[Mlir-commits] [mlir] [Linalg] Add *Conv3D* matchers (PR #172141)
Han-Chung Wang
llvmlistbot at llvm.org
Mon Dec 15 04:04:34 PST 2025
================
@@ -1088,6 +1088,114 @@ bool isaConvolutionOpOfType<linalg::Conv3DOp>(LinalgOp op,
.matchBody();
}
+// #inputMap = affine_map<(N, D, H, W, F, d, h, w, c)
+// -> (N, D + d, H + h, W + w, c)>
+// #filterMap = affine_map<(N, D, H, W, F, d, h, w, c) -> (d, h, w, c, F)>
+// #outputMap = affine_map<(N, D, H, W, F, d, h, w, c) -> (N, D, H, W, F)>
----------------
hanhanW wrote:
I did not point this out in previous reviews. It is optional. I think we can drop the comments because the code self-documents this. We can keep them for consistency.
```
.matchMaps({/*inputMap=*/{N, m.strided(D, d, 0), m.strided(H, h, 1),
m.strided(W, w, 2), c},
/*filterMap=*/{d, h, w, c, F},
/*outputMap=*/{N, D, H, W, F}})
```
https://github.com/llvm/llvm-project/pull/172141
More information about the Mlir-commits
mailing list