how about the speed of depth convolution on cpu VS original caffe?

[lyk125]

No description provided.

[farmingyard]

@lyk125
The original caffe's group conv is very slow, and the depthwise conv is more faster.

[libra7]

@farmingyard
i have try the depthwise(cpu i7 -4790hq), group conv 130ms depthwise conv 128ms
gtx1060 group conv 35ms depthwise conv 5ms
Why hasn't the CPU model accelerated
Hope to get your answer

[Laulian]

@libra7 just as confused as you are，do you have idea now?

[HolmesShuan]

Using OpenMP may help. E.g.,

#pragma omp parallel for num_threads(...)

@Laulian @libra7

[Rookielike]

Using OpenMP may help. E.g.,

#pragma omp parallel for num_threads(...)

@Laulian @libra7

@HolmesShuan can you tell me where did you add ? in conv_dw_layer.cpp?

[HolmesShuan]

@Rookielike I have tried several settings:

#pragma omp parallel for num_threads(...)
#pragma omp parallel for

in several places. However, the run-time speed becomes unstable. You may regard it as a failed attempt.
But I find ncnn utilizes this technique to boost the runtime speed of depthwise convolution on ARM CPUs. Here are some useful inference libs you may have an interest.

[AlexofNTHU]

@libra7
which customized depthconvolution layers did you use to speed up?
Did you use this one?
https://github.com/yonghenglh6/DepthwiseConvolution
Achieving 5ms for mobilenet v1 with GTX1060 sounds really good!