[clDNN] "InvalidStateError: fails: 4" occurs for DEPTHWISE_CONV_2D and SOFTMAX operations

[Christywl]

Test Env: Chromium Version: nightly build 70.0.3503.0 (48326a5) Platform: Windows(Dell XPS 13), Linux Ubuntu 16.04

Expected Result: The test should pass.

Actual Result: "InvalidStateError: fails: 4" occurs for DEPTHWISE_CONV_2D and SOFTMAX operations.

InvalidStateError: fails: 4

How to Reproduce:

1. Launch chromium and enable WebML flag
2. Visit https://brucedai.github.io/nt/test/cts-all.html

[ibelem]

It may block mobilenet example with onnx format. Raise priority.

[huningxin]

@Christywl , could you please provide a set of failure test cases instead of cts-all?

[huningxin]

@ibelem , could you please confirm the onnx mobile issue? What are the reproduce steps? Thanks!

[Christywl]

@huningxin , you could use https://brucedai.github.io/nt/test/depthwise-softmax.html.

• check result for Depthwise conv2d float example
• check result for Depthwise conv2d float relu example
• check result for Depthwise conv2d float relu1 example
• check result for Depthwise conv2d float relu6 example
• check result for Depthwise conv2d float weights as inputs example
• check result for Depthwise conv2d float weights as inputs relu example
• check result for Depthwise conv2d float weights as inputs relu1 example
• check result for Depthwise conv2d float weights as inputs relu6 example
• check result for Softmax float example/1

[pinzhenx]

Results for onnx mobilenet

As shown above, the layer outputs diverge at mobilenetv20_features_linearbottleneck0_relu1_fwd, which is a fused node of depthwise convolution. Check the model here.

Test Env

Chromium Version: nightly build 70.0.3503.0 (b7ce158) Platform: Linux Ubuntu 16.04

How to reproduce

Link: https://pinzhenx.github.io/webml-demo/examples/image_classification_onnx/testlayer?model=mobilenetv2-1.0 Set log level to Verbose, mute Info & Warnings

[Christywl]

On the nightly build efb0f90, this issue wasn't reproduced. The tests still failed due to #565 and #244. Please see the details in #565 and #244, so close this issue.