Integer vector comparison is working incorrectly on Intel based Macs with MacOS >= 11 #61
Comments
Version 0.6.4: Hot-Fixes - **Samples** apps: - Attempted replacement of `PerlinNoise` with `FastNoise2` library, but reverted it back because of no ARM support for M1 Macs (see [FastNoise2 issue](Auburn/FastNoise2#93)) - **Graphics** libraries: - Fixed runtime errors in GPU profiling builds with `METHANE_GPU_PROFILING_ENABLED=ON`: - DirectX GPU timestamps re-calibration issue leading to GPU ranges shifting was fixed. - Fixed command list execution waiting threads synchronisation in Profiling builds in Typography tutorial. - Fixed resources retaining in command lists. Retained resources were incorrectly cleared on `CommandList::Reset()`, while they should be cleared on `CommandList::Commit()`. - Fixed DirectX descriptor heaps allocations after `Context::Reset()` by always using deferred heap allocation in all cases. Deferred heap initialisation flag was removed, since it became unconditionally deferred. - Fixed sporadic hang in `CommandQueueTrackingBase::WaitForExecution()` - Fixed sporadic crash on destruction of `CommandQueueTrackingBase` with proper shutdown procedure called from destructor of derived class. - Fixed Vulkan build and some initialisation errors on MacOS. - **Tests**: - All unit tests were updated to support breaking changes in Catch v3. - `Point<T,size>` wrapper class was extended with workarounds of MacOS & ARM specific bugs in HLSL++ integer vector comparison and division operators (see [1](redorav/hlslpp#60) and [2](redorav/hlslpp#61)). - Logging of `Point`, `Rect` and `RectSize` values was added in `DataTypes` unit tests. - **External** libraries: - External dependencies management via Git submodules was replaced with [CPM.cmake](https://github.com/cpm-cmake/CPM.cmake) package manager. No submodules anymore - it greatly simplifies external library updates! - All externally dependent repositories are downloaded to CPM cache directory during CMake configuration stage, to `Build/Output/ExternalsCache` by default (it can be changed with CMake option `CPM_SOURCE_CACHE`). - When CMake project is configured under CLion, external repositories are downloaded to individual build directories of each configuration to workaround [parallel cache update collision issue of the CPM.cmake](cpm-cmake/CPM.cmake#293). - New `README.md` description of the external dependencies was added in `Externals` directory. - Almost all libraries were updated to latest version, except SPIRV-Cross and DirectX Shader Compiler pre-built binary tools. - **Build** infrastructure: - Root `CMakeLists.txt` was simplified by moving all compiler configuration options to `CMake/MethaneBuildOptions.cmake`. - Use caching of Externals CPM package sources in Azure Pipelines and GitHub Workflows to speedup builds. - MacOS builders were switched from v10.15 to v11.
|
Unfortunately I cannot seem to repro this on my Android phone (ARM 64 configuration) so I'm a bit at a loss as to what can be going wrong here. Also the equality operator is pretty simple. Can you give a bit more detail on this? |
|
This issue is reproducible on Intel x86_64 architecture (not ARM), and only on MacOS >= 11. As I mentioned, it was working fine on MacOS 10.15. It looks very strange because the same SSE-based integer vectors comparison works fine on Windows and Linux. I don't know why it happens, but it is proved with unit testing results in Azure Devops available by the links in previous message. |
|
Ah I see what you mean. I'll have to get hold of a machine like you describe somehow then, or maybe a virtual machine. |
|
@egorodet Does it still make sense to keep this open? |
Let me check this. Intel-based macOS machines are still available in GitHub hosted runners for GH Actions. |
Signed and unsigned integer vectors comparison operators (==, <, >) seem to be working incorrectly in some cases on Intel-based Macs. This is reproducible in CI system of my project which is testing wrapper class Point<T,size> around HLSL++ vector in Azure Pipelines running on Macs with x86_64 architecture. Please see build 20220716.1 (on MacOS 11) and build 20220716.2 (on MacOS 12) test results. I noticed these test failures after switching build agents to MacOS v11, while previously on MacOS 10.15 tests were passing successfully - see build 20220715.1. I don't have Intel based Mac under my hands right now, so I can't check it locally, but I believe that comparison operator tests added in PR #59 will help to reproduce this issue.
The text was updated successfully, but these errors were encountered: