[Mlir-commits] [mlir] [mlir][linalg] unfold projected permutation. (PR #114704)

[Javed Absar]

================
@@ -0,0 +1,71 @@
+// RUN: mlir-opt %s -split-input-file --linalg-specialize-generic-ops | FileCheck %s
+
+#projection = affine_map<(d0, d1, d2, d3, d4) -> (d2, d3, d1)>
+#identity   = affine_map<(d0, d1, d2, d3, d4) -> (d0, d1, d2, d3, d4)>
+
+func.func @transpose_and_broadcast(%x : tensor<7x8x9xf32>, %y:  tensor<5x9x7x8x10xf32>, %z :  tensor<5x9x7x8x10xf32>) ->  tensor<5x9x7x8x10xf32> {
+  %res = linalg.generic
+     { indexing_maps = [#projection, #identity, #identity], iterator_types = ["parallel", "parallel", "parallel", "parallel", "parallel"]} 
+     ins(%x, %y : tensor<7x8x9xf32>, tensor<5x9x7x8x10xf32>) outs(%z : tensor<5x9x7x8x10xf32>) {
+     ^bb0(%in: f32, %in_1: f32, %out: f32):
+       %div = arith.divf %in, %in_1 : f32
+       linalg.yield %div : f32
+  } -> tensor<5x9x7x8x10xf32>
+  return %res : tensor<5x9x7x8x10xf32>
+}
+
+// CHECK-LABEL: transpose_and_broadcast
+// CHECK-SAME: %[[X:.+]]: tensor<7x8x9xf32>, %[[Y:.+]]: tensor<5x9x7x8x10xf32>, %[[Z:.+]]: tensor<5x9x7x8x10xf32>) -> tensor<5x9x7x8x10xf32> {
+// CHECK: %[[E0:.+]] = tensor.empty() : tensor<9x7x8xf32>
+// CHECK: %[[X_trans:.+]] = linalg.transpose ins(%[[X]] : tensor<7x8x9xf32>) outs(%[[E0]] : tensor<9x7x8xf32>) permutation = [2, 0, 1]
----------------
javedabsar1 wrote:

Yes of course you are right! Two parts to this. 

1) Broadcast semantics is - 
dim(result, i)  == dim(input, permutation[i]) 

Therefore, for input `7x8x9` and permutation `[2,0,1]` . The output works therefore out to `9x7x8`. 

2)  working out the permutation from affine-map e.g. `(d2, d3, d1) -> (d1, d2, d3)` and vice-versa. 

https://github.com/llvm/llvm-project/pull/114704