Can you tell me more about conv2d ( ggml_conv_2d )?

[Cyberhan123]

I am not a professional, and I would like to know whether there is room for optimization of ggml_conv_2d from an industry perspective, because it is too slow at 512x512.

I searched for relevant information，I'll list them：
ggml_conv_2d: ggerganov/ggml#564 (comment)
pytorch: pytorch/pytorch#49582

[Cyberhan123]

From a lazy perspective, we can inherit third-party libraries for testing
https://github.com/oneapi-src/oneDNN
https://github.com/amd/ZenDNN

[Cyberhan123]

At present, I have found a more feasible reference: https://github.com/Tencent/ncnn/blob/master/src/layer/x86/convolution_x86.cpp#L280

[FSSRepo]

@Cyberhan123 Thank you a lot!!, best research!.

https://github.com/Tencent/ncnn/blob/master/src/layer/x86/convolution_x86.cpp#L280

I will try to implement this code to ggml_conv2d to test memory usage and speed

[Cyberhan123]

According to my research, the implementation process on the CPU needs to be completely customized instead of using the ggml implementation.

[Cyberhan123]

In CUDA, we are actually not in a hurry to implement it because it is more complicated.

[Cyberhan123]

No matter what, this is also my learning process, and I am willing to spend time studying these details with you. This is also part of our growth.