[code-reflection] Integrated: Hat ongoing buffer marking
Gary Frost
gfrost at openjdk.org
Mon Feb 17 19:22:39 UTC 2025
Lots of changes here. Specifically moving native backend data structures around.
We now cache native side buffer info (say clmem for OpenCL) in the vendor ptr of the memory segment 'state'
This has a substantial perf boost with OpenCL. Which was suprising (mandel x1.5, gol also x1.5)
Still defaulting to copying in and out for each kernel dispatch.
Close to turning on min buffer options.
-------------
Commit messages:
- clmem in vendorPtr.
- Add profiling data
- Gated fetch of dirty buffer from opencl ffi layer
- Removed Queue from all but OpenCL
- We'll use deviceDirty instead of gpuDirty
- FFI:getBuffer now getBufferFromDeviceIfDirty()
- Removed mode from BufferState. Mode now conveyed via Backend Config
- Removed mode from BufferState. Mode now conveyed via Backend Config
- We pass mode, platform and device selectors to backends.
- We pass mode, platform and device selectors to backends.
- ... and 6 more: https://git.openjdk.org/babylon/compare/14b3744a...e5e6ca56
Changes: https://git.openjdk.org/babylon/pull/321/files
Webrev: https://webrevs.openjdk.org/?repo=babylon&pr=321&range=00
Stats: 1037 lines in 30 files changed: 525 ins; 350 del; 162 mod
Patch: https://git.openjdk.org/babylon/pull/321.diff
Fetch: git fetch https://git.openjdk.org/babylon.git pull/321/head:pull/321
PR: https://git.openjdk.org/babylon/pull/321
More information about the babylon-dev
mailing list