[Mlir-commits] [mlir] [mlir][amdgpu] Lower amdgpu.make_dma_base (PR #169817)

[Erick Ochoa Lopez]

================
@@ -1222,14 +1231,12 @@ def AMDGPU_MakeDmaBaseOp :
     AMDGPU_Op<"make_dma_base", [Pure, AttrSizedOperandSegments]>,
     Arguments<(ins
                    Arg<AnyMemRef, "buffer to read from">:$src,
-                   Variadic<Index>:$srcIndices,
+                   Variadic<Index>:$src_indices,
                    Arg<AnyMemRef, "buffer to write to">:$dst,
-                   Variadic<Index>:$dstIndices)>,
+                   Variadic<Index>:$dst_indices)>,
     Results<(outs AMDGPU_TDMBaseType: $base)> {
----------------
amd-eochoalo wrote:

@krzysz00 thinking about this, maybe it would be best to instead of having src, dst memrefs here, just have lds and global memrefs. As you mentioned, the descriptor is intended to be reused across store and load operations. Then it would be easier to fold a pair of instructions:

```
%base0 = amdgpu.make_dma_base %lds[...], %global[...]
%base1 = amdgpu.make_dma_base %lds[...], %global[...]
```

Rather than

```
                                                          // src       , dst
%base0 = amdgpu.make_dma_base %lds[...], %global[...]
%base1 = amdgpu.make_dma_base %global[...], %lds[...]
```

https://github.com/llvm/llvm-project/pull/169817