[WASM] Some supported ops in wasm backend are not available

[akineeic]

Hello,
I have encountered some issues while using the tfjs-backend-wasm package, some infos are as follows:

TensorFlow.js version

• @tensorflow/tfjs-core@1.4.0

• @tensorflow/tfjs-backend-wasm@1.5.1-alpha5

Browser version

• Google Chrome: 79.0.3945.79 (Official Build) (64-bit)

Describe the problem or feature request

Since 'Softmax' is not supported in wasm backend, I want to simulate a sample softmax function with ops 'Exp', 'Sum' and 'Div'. But when I try it with the guidance of webpack and test in chrome's inspect, the following error occured.

Code to reproduce the bug / link to feature request

I used the code in the folder tfjs/tfjs-backend-wasm/starter/webpack, and I upgrade the package '@tensorflow/tfjs-backend-wasm' to version:1.5.1-alpha5 since I can't set the backend to 'wasm' while using the version in the 'yarn.lock'.

To reproduce the error, just change the code to the usage of other kernel like 'exp' or 'sum' like the followings:
const x = tf.tensor1d([1, 2, 3]);
x.sum().print(); // or tf.sum(x)

Then use commend yarn build & yarn watch to run the code in browser and the error occured. It seems like caused by the build or registry process or maybe I make some mistake in the usage. So I'm wondering are there any strategy to fix the issue?

Thanks.

[akineeic]

I have checked some other ops and found more ops that are also unavailable, like 'Greater', 'Less', 'Maximum', 'Minimum', etc. But those ops appear in the file 'all_kernels.ts' which is located in the package folder 'node_module'. Then I compared with the history of 'all_kernels.ts', it seems that the ops added after 'exp' is all unavailable. So I guess this issue occured maybe because the wasm related files in the folder 'wasm-out' didn't rebuild and update before the package published.

[dsmilkov]

Hi, I just updated our starter projects to depend on latest wasm AND latest core in #2706. The problem you were seeing is b/c tfjs-backend-wasm was at 1.5.1 but tfjs-core was at 1.4.0 (where 'exp' was not yet modularized) and that introduced a skew.