| Line | Branch | Exec | Source |
|---|---|---|---|
| 1 | /* | ||
| 2 | - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - | ||
| 3 | SLEPc - Scalable Library for Eigenvalue Problem Computations | ||
| 4 | Copyright (c) 2002-, Universitat Politecnica de Valencia, Spain | ||
| 5 | |||
| 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 (HIP version) | ||
| 12 | */ | ||
| 13 | |||
| 14 | #include <slepc/private/bvimpl.h> | ||
| 15 | #include <slepccupmblas.h> | ||
| 16 | #include "../src/sys/classes/bv/impls/mat/bvmat.h" | ||
| 17 | |||
| 18 | ✗ | PetscErrorCode BVMult_Mat_HIP(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; | |
| 24 | |||
| 25 | PetscFunctionBegin; | ||
| 26 | ✗ | if (!Y->n) PetscFunctionReturn(PETSC_SUCCESS); | |
| 27 | ✗ | PetscCall(MatDenseHIPGetArrayRead(x->A,&d_px)); | |
| 28 | ✗ | if (beta==(PetscScalar)0.0) PetscCall(MatDenseHIPGetArrayWrite(y->A,&d_py)); | |
| 29 | ✗ | else PetscCall(MatDenseHIPGetArray(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_MatDenseHIPGetArrayRead(Y,Q,&d_q)); | |
| 35 | ✗ | d_B = d_q+Y->l*ldq+X->l; | |
| 36 | ✗ | PetscCall(BVMult_BLAS_HIP(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_MatDenseHIPRestoreArrayRead(Y,Q,&d_q)); | |
| 38 | ✗ | } else PetscCall(BVAXPY_BLAS_HIP(Y,Y->n,Y->k-Y->l,alpha,d_A,X->ld,beta,d_C,Y->ld)); | |
| 39 | ✗ | PetscCall(MatDenseHIPRestoreArrayRead(x->A,&d_px)); | |
| 40 | ✗ | if (beta==(PetscScalar)0.0) PetscCall(MatDenseHIPRestoreArrayWrite(y->A,&d_py)); | |
| 41 | ✗ | else PetscCall(MatDenseHIPRestoreArray(y->A,&d_py)); | |
| 42 | ✗ | PetscFunctionReturn(PETSC_SUCCESS); | |
| 43 | ✗ | } | |
| 44 | |||
| 45 | 9785202960035239728 | PetscErrorCode BVMultVec_Mat_HIP(BV X,PetscScalar alpha,PetscScalar beta,Vec y,PetscScalar *q) | |
| 46 | { | ||
| 47 | 9785202960035239728 | BV_MAT *x = (BV_MAT*)X->data; | |
| 48 | 9785202960035239728 | PetscScalar *d_py,*d_q; | |
| 49 | 9785202960035239728 | const PetscScalar *d_px; | |
| 50 | |||
| 51 | PetscFunctionBegin; | ||
| 52 |
3/4✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 2 times.
✓ Branch 3 taken 2 times.
|
9785202960035239728 | PetscCall(MatDenseHIPGetArrayRead(x->A,&d_px)); |
| 53 |
4/6✓ Branch 0 taken 2 times.
✓ Branch 1 taken 2 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
|
14108644570652714544 | if (beta==(PetscScalar)0.0) PetscCall(VecHIPGetArrayWrite(y,&d_py)); |
| 54 |
3/4✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 1 times.
✓ Branch 3 taken 2 times.
|
6254930062798203460 | else PetscCall(VecHIPGetArray(y,&d_py)); |
| 55 |
5/6✓ Branch 0 taken 2 times.
✓ Branch 1 taken 2 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✓ Branch 4 taken 1 times.
✓ Branch 5 taken 2 times.
|
17020236053732154576 | if (!q) PetscCall(VecHIPGetArray(X->buffer,&d_q)); |
| 56 | else { | ||
| 57 | 10765305877385753228 | PetscInt k=X->k-X->l; | |
| 58 |
3/8✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✗ Branch 5 not taken.
✓ Branch 6 taken 1 times.
✓ Branch 7 taken 2 times.
|
10765305877385753228 | PetscCallHIP(hipMalloc((void**)&d_q,k*sizeof(PetscScalar))); |
| 59 |
2/8✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✗ Branch 5 not taken.
✗ Branch 6 not taken.
✓ Branch 7 taken 2 times.
|
15088747719931462028 | PetscCallHIP(hipMemcpy(d_q,q,k*sizeof(PetscScalar),hipMemcpyHostToDevice)); |
| 60 |
2/4✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✗ Branch 2 not taken.
✓ Branch 3 taken 2 times.
|
15088747719931462028 | PetscCall(PetscLogCpuToGpu(k*sizeof(PetscScalar))); |
| 61 |
2/2✓ Branch 0 taken 1 times.
✓ Branch 1 taken 2 times.
|
19412189562477170828 | } |
| 62 |
2/4✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✗ Branch 2 not taken.
✓ Branch 3 taken 2 times.
|
17020236053732154576 | PetscCall(BVMultVec_BLAS_HIP(X,X->n,X->k-X->l,alpha,d_px+(X->nc+X->l)*X->ld,X->ld,d_q,beta,d_py)); |
| 63 |
3/4✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 2 times.
✓ Branch 3 taken 2 times.
|
9785202960035239728 | PetscCall(MatDenseHIPRestoreArrayRead(x->A,&d_px)); |
| 64 |
4/6✓ Branch 0 taken 2 times.
✓ Branch 1 taken 2 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
|
14108644570652714544 | if (beta==(PetscScalar)0.0) PetscCall(VecHIPRestoreArrayWrite(y,&d_py)); |
| 65 |
3/4✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 1 times.
✓ Branch 3 taken 2 times.
|
6254930062798203460 | else PetscCall(VecHIPRestoreArray(y,&d_py)); |
| 66 |
5/6✓ Branch 0 taken 2 times.
✓ Branch 1 taken 2 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✓ Branch 4 taken 1 times.
✓ Branch 5 taken 2 times.
|
17020236053732154576 | if (!q) PetscCall(VecHIPRestoreArray(X->buffer,&d_q)); |
| 67 |
3/8✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✗ Branch 5 not taken.
✓ Branch 6 taken 1 times.
✓ Branch 7 taken 2 times.
|
10765305877385753228 | else PetscCallHIP(hipFree(d_q)); |
| 68 | 17020236053732154576 | PetscFunctionReturn(PETSC_SUCCESS); | |
| 69 | 3502570738953149040 | } | |
| 70 | |||
| 71 | 8711087579241381888 | PetscErrorCode BVMultInPlace_Mat_HIP(BV V,Mat Q,PetscInt s,PetscInt e) | |
| 72 | { | ||
| 73 | 8711087579241381888 | BV_MAT *ctx = (BV_MAT*)V->data; | |
| 74 | 8711087579241381888 | PetscScalar *d_pv; | |
| 75 | 8711087579241381888 | const PetscScalar *d_q; | |
| 76 | 8711087579241381888 | PetscInt ldq; | |
| 77 | |||
| 78 | PetscFunctionBegin; | ||
| 79 |
2/4✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✗ Branch 2 not taken.
✓ Branch 3 taken 2 times.
|
8711087579241381888 | if (s>=e || !V->n) PetscFunctionReturn(PETSC_SUCCESS); |
| 80 |
2/4✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✗ Branch 2 not taken.
✓ Branch 3 taken 2 times.
|
8711087579241381888 | PetscCall(MatDenseGetLDA(Q,&ldq)); |
| 81 |
2/4✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✗ Branch 2 not taken.
✓ Branch 3 taken 2 times.
|
8711087579241381888 | PetscCall(MatDenseHIPGetArray(ctx->A,&d_pv)); |
| 82 |
2/4✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✗ Branch 2 not taken.
✓ Branch 3 taken 2 times.
|
8711087579241381888 | PetscCall(BV_MatDenseHIPGetArrayRead(V,Q,&d_q)); |
| 83 |
2/4✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✗ Branch 2 not taken.
✓ Branch 3 taken 2 times.
|
8711087579241381888 | PetscCall(BVMultInPlace_BLAS_HIP(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 |
2/4✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✗ Branch 2 not taken.
✓ Branch 3 taken 2 times.
|
8711087579241381888 | PetscCall(BV_MatDenseHIPRestoreArrayRead(V,Q,&d_q)); |
| 85 |
2/4✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✗ Branch 2 not taken.
✓ Branch 3 taken 2 times.
|
8711087579241381888 | PetscCall(MatDenseHIPRestoreArray(ctx->A,&d_pv)); |
| 86 | 8711087579241381888 | PetscFunctionReturn(PETSC_SUCCESS); | |
| 87 | 8711087579241381888 | } | |
| 88 | |||
| 89 | ✗ | PetscErrorCode BVMultInPlaceHermitianTranspose_Mat_HIP(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; | |
| 95 | |||
| 96 | PetscFunctionBegin; | ||
| 97 | ✗ | if (s>=e || !V->n) PetscFunctionReturn(PETSC_SUCCESS); | |
| 98 | ✗ | PetscCall(MatDenseGetLDA(Q,&ldq)); | |
| 99 | ✗ | PetscCall(MatDenseHIPGetArray(ctx->A,&d_pv)); | |
| 100 | ✗ | PetscCall(BV_MatDenseHIPGetArrayRead(V,Q,&d_q)); | |
| 101 | ✗ | PetscCall(BVMultInPlace_BLAS_HIP(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_MatDenseHIPRestoreArrayRead(V,Q,&d_q)); | |
| 103 | ✗ | PetscCall(MatDenseHIPRestoreArray(ctx->A,&d_pv)); | |
| 104 | ✗ | PetscFunctionReturn(PETSC_SUCCESS); | |
| 105 | ✗ | } | |
| 106 | |||
| 107 | ✗ | PetscErrorCode BVDot_Mat_HIP(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; | |
| 113 | |||
| 114 | PetscFunctionBegin; | ||
| 115 | ✗ | PetscCall(MatDenseGetLDA(M,&ldm)); | |
| 116 | ✗ | PetscCall(MatDenseHIPGetArrayRead(x->A,&d_px)); | |
| 117 | ✗ | PetscCall(MatDenseHIPGetArrayRead(y->A,&d_py)); | |
| 118 | ✗ | PetscCall(MatDenseGetArrayWrite(M,&pm)); | |
| 119 | ✗ | PetscCall(BVDot_BLAS_HIP(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(MatDenseHIPRestoreArrayRead(x->A,&d_px)); | |
| 122 | ✗ | PetscCall(MatDenseHIPRestoreArrayRead(y->A,&d_py)); | |
| 123 | ✗ | PetscFunctionReturn(PETSC_SUCCESS); | |
| 124 | ✗ | } | |
| 125 | |||
| 126 | 6254930176346401348 | PetscErrorCode BVDotVec_Mat_HIP(BV X,Vec y,PetscScalar *q) | |
| 127 | { | ||
| 128 | 6254930176346401348 | BV_MAT *x = (BV_MAT*)X->data; | |
| 129 | 6254930176346401348 | const PetscScalar *d_px,*d_py; | |
| 130 | 6254930176346401348 | Vec z = y; | |
| 131 | |||
| 132 | PetscFunctionBegin; | ||
| 133 |
1/2✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
|
6254930176346401348 | if (PetscUnlikely(X->matrix)) { |
| 134 | ✗ | PetscCall(BV_IPMatMult(X,y)); | |
| 135 | ✗ | z = X->Bx; | |
| 136 | ✗ | } | |
| 137 |
3/4✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 1 times.
✓ Branch 3 taken 2 times.
|
6254930176346401348 | PetscCall(MatDenseHIPGetArrayRead(x->A,&d_px)); |
| 138 |
3/4✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 1 times.
✓ Branch 3 taken 2 times.
|
6254930176346401348 | PetscCall(VecHIPGetArrayRead(z,&d_py)); |
| 139 |
3/4✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 1 times.
✓ Branch 3 taken 2 times.
|
6254930176346401348 | PetscCall(BVDotVec_BLAS_HIP(X,X->n,X->k-X->l,d_px+(X->nc+X->l)*X->ld,X->ld,d_py,q,x->mpi)); |
| 140 |
3/4✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 1 times.
✓ Branch 3 taken 2 times.
|
6254930176346401348 | PetscCall(VecHIPRestoreArrayRead(z,&d_py)); |
| 141 |
3/4✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 1 times.
✓ Branch 3 taken 2 times.
|
6254930176346401348 | PetscCall(MatDenseHIPRestoreArrayRead(x->A,&d_px)); |
| 142 | 6254930176346401348 | PetscFunctionReturn(PETSC_SUCCESS); | |
| 143 | 5101953472819684932 | } | |
| 144 | |||
| 145 | ✗ | PetscErrorCode BVDotVec_Local_Mat_HIP(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; | |
| 150 | |||
| 151 | PetscFunctionBegin; | ||
| 152 | ✗ | if (PetscUnlikely(X->matrix)) { | |
| 153 | ✗ | PetscCall(BV_IPMatMult(X,y)); | |
| 154 | ✗ | z = X->Bx; | |
| 155 | ✗ | } | |
| 156 | ✗ | PetscCall(MatDenseHIPGetArrayRead(x->A,&d_px)); | |
| 157 | ✗ | PetscCall(VecHIPGetArrayRead(z,&d_py)); | |
| 158 | ✗ | PetscCall(BVDotVec_BLAS_HIP(X,X->n,X->k-X->l,d_px+(X->nc+X->l)*X->ld,X->ld,d_py,m,PETSC_FALSE)); | |
| 159 | ✗ | PetscCall(VecHIPRestoreArrayRead(z,&d_py)); | |
| 160 | ✗ | PetscCall(MatDenseHIPRestoreArrayRead(x->A,&d_px)); | |
| 161 | ✗ | PetscFunctionReturn(PETSC_SUCCESS); | |
| 162 | ✗ | } | |
| 163 | |||
| 164 | 13983673232394276876 | PetscErrorCode BVScale_Mat_HIP(BV bv,PetscInt j,PetscScalar alpha) | |
| 165 | { | ||
| 166 | 13983673232394276876 | BV_MAT *ctx = (BV_MAT*)bv->data; | |
| 167 | 13983673232394276876 | PetscScalar *d_array,*d_A; | |
| 168 | 13983673232394276876 | PetscInt n=0; | |
| 169 | |||
| 170 | PetscFunctionBegin; | ||
| 171 |
1/2✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
|
13983673232394276876 | if (!bv->n) PetscFunctionReturn(PETSC_SUCCESS); |
| 172 |
3/4✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 2 times.
✓ Branch 3 taken 2 times.
|
13983673232394276876 | PetscCall(MatDenseHIPGetArray(ctx->A,&d_array)); |
| 173 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 2 times.
|
13983673232394276876 | 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 | 13983673232394276876 | d_A = d_array+(bv->nc+j)*bv->ld; | |
| 178 | 13983673232394276876 | n = bv->n; | |
| 179 | } | ||
| 180 |
2/4✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✗ Branch 2 not taken.
✓ Branch 3 taken 2 times.
|
22909814915024826356 | PetscCall(BVScale_BLAS_HIP(bv,n,d_A,alpha)); |
| 181 |
3/4✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 2 times.
✓ Branch 3 taken 2 times.
|
13983673232394276876 | PetscCall(MatDenseHIPRestoreArray(ctx->A,&d_array)); |
| 182 | 13983673232394276876 | PetscFunctionReturn(PETSC_SUCCESS); | |
| 183 | 22315354206576373700 | } | |
| 184 | |||
| 185 | 396 | PetscErrorCode BVNorm_Mat_HIP(BV bv,PetscInt j,NormType type,PetscReal *val) | |
| 186 | { | ||
| 187 | 396 | BV_MAT *ctx = (BV_MAT*)bv->data; | |
| 188 | 396 | const PetscScalar *array,*d_array,*d_A; | |
| 189 | 396 | PetscInt n=0; | |
| 190 | |||
| 191 | PetscFunctionBegin; | ||
| 192 |
4/10✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✗ Branch 2 not taken.
✓ Branch 3 taken 2 times.
✗ Branch 4 not taken.
✗ Branch 5 not taken.
✓ Branch 6 taken 2 times.
✗ Branch 7 not taken.
✓ Branch 8 taken 2 times.
✗ Branch 9 not taken.
|
396 | if (!ctx->mpi && ((j<0 && type==NORM_FROBENIUS && bv->ld==bv->n) || (j>=0 && type==NORM_2))) { |
| 193 | /* compute on GPU with hipBLAS - TODO: include the MPI case here */ | ||
| 194 | 396 | *val = 0.0; | |
| 195 |
1/2✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
|
396 | if (!bv->n) PetscFunctionReturn(PETSC_SUCCESS); |
| 196 |
2/4✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✗ Branch 2 not taken.
✓ Branch 3 taken 2 times.
|
396 | PetscCall(MatDenseHIPGetArrayRead(ctx->A,&d_array)); |
| 197 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 2 times.
|
396 | if (PetscUnlikely(j<0)) { |
| 198 | ✗ | d_A = d_array+(bv->nc+bv->l)*bv->ld; | |
| 199 | ✗ | n = (bv->k-bv->l)*bv->ld; | |
| 200 | ✗ | } else { | |
| 201 | 396 | d_A = d_array+(bv->nc+j)*bv->ld; | |
| 202 | 396 | n = bv->n; | |
| 203 | } | ||
| 204 |
2/4✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✗ Branch 2 not taken.
✓ Branch 3 taken 2 times.
|
396 | PetscCall(BVNorm_BLAS_HIP(bv,n,d_A,val)); |
| 205 |
2/4✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✗ Branch 2 not taken.
✓ Branch 3 taken 2 times.
|
396 | PetscCall(MatDenseHIPRestoreArrayRead(ctx->A,&d_array)); |
| 206 | 396 | } else { | |
| 207 | /* compute on CPU */ | ||
| 208 | ✗ | PetscCall(MatDenseGetArrayRead(ctx->A,&array)); | |
| 209 | ✗ | if (PetscUnlikely(j<0)) PetscCall(BVNorm_LAPACK_Private(bv,bv->n,bv->k-bv->l,array+(bv->nc+bv->l)*bv->ld,bv->ld,type,val,ctx->mpi)); | |
| 210 | ✗ | else PetscCall(BVNorm_LAPACK_Private(bv,bv->n,1,array+(bv->nc+j)*bv->ld,bv->ld,type,val,ctx->mpi)); | |
| 211 | ✗ | PetscCall(MatDenseRestoreArrayRead(ctx->A,&array)); | |
| 212 | } | ||
| 213 | 396 | PetscFunctionReturn(PETSC_SUCCESS); | |
| 214 | 396 | } | |
| 215 | |||
| 216 | ✗ | PetscErrorCode BVNorm_Local_Mat_HIP(BV bv,PetscInt j,NormType type,PetscReal *val) | |
| 217 | { | ||
| 218 | ✗ | BV_MAT *ctx = (BV_MAT*)bv->data; | |
| 219 | ✗ | const PetscScalar *array,*d_array,*d_A; | |
| 220 | ✗ | PetscInt n=0; | |
| 221 | |||
| 222 | PetscFunctionBegin; | ||
| 223 | ✗ | if ((j<0 && type==NORM_FROBENIUS && bv->ld==bv->n) || (j>=0 && type==NORM_2)) { | |
| 224 | /* compute on GPU with hipBLAS */ | ||
| 225 | ✗ | *val = 0.0; | |
| 226 | ✗ | if (!bv->n) PetscFunctionReturn(PETSC_SUCCESS); | |
| 227 | ✗ | PetscCall(MatDenseHIPGetArrayRead(ctx->A,&d_array)); | |
| 228 | ✗ | if (PetscUnlikely(j<0)) { | |
| 229 | ✗ | d_A = d_array+(bv->nc+bv->l)*bv->ld; | |
| 230 | ✗ | n = (bv->k-bv->l)*bv->ld; | |
| 231 | ✗ | } else { | |
| 232 | ✗ | d_A = d_array+(bv->nc+j)*bv->ld; | |
| 233 | ✗ | n = bv->n; | |
| 234 | } | ||
| 235 | ✗ | PetscCall(BVNorm_BLAS_HIP(bv,n,d_A,val)); | |
| 236 | ✗ | PetscCall(MatDenseHIPRestoreArrayRead(ctx->A,&d_array)); | |
| 237 | ✗ | } else { | |
| 238 | /* compute on CPU */ | ||
| 239 | ✗ | PetscCall(MatDenseGetArrayRead(ctx->A,&array)); | |
| 240 | ✗ | if (PetscUnlikely(j<0)) PetscCall(BVNorm_LAPACK_Private(bv,bv->n,bv->k-bv->l,array+(bv->nc+bv->l)*bv->ld,bv->ld,type,val,PETSC_FALSE)); | |
| 241 | ✗ | else PetscCall(BVNorm_LAPACK_Private(bv,bv->n,1,array+(bv->nc+j)*bv->ld,bv->ld,type,val,PETSC_FALSE)); | |
| 242 | ✗ | PetscCall(MatDenseRestoreArrayRead(ctx->A,&array)); | |
| 243 | } | ||
| 244 | ✗ | PetscFunctionReturn(PETSC_SUCCESS); | |
| 245 | ✗ | } | |
| 246 | |||
| 247 | 6789457913237733376 | PetscErrorCode BVNormalize_Mat_HIP(BV bv,PetscScalar *eigi) | |
| 248 | { | ||
| 249 | 6789457913237733376 | BV_MAT *ctx = (BV_MAT*)bv->data; | |
| 250 | 6789457913237733376 | PetscScalar *array,*d_array,*wi=NULL; | |
| 251 | |||
| 252 | PetscFunctionBegin; | ||
| 253 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 2 times.
|
6789457913237733376 | if (eigi) wi = eigi+bv->l; |
| 254 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 2 times.
|
6789457913237733376 | if (!ctx->mpi) { |
| 255 | /* compute on GPU with hipBLAS - TODO: include the MPI case here */ | ||
| 256 |
1/2✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
|
6789457913237733376 | if (!bv->n) PetscFunctionReturn(PETSC_SUCCESS); |
| 257 |
2/4✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✗ Branch 2 not taken.
✓ Branch 3 taken 2 times.
|
6789457913237733376 | PetscCall(MatDenseHIPGetArray(ctx->A,&d_array)); |
| 258 |
2/4✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✗ Branch 2 not taken.
✓ Branch 3 taken 2 times.
|
6789457913237733376 | PetscCall(BVNormalize_BLAS_HIP(bv,bv->n,bv->k-bv->l,d_array+(bv->nc+bv->l)*bv->ld,bv->ld,wi)); |
| 259 |
2/4✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✗ Branch 2 not taken.
✓ Branch 3 taken 2 times.
|
6789457913237733376 | PetscCall(MatDenseHIPRestoreArray(ctx->A,&d_array)); |
| 260 | 6789457913237733376 | } else { | |
| 261 | /* compute on CPU */ | ||
| 262 | ✗ | PetscCall(MatDenseGetArray(ctx->A,&array)); | |
| 263 | ✗ | PetscCall(BVNormalize_LAPACK_Private(bv,bv->n,bv->k-bv->l,array+(bv->nc+bv->l)*bv->ld,bv->ld,wi,ctx->mpi)); | |
| 264 | ✗ | PetscCall(MatDenseRestoreArray(ctx->A,&array)); | |
| 265 | } | ||
| 266 | 6789457913237733376 | PetscFunctionReturn(PETSC_SUCCESS); | |
| 267 | 6789457913237733376 | } | |
| 268 | |||
| 269 | ✗ | PetscErrorCode BVMatMult_Mat_HIP(BV V,Mat A,BV W) | |
| 270 | { | ||
| 271 | ✗ | BV_MAT *v = (BV_MAT*)V->data,*w = (BV_MAT*)W->data; | |
| 272 | ✗ | Mat Vmat,Wmat; | |
| 273 | ✗ | const PetscScalar *d_pv; | |
| 274 | ✗ | PetscScalar *d_pw; | |
| 275 | ✗ | PetscInt j; | |
| 276 | |||
| 277 | PetscFunctionBegin; | ||
| 278 | ✗ | if (V->vmm) { | |
| 279 | ✗ | PetscCall(BVGetMat(V,&Vmat)); | |
| 280 | ✗ | PetscCall(BVGetMat(W,&Wmat)); | |
| 281 | ✗ | PetscCall(MatProductCreateWithMat(A,Vmat,NULL,Wmat)); | |
| 282 | ✗ | PetscCall(MatProductSetType(Wmat,MATPRODUCT_AB)); | |
| 283 | ✗ | PetscCall(MatProductSetFromOptions(Wmat)); | |
| 284 | ✗ | PetscCall(MatProductSymbolic(Wmat)); | |
| 285 | ✗ | PetscCall(MatProductNumeric(Wmat)); | |
| 286 | ✗ | PetscCall(MatProductClear(Wmat)); | |
| 287 | ✗ | PetscCall(BVRestoreMat(V,&Vmat)); | |
| 288 | ✗ | PetscCall(BVRestoreMat(W,&Wmat)); | |
| 289 | ✗ | } else { | |
| 290 | ✗ | PetscCall(MatDenseHIPGetArrayRead(v->A,&d_pv)); | |
| 291 | ✗ | PetscCall(MatDenseHIPGetArrayWrite(w->A,&d_pw)); | |
| 292 | ✗ | for (j=0;j<V->k-V->l;j++) { | |
| 293 | ✗ | PetscCall(VecHIPPlaceArray(V->cv[1],(PetscScalar *)d_pv+(V->nc+V->l+j)*V->ld)); | |
| 294 | ✗ | PetscCall(VecHIPPlaceArray(W->cv[1],d_pw+(W->nc+W->l+j)*W->ld)); | |
| 295 | ✗ | PetscCall(MatMult(A,V->cv[1],W->cv[1])); | |
| 296 | ✗ | PetscCall(VecHIPResetArray(V->cv[1])); | |
| 297 | ✗ | PetscCall(VecHIPResetArray(W->cv[1])); | |
| 298 | ✗ | } | |
| 299 | ✗ | PetscCall(MatDenseHIPRestoreArrayRead(v->A,&d_pv)); | |
| 300 | ✗ | PetscCall(MatDenseHIPRestoreArrayWrite(w->A,&d_pw)); | |
| 301 | } | ||
| 302 | ✗ | PetscFunctionReturn(PETSC_SUCCESS); | |
| 303 | ✗ | } | |
| 304 | |||
| 305 | ✗ | PetscErrorCode BVCopy_Mat_HIP(BV V,BV W) | |
| 306 | { | ||
| 307 | ✗ | BV_MAT *v = (BV_MAT*)V->data,*w = (BV_MAT*)W->data; | |
| 308 | ✗ | const PetscScalar *d_pv; | |
| 309 | ✗ | PetscScalar *d_pw; | |
| 310 | |||
| 311 | PetscFunctionBegin; | ||
| 312 | ✗ | PetscCall(MatDenseHIPGetArrayRead(v->A,&d_pv)); | |
| 313 | ✗ | PetscCall(MatDenseHIPGetArray(w->A,&d_pw)); | |
| 314 | ✗ | PetscCallHIP(hipMemcpy2D(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,hipMemcpyDeviceToDevice)); | |
| 315 | ✗ | PetscCall(MatDenseHIPRestoreArrayRead(v->A,&d_pv)); | |
| 316 | ✗ | PetscCall(MatDenseHIPRestoreArray(w->A,&d_pw)); | |
| 317 | ✗ | PetscFunctionReturn(PETSC_SUCCESS); | |
| 318 | ✗ | } | |
| 319 | |||
| 320 | 59190018048 | PetscErrorCode BVCopyColumn_Mat_HIP(BV V,PetscInt j,PetscInt i) | |
| 321 | { | ||
| 322 | 59190018048 | BV_MAT *v = (BV_MAT*)V->data; | |
| 323 | 59190018048 | PetscScalar *d_pv; | |
| 324 | |||
| 325 | PetscFunctionBegin; | ||
| 326 |
2/4✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✗ Branch 2 not taken.
✓ Branch 3 taken 2 times.
|
59190018048 | PetscCall(MatDenseHIPGetArray(v->A,&d_pv)); |
| 327 |
2/8✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✗ Branch 5 not taken.
✗ Branch 6 not taken.
✓ Branch 7 taken 2 times.
|
59190018048 | PetscCallHIP(hipMemcpy(d_pv+(V->nc+i)*V->ld,d_pv+(V->nc+j)*V->ld,V->n*sizeof(PetscScalar),hipMemcpyDeviceToDevice)); |
| 328 |
2/4✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✗ Branch 2 not taken.
✓ Branch 3 taken 2 times.
|
59190018048 | PetscCall(MatDenseHIPRestoreArray(v->A,&d_pv)); |
| 329 | 59190018048 | PetscFunctionReturn(PETSC_SUCCESS); | |
| 330 | 59190018048 | } | |
| 331 | |||
| 332 | 3889970031726905324 | PetscErrorCode BVGetColumn_Mat_HIP(BV bv,PetscInt j,Vec*) | |
| 333 | { | ||
| 334 | 3889970031726905324 | BV_MAT *ctx = (BV_MAT*)bv->data; | |
| 335 | 3889970031726905324 | PetscScalar *d_pv; | |
| 336 | 3889970031726905324 | PetscInt l; | |
| 337 | |||
| 338 | PetscFunctionBegin; | ||
| 339 |
2/2✓ Branch 0 taken 2 times.
✓ Branch 1 taken 2 times.
|
3889970031726905324 | l = BVAvailableVec; |
| 340 |
2/4✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✗ Branch 2 not taken.
✓ Branch 3 taken 2 times.
|
33003525218494388220 | PetscCall(MatDenseHIPGetArray(ctx->A,&d_pv)); |
| 341 |
3/4✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 2 times.
✓ Branch 3 taken 2 times.
|
3889962928924715012 | PetscCall(VecHIPPlaceArray(bv->cv[l],d_pv+(bv->nc+j)*bv->ld)); |
| 342 | 3889962928924715012 | PetscFunctionReturn(PETSC_SUCCESS); | |
| 343 | 3889962928924715012 | } | |
| 344 | |||
| 345 | 3889962928924715012 | PetscErrorCode BVRestoreColumn_Mat_HIP(BV bv,PetscInt j,Vec*) | |
| 346 | { | ||
| 347 | 3889962928924715012 | BV_MAT *ctx = (BV_MAT*)bv->data; | |
| 348 | 3889962928924715012 | PetscInt l; | |
| 349 | |||
| 350 | PetscFunctionBegin; | ||
| 351 | 3889962928924715012 | l = (j==bv->ci[0])? 0: 1; | |
| 352 |
3/4✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 2 times.
✓ Branch 3 taken 2 times.
|
3889962928924715012 | PetscCall(VecHIPResetArray(bv->cv[l])); |
| 353 |
3/4✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 2 times.
✓ Branch 3 taken 2 times.
|
3889962928924715012 | PetscCall(MatDenseHIPRestoreArray(ctx->A,NULL)); |
| 354 | 3889962928924715012 | PetscFunctionReturn(PETSC_SUCCESS); | |
| 355 | 11669888786774145036 | } | |
| 356 | |||
| 357 | ✗ | PetscErrorCode BVRestoreSplit_Mat_HIP(BV bv,BV *L,BV *R) | |
| 358 | { | ||
| 359 | ✗ | Mat A; | |
| 360 | ✗ | const PetscScalar *d_pv; | |
| 361 | ✗ | PetscObjectState lstate,rstate; | |
| 362 | ✗ | PetscBool change=PETSC_FALSE; | |
| 363 | |||
| 364 | PetscFunctionBegin; | ||
| 365 | /* force sync flag to PETSC_OFFLOAD_BOTH */ | ||
| 366 | ✗ | if (L) { | |
| 367 | ✗ | PetscCall(PetscObjectStateGet((PetscObject)*L,&lstate)); | |
| 368 | ✗ | if (lstate != bv->lstate) { | |
| 369 | ✗ | A = ((BV_MAT*)bv->L->data)->A; | |
| 370 | ✗ | PetscCall(MatDenseHIPGetArrayRead(A,&d_pv)); | |
| 371 | ✗ | PetscCall(MatDenseHIPRestoreArrayRead(A,&d_pv)); | |
| 372 | ✗ | change = PETSC_TRUE; | |
| 373 | ✗ | } | |
| 374 | ✗ | } | |
| 375 | ✗ | if (R) { | |
| 376 | ✗ | PetscCall(PetscObjectStateGet((PetscObject)*R,&rstate)); | |
| 377 | ✗ | if (rstate != bv->rstate) { | |
| 378 | ✗ | A = ((BV_MAT*)bv->R->data)->A; | |
| 379 | ✗ | PetscCall(MatDenseHIPGetArrayRead(A,&d_pv)); | |
| 380 | ✗ | PetscCall(MatDenseHIPRestoreArrayRead(A,&d_pv)); | |
| 381 | ✗ | change = PETSC_TRUE; | |
| 382 | ✗ | } | |
| 383 | ✗ | } | |
| 384 | ✗ | if (change) { | |
| 385 | ✗ | A = ((BV_MAT*)bv->data)->A; | |
| 386 | ✗ | PetscCall(MatDenseHIPGetArray(A,(PetscScalar **)&d_pv)); | |
| 387 | ✗ | PetscCall(MatDenseHIPRestoreArray(A,(PetscScalar **)&d_pv)); | |
| 388 | ✗ | } | |
| 389 | ✗ | PetscFunctionReturn(PETSC_SUCCESS); | |
| 390 | ✗ | } | |
| 391 | |||
| 392 | ✗ | PetscErrorCode BVRestoreSplitRows_Mat_HIP(BV bv,IS,IS,BV *U,BV *L) | |
| 393 | { | ||
| 394 | ✗ | Mat A; | |
| 395 | ✗ | const PetscScalar *d_pv; | |
| 396 | ✗ | PetscObjectState lstate,rstate; | |
| 397 | ✗ | PetscBool change=PETSC_FALSE; | |
| 398 | |||
| 399 | PetscFunctionBegin; | ||
| 400 | /* force sync flag to PETSC_OFFLOAD_BOTH */ | ||
| 401 | ✗ | if (U) { | |
| 402 | ✗ | PetscCall(PetscObjectStateGet((PetscObject)*U,&rstate)); | |
| 403 | ✗ | if (rstate != bv->rstate) { | |
| 404 | ✗ | A = ((BV_MAT*)bv->R->data)->A; | |
| 405 | ✗ | PetscCall(MatDenseHIPGetArrayRead(A,&d_pv)); | |
| 406 | ✗ | PetscCall(MatDenseHIPRestoreArrayRead(A,&d_pv)); | |
| 407 | ✗ | change = PETSC_TRUE; | |
| 408 | ✗ | } | |
| 409 | ✗ | } | |
| 410 | ✗ | if (L) { | |
| 411 | ✗ | PetscCall(PetscObjectStateGet((PetscObject)*L,&lstate)); | |
| 412 | ✗ | if (lstate != bv->lstate) { | |
| 413 | ✗ | A = ((BV_MAT*)bv->L->data)->A; | |
| 414 | ✗ | PetscCall(MatDenseHIPGetArrayRead(A,&d_pv)); | |
| 415 | ✗ | PetscCall(MatDenseHIPRestoreArrayRead(A,&d_pv)); | |
| 416 | ✗ | change = PETSC_TRUE; | |
| 417 | ✗ | } | |
| 418 | ✗ | } | |
| 419 | ✗ | if (change) { | |
| 420 | ✗ | A = ((BV_MAT*)bv->data)->A; | |
| 421 | ✗ | PetscCall(MatDenseHIPGetArray(A,(PetscScalar **)&d_pv)); | |
| 422 | ✗ | PetscCall(MatDenseHIPRestoreArray(A,(PetscScalar **)&d_pv)); | |
| 423 | ✗ | } | |
| 424 | ✗ | PetscFunctionReturn(PETSC_SUCCESS); | |
| 425 | ✗ | } | |
| 426 | |||
| 427 | ✗ | PetscErrorCode BVGetMat_Mat_HIP(BV bv,Mat *A) | |
| 428 | { | ||
| 429 | ✗ | BV_MAT *ctx = (BV_MAT*)bv->data; | |
| 430 | ✗ | PetscScalar *vv,*aa; | |
| 431 | ✗ | PetscBool create=PETSC_FALSE; | |
| 432 | ✗ | PetscInt m,cols; | |
| 433 | |||
| 434 | PetscFunctionBegin; | ||
| 435 | ✗ | m = bv->k-bv->l; | |
| 436 | ✗ | if (!bv->Aget) create=PETSC_TRUE; | |
| 437 | else { | ||
| 438 | ✗ | PetscCall(MatDenseHIPGetArray(bv->Aget,&aa)); | |
| 439 | ✗ | PetscCheck(!aa,PetscObjectComm((PetscObject)bv),PETSC_ERR_ARG_WRONGSTATE,"BVGetMat already called on this BV"); | |
| 440 | ✗ | PetscCall(MatGetSize(bv->Aget,NULL,&cols)); | |
| 441 | ✗ | if (cols!=m) { | |
| 442 | ✗ | PetscCall(MatDestroy(&bv->Aget)); | |
| 443 | ✗ | create=PETSC_TRUE; | |
| 444 | ✗ | } | |
| 445 | } | ||
| 446 | ✗ | PetscCall(MatDenseHIPGetArray(ctx->A,&vv)); | |
| 447 | ✗ | if (create) { | |
| 448 | ✗ | PetscCall(MatCreateDenseFromVecType(PetscObjectComm((PetscObject)bv),bv->vtype,bv->n,PETSC_DECIDE,bv->N,m,bv->ld,vv,&bv->Aget)); /* pass a pointer to avoid allocation of storage */ | |
| 449 | ✗ | PetscCall(MatDenseHIPReplaceArray(bv->Aget,NULL)); /* replace with a null pointer, the value after BVRestoreMat */ | |
| 450 | ✗ | } | |
| 451 | ✗ | PetscCall(MatDenseHIPPlaceArray(bv->Aget,vv+(bv->nc+bv->l)*bv->ld)); /* set the actual pointer */ | |
| 452 | ✗ | *A = bv->Aget; | |
| 453 | ✗ | PetscFunctionReturn(PETSC_SUCCESS); | |
| 454 | ✗ | } | |
| 455 | |||
| 456 | ✗ | PetscErrorCode BVRestoreMat_Mat_HIP(BV bv,Mat *A) | |
| 457 | { | ||
| 458 | ✗ | BV_MAT *ctx = (BV_MAT*)bv->data; | |
| 459 | ✗ | PetscScalar *vv,*aa; | |
| 460 | |||
| 461 | PetscFunctionBegin; | ||
| 462 | ✗ | PetscCall(MatDenseHIPGetArray(bv->Aget,&aa)); | |
| 463 | ✗ | vv = aa-(bv->nc+bv->l)*bv->ld; | |
| 464 | ✗ | PetscCall(MatDenseHIPResetArray(bv->Aget)); | |
| 465 | ✗ | PetscCall(MatDenseHIPRestoreArray(ctx->A,&vv)); | |
| 466 | ✗ | *A = NULL; | |
| 467 | ✗ | PetscFunctionReturn(PETSC_SUCCESS); | |
| 468 | ✗ | } | |
| 469 |