Capture/express repros with alternate partitioning beyond the default fusion boundaries, to study cross-op scheduling:
mm + subsequent dependent kernel (epilogue/consumer of the mm output)mm → pointwise-chain → mm (two matmuls bridged by elementwise work)
Goal: a repro form that lets us benchmark different partition choices for the same dataflow (not just the single partition the capture hook currently emits), so we can measure what a smarter scheduler boundary would buy. Relates to the 'fix realization in scheduler, not graph shape' principle.
Capture/express repros with alternate partitioning beyond the default fusion boundaries, to study cross-op scheduling:
mm+ subsequent dependent kernel (epilogue/consumer of the mm output)mm → pointwise-chain → mm(two matmuls bridged by elementwise work)Goal: a repro form that lets us benchmark different partition choices for the same dataflow (not just the single partition the capture hook currently emits), so we can measure what a smarter scheduler boundary would buy. Relates to the 'fix realization in scheduler, not graph shape' principle.