driver: corrected pztile2band

compute/pztile2band.c

@@ -21,21 +21,25 @@
  */
 #include "control/common.h"
 
-#define A(m,n)   A,  m, n
-#define B(m, n)  B, m, n
+#define A(m,n)  A, m, n
+#define B(m,n)  B, m, n
 
 /**
  *  Parallel copy of a band matrix from full NxN tile storage to band storage (LDABxN).
  */
-void chameleon_pztile2band(cham_uplo_t uplo, CHAM_desc_t *A, CHAM_desc_t *B,
-                       RUNTIME_sequence_t *sequence, RUNTIME_request_t *request)
+void chameleon_pztile2band( cham_uplo_t uplo, CHAM_desc_t *A, CHAM_desc_t *B,
+                            RUNTIME_sequence_t *sequence, RUNTIME_request_t *request )
 {
     CHAM_context_t *chamctxt;
     RUNTIME_option_t options;
+    CHAM_tile_t *tileA;
 
-    int j;
-    int tempjm, tempjn;
+    int k;
+    int tempkm, tempkn;
     int minmnt = chameleon_min(A->mt, A->nt);
+    int Bnb = B->nb;
+    int Bmb = B->mb;
+    int Amb = A->mb;
 
     chamctxt = chameleon_context_self();
     if (sequence->status != CHAMELEON_SUCCESS) {
@@ -43,71 +47,66 @@ void chameleon_pztile2band(cham_uplo_t uplo, CHAM_desc_t *A, CHAM_desc_t *B,
     }
     RUNTIME_options_init(&options, chamctxt, sequence, request);
 
-    /* The code is actually incorrect due to the removal of the ld (Need new insert_task dedicated) */
-    assert( 0 );
-
     /*
      *  ChamLower => Lower Band
      */
     if ( uplo == ChamLower ) {
-       for (j = 0; j < minmnt; j++){
-           /* Compute dimension on N with B since it is dimensioned with chameleon_min(A->m, A->n) */
-           assert( A->m == B->n );
-           assert( A->n >= B->n );
-
-           tempjm = j == A->mt-1 ? A->m - j * A->mb : A->mb;
-           tempjn = j == B->nt-1 ? B->n - j * B->nb : B->nb;
-
-           INSERT_TASK_zlaset(
-               &options,
-               ChamUpperLower, B->mb, tempjn,
-               0., 0.,
-               B(0, j) );
-
-           INSERT_TASK_zlacpy(
-               &options,
-               ChamLower, tempjm, tempjn, A->nb,
-               A(j, j),
-               B(0, j) );
-
-           if( j<minmnt-1 ){
-               tempjm = (j+1) == A->mt-1 ? A->m-(j+1)*A->mb : A->mb;
-               INSERT_TASK_zlacpyx(
-                   &options,
-                   ChamUpper, tempjm, tempjn, A->nb,
-                   0,     A(j+1, j),
-                   A->nb, B(0,   j));
-           }
-       }
+        for (k = 0; k < minmnt; k++){
+            tileA = A->get_blktile( A, k, k );
+
+            /* Computes dimension on N with B since it is dimensioned with chameleon_min(A->m, A->n) */
+            assert( A->i == B->j );
+            assert( A->j >= B->j );
+
+            tempkm = ( k == A->mt-1 ) ? A->m - k * Amb : Amb;
+            tempkn = ( k == B->nt-1 ) ? B->n - k * Bnb : Bnb;
+
+            INSERT_TASK_zlaset( &options, ChamUpperLower, Bmb, tempkn,
+                                0., 0., B, 0, k );
+
+            INSERT_TASK_zlacpyx( &options, ChamLower, tempkm, tempkn,
+                                 0, A, k, k, tileA->ld,
+                                 0, B, 0, k, Bmb-1 );
+
+            if ( k < minmnt-1 ) {
+                tileA = A->get_blktile( A, k+1, k );
+
+                tempkm = ( (k+1) == A->mt-1 ) ? A->m - (k+1) * Amb : Amb;
+
+                INSERT_TASK_zlacpyx( &options, ChamUpper, tempkm, tempkn,
+                                     0,     A, k+1, k, tileA->ld,
+                                     Bmb-1, B, 0,   k, Bmb-1 );
+            }
+        }
     }
     else if ( uplo == ChamUpper ) {
-       for (j = 0; j < minmnt; j++){
-           /* Compute dimension on M with B since it is dimensioned with chameleon_min(A->m, A->n) */
-           assert( A->n == B->n );
-           assert( A->m >= B->n );
-           tempjn = j == A->nt-1 ? A->n - j * A->nb : A->nb;
-
-           INSERT_TASK_zlaset(
-               &options,
-               ChamUpperLower, B->mb, tempjn,
-               0., 0.,
-               B(0, j) );
-
-           if(j > 0){
-               INSERT_TASK_zlacpy(
-                   &options,
-                   ChamLower, A->mb, tempjn, A->nb,
-                   A(j-1, j),
-                   B(0,   j));
-           }
-
-           tempjm = j == B->nt-1 ? B->n - j * B->nb : B->nb;
-           INSERT_TASK_zlacpyx(
-               &options,
-               ChamUpper, tempjm, tempjn, A->nb,
-               0,     A(j, j),
-               A->nb, B(0, j));
-       }
+        for (k = 0; k < minmnt; k++){
+            tileA = A->get_blktile( A, k, k );
+
+            /* Compute dimension on M with B since it is dimensioned with chameleon_min(A->m, A->n) */
+            assert( A->i == B->i );
+            assert( A->i >= B->j );
+
+            tempkm = ( k == A->mt-1 ) ? A->m - k * Amb : Amb;
+            tempkn = ( k == B->nt-1 ) ? B->n - k * Bnb : Bnb;
+
+            INSERT_TASK_zlaset( &options, ChamUpperLower, Bmb, tempkn,
+                                0., 0., B, 0, k );
+
+            INSERT_TASK_zlacpyx( &options, ChamUpper, tempkm, tempkn,
+                                 0,     A, k, k, tileA->ld,
+                                 Bmb-1, B, 0, k, Bmb-1 );
+
+            if ( k > 0 ) {
+                tileA = A->get_blktile( A, k-1, k );
+
+                tempkm = ( (k-1) == A->mt-1 ) ? A->m - (k-1) * Amb : Amb;
+
+                INSERT_TASK_zlacpyx( &options, ChamLower, tempkm, tempkn,
+                                     0, A, k-1, k, tileA->ld,
+                                     0, B, 0,   k, Bmb-1 );
+            }
+        }
     }
     RUNTIME_options_finalize(&options, chamctxt);
 }