hipBLASLt
hipBLASLt copied to clipboard
ROCm
Windows enablement
@amd-mtrifuno Is this a feature targeting 6.5? How come we're targeting release staging prior to targeting develop and then cherry picking?
Hi folks - David and I discussed a strategy for getting Windows support merged because we have two patch sets that are somewhat disjoint, and they are both in the "make it work" variety and can't be landed like this.
Here are the branches we have:
- https://github.com/ROCm/hipBLASLt/pull/1864 (by an AMD employee)
- https://github.com/ROCm/TheRock/pull/441 (by an outside contributor)
We also have another effort underway which should produce results in the next week, where we are landing a much simplified/cleaned up CMake build.
I'm in the loop on all of these, and I have a pretty good understanding of all of the changes and reasons for making them. I'm also developing on Windows and am going to help the hipBLASLt team get these patches staged and landed in a more organized fashion over the next week. As such, we are going to pause further review of 1864 and 441, using them more as reference material to assemble changes to land. Since I'm working on a Windows development environment, I'll be able to verify these more effectively than the regular hipBLASLt team. I will be prioritizing/ordering by: C++ changes, Python changes, and then CMake changes. My hope is that by holding off on the CMake changes, we can just switch to the new build files for Windows and have a better time of it all around.
I'll be staging changes on the branch https://github.com/ROCm/hipBLASLt/tree/slaurenz_therock_win_stage which I will keep rebased to develop as we land things.
Thanks for the time finding the "bottom" of what it will take to support Windows -- these exercises of cutting through in one swath are really useful to get the whole scope known.
