01:38 <MilkMann> hey everyone, I'm working on optimizing arm-based systems running linux for real-time processing tasks, specifically targeting a heterogeneous architecture with arm cortex-a72 cores and a dedicated mali-g71 gpu. the goal is to implement efficient inter-process communication (ipc) between the cpu and gpu for machine learning inference, leveraging

01:38 <MilkMann> arm's big.little architecture. the bottleneck seems to be in the cpu-gpu synchronization, especially when offloading computational tasks to the gpu using opencl, as i’m experiencing significant overhead from memory transfers between the cpu and gpu.

01:38 <MilkMann> i’ve tried fine-tuning the `cgroups` settings for cpu affinity and the `dma-buf` framework for zero-copy memory management, but the system still exhibits latency spikes when switching between cpu and gpu-heavy workloads. i'm also considering using `streaming simd extensions` (sse) to optimize the cpu’s performance, but i’m concerned that it

01:38 <MilkMann> could interfere with the gpu memory bandwidth, especially given the limited number of memory channels in the arm system.

01:38 <MilkMann> additionally, I’m exploring the possibility of utilizing arm's hardware virtualization extensions (like kvm) for isolating gpu tasks in a separate vm for better resource management. however, I'm encountering kernel panics when attempting to pass-through the gpu to a guest os using vfio, possibly due to mismatched irq routing or pcie device

01:38 <MilkMann> assignment issues.

01:38 <MilkMann> does anyone have experience optimizing cpu-gpu communication in arm linux environments? or any advice on handling real-time memory transfers and managing virtualized gpus in this architecture? any tips for reducing latency, improving resource contention, or better utilizing arm’s architecture for these use cases would be greatly appreciated

07:52 <linusw__> MilkMann: that's a tall order, but start by talking to the DRM developers. They can be found in #dri-devel on the OFTC network.

12:46 <socksins1> that guy sure is busy, having both realtime arm linux and netbsd systems to optimize :D (he asked a similar yet distinctly different qestion over there)

13:06 <phh> it's weird to jump on virtualisation, cgroups or anything, before realizing that "memory transfers" are noop

13:08 <phh> (if anyone knows /any/ product with different memory bus for Cortex-A cpu vs Mali gpu I'm curious)