projects / gcc.git / commit
? search:
summary | shortlog | log | commit | commitdiff | tree
(parent: b34b1c3) | patch
nvptx plugin: do not force JIT target SM version
authorAlexander Monakov <amonakov@ispras.ru>
Mon, 11 Jan 2016 12:55:31 +0000 (15:55 +0300)
committerAlexander Monakov <amonakov@gcc.gnu.org>
Mon, 11 Jan 2016 12:55:31 +0000 (15:55 +0300)
commit0d58938ed798882063737dedfe897a9b484295ad
tree9e5efed6bfd5c03b33fb3b973ab51fdeae168032tree
parentb34b1c343600f7587213dafee5e305a8e4000132commit | diff
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
libgomp/ChangeLog diff | blob | history
libgomp/plugin/plugin-nvptx.c diff | blob | history