tuning-hipblaslt

Capture the exact problem from your workload with HIPBLASLT_LOG_MASK=32, then run hipblaslt-bench with --algo_method all and --print_kernel_info to race every solution. Replay the winner by its solution index and deploy it through aiter, TunableOp, or the hipBLASLt extension API.

Run hipblaslt-bench with --print_kernel_info, which prints a --Solution index: line for the winner. The bracket number prefixing each result row is only an enumeration position that drifts between runs and cannot be used with --algo_method index.

Racing many solutions back-to-back keeps clocks boosted and caches warm, inflating results by up to 20% on gfx950. Replay the winning solution standalone by index and A/B it against the heuristic default to get the honest number.

No, MX data types reject --algo_method all entirely because MX GEMMs route through RocRoller rather than Tensile solutions. Use --algo_method heuristic --requested_solution -1 to enumerate candidates, but note no solution index is printed, so alternatives cannot be deployed.

Measured gains over the default heuristic range from -2% to +14% on gfx950, far below the 19-40% the race output claims. The gfx950 default uses Origami analytical selection, so always interleave-replay both candidates before adopting a winner.

Repeated iterations reuse the same A and B buffers, which stay resident in cache and inflate results by up to 43% on gfx942 for cache-resident shapes. Pass --rotating with a working set larger than last-level cache to measure realistic memory traffic.