You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
While working on #3708 I encountered several places where changing the indices of the 2D block load resulted in incorrect results that were only caught by the benchmarking infrastructure - all the core tests, language tests, and lit tests passed. We should improve local test coverage so that these issues can be found and corrected on smaller problem sizes and without taking up benchmarking infrastructure. The plan for this issue is to write ttgir which loads using a 2D block load / DPAS layout and stores using a block layout. This will allow us to check just the loads without actually running DPAS/dot operations. By using a block store instead of a DPAS layout store we can also check that any issues in the 2D block load are not hidden by a 2D block store.
The text was updated successfully, but these errors were encountered:
While working on #3708 I encountered several places where changing the indices of the 2D block load resulted in incorrect results that were only caught by the benchmarking infrastructure - all the core tests, language tests, and lit tests passed. We should improve local test coverage so that these issues can be found and corrected on smaller problem sizes and without taking up benchmarking infrastructure. The plan for this issue is to write ttgir which loads using a 2D block load / DPAS layout and stores using a block layout. This will allow us to check just the loads without actually running DPAS/dot operations. By using a block store instead of a DPAS layout store we can also check that any issues in the 2D block load are not hidden by a 2D block store.
The text was updated successfully, but these errors were encountered: