diff --git a/compute/CMakeLists.txt b/compute/CMakeLists.txt
index cc25fb7d233c7206ced2d9e5fc59766e1bef25fe..3d6a7caf4dedb592cd1aa8422212bd091c287687 100644
--- a/compute/CMakeLists.txt
+++ b/compute/CMakeLists.txt
@@ -40,11 +40,12 @@ set(CHAMELEON_CONTROL
../control/context.c
../control/control.c
../control/descriptor.c
- ../control/descriptor_rec.c
../control/descriptor_helpers.c
../control/descriptor_ipiv.c
- ../control/workspace.c
+ ../control/descriptor_rec.c
+ ../control/lamch.c
../control/tile.c
+ ../control/workspace.c
../control/chameleon_f77.c
../control/chameleon_mf77.c
map.c
diff --git a/compute/pzgepdf_qdwh.c b/compute/pzgepdf_qdwh.c
index 2f640e7a798bae6232d18bee57e035134ac81524..de7a82cb1a1f8e99c79bbea4a3551e8c5549d835 100644
--- a/compute/pzgepdf_qdwh.c
+++ b/compute/pzgepdf_qdwh.c
@@ -604,15 +604,7 @@ chameleon_pzgepdf_qdwh( cham_mtxtype_t mtxtype, CHAM_desc_t *descU, CHAM_desc_t
double normest, Unorm;
int it, itconv, facto = -1;
-#if !defined(CHAMELEON_SIMULATION)
- double eps = LAPACKE_dlamch_work('e');
-#else
-#if defined(PRECISION_z) || defined(PRECISION_d)
- double eps = 1.e-15;
-#else
- double eps = 1.e-7;
-#endif
-#endif
+ double eps = CHAMELEON_dlamch();
double tol1 = 5. * eps;
double tol3 = pow( tol1, 1./3. );
double id_flops_ratio = ( _zgepdf_qdwh_opt_id == 1 ) ? .5 : 1.5;
diff --git a/compute/pzgered.c b/compute/pzgered.c
index c43e9a0e4704b958bbe6380032f8c32ee3b060fe..e7feeed9d7ba8f0a3e433295e1d5af449035d531 100644
--- a/compute/pzgered.c
+++ b/compute/pzgered.c
@@ -28,8 +28,10 @@
#define W( desc, m, n ) (desc), (m), (n)
static inline void
-chameleon_pzgered_frb( cham_uplo_t uplo,
- CHAM_desc_t *A, CHAM_desc_t *Wnorm, CHAM_desc_t *Welt,
+chameleon_pzgered_frb( cham_uplo_t uplo,
+ CHAM_desc_t *A,
+ CHAM_desc_t *Wnorm,
+ CHAM_desc_t *Welt,
RUNTIME_option_t *options )
{
double alpha = 1.0;
@@ -155,14 +157,17 @@ chameleon_pzgered_frb( cham_uplo_t uplo,
/**
*
*/
-void chameleon_pzgered( cham_uplo_t uplo, double prec, CHAM_desc_t *A,
- RUNTIME_sequence_t *sequence, RUNTIME_request_t *request )
+void chameleon_pzgered( cham_uplo_t uplo,
+ double prec,
+ CHAM_desc_t *A,
+ RUNTIME_sequence_t *sequence,
+ RUNTIME_request_t *request )
{
CHAM_context_t *chamctxt;
RUNTIME_option_t options;
CHAM_desc_t Wcol;
CHAM_desc_t Welt;
- double gnorm, threshold, eps;
+ double gnorm, threshold, eps, eps_diag, threshold_diag;
int workmt, worknt;
int m, n;
@@ -202,37 +207,36 @@ void chameleon_pzgered( cham_uplo_t uplo, double prec, CHAM_desc_t *A,
/**
* Reduce the precision of the tiles if possible
*/
+ eps_diag = CHAMELEON_slamch();
if ( prec < 0. ) {
-#if !defined(CHAMELEON_SIMULATION)
- eps = LAPACKE_dlamch_work('e');
-#else
-#if defined(PRECISION_z) || defined(PRECISION_d)
- eps = 1.e-15;
-#else
- eps = 1.e-7;
-#endif
-#endif
+ eps = CHAMELEON_dlamch();
}
else {
eps = prec;
}
threshold = (eps * gnorm) / (double)(chameleon_min(A->mt, A->nt));
+ threshold_diag = ( eps < eps_diag ) ? threshold : (eps_diag * gnorm) / (double)(chameleon_min(A->mt, A->nt));
#if defined(CHAMELEON_DEBUG_GERED)
fprintf( stderr,
"[%2d] The norm of A is: %e\n"
"[%2d] The requested precision is: %e\n"
- "[%2d] The computed threshold is: %e\n",
+ "[%2d] The computed threshold is: %e\n"
+ "[%2d] The threshold diag is : %e\n",
A->myrank, gnorm,
A->myrank, eps,
- A->myrank, threshold );
+ A->myrank, threshold,
+ A->myrank, threshold_diag );
#endif
- for(m = 0; m < A->mt; m++) {
+
+ for(m = 0; m < A->mt; m++)
+ {
int tempmm = ( m == (A->mt-1) ) ? A->m - m * A->mb : A->mb;
int nmin = ( uplo == ChamUpper ) ? m : 0;
int nmax = ( uplo == ChamLower ) ? chameleon_min(m+1, A->nt) : A->nt;
- for(n = nmin; n < nmax; n++) {
+ for(n = nmin; n < nmax; n++)
+ {
int tempnn = ( n == (A->nt-1) ) ? A->n - n * A->nb : A->nb;
/*
@@ -241,8 +245,14 @@ void chameleon_pzgered( cham_uplo_t uplo, double prec, CHAM_desc_t *A,
* ||A_{i,j}||_F < u_{high} * || A ||_F / (nt * u_{low})
* ||A_{i,j}||_F < threshold / u_{low}
*/
- INSERT_TASK_zgered( &options, threshold,
- tempmm, tempnn, A( m, n ), W( &Wcol, m, n ) );
+ if ( m == n ) {
+ INSERT_TASK_zgered( &options, threshold_diag,
+ tempmm, tempnn, A( m, n ), W( &Wcol, m, n ) );
+ }
+ else {
+ INSERT_TASK_zgered( &options, threshold,
+ tempmm, tempnn, A( m, n ), W( &Wcol, m, n ) );
+ }
}
}
@@ -250,6 +260,6 @@ void chameleon_pzgered( cham_uplo_t uplo, double prec, CHAM_desc_t *A,
RUNTIME_sequence_wait( chamctxt, sequence );
chameleon_desc_destroy( &Wcol );
- RUNTIME_options_ws_free(&options);
- RUNTIME_options_finalize(&options, chamctxt);
+ RUNTIME_options_ws_free( &options );
+ RUNTIME_options_finalize( &options, chamctxt );
}
diff --git a/compute/pzhered.c b/compute/pzhered.c
index cc32f4243c81dd2e63656cf9972850e53dddbd83..869d748cf49c4d131ccb90292b492bb3658167da 100644
--- a/compute/pzhered.c
+++ b/compute/pzhered.c
@@ -28,8 +28,11 @@
#define W(desc, m, n) (desc), (m), (n)
static inline void
-chameleon_pzhered_frb( cham_trans_t trans, cham_uplo_t uplo,
- CHAM_desc_t *A, CHAM_desc_t *Wnorm, CHAM_desc_t *Welt,
+chameleon_pzhered_frb( cham_trans_t trans,
+ cham_uplo_t uplo,
+ CHAM_desc_t *A,
+ CHAM_desc_t *Wnorm,
+ CHAM_desc_t *Welt,
RUNTIME_option_t *options )
{
double alpha = 1.0;
@@ -84,8 +87,7 @@ chameleon_pzhered_frb( cham_trans_t trans, cham_uplo_t uplo,
{
int tempnn = (n == (NT - 1)) ? N - n * A->nb : A->nb;
- if (n == m)
- {
+ if ( n == m ) {
if ( trans == ChamConjTrans ) {
INSERT_TASK_zhessq(
options, ChamEltwise, uplo, tempmm,
@@ -97,8 +99,7 @@ chameleon_pzhered_frb( cham_trans_t trans, cham_uplo_t uplo,
A(m, n), W( Wnorm, m, n) );
}
}
- else
- {
+ else {
INSERT_TASK_zgessq(
options, ChamEltwise, tempmm, tempnn,
A(m, n), W( Wnorm, m, n ));
@@ -166,7 +167,7 @@ chameleon_pzhered_frb( cham_trans_t trans, cham_uplo_t uplo,
{
for (n = 0; n < A->q; n++)
{
- if ((m != 0) || (n != 0))
+ if ( ( m != 0 ) || ( n != 0 ) )
{
INSERT_TASK_dlacpy(
options,
@@ -180,14 +181,18 @@ chameleon_pzhered_frb( cham_trans_t trans, cham_uplo_t uplo,
/**
*
*/
-void chameleon_pzhered( cham_trans_t trans, cham_uplo_t uplo, double prec, CHAM_desc_t *A,
- RUNTIME_sequence_t *sequence, RUNTIME_request_t *request )
+void chameleon_pzhered( cham_trans_t trans,
+ cham_uplo_t uplo,
+ double prec,
+ CHAM_desc_t *A,
+ RUNTIME_sequence_t *sequence,
+ RUNTIME_request_t *request )
{
CHAM_context_t *chamctxt;
RUNTIME_option_t options;
CHAM_desc_t Wcol;
CHAM_desc_t Welt;
- double gnorm, threshold, eps;
+ double gnorm, threshold, eps, eps_diag, threshold_diag;
int workmt, worknt;
int m, n;
@@ -205,22 +210,22 @@ void chameleon_pzhered( cham_trans_t trans, cham_uplo_t uplo, double prec, CHAM_
RUNTIME_options_ws_alloc(&options, 1, 0);
/* Matrix to store the norm of each element */
- chameleon_desc_init(&Wcol, CHAMELEON_MAT_ALLOC_GLOBAL, ChamRealDouble, 2, 1, 2,
- A->mt * 2, A->nt, 0, 0, A->mt * 2, A->nt, A->p, A->q,
- NULL, NULL, A->get_rankof_init, A->get_rankof_init_arg);
+ chameleon_desc_init( &Wcol, CHAMELEON_MAT_ALLOC_GLOBAL, ChamRealDouble, 2, 1, 2,
+ A->mt * 2, A->nt, 0, 0, A->mt * 2, A->nt, A->p, A->q,
+ NULL, NULL, A->get_rankof_init, A->get_rankof_init_arg );
/* Matrix to compute the global frobenius norm */
- chameleon_desc_init(&Welt, CHAMELEON_MAT_ALLOC_GLOBAL, ChamRealDouble, 2, 1, 2,
- workmt * 2, worknt, 0, 0, workmt * 2, worknt, A->p, A->q,
- NULL, NULL, NULL, NULL);
+ chameleon_desc_init( &Welt, CHAMELEON_MAT_ALLOC_GLOBAL, ChamRealDouble, 2, 1, 2,
+ workmt * 2, worknt, 0, 0, workmt * 2, worknt, A->p, A->q,
+ NULL, NULL, NULL, NULL );
chameleon_pzhered_frb( trans, uplo, A, &Wcol, &Welt, &options );
- CHAMELEON_Desc_Flush(&Wcol, sequence);
- CHAMELEON_Desc_Flush(&Welt, sequence);
- CHAMELEON_Desc_Flush(A, sequence);
+ CHAMELEON_Desc_Flush( &Wcol, sequence );
+ CHAMELEON_Desc_Flush( &Welt, sequence );
+ CHAMELEON_Desc_Flush( A, sequence );
- RUNTIME_sequence_wait(chamctxt, sequence);
+ RUNTIME_sequence_wait( chamctxt, sequence );
gnorm = *((double *)Welt.get_blkaddr(&Welt, A->myrank / A->q, A->myrank % A->q));
chameleon_desc_destroy(&Welt);
@@ -228,33 +233,28 @@ void chameleon_pzhered( cham_trans_t trans, cham_uplo_t uplo, double prec, CHAM_
/**
* Reduce the precision of the tiles if possible
*/
- if (prec < 0.)
- {
-#if !defined(CHAMELEON_SIMULATION)
- eps = LAPACKE_dlamch_work('e');
-#else
-#if defined(PRECISION_z) || defined(PRECISION_d)
- eps = 1.e-15;
-#else
- eps = 1.e-7;
-#endif
-#endif
+ eps_diag = CHAMELEON_slamch();
+ if (prec < 0.) {
+ eps = CHAMELEON_dlamch();
}
- else
- {
+ else {
eps = prec;
}
threshold = (eps * gnorm) / (double)(chameleon_min(A->mt, A->nt));
+ threshold_diag = (eps < eps_diag) ? threshold : (eps_diag * gnorm) / (double)(chameleon_min(A->mt, A->nt));
#if defined(CHAMELEON_DEBUG_GERED)
- fprintf(stderr,
- "[%2d] The norm of A is: %e\n"
- "[%2d] The requested precision is: %e\n"
- "[%2d] The computed threshold is: %e\n",
- A->myrank, gnorm,
- A->myrank, eps,
- A->myrank, threshold);
+ fprintf( stderr,
+ "[%2d] The norm of A is: %e\n"
+ "[%2d] The requested precision is: %e\n"
+ "[%2d] The computed threshold is: %e\n"
+ "[%2d] The threshold diag is: %e\n",
+ A->myrank, gnorm,
+ A->myrank, eps,
+ A->myrank, threshold,
+ A->myrank, threshold_diag );
#endif
+
for (m = 0; m < A->mt; m++)
{
int tempmm = (m == (A->mt - 1)) ? A->m - m * A->mb : A->mb;
@@ -271,15 +271,21 @@ void chameleon_pzhered( cham_trans_t trans, cham_uplo_t uplo, double prec, CHAM_
* ||A_{i,j}||_F < u_{high} * || A ||_F / (nt * u_{low})
* ||A_{i,j}||_F < threshold / u_{low}
*/
- INSERT_TASK_zgered( &options, threshold,
- tempmm, tempnn, A( m, n ), W( &Wcol, m, n ) );
+ if ( m == n ) {
+ INSERT_TASK_zgered( &options, threshold_diag,
+ tempmm, tempnn, A( m, n ), W( &Wcol, m, n ) );
+ }
+ else {
+ INSERT_TASK_zgered( &options, threshold,
+ tempmm, tempnn, A( m, n ), W( &Wcol, m, n ) );
+ }
}
}
- CHAMELEON_Desc_Flush(A, sequence);
- RUNTIME_sequence_wait(chamctxt, sequence);
+ CHAMELEON_Desc_Flush( A, sequence );
+ RUNTIME_sequence_wait( chamctxt, sequence );
- chameleon_desc_destroy(&Wcol);
- RUNTIME_options_ws_free(&options);
- RUNTIME_options_finalize(&options, chamctxt);
+ chameleon_desc_destroy( &Wcol );
+ RUNTIME_options_ws_free( &options );
+ RUNTIME_options_finalize( &options, chamctxt );
}
diff --git a/control/lamch.c b/control/lamch.c
new file mode 100644
index 0000000000000000000000000000000000000000..7dccb23f07fd4db5ef46b5ddcb971c2ae9d03a33
--- /dev/null
+++ b/control/lamch.c
@@ -0,0 +1,56 @@
+/**
+ *
+ * @file lamch.c
+ *
+ * @copyright 2024-2024 Bordeaux INP, CNRS (LaBRI UMR 5800), Inria,
+ * Univ. Bordeaux. All rights reserved.
+ *
+ ***
+ *
+ * @brief Chameleon lamch wrapper to factorize the test with simulation compilations
+ *
+ * @version 1.3.0
+ * @author Mathieu Faverge
+ * @date 2024-09-18
+ *
+ */
+#include "control/common.h"
+#if !defined(CHAMELEON_SIMULATION)
+#include "coreblas/lapacke.h"
+#endif
+
+/**
+ *
+ * Wrapper to the LAPACKE_slamch_work('e') call that returns a pre-defined value
+ * if compile with SimGrid.
+ *
+ * @return The epsilon value for single precision
+ *
+ */
+float
+CHAMELEON_slamch( void )
+{
+#if !defined(CHAMELEON_SIMULATION)
+ return LAPACKE_slamch_work( 'e' );
+#else
+ return 1.e-7;
+#endif
+}
+
+/**
+ *
+ * Wrapper to the LAPACKE_dlamch_work('e') call that returns a pre-defined value
+ * if compile with SimGrid.
+ *
+ * @return The epsilon value for single precision
+ *
+ */
+double
+CHAMELEON_dlamch( void )
+{
+#if !defined(CHAMELEON_SIMULATION)
+ return LAPACKE_dlamch_work( 'e' );
+#else
+ return 1.e-15;
+#endif
+}
diff --git a/example/lapack_to_chameleon/step0.h b/example/lapack_to_chameleon/step0.h
index 60c98fc18c20ac9c32d7186660f194f73dc24439..1f940475fcae8b141e0da8c90a8b581b69a08c55 100644
--- a/example/lapack_to_chameleon/step0.h
+++ b/example/lapack_to_chameleon/step0.h
@@ -96,11 +96,7 @@ static void read_args(int argc, char *argv[], int *iparam){
* Print a header message to summarize main parameters
*/
static void print_header(char *prog_name, int * iparam) {
-#if defined(CHAMELEON_SIMULATION)
- double eps = 0.;
-#else
- double eps = LAPACKE_dlamch_work( 'e' );
-#endif
+ double eps = LAPACKE_dlamch_work( 'e' );
printf( "#\n"
"# CHAMELEON %s\n"
diff --git a/example/lapack_to_chameleon/step1.h b/example/lapack_to_chameleon/step1.h
index abec8c3ea598dbb972d398cdcc6f9e668063a7e1..d3bc0724abeea5322e93a84b77ea95c3c57d2029 100644
--- a/example/lapack_to_chameleon/step1.h
+++ b/example/lapack_to_chameleon/step1.h
@@ -88,11 +88,7 @@ static void read_args(int argc, char *argv[], int *iparam){
* Print a header message to summarize main parameters
*/
static void print_header(char *prog_name, int * iparam) {
-#if defined(CHAMELEON_SIMULATION)
- double eps = 0.;
-#else
- double eps = LAPACKE_dlamch_work( 'e' );
-#endif
+ double eps = LAPACKE_dlamch_work( 'e' );
printf( "#\n"
"# CHAMELEON %d.%d.%d, %s\n"
diff --git a/example/lapack_to_chameleon/step2.h b/example/lapack_to_chameleon/step2.h
index 470466c287f018a6a8e9c70318ec8b6ba53eeb19..1c43fc94446886240e588e72b82334ee76e65ada 100644
--- a/example/lapack_to_chameleon/step2.h
+++ b/example/lapack_to_chameleon/step2.h
@@ -88,11 +88,7 @@ static void read_args(int argc, char *argv[], int *iparam){
* Print a header message to summarize main parameters
*/
static void print_header(char *prog_name, int * iparam) {
-#if defined(CHAMELEON_SIMULATION)
- double eps = 0.;
-#else
- double eps = LAPACKE_dlamch_work( 'e' );
-#endif
+ double eps = LAPACKE_dlamch_work( 'e' );
printf( "#\n"
"# CHAMELEON %d.%d.%d, %s\n"
diff --git a/example/lapack_to_chameleon/step3.h b/example/lapack_to_chameleon/step3.h
index 63c1ff0f47806e36477223c6eaf593877a6072bf..cbeeccaad4bff639ef8c266789fc28dec17dcfb4 100644
--- a/example/lapack_to_chameleon/step3.h
+++ b/example/lapack_to_chameleon/step3.h
@@ -88,11 +88,7 @@ static void read_args(int argc, char *argv[], int *iparam){
* Print a header message to summarize main parameters
*/
static void print_header(char *prog_name, int * iparam) {
-#if defined(CHAMELEON_SIMULATION)
- double eps = 0.;
-#else
- double eps = LAPACKE_dlamch_work( 'e' );
-#endif
+ double eps = LAPACKE_dlamch_work( 'e' );
printf( "#\n"
"# CHAMELEON %d.%d.%d, %s\n"
diff --git a/example/lapack_to_chameleon/step4.h b/example/lapack_to_chameleon/step4.h
index 28f0c4ca3ce3960fc901f38cd199bfc602f5d04d..d5b1a1c255d73b712f56e5e6af001e9424e2bac2 100644
--- a/example/lapack_to_chameleon/step4.h
+++ b/example/lapack_to_chameleon/step4.h
@@ -88,11 +88,7 @@ static void read_args(int argc, char *argv[], int *iparam){
* Print a header message to summarize main parameters
*/
static void print_header(char *prog_name, int * iparam) {
-#if defined(CHAMELEON_SIMULATION)
- double eps = 0.;
-#else
- double eps = LAPACKE_dlamch_work( 'e' );
-#endif
+ double eps = LAPACKE_dlamch_work( 'e' );
printf( "#\n"
"# CHAMELEON %d.%d.%d, %s\n"
diff --git a/example/lapack_to_chameleon/step5.h b/example/lapack_to_chameleon/step5.h
index e2ebf14d99f2d1feed0627ac45cc5f1b02c07778..8ecf015e4a41daaec0b5d85d1f0dd57f834d8f0f 100644
--- a/example/lapack_to_chameleon/step5.h
+++ b/example/lapack_to_chameleon/step5.h
@@ -103,11 +103,7 @@ static void read_args(int argc, char *argv[], int *iparam){
* Print a header message to summarize main parameters
*/
static void print_header(char *prog_name, int * iparam) {
-#if defined(CHAMELEON_SIMULATION)
- double eps = 0.;
-#else
- double eps = LAPACKE_dlamch_work( 'e' );
-#endif
+ double eps = LAPACKE_dlamch_work( 'e' );
printf( "#\n"
"# CHAMELEON %d.%d.%d, %s\n"
diff --git a/example/lapack_to_chameleon/step6.h b/example/lapack_to_chameleon/step6.h
index d60683987471c510aafa20ce55a6a87f0e358bcb..98043e257eedaa0225d06eb5f20877bdcd43880e 100644
--- a/example/lapack_to_chameleon/step6.h
+++ b/example/lapack_to_chameleon/step6.h
@@ -115,11 +115,7 @@ static void read_args(int argc, char *argv[], int *iparam){
* Print a header message to summarize main parameters
*/
static void print_header(char *prog_name, int * iparam) {
-#if defined(CHAMELEON_SIMULATION)
- double eps = 0.;
-#else
- double eps = LAPACKE_dlamch_work( 'e' );
-#endif
+ double eps = LAPACKE_dlamch_work( 'e' );
printf( "#\n"
"# CHAMELEON %d.%d.%d, %s\n"
diff --git a/example/lapack_to_chameleon/step7.h b/example/lapack_to_chameleon/step7.h
index ad975a20ba97de204732af99a660a8f4b080638f..7622f5ddbbe669e4cdf2810d761c0aa74181f2fb 100644
--- a/example/lapack_to_chameleon/step7.h
+++ b/example/lapack_to_chameleon/step7.h
@@ -165,11 +165,7 @@ static void read_args(int argc, char *argv[], int *iparam){
* Print a header message to summarize main parameters
*/
static void print_header(char *prog_name, int * iparam) {
-#if defined(CHAMELEON_SIMULATION)
- double eps = 0.;
-#else
- double eps = LAPACKE_dlamch_work( 'e' );
-#endif
+ double eps = LAPACKE_dlamch_work( 'e' );
printf( "#\n"
"# CHAMELEON %d.%d.%d, %s\n"
diff --git a/include/chameleon.h b/include/chameleon.h
index 7aeaf7d331227c35cd34288db6a95b68fc50f2fc..cb7e23dedaf7330ab0deffef714f6d9bc01fb7b1 100644
--- a/include/chameleon.h
+++ b/include/chameleon.h
@@ -227,6 +227,10 @@ int CHAMELEON_Ipiv_Gather( CHAM_ipiv_t *ipivdesc,
int root );
void CHAMELEON_Ipiv_Print ( const CHAM_ipiv_t *ipiv );
+/* Numerical helpers */
+float CHAMELEON_slamch( void );
+double CHAMELEON_dlamch( void );
+
/**
*
* @ingroup Control
diff --git a/runtime/starpu/codelets/codelet_zgered.c b/runtime/starpu/codelets/codelet_zgered.c
index d7a132200d603dca976f5c3d90ecf69cd2a33ab3..430db00ff7881550a473fee6bf24478ddbf9b9a8 100644
--- a/runtime/starpu/codelets/codelet_zgered.c
+++ b/runtime/starpu/codelets/codelet_zgered.c
@@ -71,8 +71,7 @@ void INSERT_TASK_zgered( const RUNTIME_option_t *options,
#if defined(CHAMELEON_USE_MPI)
/* Backup the MPI tag */
- if (A->myrank == tileA->rank)
- {
+ if ( A->myrank == tileA->rank ) {
tag = starpu_mpi_data_get_tag( *handleAin );
}
#endif /* defined(CHAMELEON_USE_MPI) */
@@ -89,39 +88,41 @@ void INSERT_TASK_zgered( const RUNTIME_option_t *options,
#if defined(CHAMELEON_DEBUG_GERED)
fprintf( stderr,
"[%2d] Convert the tile ( %d, %d ) to half precision\n",
- A->myrank, Am, An);
+ A->myrank, Am, An);
#endif
- starpu_cham_tile_register( &handleAout, -1, tileA, ChamComplexHalf );
+ if ( A->myrank == tileA->rank )
+ {
+ starpu_cham_tile_register( &handleAout, -1, tileA, ChamComplexHalf );
- rt_shm_starpu_insert_task(
- &cl_dlag2h,
- STARPU_VALUE, &m, sizeof(int),
- STARPU_VALUE, &n, sizeof(int),
- STARPU_R, *handleAin,
- STARPU_W, handleAout,
- STARPU_PRIORITY, options->priority,
- STARPU_EXECUTE_ON_WORKER, options->workerid,
+ rt_shm_starpu_insert_task(
+ &cl_dlag2h,
+ STARPU_VALUE, &m, sizeof(int),
+ STARPU_VALUE, &n, sizeof(int),
+ STARPU_R, *handleAin,
+ STARPU_W, handleAout,
+ STARPU_PRIORITY, options->priority,
+ STARPU_EXECUTE_ON_WORKER, options->workerid,
#if defined(CHAMELEON_CODELETS_HAVE_NAME)
- STARPU_NAME, "dlag2h",
+ STARPU_NAME, "dlag2h",
#endif
- 0);
+ 0);
- starpu_data_unregister_no_coherency( *handleAin );
- *handleAin = handleAout;
- tileA->flttype = ChamComplexHalf;
- starpu_mpi_data_register( handleAout, tag, tileA->rank );
- }
- else
- {
- tileA->flttype = ChamComplexHalf;
- if (*handleAin != NULL)
+ starpu_data_unregister_no_coherency( *handleAin );
+ *handleAin = handleAout;
+ tileA->flttype = ChamComplexHalf;
+ starpu_mpi_data_register( handleAout, tag, tileA->rank );
+ }
+ else
{
- starpu_data_unregister_no_coherency(*handleAin);
- *handleAin = NULL;
+ tileA->flttype = ChamComplexHalf;
+ if ( *handleAin != NULL )
+ {
+ starpu_data_unregister_no_coherency( *handleAin );
+ *handleAin = NULL;
+ }
}
+ return;
}
- return;
-
}
#endif
#endif
@@ -129,11 +130,7 @@ void INSERT_TASK_zgered( const RUNTIME_option_t *options,
/*
* Check for single precision
*/
-#if !defined(CHAMELEON_SIMULATION)
- u_low = LAPACKE_slamch_work('e');
-#else
- u_low = 1e-8;
-#endif
+ u_low = CHAMELEON_slamch();
if ( lnorm < (threshold / u_low) )
{
#if defined(CHAMELEON_DEBUG_GERED)
@@ -141,34 +138,34 @@ void INSERT_TASK_zgered( const RUNTIME_option_t *options,
"[%2d] Convert the tile ( %d, %d ) to single precision\n",
A->myrank, Am, An );
#endif
- if (A->myrank == tileA->rank)
+ if ( A->myrank == tileA->rank )
{
starpu_cham_tile_register( &handleAout, -1, tileA, ChamComplexFloat );
rt_shm_starpu_insert_task(
&cl_zlag2c,
- STARPU_VALUE, &m, sizeof(int),
- STARPU_VALUE, &n, sizeof(int),
- STARPU_R, *handleAin,
- STARPU_W, handleAout,
- STARPU_PRIORITY, options->priority,
+ STARPU_VALUE, &m, sizeof(int),
+ STARPU_VALUE, &n, sizeof(int),
+ STARPU_R, *handleAin,
+ STARPU_W, handleAout,
+ STARPU_PRIORITY, options->priority,
STARPU_EXECUTE_ON_WORKER, options->workerid,
#if defined(CHAMELEON_CODELETS_HAVE_NAME)
- STARPU_NAME, "zlag2c",
+ STARPU_NAME, "zlag2c",
#endif
0);
starpu_data_unregister_no_coherency( *handleAin );
- *handleAin = handleAout;
+ *handleAin = handleAout;
tileA->flttype = ChamComplexFloat;
starpu_mpi_data_register( *handleAin, tag, tileA->rank );
}
else
{
tileA->flttype = ChamComplexFloat;
- if (*handleAin != NULL)
+ if ( *handleAin != NULL )
{
- starpu_data_unregister_no_coherency(*handleAin);
+ starpu_data_unregister_no_coherency( *handleAin );
*handleAin = NULL;
}
}
diff --git a/runtime/starpu/codelets/codelet_zgerst.c b/runtime/starpu/codelets/codelet_zgerst.c
index f0fbdc1a40cffc2f88f46f8cb32807acb0cd5720..c2faaf9398f2f5455fe3a85d6d3d5c42bfeafb8e 100644
--- a/runtime/starpu/codelets/codelet_zgerst.c
+++ b/runtime/starpu/codelets/codelet_zgerst.c
@@ -39,7 +39,8 @@ void INSERT_TASK_zgerst( const RUNTIME_option_t *options,
handleAin = A->schedopt;
handleAin += ((int64_t)A->lmt) * nn + mm;
- if ( tileA->flttype == ChamComplexDouble ) {
+ if ( tileA->flttype == ChamComplexDouble )
+ {
starpu_data_handle_t *copy = handleAin;
/* Remove first copy */
@@ -59,12 +60,12 @@ void INSERT_TASK_zgerst( const RUNTIME_option_t *options,
return;
}
- if (A->myrank != tileA->rank)
+ if ( A->myrank != tileA->rank )
{
tileA->flttype = ChamComplexDouble;
- if (*handleAin != NULL)
+ if ( *handleAin != NULL )
{
- starpu_data_unregister_no_coherency(*handleAin);
+ starpu_data_unregister_no_coherency( *handleAin );
*handleAin = NULL;
}
return;
@@ -79,9 +80,9 @@ void INSERT_TASK_zgerst( const RUNTIME_option_t *options,
switch( tileA->flttype ) {
#if defined(CHAMELEON_USE_CUDA) && (CUDA_VERSION >= 7500)
#if defined(PRECISION_d)
- /*
- * Restore from half precision
- */
+ /*
+ * Restore from half precision
+ */
case ChamComplexHalf:
assert( options->withcuda );
#if defined(CHAMELEON_DEBUG_GERED)
@@ -91,14 +92,14 @@ void INSERT_TASK_zgerst( const RUNTIME_option_t *options,
#endif
rt_shm_starpu_insert_task(
&cl_hlag2d,
- STARPU_VALUE, &m, sizeof(int),
- STARPU_VALUE, &n, sizeof(int),
- STARPU_R, *handleAin,
- STARPU_W, handleAout,
- STARPU_PRIORITY, options->priority,
+ STARPU_VALUE, &m, sizeof(int),
+ STARPU_VALUE, &n, sizeof(int),
+ STARPU_R, *handleAin,
+ STARPU_W, handleAout,
+ STARPU_PRIORITY, options->priority,
STARPU_EXECUTE_ON_WORKER, options->workerid,
#if defined(CHAMELEON_CODELETS_HAVE_NAME)
- STARPU_NAME, "hlag2d",
+ STARPU_NAME, "hlag2d",
#endif
0);
break;
@@ -108,19 +109,20 @@ void INSERT_TASK_zgerst( const RUNTIME_option_t *options,
case ChamComplexFloat:
#if defined(CHAMELEON_DEBUG_GERED)
fprintf( stderr,
- "[%2d] Convert back the tile ( %d, %d ) from half precision\n",
+ "[%2d] Convert back the tile ( %d, %d ) from single precision\n",
A->myrank, Am, An );
#endif
+
rt_shm_starpu_insert_task(
&cl_clag2z,
- STARPU_VALUE, &m, sizeof(int),
- STARPU_VALUE, &n, sizeof(int),
- STARPU_R, *handleAin,
- STARPU_W, handleAout,
- STARPU_PRIORITY, options->priority,
+ STARPU_VALUE, &m, sizeof(int),
+ STARPU_VALUE, &n, sizeof(int),
+ STARPU_R, *handleAin,
+ STARPU_W, handleAout,
+ STARPU_PRIORITY, options->priority,
STARPU_EXECUTE_ON_WORKER, options->workerid,
#if defined(CHAMELEON_CODELETS_HAVE_NAME)
- STARPU_NAME, "clag2z",
+ STARPU_NAME, "clag2z",
#endif
0);
break;
@@ -130,7 +132,7 @@ void INSERT_TASK_zgerst( const RUNTIME_option_t *options,
}
starpu_data_unregister_no_coherency( *handleAin );
- *handleAin = handleAout;
+ *handleAin = handleAout;
tileA->flttype = ChamComplexDouble;
starpu_mpi_data_register( handleAout, tag, tileA->rank );
}
diff --git a/testing/chameleon_ztesting.c b/testing/chameleon_ztesting.c
index 9b7336c18fc0803900b087f878ca287c0dc97ac5..4c2dd37c0fa0b44dfd6ba4fd66a3cd99361683ae 100644
--- a/testing/chameleon_ztesting.c
+++ b/testing/chameleon_ztesting.c
@@ -167,7 +167,6 @@ int main (int argc, char **argv) {
testing_options_init( &options );
-#if !defined(CHAMELEON_SIMULATION)
/* Let's initialize the accuracy for the checks */
{
#if (defined(PRECISION_z) || defined(PRECISION_d))
@@ -178,10 +177,9 @@ int main (int argc, char **argv) {
else
#endif
{
- testing_setaccuracy( LAPACKE_dlamch_work('e') );
+ testing_setaccuracy( CHAMELEON_dlamch() );
}
}
-#endif
rc = CHAMELEON_Init( options.threads, options.gpus );
if ( rc != CHAMELEON_SUCCESS ) {
diff --git a/testing/parameters.c b/testing/parameters.c
index 5beb021c61b6abdc0d020e53f9dd408efa938cce..9607d83e901a779375cd70703ec49dec7fc96f10 100644
--- a/testing/parameters.c
+++ b/testing/parameters.c
@@ -13,6 +13,7 @@
* @author Mathieu Faverge
* @author Alycia Lisito
* @author Lionel Eyraud-Dubois
+ * @author Lucas Barros De Assis
* @date 2023-07-05
*
*/
@@ -489,6 +490,12 @@ parameters_desc_create( const char *id, CHAM_desc_t **descptr, cham_flttype_t dt
mtxfmt = -mtxfmt; /* Inverse sign to get the defined values */
+ if ( cham_is_mixed( dtyp ) && ( (void*)mtxfmt != CHAMELEON_MAT_ALLOC_TILE ) )
+ {
+ fprintf( stderr, "parameters_desc_create: Mixed precision descriptors can only be used with tiled allocation. Please enforce '--mtxfmt=1' \n" );
+ return CHAMELEON_ERR_ILLEGAL_VALUE;
+ }
+
if ( !custom ) {
int P = parameters_getvalue_int( "P" );
int Q = parameters_compute_q( P );
diff --git a/testing/vendor_ztesting.c b/testing/vendor_ztesting.c
index 72b484b214e496a8fd84aad8a1038b915ff3bffd..fd358eed9e022e752a801d201eac831f4a028093 100644
--- a/testing/vendor_ztesting.c
+++ b/testing/vendor_ztesting.c
@@ -118,10 +118,8 @@ int main (int argc, char **argv) {
testing_options_init( &options );
-#if !defined(CHAMELEON_SIMULATION)
/* Let's initialize the accuracy for the checks */
- testing_setaccuracy( LAPACKE_dlamch_work('e') );
-#endif
+ testing_setaccuracy( CHAMELEON_dlamch() );
rc = CHAMELEON_Init( options.threads, 0 );
if ( rc != CHAMELEON_SUCCESS ) {