Hi, in caffe2/quantization/server/conv_dnnlowp_op.cc, i find it says “only 1-3d convolution supported by NHWC”, so only 3d convolution can be implemented and accelerated by fbgemm? how to apply fbgemm to 2d convolution?
Ryan95 (Ryan ) #1
dskhudia (Daya Khudia) #2
This means 1d, 2d and 3d are supported for NHWC storage. 2d convolutions (for low precision) are accelerated by fbgemm when the engine used is DNNLOWP.