Actual source code: mpiaijhipsparse.hip.cpp
1: /* Portions of this code are under:
2: Copyright (c) 2022 Advanced Micro Devices, Inc. All rights reserved.
3: */
4: #include <../src/mat/impls/aij/mpi/mpiaij.h>
5: #include <../src/mat/impls/aij/seq/seqhipsparse/hipsparsematimpl.h>
6: #include <../src/mat/impls/aij/mpi/mpihipsparse/mpihipsparsematimpl.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 VecHIPEquals {
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 MatResetPreallocationCOO_MPIAIJHIPSPARSE(Mat mat)
22: {
23: auto *aij = static_cast<Mat_MPIAIJ *>(mat->data);
24: auto *hipsparseStruct = static_cast<Mat_MPIAIJHIPSPARSE *>(aij->spptr);
26: PetscFunctionBegin;
27: if (!hipsparseStruct) PetscFunctionReturn(PETSC_SUCCESS);
28: if (hipsparseStruct->use_extended_coo) {
29: PetscCallHIP(hipFree(hipsparseStruct->Ajmap1_d));
30: PetscCallHIP(hipFree(hipsparseStruct->Aperm1_d));
31: PetscCallHIP(hipFree(hipsparseStruct->Bjmap1_d));
32: PetscCallHIP(hipFree(hipsparseStruct->Bperm1_d));
33: PetscCallHIP(hipFree(hipsparseStruct->Aimap2_d));
34: PetscCallHIP(hipFree(hipsparseStruct->Ajmap2_d));
35: PetscCallHIP(hipFree(hipsparseStruct->Aperm2_d));
36: PetscCallHIP(hipFree(hipsparseStruct->Bimap2_d));
37: PetscCallHIP(hipFree(hipsparseStruct->Bjmap2_d));
38: PetscCallHIP(hipFree(hipsparseStruct->Bperm2_d));
39: PetscCallHIP(hipFree(hipsparseStruct->Cperm1_d));
40: PetscCallHIP(hipFree(hipsparseStruct->sendbuf_d));
41: PetscCallHIP(hipFree(hipsparseStruct->recvbuf_d));
42: }
43: hipsparseStruct->use_extended_coo = PETSC_FALSE;
44: delete hipsparseStruct->coo_p;
45: delete hipsparseStruct->coo_pw;
46: hipsparseStruct->coo_p = nullptr;
47: hipsparseStruct->coo_pw = nullptr;
48: PetscFunctionReturn(PETSC_SUCCESS);
49: }
51: static PetscErrorCode MatSetValuesCOO_MPIAIJHIPSPARSE_Basic(Mat A, const PetscScalar v[], InsertMode imode)
52: {
53: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
54: Mat_MPIAIJHIPSPARSE *cusp = (Mat_MPIAIJHIPSPARSE *)a->spptr;
55: PetscInt n = cusp->coo_nd + cusp->coo_no;
57: PetscFunctionBegin;
58: if (cusp->coo_p && v) {
59: thrust::device_ptr<const PetscScalar> d_v;
60: THRUSTARRAY *w = NULL;
62: if (isHipMem(v)) {
63: d_v = thrust::device_pointer_cast(v);
64: } else {
65: w = new THRUSTARRAY(n);
66: w->assign(v, v + n);
67: PetscCall(PetscLogCpuToGpu(n * sizeof(PetscScalar)));
68: d_v = w->data();
69: }
71: auto zibit = thrust::make_zip_iterator(thrust::make_tuple(thrust::make_permutation_iterator(d_v, cusp->coo_p->begin()), cusp->coo_pw->begin()));
72: auto zieit = thrust::make_zip_iterator(thrust::make_tuple(thrust::make_permutation_iterator(d_v, cusp->coo_p->end()), cusp->coo_pw->end()));
73: PetscCall(PetscLogGpuTimeBegin());
74: thrust::for_each(zibit, zieit, VecHIPEquals());
75: PetscCall(PetscLogGpuTimeEnd());
76: delete w;
77: PetscCall(MatSetValuesCOO_SeqAIJHIPSPARSE_Basic(a->A, cusp->coo_pw->data().get(), imode));
78: PetscCall(MatSetValuesCOO_SeqAIJHIPSPARSE_Basic(a->B, cusp->coo_pw->data().get() + cusp->coo_nd, imode));
79: } else {
80: PetscCall(MatSetValuesCOO_SeqAIJHIPSPARSE_Basic(a->A, v, imode));
81: PetscCall(MatSetValuesCOO_SeqAIJHIPSPARSE_Basic(a->B, v ? v + cusp->coo_nd : nullptr, imode));
82: }
83: PetscFunctionReturn(PETSC_SUCCESS);
84: }
86: template <typename Tuple>
87: struct IsNotOffDiagT {
88: PetscInt _cstart, _cend;
90: IsNotOffDiagT(PetscInt cstart, PetscInt cend) : _cstart(cstart), _cend(cend) { }
91: __host__ __device__ bool operator()(Tuple t) { return !(thrust::get<1>(t) < _cstart || thrust::get<1>(t) >= _cend); }
92: };
94: struct IsOffDiag {
95: PetscInt _cstart, _cend;
97: IsOffDiag(PetscInt cstart, PetscInt cend) : _cstart(cstart), _cend(cend) { }
98: __host__ __device__ bool operator()(const PetscInt &c) { return c < _cstart || c >= _cend; }
99: };
101: struct GlobToLoc {
102: PetscInt _start;
104: GlobToLoc(PetscInt start) : _start(start) { }
105: __host__ __device__ PetscInt operator()(const PetscInt &c) { return c - _start; }
106: };
108: static PetscErrorCode MatSetPreallocationCOO_MPIAIJHIPSPARSE_Basic(Mat B, PetscCount n, PetscInt coo_i[], PetscInt coo_j[])
109: {
110: Mat_MPIAIJ *b = (Mat_MPIAIJ *)B->data;
111: Mat_MPIAIJHIPSPARSE *cusp = (Mat_MPIAIJHIPSPARSE *)b->spptr;
112: PetscInt N, *jj;
113: size_t noff = 0;
114: THRUSTINTARRAY d_i(n); /* on device, storing partitioned coo_i with diagonal first, and off-diag next */
115: THRUSTINTARRAY d_j(n);
116: ISLocalToGlobalMapping l2g;
118: PetscFunctionBegin;
119: PetscCall(MatDestroy(&b->A));
120: PetscCall(MatDestroy(&b->B));
122: PetscCall(PetscLogCpuToGpu(2. * n * sizeof(PetscInt)));
123: d_i.assign(coo_i, coo_i + n);
124: d_j.assign(coo_j, coo_j + n);
125: delete cusp->coo_p;
126: delete cusp->coo_pw;
127: cusp->coo_p = NULL;
128: cusp->coo_pw = NULL;
129: PetscCall(PetscLogGpuTimeBegin());
130: auto firstoffd = thrust::find_if(thrust::device, d_j.begin(), d_j.end(), IsOffDiag(B->cmap->rstart, B->cmap->rend));
131: auto firstdiag = thrust::find_if_not(thrust::device, firstoffd, d_j.end(), IsOffDiag(B->cmap->rstart, B->cmap->rend));
132: if (firstoffd != d_j.end() && firstdiag != d_j.end()) {
133: cusp->coo_p = new THRUSTINTARRAY(n);
134: cusp->coo_pw = new THRUSTARRAY(n);
135: thrust::sequence(thrust::device, cusp->coo_p->begin(), cusp->coo_p->end(), 0);
136: auto fzipp = thrust::make_zip_iterator(thrust::make_tuple(d_i.begin(), d_j.begin(), cusp->coo_p->begin()));
137: auto ezipp = thrust::make_zip_iterator(thrust::make_tuple(d_i.end(), d_j.end(), cusp->coo_p->end()));
138: auto mzipp = thrust::partition(thrust::device, fzipp, ezipp, IsNotOffDiagT<thrust::tuple<PetscInt, PetscInt, PetscInt>>(B->cmap->rstart, B->cmap->rend));
139: firstoffd = mzipp.get_iterator_tuple().get<1>();
140: }
141: cusp->coo_nd = thrust::distance(d_j.begin(), firstoffd);
142: cusp->coo_no = thrust::distance(firstoffd, d_j.end());
144: /* from global to local */
145: thrust::transform(thrust::device, d_i.begin(), d_i.end(), d_i.begin(), GlobToLoc(B->rmap->rstart));
146: thrust::transform(thrust::device, d_j.begin(), firstoffd, d_j.begin(), GlobToLoc(B->cmap->rstart));
147: PetscCall(PetscLogGpuTimeEnd());
149: /* copy offdiag column indices to map on the CPU */
150: PetscCall(PetscMalloc1(cusp->coo_no, &jj)); /* jj[] will store compacted col ids of the offdiag part */
151: PetscCallHIP(hipMemcpy(jj, d_j.data().get() + cusp->coo_nd, cusp->coo_no * sizeof(PetscInt), hipMemcpyDeviceToHost));
152: auto o_j = d_j.begin();
153: PetscCall(PetscLogGpuTimeBegin());
154: thrust::advance(o_j, cusp->coo_nd); /* sort and unique offdiag col ids */
155: thrust::sort(thrust::device, o_j, d_j.end());
156: auto wit = thrust::unique(thrust::device, o_j, d_j.end()); /* return end iter of the unique range */
157: PetscCall(PetscLogGpuTimeEnd());
158: noff = thrust::distance(o_j, wit);
159: PetscCall(PetscMalloc1(noff, &b->garray));
160: PetscCallHIP(hipMemcpy(b->garray, d_j.data().get() + cusp->coo_nd, noff * sizeof(PetscInt), hipMemcpyDeviceToHost));
161: PetscCall(PetscLogGpuToCpu((noff + cusp->coo_no) * sizeof(PetscInt)));
162: PetscCall(ISLocalToGlobalMappingCreate(PETSC_COMM_SELF, 1, noff, b->garray, PETSC_COPY_VALUES, &l2g));
163: PetscCall(ISLocalToGlobalMappingSetType(l2g, ISLOCALTOGLOBALMAPPINGHASH));
164: PetscCall(ISGlobalToLocalMappingApply(l2g, IS_GTOLM_DROP, cusp->coo_no, jj, &N, jj));
165: PetscCheck(N == cusp->coo_no, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Unexpected is size %" PetscInt_FMT " != %" PetscInt_FMT " coo size", N, cusp->coo_no);
166: PetscCall(ISLocalToGlobalMappingDestroy(&l2g));
167: PetscCall(MatCreate(PETSC_COMM_SELF, &b->A));
168: PetscCall(MatSetSizes(b->A, B->rmap->n, B->cmap->n, B->rmap->n, B->cmap->n));
169: PetscCall(MatSetType(b->A, MATSEQAIJHIPSPARSE));
170: PetscCall(MatCreate(PETSC_COMM_SELF, &b->B));
171: PetscCall(MatSetSizes(b->B, B->rmap->n, noff, B->rmap->n, noff));
172: PetscCall(MatSetType(b->B, MATSEQAIJHIPSPARSE));
174: /* GPU memory, hipsparse specific call handles it internally */
175: PetscCall(MatSetPreallocationCOO_SeqAIJHIPSPARSE_Basic(b->A, cusp->coo_nd, d_i.data().get(), d_j.data().get()));
176: PetscCall(MatSetPreallocationCOO_SeqAIJHIPSPARSE_Basic(b->B, cusp->coo_no, d_i.data().get() + cusp->coo_nd, jj));
177: PetscCall(PetscFree(jj));
179: PetscCall(MatHIPSPARSESetFormat(b->A, MAT_HIPSPARSE_MULT, cusp->diagGPUMatFormat));
180: PetscCall(MatHIPSPARSESetFormat(b->B, MAT_HIPSPARSE_MULT, cusp->offdiagGPUMatFormat));
181: PetscCall(MatBindToCPU(b->A, B->boundtocpu));
182: PetscCall(MatBindToCPU(b->B, B->boundtocpu));
183: PetscCall(MatSetUpMultiply_MPIAIJ(B));
184: PetscFunctionReturn(PETSC_SUCCESS);
185: }
187: static PetscErrorCode MatSetPreallocationCOO_MPIAIJHIPSPARSE(Mat mat, PetscCount coo_n, PetscInt coo_i[], PetscInt coo_j[])
188: {
189: Mat_MPIAIJ *mpiaij = (Mat_MPIAIJ *)mat->data;
190: Mat_MPIAIJHIPSPARSE *mpidev;
191: PetscBool coo_basic = PETSC_TRUE;
192: PetscMemType mtype = PETSC_MEMTYPE_DEVICE;
193: PetscInt rstart, rend;
195: PetscFunctionBegin;
196: PetscCall(PetscFree(mpiaij->garray));
197: PetscCall(VecDestroy(&mpiaij->lvec));
198: #if defined(PETSC_USE_CTABLE)
199: PetscCall(PetscHMapIDestroy(&mpiaij->colmap));
200: #else
201: PetscCall(PetscFree(mpiaij->colmap));
202: #endif
203: PetscCall(VecScatterDestroy(&mpiaij->Mvctx));
204: mat->assembled = PETSC_FALSE;
205: mat->was_assembled = PETSC_FALSE;
206: PetscCall(MatResetPreallocationCOO_MPIAIJ(mat));
207: PetscCall(MatResetPreallocationCOO_MPIAIJHIPSPARSE(mat));
208: if (coo_i) {
209: PetscCall(PetscLayoutGetRange(mat->rmap, &rstart, &rend));
210: PetscCall(PetscGetMemType(coo_i, &mtype));
211: if (PetscMemTypeHost(mtype)) {
212: for (PetscCount k = 0; k < coo_n; k++) { /* Are there negative indices or remote entries? */
213: if (coo_i[k] < 0 || coo_i[k] < rstart || coo_i[k] >= rend || coo_j[k] < 0) {
214: coo_basic = PETSC_FALSE;
215: break;
216: }
217: }
218: }
219: }
220: /* All ranks must agree on the value of coo_basic */
221: PetscCall(MPIU_Allreduce(MPI_IN_PLACE, &coo_basic, 1, MPIU_BOOL, MPI_LAND, PetscObjectComm((PetscObject)mat)));
222: if (coo_basic) {
223: PetscCall(MatSetPreallocationCOO_MPIAIJHIPSPARSE_Basic(mat, coo_n, coo_i, coo_j));
224: } else {
225: PetscCall(MatSetPreallocationCOO_MPIAIJ(mat, coo_n, coo_i, coo_j));
226: mat->offloadmask = PETSC_OFFLOAD_CPU;
227: /* creates the GPU memory */
228: PetscCall(MatSeqAIJHIPSPARSECopyToGPU(mpiaij->A));
229: PetscCall(MatSeqAIJHIPSPARSECopyToGPU(mpiaij->B));
230: mpidev = static_cast<Mat_MPIAIJHIPSPARSE *>(mpiaij->spptr);
231: mpidev->use_extended_coo = PETSC_TRUE;
233: PetscCallHIP(hipMalloc((void **)&mpidev->Ajmap1_d, (mpiaij->Annz + 1) * sizeof(PetscCount)));
234: PetscCallHIP(hipMalloc((void **)&mpidev->Aperm1_d, mpiaij->Atot1 * sizeof(PetscCount)));
236: PetscCallHIP(hipMalloc((void **)&mpidev->Bjmap1_d, (mpiaij->Bnnz + 1) * sizeof(PetscCount)));
237: PetscCallHIP(hipMalloc((void **)&mpidev->Bperm1_d, mpiaij->Btot1 * sizeof(PetscCount)));
239: PetscCallHIP(hipMalloc((void **)&mpidev->Aimap2_d, mpiaij->Annz2 * sizeof(PetscCount)));
240: PetscCallHIP(hipMalloc((void **)&mpidev->Ajmap2_d, (mpiaij->Annz2 + 1) * sizeof(PetscCount)));
241: PetscCallHIP(hipMalloc((void **)&mpidev->Aperm2_d, mpiaij->Atot2 * sizeof(PetscCount)));
243: PetscCallHIP(hipMalloc((void **)&mpidev->Bimap2_d, mpiaij->Bnnz2 * sizeof(PetscCount)));
244: PetscCallHIP(hipMalloc((void **)&mpidev->Bjmap2_d, (mpiaij->Bnnz2 + 1) * sizeof(PetscCount)));
245: PetscCallHIP(hipMalloc((void **)&mpidev->Bperm2_d, mpiaij->Btot2 * sizeof(PetscCount)));
247: PetscCallHIP(hipMalloc((void **)&mpidev->Cperm1_d, mpiaij->sendlen * sizeof(PetscCount)));
248: PetscCallHIP(hipMalloc((void **)&mpidev->sendbuf_d, mpiaij->sendlen * sizeof(PetscScalar)));
249: PetscCallHIP(hipMalloc((void **)&mpidev->recvbuf_d, mpiaij->recvlen * sizeof(PetscScalar)));
251: PetscCallHIP(hipMemcpy(mpidev->Ajmap1_d, mpiaij->Ajmap1, (mpiaij->Annz + 1) * sizeof(PetscCount), hipMemcpyHostToDevice));
252: PetscCallHIP(hipMemcpy(mpidev->Aperm1_d, mpiaij->Aperm1, mpiaij->Atot1 * sizeof(PetscCount), hipMemcpyHostToDevice));
254: PetscCallHIP(hipMemcpy(mpidev->Bjmap1_d, mpiaij->Bjmap1, (mpiaij->Bnnz + 1) * sizeof(PetscCount), hipMemcpyHostToDevice));
255: PetscCallHIP(hipMemcpy(mpidev->Bperm1_d, mpiaij->Bperm1, mpiaij->Btot1 * sizeof(PetscCount), hipMemcpyHostToDevice));
257: PetscCallHIP(hipMemcpy(mpidev->Aimap2_d, mpiaij->Aimap2, mpiaij->Annz2 * sizeof(PetscCount), hipMemcpyHostToDevice));
258: PetscCallHIP(hipMemcpy(mpidev->Ajmap2_d, mpiaij->Ajmap2, (mpiaij->Annz2 + 1) * sizeof(PetscCount), hipMemcpyHostToDevice));
259: PetscCallHIP(hipMemcpy(mpidev->Aperm2_d, mpiaij->Aperm2, mpiaij->Atot2 * sizeof(PetscCount), hipMemcpyHostToDevice));
261: PetscCallHIP(hipMemcpy(mpidev->Bimap2_d, mpiaij->Bimap2, mpiaij->Bnnz2 * sizeof(PetscCount), hipMemcpyHostToDevice));
262: PetscCallHIP(hipMemcpy(mpidev->Bjmap2_d, mpiaij->Bjmap2, (mpiaij->Bnnz2 + 1) * sizeof(PetscCount), hipMemcpyHostToDevice));
263: PetscCallHIP(hipMemcpy(mpidev->Bperm2_d, mpiaij->Bperm2, mpiaij->Btot2 * sizeof(PetscCount), hipMemcpyHostToDevice));
265: PetscCallHIP(hipMemcpy(mpidev->Cperm1_d, mpiaij->Cperm1, mpiaij->sendlen * sizeof(PetscCount), hipMemcpyHostToDevice));
266: }
267: PetscFunctionReturn(PETSC_SUCCESS);
268: }
270: __global__ static void MatPackCOOValues(const PetscScalar kv[], PetscCount nnz, const PetscCount perm[], PetscScalar buf[])
271: {
272: PetscCount i = blockIdx.x * blockDim.x + threadIdx.x;
273: const PetscCount grid_size = gridDim.x * blockDim.x;
274: for (; i < nnz; i += grid_size) buf[i] = kv[perm[i]];
275: }
277: __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[])
278: {
279: PetscCount i = blockIdx.x * blockDim.x + threadIdx.x;
280: const PetscCount grid_size = gridDim.x * blockDim.x;
281: for (; i < Annz + Bnnz; i += grid_size) {
282: PetscScalar sum = 0.0;
283: if (i < Annz) {
284: for (PetscCount k = Ajmap1[i]; k < Ajmap1[i + 1]; k++) sum += kv[Aperm1[k]];
285: Aa[i] = (imode == INSERT_VALUES ? 0.0 : Aa[i]) + sum;
286: } else {
287: i -= Annz;
288: for (PetscCount k = Bjmap1[i]; k < Bjmap1[i + 1]; k++) sum += kv[Bperm1[k]];
289: Ba[i] = (imode == INSERT_VALUES ? 0.0 : Ba[i]) + sum;
290: }
291: }
292: }
294: __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[])
295: {
296: PetscCount i = blockIdx.x * blockDim.x + threadIdx.x;
297: const PetscCount grid_size = gridDim.x * blockDim.x;
298: for (; i < Annz2 + Bnnz2; i += grid_size) {
299: if (i < Annz2) {
300: for (PetscCount k = Ajmap2[i]; k < Ajmap2[i + 1]; k++) Aa[Aimap2[i]] += kv[Aperm2[k]];
301: } else {
302: i -= Annz2;
303: for (PetscCount k = Bjmap2[i]; k < Bjmap2[i + 1]; k++) Ba[Bimap2[i]] += kv[Bperm2[k]];
304: }
305: }
306: }
308: static PetscErrorCode MatSetValuesCOO_MPIAIJHIPSPARSE(Mat mat, const PetscScalar v[], InsertMode imode)
309: {
310: Mat_MPIAIJ *mpiaij = static_cast<Mat_MPIAIJ *>(mat->data);
311: Mat_MPIAIJHIPSPARSE *mpidev = static_cast<Mat_MPIAIJHIPSPARSE *>(mpiaij->spptr);
312: Mat A = mpiaij->A, B = mpiaij->B;
313: PetscCount Annz = mpiaij->Annz, Annz2 = mpiaij->Annz2, Bnnz = mpiaij->Bnnz, Bnnz2 = mpiaij->Bnnz2;
314: PetscScalar *Aa, *Ba = NULL;
315: PetscScalar *vsend = mpidev->sendbuf_d, *v2 = mpidev->recvbuf_d;
316: const PetscScalar *v1 = v;
317: const PetscCount *Ajmap1 = mpidev->Ajmap1_d, *Ajmap2 = mpidev->Ajmap2_d, *Aimap2 = mpidev->Aimap2_d;
318: const PetscCount *Bjmap1 = mpidev->Bjmap1_d, *Bjmap2 = mpidev->Bjmap2_d, *Bimap2 = mpidev->Bimap2_d;
319: const PetscCount *Aperm1 = mpidev->Aperm1_d, *Aperm2 = mpidev->Aperm2_d, *Bperm1 = mpidev->Bperm1_d, *Bperm2 = mpidev->Bperm2_d;
320: const PetscCount *Cperm1 = mpidev->Cperm1_d;
321: PetscMemType memtype;
323: PetscFunctionBegin;
324: if (mpidev->use_extended_coo) {
325: PetscMPIInt size;
327: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)mat), &size));
328: PetscCall(PetscGetMemType(v, &memtype));
329: if (PetscMemTypeHost(memtype)) { /* If user gave v[] in host, we need to copy it to device */
330: PetscCallHIP(hipMalloc((void **)&v1, mpiaij->coo_n * sizeof(PetscScalar)));
331: PetscCallHIP(hipMemcpy((void *)v1, v, mpiaij->coo_n * sizeof(PetscScalar), hipMemcpyHostToDevice));
332: }
334: if (imode == INSERT_VALUES) {
335: PetscCall(MatSeqAIJHIPSPARSEGetArrayWrite(A, &Aa)); /* write matrix values */
336: PetscCall(MatSeqAIJHIPSPARSEGetArrayWrite(B, &Ba));
337: } else {
338: PetscCall(MatSeqAIJHIPSPARSEGetArray(A, &Aa)); /* read & write matrix values */
339: PetscCall(MatSeqAIJHIPSPARSEGetArray(B, &Ba));
340: }
342: /* Pack entries to be sent to remote */
343: if (mpiaij->sendlen) {
344: hipLaunchKernelGGL(HIP_KERNEL_NAME(MatPackCOOValues), dim3((mpiaij->sendlen + 255) / 256), dim3(256), 0, PetscDefaultHipStream, v1, mpiaij->sendlen, Cperm1, vsend);
345: PetscCallHIP(hipPeekAtLastError());
346: }
348: /* Send remote entries to their owner and overlap the communication with local computation */
349: PetscCall(PetscSFReduceWithMemTypeBegin(mpiaij->coo_sf, MPIU_SCALAR, PETSC_MEMTYPE_HIP, vsend, PETSC_MEMTYPE_HIP, v2, MPI_REPLACE));
350: /* Add local entries to A and B */
351: if (Annz + Bnnz > 0) {
352: hipLaunchKernelGGL(HIP_KERNEL_NAME(MatAddLocalCOOValues), dim3((Annz + Bnnz + 255) / 256), dim3(256), 0, PetscDefaultHipStream, v1, imode, Annz, Ajmap1, Aperm1, Aa, Bnnz, Bjmap1, Bperm1, Ba);
353: PetscCallHIP(hipPeekAtLastError());
354: }
355: PetscCall(PetscSFReduceEnd(mpiaij->coo_sf, MPIU_SCALAR, vsend, v2, MPI_REPLACE));
357: /* Add received remote entries to A and B */
358: if (Annz2 + Bnnz2 > 0) {
359: hipLaunchKernelGGL(HIP_KERNEL_NAME(MatAddRemoteCOOValues), dim3((Annz2 + Bnnz2 + 255) / 256), dim3(256), 0, PetscDefaultHipStream, v2, Annz2, Aimap2, Ajmap2, Aperm2, Aa, Bnnz2, Bimap2, Bjmap2, Bperm2, Ba);
360: PetscCallHIP(hipPeekAtLastError());
361: }
363: if (imode == INSERT_VALUES) {
364: PetscCall(MatSeqAIJHIPSPARSERestoreArrayWrite(A, &Aa));
365: PetscCall(MatSeqAIJHIPSPARSERestoreArrayWrite(B, &Ba));
366: } else {
367: PetscCall(MatSeqAIJHIPSPARSERestoreArray(A, &Aa));
368: PetscCall(MatSeqAIJHIPSPARSERestoreArray(B, &Ba));
369: }
370: if (PetscMemTypeHost(memtype)) PetscCallHIP(hipFree((void *)v1));
371: } else {
372: PetscCall(MatSetValuesCOO_MPIAIJHIPSPARSE_Basic(mat, v, imode));
373: }
374: mat->offloadmask = PETSC_OFFLOAD_GPU;
375: PetscFunctionReturn(PETSC_SUCCESS);
376: }
378: static PetscErrorCode MatMPIAIJGetLocalMatMerge_MPIAIJHIPSPARSE(Mat A, MatReuse scall, IS *glob, Mat *A_loc)
379: {
380: Mat Ad, Ao;
381: const PetscInt *cmap;
383: PetscFunctionBegin;
384: PetscCall(MatMPIAIJGetSeqAIJ(A, &Ad, &Ao, &cmap));
385: PetscCall(MatSeqAIJHIPSPARSEMergeMats(Ad, Ao, scall, A_loc));
386: if (glob) {
387: PetscInt cst, i, dn, on, *gidx;
389: PetscCall(MatGetLocalSize(Ad, NULL, &dn));
390: PetscCall(MatGetLocalSize(Ao, NULL, &on));
391: PetscCall(MatGetOwnershipRangeColumn(A, &cst, NULL));
392: PetscCall(PetscMalloc1(dn + on, &gidx));
393: for (i = 0; i < dn; i++) gidx[i] = cst + i;
394: for (i = 0; i < on; i++) gidx[i + dn] = cmap[i];
395: PetscCall(ISCreateGeneral(PetscObjectComm((PetscObject)Ad), dn + on, gidx, PETSC_OWN_POINTER, glob));
396: }
397: PetscFunctionReturn(PETSC_SUCCESS);
398: }
400: PetscErrorCode MatMPIAIJSetPreallocation_MPIAIJHIPSPARSE(Mat B, PetscInt d_nz, const PetscInt d_nnz[], PetscInt o_nz, const PetscInt o_nnz[])
401: {
402: Mat_MPIAIJ *b = (Mat_MPIAIJ *)B->data;
403: Mat_MPIAIJHIPSPARSE *hipsparseStruct = (Mat_MPIAIJHIPSPARSE *)b->spptr;
404: PetscInt i;
406: PetscFunctionBegin;
407: PetscCall(PetscLayoutSetUp(B->rmap));
408: PetscCall(PetscLayoutSetUp(B->cmap));
409: if (PetscDefined(USE_DEBUG) && d_nnz) {
410: 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]);
411: }
412: if (PetscDefined(USE_DEBUG) && o_nnz) {
413: 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]);
414: }
415: #if defined(PETSC_USE_CTABLE)
416: PetscCall(PetscHMapIDestroy(&b->colmap));
417: #else
418: PetscCall(PetscFree(b->colmap));
419: #endif
420: PetscCall(PetscFree(b->garray));
421: PetscCall(VecDestroy(&b->lvec));
422: PetscCall(VecScatterDestroy(&b->Mvctx));
423: /* Because the B will have been resized we simply destroy it and create a new one each time */
424: PetscCall(MatDestroy(&b->B));
425: if (!b->A) {
426: PetscCall(MatCreate(PETSC_COMM_SELF, &b->A));
427: PetscCall(MatSetSizes(b->A, B->rmap->n, B->cmap->n, B->rmap->n, B->cmap->n));
428: }
429: if (!b->B) {
430: PetscMPIInt size;
431: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)B), &size));
432: PetscCall(MatCreate(PETSC_COMM_SELF, &b->B));
433: PetscCall(MatSetSizes(b->B, B->rmap->n, size > 1 ? B->cmap->N : 0, B->rmap->n, size > 1 ? B->cmap->N : 0));
434: }
435: PetscCall(MatSetType(b->A, MATSEQAIJHIPSPARSE));
436: PetscCall(MatSetType(b->B, MATSEQAIJHIPSPARSE));
437: PetscCall(MatBindToCPU(b->A, B->boundtocpu));
438: PetscCall(MatBindToCPU(b->B, B->boundtocpu));
439: PetscCall(MatSeqAIJSetPreallocation(b->A, d_nz, d_nnz));
440: PetscCall(MatSeqAIJSetPreallocation(b->B, o_nz, o_nnz));
441: PetscCall(MatHIPSPARSESetFormat(b->A, MAT_HIPSPARSE_MULT, hipsparseStruct->diagGPUMatFormat));
442: PetscCall(MatHIPSPARSESetFormat(b->B, MAT_HIPSPARSE_MULT, hipsparseStruct->offdiagGPUMatFormat));
443: B->preallocated = PETSC_TRUE;
444: PetscFunctionReturn(PETSC_SUCCESS);
445: }
447: PetscErrorCode MatMult_MPIAIJHIPSPARSE(Mat A, Vec xx, Vec yy)
448: {
449: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
451: PetscFunctionBegin;
452: PetscCall(VecScatterBegin(a->Mvctx, xx, a->lvec, INSERT_VALUES, SCATTER_FORWARD));
453: PetscCall((*a->A->ops->mult)(a->A, xx, yy));
454: PetscCall(VecScatterEnd(a->Mvctx, xx, a->lvec, INSERT_VALUES, SCATTER_FORWARD));
455: PetscCall((*a->B->ops->multadd)(a->B, a->lvec, yy, yy));
456: PetscFunctionReturn(PETSC_SUCCESS);
457: }
459: PetscErrorCode MatZeroEntries_MPIAIJHIPSPARSE(Mat A)
460: {
461: Mat_MPIAIJ *l = (Mat_MPIAIJ *)A->data;
463: PetscFunctionBegin;
464: PetscCall(MatZeroEntries(l->A));
465: PetscCall(MatZeroEntries(l->B));
466: PetscFunctionReturn(PETSC_SUCCESS);
467: }
469: PetscErrorCode MatMultAdd_MPIAIJHIPSPARSE(Mat A, Vec xx, Vec yy, Vec zz)
470: {
471: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
473: PetscFunctionBegin;
474: PetscCall(VecScatterBegin(a->Mvctx, xx, a->lvec, INSERT_VALUES, SCATTER_FORWARD));
475: PetscCall((*a->A->ops->multadd)(a->A, xx, yy, zz));
476: PetscCall(VecScatterEnd(a->Mvctx, xx, a->lvec, INSERT_VALUES, SCATTER_FORWARD));
477: PetscCall((*a->B->ops->multadd)(a->B, a->lvec, zz, zz));
478: PetscFunctionReturn(PETSC_SUCCESS);
479: }
481: PetscErrorCode MatMultTranspose_MPIAIJHIPSPARSE(Mat A, Vec xx, Vec yy)
482: {
483: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
485: PetscFunctionBegin;
486: PetscCall((*a->B->ops->multtranspose)(a->B, xx, a->lvec));
487: PetscCall((*a->A->ops->multtranspose)(a->A, xx, yy));
488: PetscCall(VecScatterBegin(a->Mvctx, a->lvec, yy, ADD_VALUES, SCATTER_REVERSE));
489: PetscCall(VecScatterEnd(a->Mvctx, a->lvec, yy, ADD_VALUES, SCATTER_REVERSE));
490: PetscFunctionReturn(PETSC_SUCCESS);
491: }
493: PetscErrorCode MatHIPSPARSESetFormat_MPIAIJHIPSPARSE(Mat A, MatHIPSPARSEFormatOperation op, MatHIPSPARSEStorageFormat format)
494: {
495: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
496: Mat_MPIAIJHIPSPARSE *hipsparseStruct = (Mat_MPIAIJHIPSPARSE *)a->spptr;
498: PetscFunctionBegin;
499: switch (op) {
500: case MAT_HIPSPARSE_MULT_DIAG:
501: hipsparseStruct->diagGPUMatFormat = format;
502: break;
503: case MAT_HIPSPARSE_MULT_OFFDIAG:
504: hipsparseStruct->offdiagGPUMatFormat = format;
505: break;
506: case MAT_HIPSPARSE_ALL:
507: hipsparseStruct->diagGPUMatFormat = format;
508: hipsparseStruct->offdiagGPUMatFormat = format;
509: break;
510: default:
511: SETERRQ(PETSC_COMM_SELF, PETSC_ERR_SUP, "unsupported operation %d for MatHIPSPARSEFormatOperation. Only MAT_HIPSPARSE_MULT_DIAG, MAT_HIPSPARSE_MULT_DIAG, and MAT_HIPSPARSE_MULT_ALL are currently supported.", op);
512: }
513: PetscFunctionReturn(PETSC_SUCCESS);
514: }
516: PetscErrorCode MatSetFromOptions_MPIAIJHIPSPARSE(Mat A, PetscOptionItems *PetscOptionsObject)
517: {
518: MatHIPSPARSEStorageFormat format;
519: PetscBool flg;
520: Mat_MPIAIJ *a = (Mat_MPIAIJ *)A->data;
521: Mat_MPIAIJHIPSPARSE *hipsparseStruct = (Mat_MPIAIJHIPSPARSE *)a->spptr;
523: PetscFunctionBegin;
524: PetscOptionsHeadBegin(PetscOptionsObject, "MPIAIJHIPSPARSE options");
525: if (A->factortype == MAT_FACTOR_NONE) {
526: PetscCall(PetscOptionsEnum("-mat_hipsparse_mult_diag_storage_format", "sets storage format of the diagonal blocks of (mpi)aijhipsparse gpu matrices for SpMV", "MatHIPSPARSESetFormat", MatHIPSPARSEStorageFormats, (PetscEnum)hipsparseStruct->diagGPUMatFormat, (PetscEnum *)&format, &flg));
527: if (flg) PetscCall(MatHIPSPARSESetFormat(A, MAT_HIPSPARSE_MULT_DIAG, format));
528: PetscCall(PetscOptionsEnum("-mat_hipsparse_mult_offdiag_storage_format", "sets storage format of the off-diagonal blocks (mpi)aijhipsparse gpu matrices for SpMV", "MatHIPSPARSESetFormat", MatHIPSPARSEStorageFormats, (PetscEnum)hipsparseStruct->offdiagGPUMatFormat, (PetscEnum *)&format, &flg));
529: if (flg) PetscCall(MatHIPSPARSESetFormat(A, MAT_HIPSPARSE_MULT_OFFDIAG, format));
530: PetscCall(PetscOptionsEnum("-mat_hipsparse_storage_format", "sets storage format of the diagonal and off-diagonal blocks (mpi)aijhipsparse gpu matrices for SpMV", "MatHIPSPARSESetFormat", MatHIPSPARSEStorageFormats, (PetscEnum)hipsparseStruct->diagGPUMatFormat, (PetscEnum *)&format, &flg));
531: if (flg) PetscCall(MatHIPSPARSESetFormat(A, MAT_HIPSPARSE_ALL, format));
532: }
533: PetscOptionsHeadEnd();
534: PetscFunctionReturn(PETSC_SUCCESS);
535: }
537: PetscErrorCode MatAssemblyEnd_MPIAIJHIPSPARSE(Mat A, MatAssemblyType mode)
538: {
539: Mat_MPIAIJ *mpiaij = (Mat_MPIAIJ *)A->data;
540: Mat_MPIAIJHIPSPARSE *cusp = (Mat_MPIAIJHIPSPARSE *)mpiaij->spptr;
541: PetscObjectState onnz = A->nonzerostate;
543: PetscFunctionBegin;
544: PetscCall(MatAssemblyEnd_MPIAIJ(A, mode));
545: if (mpiaij->lvec) PetscCall(VecSetType(mpiaij->lvec, VECSEQHIP));
546: if (onnz != A->nonzerostate && cusp->deviceMat) {
547: PetscSplitCSRDataStructure d_mat = cusp->deviceMat, h_mat;
549: PetscCall(PetscInfo(A, "Destroy device mat since nonzerostate changed\n"));
550: PetscCall(PetscNew(&h_mat));
551: PetscCallHIP(hipMemcpy(h_mat, d_mat, sizeof(*d_mat), hipMemcpyDeviceToHost));
552: PetscCallHIP(hipFree(h_mat->colmap));
553: if (h_mat->allocated_indices) {
554: PetscCallHIP(hipFree(h_mat->diag.i));
555: PetscCallHIP(hipFree(h_mat->diag.j));
556: if (h_mat->offdiag.j) {
557: PetscCallHIP(hipFree(h_mat->offdiag.i));
558: PetscCallHIP(hipFree(h_mat->offdiag.j));
559: }
560: }
561: PetscCallHIP(hipFree(d_mat));
562: PetscCall(PetscFree(h_mat));
563: cusp->deviceMat = NULL;
564: }
565: PetscFunctionReturn(PETSC_SUCCESS);
566: }
568: PetscErrorCode MatDestroy_MPIAIJHIPSPARSE(Mat A)
569: {
570: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)A->data;
571: Mat_MPIAIJHIPSPARSE *hipsparseStruct = (Mat_MPIAIJHIPSPARSE *)aij->spptr;
573: PetscFunctionBegin;
574: PetscCheck(hipsparseStruct, PETSC_COMM_SELF, PETSC_ERR_COR, "Missing spptr");
575: if (hipsparseStruct->deviceMat) {
576: PetscSplitCSRDataStructure d_mat = hipsparseStruct->deviceMat, h_mat;
578: PetscCall(PetscInfo(A, "Have device matrix\n"));
579: PetscCall(PetscNew(&h_mat));
580: PetscCallHIP(hipMemcpy(h_mat, d_mat, sizeof(*d_mat), hipMemcpyDeviceToHost));
581: PetscCallHIP(hipFree(h_mat->colmap));
582: if (h_mat->allocated_indices) {
583: PetscCallHIP(hipFree(h_mat->diag.i));
584: PetscCallHIP(hipFree(h_mat->diag.j));
585: if (h_mat->offdiag.j) {
586: PetscCallHIP(hipFree(h_mat->offdiag.i));
587: PetscCallHIP(hipFree(h_mat->offdiag.j));
588: }
589: }
590: PetscCallHIP(hipFree(d_mat));
591: PetscCall(PetscFree(h_mat));
592: }
593: /* Free COO */
594: PetscCall(MatResetPreallocationCOO_MPIAIJHIPSPARSE(A));
595: PetscCallCXX(delete hipsparseStruct);
596: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatMPIAIJSetPreallocation_C", NULL));
597: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatMPIAIJGetLocalMatMerge_C", NULL));
598: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatSetPreallocationCOO_C", NULL));
599: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatSetValuesCOO_C", NULL));
600: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatHIPSPARSESetFormat_C", NULL));
601: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_mpiaijhipsparse_hypre_C", NULL));
602: PetscCall(MatDestroy_MPIAIJ(A));
603: PetscFunctionReturn(PETSC_SUCCESS);
604: }
606: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJHIPSPARSE(Mat B, MatType mtype, MatReuse reuse, Mat *newmat)
607: {
608: Mat_MPIAIJ *a;
609: Mat A;
611: PetscFunctionBegin;
612: PetscCall(PetscDeviceInitialize(PETSC_DEVICE_HIP));
613: if (reuse == MAT_INITIAL_MATRIX) PetscCall(MatDuplicate(B, MAT_COPY_VALUES, newmat));
614: else if (reuse == MAT_REUSE_MATRIX) PetscCall(MatCopy(B, *newmat, SAME_NONZERO_PATTERN));
615: A = *newmat;
616: A->boundtocpu = PETSC_FALSE;
617: PetscCall(PetscFree(A->defaultvectype));
618: PetscCall(PetscStrallocpy(VECHIP, &A->defaultvectype));
620: a = (Mat_MPIAIJ *)A->data;
621: if (a->A) PetscCall(MatSetType(a->A, MATSEQAIJHIPSPARSE));
622: if (a->B) PetscCall(MatSetType(a->B, MATSEQAIJHIPSPARSE));
623: if (a->lvec) PetscCall(VecSetType(a->lvec, VECSEQHIP));
625: if (reuse != MAT_REUSE_MATRIX && !a->spptr) PetscCallCXX(a->spptr = new Mat_MPIAIJHIPSPARSE);
627: A->ops->assemblyend = MatAssemblyEnd_MPIAIJHIPSPARSE;
628: A->ops->mult = MatMult_MPIAIJHIPSPARSE;
629: A->ops->multadd = MatMultAdd_MPIAIJHIPSPARSE;
630: A->ops->multtranspose = MatMultTranspose_MPIAIJHIPSPARSE;
631: A->ops->setfromoptions = MatSetFromOptions_MPIAIJHIPSPARSE;
632: A->ops->destroy = MatDestroy_MPIAIJHIPSPARSE;
633: A->ops->zeroentries = MatZeroEntries_MPIAIJHIPSPARSE;
634: A->ops->productsetfromoptions = MatProductSetFromOptions_MPIAIJBACKEND;
636: PetscCall(PetscObjectChangeTypeName((PetscObject)A, MATMPIAIJHIPSPARSE));
637: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatMPIAIJGetLocalMatMerge_C", MatMPIAIJGetLocalMatMerge_MPIAIJHIPSPARSE));
638: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatMPIAIJSetPreallocation_C", MatMPIAIJSetPreallocation_MPIAIJHIPSPARSE));
639: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatHIPSPARSESetFormat_C", MatHIPSPARSESetFormat_MPIAIJHIPSPARSE));
640: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatSetPreallocationCOO_C", MatSetPreallocationCOO_MPIAIJHIPSPARSE));
641: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatSetValuesCOO_C", MatSetValuesCOO_MPIAIJHIPSPARSE));
642: #if defined(PETSC_HAVE_HYPRE)
643: PetscCall(PetscObjectComposeFunction((PetscObject)A, "MatConvert_mpiaijhipsparse_hypre_C", MatConvert_AIJ_HYPRE));
644: #endif
645: PetscFunctionReturn(PETSC_SUCCESS);
646: }
648: PETSC_EXTERN PetscErrorCode MatCreate_MPIAIJHIPSPARSE(Mat A)
649: {
650: PetscFunctionBegin;
651: PetscCall(PetscDeviceInitialize(PETSC_DEVICE_HIP));
652: PetscCall(MatCreate_MPIAIJ(A));
653: PetscCall(MatConvert_MPIAIJ_MPIAIJHIPSPARSE(A, MATMPIAIJHIPSPARSE, MAT_INPLACE_MATRIX, &A));
654: PetscFunctionReturn(PETSC_SUCCESS);
655: }
657: /*@
658: MatCreateAIJHIPSPARSE - Creates a sparse matrix in AIJ (compressed row) format
659: (the default parallel PETSc format). This matrix will ultimately pushed down
660: to AMD GPUs and use the HIPSPARSE library for calculations. For good matrix
661: assembly performance the user should preallocate the matrix storage by setting
662: the parameter `nz` (or the array `nnz`).
664: Collective
666: Input Parameters:
667: + comm - MPI communicator, set to `PETSC_COMM_SELF`
668: . m - number of local rows (or `PETSC_DECIDE` to have calculated if `M` is given)
669: This value should be the same as the local size used in creating the
670: y vector for the matrix-vector product y = Ax.
671: . n - This value should be the same as the local size used in creating the
672: x vector for the matrix-vector product y = Ax. (or PETSC_DECIDE to have
673: calculated if `N` is given) For square matrices `n` is almost always `m`.
674: . M - number of global rows (or `PETSC_DETERMINE` to have calculated if `m` is given)
675: . N - number of global columns (or `PETSC_DETERMINE` to have calculated if `n` is given)
676: . d_nz - number of nonzeros per row (same for all rows), for the "diagonal" portion of the matrix
677: . d_nnz - array containing the number of nonzeros in the various rows (possibly different for each row) or `NULL`, for the "diagonal" portion of the matrix
678: . o_nz - number of nonzeros per row (same for all rows), for the "off-diagonal" portion of the matrix
679: - o_nnz - array containing the number of nonzeros in the various rows (possibly different for each row) or `NULL`, for the "off-diagonal" portion of the matrix
681: Output Parameter:
682: . A - the matrix
684: Level: intermediate
686: Notes:
687: It is recommended that one use the `MatCreate()`, `MatSetType()` and/or `MatSetFromOptions()`,
688: MatXXXXSetPreallocation() paradigm instead of this routine directly.
689: [MatXXXXSetPreallocation() is, for example, `MatSeqAIJSetPreallocation()`]
691: If `d_nnz` (`o_nnz`) is given then `d_nz` (`o_nz`) is ignored
693: The `MATAIJ` format (compressed row storage), is fully compatible with standard Fortran
694: storage. That is, the stored row and column indices can begin at
695: either one (as in Fortran) or zero.
697: Specify the preallocated storage with either `d_nz` (`o_nz`) or `d_nnz` (`o_nnz`) (not both).
698: Set `d_nz` (`o_nz`) = `PETSC_DEFAULT` and `d_nnz` (`o_nnz`) = `NULL` for PETSc to control dynamic memory
699: allocation.
701: .seealso: [](ch_matrices), `Mat`, `MatCreate()`, `MatCreateAIJ()`, `MatSetValues()`, `MatSeqAIJSetColumnIndices()`, `MatCreateSeqAIJWithArrays()`, `MatCreateAIJ()`, `MATMPIAIJHIPSPARSE`, `MATAIJHIPSPARSE`
702: @*/
703: PetscErrorCode MatCreateAIJHIPSPARSE(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)
704: {
705: PetscMPIInt size;
707: PetscFunctionBegin;
708: PetscCall(MatCreate(comm, A));
709: PetscCall(MatSetSizes(*A, m, n, M, N));
710: PetscCallMPI(MPI_Comm_size(comm, &size));
711: if (size > 1) {
712: PetscCall(MatSetType(*A, MATMPIAIJHIPSPARSE));
713: PetscCall(MatMPIAIJSetPreallocation(*A, d_nz, d_nnz, o_nz, o_nnz));
714: } else {
715: PetscCall(MatSetType(*A, MATSEQAIJHIPSPARSE));
716: PetscCall(MatSeqAIJSetPreallocation(*A, d_nz, d_nnz));
717: }
718: PetscFunctionReturn(PETSC_SUCCESS);
719: }
721: /*MC
722: MATAIJHIPSPARSE - A matrix type to be used for sparse matrices; it is as same as `MATMPIAIJHIPSPARSE`.
724: A matrix type type whose data resides on GPUs. These matrices can be in either
725: CSR, ELL, or Hybrid format. All matrix calculations are performed on AMD GPUs using the HIPSPARSE library.
727: This matrix type is identical to `MATSEQAIJHIPSPARSE` when constructed with a single process communicator,
728: and `MATMPIAIJHIPSPARSE` otherwise. As a result, for single process communicators,
729: `MatSeqAIJSetPreallocation()` is supported, and similarly `MatMPIAIJSetPreallocation()` is supported
730: for communicators controlling multiple processes. It is recommended that you call both of
731: the above preallocation routines for simplicity.
733: Options Database Keys:
734: + -mat_type mpiaijhipsparse - sets the matrix type to `MATMPIAIJHIPSPARSE`
735: . -mat_hipsparse_storage_format csr - sets the storage format of diagonal and off-diagonal matrices. Other options include ell (ellpack) or hyb (hybrid).
736: . -mat_hipsparse_mult_diag_storage_format csr - sets the storage format of diagonal matrix. Other options include ell (ellpack) or hyb (hybrid).
737: - -mat_hipsparse_mult_offdiag_storage_format csr - sets the storage format of off-diagonal matrix. Other options include ell (ellpack) or hyb (hybrid).
739: Level: beginner
741: .seealso: [](ch_matrices), `Mat`, `MatCreateAIJHIPSPARSE()`, `MATSEQAIJHIPSPARSE`, `MATMPIAIJHIPSPARSE`, `MatCreateSeqAIJHIPSPARSE()`, `MatHIPSPARSESetFormat()`, `MatHIPSPARSEStorageFormat`, `MatHIPSPARSEFormatOperation`
742: M*/
744: /*MC
745: MATMPIAIJHIPSPARSE - A matrix type to be used for sparse matrices; it is as same as `MATAIJHIPSPARSE`.
747: Level: beginner
749: .seealso: [](ch_matrices), `Mat`, `MATAIJHIPSPARSE`, `MATSEQAIJHIPSPARSE`
750: M*/
752: // get GPU pointers to stripped down Mat. For both seq and MPI Mat.
753: PetscErrorCode MatHIPSPARSEGetDeviceMatWrite(Mat A, PetscSplitCSRDataStructure *B)
754: {
755: PetscSplitCSRDataStructure d_mat;
756: PetscMPIInt size;
757: int *ai = NULL, *bi = NULL, *aj = NULL, *bj = NULL;
758: PetscScalar *aa = NULL, *ba = NULL;
759: Mat_SeqAIJ *jaca = NULL, *jacb = NULL;
760: Mat_SeqAIJHIPSPARSE *hipsparsestructA = NULL;
761: CsrMatrix *matrixA = NULL, *matrixB = NULL;
763: PetscFunctionBegin;
764: PetscCheck(A->assembled, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Need already assembled matrix");
765: if (A->factortype != MAT_FACTOR_NONE) {
766: *B = NULL;
767: PetscFunctionReturn(PETSC_SUCCESS);
768: }
769: PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)A), &size));
770: // get jaca
771: if (size == 1) {
772: PetscBool isseqaij;
774: PetscCall(PetscObjectBaseTypeCompare((PetscObject)A, MATSEQAIJ, &isseqaij));
775: if (isseqaij) {
776: jaca = (Mat_SeqAIJ *)A->data;
777: PetscCheck(jaca->roworiented, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Device assembly does not currently support column oriented values insertion");
778: hipsparsestructA = (Mat_SeqAIJHIPSPARSE *)A->spptr;
779: d_mat = hipsparsestructA->deviceMat;
780: PetscCall(MatSeqAIJHIPSPARSECopyToGPU(A));
781: } else {
782: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)A->data;
783: PetscCheck(aij->roworiented, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Device assembly does not currently support column oriented values insertion");
784: Mat_MPIAIJHIPSPARSE *spptr = (Mat_MPIAIJHIPSPARSE *)aij->spptr;
785: jaca = (Mat_SeqAIJ *)aij->A->data;
786: hipsparsestructA = (Mat_SeqAIJHIPSPARSE *)aij->A->spptr;
787: d_mat = spptr->deviceMat;
788: PetscCall(MatSeqAIJHIPSPARSECopyToGPU(aij->A));
789: }
790: if (hipsparsestructA->format == MAT_HIPSPARSE_CSR) {
791: Mat_SeqAIJHIPSPARSEMultStruct *matstruct = (Mat_SeqAIJHIPSPARSEMultStruct *)hipsparsestructA->mat;
792: PetscCheck(matstruct, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Missing Mat_SeqAIJHIPSPARSEMultStruct for A");
793: matrixA = (CsrMatrix *)matstruct->mat;
794: bi = NULL;
795: bj = NULL;
796: ba = NULL;
797: } else SETERRQ(PETSC_COMM_SELF, PETSC_ERR_SUP, "Device Mat needs MAT_HIPSPARSE_CSR");
798: } else {
799: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)A->data;
800: PetscCheck(aij->roworiented, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Device assembly does not currently support column oriented values insertion");
801: jaca = (Mat_SeqAIJ *)aij->A->data;
802: jacb = (Mat_SeqAIJ *)aij->B->data;
803: Mat_MPIAIJHIPSPARSE *spptr = (Mat_MPIAIJHIPSPARSE *)aij->spptr;
805: PetscCheck(A->nooffprocentries || aij->donotstash, PetscObjectComm((PetscObject)A), PETSC_ERR_SUP, "Device assembly does not currently support offproc values insertion. Use MatSetOption(A, MAT_NO_OFF_PROC_ENTRIES, PETSC_TRUE) or MatSetOption(A, MAT_IGNORE_OFF_PROC_ENTRIES, PETSC_TRUE)");
806: hipsparsestructA = (Mat_SeqAIJHIPSPARSE *)aij->A->spptr;
807: Mat_SeqAIJHIPSPARSE *hipsparsestructB = (Mat_SeqAIJHIPSPARSE *)aij->B->spptr;
808: PetscCheck(hipsparsestructA->format == MAT_HIPSPARSE_CSR, PETSC_COMM_SELF, PETSC_ERR_SUP, "Device Mat A needs MAT_HIPSPARSE_CSR");
809: if (hipsparsestructB->format == MAT_HIPSPARSE_CSR) {
810: PetscCall(MatSeqAIJHIPSPARSECopyToGPU(aij->A));
811: PetscCall(MatSeqAIJHIPSPARSECopyToGPU(aij->B));
812: Mat_SeqAIJHIPSPARSEMultStruct *matstructA = (Mat_SeqAIJHIPSPARSEMultStruct *)hipsparsestructA->mat;
813: Mat_SeqAIJHIPSPARSEMultStruct *matstructB = (Mat_SeqAIJHIPSPARSEMultStruct *)hipsparsestructB->mat;
814: PetscCheck(matstructA, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Missing Mat_SeqAIJHIPSPARSEMultStruct for A");
815: PetscCheck(matstructB, PETSC_COMM_SELF, PETSC_ERR_PLIB, "Missing Mat_SeqAIJHIPSPARSEMultStruct for B");
816: matrixA = (CsrMatrix *)matstructA->mat;
817: matrixB = (CsrMatrix *)matstructB->mat;
818: if (jacb->compressedrow.use) {
819: if (!hipsparsestructB->rowoffsets_gpu) {
820: hipsparsestructB->rowoffsets_gpu = new THRUSTINTARRAY32(A->rmap->n + 1);
821: hipsparsestructB->rowoffsets_gpu->assign(jacb->i, jacb->i + A->rmap->n + 1);
822: }
823: bi = thrust::raw_pointer_cast(hipsparsestructB->rowoffsets_gpu->data());
824: } else {
825: bi = thrust::raw_pointer_cast(matrixB->row_offsets->data());
826: }
827: bj = thrust::raw_pointer_cast(matrixB->column_indices->data());
828: ba = thrust::raw_pointer_cast(matrixB->values->data());
829: } else SETERRQ(PETSC_COMM_SELF, PETSC_ERR_SUP, "Device Mat B needs MAT_HIPSPARSE_CSR");
830: d_mat = spptr->deviceMat;
831: }
832: if (jaca->compressedrow.use) {
833: if (!hipsparsestructA->rowoffsets_gpu) {
834: hipsparsestructA->rowoffsets_gpu = new THRUSTINTARRAY32(A->rmap->n + 1);
835: hipsparsestructA->rowoffsets_gpu->assign(jaca->i, jaca->i + A->rmap->n + 1);
836: }
837: ai = thrust::raw_pointer_cast(hipsparsestructA->rowoffsets_gpu->data());
838: } else {
839: ai = thrust::raw_pointer_cast(matrixA->row_offsets->data());
840: }
841: aj = thrust::raw_pointer_cast(matrixA->column_indices->data());
842: aa = thrust::raw_pointer_cast(matrixA->values->data());
844: if (!d_mat) {
845: PetscSplitCSRDataStructure h_mat;
847: // create and populate strucy on host and copy on device
848: PetscCall(PetscInfo(A, "Create device matrix\n"));
849: PetscCall(PetscNew(&h_mat));
850: PetscCallHIP(hipMalloc((void **)&d_mat, sizeof(*d_mat)));
851: if (size > 1) { /* need the colmap array */
852: Mat_MPIAIJ *aij = (Mat_MPIAIJ *)A->data;
853: PetscInt *colmap;
854: PetscInt ii, n = aij->B->cmap->n, N = A->cmap->N;
856: PetscCheck(!n || aij->garray, PETSC_COMM_SELF, PETSC_ERR_PLIB, "MPIAIJ Matrix was assembled but is missing garray");
858: PetscCall(PetscCalloc1(N + 1, &colmap));
859: for (ii = 0; ii < n; ii++) colmap[aij->garray[ii]] = ii + 1;
860: #if defined(PETSC_USE_64BIT_INDICES)
861: { // have to make a long version of these
862: int *h_bi32, *h_bj32;
863: PetscInt *h_bi64, *h_bj64, *d_bi64, *d_bj64;
864: PetscCall(PetscCalloc4(A->rmap->n + 1, &h_bi32, jacb->nz, &h_bj32, A->rmap->n + 1, &h_bi64, jacb->nz, &h_bj64));
865: PetscCallHIP(hipMemcpy(h_bi32, bi, (A->rmap->n + 1) * sizeof(*h_bi32), hipMemcpyDeviceToHost));
866: for (int i = 0; i < A->rmap->n + 1; i++) h_bi64[i] = h_bi32[i];
867: PetscCallHIP(hipMemcpy(h_bj32, bj, jacb->nz * sizeof(*h_bj32), hipMemcpyDeviceToHost));
868: for (int i = 0; i < jacb->nz; i++) h_bj64[i] = h_bj32[i];
870: PetscCallHIP(hipMalloc((void **)&d_bi64, (A->rmap->n + 1) * sizeof(*d_bi64)));
871: PetscCallHIP(hipMemcpy(d_bi64, h_bi64, (A->rmap->n + 1) * sizeof(*d_bi64), hipMemcpyHostToDevice));
872: PetscCallHIP(hipMalloc((void **)&d_bj64, jacb->nz * sizeof(*d_bj64)));
873: PetscCallHIP(hipMemcpy(d_bj64, h_bj64, jacb->nz * sizeof(*d_bj64), hipMemcpyHostToDevice));
875: h_mat->offdiag.i = d_bi64;
876: h_mat->offdiag.j = d_bj64;
877: h_mat->allocated_indices = PETSC_TRUE;
879: PetscCall(PetscFree4(h_bi32, h_bj32, h_bi64, h_bj64));
880: }
881: #else
882: h_mat->offdiag.i = (PetscInt *)bi;
883: h_mat->offdiag.j = (PetscInt *)bj;
884: h_mat->allocated_indices = PETSC_FALSE;
885: #endif
886: h_mat->offdiag.a = ba;
887: h_mat->offdiag.n = A->rmap->n;
889: PetscCallHIP(hipMalloc((void **)&h_mat->colmap, (N + 1) * sizeof(*h_mat->colmap)));
890: PetscCallHIP(hipMemcpy(h_mat->colmap, colmap, (N + 1) * sizeof(*h_mat->colmap), hipMemcpyHostToDevice));
891: PetscCall(PetscFree(colmap));
892: }
893: h_mat->rstart = A->rmap->rstart;
894: h_mat->rend = A->rmap->rend;
895: h_mat->cstart = A->cmap->rstart;
896: h_mat->cend = A->cmap->rend;
897: h_mat->M = A->cmap->N;
898: #if defined(PETSC_USE_64BIT_INDICES)
899: {
900: int *h_ai32, *h_aj32;
901: PetscInt *h_ai64, *h_aj64, *d_ai64, *d_aj64;
902: PetscCall(PetscCalloc4(A->rmap->n + 1, &h_ai32, jaca->nz, &h_aj32, A->rmap->n + 1, &h_ai64, jaca->nz, &h_aj64));
903: PetscCallHIP(hipMemcpy(h_ai32, ai, (A->rmap->n + 1) * sizeof(*h_ai32), hipMemcpyDeviceToHost));
904: for (int i = 0; i < A->rmap->n + 1; i++) h_ai64[i] = h_ai32[i];
905: PetscCallHIP(hipMemcpy(h_aj32, aj, jaca->nz * sizeof(*h_aj32), hipMemcpyDeviceToHost));
906: for (int i = 0; i < jaca->nz; i++) h_aj64[i] = h_aj32[i];
908: PetscCallHIP(hipMalloc((void **)&d_ai64, (A->rmap->n + 1) * sizeof(*d_ai64)));
909: PetscCallHIP(hipMemcpy(d_ai64, h_ai64, (A->rmap->n + 1) * sizeof(*d_ai64), hipMemcpyHostToDevice));
910: PetscCallHIP(hipMalloc((void **)&d_aj64, jaca->nz * sizeof(*d_aj64)));
911: PetscCallHIP(hipMemcpy(d_aj64, h_aj64, jaca->nz * sizeof(*d_aj64), hipMemcpyHostToDevice));
913: h_mat->diag.i = d_ai64;
914: h_mat->diag.j = d_aj64;
915: h_mat->allocated_indices = PETSC_TRUE;
917: PetscCall(PetscFree4(h_ai32, h_aj32, h_ai64, h_aj64));
918: }
919: #else
920: h_mat->diag.i = (PetscInt *)ai;
921: h_mat->diag.j = (PetscInt *)aj;
922: h_mat->allocated_indices = PETSC_FALSE;
923: #endif
924: h_mat->diag.a = aa;
925: h_mat->diag.n = A->rmap->n;
926: h_mat->rank = PetscGlobalRank;
927: // copy pointers and metadata to device
928: PetscCallHIP(hipMemcpy(d_mat, h_mat, sizeof(*d_mat), hipMemcpyHostToDevice));
929: PetscCall(PetscFree(h_mat));
930: } else {
931: PetscCall(PetscInfo(A, "Reusing device matrix\n"));
932: }
933: *B = d_mat;
934: A->offloadmask = PETSC_OFFLOAD_GPU;
935: PetscFunctionReturn(PETSC_SUCCESS);
936: }