124943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner
224943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner#define BLAS_FUNC(NAME) CAT(CAT(SCALAR_PREFIX,NAME),_)
324943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner
424943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattnertemplate<> class blas_interface<SCALAR> : public c_interface_base<SCALAR>
524943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner{
624943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner
724943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattnerpublic :
824943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner
924943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner  static SCALAR fone;
1024943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner  static SCALAR fzero;
1124943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner
1224943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner  static inline std::string name()
1324943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner  {
1424943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner    return MAKE_STRING(CBLASNAME);
1524943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner  }
1624943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner
1724943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner  static inline void matrix_vector_product(gene_matrix & A, gene_vector & B, gene_vector & X, int N){
1824943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner    BLAS_FUNC(gemv)(&notrans,&N,&N,&fone,A,&N,B,&intone,&fzero,X,&intone);
1924943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner  }
2024943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner
2124943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner  static inline void symv(gene_matrix & A, gene_vector & B, gene_vector & X, int N){
2224943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner    BLAS_FUNC(symv)(&lower, &N,&fone,A,&N,B,&intone,&fzero,X,&intone);
2324943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner  }
2424943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner
2524943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner  static inline void syr2(gene_matrix & A, gene_vector & B, gene_vector & X, int N){
2624943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner    BLAS_FUNC(syr2)(&lower,&N,&fone,B,&intone,X,&intone,A,&N);
2724943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner  }
2824943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner
2924943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner  static inline void ger(gene_matrix & A, gene_vector & X, gene_vector & Y, int N){
3024943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner    BLAS_FUNC(ger)(&N,&N,&fone,X,&intone,Y,&intone,A,&N);
31236c1c71fab5d56a4dbebaabb969a238f11d142eGreg Clayton  }
32b01000fd063629facd45044f137446fb748ee179Greg Clayton
33236c1c71fab5d56a4dbebaabb969a238f11d142eGreg Clayton  static inline void rot(gene_vector & A,  gene_vector & B, SCALAR c, SCALAR s, int N){
34236c1c71fab5d56a4dbebaabb969a238f11d142eGreg Clayton    BLAS_FUNC(rot)(&N,A,&intone,B,&intone,&c,&s);
35236c1c71fab5d56a4dbebaabb969a238f11d142eGreg Clayton  }
36b01000fd063629facd45044f137446fb748ee179Greg Clayton
37236c1c71fab5d56a4dbebaabb969a238f11d142eGreg Clayton  static inline void atv_product(gene_matrix & A, gene_vector & B, gene_vector & X, int N){
3824943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner    BLAS_FUNC(gemv)(&trans,&N,&N,&fone,A,&N,B,&intone,&fzero,X,&intone);
3924943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner  }
4024943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner
41e005f2ce03c489ebde9110678a29cbfe8488d5b4Greg Clayton  static inline void matrix_matrix_product(gene_matrix & A, gene_matrix & B, gene_matrix & X, int N){
4224943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner    BLAS_FUNC(gemm)(&notrans,&notrans,&N,&N,&N,&fone,A,&N,B,&N,&fzero,X,&N);
4324943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner  }
4424943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner
4524943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner  static inline void transposed_matrix_matrix_product(gene_matrix & A, gene_matrix & B, gene_matrix & X, int N){
4624943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner    BLAS_FUNC(gemm)(&notrans,&notrans,&N,&N,&N,&fone,A,&N,B,&N,&fzero,X,&N);
4724943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner  }
4824943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner
4924943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner//   static inline void ata_product(gene_matrix & A, gene_matrix & X, int N){
5024943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner//     ssyrk_(&lower,&trans,&N,&N,&fone,A,&N,&fzero,X,&N);
5124943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner//   }
5224943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner
5324943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner  static inline void aat_product(gene_matrix & A, gene_matrix & X, int N){
5424943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner    BLAS_FUNC(syrk)(&lower,&notrans,&N,&N,&fone,A,&N,&fzero,X,&N);
55e005f2ce03c489ebde9110678a29cbfe8488d5b4Greg Clayton  }
5624943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner
5724943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner  static inline void axpy(SCALAR coef, const gene_vector & X, gene_vector & Y, int N){
5824943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner    BLAS_FUNC(axpy)(&N,&coef,X,&intone,Y,&intone);
5924943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner  }
6024943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner
6124943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner  static inline void axpby(SCALAR a, const gene_vector & X, SCALAR b, gene_vector & Y, int N){
6224943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner    BLAS_FUNC(scal)(&N,&b,Y,&intone);
6324943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner    BLAS_FUNC(axpy)(&N,&a,X,&intone,Y,&intone);
6424943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner  }
65b01000fd063629facd45044f137446fb748ee179Greg Clayton
66b01000fd063629facd45044f137446fb748ee179Greg Clayton  static inline void cholesky(const gene_matrix & X, gene_matrix & C, int N){
67b01000fd063629facd45044f137446fb748ee179Greg Clayton    int N2 = N*N;
68b01000fd063629facd45044f137446fb748ee179Greg Clayton    BLAS_FUNC(copy)(&N2, X, &intone, C, &intone);
69b01000fd063629facd45044f137446fb748ee179Greg Clayton    char uplo = 'L';
70b01000fd063629facd45044f137446fb748ee179Greg Clayton    int info = 0;
71b01000fd063629facd45044f137446fb748ee179Greg Clayton    BLAS_FUNC(potrf)(&uplo, &N, C, &N, &info);
7224943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner    if(info!=0) std::cerr << "potrf_ error " << info << "\n";
7324943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner  }
7424943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner
7524943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner  static inline void partial_lu_decomp(const gene_matrix & X, gene_matrix & C, int N){
7624943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner    int N2 = N*N;
7724943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner    BLAS_FUNC(copy)(&N2, X, &intone, C, &intone);
7824943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner    char uplo = 'L';
7924943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner    int info = 0;
8024943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner    int * ipiv = (int*)alloca(sizeof(int)*N);
810467c78e9a75eff9ec33d3c1f39fa83e1c5d9241Greg Clayton    BLAS_FUNC(getrf)(&N, &N, C, &N, ipiv, &info);
8224943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner    if(info!=0) std::cerr << "getrf_ error " << info << "\n";
8324943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner  }
8424943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner
85236c1c71fab5d56a4dbebaabb969a238f11d142eGreg Clayton  static inline void trisolve_lower(const gene_matrix & L, const gene_vector& B, gene_vector & X, int N){
8624943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner    BLAS_FUNC(copy)(&N, B, &intone, X, &intone);
8724943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner    BLAS_FUNC(trsv)(&lower, &notrans, &nonunit, &N, L, &N, X, &intone);
8824943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner  }
8924943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner
9024943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner  static inline void trisolve_lower_matrix(const gene_matrix & L, const gene_matrix& B, gene_matrix & X, int N){
9124943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner    BLAS_FUNC(copy)(&N, B, &intone, X, &intone);
92236c1c71fab5d56a4dbebaabb969a238f11d142eGreg Clayton    BLAS_FUNC(trsm)(&right, &lower, &notrans, &nonunit, &N, &N, &fone, L, &N, X, &N);
9324943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner  }
9424943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner
9524943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner  static inline void trmm(gene_matrix & A, gene_matrix & B, gene_matrix & X, int N){
9624943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner    BLAS_FUNC(trmm)(&left, &lower, &notrans,&nonunit, &N,&N,&fone,A,&N,B,&N);
9724943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner  }
98b01000fd063629facd45044f137446fb748ee179Greg Clayton
99b01000fd063629facd45044f137446fb748ee179Greg Clayton  #ifdef HAS_LAPACK
100b01000fd063629facd45044f137446fb748ee179Greg Clayton
101b01000fd063629facd45044f137446fb748ee179Greg Clayton  static inline void lu_decomp(const gene_matrix & X, gene_matrix & C, int N){
102b01000fd063629facd45044f137446fb748ee179Greg Clayton    int N2 = N*N;
103b01000fd063629facd45044f137446fb748ee179Greg Clayton    BLAS_FUNC(copy)(&N2, X, &intone, C, &intone);
104b01000fd063629facd45044f137446fb748ee179Greg Clayton    char uplo = 'L';
105395fc33dc4b06c048ed35047ec461bc092ef2df3Greg Clayton    int info = 0;
106395fc33dc4b06c048ed35047ec461bc092ef2df3Greg Clayton    int * ipiv = (int*)alloca(sizeof(int)*N);
107b01000fd063629facd45044f137446fb748ee179Greg Clayton    int * jpiv = (int*)alloca(sizeof(int)*N);
108b01000fd063629facd45044f137446fb748ee179Greg Clayton    BLAS_FUNC(getc2)(&N, C, &N, ipiv, jpiv, &info);
109395fc33dc4b06c048ed35047ec461bc092ef2df3Greg Clayton  }
110b01000fd063629facd45044f137446fb748ee179Greg Clayton
111b01000fd063629facd45044f137446fb748ee179Greg Clayton
112b01000fd063629facd45044f137446fb748ee179Greg Clayton
113b01000fd063629facd45044f137446fb748ee179Greg Clayton  static inline void hessenberg(const gene_matrix & X, gene_matrix & C, int N){
114b01000fd063629facd45044f137446fb748ee179Greg Clayton    {
11524943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner      int N2 = N*N;
11624943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner      int inc = 1;
11724943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner      BLAS_FUNC(copy)(&N2, X, &inc, C, &inc);
11824943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner    }
11924943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner    int info = 0;
12024943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner    int ilo = 1;
12124943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner    int ihi = N;
12224943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner    int bsize = 64;
12324943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner    int worksize = N*bsize;
12424943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner    SCALAR* d = new SCALAR[N+worksize];
12524943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner    BLAS_FUNC(gehrd)(&N, &ilo, &ihi, C, &N, d, d+N, &worksize, &info);
12624943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner    delete[] d;
12724943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner  }
12824943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner
12924943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner  static inline void tridiagonalization(const gene_matrix & X, gene_matrix & C, int N){
13024943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner    {
13124943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner      int N2 = N*N;
13224943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner      int inc = 1;
13324943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner      BLAS_FUNC(copy)(&N2, X, &inc, C, &inc);
13424943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner    }
13524943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner    char uplo = 'U';
13624943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner    int info = 0;
13724943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner    int ilo = 1;
13824943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner    int ihi = N;
13924943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner    int bsize = 64;
14024943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner    int worksize = N*bsize;
14124943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner    SCALAR* d = new SCALAR[3*N+worksize];
14224943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner    BLAS_FUNC(sytrd)(&uplo, &N, C, &N, d, d+N, d+2*N, d+3*N, &worksize, &info);
14324943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner    delete[] d;
14424943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner  }
14524943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner
14624943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner  #endif // HAS_LAPACK
14724943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner
14824943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner};
14924943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner
15024943d2ee8bfaa7cf5893e4709143924157a5c1eChris LattnerSCALAR blas_interface<SCALAR>::fone = SCALAR(1);
15124943d2ee8bfaa7cf5893e4709143924157a5c1eChris LattnerSCALAR blas_interface<SCALAR>::fzero = SCALAR(0);
15224943d2ee8bfaa7cf5893e4709143924157a5c1eChris Lattner