lds-bank-conflict

Measure LDS bank conflicts on AMD GPU kernels by running the target kernel under rocprofv3 hardware counters, parsing the output to generate a per-kernel summary that reports SQ_LDS_BANK_CONFLICT and SQ_LDS_DATA_FIFO_FULL metrics.

Yes, you can profile shared memory bank conflicts on MI300X, MI308X, and MI350 GPUs using rocprofv3 hardware counters, as this profiling workflow supports CDNA3 and CDNA4 architectures for per-dispatch and per-CTA analysis.

The best way to check LDS bank conflict severity per kernel dispatch is to execute rocprofv3 with a YAML counter configuration, which provides an automatic severity interpretation to highlight memory access bottlenecks in the generated summary.

Filter LDS bank conflict profiling results for specific kernels by supplying an optional kernel regex pattern when executing the profiling command, restricting the per-kernel summary and severity assessment to matching kernel dispatches.

Yes, you need a lds_bank_conf_counter.yaml file to define the hardware counters before executing rocprofv3, ensuring the profiler correctly captures the SQ_LDS_BANK_CONFLICT and SQ_LDS_DATA_FIFO_FULL metrics for analysis.