The batch match-mul API would be very useful for convolutions. It's also used on Nvidia GPUs.
Often you get tensors that are sliced views on one or more dimensions, with BLAS-api you need to allocate them in a contiguous buffer. BLIS does not need that, there is a repacking algorithm to overcome memory bandwidth bounds in matrix multiplication (search the "roofline model"), the slicing/realloc can be fused with repacking with BLIS API.