Conversion APIs in the in Linear Algebra Runtime spec

[anupamachandra]

Related to : https://github.com/microsoft/DirectX-Specs/blob/master/d3d/D3D12LinearAlgebraRuntimeFeatureSupport.md
In the spec, we have the following:
"SrcComponentType and DestComponentType must be identical. In other words, this conversion only modifies memory layout and not data." in the Conversion API Valid usage section .

We do want the ability for this function to convert from higher to lower precision right?

[jenatali]

I know in the cooperative vectors API, the way that you got an F8 matrix in MulOptimal layout was to pass F16 input data and have the driver convert (or not). My understanding of the reason that F8, including emulation, is part of the tier 1 constraints is so that you can have real F8 data, meaning that F8 -> F8 into a MulOptimal layout would be an unpacking step for drivers that are emulating.

I guess if we're talking about small models, then maybe F16 data that's F8-quantized might be okay as the expected input? I'm okay either way.

[anupamachandra]

I think having the API support type conversions is important for user experience than requiring the developers to do it inline with what VK supports and what we had in the earlier coopvec spec.

[jenatali]

Yeah that's fair, I can tweak the language to re-enable data type conversions.

One thing that the cooperative vector spec was pretty strict about was that the only way to get FP8 data was by converting FP16 or FP32 data, to enable drivers to emulate FP8 via FP16, and I don't think that makes sense to carry forward.