[Mlir-commits] [mlir] [mlir][tensor] Fold producer linalg transpose with consumer tensor pack (PR #75658)

[lorenzo chelini]

================
@@ -345,3 +345,170 @@ func.func @tensor_pack_linalg_transpose_fold_dynamic_outer_dims_tile_dims_tile_s
 //      CHECK:     %[[PACK:.+]] = tensor.pack %[[ARG0]] outer_dims_perm = [2, 1, 3, 0] inner_dims_pos = [3, 1, 2] inner_tiles = [%[[ARG3]], %[[ARG1]], %[[ARG2]]] into %[[INIT]] : tensor<?x?x?x?xf32> -> tensor<?x?x?x?x?x?x?xf32>
 //      CHECK:     return %[[PACK]] : tensor<?x?x?x?x?x?x?xf32>
 //      CHECK:   }
+
+// -----
+
+func.func @linalg_transpose_tensor_pack_fold(%arg0: tensor<56x57x1x64xf32>) -> tensor<1x57x56x2x32xf32> {
+  %0 = tensor.empty() : tensor<1x56x57x64xf32>
+  %transposed = linalg.transpose
+    ins(%arg0 : tensor<56x57x1x64xf32>)
+    outs(%0 : tensor<1x56x57x64xf32>)
+    permutation = [2, 0, 1, 3]
+
+  %1 = tensor.empty() : tensor<1x57x56x2x32xf32>
+  %pack = tensor.pack %transposed
+    outer_dims_perm = [0, 2, 1, 3]
+    inner_dims_pos = [3]
+    inner_tiles = [32]
+    into %1 : tensor<1x56x57x64xf32> -> tensor<1x57x56x2x32xf32>
+  return %pack : tensor<1x57x56x2x32xf32>
+}
+//      CHECK: func @linalg_transpose_tensor_pack_fold(
+// CHECK-SAME:     %[[ARG0:.+]]: tensor<56x57x1x64xf32>)
+//      CHECK:   %[[INIT:.+]] = tensor.empty() : tensor<1x57x56x2x32xf32>
+//      CHECK:   %[[PACK:.+]] = tensor.pack %[[ARG0]]
+// CHECK-SAME:      outer_dims_perm = [2, 1, 0, 3]
+// CHECK-SAME:      inner_dims_pos = [3] inner_tiles = [32] 
+// CHECK-SAME:       into %[[INIT]]
+//      CHECK:   return %[[PACK]]
+
+// -----
+
+func.func @linalg_transpose_tensor_pack_fold_with_padding(%arg0: tensor<56x57x1x55xf32>, %padding: f32) -> tensor<1x57x56x2x32xf32> {
+  %0 = tensor.empty() : tensor<1x56x57x55xf32>
+  %transpose = linalg.transpose
+    ins(%arg0 : tensor<56x57x1x55xf32>)
+    outs(%0 : tensor<1x56x57x55xf32>)
+    permutation = [2, 0, 1, 3]
+  
+  %1 = tensor.empty() : tensor<1x57x56x2x32xf32>
+  %pack = tensor.pack %transpose padding_value(%padding : f32)
+    outer_dims_perm = [0, 2, 1, 3]
+    inner_dims_pos = [3]
+    inner_tiles = [32]
+    into %1 : tensor<1x56x57x55xf32> -> tensor<1x57x56x2x32xf32>
+  return %pack : tensor<1x57x56x2x32xf32>
+}
+//      CHECK: func @linalg_transpose_tensor_pack_fold_with_padding(
+// CHECK-SAME:     %[[ARG0:.+]]: tensor<56x57x1x55xf32>, %[[PADDING:.+]]: f32)
+//      CHECK:   %[[INIT:.+]] = tensor.empty() : tensor<1x57x56x2x32xf32>
+//      CHECK:   %[[PACK:.+]] = tensor.pack %[[ARG0]] padding_value(%[[PADDING]] : f32)
+// CHECK-SAME:      outer_dims_perm = [2, 1, 0, 3]
+// CHECK-SAME:      inner_dims_pos = [3] inner_tiles = [32] 
+// CHECK-SAME:       into %[[INIT]]
+//      CHECK:   return %[[PACK]]
+
+// -----
+
+func.func @linalg_transpose_tensor_pack_fold_no_outer_dims_perm(%arg0: tensor<56x57x1x64xf32>) -> tensor<1x56x57x2x32xf32> {
+  %0 = tensor.empty() : tensor<1x56x57x64xf32>
+  %transposed = linalg.transpose
+    ins(%arg0 : tensor<56x57x1x64xf32>)
+    outs(%0 : tensor<1x56x57x64xf32>)
+    permutation = [2, 0, 1, 3]
+  
+  %1 = tensor.empty() : tensor<1x56x57x2x32xf32>
+  %pack = tensor.pack %transposed
+    inner_dims_pos = [3]
+    inner_tiles = [32]
+    into %1 : tensor<1x56x57x64xf32> -> tensor<1x56x57x2x32xf32>
+  return %pack : tensor<1x56x57x2x32xf32>
+}
+//      CHECK: func @linalg_transpose_tensor_pack_fold_no_outer_dims_perm(
+// CHECK-SAME:     %[[ARG0:.+]]: tensor<56x57x1x64xf32>)
+//      CHECK:   %[[INIT:.+]] = tensor.empty() : tensor<1x56x57x2x32xf32>
+//      CHECK:   %[[PACK:.+]] = tensor.pack %[[ARG0]]
+// CHECK-SAME:      outer_dims_perm = [2, 0, 1, 3]
+// CHECK-SAME:      inner_dims_pos = [3] inner_tiles = [32] 
+// CHECK-SAME:       into %[[INIT]]
+//      CHECK:   return %[[PACK]]
+
+// -----
+
+func.func @linalg_transpose_tensor_pack_fold_complex_inner_dims_change(%arg0: tensor<25x30x35x40xf32>, %transpose_dest: tensor<35x40x25x30xf32>, %pack_dest: tensor<3x35x5x8x5x10x5xf32>) -> tensor<3x35x5x8x5x10x5xf32> {
+  %transposed = linalg.transpose
+    ins(%arg0 : tensor<25x30x35x40xf32>)
+    outs(%transpose_dest : tensor<35x40x25x30xf32>)
+    permutation = [2, 3, 0, 1]
+  
+  %pack = tensor.pack %transposed
+    outer_dims_perm = [3, 0, 2, 1]
+    inner_dims_pos = [1, 3, 2]
+    inner_tiles = [5, 10, 5]
+    into %pack_dest : tensor<35x40x25x30xf32> -> tensor<3x35x5x8x5x10x5xf32>
+  return %pack : tensor<3x35x5x8x5x10x5xf32>
+}
+//      CHECK: module {
+//      CHECK:   func.func @linalg_transpose_tensor_pack_fold_complex_inner_dims_change(
+// CHECK-SAME:     %[[ARG0:.+]]: tensor<25x30x35x40xf32>, 
+// CHECK-SAME:     %[[ARG1:.+]]: tensor<35x40x25x30xf32>, 
+// CHECK-SAME:     %[[ARG2:.+]]: tensor<3x35x5x8x5x10x5xf32>) -> tensor<3x35x5x8x5x10x5xf32> {
+//      CHECK:     %[[VAL0:.+]] = tensor.empty() : tensor<3x35x5x8x5x10x5xf32>
+//      CHECK:     %[[PACK:.+]] = tensor.pack %[[ARG0]] 
+// CHECK-SAME:        outer_dims_perm = [1, 2, 0, 3] 
+// CHECK-SAME:        inner_dims_pos = [3, 1, 0] 
+// CHECK-SAME:        inner_tiles = [5, 10, 5] 
+// CHECK-SAME:         into %[[VAL0]] 
+//      CHECK:     return %[[PACK]]
+//      CHECK:   }
----------------
chelini wrote:

Also this last two checks can be removed.

https://github.com/llvm/llvm-project/pull/75658