[code-reflection] RFR: Hat mark buffers 1
Gary Frost
gfrost at openjdk.org
Fri Feb 14 15:37:35 UTC 2025
On Fri, 14 Feb 2025 15:33:15 GMT, Gary Frost <gfrost at openjdk.org> wrote:
> Ongoing work for moving towards buffer marking to reduce unnecessary copies.
>
> Here mostly infrastructural
>
> We moved buffer status data from the ArgSled to a trailing BufferState section at the end of each Buffer.
> The native buffers now are kept in this buffer state section along with marker bits (host dirty, gpu dirty)
> OpenCL + Mock + Spirv native backends now have event queues separated from the Kernel (in prep for the buffer to outlive the kernel dispatch)
> Changes tested on OpenCL (and Mock).
> CUDA Backends very likely broken. I will come back to those.
> HIP also. I need to work with AMD to sync the OpenCL changes to HIP backend.
>
> Buffers are still copied in and out on each dispatch.
/Integrate
-------------
PR Comment: https://git.openjdk.org/babylon/pull/318#issuecomment-2659641975
More information about the babylon-dev
mailing list