gpu: reduction: fix failing gtests #2453
Conversation
| #if DNNL_GPU_VENDOR == DNNL_VENDOR_INTEL | ||
| SKIP_IF(get_test_engine().get_kind() != engine::kind::cpu, | ||
| "Engine does not support this primitive."); | ||
| #endif |
There was a problem hiding this comment.
The Intel GPU backend should have a bunch of implementation to pick from, for both OCL and SYCL runtimes.
So even though your previous PR made those bugs visible, I don't think rehiding these issues is the proper fix here.
There was a problem hiding this comment.
My bad, I assumed those tests had been disabled on purpose and I re-added the check. From a quick look, both intel::ocl::atomic_reduction_t & intel::ocl::combined_reduction_t implementations (which are the top 2 in the implementation list) can't handle 0-dimensional inputs correctly. intel::ocl::ref_reduction_t seems to work correctly.
There was a problem hiding this comment.
For the record, I've created #2469 which fixes these failures by handling the zero-dim cases "correctly".
Description
#2243 caused some failures in the reduction gtests on intel GPU. This PR reverts a skip which was incorrectly removed which skips the tests on non-cpu engines. The
ifdeffor INTEL vendor is added as the tests should be enabled on generic backend.Checklist
General
make testandmake test_benchdnn_*) pass locally for each commit?