Jiajia Qin

@Linchenn Great investigation. Your conclusion makes sense for me. I think one possible way is to explore another efficient algorithm to process such kind of shape. But it's not easy...

> Do you have examples/hints for other algorithms? @qjia7 > > As we known, WebGL's `conv2d` has three ways (1*1, img2col and brute force). Do you mean a new way...

> error: Uncaught (in promise) Error: Kernel 'Reciprocal' not registered for backend 'webgpu' > throw new Error(`Kernel '${kernelName}' not registered for backend '${this.backendName}'`); This [op](https://github.com/tensorflow/tfjs/pull/6743) was added two weeks ago....

@Linchenn FYI. I just noticed it when I look at the webgl implementation. I don't know how much it will help the perf. But from the code design, I think...

@vladmandic @shurshilov Please have the last test with this PR. I also make it ready for review. The verified solution https://github.com/tensorflow/tfjs/issues/6822#issuecomment-1247460709: ``` vec4 nanValue = a; bvec4 isNaN = isnan(a);...

Kindly ping you guys. Also add @Linchenn in case you are in vacation.

> @qjia7 It sounds the correctness of the builtin/polyfilled `isnan` can be programmatically probed with a small snippet and then the right one could be picked up _automatically_ without "creating...

@haoyunfeix Please take a look, thanks.

@haoyunfeix @vicary I think deno is relying on an old webgpu(wgsl) point. But tfjs-webgpu is following up the latest webgpu/WGSL spec. So to resolve this issue, deno needs to upgrade...

@vicary Please let us know whether this issue can be fixed after upgrading webgpu in deno.