[Frontend] DMA-descriptor rework + masked-DMA for non-dividing shapes (#282)#285

Merged

YWHyuk merged 13 commits into

feature/codegen-ordered-stepsfrom

feature/indirect-offset-descriptor

Jul 7, 2026

Commits on Jul 7, 2026

[Frontend] Carry gather/scatter offset as an explicit transfer descriptor
YWHyuk
committed
[Frontend] Clean up indirect access: symbol-set detection, CONFIG4 stride, compute-loop multi-dim offset
YWHyuk
committed
[Frontend] Drop memref.dma_start: lower togsim.transfer directly to Gemmini
YWHyuk
committed
[Frontend] Port the trace path (build_tog) to togsim.transfer/togsim.wait
YWHyuk
committed
[Frontend] Detect the indirect offset operand by the `indirect` attr, not operand count
YWHyuk
committed
[Frontend] Fix stale build_skeleton docstrings (memref.dma_wait -> togsim.wait)
YWHyuk
committed
[Frontend] Lower togsim.transfer to a DMA descriptor + CONFIG_DESC (no CONFIG/2/3/4)
YWHyuk
committed
[Build] Bump spike to v1.0.4 (DMA descriptor / CONFIG_DESC)
YWHyuk
committed
[Build] Bump spike to v1.0.5 (masked-DMA clamp + MVOUT accumulate)
YWHyuk
committed
[Frontend] Masked-DMA dynamic clamp: drop tile divisibility, fill non-dividing tails
YWHyuk
committed
[Frontend] Extend the masked-DMA clamp to conv (explicit extents + fine-grained)
YWHyuk
committed
[Frontend] Clamp the fused-epilogue output store on non-dividing shapes
YWHyuk
committed
[Test] Golden unit test for the lower_transfer_to_gemmini pass
YWHyuk
committed