This PR is setup in tandem with a corresponding PR #17974 into o3de/development and can only be merged in tandem.
It transitions the samples in the ASV to multi-device resources.
As the change set is enormous, this feature has been developed separately on a dedicated branch o3de/o3de-atom-sampleviewer at multi-device-resources (github.com) and is now ready to be merged.
This PR does essentially two things:
It transitions existing samples to multi-device resources
It introduces two new samples, namely:
RHI/MultiGPU
Shows how to do split-frame rendering with a simple triangle pass on the RHI level, copying over the result from the second device to the first device and compositing the final output there. This sets up ScopeProducers on specific devices to submit DrawItems to different GPUs and submit the corresponding CopyItems to transfer the data between GPUs
RPI/MultiGPU
This shows how to do the same example as on the RPI layer (just without the animation) using the Pass and Pipeline setups in json.
List of PRs
The following table lists all PRs that have been merged into multi-device-resources and are part of this PR.
The final merge will also include one additional commit which changes the naming scheme (the intermediate naming used MultiDeviceResource as well as SingleDeviceResource, this now changes to Resource and DeviceResource).
The table lists all PRs into o3de as well as o3de-atom-sampleviewer.
And also many thanks to the reviewers of all these PRs, as this is a huge change, the review effort also was significant and helped improve the code significantly (listed in random order)
This PR is setup in tandem with a corresponding PR #17974 into
o3de/developmentand can only be merged in tandem. It transitions the samples in the ASV to multi-device resources. As the change set is enormous, this feature has been developed separately on a dedicated branch o3de/o3de-atom-sampleviewer at multi-device-resources (github.com) and is now ready to be merged.This PR does essentially two things:
RHI/MultiGPURHIlevel, copying over the result from the second device to the first device and compositing the final output there. This sets upScopeProducerson specific devices to submitDrawItemsto different GPUs and submit the correspondingCopyItemsto transfer the data between GPUsRPI/MultiGPURPIlayer (just without the animation) using thePassandPipelinesetups injson.List of PRs
The following table lists all PRs that have been merged into
multi-device-resourcesand are part of this PR. The final merge will also include one additional commit which changes the naming scheme (the intermediate naming usedMultiDeviceResourceas well asSingleDeviceResource, this now changes toResourceandDeviceResource). The table lists all PRs intoo3deas well aso3de-atom-sampleviewer.BufferImagePipelineStateQueryTransientAttachmentPoolModelLodMeshlets-GemRayTracingSwapChainRPI::ShaderResouceGroupuse the multi-device version ofShaderResourceGroupDrawItemDispatchItem-related resources by martinwinter-huawei · Pull Request #17520 · o3de/o3de (github.com)DispatchItemAttachmentReadbackto multi-devicedeviceIndexin theScopeclassdeviceIndexin theFrameGraphExecuteContextand utilize thedeviceIndexwherever possiblecontext.GetDeviceIndex()where possible by martinwinter-huawei · Pull Request #670 · o3de/o3de-atom-sampleviewer (github.com)FrameSchedulerandFrameGraph-related resources including the backends to support multi-device resourcesScopeProducers get adeviceIndexand consequently theRenderpassgets a deviceIndex injsonFrameGraph*CopyPassdeviceIndexin the PassViewerContributors & Reviewers
Many thanks to the contributors to this PR (listed in alphabetical order)
And also many thanks to the reviewers of all these PRs, as this is a huge change, the review effort also was significant and helped improve the code significantly (listed in random order)