search

halide / Halide

a language for fast, portable data-parallel computation
https://halide-lang.org
Other
5.84k stars 1.07k forks source link

D3D12Compute does not implement allocation cache #5000

Open shoaibkamil opened 4 years ago

shoaibkamil commented 4 years ago

We should implement an allocation cache similar to that on other GPU APIs.

steven-johnson commented 4 years ago

Have we ever refactored the cache code into something common, or are we reimplementing it in each gpu backend?

abadams commented 4 years ago

5000 :fireworks:

slomp commented 4 years ago

Looking at how the cuda allocation cache works, I think we could refactor that into a generic helper that can be used by other backends. This way there's something readily available for everybody, and if one wants to be fancy in a backend and perform sub-allocation management and things like that, they also have the option to specialize it in the backend.