Mat_MPIAIJ *)A->data;
331: PetscFunctionBegin;
332: PetscCall(MatZeroEntries(l->A));
333: PetscCall(MatZeroEntries(l->B));
334: PetscFunctionReturn(PETSC_SUCCESS);
335: }
337: static PetscErrorCode MatMultAdd_MPIAIJCUSPARSE(Mat A, Vec xx, Vec yy, Vec zz)
338: {
339: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
341: PetscFunctionBegin;
342: PetscCall(VecScatterBegin(a->Mvctx, xx, a->lvec, INSERT_VALUES, SCATTER_FORWARD));
343: PetscCall((*a->A->ops->multadd)(a->A, xx, yy, zz));
344: PetscCall(VecScatterEnd(a->Mvctx, xx, a->lvec, INSERT_VALUES, SCATTER_FORWARD));
345: PetscCall((*a->B->ops->multadd)(a->B, a->lvec, zz, zz));
346: PetscFunctionReturn(PETSC_SUCCESS);
347: }
349: static PetscErrorCode MatMultTranspose_MPIAIJCUSPARSE(Mat A, Vec xx, Vec yy)
350: {
351: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
353: PetscFunctionBegin;
354: PetscCall((*a->B->ops->multtranspose)(a->B, xx, a->lvec));
355: PetscCall((*a->A->ops->multtranspose)(a->A, xx, yy));
356: PetscCall(VecScatterBegin(a->Mvctx, a->lvec, yy, ADD_VALUES, SCATTER_REVERSE));
357: PetscCall(VecScatterEnd(a->Mvctx, a->lvec, yy, ADD_VALUES, SCATTER_REVERSE));
358: PetscFunctionReturn(PETSC_SUCCESS);
359: }
361: static PetscErrorCode MatCUSPARSESetFormat_MPIAIJCUSPARSE(Mat A, MatCUSPARSEFormatOperation op, MatCUSPARSEStorageFormat format)
362: {
363: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
364: Mat_MPIAIJCUSPARSE *cusparseStruct = (Mat_MPIAIJCUSPARSE *)a->spptr;
366: PetscFunctionBegin;
367: switch (op) {
368: case MAT_CUSPARSE_MULT_DIAG:
369: cusparseStruct->diagGPUMatFormat = format;
370: break;
371: case MAT_CUSPARSE_MULT_OFFDIAG:
372: cusparseStruct->offdiagGPUMatFormat = format;
373: break;
374: case MAT_CUSPARSE_ALL:
375: cusparseStruct->diagGPUMatFormat = format;
376: cusparseStruct->offdiagGPUMatFormat = format;
377: break;
378: default:
379: SETERRQ(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);
380: }
381: PetscFunctionReturn(PETSC_SUCCESS);
382: }
384: static PetscErrorCode MatSetFromOptions_MPIAIJCUSPARSE(Mat A, PetscOptionItems *PetscOptionsObject)
385: {
386: MatCUSPARSEStorageFormat format;
387: PetscBool flg;
388: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
389: Mat_MPIAIJCUSPARSE *cusparseStruct = (Mat_MPIAIJCUSPARSE *)a->spptr;
391: PetscFunctionBegin;
392: PetscOptionsHeadBegin(PetscOptionsObject, "MPIAIJCUSPARSE options");
393: if (A->factortype == MAT_FACTOR_NONE) {
394: PetscCall(PetscOptionsEnum("-mat_cusparse_mult_diag_storage_format", "sets storage format of the diagonal blocks of (mpi)aijcusparse gpu matrices for SpMV", "MatCUSPARSESetFormat", MatCUSPARSEStorageFormats, (PetscEnum)cusparseStruct->diagGPUMatFormat, (PetscEnum *)&format, &flg));
395: if (flg) PetscCall(MatCUSPARSESetFormat(A, MAT_CUSPARSE_MULT_DIAG, format));
396: PetscCall(PetscOptionsEnum("-mat_cusparse_mult_offdiag_storage_format", "sets storage format of the off-diagonal blocks (mpi)aijcusparse gpu matrices for SpMV", "MatCUSPARSESetFormat", MatCUSPARSEStorageFormats, (PetscEnum)cusparseStruct->offdiagGPUMatFormat, (PetscEnum *)&format, &flg));
397: if (flg) PetscCall(MatCUSPARSESetFormat(A, MAT_CUSPARSE_MULT_OFFDIAG, format));
398: PetscCall(PetscOptionsEnum("-mat_cusparse_storage_format", "sets storage format of the diagonal and off-diagonal blocks (mpi)aijcusparse gpu matrices for SpMV", "MatCUSPARSESetFormat", MatCUSPARSEStorageFormats, (PetscEnum)cusparseStruct->diagGPUMatFormat, (PetscEnum *)&format, &flg));
399: if (flg) PetscCall(MatCUSPARSESetFormat(A, MAT_CUSPARSE_ALL, format));
400: }
401: PetscOptionsHeadEnd();
402: PetscFunctionReturn(PETSC_SUCCESS);
403: }
405: static PetscErrorCode MatAssemblyEnd_MPIAIJCUSPARSE(Mat A, MatAssemblyType mode)
406: {
407: Mat_MPIAIJ *mpiaij = (Mat_MPIAIJ *)A->data;
409: PetscFunctionBegin;
410: PetscCall(MatAssemblyEnd_MPIAIJ(A, mode));
411: if (mpiaij->lvec) PetscCall(VecSetType(mpiaij->lvec, VECSEQCUDA));
412: PetscFunctionReturn(PETSC_SUCCESS);
413: }
415: static PetscErrorCode MatDestroy_MPIAIJCUSPARSE(Mat A)
416: {
417: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)A->data;
418: Mat_MPIAIJCUSPARSE *cusparseStruct = (Mat_MPIAIJCUSPARSE *)aij->spptr;
420: PetscFunctionBegin;
421: PetscCheck(cusparseStruct, PETSC_COMM_SELF, PETSC_ERR_COR, "Missing spptr");
422: PetscCallCXX(delete cusparseStruct);
423: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatMPIAIJSetPreallocation_C", NULL));
424: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatMPIAIJGetLocalMatMerge_C", NULL));
425: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatSetPreallocationCOO_C", NULL));
426: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatSetValuesCOO_C", NULL));
427: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatCUSPARSESetFormat_C", NULL));
428: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_mpiaijcusparse_hypre_C", NULL));
429: PetscCall(MatDestroy_MPIAIJ(A));
430: PetscFunctionReturn(PETSC_SUCCESS);
431: }
433: /* defines MatSetValues_MPICUSPARSE_Hash() */
434: #define TYPE AIJ
435: #define TYPE_AIJ
436: #define SUB_TYPE_CUSPARSE
437: #include "../src/mat/impls/aij/mpi/mpihashmat.h"
438: #undef TYPE
439: #undef TYPE_AIJ
440: #undef SUB_TYPE_CUSPARSE
442: static PetscErrorCode MatSetUp_MPI_HASH_CUSPARSE(Mat A)
443: {
444: Mat_MPIAIJ *b = (Mat_MPIAIJ *)A->data;
445: Mat_MPIAIJCUSPARSE *cusparseStruct = (Mat_MPIAIJCUSPARSE *)b->spptr;
447: PetscFunctionBegin;
448: PetscCall(MatSetUp_MPI_Hash(A));
449: PetscCall(MatCUSPARSESetFormat(b->A, MAT_CUSPARSE_MULT, cusparseStruct->diagGPUMatFormat));
450: PetscCall(MatCUSPARSESetFormat(b->B, MAT_CUSPARSE_MULT, cusparseStruct->offdiagGPUMatFormat));
451: A->preallocated = PETSC_TRUE;
452: PetscFunctionReturn(PETSC_SUCCESS);
453: }
455: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJCUSPARSE(Mat B, MatType, MatReuse reuse, Mat *newmat)
456: {
457: Mat_MPIAIJ *a;
458: Mat A;
460: PetscFunctionBegin;
461: PetscCall(PetscDeviceInitialize(PETSC_DEVICE_CUDA));
462: if (reuse == MAT_INITIAL_MATRIX) PetscCall(MatDuplicate(B, MAT_COPY_VALUES, newmat));
463: else if (reuse == MAT_REUSE_MATRIX) PetscCall(MatCopy(B, *newmat, SAME_NONZERO_PATTERN));
464: A = *newmat;
465: A->boundtocpu = PETSC_FALSE;
466: PetscCall(PetscFree(A->defaultvectype));
467: PetscCall(PetscStrallocpy(VECCUDA, &A->defaultvectype));
469: a = (Mat_MPIAIJ *)A->data;
470: if (a->A) PetscCall(MatSetType(a->A, MATSEQAIJCUSPARSE));
471: if (a->B) PetscCall(MatSetType(a->B, MATSEQAIJCUSPARSE));
472: if (a->lvec) PetscCall(VecSetType(a->lvec, VECSEQCUDA));
474: if (reuse != MAT_REUSE_MATRIX && !a->spptr) PetscCallCXX(a->spptr = new Mat_MPIAIJCUSPARSE);
476: A->ops->assemblyend = MatAssemblyEnd_MPIAIJCUSPARSE;
477: A->ops->mult = MatMult_MPIAIJCUSPARSE;
478: A->ops->multadd = MatMultAdd_MPIAIJCUSPARSE;
479: A->ops->multtranspose = MatMultTranspose_MPIAIJCUSPARSE;
480: A->ops->setfromoptions = MatSetFromOptions_MPIAIJCUSPARSE;
481: A->ops->destroy = MatDestroy_MPIAIJCUSPARSE;
482: A->ops->zeroentries = MatZeroEntries_MPIAIJCUSPARSE;
483: A->ops->productsetfromoptions = MatProductSetFromOptions_MPIAIJBACKEND;
484: A->ops->setup = MatSetUp_MPI_HASH_CUSPARSE;
486: PetscCall(PetscObjectChangeTypeName((PetscObject)A, MATMPIAIJCUSPARSE));
487: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatMPIAIJGetLocalMatMerge_C", MatMPIAIJGetLocalMatMerge_MPIAIJCUSPARSE));
488: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatMPIAIJSetPreallocation_C", MatMPIAIJSetPreallocation_MPIAIJCUSPARSE));
489: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatCUSPARSESetFormat_C", MatCUSPARSESetFormat_MPIAIJCUSPARSE));
490: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatSetPreallocationCOO_C", MatSetPreallocationCOO_MPIAIJCUSPARSE));
491: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatSetValuesCOO_C", MatSetValuesCOO_MPIAIJCUSPARSE));
492: #if defined(PETSC_HAVE_HYPRE)
493: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_mpiaijcusparse_hypre_C", MatConvert_AIJ_HYPRE));
494: #endif
495: PetscFunctionReturn(PETSC_SUCCESS);
496: }
498: PETSC_EXTERN PetscErrorCode MatCreate_MPIAIJCUSPARSE(Mat A)
499: {
500: PetscFunctionBegin;
501: PetscCall(PetscDeviceInitialize(PETSC_DEVICE_CUDA));
502: PetscCall(MatCreate_MPIAIJ(A));
503: PetscCall(MatConvert_MPIAIJ_MPIAIJCUSPARSE(A, MATMPIAIJCUSPARSE, MAT_INPLACE_MATRIX, &A));
504: PetscFunctionReturn(PETSC_SUCCESS);
505: }
507: /*@
508: MatCreateAIJCUSPARSE - Creates a sparse matrix in `MATAIJCUSPARSE` (compressed row) format
509: (the default parallel PETSc format). This matrix will ultimately pushed down
510: to NVIDIA GPUs and use the CuSPARSE library for calculations.
512: Collective
514: Input Parameters:
515: + comm - MPI communicator, set to `PETSC_COMM_SELF`
516: . m - number of local rows (or `PETSC_DECIDE` to have calculated if `M` is given)
517: This value should be the same as the local size used in creating the
518: y vector for the matrix-vector product y = Ax.
519: . n - This value should be the same as the local size used in creating the
520: x vector for the matrix-vector product y = Ax. (or PETSC_DECIDE to have
521: calculated if `N` is given) For square matrices `n` is almost always `m`.
522: . M - number of global rows (or `PETSC_DETERMINE` to have calculated if `m` is given)
523: . N - number of global columns (or `PETSC_DETERMINE` to have calculated if `n` is given)
524: . d_nz - number of nonzeros per row in DIAGONAL portion of local submatrix
525: (same value is used for all local rows)
526: . d_nnz - array containing the number of nonzeros in the various rows of the
527: DIAGONAL portion of the local submatrix (possibly different for each row)
528: or `NULL`, if `d_nz` is used to specify the nonzero structure.
529: The size of this array is equal to the number of local rows, i.e `m`.
530: For matrices you plan to factor you must leave room for the diagonal entry and
531: put in the entry even if it is zero.
532: . o_nz - number of nonzeros per row in the OFF-DIAGONAL portion of local
533: submatrix (same value is used for all local rows).
534: - o_nnz - array containing the number of nonzeros in the various rows of the
535: OFF-DIAGONAL portion of the local submatrix (possibly different for
536: each row) or `NULL`, if `o_nz` is used to specify the nonzero
537: structure. The size of this array is equal to the number
538: of local rows, i.e `m`.
540: Output Parameter:
541: . A - the matrix
543: Level: intermediate
545: Notes:
546: It is recommended that one use the `MatCreate()`, `MatSetType()` and/or `MatSetFromOptions()`,
547: MatXXXXSetPreallocation() paradigm instead of this routine directly.
548: [MatXXXXSetPreallocation() is, for example, `MatSeqAIJSetPreallocation()`]
550: The AIJ format, also called the
551: compressed row storage), is fully compatible with standard Fortran
552: storage. That is, the stored row and column indices can begin at
553: either one (as in Fortran) or zero.
555: .seealso: [](ch_matrices), `Mat`, `MATAIJCUSPARSE`, `MatCreate()`, `MatCreateAIJ()`, `MatSetValues()`, `MatSeqAIJSetColumnIndices()`, `MatCreateSeqAIJWithArrays()`, `MATMPIAIJCUSPARSE`
556: @*/
557: 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)
558: {
559: PetscMPIInt size;
561: PetscFunctionBegin;
562: PetscCall(MatCreate(comm, A));
563: PetscCall(MatSetSizes(*A, m, n, M, N));
564: PetscCallMPI(MPI_Comm_size(comm, &size));
565: if (size > 1) {
566: PetscCall(MatSetType(*A, MATMPIAIJCUSPARSE));
567: PetscCall(MatMPIAIJSetPreallocation(*A, d_nz, d_nnz, o_nz, o_nnz));
568: } else {
569: PetscCall(MatSetType(*A, MATSEQAIJCUSPARSE));
570: PetscCall(MatSeqAIJSetPreallocation(*A, d_nz, d_nnz));
571: }
572: PetscFunctionReturn(PETSC_SUCCESS);
573: }
575: /*MC
576: MATAIJCUSPARSE - A matrix type to be used for sparse matrices; it is as same as `MATMPIAIJCUSPARSE`.
578: A matrix type whose data resides on NVIDIA GPUs. These matrices can be in either
579: CSR, ELL, or Hybrid format. The ELL and HYB formats require CUDA 4.2 or later.
580: All matrix calculations are performed on NVIDIA GPUs using the CuSPARSE library.
582: This matrix type is identical to `MATSEQAIJCUSPARSE` when constructed with a single process communicator,
583: and `MATMPIAIJCUSPARSE` otherwise. As a result, for single process communicators,
584: `MatSeqAIJSetPreallocation()` is supported, and similarly `MatMPIAIJSetPreallocation()` is supported
585: for communicators controlling multiple processes. It is recommended that you call both of
586: the above preallocation routines for simplicity.
588: Options Database Keys:
589: + -mat_type mpiaijcusparse - sets the matrix type to `MATMPIAIJCUSPARSE`
590: . -mat_cusparse_storage_format csr - sets the storage format of diagonal and off-diagonal matrices. Other options include ell (ellpack) or hyb (hybrid).
591: . -mat_cusparse_mult_diag_storage_format csr - sets the storage format of diagonal matrix. Other options include ell (ellpack) or hyb (hybrid).
592: - -mat_cusparse_mult_offdiag_storage_format csr - sets the storage format of off-diagonal matrix. Other options include ell (ellpack) or hyb (hybrid).
594: Level: beginner
596: .seealso: [](ch_matrices), `Mat`, `MatCreateAIJCUSPARSE()`, `MATSEQAIJCUSPARSE`, `MATMPIAIJCUSPARSE`, `MatCreateSeqAIJCUSPARSE()`, `MatCUSPARSESetFormat()`, `MatCUSPARSEStorageFormat`, `MatCUSPARSEFormatOperation`
597: M*/
599: /*MC
600: MATMPIAIJCUSPARSE - A matrix type to be used for sparse matrices; it is as same as `MATAIJCUSPARSE`.
602: Level: beginner
604: .seealso: [](ch_matrices), `Mat`, `MATAIJCUSPARSE`, `MATSEQAIJCUSPARSE`
605: M*/