low performance of ConvolutionDepthWise operation

i am working on using NCNN to do a simple depthwise convolution operation. To be specific, is a depthwise convolution between a 256x26x26 tensor and a 256x4x4 tensor. Here is my naive test code.

ncnn::Option opt;
opt.num_threads = 1;

ncnn::Layer* op = ncnn::create_layer("ConvolutionDepthWise");

// set param
ncnn::ParamDict pd;
pd.set(0, 256);// num_output
pd.set(1, 4);// kernel_w
pd.set(5, 0);// bias_term
pd.set(6, 256*4*4);// weight_data_size
pd.set(7, 256);// group

op->load_param(pd);

// random input
ncnn::Mat in = RandomMat(26, 26, 256);

// random weights
ncnn::Mat weights[1];
weights[0] = RandomMat(256*4*4);// weight_data

op->load_model(ncnn::ModelBinFromMatArray(weights));

op->create_pipeline(opt);

// forward
op->forward(in, out, opt);

op->destroy_pipeline(opt);

delete op;

I got two problems.

1. The code above can not be excuted unless opt.use_packing_layout is set to be false.
2. The computation speed is very slow on a hisilocon 3559av100 test board (~35 ms), while the speed is greatly accelerated when the kernel size is set to be 3x3 or 5x5 (~1-2 ms).

any sugestions for me to do such a 4x4 depthwise convolution with a high efficiency.
thank you very much.

Start by reproducing the inline C++ test for a 256-channel 4x4 ConvolutionDepthWise operation on the Hisilicon 3559av100 setup, checking the use_packing_layout behavior. Compare the 4x4 timing with the reported 3x3 and 5x5 cases; done means explaining the failure and identifying an efficient supported path for the 4x4 operation.