Cf. the comment thread by @mdgrs in #3006 (comment)
#3006 implemented fusion of batchnorm1d into preceding convolution and linear ops. The operations that batchnorm1d lowers to are multiplication by general tensors, which only make sense to fuse into channel-wise operations.
However, some constructions (reportedly average pooling and sum pooling) reduce to multiplication by a splat after a non-channel ops like conv2d and conv1d. These splats can be fused into the filter matrix of the convolution. This issue is to add support for that.
Cf. the comment thread by @mdgrs in #3006 (comment)
#3006 implemented fusion of batchnorm1d into preceding convolution and linear ops. The operations that batchnorm1d lowers to are multiplication by general tensors, which only make sense to fuse into channel-wise operations.
However, some constructions (reportedly average pooling and sum pooling) reduce to multiplication by a splat after a non-channel ops like conv2d and conv1d. These splats can be fused into the filter matrix of the convolution. This issue is to add support for that.