From 1fdc9fe3ad6c498000df7293438009105fcf34b0 Mon Sep 17 00:00:00 2001
From: Terry Cojean <terry.cojean@inria.fr>
Date: Wed, 16 Dec 2015 14:39:49 +0000
Subject: [PATCH] Explain the restriction proposition for getrf_nopiv
---
timing/time_zgetrf_nopiv_tile.c | 6 ++++++
1 file changed, 6 insertions(+)
diff --git a/timing/time_zgetrf_nopiv_tile.c b/timing/time_zgetrf_nopiv_tile.c
index b074afc58..304d9d333 100644
--- a/timing/time_zgetrf_nopiv_tile.c
+++ b/timing/time_zgetrf_nopiv_tile.c
@@ -47,6 +47,12 @@ RunTest(int *iparam, double *dparam, morse_time_t *t_)
MORSE_zlacpy_Tile(MorseUpperLower, descA, descAC);
}
+ /**
+ * Consider this optimization on some heterogenous platforms and matrix sizes.
+ * Often, TRSM kernel on GPU yields significantly less performance rate than GEMM,
+ * while performances are similar on CPU. On this algorithm it is therefore
+ * recommended to execute all TRSMs (~low amount) on CPU to increase GPU efficiency.
+ */
//RUNTIME_zlocality_onerestrict( MORSE_TRSM, STARPU_CPU );
START_TIMING();
--
GitLab