From a372581fe7c7a026559aa9969b8445995668bcdd Mon Sep 17 00:00:00 2001
From: Mathieu Faverge <mathieu.faverge@inria.fr>
Date: Tue, 16 Mar 2021 11:31:18 +0100
Subject: [PATCH] Backport StarPU codelet restructuration from recursive
---
runtime/starpu/codelets/codelet_zgemm.c | 137 +++++++++++++---------
runtime/starpu/codelets/codelet_zherk.c | 118 +++++++++++--------
runtime/starpu/codelets/codelet_zlacpy.c | 95 +++++++++------
runtime/starpu/codelets/codelet_zlascal.c | 73 ++++++++----
runtime/starpu/codelets/codelet_zlaset.c | 78 +++++++-----
runtime/starpu/codelets/codelet_zlauum.c | 69 +++++++----
runtime/starpu/codelets/codelet_zplghe.c | 74 +++++++-----
runtime/starpu/codelets/codelet_zplgsy.c | 79 ++++++++-----
runtime/starpu/codelets/codelet_zplrnt.c | 71 +++++++----
runtime/starpu/codelets/codelet_zpotrf.c | 90 ++++++++------
runtime/starpu/codelets/codelet_zsyrk.c | 124 ++++++++++++--------
runtime/starpu/codelets/codelet_ztradd.c | 82 ++++++++-----
runtime/starpu/codelets/codelet_ztrmm.c | 126 +++++++++++---------
runtime/starpu/codelets/codelet_ztrsm.c | 119 +++++++++++--------
runtime/starpu/codelets/codelet_ztrtri.c | 87 ++++++++------
15 files changed, 855 insertions(+), 567 deletions(-)
diff --git a/runtime/starpu/codelets/codelet_zgemm.c b/runtime/starpu/codelets/codelet_zgemm.c
index 9a3b9ea3f..0567065f3 100644
--- a/runtime/starpu/codelets/codelet_zgemm.c
+++ b/runtime/starpu/codelets/codelet_zgemm.c
@@ -25,9 +25,7 @@
#include "chameleon_starpu.h"
#include "runtime_codelet_z.h"
-#if !defined(CHAMELEON_SIMULATION)
-static void cl_zgemm_cpu_func(void *descr[], void *cl_arg)
-{
+struct cl_zgemm_args_s {
cham_trans_t transA;
cham_trans_t transB;
int m;
@@ -38,47 +36,54 @@ static void cl_zgemm_cpu_func(void *descr[], void *cl_arg)
CHAM_tile_t *tileB;
CHAMELEON_Complex64_t beta;
CHAM_tile_t *tileC;
+};
+
+#if !defined(CHAMELEON_SIMULATION)
+static void
+cl_zgemm_cpu_func( void *descr[], void *cl_arg )
+{
+ struct cl_zgemm_args_s clargs;
+ CHAM_tile_t *tileA;
+ CHAM_tile_t *tileB;
+ CHAM_tile_t *tileC;
tileA = cti_interface_get(descr[0]);
tileB = cti_interface_get(descr[1]);
tileC = cti_interface_get(descr[2]);
- starpu_codelet_unpack_args(cl_arg, &transA, &transB, &m, &n, &k, &alpha, &beta);
- TCORE_zgemm( transA, transB,
- m, n, k,
- alpha, tileA, tileB,
- beta, tileC );
+ starpu_codelet_unpack_args( cl_arg, &clargs );
+ TCORE_zgemm( clargs.transA, clargs.transB,
+ clargs.m, clargs.n, clargs.k,
+ clargs.alpha, tileA, tileB,
+ clargs.beta, tileC );
}
#ifdef CHAMELEON_USE_CUDA
-static void cl_zgemm_cuda_func(void *descr[], void *cl_arg)
+static void
+cl_zgemm_cuda_func( void *descr[], void *_cl_arg )
{
- cham_trans_t transA;
- cham_trans_t transB;
- int m;
- int n;
- int k;
- cuDoubleComplex alpha;
+ struct cl_zgemm_args_s clargs;
CHAM_tile_t *tileA;
CHAM_tile_t *tileB;
- cuDoubleComplex beta;
CHAM_tile_t *tileC;
tileA = cti_interface_get(descr[0]);
tileB = cti_interface_get(descr[1]);
tileC = cti_interface_get(descr[2]);
- starpu_codelet_unpack_args(cl_arg, &transA, &transB, &m, &n, &k, &alpha, &beta);
+ starpu_codelet_unpack_args( _cl_arg, &clargs );
RUNTIME_getStream( stream );
CUDA_zgemm(
- transA, transB,
- m, n, k,
- &alpha, tileA->mat, tileA->ld,
- tileB->mat, tileB->ld,
- &beta, tileC->mat, tileC->ld,
- stream);
+ clargs.transA, clargs.transB,
+ clargs.m, clargs.n, clargs.k,
+ (cuDoubleComplex*)&(clargs.alpha),
+ tileA->mat, tileA->ld,
+ tileB->mat, tileB->ld,
+ (cuDoubleComplex*)&(clargs.beta),
+ tileC->mat, tileC->ld,
+ stream );
#ifndef STARPU_CUDA_ASYNC
cudaStreamSynchronize( stream );
@@ -92,56 +97,72 @@ static void cl_zgemm_cuda_func(void *descr[], void *cl_arg)
/*
* Codelet definition
*/
-CODELETS(zgemm, cl_zgemm_cpu_func, cl_zgemm_cuda_func, STARPU_CUDA_ASYNC)
-
-/**
- *
- * @ingroup INSERT_TASK_Complex64_t
- *
- */
-void INSERT_TASK_zgemm(const RUNTIME_option_t *options,
- cham_trans_t transA, cham_trans_t transB,
- int m, int n, int k, int nb,
- CHAMELEON_Complex64_t alpha, const CHAM_desc_t *A, int Am, int An,
- const CHAM_desc_t *B, int Bm, int Bn,
- CHAMELEON_Complex64_t beta, const CHAM_desc_t *C, int Cm, int Cn)
+CODELETS( zgemm, cl_zgemm_cpu_func, cl_zgemm_cuda_func, STARPU_CUDA_ASYNC )
+
+void INSERT_TASK_zgemm( const RUNTIME_option_t *options,
+ cham_trans_t transA, cham_trans_t transB,
+ int m, int n, int k, int nb,
+ CHAMELEON_Complex64_t alpha, const CHAM_desc_t *A, int Am, int An,
+ const CHAM_desc_t *B, int Bm, int Bn,
+ CHAMELEON_Complex64_t beta, const CHAM_desc_t *C, int Cm, int Cn )
{
if ( alpha == 0. ) {
return INSERT_TASK_zlascal( options, ChamUpperLower, m, n, nb,
beta, C, Cm, Cn );
}
- (void)nb;
- struct starpu_codelet *codelet = &cl_zgemm;
- void (*callback)(void*) = options->profiling ? cl_zgemm_callback : NULL;
- starpu_option_request_t* schedopt = (starpu_option_request_t *)(options->request->schedopt);
- int workerid = (schedopt == NULL) ? -1 : schedopt->workerid;
- int accessC = ( beta == 0. ) ? STARPU_W : STARPU_RW;
-
+ struct cl_zgemm_args_s clargs = {
+ .transA = transA,
+ .transB = transB,
+ .m = m,
+ .n = n,
+ .k = k,
+ .alpha = alpha,
+ .tileA = A->get_blktile( A, Am, An ),
+ .tileB = B->get_blktile( B, Bm, Bn ),
+ .beta = beta,
+ .tileC = C->get_blktile( C, Cm, Cn )
+ };
+ void (*callback)(void*);
+ RUNTIME_request_t *request = options->request;
+ starpu_option_request_t *schedopt = (starpu_option_request_t *)(request->schedopt);
+ int workerid, accessC;
+ char *cl_name = "zgemm";
+
+ /* Handle cache */
CHAMELEON_BEGIN_ACCESS_DECLARATION;
CHAMELEON_ACCESS_R(A, Am, An);
CHAMELEON_ACCESS_R(B, Bm, Bn);
CHAMELEON_ACCESS_RW(C, Cm, Cn);
CHAMELEON_END_ACCESS_DECLARATION;
+ /* Callback fro profiling information */
+ callback = options->profiling ? cl_zgemm_callback : NULL;
+
+ /* Fix the worker id */
+ workerid = (schedopt == NULL) ? -1 : schedopt->workerid;
+
+ /* Reduce the C access if needed */
+ accessC = ( beta == 0. ) ? STARPU_W : STARPU_RW;
+
+ /* Insert the task */
rt_starpu_insert_task(
- codelet,
- STARPU_VALUE, &transA, sizeof(int),
- STARPU_VALUE, &transB, sizeof(int),
- STARPU_VALUE, &m, sizeof(int),
- STARPU_VALUE, &n, sizeof(int),
- STARPU_VALUE, &k, sizeof(int),
- STARPU_VALUE, &alpha, sizeof(CHAMELEON_Complex64_t),
- STARPU_R, RTBLKADDR(A, CHAMELEON_Complex64_t, Am, An),
- STARPU_R, RTBLKADDR(B, CHAMELEON_Complex64_t, Bm, Bn),
- STARPU_VALUE, &beta, sizeof(CHAMELEON_Complex64_t),
- accessC, RTBLKADDR(C, CHAMELEON_Complex64_t, Cm, Cn),
- STARPU_PRIORITY, options->priority,
- STARPU_CALLBACK, callback,
+ &cl_zgemm,
+ /* Task codelet arguments */
+ STARPU_VALUE, &clargs, sizeof(struct cl_zgemm_args_s),
+ STARPU_R, RTBLKADDR(A, CHAMELEON_Complex64_t, Am, An),
+ STARPU_R, RTBLKADDR(B, CHAMELEON_Complex64_t, Bm, Bn),
+ accessC, RTBLKADDR(C, CHAMELEON_Complex64_t, Cm, Cn),
+
+ /* Common task arguments */
+ STARPU_PRIORITY, options->priority,
+ STARPU_CALLBACK, callback,
STARPU_EXECUTE_ON_WORKER, workerid,
#if defined(CHAMELEON_CODELETS_HAVE_NAME)
- STARPU_NAME, "zgemm",
+ STARPU_NAME, cl_name,
#endif
- 0);
+ 0 );
+
+ (void)nb;
}
diff --git a/runtime/starpu/codelets/codelet_zherk.c b/runtime/starpu/codelets/codelet_zherk.c
index 1b9d39818..b224473cb 100644
--- a/runtime/starpu/codelets/codelet_zherk.c
+++ b/runtime/starpu/codelets/codelet_zherk.c
@@ -25,9 +25,7 @@
#include "chameleon_starpu.h"
#include "runtime_codelet_z.h"
-#if !defined(CHAMELEON_SIMULATION)
-static void cl_zherk_cpu_func(void *descr[], void *cl_arg)
-{
+struct cl_zherk_args_s {
cham_uplo_t uplo;
cham_trans_t trans;
int n;
@@ -36,41 +34,46 @@ static void cl_zherk_cpu_func(void *descr[], void *cl_arg)
CHAM_tile_t *tileA;
double beta;
CHAM_tile_t *tileC;
+};
+
+#if !defined(CHAMELEON_SIMULATION)
+static void
+cl_zherk_cpu_func(void *descr[], void *cl_arg)
+{
+ struct cl_zherk_args_s clargs;
+ CHAM_tile_t *tileA;
+ CHAM_tile_t *tileC;
tileA = cti_interface_get(descr[0]);
tileC = cti_interface_get(descr[1]);
- starpu_codelet_unpack_args(cl_arg, &uplo, &trans, &n, &k, &alpha, &beta);
- TCORE_zherk(uplo, trans,
- n, k,
- alpha, tileA,
- beta, tileC);
+ starpu_codelet_unpack_args( cl_arg, &clargs );
+ TCORE_zherk( clargs.uplo, clargs.trans, clargs.n, clargs.k,
+ clargs.alpha, tileA, clargs.beta, tileC );
}
-#ifdef CHAMELEON_USE_CUDA
-static void cl_zherk_cuda_func(void *descr[], void *cl_arg)
+#if defined(CHAMELEON_USE_CUDA)
+static void
+cl_zherk_cuda_func(void *descr[], void *cl_arg)
{
- cham_uplo_t uplo;
- cham_trans_t trans;
- int n;
- int k;
- double alpha;
+ struct cl_zherk_args_s clargs;
CHAM_tile_t *tileA;
- double beta;
CHAM_tile_t *tileC;
tileA = cti_interface_get(descr[0]);
tileC = cti_interface_get(descr[1]);
- starpu_codelet_unpack_args(cl_arg, &uplo, &trans, &n, &k, &alpha, &beta);
+ starpu_codelet_unpack_args( cl_arg, &clargs );
RUNTIME_getStream(stream);
CUDA_zherk(
- uplo, trans, n, k,
- &alpha, tileA->mat, tileA->ld,
- &beta, tileC->mat, tileC->ld,
- stream);
+ clargs.uplo, clargs.trans, clargs.n, clargs.k,
+ (cuDoubleComplex*)&(clargs.alpha),
+ tileA->mat, tileA->ld,
+ (cuDoubleComplex*)&(clargs.beta),
+ tileC->mat, tileC->ld,
+ stream );
#ifndef STARPU_CUDA_ASYNC
cudaStreamSynchronize( stream );
@@ -78,52 +81,73 @@ static void cl_zherk_cuda_func(void *descr[], void *cl_arg)
return;
}
-#endif /* CHAMELEON_USE_CUDA */
+#endif /* defined(CHAMELEON_USE_CUDA) */
#endif /* !defined(CHAMELEON_SIMULATION) */
/*
* Codelet definition
*/
-CODELETS(zherk, cl_zherk_cpu_func, cl_zherk_cuda_func, STARPU_CUDA_ASYNC)
+CODELETS( zherk, cl_zherk_cpu_func, cl_zherk_cuda_func, STARPU_CUDA_ASYNC )
-void INSERT_TASK_zherk(const RUNTIME_option_t *options,
- cham_uplo_t uplo, cham_trans_t trans,
- int n, int k, int nb,
- double alpha, const CHAM_desc_t *A, int Am, int An,
- double beta, const CHAM_desc_t *C, int Cm, int Cn)
+void INSERT_TASK_zherk( const RUNTIME_option_t *options,
+ cham_uplo_t uplo, cham_trans_t trans,
+ int n, int k, int nb,
+ double alpha, const CHAM_desc_t *A, int Am, int An,
+ double beta, const CHAM_desc_t *C, int Cm, int Cn )
{
if ( alpha == 0. ) {
return INSERT_TASK_zlascal( options, uplo, n, n, nb,
beta, C, Cm, Cn );
}
- (void)nb;
- struct starpu_codelet *codelet = &cl_zherk;
- void (*callback)(void*) = options->profiling ? cl_zherk_callback : NULL;
- starpu_option_request_t* schedopt = (starpu_option_request_t *)(options->request->schedopt);
- int workerid = (schedopt == NULL) ? -1 : schedopt->workerid;
- int accessC = ( beta == 0. ) ? STARPU_W : STARPU_RW;
+ struct cl_zherk_args_s clargs = {
+ .uplo = uplo,
+ .trans = trans,
+ .n = n,
+ .k = k,
+ .alpha = alpha,
+ .tileA = A->get_blktile( A, Am, An ),
+ .beta = beta,
+ .tileC = C->get_blktile( C, Cm, Cn ),
+ };
+ void (*callback)(void*);
+ RUNTIME_request_t *request = options->request;
+ starpu_option_request_t *schedopt = (starpu_option_request_t *)(request->schedopt);
+ int workerid, accessC;
+ char *cl_name = "zherk";
+ /* Handle cache */
CHAMELEON_BEGIN_ACCESS_DECLARATION;
CHAMELEON_ACCESS_R(A, Am, An);
CHAMELEON_ACCESS_RW(C, Cm, Cn);
CHAMELEON_END_ACCESS_DECLARATION;
+ /* Callback fro profiling information */
+ callback = options->profiling ? cl_zherk_callback : NULL;
+
+ /* Fix the worker id */
+ workerid = (schedopt == NULL) ? -1 : schedopt->workerid;
+
+ /* Reduce the C access if needed */
+ accessC = ( beta == 0. ) ? STARPU_W : STARPU_RW;
+
+ /* Insert the task */
rt_starpu_insert_task(
- codelet,
- STARPU_VALUE, &uplo, sizeof(int),
- STARPU_VALUE, &trans, sizeof(int),
- STARPU_VALUE, &n, sizeof(int),
- STARPU_VALUE, &k, sizeof(int),
- STARPU_VALUE, &alpha, sizeof(double),
- STARPU_R, RTBLKADDR(A, CHAMELEON_Complex64_t, Am, An),
- STARPU_VALUE, &beta, sizeof(double),
- accessC, RTBLKADDR(C, CHAMELEON_Complex64_t, Cm, Cn),
- STARPU_PRIORITY, options->priority,
- STARPU_CALLBACK, callback,
+ &cl_zherk,
+ /* Task codelet arguments */
+ STARPU_VALUE, &clargs, sizeof(struct cl_zherk_args_s),
+ STARPU_R, RTBLKADDR(A, CHAMELEON_Complex64_t, Am, An),
+ accessC, RTBLKADDR(C, CHAMELEON_Complex64_t, Cm, Cn),
+
+ /* Common task arguments */
+ STARPU_PRIORITY, options->priority,
+ STARPU_CALLBACK, callback,
STARPU_EXECUTE_ON_WORKER, workerid,
#if defined(CHAMELEON_CODELETS_HAVE_NAME)
- STARPU_NAME, "zherk",
+ STARPU_NAME, cl_name,
#endif
- 0);
+
+ 0 );
+
+ (void)nb;
}
diff --git a/runtime/starpu/codelets/codelet_zlacpy.c b/runtime/starpu/codelets/codelet_zlacpy.c
index bec2d9228..3fd038bde 100644
--- a/runtime/starpu/codelets/codelet_zlacpy.c
+++ b/runtime/starpu/codelets/codelet_zlacpy.c
@@ -12,8 +12,6 @@
* @brief Chameleon zlacpy StarPU codelet
*
* @version 1.0.0
- * @comment This file has been automatically generated
- * from Plasma 2.5.0 for CHAMELEON 0.9.2
* @author Julien Langou
* @author Henricus Bouwmeester
* @author Mathieu Faverge
@@ -27,70 +25,93 @@
#include "chameleon_starpu.h"
#include "runtime_codelet_z.h"
-#if !defined(CHAMELEON_SIMULATION)
-static void cl_zlacpy_cpu_func(void *descr[], void *cl_arg)
-{
+struct cl_zlacpy_args_s {
cham_uplo_t uplo;
- int M;
- int N;
+ int m;
+ int n;
int displA;
int displB;
CHAM_tile_t *tileA;
CHAM_tile_t *tileB;
- CHAMELEON_Complex64_t *A;
- CHAMELEON_Complex64_t *B;
+};
+
+#if !defined(CHAMELEON_SIMULATION)
+static void
+cl_zlacpy_cpu_func(void *descr[], void *cl_arg)
+{
+ struct cl_zlacpy_args_s clargs;
+ CHAM_tile_t *tileA;
+ CHAM_tile_t *tileB;
tileA = cti_interface_get(descr[0]);
tileB = cti_interface_get(descr[1]);
- starpu_codelet_unpack_args(cl_arg, &uplo, &M, &N, &displA, &displB);
-
- assert( tileA->format & CHAMELEON_TILE_FULLRANK );
- assert( tileB->format & CHAMELEON_TILE_FULLRANK );
-
- A = tileA->mat;
- B = tileB->mat;
- CORE_zlacpy( uplo, M, N, A + displA, tileA->ld, B + displB, tileB->ld );
+ starpu_codelet_unpack_args( cl_arg, &clargs );
+ assert( clargs.displA == 0 );
+ assert( clargs.displB == 0 );
+ /* A = tileA->mat; */
+ /* B = tileB->mat; */
+ /* CORE_zlacpy( uplo, M, N, A + displA, tileA->ld, B + displB, tileB->ld ); */
+ TCORE_zlacpy( clargs.uplo, clargs.m, clargs.n, tileA, tileB );
}
#endif /* !defined(CHAMELEON_SIMULATION) */
/*
* Codelet definition
*/
-CODELETS_CPU(zlacpy, cl_zlacpy_cpu_func)
+CODELETS_CPU( zlacpy, cl_zlacpy_cpu_func )
void INSERT_TASK_zlacpyx( const RUNTIME_option_t *options,
cham_uplo_t uplo, int m, int n, int nb,
int displA, const CHAM_desc_t *A, int Am, int An,
int displB, const CHAM_desc_t *B, int Bm, int Bn )
{
- (void)nb;
- struct starpu_codelet *codelet = &cl_zlacpy;
- void (*callback)(void*) = options->profiling ? cl_zlacpy_callback : NULL;
- starpu_option_request_t* schedopt = (starpu_option_request_t *)(options->request->schedopt);
- int workerid = (schedopt == NULL) ? -1 : schedopt->workerid;
+ struct cl_zlacpy_args_s clargs = {
+ .uplo = uplo,
+ .m = m,
+ .n = n,
+ .displA = displA,
+ .displB = displB,
+ .tileA = A->get_blktile( A, Am, An ),
+ .tileB = B->get_blktile( B, Bm, Bn ),
+ };
+ void (*callback)(void*);
+ RUNTIME_request_t *request = options->request;
+ starpu_option_request_t *schedopt = (starpu_option_request_t *)(request->schedopt);
+ int workerid;
+ char *cl_name = "zlacpy";
+ /* Handle cache */
CHAMELEON_BEGIN_ACCESS_DECLARATION;
- CHAMELEON_ACCESS_R( A, Am, An );
- CHAMELEON_ACCESS_W( B, Bm, Bn );
+ CHAMELEON_ACCESS_R(A, Am, An);
+ CHAMELEON_ACCESS_W(B, Bm, Bn);
CHAMELEON_END_ACCESS_DECLARATION;
+ /* Callback fro profiling information */
+ callback = options->profiling ? cl_zlacpy_callback : NULL;
+
+ /* Fix the worker id */
+ workerid = (schedopt == NULL) ? -1 : schedopt->workerid;
+
+ /* Insert the task */
rt_starpu_insert_task(
- codelet,
- STARPU_VALUE, &uplo, sizeof(cham_uplo_t),
- STARPU_VALUE, &m, sizeof(int),
- STARPU_VALUE, &n, sizeof(int),
- STARPU_VALUE, &displA, sizeof(int),
- STARPU_R, RTBLKADDR(A, CHAMELEON_Complex64_t, Am, An),
- STARPU_VALUE, &displB, sizeof(int),
- STARPU_W, RTBLKADDR(B, CHAMELEON_Complex64_t, Bm, Bn),
- STARPU_PRIORITY, options->priority,
- STARPU_CALLBACK, callback,
+ &cl_zlacpy,
+ /* Task codelet arguments */
+ STARPU_VALUE, &clargs, sizeof(struct cl_zlacpy_args_s),
+ STARPU_R, RTBLKADDR(A, CHAMELEON_Complex64_t, Am, An),
+ STARPU_W, RTBLKADDR(B, CHAMELEON_Complex64_t, Bm, Bn),
+
+ /* Common task arguments */
+ STARPU_PRIORITY, options->priority,
+ STARPU_CALLBACK, callback,
STARPU_EXECUTE_ON_WORKER, workerid,
#if defined(CHAMELEON_CODELETS_HAVE_NAME)
- STARPU_NAME, "zlacpy",
+ STARPU_NAME, cl_name,
#endif
- 0);
+
+ 0 );
+
+ (void)nb;
}
void INSERT_TASK_zlacpy( const RUNTIME_option_t *options,
diff --git a/runtime/starpu/codelets/codelet_zlascal.c b/runtime/starpu/codelets/codelet_zlascal.c
index bd6823a08..eb03a8685 100644
--- a/runtime/starpu/codelets/codelet_zlascal.c
+++ b/runtime/starpu/codelets/codelet_zlascal.c
@@ -21,33 +21,38 @@
#include "chameleon_starpu.h"
#include "runtime_codelet_z.h"
-#if !defined(CHAMELEON_SIMULATION)
-static void cl_zlascal_cpu_func(void *descr[], void *cl_arg)
-{
+struct cl_zlascal_args_s {
cham_uplo_t uplo;
- int M;
- int N;
+ int m;
+ int n;
CHAMELEON_Complex64_t alpha;
CHAM_tile_t *tileA;
+};
+
+#if !defined(CHAMELEON_SIMULATION)
+static void
+cl_zlascal_cpu_func( void *descr[], void *cl_arg )
+{
+ struct cl_zlascal_args_s clargs;
+ CHAM_tile_t *tileA;
tileA = cti_interface_get(descr[0]);
- starpu_codelet_unpack_args(cl_arg, &uplo, &M, &N, &alpha);
- TCORE_zlascal(uplo, M, N, alpha, tileA);
- return;
+ starpu_codelet_unpack_args( cl_arg, &clargs );
+ TCORE_zlascal( clargs.uplo, clargs.m, clargs.n, clargs.alpha, tileA );
}
#endif /* !defined(CHAMELEON_SIMULATION) */
/*
* Codelet definition
*/
-CODELETS_CPU(zlascal, cl_zlascal_cpu_func)
+CODELETS_CPU( zlascal, cl_zlascal_cpu_func )
void INSERT_TASK_zlascal( const RUNTIME_option_t *options,
cham_uplo_t uplo,
int m, int n, int nb,
CHAMELEON_Complex64_t alpha,
- const CHAM_desc_t *A, int Am, int An)
+ const CHAM_desc_t *A, int Am, int An )
{
if ( alpha == 0. ) {
return INSERT_TASK_zlaset( options, uplo, m, n,
@@ -57,28 +62,46 @@ void INSERT_TASK_zlascal( const RUNTIME_option_t *options,
return;
}
- (void)nb;
- struct starpu_codelet *codelet = &cl_zlascal;
- void (*callback)(void*) = options->profiling ? cl_zlascal_callback : NULL;
- starpu_option_request_t* schedopt = (starpu_option_request_t *)(options->request->schedopt);
- int workerid = (schedopt == NULL) ? -1 : schedopt->workerid;
+ struct cl_zlascal_args_s clargs = {
+ .uplo = uplo,
+ .m = m,
+ .n = n,
+ .alpha = alpha,
+ .tileA = A->get_blktile( A, Am, An ),
+ };
+ void (*callback)(void*);
+ RUNTIME_request_t *request = options->request;
+ starpu_option_request_t *schedopt = (starpu_option_request_t *)(request->schedopt);
+ int workerid;
+ char *cl_name = "zlascal";
+ /* Handle cache */
CHAMELEON_BEGIN_ACCESS_DECLARATION;
CHAMELEON_ACCESS_RW(A, Am, An);
CHAMELEON_END_ACCESS_DECLARATION;
+ /* Callback fro profiling information */
+ callback = options->profiling ? cl_zlascal_callback : NULL;
+
+ /* Fix the worker id */
+ workerid = (schedopt == NULL) ? -1 : schedopt->workerid;
+
+ /* Insert the task */
rt_starpu_insert_task(
- codelet,
- STARPU_VALUE, &uplo, sizeof(int),
- STARPU_VALUE, &m, sizeof(int),
- STARPU_VALUE, &n, sizeof(int),
- STARPU_VALUE, &alpha, sizeof(CHAMELEON_Complex64_t),
- STARPU_RW, RTBLKADDR(A, CHAMELEON_Complex64_t, Am, An),
- STARPU_PRIORITY, options->priority,
- STARPU_CALLBACK, callback,
+ &cl_zlascal,
+ /* Task codelet arguments */
+ STARPU_VALUE, &clargs, sizeof(struct cl_zlascal_args_s),
+ STARPU_RW, RTBLKADDR(A, CHAMELEON_Complex64_t, Am, An),
+
+ /* Common task arguments */
+ STARPU_PRIORITY, options->priority,
+ STARPU_CALLBACK, callback,
STARPU_EXECUTE_ON_WORKER, workerid,
#if defined(CHAMELEON_CODELETS_HAVE_NAME)
- STARPU_NAME, "zlascal",
+ STARPU_NAME, cl_name,
#endif
- 0);
+
+ 0 );
+
+ (void)nb;
}
diff --git a/runtime/starpu/codelets/codelet_zlaset.c b/runtime/starpu/codelets/codelet_zlaset.c
index c7661bdb8..0636686e3 100644
--- a/runtime/starpu/codelets/codelet_zlaset.c
+++ b/runtime/starpu/codelets/codelet_zlaset.c
@@ -12,8 +12,6 @@
* @brief Chameleon zlaset StarPU codelet
*
* @version 1.0.0
- * @comment This file has been automatically generated
- * from Plasma 2.5.0 for CHAMELEON 0.9.2
* @author Hatem Ltaief
* @author Mathieu Faverge
* @author Emmanuel Agullo
@@ -26,56 +24,78 @@
#include "chameleon_starpu.h"
#include "runtime_codelet_z.h"
-#if !defined(CHAMELEON_SIMULATION)
-static void cl_zlaset_cpu_func(void *descr[], void *cl_arg)
-{
+struct cl_zlaset_args_s {
cham_uplo_t uplo;
- int M;
- int N;
+ int m;
+ int n;
CHAMELEON_Complex64_t alpha;
CHAMELEON_Complex64_t beta;
CHAM_tile_t *tileA;
+};
+
+#if !defined(CHAMELEON_SIMULATION)
+static void
+cl_zlaset_cpu_func( void *descr[], void *cl_arg )
+{
+ struct cl_zlaset_args_s clargs;
+ CHAM_tile_t *tileA;
tileA = cti_interface_get(descr[0]);
- starpu_codelet_unpack_args(cl_arg, &uplo, &M, &N, &alpha, &beta);
- TCORE_zlaset(uplo, M, N, alpha, beta, tileA);
+ starpu_codelet_unpack_args( cl_arg, &clargs );
+ TCORE_zlaset( clargs.uplo, clargs.m, clargs.n, clargs.alpha, clargs.beta, tileA );
}
#endif /* !defined(CHAMELEON_SIMULATION) */
/*
* Codelet definition
*/
-CODELETS_CPU(zlaset, cl_zlaset_cpu_func)
+CODELETS_CPU( zlaset, cl_zlaset_cpu_func )
-void INSERT_TASK_zlaset(const RUNTIME_option_t *options,
- cham_uplo_t uplo, int M, int N,
- CHAMELEON_Complex64_t alpha, CHAMELEON_Complex64_t beta,
- const CHAM_desc_t *A, int Am, int An)
+void INSERT_TASK_zlaset( const RUNTIME_option_t *options,
+ cham_uplo_t uplo, int m, int n,
+ CHAMELEON_Complex64_t alpha, CHAMELEON_Complex64_t beta,
+ const CHAM_desc_t *A, int Am, int An )
{
+ struct cl_zlaset_args_s clargs = {
+ .uplo = uplo,
+ .m = m,
+ .n = n,
+ .alpha = alpha,
+ .beta = beta,
+ .tileA = A->get_blktile( A, Am, An ),
+ };
+ void (*callback)(void*);
+ RUNTIME_request_t *request = options->request;
+ starpu_option_request_t *schedopt = (starpu_option_request_t *)(request->schedopt);
+ int workerid;
+ char *cl_name = "zlaset";
- struct starpu_codelet *codelet = &cl_zlaset;
- void (*callback)(void*) = options->profiling ? cl_zlaset_callback : NULL;
- starpu_option_request_t* schedopt = (starpu_option_request_t *)(options->request->schedopt);
- int workerid = (schedopt == NULL) ? -1 : schedopt->workerid;
-
+ /* Handle cache */
CHAMELEON_BEGIN_ACCESS_DECLARATION;
CHAMELEON_ACCESS_W(A, Am, An);
CHAMELEON_END_ACCESS_DECLARATION;
+ /* Callback fro profiling information */
+ callback = options->profiling ? cl_zlaset_callback : NULL;
+
+ /* Fix the worker id */
+ workerid = (schedopt == NULL) ? -1 : schedopt->workerid;
+
+ /* Insert the task */
rt_starpu_insert_task(
- codelet,
- STARPU_VALUE, &uplo, sizeof(int),
- STARPU_VALUE, &M, sizeof(int),
- STARPU_VALUE, &N, sizeof(int),
- STARPU_VALUE, &alpha, sizeof(CHAMELEON_Complex64_t),
- STARPU_VALUE, &beta, sizeof(CHAMELEON_Complex64_t),
+ &cl_zlaset,
+ /* Task codelet arguments */
+ STARPU_VALUE, &clargs, sizeof(struct cl_zlaset_args_s),
STARPU_W, RTBLKADDR(A, CHAMELEON_Complex64_t, Am, An),
- STARPU_PRIORITY, options->priority,
- STARPU_CALLBACK, callback,
+
+ /* Common task arguments */
+ STARPU_PRIORITY, options->priority,
+ STARPU_CALLBACK, callback,
STARPU_EXECUTE_ON_WORKER, workerid,
#if defined(CHAMELEON_CODELETS_HAVE_NAME)
- STARPU_NAME, "zlaset",
+ STARPU_NAME, cl_name,
#endif
- 0);
+
+ 0 );
}
diff --git a/runtime/starpu/codelets/codelet_zlauum.c b/runtime/starpu/codelets/codelet_zlauum.c
index 0f3f1911b..9ff3fc6e9 100644
--- a/runtime/starpu/codelets/codelet_zlauum.c
+++ b/runtime/starpu/codelets/codelet_zlauum.c
@@ -12,8 +12,6 @@
* @brief Chameleon zlauum StarPU codelet
*
* @version 1.0.0
- * @comment This file has been automatically generated
- * from Plasma 2.5.0 for CHAMELEON 0.9.2
* @author Julien Langou
* @author Henricus Bouwmeester
* @author Mathieu Faverge
@@ -27,55 +25,74 @@
#include "chameleon_starpu.h"
#include "runtime_codelet_z.h"
+struct cl_zlauum_args_s {
+ cham_uplo_t uplo;
+ int n;
+ CHAM_tile_t *tileA;
+};
+
#if !defined(CHAMELEON_SIMULATION)
-static void cl_zlauum_cpu_func(void *descr[], void *cl_arg)
+static void
+cl_zlauum_cpu_func(void *descr[], void *cl_arg)
{
- cham_uplo_t uplo;
- int N;
+ struct cl_zlauum_args_s clargs;
CHAM_tile_t *tileA;
+ int info = 0;
tileA = cti_interface_get(descr[0]);
- starpu_codelet_unpack_args(cl_arg, &uplo, &N);
- TCORE_zlauum(uplo, N, tileA);
+ starpu_codelet_unpack_args( cl_arg, &clargs );
+ TCORE_zlauum( clargs.uplo, clargs.n, tileA );
}
#endif /* !defined(CHAMELEON_SIMULATION) */
/*
* Codelet definition
*/
-CODELETS_CPU(zlauum, cl_zlauum_cpu_func)
+CODELETS_CPU( zlauum, cl_zlauum_cpu_func )
-/**
- *
- * @ingroup INSERT_TASK_Complex64_t
- *
- */
void INSERT_TASK_zlauum( const RUNTIME_option_t *options,
cham_uplo_t uplo, int n, int nb,
const CHAM_desc_t *A, int Am, int An )
{
- (void)nb;
- struct starpu_codelet *codelet = &cl_zlauum;
- void (*callback)(void*) = options->profiling ? cl_zlauum_callback : NULL;
- starpu_option_request_t* schedopt = (starpu_option_request_t *)(options->request->schedopt);
- int workerid = (schedopt == NULL) ? -1 : schedopt->workerid;
+ struct cl_zlauum_args_s clargs = {
+ .uplo = uplo,
+ .n = n,
+ .tileA = A->get_blktile( A, Am, An ),
+ };
+ void (*callback)(void*);
+ RUNTIME_request_t *request = options->request;
+ starpu_option_request_t *schedopt = (starpu_option_request_t *)(request->schedopt);
+ int workerid;
+ char *cl_name = "zlauum";
+ /* Handle cache */
CHAMELEON_BEGIN_ACCESS_DECLARATION;
CHAMELEON_ACCESS_RW(A, Am, An);
CHAMELEON_END_ACCESS_DECLARATION;
+ /* Callback fro profiling information */
+ callback = options->profiling ? cl_zlauum_callback : NULL;
+
+ /* Fix the worker id */
+ workerid = (schedopt == NULL) ? -1 : schedopt->workerid;
+
+ /* Insert the task */
rt_starpu_insert_task(
- codelet,
- STARPU_VALUE, &uplo, sizeof(int),
- STARPU_VALUE, &n, sizeof(int),
- STARPU_RW, RTBLKADDR(A, CHAMELEON_Complex64_t, Am, An),
- STARPU_PRIORITY, options->priority,
- STARPU_CALLBACK, callback,
+ &cl_zlauum,
+ /* Task codelet arguments */
+ STARPU_VALUE, &clargs, sizeof(struct cl_zlauum_args_s),
+ STARPU_RW, RTBLKADDR(A, CHAMELEON_Complex64_t, Am, An),
+
+ /* Common task arguments */
+ STARPU_PRIORITY, options->priority,
+ STARPU_CALLBACK, callback,
STARPU_EXECUTE_ON_WORKER, workerid,
#if defined(CHAMELEON_CODELETS_HAVE_NAME)
- STARPU_NAME, "zlauum",
+ STARPU_NAME, cl_name,
#endif
- 0);
+ 0 );
+
+ (void)nb;
}
diff --git a/runtime/starpu/codelets/codelet_zplghe.c b/runtime/starpu/codelets/codelet_zplghe.c
index 87c071e44..4c04611cc 100644
--- a/runtime/starpu/codelets/codelet_zplghe.c
+++ b/runtime/starpu/codelets/codelet_zplghe.c
@@ -12,8 +12,6 @@
* @brief Chameleon zplghe StarPU codelet
*
* @version 1.0.0
- * @comment This file has been automatically generated
- * from Plasma 2.5.0 for CHAMELEON 0.9.2
* @author Piotr Luszczek
* @author Pierre Lemarinier
* @author Mathieu Faverge
@@ -27,11 +25,7 @@
#include "chameleon_starpu.h"
#include "runtime_codelet_z.h"
-/* cl_zplghe_cpu_func - Generate a tile for random hermitian (positive definite if bump is large enough) matrix. */
-
-#if !defined(CHAMELEON_SIMULATION)
-static void cl_zplghe_cpu_func(void *descr[], void *cl_arg)
-{
+struct cl_zplghe_args_s {
double bump;
int m;
int n;
@@ -40,47 +34,73 @@ static void cl_zplghe_cpu_func(void *descr[], void *cl_arg)
int m0;
int n0;
unsigned long long int seed;
+};
+
+#if !defined(CHAMELEON_SIMULATION)
+static void cl_zplghe_cpu_func(void *descr[], void *cl_arg)
+{
+ struct cl_zplghe_args_s clargs;
+ CHAM_tile_t *tileA;
tileA = cti_interface_get(descr[0]);
- starpu_codelet_unpack_args(cl_arg, &bump, &m, &n, &bigM, &m0, &n0, &seed );
- TCORE_zplghe( bump, m, n, tileA, bigM, m0, n0, seed );
+ starpu_codelet_unpack_args( cl_arg, &clargs );
+ TCORE_zplghe( clargs.bump, clargs.m, clargs.n, tileA,
+ clargs.bigM, clargs.m0, clargs.n0, clargs.seed );
}
#endif /* !defined(CHAMELEON_SIMULATION) */
/*
* Codelet definition
*/
-CODELETS_CPU(zplghe, cl_zplghe_cpu_func)
+CODELETS_CPU( zplghe, cl_zplghe_cpu_func )
void INSERT_TASK_zplghe( const RUNTIME_option_t *options,
double bump, int m, int n, const CHAM_desc_t *A, int Am, int An,
int bigM, int m0, int n0, unsigned long long int seed )
{
- struct starpu_codelet *codelet = &cl_zplghe;
- void (*callback)(void*) = options->profiling ? cl_zplghe_callback : NULL;
- starpu_option_request_t* schedopt = (starpu_option_request_t *)(options->request->schedopt);
- int workerid = (schedopt == NULL) ? -1 : schedopt->workerid;
+ struct cl_zplghe_args_s clargs = {
+ .bump = bump,
+ .m = m,
+ .n = n,
+ .tileA = A->get_blktile( A, Am, An ),
+ .bigM = bigM,
+ .m0 = m0,
+ .n0 = n0,
+ .seed = seed,
+ };
+ void (*callback)(void*);
+ RUNTIME_request_t *request = options->request;
+ starpu_option_request_t *schedopt = (starpu_option_request_t *)(request->schedopt);
+ int workerid;
+ char *cl_name = "zplghe";
+ /* Handle cache */
CHAMELEON_BEGIN_ACCESS_DECLARATION;
CHAMELEON_ACCESS_W(A, Am, An);
CHAMELEON_END_ACCESS_DECLARATION;
+ /* Callback fro profiling information */
+ callback = options->profiling ? cl_zplghe_callback : NULL;
+
+ /* Fix the worker id */
+ workerid = (schedopt == NULL) ? -1 : schedopt->workerid;
+
+ /* Insert the task */
rt_starpu_insert_task(
- codelet,
- STARPU_VALUE, &bump, sizeof(double),
- STARPU_VALUE, &m, sizeof(int),
- STARPU_VALUE, &n, sizeof(int),
- STARPU_W, RTBLKADDR(A, CHAMELEON_Complex64_t, Am, An),
- STARPU_VALUE, &bigM, sizeof(int),
- STARPU_VALUE, &m0, sizeof(int),
- STARPU_VALUE, &n0, sizeof(int),
- STARPU_VALUE, &seed, sizeof(unsigned long long int),
- STARPU_PRIORITY, options->priority,
- STARPU_CALLBACK, callback,
+ &cl_zplghe,
+ /* Task codelet arguments */
+ STARPU_VALUE, &clargs, sizeof(struct cl_zplghe_args_s),
+ STARPU_W, RTBLKADDR(A, CHAMELEON_Complex64_t, Am, An),
+
+ /* Common task arguments */
+ STARPU_PRIORITY, options->priority,
+ STARPU_CALLBACK, callback,
STARPU_EXECUTE_ON_WORKER, workerid,
#if defined(CHAMELEON_CODELETS_HAVE_NAME)
- STARPU_NAME, "zplghe",
+ STARPU_NAME, cl_name,
#endif
- 0);
+
+ 0 );
+
}
diff --git a/runtime/starpu/codelets/codelet_zplgsy.c b/runtime/starpu/codelets/codelet_zplgsy.c
index 748e78d33..57aec1bf2 100644
--- a/runtime/starpu/codelets/codelet_zplgsy.c
+++ b/runtime/starpu/codelets/codelet_zplgsy.c
@@ -12,8 +12,6 @@
* @brief Chameleon zplgsy StarPU codelet
*
* @version 1.0.0
- * @comment This file has been automatically generated
- * from Plasma 2.5.0 for CHAMELEON 0.9.2
* @author Piotr Luszczek
* @author Pierre Lemarinier
* @author Mathieu Faverge
@@ -27,11 +25,7 @@
#include "chameleon_starpu.h"
#include "runtime_codelet_z.h"
-/* cl_zplgsy_cpu_func - Generate a tile for random symmetric (positive definite if 'bump' is large enough) matrix. */
-
-#if !defined(CHAMELEON_SIMULATION)
-static void cl_zplgsy_cpu_func(void *descr[], void *cl_arg)
-{
+struct cl_zplgsy_args_s {
CHAMELEON_Complex64_t bump;
int m;
int n;
@@ -40,48 +34,73 @@ static void cl_zplgsy_cpu_func(void *descr[], void *cl_arg)
int m0;
int n0;
unsigned long long int seed;
+};
+
+#if !defined(CHAMELEON_SIMULATION)
+static void cl_zplgsy_cpu_func(void *descr[], void *cl_arg)
+{
+ struct cl_zplgsy_args_s clargs;
+ CHAM_tile_t *tileA;
tileA = cti_interface_get(descr[0]);
- starpu_codelet_unpack_args(cl_arg, &bump, &m, &n, &bigM, &m0, &n0, &seed );
- TCORE_zplgsy( bump, m, n, tileA, bigM, m0, n0, seed );
+ starpu_codelet_unpack_args( cl_arg, &clargs );
+ TCORE_zplgsy( clargs.bump, clargs.m, clargs.n, tileA,
+ clargs.bigM, clargs.m0, clargs.n0, clargs.seed );
}
#endif /* !defined(CHAMELEON_SIMULATION) */
/*
* Codelet definition
*/
-CODELETS_CPU(zplgsy, cl_zplgsy_cpu_func)
+CODELETS_CPU( zplgsy, cl_zplgsy_cpu_func )
void INSERT_TASK_zplgsy( const RUNTIME_option_t *options,
- CHAMELEON_Complex64_t bump, int m, int n, const CHAM_desc_t *A, int Am, int An,
- int bigM, int m0, int n0, unsigned long long int seed )
+ CHAMELEON_Complex64_t bump, int m, int n, const CHAM_desc_t *A, int Am, int An,
+ int bigM, int m0, int n0, unsigned long long int seed )
{
+ struct cl_zplgsy_args_s clargs = {
+ .bump = bump,
+ .m = m,
+ .n = n,
+ .tileA = A->get_blktile( A, Am, An ),
+ .bigM = bigM,
+ .m0 = m0,
+ .n0 = n0,
+ .seed = seed,
+ };
+ void (*callback)(void*);
+ RUNTIME_request_t *request = options->request;
+ starpu_option_request_t *schedopt = (starpu_option_request_t *)(request->schedopt);
+ int workerid;
+ char *cl_name = "zplgsy";
- struct starpu_codelet *codelet = &cl_zplgsy;
- void (*callback)(void*) = options->profiling ? cl_zplgsy_callback : NULL;
- starpu_option_request_t* schedopt = (starpu_option_request_t *)(options->request->schedopt);
- int workerid = (schedopt == NULL) ? -1 : schedopt->workerid;
-
+ /* Handle cache */
CHAMELEON_BEGIN_ACCESS_DECLARATION;
CHAMELEON_ACCESS_W(A, Am, An);
CHAMELEON_END_ACCESS_DECLARATION;
+ /* Callback fro profiling information */
+ callback = options->profiling ? cl_zplgsy_callback : NULL;
+
+ /* Fix the worker id */
+ workerid = (schedopt == NULL) ? -1 : schedopt->workerid;
+
+ /* Insert the task */
rt_starpu_insert_task(
- codelet,
- STARPU_VALUE, &bump, sizeof(CHAMELEON_Complex64_t),
- STARPU_VALUE, &m, sizeof(int),
- STARPU_VALUE, &n, sizeof(int),
- STARPU_W, RTBLKADDR(A, CHAMELEON_Complex64_t, Am, An),
- STARPU_VALUE, &bigM, sizeof(int),
- STARPU_VALUE, &m0, sizeof(int),
- STARPU_VALUE, &n0, sizeof(int),
- STARPU_VALUE, &seed, sizeof(unsigned long long int),
- STARPU_PRIORITY, options->priority,
- STARPU_CALLBACK, callback,
+ &cl_zplgsy,
+ /* Task codelet arguments */
+ STARPU_VALUE, &clargs, sizeof(struct cl_zplgsy_args_s),
+ STARPU_W, RTBLKADDR(A, CHAMELEON_Complex64_t, Am, An),
+
+ /* Common task arguments */
+ STARPU_PRIORITY, options->priority,
+ STARPU_CALLBACK, callback,
STARPU_EXECUTE_ON_WORKER, workerid,
#if defined(CHAMELEON_CODELETS_HAVE_NAME)
- STARPU_NAME, "zplgsy",
+ STARPU_NAME, cl_name,
#endif
- 0);
+
+ 0 );
+
}
diff --git a/runtime/starpu/codelets/codelet_zplrnt.c b/runtime/starpu/codelets/codelet_zplrnt.c
index 1376e950c..a4eef0c8a 100644
--- a/runtime/starpu/codelets/codelet_zplrnt.c
+++ b/runtime/starpu/codelets/codelet_zplrnt.c
@@ -12,8 +12,6 @@
* @brief Chameleon zplrnt StarPU codelet
*
* @version 1.0.0
- * @comment This file has been automatically generated
- * from Plasma 2.5.0 for CHAMELEON 0.9.2
* @author Piotr Luszczek
* @author Pierre Lemarinier
* @author Mathieu Faverge
@@ -27,9 +25,7 @@
#include "chameleon_starpu.h"
#include "runtime_codelet_z.h"
-#if !defined(CHAMELEON_SIMULATION)
-static void cl_zplrnt_cpu_func(void *descr[], void *cl_arg)
-{
+struct cl_zplrnt_args_s {
int m;
int n;
CHAM_tile_t *tileA;
@@ -37,47 +33,72 @@ static void cl_zplrnt_cpu_func(void *descr[], void *cl_arg)
int m0;
int n0;
unsigned long long int seed;
+};
+
+#if !defined(CHAMELEON_SIMULATION)
+static void
+cl_zplrnt_cpu_func(void *descr[], void *cl_arg)
+{
+ struct cl_zplrnt_args_s clargs;
+ CHAM_tile_t *tileA;
tileA = cti_interface_get(descr[0]);
- starpu_codelet_unpack_args(cl_arg, &m, &n, &bigM, &m0, &n0, &seed );
- TCORE_zplrnt( m, n, tileA, bigM, m0, n0, seed );
+ starpu_codelet_unpack_args( cl_arg, &clargs );
+ TCORE_zplrnt( clargs.m, clargs.n, tileA,
+ clargs.bigM, clargs.m0, clargs.n0, clargs.seed );
}
#endif /* !defined(CHAMELEON_SIMULATION) */
/*
* Codelet definition
*/
-CODELETS_CPU(zplrnt, cl_zplrnt_cpu_func)
+CODELETS_CPU( zplrnt, cl_zplrnt_cpu_func )
void INSERT_TASK_zplrnt( const RUNTIME_option_t *options,
int m, int n, const CHAM_desc_t *A, int Am, int An,
int bigM, int m0, int n0, unsigned long long int seed )
{
+ struct cl_zplrnt_args_s clargs = {
+ .m = m,
+ .n = n,
+ .tileA = A->get_blktile( A, Am, An ),
+ .bigM = bigM,
+ .m0 = m0,
+ .n0 = n0,
+ .seed = seed,
+ };
+ void (*callback)(void*);
+ RUNTIME_request_t *request = options->request;
+ starpu_option_request_t *schedopt = (starpu_option_request_t *)(request->schedopt);
+ int workerid;
+ char *cl_name = "zplrnt";
- struct starpu_codelet *codelet = &cl_zplrnt;
- void (*callback)(void*) = options->profiling ? cl_zplrnt_callback : NULL;
- starpu_option_request_t* schedopt = (starpu_option_request_t *)(options->request->schedopt);
- int workerid = (schedopt == NULL) ? -1 : schedopt->workerid;
-
+ /* Handle cache */
CHAMELEON_BEGIN_ACCESS_DECLARATION;
CHAMELEON_ACCESS_W(A, Am, An);
CHAMELEON_END_ACCESS_DECLARATION;
+ /* Callback fro profiling information */
+ callback = options->profiling ? cl_zplrnt_callback : NULL;
+
+ /* Fix the worker id */
+ workerid = (schedopt == NULL) ? -1 : schedopt->workerid;
+
+ /* Insert the task */
rt_starpu_insert_task(
- codelet,
- STARPU_VALUE, &m, sizeof(int),
- STARPU_VALUE, &n, sizeof(int),
- STARPU_W, RTBLKADDR(A, CHAMELEON_Complex64_t, Am, An),
- STARPU_VALUE, &bigM, sizeof(int),
- STARPU_VALUE, &m0, sizeof(int),
- STARPU_VALUE, &n0, sizeof(int),
- STARPU_VALUE, &seed, sizeof(unsigned long long int),
- STARPU_PRIORITY, options->priority,
- STARPU_CALLBACK, callback,
+ &cl_zplrnt,
+ /* Task codelet arguments */
+ STARPU_VALUE, &clargs, sizeof(struct cl_zplrnt_args_s),
+ STARPU_W, RTBLKADDR(A, CHAMELEON_Complex64_t, Am, An),
+
+ /* Common task arguments */
+ STARPU_PRIORITY, options->priority,
+ STARPU_CALLBACK, callback,
STARPU_EXECUTE_ON_WORKER, workerid,
#if defined(CHAMELEON_CODELETS_HAVE_NAME)
- STARPU_NAME, "zplrnt",
+ STARPU_NAME, cl_name,
#endif
- 0);
+
+ 0 );
}
diff --git a/runtime/starpu/codelets/codelet_zpotrf.c b/runtime/starpu/codelets/codelet_zpotrf.c
index 2337b00fd..93f4f640d 100644
--- a/runtime/starpu/codelets/codelet_zpotrf.c
+++ b/runtime/starpu/codelets/codelet_zpotrf.c
@@ -12,8 +12,6 @@
* @brief Chameleon zpotrf StarPU codelet
*
* @version 1.0.0
- * @comment This file has been automatically generated
- * from Plasma 2.5.0 for CHAMELEON 0.9.2
* @author Hatem Ltaief
* @author Jakub Kurzak
* @author Mathieu Faverge
@@ -27,24 +25,30 @@
#include "chameleon_starpu.h"
#include "runtime_codelet_z.h"
-#if !defined(CHAMELEON_SIMULATION)
-static void cl_zpotrf_cpu_func(void *descr[], void *cl_arg)
-{
+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;
+};
+
+#if !defined(CHAMELEON_SIMULATION)
+static void
+cl_zpotrf_cpu_func(void *descr[], void *cl_arg)
+{
+ struct cl_zpotrf_args_s clargs;
+ CHAM_tile_t *tileA;
int info = 0;
tileA = cti_interface_get(descr[0]);
- starpu_codelet_unpack_args(cl_arg, &uplo, &n, &iinfo, &sequence, &request);
- TCORE_zpotrf(uplo, n, tileA, &info);
+ starpu_codelet_unpack_args( cl_arg, &clargs );
+ TCORE_zpotrf( clargs.uplo, clargs.n, tileA, &info );
- if ( (sequence->status == CHAMELEON_SUCCESS) && (info != 0) ) {
- RUNTIME_sequence_flush( NULL, sequence, request, iinfo+info );
+ if ( (clargs.sequence->status == CHAMELEON_SUCCESS) && (info != 0) ) {
+ RUNTIME_sequence_flush( NULL, clargs.sequence, clargs.request, clargs.iinfo+info );
}
}
#endif /* !defined(CHAMELEON_SIMULATION) */
@@ -52,46 +56,54 @@ static void cl_zpotrf_cpu_func(void *descr[], void *cl_arg)
/*
* Codelet definition
*/
-#if defined(CHAMELEON_SIMULATION) && defined(CHAMELEON_SIMULATION_EXTENDED)
-CODELETS( zpotrf, cl_zpotrf_cpu_func, cl_zpotrf_cuda_func, STARPU_CUDA_ASYNC )
-#else
CODELETS_CPU( zpotrf, cl_zpotrf_cpu_func )
-#endif
-/**
- *
- * @ingroup INSERT_TASK_Complex64_t
- *
- */
-void INSERT_TASK_zpotrf(const RUNTIME_option_t *options,
- cham_uplo_t uplo, int n, int nb,
- const CHAM_desc_t *A, int Am, int An,
- int iinfo)
+void INSERT_TASK_zpotrf( const RUNTIME_option_t *options,
+ cham_uplo_t uplo, int n, int nb,
+ const CHAM_desc_t *A, int Am, int An,
+ int iinfo )
{
- (void)nb;
- struct starpu_codelet *codelet = &cl_zpotrf;
- void (*callback)(void*) = options->profiling ? cl_zpotrf_callback : NULL;
- starpu_option_request_t* schedopt = (starpu_option_request_t *)(options->request->schedopt);
- int workerid = (schedopt == NULL) ? -1 : schedopt->workerid;
+ struct cl_zpotrf_args_s clargs = {
+ .uplo = uplo,
+ .n = n,
+ .tileA = A->get_blktile( A, Am, An ),
+ .iinfo = iinfo,
+ .sequence = options->sequence,
+ .request = options->request,
+ };
+ void (*callback)(void*);
+ RUNTIME_request_t *request = options->request;
+ starpu_option_request_t *schedopt = (starpu_option_request_t *)(request->schedopt);
+ int workerid;
+ char *cl_name = "zpotrf";
+ /* Handle cache */
CHAMELEON_BEGIN_ACCESS_DECLARATION;
CHAMELEON_ACCESS_RW(A, Am, An);
CHAMELEON_END_ACCESS_DECLARATION;
+ /* Callback fro profiling information */
+ callback = options->profiling ? cl_zpotrf_callback : NULL;
+
+ /* Fix the worker id */
+ workerid = (schedopt == NULL) ? -1 : schedopt->workerid;
+
+ /* Insert the task */
rt_starpu_insert_task(
- codelet,
- STARPU_VALUE, &uplo, sizeof(int),
- STARPU_VALUE, &n, sizeof(int),
- STARPU_RW, RTBLKADDR(A, CHAMELEON_Complex64_t, Am, An),
- STARPU_VALUE, &iinfo, sizeof(int),
- STARPU_VALUE, &(options->sequence), sizeof(RUNTIME_sequence_t*),
- STARPU_VALUE, &(options->request), sizeof(RUNTIME_request_t*),
- /* STARPU_SCRATCH, options->ws_worker, */
- STARPU_PRIORITY, options->priority,
- STARPU_CALLBACK, callback,
+ &cl_zpotrf,
+ /* Task codelet arguments */
+ STARPU_VALUE, &clargs, sizeof(struct cl_zpotrf_args_s),
+ STARPU_RW, RTBLKADDR(A, CHAMELEON_Complex64_t, Am, An),
+
+ /* Common task arguments */
+ STARPU_PRIORITY, options->priority,
+ STARPU_CALLBACK, callback,
STARPU_EXECUTE_ON_WORKER, workerid,
#if defined(CHAMELEON_CODELETS_HAVE_NAME)
- STARPU_NAME, "zpotrf",
+ STARPU_NAME, cl_name,
#endif
- 0);
+
+ 0 );
+
+ (void)nb;
}
diff --git a/runtime/starpu/codelets/codelet_zsyrk.c b/runtime/starpu/codelets/codelet_zsyrk.c
index 9552eee9c..0cba41cab 100644
--- a/runtime/starpu/codelets/codelet_zsyrk.c
+++ b/runtime/starpu/codelets/codelet_zsyrk.c
@@ -25,9 +25,7 @@
#include "chameleon_starpu.h"
#include "runtime_codelet_z.h"
-#if !defined(CHAMELEON_SIMULATION)
-static void cl_zsyrk_cpu_func(void *descr[], void *cl_arg)
-{
+struct cl_zsyrk_args_s {
cham_uplo_t uplo;
cham_trans_t trans;
int n;
@@ -36,41 +34,46 @@ static void cl_zsyrk_cpu_func(void *descr[], void *cl_arg)
CHAM_tile_t *tileA;
CHAMELEON_Complex64_t beta;
CHAM_tile_t *tileC;
+};
+
+#if !defined(CHAMELEON_SIMULATION)
+static void
+cl_zsyrk_cpu_func(void *descr[], void *cl_arg)
+{
+ struct cl_zsyrk_args_s clargs;
+ CHAM_tile_t *tileA;
+ CHAM_tile_t *tileC;
tileA = cti_interface_get(descr[0]);
tileC = cti_interface_get(descr[1]);
- starpu_codelet_unpack_args(cl_arg, &uplo, &trans, &n, &k, &alpha, &beta);
- TCORE_zsyrk(uplo, trans,
- n, k,
- alpha, tileA,
- beta, tileC);
+ starpu_codelet_unpack_args( cl_arg, &clargs );
+ TCORE_zsyrk( clargs.uplo, clargs.trans, clargs.n, clargs.k,
+ clargs.alpha, tileA, clargs.beta, tileC );
}
-#ifdef CHAMELEON_USE_CUDA
-static void cl_zsyrk_cuda_func(void *descr[], void *cl_arg)
+#if defined(CHAMELEON_USE_CUDA)
+static void
+cl_zsyrk_cuda_func(void *descr[], void *cl_arg)
{
- cham_uplo_t uplo;
- cham_trans_t trans;
- int n;
- int k;
- cuDoubleComplex alpha;
+ struct cl_zsyrk_args_s clargs;
CHAM_tile_t *tileA;
- cuDoubleComplex beta;
CHAM_tile_t *tileC;
tileA = cti_interface_get(descr[0]);
tileC = cti_interface_get(descr[1]);
- starpu_codelet_unpack_args(cl_arg, &uplo, &trans, &n, &k, &alpha, &beta);
+ starpu_codelet_unpack_args( cl_arg, &clargs );
RUNTIME_getStream(stream);
CUDA_zsyrk(
- uplo, trans, n, k,
- &alpha, tileA->mat, tileA->ld,
- &beta, tileC->mat, tileC->ld,
- stream);
+ clargs.uplo, clargs.trans, clargs.n, clargs.k,
+ (cuDoubleComplex*)&(clargs.alpha),
+ tileA->mat, tileA->ld,
+ (cuDoubleComplex*)&(clargs.beta),
+ tileC->mat, tileC->ld,
+ stream );
#ifndef STARPU_CUDA_ASYNC
cudaStreamSynchronize( stream );
@@ -78,57 +81,74 @@ static void cl_zsyrk_cuda_func(void *descr[], void *cl_arg)
return;
}
-#endif /* CHAMELEON_USE_CUDA */
+#endif /* defined(CHAMELEON_USE_CUDA) */
#endif /* !defined(CHAMELEON_SIMULATION) */
/*
* Codelet definition
*/
-CODELETS(zsyrk, cl_zsyrk_cpu_func, cl_zsyrk_cuda_func, STARPU_CUDA_ASYNC)
+CODELETS( zsyrk, cl_zsyrk_cpu_func, cl_zsyrk_cuda_func, STARPU_CUDA_ASYNC )
-/**
- *
- * @ingroup INSERT_TASK_Complex64_t
- *
- */
-void INSERT_TASK_zsyrk(const RUNTIME_option_t *options,
- cham_uplo_t uplo, cham_trans_t trans,
- int n, int k, int nb,
- CHAMELEON_Complex64_t alpha, const CHAM_desc_t *A, int Am, int An,
- CHAMELEON_Complex64_t beta, const CHAM_desc_t *C, int Cm, int Cn)
+void INSERT_TASK_zsyrk( const RUNTIME_option_t *options,
+ cham_uplo_t uplo, cham_trans_t trans,
+ int n, int k, int nb,
+ CHAMELEON_Complex64_t alpha, const CHAM_desc_t *A, int Am, int An,
+ CHAMELEON_Complex64_t beta, const CHAM_desc_t *C, int Cm, int Cn )
{
if ( alpha == 0. ) {
return INSERT_TASK_zlascal( options, uplo, n, n, nb,
beta, C, Cm, Cn );
}
- (void)nb;
- struct starpu_codelet *codelet = &cl_zsyrk;
- void (*callback)(void*) = options->profiling ? cl_zsyrk_callback : NULL;
- starpu_option_request_t* schedopt = (starpu_option_request_t *)(options->request->schedopt);
- int workerid = (schedopt == NULL) ? -1 : schedopt->workerid;
- int accessC = ( beta == 0. ) ? STARPU_W : STARPU_RW;
+ struct cl_zsyrk_args_s clargs = {
+ .uplo = uplo,
+ .trans = trans,
+ .n = n,
+ .k = k,
+ .alpha = alpha,
+ .tileA = A->get_blktile( A, Am, An ),
+ .beta = beta,
+ .tileC = C->get_blktile( C, Cm, Cn ),
+ };
+ void (*callback)(void*);
+ RUNTIME_request_t *request = options->request;
+ starpu_option_request_t *schedopt = (starpu_option_request_t *)(request->schedopt);
+ int workerid, accessC;
+ char *cl_name = "zsyrk";
+ /* Handle cache */
CHAMELEON_BEGIN_ACCESS_DECLARATION;
CHAMELEON_ACCESS_R(A, Am, An);
CHAMELEON_ACCESS_RW(C, Cm, Cn);
CHAMELEON_END_ACCESS_DECLARATION;
+ /* Callback fro profiling information */
+ callback = options->profiling ? cl_zsyrk_callback : NULL;
+
+ /* Fix the worker id */
+ workerid = (schedopt == NULL) ? -1 : schedopt->workerid;
+
+ /* Reduce the C access if needed */
+ accessC = ( beta == 0. ) ? STARPU_W : STARPU_RW;
+
+ /* Insert the task */
rt_starpu_insert_task(
- codelet,
- STARPU_VALUE, &uplo, sizeof(int),
- STARPU_VALUE, &trans, sizeof(int),
- STARPU_VALUE, &n, sizeof(int),
- STARPU_VALUE, &k, sizeof(int),
- STARPU_VALUE, &alpha, sizeof(CHAMELEON_Complex64_t),
- STARPU_R, RTBLKADDR(A, CHAMELEON_Complex64_t, Am, An),
- STARPU_VALUE, &beta, sizeof(CHAMELEON_Complex64_t),
- accessC, RTBLKADDR(C, CHAMELEON_Complex64_t, Cm, Cn),
- STARPU_PRIORITY, options->priority,
- STARPU_CALLBACK, callback,
+ &cl_zsyrk,
+
+ /* Task codelet arguments */
+ STARPU_VALUE, &clargs, sizeof(struct cl_zsyrk_args_s),
+ STARPU_R, RTBLKADDR(A, CHAMELEON_Complex64_t, Am, An),
+ accessC, RTBLKADDR(C, CHAMELEON_Complex64_t, Cm, Cn),
+
+ /* Common task arguments */
+ STARPU_PRIORITY, options->priority,
+ STARPU_CALLBACK, callback,
STARPU_EXECUTE_ON_WORKER, workerid,
#if defined(CHAMELEON_CODELETS_HAVE_NAME)
- STARPU_NAME, "zsyrk",
+ STARPU_NAME, cl_name,
#endif
- 0);
+
+ 0 );
+
+ (void)nb;
}
diff --git a/runtime/starpu/codelets/codelet_ztradd.c b/runtime/starpu/codelets/codelet_ztradd.c
index 689cb6254..2cbaa3aef 100644
--- a/runtime/starpu/codelets/codelet_ztradd.c
+++ b/runtime/starpu/codelets/codelet_ztradd.c
@@ -21,30 +21,38 @@
#include "chameleon_starpu.h"
#include "runtime_codelet_z.h"
-#if !defined(CHAMELEON_SIMULATION)
-static void cl_ztradd_cpu_func(void *descr[], void *cl_arg)
-{
+struct cl_ztradd_args_s {
cham_uplo_t uplo;
cham_trans_t trans;
- int M;
- int N;
+ int m;
+ int n;
CHAMELEON_Complex64_t alpha;
CHAM_tile_t *tileA;
CHAMELEON_Complex64_t beta;
CHAM_tile_t *tileB;
+};
+
+#if !defined(CHAMELEON_SIMULATION)
+static void
+cl_ztradd_cpu_func(void *descr[], void *cl_arg)
+{
+ struct cl_ztradd_args_s clargs;
+ CHAM_tile_t *tileA;
+ CHAM_tile_t *tileB;
tileA = cti_interface_get(descr[0]);
tileB = cti_interface_get(descr[1]);
- starpu_codelet_unpack_args(cl_arg, &uplo, &trans, &M, &N, &alpha, &beta);
- TCORE_ztradd(uplo, trans, M, N, alpha, tileA, beta, tileB);
- return;
+
+ starpu_codelet_unpack_args( cl_arg, &clargs );
+ TCORE_ztradd( clargs.uplo, clargs.trans, clargs.m, clargs.n,
+ clargs.alpha, tileA, clargs.beta, tileB );
}
#endif /* !defined(CHAMELEON_SIMULATION) */
/*
* Codelet definition
*/
-CODELETS_CPU(ztradd, cl_ztradd_cpu_func)
+CODELETS_CPU( ztradd, cl_ztradd_cpu_func )
void INSERT_TASK_ztradd( const RUNTIME_option_t *options,
cham_uplo_t uplo, cham_trans_t trans, int m, int n, int nb,
@@ -56,34 +64,54 @@ void INSERT_TASK_ztradd( const RUNTIME_option_t *options,
beta, B, Bm, Bn );
}
- struct starpu_codelet *codelet = &cl_ztradd;
- void (*callback)(void*) = options->profiling ? cl_zgeadd_callback : NULL;
- starpu_option_request_t* schedopt = (starpu_option_request_t *)(options->request->schedopt);
- int workerid = (schedopt == NULL) ? -1 : schedopt->workerid;
- int accessB = ( beta == 0. ) ? STARPU_W : STARPU_RW;
+ struct cl_ztradd_args_s clargs = {
+ .uplo = uplo,
+ .trans = trans,
+ .m = m,
+ .n = n,
+ .alpha = alpha,
+ .tileA = A->get_blktile( A, Am, An ),
+ .beta = beta,
+ .tileB = B->get_blktile( B, Bm, Bn ),
+ };
+ void (*callback)(void*);
+ RUNTIME_request_t *request = options->request;
+ starpu_option_request_t *schedopt = (starpu_option_request_t *)(request->schedopt);
+ int workerid, accessB;
+ char *cl_name = "ztradd";
+ /* Handle cache */
CHAMELEON_BEGIN_ACCESS_DECLARATION;
CHAMELEON_ACCESS_R(A, Am, An);
CHAMELEON_ACCESS_RW(B, Bm, Bn);
CHAMELEON_END_ACCESS_DECLARATION;
+ /* Callback fro profiling information */
+ callback = options->profiling ? cl_ztradd_callback : NULL;
+
+ /* Fix the worker id */
+ workerid = (schedopt == NULL) ? -1 : schedopt->workerid;
+
+ /* Reduce the B access if needed */
+ accessB = ( beta == 0. ) ? STARPU_W : STARPU_RW;
+
+ /* Insert the task */
rt_starpu_insert_task(
- codelet,
- STARPU_VALUE, &uplo, sizeof(int),
- STARPU_VALUE, &trans, sizeof(int),
- STARPU_VALUE, &m, sizeof(int),
- STARPU_VALUE, &n, sizeof(int),
- STARPU_VALUE, &alpha, sizeof(CHAMELEON_Complex64_t),
- STARPU_R, RTBLKADDR(A, CHAMELEON_Complex64_t, Am, An),
- STARPU_VALUE, &beta, sizeof(CHAMELEON_Complex64_t),
- accessB, RTBLKADDR(B, CHAMELEON_Complex64_t, Bm, Bn),
- STARPU_PRIORITY, options->priority,
- STARPU_CALLBACK, callback,
+ &cl_ztradd,
+ /* Task codelet arguments */
+ STARPU_VALUE, &clargs, sizeof(struct cl_ztradd_args_s),
+ STARPU_R, RTBLKADDR(A, CHAMELEON_Complex64_t, Am, An),
+ accessB, RTBLKADDR(B, CHAMELEON_Complex64_t, Bm, Bn),
+
+ /* Common task arguments */
+ STARPU_PRIORITY, options->priority,
+ STARPU_CALLBACK, callback,
STARPU_EXECUTE_ON_WORKER, workerid,
#if defined(CHAMELEON_CODELETS_HAVE_NAME)
- STARPU_NAME, "ztradd",
+ STARPU_NAME, cl_name,
#endif
- 0);
+
+ 0 );
(void)nb;
}
diff --git a/runtime/starpu/codelets/codelet_ztrmm.c b/runtime/starpu/codelets/codelet_ztrmm.c
index 354e5f454..099e5ca8a 100644
--- a/runtime/starpu/codelets/codelet_ztrmm.c
+++ b/runtime/starpu/codelets/codelet_ztrmm.c
@@ -25,51 +25,53 @@
#include "chameleon_starpu.h"
#include "runtime_codelet_z.h"
-#if !defined(CHAMELEON_SIMULATION)
-static void cl_ztrmm_cpu_func(void *descr[], void *cl_arg)
-{
+struct cl_ztrmm_args_s {
cham_side_t side;
cham_uplo_t uplo;
cham_trans_t transA;
cham_diag_t diag;
- int M;
- int N;
+ int m;
+ int n;
CHAMELEON_Complex64_t alpha;
CHAM_tile_t *tileA;
CHAM_tile_t *tileB;
+};
+
+#if !defined(CHAMELEON_SIMULATION)
+static void
+cl_ztrmm_cpu_func(void *descr[], void *cl_arg)
+{
+ struct cl_ztrmm_args_s clargs;
+ CHAM_tile_t *tileA;
+ CHAM_tile_t *tileB;
tileA = cti_interface_get(descr[0]);
tileB = cti_interface_get(descr[1]);
- starpu_codelet_unpack_args(cl_arg, &side, &uplo, &transA, &diag, &M, &N, &alpha);
- TCORE_ztrmm(side, uplo,
- transA, diag,
- M, N,
- alpha, tileA,
- tileB);
+ starpu_codelet_unpack_args( cl_arg, &clargs );
+ TCORE_ztrmm( clargs.side, clargs.uplo, clargs.transA, clargs.diag,
+ clargs.m, clargs.n, clargs.alpha, tileA, tileB );
}
#ifdef CHAMELEON_USE_CUDA
-static void cl_ztrmm_cuda_func(void *descr[], void *cl_arg)
+static void
+cl_ztrmm_cuda_func(void *descr[], void *cl_arg)
{
- cham_side_t side;
- cham_uplo_t uplo;
- cham_trans_t transA;
- cham_diag_t diag;
- int M;
- int N;
- cuDoubleComplex alpha;
+ struct cl_ztrmm_args_s clargs;
CHAM_tile_t *tileA;
CHAM_tile_t *tileB;
tileA = cti_interface_get(descr[0]);
tileB = cti_interface_get(descr[1]);
- starpu_codelet_unpack_args(cl_arg, &side, &uplo, &transA, &diag, &M, &N, &alpha);
+
+ starpu_codelet_unpack_args( cl_arg, &clargs );
RUNTIME_getStream(stream);
CUDA_ztrmm(
- side, uplo, transA, diag, M, N, &alpha,
+ clargs.side, clargs.uplo, clargs.transA, clargs.diag,
+ clargs.m, clargs.n,
+ (cuDoubleComplex*)&(clargs.alpha),
tileA->mat, tileA->ld,
tileB->mat, tileB->ld,
stream );
@@ -80,58 +82,66 @@ static void cl_ztrmm_cuda_func(void *descr[], void *cl_arg)
return;
}
-#endif /* CHAMELEON_USE_CUDA */
+#endif /* defined(CHAMELEON_USE_CUDA) */
#endif /* !defined(CHAMELEON_SIMULATION) */
-
/*
* Codelet definition
*/
-CODELETS(ztrmm, cl_ztrmm_cpu_func, cl_ztrmm_cuda_func, STARPU_CUDA_ASYNC)
+CODELETS( ztrmm, cl_ztrmm_cpu_func, cl_ztrmm_cuda_func, STARPU_CUDA_ASYNC )
-/**
- *
- * @ingroup INSERT_TASK_Complex64_t
- *
- */
-void INSERT_TASK_ztrmm(const RUNTIME_option_t *options,
- cham_side_t side, cham_uplo_t uplo, cham_trans_t transA, cham_diag_t diag,
- int m, int n, int nb,
- CHAMELEON_Complex64_t alpha, const CHAM_desc_t *A, int Am, int An,
- const CHAM_desc_t *B, int Bm, int Bn)
+void INSERT_TASK_ztrmm( const RUNTIME_option_t *options,
+ cham_side_t side, cham_uplo_t uplo, cham_trans_t transA, cham_diag_t diag,
+ int m, int n, int nb,
+ CHAMELEON_Complex64_t alpha, const CHAM_desc_t *A, int Am, int An,
+ const CHAM_desc_t *B, int Bm, int Bn )
{
- if ( alpha == 0. ) {
- return INSERT_TASK_zlaset( options, ChamUpperLower, m, n,
- alpha, alpha, B, Bm, Bn );
- }
-
- (void)nb;
- struct starpu_codelet *codelet = &cl_ztrmm;
- void (*callback)(void*) = options->profiling ? cl_ztrmm_callback : NULL;
- starpu_option_request_t* schedopt = (starpu_option_request_t *)(options->request->schedopt);
- int workerid = (schedopt == NULL) ? -1 : schedopt->workerid;
+ struct cl_ztrmm_args_s clargs = {
+ .side = side,
+ .uplo = uplo,
+ .transA = transA,
+ .diag = diag,
+ .m = m,
+ .n = n,
+ .alpha = alpha,
+ .tileA = A->get_blktile( A, Am, An ),
+ .tileB = B->get_blktile( B, Bm, Bn ),
+ };
+ void (*callback)(void*);
+ RUNTIME_request_t *request = options->request;
+ starpu_option_request_t *schedopt = (starpu_option_request_t *)(request->schedopt);
+ int workerid;
+ char *cl_name = "ztrmm";
+ /* Handle cache */
CHAMELEON_BEGIN_ACCESS_DECLARATION;
CHAMELEON_ACCESS_R(A, Am, An);
CHAMELEON_ACCESS_RW(B, Bm, Bn);
CHAMELEON_END_ACCESS_DECLARATION;
+ /* Callback fro profiling information */
+ callback = options->profiling ? cl_ztrmm_callback : NULL;
+
+ /* Fix the worker id */
+ workerid = (schedopt == NULL) ? -1 : schedopt->workerid;
+
+ /* Insert the task */
rt_starpu_insert_task(
- codelet,
- STARPU_VALUE, &side, sizeof(int),
- STARPU_VALUE, &uplo, sizeof(int),
- STARPU_VALUE, &transA, sizeof(int),
- STARPU_VALUE, &diag, sizeof(int),
- STARPU_VALUE, &m, sizeof(int),
- STARPU_VALUE, &n, sizeof(int),
- STARPU_VALUE, &alpha, sizeof(CHAMELEON_Complex64_t),
- STARPU_R, RTBLKADDR(A, CHAMELEON_Complex64_t, Am, An),
- STARPU_RW, RTBLKADDR(B, CHAMELEON_Complex64_t, Bm, Bn),
- STARPU_PRIORITY, options->priority,
- STARPU_CALLBACK, callback,
+ &cl_ztrmm,
+ /* Task codelet arguments */
+ STARPU_VALUE, &clargs, sizeof(struct cl_ztrmm_args_s),
+ STARPU_R, RTBLKADDR(A, CHAMELEON_Complex64_t, Am, An),
+ STARPU_RW, RTBLKADDR(B, CHAMELEON_Complex64_t, Bm, Bn),
+
+ /* Common task arguments */
+ STARPU_PRIORITY, options->priority,
+ STARPU_CALLBACK, callback,
STARPU_EXECUTE_ON_WORKER, workerid,
#if defined(CHAMELEON_CODELETS_HAVE_NAME)
- STARPU_NAME, "ztrmm",
+ STARPU_NAME, cl_name,
#endif
- 0);
+
+ 0 );
+
+ (void)nb;
}
diff --git a/runtime/starpu/codelets/codelet_ztrsm.c b/runtime/starpu/codelets/codelet_ztrsm.c
index 694f0a99b..e7ad01a43 100644
--- a/runtime/starpu/codelets/codelet_ztrsm.c
+++ b/runtime/starpu/codelets/codelet_ztrsm.c
@@ -12,8 +12,6 @@
* @brief Chameleon ztrsm StarPU codelet
*
* @version 1.0.0
- * @comment This file has been automatically generated
- * from Plasma 2.5.0 for CHAMELEON 0.9.2
* @author Hatem Ltaief
* @author Jakub Kurzak
* @author Mathieu Faverge
@@ -27,9 +25,7 @@
#include "chameleon_starpu.h"
#include "runtime_codelet_z.h"
-#if !defined(CHAMELEON_SIMULATION)
-static void cl_ztrsm_cpu_func(void *descr[], void *cl_arg)
-{
+struct cl_ztrsm_args_s {
cham_side_t side;
cham_uplo_t uplo;
cham_trans_t transA;
@@ -39,38 +35,43 @@ static void cl_ztrsm_cpu_func(void *descr[], void *cl_arg)
CHAMELEON_Complex64_t alpha;
CHAM_tile_t *tileA;
CHAM_tile_t *tileB;
+};
+
+#if !defined(CHAMELEON_SIMULATION)
+static void
+cl_ztrsm_cpu_func(void *descr[], void *cl_arg)
+{
+ struct cl_ztrsm_args_s clargs;
+ CHAM_tile_t *tileA;
+ CHAM_tile_t *tileB;
tileA = cti_interface_get(descr[0]);
tileB = cti_interface_get(descr[1]);
- starpu_codelet_unpack_args(cl_arg, &side, &uplo, &transA, &diag, &m, &n, &alpha);
- TCORE_ztrsm(side, uplo,
- transA, diag,
- m, n,
- alpha, tileA,
- tileB);
+
+ starpu_codelet_unpack_args( cl_arg, &clargs );
+ TCORE_ztrsm( clargs.side, clargs.uplo, clargs.transA, clargs.diag,
+ clargs.m, clargs.n, clargs.alpha, tileA, tileB );
}
#ifdef CHAMELEON_USE_CUDA
-static void cl_ztrsm_cuda_func(void *descr[], void *cl_arg)
+static void
+cl_ztrsm_cuda_func(void *descr[], void *cl_arg)
{
- cham_side_t side;
- cham_uplo_t uplo;
- cham_trans_t transA;
- cham_diag_t diag;
- int m;
- int n;
- cuDoubleComplex alpha;
+ struct cl_ztrsm_args_s clargs;
CHAM_tile_t *tileA;
CHAM_tile_t *tileB;
tileA = cti_interface_get(descr[0]);
tileB = cti_interface_get(descr[1]);
- starpu_codelet_unpack_args(cl_arg, &side, &uplo, &transA, &diag, &m, &n, &alpha);
+
+ starpu_codelet_unpack_args( cl_arg, &clargs );
RUNTIME_getStream(stream);
CUDA_ztrsm(
- side, uplo, transA, diag, m, n, &alpha,
+ clargs.side, clargs.uplo, clargs.transA, clargs.diag,
+ clargs.m, clargs.n,
+ (cuDoubleComplex*)&(clargs.alpha),
tileA->mat, tileA->ld,
tileB->mat, tileB->ld,
stream );
@@ -81,52 +82,66 @@ static void cl_ztrsm_cuda_func(void *descr[], void *cl_arg)
return;
}
-#endif /* CHAMELEON_USE_CUDA */
+#endif /* defined(CHAMELEON_USE_CUDA) */
#endif /* !defined(CHAMELEON_SIMULATION) */
/*
* Codelet definition
*/
-CODELETS(ztrsm, cl_ztrsm_cpu_func, cl_ztrsm_cuda_func, STARPU_CUDA_ASYNC)
+CODELETS( ztrsm, cl_ztrsm_cpu_func, cl_ztrsm_cuda_func, STARPU_CUDA_ASYNC )
-/**
- *
- * @ingroup INSERT_TASK_Complex64_t
- *
- */
-void INSERT_TASK_ztrsm(const RUNTIME_option_t *options,
- cham_side_t side, cham_uplo_t uplo, cham_trans_t transA, cham_diag_t diag,
- int m, int n, int nb,
- CHAMELEON_Complex64_t alpha, const CHAM_desc_t *A, int Am, int An,
- const CHAM_desc_t *B, int Bm, int Bn)
+void INSERT_TASK_ztrsm( const RUNTIME_option_t *options,
+ cham_side_t side, cham_uplo_t uplo, cham_trans_t transA, cham_diag_t diag,
+ int m, int n, int nb,
+ CHAMELEON_Complex64_t alpha, const CHAM_desc_t *A, int Am, int An,
+ const CHAM_desc_t *B, int Bm, int Bn )
{
- (void)nb;
- struct starpu_codelet *codelet = &cl_ztrsm;
- void (*callback)(void*) = options->profiling ? cl_ztrsm_callback : NULL;
- starpu_option_request_t* schedopt = (starpu_option_request_t *)(options->request->schedopt);
- int workerid = (schedopt == NULL) ? -1 : schedopt->workerid;
+ struct cl_ztrsm_args_s clargs = {
+ .side = side,
+ .uplo = uplo,
+ .transA = transA,
+ .diag = diag,
+ .m = m,
+ .n = n,
+ .alpha = alpha,
+ .tileA = A->get_blktile( A, Am, An ),
+ .tileB = B->get_blktile( B, Bm, Bn ),
+ };
+ void (*callback)(void*);
+ RUNTIME_request_t *request = options->request;
+ starpu_option_request_t *schedopt = (starpu_option_request_t *)(request->schedopt);
+ int workerid;
+ char *cl_name = "ztrsm";
+ /* Handle cache */
CHAMELEON_BEGIN_ACCESS_DECLARATION;
CHAMELEON_ACCESS_R(A, Am, An);
CHAMELEON_ACCESS_RW(B, Bm, Bn);
CHAMELEON_END_ACCESS_DECLARATION;
+ /* Callback fro profiling information */
+ callback = options->profiling ? cl_ztrsm_callback : NULL;
+
+ /* Fix the worker id */
+ workerid = (schedopt == NULL) ? -1 : schedopt->workerid;
+
+ /* Insert the task */
rt_starpu_insert_task(
- codelet,
- STARPU_VALUE, &side, sizeof(int),
- STARPU_VALUE, &uplo, sizeof(int),
- STARPU_VALUE, &transA, sizeof(int),
- STARPU_VALUE, &diag, sizeof(int),
- STARPU_VALUE, &m, sizeof(int),
- STARPU_VALUE, &n, sizeof(int),
- STARPU_VALUE, &alpha, sizeof(CHAMELEON_Complex64_t),
- STARPU_R, RTBLKADDR(A, CHAMELEON_Complex64_t, Am, An),
- STARPU_RW, RTBLKADDR(B, CHAMELEON_Complex64_t, Bm, Bn),
- STARPU_PRIORITY, options->priority,
- STARPU_CALLBACK, callback,
+ &cl_ztrsm,
+ /* Task codelet arguments */
+ STARPU_VALUE, &clargs, sizeof(struct cl_ztrsm_args_s),
+ STARPU_R, RTBLKADDR(A, CHAMELEON_Complex64_t, Am, An),
+ STARPU_RW, RTBLKADDR(B, CHAMELEON_Complex64_t, Bm, Bn),
+
+ /* Common task arguments */
+ STARPU_PRIORITY, options->priority,
+ STARPU_CALLBACK, callback,
STARPU_EXECUTE_ON_WORKER, workerid,
#if defined(CHAMELEON_CODELETS_HAVE_NAME)
- STARPU_NAME, "ztrsm",
+ STARPU_NAME, cl_name,
#endif
- 0);
+
+ 0 );
+
+ (void)nb;
}
diff --git a/runtime/starpu/codelets/codelet_ztrtri.c b/runtime/starpu/codelets/codelet_ztrtri.c
index 01b14eef4..9b2b0e5e3 100644
--- a/runtime/starpu/codelets/codelet_ztrtri.c
+++ b/runtime/starpu/codelets/codelet_ztrtri.c
@@ -12,8 +12,6 @@
* @brief Chameleon ztrtri StarPU codelet
*
* @version 1.0.0
- * @comment This file has been automatically generated
- * from Plasma 2.5.0 for CHAMELEON 0.9.2
* @author Julien Langou
* @author Henricus Bouwmeester
* @author Mathieu Faverge
@@ -27,24 +25,31 @@
#include "chameleon_starpu.h"
#include "runtime_codelet_z.h"
-#if !defined(CHAMELEON_SIMULATION)
-static void cl_ztrtri_cpu_func(void *descr[], void *cl_arg)
-{
+struct cl_ztrtri_args_s {
cham_uplo_t uplo;
cham_diag_t diag;
- int N;
+ int n;
CHAM_tile_t *tileA;
int iinfo;
RUNTIME_sequence_t *sequence;
RUNTIME_request_t *request;
+};
+
+#if !defined(CHAMELEON_SIMULATION)
+static void
+cl_ztrtri_cpu_func(void *descr[], void *cl_arg)
+{
+ struct cl_ztrtri_args_s clargs;
+ CHAM_tile_t *tileA;
int info = 0;
tileA = cti_interface_get(descr[0]);
- starpu_codelet_unpack_args(cl_arg, &uplo, &diag, &N, &iinfo, &sequence, &request);
- TCORE_ztrtri(uplo, diag, N, tileA, &info);
- if ( (sequence->status == CHAMELEON_SUCCESS) && (info != 0) ) {
- RUNTIME_sequence_flush( NULL, sequence, request, iinfo+info );
+ starpu_codelet_unpack_args( cl_arg, &clargs );
+ TCORE_ztrtri( clargs.uplo, clargs.diag, clargs.n, tileA, &info );
+
+ if ( (clargs.sequence->status == CHAMELEON_SUCCESS) && (info != 0) ) {
+ RUNTIME_sequence_flush( NULL, clargs.sequence, clargs.request, clargs.iinfo+info );
}
}
#endif /* !defined(CHAMELEON_SIMULATION) */
@@ -52,43 +57,55 @@ static void cl_ztrtri_cpu_func(void *descr[], void *cl_arg)
/*
* Codelet definition
*/
-CODELETS_CPU(ztrtri, cl_ztrtri_cpu_func)
+CODELETS_CPU( ztrtri, cl_ztrtri_cpu_func )
-/**
- *
- * @ingroup INSERT_TASK_Complex64_t
- *
- */
void INSERT_TASK_ztrtri( const RUNTIME_option_t *options,
- cham_uplo_t uplo, cham_diag_t diag,
- int n, int nb,
+ cham_uplo_t uplo, cham_diag_t diag, int n, int nb,
const CHAM_desc_t *A, int Am, int An,
int iinfo )
{
- (void)nb;
- struct starpu_codelet *codelet = &cl_ztrtri;
- void (*callback)(void*) = options->profiling ? cl_ztrtri_callback : NULL;
- starpu_option_request_t* schedopt = (starpu_option_request_t *)(options->request->schedopt);
- int workerid = (schedopt == NULL) ? -1 : schedopt->workerid;
+ struct cl_ztrtri_args_s clargs = {
+ .uplo = uplo,
+ .diag = diag,
+ .n = n,
+ .tileA = A->get_blktile( A, Am, An ),
+ .iinfo = iinfo,
+ .sequence = options->sequence,
+ .request = options->request,
+ };
+ void (*callback)(void*);
+ RUNTIME_request_t *request = options->request;
+ starpu_option_request_t *schedopt = (starpu_option_request_t *)(request->schedopt);
+ int workerid;
+ char *cl_name = "ztrtri";
+ /* Handle cache */
CHAMELEON_BEGIN_ACCESS_DECLARATION;
CHAMELEON_ACCESS_RW(A, Am, An);
CHAMELEON_END_ACCESS_DECLARATION;
+ /* Callback fro profiling information */
+ callback = options->profiling ? cl_ztrtri_callback : NULL;
+
+ /* Fix the worker id */
+ workerid = (schedopt == NULL) ? -1 : schedopt->workerid;
+
+ /* Insert the task */
rt_starpu_insert_task(
- codelet,
- STARPU_VALUE, &uplo, sizeof(int),
- STARPU_VALUE, &diag, sizeof(int),
- STARPU_VALUE, &n, sizeof(int),
- STARPU_RW, RTBLKADDR(A, CHAMELEON_Complex64_t, Am, An),
- STARPU_VALUE, &iinfo, sizeof(int),
- STARPU_VALUE, &(options->sequence), sizeof(RUNTIME_sequence_t*),
- STARPU_VALUE, &(options->request), sizeof(RUNTIME_request_t*),
- STARPU_PRIORITY, options->priority,
- STARPU_CALLBACK, callback,
+ &cl_ztrtri,
+ /* Task codelet arguments */
+ STARPU_VALUE, &clargs, sizeof(struct cl_ztrtri_args_s),
+ STARPU_RW, RTBLKADDR(A, CHAMELEON_Complex64_t, Am, An),
+
+ /* Common task arguments */
+ STARPU_PRIORITY, options->priority,
+ STARPU_CALLBACK, callback,
STARPU_EXECUTE_ON_WORKER, workerid,
#if defined(CHAMELEON_CODELETS_HAVE_NAME)
- STARPU_NAME, "ztrtri",
+ STARPU_NAME, cl_name,
#endif
- 0);
+
+ 0 );
+
+ (void)nb;
}
--
GitLab