[Mlir-commits] [mlir] [mlir][linalg] Introduce new `linalg.conv` op (PR #117688)

[Felix Schneider]

================
@@ -63,4 +63,30 @@ def TypeFn : I32EnumAttr<"TypeFn", "", [
   let cppNamespace = "::mlir::linalg";
 }
 
+
+class ConvDimEnumAttrCase<string sym, int val, string str = sym>
+    : IntEnumAttrCaseBase<I8, sym, str, val>;
+
+def ConvDimEnumAttr :
+    IntEnumAttr<I8, "ConvDimEnum", "summary", [
+      /// Batch is a dimension of input and output, indexed from a parallel loop.
+      ConvDimEnumAttrCase<"BATCH", 0, "N">,
+      /// Input channel is a dimension in all tensors, indexed from a reduction loop.
+      /// Depthwise convolutions perform no reduction across channels and therefore
+      /// do not use this.
+      ConvDimEnumAttrCase<"INPUT_CHANNEL", 1, "C">,
+      /// Output channel is a dimension in filter and output, index from a parallel loop.
+      ConvDimEnumAttrCase<"OUTPUT_CHANNEL", 2, "F">,
+      /// Group is a dimension in all tensors and indexed from a parallel loop.
+      ConvDimEnumAttrCase<"GROUP", 3, "G">,
+      /// Spatial dimensions occur in all tensors. Output is indexed from a parallel
+      /// loop, filter from a reduction loop and input from both.
+      ConvDimEnumAttrCase<"SPATIAL_0", 4, "0">,
----------------
ubfx wrote:

Agreed, the reason it's not in this PR is just so keep it representable as a simple enum for now and to keep everything testable via the old ops, which only went up to 3d. But the denotion `spatial_0, ...` was chosen to be extended to N dimensions via a more tailored attribute.

https://github.com/llvm/llvm-project/pull/117688