AMD GPU Driver Sees DC Idle Manager & Multi-SDMA Engine Optimization For Linux 7.1

Source Domain: www.phoronix.com

With Linux 7.0-rc6 having released on Sunday, we are hitting the point of the cut-off of new feature material being allowed into the Direct Rendering Manager’s DRM-Next tree of queuing new graphics/display/accelerator feature code ahead of the upcoming Linux 7.1 merge window. As presumably the last AMDGPU/AMDKFD feature pull ahead of Linux 7.1, today’s pull request from AMD contains some noteworthy final enhancements.

In recent weeks prior AMDGPU/AMDKFD pulls of new feature code for Linux 7.1 brought new hardware IP enablement in preparing for future products, color management work, and other enticing improvements for this open-source AMD GPU Linux graphics driver stack.

Notable in today’s latest pull request is adding support for using multiple engines for handling buffer fills and clears. AMD Linux engineer Pierre-Eric Pelloux-Prayer has been working on adapting the TTM memory management code to support multiple contexts for pipelined moves. Ultimately it’s allowing for multiple SDMA engines to be used for buffer fills and clears. Rather than limiting a given clear or move to one SDMA copy engine, all of the available SDMA engines on the GPU are exposed. In turn this can yield faster buffer clears and faster memory migrations in vRAM-heavy scenarios. The patches make use of a round-robin scheduling policy for using all available SDMA engines.

Another notable addition with today’s pull is introducing the DC idle state manager for the Display Core. Due to some idle/power optimization strategies like Panel Self Refresh and Panel Replay leading to increased power use if the time-in-idle period is very low, the Idle State Management “ISM” for AMDGPU DC will use a hysteresis to determine if a delay should be inserted to better handle such idle state transitions.

The intent with the DC Idle State Manager is to avoid the negative power impact for some systems when there is rapid/frequent transitions around static screen optimizations as well as possible user lag…

Source