Merge pull request #77 from nicolasbock/issue_75

Issue 75

src/C-interface/bml_scale.c

@@ -17,7 +17,7 @@
  */
 bml_matrix_t *
 bml_scale_new(
-    const double scale_factor,
+    const void *scale_factor,
     const bml_matrix_t * A)
 {
     bml_matrix_t *B = NULL;
@@ -50,7 +50,7 @@ bml_scale_new(
  */
 void
 bml_scale(
-    const double scale_factor,
+    const void *scale_factor,
     const bml_matrix_t * A,
     bml_matrix_t * B)
 {
@@ -80,7 +80,7 @@ bml_scale(
  */
 void
 bml_scale_inplace(
-    const double scale_factor,
+    const void *scale_factor,
     bml_matrix_t * A)
 {
     switch (bml_get_type(A))

src/C-interface/bml_scale.h

@@ -5,19 +5,17 @@
 
 #include "bml_types.h"
 
-// Scales A and returns a new B
 bml_matrix_t *bml_scale_new(
-    const double scale_factor,
+    const void *scale_factor,
     const bml_matrix_t * A);
 
-// Scales A and returns in existing B
 void bml_scale(
-    const double scale_factor,
+    const void *scale_factor,
     const bml_matrix_t * A,
     bml_matrix_t * B);
 
 void bml_scale_inplace(
-    const double scale_factor,
+    const void *scale_factor,
     bml_matrix_t * A);
 
 #endif

src/C-interface/dense/bml_add_dense_typed.c

@@ -146,6 +146,8 @@ void TYPED_FUNC(
     const double alpha,
     const double beta)
 {
-    bml_scale_inplace_dense(alpha, A);
+    REAL_T _alpha = (REAL_T) alpha;
+
+    bml_scale_inplace_dense(&_alpha, A);
     bml_add_identity_dense(A, beta);
 }

src/C-interface/dense/bml_adjungate_triangle_dense_typed.c

@@ -29,9 +29,7 @@ void TYPED_FUNC(
     bml_matrix_dense_t * A,
     char *triangle)
 {
-
     int N = A->N;
-    char mychar;
 
     REAL_T *A_matrix = A->matrix;
 

src/C-interface/dense/bml_copy_dense_typed.c

@@ -77,13 +77,11 @@ void TYPED_FUNC(
     }
 
     // Reorder elements in each row - just change index
-    REAL_T tmp;
 #pragma omp parallel for
     for (int i = 0; i < N; i++)
     {
         for (int j = 0; j < N; j++)
         {
-            tmp = A_matrix[ROWMAJOR(i, j, N, N)];
             A_matrix[ROWMAJOR(i, j, N, N)] =
                 B_matrix[ROWMAJOR(i, perm[j], N, N)];
         }

src/C-interface/dense/bml_multiply_dense_typed.c

@@ -44,12 +44,8 @@ void TYPED_FUNC(
     REAL_T alpha_ = (REAL_T) alpha;
     REAL_T beta_ = (REAL_T) beta;
 
-    int myRank = bml_getMyRank();
-
     C_BLAS(GEMM) ("N", "N", &A->N, &A->N, &A->N, &alpha_, B->matrix,
                   &A->N, A->matrix, &A->N, &beta_, C->matrix, &A->N);
-
-//    mkl_thread_free_buffers();
 }
 
 /** Matrix multiply.

src/C-interface/dense/bml_scale_dense.c

@@ -20,7 +20,7 @@
  */
 bml_matrix_dense_t *
 bml_scale_dense_new(
-    const double scale_factor,
+    const void *scale_factor,
     const bml_matrix_dense_t * A)
 {
     bml_matrix_dense_t *B = NULL;
@@ -57,7 +57,7 @@ bml_scale_dense_new(
  */
 void
 bml_scale_dense(
-    const double scale_factor,
+    const void *scale_factor,
     const bml_matrix_dense_t * A,
     bml_matrix_dense_t * B)
 {
@@ -85,7 +85,7 @@ bml_scale_dense(
 
 void
 bml_scale_inplace_dense(
-    const double scale_factor,
+    const void *scale_factor,
     bml_matrix_dense_t * A)
 {
     switch (A->matrix_precision)

src/C-interface/dense/bml_scale_dense.h

@@ -3,69 +3,71 @@
 
 #include "bml_types_dense.h"
 
+#include <complex.h>
+
 bml_matrix_dense_t *bml_scale_dense_new(
-    const double scale_factor,
+    const void *scale_factor,
     const bml_matrix_dense_t * A);
 
 bml_matrix_dense_t *bml_scale_dense_new_single_real(
-    const double scale_factor,
+    const float *scale_factor,
     const bml_matrix_dense_t * A);
 
 bml_matrix_dense_t *bml_scale_dense_new_double_real(
-    const double scale_factor,
+    const double *scale_factor,
     const bml_matrix_dense_t * A);
 
 bml_matrix_dense_t *bml_scale_dense_new_single_complex(
-    const double scale_factor,
+    const float complex * scale_factor,
     const bml_matrix_dense_t * A);
 
 bml_matrix_dense_t *bml_scale_dense_new_double_complex(
-    const double scale_factor,
+    const double complex * scale_factor,
     const bml_matrix_dense_t * A);
 
 void bml_scale_dense(
-    const double scale_factor,
+    const void *scale_factor,
     const bml_matrix_dense_t * A,
     bml_matrix_dense_t * B);
 
 void bml_scale_dense_single_real(
-    const double scale_factor,
+    const float *scale_factor,
     const bml_matrix_dense_t * A,
     bml_matrix_dense_t * B);
 
 void bml_scale_dense_double_real(
-    const double scale_factor,
+    const double *scale_factor,
     const bml_matrix_dense_t * A,
     bml_matrix_dense_t * B);
 
 void bml_scale_dense_single_complex(
-    const double scale_factor,
+    const float complex * scale_factor,
     const bml_matrix_dense_t * A,
     bml_matrix_dense_t * B);
 
 void bml_scale_dense_double_complex(
-    const double scale_factor,
+    const double complex * scale_factor,
     const bml_matrix_dense_t * A,
     bml_matrix_dense_t * B);
 
 void bml_scale_inplace_dense(
-    const double scale_factor,
+    const void *scale_factor,
     bml_matrix_dense_t * A);
 
 void bml_scale_inplace_dense_single_real(
-    const double scale_factor,
+    const float *scale_factor,
     bml_matrix_dense_t * A);
 
 void bml_scale_inplace_dense_double_real(
-    const double scale_factor,
+    const double *scale_factor,
     bml_matrix_dense_t * A);
 
 void bml_scale_inplace_dense_single_complex(
-    const double scale_factor,
+    const float complex * scale_factor,
     bml_matrix_dense_t * A);
 
 void bml_scale_inplace_dense_double_complex(
-    const double scale_factor,
+    const double complex * scale_factor,
     bml_matrix_dense_t * A);
 
 #endif

src/C-interface/dense/bml_scale_dense_typed.c

@@ -26,21 +26,17 @@
  */
 bml_matrix_dense_t *TYPED_FUNC(
     bml_scale_dense_new) (
-    const double scale_factor,
+    const REAL_T * scale_factor,
     const bml_matrix_dense_t * A)
 {
-    REAL_T sfactor = (REAL_T) scale_factor;
-
     bml_matrix_dense_t *B = TYPED_FUNC(bml_copy_dense_new) (A);
     REAL_T *B_matrix = B->matrix;
     int myRank = bml_getMyRank();
-    //int nElems = B->N * B->N;
     int nElems = B->domain->localRowExtent[myRank] * B->N;
     int startIndex = B->domain->localDispl[myRank];
     int inc = 1;
 
-    C_BLAS(SCAL) (&nElems, &sfactor, &(B_matrix[startIndex]), &inc);
-    //C_BLAS(SCAL) (&nElems, &sfactor, B->matrix, &inc);
+    C_BLAS(SCAL) (&nElems, scale_factor, &(B_matrix[startIndex]), &inc);
 
     return B;
 }
@@ -54,40 +50,33 @@ bml_matrix_dense_t *TYPED_FUNC(
  */
 void TYPED_FUNC(
     bml_scale_dense) (
-    const double scale_factor,
+    const REAL_T * scale_factor,
     const bml_matrix_dense_t * A,
     bml_matrix_dense_t * B)
 {
-    REAL_T sfactor = scale_factor;
-
     if (A != B)
         TYPED_FUNC(bml_copy_dense) (A, B);
 
     REAL_T *B_matrix = B->matrix;
     int myRank = bml_getMyRank();
-    //int nElems = B->N * B->N;
     int nElems = B->domain->localRowExtent[myRank] * B->N;
     int startIndex = B->domain->localDispl[myRank];
     int inc = 1;
 
-    C_BLAS(SCAL) (&nElems, &sfactor, &(B_matrix[startIndex]), &inc);
-    //C_BLAS(SCAL) (&nElems, &sfactor, B->matrix, &inc);
+    C_BLAS(SCAL) (&nElems, scale_factor, &(B_matrix[startIndex]), &inc);
 }
 
 void TYPED_FUNC(
     bml_scale_inplace_dense) (
-    const double scale_factor,
+    const REAL_T * scale_factor,
     bml_matrix_dense_t * A)
 {
-    REAL_T scale_factor_ = (REAL_T) scale_factor;
     REAL_T *A_matrix = A->matrix;
     int myRank = bml_getMyRank();
-    //int number_elements = A->N * A->N;
     int number_elements = A->domain->localRowExtent[myRank] * A->N;
     int startIndex = A->domain->localDispl[myRank];
     int inc = 1;
 
-    C_BLAS(SCAL) (&number_elements, &scale_factor_, &(A_matrix[startIndex]),
+    C_BLAS(SCAL) (&number_elements, scale_factor, &(A_matrix[startIndex]),
                   &inc);
-    //C_BLAS(SCAL) (&number_elements, &scale_factor_, A->matrix, &inc);
 }

src/C-interface/ellpack/bml_normalize_ellpack_typed.c

@@ -36,10 +36,10 @@ void TYPED_FUNC(
 {
     double maxminusmin = maxeval - mineval;
     double gershfact = maxeval / maxminusmin;
-    double scalar = (double) -1.0 / maxminusmin;
+    REAL_T scalar = (REAL_T) - 1.0 / maxminusmin;
     double threshold = 0.0;
 
-    bml_scale_inplace_ellpack(scalar, A);
+    bml_scale_inplace_ellpack(&scalar, A);
     bml_add_identity_ellpack(A, gershfact, threshold);
 }
 

src/C-interface/ellpack/bml_scale_ellpack.c

@@ -16,7 +16,7 @@
  */
 bml_matrix_ellpack_t *
 bml_scale_ellpack_new(
-    const double scale_factor,
+    const void * scale_factor,
     const bml_matrix_ellpack_t * A)
 {
     bml_matrix_ellpack_t *B = NULL;
@@ -51,7 +51,7 @@ bml_scale_ellpack_new(
  */
 void
 bml_scale_ellpack(
-    const double scale_factor,
+    const void * scale_factor,
     const bml_matrix_ellpack_t * A,
     const bml_matrix_ellpack_t * B)
 {
@@ -77,7 +77,7 @@ bml_scale_ellpack(
 
 void
 bml_scale_inplace_ellpack(
-    const double scale_factor,
+    const void * scale_factor,
     bml_matrix_ellpack_t * A)
 {
     switch (A->matrix_precision)