17:35 <jdavidberger> I have a G52 2ee that I benchmark out at ~24gflop on a compute shader (kernel 6.1.8-2-MANJARO-ARM and latest commit on main for mesa). I'm expecting ~38. It really seems like its only using 1st gen settings. The GPU itself does report prod_id 0x7402 / max_thread 768 as expected. I was wondering if anyone could give me pointers on looking into why. As far as my understanding goes, this is likely a kernel driver thing not a mesa iss

17:35 <jdavidberger> oper but figured someone here might have ideas

21:38 <cphealy> jdavidberger: You mention you are expecting ~38, is this because you experienced ~38 with Mali DDK or is ~38 based on specification numbers?

22:25 <jdavidberger> cphealy: This is based on the datasheet for the chip and specs from rockchip. I tried but failed to find a mali userspace binary that works with the G52 2EE. I can find the PDF if you want but it basically outlined 48 ops / cycle and i confirmed its running at 800mhz

22:28 <jdavidberger> It was this one: https://www.arm.com/-/media/Arm%20Developer%20Community/PDF/Mali%20GPU%20datasheet/Arm%20Mali%20GPU%20Datasheet%202021.2.pdf. The benchmark is https://github.com/krrishnarraj/clpeak/blob/master/src/kernels/compute_sp_kernels.cl but converted to opengl so it shouldnt be memory limited

22:54 <cphealy> jdavidberger: Real-world can definitely be different (slower) from the datasheet numbers. Few things to check: Make sure you are not hitting any DRAM bandwidth bottlenecks. Make sure both CPU and GPU are locked to max freq as both CPU and GPU DVFS governors can result in the cores not being at expected freq. Make sure there are no other processes running that are taking significant CPU cycles. If the CPU

22:54 <cphealy> is busy, it can easily result in lower GPU benchmark numbers.

22:58 <cphealy> jdavidberger: regarding the benchmark that you converted to opengl, can you share the source to this?

23:28 <jdavidberger> cphealy: I can; refactoring code now so it's concise but i also have a log dump with the output of BIFROST_MESA_DEBUG=shaders and the glsl. It tests single f32/vec2/vec3 etc so its a verbose log but the line i linked to is the fp32 results and you can scroll up to see the various outputs.

23:29 <cphealy> jdavidberger: Can you share which SoC this is?

23:29 <jdavidberger> I did put the GPU on performance 'governor' and it didn't make a large difference. Its this module https://wiki.pine64.org/wiki/SOQuartz

23:30 <jdavidberger> Log link is https://gist.github.com/jdavidberger/e24b9724efd946a62e9e12aaa8ebc00e#file-glminipeak-g52-log-L900

23:55 <cphealy> Ahh, the RK3566, this has quad core Cortex-A55 (little cores) only. You may want to look at what the CPU load is during the benchmark too just to make sure you are not hitting any CPU constraints.