bug #1152: Fix data race in static initialization of blas

This commit is contained in:
Benoit Jacob 2016-01-26 11:44:16 -05:00
parent 44d4674955
commit 639b1d864a
4 changed files with 388 additions and 438 deletions

View File

@ -19,19 +19,12 @@
int EIGEN_BLAS_FUNC(hemv)(char *uplo, int *n, RealScalar *palpha, RealScalar *pa, int *lda, RealScalar *px, int *incx, RealScalar *pbeta, RealScalar *py, int *incy) int EIGEN_BLAS_FUNC(hemv)(char *uplo, int *n, RealScalar *palpha, RealScalar *pa, int *lda, RealScalar *px, int *incx, RealScalar *pbeta, RealScalar *py, int *incy)
{ {
typedef void (*functype)(int, const Scalar*, int, const Scalar*, Scalar*, Scalar); typedef void (*functype)(int, const Scalar*, int, const Scalar*, Scalar*, Scalar);
static functype func[2]; static const functype func[2] = {
// array index: UP
static bool init = false; (internal::selfadjoint_matrix_vector_product<Scalar,int,ColMajor,Upper,false,false>::run),
if(!init) // array index: LO
{ (internal::selfadjoint_matrix_vector_product<Scalar,int,ColMajor,Lower,false,false>::run),
for(int k=0; k<2; ++k) };
func[k] = 0;
func[UP] = (internal::selfadjoint_matrix_vector_product<Scalar,int,ColMajor,Upper,false,false>::run);
func[LO] = (internal::selfadjoint_matrix_vector_product<Scalar,int,ColMajor,Lower,false,false>::run);
init = true;
}
Scalar* a = reinterpret_cast<Scalar*>(pa); Scalar* a = reinterpret_cast<Scalar*>(pa);
Scalar* x = reinterpret_cast<Scalar*>(px); Scalar* x = reinterpret_cast<Scalar*>(px);
@ -111,19 +104,12 @@ int EIGEN_BLAS_FUNC(hemv)(char *uplo, int *n, RealScalar *palpha, RealScalar *pa
int EIGEN_BLAS_FUNC(hpr)(char *uplo, int *n, RealScalar *palpha, RealScalar *px, int *incx, RealScalar *pap) int EIGEN_BLAS_FUNC(hpr)(char *uplo, int *n, RealScalar *palpha, RealScalar *px, int *incx, RealScalar *pap)
{ {
typedef void (*functype)(int, Scalar*, const Scalar*, RealScalar); typedef void (*functype)(int, Scalar*, const Scalar*, RealScalar);
static functype func[2]; static const functype func[2] = {
// array index: UP
static bool init = false; (internal::selfadjoint_packed_rank1_update<Scalar,int,ColMajor,Upper,false,Conj>::run),
if(!init) // array index: LO
{ (internal::selfadjoint_packed_rank1_update<Scalar,int,ColMajor,Lower,false,Conj>::run),
for(int k=0; k<2; ++k) };
func[k] = 0;
func[UP] = (internal::selfadjoint_packed_rank1_update<Scalar,int,ColMajor,Upper,false,Conj>::run);
func[LO] = (internal::selfadjoint_packed_rank1_update<Scalar,int,ColMajor,Lower,false,Conj>::run);
init = true;
}
Scalar* x = reinterpret_cast<Scalar*>(px); Scalar* x = reinterpret_cast<Scalar*>(px);
Scalar* ap = reinterpret_cast<Scalar*>(pap); Scalar* ap = reinterpret_cast<Scalar*>(pap);
@ -162,19 +148,12 @@ int EIGEN_BLAS_FUNC(hpr)(char *uplo, int *n, RealScalar *palpha, RealScalar *px,
int EIGEN_BLAS_FUNC(hpr2)(char *uplo, int *n, RealScalar *palpha, RealScalar *px, int *incx, RealScalar *py, int *incy, RealScalar *pap) int EIGEN_BLAS_FUNC(hpr2)(char *uplo, int *n, RealScalar *palpha, RealScalar *px, int *incx, RealScalar *py, int *incy, RealScalar *pap)
{ {
typedef void (*functype)(int, Scalar*, const Scalar*, const Scalar*, Scalar); typedef void (*functype)(int, Scalar*, const Scalar*, const Scalar*, Scalar);
static functype func[2]; static const functype func[2] = {
// array index: UP
static bool init = false; (internal::packed_rank2_update_selector<Scalar,int,Upper>::run),
if(!init) // array index: LO
{ (internal::packed_rank2_update_selector<Scalar,int,Lower>::run),
for(int k=0; k<2; ++k) };
func[k] = 0;
func[UP] = (internal::packed_rank2_update_selector<Scalar,int,Upper>::run);
func[LO] = (internal::packed_rank2_update_selector<Scalar,int,Lower>::run);
init = true;
}
Scalar* x = reinterpret_cast<Scalar*>(px); Scalar* x = reinterpret_cast<Scalar*>(px);
Scalar* y = reinterpret_cast<Scalar*>(py); Scalar* y = reinterpret_cast<Scalar*>(py);
@ -217,19 +196,12 @@ int EIGEN_BLAS_FUNC(hpr2)(char *uplo, int *n, RealScalar *palpha, RealScalar *px
int EIGEN_BLAS_FUNC(her)(char *uplo, int *n, RealScalar *palpha, RealScalar *px, int *incx, RealScalar *pa, int *lda) int EIGEN_BLAS_FUNC(her)(char *uplo, int *n, RealScalar *palpha, RealScalar *px, int *incx, RealScalar *pa, int *lda)
{ {
typedef void (*functype)(int, Scalar*, int, const Scalar*, const Scalar*, const Scalar&); typedef void (*functype)(int, Scalar*, int, const Scalar*, const Scalar*, const Scalar&);
static functype func[2]; static const functype func[2] = {
// array index: UP
static bool init = false; (selfadjoint_rank1_update<Scalar,int,ColMajor,Upper,false,Conj>::run),
if(!init) // array index: LO
{ (selfadjoint_rank1_update<Scalar,int,ColMajor,Lower,false,Conj>::run),
for(int k=0; k<2; ++k) };
func[k] = 0;
func[UP] = (selfadjoint_rank1_update<Scalar,int,ColMajor,Upper,false,Conj>::run);
func[LO] = (selfadjoint_rank1_update<Scalar,int,ColMajor,Lower,false,Conj>::run);
init = true;
}
Scalar* x = reinterpret_cast<Scalar*>(px); Scalar* x = reinterpret_cast<Scalar*>(px);
Scalar* a = reinterpret_cast<Scalar*>(pa); Scalar* a = reinterpret_cast<Scalar*>(pa);
@ -271,19 +243,12 @@ int EIGEN_BLAS_FUNC(her)(char *uplo, int *n, RealScalar *palpha, RealScalar *px,
int EIGEN_BLAS_FUNC(her2)(char *uplo, int *n, RealScalar *palpha, RealScalar *px, int *incx, RealScalar *py, int *incy, RealScalar *pa, int *lda) int EIGEN_BLAS_FUNC(her2)(char *uplo, int *n, RealScalar *palpha, RealScalar *px, int *incx, RealScalar *py, int *incy, RealScalar *pa, int *lda)
{ {
typedef void (*functype)(int, Scalar*, int, const Scalar*, const Scalar*, Scalar); typedef void (*functype)(int, Scalar*, int, const Scalar*, const Scalar*, Scalar);
static functype func[2]; static const functype func[2] = {
// array index: UP
static bool init = false; (internal::rank2_update_selector<Scalar,int,Upper>::run),
if(!init) // array index: LO
{ (internal::rank2_update_selector<Scalar,int,Lower>::run),
for(int k=0; k<2; ++k) };
func[k] = 0;
func[UP] = (internal::rank2_update_selector<Scalar,int,Upper>::run);
func[LO] = (internal::rank2_update_selector<Scalar,int,Lower>::run);
init = true;
}
Scalar* x = reinterpret_cast<Scalar*>(px); Scalar* x = reinterpret_cast<Scalar*>(px);
Scalar* y = reinterpret_cast<Scalar*>(py); Scalar* y = reinterpret_cast<Scalar*>(py);

View File

@ -26,20 +26,15 @@ struct general_matrix_vector_product_wrapper
int EIGEN_BLAS_FUNC(gemv)(char *opa, int *m, int *n, RealScalar *palpha, RealScalar *pa, int *lda, RealScalar *pb, int *incb, RealScalar *pbeta, RealScalar *pc, int *incc) int EIGEN_BLAS_FUNC(gemv)(char *opa, int *m, int *n, RealScalar *palpha, RealScalar *pa, int *lda, RealScalar *pb, int *incb, RealScalar *pbeta, RealScalar *pc, int *incc)
{ {
typedef void (*functype)(int, int, const Scalar *, int, const Scalar *, int , Scalar *, int, Scalar); typedef void (*functype)(int, int, const Scalar *, int, const Scalar *, int , Scalar *, int, Scalar);
static functype func[4]; static const functype func[4] = {
// array index: NOTR
static bool init = false; (general_matrix_vector_product_wrapper<int,Scalar,ColMajor,false,false>::run),
if(!init) // array index: TR
{ (general_matrix_vector_product_wrapper<int,Scalar,RowMajor,false,false>::run),
for(int k=0; k<4; ++k) // array index: ADJ
func[k] = 0; (general_matrix_vector_product_wrapper<int,Scalar,RowMajor,Conj ,false>::run),
0
func[NOTR] = (general_matrix_vector_product_wrapper<int,Scalar,ColMajor,false,false>::run); };
func[TR ] = (general_matrix_vector_product_wrapper<int,Scalar,RowMajor,false,false>::run);
func[ADJ ] = (general_matrix_vector_product_wrapper<int,Scalar,RowMajor,Conj ,false>::run);
init = true;
}
Scalar* a = reinterpret_cast<Scalar*>(pa); Scalar* a = reinterpret_cast<Scalar*>(pa);
Scalar* b = reinterpret_cast<Scalar*>(pb); Scalar* b = reinterpret_cast<Scalar*>(pb);
@ -90,32 +85,36 @@ int EIGEN_BLAS_FUNC(gemv)(char *opa, int *m, int *n, RealScalar *palpha, RealSca
int EIGEN_BLAS_FUNC(trsv)(char *uplo, char *opa, char *diag, int *n, RealScalar *pa, int *lda, RealScalar *pb, int *incb) int EIGEN_BLAS_FUNC(trsv)(char *uplo, char *opa, char *diag, int *n, RealScalar *pa, int *lda, RealScalar *pb, int *incb)
{ {
typedef void (*functype)(int, const Scalar *, int, Scalar *); typedef void (*functype)(int, const Scalar *, int, Scalar *);
static functype func[16]; static const functype func[16] = {
// array index: NOTR | (UP << 2) | (NUNIT << 3)
static bool init = false; (internal::triangular_solve_vector<Scalar,Scalar,int,OnTheLeft, Upper|0, false,ColMajor>::run),
if(!init) // array index: TR | (UP << 2) | (NUNIT << 3)
{ (internal::triangular_solve_vector<Scalar,Scalar,int,OnTheLeft, Lower|0, false,RowMajor>::run),
for(int k=0; k<16; ++k) // array index: ADJ | (UP << 2) | (NUNIT << 3)
func[k] = 0; (internal::triangular_solve_vector<Scalar,Scalar,int,OnTheLeft, Lower|0, Conj, RowMajor>::run),
0,
func[NOTR | (UP << 2) | (NUNIT << 3)] = (internal::triangular_solve_vector<Scalar,Scalar,int,OnTheLeft, Upper|0, false,ColMajor>::run); // array index: NOTR | (LO << 2) | (NUNIT << 3)
func[TR | (UP << 2) | (NUNIT << 3)] = (internal::triangular_solve_vector<Scalar,Scalar,int,OnTheLeft, Lower|0, false,RowMajor>::run); (internal::triangular_solve_vector<Scalar,Scalar,int,OnTheLeft, Lower|0, false,ColMajor>::run),
func[ADJ | (UP << 2) | (NUNIT << 3)] = (internal::triangular_solve_vector<Scalar,Scalar,int,OnTheLeft, Lower|0, Conj, RowMajor>::run); // array index: TR | (LO << 2) | (NUNIT << 3)
(internal::triangular_solve_vector<Scalar,Scalar,int,OnTheLeft, Upper|0, false,RowMajor>::run),
func[NOTR | (LO << 2) | (NUNIT << 3)] = (internal::triangular_solve_vector<Scalar,Scalar,int,OnTheLeft, Lower|0, false,ColMajor>::run); // array index: ADJ | (LO << 2) | (NUNIT << 3)
func[TR | (LO << 2) | (NUNIT << 3)] = (internal::triangular_solve_vector<Scalar,Scalar,int,OnTheLeft, Upper|0, false,RowMajor>::run); (internal::triangular_solve_vector<Scalar,Scalar,int,OnTheLeft, Upper|0, Conj, RowMajor>::run),
func[ADJ | (LO << 2) | (NUNIT << 3)] = (internal::triangular_solve_vector<Scalar,Scalar,int,OnTheLeft, Upper|0, Conj, RowMajor>::run); 0,
// array index: NOTR | (UP << 2) | (UNIT << 3)
func[NOTR | (UP << 2) | (UNIT << 3)] = (internal::triangular_solve_vector<Scalar,Scalar,int,OnTheLeft, Upper|UnitDiag,false,ColMajor>::run); (internal::triangular_solve_vector<Scalar,Scalar,int,OnTheLeft, Upper|UnitDiag,false,ColMajor>::run),
func[TR | (UP << 2) | (UNIT << 3)] = (internal::triangular_solve_vector<Scalar,Scalar,int,OnTheLeft, Lower|UnitDiag,false,RowMajor>::run); // array index: TR | (UP << 2) | (UNIT << 3)
func[ADJ | (UP << 2) | (UNIT << 3)] = (internal::triangular_solve_vector<Scalar,Scalar,int,OnTheLeft, Lower|UnitDiag,Conj, RowMajor>::run); (internal::triangular_solve_vector<Scalar,Scalar,int,OnTheLeft, Lower|UnitDiag,false,RowMajor>::run),
// array index: ADJ | (UP << 2) | (UNIT << 3)
func[NOTR | (LO << 2) | (UNIT << 3)] = (internal::triangular_solve_vector<Scalar,Scalar,int,OnTheLeft, Lower|UnitDiag,false,ColMajor>::run); (internal::triangular_solve_vector<Scalar,Scalar,int,OnTheLeft, Lower|UnitDiag,Conj, RowMajor>::run),
func[TR | (LO << 2) | (UNIT << 3)] = (internal::triangular_solve_vector<Scalar,Scalar,int,OnTheLeft, Upper|UnitDiag,false,RowMajor>::run); 0,
func[ADJ | (LO << 2) | (UNIT << 3)] = (internal::triangular_solve_vector<Scalar,Scalar,int,OnTheLeft, Upper|UnitDiag,Conj, RowMajor>::run); // array index: NOTR | (LO << 2) | (UNIT << 3)
(internal::triangular_solve_vector<Scalar,Scalar,int,OnTheLeft, Lower|UnitDiag,false,ColMajor>::run),
init = true; // array index: TR | (LO << 2) | (UNIT << 3)
} (internal::triangular_solve_vector<Scalar,Scalar,int,OnTheLeft, Upper|UnitDiag,false,RowMajor>::run),
// array index: ADJ | (LO << 2) | (UNIT << 3)
(internal::triangular_solve_vector<Scalar,Scalar,int,OnTheLeft, Upper|UnitDiag,Conj, RowMajor>::run),
0
};
Scalar* a = reinterpret_cast<Scalar*>(pa); Scalar* a = reinterpret_cast<Scalar*>(pa);
Scalar* b = reinterpret_cast<Scalar*>(pb); Scalar* b = reinterpret_cast<Scalar*>(pb);
@ -145,32 +144,36 @@ int EIGEN_BLAS_FUNC(trsv)(char *uplo, char *opa, char *diag, int *n, RealScalar
int EIGEN_BLAS_FUNC(trmv)(char *uplo, char *opa, char *diag, int *n, RealScalar *pa, int *lda, RealScalar *pb, int *incb) int EIGEN_BLAS_FUNC(trmv)(char *uplo, char *opa, char *diag, int *n, RealScalar *pa, int *lda, RealScalar *pb, int *incb)
{ {
typedef void (*functype)(int, int, const Scalar *, int, const Scalar *, int, Scalar *, int, const Scalar&); typedef void (*functype)(int, int, const Scalar *, int, const Scalar *, int, Scalar *, int, const Scalar&);
static functype func[16]; static const functype func[16] = {
// array index: NOTR | (UP << 2) | (NUNIT << 3)
static bool init = false; (internal::triangular_matrix_vector_product<int,Upper|0, Scalar,false,Scalar,false,ColMajor>::run),
if(!init) // array index: TR | (UP << 2) | (NUNIT << 3)
{ (internal::triangular_matrix_vector_product<int,Lower|0, Scalar,false,Scalar,false,RowMajor>::run),
for(int k=0; k<16; ++k) // array index: ADJ | (UP << 2) | (NUNIT << 3)
func[k] = 0; (internal::triangular_matrix_vector_product<int,Lower|0, Scalar,Conj, Scalar,false,RowMajor>::run),
0,
func[NOTR | (UP << 2) | (NUNIT << 3)] = (internal::triangular_matrix_vector_product<int,Upper|0, Scalar,false,Scalar,false,ColMajor>::run); // array index: NOTR | (LO << 2) | (NUNIT << 3)
func[TR | (UP << 2) | (NUNIT << 3)] = (internal::triangular_matrix_vector_product<int,Lower|0, Scalar,false,Scalar,false,RowMajor>::run); (internal::triangular_matrix_vector_product<int,Lower|0, Scalar,false,Scalar,false,ColMajor>::run),
func[ADJ | (UP << 2) | (NUNIT << 3)] = (internal::triangular_matrix_vector_product<int,Lower|0, Scalar,Conj, Scalar,false,RowMajor>::run); // array index: TR | (LO << 2) | (NUNIT << 3)
(internal::triangular_matrix_vector_product<int,Upper|0, Scalar,false,Scalar,false,RowMajor>::run),
func[NOTR | (LO << 2) | (NUNIT << 3)] = (internal::triangular_matrix_vector_product<int,Lower|0, Scalar,false,Scalar,false,ColMajor>::run); // array index: ADJ | (LO << 2) | (NUNIT << 3)
func[TR | (LO << 2) | (NUNIT << 3)] = (internal::triangular_matrix_vector_product<int,Upper|0, Scalar,false,Scalar,false,RowMajor>::run); (internal::triangular_matrix_vector_product<int,Upper|0, Scalar,Conj, Scalar,false,RowMajor>::run),
func[ADJ | (LO << 2) | (NUNIT << 3)] = (internal::triangular_matrix_vector_product<int,Upper|0, Scalar,Conj, Scalar,false,RowMajor>::run); 0,
// array index: NOTR | (UP << 2) | (UNIT << 3)
func[NOTR | (UP << 2) | (UNIT << 3)] = (internal::triangular_matrix_vector_product<int,Upper|UnitDiag,Scalar,false,Scalar,false,ColMajor>::run); (internal::triangular_matrix_vector_product<int,Upper|UnitDiag,Scalar,false,Scalar,false,ColMajor>::run),
func[TR | (UP << 2) | (UNIT << 3)] = (internal::triangular_matrix_vector_product<int,Lower|UnitDiag,Scalar,false,Scalar,false,RowMajor>::run); // array index: TR | (UP << 2) | (UNIT << 3)
func[ADJ | (UP << 2) | (UNIT << 3)] = (internal::triangular_matrix_vector_product<int,Lower|UnitDiag,Scalar,Conj, Scalar,false,RowMajor>::run); (internal::triangular_matrix_vector_product<int,Lower|UnitDiag,Scalar,false,Scalar,false,RowMajor>::run),
// array index: ADJ | (UP << 2) | (UNIT << 3)
func[NOTR | (LO << 2) | (UNIT << 3)] = (internal::triangular_matrix_vector_product<int,Lower|UnitDiag,Scalar,false,Scalar,false,ColMajor>::run); (internal::triangular_matrix_vector_product<int,Lower|UnitDiag,Scalar,Conj, Scalar,false,RowMajor>::run),
func[TR | (LO << 2) | (UNIT << 3)] = (internal::triangular_matrix_vector_product<int,Upper|UnitDiag,Scalar,false,Scalar,false,RowMajor>::run); 0,
func[ADJ | (LO << 2) | (UNIT << 3)] = (internal::triangular_matrix_vector_product<int,Upper|UnitDiag,Scalar,Conj, Scalar,false,RowMajor>::run); // array index: NOTR | (LO << 2) | (UNIT << 3)
(internal::triangular_matrix_vector_product<int,Lower|UnitDiag,Scalar,false,Scalar,false,ColMajor>::run),
init = true; // array index: TR | (LO << 2) | (UNIT << 3)
} (internal::triangular_matrix_vector_product<int,Upper|UnitDiag,Scalar,false,Scalar,false,RowMajor>::run),
// array index: ADJ | (LO << 2) | (UNIT << 3)
(internal::triangular_matrix_vector_product<int,Upper|UnitDiag,Scalar,Conj, Scalar,false,RowMajor>::run),
0
};
Scalar* a = reinterpret_cast<Scalar*>(pa); Scalar* a = reinterpret_cast<Scalar*>(pa);
Scalar* b = reinterpret_cast<Scalar*>(pb); Scalar* b = reinterpret_cast<Scalar*>(pb);
@ -346,32 +349,36 @@ int EIGEN_BLAS_FUNC(tbmv)(char *uplo, char *opa, char *diag, int *n, int *k, Rea
int EIGEN_BLAS_FUNC(tbsv)(char *uplo, char *op, char *diag, int *n, int *k, RealScalar *pa, int *lda, RealScalar *px, int *incx) int EIGEN_BLAS_FUNC(tbsv)(char *uplo, char *op, char *diag, int *n, int *k, RealScalar *pa, int *lda, RealScalar *px, int *incx)
{ {
typedef void (*functype)(int, int, const Scalar *, int, Scalar *); typedef void (*functype)(int, int, const Scalar *, int, Scalar *);
static functype func[16]; static const functype func[16] = {
// array index: NOTR | (UP << 2) | (NUNIT << 3)
static bool init = false; (internal::band_solve_triangular_selector<int,Upper|0, Scalar,false,Scalar,ColMajor>::run),
if(!init) // array index: TR | (UP << 2) | (NUNIT << 3)
{ (internal::band_solve_triangular_selector<int,Lower|0, Scalar,false,Scalar,RowMajor>::run),
for(int i=0; i<16; ++i) // array index: ADJ | (UP << 2) | (NUNIT << 3)
func[i] = 0; (internal::band_solve_triangular_selector<int,Lower|0, Scalar,Conj, Scalar,RowMajor>::run),
0,
func[NOTR | (UP << 2) | (NUNIT << 3)] = (internal::band_solve_triangular_selector<int,Upper|0, Scalar,false,Scalar,ColMajor>::run); // array index: NOTR | (LO << 2) | (NUNIT << 3)
func[TR | (UP << 2) | (NUNIT << 3)] = (internal::band_solve_triangular_selector<int,Lower|0, Scalar,false,Scalar,RowMajor>::run); (internal::band_solve_triangular_selector<int,Lower|0, Scalar,false,Scalar,ColMajor>::run),
func[ADJ | (UP << 2) | (NUNIT << 3)] = (internal::band_solve_triangular_selector<int,Lower|0, Scalar,Conj, Scalar,RowMajor>::run); // array index: TR | (LO << 2) | (NUNIT << 3)
(internal::band_solve_triangular_selector<int,Upper|0, Scalar,false,Scalar,RowMajor>::run),
func[NOTR | (LO << 2) | (NUNIT << 3)] = (internal::band_solve_triangular_selector<int,Lower|0, Scalar,false,Scalar,ColMajor>::run); // array index: ADJ | (LO << 2) | (NUNIT << 3)
func[TR | (LO << 2) | (NUNIT << 3)] = (internal::band_solve_triangular_selector<int,Upper|0, Scalar,false,Scalar,RowMajor>::run); (internal::band_solve_triangular_selector<int,Upper|0, Scalar,Conj, Scalar,RowMajor>::run),
func[ADJ | (LO << 2) | (NUNIT << 3)] = (internal::band_solve_triangular_selector<int,Upper|0, Scalar,Conj, Scalar,RowMajor>::run); 0,
// array index: NOTR | (UP << 2) | (UNIT << 3)
func[NOTR | (UP << 2) | (UNIT << 3)] = (internal::band_solve_triangular_selector<int,Upper|UnitDiag,Scalar,false,Scalar,ColMajor>::run); (internal::band_solve_triangular_selector<int,Upper|UnitDiag,Scalar,false,Scalar,ColMajor>::run),
func[TR | (UP << 2) | (UNIT << 3)] = (internal::band_solve_triangular_selector<int,Lower|UnitDiag,Scalar,false,Scalar,RowMajor>::run); // array index: TR | (UP << 2) | (UNIT << 3)
func[ADJ | (UP << 2) | (UNIT << 3)] = (internal::band_solve_triangular_selector<int,Lower|UnitDiag,Scalar,Conj, Scalar,RowMajor>::run); (internal::band_solve_triangular_selector<int,Lower|UnitDiag,Scalar,false,Scalar,RowMajor>::run),
// array index: ADJ | (UP << 2) | (UNIT << 3)
func[NOTR | (LO << 2) | (UNIT << 3)] = (internal::band_solve_triangular_selector<int,Lower|UnitDiag,Scalar,false,Scalar,ColMajor>::run); (internal::band_solve_triangular_selector<int,Lower|UnitDiag,Scalar,Conj, Scalar,RowMajor>::run),
func[TR | (LO << 2) | (UNIT << 3)] = (internal::band_solve_triangular_selector<int,Upper|UnitDiag,Scalar,false,Scalar,RowMajor>::run); 0,
func[ADJ | (LO << 2) | (UNIT << 3)] = (internal::band_solve_triangular_selector<int,Upper|UnitDiag,Scalar,Conj, Scalar,RowMajor>::run); // array index: NOTR | (LO << 2) | (UNIT << 3)
(internal::band_solve_triangular_selector<int,Lower|UnitDiag,Scalar,false,Scalar,ColMajor>::run),
init = true; // array index: TR | (LO << 2) | (UNIT << 3)
} (internal::band_solve_triangular_selector<int,Upper|UnitDiag,Scalar,false,Scalar,RowMajor>::run),
// array index: ADJ | (LO << 2) | (UNIT << 3)
(internal::band_solve_triangular_selector<int,Upper|UnitDiag,Scalar,Conj, Scalar,RowMajor>::run),
0,
};
Scalar* a = reinterpret_cast<Scalar*>(pa); Scalar* a = reinterpret_cast<Scalar*>(pa);
Scalar* x = reinterpret_cast<Scalar*>(px); Scalar* x = reinterpret_cast<Scalar*>(px);
@ -416,32 +423,36 @@ int EIGEN_BLAS_FUNC(tbsv)(char *uplo, char *op, char *diag, int *n, int *k, Real
int EIGEN_BLAS_FUNC(tpmv)(char *uplo, char *opa, char *diag, int *n, RealScalar *pap, RealScalar *px, int *incx) int EIGEN_BLAS_FUNC(tpmv)(char *uplo, char *opa, char *diag, int *n, RealScalar *pap, RealScalar *px, int *incx)
{ {
typedef void (*functype)(int, const Scalar*, const Scalar*, Scalar*, Scalar); typedef void (*functype)(int, const Scalar*, const Scalar*, Scalar*, Scalar);
static functype func[16]; static const functype func[16] = {
// array index: NOTR | (UP << 2) | (NUNIT << 3)
static bool init = false; (internal::packed_triangular_matrix_vector_product<int,Upper|0, Scalar,false,Scalar,false,ColMajor>::run),
if(!init) // array index: TR | (UP << 2) | (NUNIT << 3)
{ (internal::packed_triangular_matrix_vector_product<int,Lower|0, Scalar,false,Scalar,false,RowMajor>::run),
for(int k=0; k<16; ++k) // array index: ADJ | (UP << 2) | (NUNIT << 3)
func[k] = 0; (internal::packed_triangular_matrix_vector_product<int,Lower|0, Scalar,Conj, Scalar,false,RowMajor>::run),
0,
func[NOTR | (UP << 2) | (NUNIT << 3)] = (internal::packed_triangular_matrix_vector_product<int,Upper|0, Scalar,false,Scalar,false,ColMajor>::run); // array index: NOTR | (LO << 2) | (NUNIT << 3)
func[TR | (UP << 2) | (NUNIT << 3)] = (internal::packed_triangular_matrix_vector_product<int,Lower|0, Scalar,false,Scalar,false,RowMajor>::run); (internal::packed_triangular_matrix_vector_product<int,Lower|0, Scalar,false,Scalar,false,ColMajor>::run),
func[ADJ | (UP << 2) | (NUNIT << 3)] = (internal::packed_triangular_matrix_vector_product<int,Lower|0, Scalar,Conj, Scalar,false,RowMajor>::run); // array index: TR | (LO << 2) | (NUNIT << 3)
(internal::packed_triangular_matrix_vector_product<int,Upper|0, Scalar,false,Scalar,false,RowMajor>::run),
func[NOTR | (LO << 2) | (NUNIT << 3)] = (internal::packed_triangular_matrix_vector_product<int,Lower|0, Scalar,false,Scalar,false,ColMajor>::run); // array index: ADJ | (LO << 2) | (NUNIT << 3)
func[TR | (LO << 2) | (NUNIT << 3)] = (internal::packed_triangular_matrix_vector_product<int,Upper|0, Scalar,false,Scalar,false,RowMajor>::run); (internal::packed_triangular_matrix_vector_product<int,Upper|0, Scalar,Conj, Scalar,false,RowMajor>::run),
func[ADJ | (LO << 2) | (NUNIT << 3)] = (internal::packed_triangular_matrix_vector_product<int,Upper|0, Scalar,Conj, Scalar,false,RowMajor>::run); 0,
// array index: NOTR | (UP << 2) | (UNIT << 3)
func[NOTR | (UP << 2) | (UNIT << 3)] = (internal::packed_triangular_matrix_vector_product<int,Upper|UnitDiag,Scalar,false,Scalar,false,ColMajor>::run); (internal::packed_triangular_matrix_vector_product<int,Upper|UnitDiag,Scalar,false,Scalar,false,ColMajor>::run),
func[TR | (UP << 2) | (UNIT << 3)] = (internal::packed_triangular_matrix_vector_product<int,Lower|UnitDiag,Scalar,false,Scalar,false,RowMajor>::run); // array index: TR | (UP << 2) | (UNIT << 3)
func[ADJ | (UP << 2) | (UNIT << 3)] = (internal::packed_triangular_matrix_vector_product<int,Lower|UnitDiag,Scalar,Conj, Scalar,false,RowMajor>::run); (internal::packed_triangular_matrix_vector_product<int,Lower|UnitDiag,Scalar,false,Scalar,false,RowMajor>::run),
// array index: ADJ | (UP << 2) | (UNIT << 3)
func[NOTR | (LO << 2) | (UNIT << 3)] = (internal::packed_triangular_matrix_vector_product<int,Lower|UnitDiag,Scalar,false,Scalar,false,ColMajor>::run); (internal::packed_triangular_matrix_vector_product<int,Lower|UnitDiag,Scalar,Conj, Scalar,false,RowMajor>::run),
func[TR | (LO << 2) | (UNIT << 3)] = (internal::packed_triangular_matrix_vector_product<int,Upper|UnitDiag,Scalar,false,Scalar,false,RowMajor>::run); 0,
func[ADJ | (LO << 2) | (UNIT << 3)] = (internal::packed_triangular_matrix_vector_product<int,Upper|UnitDiag,Scalar,Conj, Scalar,false,RowMajor>::run); // array index: NOTR | (LO << 2) | (UNIT << 3)
(internal::packed_triangular_matrix_vector_product<int,Lower|UnitDiag,Scalar,false,Scalar,false,ColMajor>::run),
init = true; // array index: TR | (LO << 2) | (UNIT << 3)
} (internal::packed_triangular_matrix_vector_product<int,Upper|UnitDiag,Scalar,false,Scalar,false,RowMajor>::run),
// array index: ADJ | (LO << 2) | (UNIT << 3)
(internal::packed_triangular_matrix_vector_product<int,Upper|UnitDiag,Scalar,Conj, Scalar,false,RowMajor>::run),
0
};
Scalar* ap = reinterpret_cast<Scalar*>(pap); Scalar* ap = reinterpret_cast<Scalar*>(pap);
Scalar* x = reinterpret_cast<Scalar*>(px); Scalar* x = reinterpret_cast<Scalar*>(px);
@ -487,32 +498,36 @@ int EIGEN_BLAS_FUNC(tpmv)(char *uplo, char *opa, char *diag, int *n, RealScalar
int EIGEN_BLAS_FUNC(tpsv)(char *uplo, char *opa, char *diag, int *n, RealScalar *pap, RealScalar *px, int *incx) int EIGEN_BLAS_FUNC(tpsv)(char *uplo, char *opa, char *diag, int *n, RealScalar *pap, RealScalar *px, int *incx)
{ {
typedef void (*functype)(int, const Scalar*, Scalar*); typedef void (*functype)(int, const Scalar*, Scalar*);
static functype func[16]; static const functype func[16] = {
// array index: NOTR | (UP << 2) | (NUNIT << 3)
static bool init = false; (internal::packed_triangular_solve_vector<Scalar,Scalar,int,OnTheLeft, Upper|0, false,ColMajor>::run),
if(!init) // array index: TR | (UP << 2) | (NUNIT << 3)
{ (internal::packed_triangular_solve_vector<Scalar,Scalar,int,OnTheLeft, Lower|0, false,RowMajor>::run),
for(int k=0; k<16; ++k) // array index: ADJ | (UP << 2) | (NUNIT << 3)
func[k] = 0; (internal::packed_triangular_solve_vector<Scalar,Scalar,int,OnTheLeft, Lower|0, Conj, RowMajor>::run),
0,
func[NOTR | (UP << 2) | (NUNIT << 3)] = (internal::packed_triangular_solve_vector<Scalar,Scalar,int,OnTheLeft, Upper|0, false,ColMajor>::run); // array index: NOTR | (LO << 2) | (NUNIT << 3)
func[TR | (UP << 2) | (NUNIT << 3)] = (internal::packed_triangular_solve_vector<Scalar,Scalar,int,OnTheLeft, Lower|0, false,RowMajor>::run); (internal::packed_triangular_solve_vector<Scalar,Scalar,int,OnTheLeft, Lower|0, false,ColMajor>::run),
func[ADJ | (UP << 2) | (NUNIT << 3)] = (internal::packed_triangular_solve_vector<Scalar,Scalar,int,OnTheLeft, Lower|0, Conj, RowMajor>::run); // array index: TR | (LO << 2) | (NUNIT << 3)
(internal::packed_triangular_solve_vector<Scalar,Scalar,int,OnTheLeft, Upper|0, false,RowMajor>::run),
func[NOTR | (LO << 2) | (NUNIT << 3)] = (internal::packed_triangular_solve_vector<Scalar,Scalar,int,OnTheLeft, Lower|0, false,ColMajor>::run); // array index: ADJ | (LO << 2) | (NUNIT << 3)
func[TR | (LO << 2) | (NUNIT << 3)] = (internal::packed_triangular_solve_vector<Scalar,Scalar,int,OnTheLeft, Upper|0, false,RowMajor>::run); (internal::packed_triangular_solve_vector<Scalar,Scalar,int,OnTheLeft, Upper|0, Conj, RowMajor>::run),
func[ADJ | (LO << 2) | (NUNIT << 3)] = (internal::packed_triangular_solve_vector<Scalar,Scalar,int,OnTheLeft, Upper|0, Conj, RowMajor>::run); 0,
// array index: NOTR | (UP << 2) | (UNIT << 3)
func[NOTR | (UP << 2) | (UNIT << 3)] = (internal::packed_triangular_solve_vector<Scalar,Scalar,int,OnTheLeft, Upper|UnitDiag,false,ColMajor>::run); (internal::packed_triangular_solve_vector<Scalar,Scalar,int,OnTheLeft, Upper|UnitDiag,false,ColMajor>::run),
func[TR | (UP << 2) | (UNIT << 3)] = (internal::packed_triangular_solve_vector<Scalar,Scalar,int,OnTheLeft, Lower|UnitDiag,false,RowMajor>::run); // array index: TR | (UP << 2) | (UNIT << 3)
func[ADJ | (UP << 2) | (UNIT << 3)] = (internal::packed_triangular_solve_vector<Scalar,Scalar,int,OnTheLeft, Lower|UnitDiag,Conj, RowMajor>::run); (internal::packed_triangular_solve_vector<Scalar,Scalar,int,OnTheLeft, Lower|UnitDiag,false,RowMajor>::run),
// array index: ADJ | (UP << 2) | (UNIT << 3)
func[NOTR | (LO << 2) | (UNIT << 3)] = (internal::packed_triangular_solve_vector<Scalar,Scalar,int,OnTheLeft, Lower|UnitDiag,false,ColMajor>::run); (internal::packed_triangular_solve_vector<Scalar,Scalar,int,OnTheLeft, Lower|UnitDiag,Conj, RowMajor>::run),
func[TR | (LO << 2) | (UNIT << 3)] = (internal::packed_triangular_solve_vector<Scalar,Scalar,int,OnTheLeft, Upper|UnitDiag,false,RowMajor>::run); 0,
func[ADJ | (LO << 2) | (UNIT << 3)] = (internal::packed_triangular_solve_vector<Scalar,Scalar,int,OnTheLeft, Upper|UnitDiag,Conj, RowMajor>::run); // array index: NOTR | (LO << 2) | (UNIT << 3)
(internal::packed_triangular_solve_vector<Scalar,Scalar,int,OnTheLeft, Lower|UnitDiag,false,ColMajor>::run),
init = true; // array index: TR | (LO << 2) | (UNIT << 3)
} (internal::packed_triangular_solve_vector<Scalar,Scalar,int,OnTheLeft, Upper|UnitDiag,false,RowMajor>::run),
// array index: ADJ | (LO << 2) | (UNIT << 3)
(internal::packed_triangular_solve_vector<Scalar,Scalar,int,OnTheLeft, Upper|UnitDiag,Conj, RowMajor>::run),
0
};
Scalar* ap = reinterpret_cast<Scalar*>(pap); Scalar* ap = reinterpret_cast<Scalar*>(pap);
Scalar* x = reinterpret_cast<Scalar*>(px); Scalar* x = reinterpret_cast<Scalar*>(px);

View File

@ -13,19 +13,12 @@
int EIGEN_BLAS_FUNC(symv) (char *uplo, int *n, RealScalar *palpha, RealScalar *pa, int *lda, RealScalar *px, int *incx, RealScalar *pbeta, RealScalar *py, int *incy) int EIGEN_BLAS_FUNC(symv) (char *uplo, int *n, RealScalar *palpha, RealScalar *pa, int *lda, RealScalar *px, int *incx, RealScalar *pbeta, RealScalar *py, int *incy)
{ {
typedef void (*functype)(int, const Scalar*, int, const Scalar*, Scalar*, Scalar); typedef void (*functype)(int, const Scalar*, int, const Scalar*, Scalar*, Scalar);
static functype func[2]; static const functype func[2] = {
// array index: UP
static bool init = false; (internal::selfadjoint_matrix_vector_product<Scalar,int,ColMajor,Upper,false,false>::run),
if(!init) // array index: LO
{ (internal::selfadjoint_matrix_vector_product<Scalar,int,ColMajor,Lower,false,false>::run),
for(int k=0; k<2; ++k) };
func[k] = 0;
func[UP] = (internal::selfadjoint_matrix_vector_product<Scalar,int,ColMajor,Upper,false,false>::run);
func[LO] = (internal::selfadjoint_matrix_vector_product<Scalar,int,ColMajor,Lower,false,false>::run);
init = true;
}
Scalar* a = reinterpret_cast<Scalar*>(pa); Scalar* a = reinterpret_cast<Scalar*>(pa);
Scalar* x = reinterpret_cast<Scalar*>(px); Scalar* x = reinterpret_cast<Scalar*>(px);
@ -71,34 +64,13 @@ int EIGEN_BLAS_FUNC(symv) (char *uplo, int *n, RealScalar *palpha, RealScalar *p
int EIGEN_BLAS_FUNC(syr)(char *uplo, int *n, RealScalar *palpha, RealScalar *px, int *incx, RealScalar *pc, int *ldc) int EIGEN_BLAS_FUNC(syr)(char *uplo, int *n, RealScalar *palpha, RealScalar *px, int *incx, RealScalar *pc, int *ldc)
{ {
// typedef void (*functype)(int, const Scalar *, int, Scalar *, int, Scalar);
// static functype func[2];
// static bool init = false;
// if(!init)
// {
// for(int k=0; k<2; ++k)
// func[k] = 0;
//
// func[UP] = (internal::selfadjoint_product<Scalar,ColMajor,ColMajor,false,UpperTriangular>::run);
// func[LO] = (internal::selfadjoint_product<Scalar,ColMajor,ColMajor,false,LowerTriangular>::run);
// init = true;
// }
typedef void (*functype)(int, Scalar*, int, const Scalar*, const Scalar*, const Scalar&); typedef void (*functype)(int, Scalar*, int, const Scalar*, const Scalar*, const Scalar&);
static functype func[2]; static const functype func[2] = {
// array index: UP
static bool init = false; (selfadjoint_rank1_update<Scalar,int,ColMajor,Upper,false,Conj>::run),
if(!init) // array index: LO
{ (selfadjoint_rank1_update<Scalar,int,ColMajor,Lower,false,Conj>::run),
for(int k=0; k<2; ++k) };
func[k] = 0;
func[UP] = (selfadjoint_rank1_update<Scalar,int,ColMajor,Upper,false,Conj>::run);
func[LO] = (selfadjoint_rank1_update<Scalar,int,ColMajor,Lower,false,Conj>::run);
init = true;
}
Scalar* x = reinterpret_cast<Scalar*>(px); Scalar* x = reinterpret_cast<Scalar*>(px);
Scalar* c = reinterpret_cast<Scalar*>(pc); Scalar* c = reinterpret_cast<Scalar*>(pc);
@ -131,34 +103,13 @@ int EIGEN_BLAS_FUNC(syr)(char *uplo, int *n, RealScalar *palpha, RealScalar *px,
// C := alpha*x*y' + alpha*y*x' + C // C := alpha*x*y' + alpha*y*x' + C
int EIGEN_BLAS_FUNC(syr2)(char *uplo, int *n, RealScalar *palpha, RealScalar *px, int *incx, RealScalar *py, int *incy, RealScalar *pc, int *ldc) int EIGEN_BLAS_FUNC(syr2)(char *uplo, int *n, RealScalar *palpha, RealScalar *px, int *incx, RealScalar *py, int *incy, RealScalar *pc, int *ldc)
{ {
// typedef void (*functype)(int, const Scalar *, int, const Scalar *, int, Scalar *, int, Scalar);
// static functype func[2];
//
// static bool init = false;
// if(!init)
// {
// for(int k=0; k<2; ++k)
// func[k] = 0;
//
// func[UP] = (internal::selfadjoint_product<Scalar,ColMajor,ColMajor,false,UpperTriangular>::run);
// func[LO] = (internal::selfadjoint_product<Scalar,ColMajor,ColMajor,false,LowerTriangular>::run);
//
// init = true;
// }
typedef void (*functype)(int, Scalar*, int, const Scalar*, const Scalar*, Scalar); typedef void (*functype)(int, Scalar*, int, const Scalar*, const Scalar*, Scalar);
static functype func[2]; static const functype func[2] = {
// array index: UP
static bool init = false; (internal::rank2_update_selector<Scalar,int,Upper>::run),
if(!init) // array index: LO
{ (internal::rank2_update_selector<Scalar,int,Lower>::run),
for(int k=0; k<2; ++k) };
func[k] = 0;
func[UP] = (internal::rank2_update_selector<Scalar,int,Upper>::run);
func[LO] = (internal::rank2_update_selector<Scalar,int,Lower>::run);
init = true;
}
Scalar* x = reinterpret_cast<Scalar*>(px); Scalar* x = reinterpret_cast<Scalar*>(px);
Scalar* y = reinterpret_cast<Scalar*>(py); Scalar* y = reinterpret_cast<Scalar*>(py);
@ -234,19 +185,12 @@ int EIGEN_BLAS_FUNC(syr2)(char *uplo, int *n, RealScalar *palpha, RealScalar *px
int EIGEN_BLAS_FUNC(spr)(char *uplo, int *n, Scalar *palpha, Scalar *px, int *incx, Scalar *pap) int EIGEN_BLAS_FUNC(spr)(char *uplo, int *n, Scalar *palpha, Scalar *px, int *incx, Scalar *pap)
{ {
typedef void (*functype)(int, Scalar*, const Scalar*, Scalar); typedef void (*functype)(int, Scalar*, const Scalar*, Scalar);
static functype func[2]; static const functype func[2] = {
// array index: UP
static bool init = false; (internal::selfadjoint_packed_rank1_update<Scalar,int,ColMajor,Upper,false,false>::run),
if(!init) // array index: LO
{ (internal::selfadjoint_packed_rank1_update<Scalar,int,ColMajor,Lower,false,false>::run),
for(int k=0; k<2; ++k) };
func[k] = 0;
func[UP] = (internal::selfadjoint_packed_rank1_update<Scalar,int,ColMajor,Upper,false,false>::run);
func[LO] = (internal::selfadjoint_packed_rank1_update<Scalar,int,ColMajor,Lower,false,false>::run);
init = true;
}
Scalar* x = reinterpret_cast<Scalar*>(px); Scalar* x = reinterpret_cast<Scalar*>(px);
Scalar* ap = reinterpret_cast<Scalar*>(pap); Scalar* ap = reinterpret_cast<Scalar*>(pap);
@ -285,19 +229,12 @@ int EIGEN_BLAS_FUNC(spr)(char *uplo, int *n, Scalar *palpha, Scalar *px, int *in
int EIGEN_BLAS_FUNC(spr2)(char *uplo, int *n, RealScalar *palpha, RealScalar *px, int *incx, RealScalar *py, int *incy, RealScalar *pap) int EIGEN_BLAS_FUNC(spr2)(char *uplo, int *n, RealScalar *palpha, RealScalar *px, int *incx, RealScalar *py, int *incy, RealScalar *pap)
{ {
typedef void (*functype)(int, Scalar*, const Scalar*, const Scalar*, Scalar); typedef void (*functype)(int, Scalar*, const Scalar*, const Scalar*, Scalar);
static functype func[2]; static const functype func[2] = {
// array index: UP
static bool init = false; (internal::packed_rank2_update_selector<Scalar,int,Upper>::run),
if(!init) // array index: LO
{ (internal::packed_rank2_update_selector<Scalar,int,Lower>::run),
for(int k=0; k<2; ++k) };
func[k] = 0;
func[UP] = (internal::packed_rank2_update_selector<Scalar,int,Upper>::run);
func[LO] = (internal::packed_rank2_update_selector<Scalar,int,Lower>::run);
init = true;
}
Scalar* x = reinterpret_cast<Scalar*>(px); Scalar* x = reinterpret_cast<Scalar*>(px);
Scalar* y = reinterpret_cast<Scalar*>(py); Scalar* y = reinterpret_cast<Scalar*>(py);

View File

@ -13,24 +13,29 @@ int EIGEN_BLAS_FUNC(gemm)(char *opa, char *opb, int *m, int *n, int *k, RealScal
{ {
// std::cerr << "in gemm " << *opa << " " << *opb << " " << *m << " " << *n << " " << *k << " " << *lda << " " << *ldb << " " << *ldc << " " << *palpha << " " << *pbeta << "\n"; // std::cerr << "in gemm " << *opa << " " << *opb << " " << *m << " " << *n << " " << *k << " " << *lda << " " << *ldb << " " << *ldc << " " << *palpha << " " << *pbeta << "\n";
typedef void (*functype)(DenseIndex, DenseIndex, DenseIndex, const Scalar *, DenseIndex, const Scalar *, DenseIndex, Scalar *, DenseIndex, Scalar, internal::level3_blocking<Scalar,Scalar>&, Eigen::internal::GemmParallelInfo<DenseIndex>*); typedef void (*functype)(DenseIndex, DenseIndex, DenseIndex, const Scalar *, DenseIndex, const Scalar *, DenseIndex, Scalar *, DenseIndex, Scalar, internal::level3_blocking<Scalar,Scalar>&, Eigen::internal::GemmParallelInfo<DenseIndex>*);
static functype func[12]; static const functype func[12] = {
// array index: NOTR | (NOTR << 2)
static bool init = false; (internal::general_matrix_matrix_product<DenseIndex,Scalar,ColMajor,false,Scalar,ColMajor,false,ColMajor>::run),
if(!init) // array index: TR | (NOTR << 2)
{ (internal::general_matrix_matrix_product<DenseIndex,Scalar,RowMajor,false,Scalar,ColMajor,false,ColMajor>::run),
for(int i=0; i<12; ++i) // array index: ADJ | (NOTR << 2)
func[i] = 0; (internal::general_matrix_matrix_product<DenseIndex,Scalar,RowMajor,Conj, Scalar,ColMajor,false,ColMajor>::run),
func[NOTR | (NOTR << 2)] = (internal::general_matrix_matrix_product<DenseIndex,Scalar,ColMajor,false,Scalar,ColMajor,false,ColMajor>::run); 0,
func[TR | (NOTR << 2)] = (internal::general_matrix_matrix_product<DenseIndex,Scalar,RowMajor,false,Scalar,ColMajor,false,ColMajor>::run); // array index: NOTR | (TR << 2)
func[ADJ | (NOTR << 2)] = (internal::general_matrix_matrix_product<DenseIndex,Scalar,RowMajor,Conj, Scalar,ColMajor,false,ColMajor>::run); (internal::general_matrix_matrix_product<DenseIndex,Scalar,ColMajor,false,Scalar,RowMajor,false,ColMajor>::run),
func[NOTR | (TR << 2)] = (internal::general_matrix_matrix_product<DenseIndex,Scalar,ColMajor,false,Scalar,RowMajor,false,ColMajor>::run); // array index: TR | (TR << 2)
func[TR | (TR << 2)] = (internal::general_matrix_matrix_product<DenseIndex,Scalar,RowMajor,false,Scalar,RowMajor,false,ColMajor>::run); (internal::general_matrix_matrix_product<DenseIndex,Scalar,RowMajor,false,Scalar,RowMajor,false,ColMajor>::run),
func[ADJ | (TR << 2)] = (internal::general_matrix_matrix_product<DenseIndex,Scalar,RowMajor,Conj, Scalar,RowMajor,false,ColMajor>::run); // array index: ADJ | (TR << 2)
func[NOTR | (ADJ << 2)] = (internal::general_matrix_matrix_product<DenseIndex,Scalar,ColMajor,false,Scalar,RowMajor,Conj, ColMajor>::run); (internal::general_matrix_matrix_product<DenseIndex,Scalar,RowMajor,Conj, Scalar,RowMajor,false,ColMajor>::run),
func[TR | (ADJ << 2)] = (internal::general_matrix_matrix_product<DenseIndex,Scalar,RowMajor,false,Scalar,RowMajor,Conj, ColMajor>::run); 0,
func[ADJ | (ADJ << 2)] = (internal::general_matrix_matrix_product<DenseIndex,Scalar,RowMajor,Conj, Scalar,RowMajor,Conj, ColMajor>::run); // array index: NOTR | (ADJ << 2)
init = true; (internal::general_matrix_matrix_product<DenseIndex,Scalar,ColMajor,false,Scalar,RowMajor,Conj, ColMajor>::run),
} // array index: TR | (ADJ << 2)
(internal::general_matrix_matrix_product<DenseIndex,Scalar,RowMajor,false,Scalar,RowMajor,Conj, ColMajor>::run),
// array index: ADJ | (ADJ << 2)
(internal::general_matrix_matrix_product<DenseIndex,Scalar,RowMajor,Conj, Scalar,RowMajor,Conj, ColMajor>::run),
0
};
Scalar* a = reinterpret_cast<Scalar*>(pa); Scalar* a = reinterpret_cast<Scalar*>(pa);
Scalar* b = reinterpret_cast<Scalar*>(pb); Scalar* b = reinterpret_cast<Scalar*>(pb);
@ -73,49 +78,64 @@ int EIGEN_BLAS_FUNC(trsm)(char *side, char *uplo, char *opa, char *diag, int *m,
{ {
// std::cerr << "in trsm " << *side << " " << *uplo << " " << *opa << " " << *diag << " " << *m << "," << *n << " " << *palpha << " " << *lda << " " << *ldb<< "\n"; // std::cerr << "in trsm " << *side << " " << *uplo << " " << *opa << " " << *diag << " " << *m << "," << *n << " " << *palpha << " " << *lda << " " << *ldb<< "\n";
typedef void (*functype)(DenseIndex, DenseIndex, const Scalar *, DenseIndex, Scalar *, DenseIndex, internal::level3_blocking<Scalar,Scalar>&); typedef void (*functype)(DenseIndex, DenseIndex, const Scalar *, DenseIndex, Scalar *, DenseIndex, internal::level3_blocking<Scalar,Scalar>&);
static functype func[32]; static const functype func[32] = {
// array index: NOTR | (LEFT << 2) | (UP << 3) | (NUNIT << 4)
static bool init = false; (internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheLeft, Upper|0, false,ColMajor,ColMajor>::run),
if(!init) // array index: TR | (LEFT << 2) | (UP << 3) | (NUNIT << 4)
{ (internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheLeft, Lower|0, false,RowMajor,ColMajor>::run),
for(int i=0; i<32; ++i) // array index: ADJ | (LEFT << 2) | (UP << 3) | (NUNIT << 4)
func[i] = 0; (internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheLeft, Lower|0, Conj, RowMajor,ColMajor>::run),\
0,
func[NOTR | (LEFT << 2) | (UP << 3) | (NUNIT << 4)] = (internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheLeft, Upper|0, false,ColMajor,ColMajor>::run); // array index: NOTR | (RIGHT << 2) | (UP << 3) | (NUNIT << 4)
func[TR | (LEFT << 2) | (UP << 3) | (NUNIT << 4)] = (internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheLeft, Lower|0, false,RowMajor,ColMajor>::run); (internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheRight,Upper|0, false,ColMajor,ColMajor>::run),
func[ADJ | (LEFT << 2) | (UP << 3) | (NUNIT << 4)] = (internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheLeft, Lower|0, Conj, RowMajor,ColMajor>::run); // array index: TR | (RIGHT << 2) | (UP << 3) | (NUNIT << 4)
(internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheRight,Lower|0, false,RowMajor,ColMajor>::run),
func[NOTR | (RIGHT << 2) | (UP << 3) | (NUNIT << 4)] = (internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheRight,Upper|0, false,ColMajor,ColMajor>::run); // array index: ADJ | (RIGHT << 2) | (UP << 3) | (NUNIT << 4)
func[TR | (RIGHT << 2) | (UP << 3) | (NUNIT << 4)] = (internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheRight,Lower|0, false,RowMajor,ColMajor>::run); (internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheRight,Lower|0, Conj, RowMajor,ColMajor>::run),
func[ADJ | (RIGHT << 2) | (UP << 3) | (NUNIT << 4)] = (internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheRight,Lower|0, Conj, RowMajor,ColMajor>::run); 0,
// array index: NOTR | (LEFT << 2) | (LO << 3) | (NUNIT << 4)
func[NOTR | (LEFT << 2) | (LO << 3) | (NUNIT << 4)] = (internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheLeft, Lower|0, false,ColMajor,ColMajor>::run); (internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheLeft, Lower|0, false,ColMajor,ColMajor>::run),
func[TR | (LEFT << 2) | (LO << 3) | (NUNIT << 4)] = (internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheLeft, Upper|0, false,RowMajor,ColMajor>::run); // array index: TR | (LEFT << 2) | (LO << 3) | (NUNIT << 4)
func[ADJ | (LEFT << 2) | (LO << 3) | (NUNIT << 4)] = (internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheLeft, Upper|0, Conj, RowMajor,ColMajor>::run); (internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheLeft, Upper|0, false,RowMajor,ColMajor>::run),
// array index: ADJ | (LEFT << 2) | (LO << 3) | (NUNIT << 4)
func[NOTR | (RIGHT << 2) | (LO << 3) | (NUNIT << 4)] = (internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheRight,Lower|0, false,ColMajor,ColMajor>::run); (internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheLeft, Upper|0, Conj, RowMajor,ColMajor>::run),
func[TR | (RIGHT << 2) | (LO << 3) | (NUNIT << 4)] = (internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheRight,Upper|0, false,RowMajor,ColMajor>::run); 0,
func[ADJ | (RIGHT << 2) | (LO << 3) | (NUNIT << 4)] = (internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheRight,Upper|0, Conj, RowMajor,ColMajor>::run); // array index: NOTR | (RIGHT << 2) | (LO << 3) | (NUNIT << 4)
(internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheRight,Lower|0, false,ColMajor,ColMajor>::run),
// array index: TR | (RIGHT << 2) | (LO << 3) | (NUNIT << 4)
func[NOTR | (LEFT << 2) | (UP << 3) | (UNIT << 4)] = (internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheLeft, Upper|UnitDiag,false,ColMajor,ColMajor>::run); (internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheRight,Upper|0, false,RowMajor,ColMajor>::run),
func[TR | (LEFT << 2) | (UP << 3) | (UNIT << 4)] = (internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheLeft, Lower|UnitDiag,false,RowMajor,ColMajor>::run); // array index: ADJ | (RIGHT << 2) | (LO << 3) | (NUNIT << 4)
func[ADJ | (LEFT << 2) | (UP << 3) | (UNIT << 4)] = (internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheLeft, Lower|UnitDiag,Conj, RowMajor,ColMajor>::run); (internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheRight,Upper|0, Conj, RowMajor,ColMajor>::run),
0,
func[NOTR | (RIGHT << 2) | (UP << 3) | (UNIT << 4)] = (internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheRight,Upper|UnitDiag,false,ColMajor,ColMajor>::run); // array index: NOTR | (LEFT << 2) | (UP << 3) | (UNIT << 4)
func[TR | (RIGHT << 2) | (UP << 3) | (UNIT << 4)] = (internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheRight,Lower|UnitDiag,false,RowMajor,ColMajor>::run); (internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheLeft, Upper|UnitDiag,false,ColMajor,ColMajor>::run),
func[ADJ | (RIGHT << 2) | (UP << 3) | (UNIT << 4)] = (internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheRight,Lower|UnitDiag,Conj, RowMajor,ColMajor>::run); // array index: TR | (LEFT << 2) | (UP << 3) | (UNIT << 4)
(internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheLeft, Lower|UnitDiag,false,RowMajor,ColMajor>::run),
func[NOTR | (LEFT << 2) | (LO << 3) | (UNIT << 4)] = (internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheLeft, Lower|UnitDiag,false,ColMajor,ColMajor>::run); // array index: ADJ | (LEFT << 2) | (UP << 3) | (UNIT << 4)
func[TR | (LEFT << 2) | (LO << 3) | (UNIT << 4)] = (internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheLeft, Upper|UnitDiag,false,RowMajor,ColMajor>::run); (internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheLeft, Lower|UnitDiag,Conj, RowMajor,ColMajor>::run),
func[ADJ | (LEFT << 2) | (LO << 3) | (UNIT << 4)] = (internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheLeft, Upper|UnitDiag,Conj, RowMajor,ColMajor>::run); 0,
// array index: NOTR | (RIGHT << 2) | (UP << 3) | (UNIT << 4)
func[NOTR | (RIGHT << 2) | (LO << 3) | (UNIT << 4)] = (internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheRight,Lower|UnitDiag,false,ColMajor,ColMajor>::run); (internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheRight,Upper|UnitDiag,false,ColMajor,ColMajor>::run),
func[TR | (RIGHT << 2) | (LO << 3) | (UNIT << 4)] = (internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheRight,Upper|UnitDiag,false,RowMajor,ColMajor>::run); // array index: TR | (RIGHT << 2) | (UP << 3) | (UNIT << 4)
func[ADJ | (RIGHT << 2) | (LO << 3) | (UNIT << 4)] = (internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheRight,Upper|UnitDiag,Conj, RowMajor,ColMajor>::run); (internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheRight,Lower|UnitDiag,false,RowMajor,ColMajor>::run),
// array index: ADJ | (RIGHT << 2) | (UP << 3) | (UNIT << 4)
init = true; (internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheRight,Lower|UnitDiag,Conj, RowMajor,ColMajor>::run),
} 0,
// array index: NOTR | (LEFT << 2) | (LO << 3) | (UNIT << 4)
(internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheLeft, Lower|UnitDiag,false,ColMajor,ColMajor>::run),
// array index: TR | (LEFT << 2) | (LO << 3) | (UNIT << 4)
(internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheLeft, Upper|UnitDiag,false,RowMajor,ColMajor>::run),
// array index: ADJ | (LEFT << 2) | (LO << 3) | (UNIT << 4)
(internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheLeft, Upper|UnitDiag,Conj, RowMajor,ColMajor>::run),
0,
// array index: NOTR | (RIGHT << 2) | (LO << 3) | (UNIT << 4)
(internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheRight,Lower|UnitDiag,false,ColMajor,ColMajor>::run),
// array index: TR | (RIGHT << 2) | (LO << 3) | (UNIT << 4)
(internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheRight,Upper|UnitDiag,false,RowMajor,ColMajor>::run),
// array index: ADJ | (RIGHT << 2) | (LO << 3) | (UNIT << 4)
(internal::triangular_solve_matrix<Scalar,DenseIndex,OnTheRight,Upper|UnitDiag,Conj, RowMajor,ColMajor>::run),
0
};
Scalar* a = reinterpret_cast<Scalar*>(pa); Scalar* a = reinterpret_cast<Scalar*>(pa);
Scalar* b = reinterpret_cast<Scalar*>(pb); Scalar* b = reinterpret_cast<Scalar*>(pb);
@ -162,47 +182,64 @@ int EIGEN_BLAS_FUNC(trmm)(char *side, char *uplo, char *opa, char *diag, int *m,
{ {
// std::cerr << "in trmm " << *side << " " << *uplo << " " << *opa << " " << *diag << " " << *m << " " << *n << " " << *lda << " " << *ldb << " " << *palpha << "\n"; // std::cerr << "in trmm " << *side << " " << *uplo << " " << *opa << " " << *diag << " " << *m << " " << *n << " " << *lda << " " << *ldb << " " << *palpha << "\n";
typedef void (*functype)(DenseIndex, DenseIndex, DenseIndex, const Scalar *, DenseIndex, const Scalar *, DenseIndex, Scalar *, DenseIndex, const Scalar&, internal::level3_blocking<Scalar,Scalar>&); typedef void (*functype)(DenseIndex, DenseIndex, DenseIndex, const Scalar *, DenseIndex, const Scalar *, DenseIndex, Scalar *, DenseIndex, const Scalar&, internal::level3_blocking<Scalar,Scalar>&);
static functype func[32]; static const functype func[32] = {
static bool init = false; // array index: NOTR | (LEFT << 2) | (UP << 3) | (NUNIT << 4)
if(!init) (internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Upper|0, true, ColMajor,false,ColMajor,false,ColMajor>::run),
{ // array index: TR | (LEFT << 2) | (UP << 3) | (NUNIT << 4)
for(int k=0; k<32; ++k) (internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Lower|0, true, RowMajor,false,ColMajor,false,ColMajor>::run),
func[k] = 0; // array index: ADJ | (LEFT << 2) | (UP << 3) | (NUNIT << 4)
(internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Lower|0, true, RowMajor,Conj, ColMajor,false,ColMajor>::run),
func[NOTR | (LEFT << 2) | (UP << 3) | (NUNIT << 4)] = (internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Upper|0, true, ColMajor,false,ColMajor,false,ColMajor>::run); 0,
func[TR | (LEFT << 2) | (UP << 3) | (NUNIT << 4)] = (internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Lower|0, true, RowMajor,false,ColMajor,false,ColMajor>::run); // array index: NOTR | (RIGHT << 2) | (UP << 3) | (NUNIT << 4)
func[ADJ | (LEFT << 2) | (UP << 3) | (NUNIT << 4)] = (internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Lower|0, true, RowMajor,Conj, ColMajor,false,ColMajor>::run); (internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Upper|0, false,ColMajor,false,ColMajor,false,ColMajor>::run),
// array index: TR | (RIGHT << 2) | (UP << 3) | (NUNIT << 4)
func[NOTR | (RIGHT << 2) | (UP << 3) | (NUNIT << 4)] = (internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Upper|0, false,ColMajor,false,ColMajor,false,ColMajor>::run); (internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Lower|0, false,ColMajor,false,RowMajor,false,ColMajor>::run),
func[TR | (RIGHT << 2) | (UP << 3) | (NUNIT << 4)] = (internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Lower|0, false,ColMajor,false,RowMajor,false,ColMajor>::run); // array index: ADJ | (RIGHT << 2) | (UP << 3) | (NUNIT << 4)
func[ADJ | (RIGHT << 2) | (UP << 3) | (NUNIT << 4)] = (internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Lower|0, false,ColMajor,false,RowMajor,Conj, ColMajor>::run); (internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Lower|0, false,ColMajor,false,RowMajor,Conj, ColMajor>::run),
0,
func[NOTR | (LEFT << 2) | (LO << 3) | (NUNIT << 4)] = (internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Lower|0, true, ColMajor,false,ColMajor,false,ColMajor>::run); // array index: NOTR | (LEFT << 2) | (LO << 3) | (NUNIT << 4)
func[TR | (LEFT << 2) | (LO << 3) | (NUNIT << 4)] = (internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Upper|0, true, RowMajor,false,ColMajor,false,ColMajor>::run); (internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Lower|0, true, ColMajor,false,ColMajor,false,ColMajor>::run),
func[ADJ | (LEFT << 2) | (LO << 3) | (NUNIT << 4)] = (internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Upper|0, true, RowMajor,Conj, ColMajor,false,ColMajor>::run); // array index: TR | (LEFT << 2) | (LO << 3) | (NUNIT << 4)
(internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Upper|0, true, RowMajor,false,ColMajor,false,ColMajor>::run),
func[NOTR | (RIGHT << 2) | (LO << 3) | (NUNIT << 4)] = (internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Lower|0, false,ColMajor,false,ColMajor,false,ColMajor>::run); // array index: ADJ | (LEFT << 2) | (LO << 3) | (NUNIT << 4)
func[TR | (RIGHT << 2) | (LO << 3) | (NUNIT << 4)] = (internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Upper|0, false,ColMajor,false,RowMajor,false,ColMajor>::run); (internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Upper|0, true, RowMajor,Conj, ColMajor,false,ColMajor>::run),
func[ADJ | (RIGHT << 2) | (LO << 3) | (NUNIT << 4)] = (internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Upper|0, false,ColMajor,false,RowMajor,Conj, ColMajor>::run); 0,
// array index: NOTR | (RIGHT << 2) | (LO << 3) | (NUNIT << 4)
func[NOTR | (LEFT << 2) | (UP << 3) | (UNIT << 4)] = (internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Upper|UnitDiag,true, ColMajor,false,ColMajor,false,ColMajor>::run); (internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Lower|0, false,ColMajor,false,ColMajor,false,ColMajor>::run),
func[TR | (LEFT << 2) | (UP << 3) | (UNIT << 4)] = (internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Lower|UnitDiag,true, RowMajor,false,ColMajor,false,ColMajor>::run); // array index: TR | (RIGHT << 2) | (LO << 3) | (NUNIT << 4)
func[ADJ | (LEFT << 2) | (UP << 3) | (UNIT << 4)] = (internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Lower|UnitDiag,true, RowMajor,Conj, ColMajor,false,ColMajor>::run); (internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Upper|0, false,ColMajor,false,RowMajor,false,ColMajor>::run),
// array index: ADJ | (RIGHT << 2) | (LO << 3) | (NUNIT << 4)
func[NOTR | (RIGHT << 2) | (UP << 3) | (UNIT << 4)] = (internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Upper|UnitDiag,false,ColMajor,false,ColMajor,false,ColMajor>::run); (internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Upper|0, false,ColMajor,false,RowMajor,Conj, ColMajor>::run),
func[TR | (RIGHT << 2) | (UP << 3) | (UNIT << 4)] = (internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Lower|UnitDiag,false,ColMajor,false,RowMajor,false,ColMajor>::run); 0,
func[ADJ | (RIGHT << 2) | (UP << 3) | (UNIT << 4)] = (internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Lower|UnitDiag,false,ColMajor,false,RowMajor,Conj, ColMajor>::run); // array index: NOTR | (LEFT << 2) | (UP << 3) | (UNIT << 4)
(internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Upper|UnitDiag,true, ColMajor,false,ColMajor,false,ColMajor>::run),
func[NOTR | (LEFT << 2) | (LO << 3) | (UNIT << 4)] = (internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Lower|UnitDiag,true, ColMajor,false,ColMajor,false,ColMajor>::run); // array index: TR | (LEFT << 2) | (UP << 3) | (UNIT << 4)
func[TR | (LEFT << 2) | (LO << 3) | (UNIT << 4)] = (internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Upper|UnitDiag,true, RowMajor,false,ColMajor,false,ColMajor>::run); (internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Lower|UnitDiag,true, RowMajor,false,ColMajor,false,ColMajor>::run),
func[ADJ | (LEFT << 2) | (LO << 3) | (UNIT << 4)] = (internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Upper|UnitDiag,true, RowMajor,Conj, ColMajor,false,ColMajor>::run); // array index: ADJ | (LEFT << 2) | (UP << 3) | (UNIT << 4)
(internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Lower|UnitDiag,true, RowMajor,Conj, ColMajor,false,ColMajor>::run),
func[NOTR | (RIGHT << 2) | (LO << 3) | (UNIT << 4)] = (internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Lower|UnitDiag,false,ColMajor,false,ColMajor,false,ColMajor>::run); 0,
func[TR | (RIGHT << 2) | (LO << 3) | (UNIT << 4)] = (internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Upper|UnitDiag,false,ColMajor,false,RowMajor,false,ColMajor>::run); // array index: NOTR | (RIGHT << 2) | (UP << 3) | (UNIT << 4)
func[ADJ | (RIGHT << 2) | (LO << 3) | (UNIT << 4)] = (internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Upper|UnitDiag,false,ColMajor,false,RowMajor,Conj, ColMajor>::run); (internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Upper|UnitDiag,false,ColMajor,false,ColMajor,false,ColMajor>::run),
// array index: TR | (RIGHT << 2) | (UP << 3) | (UNIT << 4)
init = true; (internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Lower|UnitDiag,false,ColMajor,false,RowMajor,false,ColMajor>::run),
} // array index: ADJ | (RIGHT << 2) | (UP << 3) | (UNIT << 4)
(internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Lower|UnitDiag,false,ColMajor,false,RowMajor,Conj, ColMajor>::run),
0,
// array index: NOTR | (LEFT << 2) | (LO << 3) | (UNIT << 4)
(internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Lower|UnitDiag,true, ColMajor,false,ColMajor,false,ColMajor>::run),
// array index: TR | (LEFT << 2) | (LO << 3) | (UNIT << 4)
(internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Upper|UnitDiag,true, RowMajor,false,ColMajor,false,ColMajor>::run),
// array index: ADJ | (LEFT << 2) | (LO << 3) | (UNIT << 4)
(internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Upper|UnitDiag,true, RowMajor,Conj, ColMajor,false,ColMajor>::run),
0,
// array index: NOTR | (RIGHT << 2) | (LO << 3) | (UNIT << 4)
(internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Lower|UnitDiag,false,ColMajor,false,ColMajor,false,ColMajor>::run),
// array index: TR | (RIGHT << 2) | (LO << 3) | (UNIT << 4)
(internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Upper|UnitDiag,false,ColMajor,false,RowMajor,false,ColMajor>::run),
// array index: ADJ | (RIGHT << 2) | (LO << 3) | (UNIT << 4)
(internal::product_triangular_matrix_matrix<Scalar,DenseIndex,Upper|UnitDiag,false,ColMajor,false,RowMajor,Conj, ColMajor>::run),
0
};
Scalar* a = reinterpret_cast<Scalar*>(pa); Scalar* a = reinterpret_cast<Scalar*>(pa);
Scalar* b = reinterpret_cast<Scalar*>(pb); Scalar* b = reinterpret_cast<Scalar*>(pb);
@ -318,24 +355,22 @@ int EIGEN_BLAS_FUNC(syrk)(char *uplo, char *op, int *n, int *k, RealScalar *palp
// std::cerr << "in syrk " << *uplo << " " << *op << " " << *n << " " << *k << " " << *palpha << " " << *lda << " " << *pbeta << " " << *ldc << "\n"; // std::cerr << "in syrk " << *uplo << " " << *op << " " << *n << " " << *k << " " << *palpha << " " << *lda << " " << *pbeta << " " << *ldc << "\n";
#if !ISCOMPLEX #if !ISCOMPLEX
typedef void (*functype)(DenseIndex, DenseIndex, const Scalar *, DenseIndex, const Scalar *, DenseIndex, Scalar *, DenseIndex, const Scalar&, internal::level3_blocking<Scalar,Scalar>&); typedef void (*functype)(DenseIndex, DenseIndex, const Scalar *, DenseIndex, const Scalar *, DenseIndex, Scalar *, DenseIndex, const Scalar&, internal::level3_blocking<Scalar,Scalar>&);
static functype func[8]; static const functype func[8] = {
// array index: NOTR | (UP << 2)
static bool init = false; (internal::general_matrix_matrix_triangular_product<DenseIndex,Scalar,ColMajor,false,Scalar,RowMajor,ColMajor,Conj, Upper>::run),
if(!init) // array index: TR | (UP << 2)
{ (internal::general_matrix_matrix_triangular_product<DenseIndex,Scalar,RowMajor,false,Scalar,ColMajor,ColMajor,Conj, Upper>::run),
for(int i=0; i<8; ++i) // array index: ADJ | (UP << 2)
func[i] = 0; (internal::general_matrix_matrix_triangular_product<DenseIndex,Scalar,RowMajor,Conj, Scalar,ColMajor,ColMajor,false,Upper>::run),
0,
func[NOTR | (UP << 2)] = (internal::general_matrix_matrix_triangular_product<DenseIndex,Scalar,ColMajor,false,Scalar,RowMajor,ColMajor,Conj, Upper>::run); // array index: NOTR | (LO << 2)
func[TR | (UP << 2)] = (internal::general_matrix_matrix_triangular_product<DenseIndex,Scalar,RowMajor,false,Scalar,ColMajor,ColMajor,Conj, Upper>::run); (internal::general_matrix_matrix_triangular_product<DenseIndex,Scalar,ColMajor,false,Scalar,RowMajor,ColMajor,Conj, Lower>::run),
func[ADJ | (UP << 2)] = (internal::general_matrix_matrix_triangular_product<DenseIndex,Scalar,RowMajor,Conj, Scalar,ColMajor,ColMajor,false,Upper>::run); // array index: TR | (LO << 2)
(internal::general_matrix_matrix_triangular_product<DenseIndex,Scalar,RowMajor,false,Scalar,ColMajor,ColMajor,Conj, Lower>::run),
func[NOTR | (LO << 2)] = (internal::general_matrix_matrix_triangular_product<DenseIndex,Scalar,ColMajor,false,Scalar,RowMajor,ColMajor,Conj, Lower>::run); // array index: ADJ | (LO << 2)
func[TR | (LO << 2)] = (internal::general_matrix_matrix_triangular_product<DenseIndex,Scalar,RowMajor,false,Scalar,ColMajor,ColMajor,Conj, Lower>::run); (internal::general_matrix_matrix_triangular_product<DenseIndex,Scalar,RowMajor,Conj, Scalar,ColMajor,ColMajor,false,Lower>::run),
func[ADJ | (LO << 2)] = (internal::general_matrix_matrix_triangular_product<DenseIndex,Scalar,RowMajor,Conj, Scalar,ColMajor,ColMajor,false,Lower>::run); 0
};
init = true;
}
#endif #endif
Scalar* a = reinterpret_cast<Scalar*>(pa); Scalar* a = reinterpret_cast<Scalar*>(pa);
@ -524,22 +559,20 @@ int EIGEN_BLAS_FUNC(herk)(char *uplo, char *op, int *n, int *k, RealScalar *palp
// std::cerr << "in herk " << *uplo << " " << *op << " " << *n << " " << *k << " " << *palpha << " " << *lda << " " << *pbeta << " " << *ldc << "\n"; // std::cerr << "in herk " << *uplo << " " << *op << " " << *n << " " << *k << " " << *palpha << " " << *lda << " " << *pbeta << " " << *ldc << "\n";
typedef void (*functype)(DenseIndex, DenseIndex, const Scalar *, DenseIndex, const Scalar *, DenseIndex, Scalar *, DenseIndex, const Scalar&, internal::level3_blocking<Scalar,Scalar>&); typedef void (*functype)(DenseIndex, DenseIndex, const Scalar *, DenseIndex, const Scalar *, DenseIndex, Scalar *, DenseIndex, const Scalar&, internal::level3_blocking<Scalar,Scalar>&);
static functype func[8]; static const functype func[8] = {
// array index: NOTR | (UP << 2)
static bool init = false; (internal::general_matrix_matrix_triangular_product<DenseIndex,Scalar,ColMajor,false,Scalar,RowMajor,Conj, ColMajor,Upper>::run),
if(!init) 0,
{ // array index: ADJ | (UP << 2)
for(int i=0; i<8; ++i) (internal::general_matrix_matrix_triangular_product<DenseIndex,Scalar,RowMajor,Conj, Scalar,ColMajor,false,ColMajor,Upper>::run),
func[i] = 0; 0,
// array index: NOTR | (LO << 2)
func[NOTR | (UP << 2)] = (internal::general_matrix_matrix_triangular_product<DenseIndex,Scalar,ColMajor,false,Scalar,RowMajor,Conj, ColMajor,Upper>::run); (internal::general_matrix_matrix_triangular_product<DenseIndex,Scalar,ColMajor,false,Scalar,RowMajor,Conj, ColMajor,Lower>::run),
func[ADJ | (UP << 2)] = (internal::general_matrix_matrix_triangular_product<DenseIndex,Scalar,RowMajor,Conj, Scalar,ColMajor,false,ColMajor,Upper>::run); 0,
// array index: ADJ | (LO << 2)
func[NOTR | (LO << 2)] = (internal::general_matrix_matrix_triangular_product<DenseIndex,Scalar,ColMajor,false,Scalar,RowMajor,Conj, ColMajor,Lower>::run); (internal::general_matrix_matrix_triangular_product<DenseIndex,Scalar,RowMajor,Conj, Scalar,ColMajor,false,ColMajor,Lower>::run),
func[ADJ | (LO << 2)] = (internal::general_matrix_matrix_triangular_product<DenseIndex,Scalar,RowMajor,Conj, Scalar,ColMajor,false,ColMajor,Lower>::run); 0
};
init = true;
}
Scalar* a = reinterpret_cast<Scalar*>(pa); Scalar* a = reinterpret_cast<Scalar*>(pa);
Scalar* c = reinterpret_cast<Scalar*>(pc); Scalar* c = reinterpret_cast<Scalar*>(pc);