diff --git a/CMakeLists.txt b/CMakeLists.txt
index 2c515d7e5c5dfa0d0b0e4980a3336eb2ecdbb32b..b064cf15135af1b366763e1e8a78e9b903f02047 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -366,15 +366,6 @@ if(NOT CHAMELEON_SIMULATION)
endif()
if (CUDA_LIBRARIES)
set_target_properties(CUDA::CUDA PROPERTIES INTERFACE_LINK_LIBRARIES "${CUDA_LIBRARIES}")
- set(CMAKE_REQUIRED_INCLUDES "${CUDA_INCLUDE_DIRS}")
- set(CMAKE_REQUIRED_LIBRARIES "${CUDA_LIBRARIES}")
- if(CUDA_VERSION VERSION_LESS "4.0")
- set(CUDA_HAVE_PEER_DEVICE_MEMORY_ACCESS 0)
- else()
- check_function_exists(cuDeviceCanAccessPeer CUDA_HAVE_PEER_DEVICE_MEMORY_ACCESS)
- endif()
- unset(CMAKE_REQUIRED_INCLUDES)
- unset(CMAKE_REQUIRED_LIBRARIES)
# Add cublas if found
if (CUDA_CUBLAS_LIBRARIES)
set_target_properties(CUDA::CUBLAS PROPERTIES INTERFACE_LINK_LIBRARIES "${CUDA_CUBLAS_LIBRARIES}")
diff --git a/control/descriptor.h b/control/descriptor.h
index 6bf6d9a81ad9cb57458207b500b45b8fdc0bdc3d..dda8fa71906169a3dde202e670e27c6c1c70467b 100644
--- a/control/descriptor.h
+++ b/control/descriptor.h
@@ -233,33 +233,41 @@ inline static int chameleon_desc_islocal( const CHAM_desc_t *A, int m, int n )
* CHAMELEON_ACCESS_RW(C, Cm, Cn)
* CHAMELEON_END_ACCESS_DECLARATION
*/
-#define CHAMELEON_BEGIN_ACCESS_DECLARATION { \
- unsigned __chameleon_need_submit = 0; \
+#define CHAMELEON_BEGIN_ACCESS_DECLARATION { \
+ unsigned __chameleon_need_exec = 0; \
+ unsigned __chameleon_need_submit = 0; \
RUNTIME_BEGIN_ACCESS_DECLARATION
-#define CHAMELEON_ACCESS_R(A, Am, An) do { \
- if (chameleon_desc_islocal(A, Am, An)) __chameleon_need_submit = 1; \
- RUNTIME_ACCESS_R(A, Am, An); \
-} while(0)
-
-#define CHAMELEON_ACCESS_W(A, Am, An) do { \
- if (chameleon_desc_islocal(A, Am, An)) __chameleon_need_submit = 1; \
- RUNTIME_ACCESS_W(A, Am, An); \
-} while(0)
-
-#define CHAMELEON_ACCESS_RW(A, Am, An) do { \
- if (chameleon_desc_islocal(A, Am, An)) __chameleon_need_submit = 1; \
- RUNTIME_ACCESS_RW(A, Am, An); \
-} while(0)
-
-#define CHAMELEON_RANK_CHANGED(rank) do {\
- __chameleon_need_submit = 1; \
- RUNTIME_RANK_CHANGED(rank); \
-} while (0)
-
-#define CHAMELEON_END_ACCESS_DECLARATION \
- RUNTIME_END_ACCESS_DECLARATION; \
- if (!__chameleon_need_submit) return; \
+#define CHAMELEON_ACCESS_R(A, Am, An) do { \
+ if (chameleon_desc_islocal(A, Am, An)) __chameleon_need_submit = 1; \
+ RUNTIME_ACCESS_R(A, Am, An); \
+ } while(0)
+
+#define CHAMELEON_ACCESS_W(A, Am, An) do { \
+ if (chameleon_desc_islocal(A, Am, An)) { \
+ __chameleon_need_exec = 1; \
+ __chameleon_need_submit = 1; \
+ } \
+ RUNTIME_ACCESS_W(A, Am, An); \
+ } while(0)
+
+#define CHAMELEON_ACCESS_RW(A, Am, An) do { \
+ if (chameleon_desc_islocal(A, Am, An)) { \
+ __chameleon_need_exec = 1; \
+ __chameleon_need_submit = 1; \
+ } \
+ RUNTIME_ACCESS_RW(A, Am, An); \
+ } while(0)
+
+#define CHAMELEON_RANK_CHANGED(rank) do { \
+ __chameleon_need_submit = 1; \
+ RUNTIME_RANK_CHANGED(rank); \
+ } while (0)
+
+#define CHAMELEON_END_ACCESS_DECLARATION \
+ RUNTIME_END_ACCESS_DECLARATION; \
+ if (!__chameleon_need_submit) return; \
+ (void)__chameleon_need_exec; \
}
#ifdef __cplusplus
diff --git a/runtime/starpu/codelets/codelet_zcesca.c b/runtime/starpu/codelets/codelet_zcesca.c
index 5dbadec77774438ef701cb9d07b4f7af579094f1..1cf87fe7e330d4f636c648cb4a7886378e93a714 100644
--- a/runtime/starpu/codelets/codelet_zcesca.c
+++ b/runtime/starpu/codelets/codelet_zcesca.c
@@ -31,7 +31,7 @@ struct cl_zcesca_args_s {
#if !defined(CHAMELEON_SIMULATION)
static void cl_zcesca_cpu_func(void *descr[], void *cl_arg)
{
- struct cl_zcesca_args_s clargs;
+ struct cl_zcesca_args_s *clargs = (struct cl_zcesca_args_s *)cl_arg;
CHAM_tile_t *Gi;
CHAM_tile_t *Gj;
CHAM_tile_t *G;
@@ -46,9 +46,8 @@ static void cl_zcesca_cpu_func(void *descr[], void *cl_arg)
Dj = cti_interface_get(descr[4]);
A = cti_interface_get(descr[5]);
- starpu_codelet_unpack_args( cl_arg, &clargs );
- TCORE_zcesca( clargs.center, clargs.scale, clargs.axis,
- clargs.m, clargs.n, clargs.mt, clargs.nt,
+ TCORE_zcesca( clargs->center, clargs->scale, clargs->axis,
+ clargs->m, clargs->n, clargs->mt, clargs->nt,
Gi, Gj, G, Di, Dj, A );
}
#endif /* !defined(CHAMELEON_SIMULATION) */
@@ -68,19 +67,12 @@ void INSERT_TASK_zcesca( const RUNTIME_option_t *options,
const CHAM_desc_t *Dj, int Djm, int Djn,
CHAM_desc_t *A, int Am, int An )
{
- struct cl_zcesca_args_s clargs = {
- .center = center,
- .scale = scale,
- .axis = axis,
- .m = m,
- .n = n,
- .mt = mt,
- .nt = nt
- };
+ struct cl_zcesca_args_s *clargs = NULL;
struct starpu_codelet *codelet = &cl_zcesca;
void (*callback)(void*) = options->profiling ? cl_zcesca_callback : NULL;
starpu_option_request_t* schedopt = (starpu_option_request_t *)(options->request->schedopt);
int workerid = (schedopt == NULL) ? -1 : schedopt->workerid;
+ int exec = 0;
CHAMELEON_BEGIN_ACCESS_DECLARATION;
CHAMELEON_ACCESS_R(Gi, Gim, Gin);
@@ -89,11 +81,23 @@ void INSERT_TASK_zcesca( const RUNTIME_option_t *options,
CHAMELEON_ACCESS_R(Di, Dim, Din);
CHAMELEON_ACCESS_R(Dj, Djm, Djn);
CHAMELEON_ACCESS_RW(A, Am, An);
+ exec = __chameleon_need_exec;
CHAMELEON_END_ACCESS_DECLARATION;
+ if ( exec ) {
+ clargs = malloc( sizeof( struct cl_zcesca_args_s ) );
+ clargs->center = center;
+ clargs->scale = scale;
+ clargs->axis = axis;
+ clargs->m = m;
+ clargs->n = n;
+ clargs->mt = mt;
+ clargs->nt = nt;
+ }
+
rt_starpu_insert_task(
codelet,
- STARPU_VALUE, &clargs, sizeof(struct cl_zcesca_args_s),
+ STARPU_CL_ARGS, clargs, sizeof(struct cl_zcesca_args_s),
STARPU_R, RTBLKADDR(Gi, CHAMELEON_Complex64_t, Gim, Gin),
STARPU_R, RTBLKADDR(Gj, CHAMELEON_Complex64_t, Gjm, Gjn),
STARPU_R, RTBLKADDR(G, CHAMELEON_Complex64_t, Gm, Gn),
diff --git a/runtime/starpu/codelets/codelet_zgemm.c b/runtime/starpu/codelets/codelet_zgemm.c
index 39d242c01e4d6bf4f41eeb58a22f1b1e94612393..e66b85dedb1c17a8bff8e8ceb49f18fd662060dd 100644
--- a/runtime/starpu/codelets/codelet_zgemm.c
+++ b/runtime/starpu/codelets/codelet_zgemm.c
@@ -19,6 +19,7 @@
* @author Cedric Castagnede
* @author Lucas Barros de Assis
* @author Florent Pruvost
+ * @author Gwenole Lucas
* @date 2021-03-16
* @precisions normal z -> c d s
*
@@ -43,7 +44,7 @@ struct cl_zgemm_args_s {
static void
cl_zgemm_cpu_func( void *descr[], void *cl_arg )
{
- struct cl_zgemm_args_s clargs;
+ struct cl_zgemm_args_s *clargs = (struct cl_zgemm_args_s *)cl_arg;
CHAM_tile_t *tileA;
CHAM_tile_t *tileB;
CHAM_tile_t *tileC;
@@ -52,18 +53,17 @@ cl_zgemm_cpu_func( void *descr[], void *cl_arg )
tileB = cti_interface_get(descr[1]);
tileC = cti_interface_get(descr[2]);
- 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 );
+ 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 )
+cl_zgemm_cuda_func( void *descr[], void *cl_arg )
{
- struct cl_zgemm_args_s clargs;
+ struct cl_zgemm_args_s *clargs = (struct cl_zgemm_args_s *)cl_arg;
CHAM_tile_t *tileA;
CHAM_tile_t *tileB;
CHAM_tile_t *tileC;
@@ -72,17 +72,19 @@ cl_zgemm_cuda_func( void *descr[], void *_cl_arg )
tileB = cti_interface_get(descr[1]);
tileC = cti_interface_get(descr[2]);
- starpu_codelet_unpack_args( _cl_arg, &clargs );
-
RUNTIME_getStream( stream );
+ assert( tileA->format & CHAMELEON_TILE_FULLRANK );
+ assert( tileB->format & CHAMELEON_TILE_FULLRANK );
+ assert( tileC->format & CHAMELEON_TILE_FULLRANK );
+
CUDA_zgemm(
- clargs.transA, clargs.transB,
- clargs.m, clargs.n, clargs.k,
- (cuDoubleComplex*)&(clargs.alpha),
+ clargs->transA, clargs->transB,
+ clargs->m, clargs->n, clargs->k,
+ (cuDoubleComplex*)&(clargs->alpha),
tileA->mat, tileA->ld,
tileB->mat, tileB->ld,
- (cuDoubleComplex*)&(clargs.beta),
+ (cuDoubleComplex*)&(clargs->beta),
tileC->mat, tileC->ld,
stream );
@@ -112,22 +114,12 @@ void INSERT_TASK_zgemm( const RUNTIME_option_t *options,
beta, C, Cm, Cn );
}
- 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 )
- };
+ struct cl_zgemm_args_s *clargs = NULL;
void (*callback)(void*);
RUNTIME_request_t *request = options->request;
starpu_option_request_t *schedopt = (starpu_option_request_t *)(request->schedopt);
int workerid, accessC;
+ int exec = 0;
char *cl_name = "zgemm";
/* Handle cache */
@@ -135,8 +127,23 @@ void INSERT_TASK_zgemm( const RUNTIME_option_t *options,
CHAMELEON_ACCESS_R(A, Am, An);
CHAMELEON_ACCESS_R(B, Bm, Bn);
CHAMELEON_ACCESS_RW(C, Cm, Cn);
+ exec = __chameleon_need_exec;
CHAMELEON_END_ACCESS_DECLARATION;
+ if ( exec ) {
+ clargs = malloc( sizeof( struct cl_zgemm_args_s ) );
+ clargs->transA = transA;
+ clargs->transB = transB;
+ clargs->m = m;
+ 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 */
callback = options->profiling ? cl_zgemm_callback : NULL;
@@ -150,10 +157,12 @@ void INSERT_TASK_zgemm( const RUNTIME_option_t *options,
rt_starpu_insert_task(
&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),
+ STARPU_CL_ARGS, clargs, sizeof(struct cl_zgemm_args_s),
+
+ /* Task handles */
+ 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,
diff --git a/runtime/starpu/codelets/codelet_zgessm.c b/runtime/starpu/codelets/codelet_zgessm.c
index 5fbc6b09a9a1f2d733e9a4f52aecc7320a17b08f..5dad32c9afae12291dbc1729cd3527264d86cac8 100644
--- a/runtime/starpu/codelets/codelet_zgessm.c
+++ b/runtime/starpu/codelets/codelet_zgessm.c
@@ -43,7 +43,6 @@ static void cl_zgessm_cpu_func(void *descr[], void *cl_arg)
tileD = cti_interface_get(descr[1]);
tileA = cti_interface_get(descr[2]);
-
starpu_codelet_unpack_args(cl_arg, &m, &n, &k, &ib, &IPIV);
TCORE_zgessm(m, n, k, ib, IPIV, tileD, tileA);
}
diff --git a/runtime/starpu/codelets/codelet_zgesum.c b/runtime/starpu/codelets/codelet_zgesum.c
index 31a6ca93e5c8f9fc74a12e656309ca3059cb731d..c851c4ef555bc1ee792a712bfaa028795df4252c 100644
--- a/runtime/starpu/codelets/codelet_zgesum.c
+++ b/runtime/starpu/codelets/codelet_zgesum.c
@@ -27,15 +27,14 @@ struct cl_zgesum_args_s {
#if !defined(CHAMELEON_SIMULATION)
static void cl_zgesum_cpu_func(void *descr[], void *cl_arg)
{
- struct cl_zgesum_args_s clargs;
+ struct cl_zgesum_args_s *clargs = (struct cl_zgesum_args_s *)cl_arg;
CHAM_tile_t *tileA;
CHAM_tile_t *tileW;
tileA = cti_interface_get(descr[0]);
tileW = cti_interface_get(descr[1]);
- starpu_codelet_unpack_args( cl_arg, &clargs );
- TCORE_zgesum( clargs.storev, clargs.m, clargs.n, tileA, tileW );
+ TCORE_zgesum( clargs->storev, clargs->m, clargs->n, tileA, tileW );
}
#endif /* !defined(CHAMELEON_SIMULATION) */
@@ -49,24 +48,29 @@ void INSERT_TASK_zgesum( const RUNTIME_option_t *options,
const CHAM_desc_t *A, int Am, int An,
const CHAM_desc_t *SUMS, int SUMSm, int SUMSn )
{
- struct cl_zgesum_args_s clargs = {
- .storev = storev,
- .m = m,
- .n = n
- };
+ struct cl_zgesum_args_s *clargs = NULL;
struct starpu_codelet *codelet = &cl_zgesum;
void (*callback)(void*) = options->profiling ? cl_zgesum_callback : NULL;
starpu_option_request_t* schedopt = (starpu_option_request_t *)(options->request->schedopt);
int workerid = (schedopt == NULL) ? -1 : schedopt->workerid;
+ int exec = 0;
CHAMELEON_BEGIN_ACCESS_DECLARATION;
CHAMELEON_ACCESS_R(A, Am, An);
CHAMELEON_ACCESS_RW(SUMS, SUMSm, SUMSn);
+ exec = __chameleon_need_exec;
CHAMELEON_END_ACCESS_DECLARATION;
+ if ( exec ) {
+ clargs = malloc( sizeof( struct cl_zgesum_args_s ) );
+ clargs->storev = storev;
+ clargs->m = m;
+ clargs->n = n;
+ }
+
rt_starpu_insert_task(
codelet,
- STARPU_VALUE, &clargs, sizeof(struct cl_zgesum_args_s),
+ STARPU_CL_ARGS, clargs, sizeof(struct cl_zgesum_args_s),
STARPU_R, RTBLKADDR(A, CHAMELEON_Complex64_t, Am, An),
STARPU_RW, RTBLKADDR(SUMS, CHAMELEON_Complex64_t, SUMSm, SUMSn),
STARPU_PRIORITY, options->priority,
diff --git a/runtime/starpu/codelets/codelet_zhe2ge.c b/runtime/starpu/codelets/codelet_zhe2ge.c
index 91b62cec66d69b9ce8f6340b781a0b5ba167b99c..9c918d9ef5c2b37de710f2e766cf448de8bbfe99 100644
--- a/runtime/starpu/codelets/codelet_zhe2ge.c
+++ b/runtime/starpu/codelets/codelet_zhe2ge.c
@@ -35,7 +35,6 @@ static void cl_zhe2ge_cpu_func(void *descr[], void *cl_arg)
tileA = cti_interface_get(descr[0]);
tileB = cti_interface_get(descr[1]);
-
starpu_codelet_unpack_args(cl_arg, &uplo, &M, &N);
TCORE_zhe2ge(uplo, M, N, tileA, tileB);
}
diff --git a/runtime/starpu/codelets/codelet_zherk.c b/runtime/starpu/codelets/codelet_zherk.c
index 1b8057b1409ef597ab99892a7482f811249d4620..3cc33f4236be6b6619f91c1032df22e787d11724 100644
--- a/runtime/starpu/codelets/codelet_zherk.c
+++ b/runtime/starpu/codelets/codelet_zherk.c
@@ -19,6 +19,7 @@
* @author Cedric Castagnede
* @author Lucas Barros de Assis
* @author Florent Pruvost
+ * @author Gwenole Lucas
* @date 2021-03-16
* @precisions normal z -> c
*
@@ -41,38 +42,35 @@ struct cl_zherk_args_s {
static void
cl_zherk_cpu_func(void *descr[], void *cl_arg)
{
- struct cl_zherk_args_s clargs;
+ struct cl_zherk_args_s *clargs = (struct cl_zherk_args_s *)cl_arg;
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, &clargs );
- TCORE_zherk( clargs.uplo, clargs.trans, clargs.n, clargs.k,
- clargs.alpha, tileA, clargs.beta, tileC );
+ TCORE_zherk( clargs->uplo, clargs->trans, clargs->n, clargs->k,
+ clargs->alpha, tileA, clargs->beta, tileC );
}
#if defined(CHAMELEON_USE_CUDA)
static void
cl_zherk_cuda_func(void *descr[], void *cl_arg)
{
- struct cl_zherk_args_s clargs;
+ struct cl_zherk_args_s *clargs = (struct cl_zherk_args_s *)cl_arg;
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, &clargs );
-
RUNTIME_getStream(stream);
CUDA_zherk(
- clargs.uplo, clargs.trans, clargs.n, clargs.k,
- (cuDoubleComplex*)&(clargs.alpha),
+ clargs->uplo, clargs->trans, clargs->n, clargs->k,
+ &(clargs->alpha),
tileA->mat, tileA->ld,
- (cuDoubleComplex*)&(clargs.beta),
+ &(clargs->beta),
tileC->mat, tileC->ld,
stream );
@@ -101,28 +99,33 @@ void INSERT_TASK_zherk( const RUNTIME_option_t *options,
beta, C, Cm, Cn );
}
- 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 ),
- };
+ struct cl_zherk_args_s *clargs = NULL;
void (*callback)(void*);
RUNTIME_request_t *request = options->request;
starpu_option_request_t *schedopt = (starpu_option_request_t *)(request->schedopt);
int workerid, accessC;
+ int exec = 0;
char *cl_name = "zherk";
/* Handle cache */
CHAMELEON_BEGIN_ACCESS_DECLARATION;
CHAMELEON_ACCESS_R(A, Am, An);
CHAMELEON_ACCESS_RW(C, Cm, Cn);
+ exec = __chameleon_need_exec;
CHAMELEON_END_ACCESS_DECLARATION;
+ if ( exec ) {
+ clargs = malloc( sizeof( struct cl_zherk_args_s ) );
+ clargs->uplo = uplo;
+ clargs->trans = trans;
+ 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 */
callback = options->profiling ? cl_zherk_callback : NULL;
@@ -136,7 +139,7 @@ void INSERT_TASK_zherk( const RUNTIME_option_t *options,
rt_starpu_insert_task(
&cl_zherk,
/* Task codelet arguments */
- STARPU_VALUE, &clargs, sizeof(struct cl_zherk_args_s),
+ STARPU_CL_ARGS, clargs, sizeof(struct cl_zherk_args_s),
STARPU_R, RTBLKADDR(A, CHAMELEON_Complex64_t, Am, An),
accessC, RTBLKADDR(C, CHAMELEON_Complex64_t, Cm, Cn),
diff --git a/runtime/starpu/codelets/codelet_zlacpy.c b/runtime/starpu/codelets/codelet_zlacpy.c
index 2501eb14833810892e7a5ced85f1f7659eb1d933..b169055095f5d54d5f825921306f28150b4df275 100644
--- a/runtime/starpu/codelets/codelet_zlacpy.c
+++ b/runtime/starpu/codelets/codelet_zlacpy.c
@@ -41,20 +41,19 @@ struct cl_zlacpy_args_s {
static void
cl_zlacpy_cpu_func(void *descr[], void *cl_arg)
{
- struct cl_zlacpy_args_s clargs;
+ struct cl_zlacpy_args_s *clargs = (struct cl_zlacpy_args_s *)cl_arg;
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, &clargs );
- assert( clargs.displA == 0 );
- assert( clargs.displB == 0 );
+ 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 );
+ TCORE_zlacpy( clargs->uplo, clargs->m, clargs->n, tileA, tileB );
}
#endif /* !defined(CHAMELEON_SIMULATION) */
@@ -68,27 +67,32 @@ void INSERT_TASK_zlacpyx( const RUNTIME_option_t *options,
int displA, const CHAM_desc_t *A, int Am, int An,
int displB, const CHAM_desc_t *B, int Bm, int Bn )
{
- 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 ),
- };
+ struct cl_zlacpy_args_s *clargs = NULL;
void (*callback)(void*);
RUNTIME_request_t *request = options->request;
starpu_option_request_t *schedopt = (starpu_option_request_t *)(request->schedopt);
int workerid;
+ int exec = 0;
char *cl_name = "zlacpy";
/* Handle cache */
CHAMELEON_BEGIN_ACCESS_DECLARATION;
CHAMELEON_ACCESS_R(A, Am, An);
CHAMELEON_ACCESS_W(B, Bm, Bn);
+ exec = __chameleon_need_exec;
CHAMELEON_END_ACCESS_DECLARATION;
+ if ( exec ) {
+ clargs = malloc( sizeof( struct cl_zlacpy_args_s ) );
+ clargs->uplo = uplo;
+ clargs->m = m;
+ 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 );
+ }
+
/* Callback fro profiling information */
callback = options->profiling ? cl_zlacpy_callback : NULL;
@@ -99,7 +103,7 @@ void INSERT_TASK_zlacpyx( const RUNTIME_option_t *options,
rt_starpu_insert_task(
&cl_zlacpy,
/* Task codelet arguments */
- STARPU_VALUE, &clargs, sizeof(struct cl_zlacpy_args_s),
+ STARPU_CL_ARGS, 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),
diff --git a/runtime/starpu/codelets/codelet_zlag2c.c b/runtime/starpu/codelets/codelet_zlag2c.c
index c2204d964180306a8fd833b9024e27556779d9b3..27358b138066d2344ecedd839fd9239f86fb706f 100644
--- a/runtime/starpu/codelets/codelet_zlag2c.c
+++ b/runtime/starpu/codelets/codelet_zlag2c.c
@@ -38,7 +38,6 @@ static void cl_zlag2c_cpu_func(void *descr[], void *cl_arg)
tileA = cti_interface_get(descr[0]);
tileB = cti_interface_get(descr[1]);
-
starpu_codelet_unpack_args(cl_arg, &m, &n);
TCORE_zlag2c( m, n, tileA, tileB);
}
@@ -96,7 +95,6 @@ static void cl_clag2z_cpu_func(void *descr[], void *cl_arg)
tileA = cti_interface_get(descr[0]);
tileB = cti_interface_get(descr[1]);
-
starpu_codelet_unpack_args(cl_arg, &m, &n);
TCORE_clag2z( m, n, tileA, tileB);
}
diff --git a/runtime/starpu/codelets/codelet_zlascal.c b/runtime/starpu/codelets/codelet_zlascal.c
index 0b9620fca00ed3d3fc13245e1f670b05e1261ea4..0dcaa52d54a4033c6b41629f23e98421a5ed72a5 100644
--- a/runtime/starpu/codelets/codelet_zlascal.c
+++ b/runtime/starpu/codelets/codelet_zlascal.c
@@ -36,13 +36,12 @@ struct cl_zlascal_args_s {
static void
cl_zlascal_cpu_func( void *descr[], void *cl_arg )
{
- struct cl_zlascal_args_s clargs;
+ struct cl_zlascal_args_s *clargs = (struct cl_zlascal_args_s *)cl_arg;
CHAM_tile_t *tileA;
tileA = cti_interface_get(descr[0]);
- starpu_codelet_unpack_args( cl_arg, &clargs );
- TCORE_zlascal( clargs.uplo, clargs.m, clargs.n, clargs.alpha, tileA );
+ TCORE_zlascal( clargs->uplo, clargs->m, clargs->n, clargs->alpha, tileA );
}
#endif /* !defined(CHAMELEON_SIMULATION) */
@@ -65,24 +64,29 @@ void INSERT_TASK_zlascal( const RUNTIME_option_t *options,
return;
}
- struct cl_zlascal_args_s clargs = {
- .uplo = uplo,
- .m = m,
- .n = n,
- .alpha = alpha,
- .tileA = A->get_blktile( A, Am, An ),
- };
+ struct cl_zlascal_args_s *clargs = NULL;
void (*callback)(void*);
RUNTIME_request_t *request = options->request;
starpu_option_request_t *schedopt = (starpu_option_request_t *)(request->schedopt);
int workerid;
+ int exec = 0;
char *cl_name = "zlascal";
/* Handle cache */
CHAMELEON_BEGIN_ACCESS_DECLARATION;
CHAMELEON_ACCESS_RW(A, Am, An);
+ exec = __chameleon_need_exec;
CHAMELEON_END_ACCESS_DECLARATION;
+ if ( exec ) {
+ clargs = malloc( sizeof( struct cl_zlascal_args_s ) );
+ clargs->uplo = uplo;
+ clargs->m = m;
+ clargs->n = n;
+ clargs->alpha = alpha;
+ clargs->tileA = A->get_blktile( A, Am, An );
+ }
+
/* Callback fro profiling information */
callback = options->profiling ? cl_zlascal_callback : NULL;
@@ -93,7 +97,7 @@ void INSERT_TASK_zlascal( const RUNTIME_option_t *options,
rt_starpu_insert_task(
&cl_zlascal,
/* Task codelet arguments */
- STARPU_VALUE, &clargs, sizeof(struct cl_zlascal_args_s),
+ STARPU_CL_ARGS, clargs, sizeof(struct cl_zlascal_args_s),
STARPU_RW, RTBLKADDR(A, CHAMELEON_Complex64_t, Am, An),
/* Common task arguments */
diff --git a/runtime/starpu/codelets/codelet_zlaset.c b/runtime/starpu/codelets/codelet_zlaset.c
index c3fb53b0fdb0b7732c69d785fbf2c381a4064e60..95805c0142f2b966125bac51b9290bcf0f324bdf 100644
--- a/runtime/starpu/codelets/codelet_zlaset.c
+++ b/runtime/starpu/codelets/codelet_zlaset.c
@@ -39,13 +39,12 @@ struct cl_zlaset_args_s {
static void
cl_zlaset_cpu_func( void *descr[], void *cl_arg )
{
- struct cl_zlaset_args_s clargs;
+ struct cl_zlaset_args_s *clargs = (struct cl_zlaset_args_s *)cl_arg;
CHAM_tile_t *tileA;
tileA = cti_interface_get(descr[0]);
- starpu_codelet_unpack_args( cl_arg, &clargs );
- TCORE_zlaset( clargs.uplo, clargs.m, clargs.n, clargs.alpha, clargs.beta, tileA );
+ TCORE_zlaset( clargs->uplo, clargs->m, clargs->n, clargs->alpha, clargs->beta, tileA );
}
#endif /* !defined(CHAMELEON_SIMULATION) */
@@ -59,25 +58,30 @@ void INSERT_TASK_zlaset( const RUNTIME_option_t *options,
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 ),
- };
+ struct cl_zlaset_args_s *clargs = NULL;
void (*callback)(void*);
RUNTIME_request_t *request = options->request;
starpu_option_request_t *schedopt = (starpu_option_request_t *)(request->schedopt);
int workerid;
+ int exec = 0;
char *cl_name = "zlaset";
/* Handle cache */
CHAMELEON_BEGIN_ACCESS_DECLARATION;
CHAMELEON_ACCESS_W(A, Am, An);
+ exec = __chameleon_need_exec;
CHAMELEON_END_ACCESS_DECLARATION;
+ if ( exec ) {
+ clargs = malloc( sizeof( struct cl_zlaset_args_s ) );
+ clargs->uplo = uplo;
+ clargs->m = m;
+ clargs->n = n;
+ clargs->alpha = alpha;
+ clargs->beta = beta;
+ clargs->tileA = A->get_blktile( A, Am, An );
+ }
+
/* Callback fro profiling information */
callback = options->profiling ? cl_zlaset_callback : NULL;
@@ -88,7 +92,7 @@ void INSERT_TASK_zlaset( const RUNTIME_option_t *options,
rt_starpu_insert_task(
&cl_zlaset,
/* Task codelet arguments */
- STARPU_VALUE, &clargs, sizeof(struct cl_zlaset_args_s),
+ STARPU_CL_ARGS, clargs, sizeof(struct cl_zlaset_args_s),
STARPU_W, RTBLKADDR(A, CHAMELEON_Complex64_t, Am, An),
/* Common task arguments */
diff --git a/runtime/starpu/codelets/codelet_zlauum.c b/runtime/starpu/codelets/codelet_zlauum.c
index 740fdc0fba5bbdcde1ee0258821013d59836ed4b..700352172f44c4f3fdbda09f0474d073cb44bc44 100644
--- a/runtime/starpu/codelets/codelet_zlauum.c
+++ b/runtime/starpu/codelets/codelet_zlauum.c
@@ -37,13 +37,12 @@ struct cl_zlauum_args_s {
static void
cl_zlauum_cpu_func(void *descr[], void *cl_arg)
{
- struct cl_zlauum_args_s clargs;
+ struct cl_zlauum_args_s *clargs = (struct cl_zlauum_args_s *)cl_arg;
CHAM_tile_t *tileA;
tileA = cti_interface_get(descr[0]);
- starpu_codelet_unpack_args( cl_arg, &clargs );
- TCORE_zlauum( clargs.uplo, clargs.n, tileA );
+ TCORE_zlauum( clargs->uplo, clargs->n, tileA );
}
#endif /* !defined(CHAMELEON_SIMULATION) */
@@ -56,22 +55,27 @@ 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 )
{
- struct cl_zlauum_args_s clargs = {
- .uplo = uplo,
- .n = n,
- .tileA = A->get_blktile( A, Am, An ),
- };
+ struct cl_zlauum_args_s *clargs = NULL;
void (*callback)(void*);
RUNTIME_request_t *request = options->request;
starpu_option_request_t *schedopt = (starpu_option_request_t *)(request->schedopt);
int workerid;
+ int exec = 0;
char *cl_name = "zlauum";
/* Handle cache */
CHAMELEON_BEGIN_ACCESS_DECLARATION;
CHAMELEON_ACCESS_RW(A, Am, An);
+ exec = __chameleon_need_exec;
CHAMELEON_END_ACCESS_DECLARATION;
+ if ( exec ) {
+ 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;
@@ -82,7 +86,7 @@ void INSERT_TASK_zlauum( const RUNTIME_option_t *options,
rt_starpu_insert_task(
&cl_zlauum,
/* Task codelet arguments */
- STARPU_VALUE, &clargs, sizeof(struct cl_zlauum_args_s),
+ STARPU_CL_ARGS, clargs, sizeof(struct cl_zlauum_args_s),
STARPU_RW, RTBLKADDR(A, CHAMELEON_Complex64_t, Am, An),
/* Common task arguments */
diff --git a/runtime/starpu/codelets/codelet_zplghe.c b/runtime/starpu/codelets/codelet_zplghe.c
index a58aed69b013d4df222b681f7bee42275b98bf53..68bf93fc960e44beeed0468da28694d82562fce5 100644
--- a/runtime/starpu/codelets/codelet_zplghe.c
+++ b/runtime/starpu/codelets/codelet_zplghe.c
@@ -41,14 +41,13 @@ struct cl_zplghe_args_s {
#if !defined(CHAMELEON_SIMULATION)
static void cl_zplghe_cpu_func(void *descr[], void *cl_arg)
{
- struct cl_zplghe_args_s clargs;
+ struct cl_zplghe_args_s *clargs = (struct cl_zplghe_args_s *)cl_arg;
CHAM_tile_t *tileA;
tileA = cti_interface_get(descr[0]);
- starpu_codelet_unpack_args( cl_arg, &clargs );
- TCORE_zplghe( clargs.bump, clargs.m, clargs.n, tileA,
- clargs.bigM, clargs.m0, clargs.n0, clargs.seed );
+ TCORE_zplghe( clargs->bump, clargs->m, clargs->n, tileA,
+ clargs->bigM, clargs->m0, clargs->n0, clargs->seed );
}
#endif /* !defined(CHAMELEON_SIMULATION) */
@@ -61,27 +60,32 @@ 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 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,
- };
+ struct cl_zplghe_args_s *clargs = NULL;
void (*callback)(void*);
RUNTIME_request_t *request = options->request;
starpu_option_request_t *schedopt = (starpu_option_request_t *)(request->schedopt);
int workerid;
+ int exec = 0;
char *cl_name = "zplghe";
/* Handle cache */
CHAMELEON_BEGIN_ACCESS_DECLARATION;
CHAMELEON_ACCESS_W(A, Am, An);
+ exec = __chameleon_need_exec;
CHAMELEON_END_ACCESS_DECLARATION;
+ if ( exec ) {
+ clargs = malloc( sizeof( struct cl_zplghe_args_s ) );
+ 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;
+ clargs->seed = seed;
+ }
+
/* Callback fro profiling information */
callback = options->profiling ? cl_zplghe_callback : NULL;
@@ -92,7 +96,7 @@ void INSERT_TASK_zplghe( const RUNTIME_option_t *options,
rt_starpu_insert_task(
&cl_zplghe,
/* Task codelet arguments */
- STARPU_VALUE, &clargs, sizeof(struct cl_zplghe_args_s),
+ STARPU_CL_ARGS, clargs, sizeof(struct cl_zplghe_args_s),
STARPU_W, RTBLKADDR(A, CHAMELEON_Complex64_t, Am, An),
/* Common task arguments */
diff --git a/runtime/starpu/codelets/codelet_zplgsy.c b/runtime/starpu/codelets/codelet_zplgsy.c
index f3afa23ebea541f9989767f834a1900367ae24ff..2d342bac33f25e4e3e5ff60be624fe0d81b798c5 100644
--- a/runtime/starpu/codelets/codelet_zplgsy.c
+++ b/runtime/starpu/codelets/codelet_zplgsy.c
@@ -41,14 +41,13 @@ struct cl_zplgsy_args_s {
#if !defined(CHAMELEON_SIMULATION)
static void cl_zplgsy_cpu_func(void *descr[], void *cl_arg)
{
- struct cl_zplgsy_args_s clargs;
+ struct cl_zplgsy_args_s *clargs = (struct cl_zplgsy_args_s *)cl_arg;
CHAM_tile_t *tileA;
tileA = cti_interface_get(descr[0]);
- starpu_codelet_unpack_args( cl_arg, &clargs );
- TCORE_zplgsy( clargs.bump, clargs.m, clargs.n, tileA,
- clargs.bigM, clargs.m0, clargs.n0, clargs.seed );
+ TCORE_zplgsy( clargs->bump, clargs->m, clargs->n, tileA,
+ clargs->bigM, clargs->m0, clargs->n0, clargs->seed );
}
#endif /* !defined(CHAMELEON_SIMULATION) */
@@ -61,27 +60,32 @@ 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 )
{
- 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,
- };
+ struct cl_zplgsy_args_s *clargs = NULL;
void (*callback)(void*);
RUNTIME_request_t *request = options->request;
starpu_option_request_t *schedopt = (starpu_option_request_t *)(request->schedopt);
int workerid;
+ int exec = 0;
char *cl_name = "zplgsy";
/* Handle cache */
CHAMELEON_BEGIN_ACCESS_DECLARATION;
CHAMELEON_ACCESS_W(A, Am, An);
+ exec = __chameleon_need_exec;
CHAMELEON_END_ACCESS_DECLARATION;
+ if ( exec ) {
+ clargs = malloc( sizeof( struct cl_zplgsy_args_s ) );
+ 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;
+ clargs->seed = seed;
+ }
+
/* Callback fro profiling information */
callback = options->profiling ? cl_zplgsy_callback : NULL;
@@ -92,7 +96,7 @@ void INSERT_TASK_zplgsy( const RUNTIME_option_t *options,
rt_starpu_insert_task(
&cl_zplgsy,
/* Task codelet arguments */
- STARPU_VALUE, &clargs, sizeof(struct cl_zplgsy_args_s),
+ STARPU_CL_ARGS, clargs, sizeof(struct cl_zplgsy_args_s),
STARPU_W, RTBLKADDR(A, CHAMELEON_Complex64_t, Am, An),
/* Common task arguments */
diff --git a/runtime/starpu/codelets/codelet_zplrnt.c b/runtime/starpu/codelets/codelet_zplrnt.c
index 05df0d107f8b46bbb6df07891e5373a9308b8f1a..36191f43e2486467220f6ee4cda694cd857c1d4a 100644
--- a/runtime/starpu/codelets/codelet_zplrnt.c
+++ b/runtime/starpu/codelets/codelet_zplrnt.c
@@ -41,14 +41,13 @@ struct cl_zplrnt_args_s {
static void
cl_zplrnt_cpu_func(void *descr[], void *cl_arg)
{
- struct cl_zplrnt_args_s clargs;
+ struct cl_zplrnt_args_s *clargs = (struct cl_zplrnt_args_s *)cl_arg;
CHAM_tile_t *tileA;
tileA = cti_interface_get(descr[0]);
- starpu_codelet_unpack_args( cl_arg, &clargs );
- TCORE_zplrnt( clargs.m, clargs.n, tileA,
- clargs.bigM, clargs.m0, clargs.n0, clargs.seed );
+ TCORE_zplrnt( clargs->m, clargs->n, tileA,
+ clargs->bigM, clargs->m0, clargs->n0, clargs->seed );
}
#endif /* !defined(CHAMELEON_SIMULATION) */
@@ -61,26 +60,31 @@ 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,
- };
+ struct cl_zplrnt_args_s *clargs = NULL;
void (*callback)(void*);
RUNTIME_request_t *request = options->request;
starpu_option_request_t *schedopt = (starpu_option_request_t *)(request->schedopt);
int workerid;
+ int exec = 0;
char *cl_name = "zplrnt";
/* Handle cache */
CHAMELEON_BEGIN_ACCESS_DECLARATION;
CHAMELEON_ACCESS_W(A, Am, An);
+ exec = __chameleon_need_exec;
CHAMELEON_END_ACCESS_DECLARATION;
+ if ( exec ) {
+ 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;
+ clargs->seed = seed;
+ }
+
/* Callback fro profiling information */
callback = options->profiling ? cl_zplrnt_callback : NULL;
@@ -91,7 +95,7 @@ void INSERT_TASK_zplrnt( const RUNTIME_option_t *options,
rt_starpu_insert_task(
&cl_zplrnt,
/* Task codelet arguments */
- STARPU_VALUE, &clargs, sizeof(struct cl_zplrnt_args_s),
+ STARPU_CL_ARGS, clargs, sizeof(struct cl_zplrnt_args_s),
STARPU_W, RTBLKADDR(A, CHAMELEON_Complex64_t, Am, An),
/* Common task arguments */
diff --git a/runtime/starpu/codelets/codelet_zpotrf.c b/runtime/starpu/codelets/codelet_zpotrf.c
index 19da18c46dcac24a734ffcca15693bf13c0682ae..883645b31185692edbac2bba01005f290a738ce7 100644
--- a/runtime/starpu/codelets/codelet_zpotrf.c
+++ b/runtime/starpu/codelets/codelet_zpotrf.c
@@ -40,17 +40,16 @@ struct cl_zpotrf_args_s {
static void
cl_zpotrf_cpu_func(void *descr[], void *cl_arg)
{
- struct cl_zpotrf_args_s clargs;
+ struct cl_zpotrf_args_s *clargs = (struct cl_zpotrf_args_s *)cl_arg;
CHAM_tile_t *tileA;
int info = 0;
tileA = cti_interface_get(descr[0]);
- starpu_codelet_unpack_args( cl_arg, &clargs );
- TCORE_zpotrf( clargs.uplo, clargs.n, tileA, &info );
+ TCORE_zpotrf( clargs->uplo, clargs->n, tileA, &info );
- if ( (clargs.sequence->status == CHAMELEON_SUCCESS) && (info != 0) ) {
- RUNTIME_sequence_flush( NULL, clargs.sequence, clargs.request, clargs.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) */
@@ -65,25 +64,30 @@ void INSERT_TASK_zpotrf( const RUNTIME_option_t *options,
const CHAM_desc_t *A, int Am, int An,
int iinfo )
{
- 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,
- };
+ struct cl_zpotrf_args_s *clargs = NULL;
void (*callback)(void*);
RUNTIME_request_t *request = options->request;
starpu_option_request_t *schedopt = (starpu_option_request_t *)(request->schedopt);
int workerid;
+ int exec = 0;
char *cl_name = "zpotrf";
/* Handle cache */
CHAMELEON_BEGIN_ACCESS_DECLARATION;
CHAMELEON_ACCESS_RW(A, Am, An);
+ exec = __chameleon_need_exec;
CHAMELEON_END_ACCESS_DECLARATION;
+ if ( exec ) {
+ 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;
+ }
+
/* Callback fro profiling information */
callback = options->profiling ? cl_zpotrf_callback : NULL;
@@ -94,7 +98,7 @@ void INSERT_TASK_zpotrf( const RUNTIME_option_t *options,
rt_starpu_insert_task(
&cl_zpotrf,
/* Task codelet arguments */
- STARPU_VALUE, &clargs, sizeof(struct cl_zpotrf_args_s),
+ STARPU_CL_ARGS, clargs, sizeof(struct cl_zpotrf_args_s),
STARPU_RW, RTBLKADDR(A, CHAMELEON_Complex64_t, Am, An),
/* Common task arguments */
diff --git a/runtime/starpu/codelets/codelet_zsyrk.c b/runtime/starpu/codelets/codelet_zsyrk.c
index 9af6d9a824498c908b98eb08e8bf1dcd50c665c0..a878aaa19adfc5c60ece27a785a7dd2fd1e67986 100644
--- a/runtime/starpu/codelets/codelet_zsyrk.c
+++ b/runtime/starpu/codelets/codelet_zsyrk.c
@@ -19,6 +19,7 @@
* @author Cedric Castagnede
* @author Lucas Barros de Assis
* @author Florent Pruvost
+ * @author Gwenole Lucas
* @date 2021-03-16
* @precisions normal z -> c d s
*
@@ -41,38 +42,35 @@ struct cl_zsyrk_args_s {
static void
cl_zsyrk_cpu_func(void *descr[], void *cl_arg)
{
- struct cl_zsyrk_args_s clargs;
+ struct cl_zsyrk_args_s *clargs = (struct cl_zsyrk_args_s *)cl_arg;
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, &clargs );
- TCORE_zsyrk( clargs.uplo, clargs.trans, clargs.n, clargs.k,
- clargs.alpha, tileA, clargs.beta, tileC );
+ TCORE_zsyrk( clargs->uplo, clargs->trans, clargs->n, clargs->k,
+ clargs->alpha, tileA, clargs->beta, tileC );
}
#if defined(CHAMELEON_USE_CUDA)
static void
cl_zsyrk_cuda_func(void *descr[], void *cl_arg)
{
- struct cl_zsyrk_args_s clargs;
+ struct cl_zsyrk_args_s *clargs = (struct cl_zsyrk_args_s *)cl_arg;
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, &clargs );
-
RUNTIME_getStream(stream);
CUDA_zsyrk(
- clargs.uplo, clargs.trans, clargs.n, clargs.k,
- (cuDoubleComplex*)&(clargs.alpha),
+ clargs->uplo, clargs->trans, clargs->n, clargs->k,
+ (cuDoubleComplex*)&(clargs->alpha),
tileA->mat, tileA->ld,
- (cuDoubleComplex*)&(clargs.beta),
+ (cuDoubleComplex*)&(clargs->beta),
tileC->mat, tileC->ld,
stream );
@@ -101,28 +99,33 @@ void INSERT_TASK_zsyrk( const RUNTIME_option_t *options,
beta, C, Cm, Cn );
}
- 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 ),
- };
+ struct cl_zsyrk_args_s *clargs = NULL;
void (*callback)(void*);
RUNTIME_request_t *request = options->request;
starpu_option_request_t *schedopt = (starpu_option_request_t *)(request->schedopt);
int workerid, accessC;
+ int exec = 0;
char *cl_name = "zsyrk";
/* Handle cache */
CHAMELEON_BEGIN_ACCESS_DECLARATION;
CHAMELEON_ACCESS_R(A, Am, An);
CHAMELEON_ACCESS_RW(C, Cm, Cn);
+ exec = __chameleon_need_exec;
CHAMELEON_END_ACCESS_DECLARATION;
+ if ( exec ) {
+ clargs = malloc( sizeof( struct cl_zsyrk_args_s ) );
+ clargs->uplo = uplo;
+ clargs->trans = trans;
+ 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 */
callback = options->profiling ? cl_zsyrk_callback : NULL;
@@ -135,9 +138,8 @@ void INSERT_TASK_zsyrk( const RUNTIME_option_t *options,
/* Insert the task */
rt_starpu_insert_task(
&cl_zsyrk,
-
/* Task codelet arguments */
- STARPU_VALUE, &clargs, sizeof(struct cl_zsyrk_args_s),
+ STARPU_CL_ARGS, clargs, sizeof(struct cl_zsyrk_args_s),
STARPU_R, RTBLKADDR(A, CHAMELEON_Complex64_t, Am, An),
accessC, RTBLKADDR(C, CHAMELEON_Complex64_t, Cm, Cn),
diff --git a/runtime/starpu/codelets/codelet_ztradd.c b/runtime/starpu/codelets/codelet_ztradd.c
index 6555e6d1737be7bb41244a6539d02c8583f3b072..d59f03a5f3e4e9695c728f4eeb142add91e5192f 100644
--- a/runtime/starpu/codelets/codelet_ztradd.c
+++ b/runtime/starpu/codelets/codelet_ztradd.c
@@ -38,16 +38,15 @@ struct cl_ztradd_args_s {
static void
cl_ztradd_cpu_func(void *descr[], void *cl_arg)
{
- struct cl_ztradd_args_s clargs;
+ struct cl_ztradd_args_s *clargs = (struct cl_ztradd_args_s *)cl_arg;
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, &clargs );
- TCORE_ztradd( clargs.uplo, clargs.trans, clargs.m, clargs.n,
- clargs.alpha, tileA, clargs.beta, tileB );
+ TCORE_ztradd( clargs->uplo, clargs->trans, clargs->m, clargs->n,
+ clargs->alpha, tileA, clargs->beta, tileB );
}
#endif /* !defined(CHAMELEON_SIMULATION) */
@@ -66,28 +65,33 @@ void INSERT_TASK_ztradd( const RUNTIME_option_t *options,
beta, B, Bm, Bn );
}
- 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 ),
- };
+ struct cl_ztradd_args_s *clargs = NULL;
void (*callback)(void*);
RUNTIME_request_t *request = options->request;
starpu_option_request_t *schedopt = (starpu_option_request_t *)(request->schedopt);
int workerid, accessB;
+ int exec = 0;
char *cl_name = "ztradd";
/* Handle cache */
CHAMELEON_BEGIN_ACCESS_DECLARATION;
CHAMELEON_ACCESS_R(A, Am, An);
CHAMELEON_ACCESS_RW(B, Bm, Bn);
+ exec = __chameleon_need_exec;
CHAMELEON_END_ACCESS_DECLARATION;
+ if ( exec ) {
+ clargs = malloc( sizeof( struct cl_ztradd_args_s ) );
+ clargs->uplo = uplo;
+ clargs->trans = trans;
+ 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 */
callback = options->profiling ? cl_ztradd_callback : NULL;
@@ -101,7 +105,7 @@ void INSERT_TASK_ztradd( const RUNTIME_option_t *options,
rt_starpu_insert_task(
&cl_ztradd,
/* Task codelet arguments */
- STARPU_VALUE, &clargs, sizeof(struct cl_ztradd_args_s),
+ STARPU_CL_ARGS, clargs, sizeof(struct cl_ztradd_args_s),
STARPU_R, RTBLKADDR(A, CHAMELEON_Complex64_t, Am, An),
accessB, RTBLKADDR(B, CHAMELEON_Complex64_t, Bm, Bn),
diff --git a/runtime/starpu/codelets/codelet_ztrmm.c b/runtime/starpu/codelets/codelet_ztrmm.c
index ac31d898f7e253f3fde11b9f4df37875480eaa5b..20f86bd38af1f4842aa6e853392f00ed7adfbea3 100644
--- a/runtime/starpu/codelets/codelet_ztrmm.c
+++ b/runtime/starpu/codelets/codelet_ztrmm.c
@@ -42,37 +42,34 @@ struct cl_ztrmm_args_s {
static void
cl_ztrmm_cpu_func(void *descr[], void *cl_arg)
{
- struct cl_ztrmm_args_s clargs;
+ struct cl_ztrmm_args_s *clargs = (struct cl_ztrmm_args_s *)cl_arg;
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, &clargs );
- TCORE_ztrmm( clargs.side, clargs.uplo, clargs.transA, clargs.diag,
- clargs.m, clargs.n, clargs.alpha, tileA, tileB );
+ 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)
{
- struct cl_ztrmm_args_s clargs;
+ struct cl_ztrmm_args_s *clargs = (struct cl_ztrmm_args_s *)cl_arg;
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, &clargs );
-
RUNTIME_getStream(stream);
CUDA_ztrmm(
- clargs.side, clargs.uplo, clargs.transA, clargs.diag,
- clargs.m, clargs.n,
- (cuDoubleComplex*)&(clargs.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 );
@@ -97,29 +94,34 @@ void INSERT_TASK_ztrmm( const RUNTIME_option_t *options,
CHAMELEON_Complex64_t alpha, const CHAM_desc_t *A, int Am, int An,
const CHAM_desc_t *B, int Bm, int Bn )
{
- 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 ),
- };
+ struct cl_ztrmm_args_s *clargs = NULL;
void (*callback)(void*);
RUNTIME_request_t *request = options->request;
starpu_option_request_t *schedopt = (starpu_option_request_t *)(request->schedopt);
int workerid;
+ int exec = 0;
char *cl_name = "ztrmm";
/* Handle cache */
CHAMELEON_BEGIN_ACCESS_DECLARATION;
CHAMELEON_ACCESS_R(A, Am, An);
CHAMELEON_ACCESS_RW(B, Bm, Bn);
+ exec = __chameleon_need_exec;
CHAMELEON_END_ACCESS_DECLARATION;
+ if ( exec ) {
+ clargs = malloc( sizeof( struct cl_ztrmm_args_s ) );
+ clargs->side = side;
+ clargs->uplo = uplo;
+ clargs->transA = transA;
+ clargs->diag = diag;
+ 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;
@@ -130,7 +132,7 @@ void INSERT_TASK_ztrmm( const RUNTIME_option_t *options,
rt_starpu_insert_task(
&cl_ztrmm,
/* Task codelet arguments */
- STARPU_VALUE, &clargs, sizeof(struct cl_ztrmm_args_s),
+ STARPU_CL_ARGS, 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),
diff --git a/runtime/starpu/codelets/codelet_ztrsm.c b/runtime/starpu/codelets/codelet_ztrsm.c
index 00b45f3ebc2ce576eaaa387a055ea01099299851..e40dde763ed018a29003e39bc1b5ee377fbeb4df 100644
--- a/runtime/starpu/codelets/codelet_ztrsm.c
+++ b/runtime/starpu/codelets/codelet_ztrsm.c
@@ -19,6 +19,7 @@
* @author Cedric Castagnede
* @author Lucas Barros de Assis
* @author Florent Pruvost
+ * @author Gwenole Lucas
* @date 2021-03-16
* @precisions normal z -> c d s
*
@@ -42,37 +43,34 @@ struct cl_ztrsm_args_s {
static void
cl_ztrsm_cpu_func(void *descr[], void *cl_arg)
{
- struct cl_ztrsm_args_s clargs;
+ struct cl_ztrsm_args_s *clargs = (struct cl_ztrsm_args_s*)cl_arg;
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, &clargs );
- TCORE_ztrsm( clargs.side, clargs.uplo, clargs.transA, clargs.diag,
- clargs.m, clargs.n, clargs.alpha, tileA, tileB );
+ 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)
{
- struct cl_ztrsm_args_s clargs;
+ struct cl_ztrsm_args_s *clargs = (struct cl_ztrsm_args_s*)cl_arg;
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, &clargs );
-
RUNTIME_getStream(stream);
CUDA_ztrsm(
- clargs.side, clargs.uplo, clargs.transA, clargs.diag,
- clargs.m, clargs.n,
- (cuDoubleComplex*)&(clargs.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 );
@@ -97,29 +95,34 @@ void INSERT_TASK_ztrsm( const RUNTIME_option_t *options,
CHAMELEON_Complex64_t alpha, const CHAM_desc_t *A, int Am, int An,
const CHAM_desc_t *B, int Bm, int Bn )
{
- 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 ),
- };
+ struct cl_ztrsm_args_s *clargs = NULL;
void (*callback)(void*);
RUNTIME_request_t *request = options->request;
starpu_option_request_t *schedopt = (starpu_option_request_t *)(request->schedopt);
int workerid;
+ int exec = 0;
char *cl_name = "ztrsm";
/* Handle cache */
CHAMELEON_BEGIN_ACCESS_DECLARATION;
CHAMELEON_ACCESS_R(A, Am, An);
CHAMELEON_ACCESS_RW(B, Bm, Bn);
+ exec = __chameleon_need_exec;
CHAMELEON_END_ACCESS_DECLARATION;
+ if ( exec ) {
+ clargs = malloc( sizeof( struct cl_ztrsm_args_s ) );
+ clargs->side = side;
+ clargs->uplo = uplo;
+ clargs->transA = transA;
+ clargs->diag = diag;
+ 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;
@@ -130,7 +133,7 @@ void INSERT_TASK_ztrsm( const RUNTIME_option_t *options,
rt_starpu_insert_task(
&cl_ztrsm,
/* Task codelet arguments */
- STARPU_VALUE, &clargs, sizeof(struct cl_ztrsm_args_s),
+ STARPU_CL_ARGS, 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),
diff --git a/runtime/starpu/codelets/codelet_ztrtri.c b/runtime/starpu/codelets/codelet_ztrtri.c
index 157fa8c63a6a1d97af856659d42065a6792b7f29..9732e84c36f4c88c310ee35a8028fc66d4c7b572 100644
--- a/runtime/starpu/codelets/codelet_ztrtri.c
+++ b/runtime/starpu/codelets/codelet_ztrtri.c
@@ -41,17 +41,16 @@ struct cl_ztrtri_args_s {
static void
cl_ztrtri_cpu_func(void *descr[], void *cl_arg)
{
- struct cl_ztrtri_args_s clargs;
+ struct cl_ztrtri_args_s *clargs = (struct cl_ztrtri_args_s *)cl_arg;
CHAM_tile_t *tileA;
int info = 0;
tileA = cti_interface_get(descr[0]);
- starpu_codelet_unpack_args( cl_arg, &clargs );
- TCORE_ztrtri( clargs.uplo, clargs.diag, clargs.n, tileA, &info );
+ 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 );
+ if ( (clargs->sequence->status == CHAMELEON_SUCCESS) && (info != 0) ) {
+ RUNTIME_sequence_flush( NULL, clargs->sequence, clargs->request, clargs->iinfo+info );
}
}
#endif /* !defined(CHAMELEON_SIMULATION) */
@@ -66,26 +65,31 @@ void INSERT_TASK_ztrtri( const RUNTIME_option_t *options,
const CHAM_desc_t *A, int Am, int An,
int iinfo )
{
- 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,
- };
+ struct cl_ztrtri_args_s *clargs = NULL;
void (*callback)(void*);
RUNTIME_request_t *request = options->request;
starpu_option_request_t *schedopt = (starpu_option_request_t *)(request->schedopt);
int workerid;
+ int exec = 0;
char *cl_name = "ztrtri";
/* Handle cache */
CHAMELEON_BEGIN_ACCESS_DECLARATION;
CHAMELEON_ACCESS_RW(A, Am, An);
+ exec = __chameleon_need_exec;
CHAMELEON_END_ACCESS_DECLARATION;
+ if ( exec ) {
+ clargs = malloc( sizeof( struct cl_ztrtri_args_s ) );
+ 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;
+ }
+
/* Callback fro profiling information */
callback = options->profiling ? cl_ztrtri_callback : NULL;
@@ -96,7 +100,7 @@ void INSERT_TASK_ztrtri( const RUNTIME_option_t *options,
rt_starpu_insert_task(
&cl_ztrtri,
/* Task codelet arguments */
- STARPU_VALUE, &clargs, sizeof(struct cl_ztrtri_args_s),
+ STARPU_CL_ARGS, clargs, sizeof(struct cl_ztrtri_args_s),
STARPU_RW, RTBLKADDR(A, CHAMELEON_Complex64_t, Am, An),
/* Common task arguments */