From 15d3d3c7e1506bc0332ab87aced86a8690372cc6 Mon Sep 17 00:00:00 2001
From: Mathieu Faverge <mathieu.faverge@inria.fr>
Date: Tue, 13 Oct 2020 18:52:44 +0200
Subject: [PATCH] Add the forcegpu option to enforce possible kernels on the
GPU
---
include/chameleon/runtime_struct.h | 6 ++++++
testing/chameleon_ztesting.c | 15 ++++++++++++++-
2 files changed, 20 insertions(+), 1 deletion(-)
diff --git a/include/chameleon/runtime_struct.h b/include/chameleon/runtime_struct.h
index 96a8c5cb7..fa3fc31f5 100644
--- a/include/chameleon/runtime_struct.h
+++ b/include/chameleon/runtime_struct.h
@@ -33,6 +33,12 @@ typedef enum runtime_id_e {
RUNTIME_SCHED_OPENMP, /**< OpenMP runtime */
} RUNTIME_id_t;
+/**
+ * @brief Ids of the worker type
+ */
+#define RUNTIME_CPU ((1ULL)<<1)
+#define RUNTIME_CUDA ((1ULL)<<3)
+
/**
* @brief RUNTIME request structure
*
diff --git a/testing/chameleon_ztesting.c b/testing/chameleon_ztesting.c
index e535b6ff7..2ae0e946d 100644
--- a/testing/chameleon_ztesting.c
+++ b/testing/chameleon_ztesting.c
@@ -42,6 +42,7 @@ static parameter_t parameters[] = {
{ "nowarmup", "Disable the warmup run to load libraries", -31, PARAM_OPTION, 0, 0, TestValInt, {0}, NULL, pread_int, sprint_int },
{ "mtxfmt", "Change the way the matrix is stored (0: global, 1: tiles, 2: OOC)", -32, PARAM_OPTION | PARAM_INPUT | PARAM_OUTPUT, 1, 6, TestValInt, {0}, NULL, pread_int, sprint_int },
{ "profile", "Display the kernel profiling", -33, PARAM_OPTION, 0, 0, TestValInt, {0}, NULL, pread_int, sprint_int },
+ { "forcegpu", "Force kernels on GPU", -34, PARAM_OPTION, 0, 0, TestValInt, {0}, NULL, pread_int, sprint_int },
{ NULL, "Machine parameters", 0, PARAM_OPTION, 0, 0, 0, {0}, NULL, NULL, NULL },
{ "threads", "Number of CPU workers per node", 't', PARAM_OPTION | PARAM_OUTPUT, 1, 7, TestValInt, {-1}, NULL, pread_int, sprint_int },
@@ -490,7 +491,7 @@ parameters_destroy()
int main (int argc, char **argv) {
int ncores, ngpus, human, check, i, niter;
- int trace, nowarmup, profile;
+ int trace, nowarmup, profile, forcegpu;
int rc, info = 0;
int run_id = 0;
char *func_name;
@@ -515,6 +516,7 @@ int main (int argc, char **argv) {
trace = parameters_getvalue_int( "trace" );
nowarmup = parameters_getvalue_int( "nowarmup" );
profile = parameters_getvalue_int( "profile" );
+ forcegpu = parameters_getvalue_int( "forcegpu" );
CHAMELEON_Init( ncores, ngpus );
@@ -536,6 +538,17 @@ int main (int argc, char **argv) {
run_print_header( test, check, human );
run = runlist->head;
+ /* Force all possible kernels on GPU */
+ if ( forcegpu ) {
+ if ( ngpus == 0 ) {
+ fprintf( stderr,
+ "--forcegpu can't be enable without GPU (-g 0).\n"
+ " Please specify a larger number of GPU or disable this option\n" );
+ return EXIT_FAILURE;
+ }
+ RUNTIME_zlocality_allrestrict( RUNTIME_CUDA );
+ }
+
/* Warmup */
if ( !nowarmup ) {
run_arg_list_t copy = run_arg_list_copy( &(run->args) );
--
GitLab