diff --git a/compute/pzhetrd_he2hb.c b/compute/pzhetrd_he2hb.c
index 288cc93828c06a1d2bd1412435d59b5cc589a2b2..8d763d0bda5bf260c8f5abfb89cf438a4ed63b06 100644
--- a/compute/pzhetrd_he2hb.c
+++ b/compute/pzhetrd_he2hb.c
@@ -102,7 +102,7 @@ void chameleon_pzhetrd_he2hb(cham_uplo_t uplo,
tempkn = k == A->nt-1 ? A->n-k*A->nb : A->nb;
INSERT_TASK_zhe2ge( &options,
- uplo, tempkn, tempkn, A->mb,
+ uplo, tempkn, tempkn, A->mb,
A(k, k), D(k) );
}
diff --git a/compute/zcesca.c b/compute/zcesca.c
index 76489d42f9b1241c701703cd9e7075cf7984381f..552ec9d6e416d7bb7d7aaa46ff3fe19b862594af 100644
--- a/compute/zcesca.c
+++ b/compute/zcesca.c
@@ -290,6 +290,8 @@ int CHAMELEON_zcesca(int center, int scale, cham_store_t axis,
CHAM_desc_t *descSC = chameleon_desc_submatrix( &(ws->Wgcol), 0, 0, 1, N );
CHAMELEON_zDesc2Lap( ChamUpperLower, descSR, SR, M );
CHAMELEON_zDesc2Lap( ChamUpperLower, descSC, SC, 1 );
+ free( descSR );
+ free( descSC );
}
/* Cleanup the temporary data */
@@ -348,7 +350,7 @@ int CHAMELEON_zcesca(int center, int scale, cham_store_t axis,
*
*/
int CHAMELEON_zcesca_Tile( int center, int scale, cham_store_t axis, CHAM_desc_t *A,
- CHAMELEON_Complex64_t *SR, CHAMELEON_Complex64_t *SC)
+ CHAMELEON_Complex64_t *SR, CHAMELEON_Complex64_t *SC )
{
CHAM_context_t *chamctxt;
RUNTIME_sequence_t *sequence = NULL;
@@ -391,6 +393,8 @@ int CHAMELEON_zcesca_Tile( int center, int scale, cham_store_t axis, CHAM_desc_t
CHAM_desc_t *descSC = chameleon_desc_submatrix( &(ws->Wgcol), 0, 0, 1, A->ln );
CHAMELEON_zDesc2Lap( ChamUpperLower, descSR, SR, A->lm );
CHAMELEON_zDesc2Lap( ChamUpperLower, descSC, SC, 1 );
+ free( descSR );
+ free( descSC );
}
CHAMELEON_zcesca_WS_Free( ws );
diff --git a/compute/zgesvd.c b/compute/zgesvd.c
index 945783ccb79308d7b53b209644d91fc61dfc5e93..6d99d73b2a5d7025a0326b9a4354b3535e82ee8d 100644
--- a/compute/zgesvd.c
+++ b/compute/zgesvd.c
@@ -493,25 +493,25 @@ int CHAMELEON_zgesvd_Tile_Async( cham_job_t jobu, cham_job_t jobvt,
int info, nru, ncvt;
switch ( jobu ) {
- case ChamOVec :
+ /* case ChamOVec : */
case ChamAllVec :
nru = M;
break;
- case ChamSVec :
- nru = MINMN;
- break;
+ /* case ChamSVec : */
+ /* nru = MINMN; */
+ /* break; */
case ChamNoVec :
default:
nru = 0;
}
switch ( jobvt ) {
- case ChamOVec :
+ /* case ChamOVec : */
case ChamAllVec :
ncvt = N;
break;
- case ChamSVec :
- ncvt = MINMN;
- break;
+ /* case ChamSVec : */
+ /* ncvt = MINMN; */
+ /* break; */
case ChamNoVec :
default:
ncvt = 0;
diff --git a/compute/zgetrf.c b/compute/zgetrf.c
index 4a39cb5bba18ff489c69d81858626b5cdf955bbd..44d57502a84287db609f701b64a74f8a489aada0 100644
--- a/compute/zgetrf.c
+++ b/compute/zgetrf.c
@@ -315,6 +315,17 @@ CHAMELEON_zgetrf_Tile( CHAM_desc_t *A, CHAM_desc_t *IPIV )
* On entry, the M-by-N matrix to be factored.
* On exit, the tile factors L and U from the factorization.
*
+ * @param[in,out] IPIV
+ * On entry, the descriptor of an min(M, N)-by-1 matrix that may not
+ * have been initialized.
+ * On exit, the pivot vector generated during the factorization.
+ *
+ * @param[in,out] user_ws
+ * The opaque pointer to pre-allocated getrf workspace through
+ * CHAMELEON_zgetrf_WS_Alloc(). If user_ws is NULL, it is automatically
+ * allocated, but BE CAREFULL as it switches the call from asynchronous
+ * to synchronous call.
+ *
* @param[in] sequence
* Identifies the sequence of function calls that this call belongs to
* (for completion checks and exception handling purposes).
@@ -351,10 +362,6 @@ CHAMELEON_zgetrf_Tile_Async( CHAM_desc_t *A,
chameleon_fatal_error( "CHAMELEON_zgetrf_Tile_Async", "CHAMELEON_zgetrf_Tile_Async is only available with StarPU" );
return CHAMELEON_ERR_NOT_INITIALIZED;
}
- if ( user_ws == NULL ) {
- chameleon_fatal_error( "CHAMELEON_zgetrf_Tile_Async", "NULL user_ws" );
- return CHAMELEON_ERR_NOT_INITIALIZED;
- }
if ( sequence == NULL ) {
chameleon_fatal_error( "CHAMELEON_zgetrf_Tile_Async", "NULL sequence" );
return CHAMELEON_ERR_UNALLOCATED;
diff --git a/compute/zheevd.c b/compute/zheevd.c
index 140311a5f39ca3866dde88dc27c1a21eae3318b7..7d5421877a8d1a4a040a30963cfc81e1d3b33f61 100644
--- a/compute/zheevd.c
+++ b/compute/zheevd.c
@@ -507,7 +507,7 @@ int CHAMELEON_zheevd_Tile_Async( cham_job_t jobz, cham_uplo_t uplo,
#if defined(CHAMELEON_COPY_DIAG)
{
int m = chameleon_min(A->mt, A->nt) * A->mb;
- chameleon_zdesc_alloc(D, A->mb, A->nb, m, A->n, 0, 0, m, A->n, );
+ chameleon_zdesc_alloc(D, A->mb, A->nb, m, A->n, 0, 0, m, A->n, CHAMELEON_zgemm_WS_Free( gemm_ws ) );
Dptr = &D;
}
#endif
diff --git a/control/auxiliary.h b/control/auxiliary.h
index f4246973b61cc3b62c462a100a49d0bbc961749b..b7502ac7061a3b3a06b5c8067b2d90db28de5200 100644
--- a/control/auxiliary.h
+++ b/control/auxiliary.h
@@ -35,89 +35,6 @@
extern "C" {
#endif
-/*
- * Get environment variable
- */
-#if defined(CHAMELEON_OS_WINDOWS)
-
-static inline int
-chameleon_setenv( const char *var, const char *value, int overwrite ) {
- return !(SetEnvironmentVariable( var, value ));
-}
-
-static inline char *
-chameleon_getenv( const char *var ) {
- char *str;
- int len = 512;
- int rc;
- str = (char*)malloc(len * sizeof(char));
- rc = GetEnvironmentVariable(var, str, len);
- if (rc == 0) {
- free(str);
- str = NULL;
- }
- return str;
-}
-
-static inline void
-chameleon_cleanenv( char *str ) {
- if (str != NULL) free(str);
-}
-
-#else /* Other OS systems */
-
-static inline int
-chameleon_setenv( const char *var, const char *value, int overwrite ) {
- return setenv( var, value, overwrite );
-}
-
-static inline char *
-chameleon_getenv( const char *var ) {
- return getenv( var );
-}
-
-static inline void
-chameleon_cleanenv( char *str ) {
- (void)str;
-}
-
-#endif
-
-
-static inline int
-chameleon_env_is_set_to(char * str, char * value) {
- char * val;
- if ( (val = chameleon_getenv(str)) &&
- !strcmp(val, value))
- return 1;
- return 0;
-}
-
-static inline int
-chameleon_env_on_off( char * str, int default_value ) {
- if ( chameleon_env_is_set_to(str, "1") ) {
- return CHAMELEON_TRUE;
- }
- if ( chameleon_env_is_set_to(str, "0") ) {
- return CHAMELEON_FALSE;
- }
- return default_value;
-}
-
-static inline int
-chameleon_getenv_get_value_int(char * string, int default_value) {
- long int ret;
- char *str = chameleon_getenv(string);
- if (str == NULL) return default_value;
-
- if ( sscanf( str, "%ld", &ret ) != 1 ) {
- perror("sscanf");
- return default_value;
- }
-
- return (int)ret;
-}
-
/**
* Internal routines
*/
diff --git a/control/common.h b/control/common.h
index ef06cffb0568edde5cbc6d601ca6d4ad2a0a32c0..4c4d4acb5bfeecb9bb245122afd1ed6da1272a0d 100644
--- a/control/common.h
+++ b/control/common.h
@@ -11,11 +11,11 @@
*
* @brief Chameleon common header file
*
- * @version 1.2.0
+ * @version 1.3.0
* @author Mathieu Faverge
* @author Cedric Castagnede
* @author Florent Pruvost
- * @date 2022-02-22
+ * @date 2023-07-06
*
*/
/**
@@ -72,6 +72,7 @@
#include "control/context.h"
#include "control/descriptor.h"
#include "control/async.h"
+#include "chameleon/getenv.h"
/**
* Global shortcuts
@@ -133,6 +134,57 @@ static inline int chameleon_asprintf( char **strp, const char *fmt, ... )
return rc;
}
+#if defined(CHAMELEON_KERNELS_TRACE)
+static inline char *chameleon_codelet_name( char *kname, int nbtiles, ... )
+{
+ int nbchar = 256;
+ char output[256];
+ char *strptr = output;
+ va_list ap;
+ int rc, i;
+
+ rc = snprintf( strptr, nbchar, "%s(", kname );
+ if ( rc < 0 ) {
+ return NULL;
+ }
+ strptr += rc;
+ nbchar -= rc;
+
+ va_start( ap, nbtiles );
+ for( i=0; i<nbtiles; i++ )
+ {
+ CHAM_tile_t *tile = va_arg( ap, CHAM_tile_t* );
+ if ( i > 0 ) {
+ strptr[0] = ',';
+ strptr++;
+ nbchar--;
+ }
+ rc = snprintf( strptr, nbchar, " %s", tile->name );
+ if ( rc < 0 ) {
+ return NULL;
+ }
+ strptr += rc;
+ nbchar -= rc;
+ }
+ va_end( ap );
+
+ if ( nbchar < 3 ) {
+ return NULL;
+ }
+ strptr[0] = ' ';
+ strptr[1] = ')';
+ strptr[2] = '\0';
+
+ return strndup( output, 255 );
+}
+#else
+static inline char *chameleon_codelet_name( char *kname, int nbtiles, ... )
+{
+ (void)nbtiles;
+ return kname;
+}
+#endif
+
#ifdef __cplusplus
}
#endif
diff --git a/control/descriptor_helpers.c b/control/descriptor_helpers.c
index f11e487d32d578577534de023b37179765e263ff..87846e3a5709f9127aaf3fa4dfc72084ab3ae469 100644
--- a/control/descriptor_helpers.c
+++ b/control/descriptor_helpers.c
@@ -139,6 +139,15 @@ int chameleon_getrankof_custom_init( custom_dist_t **custom_dist,
return CHAMELEON_ERR_ILLEGAL_VALUE;
}
+ if ( (dist_m <= 0 ) || (dist_n <= 0) ) {
+ char message[300];
+ snprintf( message, 300, "Incorrect values for dist_m(%d) and/or dist_n(%d)", dist_m, dist_n );
+ chameleon_error( "chameleon_getrankof_custom_init", message );
+ free( result );
+ fclose( f );
+ return CHAMELEON_ERR_ILLEGAL_VALUE;
+ }
+
result->dist_m = dist_m;
result->dist_n = dist_n;
diff --git a/gpucublas/compute/cuda_dlag2h.cu b/gpucublas/compute/cuda_dlag2h.cu
index 8136c248fd9ff2c02a4910b251f282acdbe4cc77..000fc7454655c9a493ee970179fc28192f27f04b 100644
--- a/gpucublas/compute/cuda_dlag2h.cu
+++ b/gpucublas/compute/cuda_dlag2h.cu
@@ -133,6 +133,7 @@ CUDA_dlag2h( int m, int n,
cublasHandle_t handle )
{
cudaStream_t stream;
+ cudaError_t err;
double rmax;
if ( m < 0 ) {
@@ -166,6 +167,13 @@ CUDA_dlag2h( int m, int n,
cuda_dlag2h_kernel<<< grid, threads, 0, stream >>>( m, n, A, lda, HA, ldha, rmax );
+ err = cudaGetLastError();
+ if ( err != cudaSuccess )
+ {
+ fprintf( stderr, "CUDA_dlag2h failed to launch CUDA kernel %s\n", cudaGetErrorString(err) );
+ return CHAMELEON_ERR_UNEXPECTED;
+ }
+
return 0;
}
@@ -261,6 +269,7 @@ CUDA_hlag2d( int m, int n,
cublasHandle_t handle )
{
cudaStream_t stream;
+ cudaError_t err;
if ( m < 0 ) {
return -1;
@@ -286,5 +295,12 @@ CUDA_hlag2d( int m, int n,
cublasGetStream( handle, &stream );
cuda_hlag2d_kernel<<< grid, threads, 0, stream >>> ( m, n, HA, ldha, A, lda );
+ err = cudaGetLastError();
+ if ( err != cudaSuccess )
+ {
+ fprintf( stderr, "CUDA_hlag2d failed to launch CUDA kernel %s\n", cudaGetErrorString(err) );
+ return CHAMELEON_ERR_UNEXPECTED;
+ }
+
return 0;
}
diff --git a/gpucublas/compute/cuda_zlag2c.cu b/gpucublas/compute/cuda_zlag2c.cu
index cf3f35062fbe336c1c8f72119c659ba99e4693da..8b1fdfe7105a13cdb983a60de735d29d790eb38b 100644
--- a/gpucublas/compute/cuda_zlag2c.cu
+++ b/gpucublas/compute/cuda_zlag2c.cu
@@ -152,6 +152,7 @@ CUDA_zlag2c( int m, int n,
cublasHandle_t handle )
{
cudaStream_t stream;
+ cudaError_t err;
double rmax;
if ( m < 0 ) {
@@ -180,6 +181,13 @@ CUDA_zlag2c( int m, int n,
cuda_zlag2c_kernel<<< grid, threads, 0, stream >>>( m, n, A, lda, SA, ldsa, rmax );
+ err = cudaGetLastError();
+ if ( err != cudaSuccess )
+ {
+ fprintf( stderr, "CUDA_zlag2c failed to launch CUDA kernel %s\n", cudaGetErrorString(err) );
+ return CHAMELEON_ERR_UNEXPECTED;
+ }
+
return 0;
}
@@ -275,6 +283,7 @@ CUDA_clag2z( int m, int n,
cublasHandle_t handle )
{
cudaStream_t stream;
+ cudaError_t err;
if ( m < 0 ) {
return -1;
@@ -300,5 +309,12 @@ CUDA_clag2z( int m, int n,
cublasGetStream( handle, &stream );
cuda_clag2z_kernel<<< grid, threads, 0, stream >>> ( m, n, SA, ldsa, A, lda );
+ err = cudaGetLastError();
+ if ( err != cudaSuccess )
+ {
+ fprintf( stderr, "CUDA_clag2z failed to launch CUDA kernel %s\n", cudaGetErrorString(err) );
+ return CHAMELEON_ERR_UNEXPECTED;
+ }
+
return 0;
}
diff --git a/gpucublas/include/gpucublas.h b/gpucublas/include/gpucublas.h
index 8e7d4c3afa4d0165e2961c7791ae2dd2f12b3871..2305b4f9be07ee8f06d2ce6154fd1d55c0ece938 100644
--- a/gpucublas/include/gpucublas.h
+++ b/gpucublas/include/gpucublas.h
@@ -15,7 +15,7 @@
* @author Florent Pruvost
* @author Mathieu Faverge
* @author Nathalie Furmento
- * @date 2023-07-04
+ * @date 2023-07-06
* @precisions normal z -> c d s
*
*/
@@ -39,6 +39,7 @@
#include <cuda_fp16.h>
#endif
+#include <cuda_runtime_api.h>
#include <cublas_v2.h>
#define CUBLAS_SADDR(_a_) (&(_a_))
diff --git a/include/chameleon/getenv.h b/include/chameleon/getenv.h
new file mode 100644
index 0000000000000000000000000000000000000000..3151067446def980b186227e1a22237f211ae24a
--- /dev/null
+++ b/include/chameleon/getenv.h
@@ -0,0 +1,123 @@
+/**
+ *
+ * @file getenv.h
+ *
+ * @copyright 2009-2014 The University of Tennessee and The University of
+ * Tennessee Research Foundation. All rights reserved.
+ * @copyright 2012-2023 Bordeaux INP, CNRS (LaBRI UMR 5800), Inria,
+ * Univ. Bordeaux. All rights reserved.
+ *
+ ***
+ *
+ * @brief Chameleon auxiliary file to manage environment variables. Must be included after chameleon.h.
+ *
+ * @version 1.2.0
+ * @author Jakub Kurzak
+ * @author Piotr Luszczek
+ * @author Emmanuel Agullo
+ * @author Cedric Castagnede
+ * @author Florent Pruvost
+ * @author Mathieu Faverge
+ * @author Loris Lucido
+ * @date 2022-02-22
+ *
+ */
+#ifndef _chameleon_getenv_h_
+#define _chameleon_getenv_h_
+
+#include <string.h>
+#include <stdlib.h>
+#include <stdio.h>
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+/*
+ * Get environment variable
+ */
+#if defined(CHAMELEON_OS_WINDOWS)
+
+static inline int
+chameleon_setenv( const char *var, const char *value, int overwrite ) {
+ return !(SetEnvironmentVariable( var, value ));
+}
+
+static inline char *
+chameleon_getenv( const char *var ) {
+ char *str;
+ int len = 512;
+ int rc;
+ str = (char*)malloc(len * sizeof(char));
+ rc = GetEnvironmentVariable(var, str, len);
+ if (rc == 0) {
+ free(str);
+ str = NULL;
+ }
+ return str;
+}
+
+static inline void
+chameleon_cleanenv( char *str ) {
+ if (str != NULL) free(str);
+}
+
+#else /* Other OS systems */
+
+static inline int
+chameleon_setenv( const char *var, const char *value, int overwrite ) {
+ return setenv( var, value, overwrite );
+}
+
+static inline char *
+chameleon_getenv( const char *var ) {
+ return getenv( var );
+}
+
+static inline void
+chameleon_cleanenv( char *str ) {
+ (void)str;
+}
+
+#endif
+
+
+static inline int
+chameleon_env_is_set_to(char * str, char * value) {
+ char * val;
+ if ( (val = chameleon_getenv(str)) &&
+ !strcmp(val, value))
+ return 1;
+ return 0;
+}
+
+static inline int
+chameleon_env_on_off( char * str, int default_value ) {
+ if ( chameleon_env_is_set_to(str, "1") ) {
+ return CHAMELEON_TRUE;
+ }
+ if ( chameleon_env_is_set_to(str, "0") ) {
+ return CHAMELEON_FALSE;
+ }
+ return default_value;
+}
+
+static inline int
+chameleon_getenv_get_value_int(char * string, int default_value) {
+ long int ret;
+ char *str = chameleon_getenv(string);
+ if (str == NULL) return default_value;
+
+ if ( sscanf( str, "%ld", &ret ) != 1 ) {
+ perror("sscanf");
+ return default_value;
+ }
+
+ return (int)ret;
+}
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif /* _chameleon_getenv_h_ */
diff --git a/runtime/starpu/codelets/codelet_zgemm.c b/runtime/starpu/codelets/codelet_zgemm.c
index 53f241f899d219846396b802e54a27df9eeeb749..cf9b4b168f1abdfe222b1cd3be504b7dfb6bf38b 100644
--- a/runtime/starpu/codelets/codelet_zgemm.c
+++ b/runtime/starpu/codelets/codelet_zgemm.c
@@ -11,7 +11,7 @@
*
* @brief Chameleon zgemm StarPU codelet
*
- * @version 1.2.0
+ * @version 1.3.0
* @author Hatem Ltaief
* @author Jakub Kurzak
* @author Mathieu Faverge
@@ -24,7 +24,7 @@
* @author Lucas Nesi
* @author Loris Lucido
* @author Terry Cojean
- * @date 2023-01-30
+ * @date 2023-07-06
* @precisions normal z -> c d s
*
*/
@@ -38,10 +38,7 @@ struct cl_zgemm_args_s {
int n;
int k;
CHAMELEON_Complex64_t alpha;
- CHAM_tile_t *tileA;
- CHAM_tile_t *tileB;
CHAMELEON_Complex64_t beta;
- CHAM_tile_t *tileC;
};
#if !defined(CHAMELEON_SIMULATION)
@@ -172,10 +169,7 @@ void INSERT_TASK_zgemm_Astat( const RUNTIME_option_t *options,
clargs->n = n;
clargs->k = k;
clargs->alpha = alpha;
- clargs->tileA = A->get_blktile( A, Am, An );
- clargs->tileB = B->get_blktile( B, Bm, Bn );
clargs->beta = beta;
- clargs->tileC = C->get_blktile( C, Cm, Cn );
}
/* Callback for profiling information */
@@ -194,14 +188,11 @@ void INSERT_TASK_zgemm_Astat( const RUNTIME_option_t *options,
accessC = STARPU_RW;
}
-#if defined(CHAMELEON_KERNELS_TRACE)
- if ( clargs != NULL )
- {
- char *cl_fullname;
- chameleon_asprintf( &cl_fullname, "%s( %s, %s, %s )", cl_name, clargs->tileA->name, clargs->tileB->name, clargs->tileC->name );
- cl_name = cl_fullname;
- }
-#endif
+ /* Refine name */
+ cl_name = chameleon_codelet_name( cl_name, 3,
+ A->get_blktile( A, Am, An ),
+ B->get_blktile( B, Bm, Bn ),
+ C->get_blktile( C, Cm, Cn ) );
/* Insert the task */
rt_starpu_insert_task(
@@ -259,10 +250,7 @@ void INSERT_TASK_zgemm( const RUNTIME_option_t *options,
clargs->n = n;
clargs->k = k;
clargs->alpha = alpha;
- clargs->tileA = A->get_blktile( A, Am, An );
- clargs->tileB = B->get_blktile( B, Bm, Bn );
clargs->beta = beta;
- clargs->tileC = C->get_blktile( C, Cm, Cn );
}
/* Callback for profiling information */
@@ -271,15 +259,11 @@ void INSERT_TASK_zgemm( const RUNTIME_option_t *options,
/* Reduce the C access if needed */
accessC = ( beta == 0. ) ? STARPU_W : (STARPU_RW | ((beta == 1.) ? STARPU_COMMUTE : 0));
-#if defined(CHAMELEON_KERNELS_TRACE)
- if ( clargs != NULL )
- {
- char *cl_fullname;
- chameleon_asprintf( &cl_fullname, "%s( %s, %s, %s )", cl_name,
- clargs->tileA->name, clargs->tileB->name, clargs->tileC->name );
- cl_name = cl_fullname;
- }
-#endif
+ /* Refine name */
+ cl_name = chameleon_codelet_name( cl_name, 3,
+ A->get_blktile( A, Am, An ),
+ B->get_blktile( B, Bm, Bn ),
+ C->get_blktile( C, Cm, Cn ) );
/* Insert the task */
rt_starpu_insert_task(
diff --git a/runtime/starpu/codelets/codelet_zhemm.c b/runtime/starpu/codelets/codelet_zhemm.c
index d59d9f11586a1af9e9dbf4cedfecea4c90e3ba75..67fe91b28888f1532966332f275683e9c6ba5717 100644
--- a/runtime/starpu/codelets/codelet_zhemm.c
+++ b/runtime/starpu/codelets/codelet_zhemm.c
@@ -11,7 +11,7 @@
*
* @brief Chameleon zhemm StarPU codelet
*
- * @version 1.2.0
+ * @version 1.3.0
* @author Hatem Ltaief
* @author Jakub Kurzak
* @author Mathieu Faverge
@@ -20,7 +20,7 @@
* @author Lucas Barros de Assis
* @author Florent Pruvost
* @author Loris Lucido
- * @date 2022-02-22
+ * @date 2023-07-06
* @precisions normal z -> c
*
*/
@@ -33,10 +33,7 @@ struct cl_zhemm_args_s {
int m;
int n;
CHAMELEON_Complex64_t alpha;
- CHAM_tile_t *tileA;
- CHAM_tile_t *tileB;
CHAMELEON_Complex64_t beta;
- CHAM_tile_t *tileC;
};
#if !defined(CHAMELEON_SIMULATION)
@@ -164,10 +161,7 @@ void INSERT_TASK_zhemm_Astat( const RUNTIME_option_t *options,
clargs->m = m;
clargs->n = n;
clargs->alpha = alpha;
- clargs->tileA = A->get_blktile( A, Am, An );
- clargs->tileB = B->get_blktile( B, Bm, Bn );
clargs->beta = beta;
- clargs->tileC = C->get_blktile( C, Cm, Cn );
}
/* Callback for profiling information */
@@ -186,14 +180,11 @@ void INSERT_TASK_zhemm_Astat( const RUNTIME_option_t *options,
accessC = STARPU_RW;
}
-#if defined(CHAMELEON_KERNELS_TRACE)
- if ( clargs != NULL )
- {
- char *cl_fullname;
- chameleon_asprintf( &cl_fullname, "%s( %s, %s, %s )", cl_name, clargs->tileA->name, clargs->tileB->name, clargs->tileC->name );
- cl_name = cl_fullname;
- }
-#endif
+ /* Refine name */
+ cl_name = chameleon_codelet_name( cl_name, 3,
+ A->get_blktile( A, Am, An ),
+ B->get_blktile( B, Bm, Bn ),
+ C->get_blktile( C, Cm, Cn ) );
/* Insert the task */
rt_starpu_insert_task(
@@ -250,10 +241,7 @@ void INSERT_TASK_zhemm( const RUNTIME_option_t *options,
clargs->m = m;
clargs->n = n;
clargs->alpha = alpha;
- clargs->tileA = A->get_blktile( A, Am, An );
- clargs->tileB = B->get_blktile( B, Bm, Bn );
clargs->beta = beta;
- clargs->tileC = C->get_blktile( C, Cm, Cn );
}
/* Callback for profiling information */
@@ -262,14 +250,11 @@ void INSERT_TASK_zhemm( const RUNTIME_option_t *options,
/* Reduce the C access if needed */
accessC = ( beta == 0. ) ? STARPU_W : (STARPU_RW | ((beta == 1.) ? STARPU_COMMUTE : 0));
-#if defined(CHAMELEON_KERNELS_TRACE)
- if ( clargs != NULL )
- {
- char *cl_fullname;
- chameleon_asprintf( &cl_fullname, "%s( %s, %s, %s )", cl_name, clargs->tileA->name, clargs->tileB->name, clargs->tileC->name );
- cl_name = cl_fullname;
- }
-#endif
+ /* Refine name */
+ cl_name = chameleon_codelet_name( cl_name, 3,
+ A->get_blktile( A, Am, An ),
+ B->get_blktile( B, Bm, Bn ),
+ C->get_blktile( C, Cm, Cn ) );
/* Insert the task */
rt_starpu_insert_task(
diff --git a/runtime/starpu/codelets/codelet_zherk.c b/runtime/starpu/codelets/codelet_zherk.c
index c07ecb157fd7378d91ede9cb0f586baf154138b4..c53607f9d1034671d29bb2a4160d26e535daf598 100644
--- a/runtime/starpu/codelets/codelet_zherk.c
+++ b/runtime/starpu/codelets/codelet_zherk.c
@@ -11,7 +11,7 @@
*
* @brief Chameleon zherk StarPU codelet
*
- * @version 1.2.0
+ * @version 1.3.0
* @author Hatem Ltaief
* @author Jakub Kurzak
* @author Mathieu Faverge
@@ -21,7 +21,7 @@
* @author Florent Pruvost
* @author Gwenole Lucas
* @author Loris Lucido
- * @date 2022-02-22
+ * @date 2023-07-06
* @precisions normal z -> c
*
*/
@@ -34,9 +34,7 @@ struct cl_zherk_args_s {
int n;
int k;
double alpha;
- CHAM_tile_t *tileA;
double beta;
- CHAM_tile_t *tileC;
};
#if !defined(CHAMELEON_SIMULATION)
@@ -140,9 +138,7 @@ void INSERT_TASK_zherk( const RUNTIME_option_t *options,
clargs->n = n;
clargs->k = k;
clargs->alpha = alpha;
- clargs->tileA = A->get_blktile( A, Am, An );
clargs->beta = beta;
- clargs->tileC = C->get_blktile( C, Cm, Cn );
}
/* Callback fro profiling information */
@@ -151,14 +147,10 @@ void INSERT_TASK_zherk( const RUNTIME_option_t *options,
/* Reduce the C access if needed */
accessC = ( beta == 0. ) ? STARPU_W : STARPU_RW;
-#if defined(CHAMELEON_KERNELS_TRACE)
- if ( clargs != NULL )
- {
- char *cl_fullname;
- chameleon_asprintf( &cl_fullname, "%s( %s, %s )", cl_name, clargs->tileA->name, clargs->tileC->name );
- cl_name = cl_fullname;
- }
-#endif
+ /* Refine name */
+ cl_name = chameleon_codelet_name( cl_name, 2,
+ A->get_blktile( A, Am, An ),
+ C->get_blktile( C, Cm, Cn ) );
/* Insert the task */
rt_starpu_insert_task(
diff --git a/runtime/starpu/codelets/codelet_zlacpy.c b/runtime/starpu/codelets/codelet_zlacpy.c
index b5251bf9c285f116bb58a70419dba29bdc0a50ce..5a0deb8da5757a4c9120703de5b82459adc0da00 100644
--- a/runtime/starpu/codelets/codelet_zlacpy.c
+++ b/runtime/starpu/codelets/codelet_zlacpy.c
@@ -11,7 +11,7 @@
*
* @brief Chameleon zlacpy StarPU codelet
*
- * @version 1.2.0
+ * @version 1.3.0
* @author Julien Langou
* @author Henricus Bouwmeester
* @author Mathieu Faverge
@@ -21,7 +21,7 @@
* @author Florent Pruvost
* @author Samuel Thibault
* @author Alycia Lisito
- * @date 2022-02-22
+ * @date 2023-07-06
* @precisions normal z -> c d s
*
*/
@@ -36,8 +36,6 @@ struct cl_zlacpy_args_s {
int displB;
int lda;
int ldb;
- CHAM_tile_t *tileA;
- CHAM_tile_t *tileB;
};
#if !defined(CHAMELEON_SIMULATION)
@@ -102,8 +100,6 @@ void INSERT_TASK_zlacpyx( const RUNTIME_option_t *options,
clargs->n = n;
clargs->displA = displA;
clargs->displB = displB;
- clargs->tileA = A->get_blktile( A, Am, An );
- clargs->tileB = B->get_blktile( B, Bm, Bn );
clargs->lda = lda;
clargs->ldb = ldb;
}
@@ -154,10 +150,8 @@ void INSERT_TASK_zlacpy( const RUNTIME_option_t *options,
clargs->n = n;
clargs->displA = 0;
clargs->displB = 0;
- clargs->tileA = A->get_blktile( A, Am, An );
- clargs->tileB = B->get_blktile( B, Bm, Bn );
- clargs->lda = clargs->tileA->ld;
- clargs->ldb = clargs->tileB->ld;
+ clargs->lda = A->get_blktile( A, Am, An )->ld;
+ clargs->ldb = B->get_blktile( B, Bm, Bn )->ld;
}
/* Callback fro profiling information */
diff --git a/runtime/starpu/codelets/codelet_zlascal.c b/runtime/starpu/codelets/codelet_zlascal.c
index b76a71824883b3adc8c164286bb88d3212ed0aa1..19162d6388142d66808bd436e959aa048dc2b831 100644
--- a/runtime/starpu/codelets/codelet_zlascal.c
+++ b/runtime/starpu/codelets/codelet_zlascal.c
@@ -11,13 +11,13 @@
*
* @brief Chameleon zlascal StarPU codelet
*
- * @version 1.2.0
+ * @version 1.3.0
* @author Dalal Sukkari
* @author Lucas Barros de Assis
* @author Florent Pruvost
* @author Mathieu Faverge
* @author Samuel Thibault
- * @date 2022-02-22
+ * @date 2023-07-06
* @precisions normal z -> c d s
*
*/
@@ -29,7 +29,6 @@ struct cl_zlascal_args_s {
int m;
int n;
CHAMELEON_Complex64_t alpha;
- CHAM_tile_t *tileA;
};
#if !defined(CHAMELEON_SIMULATION)
@@ -82,7 +81,6 @@ void INSERT_TASK_zlascal( const RUNTIME_option_t *options,
clargs->m = m;
clargs->n = n;
clargs->alpha = alpha;
- clargs->tileA = A->get_blktile( A, Am, An );
}
/* Callback fro profiling information */
diff --git a/runtime/starpu/codelets/codelet_zlaset.c b/runtime/starpu/codelets/codelet_zlaset.c
index 335c1fe1017783e9a4703f516ef24f904570f209..c6acdc531ba2b8b9245aca7716885672caf0a5be 100644
--- a/runtime/starpu/codelets/codelet_zlaset.c
+++ b/runtime/starpu/codelets/codelet_zlaset.c
@@ -11,7 +11,7 @@
*
* @brief Chameleon zlaset StarPU codelet
*
- * @version 1.2.0
+ * @version 1.3.0
* @author Hatem Ltaief
* @author Mathieu Faverge
* @author Emmanuel Agullo
@@ -19,7 +19,7 @@
* @author Lucas Barros de Assis
* @author Florent Pruvost
* @author Samuel Thibault
- * @date 2022-02-22
+ * @date 2023-07-06
* @precisions normal z -> c d s
*
*/
@@ -32,7 +32,6 @@ struct cl_zlaset_args_s {
int n;
CHAMELEON_Complex64_t alpha;
CHAMELEON_Complex64_t beta;
- CHAM_tile_t *tileA;
};
#if !defined(CHAMELEON_SIMULATION)
@@ -76,7 +75,6 @@ void INSERT_TASK_zlaset( const RUNTIME_option_t *options,
clargs->n = n;
clargs->alpha = alpha;
clargs->beta = beta;
- clargs->tileA = A->get_blktile( A, Am, An );
}
/* Callback fro profiling information */
diff --git a/runtime/starpu/codelets/codelet_zlauum.c b/runtime/starpu/codelets/codelet_zlauum.c
index bbdb2899a5758606e5a518975c3b5b3bb076368b..bc49ece66b29f26d901891c3ca6900bbee749a87 100644
--- a/runtime/starpu/codelets/codelet_zlauum.c
+++ b/runtime/starpu/codelets/codelet_zlauum.c
@@ -11,7 +11,7 @@
*
* @brief Chameleon zlauum StarPU codelet
*
- * @version 1.2.0
+ * @version 1.3.0
* @author Julien Langou
* @author Henricus Bouwmeester
* @author Mathieu Faverge
@@ -20,7 +20,7 @@
* @author Lucas Barros de Assis
* @author Florent Pruvost
* @author Samuel Thibault
- * @date 2022-02-22
+ * @date 2023-07-06
* @precisions normal z -> c d s
*
*/
@@ -29,8 +29,7 @@
struct cl_zlauum_args_s {
cham_uplo_t uplo;
- int n;
- CHAM_tile_t *tileA;
+ int n;
};
#if !defined(CHAMELEON_SIMULATION)
@@ -70,19 +69,14 @@ void INSERT_TASK_zlauum( const RUNTIME_option_t *options,
clargs = malloc( sizeof( struct cl_zlauum_args_s ) );
clargs->uplo = uplo;
clargs->n = n;
- clargs->tileA = A->get_blktile( A, Am, An );
}
/* Callback fro profiling information */
callback = options->profiling ? cl_zlauum_callback : NULL;
-#if defined(CHAMELEON_KERNELS_TRACE)
- {
- char *cl_fullname;
- chameleon_asprintf( &cl_fullname, "%s( %s )", cl_name, clargs->tileA->name );
- cl_name = cl_fullname;
- }
-#endif
+ /* Refine name */
+ cl_name = chameleon_codelet_name( cl_name, 1,
+ A->get_blktile( A, Am, An ) );
/* Insert the task */
rt_starpu_insert_task(
diff --git a/runtime/starpu/codelets/codelet_zplghe.c b/runtime/starpu/codelets/codelet_zplghe.c
index c1abe29ead8e0624ea8f7c431ca2bf5459357d2a..34bee5797a5c7355a33d073c7b275ed6dcc5eadf 100644
--- a/runtime/starpu/codelets/codelet_zplghe.c
+++ b/runtime/starpu/codelets/codelet_zplghe.c
@@ -11,7 +11,7 @@
*
* @brief Chameleon zplghe StarPU codelet
*
- * @version 1.2.0
+ * @version 1.3.0
* @author Piotr Luszczek
* @author Pierre Lemarinier
* @author Mathieu Faverge
@@ -20,7 +20,7 @@
* @author Lucas Barros de Assis
* @author Florent Pruvost
* @author Samuel Thibault
- * @date 2022-02-22
+ * @date 2023-07-06
* @precisions normal z -> c
*
*/
@@ -31,7 +31,6 @@ struct cl_zplghe_args_s {
double bump;
int m;
int n;
- CHAM_tile_t *tileA;
int bigM;
int m0;
int n0;
@@ -76,7 +75,6 @@ void INSERT_TASK_zplghe( const RUNTIME_option_t *options,
clargs->bump = bump;
clargs->m = m;
clargs->n = n;
- clargs->tileA = A->get_blktile( A, Am, An );
clargs->bigM = bigM;
clargs->m0 = m0;
clargs->n0 = n0;
diff --git a/runtime/starpu/codelets/codelet_zplgsy.c b/runtime/starpu/codelets/codelet_zplgsy.c
index 3ad1ab5fa5d701d575b8ea04e0b624ecc867b25b..e39a3c69bc23ba9e16cbc9fce04734ea0332917c 100644
--- a/runtime/starpu/codelets/codelet_zplgsy.c
+++ b/runtime/starpu/codelets/codelet_zplgsy.c
@@ -11,7 +11,7 @@
*
* @brief Chameleon zplgsy StarPU codelet
*
- * @version 1.2.0
+ * @version 1.3.0
* @author Piotr Luszczek
* @author Pierre Lemarinier
* @author Mathieu Faverge
@@ -20,7 +20,7 @@
* @author Lucas Barros de Assis
* @author Florent Pruvost
* @author Samuel Thibault
- * @date 2022-02-22
+ * @date 2023-07-06
* @precisions normal z -> c d s
*
*/
@@ -31,7 +31,6 @@ struct cl_zplgsy_args_s {
CHAMELEON_Complex64_t bump;
int m;
int n;
- CHAM_tile_t *tileA;
int bigM;
int m0;
int n0;
@@ -76,7 +75,6 @@ void INSERT_TASK_zplgsy( const RUNTIME_option_t *options,
clargs->bump = bump;
clargs->m = m;
clargs->n = n;
- clargs->tileA = A->get_blktile( A, Am, An );
clargs->bigM = bigM;
clargs->m0 = m0;
clargs->n0 = n0;
diff --git a/runtime/starpu/codelets/codelet_zplrnt.c b/runtime/starpu/codelets/codelet_zplrnt.c
index df95bf5d2189bf8d481f4178ef3444d4f2cfc8f5..33a2bee23ca7b7ee93ebe9bbbc3d6f7cff2ff00b 100644
--- a/runtime/starpu/codelets/codelet_zplrnt.c
+++ b/runtime/starpu/codelets/codelet_zplrnt.c
@@ -11,7 +11,7 @@
*
* @brief Chameleon zplrnt StarPU codelet
*
- * @version 1.2.0
+ * @version 1.3.0
* @author Piotr Luszczek
* @author Pierre Lemarinier
* @author Mathieu Faverge
@@ -20,7 +20,7 @@
* @author Lucas Barros de Assis
* @author Florent Pruvost
* @author Samuel Thibault
- * @date 2022-02-22
+ * @date 2023-07-06
* @precisions normal z -> c d s
*
*/
@@ -30,7 +30,6 @@
struct cl_zplrnt_args_s {
int m;
int n;
- CHAM_tile_t *tileA;
int bigM;
int m0;
int n0;
@@ -75,7 +74,6 @@ void INSERT_TASK_zplrnt( const RUNTIME_option_t *options,
clargs = malloc( sizeof( struct cl_zplrnt_args_s ) );
clargs->m = m;
clargs->n = n;
- clargs->tileA = A->get_blktile( A, Am, An );
clargs->bigM = bigM;
clargs->m0 = m0;
clargs->n0 = n0;
diff --git a/runtime/starpu/codelets/codelet_zpotrf.c b/runtime/starpu/codelets/codelet_zpotrf.c
index dc3024575ba9ab387992b4034fb7df070a044bb8..b8b18765db71cecf6a9f79a4b3a333cb272f165d 100644
--- a/runtime/starpu/codelets/codelet_zpotrf.c
+++ b/runtime/starpu/codelets/codelet_zpotrf.c
@@ -11,7 +11,7 @@
*
* @brief Chameleon zpotrf StarPU codelet
*
- * @version 1.2.0
+ * @version 1.3.0
* @author Hatem Ltaief
* @author Jakub Kurzak
* @author Mathieu Faverge
@@ -21,7 +21,7 @@
* @author Florent Pruvost
* @author Samuel Thibault
* @author Terry Cojean
- * @date 2022-02-22
+ * @date 2023-07-06
* @precisions normal z -> c d s
*
*/
@@ -31,7 +31,6 @@
struct cl_zpotrf_args_s {
cham_uplo_t uplo;
int n;
- CHAM_tile_t *tileA;
int iinfo;
RUNTIME_sequence_t *sequence;
RUNTIME_request_t *request;
@@ -84,7 +83,6 @@ void INSERT_TASK_zpotrf( const RUNTIME_option_t *options,
clargs = malloc( sizeof( struct cl_zpotrf_args_s ) );
clargs->uplo = uplo;
clargs->n = n;
- clargs->tileA = A->get_blktile( A, Am, An );
clargs->iinfo = iinfo;
clargs->sequence = options->sequence;
clargs->request = options->request;
@@ -93,13 +91,9 @@ void INSERT_TASK_zpotrf( const RUNTIME_option_t *options,
/* Callback fro profiling information */
callback = options->profiling ? cl_zpotrf_callback : NULL;
-#if defined(CHAMELEON_KERNELS_TRACE)
- {
- char *cl_fullname;
- chameleon_asprintf( &cl_fullname, "%s( %s )", cl_name, clargs->tileA->name );
- cl_name = cl_fullname;
- }
-#endif
+ /* Refine name */
+ cl_name = chameleon_codelet_name( cl_name, 1,
+ A->get_blktile( A, Am, An ) );
/* Insert the task */
rt_starpu_insert_task(
diff --git a/runtime/starpu/codelets/codelet_zsymm.c b/runtime/starpu/codelets/codelet_zsymm.c
index 865e8563f127a8198e1b41c79b7161335939022f..ed8dea80b401d87dc8bd953c169eb11d2945cb7d 100644
--- a/runtime/starpu/codelets/codelet_zsymm.c
+++ b/runtime/starpu/codelets/codelet_zsymm.c
@@ -11,7 +11,7 @@
*
* @brief Chameleon zsymm StarPU codelet
*
- * @version 1.2.0
+ * @version 1.3.0
* @author Hatem Ltaief
* @author Jakub Kurzak
* @author Mathieu Faverge
@@ -21,7 +21,7 @@
* @author Florent Pruvost
* @author Romain Peressoni
* @author Loris Lucido
- * @date 2022-02-22
+ * @date 2023-07-06
* @precisions normal z -> c d s
*
*/
@@ -34,10 +34,7 @@ struct cl_zsymm_args_s {
int m;
int n;
CHAMELEON_Complex64_t alpha;
- CHAM_tile_t *tileA;
- CHAM_tile_t *tileB;
CHAMELEON_Complex64_t beta;
- CHAM_tile_t *tileC;
};
#if !defined(CHAMELEON_SIMULATION)
@@ -165,10 +162,7 @@ void INSERT_TASK_zsymm_Astat( const RUNTIME_option_t *options,
clargs->m = m;
clargs->n = n;
clargs->alpha = alpha;
- clargs->tileA = A->get_blktile( A, Am, An );
- clargs->tileB = B->get_blktile( B, Bm, Bn );
clargs->beta = beta;
- clargs->tileC = C->get_blktile( C, Cm, Cn );
}
/* Callback for profiling information */
@@ -187,14 +181,11 @@ void INSERT_TASK_zsymm_Astat( const RUNTIME_option_t *options,
accessC = STARPU_RW;
}
-#if defined(CHAMELEON_KERNELS_TRACE)
- if ( clargs != NULL )
- {
- char *cl_fullname;
- chameleon_asprintf( &cl_fullname, "%s( %s, %s, %s )", cl_name, clargs->tileA->name, clargs->tileB->name, clargs->tileC->name );
- cl_name = cl_fullname;
- }
-#endif
+ /* Refine name */
+ cl_name = chameleon_codelet_name( cl_name, 3,
+ A->get_blktile( A, Am, An ),
+ B->get_blktile( B, Bm, Bn ),
+ C->get_blktile( C, Cm, Cn ) );
/* Insert the task */
rt_starpu_insert_task(
@@ -251,10 +242,7 @@ void INSERT_TASK_zsymm( const RUNTIME_option_t *options,
clargs->m = m;
clargs->n = n;
clargs->alpha = alpha;
- clargs->tileA = A->get_blktile( A, Am, An );
- clargs->tileB = B->get_blktile( B, Bm, Bn );
clargs->beta = beta;
- clargs->tileC = C->get_blktile( C, Cm, Cn );
}
/* Callback for profiling information */
@@ -263,14 +251,11 @@ void INSERT_TASK_zsymm( const RUNTIME_option_t *options,
/* Reduce the C access if needed */
accessC = ( beta == 0. ) ? STARPU_W : (STARPU_RW | ((beta == 1.) ? STARPU_COMMUTE : 0));
-#if defined(CHAMELEON_KERNELS_TRACE)
- if ( clargs != NULL )
- {
- char *cl_fullname;
- chameleon_asprintf( &cl_fullname, "%s( %s, %s, %s )", cl_name, clargs->tileA->name, clargs->tileB->name, clargs->tileC->name );
- cl_name = cl_fullname;
- }
-#endif
+ /* Refine name */
+ cl_name = chameleon_codelet_name( cl_name, 3,
+ A->get_blktile( A, Am, An ),
+ B->get_blktile( B, Bm, Bn ),
+ C->get_blktile( C, Cm, Cn ) );
/* Insert the task */
rt_starpu_insert_task(
diff --git a/runtime/starpu/codelets/codelet_zsyrk.c b/runtime/starpu/codelets/codelet_zsyrk.c
index 049dfd4b889a15b1b87f80e280f281de12e13863..f11dd4158341615defeb35755d8c8dcabb28ebce 100644
--- a/runtime/starpu/codelets/codelet_zsyrk.c
+++ b/runtime/starpu/codelets/codelet_zsyrk.c
@@ -11,7 +11,7 @@
*
* @brief Chameleon zsyrk StarPU codelet
*
- * @version 1.2.0
+ * @version 1.3.0
* @author Hatem Ltaief
* @author Jakub Kurzak
* @author Mathieu Faverge
@@ -22,7 +22,7 @@
* @author Gwenole Lucas
* @author Terry Cojean
* @author Loris Lucido
- * @date 2022-02-22
+ * @date 2023-07-06
* @precisions normal z -> c d s
*
*/
@@ -35,9 +35,7 @@ struct cl_zsyrk_args_s {
int n;
int k;
CHAMELEON_Complex64_t alpha;
- CHAM_tile_t *tileA;
CHAMELEON_Complex64_t beta;
- CHAM_tile_t *tileC;
};
#if !defined(CHAMELEON_SIMULATION)
@@ -141,9 +139,7 @@ void INSERT_TASK_zsyrk( const RUNTIME_option_t *options,
clargs->n = n;
clargs->k = k;
clargs->alpha = alpha;
- clargs->tileA = A->get_blktile( A, Am, An );
clargs->beta = beta;
- clargs->tileC = C->get_blktile( C, Cm, Cn );
}
/* Callback fro profiling information */
@@ -152,14 +148,9 @@ void INSERT_TASK_zsyrk( const RUNTIME_option_t *options,
/* Reduce the C access if needed */
accessC = ( beta == 0. ) ? STARPU_W : STARPU_RW;
-#if defined(CHAMELEON_KERNELS_TRACE)
- if ( clargs != NULL )
- {
- char *cl_fullname;
- chameleon_asprintf( &cl_fullname, "%s( %s, %s )", cl_name, clargs->tileA->name, clargs->tileC->name );
- cl_name = cl_fullname;
- }
-#endif
+ cl_name = chameleon_codelet_name( cl_name, 2,
+ A->get_blktile( A, Am, An ),
+ C->get_blktile( C, Cm, Cn ) );
/* Insert the task */
rt_starpu_insert_task(
diff --git a/runtime/starpu/codelets/codelet_ztradd.c b/runtime/starpu/codelets/codelet_ztradd.c
index 8a6f28764b689581db36bbfd6e5f36175dc9add6..66fc4da3ee1e6ee84595881561b6157c6c5e2b07 100644
--- a/runtime/starpu/codelets/codelet_ztradd.c
+++ b/runtime/starpu/codelets/codelet_ztradd.c
@@ -11,12 +11,12 @@
*
* @brief Chameleon ztradd StarPU codelet
*
- * @version 1.2.0
+ * @version 1.3.0
* @author Mathieu Faverge
* @author Lucas Barros de Assis
* @author Florent Pruvost
* @author Samuel Thibault
- * @date 2022-02-22
+ * @date 2023-07-06
* @precisions normal z -> c d s
*
*/
@@ -29,9 +29,7 @@ struct cl_ztradd_args_s {
int m;
int n;
CHAMELEON_Complex64_t alpha;
- CHAM_tile_t *tileA;
CHAMELEON_Complex64_t beta;
- CHAM_tile_t *tileB;
};
#if !defined(CHAMELEON_SIMULATION)
@@ -86,9 +84,7 @@ void INSERT_TASK_ztradd( const RUNTIME_option_t *options,
clargs->m = m;
clargs->n = n;
clargs->alpha = alpha;
- clargs->tileA = A->get_blktile( A, Am, An );
clargs->beta = beta;
- clargs->tileB = B->get_blktile( B, Bm, Bn );
}
/* Callback fro profiling information */
diff --git a/runtime/starpu/codelets/codelet_ztrmm.c b/runtime/starpu/codelets/codelet_ztrmm.c
index 31e9dd3fef53ef3b389f2c78df0486f5991fbb7e..36b31640cc56bf85d19a31900af7aab3d4ca43c0 100644
--- a/runtime/starpu/codelets/codelet_ztrmm.c
+++ b/runtime/starpu/codelets/codelet_ztrmm.c
@@ -11,7 +11,7 @@
*
* @brief Chameleon ztrmm StarPU codelet
*
- * @version 1.2.0
+ * @version 1.3.0
* @author Julien Langou
* @author Henricus Bouwmeester
* @author Mathieu Faverge
@@ -20,7 +20,7 @@
* @author Lucas Barros de Assis
* @author Florent Pruvost
* @author Loris Lucido
- * @date 2022-02-22
+ * @date 2023-07-06
* @precisions normal z -> c d s
*
*/
@@ -35,8 +35,6 @@ struct cl_ztrmm_args_s {
int m;
int n;
CHAMELEON_Complex64_t alpha;
- CHAM_tile_t *tileA;
- CHAM_tile_t *tileB;
};
#if !defined(CHAMELEON_SIMULATION)
@@ -135,21 +133,15 @@ void INSERT_TASK_ztrmm( const RUNTIME_option_t *options,
clargs->m = m;
clargs->n = n;
clargs->alpha = alpha;
- clargs->tileA = A->get_blktile( A, Am, An );
- clargs->tileB = B->get_blktile( B, Bm, Bn );
}
/* Callback fro profiling information */
callback = options->profiling ? cl_ztrmm_callback : NULL;
-#if defined(CHAMELEON_KERNELS_TRACE)
- if ( clargs != NULL )
- {
- char *cl_fullname;
- chameleon_asprintf( &cl_fullname, "%s( %s, %s )", cl_name, clargs->tileA->name, clargs->tileB->name );
- cl_name = cl_fullname;
- }
-#endif
+ /* Refine name */
+ cl_name = chameleon_codelet_name( cl_name, 2,
+ A->get_blktile( A, Am, An ),
+ B->get_blktile( B, Bm, Bn ) );
/* Insert the task */
rt_starpu_insert_task(
diff --git a/runtime/starpu/codelets/codelet_ztrsm.c b/runtime/starpu/codelets/codelet_ztrsm.c
index 1ce23594d3f7928c03ba049e9ee043abee4c7c5a..901eeb7809ec81505559b2c3e83b4f7a4c7e1bb3 100644
--- a/runtime/starpu/codelets/codelet_ztrsm.c
+++ b/runtime/starpu/codelets/codelet_ztrsm.c
@@ -11,7 +11,7 @@
*
* @brief Chameleon ztrsm StarPU codelet
*
- * @version 1.2.0
+ * @version 1.3.0
* @author Hatem Ltaief
* @author Jakub Kurzak
* @author Mathieu Faverge
@@ -22,7 +22,7 @@
* @author Gwenole Lucas
* @author Loris Lucido
* @author Terry Cojean
- * @date 2023-01-30
+ * @date 2023-07-06
* @precisions normal z -> c d s
*
*/
@@ -37,8 +37,6 @@ struct cl_ztrsm_args_s {
int m;
int n;
CHAMELEON_Complex64_t alpha;
- CHAM_tile_t *tileA;
- CHAM_tile_t *tileB;
};
#if !defined(CHAMELEON_SIMULATION)
@@ -139,21 +137,15 @@ void INSERT_TASK_ztrsm( const RUNTIME_option_t *options,
clargs->m = m;
clargs->n = n;
clargs->alpha = alpha;
- clargs->tileA = A->get_blktile( A, Am, An );
- clargs->tileB = B->get_blktile( B, Bm, Bn );
}
/* Callback fro profiling information */
callback = options->profiling ? cl_ztrsm_callback : NULL;
-#if defined(CHAMELEON_KERNELS_TRACE)
- if ( clargs != NULL )
- {
- char *cl_fullname;
- chameleon_asprintf( &cl_fullname, "%s( %s, %s )", cl_name, clargs->tileA->name, clargs->tileB->name );
- cl_name = cl_fullname;
- }
-#endif
+ /* Refine name */
+ cl_name = chameleon_codelet_name( cl_name, 2,
+ A->get_blktile( A, Am, An ),
+ B->get_blktile( B, Bm, Bn ) );
/* Insert the task */
rt_starpu_insert_task(
diff --git a/runtime/starpu/codelets/codelet_ztrtri.c b/runtime/starpu/codelets/codelet_ztrtri.c
index 91680c63a3caccbb50813b62c22b4cf2b441091c..19499f058f32c493d6be3216439f18e84de15d0e 100644
--- a/runtime/starpu/codelets/codelet_ztrtri.c
+++ b/runtime/starpu/codelets/codelet_ztrtri.c
@@ -11,7 +11,7 @@
*
* @brief Chameleon ztrtri StarPU codelet
*
- * @version 1.2.0
+ * @version 1.3.0
* @author Julien Langou
* @author Henricus Bouwmeester
* @author Mathieu Faverge
@@ -20,7 +20,7 @@
* @author Lucas Barros de Assis
* @author Florent Pruvost
* @author Samuel Thibault
- * @date 2022-02-22
+ * @date 2023-07-06
* @precisions normal z -> c d s
*
*/
@@ -31,7 +31,6 @@ struct cl_ztrtri_args_s {
cham_uplo_t uplo;
cham_diag_t diag;
int n;
- CHAM_tile_t *tileA;
int iinfo;
RUNTIME_sequence_t *sequence;
RUNTIME_request_t *request;
@@ -81,7 +80,6 @@ void INSERT_TASK_ztrtri( const RUNTIME_option_t *options,
clargs->uplo = uplo;
clargs->diag = diag;
clargs->n = n;
- clargs->tileA = A->get_blktile( A, Am, An );
clargs->iinfo = iinfo;
clargs->sequence = options->sequence;
clargs->request = options->request;
@@ -90,13 +88,8 @@ void INSERT_TASK_ztrtri( const RUNTIME_option_t *options,
/* Callback fro profiling information */
callback = options->profiling ? cl_ztrtri_callback : NULL;
-#if defined(CHAMELEON_KERNELS_TRACE)
- {
- char *cl_fullname;
- chameleon_asprintf( &cl_fullname, "%s( %s )", cl_name, clargs->tileA->name );
- cl_name = cl_fullname;
- }
-#endif
+ /* Refine name */
+ cl_name = chameleon_codelet_name( cl_name, 1, A->get_blktile( A, Am, An ) );
/* Insert the task */
rt_starpu_insert_task(
diff --git a/runtime/starpu/include/chameleon_starpu.h.in b/runtime/starpu/include/chameleon_starpu.h.in
index 6eceaf063d23bfda2f7e6a2eb4ea1f402f929241..8e77d14a9dea84153c6b0672d63cecb3c71ed077 100644
--- a/runtime/starpu/include/chameleon_starpu.h.in
+++ b/runtime/starpu/include/chameleon_starpu.h.in
@@ -19,7 +19,7 @@
* @author Samuel Thibault
* @author Loris Lucido
* @author Terry Cojean
- * @date 2023-07-05
+ * @date 2023-07-06
*
*/
#ifndef _chameleon_starpu_h_
@@ -119,12 +119,6 @@ static inline int cham_to_starpu_access( cham_access_t accessA ) {
/*
* MPI Redefinitions
*/
-#if defined(CHAMELEON_RUNTIME_SYNC)
-#define TASK_SYNCHRONOUS , STARPU_TASK_SYNCHRONOUS, 1
-#else
-#define TASK_SYNCHRONOUS
-#endif
-
#if defined(CHAMELEON_USE_MPI)
#if defined(CHAMELEON_RUNTIME_SYNC)
diff --git a/runtime/starpu/interface/cham_tile_interface.c b/runtime/starpu/interface/cham_tile_interface.c
index 51e6a4b041fb50ff2f98f52f5830799875bc7af4..27eb801cb266a0038e77b19e04e729b68b3270f4 100644
--- a/runtime/starpu/interface/cham_tile_interface.c
+++ b/runtime/starpu/interface/cham_tile_interface.c
@@ -9,11 +9,11 @@
*
* @brief Chameleon tile interface for StarPU
*
- * @version 1.2.0
+ * @version 1.3.0
* @author Mathieu Faverge
* @author Gwenole Lucas
* @author Samuel Thibault
- * @date 2022-02-22
+ * @date 2023-07-06
*
*/
#include "chameleon_starpu.h"
@@ -297,16 +297,9 @@ cti_display( starpu_data_handle_t handle, FILE *f )
starpu_cham_tile_interface_t *cham_tile_interface = (starpu_cham_tile_interface_t *)
starpu_data_get_interface_on_node(handle, STARPU_MAIN_RAM);
-#if defined(CHAMELEON_KERNELS_TRACE)
- fprintf( f, "%s{.m=%u,.n=%u}",
- cham_tile_interface->tile.name,
- cham_tile_interface->tile.m,
- cham_tile_interface->tile.n );
-#else
fprintf( f, "%u\t%u\t",
cham_tile_interface->tile.m,
cham_tile_interface->tile.n );
-#endif
}
static int
diff --git a/testing/parameters.c b/testing/parameters.c
index 745f1c15648ecd3ca2bc6921e994407399724585..c3a7dcd900b63b4b002ec2d8e0132b97f7a84596 100644
--- a/testing/parameters.c
+++ b/testing/parameters.c
@@ -160,10 +160,10 @@ parameters_read_list( parameter_t *param,
previous = current;
current = current->next;
}
+ assert( current == NULL );
token = strtok_r( str, delim, &saveptr );
while ( token != NULL ) {
- assert( current == NULL );
current = calloc( 1, sizeof(vallist_t) );
/* Read the value */
@@ -179,7 +179,6 @@ parameters_read_list( parameter_t *param,
}
previous = current;
- current = NULL;
/* Move to the next token */
token = strtok_r( NULL, delim, &saveptr );
@@ -261,9 +260,9 @@ parameters_read_intrange( parameter_t *param,
previous = current;
current = current->next;
}
+ assert( current == NULL );
while ( start <= end ) {
- assert( current == NULL );
current = calloc( 1, sizeof(vallist_t) );
/* Read the value */
@@ -279,7 +278,6 @@ parameters_read_intrange( parameter_t *param,
}
previous = current;
- current = NULL;
start += step;
}
diff --git a/testing/testing_zgepdf_qr.c b/testing/testing_zgepdf_qr.c
index 5c3f0059685aa342762285003bab6b79b5b5f68a..8c6c297c08aae4b99533d1cac0047d624e6d7039 100644
--- a/testing/testing_zgepdf_qr.c
+++ b/testing/testing_zgepdf_qr.c
@@ -19,6 +19,7 @@
*
*/
#include <chameleon.h>
+#include <chameleon/getenv.h>
#if !defined(CHAMELEON_SIMULATION)
#include <coreblas/lapacke.h>
#include <coreblas/cblas.h>
@@ -58,7 +59,7 @@ testing_zgepdf_qr_desc( run_arg_list_t *args, int check )
CHAM_desc_t *descA1, *descA2, *descQ1, *descQ2;
CHAM_desc_t *TS1, *TT1, *TS2, *TT2;
libhqr_tree_t qrtreeT, qrtreeB;
- int zqdwh_opt_id = 1;
+ int zqdwh_opt_id = chameleon_env_on_off( "CHAMELEON_TESTING_GEPDF_OPTID", CHAMELEON_TRUE );
CHAMELEON_Set( CHAMELEON_TILE_SIZE, nb );
CHAMELEON_Set( CHAMELEON_INNER_BLOCK_SIZE, ib );