Added new functionality for mixed sparse-dense operations.

global/src/base.c

@@ -5897,7 +5897,6 @@ double pnga_rand(Integer iseed)
 {
   double ret;
   if (GA_Rand_seed == -1) {
-    printf("p[%d] Initializing GA_Rand seed: %d\n",GAme,iseed);
     unsigned long lseed;
     /* Choose a value for iseed if it has not already been set */
     if (iseed == 0) {

global/src/capi.c

@@ -5366,6 +5366,18 @@ int NGA_Sprs_array_create64(int64_t idim, int64_t jdim, int type)
   return (int)wnga_sprs_array_create(i,j,type,sizeof(int64_t));
 }
 
+int NGA_Sprs_array_create_from_dense(int g_a)
+{
+  Integer ga = (Integer)g_a;
+  return (int)wnga_sprs_array_create_from_dense(ga,sizeof(int),1);
+}
+
+int NGA_Sprs_array_create_from_dense64(int g_a)
+{
+  Integer ga = (Integer)g_a;
+  return (int)wnga_sprs_array_create_from_dense(ga,sizeof(int64_t),1);
+}
+
 void NGA_Sprs_array_add_element(int s_a, int idx, int jdx, void *val)
 {
   Integer sa = (Integer)s_a;
@@ -5385,7 +5397,7 @@ void NGA_Sprs_array_add_element64(int s_a, int64_t idx, int64_t jdx, void *val)
 int NGA_Sprs_array_assemble(int s_a)
 {
   Integer sa = (Integer)s_a;
-  wnga_sprs_array_assemble(sa);
+  return wnga_sprs_array_assemble(sa);
 }
 
 void NGA_Sprs_array_row_distribution(int s_a, int iproc, int *lo, int *hi)
@@ -5559,6 +5571,22 @@ int NGA_Sprs_array_matmat_multiply(int s_a, int s_b)
   return (int)wnga_sprs_array_matmat_multiply(sa, sb);
 }
 
+int NGA_Sprs_array_sprsdns_multiply(int s_a, int g_b)
+{
+  int g_c;
+  Integer sa = (Integer)s_a;
+  Integer gb = (Integer)g_b;
+  return (int)wnga_sprs_array_sprsdns_multiply(sa, gb, 1);
+}
+
+int NGA_Sprs_array_dnssprs_multiply(int g_a, int s_b)
+{
+  int g_c;
+  Integer ga = (Integer)g_a;
+  Integer sb = (Integer)s_b;
+  return (int)wnga_sprs_array_dnssprs_multiply(ga, sb, 1);
+}
+
 int NGA_Sprs_array_count_sketch(int s_a, int size_k, int *g_k, int *g_w)
 {
   Integer sa = (Integer)s_a;
@@ -5571,16 +5599,16 @@ int NGA_Sprs_array_count_sketch(int s_a, int size_k, int *g_k, int *g_w)
   return ret;
 }
 
-int NGA_Sprs_array_get_column(int g_v, int irow)
+int NGA_Sprs_array_get_column(int s_a, int icol)
 {
-  Integer gv = (Integer)g_v;
-  Integer ir = (Integer)irow;
-  return (int)wnga_sprs_array_get_column(gv, ir);
+  Integer sa = (Integer)s_a;
+  Integer ic = (Integer)icol;
+  return (int)wnga_sprs_array_get_column(sa, ic);
 }
 
-int NGA_Sprs_array_get_column64(int g_v, int64_t irow)
+int NGA_Sprs_array_get_column64(int s_a, int64_t icol)
 {
-  Integer gv = (Integer)g_v;
-  Integer ir = (Integer)irow;
-  return (int)wnga_sprs_array_get_column(gv, ir);
+  Integer sa = (Integer)s_a;
+  Integer ic = (Integer)icol;
+  return (int)wnga_sprs_array_get_column(sa, ic);
 }

global/src/cnames.h

@@ -3196,6 +3196,7 @@
 #define nga_sput_field_ F77_FUNC_(nga_sput_field,NGA_SPUT_FIELD)
 #define nga_zput_field_ F77_FUNC_(nga_zput_field,NGA_ZPUT_FIELD)
 #define nga_sprs_array_create_ F77_FUNC_(nga_sprs_array_create,NGA_SPRS_ARRAY_CREATE)
+#define nga_sprs_array_create_from_dense_ F77_FUNC_(nga_sprs_array_create_from_dense,NGA_SPRS_ARRAY_CREATE_FROM_DENSE)
 #define nga_sprs_array_add_element_ F77_FUNC_(nga_sprs_array_add_element,NGA_SPRS_ARRAY_ADD_ELEMENT)
 #define nga_sprs_array_assemble_ F77_FUNC_(nga_sprs_array_assemble,NGA_SPRS_ARRAY_ASSEMBLE)
 #define nga_sprs_array_row_distribution_ F77_FUNC_(nga_sprs_array_row_distribution,NGA_SPRS_ARRAY_ROW_DISTRIBUTION)
@@ -3210,6 +3211,8 @@
 #define nga_sprs_array_shift_diag_ F77_FUNC_(nga_sprs_array_shift_diag,NGA_SPRS_ARRAY_SHIFT_DIAG)
 #define nga_sprs_array_duplicate_ F77_FUNC_(nga_sprs_array_duplicate,NGA_SPRS_ARRAY_DUPLICATE)
 #define nga_sprs_array_matmat_multiply_ F77_FUNC_(nga_sprs_array_matmat_multiply,NGA_SPRS_ARRAY_MATMAT_MULTIPLY)
+#define nga_sprs_array_sprsdns_multiply_ F77_FUNC_(nga_sprs_array_sprsdns_multiply,NGA_SPRS_ARRAY_SPRSDNS_MULTIPLY)
+#define nga_sprs_array_dnssprs_multiply_ F77_FUNC_(nga_sprs_array_dnssprs_multiply,NGA_SPRS_ARRAY_DNSSPRS_MULTIPLY)
 #define nga_sprs_array_count_sketch_ F77_FUNC_(nga_sprs_array_count_sketch,NGA_SPRS_ARRAY_COUNT_SKETCH)
 /* the missing functions are either complex type or strangely named */
 

global/src/fapi.c

@@ -3768,6 +3768,11 @@ Integer nga_sprs_array_create_(Integer *idim, Integer *jdim, Integer *type)
   return wnga_sprs_array_create(*idim,*jdim,ctype,sizeof(Integer));
 }
 
+Integer nga_sprs_array_create_from_dense_(Integer *g_a)
+{
+  return wnga_sprs_array_create_from_dense(*g_a,sizeof(Integer),0);
+}
+
 void nga_sprs_array_add_element_(Integer *s_a, Integer *idx, Integer *jdx, void *val)
 {
   wnga_sprs_array_add_element(*s_a,*idx,*jdx,val);
@@ -3799,7 +3804,7 @@ void nga_sprs_array_access_col_block_(Integer *s_a, Integer *icol,
 void nga_sprs_array_col_block_list_(Integer *s_a, Integer *idx, Integer *n)
 {
   /* This function assumes that idx has already been allocated by calling
-   * program and that n is the length of idx on input. On output, n is
+   * program and that nn is the length of idx on input. On output, nn is
    * changed to the actual number of blocks found by the
    * wnga_sprs_array_col_block_list function
    */
@@ -3864,6 +3869,16 @@ Integer nga_sprs_array_matmat_multiply_(Integer *s_a, Integer *s_b)
   return wnga_sprs_array_matmat_multiply(*s_a, *s_b);
 }
 
+Integer nga_sprs_array_sprsdns_multiply_(Integer *s_a, Integer *g_b)
+{
+  return wnga_sprs_array_sprsdns_multiply(*s_a, *g_b, 0);
+}
+
+Integer nga_sprs_array_dnssprs_multiply_(Integer *g_a, Integer *s_b)
+{
+  return wnga_sprs_array_dnssprs_multiply(*g_a, *s_b, 0);
+}
+
 Integer nga_sprs_array_count_sketch_(Integer *s_a, Integer *size_k, Integer *g_k,
     Integer *g_w)
 {

global/src/ga-papi.h

@@ -400,7 +400,8 @@ extern void pnga_put_field(Integer g_a, Integer *lo, Integer *hi, Integer foff,
 
 /* Routines for sparse matrices */
 
-extern Integer pnga_sprs_array_create(Integer idim, Integer jdim, Integer type, Integer size);
+extern Integer pnga_sprs_array_create(Integer idim, Integer jdim, Integer type, Integer trans);
+extern Integer pnga_sprs_array_create_from_dense(Integer g_a, Integer idxsize, Integer flag);
 extern void pnga_sprs_array_add_element(Integer s_a, Integer idx, Integer jdx, void *val);
 extern logical pnga_sprs_array_assemble(Integer s_a);
 extern void pnga_sprs_array_row_distribution(Integer s_a, Integer iproc, Integer *lo, Integer *hi);
@@ -410,6 +411,8 @@ extern int pnga_sprs_array_access_col_block_idx(Integer s_a, Integer icol, Acces
 extern void pnga_sprs_array_col_block_list(Integer s_a, Integer **idx, Integer *n);
 extern Integer pnga_sprs_array_duplicate(Integer s_a);
 extern Integer pnga_sprs_array_matmat_multiply(Integer s_a, Integer s_b);
+extern Integer pnga_sprs_array_sprsdns_multiply(Integer s_a, Integer g_b, Integer trans);
+extern Integer pnga_sprs_array_dnssprs_multiply(Integer g_a, Integer s_b, Integer trans);
 extern Integer pnga_sprs_array_count_sketch(Integer s_a, Integer size_k,
     Integer *g_k, Integer *g_w, Integer trans);
 extern void pnga_sprs_array_matvec_multiply(Integer s_a, Integer g_a, Integer g_v);

global/src/ga.h

@@ -357,12 +357,14 @@ extern void          NGA_Sprs_array_access_col_block64(int s_a, int icol, long *
 extern void          NGA_Sprs_array_add_element(int s_a, int idx, int jdx, void *val);
 extern int           NGA_Sprs_array_assemble(int s_a);
 extern int           NGA_Sprs_array_create(int idim, int jdim, int type);
+extern int           NGA_Sprs_array_create_from_dense(int g_a);
 extern void          NGA_Sprs_array_column_distribution(int s_a, int iproc, int *lo, int *hi);
 extern void          NGA_Sprs_array_col_block_list(int s_a, int **idx, int *n);
 extern int           NGA_Sprs_array_count_sketch(int s_a, int size_k, int *g_k, int* g_w);
 extern int           NGA_Sprs_array_destroy(int s_a);
 extern void          NGA_Sprs_array_diag_right_multiply(int s_a, int g_d);
 extern void          NGA_Sprs_array_diag_left_multiply(int s_a, int g_d);
+extern int           NGA_Sprs_array_dnssprs_multiply(int g_a, int s_b);
 extern int           NGA_Sprs_array_duplicate(int s_a);
 extern void          NGA_Sprs_array_export(int s_a, const char* file);
 extern int           NGA_Sprs_array_get_block(int s_a, int irow, int icol,
@@ -375,6 +377,7 @@ extern int           NGA_Sprs_array_matmat_multiply(int s_a, int s_b);
 extern void          NGA_Sprs_array_matvec_multiply(int s_a, int g_a, int g_v);
 extern void          NGA_Sprs_array_row_distribution(int s_a, int iproc, int *lo, int *hi);
 extern void          NGA_Sprs_array_shift_diag(int s_a, void *shift);
+extern int           NGA_Sprs_array_sprsdns_multiply(int s_a, int g_b);
 extern void          NGA_Strided_acc(int g_a, int lo[], int hi[], int skip[], void* buf, int ld[], void *alpha); 
 extern void          NGA_Strided_get(int g_a, int lo[], int hi[], int skip[], void* buf, int ld[]); 
 extern void          NGA_Strided_put(int g_a, int lo[], int hi[], int skip[], void* buf, int ld[]); 
@@ -497,6 +500,7 @@ extern void          NGA_Strided_put64(int g_a, int64_t lo[], int64_t hi[], int6
 extern DoubleComplex NGA_Zdot_patch64(int g_a, char t_a, int64_t alo[], int64_t ahi[], int g_b, char t_b, int64_t blo[], int64_t bhi[]);
 extern void          NGA_Zero_patch64(int g_a, int64_t lo[], int64_t hi[]);
 extern int           NGA_Sprs_array_create64(int64_t idim, int64_t jdim, int type);
+extern int           NGA_Sprs_array_create_from_dense64(int g_a);
 extern void          NGA_Sprs_array_add_element64(int s_a, int64_t idx, int64_t jdx, void *val);
 extern void          NGA_Sprs_array_row_distribution64(int s_a, int iproc, int64_t *lo, int64_t *hi);
 extern void          NGA_Sprs_array_column_distribution64(int s_a, int iproc, int64_t *lo, int64_t *hi);

global/src/ga_sparse.array.h

@@ -15,6 +15,7 @@ typedef struct {
   Integer idx_size; /* size of integer indices */
   Integer grp;      /* handle for process group on which array is defined */
   Integer ilo, ihi; /* minimum and maximum row indices contained on this process */
+                    /* ilo and ihi are zero-based */
   Integer nprocs;   /* number of processors containing this array */
   Integer nblocks;  /* number of non-zero sparse blocks contained on this process */
   Integer *blkidx;  /* array containing indices of non-zero blocks */

global/src/global.fh.in

@@ -142,8 +142,12 @@
       integer          nga_spd_invert
       logical          nga_sprs_array_assemble
       integer          nga_sprs_array_create
+      integer          nga_sprs_array_create_from_dense
       integer          nga_sprs_array_duplicate
       logical          nga_sprs_array_destroy
+      integer          nga_sprs_array_dnssprs_multiply
+      integer          nga_sprs_array_matmat_multiply
+      integer          nga_sprs_array_sprsdns_multiply
       integer          nga_total_blocks
       logical          nga_update2_ghosts
       logical          nga_update3_ghosts
@@ -302,8 +306,12 @@
       external nga_spd_invert
       external nga_sprs_array_assemble
       external nga_sprs_array_create
+      external nga_sprs_array_create_from_dense
       external nga_sprs_array_duplicate
       external nga_sprs_array_destroy
+      external nga_sprs_array_dnssprs_multiply
+      external nga_sprs_array_matmat_multiply
+      external nga_sprs_array_sprsdns_multiply
       external nga_total_blocks
       external nga_update2_ghosts
       external nga_update3_ghosts