MLTensor export to GPUBuffer with custom usages?

[mwyrzykowski]

As suggested #942 (comment) it would be nice to allow custom GPUBufferUsage, e.g., GPUBufferUsage.MAP_READ or GPUBufferUsage.INDIRECT for instance

Otherwise it limits the areas where the exported buffer can be used, e.g., prevents mapAsync or indirect operations.

cc @bbernhar

[fdwr]

Mike's issue of flag usage is related to my comment (#942 (comment)). So rather than create a separate issue, maybe we can tag along with this one.

[bbernhar]

Should createExportableTensor() accept the exact GPUBufferUsageFlags for the exported buffer, or an additive set of extra usages on top of WebNN’s current baseline (STORAGE | COPY_SRC | COPY_DST)? For the exact-usage option, the current mask could remain the default. AFAIK, WebGPU has no concept of "default flags".

@mwyrzykowski @fdwr

[mwyrzykowski]

I think exact usage with the current default would be great