Hi
Where could I find more information about the GPU performance counters? (https://github.com/google/hardware-perfcounter/blob/main/include/hpc/gpu...)
Right now I need to guess their meaning from the name... but that's very limited information :-)
For example I assume SP_BUSY_CYCLES counts the cycles the SP (stream processor) is busy... however this seems to be a much higher number then the clock speed. Is it because there are multiple stream processors and the number is the sum from all of them?
How does SP_BUSY_CYCLES and SP_ALU_WORKING_CYCLES relate? Does all SP_ALU_WORKING_CYCLES count as busy cycles, so to get the load, can I use the SP_BUSY_CYCLES / SP_BUSY_CYCLES ratio?
How to compute the stall %? Is it (SP_STALL_CYCLES_VPC + SP_STALL_CYCLES_TP + SP_STALL_CYCLES_UCHE + SP_STALL_CYCLES_RB) / SP_BUSY_CYCLES?
... etc ...
Documentation of GPU performance counters
Posted: Sun, 2023-10-29 09:33