RFC Class/instance based execbuf plus more
Get series with b4
b4 am [email protected]Versions newest first
- v1 RFC Latest5 patches 7 repliesIn review
Patches in v1 No applies recorded yet
| # | Subject and files | Size | Replies | Applied |
|---|---|---|---|---|
| 1/5 | drm/i915: Select engines via class and instance in execbuffer2drivers/ include/ | +24 | 0 replies | |
| 2/5 | drm/i915: Engine capabilities uAPIdrivers/ drivers/ drivers/ +1 more | +36 | 2 replies | |
| 3/5 | drm/i915: Concurrent context uAPIdrivers/ drivers/ include/ | +31 | 3 replies | |
| 4/5 | drm/i915: Re-arrange execbuf so context is known before enginedrivers/ | +20 | 0 replies | |
| 5/5 | drm/i915: Per batch buffer VCS balancingdrivers/ drivers/ drivers/ +3 more | +78 | 2 replies |
Reviews trailers per patch
No Reviewed-by, Acked-by or Tested-by trailers for v1 yet.
Cover letter [RFC 0/5] Class/instance based execbuf plus more
From: Tvrtko Ursulin <[email protected]>
Now that the engine class concept is in, it is time to re-send the old proposal of using it for engine selection in execbuf.
Idea is primarily to fix the situation with the current VCS engine selection ABI by introducing a new, cleaner, method of selecting the VCS engine.
Then there are two new pieces of uAPI proposal, engine capabilities and concurrent contexts, which for instance enable the VA-API driver to let the i915 balance it's batch buffers dynamically.
This enables better utilization of resources on GT3/GT4 parts where:
a) a single stream can now use both engines
b) it opens the door of extending the i915 scheduler with more advanced load balancing approaches to support the multiple-streams use cases better.
For instance decoding a single H.264 stream on a GT4 part is now improved from 57 seconds to 40 seconds, with minimal VA-API code base changes:
root@sc:~/ffmpeg# VA_INTEL_CONCURRENT=0 perf stat -a -e i915/vcs0-busy/,i915/vcs1-busy/ ffmpeg -loglevel panic -hwaccel vaapi -hwaccel_output_format vaapi -i ~/bbb_sunflower_1080p_60fps_normal.mp4 -f null -
Performance counter stats for 'system wide':
57,568,097,358 ns i915/vcs0-busy/
0 ns i915/vcs1-busy/
57.585753514 seconds time elapsedroot@sc:~/ffmpeg# VA_INTEL_CONCURRENT=1 perf stat -a -e i915/vcs0-busy/,i915/vcs1-busy/ ffmpeg -loglevel panic -hwaccel vaapi -hwaccel_output_format vaapi -i ~/bbb_sunflower_1080p_60fps_normal.mp4 -f null -
Performance counter stats for 'system wide':
29,152,427,164 ns i915/vcs0-busy/
29,115,272,714 ns i915/vcs1-busy/
40.733992298 seconds time elapsedI will be sending the proof-of-concept patches for intel-vaapi-driver separately.
Tvrtko Ursulin (5):
drm/i915: Select engines via class and instance in execbuffer2 drm/i915: Engine capabilities uAPI
drm/i915: Concurrent context uAPI
drm/i915: Re-arrange execbuf so context is known before engine drm/i915: Per batch buffer VCS balancing
drivers/gpu/drm/i915/i915_drv.h | 7 +- drivers/gpu/drm/i915/i915_gem.c | 2 +- drivers/gpu/drm/i915/i915_gem_context.c | 14 +++ drivers/gpu/drm/i915/i915_gem_context.h | 20 +++++ drivers/gpu/drm/i915/i915_gem_execbuffer.c | 134 ++++++++++++++++++++++------- drivers/gpu/drm/i915/intel_engine_cs.c | 3 + drivers/gpu/drm/i915/intel_ringbuffer.h | 2 + include/uapi/drm/i915_drm.h | 34 +++++++- 8 files changed, 180 insertions(+), 36 deletions(-)
Signature · 6 lines
2.14.1 _______________________________________________ Intel-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/intel-gfx