[clDNN] deprecate clDNN backend

[huningxin]

[huningxin]

@Christywl , as https://github.com/otcshare/chromium-src/pull/27 merged, could you please share with us the status of IE-clDNN backend vs. clDNN backed in terms of

1. CTS results
2. examples
3. performance

Thanks!

[Christywl]

• CTS results: clDNN: pass 366, fail 97 IE-clDNN: pass 238, fail 225

  • main issues on IE-clDNN: ARGMAX(https://github.com/intel/webml-polyfill/issues/947), FULL_CONNECTED(https://github.com/intel/webml-polyfill/issues/750), PRELU(https://github.com/intel/webml-polyfill/issues/1093), RESIZE_BILINEAR(https://github.com/intel/webml-polyfill/issues/754), ADD/MUL broadcasting(https://github.com/intel/webml-polyfill/issues/751) aren't supported

• Examples: IE-clDNN has some issues comparing with clDNN:

  • Chromium will crash on IE-clDNN for quantized models(https://github.com/intel/webml-polyfill/issues/851), and crash for KWS CNN(TFLite) of Speech Commands (https://github.com/intel/webml-polyfill/issues/1277)
  • Some models can’t be ran on IE-clDNN since FULL_CONNECTED op isn’t supported(#750): 1. Image Classification: Inception v4(TFLite) / Inception Resnet v2(TFLite) / ResNet50 v1(ONNX) / ResNet50 v2(ONNX) / ResNet50 v1(OpenVINO) / Inception v2(OpenVINO) / Inception v4(OpenVINO) 2. Facial Landmark Detection (while the models can be ran on clDNN with the incorrect position of key points) 3. Speech Commands: KWS DNN(OpenVINO) 4. Speech Recognition: wsj_dnn5b
  • Some models can’t be ran on IE-clDNN since Broadcasting isn’t supported(#751): Image Classification: Inception v2(ONNX) / DenseNet 121(ONNX) / DenseNet 121(OpenVINO) Face Recognition: SSD MobileNet v1 Face(TFLite) + Face Reidentification(OpenVINO), SSD MobileNet v2 Face(TFLite) + Face Reidentification(OpenVINO), SSDLite MobileNet v2 Face(TFLite) + Face Reidentification(OpenVINO)
  • Segmentation doesn’t work on IE-clDNN since RESIZE_BILINEAR(https://github.com/intel/webml-polyfill/issues/754) and ARGMAX aren’t supported(https://github.com/intel/webml-polyfill/issues/947)

[ibelem]

DNNL and DNNL in IE can be taken by @lisa0314 . @mingmingtasd please investigate it.