search

CHIP-SPV / chipStar

chipStar is a tool for compiling and running HIP/CUDA on SPIR-V via OpenCL or Level Zero APIs.
Other
166 stars 27 forks source link

Add SKIP_TESTS_WITH_DOUBLES Option #826

Closed pvelesko closed 1 month ago

pjaaskel commented 3 months ago

This would be better as a check.py option as it's possible to run the same build on different devices, some of them supporting FP64, some not.

pvelesko commented 2 months ago

@pjaaskel I couldn't find another way to do this. Tried using CTEST_PRE_TEST but it didn't do anything.

Since check.py uses ctest as well, there's no simple way of doing this. In either case, just toggling the option on or off and reconfiguring cmake is enough here - you don't have to actually recompile the tests.

pvelesko commented 1 month ago

Merging this now now, we can implement a method that doesn't require CMake reconfigure later.

pjaaskel commented 1 month ago

I meant that we'd restrict the list of XPASS tests (which is currently restricted only by those cpu/gpu cl/lz toggles) by checking from clinfo if the target supports doubles, and/or have a flag in check.py that drops the tests that are known to fail with doubles.

pvelesko commented 1 month ago

XPASS tests (which is currently restricted only by those cpu/gpu cl/lz toggles

I'm sorry what is XPASS? And what toggles are you talking about?

In any case, how would we know if the test is using doubles without inspecting the SPIR-V? Inspecting the source of all the tests is not reasonable as there are thousands of tests and relying on a predefined list won't work as soon as new tests are added.

pjaaskel commented 1 month ago

Sorry for being unclear - not enough coffee consumed yet. Toggles in check.py which limit the list of tests expected to pass. You could inspect SPIR-V as you do know and dump the test cases that are expected to fail? But yep - this is not important.

pvelesko commented 1 month ago

Toggles in check.py which limit the list of tests expected to pass.

I'll add this approach to my backlog. Thank you for the suggestion.