You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
The CPU and VMVX attributes finished most of the implementation for encoding specialization. The remaining task is adding an e2e test, and I'm working on it.
In the data-tiling, we have MaterializeEncodingInfo struct to describe the layouts. Different from CPU backends, GPU has "swizzle". This field can be ignored in the interface implementation because it describes the layout of inner tiles, while it does not impact any storage size calculation. Thus, the implementation would be very similar to CPU attributes.
Here is one of examples that what is implemented for CPU backends: #19686
hanhanW
changed the title
Implement host/serialized encoding interface for GPUEncodingLayoutAttr
[DT] Implement host/serialized encoding interface for GPUEncodingLayoutAttr
Feb 18, 2025
The CPU and VMVX attributes finished most of the implementation for encoding specialization. The remaining task is adding an e2e test, and I'm working on it.
On GPU side, the path is still experimental, so I did not add the support yet. We are missing the implementation for EncodingLayoutAttrInterface and SerializedEncodingLayoutAttrInterface.
In the data-tiling, we have MaterializeEncodingInfo struct to describe the layouts. Different from CPU backends, GPU has "swizzle". This field can be ignored in the interface implementation because it describes the layout of inner tiles, while it does not impact any storage size calculation. Thus, the implementation would be very similar to CPU attributes.
Here is one of examples that what is implemented for CPU backends: #19686
For more information, please take a look at https://hackmd.io/@hwPnnvLBTB-JGVMeh-bCEA/r1dLRr1Fyl. The last section has the demo of how it works in CPU and GPU.
The text was updated successfully, but these errors were encountered: