diff --git a/compute/pzgetrf.c b/compute/pzgetrf.c
index 64552d74636c4919a6283f4a856a10eb5c2f05f2..99152bc04575e72ab83819df54472161b57406f6 100644
--- a/compute/pzgetrf.c
+++ b/compute/pzgetrf.c
@@ -118,7 +118,7 @@ chameleon_pzgetrf_panel_facto_percol( struct chameleon_pzgetrf_s *ws,
RUNTIME_option_t *options )
{
int m, h;
- int tempkm, tempkn, minmn;
+ int tempkm, tempkn, tempmm, minmn;
tempkm = k == A->mt-1 ? A->m-k*A->mb : A->mb;
tempkn = k == A->nt-1 ? A->n-k*A->nb : A->nb;
@@ -133,14 +133,15 @@ chameleon_pzgetrf_panel_facto_percol( struct chameleon_pzgetrf_s *ws,
for (h=0; h<=minmn; h++){
INSERT_TASK_zgetrf_percol_diag(
options,
- h, k * A->mb,
+ tempkm, tempkn, h, k * A->mb,
A(k, k),
ipiv );
for (m = k+1; m < A->mt; m++) {
+ tempmm = (m == (A->mt - 1)) ? A->m - m * A->mb : A->mb;
INSERT_TASK_zgetrf_percol_offdiag(
options,
- h, m * A->mb,
+ tempmm, tempkn, h, m * A->mb,
A(m, k),
ipiv );
}
@@ -164,7 +165,7 @@ chameleon_pzgetrf_panel_facto_blocked( struct chameleon_pzgetrf_s *ws,
RUNTIME_option_t *options )
{
int m, h, b, nbblock;
- int tempkm, tempkn, minmn;
+ int tempkm, tempkn, tempmm, minmn;
tempkm = k == A->mt-1 ? A->m-k*A->mb : A->mb;
tempkn = k == A->nt-1 ? A->n-k*A->nb : A->nb;
@@ -185,14 +186,15 @@ chameleon_pzgetrf_panel_facto_blocked( struct chameleon_pzgetrf_s *ws,
INSERT_TASK_zgetrf_blocked_diag(
options,
- j, k * A->mb, ws->ib,
+ tempkm, tempkn, j, k * A->mb, ws->ib,
A(k, k), Up(k, k),
ipiv );
for (m = k+1; m < A->mt; m++) {
+ tempmm = (m == (A->mt - 1)) ? A->m - m * A->mb : A->mb;
INSERT_TASK_zgetrf_blocked_offdiag(
options,
- j, m * A->mb, ws->ib,
+ tempmm, tempkn, j, m * A->mb, ws->ib,
A(m, k), Up(k, k),
ipiv );
}
diff --git a/include/chameleon/tasks_z.h b/include/chameleon/tasks_z.h
index 3d2d6de0dda97e18c0ab12a91dc680412372c610..eb855ec34be8d002a9d99ab63276b32c722edb4b 100644
--- a/include/chameleon/tasks_z.h
+++ b/include/chameleon/tasks_z.h
@@ -508,23 +508,23 @@ void INSERT_TASK_zgetrf_nopiv_percol_trsm( const RUNTIME_option_t *options,
const CHAM_desc_t *U, int Um, int Un );
void INSERT_TASK_zgetrf_percol_diag( const RUNTIME_option_t *options,
- int h, int m0,
+ int m, int n, int h, int m0,
CHAM_desc_t *A, int Am, int An,
CHAM_ipiv_t *ws );
void INSERT_TASK_zgetrf_percol_offdiag( const RUNTIME_option_t *options,
- int h, int m0,
+ int m, int n, int h, int m0,
CHAM_desc_t *A, int Am, int An,
CHAM_ipiv_t *ws );
void INSERT_TASK_zgetrf_blocked_diag( const RUNTIME_option_t *options,
- int h, int m0, int ib,
+ int m, int n, int h, int m0, int ib,
CHAM_desc_t *A, int Am, int An,
CHAM_desc_t *U, int Um, int Un,
CHAM_ipiv_t *ws );
void INSERT_TASK_zgetrf_blocked_offdiag( const RUNTIME_option_t *options,
- int h, int m0, int ib,
+ int m, int n, int h, int m0, int ib,
CHAM_desc_t *A, int Am, int An,
CHAM_desc_t *U, int Um, int Un,
CHAM_ipiv_t *ws );
diff --git a/runtime/starpu/codelets/codelet_zgetrf_blocked.c b/runtime/starpu/codelets/codelet_zgetrf_blocked.c
index 8df527007f8d004fdc73019de014e2361b6954a0..2c6daa18d9bda1f7ff433305aa98ad77f648b4b5 100644
--- a/runtime/starpu/codelets/codelet_zgetrf_blocked.c
+++ b/runtime/starpu/codelets/codelet_zgetrf_blocked.c
@@ -29,7 +29,7 @@ CHAMELEON_CL_CB( zgetrf_blocked_trsm, cti_handle_get_m(task->handles[0]), 0,
#if !defined(CHAMELEON_SIMULATION)
static void cl_zgetrf_blocked_diag_cpu_func(void *descr[], void *cl_arg)
{
- int h, m0, ib;
+ int m, n, h, m0, ib;
RUNTIME_sequence_t *sequence;
RUNTIME_request_t *request;
CHAM_tile_t *tileA;
@@ -40,7 +40,7 @@ static void cl_zgetrf_blocked_diag_cpu_func(void *descr[], void *cl_arg)
CHAMELEON_Complex64_t *U = NULL;
int ldu = -1;;
- starpu_codelet_unpack_args( cl_arg, &h, &m0, &ib,
+ starpu_codelet_unpack_args( cl_arg, &m, &n, &h, &m0, &ib,
&sequence, &request );
tileA = cti_interface_get(descr[0]);
@@ -67,7 +67,7 @@ static void cl_zgetrf_blocked_diag_cpu_func(void *descr[], void *cl_arg)
nextpiv->h = h;
nextpiv->has_diag = 1;
- CORE_zgetrf_panel_diag( tileA->m, tileA->n, h, m0, ib,
+ CORE_zgetrf_panel_diag( m, n, h, m0, ib,
CHAM_tile_get_ptr( tileA ), tileA->ld,
U, ldu,
ipiv, &(nextpiv->pivot), &(prevpiv->pivot) );
@@ -87,7 +87,7 @@ static void cl_zgetrf_blocked_diag_cpu_func(void *descr[], void *cl_arg)
CODELETS_CPU( zgetrf_blocked_diag, cl_zgetrf_blocked_diag_cpu_func );
void INSERT_TASK_zgetrf_blocked_diag( const RUNTIME_option_t *options,
- int h, int m0, int ib,
+ int m, int n, int h, int m0, int ib,
CHAM_desc_t *A, int Am, int An,
CHAM_desc_t *U, int Um, int Un,
CHAM_ipiv_t *ipiv )
@@ -123,6 +123,8 @@ void INSERT_TASK_zgetrf_blocked_diag( const RUNTIME_option_t *options,
rt_starpu_insert_task(
codelet,
+ STARPU_VALUE, &m, sizeof(int),
+ STARPU_VALUE, &n, sizeof(int),
STARPU_VALUE, &h, sizeof(int),
STARPU_VALUE, &m0, sizeof(int),
STARPU_VALUE, &ib, sizeof(int),
@@ -146,7 +148,7 @@ void INSERT_TASK_zgetrf_blocked_diag( const RUNTIME_option_t *options,
#if !defined(CHAMELEON_SIMULATION)
static void cl_zgetrf_blocked_offdiag_cpu_func(void *descr[], void *cl_arg)
{
- int h, m0, ib;
+ int m, n, h, m0, ib;
RUNTIME_sequence_t *sequence;
RUNTIME_request_t *request;
CHAM_tile_t *tileA;
@@ -156,7 +158,7 @@ static void cl_zgetrf_blocked_offdiag_cpu_func(void *descr[], void *cl_arg)
CHAMELEON_Complex64_t *U = NULL;
int ldu = -1;;
- starpu_codelet_unpack_args( cl_arg, &h, &m0, &ib, &sequence, &request );
+ starpu_codelet_unpack_args( cl_arg, &m, &n, &h, &m0, &ib, &sequence, &request );
tileA = cti_interface_get(descr[0]);
nextpiv = (cppi_interface_t*) descr[1];
@@ -169,7 +171,7 @@ static void cl_zgetrf_blocked_offdiag_cpu_func(void *descr[], void *cl_arg)
nextpiv->h = h; /* Initialize in case it uses a copy */
- CORE_zgetrf_panel_offdiag( tileA->m, tileA->n, h, m0, ib,
+ CORE_zgetrf_panel_offdiag( m, n, h, m0, ib,
CHAM_tile_get_ptr(tileA), tileA->ld,
U, ldu,
&(nextpiv->pivot), &(prevpiv->pivot) );
@@ -182,7 +184,7 @@ static void cl_zgetrf_blocked_offdiag_cpu_func(void *descr[], void *cl_arg)
CODELETS_CPU(zgetrf_blocked_offdiag, cl_zgetrf_blocked_offdiag_cpu_func)
void INSERT_TASK_zgetrf_blocked_offdiag( const RUNTIME_option_t *options,
- int h, int m0, int ib,
+ int m, int n, int h, int m0, int ib,
CHAM_desc_t *A, int Am, int An,
CHAM_desc_t *U, int Um, int Un,
CHAM_ipiv_t *ipiv )
@@ -206,6 +208,8 @@ void INSERT_TASK_zgetrf_blocked_offdiag( const RUNTIME_option_t *options,
rt_starpu_insert_task(
codelet,
+ STARPU_VALUE, &m, sizeof(int),
+ STARPU_VALUE, &n, sizeof(int),
STARPU_VALUE, &h, sizeof(int),
STARPU_VALUE, &m0, sizeof(int),
STARPU_VALUE, &ib, sizeof(int),
diff --git a/runtime/starpu/codelets/codelet_zgetrf_percol.c b/runtime/starpu/codelets/codelet_zgetrf_percol.c
index 104eee81e51ab032f78534e6b37f9b1340c23f96..5d3f83b6ce046a72135c8f513c8cc23822159595 100644
--- a/runtime/starpu/codelets/codelet_zgetrf_percol.c
+++ b/runtime/starpu/codelets/codelet_zgetrf_percol.c
@@ -28,7 +28,7 @@ CHAMELEON_CL_CB( zgetrf_percol_offdiag, cti_handle_get_m(task->handles[0]), 0, 0
#if !defined(CHAMELEON_SIMULATION)
static void cl_zgetrf_percol_diag_cpu_func(void *descr[], void *cl_arg)
{
- int h, m0;
+ int m, n, h, m0;
RUNTIME_sequence_t *sequence;
RUNTIME_request_t *request;
CHAM_tile_t *tileA;
@@ -36,8 +36,7 @@ static void cl_zgetrf_percol_diag_cpu_func(void *descr[], void *cl_arg)
cppi_interface_t *nextpiv;
cppi_interface_t *prevpiv;
- starpu_codelet_unpack_args( cl_arg, &h, &m0,
- &sequence, &request );
+ starpu_codelet_unpack_args( cl_arg, &m, &n, &h, &m0, &sequence, &request );
tileA = cti_interface_get(descr[0]);
ipiv = (int *)STARPU_VECTOR_GET_PTR(descr[1]);
@@ -58,7 +57,7 @@ static void cl_zgetrf_percol_diag_cpu_func(void *descr[], void *cl_arg)
nextpiv->h = h;
nextpiv->has_diag = 1;
- CORE_zgetrf_panel_diag( tileA->m, tileA->n, h, m0, tileA->n,
+ CORE_zgetrf_panel_diag( m, n, h, m0, tileA->n,
CHAM_tile_get_ptr( tileA ), tileA->ld,
NULL, -1,
ipiv, &(nextpiv->pivot), &(prevpiv->pivot) );
@@ -66,7 +65,7 @@ static void cl_zgetrf_percol_diag_cpu_func(void *descr[], void *cl_arg)
if ( h > 0 ) {
cppi_display_dbg( prevpiv, stderr, "Prevpiv after call: " );
}
- if ( h < tileA->n ) {
+ if ( h < n ) {
cppi_display_dbg( nextpiv, stderr, "Nextpiv after call: " );
}
}
@@ -78,7 +77,7 @@ static void cl_zgetrf_percol_diag_cpu_func(void *descr[], void *cl_arg)
CODELETS_CPU( zgetrf_percol_diag, cl_zgetrf_percol_diag_cpu_func );
void INSERT_TASK_zgetrf_percol_diag( const RUNTIME_option_t *options,
- int h, int m0,
+ int m, int n, int h, int m0,
CHAM_desc_t *A, int Am, int An,
CHAM_ipiv_t *ipiv )
{
@@ -101,6 +100,8 @@ void INSERT_TASK_zgetrf_percol_diag( const RUNTIME_option_t *options,
rt_starpu_insert_task(
codelet,
+ STARPU_VALUE, &m, sizeof(int),
+ STARPU_VALUE, &n, sizeof(int),
STARPU_VALUE, &h, sizeof(int),
STARPU_VALUE, &m0, sizeof(int),
STARPU_VALUE, &(options->sequence), sizeof(RUNTIME_sequence_t*),
@@ -122,14 +123,14 @@ void INSERT_TASK_zgetrf_percol_diag( const RUNTIME_option_t *options,
#if !defined(CHAMELEON_SIMULATION)
static void cl_zgetrf_percol_offdiag_cpu_func(void *descr[], void *cl_arg)
{
- int h, m0;
+ int m, n, h, m0;
RUNTIME_sequence_t *sequence;
RUNTIME_request_t *request;
CHAM_tile_t *tileA;
cppi_interface_t *nextpiv;
cppi_interface_t *prevpiv;
- starpu_codelet_unpack_args( cl_arg, &h, &m0, &sequence, &request );
+ starpu_codelet_unpack_args( cl_arg, &m, &n, &h, &m0, &sequence, &request );
tileA = cti_interface_get(descr[0]);
nextpiv = (cppi_interface_t*) descr[1];
@@ -137,7 +138,7 @@ static void cl_zgetrf_percol_offdiag_cpu_func(void *descr[], void *cl_arg)
nextpiv->h = h; /* Initialize in case it uses a copy */
- CORE_zgetrf_panel_offdiag( tileA->m, tileA->n, h, m0, tileA->n,
+ CORE_zgetrf_panel_offdiag( m, n, h, m0, tileA->n,
CHAM_tile_get_ptr(tileA), tileA->ld,
NULL, -1,
&(nextpiv->pivot), &(prevpiv->pivot) );
@@ -150,7 +151,7 @@ static void cl_zgetrf_percol_offdiag_cpu_func(void *descr[], void *cl_arg)
CODELETS_CPU(zgetrf_percol_offdiag, cl_zgetrf_percol_offdiag_cpu_func)
void INSERT_TASK_zgetrf_percol_offdiag( const RUNTIME_option_t *options,
- int h, int m0,
+ int m, int n, int h, int m0,
CHAM_desc_t *A, int Am, int An,
CHAM_ipiv_t *ipiv )
{
@@ -170,6 +171,8 @@ void INSERT_TASK_zgetrf_percol_offdiag( const RUNTIME_option_t *options,
rt_starpu_insert_task(
codelet,
+ STARPU_VALUE, &m, sizeof(int),
+ STARPU_VALUE, &n, sizeof(int),
STARPU_VALUE, &h, sizeof(int),
STARPU_VALUE, &m0, sizeof(int),
STARPU_VALUE, &(options->sequence), sizeof(RUNTIME_sequence_t *),