We are running Yocto Linux on an IMX8M Plus
What we have noticed is that during any GPU activity, we get huge latency spikes for memory access from the CPU (1-3ms spikes). This is a huge issue, as we are using XDP for networking.
We have tried tuning IMX8MP_ICM_A53, IMX8MP_ICM_GPU3D and IMX8MP_ICM_GPU2D, setting them to 7, 2, 2 for QoS control. This made the latency spikes less frequent, but they are still there. The only option we found was doing absolutely no GPU activity or removing the GPU from the device tree, however this is not acceptable for our use-case. Are there some other setting that could be tuned to alleviate this behavior? We cannot have the GPU locking the DRAM for more than a 50-100 μs at most
