ZhennanQin commented on a change in pull request #12530: Implement mkldnn convolution fusion and quantization. URL: https://github.com/apache/incubator-mxnet/pull/12530#discussion_r223569976
########## File path: src/operator/quantization/mkldnn/mkldnn_quantize-inl.h ########## @@ -75,6 +75,11 @@ static void MKLDNNQuantizeComputeKer(const std::vector<NDArray>& inputs, auto i_mpd = i_mem->get_primitive_desc(); auto i_desc = i_mpd.desc(); mkldnn::memory::format i_fmt = static_cast<mkldnn::memory::format>(i_desc.data.format); + if (i_fmt == mkldnn::memory::format::nchw || + i_fmt == mkldnn::memory::format::nChw8c || + i_fmt == mkldnn_nChw16c) { + i_fmt = mkldnn::memory::format::nhwc; + } Review comment: @KellenSunderland Sorry for responding late. For mkldnn quantization flow, we won't offline quantize any params, but using fp32 params(eg. weight and bias) for quantized convolution. We will online quantize convolution params in first forwarding. For the code here, it's for the default output format of 'quantize' op when using in mkldnn quantization flow, this won't effect non-mkldnn quantization flow. ---------------------------------------------------------------- This is an automated message from the Apache Git Service. To respond to the message, please log on GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: us...@infra.apache.org With regards, Apache Git Services