llvm

e3e3172b - [UR][CUDA] Fix flaky urUSMContextMemcpyExp test with optimized D2D handling (#22090)

[UR][CUDA] Fix flaky urUSMContextMemcpyExp test with optimized D2D handling (#22090) The urUSMContextMemcpyExpTestDevice::Success test was flaky on CUDA due to cuMemcpy() executing asynchronously for device-to-device copies. This caused race conditions where data was read before the copy completed. Solution uses hybrid approach: - H2D/D2H: cuMemcpy() (synchronous, fast path) - D2D: dedicated stream + cuMemcpyAsync() + cuStreamSynchronize() This provides correctness with minimal overhead (~15-70μs for stream create/destroy) while avoiding the performance cliff of context-wide synchronization (cuCtxSynchronize blocks all GPU work). Also adds comprehensive test coverage (large allocations, concurrent copies, multi-threaded sequential, unaligned pointers) and re-enables L0_V2 tests - (#19604) no repro on this configuration. Fixes #19688