Checking 3.8.0 (defaults bad, good with flag)

[wingman-jr-addon]

I gave this a shot but unfortunately the loading + first inference performance degrades heavily on my machine from 2.7.0 to 3.8.0.

Here are some numbers:
3.8.0

LIFECYCLE: TIMING LOADING: {"uploadWaitMs":3241,"downloadWaitMs":0,"kernelMs":{"error":"WebGL query timers are not supported in this environment."},"wallMs":13325}
LIFECYCLE: TIMING LOADING: {"uploadWaitMs":3262,"downloadWaitMs":0,"kernelMs":{"error":"WebGL query timers are not supported in this environment."},"wallMs":13569}
LIFECYCLE: TIMING LOADING: {"uploadWaitMs":3179,"downloadWaitMs":0,"kernelMs":{"error":"WebGL query timers are not supported in this environment."},"wallMs":13623}

2.7.0

LIFECYCLE: TIMING LOADING: {"uploadWaitMs":3349,"downloadWaitMs":0,"kernelMs":{"error":"WebGL query timers are not supported in this environment."},"wallMs":8992}
LIFECYCLE: TIMING LOADING: {"uploadWaitMs":3639,"downloadWaitMs":0,"kernelMs":{"error":"WebGL query timers are not supported in this environment."},"wallMs":9447}
LIFECYCLE: TIMING LOADING: {"uploadWaitMs":3085,"downloadWaitMs":0,"kernelMs":{"error":"WebGL query timers are not supported in this environment."},"wallMs":8448}

[wingman-jr-addon]

Update: I did some work in figuring out the core of the regression and ended up making three changes:

1. tf.env().set('WEBGL_PACK_DEPTHWISECONV', false) fixes a regression caused by a flag default changing in 3.3 -> 3.4.
2. The order of the outputs changed somehow in the TF.js upgrade, so it broke and I needed to switch the ordering on sqrxrScore.
3. TF.js team was kind and pointed out I might get a load-time performance boost from the new tf.env().set('WEBGL_USE_SHAPES_UNIFORMS', true)

So the net result is that load times are something like 500+ ms faster on my box.