topic Re: GPU profiling - shader utilisation at 200+% in i.MX Processors

GPU profiling - shader utilisation at 200+%

ATPark — Tue, 31 Aug 2021 16:02:39 GMT

I'm profiling a Qt app for an iMX6 IoT device to try to diagnose low frame rate.

I'm thinking we are probably limited by CPU not GPU, but I have one thing from vAnalyser that worries me and that I don't understand. The shader utilisation is very high.

Below is a screenshot of the frame analysis for a typical frame. I see that driver utilisation and GPU utilisation are both really low, which is fine. But the shader utilisation is at over 200%, and this is the same for ALL frames.

Can anyone explain in basic terms what it means for the shader utilisation to be above 200% consistently, and what effect that is likely to have on our application performance?

Thanks!

Re: GPU profiling - shader utilisation at 200+%

Bio_TICFSL — Wed, 01 Sep 2021 13:00:32 GMT

Hello ATPark,

That part of the advice is local to the shader core - load/store is the highest pipeline load, so that's the critical path for shader execution (ignoring external memory effects).

In addition to compute, load/store is used for reading and writing vertex data. Check your complexity (primitive count, vertex count, and number of bytes per vertex) and CPU-side culling to make sure you are not processing more geometry that the GPU can cope with.

In general for mobile aim for < 300-500K input vertices per frame, and ~32 bytes per vertex.

The driver is utilization 1% so not worry about the DRAM effects.

Regards

Re: GPU profiling - shader utilisation at 200+%

ATPark — Wed, 01 Sep 2021 17:22:20 GMT

Thanks for getting back to me on this!

Below is a typical frame with details of the vertex count and primitive count. Both are reasonably stable across frames, except for a few frames where the input primitives are '12' and the 'trivially reject count' is zero. Either way, the counts are much, much lower than 300-500k.

However, I don't know how to see the bytes per vertex; is it possible that that is very high for some reason?

With respect to 'CPU-side culling', do you know how I could check that? We aren't doing any graphics programming directly - we are drawing using Qt controls. Do you know anything we should check about our use of Qt?

Thanks,

Tony

Re: GPU profiling - shader utilisation at 200+%

Bio_TICFSL — Fri, 03 Sep 2021 14:05:33 GMT

Hi,

There is no way to see how many bytes per vertex is used, and for CPU culling you may be able to see this blog:

https://bruop.github.io/frustum_culling/

Regards