[Mlir-commits] [mlir] [MLIR][XeGPU][VectorToXeGPU] Lower vector.load/store/transfer_read/transfer_write to new offsets syntax (PR #162095)

[Dmitry Chigarev]

dchigarev wrote:

> So I think that the right solution is what you proposed: As we move the offsets to load_nd, XeGPU user must explicitly insert a memref.subview to flatten the ND tensor to 2D, and then create a 2D tensor descriptor.

@Jianhui-Li 

changed the logic to insert a memref.subview that collapses the shape to 2D where necessary. For example:

```mlir
// input:
%0 = vector.load %source[%off0, %off1, %off2] : memref<8x16x32xf32>, vector<8x16xf32>

// --vector-to-xegpu (old)
%tdesc = xegpu.create_nd_tdesc %source[%off0, %off1, %off2] : memref<8x16x32xf32> -> tdesc<8x32xf32>
%vec = xegpu.load_nd %tdesc

// --vector-to-xegpu (new)
%collapsed = memref.subview %source[%off0, 0, 0] [1, 16, 32] [1, 1, 1] :
    memref<8x16x32xf32> -> memref<16x32xf32, strided<[32, 1], offset: ?>>
%tdesc = xegpu.create_nd_tdesc %collapsed : memref<16x32xf32, ...> -> tdesc<8x32xf32>
%vec = xegpu.load_nd %tdesc[%off1, %off2]
```

https://github.com/llvm/llvm-project/pull/162095