Actual source code: matcuda.cu

slepc-3.20.2 2024-03-15
Report Typos and Errors
  1: /*
  2:    - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
  3:    SLEPc - Scalable Library for Eigenvalue Problem Computations
  4:    Copyright (c) 2002-, Universitat Politecnica de Valencia, Spain

  6:    This file is part of SLEPc.
  7:    SLEPc is distributed under a 2-clause BSD license (see LICENSE).
  8:    - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
  9: */
 10: /*
 11:    BV implemented with a dense Mat (CUDA version)
 12: */

 14: #include <slepc/private/bvimpl.h>
 15: #include <slepccublas.h>
 16: #include "../src/sys/classes/bv/impls/mat/bvmat.h"

 18: PetscErrorCode BVMult_Mat_CUDA(BV Y,PetscScalar alpha,PetscScalar beta,BV X,Mat Q)
 19: {
 20:   BV_MAT            *y = (BV_MAT*)Y->data,*x = (BV_MAT*)X->data;
 21:   const PetscScalar *d_px,*d_A,*d_B,*d_q;
 22:   PetscScalar       *d_py,*d_C;
 23:   PetscInt          ldq;

 25:   PetscFunctionBegin;
 26:   if (!Y->n) PetscFunctionReturn(PETSC_SUCCESS);
 27:   PetscCall(MatDenseCUDAGetArrayRead(x->A,&d_px));
 28:   if (beta==(PetscScalar)0.0) PetscCall(MatDenseCUDAGetArrayWrite(y->A,&d_py));
 29:   else PetscCall(MatDenseCUDAGetArray(y->A,&d_py));
 30:   d_A = d_px+(X->nc+X->l)*X->ld;
 31:   d_C = d_py+(Y->nc+Y->l)*Y->ld;
 32:   if (Q) {
 33:     PetscCall(MatDenseGetLDA(Q,&ldq));
 34:     PetscCall(BV_MatDenseCUDAGetArrayRead(Y,Q,&d_q));
 35:     d_B = d_q+Y->l*ldq+X->l;
 36:     PetscCall(BVMult_BLAS_CUDA(Y,Y->n,Y->k-Y->l,X->k-X->l,alpha,d_A,X->ld,d_B,ldq,beta,d_C,Y->ld));
 37:     PetscCall(BV_MatDenseCUDARestoreArrayRead(Y,Q,&d_q));
 38:   } else PetscCall(BVAXPY_BLAS_CUDA(Y,Y->n,Y->k-Y->l,alpha,d_A,X->ld,beta,d_C,Y->ld));
 39:   PetscCall(MatDenseCUDARestoreArrayRead(x->A,&d_px));
 40:   if (beta==(PetscScalar)0.0) PetscCall(MatDenseCUDARestoreArrayWrite(y->A,&d_py));
 41:   else PetscCall(MatDenseCUDARestoreArray(y->A,&d_py));
 42:   PetscFunctionReturn(PETSC_SUCCESS);
 43: }

 45: PetscErrorCode BVMultVec_Mat_CUDA(BV X,PetscScalar alpha,PetscScalar beta,Vec y,PetscScalar *q)
 46: {
 47:   BV_MAT            *x = (BV_MAT*)X->data;
 48:   PetscScalar       *d_py,*d_q;
 49:   const PetscScalar *d_px;

 51:   PetscFunctionBegin;
 52:   PetscCall(MatDenseCUDAGetArrayRead(x->A,&d_px));
 53:   if (beta==(PetscScalar)0.0) PetscCall(VecCUDAGetArrayWrite(y,&d_py));
 54:   else PetscCall(VecCUDAGetArray(y,&d_py));
 55:   if (!q) PetscCall(VecCUDAGetArray(X->buffer,&d_q));
 56:   else {
 57:     PetscInt k=X->k-X->l;
 58:     PetscCallCUDA(cudaMalloc((void**)&d_q,k*sizeof(PetscScalar)));
 59:     PetscCallCUDA(cudaMemcpy(d_q,q,k*sizeof(PetscScalar),cudaMemcpyHostToDevice));
 60:     PetscCall(PetscLogCpuToGpu(k*sizeof(PetscScalar)));
 61:   }
 62:   PetscCall(BVMultVec_BLAS_CUDA(X,X->n,X->k-X->l,alpha,d_px+(X->nc+X->l)*X->ld,X->ld,d_q,beta,d_py));
 63:   PetscCall(MatDenseCUDARestoreArrayRead(x->A,&d_px));
 64:   if (beta==(PetscScalar)0.0) PetscCall(VecCUDARestoreArrayWrite(y,&d_py));
 65:   else PetscCall(VecCUDARestoreArray(y,&d_py));
 66:   if (!q) PetscCall(VecCUDARestoreArray(X->buffer,&d_q));
 67:   else PetscCallCUDA(cudaFree(d_q));
 68:   PetscFunctionReturn(PETSC_SUCCESS);
 69: }

 71: PetscErrorCode BVMultInPlace_Mat_CUDA(BV V,Mat Q,PetscInt s,PetscInt e)
 72: {
 73:   BV_MAT            *ctx = (BV_MAT*)V->data;
 74:   PetscScalar       *d_pv;
 75:   const PetscScalar *d_q;
 76:   PetscInt          ldq;

 78:   PetscFunctionBegin;
 79:   if (s>=e || !V->n) PetscFunctionReturn(PETSC_SUCCESS);
 80:   PetscCall(MatDenseGetLDA(Q,&ldq));
 81:   PetscCall(MatDenseCUDAGetArray(ctx->A,&d_pv));
 82:   PetscCall(BV_MatDenseCUDAGetArrayRead(V,Q,&d_q));
 83:   PetscCall(BVMultInPlace_BLAS_CUDA(V,V->n,V->k-V->l,s-V->l,e-V->l,d_pv+(V->nc+V->l)*V->ld,V->ld,d_q+V->l*ldq+V->l,ldq,PETSC_FALSE));
 84:   PetscCall(BV_MatDenseCUDARestoreArrayRead(V,Q,&d_q));
 85:   PetscCall(MatDenseCUDARestoreArray(ctx->A,&d_pv));
 86:   PetscFunctionReturn(PETSC_SUCCESS);
 87: }

 89: PetscErrorCode BVMultInPlaceHermitianTranspose_Mat_CUDA(BV V,Mat Q,PetscInt s,PetscInt e)
 90: {
 91:   BV_MAT            *ctx = (BV_MAT*)V->data;
 92:   PetscScalar       *d_pv;
 93:   const PetscScalar *d_q;
 94:   PetscInt          ldq;

 96:   PetscFunctionBegin;
 97:   if (s>=e || !V->n) PetscFunctionReturn(PETSC_SUCCESS);
 98:   PetscCall(MatDenseGetLDA(Q,&ldq));
 99:   PetscCall(MatDenseCUDAGetArray(ctx->A,&d_pv));
100:   PetscCall(BV_MatDenseCUDAGetArrayRead(V,Q,&d_q));
101:   PetscCall(BVMultInPlace_BLAS_CUDA(V,V->n,V->k-V->l,s-V->l,e-V->l,d_pv+(V->nc+V->l)*V->ld,V->ld,d_q+V->l*ldq+V->l,ldq,PETSC_TRUE));
102:   PetscCall(BV_MatDenseCUDARestoreArrayRead(V,Q,&d_q));
103:   PetscCall(MatDenseCUDARestoreArray(ctx->A,&d_pv));
104:   PetscFunctionReturn(PETSC_SUCCESS);
105: }

107: PetscErrorCode BVDot_Mat_CUDA(BV X,BV Y,Mat M)
108: {
109:   BV_MAT            *x = (BV_MAT*)X->data,*y = (BV_MAT*)Y->data;
110:   const PetscScalar *d_px,*d_py;
111:   PetscScalar       *pm;
112:   PetscInt          ldm;

114:   PetscFunctionBegin;
115:   PetscCall(MatDenseGetLDA(M,&ldm));
116:   PetscCall(MatDenseCUDAGetArrayRead(x->A,&d_px));
117:   PetscCall(MatDenseCUDAGetArrayRead(y->A,&d_py));
118:   PetscCall(MatDenseGetArrayWrite(M,&pm));
119:   PetscCall(BVDot_BLAS_CUDA(X,Y->k-Y->l,X->k-X->l,X->n,d_py+(Y->nc+Y->l)*Y->ld,Y->ld,d_px+(X->nc+X->l)*X->ld,X->ld,pm+X->l*ldm+Y->l,ldm,x->mpi));
120:   PetscCall(MatDenseRestoreArrayWrite(M,&pm));
121:   PetscCall(MatDenseCUDARestoreArrayRead(x->A,&d_px));
122:   PetscCall(MatDenseCUDARestoreArrayRead(y->A,&d_py));
123:   PetscFunctionReturn(PETSC_SUCCESS);
124: }

126: PetscErrorCode BVDotVec_Mat_CUDA(BV X,Vec y,PetscScalar *q)
127: {
128:   BV_MAT            *x = (BV_MAT*)X->data;
129:   const PetscScalar *d_px,*d_py;
130:   Vec               z = y;

132:   PetscFunctionBegin;
133:   if (PetscUnlikely(X->matrix)) {
134:     PetscCall(BV_IPMatMult(X,y));
135:     z = X->Bx;
136:   }
137:   PetscCall(MatDenseCUDAGetArrayRead(x->A,&d_px));
138:   PetscCall(VecCUDAGetArrayRead(z,&d_py));
139:   PetscCall(BVDotVec_BLAS_CUDA(X,X->n,X->k-X->l,d_px+(X->nc+X->l)*X->ld,X->ld,d_py,q,x->mpi));
140:   PetscCall(VecCUDARestoreArrayRead(z,&d_py));
141:   PetscCall(MatDenseCUDARestoreArrayRead(x->A,&d_px));
142:   PetscFunctionReturn(PETSC_SUCCESS);
143: }

145: PetscErrorCode BVDotVec_Local_Mat_CUDA(BV X,Vec y,PetscScalar *m)
146: {
147:   BV_MAT            *x = (BV_MAT*)X->data;
148:   const PetscScalar *d_px,*d_py;
149:   Vec               z = y;

151:   PetscFunctionBegin;
152:   if (PetscUnlikely(X->matrix)) {
153:     PetscCall(BV_IPMatMult(X,y));
154:     z = X->Bx;
155:   }
156:   PetscCall(MatDenseCUDAGetArrayRead(x->A,&d_px));
157:   PetscCall(VecCUDAGetArrayRead(z,&d_py));
158:   PetscCall(BVDotVec_BLAS_CUDA(X,X->n,X->k-X->l,d_px+(X->nc+X->l)*X->ld,X->ld,d_py,m,PETSC_FALSE));
159:   PetscCall(VecCUDARestoreArrayRead(z,&d_py));
160:   PetscCall(MatDenseCUDARestoreArrayRead(x->A,&d_px));
161:   PetscFunctionReturn(PETSC_SUCCESS);
162: }

164: PetscErrorCode BVScale_Mat_CUDA(BV bv,PetscInt j,PetscScalar alpha)
165: {
166:   BV_MAT         *ctx = (BV_MAT*)bv->data;
167:   PetscScalar    *d_array,*d_A;
168:   PetscInt       n=0;

170:   PetscFunctionBegin;
171:   if (!bv->n) PetscFunctionReturn(PETSC_SUCCESS);
172:   PetscCall(MatDenseCUDAGetArray(ctx->A,&d_array));
173:   if (PetscUnlikely(j<0)) {
174:     d_A = d_array+(bv->nc+bv->l)*bv->ld;
175:     n = (bv->k-bv->l)*bv->ld;
176:   } else {
177:     d_A = d_array+(bv->nc+j)*bv->ld;
178:     n = bv->n;
179:   }
180:   PetscCall(BVScale_BLAS_CUDA(bv,n,d_A,alpha));
181:   PetscCall(MatDenseCUDARestoreArray(ctx->A,&d_array));
182:   PetscFunctionReturn(PETSC_SUCCESS);
183: }

185: PetscErrorCode BVMatMult_Mat_CUDA(BV V,Mat A,BV W)
186: {
187:   BV_MAT            *v = (BV_MAT*)V->data,*w = (BV_MAT*)W->data;
188:   Mat               Vmat,Wmat;
189:   const PetscScalar *d_pv;
190:   PetscScalar       *d_pw;
191:   PetscInt          j;

193:   PetscFunctionBegin;
194:   if (V->vmm) {
195:     PetscCall(BVGetMat(V,&Vmat));
196:     PetscCall(BVGetMat(W,&Wmat));
197:     PetscCall(MatProductCreateWithMat(A,Vmat,NULL,Wmat));
198:     PetscCall(MatProductSetType(Wmat,MATPRODUCT_AB));
199:     PetscCall(MatProductSetFromOptions(Wmat));
200:     PetscCall(MatProductSymbolic(Wmat));
201:     PetscCall(MatProductNumeric(Wmat));
202:     PetscCall(MatProductClear(Wmat));
203:     PetscCall(BVRestoreMat(V,&Vmat));
204:     PetscCall(BVRestoreMat(W,&Wmat));
205:   } else {
206:     PetscCall(MatDenseCUDAGetArrayRead(v->A,&d_pv));
207:     PetscCall(MatDenseCUDAGetArrayWrite(w->A,&d_pw));
208:     for (j=0;j<V->k-V->l;j++) {
209:       PetscCall(VecCUDAPlaceArray(V->cv[1],(PetscScalar *)d_pv+(V->nc+V->l+j)*V->ld));
210:       PetscCall(VecCUDAPlaceArray(W->cv[1],d_pw+(W->nc+W->l+j)*W->ld));
211:       PetscCall(MatMult(A,V->cv[1],W->cv[1]));
212:       PetscCall(VecCUDAResetArray(V->cv[1]));
213:       PetscCall(VecCUDAResetArray(W->cv[1]));
214:     }
215:     PetscCall(MatDenseCUDARestoreArrayRead(v->A,&d_pv));
216:     PetscCall(MatDenseCUDARestoreArrayWrite(w->A,&d_pw));
217:   }
218:   PetscFunctionReturn(PETSC_SUCCESS);
219: }

221: PetscErrorCode BVCopy_Mat_CUDA(BV V,BV W)
222: {
223:   BV_MAT            *v = (BV_MAT*)V->data,*w = (BV_MAT*)W->data;
224:   const PetscScalar *d_pv;
225:   PetscScalar       *d_pw;

227:   PetscFunctionBegin;
228:   PetscCall(MatDenseCUDAGetArrayRead(v->A,&d_pv));
229:   PetscCall(MatDenseCUDAGetArray(w->A,&d_pw));
230:   PetscCallCUDA(cudaMemcpy2D(d_pw+(W->nc+W->l)*W->ld,W->ld*sizeof(PetscScalar),d_pv+(V->nc+V->l)*V->ld,V->ld*sizeof(PetscScalar),V->n*sizeof(PetscScalar),V->k-V->l,cudaMemcpyDeviceToDevice));
231:   PetscCall(MatDenseCUDARestoreArrayRead(v->A,&d_pv));
232:   PetscCall(MatDenseCUDARestoreArray(w->A,&d_pw));
233:   PetscFunctionReturn(PETSC_SUCCESS);
234: }

236: PetscErrorCode BVCopyColumn_Mat_CUDA(BV V,PetscInt j,PetscInt i)
237: {
238:   BV_MAT         *v = (BV_MAT*)V->data;
239:   PetscScalar    *d_pv;

241:   PetscFunctionBegin;
242:   PetscCall(MatDenseCUDAGetArray(v->A,&d_pv));
243:   PetscCallCUDA(cudaMemcpy(d_pv+(V->nc+i)*V->ld,d_pv+(V->nc+j)*V->ld,V->n*sizeof(PetscScalar),cudaMemcpyDeviceToDevice));
244:   PetscCall(MatDenseCUDARestoreArray(v->A,&d_pv));
245:   PetscFunctionReturn(PETSC_SUCCESS);
246: }

248: PetscErrorCode BVGetColumn_Mat_CUDA(BV bv,PetscInt j,Vec*)
249: {
250:   BV_MAT         *ctx = (BV_MAT*)bv->data;
251:   PetscScalar    *d_pv;
252:   PetscInt       l;

254:   PetscFunctionBegin;
255:   l = BVAvailableVec;
256:   PetscCall(MatDenseCUDAGetArray(ctx->A,&d_pv));
257:   PetscCall(VecCUDAPlaceArray(bv->cv[l],d_pv+(bv->nc+j)*bv->ld));
258:   PetscFunctionReturn(PETSC_SUCCESS);
259: }

261: PetscErrorCode BVRestoreColumn_Mat_CUDA(BV bv,PetscInt j,Vec*)
262: {
263:   BV_MAT         *ctx = (BV_MAT*)bv->data;
264:   PetscInt       l;

266:   PetscFunctionBegin;
267:   l = (j==bv->ci[0])? 0: 1;
268:   PetscCall(VecCUDAResetArray(bv->cv[l]));
269:   PetscCall(MatDenseCUDARestoreArray(ctx->A,NULL));
270:   PetscFunctionReturn(PETSC_SUCCESS);
271: }

273: PetscErrorCode BVRestoreSplit_Mat_CUDA(BV bv,BV *L,BV *R)
274: {
275:   Mat               A;
276:   const PetscScalar *d_pv;
277:   PetscObjectState  lstate,rstate;
278:   PetscBool         change=PETSC_FALSE;

280:   PetscFunctionBegin;
281:   /* force sync flag to PETSC_CUDA_BOTH */
282:   if (L) {
283:     PetscCall(PetscObjectStateGet((PetscObject)*L,&lstate));
284:     if (lstate != bv->lstate) {
285:       A = ((BV_MAT*)bv->L->data)->A;
286:       PetscCall(MatDenseCUDAGetArrayRead(A,&d_pv));
287:       PetscCall(MatDenseCUDARestoreArrayRead(A,&d_pv));
288:       change = PETSC_TRUE;
289:     }
290:   }
291:   if (R) {
292:     PetscCall(PetscObjectStateGet((PetscObject)*R,&rstate));
293:     if (rstate != bv->rstate) {
294:       A = ((BV_MAT*)bv->R->data)->A;
295:       PetscCall(MatDenseCUDAGetArrayRead(A,&d_pv));
296:       PetscCall(MatDenseCUDARestoreArrayRead(A,&d_pv));
297:       change = PETSC_TRUE;
298:     }
299:   }
300:   if (change) {
301:     A = ((BV_MAT*)bv->data)->A;
302:     PetscCall(MatDenseCUDAGetArray(A,(PetscScalar **)&d_pv));
303:     PetscCall(MatDenseCUDARestoreArray(A,(PetscScalar **)&d_pv));
304:   }
305:   PetscFunctionReturn(PETSC_SUCCESS);
306: }

308: PetscErrorCode BVGetMat_Mat_CUDA(BV bv,Mat *A)
309: {
310:   BV_MAT         *ctx = (BV_MAT*)bv->data;
311:   PetscScalar    *vv,*aa;
312:   PetscBool      create=PETSC_FALSE;
313:   PetscInt       m,cols;
314:   VecType        vtype;

316:   PetscFunctionBegin;
317:   m = bv->k-bv->l;
318:   if (!bv->Aget) create=PETSC_TRUE;
319:   else {
320:     PetscCall(MatDenseCUDAGetArray(bv->Aget,&aa));
321:     PetscCheck(!aa,PetscObjectComm((PetscObject)bv),PETSC_ERR_ARG_WRONGSTATE,"BVGetMat already called on this BV");
322:     PetscCall(MatGetSize(bv->Aget,NULL,&cols));
323:     if (cols!=m) {
324:       PetscCall(MatDestroy(&bv->Aget));
325:       create=PETSC_TRUE;
326:     }
327:   }
328:   PetscCall(MatDenseCUDAGetArray(ctx->A,&vv));
329:   if (create) {
330:     PetscCall(VecGetType(bv->t,&vtype));
331:     PetscCall(MatCreateDenseFromVecType(PetscObjectComm((PetscObject)bv),vtype,bv->n,PETSC_DECIDE,bv->N,m,bv->ld,vv,&bv->Aget)); /* pass a pointer to avoid allocation of storage */
332:     PetscCall(MatDenseCUDAReplaceArray(bv->Aget,NULL));  /* replace with a null pointer, the value after BVRestoreMat */
333:   }
334:   PetscCall(MatDenseCUDAPlaceArray(bv->Aget,vv+(bv->nc+bv->l)*bv->ld));  /* set the actual pointer */
335:   *A = bv->Aget;
336:   PetscFunctionReturn(PETSC_SUCCESS);
337: }

339: PetscErrorCode BVRestoreMat_Mat_CUDA(BV bv,Mat *A)
340: {
341:   BV_MAT         *ctx = (BV_MAT*)bv->data;
342:   PetscScalar    *vv,*aa;

344:   PetscFunctionBegin;
345:   PetscCall(MatDenseCUDAGetArray(bv->Aget,&aa));
346:   vv = aa-(bv->nc+bv->l)*bv->ld;
347:   PetscCall(MatDenseCUDAResetArray(bv->Aget));
348:   PetscCall(MatDenseCUDARestoreArray(ctx->A,&vv));
349:   *A = NULL;
350:   PetscFunctionReturn(PETSC_SUCCESS);
351: }