[WebGL] Support packed Conv2DBackpropInput

[Linchenn]

With this PR:
Conv2DBackpropInput op is accelerated ~2x, partially fixed #5197.

Model	Before(ms)	After(ms)
BlazePoseDetector	23.1	17.5
ArPortraitDepth	52.3	23.9 (also set EXP_CONV to true)

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

---

This change is 

[qjia7]

Clever algorithm and great work!

[pyu10055]

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

---

tfjs-backend-webgl/src/conv_backprop_packed_gpu.ts line 55 at r1 (raw file):

        //intialize dotProd with a small epsilon seems to reduce GPU accuracy loss.
        vec4 dotProd = vec4(0.000000000000001);

use an JS constant to interpolate in this value.

---

tfjs-backend-webgl/src/conv_backprop_packed_gpu.ts line 90 at r1 (raw file):

              dotProd.xy += vec2(dot(dyValue, wValue.xy), dot(dyValue, wValue.zw)) * idyCVal;

              dySample = getDy(batch, idyR, idyC2, d2);

might be good to move all memory accesses together.

---

tfjs-backend-webgl/src/conv_backprop_packed_gpu.ts line 90 at r1 (raw file):

              dotProd.xy += vec2(dot(dyValue, wValue.xy), dot(dyValue, wValue.zw)) * idyCVal;

              dySample = getDy(batch, idyR, idyC2, d2);

can idyC2 could be the same as idyC ? only read again if idyC2 != idyC ?

[Linchenn]

Thank you Ping. After tuning the shader, the inference time for the following op is improved from ~2.8928ms to ~2.6809903999999984ms on a Linux workstation:

const x = tf.ones([1,8,6,256]);
const w = tf.ones([4,4,256,256]);

s = (await tf.profile(()=>tf.conv2dTranspose(x, w, [1,16,12,256], 1, 'valid'))).kernels[0].kernelTimeMs

After benchmarking, the improvement is mainly from add a condition for reading dySample2, from ~2.8928ms to ~2.711304960000001ms

Reviewable status: complete! 1 of 1 approvals obtained (waiting on @mattsoulanille and @pyu10055)

---

tfjs-backend-webgl/src/conv_backprop_packed_gpu.ts line 55 at r1 (raw file):

Previously, pyu10055 (Ping Yu) wrote…

use an JS constant to interpolate in this value.

Removed it as we discussed, because this initial value currently is unnecessary, but we could keep this in mind when facing some correctness issues later.

---

tfjs-backend-webgl/src/conv_backprop_packed_gpu.ts line 90 at r1 (raw file):

Previously, pyu10055 (Ping Yu) wrote…

can idyC2 could be the same as idyC ? only read again if idyC2 != idyC ?

Done.

---

tfjs-backend-webgl/src/conv_backprop_packed_gpu.ts line 90 at r1 (raw file):

Previously, pyu10055 (Ping Yu) wrote…

might be good to move all memory accesses together.

Done.

[pyu10055]

Reviewed 1 of 2 files at r1, 1 of 1 files at r4, all commit messages.
Reviewable status: complete! 2 of 1 approvals obtained (waiting on @mattsoulanille)