Fine-tune packed BackpropInput

[Linchenn]

Originally, we use idyCVal and idyCVal2 (could only be 0.0 or 1.0) to indicate if result.xy and result.zw are valid and result.xy * idyCVal and result.zw * idyCVal2 could return the results to avoid checking through if-branches.

If stride is 1, both idyCVal and idyCVal2 are always 1.0 and the original solution performs well. However, if stride is 2, whenever, only one of idyCVal and idyCVal2 is 1.0, which means either computing result.xy or result.zw is wasting of time.

As a result, this PR adds the if-branches to check idyCVal and idyCVal2 before computing, instead of always computing result.xy * idyCVal and result.zw * idyCVal2. This improves Conv2DBackpropInput ops in ArPortraitDepth ~40% and ArPortraitDepth model is improved from 23ms to 18.5ms (Benchmarked on LENOVO P620 2021).

Before the PR:

After the PR:

Reference #7371.

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

---

This change is 

[Linchenn]

Reviewable status: 0 of 1 approvals obtained

---

tfjs-backend-webgl/src/conv_backprop_packed_gpu.ts line 78 at r1 (raw file):

            if (idyCVal && idyCVal2) {
              for (int d2 = 0; d2 < ${convInfo.outChannels}; d2 += 2) {
                vec4 wValue = getW(wRPerm, wCPerm, d1, d2);

Even though the three branches have the same codes, we still should not move the if-branches into the for loop. Otherwise the performance drops (for Conv2DBackpropInput op, drops ~10%), probably because the if-branches will be executed for each iteration (then threads probably would be stalled/synced in each iteration).

Code quote:

              for (int d2 = 0; d2 < ${convInfo.outChannels}; d2 += 2) {
                vec4 wValue = getW(wRPerm, wCPerm, d1, d2);

[pyu10055]

great work, thanks!

Reviewable status: complete! 1 of 1 approvals obtained (waiting on @qjia7)

[qjia7]

LGTM, thanks!

[Linchenn]

Benchmark on Samsung Galaxy S22 Ultra:

Before this PR (with packed ConvBackpropInput):

After this PR:
ConvBackpropInput op's inference time drops (the top1 drops from 5.05 to 4.10)

[qjia7]

I just saw a big regression on Intel devices (both CFL and ADL) with this PR.
Before (CFL)


After (CFL)

Before (ADL)


After (ADL)

[qjia7]

It's surprising that there is such a big regressions for Intel devices. But the similar algorithm works well for webgpu. It may be a driver bug if it only happens on Intel devices. I can report an issue to our driver team. But it will be good if we can skip this optimization for Intel devices on TFJS level temporarily.