Hi all, Following the last set of IGT tests on the mailing list, Chris made an interesting comment about engine to engine synchronization.
Indeed with dma-fence-chain, i915 is unable to detect i915 fences given as wait dependencies to the execbuf, because those are now wrapped into a dma-fence-chain. Our submission code has an interesting optimization where it uses HW semaphores to create a dependency between 2 requests, thus doing away with the need for the host to submit a request once its dependencies have completed. This v6 tries to work around this performance degration due to dma-fence-chains by peeling them when adding fences waits to a request. It is still possible that a dma-fence-chain wraps another dma-fence-chain but I think those cases will be rare enough we don't want to care about them. Cheers, Lionel Landwerlin (3): drm/i915: introduce a mechanism to extend execbuf2 drm/i915: add syncobj timeline support drm/i915: peel dma-fence-chains wait fences .../gpu/drm/i915/gem/i915_gem_execbuffer.c | 379 +++++++++++++++--- drivers/gpu/drm/i915/i915_drv.c | 3 +- drivers/gpu/drm/i915/i915_getparam.c | 1 + include/uapi/drm/i915_drm.h | 64 ++- 4 files changed, 385 insertions(+), 62 deletions(-) -- 2.24.0 _______________________________________________ Intel-gfx mailing list Intel-gfx@lists.freedesktop.org https://lists.freedesktop.org/mailman/listinfo/intel-gfx
