On Tue, Nov 10, 2026 at 2:34 AM Americo Simoes <[email protected]> wrote: > > > > > Hello Karol, > > I have been working on a userspace zero-copy path for i915 and I have a > measured result that may be relevant to the Rusticl roadmap. I am writing to > ask whether cl_khr_external_memory_dma_buf is planned for Rusticl, and > whether the result below is useful to that work. >
I just merged support for that recently and should be supported with the iris driver in the next mesa release. > The context > > I built a small C library (chipmem + blit) that allocates an i915 GEM object, > exports it as a DMA-BUF, and submits an XY_SRC_COPY_BLT to the BLT ring with > softpin. The GPU reads and writes the same physical pages the CPU wrote. On > Skylake-U (HD Graphics 520, kernel 7.2.13) the zero-copy path passes with 0 > mismatches out of 1024. > > The Rusticl backend in the same library is the fallback for minterms the BLT > ROP cannot express, and it copies at the boundary because Rusticl does not > support cl_khr_external_memory_dma_buf. The GNOME GEGL issue on this topic > says the same thing: Rusticl "still does not support the extension" , and the > Mesa issue #12177 tracks the request . > > The measurement > > The negative control in my test removes the DMA_BUF_IOCTL_SYNC bracket and > runs the same copy 100 times per case. Across batches, skipping both syncs > produced 42–99% stale reads; skipping the output sync only produced 87–100%. > The hazard is real and timing-dependent on this hardware. > > That measurement is the reason the extension matters for my use case: the BLT > path proves the hardware can do zero-copy work, and Rusticl is the portable > backend that could do the same on hardware where the BLT command does not > exist, if the extension were implemented. > > The question > > Is cl_khr_external_memory_dma_buf on the Rusticl roadmap, and is there a > branch or MR where the work is happening? I am not asking for a feature on my > behalf. I am asking whether the measured result is useful to the people doing > that work, and whether there is a place to record it. > > The implementation and the audit are at github.com/SimoesCTT/the-level. The > audit records the paths that failed (udmabuf, SET_DOMAIN, GEM_WAIT, > relocations on Gen9) as well as the one that worked. > > Thank you for your time. > > Americo Simoes > > Tel: +65 87635603 > > email: [email protected] > >
