Merge branch 'starpu/alloc_on_the_fly' into 'master'

Starpu/alloc on the fly

See merge request solverstack/chameleon!140

compute/pzlange.c

@@ -72,11 +72,10 @@ chameleon_pzlange_one( cham_uplo_t uplo, cham_diag_t diag, CHAM_desc_t *A,
             }
 
             if ( m >= P ) {
-                INSERT_TASK_dgeadd(
-                    options,
-                    ChamNoTrans, 1, tempnn, A->nb,
-                    1.0, W( Wcol, m,   n ), 1,
-                    1.0, W( Wcol, m%P, n ), 1 );
+                INSERT_TASK_daxpy(
+                    options, tempnn, 1.,
+                    W( Wcol, m,   n ), 1,
+                    W( Wcol, m%P, n ), 1 );
             }
         }
 
@@ -85,11 +84,10 @@ chameleon_pzlange_one( cham_uplo_t uplo, cham_diag_t diag, CHAM_desc_t *A,
          *  For each i, W(i, n) = reduce( W(0..P-1, n) )
          */
         for(m = 1; m < P; m++) {
-            INSERT_TASK_dgeadd(
-                options,
-                ChamNoTrans, 1, tempnn, A->nb,
-                1.0, W( Wcol, m, n ), 1,
-                1.0, W( Wcol, 0, n ), 1 );
+            INSERT_TASK_daxpy(
+                options, tempnn, 1.,
+                W( Wcol, m, n ), 1,
+                W( Wcol, 0, n ), 1 );
         }
 
         INSERT_TASK_dlange(
@@ -165,11 +163,10 @@ chameleon_pzlange_inf( cham_uplo_t uplo, cham_diag_t diag, CHAM_desc_t *A,
             }
 
             if ( n >= Q ) {
-                INSERT_TASK_dgeadd(
-                    options,
-                    ChamNoTrans, tempmm, 1, A->mb,
-                    1.0, W( Wcol, m, n  ), tempmm,
-                    1.0, W( Wcol, m, n%Q), tempmm );
+                INSERT_TASK_daxpy(
+                    options, tempmm, 1.,
+                    W( Wcol, m, n   ), 1,
+                    W( Wcol, m, n%Q ), 1 );
             }
         }
 
@@ -178,11 +175,10 @@ chameleon_pzlange_inf( cham_uplo_t uplo, cham_diag_t diag, CHAM_desc_t *A,
          *  For each j, W(m, j) = reduce( Wcol(m, 0..Q-1) )
          */
         for(n = 1; n < Q; n++) {
-            INSERT_TASK_dgeadd(
-                options,
-                ChamNoTrans, tempmm, 1, A->mb,
-                1.0, W( Wcol, m, n), tempmm,
-                1.0, W( Wcol, m, 0), tempmm );
+            INSERT_TASK_daxpy(
+                options, tempmm, 1.,
+                W( Wcol, m, n ), 1,
+                W( Wcol, m, 0 ), 1 );
         }
 
         INSERT_TASK_dlange(
@@ -407,11 +403,14 @@ void chameleon_pzlange_generic( cham_normtype_t norm, cham_uplo_t uplo, cham_dia
     case ChamOneNorm:
         RUNTIME_options_ws_alloc( &options, 1, 0 );
 
-        chameleon_desc_init( &Wcol, CHAMELEON_MAT_ALLOC_GLOBAL, ChamRealDouble, 1, A->nb, A->nb,
+        chameleon_desc_init( &Wcol, CHAMELEON_MAT_ALLOC_TILE, ChamRealDouble, 1, A->nb, A->nb,
                              workmt, worknt * A->nb, 0, 0, workmt, worknt * A->nb, A->p, A->q,
                              NULL, NULL, NULL );
         wcol_init = 1;
 
+        /*
+         * Use the global allocator for Welt, otherwise flush may free the data before the result is read.
+         */
         chameleon_desc_init( &Welt, CHAMELEON_MAT_ALLOC_GLOBAL, ChamRealDouble, 1, 1, 1,
                              A->p, worknt, 0, 0, A->p, worknt, A->p, A->q,
                              NULL, NULL, NULL );
@@ -424,7 +423,7 @@ void chameleon_pzlange_generic( cham_normtype_t norm, cham_uplo_t uplo, cham_dia
     case ChamInfNorm:
         RUNTIME_options_ws_alloc( &options, A->mb, 0 );
 
-        chameleon_desc_init( &Wcol, CHAMELEON_MAT_ALLOC_GLOBAL, ChamRealDouble, A->mb, 1, A->mb,
+        chameleon_desc_init( &Wcol, CHAMELEON_MAT_ALLOC_TILE, ChamRealDouble, A->mb, 1, A->mb,
                              workmt * A->mb, worknt, 0, 0, workmt * A->mb, worknt, A->p, A->q,
                              NULL, NULL, NULL );
         wcol_init = 1;
@@ -522,7 +521,7 @@ void chameleon_pzlange_generic( cham_normtype_t norm, cham_uplo_t uplo, cham_dia
     }
     CHAMELEON_Desc_Flush( &Welt, sequence );
     CHAMELEON_Desc_Flush( A, sequence );
-    RUNTIME_sequence_wait(chamctxt, sequence);
+    RUNTIME_sequence_wait( chamctxt, sequence );
 
     *result = *((double *)Welt.get_blkaddr( &Welt, A->myrank / A->q, A->myrank % A->q ));
 

compute/pzlansy.c

@@ -81,11 +81,10 @@ chameleon_pzlansy_inf( cham_uplo_t uplo, CHAM_desc_t *A,
         int tempmm = ( m == (MT-1) ) ? M - m * A->mb : A->mb;
 
         for(n = Q; n < NT; n++) {
-            INSERT_TASK_dgeadd(
-                options,
-                ChamNoTrans, tempmm, 1, A->nb,
-                1.0, W( Wcol, m, n  ), tempmm,
-                1.0, W( Wcol, m, n%Q), tempmm );
+            INSERT_TASK_daxpy(
+                options, tempmm, 1.,
+                W( Wcol, m, n   ), 1,
+                W( Wcol, m, n%Q ), 1 );
         }
 
         /**
@@ -93,11 +92,10 @@ chameleon_pzlansy_inf( cham_uplo_t uplo, CHAM_desc_t *A,
          *  For each j, W(m, j) = reduce( Wcol(m, 0..Q-1) )
          */
         for(n = 1; n < Q; n++) {
-            INSERT_TASK_dgeadd(
-                options,
-                ChamNoTrans, tempmm, 1, A->mb,
-                1.0, W( Wcol, m, n), tempmm,
-                1.0, W( Wcol, m, 0), tempmm );
+            INSERT_TASK_daxpy(
+                options, tempmm, 1.,
+                W( Wcol, m, n ), 1,
+                W( Wcol, m, 0 ), 1 );
         }
 
         INSERT_TASK_dlange(
@@ -334,11 +332,14 @@ void chameleon_pzlansy_generic( cham_normtype_t norm, cham_uplo_t uplo, cham_tra
     case ChamInfNorm:
         RUNTIME_options_ws_alloc( &options, 1, 0 );
 
-        chameleon_desc_init( &Wcol, CHAMELEON_MAT_ALLOC_GLOBAL, ChamRealDouble, A->mb, 1, A->mb,
+        chameleon_desc_init( &Wcol, CHAMELEON_MAT_ALLOC_TILE, ChamRealDouble, A->mb, 1, A->mb,
                              workmt * A->mb, worknt, 0, 0, workmt * A->mb, worknt, A->p, A->q,
                              NULL, NULL, NULL );
         wcol_init = 1;
 
+        /*
+         * Use the global allocator for Welt, otherwise flush may free the data before the result is read.
+         */
         chameleon_desc_init( &Welt, CHAMELEON_MAT_ALLOC_GLOBAL, ChamRealDouble, 1, 1, 1,
                              workmt, A->q, 0, 0, workmt, A->q, A->p, A->q,
                              NULL, NULL, NULL );

compute/pzunmlq_param.c

@@ -466,6 +466,12 @@ void chameleon_pzunmlq_param( int genD, const libhqr_tree_t *qrtree,
                     RUNTIME_data_flush( sequence, T(k, n) );
                 }
 
+                /* Restore the original location of the tiles */
+                for (m = 0; m < B->mt; m++) {
+                    RUNTIME_data_migrate( sequence, B( m, k ),
+                                          B->get_rankof( B, m, k ) );
+                }
+
                 RUNTIME_iteration_pop(chamctxt);
             }
         }

compute/pzunmqr_param.c

@@ -467,6 +467,12 @@ void chameleon_pzunmqr_param( int genD, const libhqr_tree_t *qrtree,
                     RUNTIME_data_flush( sequence, T(n, k) );
                 }
 
+                /* Restore the original location of the tiles */
+                for (m = 0; m < B->mt; m++) {
+                    RUNTIME_data_migrate( sequence, B(m, k),
+                                          B->get_rankof( B, m, k ) );
+                }
+
                 RUNTIME_iteration_pop(chamctxt);
             }
         }

compute/zlaset.c

@@ -266,7 +266,7 @@ int CHAMELEON_zlaset_Tile_Async( cham_uplo_t uplo,
         return chameleon_request_fail(sequence, request, CHAMELEON_ERR_ILLEGAL_VALUE);
     }
     /* Check input arguments */
-    if (A->nb != A->mb) {
+    if ( (alpha != beta) && (A->nb != A->mb) ) {
         chameleon_error("CHAMELEON_zlaset_Tile_Async", "only square tiles supported");
         return chameleon_request_fail(sequence, request, CHAMELEON_ERR_ILLEGAL_VALUE);
     }

control/compute_z.h

@@ -134,7 +134,7 @@ void chameleon_pzungqr_param( int genD, int K, const libhqr_tree_t *qrtree,
 static inline int
 chameleon_zdesc_alloc_diag( CHAM_desc_t *descA, int nb, int m, int n, int p, int q ) {
     int diag_m = chameleon_min( m, n );
-    return chameleon_desc_init( descA, CHAMELEON_MAT_ALLOC_GLOBAL,
+    return chameleon_desc_init( descA, CHAMELEON_MAT_ALLOC_TILE,
                                 ChamComplexDouble, nb, nb, nb*nb,
                                 diag_m, nb, 0, 0, diag_m, nb, p, q,
                                 chameleon_getaddr_diag,
@@ -145,7 +145,7 @@ chameleon_zdesc_alloc_diag( CHAM_desc_t *descA, int nb, int m, int n, int p, int
 #define chameleon_zdesc_alloc( descA, mb, nb, lm, ln, i, j, m, n, free) \
     {                                                                   \
         int rc;                                                         \
-        rc = chameleon_desc_init( &(descA), CHAMELEON_MAT_ALLOC_GLOBAL, \
+        rc = chameleon_desc_init( &(descA), CHAMELEON_MAT_ALLOC_TILE, \
                                   ChamComplexDouble, (mb), (nb), ((mb)*(nb)), \
                                   (m), (n), (i), (j), (m), (n), 1, 1,   \
                                   NULL, NULL, NULL );                   \
@@ -174,7 +174,7 @@ chameleon_zlap2tile( CHAM_context_t *chamctxt,
 
     if ( CHAMELEON_TRANSLATION == ChamOutOfPlace ) {
         /* Initialize the tile descriptor */
-        chameleon_desc_init( descAt, CHAMELEON_MAT_ALLOC_GLOBAL, ChamComplexDouble, mb, nb, (mb)*(nb),
+        chameleon_desc_init( descAt, CHAMELEON_MAT_ALLOC_TILE, ChamComplexDouble, mb, nb, (mb)*(nb),
                              lm, ln, 0, 0, m, n, 1, 1,
                              chameleon_getaddr_ccrb, chameleon_getblkldd_ccrb, NULL );
 
@@ -235,6 +235,7 @@ chameleon_ztile2lap( CHAM_context_t *chamctxt, CHAM_desc_t *descAl, CHAM_desc_t
 static inline void
 chameleon_ztile2lap_cleanup( CHAM_context_t *chamctxt, CHAM_desc_t *descAl, CHAM_desc_t *descAt )
 {
+    (void)chamctxt;
     chameleon_desc_destroy( descAl );
     chameleon_desc_destroy( descAt );
 }

control/descriptor.c

@@ -226,26 +226,32 @@ int chameleon_desc_init( CHAM_desc_t *desc, void *mat,
     /* The matrix is alocated tile by tile with out of core */
     desc->ooc = 0;
 
-    // Matrix address
-    if ( mat == CHAMELEON_MAT_ALLOC_GLOBAL ) {
-        rc = chameleon_desc_mat_alloc( desc );
+    switch ( (intptr_t)mat ) {
+    case (intptr_t)CHAMELEON_MAT_ALLOC_TILE:
+        if ( chamctxt->scheduler == RUNTIME_SCHED_STARPU ) {
+            /* Let's use the allocation on the fly as in OOC */
+            desc->get_blkaddr = chameleon_getaddr_null;
+            desc->mat = NULL;
+            break;
+        }
+        /* Otherwise we switch back to the full allocation */
 
-        desc->alloc_mat = 1;
-        desc->use_mat   = 1;
-    }
-    else if ( mat == CHAMELEON_MAT_ALLOC_TILE ) {
-        //chameleon_error( "chameleon_desc_init", "CHAMELEON_MAT_ALLOC_TILE is not available yet" );
-        //desc->mat = NULL;
+    case (intptr_t)CHAMELEON_MAT_ALLOC_GLOBAL:
         rc = chameleon_desc_mat_alloc( desc );
+        desc->alloc_mat = 1;
         desc->use_mat   = 1;
+        break;
 
-        desc->alloc_mat = 1;
-    }
-    else if ( mat == CHAMELEON_MAT_OOC ) {
+    case (intptr_t)CHAMELEON_MAT_OOC:
+        if ( chamctxt->scheduler != RUNTIME_SCHED_STARPU ) {
+            chameleon_error("CHAMELEON_Desc_Create", "CHAMELEON Out-of-Core descriptors are supported only with StarPU");
+            return CHAMELEON_ERR_NOT_SUPPORTED;
+        }
         desc->mat = NULL;
         desc->ooc = 1;
-    }
-    else {
+        break;
+
+    default:
         /* memory of the matrix is handled by users */
         desc->mat     = mat;
         desc->use_mat = 1;

control/workspace.c

@@ -74,7 +74,8 @@ int chameleon_alloc_ibnb_tile(int M, int N, cham_tasktype_t func, int type, CHAM
     lm = IB * MT;
     ln = NB * NT;
 
-    return CHAMELEON_Desc_Create( desc, NULL, type, IB, NB, IB*NB, lm, ln, 0, 0, lm, ln, p, q );
+    return CHAMELEON_Desc_Create( desc, CHAMELEON_MAT_ALLOC_TILE, type, IB, NB, IB*NB,
+                                  lm, ln, 0, 0, lm, ln, p, q );
 }
 
 /**
@@ -119,7 +120,8 @@ int chameleon_alloc_ipiv(int M, int N, cham_tasktype_t func, int type, CHAM_desc
     /* TODO: Fix the distribution for IPIV */
     *IPIV = (int*)malloc( size );
 
-    return CHAMELEON_Desc_Create( desc, NULL, type, IB, NB, IB*NB, lm, ln, 0, 0, lm, ln, p, q );
+    return CHAMELEON_Desc_Create( desc, CHAMELEON_MAT_ALLOC_TILE, type, IB, NB, IB*NB,
+                                  lm, ln, 0, 0, lm, ln, p, q );
 }
 
 /**

coreblas/compute/core_zgelqt.c

@@ -67,7 +67,7 @@
  *         The leading dimension of the array A.  LDA >= max(1,M).
  *
  * @param[out] T
- *         The IB-by-N triangular factor T of the block reflector.
+ *         The IB-by-M triangular factor T of the block reflector.
  *         T is upper triangular by block (economic storage);
  *         The rest of the array is not referenced.
  *

include/chameleon/tasks_z.h

@@ -483,6 +483,8 @@ INSERT_TASK_ztsmlq( const RUNTIME_option_t *options,
                     const CHAM_desc_t *V, int Vm, int Vn, int ldv,
                     const CHAM_desc_t *T, int Tm, int Tn, int ldt )
 {
+    (void)m1;
+    (void)n1;
     return INSERT_TASK_ztpmlqt( options, side, trans, m2, n2, k, 0, ib, nb,
                                 V, Vm, Vn, ldv, T, Tm, Tn, ldt,
                                 A1, A1m, A1n, lda1, A2, A2m, A2n, lda2 );
@@ -497,6 +499,8 @@ INSERT_TASK_ztsmqr( const RUNTIME_option_t *options,
                     const CHAM_desc_t *V, int Vm, int Vn, int ldv,
                     const CHAM_desc_t *T, int Tm, int Tn, int ldt )
 {
+    (void)m1;
+    (void)n1;
     return INSERT_TASK_ztpmqrt( options, side, trans, m2, n2, k, 0, ib, nb,
                                 V, Vm, Vn, ldv, T, Tm, Tn, ldt,
                                 A1, A1m, A1n, lda1, A2, A2m, A2n, lda2 );
@@ -511,6 +515,8 @@ INSERT_TASK_zttmlq( const RUNTIME_option_t *options,
                     const CHAM_desc_t *V, int Vm, int Vn, int ldv,
                     const CHAM_desc_t *T, int Tm, int Tn, int ldt )
 {
+    (void)m1;
+    (void)n1;
     return INSERT_TASK_ztpmlqt( options, side, trans, m2, n2, k, n2, ib, nb,
                                 V, Vm, Vn, ldv, T, Tm, Tn, ldt,
                                 A1, A1m, A1n, lda1, A2, A2m, A2n, lda2 );
@@ -525,6 +531,8 @@ INSERT_TASK_zttmqr( const RUNTIME_option_t *options,
                     const CHAM_desc_t *V, int Vm, int Vn, int ldv,
                     const CHAM_desc_t *T, int Tm, int Tn, int ldt )
 {
+    (void)m1;
+    (void)n1;
     return INSERT_TASK_ztpmqrt( options, side, trans, m2, n2, k, m2, ib, nb,
                                 V, Vm, Vn, ldv, T, Tm, Tn, ldt,
                                 A1, A1m, A1n, lda1, A2, A2m, A2n, lda2 );

runtime/openmp/codelets/codelet_zgelqt.c

@@ -98,10 +98,13 @@ void INSERT_TASK_zgelqt(const RUNTIME_option_t *options,
     CHAMELEON_Complex64_t *ptrA = RTBLKADDR(A, CHAMELEON_Complex64_t, Am, An);
     CHAMELEON_Complex64_t *ptrT = RTBLKADDR(T, CHAMELEON_Complex64_t, Tm, Tn);
     int ws_size = options->ws_wsize;
-#pragma omp task firstprivate(ws_size, m, n, ib, ptrA, lda, ptrT, ldt) depend(inout:ptrA[0]) depend(inout:ptrT[0])
+
+#pragma omp task firstprivate(ws_size, m, n, ib, ptrA, lda, ptrT, ldt) depend(inout:ptrA[0]) depend(out:ptrT[0])
     {
       CHAMELEON_Complex64_t TAU[ws_size];
       CHAMELEON_Complex64_t *work = TAU + chameleon_max( m, n );
-      CORE_zgelqt(m, n, ib, ptrA, lda, ptrT, ldt, TAU, work);
+
+      CORE_zlaset( ChamUpperLower, ib, m, 0., 0., ptrT, ldt );
+      CORE_zgelqt( m, n, ib, ptrA, lda, ptrT, ldt, TAU, work );
     }
 }

runtime/openmp/codelets/codelet_zgeqrt.c

@@ -99,10 +99,13 @@ void INSERT_TASK_zgeqrt(const RUNTIME_option_t *options,
     CHAMELEON_Complex64_t *ptrA = RTBLKADDR(A, CHAMELEON_Complex64_t, Am, An);
     CHAMELEON_Complex64_t *ptrT = RTBLKADDR(T, CHAMELEON_Complex64_t, Tm, Tn);
     int ws_size = options->ws_wsize;
-#pragma omp task firstprivate(ws_size, m, n, ib, ptrA, lda, ptrT, ldt) depend(inout:ptrA[0]) depend(inout:ptrT[0])
+
+#pragma omp task firstprivate(ws_size, m, n, ib, ptrA, lda, ptrT, ldt) depend(inout:ptrA[0]) depend(out:ptrT[0])
     {
       CHAMELEON_Complex64_t TAU[ws_size];
       CHAMELEON_Complex64_t *work = TAU + chameleon_max(m, n);
-      CORE_zgeqrt(m, n, ib, ptrA, lda, ptrT, ldt, TAU, work);
+
+      CORE_zlaset( ChamUpperLower, ib, n, 0., 0., ptrT, ldt );
+      CORE_zgeqrt( m, n, ib, ptrA, lda, ptrT, ldt, TAU, work );
     }
 }

runtime/openmp/codelets/codelet_ztplqt.c

@@ -31,9 +31,13 @@ INSERT_TASK_ztplqt( const RUNTIME_option_t *options,
     CHAMELEON_Complex64_t *ptrB = RTBLKADDR(B, CHAMELEON_Complex64_t, Bm, Bn);
     CHAMELEON_Complex64_t *ptrT = RTBLKADDR(T, CHAMELEON_Complex64_t, Tm, Tn);
     int ws_size = options->ws_wsize;
-#pragma omp task firstprivate(ws_size, M, N, L, ib, ptrA, lda, ptrB, ldb, ptrT, ldt) depend(inout:ptrA[0], ptrB[0], ptrT[0])
+
+#pragma omp task firstprivate(ws_size, M, N, L, ib, ptrA, lda, ptrB, ldb, ptrT, ldt) depend(inout:ptrA[0], ptrB[0]) depend(out:ptrT[0])
     {
       CHAMELEON_Complex64_t work[ws_size];
+
+      CORE_zlaset( ChamUpperLower, ib, M, 0., 0., ptrT, ldt);
+
       CORE_ztplqt( M, N, L, ib,
                    ptrA, lda, ptrB, ldb, ptrT, ldt, work );
     }

runtime/openmp/codelets/codelet_ztpqrt.c

@@ -30,9 +30,13 @@ INSERT_TASK_ztpqrt( const RUNTIME_option_t *options,
     CHAMELEON_Complex64_t *ptrB = RTBLKADDR(B, CHAMELEON_Complex64_t, Bm, Bn);
     CHAMELEON_Complex64_t *ptrT = RTBLKADDR(T, CHAMELEON_Complex64_t, Tm, Tn);
     int ws_size = options->ws_wsize;
-#pragma omp task firstprivate(ws_size, M, N, L, ib, ptrT, ldt, ptrA, lda, ptrB, ldb) depend(in:ptrT[0]) depend(inout:ptrA[0], ptrB[0])
+
+#pragma omp task firstprivate(ws_size, M, N, L, ib, ptrT, ldt, ptrA, lda, ptrB, ldb) depend(inout:ptrA[0], ptrB[0]) depend(out:ptrT[0])
     {
       CHAMELEON_Complex64_t tmp[ws_size];
+
+      CORE_zlaset( ChamUpperLower, ib, N, 0., 0., ptrT, ldt);
+
       CORE_ztpqrt( M, N, L, ib,
           ptrA, lda, ptrB, ldb, ptrT, ldt, tmp );
     }

runtime/parsec/codelets/codelet_zgelqt.c

@@ -98,6 +98,7 @@ CORE_zgelqt_parsec( parsec_execution_stream_t *context,
     parsec_dtd_unpack_args(
         this_task, &m, &n, &ib, &A, &lda, &T, &ldt, &TAU, &WORK );
 
+    CORE_zlaset( ChamUpperLower, ib, m, 0., 0., T, ldt );
     CORE_zgelqt( m, n, ib, A, lda, T, ldt, TAU, WORK );
 
     (void)context;

runtime/parsec/codelets/codelet_zgeqrt.c

@@ -99,6 +99,7 @@ CORE_zgeqrt_parsec ( parsec_execution_stream_t *context,
     parsec_dtd_unpack_args(
         this_task, &m, &n, &ib, &A, &lda, &T, &ldt, &TAU, &WORK );
 
+    CORE_zlaset( ChamUpperLower, ib, n, 0., 0., T, ldt );
     CORE_zgeqrt( m, n, ib, A, lda, T, ldt, TAU, WORK );
 
     (void)context;

runtime/parsec/codelets/codelet_ztplqt.c

@@ -40,6 +40,7 @@ CORE_ztplqt_parsec( parsec_execution_stream_t *context,
     parsec_dtd_unpack_args(
         this_task, &M, &N, &L, &ib, &A, &lda, &B, &ldb, &T, &ldt, &WORK );
 
+    CORE_zlaset( ChamUpperLower, ib, M, 0., 0., T, ldt );
     CORE_ztplqt( M, N, L, ib,
                  A, lda, B, ldb, T, ldt, WORK );
 

runtime/parsec/codelets/codelet_ztpqrt.c

@@ -40,6 +40,7 @@ CORE_ztpqrt_parsec( parsec_execution_stream_t *context,
     parsec_dtd_unpack_args(
         this_task, &M, &N, &L, &ib, &A, &lda, &B, &ldb, &T, &ldt, &WORK );
 
+    CORE_zlaset( ChamUpperLower, ib, N, 0., 0., T, ldt );
     CORE_ztpqrt( M, N, L, ib,
                  A, lda, B, ldb, T, ldt, WORK );
 

runtime/quark/codelets/codelet_zgelqt.c

@@ -40,6 +40,7 @@ void CORE_zgelqt_quark(Quark *quark)
     CHAMELEON_Complex64_t *WORK;
 
     quark_unpack_args_9(quark, m, n, ib, A, lda, T, ldt, TAU, WORK);
+    CORE_zlaset( ChamUpperLower, ib, m, 0., 0., T, ldt );
     CORE_zgelqt(m, n, ib, A, lda, T, ldt, TAU, WORK);
 }
 

runtime/quark/codelets/codelet_zgeqrt.c

@@ -40,6 +40,7 @@ void CORE_zgeqrt_quark(Quark *quark)
     CHAMELEON_Complex64_t *WORK;
 
     quark_unpack_args_9(quark, m, n, ib, A, lda, T, ldt, TAU, WORK);
+    CORE_zlaset( ChamUpperLower, ib, n, 0., 0., T, ldt );
     CORE_zgeqrt(m, n, ib, A, lda, T, ldt, TAU, WORK);
 }
 

runtime/quark/codelets/codelet_ztplqt.c

@@ -39,6 +39,7 @@ CORE_ztplqt_quark( Quark *quark )
     quark_unpack_args_11( quark, M, N, L, ib,
                           A, lda, B, ldb, T, ldt, WORK );
 
+    CORE_zlaset( ChamUpperLower, ib, N, 0., 0., T, ldt );
     CORE_ztplqt( M, N, L, ib,
                  A, lda, B, ldb, T, ldt, WORK );
 }

runtime/quark/codelets/codelet_ztpqrt.c

@@ -39,6 +39,7 @@ CORE_ztpqrt_quark( Quark *quark )
     quark_unpack_args_11( quark, M, N, L, ib,
                           A, lda, B, ldb, T, ldt, WORK );
 
+    CORE_zlaset( ChamUpperLower, ib, N, 0., 0., T, ldt );
     CORE_ztpqrt( M, N, L, ib,
                  A, lda, B, ldb, T, ldt, WORK );
 }

runtime/starpu/codelets/codelet_zgelqt.c

@@ -26,6 +26,36 @@
 #include "chameleon_starpu.h"
 #include "runtime_codelet_z.h"
 
+#if !defined(CHAMELEON_SIMULATION)
+static void cl_zgelqt_cpu_func(void *descr[], void *cl_arg)
+{
+    CHAMELEON_starpu_ws_t *h_work;
+    int m;
+    int n;
+    int ib;
+    CHAMELEON_Complex64_t *A;
+    int lda;
+    CHAMELEON_Complex64_t *T;
+    int ldt;
+    CHAMELEON_Complex64_t *TAU, *WORK;
+
+    A   = (CHAMELEON_Complex64_t *)STARPU_MATRIX_GET_PTR(descr[0]);
+    T   = (CHAMELEON_Complex64_t *)STARPU_MATRIX_GET_PTR(descr[1]);
+    TAU = (CHAMELEON_Complex64_t *)STARPU_MATRIX_GET_PTR(descr[2]); /* max(m,n) + ib*n */
+
+    starpu_codelet_unpack_args(cl_arg, &m, &n, &ib, &lda, &ldt, &h_work);
+
+    WORK = TAU + chameleon_max( m, n );
+    CORE_zlaset( ChamUpperLower, ib, m, 0., 0., T, ldt );
+    CORE_zgelqt(m, n, ib, A, lda, T, ldt, TAU, WORK);
+}
+#endif /* !defined(CHAMELEON_SIMULATION) */
+
+/*
+ * Codelet definition
+ */
+CODELETS_CPU(zgelqt, 3, cl_zgelqt_cpu_func)
+
 /**
  *
  * @ingroup INSERT_TASK_Complex64_t
@@ -87,7 +117,6 @@
  *          \retval <0 if -i, the i-th argument had an illegal value
  *
  */
-
 void INSERT_TASK_zgelqt(const RUNTIME_option_t *options,
                        int m, int n, int ib, int nb,
                        const CHAM_desc_t *A, int Am, int An, int lda,
@@ -123,33 +152,3 @@ void INSERT_TASK_zgelqt(const RUNTIME_option_t *options,
 #endif
         0);
 }
-
-
-#if !defined(CHAMELEON_SIMULATION)
-static void cl_zgelqt_cpu_func(void *descr[], void *cl_arg)
-{
-    CHAMELEON_starpu_ws_t *h_work;
-    int m;
-    int n;
-    int ib;
-    CHAMELEON_Complex64_t *A;
-    int lda;
-    CHAMELEON_Complex64_t *T;
-    int ldt;
-    CHAMELEON_Complex64_t *TAU, *WORK;
-
-    A   = (CHAMELEON_Complex64_t *)STARPU_MATRIX_GET_PTR(descr[0]);
-    T   = (CHAMELEON_Complex64_t *)STARPU_MATRIX_GET_PTR(descr[1]);
-    TAU = (CHAMELEON_Complex64_t *)STARPU_MATRIX_GET_PTR(descr[2]); /* max(m,n) + ib*n */
-
-    starpu_codelet_unpack_args(cl_arg, &m, &n, &ib, &lda, &ldt, &h_work);
-
-    WORK = TAU + chameleon_max( m, n );
-    CORE_zgelqt(m, n, ib, A, lda, T, ldt, TAU, WORK);
-}
-#endif /* !defined(CHAMELEON_SIMULATION) */
-
-/*
- * Codelet definition
- */
-CODELETS_CPU(zgelqt, 3, cl_zgelqt_cpu_func)

runtime/starpu/codelets/codelet_zgemm.c

@@ -35,7 +35,7 @@ void INSERT_TASK_zgemm(const RUNTIME_option_t *options,
                       cham_trans_t transA, cham_trans_t transB,
                       int m, int n, int k, int nb,
                       CHAMELEON_Complex64_t alpha, const CHAM_desc_t *A, int Am, int An, int lda,
-                                               const CHAM_desc_t *B, int Bm, int Bn, int ldb,
+                                                   const CHAM_desc_t *B, int Bm, int Bn, int ldb,
                       CHAMELEON_Complex64_t beta,  const CHAM_desc_t *C, int Cm, int Cn, int ldc)
 {
     (void)nb;

runtime/starpu/codelets/codelet_zgeqrt.c

@@ -26,6 +26,37 @@
 #include "chameleon_starpu.h"
 #include "runtime_codelet_z.h"
 
+#if !defined(CHAMELEON_SIMULATION)
+static void cl_zgeqrt_cpu_func(void *descr[], void *cl_arg)
+{
+    CHAMELEON_starpu_ws_t *h_work;
+    int m;
+    int n;
+    int ib;
+    CHAMELEON_Complex64_t *A;
+    int lda;
+    CHAMELEON_Complex64_t *T;
+    int ldt;
+    CHAMELEON_Complex64_t *TAU, *WORK;
+
+    A   = (CHAMELEON_Complex64_t *)STARPU_MATRIX_GET_PTR(descr[0]);
+    T   = (CHAMELEON_Complex64_t *)STARPU_MATRIX_GET_PTR(descr[1]);
+    TAU = (CHAMELEON_Complex64_t *)STARPU_MATRIX_GET_PTR(descr[2]); /* max(m,n) + n * ib */
+
+    starpu_codelet_unpack_args(cl_arg, &m, &n, &ib, &lda, &ldt, &h_work);
+
+    WORK = TAU + chameleon_max( m, n );
+
+    CORE_zlaset( ChamUpperLower, ib, n, 0., 0., T, ldt );
+    CORE_zgeqrt(m, n, ib, A, lda, T, ldt, TAU, WORK);