wasm backend returns incorrect values as result

I have a very simple float32 model (link provided in the code below) that takes image with values in range of 0…255 as input and returns a single float32 value.
Works perfectly with webgl backend, but values explode using wasm backend.
E.g., instead of returning value in range of 1-99, it returns values like 700+.

This is just one more case that looks like incorrect cast inside wasm backend, likely related to #4326 and #4311 that I’ve reported earlier.

simple reproduction code:

const resizeT = tf.image.resizeBilinear(image, [64, 64], false); // image is any picture of a face
const normalizeT = tf.mul(resizeT, [255.0]);
resizeT.dispose();
const model = await tf.loadGraphModel('https://vladmandic.github.io/human/models/age-ssrnet-imdb.json');
const ageT = await model.predict(normalizeT);
normalizeT.dispose();
const age = ageT.dataSync();
ageT.dispose();
console.log(age);

Environment: TFJS 2.8.3 on Chrome 87 / Windows 10

[jinjingforever]

(sorry, I was on vacation last week)

Today we figured out that the error was caused by out-of-date xnnpack+emscripten. I am in the process of updating our toolchain to the latest xnnpack and emscripten. Hopefully it can be done this week. Thanks!

[jinjingforever]

BTW we further noticed that the wrong results come from some of the “div” nodes. With SIMD enabled, those nodes produce the results with the numerator and the denominator swapped… I am asking folks from the xnnpack project to investigate further. Thanks!