Mali hw counters not matching Streamline's built-in Mali Bifrost G71 profile

Hi Lorenzo,

It depends what you want to measure.

"Fragment Cycles / GPU Active Cycles" gives you a percentage of the GPU workload which is running fragment processing, but the GPU may still be idle a lot of the time so still underutilized.

"Fragment Cycles / TARGET_GPU_FREQ", where TARGET_GPU_FREQ is a statically defined constant you'll need to choose yourself, will give you a feel for how much GPU headroom you have spare (or whether you are over-using). The trick to make this work in the Streamline charts is the $ZOOM built-in variable which contains the current GUI zoom level bin size in seconds, so if you have an 400MHz GPU target your expression would be one of:

$MaliJobManagerCyclesGPUCycles / (400000000 * $ZOOM) * 100

$MaliJobManagerCyclesVertexTilignComputeCycles / (400000000 * $ZOOM) * 100

$MaliJobManagerCyclesFragmentCycles / (400000000 * $ZOOM) * 100

Each of these returns a percentage where 100% would indicate you are using all of the available performance at the target frequency, less than 100% indicates you have spare performance, and more than 100% indicates that you are over-using vs your target.

Cheers, 
Pete

Hi Lorenzo,

It depends what you want to measure.

"Fragment Cycles / GPU Active Cycles" gives you a percentage of the GPU workload which is running fragment processing, but the GPU may still be idle a lot of the time so still underutilized.

"Fragment Cycles / TARGET_GPU_FREQ", where TARGET_GPU_FREQ is a statically defined constant you'll need to choose yourself, will give you a feel for how much GPU headroom you have spare (or whether you are over-using). The trick to make this work in the Streamline charts is the $ZOOM built-in variable which contains the current GUI zoom level bin size in seconds, so if you have an 400MHz GPU target your expression would be one of:

$MaliJobManagerCyclesGPUCycles / (400000000 * $ZOOM) * 100

$MaliJobManagerCyclesVertexTilignComputeCycles / (400000000 * $ZOOM) * 100

$MaliJobManagerCyclesFragmentCycles / (400000000 * $ZOOM) * 100

Each of these returns a percentage where 100% would indicate you are using all of the available performance at the target frequency, less than 100% indicates you have spare performance, and more than 100% indicates that you are over-using vs your target.

Cheers, 
Pete

That's a great way of doing it. I did not know the $ZOOM built-in variable, that will change my game.

On the device I have, I can find the max frequency of the GPU like this:

dreamlte:/ $ cat /sys/devices/platform/13900000.mali/asv_table
GPU, vol, min, max, down_stay, mif, int, cpu
546,  662500, 78,  99, 5, 1794000, 400000, 1690000
455,  650000, 78,  85, 9, 1540000, 400000, 1456000
385,  643750, 78,  85, 1, 1352000, 400000, 1248000
338,  637500, 78,  85, 1, 1014000, 267000,  949000
260,  637500, 78,  85, 1, 421000, 178000,       0

I guess the max frequency could be lowered by the system in certain situations like thermal throttoling, but that's probably as good as it gets.