-
Notifications
You must be signed in to change notification settings - Fork 130
[SYCL] test for https://github.com/intel/llvm/pull/4405 #424
Conversation
Could you please add more information about the bug this test should trigger? |
Signed-off-by: Sergey V Maslov <[email protected]>
It was crashing due to imbalance of event retain/releases (from within Level-Zero plugin) with this special test. I've updated PR description too. |
It might be helpful to add the description of this test as a comment inside the test. |
Thanks. Can we have a test which makes that event retain/release is balanced now in a more reliable way than just crash/not crash? I believe there were are checks in the Level Zero plugin that track reference counters. |
The crash is actually a reliable "die" in L0 plugin, when release is called on an object that has 0 ref-count. I wouldn't want to spend time on crafting a more general test here for the reported regression. |
Will the test crash if the number of "retains" is bigger than the number of "retains"(imbalance in the opposite direction)?
Ok. |
? "L0" | ||
: "OpenCL"; | ||
device dev = Q.get_device(); | ||
size_t max_compute_units = dev.get_info<info::device::max_compute_units>(); |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Minor.
Information about compute units seems to be irrelevant to the test case.
Not this one, but we run entire LIT testing with ZE_DEBUG=4, mode in which L0 resources leaks are detected, we also run tests with valgrind. This test is specifically for the regression observed (and not caught by any other existing testing). |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Approving. Please, note there was a minor comment from me.
* upstream/intel: (70 commits) [SYCL][Matrix] Add the 8 bit type variants (intel#443) [SYCL] testing of linking in the presence of multiple kernels (intel#440) [NFC] Add Sergey to CODEOWNERS of USM test (intel#455) [ESIMD] E2E test for writeable simd_view subscript operator (intel#415) [SYCL] Eliminate usages of _class aliases (intel#449) [SYCL] Enable assert AOT test (intel#450) [SYCL] Disable flaky Plugin/level_zero_dynamic_batch_test (intel#453) [SYCL] Ensure JIT failure on all Intel devices (intel#452) [SYCL] Removed the remained handler::codeplay_host_task (intel#451) [SYCL] Test for Group Mask feature (intel#441) [SYCL] Update kernel_bundle test for default contexts (intel#446) [SYCL] Replace using of ONEAPI/INTEL namespaces with ext::oneapi/intel (intel#442) [SYCL] Replace deprecated has_extension (intel#434) [SYCL] test for intel/llvm#4405 (intel#424) [SYCL] Remove function pointers extension tests (intel#437) [SYCL] Added test for deprecated features (intel#435) Fix check for availability of sycl.hpp (intel#445) [SYCL] Update tests for default context extension (intel#433) Use sycl.hpp workaround only if compiler doesn't have it (intel#439) [SYCL] Allow extra command line options implicitely added by compiler (intel#432) ...
Signed-off-by: Sergey V Maslov <[email protected]>
Verify balanced event retains/releases under special circumstances.
Signed-off-by: Sergey V Maslov [email protected]