Max191
commented
3 months ago After looking into the failing test I don't think we are ready for this flip yet. We need a better way of handling shared memory.
The test failure is cause by 2 tensor.empty() ops getting tiled to the same size and the CSEd into a single empty. However, the empty ops are sort of a hacky way to represent the shared memory buffers when running GPUPromoteMatmulOperandsPass. We need a better representation of the shared memory buffers at tensor level so we don't CSE them into a single buffer. The parallelism check in bufferization has been saving us by recreating a new buffer, but it is not a good way to handle this issue.
A couple of ideas would be to create alloc_tensor ops or implement some multi-buffering with both shared memory allocs being split across a single tensor.
CC @MaheshRavishankar @qedawkins @antiagainst
MaheshRavishankar
github-actions[bot]
When there is a buffer used inside of an
scf.forallop that is defined outside of thescf.forall, bufferization will unconditionally bufferize out of place by default in order to avoid race conditions. However, handling parallel accesses to a buffer should generally be the responsibility of the source program, and if there is a race condition, then it should be handled outside of bufferization. This PR disables the parallel region check in IREE to simplify the bufferization analysis and enable more buffer reuse.It is possible that this PR could cause race conditions if data races are not handled properly, and we are relying too much on bufferization to be conservative. Turning this option off could be a good early step in diagnosing data races on GPU.