[Mlir-commits] [mlir] [MLIR][Linalg] Fix FlattenElementwiseOp for linalg.broadcast (PR #210738)

[Hugo Trachino]

================

----------------
nujaa wrote:

`linalg.broadcast` has a mandatory `dimension` attribute which must satisfy `inputRank + dimensions.size() == initRank` which suggests calling a scalar broadcast with something like `  %0 = linalg.broadcast ins(%arg0 : tensor<i32>) outs(%arg1 : tensor<32x2xi32>) dimensions = [0, 1]` but this is considered incorrect mlir input too. I tried   `%0 = linalg.broadcast ins(%arg0 : tensor<i32>) outs(%arg1 : tensor<32x2xi32>) dimensions = [0, 0]` (both dimensions as zeros) which segfaults. I will look into it and submit a patch to support this kind of linalg.broadcast.


https://github.com/llvm/llvm-project/pull/210738