drivers/gpu/drm/i915/gem/i915_gem_wait.c · fcd1b2b9c7b085e9c200f73c079b322eb8c666f9 · Kirill Smelkov / linux

We employ a "waitboost" heuristic to detect when userspace is stalled waiting for results from earlier execution. Under latency sensitive work mixed between the gpu/cpu, the GPU is typically under-utilised and so RPS sees that low utilisation as a reason to downclock the frequency, causing longer stalls and lower throughput. The user left waiting for the results is not impressed. On applying commit ("dma-buf & drm/amdgpu: remove dma_resv workaround") it was observed that deinterlacing h264 on Haswell performance dropped by 2-5x. The reason being that the natural workload was not intense enough to trigger RPS (using HW evaluation intervals) to upclock, and so it was depending on waitboosting for the throughput. Commit ("dma-buf & drm/amdgpu: remove dma_resv workaround") changes the composition of dma-resv from keeping a single write fence + multiple read fences, to a single array of multiple write and read fences (a maximum of one pair of write/read fences per context). The iteration order was also changed implicitly from all-read fences then the single write fence, to a mix of write fences followed by read fences. It is that ordering change that belied the fragility of waitboosting. Currently, a waitboost is inspected at the point of waiting on an outstanding fence. If the GPU is backlogged such that we haven't yet stated the request we need to wait on, we force the GPU to upclock until the completion of that request. By changing the order in which we waited upon requests, we ended up waiting on those requests in sequence and as such we saw that each request was already started and so not a suitable candidate for waitboosting. Instead of asking whether to boost each fence in turn, we can look at whether boosting is required for the dma-resv ensemble prior to waiting on any fence, making the heuristic more robust to the order in which fences are stored in the dma-resv. Reported-by: