[Perf] The time of Conv2DBackpropInput is very long in BlazePose/hand_detector models in WebGL

[qjia7]

hand_detector

Kernel	Time(ms)	Inputs	Output	GPUPrograms
Conv2DBackpropInput	9.73	input0: 4D[1,16,16,256]input1: 4D[2,2,128,256]	1,32,32,128	UnpackProgram: 0.026083, Conv2DDerInputProgram: 9.706294
Conv2DBackpropInput	4.93	input0: 4D[1,8,8,256]input1: 4D[2,2,256,256]	1,16,16,256	UnpackProgram: 0.011333, Conv2DDerInputProgram: 4.91648

BlazePose with inputSize 256, inputType Tensor

Kernel	Time(ms)	Inputs	Output	GPUPrograms
Conv2DBackpropInput	20.20	input0: 4D[1,7,7,1152]input1: 4D[2,2,192,1152]	1,14,14,192	UnpackProgram: 0.042083, Conv2DDerInputProgram: 20.159669

[qjia7]

fyi @pyu10055 @lina128 @jinjingforever

[lina128]

This is very informative, thank you @qjia7 ! In the GPUPrograms column, is Conv2DDerInputProgram the packed version and UnpackProgram the unpacked version, both from WebGL backend?

[pyu10055]

@lina128 the UnpackProgram is the shader to unpack outputs from the previous kernel, the Conv2DDerInputProgram is the shader for Conv2DBackpropInput, which one supports unpacked texture.

[qjia7]

FYI, #6603 use the simulated matmul_vec4 on webgpu can greatly improve the performance of Conv2DBackpropInput.

[pyu10055]

FYI @Linchenn

[Linchenn]

Thank you Jiajia @qjia7! This is a great inspiration. Let me see if we can reuse this algorithm and, at least, we could have a packed version of Conv2DDerInputProgram to improve the performance.

[Linchenn]

hand_detector is currently deprecated, while BlazePoseDetector is accelerated with packed Conv2DBackpropInput. However, Conv2DBackpropInput is still the performance bottleneck for BlazePoseDetector.

[google-ml-butler]

Are you satisfied with the resolution of your issue?
Yes
No