Handling metrics with procentage (rather than count)
dejangrubisic opened this issue · comments
All metrics are summing up to 100% which allows the user to see where the metric count is the highest. Although some metrics like GPU_BUSY show the percentage of GPU usage for a given kernel invocation. This will need to have a different mechanism, since summing percentages across whole execution doesn't make much sense.
Using some kind of DIVISION_FORMULA(GPU_KINFO_STMEM); from gpu-metrics.c would make sense