Fix condition to ignore failed LLVM GPU tests #3819
Conversation
The tests should always be ignored if `ptxas` or the ROCm dependency is not present independently of whether the corresponding target is in the build targets. At worst this will exclude too many tests but it seems the condition to run those tests is independent of whether this specific target is build.
|
I'm fine with just always adding the ignore patterns, if we don't find We need to change the |
|
Test report by @Flamefire Overview of tested easyconfigs (in order)
Build succeeded for 1 out of 1 (1 easyconfigs in total) |
|
Test report by @Flamefire Overview of tested easyconfigs (in order)
Build succeeded for 0 out of 1 (1 easyconfigs in total) |
|
Test report by @Flamefire Overview of tested easyconfigs (in order)
Build succeeded for 0 out of 1 (1 easyconfigs in total) |
|
Test report by @Flamefire Overview of tested easyconfigs (in order)
Build succeeded for 1 out of 1 (1 easyconfigs in total) |
|
Test report by @Flamefire Overview of tested easyconfigs (in order)
Build succeeded for 1 out of 1 (1 easyconfigs in total) |
|
Test report by @Flamefire Overview of tested easyconfigs (in order)
Build succeeded for 1 out of 1 (1 easyconfigs in total) |
|
Test report by @Flamefire Overview of tested easyconfigs (in order)
Build succeeded for 1 out of 1 (1 easyconfigs in total) |
|
Test report by @Flamefire Overview of tested easyconfigs (in order)
Build succeeded for 1 out of 1 (1 easyconfigs in total) |
The tests should always be ignored if
ptxasor the ROCm dependency is not present independently of whether the corresponding target is in the build targets.At worst this will exclude too many tests but it seems the condition to run those tests is independent of whether this specific target is build.
For 18.x we have
LIBOMPTARGET_BUILD_CUDA_PLUGINdefault to true and enabling the tests based on presence of GPUs (determined by runningnvptx-arch)For 19+ it seems to be similar but based on
LIBOMPTARGET_PLUGINS_TO_BUILDwhich wasn't set for 18 until #3755 and hence defaulted to "all", enabling CUDA and hence the tests@Thyre @Crivella Should we, instead of removing
self.nvptx_target_condrather check fornot self.nvptx_target_cond or ...?