From 7982e209431d8220a689d53d8852ed4e6ffe04ee Mon Sep 17 00:00:00 2001
From: Alycia Lisito <alycia.lisito@inria.fr>
Date: Fri, 23 Feb 2024 11:21:14 +0100
Subject: [PATCH] zgetrf: correct filter
---
compute/pzgetrf.c | 10 ++++++----
control/descriptor_helpers.c | 18 ++++++++++++++++++
include/chameleon/descriptor_helpers.h | 8 ++++++++
.../starpu/codelets/codelet_zgetrf_blocked.c | 6 +++---
.../starpu/codelets/codelet_zgetrf_percol.c | 5 ++---
5 files changed, 37 insertions(+), 10 deletions(-)
diff --git a/compute/pzgetrf.c b/compute/pzgetrf.c
index 83d9accfd..108c4a201 100644
--- a/compute/pzgetrf.c
+++ b/compute/pzgetrf.c
@@ -131,7 +131,6 @@ chameleon_pzgetrf_panel_facto_percol( struct chameleon_pzgetrf_s *ws,
* Algorithm per column with pivoting
*/
for (h=0; h<=minmn; h++){
-
INSERT_TASK_zgetrf_percol_diag(
options,
h, k * A->mb,
@@ -139,8 +138,6 @@ chameleon_pzgetrf_panel_facto_percol( struct chameleon_pzgetrf_s *ws,
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,
@@ -193,7 +190,6 @@ chameleon_pzgetrf_panel_facto_blocked( struct chameleon_pzgetrf_s *ws,
ipiv );
for (m = k+1; m < A->mt; m++) {
-
INSERT_TASK_zgetrf_blocked_offdiag(
options,
j, m * A->mb, ws->ib,
@@ -371,7 +367,13 @@ void chameleon_pzgetrf( struct chameleon_pzgetrf_s *ws,
RUNTIME_iteration_push( chamctxt, k );
options.priority = A->nt;
+ /*
+ * Do the panel factorization only if the current proc contributes in the
+ * block column k.
+ */
+ options.forcesub = chameleon_involved_in_panelk_2dbc( A, k );
chameleon_pzgetrf_panel_facto( ws, A, IPIV, k, &options );
+ options.forcesub = 0;
for (n = k+1; n < A->nt; n++) {
options.priority = A->nt-n;
diff --git a/control/descriptor_helpers.c b/control/descriptor_helpers.c
index 775748014..9cae18835 100644
--- a/control/descriptor_helpers.c
+++ b/control/descriptor_helpers.c
@@ -82,6 +82,24 @@ int chameleon_getrankof_2d_diag( const CHAM_desc_t *A, int m, int n )
return (mm % A->p) * A->q + (mm % A->q);
}
+/**
+ * @brief Test if the current MPI process is involved in the panel k for 2DBC distributions.
+ *
+ * @param[in] A
+ * The matrix descriptor.
+ *
+ * @param[in] k
+ * The index of the panel to test.
+ *
+ * @return 1 if the current MPI process contributes to the panel k.
+ * 0 if the current MPI process doesn't contribute to the panel k.
+ *
+ */
+int chameleon_involved_in_panelk_2dbc( const CHAM_desc_t *A, int k ) {
+ int myrank = A->myrank;
+ return ( myrank % A->q == k % A->q );
+}
+
/**
* @brief Initializes a custom distribution based on an external file.
*
diff --git a/include/chameleon/descriptor_helpers.h b/include/chameleon/descriptor_helpers.h
index 5716b2582..da79d0486 100644
--- a/include/chameleon/descriptor_helpers.h
+++ b/include/chameleon/descriptor_helpers.h
@@ -57,6 +57,14 @@ int chameleon_getrankof_custom_init ( custom_dist_t **dist, const char *filena
int chameleon_getrankof_custom_destroy( custom_dist_t **dist );
int chameleon_getrankof_custom ( const CHAM_desc_t *A, int m, int n );
+/**
+ * @}
+ * @name Panel involvement functions
+ * @{
+ */
+
+int chameleon_involved_in_panelk_2dbc( const CHAM_desc_t *A, int An );
+
/**
* @}
* @name Block address functions
diff --git a/runtime/starpu/codelets/codelet_zgetrf_blocked.c b/runtime/starpu/codelets/codelet_zgetrf_blocked.c
index 03f4e5553..832132f4f 100644
--- a/runtime/starpu/codelets/codelet_zgetrf_blocked.c
+++ b/runtime/starpu/codelets/codelet_zgetrf_blocked.c
@@ -114,7 +114,7 @@ void INSERT_TASK_zgetrf_blocked_diag( const RUNTIME_option_t *options,
/* Handle cache */
CHAMELEON_BEGIN_ACCESS_DECLARATION;
- CHAMELEON_ACCESS_RW(A, Am, An);
+ CHAMELEON_ACCESS_RW( A, Am, An );
CHAMELEON_END_ACCESS_DECLARATION;
/* Refine name */
@@ -197,7 +197,7 @@ void INSERT_TASK_zgetrf_blocked_offdiag( const RUNTIME_option_t *options,
/* Handle cache */
CHAMELEON_BEGIN_ACCESS_DECLARATION;
- CHAMELEON_ACCESS_RW(A, Am, An);
+ CHAMELEON_ACCESS_RW( A, Am, An );
CHAMELEON_END_ACCESS_DECLARATION;
/* Refine name */
@@ -275,7 +275,7 @@ void INSERT_TASK_zgetrf_blocked_trsm( const RUNTIME_option_t *options,
/* Handle cache */
CHAMELEON_BEGIN_ACCESS_DECLARATION;
- CHAMELEON_ACCESS_RW(U, Um, Un);
+ CHAMELEON_ACCESS_RW( U, Um, Un );
CHAMELEON_END_ACCESS_DECLARATION;
/* Refine name */
diff --git a/runtime/starpu/codelets/codelet_zgetrf_percol.c b/runtime/starpu/codelets/codelet_zgetrf_percol.c
index 4b2de691a..7829c3f3e 100644
--- a/runtime/starpu/codelets/codelet_zgetrf_percol.c
+++ b/runtime/starpu/codelets/codelet_zgetrf_percol.c
@@ -92,7 +92,7 @@ void INSERT_TASK_zgetrf_percol_diag( const RUNTIME_option_t *options,
/* Handle cache */
CHAMELEON_BEGIN_ACCESS_DECLARATION;
- CHAMELEON_ACCESS_RW(A, Am, An);
+ CHAMELEON_ACCESS_RW( A, Am, An );
CHAMELEON_END_ACCESS_DECLARATION;
/* Refine name */
@@ -159,10 +159,9 @@ void INSERT_TASK_zgetrf_percol_offdiag( const RUNTIME_option_t *options,
void (*callback)(void*) = options->profiling ? cl_zgetrf_percol_offdiag_callback : NULL;
char *cl_name = "zgetrf_percol_offdiag";
-
/* Handle cache */
CHAMELEON_BEGIN_ACCESS_DECLARATION;
- CHAMELEON_ACCESS_RW(A, Am, An);
+ CHAMELEON_ACCESS_RW( A, Am, An );
CHAMELEON_END_ACCESS_DECLARATION;
/* Refine name */
--
GitLab