diff --git a/.gitlab-ci.yml b/.gitlab-ci.yml
index a6fa958b65b8d4cf5028bff528f4f7291863a139..7b4b35d4c37fbe91717448d7b397df9b75295e7d 100644
--- a/.gitlab-ci.yml
+++ b/.gitlab-ci.yml
@@ -45,7 +45,7 @@ test_starpu_cuda_mpi:
- chameleon_starpu_mpi.lcov
script:
- (cd build &&
- ctest --no-compress-output -V -j 5
+ ctest --no-compress-output -V
-R "test_shm_s|mpi_s"
-D ExperimentalTest
-D ExperimentalCoverage
@@ -90,7 +90,7 @@ test_starpu_simgrid:
script:
- source .gitlab-ci-env.sh simu
- (cd build &&
- ctest --no-compress-output -V -j 5
+ ctest --no-compress-output -V
-D ExperimentalTest
-D ExperimentalCoverage
-D ExperimentalSubmit
@@ -130,7 +130,7 @@ test_quark:
- chameleon_quark.lcov
script:
- (cd build &&
- ctest --no-compress-output -V -j 5
+ ctest --no-compress-output -V
-R "test_shm_s"
-D ExperimentalTest
-D ExperimentalCoverage
diff --git a/compute/CMakeLists.txt b/compute/CMakeLists.txt
index 0034ceb13b1da10d440a280f07fd4b27bf55e401..45700455d95c5a27fd8aa2cbe3ce7f07ab1245a9 100644
--- a/compute/CMakeLists.txt
+++ b/compute/CMakeLists.txt
@@ -235,7 +235,6 @@ set(ZSRC
##################
#pzshift.c
#pzpack.c
- pztile.c
ztile.c
##################
# BUILD
diff --git a/compute/pzhetrd_he2hb.c b/compute/pzhetrd_he2hb.c
index 74c3396149dfc3a4e1d34e129ca2676dcb31f7f8..98301f2b214dc8ddfda207f630f4c324c8137ef0 100644
--- a/compute/pzhetrd_he2hb.c
+++ b/compute/pzhetrd_he2hb.c
@@ -99,6 +99,7 @@ void morse_pzhetrd_he2hb(MORSE_enum uplo,
MorseComplexDouble, A->mb, A->nb, (A->mb*A->nb),
chameleon_min(A->mt, A->nt) * A->mb, A->nb, 0, 0, chameleon_min(A->mt, A->nt) * A->mb, A->nb, 1, 1);
morse_desc_mat_alloc( AT );
+ RUNTIME_desc_create( AT );
/* Let's extract the diagonal in a temporary copy that contains A and A' */
for (k = 1; k < A->nt; k++){
diff --git a/compute/pzlange.c b/compute/pzlange.c
index f893a4d9b45d1d0f04bbc06397b7c469b69fe880..0394d6d335b3faf9247e1a80368ce308ba5758e2 100644
--- a/compute/pzlange.c
+++ b/compute/pzlange.c
@@ -154,6 +154,9 @@ void morse_pzlange( MORSE_enum norm, MORSE_desc_t *A, double *result,
VECNORMS_STEP1(m, n), 1 );
}
}
+ MORSE_Desc_Flush( VECNORMS_STEP2, sequence );
+ MORSE_Desc_Flush( VECNORMS_STEP1, sequence );
+ MORSE_Desc_Flush( RESULT, sequence );
RUNTIME_sequence_wait(morse, sequence);
MORSE_Desc_Destroy( &(VECNORMS_STEP2) );
break;
@@ -271,6 +274,9 @@ void morse_pzlange( MORSE_enum norm, MORSE_desc_t *A, double *result,
VECNORMS_STEP1(m, n), 1 );
}
}
+ MORSE_Desc_Flush( VECNORMS_STEP2, sequence );
+ MORSE_Desc_Flush( VECNORMS_STEP1, sequence );
+ MORSE_Desc_Flush( RESULT, sequence );
RUNTIME_sequence_wait(morse, sequence);
MORSE_Desc_Destroy( &(VECNORMS_STEP2) );
break;
@@ -339,6 +345,8 @@ void morse_pzlange( MORSE_enum norm, MORSE_desc_t *A, double *result,
}
}
+ MORSE_Desc_Flush( VECNORMS_STEP1, sequence );
+ MORSE_Desc_Flush( RESULT, sequence );
RUNTIME_sequence_wait(morse, sequence);
break;
@@ -400,6 +408,8 @@ void morse_pzlange( MORSE_enum norm, MORSE_desc_t *A, double *result,
}
}
+ MORSE_Desc_Flush( VECNORMS_STEP1, sequence );
+ MORSE_Desc_Flush( RESULT, sequence );
RUNTIME_sequence_wait(morse, sequence);
}
diff --git a/compute/pzlanhe.c b/compute/pzlanhe.c
index af3b5bbab0bdff3c67545b22427c59730915d31e..ef78f795fc9f33310a95affcaa84e8ee4ffb59cd 100644
--- a/compute/pzlanhe.c
+++ b/compute/pzlanhe.c
@@ -202,6 +202,9 @@ void morse_pzlanhe(MORSE_enum norm, MORSE_enum uplo, MORSE_desc_t *A, double *re
VECNORMS_STEP1(m, n), 1 );
}
}
+ MORSE_Desc_Flush( VECNORMS_STEP2, sequence );
+ MORSE_Desc_Flush( VECNORMS_STEP1, sequence );
+ MORSE_Desc_Flush( RESULT, sequence );
RUNTIME_sequence_wait(morse, sequence);
*result = *(double *)VECNORMS_STEP1->get_blkaddr(VECNORMS_STEP1, A->myrank / A->q, A->myrank % A->q );
MORSE_Desc_Destroy( &(VECNORMS_STEP1) );
@@ -338,6 +341,8 @@ void morse_pzlanhe(MORSE_enum norm, MORSE_enum uplo, MORSE_desc_t *A, double *re
}
}
+ MORSE_Desc_Flush( VECNORMS_STEP1, sequence );
+ MORSE_Desc_Flush( RESULT, sequence );
RUNTIME_sequence_wait(morse, sequence);
*result = *(double *)VECNORMS_STEP1->get_blkaddr(VECNORMS_STEP1, A->myrank / A->q, A->myrank % A->q );
MORSE_Desc_Destroy( &(VECNORMS_STEP1) );
@@ -449,6 +454,8 @@ void morse_pzlanhe(MORSE_enum norm, MORSE_enum uplo, MORSE_desc_t *A, double *re
}
}
+ MORSE_Desc_Flush( VECNORMS_STEP1, sequence );
+ MORSE_Desc_Flush( RESULT, sequence );
RUNTIME_sequence_wait(morse, sequence);
*result = *(double *)VECNORMS_STEP1->get_blkaddr(VECNORMS_STEP1, A->myrank / A->q, A->myrank % A->q );
MORSE_Desc_Destroy( &(VECNORMS_STEP1) );
diff --git a/compute/pzlansy.c b/compute/pzlansy.c
index f71eb5cb0f410729e187318be9bad6441ca876b6..f89d4302f24cb2caa16421a705f48e4997bdc729 100644
--- a/compute/pzlansy.c
+++ b/compute/pzlansy.c
@@ -202,11 +202,11 @@ void morse_pzlansy(MORSE_enum norm, MORSE_enum uplo, MORSE_desc_t *A, double *re
VECNORMS_STEP1(m, n), 1 );
}
}
+ MORSE_Desc_Flush( VECNORMS_STEP2, sequence );
+ MORSE_Desc_Flush( VECNORMS_STEP1, sequence );
+ MORSE_Desc_Flush( RESULT, sequence );
RUNTIME_sequence_wait(morse, sequence);
- *result = *(double *)VECNORMS_STEP1->get_blkaddr(VECNORMS_STEP1, A->myrank / A->q, A->myrank % A->q );
- MORSE_Desc_Destroy( &(VECNORMS_STEP1) );
MORSE_Desc_Destroy( &(VECNORMS_STEP2) );
- MORSE_Desc_Destroy( &(RESULT) );
break;
/*
* MorseFrobeniusNorm
@@ -347,10 +347,9 @@ void morse_pzlansy(MORSE_enum norm, MORSE_enum uplo, MORSE_desc_t *A, double *re
}
}
+ MORSE_Desc_Flush( VECNORMS_STEP1, sequence );
+ MORSE_Desc_Flush( RESULT, sequence );
RUNTIME_sequence_wait(morse, sequence);
- *result = *(double *)VECNORMS_STEP1->get_blkaddr(VECNORMS_STEP1, A->myrank / A->q, A->myrank % A->q );
- MORSE_Desc_Destroy( &(VECNORMS_STEP1) );
- MORSE_Desc_Destroy( &(RESULT) );
break;
/*
@@ -458,11 +457,15 @@ void morse_pzlansy(MORSE_enum norm, MORSE_enum uplo, MORSE_desc_t *A, double *re
}
}
+ MORSE_Desc_Flush( VECNORMS_STEP1, sequence );
+ MORSE_Desc_Flush( RESULT, sequence );
RUNTIME_sequence_wait(morse, sequence);
- *result = *(double *)VECNORMS_STEP1->get_blkaddr(VECNORMS_STEP1, A->myrank / A->q, A->myrank % A->q );
- MORSE_Desc_Destroy( &(VECNORMS_STEP1) );
- MORSE_Desc_Destroy( &(RESULT) );
}
+
+ *result = *(double *)VECNORMS_STEP1->get_blkaddr(VECNORMS_STEP1, A->myrank / A->q, A->myrank % A->q );
+
+ MORSE_Desc_Destroy( &(VECNORMS_STEP1) );
+ MORSE_Desc_Destroy( &(RESULT) );
RUNTIME_options_ws_free(&options);
RUNTIME_options_finalize(&options, morse);
}
diff --git a/compute/pzlantr.c b/compute/pzlantr.c
index 7c1e28b71c2c739c5cd13229f35151d544a2c10b..5f6606615b9dfde5c1968432f5c4fe919f80a4b8 100644
--- a/compute/pzlantr.c
+++ b/compute/pzlantr.c
@@ -225,6 +225,9 @@ void morse_pzlantr(MORSE_enum norm, MORSE_enum uplo, MORSE_enum diag,
VECNORMS_STEP1(m, n), 1 );
}
}
+ MORSE_Desc_Flush( VECNORMS_STEP2, sequence );
+ MORSE_Desc_Flush( VECNORMS_STEP1, sequence );
+ MORSE_Desc_Flush( RESULT, sequence );
RUNTIME_sequence_wait(morse, sequence);
*result = *(double *)VECNORMS_STEP1->get_blkaddr(VECNORMS_STEP1, A->myrank / A->q, A->myrank % A->q );
MORSE_Desc_Destroy( &(VECNORMS_STEP1) );
@@ -394,6 +397,9 @@ void morse_pzlantr(MORSE_enum norm, MORSE_enum uplo, MORSE_enum diag,
VECNORMS_STEP1(m, n), 1 );
}
}
+ MORSE_Desc_Flush( VECNORMS_STEP2, sequence );
+ MORSE_Desc_Flush( VECNORMS_STEP1, sequence );
+ MORSE_Desc_Flush( RESULT, sequence );
RUNTIME_sequence_wait(morse, sequence);
*result = *(double *)VECNORMS_STEP1->get_blkaddr(VECNORMS_STEP1, A->myrank / A->q, A->myrank % A->q );
MORSE_Desc_Destroy( &(VECNORMS_STEP1) );
@@ -534,6 +540,8 @@ void morse_pzlantr(MORSE_enum norm, MORSE_enum uplo, MORSE_enum diag,
}
}
+ MORSE_Desc_Flush( VECNORMS_STEP1, sequence );
+ MORSE_Desc_Flush( RESULT, sequence );
RUNTIME_sequence_wait(morse, sequence);
*result = *(double *)VECNORMS_STEP1->get_blkaddr(VECNORMS_STEP1, A->myrank / A->q, A->myrank % A->q );
MORSE_Desc_Destroy( &(VECNORMS_STEP1) );
@@ -658,6 +666,8 @@ void morse_pzlantr(MORSE_enum norm, MORSE_enum uplo, MORSE_enum diag,
}
}
+ MORSE_Desc_Flush( VECNORMS_STEP1, sequence );
+ MORSE_Desc_Flush( RESULT, sequence );
RUNTIME_sequence_wait(morse, sequence);
*result = *(double *)VECNORMS_STEP1->get_blkaddr(VECNORMS_STEP1, A->myrank / A->q, A->myrank % A->q );
MORSE_Desc_Destroy( &(VECNORMS_STEP1) );
diff --git a/compute/pztile.c b/compute/pztile.c
deleted file mode 100644
index 6b9b49a8233d3ebfbc8a0c967aa13b7bd0c60da8..0000000000000000000000000000000000000000
--- a/compute/pztile.c
+++ /dev/null
@@ -1,161 +0,0 @@
-/**
- *
- * @copyright (c) 2009-2014 The University of Tennessee and The University
- * of Tennessee Research Foundation.
- * All rights reserved.
- * @copyright (c) 2012-2014 Inria. All rights reserved.
- * @copyright (c) 2012-2014 Bordeaux INP, CNRS (LaBRI UMR 5800), Inria, Univ. Bordeaux. All rights reserved.
- *
- **/
-
-/**
- *
- * @file pztile.c
- *
- * MORSE auxiliary routines
- * MORSE is a software package provided by Univ. of Tennessee,
- * Univ. of California Berkeley and Univ. of Colorado Denver
- *
- * @version 0.9.0
- * @author Jakub Kurzak
- * @author Mathieu Faverge
- * @author Cedric Castagnede
- * @date 2010-11-15
- * @precisions normal z -> c d s
- *
- **/
-#include "control/common.h"
-
-#define A(m, n) A, m, n
-#define B(m, n) &B, m, n
-
-/*******************************************************************************
- * Conversion from LAPACK F77 matrix layout to tile layout - dynamic scheduling
- **/
-void morse_pzlapack_to_tile(MORSE_Complex64_t *Af77, int ldaf77, MORSE_desc_t *A,
- MORSE_sequence_t *sequence, MORSE_request_t *request)
-{
- MORSE_context_t *morse;
- MORSE_option_t options;
- MORSE_desc_t B;
- int m, n;
- int ldam;
- int tempmm, tempnn;
-
- morse = morse_context_self();
- if (sequence->status != MORSE_SUCCESS)
- return;
- RUNTIME_options_init(&options, morse, sequence, request);
-
- B = morse_desc_init(
- MorseComplexDouble, A->mb, A->nb, A->bsiz,
- ldaf77, A->n, 0, 0, A->m, A->n, 1, 1);
-
- B.get_blkaddr = morse_getaddr_cm;
- B.get_blkldd = morse_getblkldd_cm;
- B.mat = Af77;
- B.styp = MorseCM;
-
- RUNTIME_desc_create( &B );
-
- for (m = 0; m < A->mt; m++) {
- tempmm = m == A->mt-1 ? A->m-m*A->mb : A->mb;
- ldam = BLKLDD(A, m);
- for (n = 0; n < A->nt; n++) {
- tempnn = n == A->nt-1 ? A->n-n*A->nb : A->nb;
- MORSE_TASK_zlacpy(
- &options,
- MorseUpperLower,
- tempmm, tempnn, A->mb,
- B(m, n), ldaf77,
- A(m, n), ldam);
- }
- }
-
- RUNTIME_desc_flush( &B, sequence );
- RUNTIME_sequence_wait( morse, sequence );
- RUNTIME_options_finalize( &options, morse );
- RUNTIME_desc_destroy( &B );
-}
-
-/*******************************************************************************
- * Conversion from LAPACK F77 matrix layout to tile layout - dynamic scheduling
- **/
-void morse_pztile_to_lapack(MORSE_desc_t *A, MORSE_Complex64_t *Af77, int ldaf77,
- MORSE_sequence_t *sequence, MORSE_request_t *request)
-{
- MORSE_context_t *morse;
- MORSE_option_t options;
- MORSE_desc_t B;
- int m, n;
- int ldam;
- int tempmm, tempnn;
-
- morse = morse_context_self();
- if (sequence->status != MORSE_SUCCESS)
- return;
- RUNTIME_options_init(&options, morse, sequence, request);
-
- B = morse_desc_init(
- MorseComplexDouble, A->mb, A->nb, A->bsiz,
- ldaf77, A->n, 0, 0, A->m, A->n, 1, 1);
-
- B.get_blkaddr = morse_getaddr_cm;
- B.get_blkldd = morse_getblkldd_cm;
- B.mat = Af77;
- B.styp = MorseCM;
-
- RUNTIME_desc_create( &B );
-
- for (m = 0; m < A->mt; m++) {
- tempmm = m == A->mt-1 ? A->m-m*A->mb : A->mb;
- ldam = BLKLDD(A, m);
- for (n = 0; n < A->nt; n++) {
- tempnn = n == A->nt-1 ? A->n-n*A->nb : A->nb;
- MORSE_TASK_zlacpy(
- &options,
- MorseUpperLower,
- tempmm, tempnn, A->mb,
- A(m, n), ldam,
- B(m, n), ldaf77);
- }
- }
-
- RUNTIME_desc_flush( &B, sequence );
- RUNTIME_sequence_wait( morse, sequence );
- RUNTIME_options_finalize( &options, morse );
- RUNTIME_desc_destroy( &B );
-}
-
-
-/*******************************************************************************
- * Zeroes a submatrix in tile layout - dynamic scheduling
- **/
-void morse_pztile_zero(MORSE_desc_t *A, MORSE_sequence_t *sequence, MORSE_request_t *request)
-{
- MORSE_context_t *morse;
- MORSE_option_t options;
- int m, n;
- int ldam;
- int tempmm, tempnn;
-
- morse = morse_context_self();
- if (sequence->status != MORSE_SUCCESS)
- return;
- RUNTIME_options_init(&options, morse, sequence, request);
-
- for (m = 0; m < A->mt; m++) {
- tempmm = m == A->mt-1 ? A->m-m*A->mb : A->mb;
- ldam = BLKLDD(A, m);
- for (n = 0; n < A->nt; n++) {
- tempnn = n == A->nt-1 ? A->n-n*A->nb : A->nb;
- MORSE_TASK_ztile_zero(
- &options,
- 0, tempnn, 0, tempmm,
- A(m, n), ldam);
- }
- }
-
- RUNTIME_sequence_wait( morse, sequence );
- RUNTIME_options_finalize( &options, morse );
-}
diff --git a/compute/zbuild.c b/compute/zbuild.c
index 71fadaac9434b5d8357a9bddac6360fb0cfefe02..7fe753ef6ae56796ab8041ed1fdf7181727c132c 100644
--- a/compute/zbuild.c
+++ b/compute/zbuild.c
@@ -84,14 +84,14 @@
******************************************************************************/
int MORSE_zbuild( MORSE_enum uplo, int M, int N,
MORSE_Complex64_t *A, int LDA,
- void *user_data, void* user_build_callback)
+ void *user_data, void* user_build_callback )
{
int NB;
int status;
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
MORSE_request_t request = MORSE_REQUEST_INITIALIZER;
- MORSE_desc_t descA;
+ MORSE_desc_t descAl, descAt;
morse = morse_context_self();
if (morse == NULL) {
@@ -124,18 +124,26 @@ int MORSE_zbuild( MORSE_enum uplo, int M, int N,
/* Set NT */
NB = MORSE_NB;
- morse_sequence_create(morse, &sequence);
- morse_zdesc_alloc(descA, NB, NB, LDA, N, 0, 0, M, N, morse_desc_mat_free(&descA));
+ morse_sequence_create( morse, &sequence );
+
+ /* Submit the matrix conversion */
+ morse_zlap2tile( morse, &descAl, &descAt, MorseDescOutput, uplo,
+ A, NB, NB, LDA, N, M, N, sequence, &request );
/* Call the tile interface */
- MORSE_zbuild_Tile_Async(uplo, &descA, user_data, user_build_callback, sequence, &request );
+ MORSE_zbuild_Tile_Async( uplo, &descAt, user_data, user_build_callback, sequence, &request );
+
+ /* Submit the matrix conversion back */
+ morse_ztile2lap( morse, &descAl, &descAt,
+ MorseDescOutput, uplo, sequence, &request );
+
+ morse_sequence_wait( morse, sequence );
- morse_zooptile2lap(descA, A, NB, NB, LDA, N, sequence, &request);
- morse_sequence_wait(morse, sequence);
- morse_desc_mat_free(&descA);
+ /* Cleanup the temporary data */
+ morse_ztile2lap_cleanup( morse, &descAl, &descAt );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -195,11 +203,15 @@ int MORSE_zbuild_Tile( MORSE_enum uplo, MORSE_desc_t *A,
morse_fatal_error("MORSE_zbuild_Tile", "MORSE not initialized");
return MORSE_ERR_NOT_INITIALIZED;
}
- morse_sequence_create(morse, &sequence);
+ morse_sequence_create( morse, &sequence );
+
MORSE_zbuild_Tile_Async( uplo, A, user_data, user_build_callback, sequence, &request );
- morse_sequence_wait(morse, sequence);
+
+ MORSE_Desc_Flush( A, sequence );
+
+ morse_sequence_wait( morse, sequence );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -249,7 +261,7 @@ int MORSE_zbuild_Tile( MORSE_enum uplo, MORSE_desc_t *A,
int MORSE_zbuild_Tile_Async( MORSE_enum uplo, MORSE_desc_t *A,
void *user_data, void* user_build_callback,
MORSE_sequence_t *sequence,
- MORSE_request_t *request)
+ MORSE_request_t *request )
{
MORSE_context_t *morse;
@@ -267,10 +279,12 @@ int MORSE_zbuild_Tile_Async( MORSE_enum uplo, MORSE_desc_t *A,
return MORSE_ERR_UNALLOCATED;
}
/* Check sequence status */
- if (sequence->status == MORSE_SUCCESS)
+ if (sequence->status == MORSE_SUCCESS) {
request->status = MORSE_SUCCESS;
- else
+ }
+ else {
return morse_request_fail(sequence, request, MORSE_ERR_SEQUENCE_FLUSHED);
+ }
/* Check descriptors for correctness */
if (morse_desc_check(A) != MORSE_SUCCESS) {
@@ -279,10 +293,11 @@ int MORSE_zbuild_Tile_Async( MORSE_enum uplo, MORSE_desc_t *A,
}
/* Quick return */
- if (chameleon_min( A->m, A->n ) == 0)
+ if (chameleon_min( A->m, A->n ) == 0) {
return MORSE_SUCCESS;
+ }
- morse_pzbuild(uplo, A, user_data, user_build_callback, sequence, request);
+ morse_pzbuild( uplo, A, user_data, user_build_callback, sequence, request );
return MORSE_SUCCESS;
}
diff --git a/compute/zgeadd.c b/compute/zgeadd.c
index b6e85b32297b0b56bbd9b3e0d9aaa6a46230e657..5478a00aab4624a466d1df0536ed61a8fea76604 100644
--- a/compute/zgeadd.c
+++ b/compute/zgeadd.c
@@ -91,14 +91,15 @@
* @sa MORSE_sgeadd
*
******************************************************************************/
-int MORSE_zgeadd(MORSE_enum trans, int M, int N,
- MORSE_Complex64_t alpha, MORSE_Complex64_t *A, int LDA,
- MORSE_Complex64_t beta, MORSE_Complex64_t *B, int LDB)
+int MORSE_zgeadd( MORSE_enum trans, int M, int N,
+ MORSE_Complex64_t alpha, MORSE_Complex64_t *A, int LDA,
+ MORSE_Complex64_t beta, MORSE_Complex64_t *B, int LDB )
{
int NB;
int Am, An;
int status;
- MORSE_desc_t descA, descB;
+ MORSE_desc_t descAl, descAt;
+ MORSE_desc_t descBl, descBt;
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
MORSE_request_t request = MORSE_REQUEST_INITIALIZER;
@@ -151,37 +152,31 @@ int MORSE_zgeadd(MORSE_enum trans, int M, int N,
/* Set MT & NT & KT */
NB = MORSE_NB;
- morse_sequence_create(morse, &sequence);
+ morse_sequence_create( morse, &sequence );
- /* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) { */
- morse_zooplap2tile( descA, A, NB, NB, LDA, An, 0, 0, Am, An, sequence, &request,
- morse_desc_mat_free(&(descA)) );
- morse_zooplap2tile( descB, B, NB, NB, LDB, N, 0, 0, M, N, sequence, &request,
- morse_desc_mat_free(&(descA)); morse_desc_mat_free(&(descB)));
- /* } else { */
- /* morse_ziplap2tile( descA, A, NB, NB, LDA, An, 0, 0, Am, An, */
- /* sequence, &request); */
- /* morse_ziplap2tile( descB, B, NB, NB, LDB, N, 0, 0, M, N, */
- /* sequence, &request); */
- /* } */
+ /* Submit the matrix conversion */
+ morse_zlap2tile( morse, &descAl, &descAt, MorseDescInput, MorseUpperLower,
+ A, NB, NB, LDA, An, Am, An, sequence, &request );
+ morse_zlap2tile( morse, &descBl, &descBt, MorseDescInout, MorseUpperLower,
+ B, NB, NB, LDB, N, M, N, sequence, &request );
/* Call the tile interface */
- MORSE_zgeadd_Tile_Async(
- trans, alpha, &descA, beta, &descB, sequence, &request);
+ MORSE_zgeadd_Tile_Async( trans, alpha, &descAt, beta, &descBt, sequence, &request );
- /* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) { */
- morse_zooptile2lap( descB, B, NB, NB, LDB, N, sequence, &request);
- morse_sequence_wait(morse, sequence);
- morse_desc_mat_free(&descA);
- morse_desc_mat_free(&descB);
- /* } else { */
- /* morse_ziptile2lap( descA, A, NB, NB, LDA, An, sequence, &request); */
- /* morse_ziptile2lap( descB, B, NB, NB, LDB, N, sequence, &request); */
- /* morse_dynamic_sync(); */
- /* } */
+ /* Submit the matrix conversion back */
+ morse_ztile2lap( morse, &descAl, &descAt,
+ MorseDescInput, MorseUpperLower, sequence, &request );
+ morse_ztile2lap( morse, &descBl, &descBt,
+ MorseDescInout, MorseUpperLower, sequence, &request );
+
+ morse_sequence_wait( morse, sequence );
+
+ /* Cleanup the temporary data */
+ morse_ztile2lap_cleanup( morse, &descAl, &descAt );
+ morse_ztile2lap_cleanup( morse, &descBl, &descBt );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -233,9 +228,9 @@ int MORSE_zgeadd(MORSE_enum trans, int M, int N,
* @sa MORSE_sgeadd_Tile
*
******************************************************************************/
-int MORSE_zgeadd_Tile(MORSE_enum trans,
- MORSE_Complex64_t alpha, MORSE_desc_t *A,
- MORSE_Complex64_t beta, MORSE_desc_t *B)
+int MORSE_zgeadd_Tile( MORSE_enum trans,
+ MORSE_Complex64_t alpha, MORSE_desc_t *A,
+ MORSE_Complex64_t beta, MORSE_desc_t *B )
{
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
@@ -247,14 +242,16 @@ int MORSE_zgeadd_Tile(MORSE_enum trans,
morse_fatal_error("MORSE_zgeadd_Tile", "MORSE not initialized");
return MORSE_ERR_NOT_INITIALIZED;
}
- morse_sequence_create(morse, &sequence);
- MORSE_zgeadd_Tile_Async(trans, alpha, A, beta, B, sequence, &request);
- RUNTIME_desc_flush( A, sequence );
- RUNTIME_desc_flush( B, sequence );
- morse_sequence_wait(morse, sequence);
+ morse_sequence_create( morse, &sequence );
+
+ MORSE_zgeadd_Tile_Async( trans, alpha, A, beta, B, sequence, &request );
+ MORSE_Desc_Flush( A, sequence );
+ MORSE_Desc_Flush( B, sequence );
+
+ morse_sequence_wait( morse, sequence );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -287,10 +284,10 @@ int MORSE_zgeadd_Tile(MORSE_enum trans,
* @sa MORSE_sgeadd_Tile_Async
*
******************************************************************************/
-int MORSE_zgeadd_Tile_Async(MORSE_enum trans,
- MORSE_Complex64_t alpha, MORSE_desc_t *A,
- MORSE_Complex64_t beta, MORSE_desc_t *B,
- MORSE_sequence_t *sequence, MORSE_request_t *request)
+int MORSE_zgeadd_Tile_Async( MORSE_enum trans,
+ MORSE_Complex64_t alpha, MORSE_desc_t *A,
+ MORSE_Complex64_t beta, MORSE_desc_t *B,
+ MORSE_sequence_t *sequence, MORSE_request_t *request )
{
MORSE_context_t *morse;
int M, N;
@@ -310,10 +307,12 @@ int MORSE_zgeadd_Tile_Async(MORSE_enum trans,
return MORSE_ERR_UNALLOCATED;
}
/* Check sequence status */
- if (sequence->status == MORSE_SUCCESS)
+ if (sequence->status == MORSE_SUCCESS) {
request->status = MORSE_SUCCESS;
- else
+ }
+ else {
return morse_request_fail(sequence, request, MORSE_ERR_SEQUENCE_FLUSHED);
+ }
/* Check descriptors for correctness */
if (morse_desc_check(A) != MORSE_SUCCESS) {
@@ -363,11 +362,13 @@ int MORSE_zgeadd_Tile_Async(MORSE_enum trans,
N = B->n;
/* Quick return */
- if (M == 0 || N == 0 ||
- ((alpha == (MORSE_Complex64_t)0.0) && beta == (MORSE_Complex64_t)1.0))
+ if ( (M == 0) || (N == 0) ||
+ ((alpha == (MORSE_Complex64_t)0.0) && (beta == (MORSE_Complex64_t)1.0)) )
+ {
return MORSE_SUCCESS;
+ }
- morse_pztradd(MorseUpperLower, trans, alpha, A, beta, B, sequence, request);
+ morse_pztradd( MorseUpperLower, trans, alpha, A, beta, B, sequence, request );
return MORSE_SUCCESS;
}
diff --git a/compute/zgelqf.c b/compute/zgelqf.c
index 55b9284837f7af7fd60f17702fcba108942c4d1b..bcab0855fa87c58f57ba961245b9d0837e5c0559 100644
--- a/compute/zgelqf.c
+++ b/compute/zgelqf.c
@@ -75,16 +75,16 @@
* @sa MORSE_zgelqs
*
******************************************************************************/
-int MORSE_zgelqf(int M, int N,
+int MORSE_zgelqf( int M, int N,
MORSE_Complex64_t *A, int LDA,
- MORSE_desc_t *descT)
+ MORSE_desc_t *descT )
{
int NB;
int status;
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
MORSE_request_t request = MORSE_REQUEST_INITIALIZER;
- MORSE_desc_t descA;
+ MORSE_desc_t descAl, descAt;
morse = morse_context_self();
if (morse == NULL) {
@@ -120,30 +120,26 @@ int MORSE_zgelqf(int M, int N,
/* Set NT */
NB = MORSE_NB;
- morse_sequence_create(morse, &sequence);
+ morse_sequence_create( morse, &sequence );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooplap2tile( descA, A, NB, NB, LDA, N, 0, 0, M, N, sequence, &request,
- morse_desc_mat_free(&(descA)) );
-/* } else {*/
-/* morse_ziplap2tile( descA, A, NB, NB, LDA, N, 0, 0, M, N,*/
-/* sequence, &request);*/
-/* }*/
+ /* Submit the matrix conversion */
+ morse_zlap2tile( morse, &descAl, &descAt, MorseDescInout, MorseUpperLower,
+ A, NB, NB, LDA, N, M, N, sequence, &request );
/* Call the tile interface */
- MORSE_zgelqf_Tile_Async(&descA, descT, sequence, &request);
+ MORSE_zgelqf_Tile_Async( &descAt, descT, sequence, &request );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooptile2lap(descA, A, NB, NB, LDA, N, sequence, &request);
- morse_sequence_wait(morse, sequence);
- morse_desc_mat_free(&descA);
-/* } else {*/
-/* morse_ziptile2lap( descA, A, NB, NB, LDA, N, sequence, &request);*/
-/* morse_sequence_wait(morse, sequence);*/
-/* }*/
+ /* Submit the matrix conversion back */
+ morse_ztile2lap( morse, &descAl, &descAt,
+ MorseDescInout, MorseUpperLower, sequence, &request );
+
+ morse_sequence_wait( morse, sequence );
+
+ /* Cleanup the temporary data */
+ morse_ztile2lap_cleanup( morse, &descAl, &descAt );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -186,7 +182,7 @@ int MORSE_zgelqf(int M, int N,
* @sa MORSE_zgelqs_Tile
*
******************************************************************************/
-int MORSE_zgelqf_Tile(MORSE_desc_t *A, MORSE_desc_t *T)
+int MORSE_zgelqf_Tile( MORSE_desc_t *A, MORSE_desc_t *T )
{
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
@@ -198,13 +194,16 @@ int MORSE_zgelqf_Tile(MORSE_desc_t *A, MORSE_desc_t *T)
morse_fatal_error("MORSE_zgelqf_Tile", "MORSE not initialized");
return MORSE_ERR_NOT_INITIALIZED;
}
- morse_sequence_create(morse, &sequence);
- MORSE_zgelqf_Tile_Async(A, T, sequence, &request);
- RUNTIME_desc_flush( A, sequence );
- morse_sequence_wait(morse, sequence);
+ morse_sequence_create( morse, &sequence );
+
+ MORSE_zgelqf_Tile_Async( A, T, sequence, &request );
+ MORSE_Desc_Flush( A, sequence );
+ MORSE_Desc_Flush( T, sequence );
+
+ morse_sequence_wait( morse, sequence );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -237,8 +236,8 @@ int MORSE_zgelqf_Tile(MORSE_desc_t *A, MORSE_desc_t *T)
* @sa MORSE_zgelqs_Tile_Async
*
******************************************************************************/
-int MORSE_zgelqf_Tile_Async(MORSE_desc_t *A, MORSE_desc_t *T,
- MORSE_sequence_t *sequence, MORSE_request_t *request)
+int MORSE_zgelqf_Tile_Async( MORSE_desc_t *A, MORSE_desc_t *T,
+ MORSE_sequence_t *sequence, MORSE_request_t *request )
{
MORSE_context_t *morse;
MORSE_desc_t D, *Dptr = NULL;
@@ -257,10 +256,12 @@ int MORSE_zgelqf_Tile_Async(MORSE_desc_t *A, MORSE_desc_t *T,
return MORSE_ERR_UNALLOCATED;
}
/* Check sequence status */
- if (sequence->status == MORSE_SUCCESS)
+ if (sequence->status == MORSE_SUCCESS) {
request->status = MORSE_SUCCESS;
- else
+ }
+ else {
return morse_request_fail(sequence, request, MORSE_ERR_SEQUENCE_FLUSHED);
+ }
/* Check descriptors for correctness */
if (morse_desc_check(A) != MORSE_SUCCESS) {
@@ -277,10 +278,10 @@ int MORSE_zgelqf_Tile_Async(MORSE_desc_t *A, MORSE_desc_t *T,
return morse_request_fail(sequence, request, MORSE_ERR_ILLEGAL_VALUE);
}
/* Quick return */
-/*
- if (chameleon_min(M, N) == 0)
- return MORSE_SUCCESS;
-*/
+ /*
+ if (chameleon_min(M, N) == 0)
+ return MORSE_SUCCESS;
+ */
#if defined(CHAMELEON_COPY_DIAG)
{
int m = chameleon_min(A->mt, A->nt) * A->mb;
@@ -290,13 +291,17 @@ int MORSE_zgelqf_Tile_Async(MORSE_desc_t *A, MORSE_desc_t *T,
#endif
if (morse->householder == MORSE_FLAT_HOUSEHOLDER) {
- morse_pzgelqf(A, T, Dptr, sequence, request);
+ morse_pzgelqf( A, T, Dptr, sequence, request );
}
else {
- morse_pzgelqfrh(A, T, Dptr, MORSE_RHBLK, sequence, request);
+ morse_pzgelqfrh( A, T, Dptr, MORSE_RHBLK, sequence, request );
}
if (Dptr != NULL) {
- morse_desc_mat_free(Dptr);
+ MORSE_Desc_Flush( A, sequence );
+ MORSE_Desc_Flush( T, sequence );
+ MORSE_Desc_Flush( Dptr, sequence );
+ morse_sequence_wait( morse, sequence );
+ morse_desc_mat_free( Dptr );
}
(void)D;
return MORSE_SUCCESS;
diff --git a/compute/zgelqf_param.c b/compute/zgelqf_param.c
index d650ba8ad5f3b180d22f0011dd334ce27b3b55b8..fe0623cb701f9831a7bc00ca19bb0fc6754ccd5c 100644
--- a/compute/zgelqf_param.c
+++ b/compute/zgelqf_param.c
@@ -72,16 +72,16 @@
* @sa MORSE_zgelqs
*
******************************************************************************/
-int MORSE_zgelqf_param(const libhqr_tree_t *qrtree, int M, int N,
- MORSE_Complex64_t *A, int LDA,
- MORSE_desc_t *descTS, MORSE_desc_t *descTT)
+int MORSE_zgelqf_param( const libhqr_tree_t *qrtree, int M, int N,
+ MORSE_Complex64_t *A, int LDA,
+ MORSE_desc_t *descTS, MORSE_desc_t *descTT )
{
int NB;
int status;
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
MORSE_request_t request = MORSE_REQUEST_INITIALIZER;
- MORSE_desc_t descA;
+ MORSE_desc_t descAl, descAt;
morse = morse_context_self();
if (morse == NULL) {
@@ -117,30 +117,26 @@ int MORSE_zgelqf_param(const libhqr_tree_t *qrtree, int M, int N,
/* Set NT */
NB = MORSE_NB;
- morse_sequence_create(morse, &sequence);
+ morse_sequence_create( morse, &sequence );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooplap2tile( descA, A, NB, NB, LDA, N, 0, 0, M, N, sequence, &request,
- morse_desc_mat_free(&(descA)) );
-/* } else {*/
-/* morse_ziplap2tile( descA, A, NB, NB, LDA, N, 0, 0, M, N,*/
-/* sequence, &request);*/
-/* }*/
+ /* Submit the matrix conversion */
+ morse_zlap2tile( morse, &descAl, &descAt, MorseDescInout, MorseUpperLower,
+ A, NB, NB, LDA, N, M, N, sequence, &request );
/* Call the tile interface */
- MORSE_zgelqf_param_Tile_Async(qrtree, &descA, descTS, descTT, sequence, &request);
+ MORSE_zgelqf_param_Tile_Async( qrtree, &descAt, descTS, descTT, sequence, &request );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooptile2lap(descA, A, NB, NB, LDA, N, sequence, &request);
- morse_sequence_wait(morse, sequence);
- morse_desc_mat_free(&descA);
-/* } else {*/
-/* morse_ziptile2lap( descA, A, NB, NB, LDA, N, sequence, &request);*/
-/* morse_sequence_wait(morse, sequence);*/
-/* }*/
+ /* Submit the matrix conversion back */
+ morse_ztile2lap( morse, &descAl, &descAt,
+ MorseDescInout, MorseUpperLower, sequence, &request );
+
+ morse_sequence_wait( morse, sequence );
+
+ /* Cleanup the temporary data */
+ morse_ztile2lap_cleanup( morse, &descAl, &descAt );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -183,7 +179,7 @@ int MORSE_zgelqf_param(const libhqr_tree_t *qrtree, int M, int N,
* @sa MORSE_zgelqs_Tile
*
******************************************************************************/
-int MORSE_zgelqf_param_Tile(const libhqr_tree_t *qrtree, MORSE_desc_t *A, MORSE_desc_t *TS, MORSE_desc_t *TT)
+int MORSE_zgelqf_param_Tile( const libhqr_tree_t *qrtree, MORSE_desc_t *A, MORSE_desc_t *TS, MORSE_desc_t *TT )
{
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
@@ -195,13 +191,17 @@ int MORSE_zgelqf_param_Tile(const libhqr_tree_t *qrtree, MORSE_desc_t *A, MORSE_
morse_fatal_error("MORSE_zgelqf_param_Tile", "MORSE not initialized");
return MORSE_ERR_NOT_INITIALIZED;
}
- morse_sequence_create(morse, &sequence);
- MORSE_zgelqf_param_Tile_Async(qrtree, A, TS, TT, sequence, &request);
- RUNTIME_desc_flush( A, sequence );
- morse_sequence_wait(morse, sequence);
+ morse_sequence_create( morse, &sequence );
+
+ MORSE_zgelqf_param_Tile_Async( qrtree, A, TS, TT, sequence, &request );
+
+ MORSE_Desc_Flush( A, sequence );
+ MORSE_Desc_Flush( TS, sequence );
+ MORSE_Desc_Flush( TT, sequence );
+ morse_sequence_wait( morse, sequence );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -234,8 +234,8 @@ int MORSE_zgelqf_param_Tile(const libhqr_tree_t *qrtree, MORSE_desc_t *A, MORSE_
* @sa MORSE_zgelqs_Tile_Async
*
******************************************************************************/
-int MORSE_zgelqf_param_Tile_Async(const libhqr_tree_t *qrtree, MORSE_desc_t *A, MORSE_desc_t *TS, MORSE_desc_t *TT,
- MORSE_sequence_t *sequence, MORSE_request_t *request)
+int MORSE_zgelqf_param_Tile_Async( const libhqr_tree_t *qrtree, MORSE_desc_t *A, MORSE_desc_t *TS, MORSE_desc_t *TT,
+ MORSE_sequence_t *sequence, MORSE_request_t *request )
{
MORSE_context_t *morse;
MORSE_desc_t D, *Dptr = NULL;
@@ -254,10 +254,12 @@ int MORSE_zgelqf_param_Tile_Async(const libhqr_tree_t *qrtree, MORSE_desc_t *A,
return MORSE_ERR_UNALLOCATED;
}
/* Check sequence status */
- if (sequence->status == MORSE_SUCCESS)
+ if (sequence->status == MORSE_SUCCESS) {
request->status = MORSE_SUCCESS;
- else
+ }
+ else {
return morse_request_fail(sequence, request, MORSE_ERR_SEQUENCE_FLUSHED);
+ }
/* Check descriptors for correctness */
if (morse_desc_check(A) != MORSE_SUCCESS) {
@@ -278,10 +280,10 @@ int MORSE_zgelqf_param_Tile_Async(const libhqr_tree_t *qrtree, MORSE_desc_t *A,
return morse_request_fail(sequence, request, MORSE_ERR_ILLEGAL_VALUE);
}
/* Quick return */
-/*
- if (chameleon_min(M, N) == 0)
- return MORSE_SUCCESS;
-*/
+ /*
+ if (chameleon_min(M, N) == 0)
+ return MORSE_SUCCESS;
+ */
#if defined(CHAMELEON_COPY_DIAG)
{
int m = chameleon_min(A->mt, A->nt) * A->mb;
@@ -290,9 +292,14 @@ int MORSE_zgelqf_param_Tile_Async(const libhqr_tree_t *qrtree, MORSE_desc_t *A,
}
#endif
- morse_pzgelqf_param(qrtree, A, TS, TT, Dptr, sequence, request);
+ morse_pzgelqf_param( qrtree, A, TS, TT, Dptr, sequence, request );
if (Dptr != NULL) {
- morse_desc_mat_free(Dptr);
+ MORSE_Desc_Flush( A, sequence );
+ MORSE_Desc_Flush( TS, sequence );
+ MORSE_Desc_Flush( TT, sequence );
+ MORSE_Desc_Flush( Dptr, sequence );
+ morse_sequence_wait( morse, sequence );
+ morse_desc_mat_free( Dptr );
}
(void)D;
return MORSE_SUCCESS;
diff --git a/compute/zgelqs.c b/compute/zgelqs.c
index b9410fd3d107681ccdbeac7bd54ed92537ba0580..ccf896144f1af0d4f7188ac3f27d91a17d2ea8b1 100644
--- a/compute/zgelqs.c
+++ b/compute/zgelqs.c
@@ -81,17 +81,18 @@
* @sa MORSE_zgelqf
*
******************************************************************************/
-int MORSE_zgelqs(int M, int N, int NRHS,
+int MORSE_zgelqs( int M, int N, int NRHS,
MORSE_Complex64_t *A, int LDA,
MORSE_desc_t *descT,
- MORSE_Complex64_t *B, int LDB)
+ MORSE_Complex64_t *B, int LDB )
{
int NB;
int status;
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
MORSE_request_t request = MORSE_REQUEST_INITIALIZER;
- MORSE_desc_t descA, descB;
+ MORSE_desc_t descAl, descAt;
+ MORSE_desc_t descBl, descBt;
morse = morse_context_self();
if (morse == NULL) {
@@ -135,37 +136,31 @@ int MORSE_zgelqs(int M, int N, int NRHS,
/* Set NT */
NB = MORSE_NB;
- morse_sequence_create(morse, &sequence);
+ morse_sequence_create( morse, &sequence );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooplap2tile( descA, A, NB, NB, LDA, N, 0, 0, M, N, sequence, &request,
- morse_desc_mat_free(&(descA)) );
- morse_zooplap2tile( descB, B, NB, NB, LDB, NRHS, 0, 0, N, NRHS, sequence, &request,
- morse_desc_mat_free(&(descA)); morse_desc_mat_free(&(descB)));
-/* } else {*/
-/* morse_ziplap2tile( descA, A, NB, NB, LDA, N, 0, 0, M, N, */
-/* sequence, &request);*/
-/* morse_ziplap2tile( descB, B, NB, NB, LDB, NRHS, 0, 0, N, NRHS,*/
-/* sequence, &request);*/
-/* }*/
+ /* Submit the matrix conversion */
+ morse_zlap2tile( morse, &descAl, &descAt, MorseDescInput, MorseUpperLower,
+ A, NB, NB, LDA, N, M, N, sequence, &request );
+ morse_zlap2tile( morse, &descBl, &descBt, MorseDescInout, MorseUpperLower,
+ B, NB, NB, LDB, NRHS, N, NRHS, sequence, &request );
/* Call the tile interface */
- MORSE_zgelqs_Tile_Async(&descA, descT, &descB, sequence, &request);
+ MORSE_zgelqs_Tile_Async( &descAt, descT, &descBt, sequence, &request );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooptile2lap(descA, A, NB, NB, LDA, N, sequence, &request);
- morse_zooptile2lap(descB, B, NB, NB, LDB, NRHS, sequence, &request);
- morse_sequence_wait(morse, sequence);
- morse_desc_mat_free(&descA);
- morse_desc_mat_free(&descB);
-/* } else {*/
-/* morse_ziptile2lap( descA, A, NB, NB, LDA, N, sequence, &request);*/
-/* morse_ziptile2lap( descB, B, NB, NB, LDB, NRHS, sequence, &request);*/
-/* morse_sequence_wait(morse, sequence);*/
-/* }*/
+ /* Submit the matrix conversion back */
+ morse_ztile2lap( morse, &descAl, &descAt,
+ MorseDescInput, MorseUpperLower, sequence, &request );
+ morse_ztile2lap( morse, &descBl, &descBt,
+ MorseDescInout, MorseUpperLower, sequence, &request );
+
+ morse_sequence_wait( morse, sequence );
+
+ /* Cleanup the temporary data */
+ morse_ztile2lap_cleanup( morse, &descAl, &descAt );
+ morse_ztile2lap_cleanup( morse, &descBl, &descBt );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -208,7 +203,7 @@ int MORSE_zgelqs(int M, int N, int NRHS,
* @sa MORSE_zgelqf_Tile
*
******************************************************************************/
-int MORSE_zgelqs_Tile(MORSE_desc_t *A, MORSE_desc_t *T, MORSE_desc_t *B)
+int MORSE_zgelqs_Tile( MORSE_desc_t *A, MORSE_desc_t *T, MORSE_desc_t *B )
{
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
@@ -220,14 +215,17 @@ int MORSE_zgelqs_Tile(MORSE_desc_t *A, MORSE_desc_t *T, MORSE_desc_t *B)
morse_fatal_error("MORSE_zgelqs_Tile", "MORSE not initialized");
return MORSE_ERR_NOT_INITIALIZED;
}
- morse_sequence_create(morse, &sequence);
- MORSE_zgelqs_Tile_Async(A, T, B, sequence, &request);
- RUNTIME_desc_flush( A, sequence );
- RUNTIME_desc_flush( B, sequence );
- morse_sequence_wait(morse, sequence);
+ morse_sequence_create( morse, &sequence );
+
+ MORSE_zgelqs_Tile_Async( A, T, B, sequence, &request );
+ MORSE_Desc_Flush( A, sequence );
+ MORSE_Desc_Flush( T, sequence );
+ MORSE_Desc_Flush( B, sequence );
+
+ morse_sequence_wait( morse, sequence );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -261,8 +259,8 @@ int MORSE_zgelqs_Tile(MORSE_desc_t *A, MORSE_desc_t *T, MORSE_desc_t *B)
* @sa MORSE_zgelqf_Tile_Async
*
******************************************************************************/
-int MORSE_zgelqs_Tile_Async(MORSE_desc_t *A, MORSE_desc_t *T, MORSE_desc_t *B,
- MORSE_sequence_t *sequence, MORSE_request_t *request)
+int MORSE_zgelqs_Tile_Async( MORSE_desc_t *A, MORSE_desc_t *T, MORSE_desc_t *B,
+ MORSE_sequence_t *sequence, MORSE_request_t *request )
{
MORSE_desc_t *subB;
MORSE_desc_t *subA;
@@ -283,10 +281,12 @@ int MORSE_zgelqs_Tile_Async(MORSE_desc_t *A, MORSE_desc_t *T, MORSE_desc_t *B,
return MORSE_ERR_UNALLOCATED;
}
/* Check sequence status */
- if (sequence->status == MORSE_SUCCESS)
+ if (sequence->status == MORSE_SUCCESS) {
request->status = MORSE_SUCCESS;
- else
+ }
+ else {
return morse_request_fail(sequence, request, MORSE_ERR_SEQUENCE_FLUSHED);
+ }
/* Check descriptors for correctness */
if (morse_desc_check(A) != MORSE_SUCCESS) {
@@ -307,18 +307,18 @@ int MORSE_zgelqs_Tile_Async(MORSE_desc_t *A, MORSE_desc_t *T, MORSE_desc_t *B,
return morse_request_fail(sequence, request, MORSE_ERR_ILLEGAL_VALUE);
}
/* Quick return */
-/*
- if (chameleon_min(M, chameleon_min(N, NRHS)) == 0) {
- return MORSE_SUCCESS;
- }
-*/
+ /*
+ if (chameleon_min(M, chameleon_min(N, NRHS)) == 0) {
+ return MORSE_SUCCESS;
+ }
+ */
/* subB = morse_desc_submatrix(B, A->m, 0, A->n-A->m, B->n);
- morse_pztile_zero(subB, sequence, request);
- free(subB); */
+ morse_pzlaset( MorseUpperLower, 0., 0., subB, sequence, request );
+ free(subB); */
subB = morse_desc_submatrix(B, 0, 0, A->m, B->n);
subA = morse_desc_submatrix(A, 0, 0, A->m, A->m);
- morse_pztrsm(MorseLeft, MorseLower, MorseNoTrans, MorseNonUnit, 1.0, subA, subB, sequence, request);
+ morse_pztrsm( MorseLeft, MorseLower, MorseNoTrans, MorseNonUnit, 1.0, subA, subB, sequence, request );
free(subA);
free(subB);
@@ -331,14 +331,19 @@ int MORSE_zgelqs_Tile_Async(MORSE_desc_t *A, MORSE_desc_t *T, MORSE_desc_t *B,
#endif
if (morse->householder == MORSE_FLAT_HOUSEHOLDER) {
- morse_pzunmlq(MorseLeft, MorseConjTrans, A, B, T, Dptr, sequence, request);
+ morse_pzunmlq( MorseLeft, MorseConjTrans, A, B, T, Dptr, sequence, request );
}
else {
- morse_pzunmlqrh(MorseLeft, MorseConjTrans, A, B, T, Dptr, MORSE_RHBLK, sequence, request);
+ morse_pzunmlqrh( MorseLeft, MorseConjTrans, A, B, T, Dptr, MORSE_RHBLK, sequence, request );
}
if (Dptr != NULL) {
- morse_desc_mat_free(Dptr);
+ MORSE_Desc_Flush( A, sequence );
+ MORSE_Desc_Flush( B, sequence );
+ MORSE_Desc_Flush( T, sequence );
+ MORSE_Desc_Flush( Dptr, sequence );
+ morse_sequence_wait( morse, sequence );
+ morse_desc_mat_free( Dptr );
}
(void)D;
return MORSE_SUCCESS;
diff --git a/compute/zgelqs_param.c b/compute/zgelqs_param.c
index ac5c26520b0d471eabea1a527630959d135dc55a..ffac93d2ef8e3972741a96e56b75d1cf76e24dca 100644
--- a/compute/zgelqs_param.c
+++ b/compute/zgelqs_param.c
@@ -83,17 +83,18 @@
* @sa MORSE_zgelqf
*
******************************************************************************/
-int MORSE_zgelqs_param(const libhqr_tree_t *qrtree, int M, int N, int NRHS,
- MORSE_Complex64_t *A, int LDA,
- MORSE_desc_t *descTS, MORSE_desc_t *descTT,
- MORSE_Complex64_t *B, int LDB)
+int MORSE_zgelqs_param( const libhqr_tree_t *qrtree, int M, int N, int NRHS,
+ MORSE_Complex64_t *A, int LDA,
+ MORSE_desc_t *descTS, MORSE_desc_t *descTT,
+ MORSE_Complex64_t *B, int LDB )
{
int NB;
int status;
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
MORSE_request_t request = MORSE_REQUEST_INITIALIZER;
- MORSE_desc_t descA, descB;
+ MORSE_desc_t descAl, descAt;
+ MORSE_desc_t descBl, descBt;
morse = morse_context_self();
if (morse == NULL) {
@@ -137,37 +138,31 @@ int MORSE_zgelqs_param(const libhqr_tree_t *qrtree, int M, int N, int NRHS,
/* Set NT */
NB = MORSE_NB;
- morse_sequence_create(morse, &sequence);
+ morse_sequence_create( morse, &sequence );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooplap2tile( descA, A, NB, NB, LDA, N, 0, 0, M, N, sequence, &request,
- morse_desc_mat_free(&(descA)) );
- morse_zooplap2tile( descB, B, NB, NB, LDB, NRHS, 0, 0, N, NRHS, sequence, &request,
- morse_desc_mat_free(&(descA)); morse_desc_mat_free(&(descB)));
-/* } else {*/
-/* morse_ziplap2tile( descA, A, NB, NB, LDA, N, 0, 0, M, N, */
-/* sequence, &request);*/
-/* morse_ziplap2tile( descB, B, NB, NB, LDB, NRHS, 0, 0, N, NRHS,*/
-/* sequence, &request);*/
-/* }*/
+ /* Submit the matrix conversion */
+ morse_zlap2tile( morse, &descAl, &descAt, MorseDescInput, MorseUpperLower,
+ A, NB, NB, LDA, N, M, N, sequence, &request );
+ morse_zlap2tile( morse, &descBl, &descBt, MorseDescInout, MorseUpperLower,
+ B, NB, NB, LDB, NRHS, N, NRHS, sequence, &request );
/* Call the tile interface */
- MORSE_zgelqs_param_Tile_Async(qrtree, &descA, descTS, descTT, &descB, sequence, &request);
+ MORSE_zgelqs_param_Tile_Async( qrtree, &descAt, descTS, descTT, &descBt, sequence, &request );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooptile2lap(descA, A, NB, NB, LDA, N, sequence, &request);
- morse_zooptile2lap(descB, B, NB, NB, LDB, NRHS, sequence, &request);
- morse_sequence_wait(morse, sequence);
- morse_desc_mat_free(&descA);
- morse_desc_mat_free(&descB);
-/* } else {*/
-/* morse_ziptile2lap( descA, A, NB, NB, LDA, N, sequence, &request);*/
-/* morse_ziptile2lap( descB, B, NB, NB, LDB, NRHS, sequence, &request);*/
-/* morse_sequence_wait(morse, sequence);*/
-/* }*/
+ /* Submit the matrix conversion back */
+ morse_ztile2lap( morse, &descAl, &descAt,
+ MorseDescInput, MorseUpperLower, sequence, &request );
+ morse_ztile2lap( morse, &descBl, &descBt,
+ MorseDescInout, MorseUpperLower, sequence, &request );
+
+ morse_sequence_wait( morse, sequence );
+
+ /* Cleanup the temporary data */
+ morse_ztile2lap_cleanup( morse, &descAl, &descAt );
+ morse_ztile2lap_cleanup( morse, &descBl, &descBt );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -213,7 +208,8 @@ int MORSE_zgelqs_param(const libhqr_tree_t *qrtree, int M, int N, int NRHS,
* @sa MORSE_zgelqf_Tile
*
******************************************************************************/
-int MORSE_zgelqs_param_Tile(const libhqr_tree_t *qrtree, MORSE_desc_t *A, MORSE_desc_t *TS, MORSE_desc_t *TT, MORSE_desc_t *B)
+int MORSE_zgelqs_param_Tile( const libhqr_tree_t *qrtree, MORSE_desc_t *A,
+ MORSE_desc_t *TS, MORSE_desc_t *TT, MORSE_desc_t *B )
{
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
@@ -225,14 +221,18 @@ int MORSE_zgelqs_param_Tile(const libhqr_tree_t *qrtree, MORSE_desc_t *A, MORSE_
morse_fatal_error("MORSE_zgelqs_param_Tile", "MORSE not initialized");
return MORSE_ERR_NOT_INITIALIZED;
}
- morse_sequence_create(morse, &sequence);
- MORSE_zgelqs_param_Tile_Async(qrtree, A, TS, TT, B, sequence, &request);
- RUNTIME_desc_flush( A, sequence );
- RUNTIME_desc_flush( B, sequence );
- morse_sequence_wait(morse, sequence);
+ morse_sequence_create( morse, &sequence );
+
+ MORSE_zgelqs_param_Tile_Async( qrtree, A, TS, TT, B, sequence, &request );
+ MORSE_Desc_Flush( A, sequence );
+ MORSE_Desc_Flush( TS, sequence );
+ MORSE_Desc_Flush( TT, sequence );
+ MORSE_Desc_Flush( B, sequence );
+
+ morse_sequence_wait( morse, sequence );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -266,8 +266,8 @@ int MORSE_zgelqs_param_Tile(const libhqr_tree_t *qrtree, MORSE_desc_t *A, MORSE_
* @sa MORSE_zgelqf_Tile_Async
*
******************************************************************************/
-int MORSE_zgelqs_param_Tile_Async(const libhqr_tree_t *qrtree, MORSE_desc_t *A, MORSE_desc_t *TS, MORSE_desc_t *TT, MORSE_desc_t *B,
- MORSE_sequence_t *sequence, MORSE_request_t *request)
+int MORSE_zgelqs_param_Tile_Async( const libhqr_tree_t *qrtree, MORSE_desc_t *A, MORSE_desc_t *TS, MORSE_desc_t *TT, MORSE_desc_t *B,
+ MORSE_sequence_t *sequence, MORSE_request_t *request )
{
MORSE_desc_t *subB;
MORSE_desc_t *subA;
@@ -288,10 +288,12 @@ int MORSE_zgelqs_param_Tile_Async(const libhqr_tree_t *qrtree, MORSE_desc_t *A,
return MORSE_ERR_UNALLOCATED;
}
/* Check sequence status */
- if (sequence->status == MORSE_SUCCESS)
+ if (sequence->status == MORSE_SUCCESS) {
request->status = MORSE_SUCCESS;
- else
+ }
+ else {
return morse_request_fail(sequence, request, MORSE_ERR_SEQUENCE_FLUSHED);
+ }
/* Check descriptors for correctness */
if (morse_desc_check(A) != MORSE_SUCCESS) {
@@ -316,18 +318,18 @@ int MORSE_zgelqs_param_Tile_Async(const libhqr_tree_t *qrtree, MORSE_desc_t *A,
return morse_request_fail(sequence, request, MORSE_ERR_ILLEGAL_VALUE);
}
/* Quick return */
-/*
- if (chameleon_min(M, chameleon_min(N, NRHS)) == 0) {
- return MORSE_SUCCESS;
- }
-*/
+ /*
+ if (chameleon_min(M, chameleon_min(N, NRHS)) == 0) {
+ return MORSE_SUCCESS;
+ }
+ */
/* subB = morse_desc_submatrix(B, A->m, 0, A->n-A->m, B->n);
- morse_pztile_zero(subB, sequence, request);
- free(subB); */
+ morse_pzlaset( MorseUpperLower, 0., 0., subB, sequence, request );
+ free(subB); */
subB = morse_desc_submatrix(B, 0, 0, A->m, B->n);
subA = morse_desc_submatrix(A, 0, 0, A->m, A->m);
- morse_pztrsm(MorseLeft, MorseLower, MorseNoTrans, MorseNonUnit, 1.0, subA, subB, sequence, request);
+ morse_pztrsm( MorseLeft, MorseLower, MorseNoTrans, MorseNonUnit, 1.0, subA, subB, sequence, request );
free(subA);
free(subB);
@@ -339,9 +341,15 @@ int MORSE_zgelqs_param_Tile_Async(const libhqr_tree_t *qrtree, MORSE_desc_t *A,
}
#endif
- morse_pzunmlq_param(qrtree, MorseLeft, MorseConjTrans, A, B, TS, TT, Dptr, sequence, request);
+ morse_pzunmlq_param( qrtree, MorseLeft, MorseConjTrans, A, B, TS, TT, Dptr, sequence, request );
if (Dptr != NULL) {
- morse_desc_mat_free(Dptr);
+ MORSE_Desc_Flush( A, sequence );
+ MORSE_Desc_Flush( B, sequence );
+ MORSE_Desc_Flush( TS, sequence );
+ MORSE_Desc_Flush( TT, sequence );
+ MORSE_Desc_Flush( Dptr, sequence );
+ morse_sequence_wait( morse, sequence );
+ morse_desc_mat_free( Dptr );
}
(void)D;
return MORSE_SUCCESS;
diff --git a/compute/zgels.c b/compute/zgels.c
index 4ab15cb6d82b1b12fb3cdc18df73a7f04abcb674..74d65ca46ff14367be7a567f941a1f87e5e7a15b 100644
--- a/compute/zgels.c
+++ b/compute/zgels.c
@@ -107,10 +107,10 @@
* @sa MORSE_sgels
*
******************************************************************************/
-int MORSE_zgels(MORSE_enum trans, int M, int N, int NRHS,
+int MORSE_zgels( MORSE_enum trans, int M, int N, int NRHS,
MORSE_Complex64_t *A, int LDA,
MORSE_desc_t *descT,
- MORSE_Complex64_t *B, int LDB)
+ MORSE_Complex64_t *B, int LDB )
{
int i, j;
int NB;
@@ -118,7 +118,8 @@ int MORSE_zgels(MORSE_enum trans, int M, int N, int NRHS,
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
MORSE_request_t request = MORSE_REQUEST_INITIALIZER;
- MORSE_desc_t descA, descB;
+ MORSE_desc_t descAl, descAt;
+ MORSE_desc_t descBl, descBt;
morse = morse_context_self();
if (morse == NULL) {
@@ -169,51 +170,39 @@ int MORSE_zgels(MORSE_enum trans, int M, int N, int NRHS,
/* Set NT */
NB = MORSE_NB;
- morse_sequence_create(morse, &sequence);
+ morse_sequence_create( morse, &sequence );
+ /* Submit the matrix conversion */
if ( M >= N ) {
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooplap2tile( descA, A, NB, NB, LDA, N, 0, 0, M, N, sequence, &request,
- morse_desc_mat_free(&(descA)) );
- morse_zooplap2tile( descB, B, NB, NB, LDB, NRHS, 0, 0, M, NRHS, sequence, &request,
- morse_desc_mat_free(&(descA)); morse_desc_mat_free(&(descB)));
-/* } else {*/
-/* morse_ziplap2tile( descA, A, NB, NB, LDA, N, 0, 0, M, N,*/
-/* sequence, &request);*/
-/* morse_ziplap2tile( descB, B, NB, NB, LDB, NRHS, 0, 0, M, NRHS,*/
-/* sequence, &request);*/
-/* }*/
+ morse_zlap2tile( morse, &descAl, &descAt, MorseDescInout, MorseUpperLower,
+ A, NB, NB, LDA, N, M, N, sequence, &request );
+ morse_zlap2tile( morse, &descBl, &descBt, MorseDescInout, MorseUpperLower,
+ B, NB, NB, LDB, NRHS, M, NRHS, sequence, &request );
} else {
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooplap2tile( descA, A, NB, NB, LDA, N, 0, 0, M, N, sequence, &request,
- morse_desc_mat_free(&(descA)) );
- morse_zooplap2tile( descB, B, NB, NB, LDB, NRHS, 0, 0, N, NRHS, sequence, &request,
- morse_desc_mat_free(&(descA)); morse_desc_mat_free(&(descB)));
-/* } else {*/
-/* morse_ziplap2tile( descA, A, NB, NB, LDA, N, 0, 0, M, N,*/
-/* sequence, &request);*/
-/* morse_ziplap2tile( descB, B, NB, NB, LDB, NRHS, 0, 0, N, NRHS,*/
-/* sequence, &request);*/
-/* }*/
+ /* Submit the matrix conversion */
+ morse_zlap2tile( morse, &descAl, &descAt, MorseDescInout, MorseUpperLower,
+ A, NB, NB, LDA, N, M, N, sequence, &request );
+ morse_zlap2tile( morse, &descBl, &descBt, MorseDescInout, MorseUpperLower,
+ B, NB, NB, LDB, NRHS, N, NRHS, sequence, &request );
}
/* Call the tile interface */
- MORSE_zgels_Tile_Async(MorseNoTrans, &descA, descT, &descB, sequence, &request);
+ MORSE_zgels_Tile_Async( MorseNoTrans, &descAt, descT, &descBt, sequence, &request );
+
+ /* Submit the matrix conversion back */
+ morse_ztile2lap( morse, &descAl, &descAt,
+ MorseDescInout, MorseUpperLower, sequence, &request );
+ morse_ztile2lap( morse, &descBl, &descBt,
+ MorseDescInout, MorseUpperLower, sequence, &request );
+
+ morse_sequence_wait( morse, sequence );
+
+ /* Cleanup the temporary data */
+ morse_ztile2lap_cleanup( morse, &descAl, &descAt );
+ morse_ztile2lap_cleanup( morse, &descBl, &descBt );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooptile2lap(descA, A, NB, NB, LDA, N, sequence, &request);
- morse_zooptile2lap(descB, B, NB, NB, LDB, NRHS, sequence, &request);
- morse_sequence_wait(morse, sequence);
- morse_desc_mat_free(&descA);
- morse_desc_mat_free(&descB);
-/* } else {*/
-/* morse_ziptile2lap( descA, A, NB, NB, LDA, N, sequence, &request);*/
-/* morse_ziptile2lap( descB, B, NB, NB, LDB, NRHS, sequence, &request);*/
-/* morse_sequence_wait(morse, sequence);*/
-/* }*/
-
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -271,8 +260,8 @@ int MORSE_zgels(MORSE_enum trans, int M, int N, int NRHS,
* @sa MORSE_sgels_Tile
*
******************************************************************************/
-int MORSE_zgels_Tile(MORSE_enum trans, MORSE_desc_t *A,
- MORSE_desc_t *T, MORSE_desc_t *B)
+int MORSE_zgels_Tile( MORSE_enum trans, MORSE_desc_t *A,
+ MORSE_desc_t *T, MORSE_desc_t *B )
{
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
@@ -284,14 +273,17 @@ int MORSE_zgels_Tile(MORSE_enum trans, MORSE_desc_t *A,
morse_fatal_error("MORSE_zgels_Tile", "MORSE not initialized");
return MORSE_ERR_NOT_INITIALIZED;
}
- morse_sequence_create(morse, &sequence);
- MORSE_zgels_Tile_Async(trans, A, T, B, sequence, &request);
- RUNTIME_desc_flush( A, sequence );
- RUNTIME_desc_flush( B, sequence );
- morse_sequence_wait(morse, sequence);
+ morse_sequence_create( morse, &sequence );
+
+ MORSE_zgels_Tile_Async( trans, A, T, B, sequence, &request );
+ MORSE_Desc_Flush( A, sequence );
+ MORSE_Desc_Flush( T, sequence );
+ MORSE_Desc_Flush( B, sequence );
+
+ morse_sequence_wait( morse, sequence );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -324,9 +316,9 @@ int MORSE_zgels_Tile(MORSE_enum trans, MORSE_desc_t *A,
* @sa MORSE_sgels_Tile_Async
*
******************************************************************************/
-int MORSE_zgels_Tile_Async(MORSE_enum trans, MORSE_desc_t *A,
- MORSE_desc_t *T, MORSE_desc_t *B,
- MORSE_sequence_t *sequence, MORSE_request_t *request)
+int MORSE_zgels_Tile_Async( MORSE_enum trans, MORSE_desc_t *A,
+ MORSE_desc_t *T, MORSE_desc_t *B,
+ MORSE_sequence_t *sequence, MORSE_request_t *request )
{
MORSE_desc_t *subA;
MORSE_desc_t *subB;
@@ -347,10 +339,12 @@ int MORSE_zgels_Tile_Async(MORSE_enum trans, MORSE_desc_t *A,
return MORSE_ERR_UNALLOCATED;
}
/* Check sequence status */
- if (sequence->status == MORSE_SUCCESS)
+ if (sequence->status == MORSE_SUCCESS) {
request->status = MORSE_SUCCESS;
- else
+ }
+ else {
return morse_request_fail(sequence, request, MORSE_ERR_SEQUENCE_FLUSHED);
+ }
/* Check descriptors for correctness */
if (morse_desc_check(A) != MORSE_SUCCESS) {
@@ -375,12 +369,12 @@ int MORSE_zgels_Tile_Async(MORSE_enum trans, MORSE_desc_t *A,
return morse_request_fail(sequence, request, MORSE_ERR_NOT_SUPPORTED);
}
/* Quick return - currently NOT equivalent to LAPACK's:
- if (chameleon_min(M, chameleon_min(N, NRHS)) == 0) {
- for (i = 0; i < chameleon_max(M, N); i++)
- for (j = 0; j < NRHS; j++)
- B[j*LDB+i] = 0.0;
- return MORSE_SUCCESS;
- }
+ if (chameleon_min(M, chameleon_min(N, NRHS)) == 0) {
+ for (i = 0; i < chameleon_max(M, N); i++)
+ for (j = 0; j < NRHS; j++)
+ B[j*LDB+i] = 0.0;
+ return MORSE_SUCCESS;
+ }
*/
if (A->m >= A->n) {
@@ -393,24 +387,24 @@ int MORSE_zgels_Tile_Async(MORSE_enum trans, MORSE_desc_t *A,
#endif
if (morse->householder == MORSE_FLAT_HOUSEHOLDER) {
- morse_pzgeqrf(A, T, Dptr, sequence, request);
+ morse_pzgeqrf( A, T, Dptr, sequence, request );
- morse_pzunmqr(MorseLeft, MorseConjTrans, A, B, T, Dptr, sequence, request);
+ morse_pzunmqr( MorseLeft, MorseConjTrans, A, B, T, Dptr, sequence, request );
}
else {
- morse_pzgeqrfrh(A, T, Dptr, MORSE_RHBLK, sequence, request);
+ morse_pzgeqrfrh( A, T, Dptr, MORSE_RHBLK, sequence, request );
- morse_pzunmqrrh(MorseLeft, MorseConjTrans, A, B, T, Dptr, MORSE_RHBLK, sequence, request);
+ morse_pzunmqrrh( MorseLeft, MorseConjTrans, A, B, T, Dptr, MORSE_RHBLK, sequence, request );
}
subB = morse_desc_submatrix(B, 0, 0, A->n, B->n);
subA = morse_desc_submatrix(A, 0, 0, A->n, A->n);
- morse_pztrsm(MorseLeft, MorseUpper, MorseNoTrans, MorseNonUnit, 1.0, subA, subB, sequence, request);
+ morse_pztrsm( MorseLeft, MorseUpper, MorseNoTrans, MorseNonUnit, 1.0, subA, subB, sequence, request );
}
else {
/* subB = morse_desc_submatrix(B, A->m, 0, A->n-A->m, B->n);
- morse_pztile_zero(subB, sequence, request);
- free(subB); */
+ morse_pzlaset( MorseUpperLower, 0., 0., subB, sequence, request );
+ free(subB); */
#if defined(CHAMELEON_COPY_DIAG)
{
int m = chameleon_min(A->mt, A->nt) * A->mb;
@@ -419,20 +413,20 @@ int MORSE_zgels_Tile_Async(MORSE_enum trans, MORSE_desc_t *A,
}
#endif
if (morse->householder == MORSE_FLAT_HOUSEHOLDER) {
- morse_pzgelqf(A, T, Dptr, sequence, request);
+ morse_pzgelqf( A, T, Dptr, sequence, request );
}
else {
- morse_pzgelqfrh(A, T, Dptr, MORSE_RHBLK, sequence, request);
+ morse_pzgelqfrh( A, T, Dptr, MORSE_RHBLK, sequence, request );
}
subB = morse_desc_submatrix(B, 0, 0, A->m, B->n);
subA = morse_desc_submatrix(A, 0, 0, A->m, A->m);
- morse_pztrsm(MorseLeft, MorseLower, MorseNoTrans, MorseNonUnit, 1.0, subA, subB, sequence, request);
+ morse_pztrsm( MorseLeft, MorseLower, MorseNoTrans, MorseNonUnit, 1.0, subA, subB, sequence, request );
if (morse->householder == MORSE_FLAT_HOUSEHOLDER) {
- morse_pzunmlq(MorseLeft, MorseConjTrans, A, B, T, Dptr, sequence, request);
+ morse_pzunmlq( MorseLeft, MorseConjTrans, A, B, T, Dptr, sequence, request );
}
else {
- morse_pzunmlqrh(MorseLeft, MorseConjTrans, A, B, T, Dptr, MORSE_RHBLK, sequence, request);
+ morse_pzunmlqrh( MorseLeft, MorseConjTrans, A, B, T, Dptr, MORSE_RHBLK, sequence, request );
}
}
@@ -440,7 +434,12 @@ int MORSE_zgels_Tile_Async(MORSE_enum trans, MORSE_desc_t *A,
free(subB);
if (Dptr != NULL) {
- morse_desc_mat_free(Dptr);
+ MORSE_Desc_Flush( A, sequence );
+ MORSE_Desc_Flush( T, sequence );
+ MORSE_Desc_Flush( B, sequence );
+ MORSE_Desc_Flush( Dptr, sequence );
+ morse_sequence_wait( morse, sequence );
+ morse_desc_mat_free( Dptr );
}
(void)D;
return MORSE_SUCCESS;
diff --git a/compute/zgels_param.c b/compute/zgels_param.c
index 9e91871be2973d89a78be91aa3c2e788074d757f..5619b004a8cb233b87e1bb877440b836f621e1f9 100644
--- a/compute/zgels_param.c
+++ b/compute/zgels_param.c
@@ -109,10 +109,10 @@
* @sa MORSE_sgels
*
******************************************************************************/
-int MORSE_zgels_param(const libhqr_tree_t *qrtree, MORSE_enum trans, int M, int N, int NRHS,
- MORSE_Complex64_t *A, int LDA,
- MORSE_desc_t *descTS, MORSE_desc_t *descTT,
- MORSE_Complex64_t *B, int LDB)
+int MORSE_zgels_param( const libhqr_tree_t *qrtree, MORSE_enum trans, int M, int N, int NRHS,
+ MORSE_Complex64_t *A, int LDA,
+ MORSE_desc_t *descTS, MORSE_desc_t *descTT,
+ MORSE_Complex64_t *B, int LDB )
{
int i, j;
int NB;
@@ -120,7 +120,8 @@ int MORSE_zgels_param(const libhqr_tree_t *qrtree, MORSE_enum trans, int M, int
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
MORSE_request_t request = MORSE_REQUEST_INITIALIZER;
- MORSE_desc_t descA, descB;
+ MORSE_desc_t descAl, descAt;
+ MORSE_desc_t descBl, descBt;
morse = morse_context_self();
if (morse == NULL) {
@@ -171,51 +172,38 @@ int MORSE_zgels_param(const libhqr_tree_t *qrtree, MORSE_enum trans, int M, int
/* Set NT */
NB = MORSE_NB;
- morse_sequence_create(morse, &sequence);
+ morse_sequence_create( morse, &sequence );
+ /* Submit the matrix conversion */
if ( M >= N ) {
- /* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooplap2tile( descA, A, NB, NB, LDA, N, 0, 0, M, N, sequence, &request,
- morse_desc_mat_free(&(descA)) );
- morse_zooplap2tile( descB, B, NB, NB, LDB, NRHS, 0, 0, M, NRHS, sequence, &request,
- morse_desc_mat_free(&(descA)); morse_desc_mat_free(&(descB)));
- /* } else {*/
- /* morse_ziplap2tile( descA, A, NB, NB, LDA, N, 0, 0, M, N,*/
- /* sequence, &request);*/
- /* morse_ziplap2tile( descB, B, NB, NB, LDB, NRHS, 0, 0, M, NRHS,*/
- /* sequence, &request);*/
- /* }*/
+ morse_zlap2tile( morse, &descAl, &descAt, MorseDescInout, MorseUpperLower,
+ A, NB, NB, LDA, N, M, N, sequence, &request );
+ morse_zlap2tile( morse, &descBl, &descBt, MorseDescInout, MorseUpperLower,
+ B, NB, NB, LDB, NRHS, M, NRHS, sequence, &request );
} else {
- /* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooplap2tile( descA, A, NB, NB, LDA, N, 0, 0, M, N, sequence, &request,
- morse_desc_mat_free(&(descA)) );
- morse_zooplap2tile( descB, B, NB, NB, LDB, NRHS, 0, 0, N, NRHS, sequence, &request,
- morse_desc_mat_free(&(descA)); morse_desc_mat_free(&(descB)));
- /* } else {*/
- /* morse_ziplap2tile( descA, A, NB, NB, LDA, N, 0, 0, M, N,*/
- /* sequence, &request);*/
- /* morse_ziplap2tile( descB, B, NB, NB, LDB, NRHS, 0, 0, N, NRHS,*/
- /* sequence, &request);*/
- /* }*/
+ morse_zlap2tile( morse, &descAl, &descAt, MorseDescInout, MorseUpperLower,
+ A, NB, NB, LDA, N, M, N, sequence, &request );
+ morse_zlap2tile( morse, &descBl, &descBt, MorseDescInout, MorseUpperLower,
+ B, NB, NB, LDB, NRHS, N, NRHS, sequence, &request );
}
/* Call the tile interface */
- MORSE_zgels_param_Tile_Async(qrtree, MorseNoTrans, &descA, descTS, descTT, &descB, sequence, &request);
+ MORSE_zgels_param_Tile_Async( qrtree, MorseNoTrans, &descAt, descTS, descTT, &descBt, sequence, &request );
- /* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooptile2lap(descA, A, NB, NB, LDA, N, sequence, &request);
- morse_zooptile2lap(descB, B, NB, NB, LDB, NRHS, sequence, &request);
- morse_sequence_wait(morse, sequence);
- morse_desc_mat_free(&descA);
- morse_desc_mat_free(&descB);
- /* } else {*/
- /* morse_ziptile2lap( descA, A, NB, NB, LDA, N, sequence, &request);*/
- /* morse_ziptile2lap( descB, B, NB, NB, LDB, NRHS, sequence, &request);*/
- /* morse_sequence_wait(morse, sequence);*/
- /* }*/
+ /* Submit the matrix conversion back */
+ morse_ztile2lap( morse, &descAl, &descAt,
+ MorseDescInout, MorseUpperLower, sequence, &request );
+ morse_ztile2lap( morse, &descBl, &descBt,
+ MorseDescInout, MorseUpperLower, sequence, &request );
+
+ morse_sequence_wait( morse, sequence );
+
+ /* Cleanup the temporary data */
+ morse_ztile2lap_cleanup( morse, &descAl, &descAt );
+ morse_ztile2lap_cleanup( morse, &descBl, &descBt );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -276,8 +264,8 @@ int MORSE_zgels_param(const libhqr_tree_t *qrtree, MORSE_enum trans, int M, int
* @sa MORSE_sgels_Tile
*
******************************************************************************/
-int MORSE_zgels_param_Tile(const libhqr_tree_t *qrtree, MORSE_enum trans, MORSE_desc_t *A,
- MORSE_desc_t *TS, MORSE_desc_t *TT, MORSE_desc_t *B)
+int MORSE_zgels_param_Tile( const libhqr_tree_t *qrtree, MORSE_enum trans, MORSE_desc_t *A,
+ MORSE_desc_t *TS, MORSE_desc_t *TT, MORSE_desc_t *B )
{
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
@@ -289,14 +277,18 @@ int MORSE_zgels_param_Tile(const libhqr_tree_t *qrtree, MORSE_enum trans, MORSE_
morse_fatal_error("MORSE_zgels_param_Tile", "MORSE not initialized");
return MORSE_ERR_NOT_INITIALIZED;
}
- morse_sequence_create(morse, &sequence);
- MORSE_zgels_param_Tile_Async(qrtree, trans, A, TS, TT, B, sequence, &request);
- RUNTIME_desc_flush( A, sequence );
- RUNTIME_desc_flush( B, sequence );
- morse_sequence_wait(morse, sequence);
+ morse_sequence_create( morse, &sequence );
+
+ MORSE_zgels_param_Tile_Async( qrtree, trans, A, TS, TT, B, sequence, &request );
+
+ MORSE_Desc_Flush( A, sequence );
+ MORSE_Desc_Flush( TS, sequence );
+ MORSE_Desc_Flush( TT, sequence );
+ MORSE_Desc_Flush( B, sequence );
+ morse_sequence_wait( morse, sequence );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -329,9 +321,9 @@ int MORSE_zgels_param_Tile(const libhqr_tree_t *qrtree, MORSE_enum trans, MORSE_
* @sa MORSE_sgels_Tile_Async
*
******************************************************************************/
-int MORSE_zgels_param_Tile_Async(const libhqr_tree_t *qrtree, MORSE_enum trans, MORSE_desc_t *A,
- MORSE_desc_t *TS, MORSE_desc_t *TT, MORSE_desc_t *B,
- MORSE_sequence_t *sequence, MORSE_request_t *request)
+int MORSE_zgels_param_Tile_Async( const libhqr_tree_t *qrtree, MORSE_enum trans, MORSE_desc_t *A,
+ MORSE_desc_t *TS, MORSE_desc_t *TT, MORSE_desc_t *B,
+ MORSE_sequence_t *sequence, MORSE_request_t *request )
{
MORSE_desc_t *subA;
MORSE_desc_t *subB;
@@ -352,10 +344,12 @@ int MORSE_zgels_param_Tile_Async(const libhqr_tree_t *qrtree, MORSE_enum trans,
return MORSE_ERR_UNALLOCATED;
}
/* Check sequence status */
- if (sequence->status == MORSE_SUCCESS)
+ if (sequence->status == MORSE_SUCCESS) {
request->status = MORSE_SUCCESS;
- else
+ }
+ else {
return morse_request_fail(sequence, request, MORSE_ERR_SEQUENCE_FLUSHED);
+ }
/* Check descriptors for correctness */
if (morse_desc_check(A) != MORSE_SUCCESS) {
@@ -404,9 +398,9 @@ int MORSE_zgels_param_Tile_Async(const libhqr_tree_t *qrtree, MORSE_enum trans,
subB = morse_desc_submatrix(B, 0, 0, A->n, B->n);
subA = morse_desc_submatrix(A, 0, 0, A->n, A->n);
- morse_pzgeqrf_param(qrtree, A, TS, TT, Dptr, sequence, request);
- morse_pzunmqr_param(qrtree, MorseLeft, MorseConjTrans, A, B, TS, TT, Dptr, sequence, request);
- morse_pztrsm(MorseLeft, MorseUpper, MorseNoTrans, MorseNonUnit, 1.0, subA, subB, sequence, request);
+ morse_pzgeqrf_param( qrtree, A, TS, TT, Dptr, sequence, request );
+ morse_pzunmqr_param( qrtree, MorseLeft, MorseConjTrans, A, B, TS, TT, Dptr, sequence, request );
+ morse_pztrsm( MorseLeft, MorseUpper, MorseNoTrans, MorseNonUnit, 1.0, subA, subB, sequence, request );
}
else {
#if defined(CHAMELEON_COPY_DIAG)
@@ -420,16 +414,22 @@ int MORSE_zgels_param_Tile_Async(const libhqr_tree_t *qrtree, MORSE_enum trans,
subB = morse_desc_submatrix(B, 0, 0, A->m, B->n);
subA = morse_desc_submatrix(A, 0, 0, A->m, A->m);
- morse_pzgelqf_param(qrtree, A, TS, TT, Dptr, sequence, request);
- morse_pztrsm(MorseLeft, MorseLower, MorseNoTrans, MorseNonUnit, 1.0, subA, subB, sequence, request);
- morse_pzunmlq_param(qrtree, MorseLeft, MorseConjTrans, A, B, TS, TT, Dptr, sequence, request);
+ morse_pzgelqf_param( qrtree, A, TS, TT, Dptr, sequence, request );
+ morse_pztrsm( MorseLeft, MorseLower, MorseNoTrans, MorseNonUnit, 1.0, subA, subB, sequence, request );
+ morse_pzunmlq_param( qrtree, MorseLeft, MorseConjTrans, A, B, TS, TT, Dptr, sequence, request );
}
free(subA);
free(subB);
if (Dptr != NULL) {
- morse_desc_mat_free(Dptr);
+ MORSE_Desc_Flush( A, sequence );
+ MORSE_Desc_Flush( B, sequence );
+ MORSE_Desc_Flush( TS, sequence );
+ MORSE_Desc_Flush( TT, sequence );
+ MORSE_Desc_Flush( Dptr, sequence );
+ morse_sequence_wait( morse, sequence );
+ morse_desc_mat_free( Dptr );
}
(void)D;
return MORSE_SUCCESS;
diff --git a/compute/zgemm.c b/compute/zgemm.c
index 332b540061849362a78b9c4dc8d893e39ae56bbb..92101a69bb9e0f7ea48f6101e7332924e4837b1c 100644
--- a/compute/zgemm.c
+++ b/compute/zgemm.c
@@ -130,15 +130,17 @@
* @sa MORSE_sgemm
*
******************************************************************************/
-int MORSE_zgemm(MORSE_enum transA, MORSE_enum transB, int M, int N, int K,
+int MORSE_zgemm( MORSE_enum transA, MORSE_enum transB, int M, int N, int K,
MORSE_Complex64_t alpha, MORSE_Complex64_t *A, int LDA,
- MORSE_Complex64_t *B, int LDB,
- MORSE_Complex64_t beta, MORSE_Complex64_t *C, int LDC)
+ MORSE_Complex64_t *B, int LDB,
+ MORSE_Complex64_t beta, MORSE_Complex64_t *C, int LDC )
{
int NB;
int Am, An, Bm, Bn;
int status;
- MORSE_desc_t descA, descB, descC;
+ MORSE_desc_t descAl, descAt;
+ MORSE_desc_t descBl, descBt;
+ MORSE_desc_t descCl, descCt;
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
MORSE_request_t request = MORSE_REQUEST_INITIALIZER;
@@ -208,43 +210,36 @@ int MORSE_zgemm(MORSE_enum transA, MORSE_enum transB, int M, int N, int K,
/* Set MT & NT & KT */
NB = MORSE_NB;
- morse_sequence_create(morse, &sequence);
-
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooplap2tile( descA, A, NB, NB, LDA, An, 0, 0, Am, An, sequence, &request,
- morse_desc_mat_free(&(descA)) );
- morse_zooplap2tile( descB, B, NB, NB, LDB, Bn, 0, 0, Bm, Bn, sequence, &request,
- morse_desc_mat_free(&(descA)); morse_desc_mat_free(&(descB)));
- morse_zooplap2tile( descC, C, NB, NB, LDC, N, 0, 0, M, N, sequence, &request,
- morse_desc_mat_free(&(descA)); morse_desc_mat_free(&(descB)); morse_desc_mat_free(&(descC)));
-/* } else {*/
-/* morse_ziplap2tile( descA, A, NB, NB, LDA, An, 0, 0, Am, An, */
-/* sequence, &request);*/
-/* morse_ziplap2tile( descB, B, NB, NB, LDB, Bn, 0, 0, Bm, Bn, */
-/* sequence, &request);*/
-/* morse_ziplap2tile( descC, C, NB, NB, LDC, N, 0, 0, M, N, */
-/* sequence, &request);*/
-/* }*/
+ morse_sequence_create( morse, &sequence );
+
+ /* Submit the matrix conversion */
+ morse_zlap2tile( morse, &descAl, &descAt, MorseDescInput, MorseUpperLower,
+ A, NB, NB, LDA, An, Am, An, sequence, &request );
+ morse_zlap2tile( morse, &descBl, &descBt, MorseDescInput, MorseUpperLower,
+ B, NB, NB, LDB, Bn, Bm, Bn, sequence, &request );
+ morse_zlap2tile( morse, &descCl, &descCt, MorseDescInout, MorseUpperLower,
+ C, NB, NB, LDC, N, M, N, sequence, &request );
/* Call the tile interface */
- MORSE_zgemm_Tile_Async(
- transA, transB, alpha, &descA, &descB, beta, &descC, sequence, &request);
-
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooptile2lap(descC, C, NB, NB, LDC, N, sequence, &request);
- morse_sequence_wait(morse, sequence);
- morse_desc_mat_free(&descA);
- morse_desc_mat_free(&descB);
- morse_desc_mat_free(&descC);
-/* } else {*/
-/* morse_ziptile2lap( descA, A, NB, NB, LDA, An, sequence, &request);*/
-/* morse_ziptile2lap( descB, B, NB, NB, LDB, Bn, sequence, &request);*/
-/* morse_ziptile2lap( descC, C, NB, NB, LDC, N, sequence, &request);*/
-/* morse_sequence_wait(morse, sequence);*/
-/* }*/
+ MORSE_zgemm_Tile_Async( transA, transB, alpha, &descAt, &descBt, beta, &descCt, sequence, &request );
+
+ /* Submit the matrix conversion back */
+ morse_ztile2lap( morse, &descAl, &descAt,
+ MorseDescInput, MorseUpperLower, sequence, &request );
+ morse_ztile2lap( morse, &descBl, &descBt,
+ MorseDescInput, MorseUpperLower, sequence, &request );
+ morse_ztile2lap( morse, &descCl, &descCt,
+ MorseDescInout, MorseUpperLower, sequence, &request );
+
+ morse_sequence_wait( morse, sequence );
+
+ /* Cleanup the temporary data */
+ morse_ztile2lap_cleanup( morse, &descAl, &descAt );
+ morse_ztile2lap_cleanup( morse, &descBl, &descBt );
+ morse_ztile2lap_cleanup( morse, &descCl, &descCt );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -305,9 +300,9 @@ int MORSE_zgemm(MORSE_enum transA, MORSE_enum transB, int M, int N, int K,
* @sa MORSE_sgemm_Tile
*
******************************************************************************/
-int MORSE_zgemm_Tile(MORSE_enum transA, MORSE_enum transB,
- MORSE_Complex64_t alpha, MORSE_desc_t *A, MORSE_desc_t *B,
- MORSE_Complex64_t beta, MORSE_desc_t *C)
+int MORSE_zgemm_Tile( MORSE_enum transA, MORSE_enum transB,
+ MORSE_Complex64_t alpha, MORSE_desc_t *A, MORSE_desc_t *B,
+ MORSE_Complex64_t beta, MORSE_desc_t *C )
{
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
@@ -319,15 +314,17 @@ int MORSE_zgemm_Tile(MORSE_enum transA, MORSE_enum transB,
morse_fatal_error("MORSE_zgemm_Tile", "MORSE not initialized");
return MORSE_ERR_NOT_INITIALIZED;
}
- morse_sequence_create(morse, &sequence);
- MORSE_zgemm_Tile_Async(transA, transB, alpha, A, B, beta, C, sequence, &request);
- RUNTIME_desc_flush( A, sequence );
- RUNTIME_desc_flush( B, sequence );
- RUNTIME_desc_flush( C, sequence );
- morse_sequence_wait(morse, sequence);
+ morse_sequence_create( morse, &sequence );
+
+ MORSE_zgemm_Tile_Async( transA, transB, alpha, A, B, beta, C, sequence, &request );
+
+ MORSE_Desc_Flush( A, sequence );
+ MORSE_Desc_Flush( B, sequence );
+ MORSE_Desc_Flush( C, sequence );
+ morse_sequence_wait( morse, sequence );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -359,10 +356,10 @@ int MORSE_zgemm_Tile(MORSE_enum transA, MORSE_enum transB,
* @sa MORSE_sgemm_Tile_Async
*
******************************************************************************/
-int MORSE_zgemm_Tile_Async(MORSE_enum transA, MORSE_enum transB,
+int MORSE_zgemm_Tile_Async( MORSE_enum transA, MORSE_enum transB,
MORSE_Complex64_t alpha, MORSE_desc_t *A, MORSE_desc_t *B,
MORSE_Complex64_t beta, MORSE_desc_t *C,
- MORSE_sequence_t *sequence, MORSE_request_t *request)
+ MORSE_sequence_t *sequence, MORSE_request_t *request )
{
MORSE_context_t *morse;
int M, N, K;
@@ -383,10 +380,12 @@ int MORSE_zgemm_Tile_Async(MORSE_enum transA, MORSE_enum transB,
return MORSE_ERR_UNALLOCATED;
}
/* Check sequence status */
- if (sequence->status == MORSE_SUCCESS)
+ if (sequence->status == MORSE_SUCCESS) {
request->status = MORSE_SUCCESS;
- else
+ }
+ else {
return morse_request_fail(sequence, request, MORSE_ERR_SEQUENCE_FLUSHED);
+ }
/* Check descriptors for correctness */
if (morse_desc_check(A) != MORSE_SUCCESS) {
@@ -461,11 +460,13 @@ int MORSE_zgemm_Tile_Async(MORSE_enum transA, MORSE_enum transB,
K = An;
/* Quick return */
- if (M == 0 || N == 0 ||
- ((alpha == (MORSE_Complex64_t)0.0 || K == 0) && beta == (MORSE_Complex64_t)1.0))
+ if ( (M == 0) || (N == 0) ||
+ (((alpha == (MORSE_Complex64_t)0.0) || (K == 0)) && (beta == (MORSE_Complex64_t)1.0)) )
+ {
return MORSE_SUCCESS;
+ }
- morse_pzgemm(transA, transB, alpha, A, B, beta, C, sequence, request);
+ morse_pzgemm( transA, transB, alpha, A, B, beta, C, sequence, request );
return MORSE_SUCCESS;
}
diff --git a/compute/zgeqrf.c b/compute/zgeqrf.c
index 5e500472b0dd8c3ed530b94c9e230b7833bd8172..6695351de0915ddcb7d5ff6dde9ba993b9f3b197 100644
--- a/compute/zgeqrf.c
+++ b/compute/zgeqrf.c
@@ -74,16 +74,16 @@
* @sa MORSE_zgeqrs
*
******************************************************************************/
-int MORSE_zgeqrf(int M, int N,
+int MORSE_zgeqrf( int M, int N,
MORSE_Complex64_t *A, int LDA,
- MORSE_desc_t *descT)
+ MORSE_desc_t *descT )
{
int NB;
int status;
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
MORSE_request_t request = MORSE_REQUEST_INITIALIZER;
- MORSE_desc_t descA;
+ MORSE_desc_t descAl, descAt;
morse = morse_context_self();
if (morse == NULL) {
@@ -119,30 +119,26 @@ int MORSE_zgeqrf(int M, int N,
/* Set NT */
NB = MORSE_NB;
- morse_sequence_create(morse, &sequence);
+ morse_sequence_create( morse, &sequence );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooplap2tile( descA, A, NB, NB, LDA, N, 0, 0, M, N, sequence, &request,
- morse_desc_mat_free(&(descA)) );
-/* } else {*/
-/* morse_ziplap2tile( descA, A, NB, NB, LDA, N, 0, 0, M, N,*/
-/* sequence, &request);*/
-/* }*/
+ /* Submit the matrix conversion */
+ morse_zlap2tile( morse, &descAl, &descAt, MorseDescInout, MorseUpperLower,
+ A, NB, NB, LDA, N, M, N, sequence, &request );
/* Call the tile interface */
- MORSE_zgeqrf_Tile_Async(&descA, descT, sequence, &request);
+ MORSE_zgeqrf_Tile_Async( &descAt, descT, sequence, &request );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooptile2lap(descA, A, NB, NB, LDA, N, sequence, &request);
- morse_sequence_wait(morse, sequence);
- morse_desc_mat_free(&descA);
-/* } else {*/
-/* morse_ziptile2lap( descA, A, NB, NB, LDA, N, sequence, &request);*/
-/* morse_sequence_wait(morse, sequence);*/
-/* }*/
+ /* Submit the matrix conversion back */
+ morse_ztile2lap( morse, &descAl, &descAt,
+ MorseDescInout, MorseUpperLower, sequence, &request );
+
+ morse_sequence_wait( morse, sequence );
+
+ /* Cleanup the temporary data */
+ morse_ztile2lap_cleanup( morse, &descAl, &descAt );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -185,7 +181,7 @@ int MORSE_zgeqrf(int M, int N,
* @sa MORSE_zgeqrs_Tile
*
******************************************************************************/
-int MORSE_zgeqrf_Tile(MORSE_desc_t *A, MORSE_desc_t *T)
+int MORSE_zgeqrf_Tile( MORSE_desc_t *A, MORSE_desc_t *T )
{
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
@@ -197,13 +193,16 @@ int MORSE_zgeqrf_Tile(MORSE_desc_t *A, MORSE_desc_t *T)
morse_fatal_error("MORSE_zgeqrf_Tile", "MORSE not initialized");
return MORSE_ERR_NOT_INITIALIZED;
}
- morse_sequence_create(morse, &sequence);
- MORSE_zgeqrf_Tile_Async(A, T, sequence, &request);
- RUNTIME_desc_flush( A, sequence );
- morse_sequence_wait(morse, sequence);
+ morse_sequence_create( morse, &sequence );
+
+ MORSE_zgeqrf_Tile_Async( A, T, sequence, &request );
+ MORSE_Desc_Flush( A, sequence );
+ MORSE_Desc_Flush( T, sequence );
+
+ morse_sequence_wait( morse, sequence );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -236,8 +235,8 @@ int MORSE_zgeqrf_Tile(MORSE_desc_t *A, MORSE_desc_t *T)
* @sa MORSE_zgeqrs_Tile_Async
*
******************************************************************************/
-int MORSE_zgeqrf_Tile_Async(MORSE_desc_t *A, MORSE_desc_t *T,
- MORSE_sequence_t *sequence, MORSE_request_t *request)
+int MORSE_zgeqrf_Tile_Async( MORSE_desc_t *A, MORSE_desc_t *T,
+ MORSE_sequence_t *sequence, MORSE_request_t *request )
{
MORSE_context_t *morse;
MORSE_desc_t D, *Dptr = NULL;
@@ -256,10 +255,12 @@ int MORSE_zgeqrf_Tile_Async(MORSE_desc_t *A, MORSE_desc_t *T,
return MORSE_ERR_UNALLOCATED;
}
/* Check sequence status */
- if (sequence->status == MORSE_SUCCESS)
+ if (sequence->status == MORSE_SUCCESS) {
request->status = MORSE_SUCCESS;
- else
+ }
+ else {
return morse_request_fail(sequence, request, MORSE_ERR_SEQUENCE_FLUSHED);
+ }
/* Check descriptors for correctness */
if (morse_desc_check(A) != MORSE_SUCCESS) {
@@ -276,10 +277,10 @@ int MORSE_zgeqrf_Tile_Async(MORSE_desc_t *A, MORSE_desc_t *T,
return morse_request_fail(sequence, request, MORSE_ERR_ILLEGAL_VALUE);
}
/* Quick return */
-/*
- if (chameleon_min(M, N) == 0)
- return MORSE_SUCCESS;
-*/
+ /*
+ if (chameleon_min(M, N) == 0)
+ return MORSE_SUCCESS;
+ */
#if defined(CHAMELEON_COPY_DIAG)
{
int n = chameleon_min(A->mt, A->nt) * A->nb;
@@ -289,13 +290,17 @@ int MORSE_zgeqrf_Tile_Async(MORSE_desc_t *A, MORSE_desc_t *T,
#endif
if (morse->householder == MORSE_FLAT_HOUSEHOLDER) {
- morse_pzgeqrf(A, T, Dptr, sequence, request);
+ morse_pzgeqrf( A, T, Dptr, sequence, request );
}
else {
- morse_pzgeqrfrh(A, T, Dptr, MORSE_RHBLK, sequence, request);
+ morse_pzgeqrfrh( A, T, Dptr, MORSE_RHBLK, sequence, request );
}
if (Dptr != NULL) {
- morse_desc_mat_free(Dptr);
+ MORSE_Desc_Flush( A, sequence );
+ MORSE_Desc_Flush( T, sequence );
+ MORSE_Desc_Flush( Dptr, sequence );
+ morse_sequence_wait( morse, sequence );
+ morse_desc_mat_free( Dptr );
}
(void)D;
return MORSE_SUCCESS;
diff --git a/compute/zgeqrf_param.c b/compute/zgeqrf_param.c
index a67f508db78185e1264e6935a0cfa379df024c1c..5e5b220f010c1ad2872a461b9df78e5b81167f8f 100644
--- a/compute/zgeqrf_param.c
+++ b/compute/zgeqrf_param.c
@@ -77,16 +77,16 @@
* @sa MORSE_zgeqrs
*
******************************************************************************/
-int MORSE_zgeqrf_param(const libhqr_tree_t *qrtree, int M, int N,
- MORSE_Complex64_t *A, int LDA,
- MORSE_desc_t *descTS, MORSE_desc_t *descTT)
+int MORSE_zgeqrf_param( const libhqr_tree_t *qrtree, int M, int N,
+ MORSE_Complex64_t *A, int LDA,
+ MORSE_desc_t *descTS, MORSE_desc_t *descTT )
{
int NB;
int status;
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
MORSE_request_t request = MORSE_REQUEST_INITIALIZER;
- MORSE_desc_t descA;
+ MORSE_desc_t descAl, descAt;
morse = morse_context_self();
if (morse == NULL) {
@@ -122,30 +122,26 @@ int MORSE_zgeqrf_param(const libhqr_tree_t *qrtree, int M, int N,
/* Set NT */
NB = MORSE_NB;
- morse_sequence_create(morse, &sequence);
+ morse_sequence_create( morse, &sequence );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooplap2tile( descA, A, NB, NB, LDA, N, 0, 0, M, N, sequence, &request,
- morse_desc_mat_free(&(descA)) );
-/* } else {*/
-/* morse_ziplap2tile( descA, A, NB, NB, LDA, N, 0, 0, M, N,*/
-/* sequence, &request);*/
-/* }*/
+ /* Submit the matrix conversion */
+ morse_zlap2tile( morse, &descAl, &descAt, MorseDescInout, MorseUpperLower,
+ A, NB, NB, LDA, N, M, N, sequence, &request );
/* Call the tile interface */
- MORSE_zgeqrf_param_Tile_Async(qrtree, &descA, descTS, descTT, sequence, &request);
+ MORSE_zgeqrf_param_Tile_Async( qrtree, &descAt, descTS, descTT, sequence, &request );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooptile2lap(descA, A, NB, NB, LDA, N, sequence, &request);
- morse_sequence_wait(morse, sequence);
- morse_desc_mat_free(&descA);
-/* } else {*/
-/* morse_ziptile2lap( descA, A, NB, NB, LDA, N, sequence, &request);*/
-/* morse_sequence_wait(morse, sequence);*/
-/* }*/
+ /* Submit the matrix conversion back */
+ morse_ztile2lap( morse, &descAl, &descAt,
+ MorseDescInout, MorseUpperLower, sequence, &request );
+
+ morse_sequence_wait( morse, sequence );
+
+ /* Cleanup the temporary data */
+ morse_ztile2lap_cleanup( morse, &descAl, &descAt );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -192,7 +188,7 @@ int MORSE_zgeqrf_param(const libhqr_tree_t *qrtree, int M, int N,
* @sa MORSE_zgeqrs_param_Tile
*
******************************************************************************/
-int MORSE_zgeqrf_param_Tile(const libhqr_tree_t *qrtree, MORSE_desc_t *A, MORSE_desc_t *TS, MORSE_desc_t *TT)
+int MORSE_zgeqrf_param_Tile( const libhqr_tree_t *qrtree, MORSE_desc_t *A, MORSE_desc_t *TS, MORSE_desc_t *TT )
{
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
@@ -204,13 +200,17 @@ int MORSE_zgeqrf_param_Tile(const libhqr_tree_t *qrtree, MORSE_desc_t *A, MORSE_
morse_fatal_error("MORSE_zgeqrf_param_Tile", "MORSE not initialized");
return MORSE_ERR_NOT_INITIALIZED;
}
- morse_sequence_create(morse, &sequence);
- MORSE_zgeqrf_param_Tile_Async(qrtree, A, TS, TT, sequence, &request);
- RUNTIME_desc_flush( A, sequence );
- morse_sequence_wait(morse, sequence);
+ morse_sequence_create( morse, &sequence );
+
+ MORSE_zgeqrf_param_Tile_Async( qrtree, A, TS, TT, sequence, &request );
+
+ MORSE_Desc_Flush( A, sequence );
+ MORSE_Desc_Flush( TS, sequence );
+ MORSE_Desc_Flush( TT, sequence );
+ morse_sequence_wait( morse, sequence );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -243,8 +243,8 @@ int MORSE_zgeqrf_param_Tile(const libhqr_tree_t *qrtree, MORSE_desc_t *A, MORSE_
* @sa MORSE_zgeqrs_param_Tile_Async
*
******************************************************************************/
-int MORSE_zgeqrf_param_Tile_Async(const libhqr_tree_t *qrtree, MORSE_desc_t *A, MORSE_desc_t *TS, MORSE_desc_t *TT,
- MORSE_sequence_t *sequence, MORSE_request_t *request)
+int MORSE_zgeqrf_param_Tile_Async( const libhqr_tree_t *qrtree, MORSE_desc_t *A, MORSE_desc_t *TS, MORSE_desc_t *TT,
+ MORSE_sequence_t *sequence, MORSE_request_t *request )
{
MORSE_context_t *morse;
MORSE_desc_t D, *Dptr = NULL;
@@ -263,10 +263,12 @@ int MORSE_zgeqrf_param_Tile_Async(const libhqr_tree_t *qrtree, MORSE_desc_t *A,
return MORSE_ERR_UNALLOCATED;
}
/* Check sequence status */
- if (sequence->status == MORSE_SUCCESS)
+ if (sequence->status == MORSE_SUCCESS) {
request->status = MORSE_SUCCESS;
- else
+ }
+ else {
return morse_request_fail(sequence, request, MORSE_ERR_SEQUENCE_FLUSHED);
+ }
/* Check descriptors for correctness */
if (morse_desc_check(A) != MORSE_SUCCESS) {
@@ -287,10 +289,10 @@ int MORSE_zgeqrf_param_Tile_Async(const libhqr_tree_t *qrtree, MORSE_desc_t *A,
return morse_request_fail(sequence, request, MORSE_ERR_ILLEGAL_VALUE);
}
/* Quick return */
-/*
- if (chameleon_min(M, N) == 0)
- return MORSE_SUCCESS;
-*/
+ /*
+ if (chameleon_min(M, N) == 0)
+ return MORSE_SUCCESS;
+ */
#if defined(CHAMELEON_COPY_DIAG)
{
int n = chameleon_min(A->mt, A->nt) * A->nb;
@@ -299,10 +301,15 @@ int MORSE_zgeqrf_param_Tile_Async(const libhqr_tree_t *qrtree, MORSE_desc_t *A,
}
#endif
- morse_pzgeqrf_param(qrtree, A, TS, TT, Dptr, sequence, request);
+ morse_pzgeqrf_param( qrtree, A, TS, TT, Dptr, sequence, request );
if (Dptr != NULL) {
- morse_desc_mat_free(Dptr);
+ MORSE_Desc_Flush( A, sequence );
+ MORSE_Desc_Flush( TS, sequence );
+ MORSE_Desc_Flush( TT, sequence );
+ MORSE_Desc_Flush( Dptr, sequence );
+ morse_sequence_wait( morse, sequence );
+ morse_desc_mat_free( Dptr );
}
(void)D;
return MORSE_SUCCESS;
diff --git a/compute/zgeqrs.c b/compute/zgeqrs.c
index 29b962f4f6c5f20483682f54eff2822ee9402edc..d8de43c6ebff5b2c2974ca2e9c2babc10cd60d89 100644
--- a/compute/zgeqrs.c
+++ b/compute/zgeqrs.c
@@ -81,17 +81,18 @@
* @sa MORSE_zgeqrf
*
*******************************************************************************/
-int MORSE_zgeqrs(int M, int N, int NRHS,
+int MORSE_zgeqrs( int M, int N, int NRHS,
MORSE_Complex64_t *A, int LDA,
MORSE_desc_t *descT,
- MORSE_Complex64_t *B, int LDB)
+ MORSE_Complex64_t *B, int LDB )
{
int NB;
int status;
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
MORSE_request_t request = MORSE_REQUEST_INITIALIZER;
- MORSE_desc_t descA, descB;
+ MORSE_desc_t descAl, descAt;
+ MORSE_desc_t descBl, descBt;
morse = morse_context_self();
if (morse == NULL) {
@@ -135,37 +136,31 @@ int MORSE_zgeqrs(int M, int N, int NRHS,
/* Set NT */
NB = MORSE_NB;
- morse_sequence_create(morse, &sequence);
+ morse_sequence_create( morse, &sequence );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooplap2tile( descA, A, NB, NB, LDA, N, 0, 0, M, N, sequence, &request,
- morse_desc_mat_free(&(descA)) );
- morse_zooplap2tile( descB, B, NB, NB, LDB, NRHS, 0, 0, M, NRHS, sequence, &request,
- morse_desc_mat_free(&(descA)); morse_desc_mat_free(&(descB)));
-/* } else {*/
-/* morse_ziplap2tile( descA, A, NB, NB, LDA, N, 0, 0, M, N, */
-/* sequence, &request);*/
-/* morse_ziplap2tile( descB, B, NB, NB, LDB, NRHS, 0, 0, M, NRHS,*/
-/* sequence, &request);*/
-/* }*/
+ /* Submit the matrix conversion */
+ morse_zlap2tile( morse, &descAl, &descAt, MorseDescInput, MorseUpperLower,
+ A, NB, NB, LDA, N, M, N, sequence, &request );
+ morse_zlap2tile( morse, &descBl, &descBt, MorseDescInout, MorseUpperLower,
+ B, NB, NB, LDB, NRHS, M, NRHS, sequence, &request );
/* Call the tile interface */
- MORSE_zgeqrs_Tile_Async(&descA, descT, &descB, sequence, &request);
+ MORSE_zgeqrs_Tile_Async( &descAt, descT, &descBt, sequence, &request );
+
+ /* Submit the matrix conversion back */
+ morse_ztile2lap( morse, &descAl, &descAt,
+ MorseDescInput, MorseUpperLower, sequence, &request );
+ morse_ztile2lap( morse, &descBl, &descBt,
+ MorseDescInout, MorseUpperLower, sequence, &request );
+
+ morse_sequence_wait( morse, sequence );
+
+ /* Cleanup the temporary data */
+ morse_ztile2lap_cleanup( morse, &descAl, &descAt );
+ morse_ztile2lap_cleanup( morse, &descBl, &descBt );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooptile2lap(descA, A, NB, NB, LDA, N, sequence, &request);
- morse_zooptile2lap(descB, B, NB, NB, LDB, NRHS, sequence, &request);
- morse_sequence_wait(morse, sequence);
- morse_desc_mat_free(&descA);
- morse_desc_mat_free(&descB);
-/* } else {*/
-/* morse_ziptile2lap( descA, A, NB, NB, LDA, N, sequence, &request);*/
-/* morse_ziptile2lap( descB, B, NB, NB, LDB, NRHS, sequence, &request);*/
-/* morse_sequence_wait(morse, sequence);*/
-/* }*/
-
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -207,7 +202,7 @@ int MORSE_zgeqrs(int M, int N, int NRHS,
* @sa MORSE_zgeqrf_Tile
*
******************************************************************************/
-int MORSE_zgeqrs_Tile(MORSE_desc_t *A, MORSE_desc_t *T, MORSE_desc_t *B)
+int MORSE_zgeqrs_Tile( MORSE_desc_t *A, MORSE_desc_t *T, MORSE_desc_t *B )
{
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
@@ -219,15 +214,17 @@ int MORSE_zgeqrs_Tile(MORSE_desc_t *A, MORSE_desc_t *T, MORSE_desc_t *B)
morse_fatal_error("MORSE_zgeqrs_Tile", "MORSE not initialized");
return MORSE_ERR_NOT_INITIALIZED;
}
- morse_sequence_create(morse, &sequence);
- MORSE_zgeqrs_Tile_Async(A, T, B, sequence, &request);
- RUNTIME_desc_flush( A, sequence );
- RUNTIME_desc_flush( T, sequence );
- RUNTIME_desc_flush( B, sequence );
- morse_sequence_wait(morse, sequence);
+ morse_sequence_create( morse, &sequence );
+ MORSE_zgeqrs_Tile_Async( A, T, B, sequence, &request );
+
+ MORSE_Desc_Flush( A, sequence );
+ MORSE_Desc_Flush( T, sequence );
+ MORSE_Desc_Flush( B, sequence );
+
+ morse_sequence_wait( morse, sequence );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -261,8 +258,8 @@ int MORSE_zgeqrs_Tile(MORSE_desc_t *A, MORSE_desc_t *T, MORSE_desc_t *B)
* @sa MORSE_zgeqrf_Tile_Async
*
******************************************************************************/
-int MORSE_zgeqrs_Tile_Async(MORSE_desc_t *A, MORSE_desc_t *T, MORSE_desc_t *B,
- MORSE_sequence_t *sequence, MORSE_request_t *request)
+int MORSE_zgeqrs_Tile_Async( MORSE_desc_t *A, MORSE_desc_t *T, MORSE_desc_t *B,
+ MORSE_sequence_t *sequence, MORSE_request_t *request )
{
MORSE_desc_t *subA;
MORSE_desc_t *subB;
@@ -283,10 +280,12 @@ int MORSE_zgeqrs_Tile_Async(MORSE_desc_t *A, MORSE_desc_t *T, MORSE_desc_t *B,
return MORSE_ERR_UNALLOCATED;
}
/* Check sequence status */
- if (sequence->status == MORSE_SUCCESS)
+ if (sequence->status == MORSE_SUCCESS) {
request->status = MORSE_SUCCESS;
- else
+ }
+ else {
return morse_request_fail(sequence, request, MORSE_ERR_SEQUENCE_FLUSHED);
+ }
/* Check descriptors for correctness */
if (morse_desc_check(A) != MORSE_SUCCESS) {
@@ -307,11 +306,11 @@ int MORSE_zgeqrs_Tile_Async(MORSE_desc_t *A, MORSE_desc_t *T, MORSE_desc_t *B,
return morse_request_fail(sequence, request, MORSE_ERR_ILLEGAL_VALUE);
}
/* Quick return */
-/*
- if (chameleon_min(M, chameleon_min(N, NRHS)) == 0) {
- return MORSE_SUCCESS;
- }
-*/
+ /*
+ if (chameleon_min(M, chameleon_min(N, NRHS)) == 0) {
+ return MORSE_SUCCESS;
+ }
+ */
#if defined(CHAMELEON_COPY_DIAG)
{
int n = chameleon_min(A->mt, A->nt) * A->nb;
@@ -321,20 +320,25 @@ int MORSE_zgeqrs_Tile_Async(MORSE_desc_t *A, MORSE_desc_t *T, MORSE_desc_t *B,
#endif
if (morse->householder == MORSE_FLAT_HOUSEHOLDER) {
- morse_pzunmqr(MorseLeft, MorseConjTrans, A, B, T, Dptr, sequence, request);
+ morse_pzunmqr( MorseLeft, MorseConjTrans, A, B, T, Dptr, sequence, request );
}
else {
- morse_pzunmqrrh(MorseLeft, MorseConjTrans, A, B, T, Dptr, MORSE_RHBLK, sequence, request);
+ morse_pzunmqrrh( MorseLeft, MorseConjTrans, A, B, T, Dptr, MORSE_RHBLK, sequence, request );
}
subB = morse_desc_submatrix(B, 0, 0, A->n, B->n);
subA = morse_desc_submatrix(A, 0, 0, A->n, A->n);
- morse_pztrsm(MorseLeft, MorseUpper, MorseNoTrans, MorseNonUnit, 1.0, subA, subB, sequence, request);
+ morse_pztrsm( MorseLeft, MorseUpper, MorseNoTrans, MorseNonUnit, 1.0, subA, subB, sequence, request );
free(subA);
free(subB);
if (Dptr != NULL) {
- morse_desc_mat_free(Dptr);
+ MORSE_Desc_Flush( A, sequence );
+ MORSE_Desc_Flush( B, sequence );
+ MORSE_Desc_Flush( T, sequence );
+ MORSE_Desc_Flush( Dptr, sequence );
+ morse_sequence_wait( morse, sequence );
+ morse_desc_mat_free( Dptr );
}
(void)D;
return MORSE_SUCCESS;
diff --git a/compute/zgeqrs_param.c b/compute/zgeqrs_param.c
index 6e6cb28b394e3b1d815757968cd6dcbf5ae56ae6..cd7f17b6b5ec4601db4ac7457bea27eca6f8fb80 100644
--- a/compute/zgeqrs_param.c
+++ b/compute/zgeqrs_param.c
@@ -76,17 +76,18 @@
* @sa MORSE_zgeqrf
*
*******************************************************************************/
-int MORSE_zgeqrs_param(const libhqr_tree_t *qrtree, int M, int N, int NRHS,
- MORSE_Complex64_t *A, int LDA,
- MORSE_desc_t *descTS, MORSE_desc_t *descTT,
- MORSE_Complex64_t *B, int LDB)
+int MORSE_zgeqrs_param( const libhqr_tree_t *qrtree, int M, int N, int NRHS,
+ MORSE_Complex64_t *A, int LDA,
+ MORSE_desc_t *descTS, MORSE_desc_t *descTT,
+ MORSE_Complex64_t *B, int LDB )
{
int NB;
int status;
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
MORSE_request_t request = MORSE_REQUEST_INITIALIZER;
- MORSE_desc_t descA, descB;
+ MORSE_desc_t descAl, descAt;
+ MORSE_desc_t descBl, descBt;
morse = morse_context_self();
if (morse == NULL) {
@@ -130,37 +131,31 @@ int MORSE_zgeqrs_param(const libhqr_tree_t *qrtree, int M, int N, int NRHS,
/* Set NT */
NB = MORSE_NB;
- morse_sequence_create(morse, &sequence);
+ morse_sequence_create( morse, &sequence );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooplap2tile( descA, A, NB, NB, LDA, N, 0, 0, M, N, sequence, &request,
- morse_desc_mat_free(&(descA)) );
- morse_zooplap2tile( descB, B, NB, NB, LDB, NRHS, 0, 0, M, NRHS, sequence, &request,
- morse_desc_mat_free(&(descA)); morse_desc_mat_free(&(descB)));
-/* } else {*/
-/* morse_ziplap2tile( descA, A, NB, NB, LDA, N, 0, 0, M, N, */
-/* sequence, &request);*/
-/* morse_ziplap2tile( descB, B, NB, NB, LDB, NRHS, 0, 0, M, NRHS,*/
-/* sequence, &request);*/
-/* }*/
+ /* Submit the matrix conversion */
+ morse_zlap2tile( morse, &descAl, &descAt, MorseDescInput, MorseUpperLower,
+ A, NB, NB, LDA, N, M, N, sequence, &request );
+ morse_zlap2tile( morse, &descBl, &descBt, MorseDescInout, MorseUpperLower,
+ B, NB, NB, LDB, NRHS, M, NRHS, sequence, &request );
/* Call the tile interface */
- MORSE_zgeqrs_param_Tile_Async(qrtree, &descA, descTS, descTT, &descB, sequence, &request);
+ MORSE_zgeqrs_param_Tile_Async( qrtree, &descAt, descTS, descTT, &descBt, sequence, &request );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooptile2lap(descA, A, NB, NB, LDA, N, sequence, &request);
- morse_zooptile2lap(descB, B, NB, NB, LDB, NRHS, sequence, &request);
- morse_sequence_wait(morse, sequence);
- morse_desc_mat_free(&descA);
- morse_desc_mat_free(&descB);
-/* } else {*/
-/* morse_ziptile2lap( descA, A, NB, NB, LDA, N, sequence, &request);*/
-/* morse_ziptile2lap( descB, B, NB, NB, LDB, NRHS, sequence, &request);*/
-/* morse_sequence_wait(morse, sequence);*/
-/* }*/
+ /* Submit the matrix conversion back */
+ morse_ztile2lap( morse, &descAl, &descAt,
+ MorseDescInput, MorseUpperLower, sequence, &request );
+ morse_ztile2lap( morse, &descBl, &descBt,
+ MorseDescInout, MorseUpperLower, sequence, &request );
+
+ morse_sequence_wait( morse, sequence );
+
+ /* Cleanup the temporary data */
+ morse_ztile2lap_cleanup( morse, &descAl, &descAt );
+ morse_ztile2lap_cleanup( morse, &descBl, &descBt );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -202,7 +197,8 @@ int MORSE_zgeqrs_param(const libhqr_tree_t *qrtree, int M, int N, int NRHS,
* @sa MORSE_zgeqrf_Tile
*
******************************************************************************/
-int MORSE_zgeqrs_param_Tile(const libhqr_tree_t *qrtree, MORSE_desc_t *A, MORSE_desc_t *TS, MORSE_desc_t *TT, MORSE_desc_t *B)
+int MORSE_zgeqrs_param_Tile( const libhqr_tree_t *qrtree, MORSE_desc_t *A,
+ MORSE_desc_t *TS, MORSE_desc_t *TT, MORSE_desc_t *B )
{
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
@@ -214,16 +210,18 @@ int MORSE_zgeqrs_param_Tile(const libhqr_tree_t *qrtree, MORSE_desc_t *A, MORSE_
morse_fatal_error("MORSE_zgeqrs_param_Tile", "MORSE not initialized");
return MORSE_ERR_NOT_INITIALIZED;
}
- morse_sequence_create(morse, &sequence);
- MORSE_zgeqrs_param_Tile_Async(qrtree, A, TS, TT, B, sequence, &request);
- RUNTIME_desc_flush( A, sequence );
- RUNTIME_desc_flush( TS, sequence );
- RUNTIME_desc_flush( TT, sequence );
- RUNTIME_desc_flush( B, sequence );
- morse_sequence_wait(morse, sequence);
+ morse_sequence_create( morse, &sequence );
+
+ MORSE_zgeqrs_param_Tile_Async( qrtree, A, TS, TT, B, sequence, &request );
+ MORSE_Desc_Flush( A, sequence );
+ MORSE_Desc_Flush( TS, sequence );
+ MORSE_Desc_Flush( TT, sequence );
+ MORSE_Desc_Flush( B, sequence );
+
+ morse_sequence_wait( morse, sequence );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -257,9 +255,9 @@ int MORSE_zgeqrs_param_Tile(const libhqr_tree_t *qrtree, MORSE_desc_t *A, MORSE_
* @sa MORSE_zgeqrf_Tile_Async
*
******************************************************************************/
-int MORSE_zgeqrs_param_Tile_Async(const libhqr_tree_t *qrtree,
- MORSE_desc_t *A, MORSE_desc_t *TS, MORSE_desc_t *TT, MORSE_desc_t *B,
- MORSE_sequence_t *sequence, MORSE_request_t *request)
+int MORSE_zgeqrs_param_Tile_Async( const libhqr_tree_t *qrtree,
+ MORSE_desc_t *A, MORSE_desc_t *TS, MORSE_desc_t *TT, MORSE_desc_t *B,
+ MORSE_sequence_t *sequence, MORSE_request_t *request )
{
MORSE_desc_t *subA;
MORSE_desc_t *subB;
@@ -280,10 +278,12 @@ int MORSE_zgeqrs_param_Tile_Async(const libhqr_tree_t *qrtree,
return MORSE_ERR_UNALLOCATED;
}
/* Check sequence status */
- if (sequence->status == MORSE_SUCCESS)
+ if (sequence->status == MORSE_SUCCESS) {
request->status = MORSE_SUCCESS;
- else
+ }
+ else {
return morse_request_fail(sequence, request, MORSE_ERR_SEQUENCE_FLUSHED);
+ }
/* Check descriptors for correctness */
if (morse_desc_check(A) != MORSE_SUCCESS) {
@@ -324,14 +324,20 @@ int MORSE_zgeqrs_param_Tile_Async(const libhqr_tree_t *qrtree,
subB = morse_desc_submatrix(B, 0, 0, A->n, B->n);
subA = morse_desc_submatrix(A, 0, 0, A->n, A->n);
- morse_pzunmqr_param(qrtree, MorseLeft, MorseConjTrans, A, B, TS, TT, Dptr, sequence, request);
- morse_pztrsm(MorseLeft, MorseUpper, MorseNoTrans, MorseNonUnit, 1.0, subA, subB, sequence, request);
+ morse_pzunmqr_param( qrtree, MorseLeft, MorseConjTrans, A, B, TS, TT, Dptr, sequence, request );
+ morse_pztrsm( MorseLeft, MorseUpper, MorseNoTrans, MorseNonUnit, 1.0, subA, subB, sequence, request );
free(subA);
free(subB);
if (Dptr != NULL) {
- morse_desc_mat_free(Dptr);
+ MORSE_Desc_Flush( A, sequence );
+ MORSE_Desc_Flush( B, sequence );
+ MORSE_Desc_Flush( TS, sequence );
+ MORSE_Desc_Flush( TT, sequence );
+ MORSE_Desc_Flush( Dptr, sequence );
+ morse_sequence_wait( morse, sequence );
+ morse_desc_mat_free( Dptr );
}
(void)D;
return MORSE_SUCCESS;
diff --git a/compute/zgesv_incpiv.c b/compute/zgesv_incpiv.c
index ca0ca1e6a74dfa96cce66e4616ebd1be54f4af4c..700f6b02e960c8723090f567f33107b467f6dddc 100644
--- a/compute/zgesv_incpiv.c
+++ b/compute/zgesv_incpiv.c
@@ -86,17 +86,18 @@
* @sa MORSE_sgesv_incpiv
*
******************************************************************************/
-int MORSE_zgesv_incpiv(int N, int NRHS,
- MORSE_Complex64_t *A, int LDA,
- MORSE_desc_t *descL, int *IPIV,
- MORSE_Complex64_t *B, int LDB)
+int MORSE_zgesv_incpiv( int N, int NRHS,
+ MORSE_Complex64_t *A, int LDA,
+ MORSE_desc_t *descL, int *IPIV,
+ MORSE_Complex64_t *B, int LDB )
{
int NB;
int status;
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
MORSE_request_t request = MORSE_REQUEST_INITIALIZER;
- MORSE_desc_t descA, descB;
+ MORSE_desc_t descAl, descAt;
+ MORSE_desc_t descBl, descBt;
morse = morse_context_self();
if (morse == NULL) {
@@ -134,37 +135,31 @@ int MORSE_zgesv_incpiv(int N, int NRHS,
/* Set NT & NTRHS */
NB = MORSE_NB;
- morse_sequence_create(morse, &sequence);
+ morse_sequence_create( morse, &sequence );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooplap2tile( descA, A, NB, NB, LDA, N, 0, 0, N, N, sequence, &request,
- morse_desc_mat_free(&(descA)) );
- morse_zooplap2tile( descB, B, NB, NB, LDB, NRHS, 0, 0, N, NRHS, sequence, &request,
- morse_desc_mat_free(&(descA)); morse_desc_mat_free(&(descB)));
-/* } else {*/
-/* morse_ziplap2tile( descA, A, NB, NB, LDA, N, 0, 0, N, N, */
-/* sequence, &request);*/
-/* morse_ziplap2tile( descB, B, NB, NB, LDB, NRHS, 0, 0, N, NRHS,*/
-/* sequence, &request);*/
-/* }*/
+ /* Submit the matrix conversion */
+ morse_zlap2tile( morse, &descAl, &descAt, MorseDescInout, MorseUpperLower,
+ A, NB, NB, LDA, N, N, N, sequence, &request );
+ morse_zlap2tile( morse, &descBl, &descBt, MorseDescInout, MorseUpperLower,
+ B, NB, NB, LDB, NRHS, N, NRHS, sequence, &request );
/* Call the tile interface */
- MORSE_zgesv_incpiv_Tile_Async(&descA, descL, IPIV, &descB, sequence, &request);
+ MORSE_zgesv_incpiv_Tile_Async( &descAt, descL, IPIV, &descBt, sequence, &request );
+
+ /* Submit the matrix conversion back */
+ morse_ztile2lap( morse, &descAl, &descAt,
+ MorseDescInout, MorseUpperLower, sequence, &request );
+ morse_ztile2lap( morse, &descBl, &descBt,
+ MorseDescInout, MorseUpperLower, sequence, &request );
+
+ morse_sequence_wait( morse, sequence );
+
+ /* Cleanup the temporary data */
+ morse_ztile2lap_cleanup( morse, &descAl, &descAt );
+ morse_ztile2lap_cleanup( morse, &descBl, &descBt );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooptile2lap(descA, A, NB, NB, LDA, N, sequence, &request);
- morse_zooptile2lap(descB, B, NB, NB, LDB, NRHS, sequence, &request);
- morse_sequence_wait(morse, sequence);
- morse_desc_mat_free(&descA);
- morse_desc_mat_free(&descB);
-/* } else {*/
-/* morse_ziptile2lap( descA, A, NB, NB, LDA, N, sequence, &request);*/
-/* morse_ziptile2lap( descB, B, NB, NB, LDB, NRHS, sequence, &request);*/
-/* morse_sequence_wait(morse, sequence);*/
-/* }*/
-
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
/**
@@ -212,7 +207,7 @@ int MORSE_zgesv_incpiv(int N, int NRHS,
* @sa MORSE_zcgesv_Tile
*
******************************************************************************/
-int MORSE_zgesv_incpiv_Tile(MORSE_desc_t *A, MORSE_desc_t *L, int *IPIV, MORSE_desc_t *B)
+int MORSE_zgesv_incpiv_Tile( MORSE_desc_t *A, MORSE_desc_t *L, int *IPIV, MORSE_desc_t *B )
{
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
@@ -224,14 +219,17 @@ int MORSE_zgesv_incpiv_Tile(MORSE_desc_t *A, MORSE_desc_t *L, int *IPIV, MORSE_d
morse_fatal_error("MORSE_zgesv_incpiv_Tile", "MORSE not initialized");
return MORSE_ERR_NOT_INITIALIZED;
}
- morse_sequence_create(morse, &sequence);
- MORSE_zgesv_incpiv_Tile_Async(A, L, IPIV, B, sequence, &request);
- RUNTIME_desc_flush( A, sequence );
- RUNTIME_desc_flush( B, sequence );
- morse_sequence_wait(morse, sequence);
-
+ morse_sequence_create( morse, &sequence );
+
+ MORSE_zgesv_incpiv_Tile_Async( A, L, IPIV, B, sequence, &request );
+
+ MORSE_Desc_Flush( A, sequence );
+ MORSE_Desc_Flush( L, sequence );
+ MORSE_Desc_Flush( B, sequence );
+
+ morse_sequence_wait( morse, sequence );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -265,8 +263,8 @@ int MORSE_zgesv_incpiv_Tile(MORSE_desc_t *A, MORSE_desc_t *L, int *IPIV, MORSE_d
* @sa MORSE_zcgesv_Tile_Async
*
******************************************************************************/
-int MORSE_zgesv_incpiv_Tile_Async(MORSE_desc_t *A, MORSE_desc_t *L, int *IPIV, MORSE_desc_t *B,
- MORSE_sequence_t *sequence, MORSE_request_t *request)
+int MORSE_zgesv_incpiv_Tile_Async( MORSE_desc_t *A, MORSE_desc_t *L, int *IPIV, MORSE_desc_t *B,
+ MORSE_sequence_t *sequence, MORSE_request_t *request )
{
MORSE_context_t *morse;
@@ -284,10 +282,12 @@ int MORSE_zgesv_incpiv_Tile_Async(MORSE_desc_t *A, MORSE_desc_t *L, int *IPIV, M
return MORSE_ERR_UNALLOCATED;
}
/* Check sequence status */
- if (sequence->status == MORSE_SUCCESS)
+ if (sequence->status == MORSE_SUCCESS) {
request->status = MORSE_SUCCESS;
- else
+ }
+ else {
return morse_request_fail(sequence, request, MORSE_ERR_SEQUENCE_FLUSHED);
+ }
/* Check descriptors for correctness */
if (morse_desc_check(A) != MORSE_SUCCESS) {
@@ -308,16 +308,16 @@ int MORSE_zgesv_incpiv_Tile_Async(MORSE_desc_t *A, MORSE_desc_t *L, int *IPIV, M
return morse_request_fail(sequence, request, MORSE_ERR_ILLEGAL_VALUE);
}
/* Quick return */
-/*
- if (chameleon_min(N, NRHS) == 0)
- return MORSE_SUCCESS;
-*/
+ /*
+ if (chameleon_min(N, NRHS) == 0)
+ return MORSE_SUCCESS;
+ */
- morse_pzgetrf_incpiv(A, L, IPIV, sequence, request);
+ morse_pzgetrf_incpiv( A, L, IPIV, sequence, request );
- morse_pztrsmpl(A, B, L, IPIV, sequence, request);
+ morse_pztrsmpl( A, B, L, IPIV, sequence, request );
- morse_pztrsm(MorseLeft, MorseUpper, MorseNoTrans, MorseNonUnit, (MORSE_Complex64_t)1.0, A, B, sequence, request);
+ morse_pztrsm( MorseLeft, MorseUpper, MorseNoTrans, MorseNonUnit, (MORSE_Complex64_t)1.0, A, B, sequence, request );
return MORSE_SUCCESS;
}
diff --git a/compute/zgesv_nopiv.c b/compute/zgesv_nopiv.c
index bc0d9bdf9d3fd297954c2a6cc14d692c59002705..0a5753ad3753d7cef81d1f2a6f91be709d16df02 100644
--- a/compute/zgesv_nopiv.c
+++ b/compute/zgesv_nopiv.c
@@ -86,16 +86,17 @@
* @sa MORSE_sgesv_nopiv
*
******************************************************************************/
-int MORSE_zgesv_nopiv(int N, int NRHS,
- MORSE_Complex64_t *A, int LDA,
- MORSE_Complex64_t *B, int LDB)
+int MORSE_zgesv_nopiv( int N, int NRHS,
+ MORSE_Complex64_t *A, int LDA,
+ MORSE_Complex64_t *B, int LDB )
{
int NB;
int status;
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
MORSE_request_t request = MORSE_REQUEST_INITIALIZER;
- MORSE_desc_t descA, descB;
+ MORSE_desc_t descAl, descAt;
+ MORSE_desc_t descBl, descBt;
morse = morse_context_self();
if (morse == NULL) {
@@ -133,37 +134,31 @@ int MORSE_zgesv_nopiv(int N, int NRHS,
/* Set NT & NTRHS */
NB = MORSE_NB;
- morse_sequence_create(morse, &sequence);
+ morse_sequence_create( morse, &sequence );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooplap2tile( descA, A, NB, NB, LDA, N, 0, 0, N, N, sequence, &request,
- morse_desc_mat_free(&(descA)) );
- morse_zooplap2tile( descB, B, NB, NB, LDB, NRHS, 0, 0, N, NRHS, sequence, &request,
- morse_desc_mat_free(&(descA)); morse_desc_mat_free(&(descB)));
-/* } else {*/
-/* morse_ziplap2tile( descA, A, NB, NB, LDA, N, 0, 0, N, N, */
-/* sequence, &request);*/
-/* morse_ziplap2tile( descB, B, NB, NB, LDB, NRHS, 0, 0, N, NRHS,*/
-/* sequence, &request);*/
-/* }*/
+ /* Submit the matrix conversion */
+ morse_zlap2tile( morse, &descAl, &descAt, MorseDescInout, MorseUpperLower,
+ A, NB, NB, LDA, N, N, N, sequence, &request );
+ morse_zlap2tile( morse, &descBl, &descBt, MorseDescInout, MorseUpperLower,
+ B, NB, NB, LDB, NRHS, N, NRHS, sequence, &request );
/* Call the tile interface */
- MORSE_zgesv_nopiv_Tile_Async(&descA, &descB, sequence, &request);
+ MORSE_zgesv_nopiv_Tile_Async( &descAt, &descBt, sequence, &request );
+
+ /* Submit the matrix conversion back */
+ morse_ztile2lap( morse, &descAl, &descAt,
+ MorseDescInout, MorseUpperLower, sequence, &request );
+ morse_ztile2lap( morse, &descBl, &descBt,
+ MorseDescInout, MorseUpperLower, sequence, &request );
+
+ morse_sequence_wait( morse, sequence );
+
+ /* Cleanup the temporary data */
+ morse_ztile2lap_cleanup( morse, &descAl, &descAt );
+ morse_ztile2lap_cleanup( morse, &descBl, &descBt );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooptile2lap(descA, A, NB, NB, LDA, N, sequence, &request);
- morse_zooptile2lap(descB, B, NB, NB, LDB, NRHS, sequence, &request);
- morse_sequence_wait(morse, sequence);
- morse_desc_mat_free(&descA);
- morse_desc_mat_free(&descB);
-/* } else {*/
-/* morse_ziptile2lap( descA, A, NB, NB, LDA, N, sequence, &request);*/
-/* morse_ziptile2lap( descB, B, NB, NB, LDB, NRHS, sequence, &request);*/
-/* morse_sequence_wait(morse, sequence);*/
-/* }*/
-
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
/**
@@ -204,7 +199,7 @@ int MORSE_zgesv_nopiv(int N, int NRHS,
* @sa MORSE_zcgesv_Tile
*
******************************************************************************/
-int MORSE_zgesv_nopiv_Tile(MORSE_desc_t *A, MORSE_desc_t *B)
+int MORSE_zgesv_nopiv_Tile( MORSE_desc_t *A, MORSE_desc_t *B )
{
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
@@ -216,14 +211,16 @@ int MORSE_zgesv_nopiv_Tile(MORSE_desc_t *A, MORSE_desc_t *B)
morse_fatal_error("MORSE_zgesv_nopiv_Tile", "MORSE not initialized");
return MORSE_ERR_NOT_INITIALIZED;
}
- morse_sequence_create(morse, &sequence);
- MORSE_zgesv_nopiv_Tile_Async(A, B, sequence, &request);
- RUNTIME_desc_flush( A, sequence );
- RUNTIME_desc_flush( B, sequence );
- morse_sequence_wait(morse, sequence);
-
+ morse_sequence_create( morse, &sequence );
+
+ MORSE_zgesv_nopiv_Tile_Async( A, B, sequence, &request );
+
+ MORSE_Desc_Flush( A, sequence );
+ MORSE_Desc_Flush( B, sequence );
+
+ morse_sequence_wait( morse, sequence );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -257,8 +254,8 @@ int MORSE_zgesv_nopiv_Tile(MORSE_desc_t *A, MORSE_desc_t *B)
* @sa MORSE_zcgesv_Tile_Async
*
******************************************************************************/
-int MORSE_zgesv_nopiv_Tile_Async(MORSE_desc_t *A, MORSE_desc_t *B,
- MORSE_sequence_t *sequence, MORSE_request_t *request)
+int MORSE_zgesv_nopiv_Tile_Async( MORSE_desc_t *A, MORSE_desc_t *B,
+ MORSE_sequence_t *sequence, MORSE_request_t *request )
{
MORSE_context_t *morse;
@@ -276,10 +273,12 @@ int MORSE_zgesv_nopiv_Tile_Async(MORSE_desc_t *A, MORSE_desc_t *B,
return MORSE_ERR_UNALLOCATED;
}
/* Check sequence status */
- if (sequence->status == MORSE_SUCCESS)
+ if (sequence->status == MORSE_SUCCESS) {
request->status = MORSE_SUCCESS;
- else
+ }
+ else {
return morse_request_fail(sequence, request, MORSE_ERR_SEQUENCE_FLUSHED);
+ }
/* Check descriptors for correctness */
if (morse_desc_check(A) != MORSE_SUCCESS) {
@@ -296,16 +295,16 @@ int MORSE_zgesv_nopiv_Tile_Async(MORSE_desc_t *A, MORSE_desc_t *B,
return morse_request_fail(sequence, request, MORSE_ERR_ILLEGAL_VALUE);
}
/* Quick return */
-/*
- if (chameleon_min(N, NRHS) == 0)
- return MORSE_SUCCESS;
-*/
+ /*
+ if (chameleon_min(N, NRHS) == 0)
+ return MORSE_SUCCESS;
+ */
- morse_pzgetrf_nopiv(A, sequence, request);
+ morse_pzgetrf_nopiv( A, sequence, request );
- morse_pztrsm(MorseLeft, MorseLower, MorseNoTrans, MorseUnit, (MORSE_Complex64_t)1.0, A, B, sequence, request);
+ morse_pztrsm( MorseLeft, MorseLower, MorseNoTrans, MorseUnit, (MORSE_Complex64_t)1.0, A, B, sequence, request );
- morse_pztrsm(MorseLeft, MorseUpper, MorseNoTrans, MorseNonUnit, (MORSE_Complex64_t)1.0, A, B, sequence, request);
+ morse_pztrsm( MorseLeft, MorseUpper, MorseNoTrans, MorseNonUnit, (MORSE_Complex64_t)1.0, A, B, sequence, request );
return MORSE_SUCCESS;
}
diff --git a/compute/zgesvd.c b/compute/zgesvd.c
index 968c0a01b427303b55a50c8574231d407222a849..ba35c4da4e0164dbe0bb319ffea38209cbaa41b1 100644
--- a/compute/zgesvd.c
+++ b/compute/zgesvd.c
@@ -147,20 +147,20 @@
* @sa MORSE_sgesvd
*
******************************************************************************/
-int MORSE_zgesvd(MORSE_enum jobu, MORSE_enum jobvt,
+int MORSE_zgesvd( MORSE_enum jobu, MORSE_enum jobvt,
int M, int N,
MORSE_Complex64_t *A, int LDA,
double *S,
MORSE_desc_t *descT,
MORSE_Complex64_t *U, int LDU,
- MORSE_Complex64_t *VT, int LDVT)
+ MORSE_Complex64_t *VT, int LDVT )
{
int NB;
int status;
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
MORSE_request_t request = MORSE_REQUEST_INITIALIZER;
- MORSE_desc_t descA;
+ MORSE_desc_t descAl, descAt;
morse = morse_context_self();
if (morse == NULL) {
@@ -212,30 +212,26 @@ int MORSE_zgesvd(MORSE_enum jobu, MORSE_enum jobvt,
/* Set MT, NT */
NB = MORSE_NB;
- morse_sequence_create(morse, &sequence);
+ morse_sequence_create( morse, &sequence );
- /* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) { */
- morse_zooplap2tile( descA, A, NB, NB, LDA, N, 0, 0, M, N, sequence, &request,
- morse_desc_mat_free(&(descA)) );
- /* } else { */
- /* morse_ziplap2tile( descA, A, NB, NB, LDA, N, 0, 0, M, N, */
- /* sequence, &request); */
- /* } */
+ /* Submit the matrix conversion */
+ morse_zlap2tile( morse, &descAl, &descAt, MorseDescInout, MorseUpperLower,
+ A, NB, NB, LDA, N, M, N, sequence, &request );
/* Call the tile interface */
- MORSE_zgesvd_Tile_Async(jobu, jobvt, &descA, S, descT, U, LDU, VT, LDVT, sequence, &request);
+ MORSE_zgesvd_Tile_Async( jobu, jobvt, &descAt, S, descT, U, LDU, VT, LDVT, sequence, &request );
+
+ /* Submit the matrix conversion back */
+ morse_ztile2lap( morse, &descAl, &descAt,
+ MorseDescInout, MorseUpperLower, sequence, &request );
- /* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) { */
- morse_zooptile2lap( descA, A, NB, NB, LDA, N, sequence, &request);
- morse_sequence_wait(morse, sequence);
- morse_desc_mat_free(&descA);
- /* } else { */
- /* morse_ziptile2lap( descA, A, NB, NB, LDA, N, sequence, &request); */
- /* morse_sequence_wait(morse, sequence); */
- /* } */
+ morse_sequence_wait( morse, sequence );
+
+ /* Cleanup the temporary data */
+ morse_ztile2lap_cleanup( morse, &descAl, &descAt );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -339,16 +335,16 @@ int MORSE_zgesvd(MORSE_enum jobu, MORSE_enum jobvt,
* @sa MORSE_sgesvd_Tile
*
******************************************************************************/
-int MORSE_zgesvd_Tile(MORSE_enum jobu, MORSE_enum jobvt,
- MORSE_desc_t *A,
- double *S,
- MORSE_desc_t *T,
- MORSE_Complex64_t *U, int LDU,
- MORSE_Complex64_t *VT, int LDVT)
+int MORSE_zgesvd_Tile( MORSE_enum jobu, MORSE_enum jobvt,
+ MORSE_desc_t *A,
+ double *S,
+ MORSE_desc_t *T,
+ MORSE_Complex64_t *U, int LDU,
+ MORSE_Complex64_t *VT, int LDVT )
{
- MORSE_context_t *morse;
+ MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
- MORSE_request_t request = MORSE_REQUEST_INITIALIZER;
+ MORSE_request_t request = MORSE_REQUEST_INITIALIZER;
int status;
morse = morse_context_self();
@@ -356,11 +352,16 @@ int MORSE_zgesvd_Tile(MORSE_enum jobu, MORSE_enum jobvt,
morse_fatal_error("MORSE_zgesvd_Tile", "MORSE not initialized");
return MORSE_ERR_NOT_INITIALIZED;
}
- morse_sequence_create(morse, &sequence);
- MORSE_zgesvd_Tile_Async(jobu, jobvt, A, S, T, U, LDU, VT, LDVT, sequence, &request);
- morse_sequence_wait(morse, sequence);
+ morse_sequence_create( morse, &sequence );
+
+ MORSE_zgesvd_Tile_Async( jobu, jobvt, A, S, T, U, LDU, VT, LDVT, sequence, &request );
+
+ MORSE_Desc_Flush( A, sequence );
+ MORSE_Desc_Flush( T, sequence );
+
+ morse_sequence_wait( morse, sequence );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -394,17 +395,18 @@ int MORSE_zgesvd_Tile(MORSE_enum jobu, MORSE_enum jobvt,
* @sa MORSE_sgesvd_Tile_Async
*
******************************************************************************/
-int MORSE_zgesvd_Tile_Async(MORSE_enum jobu, MORSE_enum jobvt,
- MORSE_desc_t *A,
- double *S,
- MORSE_desc_t *T,
- MORSE_Complex64_t *U, int LDU,
- MORSE_Complex64_t *VT, int LDVT,
- MORSE_sequence_t *sequence, MORSE_request_t *request)
+int MORSE_zgesvd_Tile_Async( MORSE_enum jobu, MORSE_enum jobvt,
+ MORSE_desc_t *A,
+ double *S,
+ MORSE_desc_t *T,
+ MORSE_Complex64_t *U, int LDU,
+ MORSE_Complex64_t *VT, int LDVT,
+ MORSE_sequence_t *sequence, MORSE_request_t *request )
{
MORSE_desc_t descA;
MORSE_desc_t descT;
- MORSE_desc_t descU, descVT;
+ MORSE_desc_t descU;
+ MORSE_desc_t descVT;
MORSE_desc_t descAB;
MORSE_desc_t D, *Dptr = NULL;
MORSE_desc_t *subA, *subT, *subUVT;
@@ -430,10 +432,12 @@ int MORSE_zgesvd_Tile_Async(MORSE_enum jobu, MORSE_enum jobvt,
return MORSE_ERR_UNALLOCATED;
}
/* Check sequence status */
- if (sequence->status == MORSE_SUCCESS)
+ if (sequence->status == MORSE_SUCCESS) {
request->status = MORSE_SUCCESS;
- else
+ }
+ else {
return morse_request_fail(sequence, request, MORSE_ERR_SEQUENCE_FLUSHED);
+ }
/* Check descriptors for correctness */
if (morse_desc_check(A) != MORSE_SUCCESS) {
@@ -534,7 +538,7 @@ int MORSE_zgesvd_Tile_Async(MORSE_enum jobu, MORSE_enum jobvt,
}
}
- morse_sequence_wait(morse, sequence);
+ morse_sequence_wait( morse, sequence );
#if !defined(CHAMELEON_SIMULATION)
info = LAPACKE_zgbbrd( LAPACK_COL_MAJOR,
@@ -550,16 +554,17 @@ int MORSE_zgesvd_Tile_Async(MORSE_enum jobu, MORSE_enum jobvt,
fprintf(stderr, "MORSE_zgesvd_Tile_Async: LAPACKE_zgbbrd = %d\n", info );
}
#endif /* !defined(CHAMELEON_SIMULATION) */
- morse_desc_mat_free(&descAB);
+ morse_desc_mat_free( &descAB );
/* Transform U and Vt into tile format */
if ( jobu != MorseNoVec ) {
- morse_zooplap2tile( descU, U, NB, NB, LDU, M, 0, 0, M, M, sequence, request, morse_desc_mat_free(&(descU)) );
+ //morse_zooplap2tile( descU, U, NB, NB, LDU, M, 0, 0, M, M, sequence, request, morse_desc_mat_free(&(descU)) );
}
if ( jobvt != MorseNoVec ) {
- morse_zooplap2tile( descVT, VT, NB, NB, LDVT, N, 0, 0, N, N, sequence, request, morse_desc_mat_free(&(descVT)) );
+ //morse_zooplap2tile( descVT, VT, NB, NB, LDVT, N, 0, 0, N, N, sequence, request, morse_desc_mat_free(&(descVT)) );
}
- morse_sequence_wait(morse, sequence);
+
+ morse_sequence_wait( morse, sequence );
subA = NULL;
subT = NULL;
@@ -598,13 +603,14 @@ int MORSE_zgesvd_Tile_Async(MORSE_enum jobu, MORSE_enum jobvt,
/* Transform U and VT into lapack layout */
if ( jobu != MorseNoVec ) {
- morse_zooptile2lap( descU, U, NB, NB, LDU, M, sequence, request );
+ //morse_zooptile2lap( descU, U, NB, NB, LDU, M, sequence, request );
}
if ( jobvt != MorseNoVec ) {
- morse_zooptile2lap( descVT, VT, NB, NB, LDVT, N, sequence, request );
+ //morse_zooptile2lap( descVT, VT, NB, NB, LDVT, N, sequence, request );
}
- morse_sequence_wait(morse, sequence);
+ morse_sequence_wait( morse, sequence );
+
if (subA) {
free(subA); free(subUVT); free(subT);
}
@@ -627,7 +633,7 @@ int MORSE_zgesvd_Tile_Async(MORSE_enum jobu, MORSE_enum jobvt,
morse_desc_mat_free( &descVT );
free(E);
if (Dptr != NULL) {
- morse_desc_mat_free(Dptr);
+ morse_desc_mat_free( Dptr );
}
(void)D;
return MORSE_SUCCESS;
diff --git a/compute/zgetrf_incpiv.c b/compute/zgetrf_incpiv.c
index aaf96f147bfdf88424c2200fd29eb645e005aa47..a86915d081d5b92b07269f534348dc985c023f7d 100644
--- a/compute/zgetrf_incpiv.c
+++ b/compute/zgetrf_incpiv.c
@@ -78,16 +78,16 @@
* @sa MORSE_zgetrs_incpiv
*
******************************************************************************/
-int MORSE_zgetrf_incpiv(int M, int N,
- MORSE_Complex64_t *A, int LDA,
- MORSE_desc_t *descL, int *IPIV)
+int MORSE_zgetrf_incpiv( int M, int N,
+ MORSE_Complex64_t *A, int LDA,
+ MORSE_desc_t *descL, int *IPIV )
{
int NB;
int status;
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
MORSE_request_t request = MORSE_REQUEST_INITIALIZER;
- MORSE_desc_t descA;
+ MORSE_desc_t descAl, descAt;
morse = morse_context_self();
if (morse == NULL) {
@@ -121,30 +121,26 @@ int MORSE_zgetrf_incpiv(int M, int N,
/* Set NT & NTRHS */
NB = MORSE_NB;
- morse_sequence_create(morse, &sequence);
+ morse_sequence_create( morse, &sequence );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooplap2tile( descA, A, NB, NB, LDA, N, 0, 0, M, N, sequence, &request,
- morse_desc_mat_free(&(descA)) );
-/* } else {*/
-/* morse_ziplap2tile( descA, A, NB, NB, LDA, N, 0, 0, M, N,*/
-/* sequence, &request);*/
-/* }*/
+ /* Submit the matrix conversion */
+ morse_zlap2tile( morse, &descAl, &descAt, MorseDescInout, MorseUpperLower,
+ A, NB, NB, LDA, N, M, N, sequence, &request );
/* Call the tile interface */
- MORSE_zgetrf_incpiv_Tile_Async(&descA, descL, IPIV, sequence, &request);
+ MORSE_zgetrf_incpiv_Tile_Async( &descAt, descL, IPIV, sequence, &request );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooptile2lap(descA, A, NB, NB, LDA, N, sequence, &request);
- morse_sequence_wait(morse, sequence);
- morse_desc_mat_free(&descA);
-/* } else {*/
-/* morse_ziptile2lap( descA, A, NB, NB, LDA, N, sequence, &request);*/
-/* morse_sequence_wait(morse, sequence);*/
-/* }*/
+ /* Submit the matrix conversion back */
+ morse_ztile2lap( morse, &descAl, &descAt,
+ MorseDescInout, MorseUpperLower, sequence, &request );
+
+ morse_sequence_wait( morse, sequence );
+
+ /* Cleanup the temporary data */
+ morse_ztile2lap_cleanup( morse, &descAl, &descAt );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -190,7 +186,7 @@ int MORSE_zgetrf_incpiv(int M, int N,
* @sa MORSE_zgetrs_incpiv_Tile
*
******************************************************************************/
-int MORSE_zgetrf_incpiv_Tile(MORSE_desc_t *A, MORSE_desc_t *L, int *IPIV)
+int MORSE_zgetrf_incpiv_Tile( MORSE_desc_t *A, MORSE_desc_t *L, int *IPIV )
{
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
@@ -202,13 +198,16 @@ int MORSE_zgetrf_incpiv_Tile(MORSE_desc_t *A, MORSE_desc_t *L, int *IPIV)
morse_fatal_error("MORSE_zgetrf_incpiv_Tile", "MORSE not initialized");
return MORSE_ERR_NOT_INITIALIZED;
}
- morse_sequence_create(morse, &sequence);
- MORSE_zgetrf_incpiv_Tile_Async(A, L, IPIV, sequence, &request);
- RUNTIME_desc_flush( A, sequence );
- morse_sequence_wait(morse, sequence);
+ morse_sequence_create( morse, &sequence );
+
+ MORSE_zgetrf_incpiv_Tile_Async( A, L, IPIV, sequence, &request );
+
+ MORSE_Desc_Flush( A, sequence );
+ MORSE_Desc_Flush( L, sequence );
+ morse_sequence_wait( morse, sequence );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -241,8 +240,8 @@ int MORSE_zgetrf_incpiv_Tile(MORSE_desc_t *A, MORSE_desc_t *L, int *IPIV)
* @sa MORSE_zgetrs_incpiv_Tile_Async
*
******************************************************************************/
-int MORSE_zgetrf_incpiv_Tile_Async(MORSE_desc_t *A, MORSE_desc_t *L, int *IPIV,
- MORSE_sequence_t *sequence, MORSE_request_t *request)
+int MORSE_zgetrf_incpiv_Tile_Async( MORSE_desc_t *A, MORSE_desc_t *L, int *IPIV,
+ MORSE_sequence_t *sequence, MORSE_request_t *request )
{
MORSE_context_t *morse;
@@ -260,10 +259,12 @@ int MORSE_zgetrf_incpiv_Tile_Async(MORSE_desc_t *A, MORSE_desc_t *L, int *IPIV,
return MORSE_ERR_UNALLOCATED;
}
/* Check sequence status */
- if (sequence->status == MORSE_SUCCESS)
+ if (sequence->status == MORSE_SUCCESS) {
request->status = MORSE_SUCCESS;
- else
+ }
+ else {
return morse_request_fail(sequence, request, MORSE_ERR_SEQUENCE_FLUSHED);
+ }
/* Check descriptors for correctness */
if (morse_desc_check(A) != MORSE_SUCCESS) {
@@ -280,12 +281,12 @@ int MORSE_zgetrf_incpiv_Tile_Async(MORSE_desc_t *A, MORSE_desc_t *L, int *IPIV,
return morse_request_fail(sequence, request, MORSE_ERR_ILLEGAL_VALUE);
}
/* Quick return */
-/*
- if (chameleon_min(M, N) == 0)
- return MORSE_SUCCESS;
-*/
+ /*
+ if (chameleon_min(M, N) == 0)
+ return MORSE_SUCCESS;
+ */
- morse_pzgetrf_incpiv(A, L, IPIV, sequence, request);
+ morse_pzgetrf_incpiv( A, L, IPIV, sequence, request );
return MORSE_SUCCESS;
}
diff --git a/compute/zgetrf_nopiv.c b/compute/zgetrf_nopiv.c
index 5e56c4a8b6bd8dffd399385ebc61bcf2ac52ea36..cf44857a8df9ac437e6b9bd4ef6b5c84de571d94 100644
--- a/compute/zgetrf_nopiv.c
+++ b/compute/zgetrf_nopiv.c
@@ -72,12 +72,12 @@
* @sa MORSE_sgetrf_nopiv
*
******************************************************************************/
-int MORSE_zgetrf_nopiv(int M, int N,
- MORSE_Complex64_t *A, int LDA)
+int MORSE_zgetrf_nopiv( int M, int N,
+ MORSE_Complex64_t *A, int LDA )
{
int NB;
int status;
- MORSE_desc_t descA ;
+ MORSE_desc_t descAl, descAt;
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
MORSE_request_t request = MORSE_REQUEST_INITIALIZER;
@@ -114,32 +114,26 @@ int MORSE_zgetrf_nopiv(int M, int N,
/* Set NT & NTRHS */
NB = MORSE_NB;
- morse_sequence_create(morse, &sequence);
+ morse_sequence_create( morse, &sequence );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooplap2tile( descA, A, NB, NB, LDA, N, 0, 0, N, N,
- sequence, &request,
- morse_desc_mat_free(&(descA)) );
-/* } else {*/
-/* morse_ziplap2tile( descA, A, NB, NB, LDA, N, 0, 0, N, N,*/
-/* sequence, &request);*/
-/* }*/
+ /* Submit the matrix conversion */
+ morse_zlap2tile( morse, &descAl, &descAt, MorseDescInout, MorseUpperLower,
+ A, NB, NB, LDA, N, N, N, sequence, &request );
/* Call the tile interface */
- MORSE_zgetrf_nopiv_Tile_Async(&descA, sequence, &request);
+ MORSE_zgetrf_nopiv_Tile_Async( &descAt, sequence, &request );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooptile2lap(descA, A, NB, NB, LDA, N, sequence, &request );
- morse_sequence_wait(morse, sequence);
- morse_desc_mat_free(&descA);
-/* } else {*/
-/* morse_ziptile2lap( descA, A, NB, NB, LDA, N,*/
-/* sequence, &request );*/
-/* morse_sequence_wait(morse, sequence);*/
-/* }*/
+ /* Submit the matrix conversion back */
+ morse_ztile2lap( morse, &descAl, &descAt,
+ MorseDescInout, MorseUpperLower, sequence, &request );
+
+ morse_sequence_wait( morse, sequence );
+
+ /* Cleanup the temporary data */
+ morse_ztile2lap_cleanup( morse, &descAl, &descAt );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -178,7 +172,7 @@ int MORSE_zgetrf_nopiv(int M, int N,
* @sa MORSE_zgetrs_Tile
*
******************************************************************************/
-int MORSE_zgetrf_nopiv_Tile(MORSE_desc_t *A)
+int MORSE_zgetrf_nopiv_Tile( MORSE_desc_t *A )
{
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
@@ -190,13 +184,15 @@ int MORSE_zgetrf_nopiv_Tile(MORSE_desc_t *A)
morse_fatal_error("MORSE_zgetrf_nopiv_Tile", "MORSE not initialized");
return MORSE_ERR_NOT_INITIALIZED;
}
- morse_sequence_create(morse, &sequence);
- MORSE_zgetrf_nopiv_Tile_Async(A, sequence, &request);
- RUNTIME_desc_flush( A, sequence );
- morse_sequence_wait(morse, sequence);
-
+ morse_sequence_create( morse, &sequence );
+
+ MORSE_zgetrf_nopiv_Tile_Async( A, sequence, &request );
+
+ MORSE_Desc_Flush( A, sequence );
+
+ morse_sequence_wait( morse, sequence );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -233,9 +229,9 @@ int MORSE_zgetrf_nopiv_Tile(MORSE_desc_t *A)
* @sa MORSE_zgetrs_Tile_Async
*
******************************************************************************/
-int MORSE_zgetrf_nopiv_Tile_Async(MORSE_desc_t *A,
- MORSE_sequence_t *sequence,
- MORSE_request_t *request)
+int MORSE_zgetrf_nopiv_Tile_Async( MORSE_desc_t *A,
+ MORSE_sequence_t *sequence,
+ MORSE_request_t *request )
{
MORSE_context_t *morse;
@@ -253,10 +249,12 @@ int MORSE_zgetrf_nopiv_Tile_Async(MORSE_desc_t *A,
return MORSE_ERR_UNALLOCATED;
}
/* Check sequence status */
- if (sequence->status == MORSE_SUCCESS)
+ if (sequence->status == MORSE_SUCCESS) {
request->status = MORSE_SUCCESS;
- else
+ }
+ else {
return morse_request_fail(sequence, request, MORSE_ERR_SEQUENCE_FLUSHED);
+ }
/* Check descriptors for correctness */
if (morse_desc_check(A) != MORSE_SUCCESS) {
@@ -270,7 +268,7 @@ int MORSE_zgetrf_nopiv_Tile_Async(MORSE_desc_t *A,
return morse_request_fail(sequence, request, MORSE_ERR_ILLEGAL_VALUE);
}
- morse_pzgetrf_nopiv(A, sequence, request);
+ morse_pzgetrf_nopiv( A, sequence, request );
return MORSE_SUCCESS;
}
diff --git a/compute/zgetrs_incpiv.c b/compute/zgetrs_incpiv.c
index 195b0895d444eb00ca8e7b5ad70bff8bd06a5800..f877c8aa38b8d1ab39c86809842803384e53ba85 100644
--- a/compute/zgetrs_incpiv.c
+++ b/compute/zgetrs_incpiv.c
@@ -88,17 +88,18 @@
* @sa MORSE_zgetrf_incpiv
*
******************************************************************************/
-int MORSE_zgetrs_incpiv(MORSE_enum trans, int N, int NRHS,
- MORSE_Complex64_t *A, int LDA,
- MORSE_desc_t *descL, int *IPIV,
- MORSE_Complex64_t *B, int LDB)
+int MORSE_zgetrs_incpiv( MORSE_enum trans, int N, int NRHS,
+ MORSE_Complex64_t *A, int LDA,
+ MORSE_desc_t *descL, int *IPIV,
+ MORSE_Complex64_t *B, int LDB )
{
int NB;
int status;
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
MORSE_request_t request = MORSE_REQUEST_INITIALIZER;
- MORSE_desc_t descA, descB;
+ MORSE_desc_t descAl, descAt;
+ MORSE_desc_t descBl, descBt;
morse = morse_context_self();
if (morse == NULL) {
@@ -140,36 +141,31 @@ int MORSE_zgetrs_incpiv(MORSE_enum trans, int N, int NRHS,
/* Set NT & NTRHS */
NB = MORSE_NB;
- morse_sequence_create(morse, &sequence);
+ morse_sequence_create( morse, &sequence );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooplap2tile( descA, A, NB, NB, LDA, N, 0, 0, N, N, sequence, &request,
- morse_desc_mat_free(&(descA)) );
- morse_zooplap2tile( descB, B, NB, NB, LDB, NRHS, 0, 0, N, NRHS, sequence, &request,
- morse_desc_mat_free(&(descA)); morse_desc_mat_free(&(descB)));
-/* } else {*/
-/* morse_ziplap2tile( descA, A, NB, NB, LDA, N, 0, 0, N, N, */
-/* sequence, &request);*/
-/* morse_ziplap2tile( descB, B, NB, NB, LDB, NRHS, 0, 0, N, NRHS,*/
-/* sequence, &request);*/
-/* }*/
+ /* Submit the matrix conversion */
+ morse_zlap2tile( morse, &descAl, &descAt, MorseDescInput, MorseUpperLower,
+ A, NB, NB, LDA, N, N, N, sequence, &request );
+ morse_zlap2tile( morse, &descBl, &descBt, MorseDescInout, MorseUpperLower,
+ B, NB, NB, LDB, NRHS, N, NRHS, sequence, &request );
/* Call the tile interface */
- MORSE_zgetrs_incpiv_Tile_Async(&descA, descL, IPIV, &descB, sequence, &request);
+ MORSE_zgetrs_incpiv_Tile_Async( &descAt, descL, IPIV, &descBt, sequence, &request );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooptile2lap(descB, B, NB, NB, LDB, NRHS, sequence, &request);
- morse_sequence_wait(morse, sequence);
- morse_desc_mat_free(&descA);
- morse_desc_mat_free(&descB);
-/* } else {*/
-/* morse_ziptile2lap( descA, A, NB, NB, LDA, N, sequence, &request);*/
-/* morse_ziptile2lap( descB, B, NB, NB, LDB, NRHS, sequence, &request);*/
-/* morse_sequence_wait(morse, sequence);*/
-/* }*/
+ /* Submit the matrix conversion back */
+ morse_ztile2lap( morse, &descAl, &descAt,
+ MorseDescInput, MorseUpperLower, sequence, &request );
+ morse_ztile2lap( morse, &descBl, &descBt,
+ MorseDescInout, MorseUpperLower, sequence, &request );
+
+ morse_sequence_wait( morse, sequence );
+
+ /* Cleanup the temporary data */
+ morse_ztile2lap_cleanup( morse, &descAl, &descAt );
+ morse_ztile2lap_cleanup( morse, &descBl, &descBt );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -215,7 +211,7 @@ int MORSE_zgetrs_incpiv(MORSE_enum trans, int N, int NRHS,
* @sa MORSE_zgetrf_incpiv_Tile
*
******************************************************************************/
-int MORSE_zgetrs_incpiv_Tile(MORSE_desc_t *A, MORSE_desc_t *L, int *IPIV, MORSE_desc_t *B)
+int MORSE_zgetrs_incpiv_Tile( MORSE_desc_t *A, MORSE_desc_t *L, int *IPIV, MORSE_desc_t *B )
{
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
@@ -227,14 +223,17 @@ int MORSE_zgetrs_incpiv_Tile(MORSE_desc_t *A, MORSE_desc_t *L, int *IPIV, MORSE_
morse_fatal_error("MORSE_zgetrs_incpiv_Tile", "MORSE not initialized");
return MORSE_ERR_NOT_INITIALIZED;
}
- morse_sequence_create(morse, &sequence);
- MORSE_zgetrs_incpiv_Tile_Async(A, L, IPIV, B, sequence, &request);
- RUNTIME_desc_flush( A, sequence );
- RUNTIME_desc_flush( B, sequence );
- morse_sequence_wait(morse, sequence);
-
+ morse_sequence_create( morse, &sequence );
+
+ MORSE_zgetrs_incpiv_Tile_Async( A, L, IPIV, B, sequence, &request );
+
+ MORSE_Desc_Flush( A, sequence );
+ MORSE_Desc_Flush( L, sequence );
+ MORSE_Desc_Flush( B, sequence );
+
+ morse_sequence_wait( morse, sequence );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -268,8 +267,8 @@ int MORSE_zgetrs_incpiv_Tile(MORSE_desc_t *A, MORSE_desc_t *L, int *IPIV, MORSE_
* @sa MORSE_zgetrf_incpiv_Tile_Async
*
******************************************************************************/
-int MORSE_zgetrs_incpiv_Tile_Async(MORSE_desc_t *A, MORSE_desc_t *L, int *IPIV, MORSE_desc_t *B,
- MORSE_sequence_t *sequence, MORSE_request_t *request)
+int MORSE_zgetrs_incpiv_Tile_Async( MORSE_desc_t *A, MORSE_desc_t *L, int *IPIV, MORSE_desc_t *B,
+ MORSE_sequence_t *sequence, MORSE_request_t *request )
{
MORSE_context_t *morse;
@@ -287,10 +286,12 @@ int MORSE_zgetrs_incpiv_Tile_Async(MORSE_desc_t *A, MORSE_desc_t *L, int *IPIV,
return MORSE_ERR_UNALLOCATED;
}
/* Check sequence status */
- if (sequence->status == MORSE_SUCCESS)
+ if (sequence->status == MORSE_SUCCESS) {
request->status = MORSE_SUCCESS;
- else
+ }
+ else {
return morse_request_fail(sequence, request, MORSE_ERR_SEQUENCE_FLUSHED);
+ }
/* Check descriptors for correctness */
if (morse_desc_check(A) != MORSE_SUCCESS) {
@@ -311,13 +312,13 @@ int MORSE_zgetrs_incpiv_Tile_Async(MORSE_desc_t *A, MORSE_desc_t *L, int *IPIV,
return morse_request_fail(sequence, request, MORSE_ERR_ILLEGAL_VALUE);
}
/* Quick return */
-/*
- if (chameleon_min(N, NRHS) == 0)
- return MORSE_SUCCESS;
-*/
- morse_pztrsmpl(A, B, L, IPIV, sequence, request);
+ /*
+ if (chameleon_min(N, NRHS) == 0)
+ return MORSE_SUCCESS;
+ */
+ morse_pztrsmpl( A, B, L, IPIV, sequence, request );
- morse_pztrsm(MorseLeft, MorseUpper, MorseNoTrans, MorseNonUnit, 1.0, A, B, sequence, request);
+ morse_pztrsm( MorseLeft, MorseUpper, MorseNoTrans, MorseNonUnit, 1.0, A, B, sequence, request );
return MORSE_SUCCESS;
}
diff --git a/compute/zgetrs_nopiv.c b/compute/zgetrs_nopiv.c
index f92c66abf05da77ee81aa117a5ac2f1d21411e03..7833e3eafad5da98d755e1af57c94b89dd3c94d0 100644
--- a/compute/zgetrs_nopiv.c
+++ b/compute/zgetrs_nopiv.c
@@ -83,16 +83,17 @@
* @sa MORSE_zgetrf_nopiv
*
******************************************************************************/
-int MORSE_zgetrs_nopiv(MORSE_enum trans, int N, int NRHS,
- MORSE_Complex64_t *A, int LDA,
- MORSE_Complex64_t *B, int LDB)
+int MORSE_zgetrs_nopiv( MORSE_enum trans, int N, int NRHS,
+ MORSE_Complex64_t *A, int LDA,
+ MORSE_Complex64_t *B, int LDB )
{
int NB;
int status;
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
MORSE_request_t request = MORSE_REQUEST_INITIALIZER;
- MORSE_desc_t descA, descB;
+ MORSE_desc_t descAl, descAt;
+ MORSE_desc_t descBl, descBt;
morse = morse_context_self();
if (morse == NULL) {
@@ -134,36 +135,31 @@ int MORSE_zgetrs_nopiv(MORSE_enum trans, int N, int NRHS,
/* Set NT & NTRHS */
NB = MORSE_NB;
- morse_sequence_create(morse, &sequence);
+ morse_sequence_create( morse, &sequence );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooplap2tile( descA, A, NB, NB, LDA, N, 0, 0, N, N, sequence, &request,
- morse_desc_mat_free(&(descA)) );
- morse_zooplap2tile( descB, B, NB, NB, LDB, NRHS, 0, 0, N, NRHS, sequence, &request,
- morse_desc_mat_free(&(descA)); morse_desc_mat_free(&(descB)));
-/* } else {*/
-/* morse_ziplap2tile( descA, A, NB, NB, LDA, N, 0, 0, N, N, */
-/* sequence, &request);*/
-/* morse_ziplap2tile( descB, B, NB, NB, LDB, NRHS, 0, 0, N, NRHS,*/
-/* sequence, &request);*/
-/* }*/
+ /* Submit the matrix conversion */
+ morse_zlap2tile( morse, &descAl, &descAt, MorseDescInput, MorseUpperLower,
+ A, NB, NB, LDA, N, N, N, sequence, &request );
+ morse_zlap2tile( morse, &descBl, &descBt, MorseDescInout, MorseUpperLower,
+ B, NB, NB, LDB, NRHS, N, NRHS, sequence, &request );
/* Call the tile interface */
- MORSE_zgetrs_nopiv_Tile_Async(&descA, &descB, sequence, &request);
+ MORSE_zgetrs_nopiv_Tile_Async( &descAt, &descBt, sequence, &request );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooptile2lap(descB, B, NB, NB, LDB, NRHS, sequence, &request);
- morse_sequence_wait(morse, sequence);
- morse_desc_mat_free(&descA);
- morse_desc_mat_free(&descB);
-/* } else {*/
-/* morse_ziptile2lap( descA, A, NB, NB, LDA, N, sequence, &request);*/
-/* morse_ziptile2lap( descB, B, NB, NB, LDB, NRHS, sequence, &request);*/
-/* morse_sequence_wait(morse, sequence);*/
-/* }*/
+ /* Submit the matrix conversion back */
+ morse_ztile2lap( morse, &descAl, &descAt,
+ MorseDescInput, MorseUpperLower, sequence, &request );
+ morse_ztile2lap( morse, &descBl, &descBt,
+ MorseDescInout, MorseUpperLower, sequence, &request );
+
+ morse_sequence_wait( morse, sequence );
+
+ /* Cleanup the temporary data */
+ morse_ztile2lap_cleanup( morse, &descAl, &descAt );
+ morse_ztile2lap_cleanup( morse, &descBl, &descBt );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -203,7 +199,7 @@ int MORSE_zgetrs_nopiv(MORSE_enum trans, int N, int NRHS,
* @sa MORSE_zgetrf_nopiv_Tile
*
******************************************************************************/
-int MORSE_zgetrs_nopiv_Tile(MORSE_desc_t *A, MORSE_desc_t *B)
+int MORSE_zgetrs_nopiv_Tile( MORSE_desc_t *A, MORSE_desc_t *B )
{
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
@@ -215,14 +211,16 @@ int MORSE_zgetrs_nopiv_Tile(MORSE_desc_t *A, MORSE_desc_t *B)
morse_fatal_error("MORSE_zgetrs_nopiv_Tile", "MORSE not initialized");
return MORSE_ERR_NOT_INITIALIZED;
}
- morse_sequence_create(morse, &sequence);
- MORSE_zgetrs_nopiv_Tile_Async(A, B, sequence, &request);
- RUNTIME_desc_flush( A, sequence );
- RUNTIME_desc_flush( B, sequence );
- morse_sequence_wait(morse, sequence);
+ morse_sequence_create( morse, &sequence );
+
+ MORSE_zgetrs_nopiv_Tile_Async( A, B, sequence, &request );
+
+ MORSE_Desc_Flush( A, sequence );
+ MORSE_Desc_Flush( B, sequence );
+ morse_sequence_wait( morse, sequence );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -256,8 +254,8 @@ int MORSE_zgetrs_nopiv_Tile(MORSE_desc_t *A, MORSE_desc_t *B)
* @sa MORSE_zgetrf_nopiv_Tile_Async
*
******************************************************************************/
-int MORSE_zgetrs_nopiv_Tile_Async(MORSE_desc_t *A, MORSE_desc_t *B,
- MORSE_sequence_t *sequence, MORSE_request_t *request)
+int MORSE_zgetrs_nopiv_Tile_Async( MORSE_desc_t *A, MORSE_desc_t *B,
+ MORSE_sequence_t *sequence, MORSE_request_t *request )
{
MORSE_context_t *morse;
@@ -275,10 +273,12 @@ int MORSE_zgetrs_nopiv_Tile_Async(MORSE_desc_t *A, MORSE_desc_t *B,
return MORSE_ERR_UNALLOCATED;
}
/* Check sequence status */
- if (sequence->status == MORSE_SUCCESS)
+ if (sequence->status == MORSE_SUCCESS) {
request->status = MORSE_SUCCESS;
- else
+ }
+ else {
return morse_request_fail(sequence, request, MORSE_ERR_SEQUENCE_FLUSHED);
+ }
/* Check descriptors for correctness */
if (morse_desc_check(A) != MORSE_SUCCESS) {
@@ -295,13 +295,13 @@ int MORSE_zgetrs_nopiv_Tile_Async(MORSE_desc_t *A, MORSE_desc_t *B,
return morse_request_fail(sequence, request, MORSE_ERR_ILLEGAL_VALUE);
}
/* Quick return */
-/*
- if (chameleon_min(N, NRHS) == 0)
- return MORSE_SUCCESS;
-*/
- morse_pztrsm(MorseLeft, MorseLower, MorseNoTrans, MorseUnit, (MORSE_Complex64_t)1.0, A, B, sequence, request);
+ /*
+ if (chameleon_min(N, NRHS) == 0)
+ return MORSE_SUCCESS;
+ */
+ morse_pztrsm( MorseLeft, MorseLower, MorseNoTrans, MorseUnit, (MORSE_Complex64_t)1.0, A, B, sequence, request );
- morse_pztrsm(MorseLeft, MorseUpper, MorseNoTrans, MorseNonUnit, (MORSE_Complex64_t)1.0, A, B, sequence, request);
+ morse_pztrsm( MorseLeft, MorseUpper, MorseNoTrans, MorseNonUnit, (MORSE_Complex64_t)1.0, A, B, sequence, request );
return MORSE_SUCCESS;
}
diff --git a/compute/zheevd.c b/compute/zheevd.c
index 8094b590b35474f80f0740ba8a04389eeb557b6b..7346739297368b2bdc181beeaf0a99385fe418ee 100644
--- a/compute/zheevd.c
+++ b/compute/zheevd.c
@@ -100,17 +100,17 @@
* @sa MORSE_ssyev
*
******************************************************************************/
-int MORSE_zheevd(MORSE_enum jobz, MORSE_enum uplo, int N,
- MORSE_Complex64_t *A, int LDA,
- double *W,
- MORSE_desc_t *descT)
+int MORSE_zheevd( MORSE_enum jobz, MORSE_enum uplo, int N,
+ MORSE_Complex64_t *A, int LDA,
+ double *W,
+ MORSE_desc_t *descT )
{
int NB;
int status;
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
MORSE_request_t request = MORSE_REQUEST_INITIALIZER;
- MORSE_desc_t descA;
+ MORSE_desc_t descAl, descAt;
morse = morse_context_self();
if (morse == NULL) {
@@ -123,7 +123,7 @@ int MORSE_zheevd(MORSE_enum jobz, MORSE_enum uplo, int N,
morse_error("MORSE_zheevd", "illegal value of jobz");
return -1;
}
- if (uplo != MorseLower && uplo != MorseUpper) {
+ if ((uplo != MorseLower) && (uplo != MorseUpper)) {
morse_error("MORSE_zheevd", "illegal value of uplo");
return -2;
}
@@ -150,30 +150,26 @@ int MORSE_zheevd(MORSE_enum jobz, MORSE_enum uplo, int N,
/* Set NT */
NB = MORSE_NB;
- morse_sequence_create(morse, &sequence);
+ morse_sequence_create( morse, &sequence );
- /* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) { */
- morse_zooplap2tile( descA, A, NB, NB, LDA, N, 0, 0, N, N, sequence, &request,
- morse_desc_mat_free(&(descA)) );
- /* } else { */
- /* morse_ziplap2tile( descA, A, NB, NB, LDA, N, 0, 0, N, N, */
- /* sequence, &requoest); */
- /* } */
+ /* Submit the matrix conversion */
+ morse_zlap2tile( morse, &descAl, &descAt, MorseDescInout, uplo,
+ A, NB, NB, LDA, N, N, N, sequence, &request );
/* Call the tile interface */
- MORSE_zheevd_Tile_Async(jobz, uplo, &descA, W, descT, sequence, &request);
+ MORSE_zheevd_Tile_Async( jobz, uplo, &descAt, W, descT, sequence, &request );
- /* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) { */
- morse_zooptile2lap( descA, A, NB, NB, LDA, N, sequence, &request);
- morse_sequence_wait(morse, sequence);
- morse_desc_mat_free(&descA);
- /* } else { */
- /* morse_ziptile2lap( descA, A, NB, NB, LDA, N, sequence, &request); */
- /* morse_sequence_wait(morse, sequence); */
- /* } */
+ /* Submit the matrix conversion back */
+ morse_ztile2lap( morse, &descAl, &descAt,
+ MorseDescInout, uplo, sequence, &request );
+
+ morse_sequence_wait( morse, sequence );
+
+ /* Cleanup the temporary data */
+ morse_ztile2lap_cleanup( morse, &descAl, &descAt );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
/**
@@ -244,12 +240,12 @@ int MORSE_zheevd(MORSE_enum jobz, MORSE_enum uplo, int N,
* @sa MORSE_ssyev
*
******************************************************************************/
-int MORSE_zheevd_Tile(MORSE_enum jobz, MORSE_enum uplo,
- MORSE_desc_t *A, double *W, MORSE_desc_t *T)
+int MORSE_zheevd_Tile( MORSE_enum jobz, MORSE_enum uplo,
+ MORSE_desc_t *A, double *W, MORSE_desc_t *T )
{
- MORSE_context_t *morse;
+ MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
- MORSE_request_t request = MORSE_REQUEST_INITIALIZER;
+ MORSE_request_t request = MORSE_REQUEST_INITIALIZER;
int status;
morse = morse_context_self();
@@ -257,14 +253,16 @@ int MORSE_zheevd_Tile(MORSE_enum jobz, MORSE_enum uplo,
morse_fatal_error("MORSE_zheevd_Tile", "MORSE not initialized");
return MORSE_ERR_NOT_INITIALIZED;
}
- morse_sequence_create(morse, &sequence);
- MORSE_zheevd_Tile_Async(jobz, uplo, A, W, T, sequence, &request);
- RUNTIME_desc_flush( A, sequence );
- RUNTIME_desc_flush( T, sequence );
- morse_sequence_wait(morse, sequence);
+ morse_sequence_create( morse, &sequence );
+ MORSE_zheevd_Tile_Async( jobz, uplo, A, W, T, sequence, &request );
+
+ MORSE_Desc_Flush( A, sequence );
+ MORSE_Desc_Flush( T, sequence );
+
+ morse_sequence_wait( morse, sequence );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -334,19 +332,20 @@ int MORSE_zheevd_Tile(MORSE_enum jobz, MORSE_enum uplo,
* @sa MORSE_ssyev_Tile_Async
*
******************************************************************************/
-int MORSE_zheevd_Tile_Async(MORSE_enum jobz, MORSE_enum uplo,
- MORSE_desc_t *A, double *W, MORSE_desc_t *T,
- MORSE_sequence_t *sequence, MORSE_request_t *request)
+int MORSE_zheevd_Tile_Async( MORSE_enum jobz, MORSE_enum uplo,
+ MORSE_desc_t *A, double *W, MORSE_desc_t *T,
+ MORSE_sequence_t *sequence, MORSE_request_t *request )
{
MORSE_context_t *morse;
MORSE_desc_t descA;
MORSE_desc_t descT;
MORSE_desc_t D, *Dptr = NULL;
MORSE_Complex64_t *Q2;
- int N, NB, status;
+ int N, status;
double *E;
MORSE_Complex64_t *V;
- MORSE_desc_t descQ2, descV;
+ MORSE_desc_t descQ2;
+ MORSE_desc_t descV;
MORSE_desc_t *subA, *subQ, *subT;
morse = morse_context_self();
@@ -363,10 +362,12 @@ int MORSE_zheevd_Tile_Async(MORSE_enum jobz, MORSE_enum uplo,
return MORSE_ERR_UNALLOCATED;
}
/* Check sequence status */
- if (sequence->status == MORSE_SUCCESS)
+ if (sequence->status == MORSE_SUCCESS) {
request->status = MORSE_SUCCESS;
- else
+ }
+ else {
return morse_request_fail(sequence, request, MORSE_ERR_SEQUENCE_FLUSHED);
+ }
/* Check descriptors for correctness */
if (morse_desc_check(A) != MORSE_SUCCESS) {
@@ -386,7 +387,7 @@ int MORSE_zheevd_Tile_Async(MORSE_enum jobz, MORSE_enum uplo,
morse_error("MORSE_zheevd_Tile_Async", "illegal value of jobz");
return morse_request_fail(sequence, request, MORSE_ERR_ILLEGAL_VALUE);
}
- if (uplo != MorseLower && uplo != MorseUpper) {
+ if ((uplo != MorseLower) && (uplo != MorseUpper)) {
morse_error("MORSE_zheevd_Tile_Async", "illegal value of uplo");
return morse_request_fail(sequence, request, MORSE_ERR_ILLEGAL_VALUE);
}
@@ -399,8 +400,7 @@ int MORSE_zheevd_Tile_Async(MORSE_enum jobz, MORSE_enum uplo,
return morse_request_fail(sequence, request, MORSE_ERR_ILLEGAL_VALUE);
}
- N = descA.m;
- NB = chameleon_min(descA.mb,descA.m);
+ N = descA.m;
/* Allocate data structures for reduction to tridiagonal form */
E = malloc( (N - 1) * sizeof(double) );
@@ -425,7 +425,6 @@ int MORSE_zheevd_Tile_Async(MORSE_enum jobz, MORSE_enum uplo,
morse_error("MORSE_zheevd_Tile", "MORSE_zhetrd failed");
}
-
if (jobz == MorseNoVec){
#if !defined(CHAMELEON_SIMULATION)
/* Tridiagonal eigensolver */
@@ -469,21 +468,21 @@ int MORSE_zheevd_Tile_Async(MORSE_enum jobz, MORSE_enum uplo,
/* Q from MORSE_zhetrd refers to Q2 (lapack layout) */
/* V from LAPACKE_zstedc refers to V (lapack layout) */
/* The final eigenvectors are (Q1 Q2 V) or (Q1^h Q2 V) */
- morse_zooplap2tile( descQ2, Q2, NB, NB, N, N, 0, 0, N, N, sequence, request,
- morse_desc_mat_free(&(descQ2)) );
- morse_zooplap2tile( descV, V, NB, NB, N, N, 0, 0, N, N, sequence, request,
- morse_desc_mat_free(&(descQ2)); morse_desc_mat_free(&(descV)) );
+ /* morse_zooplap2tile( descQ2, Q2, NB, NB, N, N, 0, 0, N, N, sequence, request, */
+ /* morse_desc_mat_free( &descQ2 ) ); */
+ /* morse_zooplap2tile( descV, V, NB, NB, N, N, 0, 0, N, N, sequence, request, */
+ /* morse_desc_mat_free(&(descQ2)); morse_desc_mat_free(&(descV)) ); */
if (uplo == MorseLower)
{
#if defined(CHAMELEON_COPY_DIAG)
- {
- int n = chameleon_min(A->mt, A->nt) * A->nb;
- morse_zdesc_alloc(D, A->mb, A->nb, A->m, n, 0, 0, A->m, n, );
- Dptr = &D;
- }
+ {
+ int n = chameleon_min(A->mt, A->nt) * A->nb;
+ morse_zdesc_alloc(D, A->mb, A->nb, A->m, n, 0, 0, A->m, n, );
+ Dptr = &D;
+ }
#endif
subA = morse_desc_submatrix(&descA, descA.mb, 0, descA.m -descA.mb, descA.n-descA.nb);
- subQ = morse_desc_submatrix(&descQ2, descQ2.mb, 0, descQ2.m-descQ2.mb, descQ2.n );
+ subQ = morse_desc_submatrix(&descQ2, descQ2.mb, 0, descQ2.m-descQ2.mb, descQ2.n );
subT = morse_desc_submatrix(&descT, descT.mb, 0, descT.m -descT.mb, descT.n-descT.nb);
/* Compute Q2 = Q1 * Q2 */
@@ -495,19 +494,19 @@ int MORSE_zheevd_Tile_Async(MORSE_enum jobz, MORSE_enum uplo,
morse_pzgemm( MorseNoTrans, MorseNoTrans,
1.0, &descQ2, &descV,
0.0, &descA,
- sequence, request);
+ sequence, request );
}
else {
#if defined(CHAMELEON_COPY_DIAG)
- {
- int m = chameleon_min(A->mt, A->nt) * A->mb;
- morse_zdesc_alloc(D, A->mb, A->nb, m, A->n, 0, 0, m, A->n, );
- Dptr = &D;
- }
+ {
+ int m = chameleon_min(A->mt, A->nt) * A->mb;
+ morse_zdesc_alloc(D, A->mb, A->nb, m, A->n, 0, 0, m, A->n, );
+ Dptr = &D;
+ }
#endif
subA = morse_desc_submatrix(&descA, 0, descA.nb, descA.m -descA.mb, descA.n -descA.nb );
- subQ = morse_desc_submatrix(&descQ2, descQ2.mb, 0, descQ2.m-descQ2.mb, descQ2.n );
+ subQ = morse_desc_submatrix(&descQ2, descQ2.mb, 0, descQ2.m-descQ2.mb, descQ2.n );
subT = morse_desc_submatrix(&descT, 0, descT.nb, descT.m -descT.mb, descT.n -descT.nb );
/* Compute Q2 = Q1^h * Q2 */
@@ -519,21 +518,22 @@ int MORSE_zheevd_Tile_Async(MORSE_enum jobz, MORSE_enum uplo,
morse_pzgemm( MorseNoTrans, MorseNoTrans,
1.0, &descQ2, &descV,
0.0, &descA,
- sequence, request);
+ sequence, request );
}
- morse_sequence_wait(morse, sequence);
+ morse_sequence_wait( morse, sequence );
free(subA); free(subQ); free(subT);
- morse_desc_mat_free(&descQ2);
+ morse_desc_mat_free( &descQ2 );
free(Q2);
- morse_desc_mat_free(&descV);
+ /* Cleanup the temporary data */
+ morse_desc_mat_free( &descV );
free(V);
free(E);
if (Dptr != NULL) {
- morse_desc_mat_free(Dptr);
+ morse_desc_mat_free( Dptr );
}
(void)D;
return MORSE_SUCCESS;
diff --git a/compute/zhemm.c b/compute/zhemm.c
index 5b6b31852f6335f10dde0a428acb5988abb6338d..ce06a71dab415bba74aa660e34c21b270d79bfab 100644
--- a/compute/zhemm.c
+++ b/compute/zhemm.c
@@ -106,15 +106,17 @@
* @sa MORSE_shemm
*
******************************************************************************/
-int MORSE_zhemm(MORSE_enum side, MORSE_enum uplo, int M, int N,
+int MORSE_zhemm( MORSE_enum side, MORSE_enum uplo, int M, int N,
MORSE_Complex64_t alpha, MORSE_Complex64_t *A, int LDA,
- MORSE_Complex64_t *B, int LDB,
- MORSE_Complex64_t beta, MORSE_Complex64_t *C, int LDC)
+ MORSE_Complex64_t *B, int LDB,
+ MORSE_Complex64_t beta, MORSE_Complex64_t *C, int LDC )
{
int NB;
int Am;
int status;
- MORSE_desc_t descA, descB, descC;
+ MORSE_desc_t descAl, descAt;
+ MORSE_desc_t descBl, descBt;
+ MORSE_desc_t descCl, descCt;
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
MORSE_request_t request = MORSE_REQUEST_INITIALIZER;
@@ -171,43 +173,36 @@ int MORSE_zhemm(MORSE_enum side, MORSE_enum uplo, int M, int N,
/* Set MT & NT & KT */
NB = MORSE_NB;
- morse_sequence_create(morse, &sequence);
+ morse_sequence_create( morse, &sequence );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooplap2tile( descA, A, NB, NB, LDA, Am, 0, 0, Am, Am, sequence, &request,
- morse_desc_mat_free(&(descA)) );
- morse_zooplap2tile( descB, B, NB, NB, LDB, N, 0, 0, M, N, sequence, &request,
- morse_desc_mat_free(&(descA)); morse_desc_mat_free(&(descB)));
- morse_zooplap2tile( descC, C, NB, NB, LDC, N, 0, 0, M, N, sequence, &request,
- morse_desc_mat_free(&(descA)); morse_desc_mat_free(&(descB)); morse_desc_mat_free(&(descC)));
-/* } else {*/
-/* morse_ziplap2tile( descA, A, NB, NB, LDA, Am, 0, 0, Am, Am,*/
-/* sequence, &request);*/
-/* morse_ziplap2tile( descB, B, NB, NB, LDB, N, 0, 0, M, N,*/
-/* sequence, &request);*/
-/* morse_ziplap2tile( descC, C, NB, NB, LDC, N, 0, 0, M, N,*/
-/* sequence, &request);*/
-/* }*/
+ /* Submit the matrix conversion */
+ morse_zlap2tile( morse, &descAl, &descAt, MorseDescInput, uplo,
+ A, NB, NB, LDA, Am, Am, Am, sequence, &request );
+ morse_zlap2tile( morse, &descBl, &descBt, MorseDescInput, MorseUpperLower,
+ B, NB, NB, LDB, N, M, N, sequence, &request );
+ morse_zlap2tile( morse, &descCl, &descCt, MorseDescInout, MorseUpperLower,
+ C, NB, NB, LDC, N, M, N, sequence, &request );
/* Call the tile interface */
- MORSE_zhemm_Tile_Async(
- side, uplo, alpha, &descA, &descB, beta, &descC, sequence, &request);
+ MORSE_zhemm_Tile_Async( side, uplo, alpha, &descAt, &descBt, beta, &descCt, sequence, &request );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooptile2lap(descC, C, NB, NB, LDC, N, sequence, &request);
- morse_sequence_wait(morse, sequence);
- morse_desc_mat_free(&descA);
- morse_desc_mat_free(&descB);
- morse_desc_mat_free(&descC);
-/* } else {*/
-/* morse_ziptile2lap( descA, A, NB, NB, LDA, Am, sequence, &request);*/
-/* morse_ziptile2lap( descB, B, NB, NB, LDB, N, sequence, &request);*/
-/* morse_ziptile2lap( descC, C, NB, NB, LDC, N, sequence, &request);*/
-/* morse_sequence_wait(morse, sequence);*/
-/* }*/
+ /* Submit the matrix conversion back */
+ morse_ztile2lap( morse, &descAl, &descAt,
+ MorseDescInput, uplo, sequence, &request );
+ morse_ztile2lap( morse, &descBl, &descBt,
+ MorseDescInput, MorseUpperLower, sequence, &request );
+ morse_ztile2lap( morse, &descCl, &descCt,
+ MorseDescInout, MorseUpperLower, sequence, &request );
+
+ morse_sequence_wait( morse, sequence );
+
+ /* Cleanup the temporary data */
+ morse_ztile2lap_cleanup( morse, &descAl, &descAt );
+ morse_ztile2lap_cleanup( morse, &descBl, &descBt );
+ morse_ztile2lap_cleanup( morse, &descCl, &descCt );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -270,9 +265,9 @@ int MORSE_zhemm(MORSE_enum side, MORSE_enum uplo, int M, int N,
* @sa MORSE_shemm_Tile
*
******************************************************************************/
-int MORSE_zhemm_Tile(MORSE_enum side, MORSE_enum uplo,
+int MORSE_zhemm_Tile( MORSE_enum side, MORSE_enum uplo,
MORSE_Complex64_t alpha, MORSE_desc_t *A, MORSE_desc_t *B,
- MORSE_Complex64_t beta, MORSE_desc_t *C)
+ MORSE_Complex64_t beta, MORSE_desc_t *C )
{
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
@@ -284,15 +279,17 @@ int MORSE_zhemm_Tile(MORSE_enum side, MORSE_enum uplo,
morse_fatal_error("MORSE_zhemm_Tile", "MORSE not initialized");
return MORSE_ERR_NOT_INITIALIZED;
}
- morse_sequence_create(morse, &sequence);
- MORSE_zhemm_Tile_Async(side, uplo, alpha, A, B, beta, C, sequence, &request);
- RUNTIME_desc_flush( A, sequence );
- RUNTIME_desc_flush( B, sequence );
- RUNTIME_desc_flush( C, sequence );
- morse_sequence_wait(morse, sequence);
-
+ morse_sequence_create( morse, &sequence );
+
+ MORSE_zhemm_Tile_Async(side, uplo, alpha, A, B, beta, C, sequence, &request );
+
+ MORSE_Desc_Flush( A, sequence );
+ MORSE_Desc_Flush( B, sequence );
+ MORSE_Desc_Flush( C, sequence );
+
+ morse_sequence_wait( morse, sequence );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -324,10 +321,10 @@ int MORSE_zhemm_Tile(MORSE_enum side, MORSE_enum uplo,
* @sa MORSE_shemm_Tile_Async
*
******************************************************************************/
-int MORSE_zhemm_Tile_Async(MORSE_enum side, MORSE_enum uplo,
+int MORSE_zhemm_Tile_Async( MORSE_enum side, MORSE_enum uplo,
MORSE_Complex64_t alpha, MORSE_desc_t *A, MORSE_desc_t *B,
MORSE_Complex64_t beta, MORSE_desc_t *C,
- MORSE_sequence_t *sequence, MORSE_request_t *request)
+ MORSE_sequence_t *sequence, MORSE_request_t *request )
{
MORSE_context_t *morse;
@@ -345,10 +342,12 @@ int MORSE_zhemm_Tile_Async(MORSE_enum side, MORSE_enum uplo,
return MORSE_ERR_UNALLOCATED;
}
/* Check sequence status */
- if (sequence->status == MORSE_SUCCESS)
+ if (sequence->status == MORSE_SUCCESS) {
request->status = MORSE_SUCCESS;
- else
+ }
+ else {
return morse_request_fail(sequence, request, MORSE_ERR_SEQUENCE_FLUSHED);
+ }
/* Check descriptors for correctness */
if (morse_desc_check(A) != MORSE_SUCCESS) {
@@ -416,11 +415,13 @@ int MORSE_zhemm_Tile_Async(MORSE_enum side, MORSE_enum uplo,
}
/* Quick return */
- if (C->m == 0 || C->n == 0 ||
- ( (alpha == (MORSE_Complex64_t)0.0) && (beta == (MORSE_Complex64_t)1.0) ))
+ if ( (C->m == 0) || (C->n == 0) ||
+ ( (alpha == (MORSE_Complex64_t)0.0) && (beta == (MORSE_Complex64_t)1.0) ) )
+ {
return MORSE_SUCCESS;
+ }
- morse_pzhemm(side, uplo, alpha, A, B, beta, C, sequence, request);
+ morse_pzhemm( side, uplo, alpha, A, B, beta, C, sequence, request );
return MORSE_SUCCESS;
}
diff --git a/compute/zher2k.c b/compute/zher2k.c
index e8f2e6d892b9bf8be8b7c8f3fff666afce446ece..71608495f399ad90c52a5b4dd4c00a9366a5f3b2 100644
--- a/compute/zher2k.c
+++ b/compute/zher2k.c
@@ -83,7 +83,7 @@
* @param[in] LDB
* The leading dimension of the array B. LDB must be at least
* max( 1, N ), otherwise LDB must be at least max( 1, K ).
- *
+ *
* @param[in] beta
* beta specifies the scalar beta.
*
@@ -108,14 +108,16 @@
* @sa MORSE_sher2k
*
******************************************************************************/
-int MORSE_zher2k(MORSE_enum uplo, MORSE_enum trans, int N, int K,
- MORSE_Complex64_t alpha, MORSE_Complex64_t *A, int LDA, MORSE_Complex64_t *B, int LDB,
- double beta, MORSE_Complex64_t *C, int LDC)
+int MORSE_zher2k( MORSE_enum uplo, MORSE_enum trans, int N, int K,
+ MORSE_Complex64_t alpha, MORSE_Complex64_t *A, int LDA, MORSE_Complex64_t *B, int LDB,
+ double beta, MORSE_Complex64_t *C, int LDC )
{
int NB;
int Am, An;
int status;
- MORSE_desc_t descA, descB, descC;
+ MORSE_desc_t descAl, descAt;
+ MORSE_desc_t descBl, descBt;
+ MORSE_desc_t descCl, descCt;
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
MORSE_request_t request = MORSE_REQUEST_INITIALIZER;
@@ -135,7 +137,7 @@ int MORSE_zher2k(MORSE_enum uplo, MORSE_enum trans, int N, int K,
morse_error("MORSE_zher2k", "illegal value of trans");
return -2;
}
- if ( trans == MorseNoTrans ) {
+ if ( trans == MorseNoTrans ) {
Am = N; An = K;
} else {
Am = K; An = N;
@@ -176,42 +178,36 @@ int MORSE_zher2k(MORSE_enum uplo, MORSE_enum trans, int N, int K,
/* Set MT & NT & KT */
NB = MORSE_NB;
- morse_sequence_create(morse, &sequence);
+ morse_sequence_create( morse, &sequence );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooplap2tile( descA, A, NB, NB, LDA, An, 0, 0, Am, An, sequence, &request,
- morse_desc_mat_free(&(descA)) );
- morse_zooplap2tile( descB, B, NB, NB, LDB, An, 0, 0, Am, An, sequence, &request,
- morse_desc_mat_free(&(descA)); morse_desc_mat_free(&(descB)));
- morse_zooplap2tile( descC, C, NB, NB, LDC, N, 0, 0, N, N, sequence, &request,
- morse_desc_mat_free(&(descA)); morse_desc_mat_free(&(descB)); morse_desc_mat_free(&(descC)));
-/* } else {*/
-/* morse_ziplap2tile( descA, A, NB, NB, LDA, An, 0, 0, Am, An, */
-/* sequence, &request);*/
-/* morse_ziplap2tile( descB, B, NB, NB, LDB, An, 0, 0, Am, An, */
-/* sequence, &request);*/
-/* morse_ziplap2tile( descC, C, NB, NB, LDC, N, 0, 0, N, N, */
-/* sequence, &request);*/
-/* }*/
+ /* Submit the matrix conversion */
+ morse_zlap2tile( morse, &descAl, &descAt, MorseDescInput, MorseUpperLower,
+ A, NB, NB, LDA, An, Am, An, sequence, &request );
+ morse_zlap2tile( morse, &descBl, &descBt, MorseDescInput, MorseUpperLower,
+ B, NB, NB, LDB, An, Am, An, sequence, &request );
+ morse_zlap2tile( morse, &descCl, &descCt, MorseDescInout, uplo,
+ C, NB, NB, LDC, N, N, N, sequence, &request );
/* Call the tile interface */
- MORSE_zher2k_Tile_Async(uplo, trans, alpha, &descA, &descB, beta, &descC, sequence, &request);
+ MORSE_zher2k_Tile_Async( uplo, trans, alpha, &descAt, &descBt, beta, &descCt, sequence, &request );
+
+ /* Submit the matrix conversion back */
+ morse_ztile2lap( morse, &descAl, &descAt,
+ MorseDescInput, MorseUpperLower, sequence, &request );
+ morse_ztile2lap( morse, &descBl, &descBt,
+ MorseDescInput, MorseUpperLower, sequence, &request );
+ morse_ztile2lap( morse, &descCl, &descCt,
+ MorseDescInout, uplo, sequence, &request );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooptile2lap(descC, C, NB, NB, LDC, N, sequence, &request);
- morse_sequence_wait(morse, sequence);
- morse_desc_mat_free(&descA);
- morse_desc_mat_free(&descB);
- morse_desc_mat_free(&descC);
-/* } else {*/
-/* morse_ziptile2lap( descA, A, NB, NB, LDA, An, sequence, &request);*/
-/* morse_ziptile2lap( descB, B, NB, NB, LDB, An, sequence, &request);*/
-/* morse_ziptile2lap( descC, C, NB, NB, LDC, N, sequence, &request);*/
-/* morse_sequence_wait(morse, sequence);*/
-/* }*/
+ morse_sequence_wait( morse, sequence );
+
+ /* Cleanup the temporary data */
+ morse_ztile2lap_cleanup( morse, &descAl, &descAt );
+ morse_ztile2lap_cleanup( morse, &descBl, &descBt );
+ morse_ztile2lap_cleanup( morse, &descCl, &descCt );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -269,9 +265,9 @@ int MORSE_zher2k(MORSE_enum uplo, MORSE_enum trans, int N, int K,
* @sa MORSE_sher2k
*
******************************************************************************/
-int MORSE_zher2k_Tile(MORSE_enum uplo, MORSE_enum trans,
- MORSE_Complex64_t alpha, MORSE_desc_t *A, MORSE_desc_t *B,
- double beta, MORSE_desc_t *C)
+int MORSE_zher2k_Tile( MORSE_enum uplo, MORSE_enum trans,
+ MORSE_Complex64_t alpha, MORSE_desc_t *A, MORSE_desc_t *B,
+ double beta, MORSE_desc_t *C )
{
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
@@ -283,15 +279,17 @@ int MORSE_zher2k_Tile(MORSE_enum uplo, MORSE_enum trans,
morse_fatal_error("MORSE_zher2k_Tile", "MORSE not initialized");
return MORSE_ERR_NOT_INITIALIZED;
}
- morse_sequence_create(morse, &sequence);
- MORSE_zher2k_Tile_Async(uplo, trans, alpha, A, B, beta, C, sequence, &request);
- RUNTIME_desc_flush( A, sequence );
- RUNTIME_desc_flush( B, sequence );
- RUNTIME_desc_flush( C, sequence );
- morse_sequence_wait(morse, sequence);
-
+ morse_sequence_create( morse, &sequence );
+
+ MORSE_zher2k_Tile_Async( uplo, trans, alpha, A, B, beta, C, sequence, &request );
+
+ MORSE_Desc_Flush( A, sequence );
+ MORSE_Desc_Flush( B, sequence );
+ MORSE_Desc_Flush( C, sequence );
+
+ morse_sequence_wait( morse, sequence );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -323,10 +321,10 @@ int MORSE_zher2k_Tile(MORSE_enum uplo, MORSE_enum trans,
* @sa MORSE_sher2k_Tile_Async
*
******************************************************************************/
-int MORSE_zher2k_Tile_Async(MORSE_enum uplo, MORSE_enum trans,
- MORSE_Complex64_t alpha, MORSE_desc_t *A, MORSE_desc_t *B,
- double beta, MORSE_desc_t *C,
- MORSE_sequence_t *sequence, MORSE_request_t *request)
+int MORSE_zher2k_Tile_Async( MORSE_enum uplo, MORSE_enum trans,
+ MORSE_Complex64_t alpha, MORSE_desc_t *A, MORSE_desc_t *B,
+ double beta, MORSE_desc_t *C,
+ MORSE_sequence_t *sequence, MORSE_request_t *request )
{
MORSE_context_t *morse;
int N, K;
@@ -346,10 +344,12 @@ int MORSE_zher2k_Tile_Async(MORSE_enum uplo, MORSE_enum trans,
return MORSE_ERR_UNALLOCATED;
}
/* Check sequence status */
- if (sequence->status == MORSE_SUCCESS)
+ if (sequence->status == MORSE_SUCCESS) {
request->status = MORSE_SUCCESS;
- else
+ }
+ else {
return morse_request_fail(sequence, request, MORSE_ERR_SEQUENCE_FLUSHED);
+ }
/* Check descriptors for correctness */
if (morse_desc_check(A) != MORSE_SUCCESS) {
@@ -405,11 +405,13 @@ int MORSE_zher2k_Tile_Async(MORSE_enum uplo, MORSE_enum trans,
K = An;
/* Quick return */
- if ( N == 0 ||
- ((alpha == (MORSE_Complex64_t)0.0 || K == 0) && beta == (double)1.0))
+ if ( (N == 0) ||
+ (((alpha == (MORSE_Complex64_t)0.0) || (K == 0)) && (beta == (double)1.0)) )
+ {
return MORSE_SUCCESS;
+ }
- morse_pzher2k(uplo, trans, alpha, A, B, beta, C, sequence, request);
+ morse_pzher2k( uplo, trans, alpha, A, B, beta, C, sequence, request );
return MORSE_SUCCESS;
}
diff --git a/compute/zherk.c b/compute/zherk.c
index 6fbe72d552b780604b3b3026e7c68e3cb7452227..b99dd19884401cb400d435abc60dea745f9e333f 100644
--- a/compute/zherk.c
+++ b/compute/zherk.c
@@ -98,14 +98,15 @@
* @sa MORSE_sherk
*
******************************************************************************/
-int MORSE_zherk(MORSE_enum uplo, MORSE_enum trans, int N, int K,
+int MORSE_zherk( MORSE_enum uplo, MORSE_enum trans, int N, int K,
double alpha, MORSE_Complex64_t *A, int LDA,
- double beta, MORSE_Complex64_t *C, int LDC)
+ double beta, MORSE_Complex64_t *C, int LDC )
{
int NB;
int Am, An;
int status;
- MORSE_desc_t descA, descC;
+ MORSE_desc_t descAl, descAt;
+ MORSE_desc_t descCl, descCt;
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
MORSE_request_t request = MORSE_REQUEST_INITIALIZER;
@@ -125,7 +126,7 @@ int MORSE_zherk(MORSE_enum uplo, MORSE_enum trans, int N, int K,
morse_error("MORSE_zherk", "illegal value of trans");
return -2;
}
- if ( trans == MorseNoTrans ) {
+ if ( trans == MorseNoTrans ) {
Am = N; An = K;
} else {
Am = K; An = N;
@@ -162,36 +163,31 @@ int MORSE_zherk(MORSE_enum uplo, MORSE_enum trans, int N, int K,
/* Set MT & NT & KT */
NB = MORSE_NB;
- morse_sequence_create(morse, &sequence);
+ morse_sequence_create( morse, &sequence );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooplap2tile( descA, A, NB, NB, LDA, An, 0, 0, Am, An, sequence, &request,
- morse_desc_mat_free(&(descA)) );
- morse_zooplap2tile( descC, C, NB, NB, LDC, N, 0, 0, N, N, sequence, &request,
- morse_desc_mat_free(&(descA)); morse_desc_mat_free(&(descC)));
-/* } else {*/
-/* morse_ziplap2tile( descA, A, NB, NB, LDA, An, 0, 0, Am, An, */
-/* sequence, &request);*/
-/* morse_ziplap2tile( descC, C, NB, NB, LDC, N, 0, 0, N, N, */
-/* sequence, &request);*/
-/* }*/
+ /* Submit the matrix conversion */
+ morse_zlap2tile( morse, &descAl, &descAt, MorseDescInput, MorseUpperLower,
+ A, NB, NB, LDA, An, Am, An, sequence, &request );
+ morse_zlap2tile( morse, &descCl, &descCt, MorseDescInout, uplo,
+ C, NB, NB, LDC, N, N, N, sequence, &request );
/* Call the tile interface */
- MORSE_zherk_Tile_Async(uplo, trans, alpha, &descA, beta, &descC, sequence, &request);
+ MORSE_zherk_Tile_Async( uplo, trans, alpha, &descAt, beta, &descCt, sequence, &request );
+
+ /* Submit the matrix conversion back */
+ morse_ztile2lap( morse, &descAl, &descAt,
+ MorseDescInput, MorseUpperLower, sequence, &request );
+ morse_ztile2lap( morse, &descCl, &descCt,
+ MorseDescInout, uplo, sequence, &request );
+
+ morse_sequence_wait( morse, sequence );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooptile2lap(descC, C, NB, NB, LDC, N, sequence, &request);
- morse_sequence_wait(morse, sequence);
- morse_desc_mat_free(&descA);
- morse_desc_mat_free(&descC);
-/* } else {*/
-/* morse_ziptile2lap( descA, A, NB, NB, LDA, An, sequence, &request);*/
-/* morse_ziptile2lap( descC, C, NB, NB, LDC, N, sequence, &request);*/
-/* morse_sequence_wait(morse, sequence);*/
-/* }*/
+ /* Cleanup the temporary data */
+ morse_ztile2lap_cleanup( morse, &descAl, &descAt );
+ morse_ztile2lap_cleanup( morse, &descCl, &descCt );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -245,9 +241,9 @@ int MORSE_zherk(MORSE_enum uplo, MORSE_enum trans, int N, int K,
* @sa MORSE_sherk
*
******************************************************************************/
-int MORSE_zherk_Tile(MORSE_enum uplo, MORSE_enum trans,
+int MORSE_zherk_Tile( MORSE_enum uplo, MORSE_enum trans,
double alpha, MORSE_desc_t *A,
- double beta, MORSE_desc_t *C)
+ double beta, MORSE_desc_t *C )
{
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
@@ -259,14 +255,16 @@ int MORSE_zherk_Tile(MORSE_enum uplo, MORSE_enum trans,
morse_fatal_error("MORSE_zherk_Tile", "MORSE not initialized");
return MORSE_ERR_NOT_INITIALIZED;
}
- morse_sequence_create(morse, &sequence);
- MORSE_zherk_Tile_Async(uplo, trans, alpha, A, beta, C, sequence, &request);
- RUNTIME_desc_flush( A, sequence );
- RUNTIME_desc_flush( C, sequence );
- morse_sequence_wait(morse, sequence);
-
+ morse_sequence_create( morse, &sequence );
+
+ MORSE_zherk_Tile_Async( uplo, trans, alpha, A, beta, C, sequence, &request );
+
+ MORSE_Desc_Flush( A, sequence );
+ MORSE_Desc_Flush( C, sequence );
+
+ morse_sequence_wait( morse, sequence );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -298,10 +296,10 @@ int MORSE_zherk_Tile(MORSE_enum uplo, MORSE_enum trans,
* @sa MORSE_sherk_Tile_Async
*
******************************************************************************/
-int MORSE_zherk_Tile_Async(MORSE_enum uplo, MORSE_enum trans,
+int MORSE_zherk_Tile_Async( MORSE_enum uplo, MORSE_enum trans,
double alpha, MORSE_desc_t *A,
double beta, MORSE_desc_t *C,
- MORSE_sequence_t *sequence, MORSE_request_t *request)
+ MORSE_sequence_t *sequence, MORSE_request_t *request )
{
MORSE_context_t *morse;
int N, K;
@@ -321,10 +319,12 @@ int MORSE_zherk_Tile_Async(MORSE_enum uplo, MORSE_enum trans,
return MORSE_ERR_UNALLOCATED;
}
/* Check sequence status */
- if (sequence->status == MORSE_SUCCESS)
+ if (sequence->status == MORSE_SUCCESS) {
request->status = MORSE_SUCCESS;
- else
+ }
+ else {
return morse_request_fail(sequence, request, MORSE_ERR_SEQUENCE_FLUSHED);
+ }
/* Check descriptors for correctness */
if (morse_desc_check(A) != MORSE_SUCCESS) {
@@ -376,11 +376,13 @@ int MORSE_zherk_Tile_Async(MORSE_enum uplo, MORSE_enum trans,
K = An;
/* Quick return */
- if ( N == 0 ||
- ((alpha == (double)0.0 || K == 0) && beta == (double)1.0))
+ if ( (N == 0) ||
+ (((alpha == (double)0.0) || (K == 0)) && (beta == (double)1.0)) )
+ {
return MORSE_SUCCESS;
+ }
- morse_pzherk(uplo, trans, alpha, A, beta, C, sequence, request);
+ morse_pzherk( uplo, trans, alpha, A, beta, C, sequence, request );
return MORSE_SUCCESS;
}
diff --git a/compute/zhetrd.c b/compute/zhetrd.c
index e69bce8a9eb5f25e76f796a2cd2857f65df93daf..5593d45a4b18c68d3403bdf7bee88be86f2449eb 100644
--- a/compute/zhetrd.c
+++ b/compute/zhetrd.c
@@ -113,19 +113,19 @@
* @sa MORSE_ssytrd
*
******************************************************************************/
-int MORSE_zhetrd(MORSE_enum jobz, MORSE_enum uplo, int N,
- MORSE_Complex64_t *A, int LDA,
- double *D,
- double *E,
- MORSE_desc_t *descT,
- MORSE_Complex64_t *Q, int LDQ)
+int MORSE_zhetrd( MORSE_enum jobz, MORSE_enum uplo, int N,
+ MORSE_Complex64_t *A, int LDA,
+ double *D,
+ double *E,
+ MORSE_desc_t *descT,
+ MORSE_Complex64_t *Q, int LDQ )
{
int NB;
int status;
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
MORSE_request_t request = MORSE_REQUEST_INITIALIZER;
- MORSE_desc_t descA;
+ MORSE_desc_t descAl, descAt;
morse = morse_context_self();
if (morse == NULL) {
@@ -138,7 +138,7 @@ int MORSE_zhetrd(MORSE_enum jobz, MORSE_enum uplo, int N,
morse_error("MORSE_zhetrd", "illegal value of jobz");
return -1;
}
- if (uplo != MorseLower && uplo != MorseUpper) {
+ if ((uplo != MorseLower) && (uplo != MorseUpper)) {
morse_error("MORSE_zhetrd", "illegal value of uplo");
return -1;
}
@@ -164,30 +164,26 @@ int MORSE_zhetrd(MORSE_enum jobz, MORSE_enum uplo, int N,
/* Set NT */
NB = MORSE_NB;
- morse_sequence_create(morse, &sequence);
+ morse_sequence_create( morse, &sequence );
- /* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) { */
- morse_zooplap2tile( descA, A, NB, NB, LDA, N, 0, 0, N, N, sequence, &request,
- morse_desc_mat_free(&(descA)) );
- /* } else { */
- /* morse_ziplap2tile( descA, A, NB, NB, LDA, N, 0, 0, N, N, */
- /* sequence, &request); */
- /* } */
+ /* Submit the matrix conversion */
+ morse_zlap2tile( morse, &descAl, &descAt, MorseDescInout, uplo,
+ A, NB, NB, LDA, N, N, N, sequence, &request );
/* Call the tile interface */
- MORSE_zhetrd_Tile_Async(jobz, uplo, &descA, D, E, descT, Q, LDQ, sequence, &request);
+ MORSE_zhetrd_Tile_Async( jobz, uplo, &descAt, D, E, descT, Q, LDQ, sequence, &request );
+
+ /* Submit the matrix conversion back */
+ morse_ztile2lap( morse, &descAl, &descAt,
+ MorseDescInout, uplo, sequence, &request );
+
+ morse_sequence_wait( morse, sequence );
- /* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) { */
- morse_zooptile2lap( descA, A, NB, NB, LDA, N, sequence, &request);
- morse_sequence_wait(morse, sequence);
- morse_desc_mat_free(&descA);
- /* } else { */
- /* morse_ziptile2lap( descA, A, NB, NB, LDA, N, sequence, &request); */
- /* morse_sequence_wait(morse, sequence); */
- /* } */
+ /* Cleanup the temporary data */
+ morse_ztile2lap_cleanup( morse, &descAl, &descAt );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
/**
@@ -271,9 +267,9 @@ int MORSE_zhetrd(MORSE_enum jobz, MORSE_enum uplo, int N,
* @sa MORSE_zhetrd_Tile
*
******************************************************************************/
-int MORSE_zhetrd_Tile(MORSE_enum jobz, MORSE_enum uplo,
- MORSE_desc_t *A, double *D, double *E,
- MORSE_desc_t *T, MORSE_Complex64_t *Q, int LDQ)
+int MORSE_zhetrd_Tile( MORSE_enum jobz, MORSE_enum uplo,
+ MORSE_desc_t *A, double *D, double *E,
+ MORSE_desc_t *T, MORSE_Complex64_t *Q, int LDQ )
{
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
@@ -285,11 +281,16 @@ int MORSE_zhetrd_Tile(MORSE_enum jobz, MORSE_enum uplo,
morse_fatal_error("MORSE_zhetrd_Tile", "MORSE not initialized");
return MORSE_ERR_NOT_INITIALIZED;
}
- morse_sequence_create(morse, &sequence);
- MORSE_zhetrd_Tile_Async(jobz, uplo, A, D, E, T, Q, LDQ, sequence, &request);
- morse_sequence_wait(morse, sequence);
+ morse_sequence_create( morse, &sequence );
+
+ MORSE_zhetrd_Tile_Async( jobz, uplo, A, D, E, T, Q, LDQ, sequence, &request );
+
+ MORSE_Desc_Flush( A, sequence );
+ MORSE_Desc_Flush( T, sequence );
+
+ morse_sequence_wait( morse, sequence );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -325,14 +326,14 @@ int MORSE_zhetrd_Tile(MORSE_enum jobz, MORSE_enum uplo,
* @sa MORSE_ssytrd_Tile_Async
*
******************************************************************************/
-int MORSE_zhetrd_Tile_Async(MORSE_enum jobz,
- MORSE_enum uplo,
- MORSE_desc_t *A,
- double *W,
- double *E,
- MORSE_desc_t *T,
- MORSE_Complex64_t *Q, int LDQ,
- MORSE_sequence_t *sequence, MORSE_request_t *request)
+int MORSE_zhetrd_Tile_Async( MORSE_enum jobz,
+ MORSE_enum uplo,
+ MORSE_desc_t *A,
+ double *W,
+ double *E,
+ MORSE_desc_t *T,
+ MORSE_Complex64_t *Q, int LDQ,
+ MORSE_sequence_t *sequence, MORSE_request_t *request )
{
MORSE_context_t *morse;
MORSE_desc_t descA;
@@ -356,10 +357,12 @@ int MORSE_zhetrd_Tile_Async(MORSE_enum jobz,
}
/* Check sequence status */
- if (sequence->status == MORSE_SUCCESS)
+ if (sequence->status == MORSE_SUCCESS) {
request->status = MORSE_SUCCESS;
- else
+ }
+ else {
return morse_request_fail(sequence, request, MORSE_ERR_SEQUENCE_FLUSHED);
+ }
/* Check descriptors for correctness */
if (morse_desc_check(A) != MORSE_SUCCESS) {
@@ -378,7 +381,7 @@ int MORSE_zhetrd_Tile_Async(MORSE_enum jobz,
morse_error("MORSE_zhetrd_Tile_Async", "illegal value of jobz");
return morse_request_fail(sequence, request, MORSE_ERR_ILLEGAL_VALUE);
}
- if (uplo != MorseLower && uplo != MorseUpper) {
+ if ((uplo != MorseLower) && (uplo != MorseUpper)) {
morse_error("MORSE_zhetrd_Tile_Async", "illegal value of uplo");
return morse_request_fail(sequence, request, MORSE_ERR_ILLEGAL_VALUE);
}
@@ -416,7 +419,8 @@ int MORSE_zhetrd_Tile_Async(MORSE_enum jobz,
/* Copy data into band structure */
morse_pztile2band( uplo, A, &descAB,
sequence, request );
- morse_sequence_wait(morse, sequence);
+
+ morse_sequence_wait( morse, sequence );
/* Reduce band matrix to tridiagonal matrix */
#if !defined(CHAMELEON_SIMULATION)
@@ -431,9 +435,9 @@ int MORSE_zhetrd_Tile_Async(MORSE_enum jobz,
}
#endif /* !defined(CHAMELEON_SIMULATION) */
if (Dptr != NULL) {
- morse_desc_mat_free(Dptr);
+ morse_desc_mat_free( Dptr );
}
- morse_desc_mat_free(&descAB);
+ morse_desc_mat_free( &descAB );
(void)D;
return MORSE_SUCCESS;
}
diff --git a/compute/zlacpy.c b/compute/zlacpy.c
index 3088442702a338f26aa23fbba300a5aa8e9bd9a9..d47417ac4d7bdc868db32dae4d94bdac3fad0ae3 100644
--- a/compute/zlacpy.c
+++ b/compute/zlacpy.c
@@ -74,16 +74,17 @@
* @sa MORSE_slacpy
*
******************************************************************************/
-int MORSE_zlacpy(MORSE_enum uplo, int M, int N,
+int MORSE_zlacpy( MORSE_enum uplo, int M, int N,
MORSE_Complex64_t *A, int LDA,
- MORSE_Complex64_t *B, int LDB)
+ MORSE_Complex64_t *B, int LDB )
{
int NB;
int status;
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
MORSE_request_t request = MORSE_REQUEST_INITIALIZER;
- MORSE_desc_t descA, descB;
+ MORSE_desc_t descAl, descAt;
+ MORSE_desc_t descBl, descBt;
morse = morse_context_self();
if (morse == NULL) {
@@ -116,7 +117,7 @@ int MORSE_zlacpy(MORSE_enum uplo, int M, int N,
/* Quick return */
if (chameleon_min(N, M) == 0)
- return (double)0.0;
+ return (double)0.0;
/* Tune NB depending on M, N & NRHS; Set NBNB */
status = morse_tune(MORSE_FUNC_ZGEMM, M, N, 0);
@@ -128,34 +129,30 @@ int MORSE_zlacpy(MORSE_enum uplo, int M, int N,
/* Set NT */
NB = MORSE_NB;
- morse_sequence_create(morse, &sequence);
+ morse_sequence_create( morse, &sequence );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooplap2tile( descA, A, NB, NB, LDA, N, 0, 0, M, N, sequence, &request,
- morse_desc_mat_free(&(descA)) );
- morse_zooplap2tile( descB, B, NB, NB, LDB, N, 0, 0, M, N, sequence, &request,
- morse_desc_mat_free(&(descA)); morse_desc_mat_free(&(descB)) );
-/* } else {*/
-/* morse_ziplap2tile( descA, A, NB, NB, LDA, N, 0, 0, M, N,*/
-/* sequence, &request);*/
-/* morse_ziplap2tile( descB, B, NB, NB, LDA, N, 0, 0, M, N,*/
-/* sequence, &request);*/
-/* }*/
+ /* Submit the matrix conversion */
+ morse_zlap2tile( morse, &descAl, &descAt, MorseDescInput, uplo,
+ A, NB, NB, LDA, N, M, N, sequence, &request );
+ morse_zlap2tile( morse, &descBl, &descBt, MorseDescInout, uplo,
+ B, NB, NB, LDB, N, M, N, sequence, &request );
/* Call the tile interface */
- MORSE_zlacpy_Tile_Async(uplo, &descA, &descB, sequence, &request);
+ MORSE_zlacpy_Tile_Async( uplo, &descAt, &descBt, sequence, &request );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooptile2lap(descB, B, NB, NB, LDB, N, sequence, &request);
- morse_sequence_wait(morse, sequence);
- morse_desc_mat_free(&descA);
- morse_desc_mat_free(&descB);
-/* } else {*/
-/* morse_ziptile2lap( descB, B, NB, NB, LDB, N, sequence, &request);*/
-/* morse_sequence_wait(morse, sequence);*/
-/* }*/
+ /* Submit the matrix conversion back */
+ morse_ztile2lap( morse, &descAl, &descAt,
+ MorseDescInput, uplo, sequence, &request );
+ morse_ztile2lap( morse, &descBl, &descBt,
+ MorseDescInout, uplo, sequence, &request );
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_wait( morse, sequence );
+
+ /* Cleanup the temporary data */
+ morse_ztile2lap_cleanup( morse, &descAl, &descAt );
+ morse_ztile2lap_cleanup( morse, &descBl, &descBt );
+
+ morse_sequence_destroy( morse, sequence );
return MORSE_SUCCESS;
}
@@ -200,22 +197,29 @@ int MORSE_zlacpy(MORSE_enum uplo, int M, int N,
* @sa MORSE_slacpy_Tile
*
******************************************************************************/
-int MORSE_zlacpy_Tile(MORSE_enum uplo, MORSE_desc_t *A, MORSE_desc_t *B)
+int MORSE_zlacpy_Tile( MORSE_enum uplo, MORSE_desc_t *A, MORSE_desc_t *B )
{
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
MORSE_request_t request = MORSE_REQUEST_INITIALIZER;
+ int status;
morse = morse_context_self();
if (morse == NULL) {
morse_fatal_error("MORSE_zlacpy_Tile", "MORSE not initialized");
return MORSE_ERR_NOT_INITIALIZED;
}
- morse_sequence_create(morse, &sequence);
- MORSE_zlacpy_Tile_Async(uplo, A, B, sequence, &request);
- morse_sequence_wait(morse, sequence);
- morse_sequence_destroy(morse, sequence);
- return MORSE_SUCCESS;
+ morse_sequence_create( morse, &sequence );
+
+ MORSE_zlacpy_Tile_Async( uplo, A, B, sequence, &request );
+
+ MORSE_Desc_Flush( A, sequence );
+ MORSE_Desc_Flush( B, sequence );
+
+ morse_sequence_wait( morse, sequence );
+ status = sequence->status;
+ morse_sequence_destroy( morse, sequence );
+ return status;
}
/**
@@ -245,8 +249,8 @@ int MORSE_zlacpy_Tile(MORSE_enum uplo, MORSE_desc_t *A, MORSE_desc_t *B)
* @sa MORSE_slacpy_Tile_Async
*
******************************************************************************/
-int MORSE_zlacpy_Tile_Async(MORSE_enum uplo, MORSE_desc_t *A, MORSE_desc_t *B,
- MORSE_sequence_t *sequence, MORSE_request_t *request)
+int MORSE_zlacpy_Tile_Async( MORSE_enum uplo, MORSE_desc_t *A, MORSE_desc_t *B,
+ MORSE_sequence_t *sequence, MORSE_request_t *request )
{
MORSE_context_t *morse;
@@ -264,10 +268,12 @@ int MORSE_zlacpy_Tile_Async(MORSE_enum uplo, MORSE_desc_t *A, MORSE_desc_t *B,
return MORSE_ERR_UNALLOCATED;
}
/* Check sequence status */
- if (sequence->status == MORSE_SUCCESS)
+ if (sequence->status == MORSE_SUCCESS) {
request->status = MORSE_SUCCESS;
- else
+ }
+ else {
return morse_request_fail(sequence, request, MORSE_ERR_SEQUENCE_FLUSHED);
+ }
/* Check descriptors for correctness */
if (morse_desc_check(A) != MORSE_SUCCESS) {
@@ -295,7 +301,7 @@ int MORSE_zlacpy_Tile_Async(MORSE_enum uplo, MORSE_desc_t *A, MORSE_desc_t *B,
return MORSE_SUCCESS;
}
- morse_pzlacpy(uplo, A, B, sequence, request);
+ morse_pzlacpy( uplo, A, B, sequence, request );
return MORSE_SUCCESS;
}
diff --git a/compute/zlange.c b/compute/zlange.c
index 82831613fd334461bc4f2a1e1cf437d6eb328a03..e351bb5c7672c141f0e357cb53213f06c39fc6dc 100644
--- a/compute/zlange.c
+++ b/compute/zlange.c
@@ -84,7 +84,7 @@
*
******************************************************************************/
double MORSE_zlange(MORSE_enum norm, int M, int N,
- MORSE_Complex64_t *A, int LDA)
+ MORSE_Complex64_t *A, int LDA )
{
int NB;
int status;
@@ -92,7 +92,7 @@ double MORSE_zlange(MORSE_enum norm, int M, int N,
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
MORSE_request_t request = MORSE_REQUEST_INITIALIZER;
- MORSE_desc_t descA;
+ MORSE_desc_t descAl, descAt;
morse = morse_context_self();
if (morse == NULL) {
@@ -101,7 +101,7 @@ double MORSE_zlange(MORSE_enum norm, int M, int N,
}
/* Check input arguments */
if ( (norm != MorseMaxNorm) && (norm != MorseOneNorm)
- && (norm != MorseInfNorm) && (norm != MorseFrobeniusNorm) ) {
+ && (norm != MorseInfNorm) && (norm != MorseFrobeniusNorm) ) {
morse_error("MORSE_zlange", "illegal value of norm");
return -1;
}
@@ -120,7 +120,7 @@ double MORSE_zlange(MORSE_enum norm, int M, int N,
/* Quick return */
if (chameleon_min(N, M) == 0)
- return (double)0.0;
+ return (double)0.0;
/* Tune NB depending on M, N & NRHS; Set NBNB */
status = morse_tune(MORSE_FUNC_ZGEMM, M, N, 0);
@@ -132,30 +132,25 @@ double MORSE_zlange(MORSE_enum norm, int M, int N,
/* Set NT */
NB = MORSE_NB;
- morse_sequence_create(morse, &sequence);
+ morse_sequence_create( morse, &sequence );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooplap2tile( descA, A, NB, NB, LDA, N, 0, 0, M, N, sequence, &request,
- morse_desc_mat_free(&(descA)) );
-/* } else {*/
-/* morse_ziplap2tile( descA, A, NB, NB, LDA, N, 0, 0, M, N,*/
-/* sequence, &request);*/
-/* }*/
+ /* Submit the matrix conversion */
+ morse_zlap2tile( morse, &descAl, &descAt, MorseDescInput, MorseUpperLower,
+ A, NB, NB, LDA, N, M, N, sequence, &request );
/* Call the tile interface */
- MORSE_zlange_Tile_Async(norm, &descA, &value, sequence, &request);
+ MORSE_zlange_Tile_Async( norm, &descAt, &value, sequence, &request );
+
+ /* Submit the matrix conversion back */
+ morse_ztile2lap( morse, &descAl, &descAt,
+ MorseDescInput, MorseUpperLower, sequence, &request );
+
+ morse_sequence_wait( morse, sequence );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooptile2lap(descA, A, NB, NB, LDA, N, sequence, &request);
- RUNTIME_desc_flush( &descA, sequence );
- morse_sequence_wait(morse, sequence);
- morse_desc_mat_free(&descA);
-/* } else {*/
-/* morse_ziptile2lap( descA, A, NB, NB, LDA, N, sequence, &request);*/
-/* morse_sequence_wait(morse, sequence);*/
-/* }*/
+ /* Cleanup the temporary data */
+ morse_ztile2lap_cleanup( morse, &descAl, &descAt );
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return value;
}
@@ -198,11 +193,12 @@ double MORSE_zlange(MORSE_enum norm, int M, int N,
* @sa MORSE_slange_Tile
*
******************************************************************************/
-double MORSE_zlange_Tile(MORSE_enum norm, MORSE_desc_t *A)
+double MORSE_zlange_Tile(MORSE_enum norm, MORSE_desc_t *A )
{
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
MORSE_request_t request = MORSE_REQUEST_INITIALIZER;
+ int status;
double value;
morse = morse_context_self();
@@ -210,11 +206,16 @@ double MORSE_zlange_Tile(MORSE_enum norm, MORSE_desc_t *A)
morse_fatal_error("MORSE_zlange_Tile", "MORSE not initialized");
return MORSE_ERR_NOT_INITIALIZED;
}
- morse_sequence_create(morse, &sequence);
- MORSE_zlange_Tile_Async(norm, A, &value, sequence, &request);
- morse_sequence_wait(morse, sequence);
- morse_sequence_destroy(morse, sequence);
- return value;
+ morse_sequence_create( morse, &sequence );
+
+ MORSE_zlange_Tile_Async( norm, A, &value, sequence, &request );
+
+ MORSE_Desc_Flush( A, sequence );
+
+ morse_sequence_wait( morse, sequence );
+ status = sequence->status;
+ morse_sequence_destroy( morse, sequence );
+ return ( status == MORSE_SUCCESS ) ? value : (double)status;
}
/**
@@ -244,8 +245,8 @@ double MORSE_zlange_Tile(MORSE_enum norm, MORSE_desc_t *A)
* @sa MORSE_slange_Tile_Async
*
******************************************************************************/
-int MORSE_zlange_Tile_Async(MORSE_enum norm, MORSE_desc_t *A, double *value,
- MORSE_sequence_t *sequence, MORSE_request_t *request)
+int MORSE_zlange_Tile_Async( MORSE_enum norm, MORSE_desc_t *A, double *value,
+ MORSE_sequence_t *sequence, MORSE_request_t *request )
{
MORSE_context_t *morse;
@@ -263,10 +264,12 @@ int MORSE_zlange_Tile_Async(MORSE_enum norm, MORSE_desc_t *A, double *value,
return MORSE_ERR_UNALLOCATED;
}
/* Check sequence status */
- if (sequence->status == MORSE_SUCCESS)
+ if (sequence->status == MORSE_SUCCESS) {
request->status = MORSE_SUCCESS;
- else
+ }
+ else {
return morse_request_fail(sequence, request, MORSE_ERR_SEQUENCE_FLUSHED);
+ }
/* Check descriptors for correctness */
if (morse_desc_check(A) != MORSE_SUCCESS) {
@@ -289,7 +292,7 @@ int MORSE_zlange_Tile_Async(MORSE_enum norm, MORSE_desc_t *A, double *value,
return MORSE_SUCCESS;
}
- morse_pzlange(norm, A, value, sequence, request);
+ morse_pzlange( norm, A, value, sequence, request );
return MORSE_SUCCESS;
}
diff --git a/compute/zlanhe.c b/compute/zlanhe.c
index e666221ace06405bf247ff7faab0d9b2cc7ec18d..e43a8111d29f62922cd24050badb702d29875b32 100644
--- a/compute/zlanhe.c
+++ b/compute/zlanhe.c
@@ -84,7 +84,7 @@
*
******************************************************************************/
double MORSE_zlanhe(MORSE_enum norm, MORSE_enum uplo, int N,
- MORSE_Complex64_t *A, int LDA)
+ MORSE_Complex64_t *A, int LDA )
{
int NB;
int status;
@@ -92,7 +92,7 @@ double MORSE_zlanhe(MORSE_enum norm, MORSE_enum uplo, int N,
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
MORSE_request_t request = MORSE_REQUEST_INITIALIZER;
- MORSE_desc_t descA;
+ MORSE_desc_t descAl, descAt;
morse = morse_context_self();
if (morse == NULL) {
@@ -101,7 +101,7 @@ double MORSE_zlanhe(MORSE_enum norm, MORSE_enum uplo, int N,
}
/* Check input arguments */
if ( (norm != MorseMaxNorm) && (norm != MorseOneNorm)
- && (norm != MorseInfNorm) && (norm != MorseFrobeniusNorm) ) {
+ && (norm != MorseInfNorm) && (norm != MorseFrobeniusNorm) ) {
morse_error("MORSE_zlanhe", "illegal value of norm");
return -1;
}
@@ -120,7 +120,7 @@ double MORSE_zlanhe(MORSE_enum norm, MORSE_enum uplo, int N,
/* Quick return */
if ( N == 0)
- return (double)0.0;
+ return (double)0.0;
/* Tune NB depending on M, N & NRHS; Set NBNB */
status = morse_tune(MORSE_FUNC_ZGEMM, N, N, 0);
@@ -132,30 +132,25 @@ double MORSE_zlanhe(MORSE_enum norm, MORSE_enum uplo, int N,
/* Set NT */
NB = MORSE_NB;
- morse_sequence_create(morse, &sequence);
+ morse_sequence_create( morse, &sequence );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooplap2tile( descA, A, NB, NB, LDA, N, 0, 0, N, N, sequence, &request,
- morse_desc_mat_free(&(descA)) );
-/* } else {*/
-/* morse_ziplap2tile( descA, A, NB, NB, LDA, N, 0, 0, N, N,*/
-/* sequence, &request);*/
-/* }*/
+ /* Submit the matrix conversion */
+ morse_zlap2tile( morse, &descAl, &descAt, MorseDescInput, uplo,
+ A, NB, NB, LDA, N, N, N, sequence, &request );
/* Call the tile interface */
- MORSE_zlanhe_Tile_Async(norm, uplo, &descA, &value, sequence, &request);
+ MORSE_zlanhe_Tile_Async( norm, uplo, &descAt, &value, sequence, &request );
+
+ /* Submit the matrix conversion back */
+ morse_ztile2lap( morse, &descAl, &descAt,
+ MorseDescInput, uplo, sequence, &request );
+
+ morse_sequence_wait( morse, sequence );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooptile2lap(descA, A, NB, NB, LDA, N, sequence, &request);
- RUNTIME_desc_flush( &descA, sequence );
- morse_sequence_wait(morse, sequence);
- morse_desc_mat_free(&descA);
-/* } else {*/
-/* morse_ziptile2lap( descA, A, NB, NB, LDA, N, sequence, &request);*/
-/* morse_sequence_wait(morse, sequence);*/
-/* }*/
+ /* Cleanup the temporary data */
+ morse_ztile2lap_cleanup( morse, &descAl, &descAt );
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return value;
}
@@ -202,11 +197,12 @@ double MORSE_zlanhe(MORSE_enum norm, MORSE_enum uplo, int N,
* @sa MORSE_slanhe_Tile
*
******************************************************************************/
-double MORSE_zlanhe_Tile(MORSE_enum norm, MORSE_enum uplo, MORSE_desc_t *A)
+double MORSE_zlanhe_Tile( MORSE_enum norm, MORSE_enum uplo, MORSE_desc_t *A )
{
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
MORSE_request_t request = MORSE_REQUEST_INITIALIZER;
+ int status;
double value;
morse = morse_context_self();
@@ -214,11 +210,16 @@ double MORSE_zlanhe_Tile(MORSE_enum norm, MORSE_enum uplo, MORSE_desc_t *A)
morse_fatal_error("MORSE_zlanhe_Tile", "MORSE not initialized");
return MORSE_ERR_NOT_INITIALIZED;
}
- morse_sequence_create(morse, &sequence);
- MORSE_zlanhe_Tile_Async(norm, uplo, A, &value, sequence, &request);
- morse_sequence_wait(morse, sequence);
- morse_sequence_destroy(morse, sequence);
- return value;
+ morse_sequence_create( morse, &sequence );
+
+ MORSE_zlanhe_Tile_Async( norm, uplo, A, &value, sequence, &request );
+
+ MORSE_Desc_Flush( A, sequence );
+
+ morse_sequence_wait( morse, sequence );
+ status = sequence->status;
+ morse_sequence_destroy( morse, sequence );
+ return ( status == MORSE_SUCCESS ) ? value : (double)status;
}
/**
@@ -248,8 +249,8 @@ double MORSE_zlanhe_Tile(MORSE_enum norm, MORSE_enum uplo, MORSE_desc_t *A)
* @sa MORSE_slanhe_Tile_Async
*
******************************************************************************/
-int MORSE_zlanhe_Tile_Async(MORSE_enum norm, MORSE_enum uplo, MORSE_desc_t *A, double *value,
- MORSE_sequence_t *sequence, MORSE_request_t *request)
+int MORSE_zlanhe_Tile_Async( MORSE_enum norm, MORSE_enum uplo, MORSE_desc_t *A, double *value,
+ MORSE_sequence_t *sequence, MORSE_request_t *request )
{
MORSE_context_t *morse;
@@ -267,10 +268,12 @@ int MORSE_zlanhe_Tile_Async(MORSE_enum norm, MORSE_enum uplo, MORSE_desc_t *A, d
return MORSE_ERR_UNALLOCATED;
}
/* Check sequence status */
- if (sequence->status == MORSE_SUCCESS)
+ if (sequence->status == MORSE_SUCCESS) {
request->status = MORSE_SUCCESS;
- else
+ }
+ else {
return morse_request_fail(sequence, request, MORSE_ERR_SEQUENCE_FLUSHED);
+ }
/* Check descriptors for correctness */
if (morse_desc_check(A) != MORSE_SUCCESS) {
@@ -297,7 +300,7 @@ int MORSE_zlanhe_Tile_Async(MORSE_enum norm, MORSE_enum uplo, MORSE_desc_t *A, d
return MORSE_SUCCESS;
}
- morse_pzlanhe(norm, uplo, A, value, sequence, request);
+ morse_pzlanhe( norm, uplo, A, value, sequence, request );
return MORSE_SUCCESS;
}
diff --git a/compute/zlansy.c b/compute/zlansy.c
index 35df4fc8912e33279b04fa6af22974ec005a1e11..cb97a551881bb1cdb44580e6f3ca7d6631500964 100644
--- a/compute/zlansy.c
+++ b/compute/zlansy.c
@@ -84,7 +84,7 @@
*
******************************************************************************/
double MORSE_zlansy(MORSE_enum norm, MORSE_enum uplo, int N,
- MORSE_Complex64_t *A, int LDA)
+ MORSE_Complex64_t *A, int LDA )
{
int NB;
int status;
@@ -92,7 +92,7 @@ double MORSE_zlansy(MORSE_enum norm, MORSE_enum uplo, int N,
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
MORSE_request_t request = MORSE_REQUEST_INITIALIZER;
- MORSE_desc_t descA;
+ MORSE_desc_t descAl, descAt;
morse = morse_context_self();
if (morse == NULL) {
@@ -101,7 +101,7 @@ double MORSE_zlansy(MORSE_enum norm, MORSE_enum uplo, int N,
}
/* Check input arguments */
if ( (norm != MorseMaxNorm) && (norm != MorseOneNorm)
- && (norm != MorseInfNorm) && (norm != MorseFrobeniusNorm) ) {
+ && (norm != MorseInfNorm) && (norm != MorseFrobeniusNorm) ) {
morse_error("MORSE_zlansy", "illegal value of norm");
return -1;
}
@@ -120,7 +120,7 @@ double MORSE_zlansy(MORSE_enum norm, MORSE_enum uplo, int N,
/* Quick return */
if ( N == 0)
- return (double)0.0;
+ return (double)0.0;
/* Tune NB depending on M, N & NRHS; Set NBNB */
status = morse_tune(MORSE_FUNC_ZGEMM, N, N, 0);
@@ -132,28 +132,25 @@ double MORSE_zlansy(MORSE_enum norm, MORSE_enum uplo, int N,
/* Set NT */
NB = MORSE_NB;
- morse_sequence_create(morse, &sequence);
+ morse_sequence_create( morse, &sequence );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooplap2tile( descA, A, NB, NB, LDA, N, 0, 0, N, N, sequence, &request,
- morse_desc_mat_free(&(descA)) );
-/* } else {*/
-/* morse_ziplap2tile( descA, A, NB, NB, LDA, N, 0, 0, N, N,*/
-/* sequence, &request);*/
-/* }*/
+ /* Submit the matrix conversion */
+ morse_zlap2tile( morse, &descAl, &descAt, MorseDescInput, uplo,
+ A, NB, NB, LDA, N, N, N, sequence, &request );
/* Call the tile interface */
- MORSE_zlansy_Tile_Async(norm, uplo, &descA, &value, sequence, &request);
+ MORSE_zlansy_Tile_Async( norm, uplo, &descAt, &value, sequence, &request );
+
+ /* Submit the matrix conversion back */
+ morse_ztile2lap( morse, &descAl, &descAt,
+ MorseDescInput, uplo, sequence, &request );
+
+ morse_sequence_wait( morse, sequence );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_sequence_wait(morse, sequence);
- morse_desc_mat_free(&descA);
-/* } else {*/
-/* morse_ziptile2lap( descA, A, NB, NB, LDA, N, sequence, &request);*/
-/* morse_sequence_wait(morse, sequence);*/
-/* }*/
+ /* Cleanup the temporary data */
+ morse_ztile2lap_cleanup( morse, &descAl, &descAt );
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return value;
}
@@ -200,11 +197,12 @@ double MORSE_zlansy(MORSE_enum norm, MORSE_enum uplo, int N,
* @sa MORSE_slansy_Tile
*
******************************************************************************/
-double MORSE_zlansy_Tile(MORSE_enum norm, MORSE_enum uplo, MORSE_desc_t *A)
+double MORSE_zlansy_Tile( MORSE_enum norm, MORSE_enum uplo, MORSE_desc_t *A )
{
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
MORSE_request_t request = MORSE_REQUEST_INITIALIZER;
+ int status;
double value;
morse = morse_context_self();
@@ -212,11 +210,16 @@ double MORSE_zlansy_Tile(MORSE_enum norm, MORSE_enum uplo, MORSE_desc_t *A)
morse_fatal_error("MORSE_zlansy_Tile", "MORSE not initialized");
return MORSE_ERR_NOT_INITIALIZED;
}
- morse_sequence_create(morse, &sequence);
- MORSE_zlansy_Tile_Async(norm, uplo, A, &value, sequence, &request);
- morse_sequence_wait(morse, sequence);
- morse_sequence_destroy(morse, sequence);
- return value;
+ morse_sequence_create( morse, &sequence );
+
+ MORSE_zlansy_Tile_Async( norm, uplo, A, &value, sequence, &request );
+
+ MORSE_Desc_Flush( A, sequence );
+
+ morse_sequence_wait( morse, sequence );
+ status = sequence->status;
+ morse_sequence_destroy( morse, sequence );
+ return ( status == MORSE_SUCCESS ) ? value : (double)status;
}
/**
@@ -246,8 +249,8 @@ double MORSE_zlansy_Tile(MORSE_enum norm, MORSE_enum uplo, MORSE_desc_t *A)
* @sa MORSE_slansy_Tile_Async
*
******************************************************************************/
-int MORSE_zlansy_Tile_Async(MORSE_enum norm, MORSE_enum uplo, MORSE_desc_t *A, double *value,
- MORSE_sequence_t *sequence, MORSE_request_t *request)
+int MORSE_zlansy_Tile_Async( MORSE_enum norm, MORSE_enum uplo, MORSE_desc_t *A, double *value,
+ MORSE_sequence_t *sequence, MORSE_request_t *request )
{
MORSE_context_t *morse;
@@ -265,10 +268,12 @@ int MORSE_zlansy_Tile_Async(MORSE_enum norm, MORSE_enum uplo, MORSE_desc_t *A, d
return MORSE_ERR_UNALLOCATED;
}
/* Check sequence status */
- if (sequence->status == MORSE_SUCCESS)
+ if (sequence->status == MORSE_SUCCESS) {
request->status = MORSE_SUCCESS;
- else
+ }
+ else {
return morse_request_fail(sequence, request, MORSE_ERR_SEQUENCE_FLUSHED);
+ }
/* Check descriptors for correctness */
if (morse_desc_check(A) != MORSE_SUCCESS) {
@@ -295,7 +300,7 @@ int MORSE_zlansy_Tile_Async(MORSE_enum norm, MORSE_enum uplo, MORSE_desc_t *A, d
return MORSE_SUCCESS;
}
- morse_pzlansy(norm, uplo, A, value, sequence, request);
+ morse_pzlansy( norm, uplo, A, value, sequence, request );
return MORSE_SUCCESS;
}
diff --git a/compute/zlantr.c b/compute/zlantr.c
index 3f5e8f1eda3637f3c91541e8da96c816a732dbf6..f803abd84474cbe35a720c2bd5e98d20bd727a5d 100644
--- a/compute/zlantr.c
+++ b/compute/zlantr.c
@@ -96,7 +96,7 @@
*
******************************************************************************/
double MORSE_zlantr(MORSE_enum norm, MORSE_enum uplo, MORSE_enum diag,
- int M, int N, MORSE_Complex64_t *A, int LDA)
+ int M, int N, MORSE_Complex64_t *A, int LDA )
{
int NB;
int status;
@@ -104,7 +104,7 @@ double MORSE_zlantr(MORSE_enum norm, MORSE_enum uplo, MORSE_enum diag,
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
MORSE_request_t request = MORSE_REQUEST_INITIALIZER;
- MORSE_desc_t descA;
+ MORSE_desc_t descAl, descAt;
morse = morse_context_self();
if (morse == NULL) {
@@ -113,15 +113,15 @@ double MORSE_zlantr(MORSE_enum norm, MORSE_enum uplo, MORSE_enum diag,
}
/* Check input arguments */
if ( (norm != MorseMaxNorm) && (norm != MorseOneNorm)
- && (norm != MorseInfNorm) && (norm != MorseFrobeniusNorm) ) {
+ && (norm != MorseInfNorm) && (norm != MorseFrobeniusNorm) ) {
morse_error("MORSE_zlantr", "illegal value of norm");
return -1;
}
- if (uplo != MorseUpper && uplo != MorseLower) {
+ if ((uplo != MorseUpper) && (uplo != MorseLower)) {
morse_error("MORSE_zlantr", "illegal value of uplo");
return -2;
}
- if (diag != MorseUnit && diag != MorseNonUnit) {
+ if ((diag != MorseUnit) && (diag != MorseNonUnit)) {
morse_error("MORSE_zlantr", "illegal value of diag");
return -3;
}
@@ -140,7 +140,7 @@ double MORSE_zlantr(MORSE_enum norm, MORSE_enum uplo, MORSE_enum diag,
/* Quick return */
if (chameleon_min(N, M) == 0)
- return (double)0.0;
+ return (double)0.0;
/* Tune NB depending on M, N & NRHS; Set NBNB */
status = morse_tune(MORSE_FUNC_ZGEMM, M, N, 0);
@@ -152,28 +152,25 @@ double MORSE_zlantr(MORSE_enum norm, MORSE_enum uplo, MORSE_enum diag,
/* Set NT */
NB = MORSE_NB;
- morse_sequence_create(morse, &sequence);
+ morse_sequence_create( morse, &sequence );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooplap2tile( descA, A, NB, NB, LDA, N, 0, 0, M, N, sequence, &request,
- morse_desc_mat_free(&(descA)) );
-/* } else {*/
-/* morse_ziplap2tile( descA, A, NB, NB, LDA, N, 0, 0, M, N,*/
-/* sequence, &request);*/
-/* }*/
+ /* Submit the matrix conversion */
+ morse_zlap2tile( morse, &descAl, &descAt, MorseDescInput, uplo,
+ A, NB, NB, LDA, N, M, N, sequence, &request );
/* Call the tile interface */
- MORSE_zlantr_Tile_Async(norm, uplo, diag, &descA, &value, sequence, &request);
+ MORSE_zlantr_Tile_Async( norm, uplo, diag, &descAt, &value, sequence, &request );
+
+ /* Submit the matrix conversion back */
+ morse_ztile2lap( morse, &descAl, &descAt,
+ MorseDescInput, uplo, sequence, &request );
+
+ morse_sequence_wait( morse, sequence );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_sequence_wait(morse, sequence);
- morse_desc_mat_free(&descA);
-/* } else {*/
-/* morse_ziptile2lap( descA, A, NB, NB, LDA, N, sequence, &request);*/
-/* morse_sequence_wait(morse, sequence);*/
-/* }*/
+ /* Cleanup the temporary data */
+ morse_ztile2lap_cleanup( morse, &descAl, &descAt );
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return value;
}
@@ -222,11 +219,12 @@ double MORSE_zlantr(MORSE_enum norm, MORSE_enum uplo, MORSE_enum diag,
* @sa MORSE_slantr_Tile
*
******************************************************************************/
-double MORSE_zlantr_Tile(MORSE_enum norm, MORSE_enum uplo, MORSE_enum diag, MORSE_desc_t *A)
+double MORSE_zlantr_Tile(MORSE_enum norm, MORSE_enum uplo, MORSE_enum diag, MORSE_desc_t *A )
{
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
MORSE_request_t request = MORSE_REQUEST_INITIALIZER;
+ int status;
double value;
morse = morse_context_self();
@@ -234,11 +232,16 @@ double MORSE_zlantr_Tile(MORSE_enum norm, MORSE_enum uplo, MORSE_enum diag, MORS
morse_fatal_error("MORSE_zlantr_Tile", "MORSE not initialized");
return MORSE_ERR_NOT_INITIALIZED;
}
- morse_sequence_create(morse, &sequence);
- MORSE_zlantr_Tile_Async(norm, uplo, diag, A, &value, sequence, &request);
- morse_sequence_wait(morse, sequence);
- morse_sequence_destroy(morse, sequence);
- return value;
+ morse_sequence_create( morse, &sequence );
+
+ MORSE_zlantr_Tile_Async( norm, uplo, diag, A, &value, sequence, &request );
+
+ MORSE_Desc_Flush( A, sequence );
+
+ morse_sequence_wait( morse, sequence );
+ status = sequence->status;
+ morse_sequence_destroy( morse, sequence );
+ return ( status == MORSE_SUCCESS ) ? value : (double)status;
}
/**
@@ -268,9 +271,9 @@ double MORSE_zlantr_Tile(MORSE_enum norm, MORSE_enum uplo, MORSE_enum diag, MORS
* @sa MORSE_slantr_Tile_Async
*
******************************************************************************/
-int MORSE_zlantr_Tile_Async(MORSE_enum norm, MORSE_enum uplo, MORSE_enum diag,
- MORSE_desc_t *A, double *value,
- MORSE_sequence_t *sequence, MORSE_request_t *request)
+int MORSE_zlantr_Tile_Async( MORSE_enum norm, MORSE_enum uplo, MORSE_enum diag,
+ MORSE_desc_t *A, double *value,
+ MORSE_sequence_t *sequence, MORSE_request_t *request )
{
MORSE_context_t *morse;
@@ -288,10 +291,12 @@ int MORSE_zlantr_Tile_Async(MORSE_enum norm, MORSE_enum uplo, MORSE_enum diag,
return MORSE_ERR_UNALLOCATED;
}
/* Check sequence status */
- if (sequence->status == MORSE_SUCCESS)
+ if (sequence->status == MORSE_SUCCESS) {
request->status = MORSE_SUCCESS;
- else
+ }
+ else {
return morse_request_fail(sequence, request, MORSE_ERR_SEQUENCE_FLUSHED);
+ }
/* Check descriptors for correctness */
if (morse_desc_check(A) != MORSE_SUCCESS) {
@@ -308,11 +313,11 @@ int MORSE_zlantr_Tile_Async(MORSE_enum norm, MORSE_enum uplo, MORSE_enum diag,
morse_error("MORSE_zlantr_Tile", "illegal value of norm");
return morse_request_fail(sequence, request, MORSE_ERR_ILLEGAL_VALUE);
}
- if (uplo != MorseUpper && uplo != MorseLower) {
+ if ((uplo != MorseUpper) && (uplo != MorseLower)) {
morse_error("MORSE_zlantr_Tile", "illegal value of uplo");
return morse_request_fail(sequence, request, MORSE_ERR_ILLEGAL_VALUE);
}
- if (diag != MorseUnit && diag != MorseNonUnit) {
+ if ((diag != MorseUnit) && (diag != MorseNonUnit)) {
morse_error("MORSE_zlantr_Tile", "illegal value of diag");
return morse_request_fail(sequence, request, MORSE_ERR_ILLEGAL_VALUE);
}
@@ -323,7 +328,7 @@ int MORSE_zlantr_Tile_Async(MORSE_enum norm, MORSE_enum uplo, MORSE_enum diag,
return MORSE_SUCCESS;
}
- morse_pzlantr(norm, uplo, diag, A, value, sequence, request);
+ morse_pzlantr( norm, uplo, diag, A, value, sequence, request );
return MORSE_SUCCESS;
}
diff --git a/compute/zlascal.c b/compute/zlascal.c
index c86ca1da02ebad9066a30f628521088f95ad0d94..b37d8f61881884d7b7dcd943f7c7b5b44c55f0be 100644
--- a/compute/zlascal.c
+++ b/compute/zlascal.c
@@ -62,12 +62,12 @@
* @sa MORSE_slascal
*
******************************************************************************/
-int MORSE_zlascal(MORSE_enum uplo, int M, int N,
- MORSE_Complex64_t alpha, MORSE_Complex64_t *A, int LDA)
+int MORSE_zlascal( MORSE_enum uplo, int M, int N,
+ MORSE_Complex64_t alpha, MORSE_Complex64_t *A, int LDA )
{
int NB;
int status;
- MORSE_desc_t descA;
+ MORSE_desc_t descAl, descAt;
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
MORSE_request_t request = MORSE_REQUEST_INITIALIZER;
@@ -79,7 +79,7 @@ int MORSE_zlascal(MORSE_enum uplo, int M, int N,
}
/* Check input arguments */
- if (uplo != MorseUpper && uplo != MorseLower && uplo != MorseUpperLower) {
+ if ((uplo != MorseUpper) && (uplo != MorseLower) && (uplo != MorseUpperLower)) {
morse_error("MORSE_zlascal", "illegal value of uplo");
return -1;
}
@@ -111,31 +111,26 @@ int MORSE_zlascal(MORSE_enum uplo, int M, int N,
/* Set MT & NT & KT */
NB = MORSE_NB;
- morse_sequence_create(morse, &sequence);
+ morse_sequence_create( morse, &sequence );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooplap2tile( descA, A, NB, NB, LDA, N, 0, 0, M, N, sequence, &request,
- morse_desc_mat_free(&(descA)) );
-/* } else {*/
-/* morse_ziplap2tile( descA, A, NB, NB, LDA, N , 0, 0, M, N,*/
-/* sequence, &request);*/
-/* }*/
+ /* Submit the matrix conversion */
+ morse_zlap2tile( morse, &descAl, &descAt, MorseDescInout, uplo,
+ A, NB, NB, LDA, N, M, N, sequence, &request );
/* Call the tile interface */
- MORSE_zlascal_Tile_Async(
- uplo, alpha, &descA, sequence, &request);
+ MORSE_zlascal_Tile_Async( uplo, alpha, &descAt, sequence, &request );
+
+ /* Submit the matrix conversion back */
+ morse_ztile2lap( morse, &descAl, &descAt,
+ MorseDescInout, uplo, sequence, &request );
+
+ morse_sequence_wait( morse, sequence );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooptile2lap( descA, A, NB, NB, LDA, N, sequence, &request);
- RUNTIME_sequence_wait(morse, sequence);
- morse_desc_mat_free(&descA);
-/* } else {*/
-/* morse_ziptile2lap( descA, A, NB, NB, LDA, N, sequence, &request);*/
-/* morse_dynamic_sync();*/
-/* }*/
+ /* Cleanup the temporary data */
+ morse_ztile2lap_cleanup( morse, &descAl, &descAt );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -179,8 +174,8 @@ int MORSE_zlascal(MORSE_enum uplo, int M, int N,
* @sa MORSE_slascal_Tile
*
******************************************************************************/
-int MORSE_zlascal_Tile(MORSE_enum uplo,
- MORSE_Complex64_t alpha, MORSE_desc_t *A)
+int MORSE_zlascal_Tile( MORSE_enum uplo,
+ MORSE_Complex64_t alpha, MORSE_desc_t *A )
{
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
@@ -192,11 +187,15 @@ int MORSE_zlascal_Tile(MORSE_enum uplo,
morse_fatal_error("MORSE_zlascal_Tile", "MORSE not initialized");
return MORSE_ERR_NOT_INITIALIZED;
}
- morse_sequence_create(morse, &sequence);
- MORSE_zlascal_Tile_Async(uplo, alpha, A, sequence, &request);
- RUNTIME_sequence_wait(morse, sequence);
+ morse_sequence_create( morse, &sequence );
+
+ MORSE_zlascal_Tile_Async( uplo, alpha, A, sequence, &request );
+
+ MORSE_Desc_Flush( A, sequence );
+
+ morse_sequence_wait( morse, sequence );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -229,9 +228,9 @@ int MORSE_zlascal_Tile(MORSE_enum uplo,
* @sa MORSE_slascal_Tile_Async
*
******************************************************************************/
-int MORSE_zlascal_Tile_Async(MORSE_enum uplo,
+int MORSE_zlascal_Tile_Async( MORSE_enum uplo,
MORSE_Complex64_t alpha, MORSE_desc_t *A,
- MORSE_sequence_t *sequence, MORSE_request_t *request)
+ MORSE_sequence_t *sequence, MORSE_request_t *request )
{
MORSE_context_t *morse;
MORSE_desc_t descA;
@@ -250,10 +249,12 @@ int MORSE_zlascal_Tile_Async(MORSE_enum uplo,
return MORSE_ERR_UNALLOCATED;
}
/* Check sequence status */
- if (sequence->status == MORSE_SUCCESS)
+ if (sequence->status == MORSE_SUCCESS) {
request->status = MORSE_SUCCESS;
- else
+ }
+ else {
return morse_request_fail(sequence, request, MORSE_ERR_SEQUENCE_FLUSHED);
+ }
/* Check descriptors for correctness */
if (morse_desc_check(A) != MORSE_SUCCESS) {
@@ -263,7 +264,7 @@ int MORSE_zlascal_Tile_Async(MORSE_enum uplo,
descA = *A;
}
/* Check input arguments */
- if (uplo != MorseUpper && uplo != MorseLower && uplo != MorseUpperLower) {
+ if ((uplo != MorseUpper) && (uplo != MorseLower) && (uplo != MorseUpperLower)) {
morse_error("MORSE_zlascal", "illegal value of uplo");
return morse_request_fail(sequence, request, MORSE_ERR_ILLEGAL_VALUE);
}
@@ -276,9 +277,11 @@ int MORSE_zlascal_Tile_Async(MORSE_enum uplo,
/* Quick return */
if ( (descA.m == 0) || (descA.n == 0) ||
(alpha == (MORSE_Complex64_t)1.0) )
+ {
return MORSE_SUCCESS;
+ }
- morse_pzlascal( uplo, alpha, A, sequence, request);
+ morse_pzlascal( uplo, alpha, A, sequence, request );
return MORSE_SUCCESS;
}
diff --git a/compute/zlaset.c b/compute/zlaset.c
index fab26797f9a563522dcb8d744d81070a9bc5dc0b..41093bc7d0468aa7cf7e5b0c7df96d7da21b69e8 100644
--- a/compute/zlaset.c
+++ b/compute/zlaset.c
@@ -75,16 +75,16 @@
* @sa MORSE_slaset
*
******************************************************************************/
-int MORSE_zlaset(MORSE_enum uplo, int M, int N,
+int MORSE_zlaset( MORSE_enum uplo, int M, int N,
MORSE_Complex64_t alpha, MORSE_Complex64_t beta,
- MORSE_Complex64_t *A, int LDA)
+ MORSE_Complex64_t *A, int LDA )
{
int NB;
int status;
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
MORSE_request_t request = MORSE_REQUEST_INITIALIZER;
- MORSE_desc_t descA;
+ MORSE_desc_t descAl, descAt;
morse = morse_context_self();
if (morse == NULL) {
@@ -113,7 +113,7 @@ int MORSE_zlaset(MORSE_enum uplo, int M, int N,
/* Quick return */
if (chameleon_min(N, M) == 0)
- return (double)0.0;
+ return (double)0.0;
/* Tune NB depending on M, N & NRHS; Set NBNB */
status = morse_tune(MORSE_FUNC_ZGEMM, M, N, 0);
@@ -125,28 +125,25 @@ int MORSE_zlaset(MORSE_enum uplo, int M, int N,
/* Set NT */
NB = MORSE_NB;
- morse_sequence_create(morse, &sequence);
+ morse_sequence_create( morse, &sequence );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooplap2tile( descA, A, NB, NB, LDA, N, 0, 0, M, N, sequence, &request,
- morse_desc_mat_free(&(descA)) );
-/* } else {*/
-/* morse_ziplap2tile( descA, A, NB, NB, LDA, N, 0, 0, M, N,*/
-/* sequence, &request);*/
-/* }*/
+ /* Submit the matrix conversion */
+ morse_zlap2tile( morse, &descAl, &descAt, MorseDescInout, uplo,
+ A, NB, NB, LDA, N, M, N, sequence, &request );
/* Call the tile interface */
- MORSE_zlaset_Tile_Async(uplo, alpha, beta, &descA, sequence, &request);
+ MORSE_zlaset_Tile_Async( uplo, alpha, beta, &descAt, sequence, &request );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_sequence_wait(morse, sequence);
- morse_desc_mat_free(&descA);
-/* } else {*/
-/* morse_ziptile2lap( descA, A, NB, NB, LDA, N, sequence, &request);*/
-/* morse_sequence_wait(morse, sequence);*/
-/* }*/
+ /* Submit the matrix conversion back */
+ morse_ztile2lap( morse, &descAl, &descAt,
+ MorseDescInout, uplo, sequence, &request );
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_wait( morse, sequence );
+
+ /* Cleanup the temporary data */
+ morse_ztile2lap_cleanup( morse, &descAl, &descAt );
+
+ morse_sequence_destroy( morse, sequence );
return MORSE_SUCCESS;
}
@@ -187,24 +184,30 @@ int MORSE_zlaset(MORSE_enum uplo, int M, int N,
* @sa MORSE_slaset_Tile
*
******************************************************************************/
-int MORSE_zlaset_Tile(MORSE_enum uplo,
+int MORSE_zlaset_Tile( MORSE_enum uplo,
MORSE_Complex64_t alpha, MORSE_Complex64_t beta,
- MORSE_desc_t *A)
+ MORSE_desc_t *A )
{
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
MORSE_request_t request = MORSE_REQUEST_INITIALIZER;
+ int status;
morse = morse_context_self();
if (morse == NULL) {
morse_fatal_error("MORSE_zlaset_Tile", "MORSE not initialized");
return MORSE_ERR_NOT_INITIALIZED;
}
- morse_sequence_create(morse, &sequence);
- MORSE_zlaset_Tile_Async(uplo, alpha, beta, A, sequence, &request);
- morse_sequence_wait(morse, sequence);
- morse_sequence_destroy(morse, sequence);
- return MORSE_SUCCESS;
+ morse_sequence_create( morse, &sequence );
+
+ MORSE_zlaset_Tile_Async( uplo, alpha, beta, A, sequence, &request );
+
+ MORSE_Desc_Flush( A, sequence );
+
+ morse_sequence_wait( morse, sequence );
+ status = sequence->status;
+ morse_sequence_destroy( morse, sequence );
+ return status;
}
/**
@@ -234,10 +237,10 @@ int MORSE_zlaset_Tile(MORSE_enum uplo,
* @sa MORSE_slaset_Tile_Async
*
******************************************************************************/
-int MORSE_zlaset_Tile_Async(MORSE_enum uplo,
+int MORSE_zlaset_Tile_Async( MORSE_enum uplo,
MORSE_Complex64_t alpha, MORSE_Complex64_t beta,
MORSE_desc_t *A,
- MORSE_sequence_t *sequence, MORSE_request_t *request)
+ MORSE_sequence_t *sequence, MORSE_request_t *request )
{
MORSE_context_t *morse;
@@ -255,10 +258,12 @@ int MORSE_zlaset_Tile_Async(MORSE_enum uplo,
return MORSE_ERR_UNALLOCATED;
}
/* Check sequence status */
- if (sequence->status == MORSE_SUCCESS)
+ if (sequence->status == MORSE_SUCCESS) {
request->status = MORSE_SUCCESS;
- else
+ }
+ else {
return morse_request_fail(sequence, request, MORSE_ERR_SEQUENCE_FLUSHED);
+ }
/* Check descriptors for correctness */
if (morse_desc_check(A) != MORSE_SUCCESS) {
@@ -282,7 +287,7 @@ int MORSE_zlaset_Tile_Async(MORSE_enum uplo,
return MORSE_SUCCESS;
}
- morse_pzlaset(uplo, alpha, beta, A, sequence, request);
+ morse_pzlaset( uplo, alpha, beta, A, sequence, request );
return MORSE_SUCCESS;
}
diff --git a/compute/zlauum.c b/compute/zlauum.c
index 8ef3f032d56c4805b9e8fad10267a7cc1eb97a26..7c29cf3908488ddd67fa090718b3daaab92556dd 100644
--- a/compute/zlauum.c
+++ b/compute/zlauum.c
@@ -77,23 +77,23 @@
* @sa MORSE_zpotri
*
******************************************************************************/
-int MORSE_zlauum(MORSE_enum uplo, int N,
- MORSE_Complex64_t *A, int LDA)
+int MORSE_zlauum( MORSE_enum uplo, int N,
+ MORSE_Complex64_t *A, int LDA )
{
int NB;
int status;
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
MORSE_request_t request = MORSE_REQUEST_INITIALIZER;
- MORSE_desc_t descA;
-
+ MORSE_desc_t descAl, descAt;
+
morse = morse_context_self();
if (morse == NULL) {
morse_fatal_error("MORSE_zlauum", "MORSE not initialized");
return MORSE_ERR_NOT_INITIALIZED;
}
/* Check input arguments */
- if (uplo != MorseUpper && uplo != MorseLower) {
+ if ((uplo != MorseUpper) && (uplo != MorseLower)) {
morse_error("MORSE_zlauum", "illegal value of uplo");
return -1;
}
@@ -119,30 +119,26 @@ int MORSE_zlauum(MORSE_enum uplo, int N,
/* Set NT */
NB = MORSE_NB;
- morse_sequence_create(morse, &sequence);
+ morse_sequence_create( morse, &sequence );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooplap2tile( descA, A, NB, NB, LDA, N, 0, 0, N, N, sequence, &request,
- morse_desc_mat_free(&(descA)) );
-/* } else {*/
-/* morse_ziplap2tile( descA, A, NB, NB, LDA, N, 0, 0, N, N,*/
-/* sequence, &request);*/
-/* }*/
+ /* Submit the matrix conversion */
+ morse_zlap2tile( morse, &descAl, &descAt, MorseDescInout, uplo,
+ A, NB, NB, LDA, N, N, N, sequence, &request );
/* Call the tile interface */
- MORSE_zlauum_Tile_Async(uplo, &descA, sequence, &request);
+ MORSE_zlauum_Tile_Async( uplo, &descAt, sequence, &request );
+
+ /* Submit the matrix conversion back */
+ morse_ztile2lap( morse, &descAl, &descAt,
+ MorseDescInout, uplo, sequence, &request );
+
+ morse_sequence_wait( morse, sequence );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooptile2lap(descA, A, NB, NB, LDA, N, sequence, &request);
- morse_sequence_wait(morse, sequence);
- morse_desc_mat_free(&descA);
-/* } else {*/
-/* morse_ziptile2lap( descA, A, NB, NB, LDA, N, sequence, &request);*/
-/* morse_sequence_wait(morse, sequence);*/
-/* }*/
+ /* Cleanup the temporary data */
+ morse_ztile2lap_cleanup( morse, &descAl, &descAt );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -187,7 +183,7 @@ int MORSE_zlauum(MORSE_enum uplo, int N,
* @sa MORSE_zpotri_Tile
*
******************************************************************************/
-int MORSE_zlauum_Tile(MORSE_enum uplo, MORSE_desc_t *A)
+int MORSE_zlauum_Tile( MORSE_enum uplo, MORSE_desc_t *A )
{
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
@@ -199,13 +195,15 @@ int MORSE_zlauum_Tile(MORSE_enum uplo, MORSE_desc_t *A)
morse_fatal_error("MORSE_zlauum_Tile", "MORSE not initialized");
return MORSE_ERR_NOT_INITIALIZED;
}
- morse_sequence_create(morse, &sequence);
- MORSE_zlauum_Tile_Async(uplo, A, sequence, &request);
- RUNTIME_desc_flush( A, sequence );
- morse_sequence_wait(morse, sequence);
-
+ morse_sequence_create( morse, &sequence );
+
+ MORSE_zlauum_Tile_Async( uplo, A, sequence, &request );
+
+ MORSE_Desc_Flush( A, sequence );
+
+ morse_sequence_wait( morse, sequence );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -240,8 +238,8 @@ int MORSE_zlauum_Tile(MORSE_enum uplo, MORSE_desc_t *A)
* @sa MORSE_zpotri_Tile_Async
*
******************************************************************************/
-int MORSE_zlauum_Tile_Async(MORSE_enum uplo, MORSE_desc_t *A,
- MORSE_sequence_t *sequence, MORSE_request_t *request)
+int MORSE_zlauum_Tile_Async( MORSE_enum uplo, MORSE_desc_t *A,
+ MORSE_sequence_t *sequence, MORSE_request_t *request )
{
MORSE_context_t *morse;
@@ -259,10 +257,12 @@ int MORSE_zlauum_Tile_Async(MORSE_enum uplo, MORSE_desc_t *A,
return MORSE_ERR_UNALLOCATED;
}
/* Check sequence status */
- if (sequence->status == MORSE_SUCCESS)
+ if (sequence->status == MORSE_SUCCESS) {
request->status = MORSE_SUCCESS;
- else
+ }
+ else {
return morse_request_fail(sequence, request, MORSE_ERR_SEQUENCE_FLUSHED);
+ }
/* Check descriptors for correctness */
if (morse_desc_check(A) != MORSE_SUCCESS) {
@@ -274,16 +274,16 @@ int MORSE_zlauum_Tile_Async(MORSE_enum uplo, MORSE_desc_t *A,
morse_error("MORSE_zlauum_Tile", "only square tiles supported");
return morse_request_fail(sequence, request, MORSE_ERR_ILLEGAL_VALUE);
}
- if (uplo != MorseUpper && uplo != MorseLower) {
+ if ((uplo != MorseUpper) && (uplo != MorseLower)) {
morse_error("MORSE_zlauum_Tile", "illegal value of uplo");
return morse_request_fail(sequence, request, -1);
}
/* Quick return */
-/*
- if (chameleon_max(N, 0) == 0)
- return MORSE_SUCCESS;
-*/
- morse_pzlauum(uplo, A, sequence, request);
+ /*
+ if (chameleon_max(N, 0) == 0)
+ return MORSE_SUCCESS;
+ */
+ morse_pzlauum( uplo, A, sequence, request );
return MORSE_SUCCESS;
}
diff --git a/compute/zplghe.c b/compute/zplghe.c
index 0d7af992710735c1e17462eb6e5d6d8f71bb6c36..5f8deeee1f8174ba6eb659b97d140243beefb537 100644
--- a/compute/zplghe.c
+++ b/compute/zplghe.c
@@ -84,7 +84,7 @@ int MORSE_zplghe( double bump, MORSE_enum uplo, int N,
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
MORSE_request_t request = MORSE_REQUEST_INITIALIZER;
- MORSE_desc_t descA;
+ MORSE_desc_t descAl, descAt;
morse = morse_context_self();
if (morse == NULL) {
@@ -113,19 +113,26 @@ int MORSE_zplghe( double bump, MORSE_enum uplo, int N,
/* Set NT */
NB = MORSE_NB;
- morse_sequence_create(morse, &sequence);
+ morse_sequence_create( morse, &sequence );
- morse_zdesc_alloc(descA, NB, NB, LDA, N, 0, 0, N, N, morse_desc_mat_free(&descA));
+ /* Submit the matrix conversion */
+ morse_zlap2tile( morse, &descAl, &descAt, MorseDescOutput, uplo,
+ A, NB, NB, LDA, N, N, N, sequence, &request );
/* Call the tile interface */
- MORSE_zplghe_Tile_Async( bump, uplo, &descA, seed, sequence, &request );
+ MORSE_zplghe_Tile_Async( bump, uplo, &descAt, seed, sequence, &request );
- morse_zooptile2lap(descA, A, NB, NB, LDA, N, sequence, &request);
- morse_sequence_wait(morse, sequence);
- morse_desc_mat_free(&descA);
+ /* Submit the matrix conversion back */
+ morse_ztile2lap( morse, &descAl, &descAt,
+ MorseDescOutput, uplo, sequence, &request );
+
+ morse_sequence_wait( morse, sequence );
+
+ /* Cleanup the temporary data */
+ morse_ztile2lap_cleanup( morse, &descAl, &descAt );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -184,11 +191,15 @@ int MORSE_zplghe_Tile( double bump, MORSE_enum uplo, MORSE_desc_t *A,
morse_fatal_error("MORSE_zplghe_Tile", "MORSE not initialized");
return MORSE_ERR_NOT_INITIALIZED;
}
- morse_sequence_create(morse, &sequence);
+ morse_sequence_create( morse, &sequence );
+
MORSE_zplghe_Tile_Async( bump, uplo, A, seed, sequence, &request );
- morse_sequence_wait(morse, sequence);
+
+ MORSE_Desc_Flush( A, sequence );
+
+ morse_sequence_wait( morse, sequence );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -244,10 +255,12 @@ int MORSE_zplghe_Tile_Async( double bump,
return MORSE_ERR_UNALLOCATED;
}
/* Check sequence status */
- if (sequence->status == MORSE_SUCCESS)
+ if (sequence->status == MORSE_SUCCESS) {
request->status = MORSE_SUCCESS;
- else
+ }
+ else {
return morse_request_fail(sequence, request, MORSE_ERR_SEQUENCE_FLUSHED);
+ }
/* Check descriptors for correctness */
if (morse_desc_check(A) != MORSE_SUCCESS) {
@@ -264,7 +277,7 @@ int MORSE_zplghe_Tile_Async( double bump,
if (chameleon_min( A->m, A->n ) == 0)
return MORSE_SUCCESS;
- morse_pzplghe(bump, uplo, A, seed, sequence, request);
+ morse_pzplghe( bump, uplo, A, seed, sequence, request );
return MORSE_SUCCESS;
}
diff --git a/compute/zplgsy.c b/compute/zplgsy.c
index 7c8e8aeda0f5d1e96d384d40b529751a266875c8..31ec6fdc714d98fd2d8e0ca6a5fd56c2c13be896 100644
--- a/compute/zplgsy.c
+++ b/compute/zplgsy.c
@@ -84,7 +84,7 @@ int MORSE_zplgsy( MORSE_Complex64_t bump, MORSE_enum uplo, int N,
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
MORSE_request_t request = MORSE_REQUEST_INITIALIZER;
- MORSE_desc_t descA;
+ MORSE_desc_t descAl, descAt;
morse = morse_context_self();
if (morse == NULL) {
@@ -113,19 +113,26 @@ int MORSE_zplgsy( MORSE_Complex64_t bump, MORSE_enum uplo, int N,
/* Set NT */
NB = MORSE_NB;
- morse_sequence_create(morse, &sequence);
+ morse_sequence_create( morse, &sequence );
- morse_zdesc_alloc(descA, NB, NB, LDA, N, 0, 0, N, N, morse_desc_mat_free(&descA));
+ /* Submit the matrix conversion */
+ morse_zlap2tile( morse, &descAl, &descAt, MorseDescOutput, uplo,
+ A, NB, NB, LDA, N, N, N, sequence, &request );
/* Call the tile interface */
- MORSE_zplgsy_Tile_Async( bump, uplo, &descA, seed, sequence, &request );
+ MORSE_zplgsy_Tile_Async( bump, uplo, &descAt, seed, sequence, &request );
- morse_zooptile2lap(descA, A, NB, NB, LDA, N, sequence, &request);
- morse_sequence_wait(morse, sequence);
- morse_desc_mat_free(&descA);
+ /* Submit the matrix conversion back */
+ morse_ztile2lap( morse, &descAl, &descAt,
+ MorseDescOutput, uplo, sequence, &request );
+
+ morse_sequence_wait( morse, sequence );
+
+ /* Cleanup the temporary data */
+ morse_ztile2lap_cleanup( morse, &descAl, &descAt );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -172,7 +179,7 @@ int MORSE_zplgsy( MORSE_Complex64_t bump, MORSE_enum uplo, int N,
*
******************************************************************************/
int MORSE_zplgsy_Tile( MORSE_Complex64_t bump, MORSE_enum uplo,
- MORSE_desc_t *A,
+ MORSE_desc_t *A,
unsigned long long int seed )
{
MORSE_context_t *morse;
@@ -185,11 +192,15 @@ int MORSE_zplgsy_Tile( MORSE_Complex64_t bump, MORSE_enum uplo,
morse_fatal_error("MORSE_zplgsy_Tile", "MORSE not initialized");
return MORSE_ERR_NOT_INITIALIZED;
}
- morse_sequence_create(morse, &sequence);
+ morse_sequence_create( morse, &sequence );
+
MORSE_zplgsy_Tile_Async( bump, uplo, A, seed, sequence, &request );
- morse_sequence_wait(morse, sequence);
+
+ MORSE_Desc_Flush( A, sequence );
+
+ morse_sequence_wait( morse, sequence );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -246,10 +257,12 @@ int MORSE_zplgsy_Tile_Async( MORSE_Complex64_t bump,
return MORSE_ERR_UNALLOCATED;
}
/* Check sequence status */
- if (sequence->status == MORSE_SUCCESS)
+ if (sequence->status == MORSE_SUCCESS) {
request->status = MORSE_SUCCESS;
- else
+ }
+ else {
return morse_request_fail(sequence, request, MORSE_ERR_SEQUENCE_FLUSHED);
+ }
/* Check descriptors for correctness */
if (morse_desc_check(A) != MORSE_SUCCESS) {
@@ -266,7 +279,7 @@ int MORSE_zplgsy_Tile_Async( MORSE_Complex64_t bump,
if (chameleon_min( A->m, A->n ) == 0)
return MORSE_SUCCESS;
- morse_pzplgsy(bump, uplo, A, seed, sequence, request);
+ morse_pzplgsy( bump, uplo, A, seed, sequence, request );
return MORSE_SUCCESS;
}
diff --git a/compute/zplrnt.c b/compute/zplrnt.c
index 45e141e11fb5191fe9b2d40b94c7f7a58904f278..17edaa9d72de578abfb3e913b2fe1157494fc028 100644
--- a/compute/zplrnt.c
+++ b/compute/zplrnt.c
@@ -78,7 +78,7 @@ int MORSE_zplrnt( int M, int N,
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
MORSE_request_t request = MORSE_REQUEST_INITIALIZER;
- MORSE_desc_t descA;
+ MORSE_desc_t descAl, descAt;
morse = morse_context_self();
if (morse == NULL) {
@@ -111,19 +111,26 @@ int MORSE_zplrnt( int M, int N,
/* Set NT */
NB = MORSE_NB;
- morse_sequence_create(morse, &sequence);
- morse_zdesc_alloc(descA, NB, NB, LDA, N, 0, 0, M, N, morse_desc_mat_free(&descA));
+ morse_sequence_create( morse, &sequence );
+
+ /* Submit the matrix conversion */
+ morse_zlap2tile( morse, &descAl, &descAt, MorseDescOutput, MorseUpperLower,
+ A, NB, NB, LDA, N, M, N, sequence, &request );
/* Call the tile interface */
- MORSE_zplrnt_Tile_Async( &descA, seed, sequence, &request );
+ MORSE_zplrnt_Tile_Async( &descAt, seed, sequence, &request );
- morse_zooptile2lap(descA, A, NB, NB, LDA, N, sequence, &request);
- morse_sequence_wait(morse, sequence);
- morse_desc_mat_free(&descA);
+ /* Submit the matrix conversion back */
+ morse_ztile2lap( morse, &descAl, &descAt,
+ MorseDescOutput, MorseUpperLower, sequence, &request );
- status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_wait( morse, sequence );
+
+ /* Cleanup the temporary data */
+ morse_ztile2lap_cleanup( morse, &descAl, &descAt );
+ status = sequence->status;
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -175,11 +182,15 @@ int MORSE_zplrnt_Tile( MORSE_desc_t *A,
morse_fatal_error("MORSE_zplrnt_Tile", "MORSE not initialized");
return MORSE_ERR_NOT_INITIALIZED;
}
- morse_sequence_create(morse, &sequence);
+ morse_sequence_create( morse, &sequence );
+
MORSE_zplrnt_Tile_Async( A, seed, sequence, &request );
- morse_sequence_wait(morse, sequence);
+
+ MORSE_Desc_Flush( A, sequence );
+
+ morse_sequence_wait( morse, sequence );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -216,7 +227,7 @@ int MORSE_zplrnt_Tile( MORSE_desc_t *A,
int MORSE_zplrnt_Tile_Async( MORSE_desc_t *A,
unsigned long long int seed,
MORSE_sequence_t *sequence,
- MORSE_request_t *request)
+ MORSE_request_t *request )
{
MORSE_context_t *morse;
@@ -234,10 +245,12 @@ int MORSE_zplrnt_Tile_Async( MORSE_desc_t *A,
return MORSE_ERR_UNALLOCATED;
}
/* Check sequence status */
- if (sequence->status == MORSE_SUCCESS)
+ if (sequence->status == MORSE_SUCCESS) {
request->status = MORSE_SUCCESS;
- else
+ }
+ else {
return morse_request_fail(sequence, request, MORSE_ERR_SEQUENCE_FLUSHED);
+ }
/* Check descriptors for correctness */
if (morse_desc_check(A) != MORSE_SUCCESS) {
@@ -254,7 +267,7 @@ int MORSE_zplrnt_Tile_Async( MORSE_desc_t *A,
if (chameleon_min( A->m, A->n ) == 0)
return MORSE_SUCCESS;
- morse_pzplrnt(A, seed, sequence, request);
+ morse_pzplrnt( A, seed, sequence, request );
return MORSE_SUCCESS;
}
diff --git a/compute/zposv.c b/compute/zposv.c
index 5fe1ab279a4feb38ac097f80ca68bfed70807a54..6f4615ce69fbe0b6ae145d5bbe7ef5ba3ec94941 100644
--- a/compute/zposv.c
+++ b/compute/zposv.c
@@ -95,16 +95,17 @@
* @sa MORSE_sposv
*
******************************************************************************/
-int MORSE_zposv(MORSE_enum uplo, int N, int NRHS,
+int MORSE_zposv( MORSE_enum uplo, int N, int NRHS,
MORSE_Complex64_t *A, int LDA,
- MORSE_Complex64_t *B, int LDB)
+ MORSE_Complex64_t *B, int LDB )
{
int NB;
int status;
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
MORSE_request_t request = MORSE_REQUEST_INITIALIZER;
- MORSE_desc_t descA, descB;
+ MORSE_desc_t descAl, descAt;
+ MORSE_desc_t descBl, descBt;
morse = morse_context_self();
if (morse == NULL) {
@@ -112,7 +113,7 @@ int MORSE_zposv(MORSE_enum uplo, int N, int NRHS,
return MORSE_ERR_NOT_INITIALIZED;
}
/* Check input arguments */
- if (uplo != MorseUpper && uplo != MorseLower) {
+ if ((uplo != MorseUpper) && (uplo != MorseLower)) {
morse_error("MORSE_zposv", "illegal value of uplo");
return -1;
}
@@ -145,39 +146,33 @@ int MORSE_zposv(MORSE_enum uplo, int N, int NRHS,
}
/* Set NT & NTRHS */
- NB = MORSE_NB;
+ NB = MORSE_NB;
- morse_sequence_create(morse, &sequence);
+ morse_sequence_create( morse, &sequence );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooplap2tile( descA, A, NB, NB, LDA, N, 0, 0, N, N, sequence, &request,
- morse_desc_mat_free(&(descA)) );
- morse_zooplap2tile( descB, B, NB, NB, LDB, NRHS, 0, 0, N, NRHS, sequence, &request,
- morse_desc_mat_free(&(descA)); morse_desc_mat_free(&(descB)));
-/* } else {*/
-/* morse_ziplap2tile( descA, A, NB, NB, LDA, N, 0, 0, N, N, */
-/* sequence, &request);*/
-/* morse_ziplap2tile( descB, B, NB, NB, LDB, NRHS, 0, 0, N, NRHS,*/
-/* sequence, &request);*/
-/* }*/
+ /* Submit the matrix conversion */
+ morse_zlap2tile( morse, &descAl, &descAt, MorseDescInout, uplo,
+ A, NB, NB, LDA, N, N, N, sequence, &request );
+ morse_zlap2tile( morse, &descBl, &descBt, MorseDescInout, MorseUpperLower,
+ B, NB, NB, LDB, NRHS, N, NRHS, sequence, &request );
/* Call the tile interface */
- MORSE_zposv_Tile_Async(uplo, &descA, &descB, sequence, &request);
+ MORSE_zposv_Tile_Async( uplo, &descAt, &descBt, sequence, &request );
+
+ /* Submit the matrix conversion back */
+ morse_ztile2lap( morse, &descAl, &descAt,
+ MorseDescInout, uplo, sequence, &request );
+ morse_ztile2lap( morse, &descBl, &descBt,
+ MorseDescInout, MorseUpperLower, sequence, &request );
+
+ morse_sequence_wait( morse, sequence );
+
+ /* Cleanup the temporary data */
+ morse_ztile2lap_cleanup( morse, &descAl, &descAt );
+ morse_ztile2lap_cleanup( morse, &descBl, &descBt );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooptile2lap(descA, A, NB, NB, LDA, N, sequence, &request);
- morse_zooptile2lap(descB, B, NB, NB, LDB, NRHS, sequence, &request);
- morse_sequence_wait(morse, sequence);
- morse_desc_mat_free(&descA);
- morse_desc_mat_free(&descB);
-/* } else {*/
-/* morse_ziptile2lap( descA, A, NB, NB, LDA, N, sequence, &request);*/
-/* morse_ziptile2lap( descB, B, NB, NB, LDB, NRHS, sequence, &request);*/
-/* morse_sequence_wait(morse, sequence);*/
-/* }*/
-
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
/**
@@ -228,9 +223,9 @@ int MORSE_zposv(MORSE_enum uplo, int N, int NRHS,
* @sa MORSE_cposv_Tile
* @sa MORSE_dposv_Tile
* @sa MORSE_sposv_Tile
-*
+ *
******************************************************************************/
-int MORSE_zposv_Tile(MORSE_enum uplo, MORSE_desc_t *A, MORSE_desc_t *B)
+int MORSE_zposv_Tile( MORSE_enum uplo, MORSE_desc_t *A, MORSE_desc_t *B )
{
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
@@ -242,14 +237,16 @@ int MORSE_zposv_Tile(MORSE_enum uplo, MORSE_desc_t *A, MORSE_desc_t *B)
morse_fatal_error("MORSE_zposv_Tile", "MORSE not initialized");
return MORSE_ERR_NOT_INITIALIZED;
}
- morse_sequence_create(morse, &sequence);
- MORSE_zposv_Tile_Async(uplo, A, B, sequence, &request);
- RUNTIME_desc_flush( A, sequence );
- RUNTIME_desc_flush( B, sequence );
- morse_sequence_wait(morse, sequence);
-
+ morse_sequence_create( morse, &sequence );
+
+ MORSE_zposv_Tile_Async( uplo, A, B, sequence, &request );
+
+ MORSE_Desc_Flush( A, sequence );
+ MORSE_Desc_Flush( B, sequence );
+
+ morse_sequence_wait( morse, sequence );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -282,8 +279,8 @@ int MORSE_zposv_Tile(MORSE_enum uplo, MORSE_desc_t *A, MORSE_desc_t *B)
* @sa MORSE_sposv_Tile_Async
*
******************************************************************************/
-int MORSE_zposv_Tile_Async(MORSE_enum uplo, MORSE_desc_t *A, MORSE_desc_t *B,
- MORSE_sequence_t *sequence, MORSE_request_t *request)
+int MORSE_zposv_Tile_Async( MORSE_enum uplo, MORSE_desc_t *A, MORSE_desc_t *B,
+ MORSE_sequence_t *sequence, MORSE_request_t *request )
{
MORSE_context_t *morse;
@@ -301,10 +298,12 @@ int MORSE_zposv_Tile_Async(MORSE_enum uplo, MORSE_desc_t *A, MORSE_desc_t *B,
return MORSE_ERR_UNALLOCATED;
}
/* Check sequence status */
- if (sequence->status == MORSE_SUCCESS)
+ if (sequence->status == MORSE_SUCCESS) {
request->status = MORSE_SUCCESS;
- else
+ }
+ else {
return morse_request_fail(sequence, request, MORSE_ERR_SEQUENCE_FLUSHED);
+ }
/* Check descriptors for correctness */
if (morse_desc_check(A) != MORSE_SUCCESS) {
@@ -320,21 +319,21 @@ int MORSE_zposv_Tile_Async(MORSE_enum uplo, MORSE_desc_t *A, MORSE_desc_t *B,
morse_error("MORSE_zposv_Tile", "only square tiles supported");
return morse_request_fail(sequence, request, MORSE_ERR_ILLEGAL_VALUE);
}
- if (uplo != MorseUpper && uplo != MorseLower) {
+ if ((uplo != MorseUpper) && (uplo != MorseLower)) {
morse_error("MORSE_zposv_Tile", "illegal value of uplo");
return morse_request_fail(sequence, request, -1);
}
/* Quick return - currently NOT equivalent to LAPACK's
* LAPACK does not have such check for DPOSV */
-/*
- if (chameleon_min(N, NRHS) == 0)
- return MORSE_SUCCESS;
-*/
- morse_pzpotrf(uplo, A, sequence, request);
+ /*
+ if (chameleon_min(N, NRHS) == 0)
+ return MORSE_SUCCESS;
+ */
+ morse_pzpotrf( uplo, A, sequence, request );
- morse_pztrsm(MorseLeft, uplo, uplo == MorseUpper ? MorseConjTrans : MorseNoTrans, MorseNonUnit, 1.0, A, B, sequence, request);
+ morse_pztrsm( MorseLeft, uplo, uplo == MorseUpper ? MorseConjTrans : MorseNoTrans, MorseNonUnit, 1.0, A, B, sequence, request );
- morse_pztrsm(MorseLeft, uplo, uplo == MorseUpper ? MorseNoTrans : MorseConjTrans, MorseNonUnit, 1.0, A, B, sequence, request);
+ morse_pztrsm( MorseLeft, uplo, uplo == MorseUpper ? MorseNoTrans : MorseConjTrans, MorseNonUnit, 1.0, A, B, sequence, request );
return MORSE_SUCCESS;
}
diff --git a/compute/zpotrf.c b/compute/zpotrf.c
index 08261848912dfe3913b30bc45f3e0ffd3adb1d9d..f37ee389b55fdc2ab7d11c0c02560f4b91b9eda4 100644
--- a/compute/zpotrf.c
+++ b/compute/zpotrf.c
@@ -83,15 +83,15 @@
* @sa MORSE_zpotrs
*
******************************************************************************/
-int MORSE_zpotrf(MORSE_enum uplo, int N,
- MORSE_Complex64_t *A, int LDA)
+int MORSE_zpotrf( MORSE_enum uplo, int N,
+ MORSE_Complex64_t *A, int LDA )
{
int NB;
int status;
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
MORSE_request_t request = MORSE_REQUEST_INITIALIZER;
- MORSE_desc_t descA;
+ MORSE_desc_t descAl, descAt;
morse = morse_context_self();
if (morse == NULL) {
@@ -99,7 +99,7 @@ int MORSE_zpotrf(MORSE_enum uplo, int N,
return MORSE_ERR_NOT_INITIALIZED;
}
/* Check input arguments */
- if (uplo != MorseUpper && uplo != MorseLower) {
+ if ((uplo != MorseUpper) && (uplo != MorseLower)) {
morse_error("MORSE_zpotrf", "illegal value of uplo");
return -1;
}
@@ -123,32 +123,28 @@ int MORSE_zpotrf(MORSE_enum uplo, int N,
}
/* Set NT */
- NB = MORSE_NB;
+ NB = MORSE_NB;
- morse_sequence_create(morse, &sequence);
+ morse_sequence_create( morse, &sequence );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooplap2tile( descA, A, NB, NB, LDA, N, 0, 0, N, N, sequence, &request,
- morse_desc_mat_free(&(descA)) );
-/* } else {*/
-/* morse_ziplap2tile( descA, A, NB, NB, LDA, N, 0, 0, N, N,*/
-/* sequence, &request);*/
-/* }*/
+ /* Submit the matrix conversion */
+ morse_zlap2tile( morse, &descAl, &descAt, MorseDescInout, uplo,
+ A, NB, NB, LDA, N, N, N, sequence, &request );
/* Call the tile interface */
- MORSE_zpotrf_Tile_Async(uplo, &descA, sequence, &request);
+ MORSE_zpotrf_Tile_Async( uplo, &descAt, sequence, &request );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooptile2lap(descA, A, NB, NB, LDA, N, sequence, &request);
- morse_sequence_wait(morse, sequence);
- morse_desc_mat_free(&descA);
-/* } else {*/
-/* morse_ziptile2lap( descA, A, NB, NB, LDA, N, sequence, &request);*/
-/* morse_sequence_wait(morse, sequence);*/
-/* }*/
+ /* Submit the matrix conversion back */
+ morse_ztile2lap( morse, &descAl, &descAt,
+ MorseDescInout, uplo, sequence, &request );
+
+ morse_sequence_wait( morse, sequence );
+
+ /* Cleanup the temporary data */
+ morse_ztile2lap_cleanup( morse, &descAl, &descAt );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -199,7 +195,7 @@ int MORSE_zpotrf(MORSE_enum uplo, int N,
* @sa MORSE_zpotrs_Tile
*
******************************************************************************/
-int MORSE_zpotrf_Tile(MORSE_enum uplo, MORSE_desc_t *A)
+int MORSE_zpotrf_Tile( MORSE_enum uplo, MORSE_desc_t *A )
{
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
@@ -211,13 +207,15 @@ int MORSE_zpotrf_Tile(MORSE_enum uplo, MORSE_desc_t *A)
morse_fatal_error("MORSE_zpotrf_Tile", "MORSE not initialized");
return MORSE_ERR_NOT_INITIALIZED;
}
- morse_sequence_create(morse, &sequence);
- MORSE_zpotrf_Tile_Async(uplo, A, sequence, &request);
- RUNTIME_desc_flush( A, sequence );
- morse_sequence_wait(morse, sequence);
+ morse_sequence_create( morse, &sequence );
+
+ MORSE_zpotrf_Tile_Async( uplo, A, sequence, &request );
+
+ MORSE_Desc_Flush( A, sequence );
+ morse_sequence_wait( morse, sequence );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -251,8 +249,8 @@ int MORSE_zpotrf_Tile(MORSE_enum uplo, MORSE_desc_t *A)
* @sa MORSE_zpotrs_Tile_Async
*
******************************************************************************/
-int MORSE_zpotrf_Tile_Async(MORSE_enum uplo, MORSE_desc_t *A,
- MORSE_sequence_t *sequence, MORSE_request_t *request)
+int MORSE_zpotrf_Tile_Async( MORSE_enum uplo, MORSE_desc_t *A,
+ MORSE_sequence_t *sequence, MORSE_request_t *request )
{
MORSE_context_t *morse;
@@ -270,10 +268,12 @@ int MORSE_zpotrf_Tile_Async(MORSE_enum uplo, MORSE_desc_t *A,
return MORSE_ERR_UNALLOCATED;
}
/* Check sequence status */
- if (sequence->status == MORSE_SUCCESS)
+ if (sequence->status == MORSE_SUCCESS) {
request->status = MORSE_SUCCESS;
- else
+ }
+ else {
return morse_request_fail(sequence, request, MORSE_ERR_SEQUENCE_FLUSHED);
+ }
/* Check descriptors for correctness */
if (morse_desc_check(A) != MORSE_SUCCESS) {
@@ -285,16 +285,16 @@ int MORSE_zpotrf_Tile_Async(MORSE_enum uplo, MORSE_desc_t *A,
morse_error("MORSE_zpotrf_Tile_Async", "only square tiles supported");
return morse_request_fail(sequence, request, MORSE_ERR_ILLEGAL_VALUE);
}
- if (uplo != MorseUpper && uplo != MorseLower) {
+ if ((uplo != MorseUpper) && (uplo != MorseLower)) {
morse_error("MORSE_zpotrf_Tile_Async", "illegal value of uplo");
return morse_request_fail(sequence, request, -1);
}
/* Quick return */
-/*
- if (chameleon_max(N, 0) == 0)
- return MORSE_SUCCESS;
-*/
- morse_pzpotrf(uplo, A, sequence, request);
+ /*
+ if (chameleon_max(N, 0) == 0)
+ return MORSE_SUCCESS;
+ */
+ morse_pzpotrf( uplo, A, sequence, request );
return MORSE_SUCCESS;
}
diff --git a/compute/zpotri.c b/compute/zpotri.c
index 801f077cc45b83a80792ea563a91551e4be1f9e2..6662a1c661608b63269d537b8def7ed4c54c4f77 100644
--- a/compute/zpotri.c
+++ b/compute/zpotri.c
@@ -74,15 +74,15 @@
* @sa MORSE_zpotrf
*
******************************************************************************/
-int MORSE_zpotri(MORSE_enum uplo, int N,
- MORSE_Complex64_t *A, int LDA)
+int MORSE_zpotri( MORSE_enum uplo, int N,
+ MORSE_Complex64_t *A, int LDA )
{
int NB;
int status;
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
MORSE_request_t request = MORSE_REQUEST_INITIALIZER;
- MORSE_desc_t descA;
+ MORSE_desc_t descAl, descAt;
morse = morse_context_self();
if (morse == NULL) {
@@ -90,7 +90,7 @@ int MORSE_zpotri(MORSE_enum uplo, int N,
return MORSE_ERR_NOT_INITIALIZED;
}
/* Check input arguments */
- if (uplo != MorseUpper && uplo != MorseLower) {
+ if ((uplo != MorseUpper) && (uplo != MorseLower)) {
morse_error("MORSE_zpotri", "illegal value of uplo");
return -1;
}
@@ -116,30 +116,26 @@ int MORSE_zpotri(MORSE_enum uplo, int N,
/* Set NT */
NB = MORSE_NB;
- morse_sequence_create(morse, &sequence);
+ morse_sequence_create( morse, &sequence );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooplap2tile( descA, A, NB, NB, LDA, N, 0, 0, N, N, sequence, &request,
- morse_desc_mat_free(&(descA)) );
-/* } else {*/
-/* morse_ziplap2tile( descA, A, NB, NB, LDA, N, 0, 0, N, N,*/
-/* sequence, &request);*/
-/* }*/
+ /* Submit the matrix conversion */
+ morse_zlap2tile( morse, &descAl, &descAt, MorseDescInout, uplo,
+ A, NB, NB, LDA, N, N, N, sequence, &request );
/* Call the tile interface */
- MORSE_zpotri_Tile_Async(uplo, &descA, sequence, &request);
+ MORSE_zpotri_Tile_Async( uplo, &descAt, sequence, &request );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooptile2lap(descA, A, NB, NB, LDA, N, sequence, &request);
- morse_sequence_wait(morse, sequence);
- morse_desc_mat_free(&descA);
-/* } else {*/
-/* morse_ziptile2lap( descA, A, NB, NB, LDA, N, sequence, &request);*/
-/* morse_sequence_wait(morse, sequence);*/
-/* }*/
+ /* Submit the matrix conversion back */
+ morse_ztile2lap( morse, &descAl, &descAt,
+ MorseDescInout, uplo, sequence, &request );
+
+ morse_sequence_wait( morse, sequence );
+
+ /* Cleanup the temporary data */
+ morse_ztile2lap_cleanup( morse, &descAl, &descAt );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -187,7 +183,7 @@ int MORSE_zpotri(MORSE_enum uplo, int N,
* @sa MORSE_zpotrf_Tile
*
******************************************************************************/
-int MORSE_zpotri_Tile(MORSE_enum uplo, MORSE_desc_t *A)
+int MORSE_zpotri_Tile( MORSE_enum uplo, MORSE_desc_t *A )
{
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
@@ -199,13 +195,15 @@ int MORSE_zpotri_Tile(MORSE_enum uplo, MORSE_desc_t *A)
morse_fatal_error("MORSE_zpotri_Tile", "MORSE not initialized");
return MORSE_ERR_NOT_INITIALIZED;
}
- morse_sequence_create(morse, &sequence);
- MORSE_zpotri_Tile_Async(uplo, A, sequence, &request);
- RUNTIME_desc_flush( A, sequence );
- morse_sequence_wait(morse, sequence);
-
+ morse_sequence_create( morse, &sequence );
+
+ MORSE_zpotri_Tile_Async( uplo, A, sequence, &request );
+
+ MORSE_Desc_Flush( A, sequence );
+
+ morse_sequence_wait( morse, sequence );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -240,8 +238,8 @@ int MORSE_zpotri_Tile(MORSE_enum uplo, MORSE_desc_t *A)
* @sa MORSE_zpotrf_Tile_Async
*
******************************************************************************/
-int MORSE_zpotri_Tile_Async(MORSE_enum uplo, MORSE_desc_t *A,
- MORSE_sequence_t *sequence, MORSE_request_t *request)
+int MORSE_zpotri_Tile_Async( MORSE_enum uplo, MORSE_desc_t *A,
+ MORSE_sequence_t *sequence, MORSE_request_t *request )
{
MORSE_context_t *morse;
@@ -259,10 +257,12 @@ int MORSE_zpotri_Tile_Async(MORSE_enum uplo, MORSE_desc_t *A,
return MORSE_ERR_UNALLOCATED;
}
/* Check sequence status */
- if (sequence->status == MORSE_SUCCESS)
+ if (sequence->status == MORSE_SUCCESS) {
request->status = MORSE_SUCCESS;
- else
+ }
+ else {
return morse_request_fail(sequence, request, MORSE_ERR_SEQUENCE_FLUSHED);
+ }
/* Check descriptors for correctness */
if (morse_desc_check(A) != MORSE_SUCCESS) {
@@ -274,18 +274,18 @@ int MORSE_zpotri_Tile_Async(MORSE_enum uplo, MORSE_desc_t *A,
morse_error("MORSE_zpotri_Tile_Async", "only square tiles supported");
return morse_request_fail(sequence, request, MORSE_ERR_ILLEGAL_VALUE);
}
- if (uplo != MorseUpper && uplo != MorseLower) {
+ if ((uplo != MorseUpper) && (uplo != MorseLower)) {
morse_error("MORSE_zpotri_Tile_Async", "illegal value of uplo");
return morse_request_fail(sequence, request, -1);
}
/* Quick return */
-/*
- if (chameleon_max(N, 0) == 0)
- return MORSE_SUCCESS;
-*/
- morse_pztrtri(uplo, MorseNonUnit, A, sequence, request);
+ /*
+ if (chameleon_max(N, 0) == 0)
+ return MORSE_SUCCESS;
+ */
+ morse_pztrtri( uplo, MorseNonUnit, A, sequence, request );
- morse_pzlauum(uplo, A, sequence, request);
+ morse_pzlauum( uplo, A, sequence, request );
return MORSE_SUCCESS;
}
diff --git a/compute/zpotrimm.c b/compute/zpotrimm.c
index 180725746aa845cb28b197c915d64304e6e226e1..ec2f8759a6a6dd690545751c165cbfd71044c36b 100644
--- a/compute/zpotrimm.c
+++ b/compute/zpotrimm.c
@@ -74,17 +74,19 @@
* @sa MORSE_zpotrf
*
******************************************************************************/
-int MORSE_zpotrimm(MORSE_enum uplo, int N,
- MORSE_Complex64_t *A, int LDA,
- MORSE_Complex64_t *B, int LDB,
- MORSE_Complex64_t *C, int LDC)
+int MORSE_zpotrimm( MORSE_enum uplo, int N,
+ MORSE_Complex64_t *A, int LDA,
+ MORSE_Complex64_t *B, int LDB,
+ MORSE_Complex64_t *C, int LDC )
{
int NB;
int status;
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
MORSE_request_t request = MORSE_REQUEST_INITIALIZER;
- MORSE_desc_t descA, descB, descC;
+ MORSE_desc_t descAl, descAt;
+ MORSE_desc_t descBl, descBt;
+ MORSE_desc_t descCl, descCt;
morse = morse_context_self();
if (morse == NULL) {
@@ -92,7 +94,7 @@ int MORSE_zpotrimm(MORSE_enum uplo, int N,
return MORSE_ERR_NOT_INITIALIZED;
}
/* Check input arguments */
- if (uplo != MorseUpper && uplo != MorseLower) {
+ if ((uplo != MorseUpper) && (uplo != MorseLower)) {
morse_error("MORSE_zpotrimm", "illegal value of uplo");
return -1;
}
@@ -126,38 +128,36 @@ int MORSE_zpotrimm(MORSE_enum uplo, int N,
/* Set NT */
NB = MORSE_NB;
- morse_sequence_create(morse, &sequence);
+ morse_sequence_create( morse, &sequence );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooplap2tile( descA, A, NB, NB, LDA, N, 0, 0, N, N, sequence, &request,
- morse_desc_mat_free(&(descA)) );
- morse_zooplap2tile( descB, B, NB, NB, LDB, N, 0, 0, N, N, sequence, &request,
- morse_desc_mat_free(&(descA)); morse_desc_mat_free(&(descB)));
- morse_zooplap2tile( descC, C, NB, NB, LDC, N, 0, 0, N, N, sequence, &request,
- morse_desc_mat_free(&(descA)); morse_desc_mat_free(&(descB)); morse_desc_mat_free(&(descC)));
-/* } else {*/
-/* morse_ziplap2tile( descA, A, NB, NB, LDA, N, 0, 0, N, N,*/
-/* sequence, &request);*/
-/* }*/
+ /* Submit the matrix conversion */
+ morse_zlap2tile( morse, &descAl, &descAt, MorseDescInout, uplo,
+ A, NB, NB, LDA, N, N, N, sequence, &request );
+ morse_zlap2tile( morse, &descBl, &descBt, MorseDescInput, MorseUpperLower,
+ B, NB, NB, LDB, N, N, N, sequence, &request );
+ morse_zlap2tile( morse, &descCl, &descCt, MorseDescInout, MorseUpperLower,
+ C, NB, NB, LDC, N, N, N, sequence, &request );
/* Call the tile interface */
- MORSE_zpotrimm_Tile_Async(uplo, &descA, &descB, &descC, sequence, &request);
+ MORSE_zpotrimm_Tile_Async( uplo, &descAt, &descBt, &descCt, sequence, &request );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooptile2lap(descA, A, NB, NB, LDA, N, sequence, &request);
- morse_zooptile2lap(descB, A, NB, NB, LDB, N, sequence, &request);
- morse_zooptile2lap(descC, C, NB, NB, LDC, N, sequence, &request);
- morse_sequence_wait(morse, sequence);
- morse_desc_mat_free(&descA);
- morse_desc_mat_free(&descB);
- morse_desc_mat_free(&descC);
-/* } else {*/
-/* morse_ziptile2lap( descA, A, NB, NB, LDA, N, sequence, &request);*/
-/* morse_sequence_wait(morse, sequence);*/
-/* }*/
+ /* Submit the matrix conversion back */
+ morse_ztile2lap( morse, &descAl, &descAt,
+ MorseDescInout, uplo, sequence, &request );
+ morse_ztile2lap( morse, &descBl, &descBt,
+ MorseDescInput, MorseUpperLower, sequence, &request );
+ morse_ztile2lap( morse, &descCl, &descCt,
+ MorseDescInout, MorseUpperLower, sequence, &request );
+
+ morse_sequence_wait( morse, sequence );
+
+ /* Cleanup the temporary data */
+ morse_ztile2lap_cleanup( morse, &descAl, &descAt );
+ morse_ztile2lap_cleanup( morse, &descBl, &descBt );
+ morse_ztile2lap_cleanup( morse, &descCl, &descCt );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -205,7 +205,7 @@ int MORSE_zpotrimm(MORSE_enum uplo, int N,
* @sa MORSE_zpotrf_Tile
*
******************************************************************************/
-int MORSE_zpotrimm_Tile(MORSE_enum uplo, MORSE_desc_t *A, MORSE_desc_t *B, MORSE_desc_t *C)
+int MORSE_zpotrimm_Tile( MORSE_enum uplo, MORSE_desc_t *A, MORSE_desc_t *B, MORSE_desc_t *C )
{
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
@@ -217,13 +217,17 @@ int MORSE_zpotrimm_Tile(MORSE_enum uplo, MORSE_desc_t *A, MORSE_desc_t *B, MORSE
morse_fatal_error("MORSE_zpotrimm_Tile", "MORSE not initialized");
return MORSE_ERR_NOT_INITIALIZED;
}
- morse_sequence_create(morse, &sequence);
- MORSE_zpotrimm_Tile_Async(uplo, A, B, C, sequence, &request);
- RUNTIME_desc_flush( C, sequence );
- morse_sequence_wait(morse, sequence);
+ morse_sequence_create( morse, &sequence );
+
+ MORSE_zpotrimm_Tile_Async( uplo, A, B, C, sequence, &request );
+
+ MORSE_Desc_Flush( A, sequence );
+ MORSE_Desc_Flush( B, sequence );
+ MORSE_Desc_Flush( C, sequence );
+ morse_sequence_wait( morse, sequence );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -258,8 +262,8 @@ int MORSE_zpotrimm_Tile(MORSE_enum uplo, MORSE_desc_t *A, MORSE_desc_t *B, MORSE
* @sa MORSE_zpotrf_Tile_Async
*
******************************************************************************/
-int MORSE_zpotrimm_Tile_Async(MORSE_enum uplo, MORSE_desc_t *A, MORSE_desc_t *B, MORSE_desc_t *C,
- MORSE_sequence_t *sequence, MORSE_request_t *request)
+int MORSE_zpotrimm_Tile_Async( MORSE_enum uplo, MORSE_desc_t *A, MORSE_desc_t *B, MORSE_desc_t *C,
+ MORSE_sequence_t *sequence, MORSE_request_t *request )
{
MORSE_context_t *morse;
@@ -277,10 +281,12 @@ int MORSE_zpotrimm_Tile_Async(MORSE_enum uplo, MORSE_desc_t *A, MORSE_desc_t *B,
return MORSE_ERR_UNALLOCATED;
}
/* Check sequence status */
- if (sequence->status == MORSE_SUCCESS)
+ if (sequence->status == MORSE_SUCCESS) {
request->status = MORSE_SUCCESS;
- else
+ }
+ else {
return morse_request_fail(sequence, request, MORSE_ERR_SEQUENCE_FLUSHED);
+ }
/* Check descriptors for correctness */
if (morse_desc_check(A) != MORSE_SUCCESS) {
@@ -308,7 +314,7 @@ int MORSE_zpotrimm_Tile_Async(MORSE_enum uplo, MORSE_desc_t *A, MORSE_desc_t *B,
morse_error("MORSE_zpotrimm_Tile_Async", "only square tiles supported");
return morse_request_fail(sequence, request, MORSE_ERR_ILLEGAL_VALUE);
}
- if (uplo != MorseUpper && uplo != MorseLower) {
+ if ((uplo != MorseUpper) && (uplo != MorseLower)) {
morse_error("MORSE_zpotrimm_Tile_Async", "illegal value of uplo");
return morse_request_fail(sequence, request, -1);
}
@@ -317,11 +323,11 @@ int MORSE_zpotrimm_Tile_Async(MORSE_enum uplo, MORSE_desc_t *A, MORSE_desc_t *B,
if (chameleon_max(N, 0) == 0)
return MORSE_SUCCESS;
*/
- morse_pzpotrimm(uplo, A, B, C, sequence, request);
+ morse_pzpotrimm( uplo, A, B, C, sequence, request );
/*
- morse_pztrtri(uplo, MorseNonUnit, A, sequence, request);
- morse_pzlauum(uplo, A, sequence, request);
- */
+ morse_pztrtri( uplo, MorseNonUnit, A, sequence, request );
+ morse_pzlauum( uplo, A, sequence, request );
+ */
return MORSE_SUCCESS;
}
diff --git a/compute/zpotrs.c b/compute/zpotrs.c
index 0494b688426132276bb426e911b8645bbd439e35..7ff43c5698b9f6eda87fe6342edf3abdca8ed95d 100644
--- a/compute/zpotrs.c
+++ b/compute/zpotrs.c
@@ -80,16 +80,17 @@
* @sa MORSE_zpotrf
*
******************************************************************************/
-int MORSE_zpotrs(MORSE_enum uplo, int N, int NRHS,
+int MORSE_zpotrs( MORSE_enum uplo, int N, int NRHS,
MORSE_Complex64_t *A, int LDA,
- MORSE_Complex64_t *B, int LDB)
+ MORSE_Complex64_t *B, int LDB )
{
int NB;
int status;
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
MORSE_request_t request = MORSE_REQUEST_INITIALIZER;
- MORSE_desc_t descA, descB;
+ MORSE_desc_t descAl, descAt;
+ MORSE_desc_t descBl, descBt;
morse = morse_context_self();
if (morse == NULL) {
@@ -97,7 +98,7 @@ int MORSE_zpotrs(MORSE_enum uplo, int N, int NRHS,
return MORSE_ERR_NOT_INITIALIZED;
}
/* Check input arguments */
- if (uplo != MorseUpper && uplo != MorseLower) {
+ if ((uplo != MorseUpper) && (uplo != MorseLower)) {
morse_error("MORSE_zpotrs", "illegal value of uplo");
return -1;
}
@@ -131,36 +132,31 @@ int MORSE_zpotrs(MORSE_enum uplo, int N, int NRHS,
/* Set NT & NTRHS */
NB = MORSE_NB;
- morse_sequence_create(morse, &sequence);
+ morse_sequence_create( morse, &sequence );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooplap2tile( descA, A, NB, NB, LDA, N, 0, 0, N, N, sequence, &request,
- morse_desc_mat_free(&(descA)) );
- morse_zooplap2tile( descB, B, NB, NB, LDB, NRHS, 0, 0, N, NRHS, sequence, &request,
- morse_desc_mat_free(&(descA)); morse_desc_mat_free(&(descB)));
-/* } else {*/
-/* morse_ziplap2tile( descA, A, NB, NB, LDA, N, 0, 0, N, N, */
-/* sequence, &request);*/
-/* morse_ziplap2tile( descB, B, NB, NB, LDB, NRHS, 0, 0, N, NRHS,*/
-/* sequence, &request);*/
-/* }*/
+ /* Submit the matrix conversion */
+ morse_zlap2tile( morse, &descAl, &descAt, MorseDescInput, uplo,
+ A, NB, NB, LDA, N, N, N, sequence, &request );
+ morse_zlap2tile( morse, &descBl, &descBt, MorseDescInout, MorseUpperLower,
+ B, NB, NB, LDB, NRHS, N, NRHS, sequence, &request );
/* Call the tile interface */
- MORSE_zpotrs_Tile_Async(uplo, &descA, &descB, sequence, &request);
+ MORSE_zpotrs_Tile_Async( uplo, &descAt, &descBt, sequence, &request );
+
+ /* Submit the matrix conversion back */
+ morse_ztile2lap( morse, &descAl, &descAt,
+ MorseDescInput, uplo, sequence, &request );
+ morse_ztile2lap( morse, &descBl, &descBt,
+ MorseDescInout, MorseUpperLower, sequence, &request );
+
+ morse_sequence_wait( morse, sequence );
+
+ /* Cleanup the temporary data */
+ morse_ztile2lap_cleanup( morse, &descAl, &descAt );
+ morse_ztile2lap_cleanup( morse, &descBl, &descBt );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooptile2lap(descB, B, NB, NB, LDB, NRHS, sequence, &request);
- morse_sequence_wait(morse, sequence);
- morse_desc_mat_free(&descA);
- morse_desc_mat_free(&descB);
-/* } else {*/
-/* morse_ziptile2lap( descA, A, NB, NB, LDA, N, sequence, &request);*/
-/* morse_ziptile2lap( descB, B, NB, NB, LDB, NRHS, sequence, &request);*/
-/* morse_sequence_wait(morse, sequence);*/
-/* }*/
-
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -205,7 +201,7 @@ int MORSE_zpotrs(MORSE_enum uplo, int N, int NRHS,
* @sa MORSE_zpotrf_Tile
*
******************************************************************************/
-int MORSE_zpotrs_Tile(MORSE_enum uplo, MORSE_desc_t *A, MORSE_desc_t *B)
+int MORSE_zpotrs_Tile( MORSE_enum uplo, MORSE_desc_t *A, MORSE_desc_t *B )
{
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
@@ -217,14 +213,16 @@ int MORSE_zpotrs_Tile(MORSE_enum uplo, MORSE_desc_t *A, MORSE_desc_t *B)
morse_fatal_error("MORSE_zpotrs_Tile", "MORSE not initialized");
return MORSE_ERR_NOT_INITIALIZED;
}
- morse_sequence_create(morse, &sequence);
- MORSE_zpotrs_Tile_Async(uplo, A, B, sequence, &request);
- RUNTIME_desc_flush( A, sequence );
- RUNTIME_desc_flush( B, sequence );
- morse_sequence_wait(morse, sequence);
+ morse_sequence_create( morse, &sequence );
+
+ MORSE_zpotrs_Tile_Async( uplo, A, B, sequence, &request );
+
+ MORSE_Desc_Flush( A, sequence );
+ MORSE_Desc_Flush( B, sequence );
+ morse_sequence_wait( morse, sequence );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -258,8 +256,8 @@ int MORSE_zpotrs_Tile(MORSE_enum uplo, MORSE_desc_t *A, MORSE_desc_t *B)
* @sa MORSE_zpotrf_Tile_Async
*
******************************************************************************/
-int MORSE_zpotrs_Tile_Async(MORSE_enum uplo, MORSE_desc_t *A, MORSE_desc_t *B,
- MORSE_sequence_t *sequence, MORSE_request_t *request)
+int MORSE_zpotrs_Tile_Async( MORSE_enum uplo, MORSE_desc_t *A, MORSE_desc_t *B,
+ MORSE_sequence_t *sequence, MORSE_request_t *request )
{
MORSE_context_t *morse;
@@ -277,10 +275,12 @@ int MORSE_zpotrs_Tile_Async(MORSE_enum uplo, MORSE_desc_t *A, MORSE_desc_t *B,
return MORSE_ERR_UNALLOCATED;
}
/* Check sequence status */
- if (sequence->status == MORSE_SUCCESS)
+ if (sequence->status == MORSE_SUCCESS) {
request->status = MORSE_SUCCESS;
- else
+ }
+ else {
return morse_request_fail(sequence, request, MORSE_ERR_SEQUENCE_FLUSHED);
+ }
/* Check descriptors for correctness */
if (morse_desc_check(A) != MORSE_SUCCESS) {
@@ -296,18 +296,18 @@ int MORSE_zpotrs_Tile_Async(MORSE_enum uplo, MORSE_desc_t *A, MORSE_desc_t *B,
morse_error("MORSE_zpotrs_Tile", "only square tiles supported");
return morse_request_fail(sequence, request, MORSE_ERR_ILLEGAL_VALUE);
}
- if (uplo != MorseUpper && uplo != MorseLower) {
+ if ((uplo != MorseUpper) && (uplo != MorseLower)) {
morse_error("MORSE_zpotrs_Tile", "illegal value of uplo");
return morse_request_fail(sequence, request, -1);
}
/* Quick return */
-/*
- if (chameleon_min(N, NRHS) == 0)
- return MORSE_SUCCESS;
-*/
- morse_pztrsm(MorseLeft, uplo, uplo == MorseUpper ? MorseConjTrans : MorseNoTrans, MorseNonUnit, 1.0, A, B, sequence, request);
+ /*
+ if (chameleon_min(N, NRHS) == 0)
+ return MORSE_SUCCESS;
+ */
+ morse_pztrsm( MorseLeft, uplo, uplo == MorseUpper ? MorseConjTrans : MorseNoTrans, MorseNonUnit, 1.0, A, B, sequence, request );
- morse_pztrsm(MorseLeft, uplo, uplo == MorseUpper ? MorseNoTrans : MorseConjTrans, MorseNonUnit, 1.0, A, B, sequence, request);
+ morse_pztrsm( MorseLeft, uplo, uplo == MorseUpper ? MorseNoTrans : MorseConjTrans, MorseNonUnit, 1.0, A, B, sequence, request );
return MORSE_SUCCESS;
}
diff --git a/compute/zsymm.c b/compute/zsymm.c
index 344f04da6f07ecd8b3be760a95e8d3621aaee504..a6662f3e9adeb3b6c2468e4ee859f325c94395af 100644
--- a/compute/zsymm.c
+++ b/compute/zsymm.c
@@ -106,15 +106,17 @@
* @sa MORSE_ssymm
*
******************************************************************************/
-int MORSE_zsymm(MORSE_enum side, MORSE_enum uplo, int M, int N,
+int MORSE_zsymm( MORSE_enum side, MORSE_enum uplo, int M, int N,
MORSE_Complex64_t alpha, MORSE_Complex64_t *A, int LDA,
- MORSE_Complex64_t *B, int LDB,
- MORSE_Complex64_t beta, MORSE_Complex64_t *C, int LDC)
+ MORSE_Complex64_t *B, int LDB,
+ MORSE_Complex64_t beta, MORSE_Complex64_t *C, int LDC )
{
int NB;
int Am;
int status;
- MORSE_desc_t descA, descB, descC;
+ MORSE_desc_t descAl, descAt;
+ MORSE_desc_t descBl, descBt;
+ MORSE_desc_t descCl, descCt;
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
MORSE_request_t request = MORSE_REQUEST_INITIALIZER;
@@ -171,43 +173,36 @@ int MORSE_zsymm(MORSE_enum side, MORSE_enum uplo, int M, int N,
/* Set MT & NT & KT */
NB = MORSE_NB;
- morse_sequence_create(morse, &sequence);
+ morse_sequence_create( morse, &sequence );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooplap2tile( descA, A, NB, NB, LDA, Am, 0, 0, Am, Am, sequence, &request,
- morse_desc_mat_free(&(descA)) );
- morse_zooplap2tile( descB, B, NB, NB, LDB, N, 0, 0, M, N, sequence, &request,
- morse_desc_mat_free(&(descA)); morse_desc_mat_free(&(descB)));
- morse_zooplap2tile( descC, C, NB, NB, LDC, N, 0, 0, M, N, sequence, &request,
- morse_desc_mat_free(&(descA)); morse_desc_mat_free(&(descB)); morse_desc_mat_free(&(descC)));
-/* } else {*/
-/* morse_ziplap2tile( descA, A, NB, NB, LDA, Am, 0, 0, Am, Am,*/
-/* sequence, &request);*/
-/* morse_ziplap2tile( descB, B, NB, NB, LDB, N, 0, 0, M, N,*/
-/* sequence, &request);*/
-/* morse_ziplap2tile( descC, C, NB, NB, LDC, N, 0, 0, M, N,*/
-/* sequence, &request);*/
-/* }*/
+ /* Submit the matrix conversion */
+ morse_zlap2tile( morse, &descAl, &descAt, MorseDescInput, uplo,
+ A, NB, NB, LDA, Am, Am, Am, sequence, &request );
+ morse_zlap2tile( morse, &descBl, &descBt, MorseDescInput, MorseUpperLower,
+ B, NB, NB, LDB, N, M, N, sequence, &request );
+ morse_zlap2tile( morse, &descCl, &descCt, MorseDescInout, MorseUpperLower,
+ C, NB, NB, LDC, N, M, N, sequence, &request );
/* Call the tile interface */
- MORSE_zsymm_Tile_Async(
- side, uplo, alpha, &descA, &descB, beta, &descC, sequence, &request);
+ MORSE_zsymm_Tile_Async( side, uplo, alpha, &descAt, &descBt, beta, &descCt, sequence, &request );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooptile2lap(descC, C, NB, NB, LDC, N, sequence, &request);
- morse_sequence_wait(morse, sequence);
- morse_desc_mat_free(&descA);
- morse_desc_mat_free(&descB);
- morse_desc_mat_free(&descC);
-/* } else {*/
-/* morse_ziptile2lap( descA, A, NB, NB, LDA, Am, sequence, &request);*/
-/* morse_ziptile2lap( descB, B, NB, NB, LDB, N, sequence, &request);*/
-/* morse_ziptile2lap( descC, C, NB, NB, LDC, N, sequence, &request);*/
-/* morse_sequence_wait(morse, sequence);*/
-/* }*/
+ /* Submit the matrix conversion back */
+ morse_ztile2lap( morse, &descAl, &descAt,
+ MorseDescInput, uplo, sequence, &request );
+ morse_ztile2lap( morse, &descBl, &descBt,
+ MorseDescInput, MorseUpperLower, sequence, &request );
+ morse_ztile2lap( morse, &descCl, &descCt,
+ MorseDescInout, MorseUpperLower, sequence, &request );
+
+ morse_sequence_wait( morse, sequence );
+
+ /* Cleanup the temporary data */
+ morse_ztile2lap_cleanup( morse, &descAl, &descAt );
+ morse_ztile2lap_cleanup( morse, &descBl, &descBt );
+ morse_ztile2lap_cleanup( morse, &descCl, &descCt );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -270,9 +265,9 @@ int MORSE_zsymm(MORSE_enum side, MORSE_enum uplo, int M, int N,
* @sa MORSE_ssymm_Tile
*
******************************************************************************/
-int MORSE_zsymm_Tile(MORSE_enum side, MORSE_enum uplo,
+int MORSE_zsymm_Tile( MORSE_enum side, MORSE_enum uplo,
MORSE_Complex64_t alpha, MORSE_desc_t *A, MORSE_desc_t *B,
- MORSE_Complex64_t beta, MORSE_desc_t *C)
+ MORSE_Complex64_t beta, MORSE_desc_t *C )
{
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
@@ -284,15 +279,17 @@ int MORSE_zsymm_Tile(MORSE_enum side, MORSE_enum uplo,
morse_fatal_error("MORSE_zsymm_Tile", "MORSE not initialized");
return MORSE_ERR_NOT_INITIALIZED;
}
- morse_sequence_create(morse, &sequence);
- MORSE_zsymm_Tile_Async(side, uplo, alpha, A, B, beta, C, sequence, &request);
- RUNTIME_desc_flush( A, sequence );
- RUNTIME_desc_flush( B, sequence );
- RUNTIME_desc_flush( C, sequence );
- morse_sequence_wait(morse, sequence);
+ morse_sequence_create( morse, &sequence );
+
+ MORSE_zsymm_Tile_Async(side, uplo, alpha, A, B, beta, C, sequence, &request );
+
+ MORSE_Desc_Flush( A, sequence );
+ MORSE_Desc_Flush( B, sequence );
+ MORSE_Desc_Flush( C, sequence );
+ morse_sequence_wait( morse, sequence );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -324,10 +321,10 @@ int MORSE_zsymm_Tile(MORSE_enum side, MORSE_enum uplo,
* @sa MORSE_ssymm_Tile_Async
*
******************************************************************************/
-int MORSE_zsymm_Tile_Async(MORSE_enum side, MORSE_enum uplo,
+int MORSE_zsymm_Tile_Async( MORSE_enum side, MORSE_enum uplo,
MORSE_Complex64_t alpha, MORSE_desc_t *A, MORSE_desc_t *B,
MORSE_Complex64_t beta, MORSE_desc_t *C,
- MORSE_sequence_t *sequence, MORSE_request_t *request)
+ MORSE_sequence_t *sequence, MORSE_request_t *request )
{
MORSE_context_t *morse;
@@ -345,10 +342,12 @@ int MORSE_zsymm_Tile_Async(MORSE_enum side, MORSE_enum uplo,
return MORSE_ERR_UNALLOCATED;
}
/* Check sequence status */
- if (sequence->status == MORSE_SUCCESS)
+ if (sequence->status == MORSE_SUCCESS) {
request->status = MORSE_SUCCESS;
- else
+ }
+ else {
return morse_request_fail(sequence, request, MORSE_ERR_SEQUENCE_FLUSHED);
+ }
/* Check descriptors for correctness */
if (morse_desc_check(A) != MORSE_SUCCESS) {
@@ -416,11 +415,13 @@ int MORSE_zsymm_Tile_Async(MORSE_enum side, MORSE_enum uplo,
}
/* Quick return */
- if (C->m == 0 || C->n == 0 ||
- ( (alpha == (MORSE_Complex64_t)0.0) && (beta == (MORSE_Complex64_t)1.0) ))
+ if ( (C->m == 0) || (C->n == 0) ||
+ ( (alpha == (MORSE_Complex64_t)0.0) && (beta == (MORSE_Complex64_t)1.0) ) )
+ {
return MORSE_SUCCESS;
+ }
- morse_pzsymm(side, uplo, alpha, A, B, beta, C, sequence, request);
+ morse_pzsymm( side, uplo, alpha, A, B, beta, C, sequence, request );
return MORSE_SUCCESS;
}
diff --git a/compute/zsyr2k.c b/compute/zsyr2k.c
index 50c9b234e36efeb66d03c3347ebc1c6e88a56f2d..84a501a4808176c987358ebbca2a9d9a0a5a2d27 100644
--- a/compute/zsyr2k.c
+++ b/compute/zsyr2k.c
@@ -83,7 +83,7 @@
* @param[in] LDB
* The leading dimension of the array B. LDB must be at least
* max( 1, N ), otherwise LDB must be at least max( 1, K ).
- *
+ *
* @param[in] beta
* beta specifies the scalar beta.
*
@@ -108,14 +108,16 @@
* @sa MORSE_ssyr2k
*
******************************************************************************/
-int MORSE_zsyr2k(MORSE_enum uplo, MORSE_enum trans, int N, int K,
- MORSE_Complex64_t alpha, MORSE_Complex64_t *A, int LDA, MORSE_Complex64_t *B, int LDB,
- MORSE_Complex64_t beta, MORSE_Complex64_t *C, int LDC)
+int MORSE_zsyr2k( MORSE_enum uplo, MORSE_enum trans, int N, int K,
+ MORSE_Complex64_t alpha, MORSE_Complex64_t *A, int LDA, MORSE_Complex64_t *B, int LDB,
+ MORSE_Complex64_t beta, MORSE_Complex64_t *C, int LDC )
{
int NB;
int Am, An;
int status;
- MORSE_desc_t descA, descB, descC;
+ MORSE_desc_t descAl, descAt;
+ MORSE_desc_t descBl, descBt;
+ MORSE_desc_t descCl, descCt;
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
MORSE_request_t request = MORSE_REQUEST_INITIALIZER;
@@ -135,7 +137,7 @@ int MORSE_zsyr2k(MORSE_enum uplo, MORSE_enum trans, int N, int K,
morse_error("MORSE_zsyr2k", "illegal value of trans");
return -2;
}
- if ( trans == MorseNoTrans ) {
+ if ( trans == MorseNoTrans ) {
Am = N; An = K;
} else {
Am = K; An = N;
@@ -176,42 +178,36 @@ int MORSE_zsyr2k(MORSE_enum uplo, MORSE_enum trans, int N, int K,
/* Set MT & NT & KT */
NB = MORSE_NB;
- morse_sequence_create(morse, &sequence);
+ morse_sequence_create( morse, &sequence );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooplap2tile( descA, A, NB, NB, LDA, An, 0, 0, Am, An, sequence, &request,
- morse_desc_mat_free(&(descA)) );
- morse_zooplap2tile( descB, B, NB, NB, LDB, An, 0, 0, Am, An, sequence, &request,
- morse_desc_mat_free(&(descA)); morse_desc_mat_free(&(descB)));
- morse_zooplap2tile( descC, C, NB, NB, LDC, N, 0, 0, N, N, sequence, &request,
- morse_desc_mat_free(&(descA)); morse_desc_mat_free(&(descB)); morse_desc_mat_free(&(descC)));
-/* } else {*/
-/* morse_ziplap2tile( descA, A, NB, NB, LDA, An, 0, 0, Am, An, */
-/* sequence, &request);*/
-/* morse_ziplap2tile( descB, B, NB, NB, LDB, An, 0, 0, Am, An, */
-/* sequence, &request);*/
-/* morse_ziplap2tile( descC, C, NB, NB, LDC, N, 0, 0, N, N, */
-/* sequence, &request);*/
-/* }*/
+ /* Submit the matrix conversion */
+ morse_zlap2tile( morse, &descAl, &descAt, MorseDescInput, MorseUpperLower,
+ A, NB, NB, LDA, An, Am, An, sequence, &request );
+ morse_zlap2tile( morse, &descBl, &descBt, MorseDescInput, MorseUpperLower,
+ B, NB, NB, LDB, An, Am, An, sequence, &request );
+ morse_zlap2tile( morse, &descCl, &descCt, MorseDescInout, uplo,
+ C, NB, NB, LDC, N, N, N, sequence, &request );
/* Call the tile interface */
- MORSE_zsyr2k_Tile_Async(uplo, trans, alpha, &descA, &descB, beta, &descC, sequence, &request);
+ MORSE_zsyr2k_Tile_Async( uplo, trans, alpha, &descAt, &descBt, beta, &descCt, sequence, &request );
+
+ /* Submit the matrix conversion back */
+ morse_ztile2lap( morse, &descAl, &descAt,
+ MorseDescInput, MorseUpperLower, sequence, &request );
+ morse_ztile2lap( morse, &descBl, &descBt,
+ MorseDescInput, MorseUpperLower, sequence, &request );
+ morse_ztile2lap( morse, &descCl, &descCt,
+ MorseDescInout, uplo, sequence, &request );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooptile2lap(descC, C, NB, NB, LDC, N, sequence, &request);
- morse_sequence_wait(morse, sequence);
- morse_desc_mat_free(&descA);
- morse_desc_mat_free(&descB);
- morse_desc_mat_free(&descC);
-/* } else {*/
-/* morse_ziptile2lap( descA, A, NB, NB, LDA, An, sequence, &request);*/
-/* morse_ziptile2lap( descB, B, NB, NB, LDB, An, sequence, &request);*/
-/* morse_ziptile2lap( descC, C, NB, NB, LDC, N, sequence, &request);*/
-/* morse_sequence_wait(morse, sequence);*/
-/* }*/
+ morse_sequence_wait( morse, sequence );
+
+ /* Cleanup the temporary data */
+ morse_ztile2lap_cleanup( morse, &descAl, &descAt );
+ morse_ztile2lap_cleanup( morse, &descBl, &descBt );
+ morse_ztile2lap_cleanup( morse, &descCl, &descCt );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -269,9 +265,9 @@ int MORSE_zsyr2k(MORSE_enum uplo, MORSE_enum trans, int N, int K,
* @sa MORSE_ssyr2k
*
******************************************************************************/
-int MORSE_zsyr2k_Tile(MORSE_enum uplo, MORSE_enum trans,
- MORSE_Complex64_t alpha, MORSE_desc_t *A, MORSE_desc_t *B,
- MORSE_Complex64_t beta, MORSE_desc_t *C)
+int MORSE_zsyr2k_Tile( MORSE_enum uplo, MORSE_enum trans,
+ MORSE_Complex64_t alpha, MORSE_desc_t *A, MORSE_desc_t *B,
+ MORSE_Complex64_t beta, MORSE_desc_t *C )
{
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
@@ -283,15 +279,17 @@ int MORSE_zsyr2k_Tile(MORSE_enum uplo, MORSE_enum trans,
morse_fatal_error("MORSE_zsyr2k_Tile", "MORSE not initialized");
return MORSE_ERR_NOT_INITIALIZED;
}
- morse_sequence_create(morse, &sequence);
- MORSE_zsyr2k_Tile_Async(uplo, trans, alpha, A, B, beta, C, sequence, &request);
- RUNTIME_desc_flush( A, sequence );
- RUNTIME_desc_flush( B, sequence );
- RUNTIME_desc_flush( C, sequence );
- morse_sequence_wait(morse, sequence);
-
+ morse_sequence_create( morse, &sequence );
+
+ MORSE_zsyr2k_Tile_Async( uplo, trans, alpha, A, B, beta, C, sequence, &request );
+
+ MORSE_Desc_Flush( A, sequence );
+ MORSE_Desc_Flush( B, sequence );
+ MORSE_Desc_Flush( C, sequence );
+
+ morse_sequence_wait( morse, sequence );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -323,10 +321,10 @@ int MORSE_zsyr2k_Tile(MORSE_enum uplo, MORSE_enum trans,
* @sa MORSE_ssyr2k_Tile_Async
*
******************************************************************************/
-int MORSE_zsyr2k_Tile_Async(MORSE_enum uplo, MORSE_enum trans,
- MORSE_Complex64_t alpha, MORSE_desc_t *A, MORSE_desc_t *B,
- MORSE_Complex64_t beta, MORSE_desc_t *C,
- MORSE_sequence_t *sequence, MORSE_request_t *request)
+int MORSE_zsyr2k_Tile_Async( MORSE_enum uplo, MORSE_enum trans,
+ MORSE_Complex64_t alpha, MORSE_desc_t *A, MORSE_desc_t *B,
+ MORSE_Complex64_t beta, MORSE_desc_t *C,
+ MORSE_sequence_t *sequence, MORSE_request_t *request )
{
MORSE_context_t *morse;
int N, K;
@@ -346,10 +344,12 @@ int MORSE_zsyr2k_Tile_Async(MORSE_enum uplo, MORSE_enum trans,
return MORSE_ERR_UNALLOCATED;
}
/* Check sequence status */
- if (sequence->status == MORSE_SUCCESS)
+ if (sequence->status == MORSE_SUCCESS) {
request->status = MORSE_SUCCESS;
- else
+ }
+ else {
return morse_request_fail(sequence, request, MORSE_ERR_SEQUENCE_FLUSHED);
+ }
/* Check descriptors for correctness */
if (morse_desc_check(A) != MORSE_SUCCESS) {
@@ -406,10 +406,10 @@ int MORSE_zsyr2k_Tile_Async(MORSE_enum uplo, MORSE_enum trans,
/* Quick return */
if ( N == 0 ||
- ((alpha == (MORSE_Complex64_t)0.0 || K == 0) && beta == (MORSE_Complex64_t)1.0))
+ ((alpha == (MORSE_Complex64_t)0.0 || K == 0) && beta == (MORSE_Complex64_t)1.0))
return MORSE_SUCCESS;
- morse_pzsyr2k(uplo, trans, alpha, A, B, beta, C, sequence, request);
+ morse_pzsyr2k( uplo, trans, alpha, A, B, beta, C, sequence, request );
return MORSE_SUCCESS;
}
diff --git a/compute/zsyrk.c b/compute/zsyrk.c
index 4b29d6b7f93eb1e267d901ae11a897b6119bc634..2dd3e6ac15d9a225cbc45c5b79fde39396386a61 100644
--- a/compute/zsyrk.c
+++ b/compute/zsyrk.c
@@ -98,14 +98,15 @@
* @sa MORSE_ssyrk
*
******************************************************************************/
-int MORSE_zsyrk(MORSE_enum uplo, MORSE_enum trans, int N, int K,
+int MORSE_zsyrk( MORSE_enum uplo, MORSE_enum trans, int N, int K,
MORSE_Complex64_t alpha, MORSE_Complex64_t *A, int LDA,
- MORSE_Complex64_t beta, MORSE_Complex64_t *C, int LDC)
+ MORSE_Complex64_t beta, MORSE_Complex64_t *C, int LDC )
{
int NB;
int Am, An;
int status;
- MORSE_desc_t descA, descC;
+ MORSE_desc_t descAl, descAt;
+ MORSE_desc_t descCl, descCt;
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
MORSE_request_t request = MORSE_REQUEST_INITIALIZER;
@@ -125,7 +126,7 @@ int MORSE_zsyrk(MORSE_enum uplo, MORSE_enum trans, int N, int K,
morse_error("MORSE_zsyrk", "illegal value of trans");
return -2;
}
- if ( trans == MorseNoTrans ) {
+ if ( trans == MorseNoTrans ) {
Am = N; An = K;
} else {
Am = K; An = N;
@@ -162,36 +163,31 @@ int MORSE_zsyrk(MORSE_enum uplo, MORSE_enum trans, int N, int K,
/* Set MT & NT & KT */
NB = MORSE_NB;
- morse_sequence_create(morse, &sequence);
+ morse_sequence_create( morse, &sequence );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooplap2tile( descA, A, NB, NB, LDA, An, 0, 0, Am, An, sequence, &request,
- morse_desc_mat_free(&(descA)) );
- morse_zooplap2tile( descC, C, NB, NB, LDC, N, 0, 0, N, N, sequence, &request,
- morse_desc_mat_free(&(descA)); morse_desc_mat_free(&(descC)));
-/* } else {*/
-/* morse_ziplap2tile( descA, A, NB, NB, LDA, An, 0, 0, Am, An, */
-/* sequence, &request);*/
-/* morse_ziplap2tile( descC, C, NB, NB, LDC, N, 0, 0, N, N, */
-/* sequence, &request);*/
-/* }*/
+ /* Submit the matrix conversion */
+ morse_zlap2tile( morse, &descAl, &descAt, MorseDescInput, MorseUpperLower,
+ A, NB, NB, LDA, An, Am, An, sequence, &request );
+ morse_zlap2tile( morse, &descCl, &descCt, MorseDescInout, uplo,
+ C, NB, NB, LDC, N, N, N, sequence, &request );
/* Call the tile interface */
- MORSE_zsyrk_Tile_Async(uplo, trans, alpha, &descA, beta, &descC, sequence, &request);
+ MORSE_zsyrk_Tile_Async( uplo, trans, alpha, &descAt, beta, &descCt, sequence, &request );
+
+ /* Submit the matrix conversion back */
+ morse_ztile2lap( morse, &descAl, &descAt,
+ MorseDescInput, MorseUpperLower, sequence, &request );
+ morse_ztile2lap( morse, &descCl, &descCt,
+ MorseDescInout, uplo, sequence, &request );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooptile2lap(descC, C, NB, NB, LDC, N, sequence, &request);
- morse_sequence_wait(morse, sequence);
- morse_desc_mat_free(&descA);
- morse_desc_mat_free(&descC);
-/* } else {*/
-/* morse_ziptile2lap( descA, A, NB, NB, LDA, An, sequence, &request);*/
-/* morse_ziptile2lap( descC, C, NB, NB, LDC, N, sequence, &request);*/
-/* morse_sequence_wait(morse, sequence);*/
-/* }*/
+ morse_sequence_wait( morse, sequence );
+
+ /* Cleanup the temporary data */
+ morse_ztile2lap_cleanup( morse, &descAl, &descAt );
+ morse_ztile2lap_cleanup( morse, &descCl, &descCt );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -245,9 +241,9 @@ int MORSE_zsyrk(MORSE_enum uplo, MORSE_enum trans, int N, int K,
* @sa MORSE_ssyrk
*
******************************************************************************/
-int MORSE_zsyrk_Tile(MORSE_enum uplo, MORSE_enum trans,
+int MORSE_zsyrk_Tile( MORSE_enum uplo, MORSE_enum trans,
MORSE_Complex64_t alpha, MORSE_desc_t *A,
- MORSE_Complex64_t beta, MORSE_desc_t *C)
+ MORSE_Complex64_t beta, MORSE_desc_t *C )
{
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
@@ -259,14 +255,16 @@ int MORSE_zsyrk_Tile(MORSE_enum uplo, MORSE_enum trans,
morse_fatal_error("MORSE_zsyrk_Tile", "MORSE not initialized");
return MORSE_ERR_NOT_INITIALIZED;
}
- morse_sequence_create(morse, &sequence);
- MORSE_zsyrk_Tile_Async(uplo, trans, alpha, A, beta, C, sequence, &request);
- RUNTIME_desc_flush( A, sequence );
- RUNTIME_desc_flush( C, sequence );
- morse_sequence_wait(morse, sequence);
-
+ morse_sequence_create( morse, &sequence );
+
+ MORSE_zsyrk_Tile_Async( uplo, trans, alpha, A, beta, C, sequence, &request );
+
+ MORSE_Desc_Flush( A, sequence );
+ MORSE_Desc_Flush( C, sequence );
+
+ morse_sequence_wait( morse, sequence );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -298,10 +296,10 @@ int MORSE_zsyrk_Tile(MORSE_enum uplo, MORSE_enum trans,
* @sa MORSE_ssyrk_Tile_Async
*
******************************************************************************/
-int MORSE_zsyrk_Tile_Async(MORSE_enum uplo, MORSE_enum trans,
+int MORSE_zsyrk_Tile_Async( MORSE_enum uplo, MORSE_enum trans,
MORSE_Complex64_t alpha, MORSE_desc_t *A,
MORSE_Complex64_t beta, MORSE_desc_t *C,
- MORSE_sequence_t *sequence, MORSE_request_t *request)
+ MORSE_sequence_t *sequence, MORSE_request_t *request )
{
MORSE_context_t *morse;
int N, K;
@@ -321,10 +319,12 @@ int MORSE_zsyrk_Tile_Async(MORSE_enum uplo, MORSE_enum trans,
return MORSE_ERR_UNALLOCATED;
}
/* Check sequence status */
- if (sequence->status == MORSE_SUCCESS)
+ if (sequence->status == MORSE_SUCCESS) {
request->status = MORSE_SUCCESS;
- else
+ }
+ else {
return morse_request_fail(sequence, request, MORSE_ERR_SEQUENCE_FLUSHED);
+ }
/* Check descriptors for correctness */
if (morse_desc_check(A) != MORSE_SUCCESS) {
@@ -377,10 +377,10 @@ int MORSE_zsyrk_Tile_Async(MORSE_enum uplo, MORSE_enum trans,
/* Quick return */
if ( N == 0 ||
- ((alpha == (MORSE_Complex64_t)0.0 || K == 0) && beta == (MORSE_Complex64_t)1.0))
+ ((alpha == (MORSE_Complex64_t)0.0 || K == 0) && beta == (MORSE_Complex64_t)1.0))
return MORSE_SUCCESS;
- morse_pzsyrk(uplo, trans, alpha, A, beta, C, sequence, request);
+ morse_pzsyrk( uplo, trans, alpha, A, beta, C, sequence, request );
return MORSE_SUCCESS;
}
diff --git a/compute/zsysv.c b/compute/zsysv.c
index d39bda8bfff3341c1453dcd894478e17d9de50e2..a53b8cb57eb575da39f9b05387dec10f8483b09f 100644
--- a/compute/zsysv.c
+++ b/compute/zsysv.c
@@ -92,16 +92,17 @@
* @sa MORSE_csysv
*
******************************************************************************/
-int MORSE_zsysv(MORSE_enum uplo, int N, int NRHS,
- MORSE_Complex64_t *A, int LDA,
- MORSE_Complex64_t *B, int LDB)
+int MORSE_zsysv( MORSE_enum uplo, int N, int NRHS,
+ MORSE_Complex64_t *A, int LDA,
+ MORSE_Complex64_t *B, int LDB )
{
int NB;
int status;
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
MORSE_request_t request = MORSE_REQUEST_INITIALIZER;
- MORSE_desc_t descA, descB;
+ MORSE_desc_t descAl, descAt;
+ MORSE_desc_t descBl, descBt;
morse = morse_context_self();
if (morse == NULL) {
@@ -109,7 +110,7 @@ int MORSE_zsysv(MORSE_enum uplo, int N, int NRHS,
return MORSE_ERR_NOT_INITIALIZED;
}
/* Check input arguments */
- if (uplo != MorseUpper && uplo != MorseLower) {
+ if ((uplo != MorseUpper) && (uplo != MorseLower)) {
morse_error("MORSE_zsysv", "illegal value of uplo");
return -1;
}
@@ -144,37 +145,31 @@ int MORSE_zsysv(MORSE_enum uplo, int N, int NRHS,
/* Set NT & NTRHS */
NB = MORSE_NB;
- morse_sequence_create(morse, &sequence);
+ morse_sequence_create( morse, &sequence );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooplap2tile( descA, A, NB, NB, LDA, N, 0, 0, N, N, sequence, &request,
- morse_desc_mat_free(&(descA)) );
- morse_zooplap2tile( descB, B, NB, NB, LDB, NRHS, 0, 0, N, NRHS, sequence, &request,
- morse_desc_mat_free(&(descA)); morse_desc_mat_free(&(descB)));
-/* } else {*/
-/* morse_ziplap2tile( descA, A, NB, NB, LDA, N, 0, 0, N, N, */
-/* sequence, &request);*/
-/* morse_ziplap2tile( descB, B, NB, NB, LDB, NRHS, 0, 0, N, NRHS,*/
-/* sequence, &request);*/
-/* }*/
+ /* Submit the matrix conversion */
+ morse_zlap2tile( morse, &descAl, &descAt, MorseDescInout, uplo,
+ A, NB, NB, LDA, N, N, N, sequence, &request );
+ morse_zlap2tile( morse, &descBl, &descBt, MorseDescInout, MorseUpperLower,
+ B, NB, NB, LDB, NRHS, N, NRHS, sequence, &request );
/* Call the tile interface */
- MORSE_zsysv_Tile_Async(uplo, &descA, &descB, sequence, &request);
+ MORSE_zsysv_Tile_Async( uplo, &descAt, &descBt, sequence, &request );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooptile2lap(descA, A, NB, NB, LDA, N, sequence, &request);
- morse_zooptile2lap(descB, B, NB, NB, LDB, NRHS, sequence, &request);
- morse_sequence_wait(morse, sequence);
- morse_desc_mat_free(&descA);
- morse_desc_mat_free(&descB);
-/* } else {*/
-/* morse_ziptile2lap( descA, A, NB, NB, LDA, N, sequence, &request);*/
-/* morse_ziptile2lap( descB, B, NB, NB, LDB, NRHS, sequence, &request);*/
-/* morse_sequence_wait(morse, sequence);*/
-/* }*/
+ /* Submit the matrix conversion back */
+ morse_ztile2lap( morse, &descAl, &descAt,
+ MorseDescInout, uplo, sequence, &request );
+ morse_ztile2lap( morse, &descBl, &descBt,
+ MorseDescInout, MorseUpperLower, sequence, &request );
+
+ morse_sequence_wait( morse, sequence );
+
+ /* Cleanup the temporary data */
+ morse_ztile2lap_cleanup( morse, &descAl, &descAt );
+ morse_ztile2lap_cleanup( morse, &descBl, &descBt );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
/**
@@ -221,9 +216,9 @@ int MORSE_zsysv(MORSE_enum uplo, int N, int NRHS,
* @sa MORSE_zsysv
* @sa MORSE_zsysv_Tile_Async
* @sa MORSE_csysv_Tile
-*
+ *
******************************************************************************/
-int MORSE_zsysv_Tile(MORSE_enum uplo, MORSE_desc_t *A, MORSE_desc_t *B)
+int MORSE_zsysv_Tile( MORSE_enum uplo, MORSE_desc_t *A, MORSE_desc_t *B )
{
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
@@ -235,14 +230,16 @@ int MORSE_zsysv_Tile(MORSE_enum uplo, MORSE_desc_t *A, MORSE_desc_t *B)
morse_fatal_error("MORSE_zsysv_Tile", "MORSE not initialized");
return MORSE_ERR_NOT_INITIALIZED;
}
- morse_sequence_create(morse, &sequence);
- MORSE_zsysv_Tile_Async(uplo, A, B, sequence, &request);
- RUNTIME_desc_flush( A, sequence );
- RUNTIME_desc_flush( B, sequence );
- morse_sequence_wait(morse, sequence);
+ morse_sequence_create( morse, &sequence );
+
+ MORSE_zsysv_Tile_Async( uplo, A, B, sequence, &request );
+
+ MORSE_Desc_Flush( A, sequence );
+ MORSE_Desc_Flush( B, sequence );
+ morse_sequence_wait( morse, sequence );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -273,8 +270,8 @@ int MORSE_zsysv_Tile(MORSE_enum uplo, MORSE_desc_t *A, MORSE_desc_t *B)
* @sa MORSE_csysv_Tile_Async
*
******************************************************************************/
-int MORSE_zsysv_Tile_Async(MORSE_enum uplo, MORSE_desc_t *A, MORSE_desc_t *B,
- MORSE_sequence_t *sequence, MORSE_request_t *request)
+int MORSE_zsysv_Tile_Async( MORSE_enum uplo, MORSE_desc_t *A, MORSE_desc_t *B,
+ MORSE_sequence_t *sequence, MORSE_request_t *request )
{
MORSE_context_t *morse;
@@ -292,10 +289,12 @@ int MORSE_zsysv_Tile_Async(MORSE_enum uplo, MORSE_desc_t *A, MORSE_desc_t *B,
return MORSE_ERR_UNALLOCATED;
}
/* Check sequence status */
- if (sequence->status == MORSE_SUCCESS)
+ if (sequence->status == MORSE_SUCCESS) {
request->status = MORSE_SUCCESS;
- else
+ }
+ else {
return morse_request_fail(sequence, request, MORSE_ERR_SEQUENCE_FLUSHED);
+ }
/* Check descriptors for correctness */
if (morse_desc_check(A) != MORSE_SUCCESS) {
@@ -311,21 +310,21 @@ int MORSE_zsysv_Tile_Async(MORSE_enum uplo, MORSE_desc_t *A, MORSE_desc_t *B,
morse_error("MORSE_zsysv_Tile", "only square tiles supported");
return morse_request_fail(sequence, request, MORSE_ERR_ILLEGAL_VALUE);
}
- if (uplo != MorseUpper && uplo != MorseLower) {
+ if ((uplo != MorseUpper) && (uplo != MorseLower)) {
morse_error("MORSE_zsysv_Tile", "illegal value of uplo");
return morse_request_fail(sequence, request, -1);
}
/* Quick return - currently NOT equivalent to LAPACK's
* LAPACK does not have such check for Dsysv */
-/*
- if (chameleon_min(N, NRHS) == 0)
- return MORSE_SUCCESS;
-*/
- morse_pzsytrf(uplo, A, sequence, request);
+ /*
+ if (chameleon_min(N, NRHS) == 0)
+ return MORSE_SUCCESS;
+ */
+ morse_pzsytrf( uplo, A, sequence, request );
- morse_pztrsm(MorseLeft, uplo, uplo == MorseUpper ? MorseTrans : MorseNoTrans, MorseNonUnit, 1.0, A, B, sequence, request);
+ morse_pztrsm( MorseLeft, uplo, uplo == MorseUpper ? MorseTrans : MorseNoTrans, MorseNonUnit, 1.0, A, B, sequence, request );
- morse_pztrsm(MorseLeft, uplo, uplo == MorseUpper ? MorseNoTrans : MorseTrans, MorseNonUnit, 1.0, A, B, sequence, request);
+ morse_pztrsm( MorseLeft, uplo, uplo == MorseUpper ? MorseNoTrans : MorseTrans, MorseNonUnit, 1.0, A, B, sequence, request );
return MORSE_SUCCESS;
}
diff --git a/compute/zsytrf.c b/compute/zsytrf.c
index c2e3eeead99ea9237859064a13a66558fdc25360..11d74b77576fe3474cfd29ef9396a1788abd88c9 100644
--- a/compute/zsytrf.c
+++ b/compute/zsytrf.c
@@ -78,15 +78,15 @@
* @sa MORSE_zpotrs
*
******************************************************************************/
-int MORSE_zsytrf(MORSE_enum uplo, int N,
- MORSE_Complex64_t *A, int LDA)
+int MORSE_zsytrf( MORSE_enum uplo, int N,
+ MORSE_Complex64_t *A, int LDA )
{
int NB;
int status;
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
MORSE_request_t request = MORSE_REQUEST_INITIALIZER;
- MORSE_desc_t descA;
+ MORSE_desc_t descAl, descAt;
morse = morse_context_self();
if (morse == NULL) {
@@ -94,7 +94,7 @@ int MORSE_zsytrf(MORSE_enum uplo, int N,
return MORSE_ERR_NOT_INITIALIZED;
}
/* Check input arguments */
- if (uplo != MorseUpper && uplo != MorseLower) {
+ if ((uplo != MorseUpper) && (uplo != MorseLower)) {
morse_error("MORSE_zsytrf", "illegal value of uplo");
return -1;
}
@@ -120,30 +120,26 @@ int MORSE_zsytrf(MORSE_enum uplo, int N,
/* Set NT */
NB = MORSE_NB;
- morse_sequence_create(morse, &sequence);
+ morse_sequence_create( morse, &sequence );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooplap2tile( descA, A, NB, NB, LDA, N, 0, 0, N, N, sequence, &request,
- morse_desc_mat_free(&(descA)) );
-/* } else {*/
-/* morse_ziplap2tile( descA, A, NB, NB, LDA, N, 0, 0, N, N,*/
-/* sequence, &request);*/
-/* }*/
+ /* Submit the matrix conversion */
+ morse_zlap2tile( morse, &descAl, &descAt, MorseDescInout, uplo,
+ A, NB, NB, LDA, N, N, N, sequence, &request );
/* Call the tile interface */
- MORSE_zsytrf_Tile_Async(uplo, &descA, sequence, &request);
+ MORSE_zsytrf_Tile_Async( uplo, &descAt, sequence, &request );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooptile2lap(descA, A, NB, NB, LDA, N, sequence, &request);
- morse_sequence_wait(morse, sequence);
- morse_desc_mat_free(&descA);
-/* } else {*/
-/* morse_ziptile2lap( descA, A, NB, NB, LDA, N, sequence, &request);*/
-/* morse_sequence_wait(morse, sequence);*/
-/* }*/
+ /* Submit the matrix conversion back */
+ morse_ztile2lap( morse, &descAl, &descAt,
+ MorseDescInout, uplo, sequence, &request );
+
+ morse_sequence_wait( morse, sequence );
+
+ /* Cleanup the temporary data */
+ morse_ztile2lap_cleanup( morse, &descAl, &descAt );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -188,7 +184,7 @@ int MORSE_zsytrf(MORSE_enum uplo, int N,
* @sa MORSE_csytrf_Tile
*
******************************************************************************/
-int MORSE_zsytrf_Tile(MORSE_enum uplo, MORSE_desc_t *A)
+int MORSE_zsytrf_Tile( MORSE_enum uplo, MORSE_desc_t *A )
{
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
@@ -200,13 +196,15 @@ int MORSE_zsytrf_Tile(MORSE_enum uplo, MORSE_desc_t *A)
morse_fatal_error("MORSE_zsytrf_Tile", "MORSE not initialized");
return MORSE_ERR_NOT_INITIALIZED;
}
- morse_sequence_create(morse, &sequence);
- MORSE_zsytrf_Tile_Async(uplo, A, sequence, &request);
- RUNTIME_desc_flush( A, sequence );
- morse_sequence_wait(morse, sequence);
+ morse_sequence_create( morse, &sequence );
+
+ MORSE_zsytrf_Tile_Async( uplo, A, sequence, &request );
+
+ MORSE_Desc_Flush( A, sequence );
+ morse_sequence_wait( morse, sequence );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -237,8 +235,8 @@ int MORSE_zsytrf_Tile(MORSE_enum uplo, MORSE_desc_t *A)
* @sa MORSE_csytrf_Tile_Async
*
******************************************************************************/
-int MORSE_zsytrf_Tile_Async(MORSE_enum uplo, MORSE_desc_t *A,
- MORSE_sequence_t *sequence, MORSE_request_t *request)
+int MORSE_zsytrf_Tile_Async( MORSE_enum uplo, MORSE_desc_t *A,
+ MORSE_sequence_t *sequence, MORSE_request_t *request )
{
MORSE_context_t *morse;
@@ -256,10 +254,12 @@ int MORSE_zsytrf_Tile_Async(MORSE_enum uplo, MORSE_desc_t *A,
return MORSE_ERR_UNALLOCATED;
}
/* Check sequence status */
- if (sequence->status == MORSE_SUCCESS)
+ if (sequence->status == MORSE_SUCCESS) {
request->status = MORSE_SUCCESS;
- else
+ }
+ else {
return morse_request_fail(sequence, request, MORSE_ERR_SEQUENCE_FLUSHED);
+ }
/* Check descriptors for correctness */
if (morse_desc_check(A) != MORSE_SUCCESS) {
@@ -271,16 +271,16 @@ int MORSE_zsytrf_Tile_Async(MORSE_enum uplo, MORSE_desc_t *A,
morse_error("MORSE_zsytrf_Tile_Async", "only square tiles supported");
return morse_request_fail(sequence, request, MORSE_ERR_ILLEGAL_VALUE);
}
- if (uplo != MorseUpper && uplo != MorseLower) {
+ if ((uplo != MorseUpper) && (uplo != MorseLower)) {
morse_error("MORSE_zsytrf_Tile_Async", "illegal value of uplo");
return morse_request_fail(sequence, request, -1);
}
/* Quick return */
-/*
- if (chameleon_max(N, 0) == 0)
- return MORSE_SUCCESS;
-*/
- morse_pzsytrf(uplo, A, sequence, request);
+ /*
+ if (chameleon_max(N, 0) == 0)
+ return MORSE_SUCCESS;
+ */
+ morse_pzsytrf( uplo, A, sequence, request );
return MORSE_SUCCESS;
}
diff --git a/compute/zsytrs.c b/compute/zsytrs.c
index df6f7d9dea81ba12ef7e6f443d14f3f9105b2796..7223993d95d2b857c8eebdb664f1c318fecf9333 100644
--- a/compute/zsytrs.c
+++ b/compute/zsytrs.c
@@ -79,16 +79,17 @@
* @sa MORSE_csytrs
*
******************************************************************************/
-int MORSE_zsytrs(MORSE_enum uplo, int N, int NRHS,
- MORSE_Complex64_t *A, int LDA,
- MORSE_Complex64_t *B, int LDB)
+int MORSE_zsytrs( MORSE_enum uplo, int N, int NRHS,
+ MORSE_Complex64_t *A, int LDA,
+ MORSE_Complex64_t *B, int LDB )
{
int NB;
int status;
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
MORSE_request_t request = MORSE_REQUEST_INITIALIZER;
- MORSE_desc_t descA, descB;
+ MORSE_desc_t descAl, descAt;
+ MORSE_desc_t descBl, descBt;
morse = morse_context_self();
if (morse == NULL) {
@@ -96,7 +97,7 @@ int MORSE_zsytrs(MORSE_enum uplo, int N, int NRHS,
return MORSE_ERR_NOT_INITIALIZED;
}
/* Check input arguments */
- if (uplo != MorseUpper && uplo != MorseLower) {
+ if ((uplo != MorseUpper) && (uplo != MorseLower)) {
morse_error("MORSE_zsytrs", "illegal value of uplo");
return -1;
}
@@ -130,36 +131,31 @@ int MORSE_zsytrs(MORSE_enum uplo, int N, int NRHS,
/* Set NT & NTRHS */
NB = MORSE_NB;
- morse_sequence_create(morse, &sequence);
+ morse_sequence_create( morse, &sequence );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooplap2tile( descA, A, NB, NB, LDA, N, 0, 0, N, N, sequence, &request,
- morse_desc_mat_free(&(descA)) );
- morse_zooplap2tile( descB, B, NB, NB, LDB, NRHS, 0, 0, N, NRHS, sequence, &request,
- morse_desc_mat_free(&(descA)); morse_desc_mat_free(&(descB)));
-/* } else {*/
-/* morse_ziplap2tile( descA, A, NB, NB, LDA, N, 0, 0, N, N, */
-/* sequence, &request);*/
-/* morse_ziplap2tile( descB, B, NB, NB, LDB, NRHS, 0, 0, N, NRHS,*/
-/* sequence, &request);*/
-/* }*/
+ /* Submit the matrix conversion */
+ morse_zlap2tile( morse, &descAl, &descAt, MorseDescInput, uplo,
+ A, NB, NB, LDA, N, N, N, sequence, &request );
+ morse_zlap2tile( morse, &descBl, &descBt, MorseDescInout, MorseUpperLower,
+ B, NB, NB, LDB, NRHS, N, NRHS, sequence, &request );
/* Call the tile interface */
- MORSE_zsytrs_Tile_Async(uplo, &descA, &descB, sequence, &request);
+ MORSE_zsytrs_Tile_Async( uplo, &descAt, &descBt, sequence, &request );
+
+ /* Submit the matrix conversion back */
+ morse_ztile2lap( morse, &descAl, &descAt,
+ MorseDescInput, uplo, sequence, &request );
+ morse_ztile2lap( morse, &descBl, &descBt,
+ MorseDescInout, MorseUpperLower, sequence, &request );
+
+ morse_sequence_wait( morse, sequence );
+
+ /* Cleanup the temporary data */
+ morse_ztile2lap_cleanup( morse, &descAl, &descAt );
+ morse_ztile2lap_cleanup( morse, &descBl, &descBt );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooptile2lap(descB, B, NB, NB, LDB, NRHS, sequence, &request);
- morse_sequence_wait(morse, sequence);
- morse_desc_mat_free(&descA);
- morse_desc_mat_free(&descB);
-/* } else {*/
-/* morse_ziptile2lap( descA, A, NB, NB, LDA, N, sequence, &request);*/
-/* morse_ziptile2lap( descB, B, NB, NB, LDB, NRHS, sequence, &request);*/
-/* morse_sequence_wait(morse, sequence);*/
-/* }*/
-
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -201,7 +197,7 @@ int MORSE_zsytrs(MORSE_enum uplo, int N, int NRHS,
* @sa MORSE_csytrs_Tile
*
******************************************************************************/
-int MORSE_zsytrs_Tile(MORSE_enum uplo, MORSE_desc_t *A, MORSE_desc_t *B)
+int MORSE_zsytrs_Tile( MORSE_enum uplo, MORSE_desc_t *A, MORSE_desc_t *B )
{
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
@@ -213,14 +209,16 @@ int MORSE_zsytrs_Tile(MORSE_enum uplo, MORSE_desc_t *A, MORSE_desc_t *B)
morse_fatal_error("MORSE_zsytrs_Tile", "MORSE not initialized");
return MORSE_ERR_NOT_INITIALIZED;
}
- morse_sequence_create(morse, &sequence);
- MORSE_zsytrs_Tile_Async(uplo, A, B, sequence, &request);
- RUNTIME_desc_flush( A, sequence );
- RUNTIME_desc_flush( B, sequence );
- morse_sequence_wait(morse, sequence);
+ morse_sequence_create( morse, &sequence );
+
+ MORSE_zsytrs_Tile_Async( uplo, A, B, sequence, &request );
+
+ MORSE_Desc_Flush( A, sequence );
+ MORSE_Desc_Flush( B, sequence );
+ morse_sequence_wait( morse, sequence );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -251,8 +249,8 @@ int MORSE_zsytrs_Tile(MORSE_enum uplo, MORSE_desc_t *A, MORSE_desc_t *B)
* @sa MORSE_csytrs_Tile_Async
*
******************************************************************************/
-int MORSE_zsytrs_Tile_Async(MORSE_enum uplo, MORSE_desc_t *A, MORSE_desc_t *B,
- MORSE_sequence_t *sequence, MORSE_request_t *request)
+int MORSE_zsytrs_Tile_Async( MORSE_enum uplo, MORSE_desc_t *A, MORSE_desc_t *B,
+ MORSE_sequence_t *sequence, MORSE_request_t *request )
{
MORSE_context_t *morse;
@@ -270,10 +268,12 @@ int MORSE_zsytrs_Tile_Async(MORSE_enum uplo, MORSE_desc_t *A, MORSE_desc_t *B,
return MORSE_ERR_UNALLOCATED;
}
/* Check sequence status */
- if (sequence->status == MORSE_SUCCESS)
+ if (sequence->status == MORSE_SUCCESS) {
request->status = MORSE_SUCCESS;
- else
+ }
+ else {
return morse_request_fail(sequence, request, MORSE_ERR_SEQUENCE_FLUSHED);
+ }
/* Check descriptors for correctness */
if (morse_desc_check(A) != MORSE_SUCCESS) {
@@ -289,18 +289,18 @@ int MORSE_zsytrs_Tile_Async(MORSE_enum uplo, MORSE_desc_t *A, MORSE_desc_t *B,
morse_error("MORSE_zsytrs_Tile", "only square tiles supported");
return morse_request_fail(sequence, request, MORSE_ERR_ILLEGAL_VALUE);
}
- if (uplo != MorseUpper && uplo != MorseLower) {
+ if ((uplo != MorseUpper) && (uplo != MorseLower)) {
morse_error("MORSE_zsytrs_Tile", "illegal value of uplo");
return morse_request_fail(sequence, request, -1);
}
/* Quick return */
-/*
- if (chameleon_min(N, NRHS) == 0)
- return MORSE_SUCCESS;
-*/
- morse_pztrsm(MorseLeft, uplo, uplo == MorseUpper ? MorseTrans : MorseNoTrans, MorseNonUnit, 1.0, A, B, sequence, request);
+ /*
+ if (chameleon_min(N, NRHS) == 0)
+ return MORSE_SUCCESS;
+ */
+ morse_pztrsm( MorseLeft, uplo, uplo == MorseUpper ? MorseTrans : MorseNoTrans, MorseNonUnit, 1.0, A, B, sequence, request );
- morse_pztrsm(MorseLeft, uplo, uplo == MorseUpper ? MorseNoTrans : MorseTrans, MorseNonUnit, 1.0, A, B, sequence, request);
+ morse_pztrsm( MorseLeft, uplo, uplo == MorseUpper ? MorseNoTrans : MorseTrans, MorseNonUnit, 1.0, A, B, sequence, request );
return MORSE_SUCCESS;
}
diff --git a/compute/ztile.c b/compute/ztile.c
index 5377a78643d356f397c3e46011c097a8113a157d..868a565e4178542648efe0234f848cd32a651df6 100644
--- a/compute/ztile.c
+++ b/compute/ztile.c
@@ -55,18 +55,18 @@
*
*******************************************************************************
*
- * @sa MORSE_zLapack_to_Tile_Async
* @sa MORSE_zTile_to_Lapack
* @sa MORSE_cLapack_to_Tile
* @sa MORSE_dLapack_to_Tile
* @sa MORSE_sLapack_to_Tile
*
******************************************************************************/
-int MORSE_zLapack_to_Tile(MORSE_Complex64_t *Af77, int LDA, MORSE_desc_t *A)
+int MORSE_zLapack_to_Tile( MORSE_Complex64_t *Af77, int LDA, MORSE_desc_t *A )
{
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
MORSE_request_t request;
+ MORSE_desc_t *B;
int status;
morse = morse_context_self();
@@ -79,79 +79,28 @@ int MORSE_zLapack_to_Tile(MORSE_Complex64_t *Af77, int LDA, MORSE_desc_t *A)
morse_error("MORSE_zLapack_to_Tile", "invalid descriptor");
return MORSE_ERR_ILLEGAL_VALUE;
}
- morse_sequence_create(morse, &sequence);
- morse_pzlapack_to_tile( Af77, LDA, A, sequence, &request);
+ /* Create the B descriptor to handle the Lapack format matrix */
+ MORSE_Desc_Create_User( &B, Af77, MorseComplexDouble, A->mb, A->nb, A->bsiz,
+ LDA, A->n, 0, 0, A->m, A->n, 1, 1,
+ morse_getaddr_cm, morse_getblkldd_cm, NULL );
- RUNTIME_desc_flush( A, sequence );
- RUNTIME_sequence_wait( morse, sequence );
+ /* Start the computation */
+ morse_sequence_create( morse, &sequence );
- status = sequence->status;
- morse_sequence_destroy(morse, sequence);
- return status;
-}
+ morse_pzlacpy( MorseUpperLower, B, A, sequence, &request );
-/**
- ********************************************************************************
- *
- * @ingroup MORSE_Complex64_t_Tile_Async
- *
- * MORSE_zLapack_to_Tile_Async - Conversion from LAPACK layout to tile layout.
- * Non-blocking equivalent of MORSE_zLapack_to_Tile().
- * May return before the computation is finished.
- * Allows for pipelining of operations ar runtime.
- *
- *
- *******************************************************************************
- *
- * @param[in] Af77
- * LAPACK matrix.
- *
- * @param[in] LDA
- * The leading dimension of the matrix Af77.
- *
- * @param[in,out] A
- * Descriptor of the MORSE matrix in tile layout.
- * If MORSE_TRANSLATION_MODE is set to MORSE_INPLACE,
- * A->mat is not used and set to Af77 when returns, else if
- * MORSE_TRANSLATION_MODE is set to MORSE_OUTOFPLACE,
- * A->mat has to be allocated before.
- *
- * @param[in] sequence
- * Identifies the sequence of function calls that this call belongs to
- * (for completion checks and exception handling purposes).
- *
- * @param[out] request
- * Identifies this function call (for exception handling purposes).
- *
- *******************************************************************************
- *
- * @sa MORSE_zTile_to_Lapack_Async
- * @sa MORSE_zLapack_to_Tile
- * @sa MORSE_cLapack_to_Tile_Async
- * @sa MORSE_dLapack_to_Tile_Async
- * @sa MORSE_sLapack_to_Tile_Async
- *
- ******************************************************************************/
-int MORSE_zLapack_to_Tile_Async(MORSE_Complex64_t *Af77, int LDA, MORSE_desc_t *A,
- MORSE_sequence_t *sequence, MORSE_request_t *request)
-{
- MORSE_context_t *morse;
+ MORSE_Desc_Flush( B, sequence );
+ MORSE_Desc_Flush( A, sequence );
- morse = morse_context_self();
- if (morse == NULL) {
- morse_fatal_error("MORSE_zLapack_to_Tile", "MORSE not initialized");
- return MORSE_ERR_NOT_INITIALIZED;
- }
- /* Check descriptor for correctness */
- if (morse_desc_check( A ) != MORSE_SUCCESS) {
- morse_error("MORSE_zLapack_to_Tile", "invalid descriptor");
- return MORSE_ERR_ILLEGAL_VALUE;
- }
+ morse_sequence_wait( morse, sequence );
- morse_pzlapack_to_tile( Af77, LDA, A, sequence, request);
+ /* Destroy temporary B descriptor */
+ MORSE_Desc_Destroy( &B );
- return MORSE_SUCCESS;
+ status = sequence->status;
+ morse_sequence_destroy( morse, sequence );
+ return status;
}
/**
@@ -183,18 +132,18 @@ int MORSE_zLapack_to_Tile_Async(MORSE_Complex64_t *Af77, int LDA, MORSE_desc_t *
*
*******************************************************************************
*
- * @sa MORSE_zTile_to_Lapack_Async
* @sa MORSE_zLapack_to_Tile
* @sa MORSE_cTile_to_Lapack
* @sa MORSE_dTile_to_Lapack
* @sa MORSE_sTile_to_Lapack
*
-******************************************************************************/
-int MORSE_zTile_to_Lapack(MORSE_desc_t *A, MORSE_Complex64_t *Af77, int LDA)
+ ******************************************************************************/
+int MORSE_zTile_to_Lapack( MORSE_desc_t *A, MORSE_Complex64_t *Af77, int LDA )
{
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
MORSE_request_t request;
+ MORSE_desc_t *B;
int status;
morse = morse_context_self();
@@ -207,76 +156,25 @@ int MORSE_zTile_to_Lapack(MORSE_desc_t *A, MORSE_Complex64_t *Af77, int LDA)
morse_error("MORSE_zTile_to_Lapack", "invalid descriptor");
return MORSE_ERR_ILLEGAL_VALUE;
}
- morse_sequence_create(morse, &sequence);
- morse_pztile_to_lapack( A, Af77, LDA, sequence, &request);
- RUNTIME_desc_flush( A, sequence );
- RUNTIME_sequence_wait( morse, sequence );
+ /* Create the B descriptor to handle the Lapack format matrix */
+ MORSE_Desc_Create_User( &B, Af77, MorseComplexDouble, A->mb, A->nb, A->bsiz,
+ LDA, A->n, 0, 0, A->m, A->n, 1, 1,
+ morse_getaddr_cm, morse_getblkldd_cm, NULL );
- status = sequence->status;
- morse_sequence_destroy(morse, sequence);
- return status;
-}
+ /* Start the computation */
+ morse_sequence_create( morse, &sequence );
-/**
- ********************************************************************************
- *
- * @ingroup MORSE_Complex64_t_Tile_Async
- *
- * MORSE_zTile_to_Lapack_Async - Conversion from LAPACK layout to tile layout.
- * Non-blocking equivalent of MORSE_zTile_to_Lapack().
- * May return before the computation is finished.
- * Allows for pipelining of operations ar runtime.
- *
- *
- *******************************************************************************
- *
- * @param[in] A
- * Descriptor of the MORSE matrix in tile layout.
- *
- * @param[in,out] Af77
- * LAPACK matrix.
- * If MORSE_TRANSLATION_MODE is set to MORSE_INPLACE,
- * Af77 has to be A->mat, else if
- * MORSE_TRANSLATION_MODE is set to MORSE_OUTOFPLACE,
- * Af77 has to be allocated before.
- *
- * @param[in] LDA
- * The leading dimension of the matrix Af77.
- *
- * @param[in] sequence
- * Identifies the sequence of function calls that this call belongs to
- * (for completion checks and exception handling purposes).
- *
- * @param[out] request
- * Identifies this function call (for exception handling purposes).
- *
- *******************************************************************************
- *
- * @sa MORSE_zLapack_to_Tile_Async
- * @sa MORSE_zTile_to_Lapack
- * @sa MORSE_cTile_to_Lapack_Async
- * @sa MORSE_dTile_to_Lapack_Async
- * @sa MORSE_sTile_to_Lapack_Async
- *
- ******************************************************************************/
-int MORSE_zTile_to_Lapack_Async(MORSE_desc_t *A, MORSE_Complex64_t *Af77, int LDA,
- MORSE_sequence_t *sequence, MORSE_request_t *request)
-{
- MORSE_context_t *morse;
+ morse_pzlacpy( MorseUpperLower, A, B, sequence, &request );
- morse = morse_context_self();
- if (morse == NULL) {
- morse_fatal_error("MORSE_zTile_to_Lapack", "MORSE not initialized");
- return MORSE_ERR_NOT_INITIALIZED;
- }
- /* Check descriptor for correctness */
- if (morse_desc_check( A ) != MORSE_SUCCESS) {
- morse_error("MORSE_zTile_to_Lapack", "invalid descriptor");
- return MORSE_ERR_ILLEGAL_VALUE;
- }
+ MORSE_Desc_Flush( A, sequence );
+ MORSE_Desc_Flush( B, sequence );
+
+ morse_sequence_wait( morse, sequence );
- morse_pztile_to_lapack( A, Af77, LDA, sequence, request );
+ MORSE_Desc_Destroy( &B );
- return MORSE_SUCCESS;
+ status = sequence->status;
+ morse_sequence_destroy( morse, sequence );
+ return status;
}
diff --git a/compute/ztpgqrt.c b/compute/ztpgqrt.c
index bf6fee28d640f5e6345299766d1e983c74a5a7d3..d28a530995221f3bd0181658f6e23c836712fd1a 100644
--- a/compute/ztpgqrt.c
+++ b/compute/ztpgqrt.c
@@ -144,7 +144,10 @@ int MORSE_ztpgqrt( int M, int N, int K, int L,
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
MORSE_request_t request = MORSE_REQUEST_INITIALIZER;
- MORSE_desc_t descQ1, descQ2, descV1, descV2;
+ MORSE_desc_t descQ1l, descQ1t;
+ MORSE_desc_t descQ2l, descQ2t;
+ MORSE_desc_t descV1l, descV1t;
+ MORSE_desc_t descV2l, descV2t;
int minMK = chameleon_min( M, K );
morse = morse_context_self();
@@ -205,49 +208,41 @@ int MORSE_ztpgqrt( int M, int N, int K, int L,
/* Set NT */
NB = MORSE_NB;
- morse_sequence_create(morse, &sequence);
+ morse_sequence_create( morse, &sequence );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooplap2tile( descV1, V1, NB, NB, LDV1, K, 0, 0, M, K, sequence, &request,
- morse_desc_mat_free(&(descV1)) );
- morse_zooplap2tile( descV2, V2, NB, NB, LDV2, K, 0, 0, M, K, sequence, &request,
- (morse_desc_mat_free(&(descV1)),
- morse_desc_mat_free(&(descV2))) );
- morse_zooplap2tile( descQ1, Q1, NB, NB, LDQ1, N, 0, 0, K, N, sequence, &request,
- (morse_desc_mat_free(&(descV1)),
- morse_desc_mat_free(&(descV2)),
- morse_desc_mat_free(&(descQ1))) );
- morse_zooplap2tile( descQ2, Q2, NB, NB, LDQ2, N, 0, 0, M, N, sequence, &request,
- (morse_desc_mat_free(&(descV1)),
- morse_desc_mat_free(&(descV2)),
- morse_desc_mat_free(&(descQ1)),
- morse_desc_mat_free(&(descQ2))) );
-/* } else {*/
-/* morse_ziplap2tile( descQ1, Q1, NB, NB, LDQ1, N, 0, 0, M, N,*/
-/* sequence, &request);*/
-/* }*/
+ /* Submit the matrix conversion */
+ morse_zlap2tile( morse, &descV1l, &descV1t, MorseDescInput, MorseUpperLower,
+ V1, NB, NB, LDV1, K, M, K, sequence, &request );
+ morse_zlap2tile( morse, &descV2l, &descV2t, MorseDescInput, MorseUpperLower,
+ V2, NB, NB, LDV2, K, M, K, sequence, &request );
+ morse_zlap2tile( morse, &descQ1l, &descQ1t, MorseDescInout, MorseUpperLower,
+ Q1, NB, NB, LDQ1, N, K, N, sequence, &request );
+ morse_zlap2tile( morse, &descQ2l, &descQ2t, MorseDescInout, MorseUpperLower,
+ Q2, NB, NB, LDQ2, N, M, N, sequence, &request );
/* Call the tile interface */
- MORSE_ztpgqrt_Tile_Async(L, &descV1, descT1, &descV2, descT2, &descQ1, &descQ2, sequence, &request);
+ MORSE_ztpgqrt_Tile_Async( L, &descV1t, descT1, &descV2t, descT2, &descQ1t, &descQ2t, sequence, &request );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooptile2lap(descQ1, Q1, NB, NB, LDQ1, N, sequence, &request);
- morse_zooptile2lap(descQ2, Q2, NB, NB, LDQ2, N, sequence, &request);
- morse_sequence_wait(morse, sequence);
- morse_desc_mat_free(&descV1);
- morse_desc_mat_free(&descV2);
- morse_desc_mat_free(&descQ1);
- morse_desc_mat_free(&descQ2);
-/* } else {*/
-/* morse_ziptile2lap( descV1, V1, NB, NB, LDV1, K, sequence, &request);*/
-/* morse_ziptile2lap( descV2, V2, NB, NB, LDV2, K, sequence, &request);*/
-/* morse_ziptile2lap( descQ1, Q1, NB, NB, LDQ1, N, sequence, &request);*/
-/* morse_ziptile2lap( descQ2, Q2, NB, NB, LDQ2, N, sequence, &request);*/
-/* morse_sequence_wait(morse, sequence);*/
-/* }*/
+ /* Submit the matrix conversion back */
+ morse_ztile2lap( morse, &descV1l, &descV1t,
+ MorseDescInput, MorseUpperLower, sequence, &request );
+ morse_ztile2lap( morse, &descV2l, &descV2t,
+ MorseDescInput, MorseUpperLower, sequence, &request );
+ morse_ztile2lap( morse, &descQ1l, &descQ1t,
+ MorseDescInout, MorseUpperLower, sequence, &request );
+ morse_ztile2lap( morse, &descQ2l, &descQ2t,
+ MorseDescInout, MorseUpperLower, sequence, &request );
+
+ morse_sequence_wait( morse, sequence );
+
+ /* Cleanup the temporary data */
+ morse_ztile2lap_cleanup( morse, &descV1l, &descV1t );
+ morse_ztile2lap_cleanup( morse, &descV2l, &descV2t );
+ morse_ztile2lap_cleanup( morse, &descQ1l, &descQ1t );
+ morse_ztile2lap_cleanup( morse, &descQ2l, &descQ2t );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -294,14 +289,20 @@ int MORSE_ztpgqrt_Tile( int L,
morse_fatal_error("MORSE_ztpgqrt_Tile", "MORSE not initialized");
return MORSE_ERR_NOT_INITIALIZED;
}
- morse_sequence_create(morse, &sequence);
- MORSE_ztpgqrt_Tile_Async(L, V1, T1, V2, T2, Q1, Q2, sequence, &request);
- RUNTIME_desc_flush( Q1, sequence );
- RUNTIME_desc_flush( Q2, sequence );
- morse_sequence_wait(morse, sequence);
+ morse_sequence_create( morse, &sequence );
+
+ MORSE_ztpgqrt_Tile_Async( L, V1, T1, V2, T2, Q1, Q2, sequence, &request );
+ MORSE_Desc_Flush( V1, sequence );
+ MORSE_Desc_Flush( T1, sequence );
+ MORSE_Desc_Flush( V2, sequence );
+ MORSE_Desc_Flush( T2, sequence );
+ MORSE_Desc_Flush( Q1, sequence );
+ MORSE_Desc_Flush( Q2, sequence );
+
+ morse_sequence_wait( morse, sequence );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -357,10 +358,12 @@ int MORSE_ztpgqrt_Tile_Async( int L,
return MORSE_ERR_UNALLOCATED;
}
/* Check sequence status */
- if (sequence->status == MORSE_SUCCESS)
+ if (sequence->status == MORSE_SUCCESS) {
request->status = MORSE_SUCCESS;
- else
+ }
+ else {
return morse_request_fail(sequence, request, MORSE_ERR_SEQUENCE_FLUSHED);
+ }
/* Check descriptors for correctness */
if (morse_desc_check(V1) != MORSE_SUCCESS) {
@@ -400,7 +403,7 @@ int MORSE_ztpgqrt_Tile_Async( int L,
{
int minMT;
if (V1->m > V1->n) {
- minMT = V1->nt;
+ minMT = V1->nt;
} else {
minMT = V1->mt;
}
@@ -415,10 +418,18 @@ int MORSE_ztpgqrt_Tile_Async( int L,
morse_pztpgqrt( L, V1, T1, V2, T2, Q1, Q2, Dptr, sequence, request );
/* } */
/* else { */
- /* morse_pztpgqrtrh(Q1, T, MORSE_RHBLK, sequence, request); */
+ /* morse_pztpgqrtrh( Q1, T, MORSE_RHBLK, sequence, request ); */
/* } */
if (Dptr != NULL) {
- morse_desc_mat_free(Dptr);
+ MORSE_Desc_Flush( V1, sequence );
+ MORSE_Desc_Flush( T1, sequence );
+ MORSE_Desc_Flush( V2, sequence );
+ MORSE_Desc_Flush( T2, sequence );
+ MORSE_Desc_Flush( Q1, sequence );
+ MORSE_Desc_Flush( Q2, sequence );
+ MORSE_Desc_Flush( Dptr, sequence );
+ morse_sequence_wait( morse, sequence );
+ morse_desc_mat_free( Dptr );
}
(void)D;
return MORSE_SUCCESS;
diff --git a/compute/ztpqrt.c b/compute/ztpqrt.c
index 3298638e02705a80365fc645a57c1c906e76c021..9f5f60bbffdeac37cbad20448128fae9a053466d 100644
--- a/compute/ztpqrt.c
+++ b/compute/ztpqrt.c
@@ -138,7 +138,8 @@ int MORSE_ztpqrt( int M, int N, int L,
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
MORSE_request_t request = MORSE_REQUEST_INITIALIZER;
- MORSE_desc_t descA, descB;
+ MORSE_desc_t descAl, descAt;
+ MORSE_desc_t descBl, descBt;
int minMN = chameleon_min( M, N );
morse = morse_context_self();
@@ -183,35 +184,31 @@ int MORSE_ztpqrt( int M, int N, int L,
/* Set NT */
NB = MORSE_NB;
- morse_sequence_create(morse, &sequence);
+ morse_sequence_create( morse, &sequence );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooplap2tile( descA, A, NB, NB, LDA, N, 0, 0, N, N, sequence, &request,
- morse_desc_mat_free(&(descA)) );
- morse_zooplap2tile( descB, B, NB, NB, LDB, N, 0, 0, M, N, sequence, &request,
- (morse_desc_mat_free(&(descA)), morse_desc_mat_free(&(descB))) );
-/* } else {*/
-/* morse_ziplap2tile( descA, A, NB, NB, LDA, N, 0, 0, M, N,*/
-/* sequence, &request);*/
-/* }*/
+ /* Submit the matrix conversion */
+ morse_zlap2tile( morse, &descAl, &descAt, MorseDescInout, MorseUpper,
+ A, NB, NB, LDA, N, N, N, sequence, &request );
+ morse_zlap2tile( morse, &descBl, &descBt, MorseDescInout, MorseUpperLower,
+ B, NB, NB, LDB, N, M, N, sequence, &request );
/* Call the tile interface */
- MORSE_ztpqrt_Tile_Async(L, &descA, &descB, descT, sequence, &request);
+ MORSE_ztpqrt_Tile_Async( L, &descAt, &descBt, descT, sequence, &request );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooptile2lap(descA, A, NB, NB, LDA, N, sequence, &request);
- morse_zooptile2lap(descB, B, NB, NB, LDB, N, sequence, &request);
- morse_sequence_wait(morse, sequence);
- morse_desc_mat_free(&descA);
- morse_desc_mat_free(&descB);
-/* } else {*/
-/* morse_ziptile2lap( descA, A, NB, NB, LDA, N, sequence, &request);*/
-/* morse_ziptile2lap( descB, B, NB, NB, LDB, N, sequence, &request);*/
-/* morse_sequence_wait(morse, sequence);*/
-/* }*/
+ /* Submit the matrix conversion back */
+ morse_ztile2lap( morse, &descAl, &descAt,
+ MorseDescInout, MorseUpper, sequence, &request );
+ morse_ztile2lap( morse, &descBl, &descBt,
+ MorseDescInout, MorseUpperLower, sequence, &request );
+
+ morse_sequence_wait( morse, sequence );
+
+ /* Cleanup the temporary data */
+ morse_ztile2lap_cleanup( morse, &descAl, &descAt );
+ morse_ztile2lap_cleanup( morse, &descBl, &descBt );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -266,14 +263,17 @@ int MORSE_ztpqrt_Tile( int L, MORSE_desc_t *A, MORSE_desc_t *B, MORSE_desc_t *T
morse_fatal_error("MORSE_ztpqrt_Tile", "MORSE not initialized");
return MORSE_ERR_NOT_INITIALIZED;
}
- morse_sequence_create(morse, &sequence);
- MORSE_ztpqrt_Tile_Async(L, A, B, T, sequence, &request);
- RUNTIME_desc_flush( A, sequence );
- RUNTIME_desc_flush( B, sequence );
- morse_sequence_wait(morse, sequence);
+ morse_sequence_create( morse, &sequence );
+
+ MORSE_ztpqrt_Tile_Async( L, A, B, T, sequence, &request );
+ MORSE_Desc_Flush( A, sequence );
+ MORSE_Desc_Flush( B, sequence );
+ MORSE_Desc_Flush( T, sequence );
+
+ morse_sequence_wait( morse, sequence );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -325,10 +325,12 @@ int MORSE_ztpqrt_Tile_Async( int L, MORSE_desc_t *A, MORSE_desc_t *B, MORSE_desc
return MORSE_ERR_UNALLOCATED;
}
/* Check sequence status */
- if (sequence->status == MORSE_SUCCESS)
+ if (sequence->status == MORSE_SUCCESS) {
request->status = MORSE_SUCCESS;
- else
+ }
+ else {
return morse_request_fail(sequence, request, MORSE_ERR_SEQUENCE_FLUSHED);
+ }
/* Check descriptors for correctness */
if (morse_desc_check(A) != MORSE_SUCCESS) {
@@ -354,10 +356,10 @@ int MORSE_ztpqrt_Tile_Async( int L, MORSE_desc_t *A, MORSE_desc_t *B, MORSE_desc
}
/* if (morse->householder == MORSE_FLAT_HOUSEHOLDER) { */
- morse_pztpqrt(L, A, B, T, sequence, request);
+ morse_pztpqrt( L, A, B, T, sequence, request );
/* } */
/* else { */
- /* morse_pztpqrtrh(A, T, MORSE_RHBLK, sequence, request); */
+ /* morse_pztpqrtrh( A, T, MORSE_RHBLK, sequence, request ); */
/* } */
return MORSE_SUCCESS;
diff --git a/compute/ztradd.c b/compute/ztradd.c
index 7bbf05f99cc7828f7c8f7201cf1d63258c5a605d..bd88723f3df62ac702ab43c91ee3899661725b9c 100644
--- a/compute/ztradd.c
+++ b/compute/ztradd.c
@@ -97,14 +97,15 @@
* @sa MORSE_stradd
*
******************************************************************************/
-int MORSE_ztradd(MORSE_enum uplo, MORSE_enum trans, int M, int N,
- MORSE_Complex64_t alpha, MORSE_Complex64_t *A, int LDA,
- MORSE_Complex64_t beta, MORSE_Complex64_t *B, int LDB)
+int MORSE_ztradd( MORSE_enum uplo, MORSE_enum trans, int M, int N,
+ MORSE_Complex64_t alpha, MORSE_Complex64_t *A, int LDA,
+ MORSE_Complex64_t beta, MORSE_Complex64_t *B, int LDB )
{
int NB;
int Am, An;
int status;
- MORSE_desc_t descA, descB;
+ MORSE_desc_t descAl, descAt;
+ MORSE_desc_t descBl, descBt;
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
MORSE_request_t request = MORSE_REQUEST_INITIALIZER;
@@ -161,37 +162,31 @@ int MORSE_ztradd(MORSE_enum uplo, MORSE_enum trans, int M, int N,
/* Set MT & NT & KT */
NB = MORSE_NB;
- morse_sequence_create(morse, &sequence);
+ morse_sequence_create( morse, &sequence );
- /* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) { */
- morse_zooplap2tile( descA, A, NB, NB, LDA, An, 0, 0, Am, An, sequence, &request,
- morse_desc_mat_free(&(descA)) );
- morse_zooplap2tile( descB, B, NB, NB, LDB, N, 0, 0, M, N, sequence, &request,
- morse_desc_mat_free(&(descA)); morse_desc_mat_free(&(descB)));
- /* } else { */
- /* morse_ziplap2tile( descA, A, NB, NB, LDA, An, 0, 0, Am, An, */
- /* sequence, &request); */
- /* morse_ziplap2tile( descB, B, NB, NB, LDB, N, 0, 0, M, N, */
- /* sequence, &request); */
- /* } */
+ /* Submit the matrix conversion */
+ morse_zlap2tile( morse, &descAl, &descAt, MorseDescInput, uplo,
+ A, NB, NB, LDA, An, Am, An, sequence, &request );
+ morse_zlap2tile( morse, &descBl, &descBt, MorseDescInout, uplo,
+ B, NB, NB, LDB, N, M, N, sequence, &request );
/* Call the tile interface */
- MORSE_ztradd_Tile_Async(
- uplo, trans, alpha, &descA, beta, &descB, sequence, &request);
+ MORSE_ztradd_Tile_Async( uplo, trans, alpha, &descAt, beta, &descBt, sequence, &request );
- /* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) { */
- morse_zooptile2lap( descB, B, NB, NB, LDB, N, sequence, &request);
- morse_sequence_wait(morse, sequence);
- morse_desc_mat_free(&descA);
- morse_desc_mat_free(&descB);
- /* } else { */
- /* morse_ziptile2lap( descA, A, NB, NB, LDA, An, sequence, &request); */
- /* morse_ziptile2lap( descB, B, NB, NB, LDB, N, sequence, &request); */
- /* morse_dynamic_sync(); */
- /* } */
+ /* Submit the matrix conversion back */
+ morse_ztile2lap( morse, &descAl, &descAt,
+ MorseDescInput, uplo, sequence, &request );
+ morse_ztile2lap( morse, &descBl, &descBt,
+ MorseDescInout, uplo, sequence, &request );
+
+ morse_sequence_wait( morse, sequence );
+
+ /* Cleanup the temporary data */
+ morse_ztile2lap_cleanup( morse, &descAl, &descAt );
+ morse_ztile2lap_cleanup( morse, &descBl, &descBt );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -249,9 +244,9 @@ int MORSE_ztradd(MORSE_enum uplo, MORSE_enum trans, int M, int N,
* @sa MORSE_stradd_Tile
*
******************************************************************************/
-int MORSE_ztradd_Tile(MORSE_enum uplo, MORSE_enum trans,
+int MORSE_ztradd_Tile( MORSE_enum uplo, MORSE_enum trans,
MORSE_Complex64_t alpha, MORSE_desc_t *A,
- MORSE_Complex64_t beta, MORSE_desc_t *B)
+ MORSE_Complex64_t beta, MORSE_desc_t *B )
{
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
@@ -263,14 +258,16 @@ int MORSE_ztradd_Tile(MORSE_enum uplo, MORSE_enum trans,
morse_fatal_error("MORSE_ztradd_Tile", "MORSE not initialized");
return MORSE_ERR_NOT_INITIALIZED;
}
- morse_sequence_create(morse, &sequence);
- MORSE_ztradd_Tile_Async(uplo, trans, alpha, A, beta, B, sequence, &request);
- RUNTIME_desc_flush( A, sequence );
- RUNTIME_desc_flush( B, sequence );
- morse_sequence_wait(morse, sequence);
+ morse_sequence_create( morse, &sequence );
+
+ MORSE_ztradd_Tile_Async( uplo, trans, alpha, A, beta, B, sequence, &request );
+ MORSE_Desc_Flush( A, sequence );
+ MORSE_Desc_Flush( B, sequence );
+
+ morse_sequence_wait( morse, sequence );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -303,10 +300,10 @@ int MORSE_ztradd_Tile(MORSE_enum uplo, MORSE_enum trans,
* @sa MORSE_stradd_Tile_Async
*
******************************************************************************/
-int MORSE_ztradd_Tile_Async(MORSE_enum uplo, MORSE_enum trans,
- MORSE_Complex64_t alpha, MORSE_desc_t *A,
- MORSE_Complex64_t beta, MORSE_desc_t *B,
- MORSE_sequence_t *sequence, MORSE_request_t *request)
+int MORSE_ztradd_Tile_Async( MORSE_enum uplo, MORSE_enum trans,
+ MORSE_Complex64_t alpha, MORSE_desc_t *A,
+ MORSE_Complex64_t beta, MORSE_desc_t *B,
+ MORSE_sequence_t *sequence, MORSE_request_t *request )
{
MORSE_context_t *morse;
int M, N;
@@ -326,10 +323,12 @@ int MORSE_ztradd_Tile_Async(MORSE_enum uplo, MORSE_enum trans,
return MORSE_ERR_UNALLOCATED;
}
/* Check sequence status */
- if (sequence->status == MORSE_SUCCESS)
+ if (sequence->status == MORSE_SUCCESS) {
request->status = MORSE_SUCCESS;
- else
+ }
+ else {
return morse_request_fail(sequence, request, MORSE_ERR_SEQUENCE_FLUSHED);
+ }
/* Check descriptors for correctness */
if (morse_desc_check(A) != MORSE_SUCCESS) {
@@ -383,7 +382,7 @@ int MORSE_ztradd_Tile_Async(MORSE_enum uplo, MORSE_enum trans,
((alpha == (MORSE_Complex64_t)0.0) && beta == (MORSE_Complex64_t)1.0))
return MORSE_SUCCESS;
- morse_pztradd(uplo, trans, alpha, A, beta, B, sequence, request);
+ morse_pztradd( uplo, trans, alpha, A, beta, B, sequence, request );
return MORSE_SUCCESS;
}
diff --git a/compute/ztrmm.c b/compute/ztrmm.c
index 434621d6da5963f885d2e35142b9d1db28b8f22f..3496308a72a917e781343b596c7836c4fe9a4d59 100644
--- a/compute/ztrmm.c
+++ b/compute/ztrmm.c
@@ -100,18 +100,19 @@
* @sa MORSE_strmm
*
******************************************************************************/
-int MORSE_ztrmm(MORSE_enum side, MORSE_enum uplo,
+int MORSE_ztrmm( MORSE_enum side, MORSE_enum uplo,
MORSE_enum transA, MORSE_enum diag,
int N, int NRHS, MORSE_Complex64_t alpha,
MORSE_Complex64_t *A, int LDA,
- MORSE_Complex64_t *B, int LDB)
+ MORSE_Complex64_t *B, int LDB )
{
int NB, NA;
int status;
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
MORSE_request_t request = MORSE_REQUEST_INITIALIZER;
- MORSE_desc_t descA, descB;
+ MORSE_desc_t descAl, descAt;
+ MORSE_desc_t descBl, descBt;
morse = morse_context_self();
if (morse == NULL) {
@@ -120,9 +121,9 @@ int MORSE_ztrmm(MORSE_enum side, MORSE_enum uplo,
}
if (side == MorseLeft) {
- NA = N;
+ NA = N;
} else {
- NA = NRHS;
+ NA = NRHS;
}
/* Check input arguments */
@@ -130,7 +131,7 @@ int MORSE_ztrmm(MORSE_enum side, MORSE_enum uplo,
morse_error("MORSE_ztrmm", "illegal value of side");
return -1;
}
- if (uplo != MorseUpper && uplo != MorseLower) {
+ if ((uplo != MorseUpper) && (uplo != MorseLower)) {
morse_error("MORSE_ztrmm", "illegal value of uplo");
return -2;
}
@@ -141,7 +142,7 @@ int MORSE_ztrmm(MORSE_enum side, MORSE_enum uplo,
morse_error("MORSE_ztrmm", "illegal value of transA");
return -3;
}
- if (diag != MorseUnit && diag != MorseNonUnit) {
+ if ((diag != MorseUnit) && (diag != MorseNonUnit)) {
morse_error("MORSE_ztrmm", "illegal value of diag");
return -4;
}
@@ -175,37 +176,31 @@ int MORSE_ztrmm(MORSE_enum side, MORSE_enum uplo,
/* Set NT & NTRHS */
NB = MORSE_NB;
- morse_sequence_create(morse, &sequence);
+ morse_sequence_create( morse, &sequence );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooplap2tile( descA, A, NB, NB, LDA, NA, 0, 0, NA, NA, sequence, &request,
- morse_desc_mat_free(&(descA)) );
- morse_zooplap2tile( descB, B, NB, NB, LDB, NRHS, 0, 0, N, NRHS, sequence, &request,
- morse_desc_mat_free(&(descA)); morse_desc_mat_free(&(descB)));
-/* } else {*/
-/* morse_ziplap2tile( descA, A, NB, NB, LDA, NA, 0, 0, NA, NA, */
-/* sequence, &request);*/
-/* morse_ziplap2tile( descB, B, NB, NB, LDB, NRHS, 0, 0, N, NRHS,*/
-/* sequence, &request);*/
-/* }*/
+ /* Submit the matrix conversion */
+ morse_zlap2tile( morse, &descAl, &descAt, MorseDescInput, uplo,
+ A, NB, NB, LDA, NA, NA, NA, sequence, &request );
+ morse_zlap2tile( morse, &descBl, &descBt, MorseDescInout, MorseUpperLower,
+ B, NB, NB, LDB, NRHS, N, NRHS, sequence, &request );
/* Call the tile interface */
- MORSE_ztrmm_Tile_Async(
- side, uplo, transA, diag, alpha, &descA, &descB, sequence, &request);
+ MORSE_ztrmm_Tile_Async( side, uplo, transA, diag, alpha, &descAt, &descBt, sequence, &request );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooptile2lap(descB, B, NB, NB, LDB, NRHS, sequence, &request);
- morse_sequence_wait(morse, sequence);
- morse_desc_mat_free(&descA);
- morse_desc_mat_free(&descB);
-/* } else {*/
-/* morse_ziptile2lap( descA, A, NB, NB, LDA, NA, sequence, &request);*/
-/* morse_ziptile2lap( descB, B, NB, NB, LDB, NRHS, sequence, &request);*/
-/* morse_sequence_wait(morse, sequence);*/
-/* }*/
+ /* Submit the matrix conversion back */
+ morse_ztile2lap( morse, &descAl, &descAt,
+ MorseDescInput, uplo, sequence, &request );
+ morse_ztile2lap( morse, &descBl, &descBt,
+ MorseDescInout, MorseUpperLower, sequence, &request );
+
+ morse_sequence_wait( morse, sequence );
+
+ /* Cleanup the temporary data */
+ morse_ztile2lap_cleanup( morse, &descAl, &descAt );
+ morse_ztile2lap_cleanup( morse, &descBl, &descBt );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
/**
@@ -271,9 +266,9 @@ int MORSE_ztrmm(MORSE_enum side, MORSE_enum uplo,
* @sa MORSE_strmm_Tile
*
******************************************************************************/
-int MORSE_ztrmm_Tile(MORSE_enum side, MORSE_enum uplo,
+int MORSE_ztrmm_Tile( MORSE_enum side, MORSE_enum uplo,
MORSE_enum transA, MORSE_enum diag,
- MORSE_Complex64_t alpha, MORSE_desc_t *A, MORSE_desc_t *B)
+ MORSE_Complex64_t alpha, MORSE_desc_t *A, MORSE_desc_t *B )
{
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
@@ -285,14 +280,16 @@ int MORSE_ztrmm_Tile(MORSE_enum side, MORSE_enum uplo,
morse_fatal_error("MORSE_ztrmm_Tile", "MORSE not initialized");
return MORSE_ERR_NOT_INITIALIZED;
}
- morse_sequence_create(morse, &sequence);
- MORSE_ztrmm_Tile_Async(side, uplo, transA, diag, alpha, A, B, sequence, &request);
- RUNTIME_desc_flush( A, sequence );
- RUNTIME_desc_flush( B, sequence );
- morse_sequence_wait(morse, sequence);
+ morse_sequence_create( morse, &sequence );
+
+ MORSE_ztrmm_Tile_Async(side, uplo, transA, diag, alpha, A, B, sequence, &request );
+ MORSE_Desc_Flush( A, sequence );
+ MORSE_Desc_Flush( B, sequence );
+
+ morse_sequence_wait( morse, sequence );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -324,10 +321,10 @@ int MORSE_ztrmm_Tile(MORSE_enum side, MORSE_enum uplo,
* @sa MORSE_strmm_Tile_Async
*
******************************************************************************/
-int MORSE_ztrmm_Tile_Async(MORSE_enum side, MORSE_enum uplo,
+int MORSE_ztrmm_Tile_Async( MORSE_enum side, MORSE_enum uplo,
MORSE_enum transA, MORSE_enum diag,
MORSE_Complex64_t alpha, MORSE_desc_t *A, MORSE_desc_t *B,
- MORSE_sequence_t *sequence, MORSE_request_t *request)
+ MORSE_sequence_t *sequence, MORSE_request_t *request )
{
MORSE_context_t *morse;
@@ -345,10 +342,12 @@ int MORSE_ztrmm_Tile_Async(MORSE_enum side, MORSE_enum uplo,
return MORSE_ERR_UNALLOCATED;
}
/* Check sequence status */
- if (sequence->status == MORSE_SUCCESS)
+ if (sequence->status == MORSE_SUCCESS) {
request->status = MORSE_SUCCESS;
- else
+ }
+ else {
return morse_request_fail(sequence, request, MORSE_ERR_SEQUENCE_FLUSHED);
+ }
/* Check descriptors for correctness */
if (morse_desc_check(A) != MORSE_SUCCESS) {
@@ -368,7 +367,7 @@ int MORSE_ztrmm_Tile_Async(MORSE_enum side, MORSE_enum uplo,
morse_error("MORSE_ztrmm_Tile", "illegal value of side");
return morse_request_fail(sequence, request, -1);
}
- if (uplo != MorseUpper && uplo != MorseLower) {
+ if ((uplo != MorseUpper) && (uplo != MorseLower)) {
morse_error("MORSE_ztrmm_Tile", "illegal value of uplo");
return morse_request_fail(sequence, request, -2);
}
@@ -376,13 +375,13 @@ int MORSE_ztrmm_Tile_Async(MORSE_enum side, MORSE_enum uplo,
morse_error("MORSE_ztrmm_Tile", "illegal value of transA");
return morse_request_fail(sequence, request, -3);
}
- if (diag != MorseUnit && diag != MorseNonUnit) {
+ if ((diag != MorseUnit) && (diag != MorseNonUnit)) {
morse_error("MORSE_ztrmm_Tile", "illegal value of diag");
return morse_request_fail(sequence, request, -4);
}
/* Quick return */
- morse_pztrmm(side, uplo, transA, diag, alpha, A, B, sequence, request);
+ morse_pztrmm( side, uplo, transA, diag, alpha, A, B, sequence, request );
return MORSE_SUCCESS;
}
diff --git a/compute/ztrsm.c b/compute/ztrsm.c
index db4e89dec01cd316b7933dd5faf1c798923cf1e1..053be9e3e2774c022818fbd7fdc6c5c783258285 100644
--- a/compute/ztrsm.c
+++ b/compute/ztrsm.c
@@ -101,18 +101,19 @@
* @sa MORSE_strsm
*
******************************************************************************/
-int MORSE_ztrsm(MORSE_enum side, MORSE_enum uplo,
+int MORSE_ztrsm( MORSE_enum side, MORSE_enum uplo,
MORSE_enum transA, MORSE_enum diag,
int N, int NRHS, MORSE_Complex64_t alpha,
MORSE_Complex64_t *A, int LDA,
- MORSE_Complex64_t *B, int LDB)
+ MORSE_Complex64_t *B, int LDB )
{
int NB, NA;
int status;
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
MORSE_request_t request = MORSE_REQUEST_INITIALIZER;
- MORSE_desc_t descA, descB;
+ MORSE_desc_t descAl, descAt;
+ MORSE_desc_t descBl, descBt;
morse = morse_context_self();
if (morse == NULL) {
@@ -121,9 +122,9 @@ int MORSE_ztrsm(MORSE_enum side, MORSE_enum uplo,
}
if (side == MorseLeft) {
- NA = N;
+ NA = N;
} else {
- NA = NRHS;
+ NA = NRHS;
}
/* Check input arguments */
@@ -131,7 +132,7 @@ int MORSE_ztrsm(MORSE_enum side, MORSE_enum uplo,
morse_error("MORSE_ztrsm", "illegal value of side");
return -1;
}
- if (uplo != MorseUpper && uplo != MorseLower) {
+ if ((uplo != MorseUpper) && (uplo != MorseLower)) {
morse_error("MORSE_ztrsm", "illegal value of uplo");
return -2;
}
@@ -139,7 +140,7 @@ int MORSE_ztrsm(MORSE_enum side, MORSE_enum uplo,
morse_error("MORSE_ztrsm", "illegal value of transA");
return -3;
}
- if (diag != MorseUnit && diag != MorseNonUnit) {
+ if ((diag != MorseUnit) && (diag != MorseNonUnit)) {
morse_error("MORSE_ztrsm", "illegal value of diag");
return -4;
}
@@ -173,37 +174,31 @@ int MORSE_ztrsm(MORSE_enum side, MORSE_enum uplo,
/* Set NT & NTRHS */
NB = MORSE_NB;
- morse_sequence_create(morse, &sequence);
+ morse_sequence_create( morse, &sequence );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooplap2tile( descA, A, NB, NB, LDA, NA, 0, 0, NA, NA, sequence, &request,
- morse_desc_mat_free(&(descA)) );
- morse_zooplap2tile( descB, B, NB, NB, LDB, NRHS, 0, 0, N, NRHS, sequence, &request,
- morse_desc_mat_free(&(descA)); morse_desc_mat_free(&(descB)));
-/* } else {*/
-/* morse_ziplap2tile( descA, A, NB, NB, LDA, NA, 0, 0, NA, NA, */
-/* sequence, &request);*/
-/* morse_ziplap2tile( descB, B, NB, NB, LDB, NRHS, 0, 0, N, NRHS,*/
-/* sequence, &request);*/
-/* }*/
+ /* Submit the matrix conversion */
+ morse_zlap2tile( morse, &descAl, &descAt, MorseDescInput, uplo,
+ A, NB, NB, LDA, NA, NA, NA, sequence, &request );
+ morse_zlap2tile( morse, &descBl, &descBt, MorseDescInout, MorseUpperLower,
+ B, NB, NB, LDB, NRHS, N, NRHS, sequence, &request );
/* Call the tile interface */
- MORSE_ztrsm_Tile_Async(
- side, uplo, transA, diag, alpha, &descA, &descB, sequence, &request);
+ MORSE_ztrsm_Tile_Async( side, uplo, transA, diag, alpha, &descAt, &descBt, sequence, &request );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooptile2lap(descB, B, NB, NB, LDB, NRHS, sequence, &request);
- morse_sequence_wait(morse, sequence);
- morse_desc_mat_free(&descA);
- morse_desc_mat_free(&descB);
-/* } else {*/
-/* morse_ziptile2lap( descA, A, NB, NB, LDA, NA, sequence, &request);*/
-/* morse_ziptile2lap( descB, B, NB, NB, LDB, NRHS, sequence, &request);*/
-/* morse_sequence_wait(morse, sequence);*/
-/* }*/
+ /* Submit the matrix conversion back */
+ morse_ztile2lap( morse, &descAl, &descAt,
+ MorseDescInput, uplo, sequence, &request );
+ morse_ztile2lap( morse, &descBl, &descBt,
+ MorseDescInout, MorseUpperLower, sequence, &request );
+
+ morse_sequence_wait( morse, sequence );
+
+ /* Cleanup the temporary data */
+ morse_ztile2lap_cleanup( morse, &descAl, &descAt );
+ morse_ztile2lap_cleanup( morse, &descBl, &descBt );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
/**
@@ -269,9 +264,9 @@ int MORSE_ztrsm(MORSE_enum side, MORSE_enum uplo,
* @sa MORSE_strsm_Tile
*
******************************************************************************/
-int MORSE_ztrsm_Tile(MORSE_enum side, MORSE_enum uplo,
+int MORSE_ztrsm_Tile( MORSE_enum side, MORSE_enum uplo,
MORSE_enum transA, MORSE_enum diag,
- MORSE_Complex64_t alpha, MORSE_desc_t *A, MORSE_desc_t *B)
+ MORSE_Complex64_t alpha, MORSE_desc_t *A, MORSE_desc_t *B )
{
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
@@ -283,14 +278,16 @@ int MORSE_ztrsm_Tile(MORSE_enum side, MORSE_enum uplo,
morse_fatal_error("MORSE_ztrsm_Tile", "MORSE not initialized");
return MORSE_ERR_NOT_INITIALIZED;
}
- morse_sequence_create(morse, &sequence);
- MORSE_ztrsm_Tile_Async(side, uplo, transA, diag, alpha, A, B, sequence, &request);
- RUNTIME_desc_flush( A, sequence );
- RUNTIME_desc_flush( B, sequence );
- morse_sequence_wait(morse, sequence);
-
+ morse_sequence_create( morse, &sequence );
+
+ MORSE_ztrsm_Tile_Async(side, uplo, transA, diag, alpha, A, B, sequence, &request );
+
+ MORSE_Desc_Flush( A, sequence );
+ MORSE_Desc_Flush( B, sequence );
+
+ morse_sequence_wait( morse, sequence );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -322,10 +319,10 @@ int MORSE_ztrsm_Tile(MORSE_enum side, MORSE_enum uplo,
* @sa MORSE_strsm_Tile_Async
*
******************************************************************************/
-int MORSE_ztrsm_Tile_Async(MORSE_enum side, MORSE_enum uplo,
+int MORSE_ztrsm_Tile_Async( MORSE_enum side, MORSE_enum uplo,
MORSE_enum transA, MORSE_enum diag,
MORSE_Complex64_t alpha, MORSE_desc_t *A, MORSE_desc_t *B,
- MORSE_sequence_t *sequence, MORSE_request_t *request)
+ MORSE_sequence_t *sequence, MORSE_request_t *request )
{
MORSE_context_t *morse;
@@ -343,10 +340,12 @@ int MORSE_ztrsm_Tile_Async(MORSE_enum side, MORSE_enum uplo,
return MORSE_ERR_UNALLOCATED;
}
/* Check sequence status */
- if (sequence->status == MORSE_SUCCESS)
+ if (sequence->status == MORSE_SUCCESS) {
request->status = MORSE_SUCCESS;
- else
+ }
+ else {
return morse_request_fail(sequence, request, MORSE_ERR_SEQUENCE_FLUSHED);
+ }
/* Check descriptors for correctness */
if (morse_desc_check(A) != MORSE_SUCCESS) {
@@ -366,7 +365,7 @@ int MORSE_ztrsm_Tile_Async(MORSE_enum side, MORSE_enum uplo,
morse_error("MORSE_ztrsm_Tile", "illegal value of side");
return morse_request_fail(sequence, request, -1);
}
- if (uplo != MorseUpper && uplo != MorseLower) {
+ if ((uplo != MorseUpper) && (uplo != MorseLower)) {
morse_error("MORSE_ztrsm_Tile", "illegal value of uplo");
return morse_request_fail(sequence, request, -2);
}
@@ -374,13 +373,13 @@ int MORSE_ztrsm_Tile_Async(MORSE_enum side, MORSE_enum uplo,
morse_error("MORSE_ztrsm_Tile", "illegal value of transA");
return morse_request_fail(sequence, request, -3);
}
- if (diag != MorseUnit && diag != MorseNonUnit) {
+ if ((diag != MorseUnit) && (diag != MorseNonUnit)) {
morse_error("MORSE_ztrsm_Tile", "illegal value of diag");
return morse_request_fail(sequence, request, -4);
}
/* Quick return */
- morse_pztrsm(side, uplo, transA, diag, alpha, A, B, sequence, request);
+ morse_pztrsm( side, uplo, transA, diag, alpha, A, B, sequence, request );
return MORSE_SUCCESS;
}
diff --git a/compute/ztrsmpl.c b/compute/ztrsmpl.c
index 0439834108f356cdb5ed263047c219f069596656..d84670ad98dc9fcf87bcccfd63c73b3538380f2c 100644
--- a/compute/ztrsmpl.c
+++ b/compute/ztrsmpl.c
@@ -80,17 +80,18 @@
* @sa MORSE_zgetrf_incpiv
*
******************************************************************************/
-int MORSE_ztrsmpl(int N, int NRHS,
+int MORSE_ztrsmpl( int N, int NRHS,
MORSE_Complex64_t *A, int LDA,
MORSE_desc_t *descL, int *IPIV,
- MORSE_Complex64_t *B, int LDB)
+ MORSE_Complex64_t *B, int LDB )
{
int NB;
int status;
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
MORSE_request_t request = MORSE_REQUEST_INITIALIZER;
- MORSE_desc_t descA, descB;
+ MORSE_desc_t descAl, descAt;
+ MORSE_desc_t descBl, descBt;
morse = morse_context_self();
if (morse == NULL) {
@@ -128,36 +129,31 @@ int MORSE_ztrsmpl(int N, int NRHS,
/* Set Mt, NT & NTRHS */
NB = MORSE_NB;
- morse_sequence_create(morse, &sequence);
+ morse_sequence_create( morse, &sequence );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooplap2tile( descA, A, NB, NB, LDA, N, 0, 0, N, N, sequence, &request,
- morse_desc_mat_free(&(descA)) );
- morse_zooplap2tile( descB, B, NB, NB, LDB, NRHS, 0, 0, N, NRHS, sequence, &request,
- morse_desc_mat_free(&(descA)); morse_desc_mat_free(&(descB)));
-/* } else {*/
-/* morse_ziplap2tile( descA, A, NB, NB, LDA, N, 0, 0, N, N, */
-/* sequence, &request);*/
-/* morse_ziplap2tile( descB, B, NB, NB, LDB, NRHS, 0, 0, N, NRHS,*/
-/* sequence, &request);*/
-/* }*/
+ /* Submit the matrix conversion */
+ morse_zlap2tile( morse, &descAl, &descAt, MorseDescInput, MorseLower,
+ A, NB, NB, LDA, N, N, N, sequence, &request );
+ morse_zlap2tile( morse, &descBl, &descBt, MorseDescInout, MorseUpperLower,
+ B, NB, NB, LDB, NRHS, N, NRHS, sequence, &request );
/* Call the tile interface */
- MORSE_ztrsmpl_Tile_Async(&descA, descL, IPIV, &descB, sequence, &request);
+ MORSE_ztrsmpl_Tile_Async( &descAt, descL, IPIV, &descBt, sequence, &request );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooptile2lap(descB, B, NB, NB, LDB, NRHS, sequence, &request);
- morse_sequence_wait(morse, sequence);
- morse_desc_mat_free(&descA);
- morse_desc_mat_free(&descB);
-/* } else {*/
-/* morse_ziptile2lap( descA, A, NB, NB, LDA, N, sequence, &request);*/
-/* morse_ziptile2lap( descB, B, NB, NB, LDB, NRHS, sequence, &request);*/
-/* morse_sequence_wait(morse, sequence);*/
-/* }*/
+ /* Submit the matrix conversion back */
+ morse_ztile2lap( morse, &descAl, &descAt,
+ MorseDescInput, MorseLower, sequence, &request );
+ morse_ztile2lap( morse, &descBl, &descBt,
+ MorseDescInout, MorseUpperLower, sequence, &request );
+
+ morse_sequence_wait( morse, sequence );
+
+ /* Cleanup the temporary data */
+ morse_ztile2lap_cleanup( morse, &descAl, &descAt );
+ morse_ztile2lap_cleanup( morse, &descBl, &descBt );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -200,7 +196,7 @@ int MORSE_ztrsmpl(int N, int NRHS,
* @sa MORSE_zgetrf_incpiv_Tile
*
******************************************************************************/
-int MORSE_ztrsmpl_Tile(MORSE_desc_t *A, MORSE_desc_t *L, int *IPIV, MORSE_desc_t *B)
+int MORSE_ztrsmpl_Tile( MORSE_desc_t *A, MORSE_desc_t *L, int *IPIV, MORSE_desc_t *B )
{
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
@@ -212,14 +208,17 @@ int MORSE_ztrsmpl_Tile(MORSE_desc_t *A, MORSE_desc_t *L, int *IPIV, MORSE_desc_t
morse_fatal_error("MORSE_ztrsmpl_Tile", "MORSE not initialized");
return MORSE_ERR_NOT_INITIALIZED;
}
- morse_sequence_create(morse, &sequence);
- MORSE_ztrsmpl_Tile_Async(A, L, IPIV, B, sequence, &request);
- RUNTIME_desc_flush( A, sequence );
- RUNTIME_desc_flush( B, sequence );
- morse_sequence_wait(morse, sequence);
-
+ morse_sequence_create( morse, &sequence );
+
+ MORSE_ztrsmpl_Tile_Async( A, L, IPIV, B, sequence, &request );
+
+ MORSE_Desc_Flush( A, sequence );
+ MORSE_Desc_Flush( L, sequence );
+ MORSE_Desc_Flush( B, sequence );
+
+ morse_sequence_wait( morse, sequence );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -253,8 +252,8 @@ int MORSE_ztrsmpl_Tile(MORSE_desc_t *A, MORSE_desc_t *L, int *IPIV, MORSE_desc_t
* @sa MORSE_zgetrf_incpiv_Tile_Async
*
******************************************************************************/
-int MORSE_ztrsmpl_Tile_Async(MORSE_desc_t *A, MORSE_desc_t *L, int *IPIV, MORSE_desc_t *B,
- MORSE_sequence_t *sequence, MORSE_request_t *request)
+int MORSE_ztrsmpl_Tile_Async( MORSE_desc_t *A, MORSE_desc_t *L, int *IPIV, MORSE_desc_t *B,
+ MORSE_sequence_t *sequence, MORSE_request_t *request )
{
MORSE_context_t *morse;
@@ -272,10 +271,12 @@ int MORSE_ztrsmpl_Tile_Async(MORSE_desc_t *A, MORSE_desc_t *L, int *IPIV, MORSE_
return MORSE_ERR_UNALLOCATED;
}
/* Check sequence status */
- if (sequence->status == MORSE_SUCCESS)
+ if (sequence->status == MORSE_SUCCESS) {
request->status = MORSE_SUCCESS;
- else
+ }
+ else {
return morse_request_fail(sequence, request, MORSE_ERR_SEQUENCE_FLUSHED);
+ }
/* Check descriptors for correctness */
if (morse_desc_check(A) != MORSE_SUCCESS) {
@@ -296,11 +297,11 @@ int MORSE_ztrsmpl_Tile_Async(MORSE_desc_t *A, MORSE_desc_t *L, int *IPIV, MORSE_
return morse_request_fail(sequence, request, MORSE_ERR_ILLEGAL_VALUE);
}
/* Quick return */
-/*
- if (chameleon_min(N, NRHS) == 0)
- return MORSE_SUCCESS;
-*/
- morse_pztrsmpl(A, B, L, IPIV, sequence, request);
+ /*
+ if (chameleon_min(N, NRHS) == 0)
+ return MORSE_SUCCESS;
+ */
+ morse_pztrsmpl( A, B, L, IPIV, sequence, request );
return MORSE_SUCCESS;
}
diff --git a/compute/ztrtri.c b/compute/ztrtri.c
index 2f2d36c6040991076ab17946577becfde41f311a..e4d822d94fce6c2b348925b26faf29c0bed71cfa 100644
--- a/compute/ztrtri.c
+++ b/compute/ztrtri.c
@@ -82,15 +82,15 @@
* @sa MORSE_zpotri
*
******************************************************************************/
-int MORSE_ztrtri(MORSE_enum uplo, MORSE_enum diag, int N,
- MORSE_Complex64_t *A, int LDA)
+int MORSE_ztrtri( MORSE_enum uplo, MORSE_enum diag, int N,
+ MORSE_Complex64_t *A, int LDA )
{
int NB;
int status;
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
MORSE_request_t request = MORSE_REQUEST_INITIALIZER;
- MORSE_desc_t descA;
+ MORSE_desc_t descAl, descAt;
morse = morse_context_self();
if (morse == NULL) {
@@ -98,11 +98,11 @@ int MORSE_ztrtri(MORSE_enum uplo, MORSE_enum diag, int N,
return MORSE_ERR_NOT_INITIALIZED;
}
/* Check input arguments */
- if (uplo != MorseUpper && uplo != MorseLower) {
+ if ((uplo != MorseUpper) && (uplo != MorseLower)) {
morse_error("MORSE_ztrtri", "illegal value of uplo");
return -1;
}
- if (diag != MorseUnit && diag != MorseNonUnit) {
+ if ((diag != MorseUnit) && (diag != MorseNonUnit)) {
morse_error("MORSE_ztrtri", "illegal value of diag");
return -2;
}
@@ -126,32 +126,28 @@ int MORSE_ztrtri(MORSE_enum uplo, MORSE_enum diag, int N,
}
/* Set NT */
- NB = MORSE_NB;
+ NB = MORSE_NB;
- morse_sequence_create(morse, &sequence);
+ morse_sequence_create( morse, &sequence );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooplap2tile( descA, A, NB, NB, LDA, N, 0, 0, N, N, sequence, &request,
- morse_desc_mat_free(&(descA)) );
-/* } else {*/
-/* morse_ziplap2tile( descA, A, NB, NB, LDA, N, 0, 0, N, N,*/
-/* sequence, &request);*/
-/* }*/
+ /* Submit the matrix conversion */
+ morse_zlap2tile( morse, &descAl, &descAt, MorseDescInout, uplo,
+ A, NB, NB, LDA, N, N, N, sequence, &request );
/* Call the tile interface */
- MORSE_ztrtri_Tile_Async(uplo, diag, &descA, sequence, &request);
+ MORSE_ztrtri_Tile_Async( uplo, diag, &descAt, sequence, &request );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooptile2lap(descA, A, NB, NB, LDA, N, sequence, &request);
- morse_sequence_wait(morse, sequence);
- morse_desc_mat_free(&descA);
-/* } else {*/
-/* morse_ziptile2lap( descA, A, NB, NB, LDA, N, sequence, &request);*/
-/* morse_sequence_wait(morse, sequence);*/
-/* }*/
+ /* Submit the matrix conversion back */
+ morse_ztile2lap( morse, &descAl, &descAt,
+ MorseDescInout, uplo, sequence, &request );
+
+ morse_sequence_wait( morse, sequence );
+
+ /* Cleanup the temporary data */
+ morse_ztile2lap_cleanup( morse, &descAl, &descAt );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -206,7 +202,7 @@ int MORSE_ztrtri(MORSE_enum uplo, MORSE_enum diag, int N,
* @sa MORSE_zpotri_Tile
*
******************************************************************************/
-int MORSE_ztrtri_Tile(MORSE_enum uplo, MORSE_enum diag, MORSE_desc_t *A)
+int MORSE_ztrtri_Tile( MORSE_enum uplo, MORSE_enum diag, MORSE_desc_t *A )
{
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
@@ -218,13 +214,15 @@ int MORSE_ztrtri_Tile(MORSE_enum uplo, MORSE_enum diag, MORSE_desc_t *A)
morse_fatal_error("MORSE_ztrtri_Tile", "MORSE not initialized");
return MORSE_ERR_NOT_INITIALIZED;
}
- morse_sequence_create(morse, &sequence);
- MORSE_ztrtri_Tile_Async(uplo, diag, A, sequence, &request);
- RUNTIME_desc_flush( A, sequence );
- morse_sequence_wait(morse, sequence);
-
+ morse_sequence_create( morse, &sequence );
+
+ MORSE_ztrtri_Tile_Async( uplo, diag, A, sequence, &request );
+
+ MORSE_Desc_Flush( A, sequence );
+
+ morse_sequence_wait( morse, sequence );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -258,8 +256,8 @@ int MORSE_ztrtri_Tile(MORSE_enum uplo, MORSE_enum diag, MORSE_desc_t *A)
* @sa MORSE_zpotri_Tile_Async
*
******************************************************************************/
-int MORSE_ztrtri_Tile_Async(MORSE_enum uplo, MORSE_enum diag, MORSE_desc_t *A,
- MORSE_sequence_t *sequence, MORSE_request_t *request)
+int MORSE_ztrtri_Tile_Async( MORSE_enum uplo, MORSE_enum diag, MORSE_desc_t *A,
+ MORSE_sequence_t *sequence, MORSE_request_t *request )
{
MORSE_context_t *morse;
@@ -277,10 +275,12 @@ int MORSE_ztrtri_Tile_Async(MORSE_enum uplo, MORSE_enum diag, MORSE_desc_t *A,
return MORSE_ERR_UNALLOCATED;
}
/* Check sequence status */
- if (sequence->status == MORSE_SUCCESS)
+ if (sequence->status == MORSE_SUCCESS) {
request->status = MORSE_SUCCESS;
- else
+ }
+ else {
return morse_request_fail(sequence, request, MORSE_ERR_SEQUENCE_FLUSHED);
+ }
/* Check descriptors for correctness */
if (morse_desc_check(A) != MORSE_SUCCESS) {
@@ -292,20 +292,20 @@ int MORSE_ztrtri_Tile_Async(MORSE_enum uplo, MORSE_enum diag, MORSE_desc_t *A,
morse_error("MORSE_ztrtri_Tile", "only square tiles supported");
return morse_request_fail(sequence, request, MORSE_ERR_ILLEGAL_VALUE);
}
- if (uplo != MorseUpper && uplo != MorseLower) {
+ if ((uplo != MorseUpper) && (uplo != MorseLower)) {
morse_error("MORSE_ztrtri_Tile", "illegal value of uplo");
return morse_request_fail(sequence, request, -1);
}
- if (diag != MorseUnit && diag != MorseNonUnit) {
+ if ((diag != MorseUnit) && (diag != MorseNonUnit)) {
morse_error("MORSE_ztrtri_Tile", "illegal value of diag");
return morse_request_fail(sequence, request, -2);
}
/* Quick return */
-/*
- if (chameleon_max(N, 0) == 0)
- return MORSE_SUCCESS;
-*/
- morse_pztrtri(uplo, diag, A, sequence, request);
+ /*
+ if (chameleon_max(N, 0) == 0)
+ return MORSE_SUCCESS;
+ */
+ morse_pztrtri( uplo, diag, A, sequence, request );
return MORSE_SUCCESS;
}
diff --git a/compute/zunglq.c b/compute/zunglq.c
index 7ececcb759dc1614ec1ec3b72f2691c541c6c841..f8738d76871d662731b63610adc717d64bd31ffe 100644
--- a/compute/zunglq.c
+++ b/compute/zunglq.c
@@ -81,17 +81,18 @@
* @sa MORSE_zgelqf
*
******************************************************************************/
-int MORSE_zunglq(int M, int N, int K,
+int MORSE_zunglq( int M, int N, int K,
MORSE_Complex64_t *A, int LDA,
MORSE_desc_t *descT,
- MORSE_Complex64_t *Q, int LDQ)
+ MORSE_Complex64_t *Q, int LDQ )
{
int NB;
int status;
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
MORSE_request_t request = MORSE_REQUEST_INITIALIZER;
- MORSE_desc_t descA, descQ;
+ MORSE_desc_t descAl, descAt;
+ MORSE_desc_t descQl, descQt;
morse = morse_context_self();
if (morse == NULL) {
@@ -134,36 +135,31 @@ int MORSE_zunglq(int M, int N, int K,
/* Set NT */
NB = MORSE_NB;
- morse_sequence_create(morse, &sequence);
+ morse_sequence_create( morse, &sequence );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooplap2tile( descA, A, NB, NB, LDA, N, 0, 0, K, N, sequence, &request,
- morse_desc_mat_free(&(descA)) );
- morse_zooplap2tile( descQ, Q, NB, NB, LDQ, N, 0, 0, M, N, sequence, &request,
- morse_desc_mat_free(&(descA)); morse_desc_mat_free(&(descQ)));
-/* } else {*/
-/* morse_ziplap2tile( descA, A, NB, NB, LDA, N, 0, 0, K, N,*/
-/* sequence, &request);*/
-/* morse_ziplap2tile( descQ, Q, NB, NB, LDQ, N, 0, 0, M, N,*/
-/* sequence, &request);*/
-/* }*/
+ /* Submit the matrix conversion */
+ morse_zlap2tile( morse, &descAl, &descAt, MorseDescInput, MorseUpper,
+ A, NB, NB, LDA, N, K, N, sequence, &request );
+ morse_zlap2tile( morse, &descQl, &descQt, MorseDescInout, MorseUpperLower,
+ Q, NB, NB, LDQ, N, M, N, sequence, &request );
/* Call the tile interface */
- MORSE_zunglq_Tile_Async(&descA, descT, &descQ, sequence, &request);
+ MORSE_zunglq_Tile_Async( &descAt, descT, &descQt, sequence, &request );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooptile2lap(descQ, Q, NB, NB, LDQ, N, sequence, &request);
- morse_sequence_wait(morse, sequence);
- morse_desc_mat_free(&descA);
- morse_desc_mat_free(&descQ);
-/* } else {*/
-/* morse_ziptile2lap( descA, A, NB, NB, LDA, N, sequence, &request);*/
-/* morse_ziptile2lap( descQ, Q, NB, NB, LDQ, N, sequence, &request);*/
-/* morse_sequence_wait(morse, sequence);*/
-/* }*/
+ /* Submit the matrix conversion back */
+ morse_ztile2lap( morse, &descAl, &descAt,
+ MorseDescInput, MorseUpper, sequence, &request );
+ morse_ztile2lap( morse, &descQl, &descQt,
+ MorseDescInout, MorseUpperLower, sequence, &request );
+
+ morse_sequence_wait( morse, sequence );
+
+ /* Cleanup the temporary data */
+ morse_ztile2lap_cleanup( morse, &descAl, &descAt );
+ morse_ztile2lap_cleanup( morse, &descQl, &descQt );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -202,7 +198,7 @@ int MORSE_zunglq(int M, int N, int K,
* @sa MORSE_zgelqf_Tile
*
******************************************************************************/
-int MORSE_zunglq_Tile(MORSE_desc_t *A, MORSE_desc_t *T, MORSE_desc_t *Q)
+int MORSE_zunglq_Tile( MORSE_desc_t *A, MORSE_desc_t *T, MORSE_desc_t *Q )
{
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
@@ -214,14 +210,17 @@ int MORSE_zunglq_Tile(MORSE_desc_t *A, MORSE_desc_t *T, MORSE_desc_t *Q)
morse_fatal_error("MORSE_zunglq_Tile", "MORSE not initialized");
return MORSE_ERR_NOT_INITIALIZED;
}
- morse_sequence_create(morse, &sequence);
- MORSE_zunglq_Tile_Async(A, T, Q, sequence, &request);
- RUNTIME_desc_flush( A, sequence );
- RUNTIME_desc_flush( Q, sequence );
- morse_sequence_wait(morse, sequence);
+ morse_sequence_create( morse, &sequence );
+
+ MORSE_zunglq_Tile_Async( A, T, Q, sequence, &request );
+
+ MORSE_Desc_Flush( A, sequence );
+ MORSE_Desc_Flush( T, sequence );
+ MORSE_Desc_Flush( Q, sequence );
+ morse_sequence_wait( morse, sequence );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -253,8 +252,8 @@ int MORSE_zunglq_Tile(MORSE_desc_t *A, MORSE_desc_t *T, MORSE_desc_t *Q)
* @sa MORSE_zgelqf_Tile_Async
*
******************************************************************************/
-int MORSE_zunglq_Tile_Async(MORSE_desc_t *A, MORSE_desc_t *T, MORSE_desc_t *Q,
- MORSE_sequence_t *sequence, MORSE_request_t *request)
+int MORSE_zunglq_Tile_Async( MORSE_desc_t *A, MORSE_desc_t *T, MORSE_desc_t *Q,
+ MORSE_sequence_t *sequence, MORSE_request_t *request )
{
MORSE_context_t *morse;
MORSE_desc_t D, *Dptr = NULL;
@@ -273,10 +272,12 @@ int MORSE_zunglq_Tile_Async(MORSE_desc_t *A, MORSE_desc_t *T, MORSE_desc_t *Q,
return MORSE_ERR_UNALLOCATED;
}
/* Check sequence status */
- if (sequence->status == MORSE_SUCCESS)
+ if (sequence->status == MORSE_SUCCESS) {
request->status = MORSE_SUCCESS;
- else
+ }
+ else {
return morse_request_fail(sequence, request, MORSE_ERR_SEQUENCE_FLUSHED);
+ }
/* Check descriptors for correctness */
if (morse_desc_check(A) != MORSE_SUCCESS) {
@@ -298,10 +299,10 @@ int MORSE_zunglq_Tile_Async(MORSE_desc_t *A, MORSE_desc_t *T, MORSE_desc_t *Q,
}
/* Quick return - currently NOT equivalent to LAPACK's:
* CALL DLASET( 'Full', MAX( M, N ), NRHS, ZERO, ZERO, Q, LDQ ) */
-/*
- if (chameleon_min(M, N) == 0)
- return MORSE_SUCCESS;
-*/
+ /*
+ if (chameleon_min(M, N) == 0)
+ return MORSE_SUCCESS;
+ */
#if defined(CHAMELEON_COPY_DIAG)
{
@@ -311,16 +312,21 @@ int MORSE_zunglq_Tile_Async(MORSE_desc_t *A, MORSE_desc_t *T, MORSE_desc_t *Q,
}
#endif
- morse_pzlaset(MorseUpperLower, 0., 1., Q, sequence, request);
+ morse_pzlaset( MorseUpperLower, 0., 1., Q, sequence, request );
if (morse->householder == MORSE_FLAT_HOUSEHOLDER) {
- morse_pzunglq(A, Q, T, Dptr, sequence, request);
+ morse_pzunglq( A, Q, T, Dptr, sequence, request );
}
else {
- morse_pzunglqrh(A, Q, T, Dptr, MORSE_RHBLK, sequence, request);
+ morse_pzunglqrh( A, Q, T, Dptr, MORSE_RHBLK, sequence, request );
}
if (Dptr != NULL) {
- morse_desc_mat_free(Dptr);
+ MORSE_Desc_Flush( A, sequence );
+ MORSE_Desc_Flush( Q, sequence );
+ MORSE_Desc_Flush( T, sequence );
+ MORSE_Desc_Flush( Dptr, sequence );
+ morse_sequence_wait( morse, sequence );
+ morse_desc_mat_free( Dptr );
}
(void)D;
return MORSE_SUCCESS;
diff --git a/compute/zunglq_param.c b/compute/zunglq_param.c
index 5b6d30e6065194ea3f825b7304f889c8f9be2bb9..45110d3f79934e99119b81a80f52fe27e6c8d31b 100644
--- a/compute/zunglq_param.c
+++ b/compute/zunglq_param.c
@@ -78,17 +78,18 @@
* @sa MORSE_zgelqf
*
******************************************************************************/
-int MORSE_zunglq_param(const libhqr_tree_t *qrtree, int M, int N, int K,
- MORSE_Complex64_t *A, int LDA,
- MORSE_desc_t *descTS, MORSE_desc_t *descTT,
- MORSE_Complex64_t *Q, int LDQ)
+int MORSE_zunglq_param( const libhqr_tree_t *qrtree, int M, int N, int K,
+ MORSE_Complex64_t *A, int LDA,
+ MORSE_desc_t *descTS, MORSE_desc_t *descTT,
+ MORSE_Complex64_t *Q, int LDQ )
{
int NB;
int status;
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
MORSE_request_t request = MORSE_REQUEST_INITIALIZER;
- MORSE_desc_t descA, descQ;
+ MORSE_desc_t descAl, descAt;
+ MORSE_desc_t descQl, descQt;
morse = morse_context_self();
if (morse == NULL) {
@@ -131,36 +132,31 @@ int MORSE_zunglq_param(const libhqr_tree_t *qrtree, int M, int N, int K,
/* Set NT */
NB = MORSE_NB;
- morse_sequence_create(morse, &sequence);
+ morse_sequence_create( morse, &sequence );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooplap2tile( descA, A, NB, NB, LDA, N, 0, 0, K, N, sequence, &request,
- morse_desc_mat_free(&(descA)) );
- morse_zooplap2tile( descQ, Q, NB, NB, LDQ, N, 0, 0, M, N, sequence, &request,
- morse_desc_mat_free(&(descA)); morse_desc_mat_free(&(descQ)));
-/* } else {*/
-/* morse_ziplap2tile( descA, A, NB, NB, LDA, N, 0, 0, K, N,*/
-/* sequence, &request);*/
-/* morse_ziplap2tile( descQ, Q, NB, NB, LDQ, N, 0, 0, M, N,*/
-/* sequence, &request);*/
-/* }*/
+ /* Submit the matrix conversion */
+ morse_zlap2tile( morse, &descAl, &descAt, MorseDescInput, MorseUpper,
+ A, NB, NB, LDA, N, K, N, sequence, &request );
+ morse_zlap2tile( morse, &descQl, &descQt, MorseDescInout, MorseUpperLower,
+ Q, NB, NB, LDQ, N, M, N, sequence, &request );
/* Call the tile interface */
- MORSE_zunglq_param_Tile_Async(qrtree, &descA, descTS, descTT, &descQ, sequence, &request);
+ MORSE_zunglq_param_Tile_Async( qrtree, &descAt, descTS, descTT, &descQt, sequence, &request );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooptile2lap(descQ, Q, NB, NB, LDQ, N, sequence, &request);
- morse_sequence_wait(morse, sequence);
- morse_desc_mat_free(&descA);
- morse_desc_mat_free(&descQ);
-/* } else {*/
-/* morse_ziptile2lap( descA, A, NB, NB, LDA, N, sequence, &request);*/
-/* morse_ziptile2lap( descQ, Q, NB, NB, LDQ, N, sequence, &request);*/
-/* morse_sequence_wait(morse, sequence);*/
-/* }*/
+ /* Submit the matrix conversion back */
+ morse_ztile2lap( morse, &descAl, &descAt,
+ MorseDescInput, MorseUpper, sequence, &request );
+ morse_ztile2lap( morse, &descQl, &descQt,
+ MorseDescInout, MorseUpperLower, sequence, &request );
+
+ morse_sequence_wait( morse, sequence );
+
+ /* Cleanup the temporary data */
+ morse_ztile2lap_cleanup( morse, &descAl, &descAt );
+ morse_ztile2lap_cleanup( morse, &descQl, &descQt );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -199,7 +195,7 @@ int MORSE_zunglq_param(const libhqr_tree_t *qrtree, int M, int N, int K,
* @sa MORSE_zgelqf_Tile
*
******************************************************************************/
-int MORSE_zunglq_param_Tile(const libhqr_tree_t *qrtree, MORSE_desc_t *A, MORSE_desc_t *TS, MORSE_desc_t *TT, MORSE_desc_t *Q)
+int MORSE_zunglq_param_Tile( const libhqr_tree_t *qrtree, MORSE_desc_t *A, MORSE_desc_t *TS, MORSE_desc_t *TT, MORSE_desc_t *Q )
{
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
@@ -211,14 +207,18 @@ int MORSE_zunglq_param_Tile(const libhqr_tree_t *qrtree, MORSE_desc_t *A, MORSE_
morse_fatal_error("MORSE_zunglq_param_Tile", "MORSE not initialized");
return MORSE_ERR_NOT_INITIALIZED;
}
- morse_sequence_create(morse, &sequence);
- MORSE_zunglq_param_Tile_Async(qrtree, A, TS, TT, Q, sequence, &request);
- RUNTIME_desc_flush( A, sequence );
- RUNTIME_desc_flush( Q, sequence );
- morse_sequence_wait(morse, sequence);
+ morse_sequence_create( morse, &sequence );
+
+ MORSE_zunglq_param_Tile_Async( qrtree, A, TS, TT, Q, sequence, &request );
+
+ MORSE_Desc_Flush( A, sequence );
+ MORSE_Desc_Flush( TS, sequence );
+ MORSE_Desc_Flush( TT, sequence );
+ MORSE_Desc_Flush( Q, sequence );
+ morse_sequence_wait( morse, sequence );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -250,8 +250,8 @@ int MORSE_zunglq_param_Tile(const libhqr_tree_t *qrtree, MORSE_desc_t *A, MORSE_
* @sa MORSE_zgelqf_Tile_Async
*
******************************************************************************/
-int MORSE_zunglq_param_Tile_Async(const libhqr_tree_t *qrtree, MORSE_desc_t *A, MORSE_desc_t *TS, MORSE_desc_t *TT, MORSE_desc_t *Q,
- MORSE_sequence_t *sequence, MORSE_request_t *request)
+int MORSE_zunglq_param_Tile_Async( const libhqr_tree_t *qrtree, MORSE_desc_t *A, MORSE_desc_t *TS, MORSE_desc_t *TT, MORSE_desc_t *Q,
+ MORSE_sequence_t *sequence, MORSE_request_t *request )
{
MORSE_context_t *morse;
MORSE_desc_t D, *Dptr = NULL;
@@ -270,10 +270,12 @@ int MORSE_zunglq_param_Tile_Async(const libhqr_tree_t *qrtree, MORSE_desc_t *A,
return MORSE_ERR_UNALLOCATED;
}
/* Check sequence status */
- if (sequence->status == MORSE_SUCCESS)
+ if (sequence->status == MORSE_SUCCESS) {
request->status = MORSE_SUCCESS;
- else
+ }
+ else {
return morse_request_fail(sequence, request, MORSE_ERR_SEQUENCE_FLUSHED);
+ }
/* Check descriptors for correctness */
if (morse_desc_check(A) != MORSE_SUCCESS) {
@@ -299,10 +301,10 @@ int MORSE_zunglq_param_Tile_Async(const libhqr_tree_t *qrtree, MORSE_desc_t *A,
}
/* Quick return - currently NOT equivalent to LAPACK's:
* CALL DLASET( 'Full', MAX( M, N ), NRHS, ZERO, ZERO, Q, LDQ ) */
-/*
- if (chameleon_min(M, N) == 0)
- return MORSE_SUCCESS;
-*/
+ /*
+ if (chameleon_min(M, N) == 0)
+ return MORSE_SUCCESS;
+ */
#if defined(CHAMELEON_COPY_DIAG)
{
int m = chameleon_min(A->mt, A->nt) * A->mb;
@@ -311,11 +313,17 @@ int MORSE_zunglq_param_Tile_Async(const libhqr_tree_t *qrtree, MORSE_desc_t *A,
}
#endif
- morse_pzlaset(MorseUpperLower, 0., 1., Q, sequence, request);
- morse_pzunglq_param(qrtree, A, Q, TS, TT, Dptr, sequence, request);
+ morse_pzlaset( MorseUpperLower, 0., 1., Q, sequence, request );
+ morse_pzunglq_param( qrtree, A, Q, TS, TT, Dptr, sequence, request );
if (Dptr != NULL) {
- morse_desc_mat_free(Dptr);
+ MORSE_Desc_Flush( A, sequence );
+ MORSE_Desc_Flush( Q, sequence );
+ MORSE_Desc_Flush( TS, sequence );
+ MORSE_Desc_Flush( TT, sequence );
+ MORSE_Desc_Flush( Dptr, sequence );
+ morse_sequence_wait( morse, sequence );
+ morse_desc_mat_free( Dptr );
}
(void)D;
return MORSE_SUCCESS;
diff --git a/compute/zungqr.c b/compute/zungqr.c
index 822dc4e568f7c9bca64af7c37746b57105aeba7f..954f2da51d54070bcc6feb12fc1c66aa6d8664ae 100644
--- a/compute/zungqr.c
+++ b/compute/zungqr.c
@@ -81,17 +81,18 @@
* @sa MORSE_zgeqrf
*
******************************************************************************/
-int MORSE_zungqr(int M, int N, int K,
+int MORSE_zungqr( int M, int N, int K,
MORSE_Complex64_t *A, int LDA,
MORSE_desc_t *descT,
- MORSE_Complex64_t *Q, int LDQ)
+ MORSE_Complex64_t *Q, int LDQ )
{
int NB;
int status;
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
MORSE_request_t request = MORSE_REQUEST_INITIALIZER;
- MORSE_desc_t descA, descQ;
+ MORSE_desc_t descAl, descAt;
+ MORSE_desc_t descQl, descQt;
morse = morse_context_self();
if (morse == NULL) {
@@ -133,36 +134,31 @@ int MORSE_zungqr(int M, int N, int K,
/* Set NT */
NB = MORSE_NB;
- morse_sequence_create(morse, &sequence);
+ morse_sequence_create( morse, &sequence );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooplap2tile( descA, A, NB, NB, LDA, N, 0, 0, M, K, sequence, &request,
- morse_desc_mat_free(&(descA)) );
- morse_zooplap2tile( descQ, Q, NB, NB, LDQ, N, 0, 0, M, N, sequence, &request,
- morse_desc_mat_free(&(descA)); morse_desc_mat_free(&(descQ)));
-/* } else {*/
-/* morse_ziplap2tile( descA, A, NB, NB, LDA, N, 0, 0, M, K,*/
-/* sequence, &request);*/
-/* morse_ziplap2tile( descQ, Q, NB, NB, LDQ, N, 0, 0, M, N,*/
-/* sequence, &request);*/
-/* }*/
+ /* Submit the matrix conversion */
+ morse_zlap2tile( morse, &descAl, &descAt, MorseDescInput, MorseLower,
+ A, NB, NB, LDA, N, M, K, sequence, &request );
+ morse_zlap2tile( morse, &descQl, &descQt, MorseDescInout, MorseUpperLower,
+ Q, NB, NB, LDQ, N, M, N, sequence, &request );
/* Call the tile interface */
- MORSE_zungqr_Tile_Async(&descA, descT, &descQ, sequence, &request);
+ MORSE_zungqr_Tile_Async( &descAt, descT, &descQt, sequence, &request );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooptile2lap(descQ, Q, NB, NB, LDQ, N, sequence, &request);
- morse_sequence_wait(morse, sequence);
- morse_desc_mat_free(&descA);
- morse_desc_mat_free(&descQ);
-/* } else {*/
-/* morse_ziptile2lap( descA, A, NB, NB, LDA, K, sequence, &request);*/
-/* morse_ziptile2lap( descQ, Q, NB, NB, LDQ, N, sequence, &request);*/
-/* morse_sequence_wait(morse, sequence);*/
-/* }*/
+ /* Submit the matrix conversion back */
+ morse_ztile2lap( morse, &descAl, &descAt,
+ MorseDescInput, MorseLower, sequence, &request );
+ morse_ztile2lap( morse, &descQl, &descQt,
+ MorseDescInout, MorseUpperLower, sequence, &request );
+
+ morse_sequence_wait( morse, sequence );
+
+ /* Cleanup the temporary data */
+ morse_ztile2lap_cleanup( morse, &descAl, &descAt );
+ morse_ztile2lap_cleanup( morse, &descQl, &descQt );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -201,7 +197,7 @@ int MORSE_zungqr(int M, int N, int K,
* @sa MORSE_zgeqrf_Tile
*
******************************************************************************/
-int MORSE_zungqr_Tile(MORSE_desc_t *A, MORSE_desc_t *T, MORSE_desc_t *Q)
+int MORSE_zungqr_Tile( MORSE_desc_t *A, MORSE_desc_t *T, MORSE_desc_t *Q )
{
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
@@ -213,14 +209,17 @@ int MORSE_zungqr_Tile(MORSE_desc_t *A, MORSE_desc_t *T, MORSE_desc_t *Q)
morse_fatal_error("MORSE_zungqr_Tile", "MORSE not initialized");
return MORSE_ERR_NOT_INITIALIZED;
}
- morse_sequence_create(morse, &sequence);
- MORSE_zungqr_Tile_Async(A, T, Q, sequence, &request);
- RUNTIME_desc_flush( A, sequence );
- RUNTIME_desc_flush( Q, sequence );
- morse_sequence_wait(morse, sequence);
-
+ morse_sequence_create( morse, &sequence );
+
+ MORSE_zungqr_Tile_Async( A, T, Q, sequence, &request );
+
+ MORSE_Desc_Flush( A, sequence );
+ MORSE_Desc_Flush( T, sequence );
+ MORSE_Desc_Flush( Q, sequence );
+
+ morse_sequence_wait( morse, sequence );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -252,8 +251,8 @@ int MORSE_zungqr_Tile(MORSE_desc_t *A, MORSE_desc_t *T, MORSE_desc_t *Q)
* @sa MORSE_zgeqrf_Tile_Async
*
******************************************************************************/
-int MORSE_zungqr_Tile_Async(MORSE_desc_t *A, MORSE_desc_t *T, MORSE_desc_t *Q,
- MORSE_sequence_t *sequence, MORSE_request_t *request)
+int MORSE_zungqr_Tile_Async( MORSE_desc_t *A, MORSE_desc_t *T, MORSE_desc_t *Q,
+ MORSE_sequence_t *sequence, MORSE_request_t *request )
{
MORSE_context_t *morse;
MORSE_desc_t D, *Dptr = NULL;
@@ -272,10 +271,12 @@ int MORSE_zungqr_Tile_Async(MORSE_desc_t *A, MORSE_desc_t *T, MORSE_desc_t *Q,
return MORSE_ERR_UNALLOCATED;
}
/* Check sequence status */
- if (sequence->status == MORSE_SUCCESS)
+ if (sequence->status == MORSE_SUCCESS) {
request->status = MORSE_SUCCESS;
- else
+ }
+ else {
return morse_request_fail(sequence, request, MORSE_ERR_SEQUENCE_FLUSHED);
+ }
/* Check descriptors for correctness */
if (morse_desc_check(A) != MORSE_SUCCESS) {
@@ -296,10 +297,10 @@ int MORSE_zungqr_Tile_Async(MORSE_desc_t *A, MORSE_desc_t *T, MORSE_desc_t *Q,
return morse_request_fail(sequence, request, MORSE_ERR_ILLEGAL_VALUE);
}
/* Quick return */
-/*
- if (N <= 0)
- return MORSE_SUCCESS;
-*/
+ /*
+ if (N <= 0)
+ return MORSE_SUCCESS;
+ */
#if defined(CHAMELEON_COPY_DIAG)
{
int n = chameleon_min(A->mt, A->nt) * A->nb;
@@ -308,16 +309,21 @@ int MORSE_zungqr_Tile_Async(MORSE_desc_t *A, MORSE_desc_t *T, MORSE_desc_t *Q,
}
#endif
- morse_pzlaset(MorseUpperLower, 0., 1., Q, sequence, request);
+ morse_pzlaset( MorseUpperLower, 0., 1., Q, sequence, request );
if (morse->householder == MORSE_FLAT_HOUSEHOLDER) {
- morse_pzungqr(A, Q, T, Dptr, sequence, request);
+ morse_pzungqr( A, Q, T, Dptr, sequence, request );
}
else {
- morse_pzungqrrh(A, Q, T, Dptr, MORSE_RHBLK, sequence, request);
+ morse_pzungqrrh( A, Q, T, Dptr, MORSE_RHBLK, sequence, request );
}
if (Dptr != NULL) {
- morse_desc_mat_free(Dptr);
+ MORSE_Desc_Flush( A, sequence );
+ MORSE_Desc_Flush( Q, sequence );
+ MORSE_Desc_Flush( T, sequence );
+ MORSE_Desc_Flush( Dptr, sequence );
+ morse_sequence_wait( morse, sequence );
+ morse_desc_mat_free( Dptr );
}
(void)D;
return MORSE_SUCCESS;
diff --git a/compute/zungqr_param.c b/compute/zungqr_param.c
index af88ea598d0b1f9c13bfd1c328f5f5ec2e1170d3..8621b7816829436859c3cafe994ad6bbb0f9fc6a 100644
--- a/compute/zungqr_param.c
+++ b/compute/zungqr_param.c
@@ -78,19 +78,20 @@
* @sa MORSE_zgeqrf
*
******************************************************************************/
-int MORSE_zungqr_param(const libhqr_tree_t *qrtree,
- int M, int N, int K,
- MORSE_Complex64_t *A, int LDA,
- MORSE_desc_t *descTS,
- MORSE_desc_t *descTT,
- MORSE_Complex64_t *Q, int LDQ)
+int MORSE_zungqr_param( const libhqr_tree_t *qrtree,
+ int M, int N, int K,
+ MORSE_Complex64_t *A, int LDA,
+ MORSE_desc_t *descTS,
+ MORSE_desc_t *descTT,
+ MORSE_Complex64_t *Q, int LDQ )
{
int NB;
int status;
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
MORSE_request_t request = MORSE_REQUEST_INITIALIZER;
- MORSE_desc_t descA, descQ;
+ MORSE_desc_t descAl, descAt;
+ MORSE_desc_t descQl, descQt;
morse = morse_context_self();
if (morse == NULL) {
@@ -132,36 +133,31 @@ int MORSE_zungqr_param(const libhqr_tree_t *qrtree,
/* Set NT */
NB = MORSE_NB;
- morse_sequence_create(morse, &sequence);
+ morse_sequence_create( morse, &sequence );
- /* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooplap2tile( descA, A, NB, NB, LDA, N, 0, 0, M, K, sequence, &request,
- morse_desc_mat_free(&(descA)) );
- morse_zooplap2tile( descQ, Q, NB, NB, LDQ, N, 0, 0, M, N, sequence, &request,
- morse_desc_mat_free(&(descA)); morse_desc_mat_free(&(descQ)));
- /* } else {*/
- /* morse_ziplap2tile( descA, A, NB, NB, LDA, N, 0, 0, M, K,*/
- /* sequence, &request);*/
- /* morse_ziplap2tile( descQ, Q, NB, NB, LDQ, N, 0, 0, M, N,*/
- /* sequence, &request);*/
- /* }*/
+ /* Submit the matrix conversion */
+ morse_zlap2tile( morse, &descAl, &descAt, MorseDescInput, MorseLower,
+ A, NB, NB, LDA, N, M, K, sequence, &request );
+ morse_zlap2tile( morse, &descQl, &descQt, MorseDescInout, MorseUpperLower,
+ Q, NB, NB, LDQ, N, M, N, sequence, &request );
/* Call the tile interface */
- MORSE_zungqr_param_Tile_Async(qrtree, &descA, descTS, descTT, &descQ, sequence, &request);
+ MORSE_zungqr_param_Tile_Async( qrtree, &descAt, descTS, descTT, &descQt, sequence, &request );
- /* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooptile2lap(descQ, Q, NB, NB, LDQ, N, sequence, &request);
- morse_sequence_wait(morse, sequence);
- morse_desc_mat_free(&descA);
- morse_desc_mat_free(&descQ);
- /* } else {*/
- /* morse_ziptile2lap( descA, A, NB, NB, LDA, K, sequence, &request);*/
- /* morse_ziptile2lap( descQ, Q, NB, NB, LDQ, N, sequence, &request);*/
- /* morse_sequence_wait(morse, sequence);*/
- /* }*/
+ /* Submit the matrix conversion back */
+ morse_ztile2lap( morse, &descAl, &descAt,
+ MorseDescInput, MorseLower, sequence, &request );
+ morse_ztile2lap( morse, &descQl, &descQt,
+ MorseDescInout, MorseUpperLower, sequence, &request );
+
+ morse_sequence_wait( morse, sequence );
+
+ /* Cleanup the temporary data */
+ morse_ztile2lap_cleanup( morse, &descAl, &descAt );
+ morse_ztile2lap_cleanup( morse, &descQl, &descQt );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -200,7 +196,7 @@ int MORSE_zungqr_param(const libhqr_tree_t *qrtree,
* @sa MORSE_zgeqrf_Tile
*
******************************************************************************/
-int MORSE_zungqr_param_Tile(const libhqr_tree_t *qrtree, MORSE_desc_t *A, MORSE_desc_t *TS, MORSE_desc_t *TT, MORSE_desc_t *Q)
+int MORSE_zungqr_param_Tile( const libhqr_tree_t *qrtree, MORSE_desc_t *A, MORSE_desc_t *TS, MORSE_desc_t *TT, MORSE_desc_t *Q )
{
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
@@ -212,14 +208,18 @@ int MORSE_zungqr_param_Tile(const libhqr_tree_t *qrtree, MORSE_desc_t *A, MORSE_
morse_fatal_error("MORSE_zungqr_param_Tile", "MORSE not initialized");
return MORSE_ERR_NOT_INITIALIZED;
}
- morse_sequence_create(morse, &sequence);
- MORSE_zungqr_param_Tile_Async(qrtree, A, TS, TT, Q, sequence, &request);
- RUNTIME_desc_flush( A, sequence );
- RUNTIME_desc_flush( Q, sequence );
- morse_sequence_wait(morse, sequence);
+ morse_sequence_create( morse, &sequence );
+
+ MORSE_zungqr_param_Tile_Async( qrtree, A, TS, TT, Q, sequence, &request );
+
+ MORSE_Desc_Flush( A, sequence );
+ MORSE_Desc_Flush( TS, sequence );
+ MORSE_Desc_Flush( TT, sequence );
+ MORSE_Desc_Flush( Q, sequence );
+ morse_sequence_wait( morse, sequence );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -251,7 +251,7 @@ int MORSE_zungqr_param_Tile(const libhqr_tree_t *qrtree, MORSE_desc_t *A, MORSE_
* @sa MORSE_zgeqrf_Tile_Async
*
******************************************************************************/
-int MORSE_zungqr_param_Tile_Async(const libhqr_tree_t *qrtree, MORSE_desc_t *A, MORSE_desc_t *TS, MORSE_desc_t *TT, MORSE_desc_t *Q, MORSE_sequence_t *sequence, MORSE_request_t *request)
+int MORSE_zungqr_param_Tile_Async( const libhqr_tree_t *qrtree, MORSE_desc_t *A, MORSE_desc_t *TS, MORSE_desc_t *TT, MORSE_desc_t *Q, MORSE_sequence_t *sequence, MORSE_request_t *request )
{
MORSE_context_t *morse;
MORSE_desc_t D, *Dptr = NULL;
@@ -269,10 +269,12 @@ int MORSE_zungqr_param_Tile_Async(const libhqr_tree_t *qrtree, MORSE_desc_t *A,
return MORSE_ERR_UNALLOCATED;
}
/* Check sequence status */
- if (sequence->status == MORSE_SUCCESS)
+ if (sequence->status == MORSE_SUCCESS) {
request->status = MORSE_SUCCESS;
- else
+ }
+ else {
return morse_request_fail(sequence, request, MORSE_ERR_SEQUENCE_FLUSHED);
+ }
/* Check descriptors for correctness */
if (morse_desc_check(A) != MORSE_SUCCESS) {
@@ -297,10 +299,10 @@ int MORSE_zungqr_param_Tile_Async(const libhqr_tree_t *qrtree, MORSE_desc_t *A,
return morse_request_fail(sequence, request, MORSE_ERR_ILLEGAL_VALUE);
}
/* Quick return */
-/*
- if (N <= 0)
- return MORSE_SUCCESS;
- */
+ /*
+ if (N <= 0)
+ return MORSE_SUCCESS;
+ */
#if defined(CHAMELEON_COPY_DIAG)
{
int n = chameleon_min(A->mt, A->nt) * A->nb;
@@ -309,11 +311,17 @@ int MORSE_zungqr_param_Tile_Async(const libhqr_tree_t *qrtree, MORSE_desc_t *A,
}
#endif
- morse_pzlaset(MorseUpperLower, 0., 1., Q, sequence, request);
- morse_pzungqr_param(qrtree, A, Q, TS, TT, Dptr, sequence, request);
+ morse_pzlaset( MorseUpperLower, 0., 1., Q, sequence, request );
+ morse_pzungqr_param( qrtree, A, Q, TS, TT, Dptr, sequence, request );
if (Dptr != NULL) {
- morse_desc_mat_free(Dptr);
+ MORSE_Desc_Flush( A, sequence );
+ MORSE_Desc_Flush( Q, sequence );
+ MORSE_Desc_Flush( TS, sequence );
+ MORSE_Desc_Flush( TT, sequence );
+ MORSE_Desc_Flush( Dptr, sequence );
+ morse_sequence_wait( morse, sequence );
+ morse_desc_mat_free( Dptr );
}
(void)D;
return MORSE_SUCCESS;
diff --git a/compute/zunmlq.c b/compute/zunmlq.c
index 81535c19d79ff71a24345609b2e319ad90c4056a..5901868c47a07b34ecc01815c8532909387c8d0d 100644
--- a/compute/zunmlq.c
+++ b/compute/zunmlq.c
@@ -105,17 +105,18 @@
* @sa MORSE_zgelqf
*
******************************************************************************/
-int MORSE_zunmlq(MORSE_enum side, MORSE_enum trans, int M, int N, int K,
+int MORSE_zunmlq( MORSE_enum side, MORSE_enum trans, int M, int N, int K,
MORSE_Complex64_t *A, int LDA,
MORSE_desc_t *descT,
- MORSE_Complex64_t *C, int LDC)
+ MORSE_Complex64_t *C, int LDC )
{
int NB, An;
int status;
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
MORSE_request_t request = MORSE_REQUEST_INITIALIZER;
- MORSE_desc_t descA, descC;
+ MORSE_desc_t descAl, descAt;
+ MORSE_desc_t descCl, descCt;
morse = morse_context_self();
if (morse == NULL) {
@@ -171,37 +172,31 @@ int MORSE_zunmlq(MORSE_enum side, MORSE_enum trans, int M, int N, int K,
/* Set MT, NT & NTRHS */
NB = MORSE_NB;
- morse_sequence_create(morse, &sequence);
+ morse_sequence_create( morse, &sequence );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooplap2tile( descA, A, NB, NB, LDA, An, 0, 0, K, An, sequence, &request,
- morse_desc_mat_free(&(descA)) );
- morse_zooplap2tile( descC, C, NB, NB, LDC, N, 0, 0, M, N, sequence, &request,
- morse_desc_mat_free(&(descA)); morse_desc_mat_free(&(descC)));
-/* } else {*/
-/* morse_ziplap2tile( descA, A, NB, NB, LDA, An, 0, 0, K, An,*/
-/* sequence, &request);*/
-/* morse_ziplap2tile( descC, C, NB, NB, LDC, N, 0, 0, M, N,*/
-/* sequence, &request);*/
-/* }*/
+ /* Submit the matrix conversion */
+ morse_zlap2tile( morse, &descAl, &descAt, MorseDescInput, MorseUpper,
+ A, NB, NB, LDA, An, K, An, sequence, &request );
+ morse_zlap2tile( morse, &descCl, &descCt, MorseDescInout, MorseUpperLower,
+ C, NB, NB, LDC, N, M, N, sequence, &request );
/* Call the tile interface */
- MORSE_zunmlq_Tile_Async(
- side, trans, &descA, descT, &descC, sequence, &request);
+ MORSE_zunmlq_Tile_Async( side, trans, &descAt, descT, &descCt, sequence, &request );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooptile2lap(descC, C, NB, NB, LDC, N, sequence, &request);
- morse_sequence_wait(morse, sequence);
- morse_desc_mat_free(&descA);
- morse_desc_mat_free(&descC);
-/* } else {*/
-/* morse_ziptile2lap( descA, A, NB, NB, LDA, An, sequence, &request);*/
-/* morse_ziptile2lap( descC, C, NB, NB, LDC, N, sequence, &request);*/
-/* morse_sequence_wait(morse, sequence);*/
-/* }*/
+ /* Submit the matrix conversion back */
+ morse_ztile2lap( morse, &descAl, &descAt,
+ MorseDescInput, MorseUpper, sequence, &request );
+ morse_ztile2lap( morse, &descCl, &descCt,
+ MorseDescInout, MorseUpperLower, sequence, &request );
+
+ morse_sequence_wait( morse, sequence );
+
+ /* Cleanup the temporary data */
+ morse_ztile2lap_cleanup( morse, &descAl, &descAt );
+ morse_ztile2lap_cleanup( morse, &descCl, &descCt );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -254,8 +249,8 @@ int MORSE_zunmlq(MORSE_enum side, MORSE_enum trans, int M, int N, int K,
* @sa MORSE_zgelqf_Tile
*
******************************************************************************/
-int MORSE_zunmlq_Tile(MORSE_enum side, MORSE_enum trans,
- MORSE_desc_t *A, MORSE_desc_t *T, MORSE_desc_t *C)
+int MORSE_zunmlq_Tile( MORSE_enum side, MORSE_enum trans,
+ MORSE_desc_t *A, MORSE_desc_t *T, MORSE_desc_t *C )
{
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
@@ -267,14 +262,17 @@ int MORSE_zunmlq_Tile(MORSE_enum side, MORSE_enum trans,
morse_fatal_error("MORSE_zunmlq_Tile", "MORSE not initialized");
return MORSE_ERR_NOT_INITIALIZED;
}
- morse_sequence_create(morse, &sequence);
- MORSE_zunmlq_Tile_Async(side, trans, A, T, C, sequence, &request);
- RUNTIME_desc_flush( A, sequence );
- RUNTIME_desc_flush( C, sequence );
- morse_sequence_wait(morse, sequence);
+ morse_sequence_create( morse, &sequence );
+
+ MORSE_zunmlq_Tile_Async(side, trans, A, T, C, sequence, &request );
+ MORSE_Desc_Flush( A, sequence );
+ MORSE_Desc_Flush( T, sequence );
+ MORSE_Desc_Flush( C, sequence );
+
+ morse_sequence_wait( morse, sequence );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -306,9 +304,9 @@ int MORSE_zunmlq_Tile(MORSE_enum side, MORSE_enum trans,
* @sa MORSE_zgelqf_Tile_Async
*
******************************************************************************/
-int MORSE_zunmlq_Tile_Async(MORSE_enum side, MORSE_enum trans,
+int MORSE_zunmlq_Tile_Async( MORSE_enum side, MORSE_enum trans,
MORSE_desc_t *A, MORSE_desc_t *T, MORSE_desc_t *C,
- MORSE_sequence_t *sequence, MORSE_request_t *request)
+ MORSE_sequence_t *sequence, MORSE_request_t *request )
{
MORSE_context_t *morse;
MORSE_desc_t D, *Dptr = NULL;
@@ -327,10 +325,12 @@ int MORSE_zunmlq_Tile_Async(MORSE_enum side, MORSE_enum trans,
return MORSE_ERR_UNALLOCATED;
}
/* Check sequence status */
- if (sequence->status == MORSE_SUCCESS)
+ if (sequence->status == MORSE_SUCCESS) {
request->status = MORSE_SUCCESS;
- else
+ }
+ else {
return morse_request_fail(sequence, request, MORSE_ERR_SEQUENCE_FLUSHED);
+ }
/* Check descriptors for correctness */
if (morse_desc_check(A) != MORSE_SUCCESS) {
@@ -358,10 +358,10 @@ int MORSE_zunmlq_Tile_Async(MORSE_enum side, MORSE_enum trans,
}
/* Quick return - currently NOT equivalent to LAPACK's:
* CALL DLASET( 'Full', MAX( M, N ), NRHS, ZERO, ZERO, C, LDC ) */
-/*
- if (chameleon_min(M, chameleon_min(N, K)) == 0)
- return MORSE_SUCCESS;
-*/
+ /*
+ if (chameleon_min(M, chameleon_min(N, K)) == 0)
+ return MORSE_SUCCESS;
+ */
#if defined(CHAMELEON_COPY_DIAG)
{
int m = chameleon_min(A->mt, A->nt) * A->mb;
@@ -373,16 +373,21 @@ int MORSE_zunmlq_Tile_Async(MORSE_enum side, MORSE_enum trans,
if (morse->householder == MORSE_FLAT_HOUSEHOLDER) {
if ( (trans == MorseConjTrans) &&
(side == MorseLeft) ) {
- morse_pzunmlq(side, trans, A, C, T, Dptr, sequence, request);
+ morse_pzunmlq( side, trans, A, C, T, Dptr, sequence, request );
} else {
- morse_pzunmlq(side, trans, A, C, T, Dptr, sequence, request);
+ morse_pzunmlq( side, trans, A, C, T, Dptr, sequence, request );
}
}
else {
- morse_pzunmlqrh(side, trans, A, C, T, Dptr, MORSE_RHBLK, sequence, request);
+ morse_pzunmlqrh( side, trans, A, C, T, Dptr, MORSE_RHBLK, sequence, request );
}
if (Dptr != NULL) {
- morse_desc_mat_free(Dptr);
+ MORSE_Desc_Flush( A, sequence );
+ MORSE_Desc_Flush( C, sequence );
+ MORSE_Desc_Flush( T, sequence );
+ MORSE_Desc_Flush( Dptr, sequence );
+ morse_sequence_wait( morse, sequence );
+ morse_desc_mat_free( Dptr );
}
(void)D;
return MORSE_SUCCESS;
diff --git a/compute/zunmlq_param.c b/compute/zunmlq_param.c
index b6f746d02e23a78cbfdd0853a34953d6d04fdac5..4a637513666d3b5f5485712645b7cf32d061e6c0 100644
--- a/compute/zunmlq_param.c
+++ b/compute/zunmlq_param.c
@@ -104,17 +104,18 @@
* @sa MORSE_zgelqf
*
******************************************************************************/
-int MORSE_zunmlq_param(const libhqr_tree_t *qrtree, MORSE_enum side, MORSE_enum trans, int M, int N, int K,
- MORSE_Complex64_t *A, int LDA,
- MORSE_desc_t *descTS, MORSE_desc_t *descTT,
- MORSE_Complex64_t *C, int LDC)
+int MORSE_zunmlq_param( const libhqr_tree_t *qrtree, MORSE_enum side, MORSE_enum trans, int M, int N, int K,
+ MORSE_Complex64_t *A, int LDA,
+ MORSE_desc_t *descTS, MORSE_desc_t *descTT,
+ MORSE_Complex64_t *C, int LDC )
{
int NB, An;
int status;
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
MORSE_request_t request = MORSE_REQUEST_INITIALIZER;
- MORSE_desc_t descA, descC;
+ MORSE_desc_t descAl, descAt;
+ MORSE_desc_t descCl, descCt;
morse = morse_context_self();
if (morse == NULL) {
@@ -170,37 +171,31 @@ int MORSE_zunmlq_param(const libhqr_tree_t *qrtree, MORSE_enum side, MORSE_enum
/* Set MT, NT & NTRHS */
NB = MORSE_NB;
- morse_sequence_create(morse, &sequence);
+ morse_sequence_create( morse, &sequence );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooplap2tile( descA, A, NB, NB, LDA, An, 0, 0, K, An, sequence, &request,
- morse_desc_mat_free(&(descA)) );
- morse_zooplap2tile( descC, C, NB, NB, LDC, N, 0, 0, M, N, sequence, &request,
- morse_desc_mat_free(&(descA)); morse_desc_mat_free(&(descC)));
-/* } else {*/
-/* morse_ziplap2tile( descA, A, NB, NB, LDA, An, 0, 0, K, An,*/
-/* sequence, &request);*/
-/* morse_ziplap2tile( descC, C, NB, NB, LDC, N, 0, 0, M, N,*/
-/* sequence, &request);*/
-/* }*/
+ /* Submit the matrix conversion */
+ morse_zlap2tile( morse, &descAl, &descAt, MorseDescInput, MorseUpper,
+ A, NB, NB, LDA, An, K, An, sequence, &request );
+ morse_zlap2tile( morse, &descCl, &descCt, MorseDescInout, MorseUpperLower,
+ C, NB, NB, LDC, N, M, N, sequence, &request );
/* Call the tile interface */
- MORSE_zunmlq_param_Tile_Async(
- qrtree, side, trans, &descA, descTS, descTT, &descC, sequence, &request);
+ MORSE_zunmlq_param_Tile_Async( qrtree, side, trans, &descAt, descTS, descTT, &descCt, sequence, &request );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooptile2lap(descC, C, NB, NB, LDC, N, sequence, &request);
- morse_sequence_wait(morse, sequence);
- morse_desc_mat_free(&descA);
- morse_desc_mat_free(&descC);
-/* } else {*/
-/* morse_ziptile2lap( descA, A, NB, NB, LDA, An, sequence, &request);*/
-/* morse_ziptile2lap( descC, C, NB, NB, LDC, N, sequence, &request);*/
-/* morse_sequence_wait(morse, sequence);*/
-/* }*/
+ /* Submit the matrix conversion back */
+ morse_ztile2lap( morse, &descAl, &descAt,
+ MorseDescInput, MorseUpper, sequence, &request );
+ morse_ztile2lap( morse, &descCl, &descCt,
+ MorseDescInout, MorseUpperLower, sequence, &request );
+
+ morse_sequence_wait( morse, sequence );
+
+ /* Cleanup the temporary data */
+ morse_ztile2lap_cleanup( morse, &descAl, &descAt );
+ morse_ztile2lap_cleanup( morse, &descCl, &descCt );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -253,8 +248,8 @@ int MORSE_zunmlq_param(const libhqr_tree_t *qrtree, MORSE_enum side, MORSE_enum
* @sa MORSE_zgelqf_Tile
*
******************************************************************************/
-int MORSE_zunmlq_param_Tile(const libhqr_tree_t *qrtree, MORSE_enum side, MORSE_enum trans,
- MORSE_desc_t *A, MORSE_desc_t *TS, MORSE_desc_t *TT, MORSE_desc_t *C)
+int MORSE_zunmlq_param_Tile( const libhqr_tree_t *qrtree, MORSE_enum side, MORSE_enum trans,
+ MORSE_desc_t *A, MORSE_desc_t *TS, MORSE_desc_t *TT, MORSE_desc_t *C )
{
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
@@ -266,14 +261,18 @@ int MORSE_zunmlq_param_Tile(const libhqr_tree_t *qrtree, MORSE_enum side, MORSE_
morse_fatal_error("MORSE_zunmlq_param_Tile", "MORSE not initialized");
return MORSE_ERR_NOT_INITIALIZED;
}
- morse_sequence_create(morse, &sequence);
- MORSE_zunmlq_param_Tile_Async(qrtree, side, trans, A, TS, TT, C, sequence, &request);
- RUNTIME_desc_flush( A, sequence );
- RUNTIME_desc_flush( C, sequence );
- morse_sequence_wait(morse, sequence);
+ morse_sequence_create( morse, &sequence );
+
+ MORSE_zunmlq_param_Tile_Async( qrtree, side, trans, A, TS, TT, C, sequence, &request );
+ MORSE_Desc_Flush( A, sequence );
+ MORSE_Desc_Flush( TS, sequence );
+ MORSE_Desc_Flush( TT, sequence );
+ MORSE_Desc_Flush( C, sequence );
+
+ morse_sequence_wait( morse, sequence );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -305,9 +304,9 @@ int MORSE_zunmlq_param_Tile(const libhqr_tree_t *qrtree, MORSE_enum side, MORSE_
* @sa MORSE_zgelqf_Tile_Async
*
******************************************************************************/
-int MORSE_zunmlq_param_Tile_Async(const libhqr_tree_t *qrtree, MORSE_enum side, MORSE_enum trans,
- MORSE_desc_t *A, MORSE_desc_t *TS, MORSE_desc_t *TT, MORSE_desc_t *C,
- MORSE_sequence_t *sequence, MORSE_request_t *request)
+int MORSE_zunmlq_param_Tile_Async( const libhqr_tree_t *qrtree, MORSE_enum side, MORSE_enum trans,
+ MORSE_desc_t *A, MORSE_desc_t *TS, MORSE_desc_t *TT, MORSE_desc_t *C,
+ MORSE_sequence_t *sequence, MORSE_request_t *request )
{
MORSE_context_t *morse;
MORSE_desc_t D, *Dptr = NULL;
@@ -326,10 +325,12 @@ int MORSE_zunmlq_param_Tile_Async(const libhqr_tree_t *qrtree, MORSE_enum side,
return MORSE_ERR_UNALLOCATED;
}
/* Check sequence status */
- if (sequence->status == MORSE_SUCCESS)
+ if (sequence->status == MORSE_SUCCESS) {
request->status = MORSE_SUCCESS;
- else
+ }
+ else {
return morse_request_fail(sequence, request, MORSE_ERR_SEQUENCE_FLUSHED);
+ }
/* Check descriptors for correctness */
if (morse_desc_check(A) != MORSE_SUCCESS) {
@@ -361,10 +362,10 @@ int MORSE_zunmlq_param_Tile_Async(const libhqr_tree_t *qrtree, MORSE_enum side,
}
/* Quick return - currently NOT equivalent to LAPACK's:
* CALL DLASET( 'Full', MAX( M, N ), NRHS, ZERO, ZERO, C, LDC ) */
-/*
- if (chameleon_min(M, chameleon_min(N, K)) == 0)
- return MORSE_SUCCESS;
-*/
+ /*
+ if (chameleon_min(M, chameleon_min(N, K)) == 0)
+ return MORSE_SUCCESS;
+ */
#if defined(CHAMELEON_COPY_DIAG)
{
@@ -374,10 +375,16 @@ int MORSE_zunmlq_param_Tile_Async(const libhqr_tree_t *qrtree, MORSE_enum side,
}
#endif
- morse_pzunmlq_param(qrtree, side, trans, A, C, TS, TT, Dptr, sequence, request);
+ morse_pzunmlq_param( qrtree, side, trans, A, C, TS, TT, Dptr, sequence, request );
if (Dptr != NULL) {
- morse_desc_mat_free(Dptr);
+ MORSE_Desc_Flush( A, sequence );
+ MORSE_Desc_Flush( C, sequence );
+ MORSE_Desc_Flush( TS, sequence );
+ MORSE_Desc_Flush( TT, sequence );
+ MORSE_Desc_Flush( Dptr, sequence );
+ morse_sequence_wait( morse, sequence );
+ morse_desc_mat_free( Dptr );
}
(void)D;
return MORSE_SUCCESS;
diff --git a/compute/zunmqr.c b/compute/zunmqr.c
index 4464aaecbe90f5d6bbf86bf4c644aa89ccbe87d5..7e836f91f1ff8e026496681e9db7d5855aa65e6f 100644
--- a/compute/zunmqr.c
+++ b/compute/zunmqr.c
@@ -107,17 +107,18 @@
* @sa MORSE_zgeqrf
*
******************************************************************************/
-int MORSE_zunmqr(MORSE_enum side, MORSE_enum trans, int M, int N, int K,
+int MORSE_zunmqr( MORSE_enum side, MORSE_enum trans, int M, int N, int K,
MORSE_Complex64_t *A, int LDA,
MORSE_desc_t *descT,
- MORSE_Complex64_t *C, int LDC)
+ MORSE_Complex64_t *C, int LDC )
{
int NB, Am;
int status;
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
MORSE_request_t request = MORSE_REQUEST_INITIALIZER;
- MORSE_desc_t descA, descC;
+ MORSE_desc_t descAl, descAt;
+ MORSE_desc_t descCl, descCt;
morse = morse_context_self();
if (morse == NULL) {
@@ -174,37 +175,31 @@ int MORSE_zunmqr(MORSE_enum side, MORSE_enum trans, int M, int N, int K,
/* Set MT, NT & NTRHS */
NB = MORSE_NB;
- morse_sequence_create(morse, &sequence);
+ morse_sequence_create( morse, &sequence );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooplap2tile( descA, A, NB, NB, LDA, K, 0, 0, Am, K, sequence, &request,
- morse_desc_mat_free(&(descA)) );
- morse_zooplap2tile( descC, C, NB, NB, LDC, N, 0, 0, M, N, sequence, &request,
- morse_desc_mat_free(&(descA)); morse_desc_mat_free(&(descC)));
-/* } else {*/
-/* morse_ziplap2tile( descA, A, NB, NB, LDA, K, 0, 0, Am, K,*/
-/* sequence, &request);*/
-/* morse_ziplap2tile( descC, C, NB, NB, LDC, N, 0, 0, M, N,*/
-/* sequence, &request);*/
-/* }*/
+ /* Submit the matrix conversion */
+ morse_zlap2tile( morse, &descAl, &descAt, MorseDescInput, MorseLower,
+ A, NB, NB, LDA, K, Am, K, sequence, &request );
+ morse_zlap2tile( morse, &descCl, &descCt, MorseDescInout, MorseUpperLower,
+ C, NB, NB, LDC, N, M, N, sequence, &request );
/* Call the tile interface */
- MORSE_zunmqr_Tile_Async(
- side, trans, &descA, descT, &descC, sequence, &request);
+ MORSE_zunmqr_Tile_Async( side, trans, &descAt, descT, &descCt, sequence, &request );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooptile2lap(descC, C, NB, NB, LDC, N, sequence, &request);
- morse_sequence_wait(morse, sequence);
- morse_desc_mat_free(&descA);
- morse_desc_mat_free(&descC);
-/* } else {*/
-/* morse_ziptile2lap( descA, A, NB, NB, LDA, K, sequence, &request);*/
-/* morse_ziptile2lap( descC, C, NB, NB, LDC, N, sequence, &request);*/
-/* morse_sequence_wait(morse, sequence);*/
-/* }*/
+ /* Submit the matrix conversion back */
+ morse_ztile2lap( morse, &descAl, &descAt,
+ MorseDescInput, MorseLower, sequence, &request );
+ morse_ztile2lap( morse, &descCl, &descCt,
+ MorseDescInout, MorseUpperLower, sequence, &request );
+
+ morse_sequence_wait( morse, sequence );
+
+ /* Cleanup the temporary data */
+ morse_ztile2lap_cleanup( morse, &descAl, &descAt );
+ morse_ztile2lap_cleanup( morse, &descCl, &descCt );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -258,8 +253,8 @@ int MORSE_zunmqr(MORSE_enum side, MORSE_enum trans, int M, int N, int K,
* @sa MORSE_zgeqrf_Tile
*
******************************************************************************/
-int MORSE_zunmqr_Tile(MORSE_enum side, MORSE_enum trans,
- MORSE_desc_t *A, MORSE_desc_t *T, MORSE_desc_t *C)
+int MORSE_zunmqr_Tile( MORSE_enum side, MORSE_enum trans,
+ MORSE_desc_t *A, MORSE_desc_t *T, MORSE_desc_t *C )
{
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
@@ -271,14 +266,17 @@ int MORSE_zunmqr_Tile(MORSE_enum side, MORSE_enum trans,
morse_fatal_error("MORSE_zunmqr_Tile", "MORSE not initialized");
return MORSE_ERR_NOT_INITIALIZED;
}
- morse_sequence_create(morse, &sequence);
- MORSE_zunmqr_Tile_Async(side, trans, A, T, C, sequence, &request);
- RUNTIME_desc_flush( A, sequence );
- RUNTIME_desc_flush( C, sequence );
- morse_sequence_wait(morse, sequence);
+ morse_sequence_create( morse, &sequence );
+
+ MORSE_zunmqr_Tile_Async(side, trans, A, T, C, sequence, &request );
+ MORSE_Desc_Flush( A, sequence );
+ MORSE_Desc_Flush( T, sequence );
+ MORSE_Desc_Flush( C, sequence );
+
+ morse_sequence_wait( morse, sequence );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -310,9 +308,9 @@ int MORSE_zunmqr_Tile(MORSE_enum side, MORSE_enum trans,
* @sa MORSE_zgeqrf_Tile_Async
*
******************************************************************************/
-int MORSE_zunmqr_Tile_Async(MORSE_enum side, MORSE_enum trans,
+int MORSE_zunmqr_Tile_Async( MORSE_enum side, MORSE_enum trans,
MORSE_desc_t *A, MORSE_desc_t *T, MORSE_desc_t *C,
- MORSE_sequence_t *sequence, MORSE_request_t *request)
+ MORSE_sequence_t *sequence, MORSE_request_t *request )
{
MORSE_context_t *morse;
MORSE_desc_t D, *Dptr = NULL;
@@ -331,10 +329,12 @@ int MORSE_zunmqr_Tile_Async(MORSE_enum side, MORSE_enum trans,
return MORSE_ERR_UNALLOCATED;
}
/* Check sequence status */
- if (sequence->status == MORSE_SUCCESS)
+ if (sequence->status == MORSE_SUCCESS) {
request->status = MORSE_SUCCESS;
- else
+ }
+ else {
return morse_request_fail(sequence, request, MORSE_ERR_SEQUENCE_FLUSHED);
+ }
/* Check descriptors for correctness */
if (morse_desc_check(A) != MORSE_SUCCESS) {
@@ -362,10 +362,10 @@ int MORSE_zunmqr_Tile_Async(MORSE_enum side, MORSE_enum trans,
}
/* Quick return - currently NOT equivalent to LAPACK's:
* CALL DLASET( 'Full', MAX( M, N ), NRHS, ZERO, ZERO, C, LDC ) */
-/*
- if (chameleon_min(M, chameleon_min(N, K)) == 0)
- return MORSE_SUCCESS;
-*/
+ /*
+ if (chameleon_min(M, chameleon_min(N, K)) == 0)
+ return MORSE_SUCCESS;
+ */
#if defined(CHAMELEON_COPY_DIAG)
{
@@ -375,21 +375,26 @@ int MORSE_zunmqr_Tile_Async(MORSE_enum side, MORSE_enum trans,
}
#endif
- if (morse->householder == MORSE_FLAT_HOUSEHOLDER) {
+ if (morse->householder == MORSE_FLAT_HOUSEHOLDER) {
if ( (trans == MorseConjTrans) &&
(side == MorseLeft) ) {
- morse_pzunmqr(side, trans, A, C, T, Dptr, sequence, request);
+ morse_pzunmqr( side, trans, A, C, T, Dptr, sequence, request );
}
else {
- morse_pzunmqr(side, trans, A, C, T, Dptr, sequence, request);
+ morse_pzunmqr( side, trans, A, C, T, Dptr, sequence, request );
}
}
else {
- morse_pzunmqrrh(side, trans, A, C, T, Dptr, MORSE_RHBLK, sequence, request);
+ morse_pzunmqrrh( side, trans, A, C, T, Dptr, MORSE_RHBLK, sequence, request );
}
if (Dptr != NULL) {
- morse_desc_mat_free(Dptr);
+ MORSE_Desc_Flush( A, sequence );
+ MORSE_Desc_Flush( C, sequence );
+ MORSE_Desc_Flush( T, sequence );
+ MORSE_Desc_Flush( Dptr, sequence );
+ morse_sequence_wait( morse, sequence );
+ morse_desc_mat_free( Dptr );
}
(void)D;
return MORSE_SUCCESS;
diff --git a/compute/zunmqr_param.c b/compute/zunmqr_param.c
index d285caabf0ce422062a6b621355c85bf309c4de4..45c88e5750f2c250327db48c6a6d2bc714faec2b 100644
--- a/compute/zunmqr_param.c
+++ b/compute/zunmqr_param.c
@@ -107,18 +107,19 @@
* @sa MORSE_zgeqrf
*
******************************************************************************/
-int MORSE_zunmqr_param(const libhqr_tree_t *qrtree,
- MORSE_enum side, MORSE_enum trans, int M, int N, int K,
- MORSE_Complex64_t *A, int LDA,
- MORSE_desc_t *descTS, MORSE_desc_t *descTT,
- MORSE_Complex64_t *C, int LDC)
+int MORSE_zunmqr_param( const libhqr_tree_t *qrtree,
+ MORSE_enum side, MORSE_enum trans, int M, int N, int K,
+ MORSE_Complex64_t *A, int LDA,
+ MORSE_desc_t *descTS, MORSE_desc_t *descTT,
+ MORSE_Complex64_t *C, int LDC )
{
int NB, Am;
int status;
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
MORSE_request_t request = MORSE_REQUEST_INITIALIZER;
- MORSE_desc_t descA, descC;
+ MORSE_desc_t descAl, descAt;
+ MORSE_desc_t descCl, descCt;
morse = morse_context_self();
if (morse == NULL) {
@@ -175,37 +176,31 @@ int MORSE_zunmqr_param(const libhqr_tree_t *qrtree,
/* Set MT, NT & NTRHS */
NB = MORSE_NB;
- morse_sequence_create(morse, &sequence);
+ morse_sequence_create( morse, &sequence );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooplap2tile( descA, A, NB, NB, LDA, K, 0, 0, Am, K, sequence, &request,
- morse_desc_mat_free(&(descA)) );
- morse_zooplap2tile( descC, C, NB, NB, LDC, N, 0, 0, M, N, sequence, &request,
- morse_desc_mat_free(&(descA)); morse_desc_mat_free(&(descC)));
-/* } else {*/
-/* morse_ziplap2tile( descA, A, NB, NB, LDA, K, 0, 0, Am, K,*/
-/* sequence, &request);*/
-/* morse_ziplap2tile( descC, C, NB, NB, LDC, N, 0, 0, M, N,*/
-/* sequence, &request);*/
-/* }*/
+ /* Submit the matrix conversion */
+ morse_zlap2tile( morse, &descAl, &descAt, MorseDescInput, MorseLower,
+ A, NB, NB, LDA, K, Am, K, sequence, &request );
+ morse_zlap2tile( morse, &descCl, &descCt, MorseDescInout, MorseUpperLower,
+ C, NB, NB, LDC, N, M, N, sequence, &request );
/* Call the tile interface */
- MORSE_zunmqr_param_Tile_Async(
- qrtree, side, trans, &descA, descTS, descTT, &descC, sequence, &request);
+ MORSE_zunmqr_param_Tile_Async( qrtree, side, trans, &descAt, descTS, descTT, &descCt, sequence, &request );
-/* if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {*/
- morse_zooptile2lap(descC, C, NB, NB, LDC, N, sequence, &request);
- morse_sequence_wait(morse, sequence);
- morse_desc_mat_free(&descA);
- morse_desc_mat_free(&descC);
-/* } else {*/
-/* morse_ziptile2lap( descA, A, NB, NB, LDA, K, sequence, &request);*/
-/* morse_ziptile2lap( descC, C, NB, NB, LDC, N, sequence, &request);*/
-/* morse_sequence_wait(morse, sequence);*/
-/* }*/
+ /* Submit the matrix conversion back */
+ morse_ztile2lap( morse, &descAl, &descAt,
+ MorseDescInput, MorseLower, sequence, &request );
+ morse_ztile2lap( morse, &descCl, &descCt,
+ MorseDescInout, MorseUpperLower, sequence, &request );
+
+ morse_sequence_wait( morse, sequence );
+
+ /* Cleanup the temporary data */
+ morse_ztile2lap_cleanup( morse, &descAl, &descAt );
+ morse_ztile2lap_cleanup( morse, &descCl, &descCt );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -259,8 +254,8 @@ int MORSE_zunmqr_param(const libhqr_tree_t *qrtree,
* @sa MORSE_zgeqrf_Tile
*
******************************************************************************/
-int MORSE_zunmqr_param_Tile(const libhqr_tree_t *qrtree, MORSE_enum side, MORSE_enum trans,
- MORSE_desc_t *A, MORSE_desc_t *TS, MORSE_desc_t *TT, MORSE_desc_t *C)
+int MORSE_zunmqr_param_Tile( const libhqr_tree_t *qrtree, MORSE_enum side, MORSE_enum trans,
+ MORSE_desc_t *A, MORSE_desc_t *TS, MORSE_desc_t *TT, MORSE_desc_t *C )
{
MORSE_context_t *morse;
MORSE_sequence_t *sequence = NULL;
@@ -272,14 +267,18 @@ int MORSE_zunmqr_param_Tile(const libhqr_tree_t *qrtree, MORSE_enum side, MORSE_
morse_fatal_error("MORSE_zunmqr_param_Tile", "MORSE not initialized");
return MORSE_ERR_NOT_INITIALIZED;
}
- morse_sequence_create(morse, &sequence);
- MORSE_zunmqr_param_Tile_Async(qrtree, side, trans, A, TS, TT, C, sequence, &request);
- RUNTIME_desc_flush( A, sequence );
- RUNTIME_desc_flush( C, sequence );
- morse_sequence_wait(morse, sequence);
+ morse_sequence_create( morse, &sequence );
+
+ MORSE_zunmqr_param_Tile_Async( qrtree, side, trans, A, TS, TT, C, sequence, &request );
+
+ MORSE_Desc_Flush( A, sequence );
+ MORSE_Desc_Flush( TS, sequence );
+ MORSE_Desc_Flush( TT, sequence );
+ MORSE_Desc_Flush( C, sequence );
+ morse_sequence_wait( morse, sequence );
status = sequence->status;
- morse_sequence_destroy(morse, sequence);
+ morse_sequence_destroy( morse, sequence );
return status;
}
@@ -311,10 +310,10 @@ int MORSE_zunmqr_param_Tile(const libhqr_tree_t *qrtree, MORSE_enum side, MORSE_
* @sa MORSE_zgeqrf_Tile_Async
*
******************************************************************************/
-int MORSE_zunmqr_param_Tile_Async(const libhqr_tree_t *qrtree,
- MORSE_enum side, MORSE_enum trans,
- MORSE_desc_t *A, MORSE_desc_t *TS, MORSE_desc_t *TT, MORSE_desc_t *C,
- MORSE_sequence_t *sequence, MORSE_request_t *request)
+int MORSE_zunmqr_param_Tile_Async( const libhqr_tree_t *qrtree,
+ MORSE_enum side, MORSE_enum trans,
+ MORSE_desc_t *A, MORSE_desc_t *TS, MORSE_desc_t *TT, MORSE_desc_t *C,
+ MORSE_sequence_t *sequence, MORSE_request_t *request )
{
MORSE_context_t *morse;
MORSE_desc_t D, *Dptr = NULL;
@@ -333,10 +332,12 @@ int MORSE_zunmqr_param_Tile_Async(const libhqr_tree_t *qrtree,
return MORSE_ERR_UNALLOCATED;
}
/* Check sequence status */
- if (sequence->status == MORSE_SUCCESS)
+ if (sequence->status == MORSE_SUCCESS) {
request->status = MORSE_SUCCESS;
- else
+ }
+ else {
return morse_request_fail(sequence, request, MORSE_ERR_SEQUENCE_FLUSHED);
+ }
/* Check descriptors for correctness */
if (morse_desc_check(A) != MORSE_SUCCESS) {
@@ -368,10 +369,10 @@ int MORSE_zunmqr_param_Tile_Async(const libhqr_tree_t *qrtree,
}
/* Quick return - currently NOT equivalent to LAPACK's:
* CALL DLASET( 'Full', MAX( M, N ), NRHS, ZERO, ZERO, C, LDC ) */
-/*
- if (chameleon_min(M, chameleon_min(N, K)) == 0)
- return MORSE_SUCCESS;
-*/
+ /*
+ if (chameleon_min(M, chameleon_min(N, K)) == 0)
+ return MORSE_SUCCESS;
+ */
#if defined(CHAMELEON_COPY_DIAG)
{
@@ -381,10 +382,16 @@ int MORSE_zunmqr_param_Tile_Async(const libhqr_tree_t *qrtree,
}
#endif
- morse_pzunmqr_param(qrtree, side, trans, A, C, TS, TT, Dptr, sequence, request);
+ morse_pzunmqr_param( qrtree, side, trans, A, C, TS, TT, Dptr, sequence, request );
if (Dptr != NULL) {
- morse_desc_mat_free(Dptr);
+ MORSE_Desc_Flush( A, sequence );
+ MORSE_Desc_Flush( C, sequence );
+ MORSE_Desc_Flush( TS, sequence );
+ MORSE_Desc_Flush( TT, sequence );
+ MORSE_Desc_Flush( Dptr, sequence );
+ morse_sequence_wait( morse, sequence );
+ morse_desc_mat_free( Dptr );
}
(void)D;
return MORSE_SUCCESS;
diff --git a/control/CMakeLists.txt b/control/CMakeLists.txt
index 0fd19f58836468a32c90843685864374aa455d46..5cc1b6b04e00f3351d4d80b872c6b3ed7f121d59 100644
--- a/control/CMakeLists.txt
+++ b/control/CMakeLists.txt
@@ -50,7 +50,6 @@ set(CHAMELEON_HDRS
global.h
morse_f77.h
morsewinthread.h
- tile.h
workspace.h
)
diff --git a/control/common.h b/control/common.h
index 81bce72c4f73839e0f48a2bcfaba627df088d34f..1df9385bafadfcc64fde9eb5a603035ad1e67915 100644
--- a/control/common.h
+++ b/control/common.h
@@ -78,7 +78,6 @@
#include "control/auxiliary.h"
#include "control/context.h"
#include "control/descriptor.h"
-#include "control/tile.h"
#include "control/async.h"
/*******************************************************************************
@@ -108,7 +107,6 @@
#define MorseIPT_Panel 1
#define MorseIPT_All 2
-
/*******************************************************************************
* Global array of LAPACK constants
**/
diff --git a/control/compute_z.h b/control/compute_z.h
index 2ac2d5090918260460aad57952c551f59ff4e433..168277b4be24bcdaf37024488cfc93aef9e0cfb1 100644
--- a/control/compute_z.h
+++ b/control/compute_z.h
@@ -28,14 +28,22 @@
*
**/
+/*******************************************************************************
+ * LAPACK/Tile Descriptor accesses
+ **/
+#define MorseDescInput 1
+#define MorseDescOutput 2
+#define MorseDescInout (MorseDescInput | MorseDescOutput)
+
/***************************************************************************//**
* Macro for matrix conversion / Lapack interface
**/
-#define morse_zdesc_alloc_diag(descA, mb, nb, lm, ln, i, j, m, n, p, q) \
+#define morse_zdesc_alloc_diag( descA, mb, nb, lm, ln, i, j, m, n, p, q) \
descA = morse_desc_init_diag( \
MorseComplexDouble, (mb), (nb), ((mb)*(nb)), \
(m), (n), (i), (j), (m), (n), p, q); \
- morse_desc_mat_alloc( &(descA) );
+ morse_desc_mat_alloc( &(descA) ); \
+ RUNTIME_desc_create( &(descA) );
#define morse_zdesc_alloc( descA, mb, nb, lm, ln, i, j, m, n, free) \
descA = morse_desc_init( \
@@ -45,34 +53,8 @@
morse_error( __func__, "morse_desc_mat_alloc() failed"); \
{free;}; \
return MORSE_ERR_OUT_OF_RESOURCES; \
- }
-
-#define morse_zooplap2tile( descA, A, mb, nb, lm, ln, i, j, m, n, seq, req, free) \
- descA = morse_desc_init( \
- MorseComplexDouble, (mb), (nb), ((mb)*(nb)), \
- (lm), (ln), (i), (j), (m), (n), 1, 1); \
- if ( morse_desc_mat_alloc( &(descA) ) ) { \
- morse_error( __func__, "morse_desc_mat_alloc() failed"); \
- {free;}; \
- return MORSE_ERR_OUT_OF_RESOURCES; \
} \
- morse_pzlapack_to_tile(A, lm, &descA, seq, req);
-
-#define morse_ziplap2tile( descA, A, mb, nb, lm, ln, i, j, m, n, seq, req) \
- descA = morse_desc_init( \
- MorseComplexDouble, (mb), (nb), ((mb)*(nb)), \
- (lm), (ln), (i), (j), (m), (n), 1, 1); \
- descA.mat = A; \
- MORSE_zgecfi_Async((lm), (ln), (A), MorseCM, (mb), (nb), \
- MorseCCRB, (mb), (nb), (seq), (req));
-
-
-#define morse_zooptile2lap( descA, A, mb, nb, lm, ln, seq, req) \
- morse_pztile_to_lapack(&descA, A, lm, seq, req);
-
-#define morse_ziptile2lap( descA, A, mb, nb, lm, ln, seq, req) \
- MORSE_zgecfi_Async((lm), (ln), (A), MorseCCRB, (mb), (nb), \
- MorseCM, (mb), (nb), (seq), (req));
+ RUNTIME_desc_create( &(descA) );
/***************************************************************************//**
* Declarations of internal sequential functions
@@ -171,3 +153,98 @@ void morse_pzunglq_param(const libhqr_tree_t *qrtree, MORSE_desc_t *A, MORSE_des
void morse_pzungqr_param(const libhqr_tree_t *qrtree, MORSE_desc_t *A, MORSE_desc_t *Q,
MORSE_desc_t *TS, MORSE_desc_t *TT, MORSE_desc_t *D,
MORSE_sequence_t *sequence, MORSE_request_t *request);
+
+
+/**
+ * @brief Internal function to convert the lapack format to tile format in
+ * LAPACK interface calls
+ */
+static inline int
+morse_zlap2tile( MORSE_context_t *morse,
+ MORSE_desc_t *descAl, MORSE_desc_t *descAt,
+ MORSE_enum mode, MORSE_enum uplo,
+ MORSE_Complex64_t *A, int mb, int nb, int lm, int ln, int m, int n,
+ MORSE_sequence_t *seq, MORSE_request_t *req )
+{
+ /* Initialize the Lapack descriptor */
+ *descAl = morse_desc_init_user( MorseComplexDouble, mb, nb, (mb)*(nb),
+ lm, ln, 0, 0, m, n, 1, 1,
+ morse_getaddr_cm, morse_getblkldd_cm, NULL );
+ descAl->mat = A;
+ descAl->styp = MorseCM;
+
+ /* Initialize the tile descriptor */
+ *descAt = morse_desc_init( MorseComplexDouble, mb, nb, (mb)*(nb),
+ lm, ln, 0, 0, m, n, 1, 1 );
+
+ if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {
+ if ( morse_desc_mat_alloc( descAt ) ) {
+ morse_error( "morse_zlap2tile", "morse_desc_mat_alloc() failed");
+ return MORSE_ERR_OUT_OF_RESOURCES;
+ }
+
+ RUNTIME_desc_create( descAl );
+ RUNTIME_desc_create( descAt );
+
+ if ( mode & MorseDescInput ) {
+ morse_pzlacpy( uplo, descAl, descAt, seq, req );
+ }
+ }
+ else {
+ morse_fatal_error( "morse_zlap2tile", "INPLACE translation not supported yet");
+ descAt->mat = A;
+
+ RUNTIME_desc_create( descAl );
+ RUNTIME_desc_create( descAt );
+
+ if ( mode & MorseDescInput ) {
+ /* MORSE_zgecfi_Async( lm, ln, A, MorseCM, mb, nb, */
+ /* MorseCCRB, mb, nb, seq, req ); */
+ }
+ return MORSE_ERR_NOT_SUPPORTED;
+ }
+
+ return MORSE_SUCCESS;
+}
+
+/**
+ * @brief Internal function to convert back the tile format to the lapack format
+ * in LAPACK interface calls
+ */
+static inline int
+morse_ztile2lap( MORSE_context_t *morse, MORSE_desc_t *descAl, MORSE_desc_t *descAt,
+ MORSE_enum mode, MORSE_enum uplo, MORSE_sequence_t *seq, MORSE_request_t *req )
+{
+ if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {
+ if ( mode & MorseDescOutput ) {
+ morse_pzlacpy( uplo, descAt, descAl, seq, req );
+ }
+ }
+ else {
+ morse_fatal_error( "morse_ztile2lap", "INPLACE translation not supported yet");
+ if ( mode & MorseDescOutput ) {
+ /* MORSE_zgecfi_Async( descAl->lm, descAl->ln, descAl->mat, */
+ /* MorseCCRB, descAl->mb, descAl->nb, */
+ /* MorseCM, descAl->mb, descAl->nb, seq, req ); */
+ }
+ return MORSE_ERR_NOT_SUPPORTED;
+ }
+ RUNTIME_desc_flush( descAl, seq );
+ RUNTIME_desc_flush( descAt, seq );
+
+ return MORSE_SUCCESS;
+}
+
+/**
+ * @brief Internal function to cleanup the temporary data from the layout
+ * conversions in LAPACK interface calls
+ */
+static inline void
+morse_ztile2lap_cleanup( MORSE_context_t *morse, MORSE_desc_t *descAl, MORSE_desc_t *descAt )
+{
+ if ( MORSE_TRANSLATION == MORSE_OUTOFPLACE ) {
+ morse_desc_mat_free( descAt );
+ }
+ RUNTIME_desc_destroy( descAl );
+ RUNTIME_desc_destroy( descAt );
+}
diff --git a/control/descriptor.c b/control/descriptor.c
index de7d216bdda0750193727395791c2fe9d3631479..2459329895893f8f9435f4433db6c35df3c1ba83 100644
--- a/control/descriptor.c
+++ b/control/descriptor.c
@@ -258,7 +258,7 @@ MORSE_desc_t* morse_desc_submatrix(MORSE_desc_t *descA, int i, int j, int m, int
/*******************************************************************************
* Check for descriptor correctness
**/
-int morse_desc_check(MORSE_desc_t *desc)
+int morse_desc_check(const MORSE_desc_t *desc)
{
if (desc == NULL) {
morse_error("morse_desc_check", "NULL descriptor");
@@ -307,7 +307,6 @@ int morse_desc_check(MORSE_desc_t *desc)
**/
int morse_desc_mat_alloc( MORSE_desc_t *desc )
{
-
size_t size = (size_t)(desc->llm) * (size_t)(desc->lln)
* (size_t)MORSE_Element_Size(desc->dtyp);
if ((desc->mat = RUNTIME_malloc(size)) == NULL) {
@@ -317,7 +316,6 @@ int morse_desc_mat_alloc( MORSE_desc_t *desc )
/* The matrix has already been registered by the Runtime alloc */
desc->register_mat = 0;
- RUNTIME_desc_create(desc);
return MORSE_SUCCESS;
}
@@ -327,8 +325,6 @@ int morse_desc_mat_alloc( MORSE_desc_t *desc )
**/
int morse_desc_mat_free( MORSE_desc_t *desc )
{
- RUNTIME_desc_destroy( desc );
-
if ( (desc->mat != NULL) &&
(desc->use_mat == 1 ) &&
(desc->alloc_mat == 1 ) )
@@ -751,6 +747,7 @@ int MORSE_Desc_Destroy(MORSE_desc_t **desc)
return MORSE_ERR_UNALLOCATED;
}
+ RUNTIME_desc_destroy( *desc );
morse_desc_mat_free( *desc );
free(*desc);
*desc = NULL;
diff --git a/control/descriptor.h b/control/descriptor.h
index 2f6c6afe937bfe04fa8d93f6f4c4a26154d4ece4..37cb2b19a026eabd6e740673b8b6a2416a37bbce 100644
--- a/control/descriptor.h
+++ b/control/descriptor.h
@@ -63,7 +63,7 @@ MORSE_desc_t morse_desc_init_user(MORSE_enum dtyp, int mb, int nb, int bsiz,
int (*get_rankof)( const MORSE_desc_t*, int, int ));
MORSE_desc_t* morse_desc_submatrix(MORSE_desc_t *descA, int i, int j, int m, int n);
-int morse_desc_check (MORSE_desc_t *desc);
+int morse_desc_check (const MORSE_desc_t *desc);
int morse_desc_mat_alloc(MORSE_desc_t *desc);
int morse_desc_mat_free (MORSE_desc_t *desc);
diff --git a/control/morse_zf77.c b/control/morse_zf77.c
index 8ea877f355cf352aebc047f2d8d5a65b79a47501..c9013dcc4c3675709ddb2cc4dd690da3ee1cc85e 100644
--- a/control/morse_zf77.c
+++ b/control/morse_zf77.c
@@ -40,9 +40,7 @@
* FORTRAN API - math functions (simple interface)
**/
#define MORSE_ZLAPACK_TO_TILE MORSE_FNAME(zlapack_to_tile, ZLAPACK_TO_TILE)
-#define MORSE_ZLAPACK_TO_TILE_ASYNC MORSE_FNAME(zlapack_to_tile_async, ZLAPACK_TO_TILE_ASYNC)
#define MORSE_ZTILE_TO_LAPACK MORSE_FNAME(ztile_to_lapack, ZTILE_TO_LAPACK)
-#define MORSE_ZTILE_TO_LAPACK_ASYNC MORSE_FNAME(ztile_to_lapack_async, ZTILE_TO_LAPACK_ASYNC)
//#define MORSE_ZGEBRD MORSE_FNAME(zgebrd , ZGEBRD )
//#define MORSE_ZGECFI MORSE_FNAME(zgecfi , ZGECFI )
//#define MORSE_ZGECFI_ASYNC MORSE_FNAME(zgecfi_async , ZGECFI_ASYNC )
@@ -293,15 +291,9 @@
void MORSE_ZLAPACK_TO_TILE(MORSE_Complex64_t *Af77, int *LDA, MORSE_desc_t *A, int *info)
{ *info = MORSE_zLapack_to_Tile(Af77, *LDA, A); }
-void MORSE_ZLAPACK_TO_TILE_ASYNC(MORSE_Complex64_t *Af77, int *LDA, MORSE_desc_t *A, MORSE_sequence_t *sequence, MORSE_request_t *request, int *info)
-{ *info = MORSE_zLapack_to_Tile_Async(Af77, *LDA, A, sequence, request); }
-
void MORSE_ZTILE_TO_LAPACK(MORSE_desc_t *A, MORSE_Complex64_t *Af77, int *LDA, int *info)
{ *info = MORSE_zTile_to_Lapack(A, Af77, *LDA); }
-void MORSE_ZTILE_TO_LAPACK_ASYNC(MORSE_desc_t *A, MORSE_Complex64_t *Af77, int *LDA, MORSE_sequence_t *sequence, MORSE_request_t *request, int *info)
-{ *info = MORSE_zTile_to_Lapack_Async(A, Af77, *LDA, sequence, request); }
-
//void MORSE_ZGEBRD(int *M, int *N, MORSE_Complex64_t *A, int *LDA, double *D, double *E, MORSE_desc_t **descT, int *info)
//{ *info = MORSE_zgebrd(*M, *N, A, *LDA, D, E, *descT); }
//
diff --git a/control/morse_zf90.F90 b/control/morse_zf90.F90
index 78c4631dd3c84fc1a76011e71865de0e4387184c..c9bf4035a99c4ac19d5db88427f5569126c67ab4 100644
--- a/control/morse_zf90.F90
+++ b/control/morse_zf90.F90
@@ -96,20 +96,6 @@ module morse_z
end function MORSE_zLapack_to_Tile_c
end interface
- interface
- function MORSE_zLapack_to_Tile_Async_c(Af77,LDA,A,sequence,request) &
- & bind(c, name='MORSE_zLapack_to_Tile_Async')
- use iso_c_binding
- implicit none
- integer(kind=c_int) :: MORSE_zLapack_to_Tile_Async_c
- type(c_ptr), value :: Af77
- integer(kind=c_int), value :: LDA
- type(c_ptr), value :: A
- type(c_ptr), value :: sequence
- type(c_ptr), value :: request
- end function MORSE_zLapack_to_Tile_Async_c
- end interface
-
interface
function MORSE_zTile_to_Lapack_c(A,Af77,LDA) &
& bind(c, name='MORSE_zTile_to_Lapack')
@@ -122,20 +108,6 @@ module morse_z
end function MORSE_zTile_to_Lapack_c
end interface
- interface
- function MORSE_zTile_to_Lapack_Async_c(A,Af77,LDA,sequence,request) &
- & bind(c, name='MORSE_zTile_to_Lapack_Async')
- use iso_c_binding
- implicit none
- integer(kind=c_int) :: MORSE_zTile_to_Lapack_Async_c
- type(c_ptr), value :: A
- type(c_ptr), value :: Af77
- integer(kind=c_int), value :: LDA
- type(c_ptr), value :: sequence
- type(c_ptr), value :: request
- end function MORSE_zTile_to_Lapack_Async_c
- end interface
-
interface
function MORSE_zgebrd_c(M,N,A,LDA,D,E,descT) &
& bind(c, name='MORSE_zgebrd')
@@ -5507,28 +5479,4 @@ module morse_z
info = MORSE_zTile_to_Lapack_c(A,c_loc(Af77),LDA)
end subroutine MORSE_zTile_to_Lapack
- subroutine MORSE_zLapack_to_Tile_Async(Af77,LDA,A,sequence,request,info)
- use iso_c_binding
- implicit none
- integer(kind=c_int), intent(out) :: info
- integer(kind=c_int), intent(in) :: LDA
- complex(kind=c_double_complex), intent(in), target :: Af77(LDA,*)
- type(c_ptr), value :: A ! Arg managed by MORSE: opaque to Fortran
- type(c_ptr), value :: request ! Arg managed by MORSE: opaque to Fortran
- type(c_ptr), value :: sequence ! Arg managed by MORSE: opaque to Fortran
- info = MORSE_zLapack_to_Tile_Async_c(c_loc(Af77),LDA,A,sequence,request)
- end subroutine MORSE_zLapack_to_Tile_Async
-
- subroutine MORSE_zTile_to_Lapack_Async(A,Af77,LDA,sequence,request,info)
- use iso_c_binding
- implicit none
- integer(kind=c_int), intent(out) :: info
- integer(kind=c_int), intent(in) :: LDA
- complex(kind=c_double_complex), intent(out), target :: Af77(LDA,*)
- type(c_ptr), value :: A ! Arg managed by MORSE: opaque to Fortran
- type(c_ptr), value :: request ! Arg managed by MORSE: opaque to Fortran
- type(c_ptr), value :: sequence ! Arg managed by MORSE: opaque to Fortran
- info = MORSE_zTile_to_Lapack_Async_c(A,c_loc(Af77),LDA,sequence,request)
- end subroutine MORSE_zTile_to_Lapack_Async
-
end module morse_z
diff --git a/control/tile.c b/control/tile.c
index ea216a770a6767745e2659cc27e4f312a3161808..2d2a05270486126208ccaa885c6f87ed56ed9078 100644
--- a/control/tile.c
+++ b/control/tile.c
@@ -29,10 +29,8 @@
* @brief Group routines exposed to users for matrices conversion LAPACK-Tile
*
*/
-
#include "control/common.h"
#include "control/auxiliary.h"
-#include "control/tile.h"
/** ***************************************************************************
*
@@ -59,54 +57,21 @@
*****************************************************************************/
int MORSE_Lapack_to_Tile(void *Af77, int LDA, MORSE_desc_t *A)
{
- MORSE_context_t *morse;
- MORSE_sequence_t *sequence = NULL;
- MORSE_request_t request;
- int status;
-
- morse = morse_context_self();
- if (morse == NULL) {
- morse_fatal_error("MORSE_Lapack_to_Tile", "MORSE not initialized");
- return MORSE_ERR_NOT_INITIALIZED;
- }
- /* Check descriptor for correctness */
- if (morse_desc_check(A) != MORSE_SUCCESS) {
- morse_error("MORSE_Lapack_to_Tile", "invalid descriptor");
- return MORSE_ERR_ILLEGAL_VALUE;
- }
- morse_sequence_create(morse, &sequence);
switch( A->dtyp ) {
-#if defined(PRECISION_s)
- case MorseRealFloat:
- morse_pslapack_to_tile(Af77, LDA, A, sequence, &request);
- break;
-#endif
-
-#if defined(PRECISION_d)
- case MorseRealDouble:
- morse_pdlapack_to_tile(Af77, LDA, A, sequence, &request);
+ case MorseComplexDouble:
+ return MORSE_zLapack_to_Tile( (MORSE_Complex64_t *)Af77, LDA, A );
break;
-#endif
-
-#if defined(PRECISION_c)
case MorseComplexFloat:
- morse_pclapack_to_tile(Af77, LDA, A, sequence, &request);
+ return MORSE_cLapack_to_Tile( (MORSE_Complex32_t *)Af77, LDA, A );
break;
-#endif
-
-#if defined(PRECISION_z)
- case MorseComplexDouble:
- morse_pzlapack_to_tile(Af77, LDA, A, sequence, &request);
+ case MorseRealFloat:
+ return MORSE_sLapack_to_Tile( (float *)Af77, LDA, A );
break;
-#endif
-
+ case MorseRealDouble:
default:
- morse_error("MORSE_Lapack_to_Tile", "Type unknown");
+ return MORSE_dLapack_to_Tile( (double *)Af77, LDA, A );
}
- RUNTIME_barrier(morse);
- status = sequence->status;
- morse_sequence_destroy(morse, sequence);
- return status;
+ return MORSE_ERR_ILLEGAL_VALUE;
}
/** ***************************************************************************
@@ -134,52 +99,19 @@ int MORSE_Lapack_to_Tile(void *Af77, int LDA, MORSE_desc_t *A)
*****************************************************************************/
int MORSE_Tile_to_Lapack(MORSE_desc_t *A, void *Af77, int LDA)
{
- MORSE_context_t *morse;
- MORSE_sequence_t *sequence = NULL;
- MORSE_request_t request;
- int status;
-
- morse = morse_context_self();
- if (morse == NULL) {
- morse_fatal_error("MORSE_Tile_to_Lapack", "MORSE not initialized");
- return MORSE_ERR_NOT_INITIALIZED;
- }
- /* Check descriptor for correctness */
- if (morse_desc_check(A) != MORSE_SUCCESS) {
- morse_error("MORSE_Tile_to_Lapack", "invalid descriptor");
- return MORSE_ERR_ILLEGAL_VALUE;
- }
- morse_sequence_create(morse, &sequence);
switch( A->dtyp ) {
-#if defined(PRECISION_s)
- case MorseRealFloat:
- morse_pstile_to_lapack(A, Af77, LDA, sequence, &request);
- break;
-#endif
-
-#if defined(PRECISION_d)
- case MorseRealDouble:
- morse_pdtile_to_lapack(A, Af77, LDA, sequence, &request);
+ case MorseComplexDouble:
+ return MORSE_zTile_to_Lapack( A, (MORSE_Complex64_t *)Af77, LDA );
break;
-#endif
-
-#if defined(PRECISION_c)
case MorseComplexFloat:
- morse_pctile_to_lapack(A, Af77, LDA, sequence, &request);
+ return MORSE_cTile_to_Lapack( A, (MORSE_Complex32_t *)Af77, LDA );
break;
-#endif
-
-#if defined(PRECISION_z)
- case MorseComplexDouble:
- morse_pztile_to_lapack(A, Af77, LDA, sequence, &request);
+ case MorseRealFloat:
+ return MORSE_sTile_to_Lapack( A, (float *)Af77, LDA );
break;
-#endif
-
+ case MorseRealDouble:
default:
- morse_error("MORSE_Tile_to_Lapack", "Type unknown");
+ return MORSE_dTile_to_Lapack( A, (double *)Af77, LDA );
}
- RUNTIME_barrier(morse);
- status = sequence->status;
- morse_sequence_destroy(morse, sequence);
- return status;
+ return MORSE_ERR_ILLEGAL_VALUE;
}
diff --git a/control/tile.h b/control/tile.h
deleted file mode 100644
index 0cde00ce27af81b6fede36e1fb99ec83fe6beb01..0000000000000000000000000000000000000000
--- a/control/tile.h
+++ /dev/null
@@ -1,63 +0,0 @@
-/**
- *
- * @copyright (c) 2009-2014 The University of Tennessee and The University
- * of Tennessee Research Foundation.
- * All rights reserved.
- * @copyright (c) 2012-2014 Inria. All rights reserved.
- * @copyright (c) 2012-2014 Bordeaux INP, CNRS (LaBRI UMR 5800), Inria, Univ. Bordeaux. All rights reserved.
- *
- **/
-
-/**
- *
- * @file tile.h
- *
- * MORSE auxiliary routines
- * MORSE is a software package provided by Univ. of Tennessee,
- * Univ. of California Berkeley and Univ. of Colorado Denver
- *
- * @version 0.9.0
- * @author Jakub Kurzak
- * @author Cedric Castagnede
- * @date 2010-11-15
- *
- **/
-#ifndef _MORSE_TILE_H_
-#define _MORSE_TILE_H_
-
-#if defined( _WIN32 ) || defined( _WIN64 )
-typedef __int64 int64_t;
-#else
-#include <inttypes.h>
-#endif
-
-#define ELTADDR(A, type, m, n) (type *)morse_geteltaddr(A, m, n)
-#define ELTLDD(A, k) ( ( (((k)-1)/(A).mb) + (A).i/(A).mb) < (A).lm1 ? (A).mb : (A).lm%(A).mb )
-
-#ifdef __cplusplus
-extern "C" {
-#endif
-
-/*******************************************************************************
- * Internal routines - dynamic scheduling
- **/
-void morse_pztile_to_lapack(MORSE_desc_t*, MORSE_Complex64_t*, int, MORSE_sequence_t *sequence, MORSE_request_t *request);
-void morse_pctile_to_lapack(MORSE_desc_t*, MORSE_Complex32_t*, int, MORSE_sequence_t *sequence, MORSE_request_t *request);
-void morse_pdtile_to_lapack(MORSE_desc_t*, double*, int, MORSE_sequence_t *sequence, MORSE_request_t *request);
-void morse_pstile_to_lapack(MORSE_desc_t*, float*, int, MORSE_sequence_t *sequence, MORSE_request_t *request);
-
-void morse_pzlapack_to_tile(MORSE_Complex64_t*, int, MORSE_desc_t*, MORSE_sequence_t *sequence, MORSE_request_t *request);
-void morse_pclapack_to_tile(MORSE_Complex32_t*, int, MORSE_desc_t*, MORSE_sequence_t *sequence, MORSE_request_t *request);
-void morse_pdlapack_to_tile(double*, int, MORSE_desc_t*, MORSE_sequence_t *sequence, MORSE_request_t *request);
-void morse_pslapack_to_tile(float*, int, MORSE_desc_t*, MORSE_sequence_t *sequence, MORSE_request_t *request);
-
-void morse_pztile_zero(MORSE_desc_t *dA, MORSE_sequence_t *sequence, MORSE_request_t *request);
-void morse_pctile_zero(MORSE_desc_t *dA, MORSE_sequence_t *sequence, MORSE_request_t *request);
-void morse_pdtile_zero(MORSE_desc_t *dA, MORSE_sequence_t *sequence, MORSE_request_t *request);
-void morse_pstile_zero(MORSE_desc_t *dA, MORSE_sequence_t *sequence, MORSE_request_t *request);
-
-#ifdef __cplusplus
-}
-#endif
-
-#endif
diff --git a/control/workspace.c b/control/workspace.c
index efcb49f8c9f5be51ed3f9cbf6cdc3bcaff815b69..852402979c754194e26fec04c906ea8dfe47b941 100644
--- a/control/workspace.c
+++ b/control/workspace.c
@@ -94,6 +94,8 @@ int morse_alloc_ibnb_tile(int M, int N, MORSE_enum func, int type, MORSE_desc_t
return MORSE_ERR_OUT_OF_RESOURCES;
}
+ RUNTIME_desc_create( *desc );
+
/* Check that everything is ok */
status = morse_desc_check(*desc);
if (status != MORSE_SUCCESS) {
@@ -156,6 +158,8 @@ int morse_alloc_ipiv(int M, int N, MORSE_enum func, int type, MORSE_desc_t **des
return MORSE_ERR_OUT_OF_RESOURCES;
}
+ RUNTIME_desc_create( *desc );
+
return MORSE_SUCCESS;
}
diff --git a/coreblas/compute/core_zgetrf_incpiv.c b/coreblas/compute/core_zgetrf_incpiv.c
index 220d8794e04617a0632f892ac51ec1e8ed6a6724..52456b02751661c7d2cf7bea897fb8d09ea23b7f 100644
--- a/coreblas/compute/core_zgetrf_incpiv.c
+++ b/coreblas/compute/core_zgetrf_incpiv.c
@@ -86,9 +86,9 @@
*
******************************************************************************/
-int CORE_zgetrf_incpiv(int M, int N, int IB,
- MORSE_Complex64_t *A, int LDA,
- int *IPIV, int *INFO)
+int CORE_zgetrf_incpiv( int M, int N, int IB,
+ MORSE_Complex64_t *A, int LDA,
+ int *IPIV, int *INFO )
{
int i, j, k, sb;
int iinfo;
diff --git a/include/chameleon/morse_constants.h b/include/chameleon/morse_constants.h
index 8ffe90dc0868c8d574dd5fde01cc92fe63724fba..5638145bd06671d87a12bd4fb17c540b9529beb1 100644
--- a/include/chameleon/morse_constants.h
+++ b/include/chameleon/morse_constants.h
@@ -109,7 +109,6 @@
#define morse_const_neg(const) (((const-1)^0x01)+1)
-
/** ****************************************************************************
* MORSE constants - boolean
**/
@@ -119,7 +118,6 @@
#define MORSE_CPU ((1ULL)<<1)
#define MORSE_CUDA ((1ULL)<<3)
-
/** ****************************************************************************
* State machine switches
**/
@@ -149,7 +147,6 @@
#define MORSE_INPLACE 1
#define MORSE_OUTOFPLACE 2
-
/** ****************************************************************************
* MORSE constants - success & error codes
**/
@@ -166,7 +163,6 @@
#define MORSE_ERR_UNEXPECTED -110
#define MORSE_ERR_SEQUENCE_FLUSHED -111
-
/** ****************************************************************************
* Kernels options
**/
@@ -181,7 +177,6 @@
#define CALLBACK 17
#define REDUX 18
-
/** ****************************************************************************
* MORSE ???
**/
diff --git a/include/chameleon/morse_z.h b/include/chameleon/morse_z.h
index 89e0be9c89cd3af78fa66f8f73a71d89150578d0..2c504d54445271296b67344f4a4f06e8a6db3b97 100644
--- a/include/chameleon/morse_z.h
+++ b/include/chameleon/morse_z.h
@@ -349,8 +349,6 @@ int MORSE_Alloc_Workspace_zgetrf_incpiv_Tile(int N, MORSE_desc_t **descL, int **
**/
int MORSE_zLapack_to_Tile(MORSE_Complex64_t *Af77, int LDA, MORSE_desc_t *A);
int MORSE_zTile_to_Lapack(MORSE_desc_t *A, MORSE_Complex64_t *Af77, int LDA);
-int MORSE_zLapack_to_Tile_Async(MORSE_Complex64_t *Af77, int LDA, MORSE_desc_t *A, MORSE_sequence_t *sequence, MORSE_request_t *request);
-int MORSE_zTile_to_Lapack_Async(MORSE_desc_t *A, MORSE_Complex64_t *Af77, int LDA, MORSE_sequence_t *sequence, MORSE_request_t *request);
/** ****************************************************************************
* User Builder function prototypes