fast Depthwise mxnet operator

[vsooda]

@KeyKy Would you mind update the fast depthwise operator. or could you describe the key for implement fast operator. Thank you.

[KeyKy]

@vsooda fast depthwise operator is implemented by just copy the cuda code from ^-^ url. And then you can get the 2ms/image.

[ccJia]

@KeyKy Have some good idea or implement on CPU T-T ??

[KeyKy]

@ccJia fast depthwise operator (which num_group == num_outputs) should be calculated in multi-threaded channel-by-channel. According to tencent's ncnn, I think sliding window (highly optimized, which also using in BVLC/caffe#5665) maybe faster than im2col+GEMM.

[ccJia]

@KeyKy OK, thank you . Let me have a try ^-^.

[KeyKy]

@ccJia looking forward to your job. Could you share with me after you have done the work? ^-^

[ccJia]

@KeyKy Some of my friends told me that the direct convolution like NCNN is not work.....

[KeyKy]

@ccJia Did they explain the reason or they try?

[ccJia]

@KeyKy They tried on Win.

[KeyKy]

@ccJia OK. I will try it in my spare time.

[ccJia]

@KeyKy waiting your good news .

[vsooda]

@KeyKy thank you