WebGL Followup

[tqchen]

#672 Provide basic support for WebGL features. This Issue lists the followup items to be done.

• Store large data into 2D texture
  - We can use canonical form of [1024, y] (always make lowest dimension 1024).
• GEMM 1024x1024
• TOPI support of gl functions
• End to end example of resnet/mobielnet
• Optimize the bottleneck operators(conv2d_nchw), possibly via intermediate layout change

[tqchen]

@phisiart

[phisiart]

• conv2d_nchw
• dense
• injective
• pool
• softmax

[tqchen]

#836 #817

[tqchen]

dmlc/nnvm#369

[tqchen]

Now the only item that is left is the perf improvement of conv2d

[phisiart]

I feel like I can't continue working on this. My full-time job starts on Monday.

I took a look at deeplearn.js today. They don't seem to have special treatment for their WebGL conv2d operator: it's just several layers of for loops. There's no layout transformation either (no transpose or tiling) in model construction.

Last semester when we were doing the course project, we experimented matrix multiplication, and transpose doesn't really change performance. We didn't try tiling though.

[tqchen]

close this for now as most of stuffs are checked in thanks to @phisiart , followup with perf on another issue