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 private kernels that use the LAPACK | ||
12 | */ | ||
13 | |||
14 | #include <slepc/private/bvimpl.h> | ||
15 | #include <slepcblaslapack.h> | ||
16 | |||
17 | /* | ||
18 | Reduction operation to compute sqrt(x**2+y**2) when normalizing vectors | ||
19 | */ | ||
20 | 15246 | SLEPC_EXTERN void MPIAPI SlepcPythag(void *in,void *inout,PetscMPIInt *len,MPI_Datatype *datatype) | |
21 | { | ||
22 | 15246 | PetscBLASInt i,n=*len; | |
23 | 15246 | PetscReal *x = (PetscReal*)in,*y = (PetscReal*)inout; | |
24 | |||
25 |
1/2✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
|
15246 | PetscFunctionBegin; |
26 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 10 times.
|
15246 | if (PetscUnlikely(*datatype!=MPIU_REAL)) { |
27 | ✗ | (void)(*PetscErrorPrintf)("Only implemented for MPIU_REAL data type"); | |
28 | ✗ | MPI_Abort(PETSC_COMM_WORLD,1); | |
29 | } | ||
30 |
2/2✓ Branch 0 taken 10 times.
✓ Branch 1 taken 10 times.
|
33222 | for (i=0;i<n;i++) y[i] = SlepcAbs(x[i],y[i]); |
31 |
6/12✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✗ Branch 2 not taken.
✓ Branch 3 taken 2 times.
✓ Branch 4 taken 2 times.
✗ Branch 5 not taken.
✓ Branch 6 taken 2 times.
✗ Branch 7 not taken.
✓ Branch 8 taken 2 times.
✗ Branch 9 not taken.
✗ Branch 10 not taken.
✓ Branch 11 taken 2 times.
|
15246 | PetscFunctionReturnVoid(); |
32 | } | ||
33 | |||
34 | /* | ||
35 | Compute ||A|| for an mxn matrix | ||
36 | */ | ||
37 | 102569 | PetscErrorCode BVNorm_LAPACK_Private(BV bv,PetscInt m_,PetscInt n_,const PetscScalar *A,PetscInt lda_,NormType type,PetscReal *nrm,PetscBool mpi) | |
38 | { | ||
39 | 102569 | PetscBLASInt m,n,lda,i,j; | |
40 | 102569 | PetscMPIInt len; | |
41 | 102569 | PetscReal lnrm,*rwork=NULL,*rwork2=NULL; | |
42 | |||
43 |
1/2✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
|
102569 | PetscFunctionBegin; |
44 |
4/6✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
|
102569 | PetscCall(PetscFPTrapPush(PETSC_FP_TRAP_OFF)); |
45 |
4/6✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
|
102569 | PetscCall(PetscBLASIntCast(m_,&m)); |
46 |
4/6✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
|
102569 | PetscCall(PetscBLASIntCast(n_,&n)); |
47 |
4/6✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
|
102569 | PetscCall(PetscBLASIntCast(lda_,&lda)); |
48 |
2/2✓ Branch 0 taken 10 times.
✓ Branch 1 taken 10 times.
|
102569 | if (type==NORM_FROBENIUS || type==NORM_2) { |
49 | 102465 | lnrm = LAPACKlange_("F",&m,&n,(PetscScalar*)A,&lda,rwork); | |
50 |
30/60✓ Branch 0 taken 10 times.
✓ Branch 1 taken 10 times.
✓ Branch 2 taken 2 times.
✓ Branch 3 taken 8 times.
✓ Branch 4 taken 2 times.
✓ Branch 5 taken 8 times.
✓ Branch 6 taken 2 times.
✗ Branch 7 not taken.
✓ Branch 8 taken 2 times.
✗ Branch 9 not taken.
✓ Branch 10 taken 2 times.
✗ Branch 11 not taken.
✓ Branch 12 taken 2 times.
✗ Branch 13 not taken.
✗ Branch 14 not taken.
✓ Branch 15 taken 2 times.
✗ Branch 16 not taken.
✓ Branch 17 taken 2 times.
✓ Branch 18 taken 2 times.
✗ Branch 19 not taken.
✗ Branch 20 not taken.
✓ Branch 21 taken 2 times.
✓ Branch 22 taken 2 times.
✗ Branch 23 not taken.
✗ Branch 24 not taken.
✓ Branch 25 taken 2 times.
✓ Branch 26 taken 2 times.
✗ Branch 27 not taken.
✗ Branch 28 not taken.
✓ Branch 29 taken 2 times.
✓ Branch 30 taken 2 times.
✗ Branch 31 not taken.
✗ Branch 32 not taken.
✓ Branch 33 taken 2 times.
✓ Branch 34 taken 2 times.
✗ Branch 35 not taken.
✓ Branch 36 taken 2 times.
✗ Branch 37 not taken.
✗ Branch 38 not taken.
✓ Branch 39 taken 2 times.
✗ Branch 40 not taken.
✗ Branch 41 not taken.
✗ Branch 42 not taken.
✓ Branch 43 taken 2 times.
✗ Branch 44 not taken.
✗ Branch 45 not taken.
✗ Branch 46 not taken.
✓ Branch 47 taken 2 times.
✓ Branch 48 taken 2 times.
✗ Branch 49 not taken.
✗ Branch 50 not taken.
✓ Branch 51 taken 2 times.
✓ Branch 52 taken 2 times.
✗ Branch 53 not taken.
✓ Branch 54 taken 2 times.
✗ Branch 55 not taken.
✗ Branch 56 not taken.
✓ Branch 57 taken 2 times.
✗ Branch 58 not taken.
✗ Branch 59 not taken.
|
102465 | if (mpi) PetscCallMPI(MPIU_Allreduce(&lnrm,nrm,1,MPIU_REAL,MPIU_LAPY2,PetscObjectComm((PetscObject)bv))); |
51 | 86657 | else *nrm = lnrm; | |
52 |
3/6✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
|
102465 | PetscCall(PetscLogFlops(2.0*m*n)); |
53 |
1/2✓ Branch 0 taken 10 times.
✗ Branch 1 not taken.
|
104 | } else if (type==NORM_1) { |
54 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 10 times.
|
104 | if (mpi) { |
55 | ✗ | PetscCall(BVAllocateWork_Private(bv,2*n_)); | |
56 | ✗ | rwork = (PetscReal*)bv->work; | |
57 | ✗ | rwork2 = rwork+n_; | |
58 | ✗ | PetscCall(PetscArrayzero(rwork,n_)); | |
59 | ✗ | PetscCall(PetscArrayzero(rwork2,n_)); | |
60 | ✗ | for (j=0;j<n_;j++) { | |
61 | ✗ | for (i=0;i<m_;i++) { | |
62 | ✗ | rwork[j] += PetscAbsScalar(A[i+j*lda_]); | |
63 | } | ||
64 | } | ||
65 | ✗ | PetscCall(PetscMPIIntCast(n_,&len)); | |
66 | ✗ | PetscCallMPI(MPIU_Allreduce(rwork,rwork2,len,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)bv))); | |
67 | ✗ | *nrm = 0.0; | |
68 | ✗ | for (j=0;j<n_;j++) if (rwork2[j] > *nrm) *nrm = rwork2[j]; | |
69 | } else { | ||
70 | 104 | *nrm = LAPACKlange_("O",&m,&n,(PetscScalar*)A,&lda,rwork); | |
71 | } | ||
72 |
3/6✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
|
104 | PetscCall(PetscLogFlops(1.0*m*n)); |
73 | ✗ | } else if (type==NORM_INFINITY) { | |
74 | ✗ | PetscCall(BVAllocateWork_Private(bv,m_)); | |
75 | ✗ | rwork = (PetscReal*)bv->work; | |
76 | ✗ | lnrm = LAPACKlange_("I",&m,&n,(PetscScalar*)A,&lda,rwork); | |
77 | ✗ | if (mpi) PetscCallMPI(MPIU_Allreduce(&lnrm,nrm,1,MPIU_REAL,MPIU_MAX,PetscObjectComm((PetscObject)bv))); | |
78 | ✗ | else *nrm = lnrm; | |
79 | ✗ | PetscCall(PetscLogFlops(1.0*m*n)); | |
80 | } | ||
81 |
4/6✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
|
102569 | PetscCall(PetscFPTrapPop()); |
82 |
6/12✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✗ Branch 2 not taken.
✓ Branch 3 taken 2 times.
✓ Branch 4 taken 2 times.
✗ Branch 5 not taken.
✓ Branch 6 taken 2 times.
✗ Branch 7 not taken.
✓ Branch 8 taken 2 times.
✗ Branch 9 not taken.
✗ Branch 10 not taken.
✓ Branch 11 taken 2 times.
|
19421 | PetscFunctionReturn(PETSC_SUCCESS); |
83 | } | ||
84 | |||
85 | /* | ||
86 | Normalize the columns of an mxn matrix A | ||
87 | */ | ||
88 | 37441 | PetscErrorCode BVNormalize_LAPACK_Private(BV bv,PetscInt m_,PetscInt n_,const PetscScalar *A,PetscInt lda_,PetscScalar *eigi,PetscBool mpi) | |
89 | { | ||
90 | 37441 | PetscBLASInt m,lda,j,k,info,zero=0; | |
91 | 37441 | PetscMPIInt len; | |
92 | 37441 | PetscReal *norms,*rwork=NULL,*rwork2=NULL,done=1.0; | |
93 | |||
94 |
1/2✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
|
37441 | PetscFunctionBegin; |
95 |
4/6✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
|
37441 | PetscCall(PetscFPTrapPush(PETSC_FP_TRAP_OFF)); |
96 |
4/6✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
|
37441 | PetscCall(PetscBLASIntCast(m_,&m)); |
97 |
4/6✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
|
37441 | PetscCall(PetscBLASIntCast(lda_,&lda)); |
98 |
4/6✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
|
37441 | PetscCall(BVAllocateWork_Private(bv,2*n_)); |
99 | 37441 | rwork = (PetscReal*)bv->work; | |
100 | 37441 | rwork2 = rwork+n_; | |
101 | /* compute local norms */ | ||
102 |
2/2✓ Branch 0 taken 10 times.
✓ Branch 1 taken 10 times.
|
407895 | for (j=0;j<n_;j++) { |
103 | 370454 | k = 1; | |
104 | #if !defined(PETSC_USE_COMPLEX) | ||
105 |
4/4✓ Branch 0 taken 5 times.
✓ Branch 1 taken 5 times.
✓ Branch 2 taken 5 times.
✓ Branch 3 taken 5 times.
|
172077 | if (eigi && eigi[j] != 0.0) k = 2; |
106 | #endif | ||
107 | 370454 | rwork[j] = LAPACKlange_("F",&m,&k,(PetscScalar*)(A+j*lda_),&lda,rwork2); | |
108 |
2/2✓ Branch 0 taken 5 times.
✓ Branch 1 taken 10 times.
|
370454 | if (k==2) { rwork[j+1] = rwork[j]; j++; } |
109 | } | ||
110 | /* reduction to get global norms */ | ||
111 |
2/2✓ Branch 0 taken 10 times.
✓ Branch 1 taken 10 times.
|
37441 | if (mpi) { |
112 |
4/6✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
|
996 | PetscCall(PetscMPIIntCast(n_,&len)); |
113 |
4/6✓ Branch 0 taken 5 times.
✓ Branch 1 taken 5 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
|
996 | PetscCall(PetscArrayzero(rwork2,n_)); |
114 |
28/58✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✓ Branch 3 taken 8 times.
✓ Branch 4 taken 2 times.
✗ Branch 5 not taken.
✓ Branch 6 taken 2 times.
✗ Branch 7 not taken.
✓ Branch 8 taken 2 times.
✗ Branch 9 not taken.
✓ Branch 10 taken 2 times.
✗ Branch 11 not taken.
✗ Branch 12 not taken.
✓ Branch 13 taken 2 times.
✗ Branch 14 not taken.
✓ Branch 15 taken 2 times.
✓ Branch 16 taken 2 times.
✗ Branch 17 not taken.
✗ Branch 18 not taken.
✓ Branch 19 taken 2 times.
✓ Branch 20 taken 2 times.
✗ Branch 21 not taken.
✗ Branch 22 not taken.
✓ Branch 23 taken 2 times.
✓ Branch 24 taken 2 times.
✗ Branch 25 not taken.
✗ Branch 26 not taken.
✓ Branch 27 taken 2 times.
✓ Branch 28 taken 2 times.
✗ Branch 29 not taken.
✗ Branch 30 not taken.
✓ Branch 31 taken 2 times.
✓ Branch 32 taken 2 times.
✗ Branch 33 not taken.
✓ Branch 34 taken 2 times.
✗ Branch 35 not taken.
✗ Branch 36 not taken.
✓ Branch 37 taken 2 times.
✗ Branch 38 not taken.
✗ Branch 39 not taken.
✗ Branch 40 not taken.
✓ Branch 41 taken 2 times.
✗ Branch 42 not taken.
✗ Branch 43 not taken.
✗ Branch 44 not taken.
✓ Branch 45 taken 2 times.
✓ Branch 46 taken 2 times.
✗ Branch 47 not taken.
✗ Branch 48 not taken.
✓ Branch 49 taken 2 times.
✓ Branch 50 taken 2 times.
✗ Branch 51 not taken.
✓ Branch 52 taken 2 times.
✗ Branch 53 not taken.
✗ Branch 54 not taken.
✓ Branch 55 taken 2 times.
✗ Branch 56 not taken.
✗ Branch 57 not taken.
|
996 | PetscCallMPI(MPIU_Allreduce(rwork,rwork2,len,MPIU_REAL,MPIU_LAPY2,PetscObjectComm((PetscObject)bv))); |
115 | 180 | norms = rwork2; | |
116 | } else norms = rwork; | ||
117 | /* scale columns */ | ||
118 |
2/2✓ Branch 0 taken 10 times.
✓ Branch 1 taken 10 times.
|
407895 | for (j=0;j<n_;j++) { |
119 | 370454 | k = 1; | |
120 | #if !defined(PETSC_USE_COMPLEX) | ||
121 |
4/4✓ Branch 0 taken 5 times.
✓ Branch 1 taken 5 times.
✓ Branch 2 taken 5 times.
✓ Branch 3 taken 5 times.
|
172077 | if (eigi && eigi[j] != 0.0) k = 2; |
122 | #endif | ||
123 |
10/20✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✓ Branch 4 taken 2 times.
✗ Branch 5 not taken.
✗ Branch 6 not taken.
✓ Branch 7 taken 2 times.
✓ Branch 8 taken 2 times.
✗ Branch 9 not taken.
✗ Branch 10 not taken.
✓ Branch 11 taken 2 times.
✓ Branch 12 taken 2 times.
✗ Branch 13 not taken.
✓ Branch 14 taken 2 times.
✗ Branch 15 not taken.
✗ Branch 16 not taken.
✓ Branch 17 taken 2 times.
✗ Branch 18 not taken.
✗ Branch 19 not taken.
|
370454 | PetscCallBLAS("LAPACKlascl",LAPACKlascl_("G",&zero,&zero,norms+j,&done,&m,&k,(PetscScalar*)(A+j*lda_),&lda,&info)); |
124 |
1/4✗ Branch 0 not taken.
✓ Branch 1 taken 10 times.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
|
370454 | SlepcCheckLapackInfo("lascl",info); |
125 |
2/2✓ Branch 0 taken 5 times.
✓ Branch 1 taken 10 times.
|
370454 | if (k==2) j++; |
126 | } | ||
127 |
3/6✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
|
37441 | PetscCall(PetscLogFlops(3.0*m*n_)); |
128 |
4/6✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
|
37441 | PetscCall(PetscFPTrapPop()); |
129 |
6/12✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✗ Branch 2 not taken.
✓ Branch 3 taken 2 times.
✓ Branch 4 taken 2 times.
✗ Branch 5 not taken.
✓ Branch 6 taken 2 times.
✗ Branch 7 not taken.
✓ Branch 8 taken 2 times.
✗ Branch 9 not taken.
✗ Branch 10 not taken.
✓ Branch 11 taken 2 times.
|
7465 | PetscFunctionReturn(PETSC_SUCCESS); |
130 | } | ||
131 | |||
132 | /* | ||
133 | Compute the upper Cholesky factor in R and its inverse in S. | ||
134 | If S == R then the inverse overwrites the Cholesky factor. | ||
135 | */ | ||
136 | 1178 | PetscErrorCode BVMatCholInv_LAPACK_Private(BV bv,Mat R,Mat S) | |
137 | { | ||
138 | 1178 | PetscInt i,k,l,n,m,ld,lds; | |
139 | 1178 | PetscScalar *pR,*pS; | |
140 | 1178 | PetscBLASInt info,n_ = 0,m_ = 0,ld_,lds_; | |
141 | |||
142 |
1/2✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
|
1178 | PetscFunctionBegin; |
143 | 1178 | l = bv->l; | |
144 | 1178 | k = bv->k; | |
145 |
4/6✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
|
1178 | PetscCall(MatGetSize(R,&m,NULL)); |
146 | 1178 | n = k-l; | |
147 |
4/6✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
|
1178 | PetscCall(PetscBLASIntCast(m,&m_)); |
148 |
4/6✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
|
1178 | PetscCall(PetscBLASIntCast(n,&n_)); |
149 | 1178 | ld = m; | |
150 | 1178 | ld_ = m_; | |
151 |
4/6✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
|
1178 | PetscCall(MatDenseGetArray(R,&pR)); |
152 | |||
153 |
2/2✓ Branch 0 taken 10 times.
✓ Branch 1 taken 10 times.
|
1178 | if (S==R) { |
154 |
4/6✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
|
794 | PetscCall(BVAllocateWork_Private(bv,m*k)); |
155 | 794 | pS = bv->work; | |
156 | 794 | lds = ld; | |
157 | 794 | lds_ = ld_; | |
158 | } else { | ||
159 |
4/6✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
|
384 | PetscCall(MatDenseGetArray(S,&pS)); |
160 |
4/6✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
|
384 | PetscCall(MatGetSize(S,&lds,NULL)); |
161 |
5/6✓ Branch 0 taken 9 times.
✓ Branch 1 taken 1 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✓ Branch 4 taken 1 times.
✓ Branch 5 taken 1 times.
|
384 | PetscCall(PetscBLASIntCast(lds,&lds_)); |
162 | } | ||
163 | |||
164 | /* save a copy of matrix in S */ | ||
165 |
7/8✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 10 times.
✓ Branch 3 taken 8 times.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
✓ Branch 6 taken 2 times.
✓ Branch 7 taken 2 times.
|
10810 | for (i=l;i<k;i++) PetscCall(PetscArraycpy(pS+i*lds+l,pR+i*ld+l,n)); |
166 | |||
167 | /* compute upper Cholesky factor in R */ | ||
168 |
4/6✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
|
1178 | PetscCall(PetscFPTrapPush(PETSC_FP_TRAP_OFF)); |
169 |
10/20✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✓ Branch 4 taken 2 times.
✗ Branch 5 not taken.
✗ Branch 6 not taken.
✓ Branch 7 taken 2 times.
✓ Branch 8 taken 2 times.
✗ Branch 9 not taken.
✗ Branch 10 not taken.
✓ Branch 11 taken 2 times.
✓ Branch 12 taken 2 times.
✗ Branch 13 not taken.
✓ Branch 14 taken 2 times.
✗ Branch 15 not taken.
✗ Branch 16 not taken.
✓ Branch 17 taken 2 times.
✗ Branch 18 not taken.
✗ Branch 19 not taken.
|
1178 | PetscCallBLAS("LAPACKpotrf",LAPACKpotrf_("U",&n_,pR+l*ld+l,&ld_,&info)); |
170 |
4/6✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
|
1178 | PetscCall(PetscLogFlops((1.0*n*n*n)/3.0)); |
171 | |||
172 |
1/2✗ Branch 0 not taken.
✓ Branch 1 taken 10 times.
|
1178 | if (info) { /* LAPACKpotrf failed, retry on diagonally perturbed matrix */ |
173 | ✗ | for (i=l;i<k;i++) { | |
174 | ✗ | PetscCall(PetscArraycpy(pR+i*ld+l,pS+i*lds+l,n)); | |
175 | ✗ | pR[i+i*ld] += 50.0*PETSC_MACHINE_EPSILON; | |
176 | } | ||
177 | ✗ | PetscCallBLAS("LAPACKpotrf",LAPACKpotrf_("U",&n_,pR+l*ld+l,&ld_,&info)); | |
178 | ✗ | SlepcCheckLapackInfo("potrf",info); | |
179 | ✗ | PetscCall(PetscLogFlops((1.0*n*n*n)/3.0)); | |
180 | } | ||
181 | |||
182 | /* compute S = inv(R) */ | ||
183 |
2/2✓ Branch 0 taken 10 times.
✓ Branch 1 taken 10 times.
|
1178 | if (S==R) { |
184 |
10/20✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✓ Branch 4 taken 2 times.
✗ Branch 5 not taken.
✗ Branch 6 not taken.
✓ Branch 7 taken 2 times.
✓ Branch 8 taken 2 times.
✗ Branch 9 not taken.
✗ Branch 10 not taken.
✓ Branch 11 taken 2 times.
✓ Branch 12 taken 2 times.
✗ Branch 13 not taken.
✓ Branch 14 taken 2 times.
✗ Branch 15 not taken.
✗ Branch 16 not taken.
✓ Branch 17 taken 2 times.
✗ Branch 18 not taken.
✗ Branch 19 not taken.
|
794 | PetscCallBLAS("LAPACKtrtri",LAPACKtrtri_("U","N",&n_,pR+l*ld+l,&ld_,&info)); |
185 | } else { | ||
186 |
4/6✓ Branch 0 taken 5 times.
✓ Branch 1 taken 5 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
|
384 | PetscCall(PetscArrayzero(pS+l*lds,(k-l)*k)); |
187 |
7/8✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 10 times.
✓ Branch 3 taken 8 times.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
✓ Branch 6 taken 2 times.
✓ Branch 7 taken 2 times.
|
1920 | for (i=l;i<k;i++) PetscCall(PetscArraycpy(pS+i*lds+l,pR+i*ld+l,n)); |
188 |
10/20✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✓ Branch 4 taken 2 times.
✗ Branch 5 not taken.
✗ Branch 6 not taken.
✓ Branch 7 taken 2 times.
✓ Branch 8 taken 2 times.
✗ Branch 9 not taken.
✗ Branch 10 not taken.
✓ Branch 11 taken 2 times.
✓ Branch 12 taken 2 times.
✗ Branch 13 not taken.
✓ Branch 14 taken 2 times.
✗ Branch 15 not taken.
✗ Branch 16 not taken.
✓ Branch 17 taken 2 times.
✗ Branch 18 not taken.
✗ Branch 19 not taken.
|
384 | PetscCallBLAS("LAPACKtrtri",LAPACKtrtri_("U","N",&n_,pS+l*lds+l,&lds_,&info)); |
189 | } | ||
190 |
1/4✗ Branch 0 not taken.
✓ Branch 1 taken 10 times.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
|
1178 | SlepcCheckLapackInfo("trtri",info); |
191 |
4/6✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
|
1178 | PetscCall(PetscFPTrapPop()); |
192 |
3/6✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
|
1178 | PetscCall(PetscLogFlops(0.33*n*n*n)); |
193 | |||
194 | /* Zero out entries below the diagonal */ | ||
195 |
2/2✓ Branch 0 taken 10 times.
✓ Branch 1 taken 10 times.
|
9632 | for (i=l;i<k-1;i++) { |
196 |
4/6✓ Branch 0 taken 5 times.
✓ Branch 1 taken 5 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
|
8454 | PetscCall(PetscArrayzero(pR+i*ld+i+1,(k-i-1))); |
197 |
6/8✓ Branch 0 taken 10 times.
✓ Branch 1 taken 10 times.
✓ Branch 2 taken 5 times.
✓ Branch 3 taken 5 times.
✓ Branch 4 taken 2 times.
✗ Branch 5 not taken.
✗ Branch 6 not taken.
✓ Branch 7 taken 2 times.
|
8838 | if (S!=R) PetscCall(PetscArrayzero(pS+i*lds+i+1,(k-i-1))); |
198 | } | ||
199 |
4/6✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
|
1178 | PetscCall(MatDenseRestoreArray(R,&pR)); |
200 |
6/8✓ Branch 0 taken 10 times.
✓ Branch 1 taken 10 times.
✓ Branch 2 taken 2 times.
✓ Branch 3 taken 8 times.
✓ Branch 4 taken 2 times.
✗ Branch 5 not taken.
✗ Branch 6 not taken.
✓ Branch 7 taken 2 times.
|
1178 | if (S!=R) PetscCall(MatDenseRestoreArray(S,&pS)); |
201 |
6/12✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✗ Branch 2 not taken.
✓ Branch 3 taken 2 times.
✓ Branch 4 taken 2 times.
✗ Branch 5 not taken.
✓ Branch 6 taken 2 times.
✗ Branch 7 not taken.
✓ Branch 8 taken 2 times.
✗ Branch 9 not taken.
✗ Branch 10 not taken.
✓ Branch 11 taken 2 times.
|
210 | PetscFunctionReturn(PETSC_SUCCESS); |
202 | } | ||
203 | |||
204 | /* | ||
205 | Compute the inverse of an upper triangular matrix R, store it in S. | ||
206 | If S == R then the inverse overwrites R. | ||
207 | */ | ||
208 | 794 | PetscErrorCode BVMatTriInv_LAPACK_Private(BV bv,Mat R,Mat S) | |
209 | { | ||
210 | 794 | PetscInt i,k,l,n,m,ld,lds; | |
211 | 794 | PetscScalar *pR,*pS; | |
212 | 794 | PetscBLASInt info,n_,m_ = 0,ld_,lds_; | |
213 | |||
214 |
1/2✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
|
794 | PetscFunctionBegin; |
215 | 794 | l = bv->l; | |
216 | 794 | k = bv->k; | |
217 |
4/6✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
|
794 | PetscCall(MatGetSize(R,&m,NULL)); |
218 | 794 | n = k-l; | |
219 |
4/6✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
|
794 | PetscCall(PetscBLASIntCast(m,&m_)); |
220 |
4/6✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
|
794 | PetscCall(PetscBLASIntCast(n,&n_)); |
221 | 794 | ld = m; | |
222 | 794 | ld_ = m_; | |
223 |
4/6✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
|
794 | PetscCall(MatDenseGetArray(R,&pR)); |
224 | |||
225 |
2/2✓ Branch 0 taken 10 times.
✓ Branch 1 taken 10 times.
|
794 | if (S==R) { |
226 |
4/6✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
|
602 | PetscCall(BVAllocateWork_Private(bv,m*k)); |
227 | 602 | pS = bv->work; | |
228 | 602 | lds = ld; | |
229 | 602 | lds_ = ld_; | |
230 | } else { | ||
231 |
4/6✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
|
192 | PetscCall(MatDenseGetArray(S,&pS)); |
232 |
4/6✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
|
192 | PetscCall(MatGetSize(S,&lds,NULL)); |
233 |
4/6✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
|
192 | PetscCall(PetscBLASIntCast(lds,&lds_)); |
234 | } | ||
235 | |||
236 | /* compute S = inv(R) */ | ||
237 |
4/6✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
|
794 | PetscCall(PetscFPTrapPush(PETSC_FP_TRAP_OFF)); |
238 |
2/2✓ Branch 0 taken 10 times.
✓ Branch 1 taken 10 times.
|
794 | if (S==R) { |
239 |
10/20✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✓ Branch 4 taken 2 times.
✗ Branch 5 not taken.
✗ Branch 6 not taken.
✓ Branch 7 taken 2 times.
✓ Branch 8 taken 2 times.
✗ Branch 9 not taken.
✗ Branch 10 not taken.
✓ Branch 11 taken 2 times.
✓ Branch 12 taken 2 times.
✗ Branch 13 not taken.
✓ Branch 14 taken 2 times.
✗ Branch 15 not taken.
✗ Branch 16 not taken.
✓ Branch 17 taken 2 times.
✗ Branch 18 not taken.
✗ Branch 19 not taken.
|
602 | PetscCallBLAS("LAPACKtrtri",LAPACKtrtri_("U","N",&n_,pR+l*ld+l,&ld_,&info)); |
240 | } else { | ||
241 |
4/6✓ Branch 0 taken 5 times.
✓ Branch 1 taken 5 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
|
192 | PetscCall(PetscArrayzero(pS+l*lds,(k-l)*k)); |
242 |
7/8✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 10 times.
✓ Branch 3 taken 8 times.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
✓ Branch 6 taken 2 times.
✓ Branch 7 taken 2 times.
|
960 | for (i=l;i<k;i++) PetscCall(PetscArraycpy(pS+i*lds+l,pR+i*ld+l,n)); |
243 |
10/20✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✓ Branch 4 taken 2 times.
✗ Branch 5 not taken.
✗ Branch 6 not taken.
✓ Branch 7 taken 2 times.
✓ Branch 8 taken 2 times.
✗ Branch 9 not taken.
✗ Branch 10 not taken.
✓ Branch 11 taken 2 times.
✓ Branch 12 taken 2 times.
✗ Branch 13 not taken.
✓ Branch 14 taken 2 times.
✗ Branch 15 not taken.
✗ Branch 16 not taken.
✓ Branch 17 taken 2 times.
✗ Branch 18 not taken.
✗ Branch 19 not taken.
|
192 | PetscCallBLAS("LAPACKtrtri",LAPACKtrtri_("U","N",&n_,pS+l*lds+l,&lds_,&info)); |
244 | } | ||
245 |
1/4✗ Branch 0 not taken.
✓ Branch 1 taken 10 times.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
|
794 | SlepcCheckLapackInfo("trtri",info); |
246 |
4/6✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
|
794 | PetscCall(PetscFPTrapPop()); |
247 |
3/6✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
|
794 | PetscCall(PetscLogFlops(0.33*n*n*n)); |
248 | |||
249 |
4/6✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
|
794 | PetscCall(MatDenseRestoreArray(R,&pR)); |
250 |
6/8✓ Branch 0 taken 10 times.
✓ Branch 1 taken 10 times.
✓ Branch 2 taken 2 times.
✓ Branch 3 taken 8 times.
✓ Branch 4 taken 2 times.
✗ Branch 5 not taken.
✗ Branch 6 not taken.
✓ Branch 7 taken 2 times.
|
794 | if (S!=R) PetscCall(MatDenseRestoreArray(S,&pS)); |
251 |
6/12✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✗ Branch 2 not taken.
✓ Branch 3 taken 2 times.
✓ Branch 4 taken 2 times.
✗ Branch 5 not taken.
✓ Branch 6 taken 2 times.
✗ Branch 7 not taken.
✓ Branch 8 taken 2 times.
✗ Branch 9 not taken.
✗ Branch 10 not taken.
✓ Branch 11 taken 2 times.
|
146 | PetscFunctionReturn(PETSC_SUCCESS); |
252 | } | ||
253 | |||
254 | /* | ||
255 | Compute the matrix to be used for post-multiplying the basis in the SVQB | ||
256 | block orthogonalization method. | ||
257 | On input R = V'*V, on output S = D*U*Lambda^{-1/2} where (U,Lambda) is | ||
258 | the eigendecomposition of D*R*D with D=diag(R)^{-1/2}. | ||
259 | If S == R then the result overwrites R. | ||
260 | */ | ||
261 | 1178 | PetscErrorCode BVMatSVQB_LAPACK_Private(BV bv,Mat R,Mat S) | |
262 | { | ||
263 | 1178 | PetscInt i,j,k,l,n,m,ld,lds; | |
264 | 1178 | PetscScalar *pR,*pS,*D,*work,a; | |
265 | 1178 | PetscReal *eig,dummy; | |
266 | 1178 | PetscBLASInt info,lwork,n_,m_ = 0,ld_,lds_; | |
267 | #if defined(PETSC_USE_COMPLEX) | ||
268 | 589 | PetscReal *rwork,rdummy; | |
269 | #endif | ||
270 | |||
271 |
1/2✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
|
1178 | PetscFunctionBegin; |
272 | 1178 | l = bv->l; | |
273 | 1178 | k = bv->k; | |
274 |
4/6✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
|
1178 | PetscCall(MatGetSize(R,&m,NULL)); |
275 |
4/6✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
|
1178 | PetscCall(MatDenseGetLDA(R,&ld)); |
276 | 1178 | n = k-l; | |
277 |
4/6✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
|
1178 | PetscCall(PetscBLASIntCast(m,&m_)); |
278 |
4/6✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
|
1178 | PetscCall(PetscBLASIntCast(n,&n_)); |
279 | 1178 | ld_ = m_; | |
280 |
4/6✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
|
1178 | PetscCall(MatDenseGetArray(R,&pR)); |
281 | |||
282 |
2/2✓ Branch 0 taken 10 times.
✓ Branch 1 taken 10 times.
|
1178 | if (S==R) { |
283 | 794 | pS = pR; | |
284 | 794 | lds = ld; | |
285 | 794 | lds_ = ld_; | |
286 | } else { | ||
287 |
4/6✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
|
384 | PetscCall(MatDenseGetArray(S,&pS)); |
288 |
4/6✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
|
384 | PetscCall(MatDenseGetLDA(S,&lds)); |
289 |
4/6✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
|
384 | PetscCall(PetscBLASIntCast(lds,&lds_)); |
290 | } | ||
291 |
4/6✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
|
1178 | PetscCall(PetscFPTrapPush(PETSC_FP_TRAP_OFF)); |
292 | |||
293 | /* workspace query and memory allocation */ | ||
294 | 1178 | lwork = -1; | |
295 | #if defined(PETSC_USE_COMPLEX) | ||
296 |
10/20✓ Branch 0 taken 1 times.
✓ Branch 1 taken 4 times.
✓ Branch 2 taken 1 times.
✗ Branch 3 not taken.
✓ Branch 4 taken 1 times.
✗ Branch 5 not taken.
✗ Branch 6 not taken.
✓ Branch 7 taken 1 times.
✓ Branch 8 taken 1 times.
✗ Branch 9 not taken.
✗ Branch 10 not taken.
✓ Branch 11 taken 1 times.
✓ Branch 12 taken 1 times.
✗ Branch 13 not taken.
✓ Branch 14 taken 1 times.
✗ Branch 15 not taken.
✗ Branch 16 not taken.
✓ Branch 17 taken 1 times.
✗ Branch 18 not taken.
✗ Branch 19 not taken.
|
589 | PetscCallBLAS("LAPACKsyev",LAPACKsyev_("V","L",&n_,pS,&lds_,&dummy,&a,&lwork,&rdummy,&info)); |
297 |
4/6✓ Branch 0 taken 1 times.
✓ Branch 1 taken 4 times.
✓ Branch 2 taken 1 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 1 times.
|
589 | PetscCall(PetscBLASIntCast((PetscInt)PetscRealPart(a),&lwork)); |
298 |
5/8✓ Branch 0 taken 5 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 1 times.
✓ Branch 3 taken 4 times.
✓ Branch 4 taken 1 times.
✗ Branch 5 not taken.
✗ Branch 6 not taken.
✓ Branch 7 taken 1 times.
|
589 | PetscCall(PetscMalloc4(n,&eig,n,&D,lwork,&work,PetscMax(1,3*n-2),&rwork)); |
299 | #else | ||
300 |
10/20✓ Branch 0 taken 1 times.
✓ Branch 1 taken 4 times.
✓ Branch 2 taken 1 times.
✗ Branch 3 not taken.
✓ Branch 4 taken 1 times.
✗ Branch 5 not taken.
✗ Branch 6 not taken.
✓ Branch 7 taken 1 times.
✓ Branch 8 taken 1 times.
✗ Branch 9 not taken.
✗ Branch 10 not taken.
✓ Branch 11 taken 1 times.
✓ Branch 12 taken 1 times.
✗ Branch 13 not taken.
✓ Branch 14 taken 1 times.
✗ Branch 15 not taken.
✗ Branch 16 not taken.
✓ Branch 17 taken 1 times.
✗ Branch 18 not taken.
✗ Branch 19 not taken.
|
589 | PetscCallBLAS("LAPACKsyev",LAPACKsyev_("V","L",&n_,pS,&lds_,&dummy,&a,&lwork,&info)); |
301 |
4/6✓ Branch 0 taken 1 times.
✓ Branch 1 taken 4 times.
✓ Branch 2 taken 1 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 1 times.
|
589 | PetscCall(PetscBLASIntCast((PetscInt)a,&lwork)); |
302 |
4/6✓ Branch 0 taken 1 times.
✓ Branch 1 taken 4 times.
✓ Branch 2 taken 1 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 1 times.
|
589 | PetscCall(PetscMalloc3(n,&eig,n,&D,lwork,&work)); |
303 | #endif | ||
304 | |||
305 | /* copy and scale matrix */ | ||
306 |
2/2✓ Branch 0 taken 10 times.
✓ Branch 1 taken 10 times.
|
10810 | for (i=l;i<k;i++) D[i-l] = 1.0/PetscSqrtReal(PetscRealPart(pR[i+i*ld])); |
307 |
4/4✓ Branch 0 taken 10 times.
✓ Branch 1 taken 10 times.
✓ Branch 2 taken 10 times.
✓ Branch 3 taken 10 times.
|
131130 | for (i=l;i<k;i++) for (j=l;j<k;j++) pS[i+j*lds] = pR[i+j*ld]*D[i-l]; |
308 |
4/4✓ Branch 0 taken 10 times.
✓ Branch 1 taken 10 times.
✓ Branch 2 taken 10 times.
✓ Branch 3 taken 10 times.
|
131130 | for (j=l;j<k;j++) for (i=l;i<k;i++) pS[i+j*lds] *= D[j-l]; |
309 | |||
310 | /* compute eigendecomposition */ | ||
311 | #if defined(PETSC_USE_COMPLEX) | ||
312 |
10/20✓ Branch 0 taken 1 times.
✓ Branch 1 taken 4 times.
✓ Branch 2 taken 1 times.
✗ Branch 3 not taken.
✓ Branch 4 taken 1 times.
✗ Branch 5 not taken.
✗ Branch 6 not taken.
✓ Branch 7 taken 1 times.
✓ Branch 8 taken 1 times.
✗ Branch 9 not taken.
✗ Branch 10 not taken.
✓ Branch 11 taken 1 times.
✓ Branch 12 taken 1 times.
✗ Branch 13 not taken.
✓ Branch 14 taken 1 times.
✗ Branch 15 not taken.
✗ Branch 16 not taken.
✓ Branch 17 taken 1 times.
✗ Branch 18 not taken.
✗ Branch 19 not taken.
|
589 | PetscCallBLAS("LAPACKsyev",LAPACKsyev_("V","L",&n_,pS+l*lds+l,&lds_,eig,work,&lwork,rwork,&info)); |
313 | #else | ||
314 |
10/20✓ Branch 0 taken 1 times.
✓ Branch 1 taken 4 times.
✓ Branch 2 taken 1 times.
✗ Branch 3 not taken.
✓ Branch 4 taken 1 times.
✗ Branch 5 not taken.
✗ Branch 6 not taken.
✓ Branch 7 taken 1 times.
✓ Branch 8 taken 1 times.
✗ Branch 9 not taken.
✗ Branch 10 not taken.
✓ Branch 11 taken 1 times.
✓ Branch 12 taken 1 times.
✗ Branch 13 not taken.
✓ Branch 14 taken 1 times.
✗ Branch 15 not taken.
✗ Branch 16 not taken.
✓ Branch 17 taken 1 times.
✗ Branch 18 not taken.
✗ Branch 19 not taken.
|
589 | PetscCallBLAS("LAPACKsyev",LAPACKsyev_("V","L",&n_,pS+l*lds+l,&lds_,eig,work,&lwork,&info)); |
315 | #endif | ||
316 |
1/4✗ Branch 0 not taken.
✓ Branch 1 taken 10 times.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
|
1178 | SlepcCheckLapackInfo("syev",info); |
317 | |||
318 |
2/2✓ Branch 0 taken 10 times.
✓ Branch 1 taken 10 times.
|
1178 | if (S!=R) { /* R = U' */ |
319 |
4/4✓ Branch 0 taken 10 times.
✓ Branch 1 taken 10 times.
✓ Branch 2 taken 10 times.
✓ Branch 3 taken 10 times.
|
9600 | for (i=l;i<k;i++) for (j=l;j<k;j++) pR[i+j*ld] = pS[j+i*lds]; |
320 | } | ||
321 | |||
322 | /* compute S = D*U*Lambda^{-1/2} */ | ||
323 |
4/4✓ Branch 0 taken 10 times.
✓ Branch 1 taken 10 times.
✓ Branch 2 taken 10 times.
✓ Branch 3 taken 10 times.
|
131130 | for (i=l;i<k;i++) for (j=l;j<k;j++) pS[i+j*lds] *= D[i-l]; |
324 |
4/4✓ Branch 0 taken 10 times.
✓ Branch 1 taken 10 times.
✓ Branch 2 taken 10 times.
✓ Branch 3 taken 10 times.
|
131130 | for (j=l;j<k;j++) for (i=l;i<k;i++) pS[i+j*lds] /= PetscSqrtReal(eig[j-l]); |
325 | |||
326 |
2/2✓ Branch 0 taken 10 times.
✓ Branch 1 taken 10 times.
|
1178 | if (S!=R) { /* compute R = inv(S) = Lambda^{1/2}*U'/D */ |
327 |
4/4✓ Branch 0 taken 10 times.
✓ Branch 1 taken 10 times.
✓ Branch 2 taken 10 times.
✓ Branch 3 taken 10 times.
|
9600 | for (i=l;i<k;i++) for (j=l;j<k;j++) pR[i+j*ld] *= PetscSqrtReal(eig[i-l]); |
328 |
4/4✓ Branch 0 taken 10 times.
✓ Branch 1 taken 10 times.
✓ Branch 2 taken 10 times.
✓ Branch 3 taken 10 times.
|
9600 | for (j=l;j<k;j++) for (i=l;i<k;i++) pR[i+j*ld] /= D[j-l]; |
329 | } | ||
330 | |||
331 | #if defined(PETSC_USE_COMPLEX) | ||
332 |
4/6✓ Branch 0 taken 1 times.
✓ Branch 1 taken 4 times.
✓ Branch 2 taken 1 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 1 times.
|
589 | PetscCall(PetscFree4(eig,D,work,rwork)); |
333 | #else | ||
334 |
4/6✓ Branch 0 taken 1 times.
✓ Branch 1 taken 4 times.
✓ Branch 2 taken 1 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 1 times.
|
589 | PetscCall(PetscFree3(eig,D,work)); |
335 | #endif | ||
336 |
3/6✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
|
1178 | PetscCall(PetscLogFlops(9.0*n*n*n)); |
337 |
4/6✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
|
1178 | PetscCall(PetscFPTrapPop()); |
338 | |||
339 |
4/6✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
|
1178 | PetscCall(MatDenseRestoreArray(R,&pR)); |
340 |
6/8✓ Branch 0 taken 10 times.
✓ Branch 1 taken 10 times.
✓ Branch 2 taken 2 times.
✓ Branch 3 taken 8 times.
✓ Branch 4 taken 2 times.
✗ Branch 5 not taken.
✗ Branch 6 not taken.
✓ Branch 7 taken 2 times.
|
1178 | if (S!=R) PetscCall(MatDenseRestoreArray(S,&pS)); |
341 |
6/12✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✗ Branch 2 not taken.
✓ Branch 3 taken 2 times.
✓ Branch 4 taken 2 times.
✗ Branch 5 not taken.
✓ Branch 6 taken 2 times.
✗ Branch 7 not taken.
✓ Branch 8 taken 2 times.
✗ Branch 9 not taken.
✗ Branch 10 not taken.
✓ Branch 11 taken 2 times.
|
210 | PetscFunctionReturn(PETSC_SUCCESS); |
342 | } | ||
343 | |||
344 | /* | ||
345 | QR factorization of an mxn matrix via parallel TSQR | ||
346 | */ | ||
347 | 2101 | PetscErrorCode BVOrthogonalize_LAPACK_TSQR(BV bv,PetscInt m_,PetscInt n_,PetscScalar *Q,PetscInt ldq_,PetscScalar *R,PetscInt ldr) | |
348 | { | ||
349 | 2101 | PetscInt level,plevel,nlevels,lda,worklen; | |
350 | 2101 | PetscBLASInt m,n,ldq,i,j,k,l,nb,sz,lwork,info; | |
351 | 2101 | PetscScalar *tau,*work,*A=NULL,*QQ=NULL,*Qhalf,*C=NULL,one=1.0,zero=0.0; | |
352 | 2101 | PetscMPIInt rank,size,count,stride,powtwo,s = 0; | |
353 | 2101 | MPI_Datatype tmat; | |
354 | |||
355 |
1/2✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
|
2101 | PetscFunctionBegin; |
356 |
4/6✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
|
2101 | PetscCall(PetscFPTrapPush(PETSC_FP_TRAP_OFF)); |
357 |
4/6✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
|
2101 | PetscCall(PetscBLASIntCast(m_,&m)); |
358 |
4/6✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
|
2101 | PetscCall(PetscBLASIntCast(n_,&n)); |
359 |
4/6✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
|
2101 | PetscCall(PetscBLASIntCast(ldq_,&ldq)); |
360 | 2101 | k = PetscMin(m,n); | |
361 | 2101 | nb = 16; | |
362 | 2101 | lda = 2*n; | |
363 |
14/28✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✓ Branch 3 taken 8 times.
✓ Branch 4 taken 2 times.
✗ Branch 5 not taken.
✓ Branch 6 taken 2 times.
✗ Branch 7 not taken.
✓ Branch 8 taken 2 times.
✗ Branch 9 not taken.
✗ Branch 10 not taken.
✓ Branch 11 taken 2 times.
✓ Branch 12 taken 2 times.
✗ Branch 13 not taken.
✗ Branch 14 not taken.
✓ Branch 15 taken 2 times.
✓ Branch 16 taken 2 times.
✗ Branch 17 not taken.
✓ Branch 18 taken 2 times.
✗ Branch 19 not taken.
✗ Branch 20 not taken.
✓ Branch 21 taken 2 times.
✗ Branch 22 not taken.
✗ Branch 23 not taken.
✗ Branch 24 not taken.
✓ Branch 25 taken 2 times.
✗ Branch 26 not taken.
✗ Branch 27 not taken.
|
2101 | PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)bv),&size)); |
364 |
14/28✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✓ Branch 3 taken 8 times.
✓ Branch 4 taken 2 times.
✗ Branch 5 not taken.
✓ Branch 6 taken 2 times.
✗ Branch 7 not taken.
✓ Branch 8 taken 2 times.
✗ Branch 9 not taken.
✗ Branch 10 not taken.
✓ Branch 11 taken 2 times.
✓ Branch 12 taken 2 times.
✗ Branch 13 not taken.
✗ Branch 14 not taken.
✓ Branch 15 taken 2 times.
✓ Branch 16 taken 2 times.
✗ Branch 17 not taken.
✓ Branch 18 taken 2 times.
✗ Branch 19 not taken.
✗ Branch 20 not taken.
✓ Branch 21 taken 2 times.
✗ Branch 22 not taken.
✗ Branch 23 not taken.
✗ Branch 24 not taken.
✓ Branch 25 taken 2 times.
✗ Branch 26 not taken.
✗ Branch 27 not taken.
|
2101 | PetscCallMPI(MPI_Comm_rank(PetscObjectComm((PetscObject)bv),&rank)); |
365 | 2101 | nlevels = (PetscInt)PetscCeilReal(PetscLog2Real((PetscReal)size)); | |
366 |
4/6✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
|
4202 | PetscCall(PetscMPIIntCast(PetscPowInt(2,(PetscInt)PetscFloorReal(PetscLog2Real((PetscReal)size))),&powtwo)); |
367 | 2101 | worklen = n+n*nb; | |
368 |
2/2✓ Branch 0 taken 10 times.
✓ Branch 1 taken 10 times.
|
2101 | if (nlevels) worklen += n*lda+n*lda*nlevels+n*lda; |
369 |
4/6✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
|
2101 | PetscCall(BVAllocateWork_Private(bv,worklen)); |
370 | 2101 | tau = bv->work; | |
371 | 2101 | work = bv->work+n; | |
372 |
4/6✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
|
2101 | PetscCall(PetscBLASIntCast(n*nb,&lwork)); |
373 |
2/2✓ Branch 0 taken 10 times.
✓ Branch 1 taken 10 times.
|
2101 | if (nlevels) { |
374 | 1224 | A = bv->work+n+n*nb; | |
375 | 1224 | QQ = bv->work+n+n*nb+n*lda; | |
376 | 1224 | C = bv->work+n+n*nb+n*lda+n*lda*nlevels; | |
377 | } | ||
378 | |||
379 | /* Compute QR */ | ||
380 |
10/20✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✓ Branch 4 taken 2 times.
✗ Branch 5 not taken.
✗ Branch 6 not taken.
✓ Branch 7 taken 2 times.
✓ Branch 8 taken 2 times.
✗ Branch 9 not taken.
✗ Branch 10 not taken.
✓ Branch 11 taken 2 times.
✓ Branch 12 taken 2 times.
✗ Branch 13 not taken.
✓ Branch 14 taken 2 times.
✗ Branch 15 not taken.
✗ Branch 16 not taken.
✓ Branch 17 taken 2 times.
✗ Branch 18 not taken.
✗ Branch 19 not taken.
|
2101 | PetscCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&m,&n,Q,&ldq,tau,work,&lwork,&info)); |
381 |
1/4✗ Branch 0 not taken.
✓ Branch 1 taken 10 times.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
|
2101 | SlepcCheckLapackInfo("geqrf",info); |
382 | |||
383 | /* Extract R */ | ||
384 |
1/2✓ Branch 0 taken 10 times.
✗ Branch 1 not taken.
|
2101 | if (R || nlevels) { |
385 |
2/2✓ Branch 0 taken 10 times.
✓ Branch 1 taken 10 times.
|
26697 | for (j=0;j<n;j++) { |
386 |
2/2✓ Branch 0 taken 10 times.
✓ Branch 1 taken 10 times.
|
273234 | for (i=0;i<=PetscMin(j,m-1);i++) { |
387 |
2/2✓ Branch 0 taken 10 times.
✓ Branch 1 taken 10 times.
|
248638 | if (nlevels) A[i+j*lda] = Q[i+j*ldq]; |
388 | 231310 | else R[i+j*ldr] = Q[i+j*ldq]; | |
389 | } | ||
390 |
2/2✓ Branch 0 taken 10 times.
✓ Branch 1 taken 10 times.
|
252318 | for (i=PetscMin(j,m-1)+1;i<n;i++) { |
391 |
2/2✓ Branch 0 taken 10 times.
✓ Branch 1 taken 10 times.
|
227722 | if (nlevels) A[i+j*lda] = 0.0; |
392 | 212450 | else R[i+j*ldr] = 0.0; | |
393 | } | ||
394 | } | ||
395 | } | ||
396 | |||
397 | /* Compute orthogonal matrix in Q */ | ||
398 |
10/20✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✓ Branch 4 taken 2 times.
✗ Branch 5 not taken.
✗ Branch 6 not taken.
✓ Branch 7 taken 2 times.
✓ Branch 8 taken 2 times.
✗ Branch 9 not taken.
✗ Branch 10 not taken.
✓ Branch 11 taken 2 times.
✓ Branch 12 taken 2 times.
✗ Branch 13 not taken.
✓ Branch 14 taken 2 times.
✗ Branch 15 not taken.
✗ Branch 16 not taken.
✓ Branch 17 taken 2 times.
✗ Branch 18 not taken.
✗ Branch 19 not taken.
|
2101 | PetscCallBLAS("LAPACKorgqr",LAPACKorgqr_(&m,&k,&k,Q,&ldq,tau,work,&lwork,&info)); |
399 |
1/4✗ Branch 0 not taken.
✓ Branch 1 taken 10 times.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
|
2101 | SlepcCheckLapackInfo("orgqr",info); |
400 | |||
401 |
2/2✓ Branch 0 taken 10 times.
✓ Branch 1 taken 10 times.
|
2101 | if (nlevels) { |
402 | |||
403 |
4/6✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
|
1224 | PetscCall(PetscMPIIntCast(n,&count)); |
404 |
4/6✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
|
1224 | PetscCall(PetscMPIIntCast(lda,&stride)); |
405 |
4/6✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
|
1224 | PetscCall(PetscBLASIntCast(lda,&l)); |
406 |
14/28✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✓ Branch 3 taken 8 times.
✓ Branch 4 taken 2 times.
✗ Branch 5 not taken.
✓ Branch 6 taken 2 times.
✗ Branch 7 not taken.
✓ Branch 8 taken 2 times.
✗ Branch 9 not taken.
✗ Branch 10 not taken.
✓ Branch 11 taken 2 times.
✓ Branch 12 taken 2 times.
✗ Branch 13 not taken.
✗ Branch 14 not taken.
✓ Branch 15 taken 2 times.
✓ Branch 16 taken 2 times.
✗ Branch 17 not taken.
✓ Branch 18 taken 2 times.
✗ Branch 19 not taken.
✗ Branch 20 not taken.
✓ Branch 21 taken 2 times.
✗ Branch 22 not taken.
✗ Branch 23 not taken.
✗ Branch 24 not taken.
✓ Branch 25 taken 2 times.
✗ Branch 26 not taken.
✗ Branch 27 not taken.
|
1224 | PetscCallMPI(MPI_Type_vector(count,count,stride,MPIU_SCALAR,&tmat)); |
407 |
14/28✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✓ Branch 3 taken 8 times.
✓ Branch 4 taken 2 times.
✗ Branch 5 not taken.
✓ Branch 6 taken 2 times.
✗ Branch 7 not taken.
✓ Branch 8 taken 2 times.
✗ Branch 9 not taken.
✗ Branch 10 not taken.
✓ Branch 11 taken 2 times.
✓ Branch 12 taken 2 times.
✗ Branch 13 not taken.
✗ Branch 14 not taken.
✓ Branch 15 taken 2 times.
✓ Branch 16 taken 2 times.
✗ Branch 17 not taken.
✓ Branch 18 taken 2 times.
✗ Branch 19 not taken.
✗ Branch 20 not taken.
✓ Branch 21 taken 2 times.
✗ Branch 22 not taken.
✗ Branch 23 not taken.
✗ Branch 24 not taken.
✓ Branch 25 taken 2 times.
✗ Branch 26 not taken.
✗ Branch 27 not taken.
|
1224 | PetscCallMPI(MPI_Type_commit(&tmat)); |
408 | |||
409 |
2/2✓ Branch 0 taken 10 times.
✓ Branch 1 taken 10 times.
|
4128 | for (level=nlevels;level>=1;level--) { |
410 | |||
411 | 2904 | plevel = PetscPowInt(2,level); | |
412 |
4/6✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
|
5808 | PetscCall(PetscMPIIntCast(plevel*PetscFloorReal(rank/(PetscReal)plevel)+(rank+PetscPowInt(2,level-1))%plevel,&s)); |
413 | |||
414 | /* Stack triangular matrices */ | ||
415 |
4/4✓ Branch 0 taken 10 times.
✓ Branch 1 taken 10 times.
✓ Branch 2 taken 10 times.
✓ Branch 3 taken 10 times.
|
2904 | if (rank<s && s<size) { /* send top part, receive bottom part */ |
416 |
14/28✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✓ Branch 3 taken 8 times.
✓ Branch 4 taken 2 times.
✗ Branch 5 not taken.
✓ Branch 6 taken 2 times.
✗ Branch 7 not taken.
✓ Branch 8 taken 2 times.
✗ Branch 9 not taken.
✗ Branch 10 not taken.
✓ Branch 11 taken 2 times.
✓ Branch 12 taken 2 times.
✗ Branch 13 not taken.
✗ Branch 14 not taken.
✓ Branch 15 taken 2 times.
✓ Branch 16 taken 2 times.
✗ Branch 17 not taken.
✓ Branch 18 taken 2 times.
✗ Branch 19 not taken.
✗ Branch 20 not taken.
✓ Branch 21 taken 2 times.
✗ Branch 22 not taken.
✗ Branch 23 not taken.
✗ Branch 24 not taken.
✓ Branch 25 taken 2 times.
✗ Branch 26 not taken.
✗ Branch 27 not taken.
|
1272 | PetscCallMPI(MPI_Sendrecv(A,1,tmat,s,111,A+n,1,tmat,s,111,PetscObjectComm((PetscObject)bv),MPI_STATUS_IGNORE)); |
417 |
2/2✓ Branch 0 taken 10 times.
✓ Branch 1 taken 10 times.
|
1632 | } else if (s<size) { /* copy top to bottom, receive top part */ |
418 |
14/28✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✓ Branch 3 taken 8 times.
✓ Branch 4 taken 2 times.
✗ Branch 5 not taken.
✓ Branch 6 taken 2 times.
✗ Branch 7 not taken.
✓ Branch 8 taken 2 times.
✗ Branch 9 not taken.
✗ Branch 10 not taken.
✓ Branch 11 taken 2 times.
✓ Branch 12 taken 2 times.
✗ Branch 13 not taken.
✗ Branch 14 not taken.
✓ Branch 15 taken 2 times.
✓ Branch 16 taken 2 times.
✗ Branch 17 not taken.
✓ Branch 18 taken 2 times.
✗ Branch 19 not taken.
✗ Branch 20 not taken.
✓ Branch 21 taken 2 times.
✗ Branch 22 not taken.
✗ Branch 23 not taken.
✗ Branch 24 not taken.
✓ Branch 25 taken 2 times.
✗ Branch 26 not taken.
✗ Branch 27 not taken.
|
1272 | PetscCallMPI(MPI_Sendrecv(A,1,tmat,rank,111,A+n,1,tmat,rank,111,PetscObjectComm((PetscObject)bv),MPI_STATUS_IGNORE)); |
419 |
14/28✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✓ Branch 3 taken 8 times.
✓ Branch 4 taken 2 times.
✗ Branch 5 not taken.
✓ Branch 6 taken 2 times.
✗ Branch 7 not taken.
✓ Branch 8 taken 2 times.
✗ Branch 9 not taken.
✗ Branch 10 not taken.
✓ Branch 11 taken 2 times.
✓ Branch 12 taken 2 times.
✗ Branch 13 not taken.
✗ Branch 14 not taken.
✓ Branch 15 taken 2 times.
✓ Branch 16 taken 2 times.
✗ Branch 17 not taken.
✓ Branch 18 taken 2 times.
✗ Branch 19 not taken.
✗ Branch 20 not taken.
✓ Branch 21 taken 2 times.
✗ Branch 22 not taken.
✗ Branch 23 not taken.
✗ Branch 24 not taken.
✓ Branch 25 taken 2 times.
✗ Branch 26 not taken.
✗ Branch 27 not taken.
|
1272 | PetscCallMPI(MPI_Sendrecv(A+n,1,tmat,s,111,A,1,tmat,s,111,PetscObjectComm((PetscObject)bv),MPI_STATUS_IGNORE)); |
420 | } | ||
421 |
3/4✓ Branch 0 taken 10 times.
✓ Branch 1 taken 10 times.
✓ Branch 2 taken 10 times.
✗ Branch 3 not taken.
|
2904 | if (level<nlevels && size!=powtwo) { /* for cases when size is not a power of 2 */ |
422 |
2/2✓ Branch 0 taken 10 times.
✓ Branch 1 taken 10 times.
|
1680 | if (rank<size-powtwo) { /* send bottom part */ |
423 |
17/32✓ Branch 0 taken 10 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 2 times.
✓ Branch 3 taken 8 times.
✓ Branch 4 taken 2 times.
✓ Branch 5 taken 8 times.
✓ Branch 6 taken 2 times.
✓ Branch 7 taken 8 times.
✓ Branch 8 taken 2 times.
✗ Branch 9 not taken.
✓ Branch 10 taken 2 times.
✗ Branch 11 not taken.
✗ Branch 12 not taken.
✓ Branch 13 taken 2 times.
✗ Branch 14 not taken.
✓ Branch 15 taken 2 times.
✓ Branch 16 taken 2 times.
✗ Branch 17 not taken.
✗ Branch 18 not taken.
✓ Branch 19 taken 2 times.
✓ Branch 20 taken 2 times.
✗ Branch 21 not taken.
✓ Branch 22 taken 2 times.
✗ Branch 23 not taken.
✗ Branch 24 not taken.
✓ Branch 25 taken 2 times.
✗ Branch 26 not taken.
✗ Branch 27 not taken.
✗ Branch 28 not taken.
✓ Branch 29 taken 2 times.
✗ Branch 30 not taken.
✗ Branch 31 not taken.
|
720 | PetscCallMPI(MPI_Send(A+n,1,tmat,rank+powtwo,111,PetscObjectComm((PetscObject)bv))); |
424 |
2/2✓ Branch 0 taken 10 times.
✓ Branch 1 taken 10 times.
|
960 | } else if (rank>=powtwo) { /* receive bottom part */ |
425 |
17/32✓ Branch 0 taken 10 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 2 times.
✓ Branch 3 taken 8 times.
✓ Branch 4 taken 2 times.
✓ Branch 5 taken 8 times.
✓ Branch 6 taken 2 times.
✓ Branch 7 taken 8 times.
✓ Branch 8 taken 2 times.
✗ Branch 9 not taken.
✓ Branch 10 taken 2 times.
✗ Branch 11 not taken.
✗ Branch 12 not taken.
✓ Branch 13 taken 2 times.
✗ Branch 14 not taken.
✓ Branch 15 taken 2 times.
✓ Branch 16 taken 2 times.
✗ Branch 17 not taken.
✗ Branch 18 not taken.
✓ Branch 19 taken 2 times.
✓ Branch 20 taken 2 times.
✗ Branch 21 not taken.
✓ Branch 22 taken 2 times.
✗ Branch 23 not taken.
✗ Branch 24 not taken.
✓ Branch 25 taken 2 times.
✗ Branch 26 not taken.
✗ Branch 27 not taken.
✗ Branch 28 not taken.
✓ Branch 29 taken 2 times.
✗ Branch 30 not taken.
✗ Branch 31 not taken.
|
720 | PetscCallMPI(MPI_Recv(A+n,1,tmat,rank-powtwo,111,PetscObjectComm((PetscObject)bv),MPI_STATUS_IGNORE)); |
426 | } | ||
427 | } | ||
428 | /* Compute QR and build orthogonal matrix */ | ||
429 |
5/6✓ Branch 0 taken 10 times.
✓ Branch 1 taken 10 times.
✓ Branch 2 taken 10 times.
✗ Branch 3 not taken.
✓ Branch 4 taken 10 times.
✓ Branch 5 taken 10 times.
|
2904 | if (level<nlevels || (level==nlevels && s<size)) { |
430 |
10/20✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✓ Branch 4 taken 2 times.
✗ Branch 5 not taken.
✗ Branch 6 not taken.
✓ Branch 7 taken 2 times.
✓ Branch 8 taken 2 times.
✗ Branch 9 not taken.
✗ Branch 10 not taken.
✓ Branch 11 taken 2 times.
✓ Branch 12 taken 2 times.
✗ Branch 13 not taken.
✓ Branch 14 taken 2 times.
✗ Branch 15 not taken.
✗ Branch 16 not taken.
✓ Branch 17 taken 2 times.
✗ Branch 18 not taken.
✗ Branch 19 not taken.
|
2784 | PetscCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&l,&n,A,&l,tau,work,&lwork,&info)); |
431 |
1/4✗ Branch 0 not taken.
✓ Branch 1 taken 10 times.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
|
2784 | SlepcCheckLapackInfo("geqrf",info); |
432 |
4/6✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
|
2784 | PetscCall(PetscArraycpy(QQ+(level-1)*n*lda,A,n*lda)); |
433 |
10/20✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✓ Branch 4 taken 2 times.
✗ Branch 5 not taken.
✗ Branch 6 not taken.
✓ Branch 7 taken 2 times.
✓ Branch 8 taken 2 times.
✗ Branch 9 not taken.
✗ Branch 10 not taken.
✓ Branch 11 taken 2 times.
✓ Branch 12 taken 2 times.
✗ Branch 13 not taken.
✓ Branch 14 taken 2 times.
✗ Branch 15 not taken.
✗ Branch 16 not taken.
✓ Branch 17 taken 2 times.
✗ Branch 18 not taken.
✗ Branch 19 not taken.
|
2784 | PetscCallBLAS("LAPACKorgqr",LAPACKorgqr_(&l,&n,&n,QQ+(level-1)*n*lda,&l,tau,work,&lwork,&info)); |
434 |
1/4✗ Branch 0 not taken.
✓ Branch 1 taken 10 times.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
|
2784 | SlepcCheckLapackInfo("orgqr",info); |
435 |
2/2✓ Branch 0 taken 10 times.
✓ Branch 1 taken 10 times.
|
16320 | for (j=0;j<n;j++) { |
436 |
2/2✓ Branch 0 taken 10 times.
✓ Branch 1 taken 10 times.
|
46208 | for (i=j+1;i<n;i++) A[i+j*lda] = 0.0; |
437 | } | ||
438 |
1/2✓ Branch 0 taken 10 times.
✗ Branch 1 not taken.
|
120 | } else if (level==nlevels) { /* only one triangular matrix, set Q=I */ |
439 |
4/6✓ Branch 0 taken 5 times.
✓ Branch 1 taken 5 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
|
120 | PetscCall(PetscArrayzero(QQ+(level-1)*n*lda,n*lda)); |
440 |
2/2✓ Branch 0 taken 10 times.
✓ Branch 1 taken 10 times.
|
720 | for (j=0;j<n;j++) QQ[j+j*lda+(level-1)*n*lda] = 1.0; |
441 | } | ||
442 | } | ||
443 | |||
444 | /* Extract R */ | ||
445 |
1/2✓ Branch 0 taken 10 times.
✗ Branch 1 not taken.
|
1224 | if (R) { |
446 |
2/2✓ Branch 0 taken 10 times.
✓ Branch 1 taken 10 times.
|
6960 | for (j=0;j<n;j++) { |
447 |
2/2✓ Branch 0 taken 10 times.
✓ Branch 1 taken 10 times.
|
24904 | for (i=0;i<=j;i++) R[i+j*ldr] = A[i+j*lda]; |
448 |
2/2✓ Branch 0 taken 10 times.
✓ Branch 1 taken 10 times.
|
19168 | for (i=j+1;i<n;i++) R[i+j*ldr] = 0.0; |
449 | } | ||
450 | } | ||
451 | |||
452 | /* Accumulate orthogonal matrices */ | ||
453 |
2/2✓ Branch 0 taken 10 times.
✓ Branch 1 taken 10 times.
|
4128 | for (level=1;level<=nlevels;level++) { |
454 | 2904 | plevel = PetscPowInt(2,level); | |
455 |
4/6✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
|
5808 | PetscCall(PetscMPIIntCast(plevel*PetscFloorReal(rank/(PetscReal)plevel)+(rank+PetscPowInt(2,level-1))%plevel,&s)); |
456 |
2/2✓ Branch 0 taken 10 times.
✓ Branch 1 taken 10 times.
|
2904 | Qhalf = (rank<s)? QQ+(level-1)*n*lda: QQ+(level-1)*n*lda+n; |
457 |
2/2✓ Branch 0 taken 10 times.
✓ Branch 1 taken 10 times.
|
2904 | if (level<nlevels) { |
458 |
10/20✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✓ Branch 4 taken 2 times.
✗ Branch 5 not taken.
✗ Branch 6 not taken.
✓ Branch 7 taken 2 times.
✓ Branch 8 taken 2 times.
✗ Branch 9 not taken.
✗ Branch 10 not taken.
✓ Branch 11 taken 2 times.
✓ Branch 12 taken 2 times.
✗ Branch 13 not taken.
✓ Branch 14 taken 2 times.
✗ Branch 15 not taken.
✗ Branch 16 not taken.
✓ Branch 17 taken 2 times.
✗ Branch 18 not taken.
✗ Branch 19 not taken.
|
1680 | PetscCallBLAS("BLASgemm",BLASgemm_("N","N",&l,&n,&n,&one,QQ+level*n*lda,&l,Qhalf,&l,&zero,C,&l)); |
459 |
4/6✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
|
1680 | PetscCall(PetscArraycpy(QQ+level*n*lda,C,n*lda)); |
460 | } else { | ||
461 |
2/2✓ Branch 0 taken 10 times.
✓ Branch 1 taken 10 times.
|
1888 | for (i=0;i<m/l;i++) { |
462 |
10/20✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✓ Branch 4 taken 2 times.
✗ Branch 5 not taken.
✗ Branch 6 not taken.
✓ Branch 7 taken 2 times.
✓ Branch 8 taken 2 times.
✗ Branch 9 not taken.
✗ Branch 10 not taken.
✓ Branch 11 taken 2 times.
✓ Branch 12 taken 2 times.
✗ Branch 13 not taken.
✓ Branch 14 taken 2 times.
✗ Branch 15 not taken.
✗ Branch 16 not taken.
✓ Branch 17 taken 2 times.
✗ Branch 18 not taken.
✗ Branch 19 not taken.
|
664 | PetscCallBLAS("BLASgemm",BLASgemm_("N","N",&l,&n,&n,&one,Q+i*l,&ldq,Qhalf,&l,&zero,C,&l)); |
463 |
7/8✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 10 times.
✓ Branch 3 taken 8 times.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
✓ Branch 6 taken 2 times.
✓ Branch 7 taken 2 times.
|
3392 | for (j=0;j<n;j++) PetscCall(PetscArraycpy(Q+i*l+j*ldq,C+j*l,l)); |
464 | } | ||
465 | 1224 | sz = m%l; | |
466 |
1/2✓ Branch 0 taken 10 times.
✗ Branch 1 not taken.
|
1224 | if (sz) { |
467 |
10/20✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✓ Branch 4 taken 2 times.
✗ Branch 5 not taken.
✗ Branch 6 not taken.
✓ Branch 7 taken 2 times.
✓ Branch 8 taken 2 times.
✗ Branch 9 not taken.
✗ Branch 10 not taken.
✓ Branch 11 taken 2 times.
✓ Branch 12 taken 2 times.
✗ Branch 13 not taken.
✓ Branch 14 taken 2 times.
✗ Branch 15 not taken.
✗ Branch 16 not taken.
✓ Branch 17 taken 2 times.
✗ Branch 18 not taken.
✗ Branch 19 not taken.
|
1224 | PetscCallBLAS("BLASgemm",BLASgemm_("N","N",&sz,&n,&n,&one,Q+(m/l)*l,&ldq,Qhalf,&l,&zero,C,&l)); |
468 |
7/8✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 10 times.
✓ Branch 3 taken 8 times.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
✓ Branch 6 taken 2 times.
✓ Branch 7 taken 2 times.
|
6960 | for (j=0;j<n;j++) PetscCall(PetscArraycpy(Q+(m/l)*l+j*ldq,C+j*l,sz)); |
469 | } | ||
470 | } | ||
471 | } | ||
472 | |||
473 |
14/28✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✓ Branch 3 taken 8 times.
✓ Branch 4 taken 2 times.
✗ Branch 5 not taken.
✓ Branch 6 taken 2 times.
✗ Branch 7 not taken.
✓ Branch 8 taken 2 times.
✗ Branch 9 not taken.
✗ Branch 10 not taken.
✓ Branch 11 taken 2 times.
✓ Branch 12 taken 2 times.
✗ Branch 13 not taken.
✗ Branch 14 not taken.
✓ Branch 15 taken 2 times.
✓ Branch 16 taken 2 times.
✗ Branch 17 not taken.
✓ Branch 18 taken 2 times.
✗ Branch 19 not taken.
✗ Branch 20 not taken.
✓ Branch 21 taken 2 times.
✗ Branch 22 not taken.
✗ Branch 23 not taken.
✗ Branch 24 not taken.
✓ Branch 25 taken 2 times.
✗ Branch 26 not taken.
✗ Branch 27 not taken.
|
1224 | PetscCallMPI(MPI_Type_free(&tmat)); |
474 | } | ||
475 | |||
476 |
3/6✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
|
2101 | PetscCall(PetscLogFlops(3.0*m*n*n)); |
477 |
4/6✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
|
2101 | PetscCall(PetscFPTrapPop()); |
478 |
6/12✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✗ Branch 2 not taken.
✓ Branch 3 taken 2 times.
✓ Branch 4 taken 2 times.
✗ Branch 5 not taken.
✓ Branch 6 taken 2 times.
✗ Branch 7 not taken.
✓ Branch 8 taken 2 times.
✗ Branch 9 not taken.
✗ Branch 10 not taken.
✓ Branch 11 taken 2 times.
|
380 | PetscFunctionReturn(PETSC_SUCCESS); |
479 | } | ||
480 | |||
481 | /* | ||
482 | Reduction operation to compute [~,Rout]=qr([Rin1;Rin2]) in the TSQR algorithm; | ||
483 | all matrices are upper triangular stored in packed format | ||
484 | */ | ||
485 | 384 | SLEPC_EXTERN void MPIAPI SlepcGivensPacked(void *in,void *inout,PetscMPIInt *len,MPI_Datatype *datatype) | |
486 | { | ||
487 | 384 | PetscBLASInt n,i,j,k,one=1; | |
488 | 384 | PetscMPIInt tsize; | |
489 | 384 | PetscScalar v,s,*R2=(PetscScalar*)in,*R1=(PetscScalar*)inout; | |
490 | 384 | PetscReal c; | |
491 | |||
492 |
1/2✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
|
384 | PetscFunctionBegin; |
493 |
12/24✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✓ Branch 4 taken 2 times.
✗ Branch 5 not taken.
✓ Branch 6 taken 2 times.
✗ Branch 7 not taken.
✓ Branch 8 taken 2 times.
✗ Branch 9 not taken.
✓ Branch 10 taken 2 times.
✗ Branch 11 not taken.
✗ Branch 12 not taken.
✓ Branch 13 taken 2 times.
✓ Branch 14 taken 2 times.
✗ Branch 15 not taken.
✓ Branch 16 taken 2 times.
✗ Branch 17 not taken.
✗ Branch 18 not taken.
✓ Branch 19 taken 2 times.
✗ Branch 20 not taken.
✗ Branch 21 not taken.
✗ Branch 22 not taken.
✓ Branch 23 taken 2 times.
|
384 | PetscCallMPIAbort(PETSC_COMM_SELF,MPI_Type_size(*datatype,&tsize)); /* we assume len=1 */ |
494 | 384 | tsize /= sizeof(PetscScalar); | |
495 | 384 | n = (-1+(PetscBLASInt)PetscSqrtReal(1+8*tsize))/2; | |
496 |
2/2✓ Branch 0 taken 10 times.
✓ Branch 1 taken 10 times.
|
1920 | for (j=0;j<n;j++) { |
497 |
2/2✓ Branch 0 taken 10 times.
✓ Branch 1 taken 10 times.
|
6144 | for (i=0;i<=j;i++) { |
498 | 4608 | LAPACKlartg_(R1+(2*n-j-1)*j/2+j,R2+(2*n-i-1)*i/2+j,&c,&s,&v); | |
499 | 4608 | R1[(2*n-j-1)*j/2+j] = v; | |
500 | 4608 | k = n-j-1; | |
501 |
2/2✓ Branch 0 taken 10 times.
✓ Branch 1 taken 10 times.
|
4608 | if (k) BLASrot_(&k,R1+(2*n-j-1)*j/2+j+1,&one,R2+(2*n-i-1)*i/2+j+1,&one,&c,&s); |
502 | } | ||
503 | } | ||
504 |
6/12✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✗ Branch 2 not taken.
✓ Branch 3 taken 2 times.
✓ Branch 4 taken 2 times.
✗ Branch 5 not taken.
✓ Branch 6 taken 2 times.
✗ Branch 7 not taken.
✓ Branch 8 taken 2 times.
✗ Branch 9 not taken.
✗ Branch 10 not taken.
✓ Branch 11 taken 2 times.
|
384 | PetscFunctionReturnVoid(); |
505 | } | ||
506 | |||
507 | /* | ||
508 | Computes the R factor of the QR factorization of an mxn matrix via parallel TSQR | ||
509 | */ | ||
510 | 794 | PetscErrorCode BVOrthogonalize_LAPACK_TSQR_OnlyR(BV bv,PetscInt m_,PetscInt n_,PetscScalar *Q,PetscInt ldq_,PetscScalar *R,PetscInt ldr) | |
511 | { | ||
512 | 794 | PetscInt worklen; | |
513 | 794 | PetscBLASInt m,n,ldq,i,j,s,nb,lwork,info; | |
514 | 794 | PetscScalar *tau,*work,*A=NULL,*R1=NULL,*R2=NULL; | |
515 | 794 | PetscMPIInt size,count; | |
516 | 794 | MPI_Datatype tmat; | |
517 | |||
518 |
1/2✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
|
794 | PetscFunctionBegin; |
519 |
4/6✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
|
794 | PetscCall(PetscFPTrapPush(PETSC_FP_TRAP_OFF)); |
520 |
4/6✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
|
794 | PetscCall(PetscBLASIntCast(m_,&m)); |
521 |
4/6✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
|
794 | PetscCall(PetscBLASIntCast(n_,&n)); |
522 |
4/6✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
|
794 | PetscCall(PetscBLASIntCast(ldq_,&ldq)); |
523 | 794 | nb = 16; | |
524 | 794 | s = n+n*(n-1)/2; /* length of packed triangular matrix */ | |
525 |
14/28✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✓ Branch 3 taken 8 times.
✓ Branch 4 taken 2 times.
✗ Branch 5 not taken.
✓ Branch 6 taken 2 times.
✗ Branch 7 not taken.
✓ Branch 8 taken 2 times.
✗ Branch 9 not taken.
✗ Branch 10 not taken.
✓ Branch 11 taken 2 times.
✓ Branch 12 taken 2 times.
✗ Branch 13 not taken.
✗ Branch 14 not taken.
✓ Branch 15 taken 2 times.
✓ Branch 16 taken 2 times.
✗ Branch 17 not taken.
✓ Branch 18 taken 2 times.
✗ Branch 19 not taken.
✗ Branch 20 not taken.
✓ Branch 21 taken 2 times.
✗ Branch 22 not taken.
✗ Branch 23 not taken.
✗ Branch 24 not taken.
✓ Branch 25 taken 2 times.
✗ Branch 26 not taken.
✗ Branch 27 not taken.
|
794 | PetscCallMPI(MPI_Comm_size(PetscObjectComm((PetscObject)bv),&size)); |
526 | 794 | worklen = n+n*nb+2*s+ldq*n; | |
527 |
4/6✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
|
794 | PetscCall(BVAllocateWork_Private(bv,worklen)); |
528 | 794 | tau = bv->work; | |
529 | 794 | work = bv->work+n; | |
530 | 794 | R1 = bv->work+n+n*nb; | |
531 | 794 | R2 = bv->work+n+n*nb+s; | |
532 | 794 | A = bv->work+n+n*nb+2*s; | |
533 |
4/6✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
|
794 | PetscCall(PetscBLASIntCast(n*nb,&lwork)); |
534 |
4/6✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
|
794 | PetscCall(PetscArraycpy(A,Q,ldq*n)); |
535 | |||
536 | /* Compute QR */ | ||
537 |
10/20✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✓ Branch 4 taken 2 times.
✗ Branch 5 not taken.
✗ Branch 6 not taken.
✓ Branch 7 taken 2 times.
✓ Branch 8 taken 2 times.
✗ Branch 9 not taken.
✗ Branch 10 not taken.
✓ Branch 11 taken 2 times.
✓ Branch 12 taken 2 times.
✗ Branch 13 not taken.
✓ Branch 14 taken 2 times.
✗ Branch 15 not taken.
✗ Branch 16 not taken.
✓ Branch 17 taken 2 times.
✗ Branch 18 not taken.
✗ Branch 19 not taken.
|
794 | PetscCallBLAS("LAPACKgeqrf",LAPACKgeqrf_(&m,&n,A,&ldq,tau,work,&lwork,&info)); |
538 |
1/4✗ Branch 0 not taken.
✓ Branch 1 taken 10 times.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
|
794 | SlepcCheckLapackInfo("geqrf",info); |
539 | |||
540 |
2/2✓ Branch 0 taken 10 times.
✓ Branch 1 taken 10 times.
|
794 | if (size==1) { |
541 | /* Extract R */ | ||
542 |
2/2✓ Branch 0 taken 10 times.
✓ Branch 1 taken 10 times.
|
6970 | for (j=0;j<n;j++) { |
543 |
2/2✓ Branch 0 taken 10 times.
✓ Branch 1 taken 10 times.
|
62320 | for (i=0;i<=PetscMin(j,m-1);i++) R[i+j*ldr] = A[i+j*ldq]; |
544 |
2/2✓ Branch 0 taken 10 times.
✓ Branch 1 taken 10 times.
|
55760 | for (i=PetscMin(j,m-1)+1;i<n;i++) R[i+j*ldr] = 0.0; |
545 | } | ||
546 | } else { | ||
547 | /* Use MPI reduction operation to obtain global R */ | ||
548 |
4/6✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
|
384 | PetscCall(PetscMPIIntCast(s,&count)); |
549 |
14/28✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✓ Branch 3 taken 8 times.
✓ Branch 4 taken 2 times.
✗ Branch 5 not taken.
✓ Branch 6 taken 2 times.
✗ Branch 7 not taken.
✓ Branch 8 taken 2 times.
✗ Branch 9 not taken.
✗ Branch 10 not taken.
✓ Branch 11 taken 2 times.
✓ Branch 12 taken 2 times.
✗ Branch 13 not taken.
✗ Branch 14 not taken.
✓ Branch 15 taken 2 times.
✓ Branch 16 taken 2 times.
✗ Branch 17 not taken.
✓ Branch 18 taken 2 times.
✗ Branch 19 not taken.
✗ Branch 20 not taken.
✓ Branch 21 taken 2 times.
✗ Branch 22 not taken.
✗ Branch 23 not taken.
✗ Branch 24 not taken.
✓ Branch 25 taken 2 times.
✗ Branch 26 not taken.
✗ Branch 27 not taken.
|
384 | PetscCallMPI(MPI_Type_contiguous(count,MPIU_SCALAR,&tmat)); |
550 |
14/28✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✓ Branch 3 taken 8 times.
✓ Branch 4 taken 2 times.
✗ Branch 5 not taken.
✓ Branch 6 taken 2 times.
✗ Branch 7 not taken.
✓ Branch 8 taken 2 times.
✗ Branch 9 not taken.
✗ Branch 10 not taken.
✓ Branch 11 taken 2 times.
✓ Branch 12 taken 2 times.
✗ Branch 13 not taken.
✗ Branch 14 not taken.
✓ Branch 15 taken 2 times.
✓ Branch 16 taken 2 times.
✗ Branch 17 not taken.
✓ Branch 18 taken 2 times.
✗ Branch 19 not taken.
✗ Branch 20 not taken.
✓ Branch 21 taken 2 times.
✗ Branch 22 not taken.
✗ Branch 23 not taken.
✗ Branch 24 not taken.
✓ Branch 25 taken 2 times.
✗ Branch 26 not taken.
✗ Branch 27 not taken.
|
384 | PetscCallMPI(MPI_Type_commit(&tmat)); |
551 |
2/2✓ Branch 0 taken 10 times.
✓ Branch 1 taken 10 times.
|
1920 | for (i=0;i<n;i++) { |
552 |
3/4✓ Branch 0 taken 10 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 10 times.
✓ Branch 3 taken 10 times.
|
6144 | for (j=i;j<n;j++) R1[(2*n-i-1)*i/2+j] = (i<m)?A[i+j*ldq]:0.0; |
553 | } | ||
554 |
28/58✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✓ Branch 3 taken 8 times.
✓ Branch 4 taken 2 times.
✗ Branch 5 not taken.
✓ Branch 6 taken 2 times.
✗ Branch 7 not taken.
✓ Branch 8 taken 2 times.
✗ Branch 9 not taken.
✓ Branch 10 taken 2 times.
✗ Branch 11 not taken.
✗ Branch 12 not taken.
✓ Branch 13 taken 2 times.
✗ Branch 14 not taken.
✓ Branch 15 taken 2 times.
✓ Branch 16 taken 2 times.
✗ Branch 17 not taken.
✗ Branch 18 not taken.
✓ Branch 19 taken 2 times.
✓ Branch 20 taken 2 times.
✗ Branch 21 not taken.
✗ Branch 22 not taken.
✓ Branch 23 taken 2 times.
✓ Branch 24 taken 2 times.
✗ Branch 25 not taken.
✗ Branch 26 not taken.
✓ Branch 27 taken 2 times.
✓ Branch 28 taken 2 times.
✗ Branch 29 not taken.
✗ Branch 30 not taken.
✓ Branch 31 taken 2 times.
✓ Branch 32 taken 2 times.
✗ Branch 33 not taken.
✓ Branch 34 taken 2 times.
✗ Branch 35 not taken.
✗ Branch 36 not taken.
✓ Branch 37 taken 2 times.
✗ Branch 38 not taken.
✗ Branch 39 not taken.
✗ Branch 40 not taken.
✓ Branch 41 taken 2 times.
✗ Branch 42 not taken.
✗ Branch 43 not taken.
✗ Branch 44 not taken.
✓ Branch 45 taken 2 times.
✓ Branch 46 taken 2 times.
✗ Branch 47 not taken.
✗ Branch 48 not taken.
✓ Branch 49 taken 2 times.
✓ Branch 50 taken 2 times.
✗ Branch 51 not taken.
✓ Branch 52 taken 2 times.
✗ Branch 53 not taken.
✗ Branch 54 not taken.
✓ Branch 55 taken 2 times.
✗ Branch 56 not taken.
✗ Branch 57 not taken.
|
384 | PetscCallMPI(MPIU_Allreduce(R1,R2,1,tmat,MPIU_TSQR,PetscObjectComm((PetscObject)bv))); |
555 |
2/2✓ Branch 0 taken 10 times.
✓ Branch 1 taken 10 times.
|
1920 | for (i=0;i<n;i++) { |
556 |
2/2✓ Branch 0 taken 10 times.
✓ Branch 1 taken 10 times.
|
4608 | for (j=0;j<i;j++) R[i+j*ldr] = 0.0; |
557 |
2/2✓ Branch 0 taken 10 times.
✓ Branch 1 taken 10 times.
|
6144 | for (j=i;j<n;j++) R[i+j*ldr] = R2[(2*n-i-1)*i/2+j]; |
558 | } | ||
559 |
14/28✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✓ Branch 3 taken 8 times.
✓ Branch 4 taken 2 times.
✗ Branch 5 not taken.
✓ Branch 6 taken 2 times.
✗ Branch 7 not taken.
✓ Branch 8 taken 2 times.
✗ Branch 9 not taken.
✗ Branch 10 not taken.
✓ Branch 11 taken 2 times.
✓ Branch 12 taken 2 times.
✗ Branch 13 not taken.
✗ Branch 14 not taken.
✓ Branch 15 taken 2 times.
✓ Branch 16 taken 2 times.
✗ Branch 17 not taken.
✓ Branch 18 taken 2 times.
✗ Branch 19 not taken.
✗ Branch 20 not taken.
✓ Branch 21 taken 2 times.
✗ Branch 22 not taken.
✗ Branch 23 not taken.
✗ Branch 24 not taken.
✓ Branch 25 taken 2 times.
✗ Branch 26 not taken.
✗ Branch 27 not taken.
|
384 | PetscCallMPI(MPI_Type_free(&tmat)); |
560 | } | ||
561 | |||
562 |
3/6✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
|
794 | PetscCall(PetscLogFlops(3.0*m*n*n)); |
563 |
4/6✓ Branch 0 taken 2 times.
✓ Branch 1 taken 8 times.
✓ Branch 2 taken 2 times.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✓ Branch 5 taken 2 times.
|
794 | PetscCall(PetscFPTrapPop()); |
564 |
6/12✓ Branch 0 taken 2 times.
✗ Branch 1 not taken.
✗ Branch 2 not taken.
✓ Branch 3 taken 2 times.
✓ Branch 4 taken 2 times.
✗ Branch 5 not taken.
✓ Branch 6 taken 2 times.
✗ Branch 7 not taken.
✓ Branch 8 taken 2 times.
✗ Branch 9 not taken.
✗ Branch 10 not taken.
✓ Branch 11 taken 2 times.
|
146 | PetscFunctionReturn(PETSC_SUCCESS); |
565 | } | ||
566 |