Interleaving columns

[davisyoshida]

I'm working with 4-bit quantized matrices, meaning that each value in the matrix is a uint8 containing two 4 bit values. If I have an mxn block of a matrix, I can use bitwise ops to get two mxn blocks, whose columns need to be interleaved into an mx2n matrix.

Is there any way to accomplish this interleaving in triton? It seems like for the most part slicing/concatenating is not the intended use case, so is it necessary to handwrite a kernel?

[michaelfeil]

Would also be interested in this.

[jlebar]

You can accomplish this using tl.join.