From fb6ccd8c0fde2b3abf284d518e93fb7511303edb Mon Sep 17 00:00:00 2001
From: Alycia Lisito <alycia.lisito@inria.fr>
Date: Wed, 14 Feb 2024 16:17:40 +0100
Subject: [PATCH] zgetrf batched: add batched percol algorithm
---
compute/pzgetrf.c | 58 +++++++++++++++++++++++++++++++++++++++-
testing/CTestLists.cmake | 6 +++++
2 files changed, 63 insertions(+), 1 deletion(-)
diff --git a/compute/pzgetrf.c b/compute/pzgetrf.c
index 99152bc04..000209624 100644
--- a/compute/pzgetrf.c
+++ b/compute/pzgetrf.c
@@ -157,6 +157,57 @@ chameleon_pzgetrf_panel_facto_percol( struct chameleon_pzgetrf_s *ws,
RUNTIME_ipiv_flushk( options->sequence, ipiv, k );
}
+/*
+ * Factorization of panel k - dynamic scheduling - batched version / stock
+ */
+static inline void
+chameleon_pzgetrf_panel_facto_percol_batched( struct chameleon_pzgetrf_s *ws,
+ CHAM_desc_t *A,
+ CHAM_ipiv_t *ipiv,
+ int k,
+ RUNTIME_option_t *options )
+{
+ int m, h;
+ int tempkm, tempkn, tempmm, minmn;
+ void **clargs = malloc( sizeof(char *) );
+ memset( clargs, 0, sizeof(char *) );
+
+ tempkm = k == A->mt-1 ? A->m-k*A->mb : A->mb;
+ tempkn = k == A->nt-1 ? A->n-k*A->nb : A->nb;
+ minmn = chameleon_min( tempkm, tempkn );
+
+ /* Update the number of column */
+ ipiv->n = minmn;
+
+ /*
+ * Algorithm per column with pivoting (no recursion)
+ */
+ /* Iterate on current panel column */
+ /* Since index h scales column h-1, we need to iterate up to minmn (included) */
+ for ( h = 0; h <= minmn; h++ ) {
+
+ INSERT_TASK_zgetrf_percol_diag( options, 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_panel_offdiag_batched( options, tempmm, tempkn, h, m * A->mb,
+ (void *)ws, A(m, k), clargs, ipiv );
+ }
+ INSERT_TASK_zgetrf_panel_offdiag_batched_flush( options, A, k, clargs, ipiv );
+
+ if ( h < minmn ) {
+ /* Reduce globally (between MPI processes) */
+ INSERT_TASK_ipiv_reducek( options, ipiv, k, h );
+ }
+ }
+
+ free( clargs );
+
+ /* Flush temporary data used for the pivoting */
+ INSERT_TASK_ipiv_to_perm( options, k * A->mb, tempkm, minmn, ipiv, k );
+ RUNTIME_ipiv_flushk( options->sequence, ipiv, k );
+}
+
static inline void
chameleon_pzgetrf_panel_facto_blocked( struct chameleon_pzgetrf_s *ws,
CHAM_desc_t *A,
@@ -235,7 +286,12 @@ chameleon_pzgetrf_panel_facto( struct chameleon_pzgetrf_s *ws,
break;
case ChamGetrfPPivPerColumn:
- chameleon_pzgetrf_panel_facto_percol( ws, A, ipiv, k, options );
+ if ( ws->batch_size > 1 ) {
+ chameleon_pzgetrf_panel_facto_percol_batched( ws, A, ipiv, k, options );
+ }
+ else {
+ chameleon_pzgetrf_panel_facto_percol( ws, A, ipiv, k, options );
+ }
break;
case ChamGetrfPPiv:
diff --git a/testing/CTestLists.cmake b/testing/CTestLists.cmake
index b429b5c02..98bdb1939 100644
--- a/testing/CTestLists.cmake
+++ b/testing/CTestLists.cmake
@@ -94,6 +94,12 @@ if (NOT CHAMELEON_SIMULATION)
set_tests_properties( test_${cat}_${prec}getrf_ppivpercol
PROPERTIES ENVIRONMENT "CHAMELEON_GETRF_ALGO=ppivpercolumn;CHAMELEON_GETRF_BATCH_SIZE=1" )
+ if ( ${cat} STREQUAL "shm" )
+ add_test( test_${cat}_${prec}getrf_ppivpercol_batch ${PREFIX} ${CMD} -c -t ${THREADS} -g ${gpus} -P 1 -f input/getrf_nopiv.in )
+ set_tests_properties( test_${cat}_${prec}getrf_ppivpercol_batch
+ PROPERTIES ENVIRONMENT "CHAMELEON_GETRF_ALGO=ppivpercolumn;CHAMELEON_GETRF_BATCH_SIZE=6" )
+ endif()
+
add_test( test_${cat}_${prec}getrf_ppiv ${PREFIX} ${CMD} -c -t ${THREADS} -g ${gpus} -P 1 -f input/getrf.in )
set_tests_properties( test_${cat}_${prec}getrf_ppiv
PROPERTIES ENVIRONMENT "CHAMELEON_GETRF_ALGO=ppiv;CHAMELEON_GETRF_BATCH_SIZE=1" )
--
GitLab