[Mlir-commits] [mlir] [mlir][memref]: Add expand/collapse rewrite pattern to MemRef::CopyOp (PR #67808)

[Aviad Cohen]

AviadCo wrote:

@nicolasvasilache thanks for the ideas!

Regarding `linalg::collapseGenericOpIterationDims` - this may work with some refactoring, I'll see what I can do to make some pattern match based on `linalg.generic` to collapse a `linalg.copy` and base the DMA on `linalg.copy` instead.

I also investigated `bufferization.bufferize_to_alloc_tensor`. We expect the inputs to first exist on   global memory and to be able to work with them to promote (DMA import/export) to the local memory. So it seems that we better bufferize to allocation on global memory first and then use the `promote` transform to DMA it to local memory.

https://github.com/llvm/llvm-project/pull/67808