Pull Request Readiness Checklist

Mish (+ swish and sigmoid):

This activation is used in YOLOv4. Surprisingly, the mish kernels are compute-bound instead of being bandwidth-bound. The first convolution in YOLOv4 takes less time than the fused bias mish activation step.

Mish is a composition of several functions: tanh(log(1 + exp(x))). Hence, it is very likely that there is a very good approximation or simplification to the mish function (clearly the log isn't necessary since tanh is a composition over e^x). This PR introduces a fast numerically stable implementation of the mish activation.

Details of the approximation can be found here.

Based on similar reasoning, good fast approximations to swish and sigmoid were also added.

Identity is fused with convolution layer:

Activations would not be fused with convolution if an identity layer is present between them. This happens in the case of Mask RCNN: the final sigmoid operation is not fused with the preceding convolution operation.

This kernel is compute-bound if the sigmoid approximation mentioned above is not used. The timings reported here are with the approximation.

Others:

See details at https://github.com/opencv/opencv/wiki/How_to_contribute#making-a-good-pull-request

• I agree to contribute to the project under OpenCV (BSD) License.
• To the best of my knowledge, the proposed patch is not based on a code under GPL or other license that is incompatible with OpenCV
• The PR is proposed to proper branch
• There is reference to original bug report and related work
• There is accuracy test, performance test and test data in opencv_extra repository, if applicable
  Patch to opencv_extra has the same branch name.
• The feature is well documented and sample code can be built with the project CMake

force_builders=Custom
buildworker:Custom=linux-4
build_image:Custom=ubuntu-cuda:18.04