On Wed, 29 Mar 2023 14:39:12 GMT, Alexey Ushakov <a...@openjdk.org> wrote:
> Use one display link thread per MTLContext. Adjust the refresh rate with the > current display. Those are the test batches run on internal servers. I will see if I can construct and share the J2DBench option file that shows the difference. The results that prompted my review comments- **J2DBench test related to opaque image rendering ** ------------------------------------------------------ Compared to the current mainline metal rendering pipeline performance - | Architecture | Your fix | Your fix + adding back BlitCommandQueue | | ------------ | ------------- | ------------- | | x64 | -88% degradation | -20% degradation | | aarch64 | -4% degradation | -1% degradation (negligible) | **Swingmark test** shows - ------------------------------------------------------ | Architecture | Your fix | Your fix + adding back BlitCommandQueue | | ------------ | ------------- | ------------- | | x64 | > -100% degradation | Equal to mainline | | aarch64 | -20% degradation | -15% degradation | ------------- PR Comment: https://git.openjdk.org/jdk/pull/13230#issuecomment-1498951287