git
/
gcc
mirror of git://gcc.gnu.org/git/gcc.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

nvptx plugin: do not force JIT target SM version 

When link_ptx runs, a CUDA device is already bound to current thread, so the
driver library knows the target architecture.  There isn't any benefit from
forcing a specific target here; on the contrary, hardcoding sm_30 breaks
offloading on later (Maxwell, sm_5x) devices.
    
	* plugin/plugin-nvptx.c (link_ptx): Do not set CU_JIT_TARGET.

From-SVN: r232227
This commit is contained in:
Alexander Monakov 2016-01-11 15:55:31 +03:00 committed by Alexander Monakov
parent b34b1c3436
commit 0d58938ed7
2 changed files with 7 additions and 6 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
libgomp/ChangeLog
View File

@ -1,3 +1,7 @@
2016-01-11 Alexander Monakov <amonakov@ispras.ru>
* plugin/plugin-nvptx.c (link_ptx): Do not set CU_JIT_TARGET.
2016-01-07 H.J. Lu <hongjiu.lu@intel.com>
PR fortran/66680

9
libgomp/plugin/plugin-nvptx.c
View File

@ -712,8 +712,8 @@ static void
link_ptx (CUmodule *module, const struct targ_ptx_obj *ptx_objs,
unsigned num_objs)
{
CUjit_option opts[7];
void *optvals[7];
CUjit_option opts[6];
void *optvals[6];
float elapsed = 0.0;
#define LOGSIZE 8192
char elog[LOGSIZE];
@ -742,10 +742,7 @@ link_ptx (CUmodule *module, const struct targ_ptx_obj *ptx_objs,
opts[5] = CU_JIT_LOG_VERBOSE;
optvals[5] = (void *) 1;
opts[6] = CU_JIT_TARGET;
optvals[6] = (void *) CU_TARGET_COMPUTE_30;
r = cuLinkCreate (7, opts, optvals, &linkstate);
r = cuLinkCreate (6, opts, optvals, &linkstate);
if (r != CUDA_SUCCESS)
GOMP_PLUGIN_fatal ("cuLinkCreate error: %s", cuda_error (r));