[WebGPU] Fix shader key for ScatterProgram

[Linchenn]

Issue

This PR fixes the following error because a TensorScatterUpdate op is trying to reuse a cached pipeline but it should not, because the size of their uniforms (the strides, the length of which is dynamic) are different.

Solution

We could also add such uniform information into TensorScatterUpdate kernel's shader key instead of all kernel's shader key.

Alternatives considered

We could also add uniform informations into shader key to fix it, but it's unnecessary to apply it to all shader keys since this uniform problem only happens, from my perspective, for scatter program.

To see the logs from the Cloud Build CI, please join either our discussion or announcement mailing list.

[qjia7]

In theory, we could also add such uniform information into TensorScatterUpdate kernel's shader key instead of all kernel's shader key. However, the uniformData is passed directly into runWebGPUProgram, and is not passed to ScatterProgram constructor, so TensorScatterUpdate kernel's shader key could not know the size of uniforms at runtime and thus could not add such uniform information.

Does it work to change it as below?

this.shaderKey = `scatter_${indicesRank}_${updatesRank}_${
        this.sliceDimGreaterThanOne}_${outputDtype}_${sumDupeIndices}`;

to

this.shaderKey = `scatter_${indicesRank}_${updatesRank}_${
        this.sliceDimGreaterThanOne}_${outputDtype}_${sumDupeIndices}_${strides.length}`;

[Linchenn]

Does it work to change it as below?

Nice catch! It works and just updated the PR.

[qjia7]

LGTM, thanks!