Mirror intel/llvm commits #2784
Conversation
Current "submission time" calculation is inaccurate because we don't use both synchronized timestamps returned by zeDeviceGetGlobalTimestamps but using only device timestamp from that call and use std::chrono "close" to that call to record the host time. This estimation becomes inaccurate pretty quickly. This PR fixes this problem using the known fact that L0 runtime implementation uses CLOCK_MONOTONIC_RAW on Linux and QueryPerformanceCounter on Windows. So, with this fix, at the first call, we use both device and host timestamps from zeDeviceGetGlobalTimestamps, subsequent calls (when device timestamp is not requested) will return corresponding host timestamp only, without making the `zeDeviceGetGlobalTimestamps` call which has high latency. Even though this approach improves accuracy and submit time doesn't become "invalid" (submit time > start_time) fast, it still doesn't guarantee that it will not happen. So, there will be additional fix done in intel/llvm#18717 to fix that. Test is also updated there to check larger number of iterations. Also apply the same fix as intel/llvm@138bef7 for cuda adapter (i.e. record host time after cuEventSynchronize for more precise measurement)
github-actions
bot
added
ci/cd
Unified Runtime -> intel/llvm Repo Move NoticeInformationThe source code of Unified Runtime has been moved to intel/llvm under the unified-runtime top-level directory, The code will be mirrored to oneapi-src/unified-runtime and the specification will continue to be hosted at oneapi-src.github.io/unified-runtime. The contribution guide will be updated with new instructions for contributing to Unified Runtime. PR MigrationAll open PRs including this one will be marked with the Should you wish to continue with your PR you will need to migrate it to intel/llvm. If your PR should remain open and not be closed automatically, you can remove the This is an automated comment. |
Automated changes by create-pull-request GitHub action