Skip to content

perf: generate padded hit responses on GPU

徐闯 requested to merge perf/padded-genr into J25.7.3

Replace host expansion and round trips in Probe.genR with a direct padded HIP lookup while preserving the historical float32 interpolation order. MI100 job 5336110 produced byte-identical 200-step chains and reduced synchronized ReconBlock time from 41.094 s to 37.056 s.

Merge request reports