[MLIR] Specify new padOp's output type in DropPadUnitDims (#150706)

Previously when dropping unit dim from a pad with mixed dynamic/static
input/output shapes, the resulting shape would take on the Type of the
input, resulting in invalid IR.

Also did some minor cleanup to the formatting of the
`drop_unit_dim_corresponding_to_dynamic_dim` test to make it match the
rest of the file.

---------

Signed-off-by: dan <[email protected]>

Author: dan-garvey

mlir/lib/Dialect/Linalg/Transforms/DropUnitDims.cpp

@@ -637,6 +637,7 @@ struct DropPadUnitDims : public OpRewritePattern<tensor::PadOp> {
     }
 
     ArrayRef<int64_t> sourceShape = padOp.getSourceType().getShape();
+    ArrayRef<int64_t> resultShape = padOp.getResultType().getShape();
     int64_t padRank = sourceShape.size();
 
     auto isStaticZero = [](OpFoldResult f) {
@@ -647,16 +648,18 @@ struct DropPadUnitDims : public OpRewritePattern<tensor::PadOp> {
                                                  allowedUnitDims.end());
     llvm::SmallDenseSet<unsigned> unitDims;
     SmallVector<int64_t> newShape;
+    SmallVector<int64_t> newResultShape;
     SmallVector<OpFoldResult> newLowPad;
     SmallVector<OpFoldResult> newHighPad;
-    for (const auto [dim, size, low, high] :
-         zip_equal(llvm::seq(static_cast<int64_t>(0), padRank), sourceShape,
-                   padOp.getMixedLowPad(), padOp.getMixedHighPad())) {
+    for (const auto [dim, size, outSize, low, high] : zip_equal(
+             llvm::seq(static_cast<int64_t>(0), padRank), sourceShape,
+             resultShape, padOp.getMixedLowPad(), padOp.getMixedHighPad())) {
       if (unitDimsFilter.contains(dim) && size == 1 && isStaticZero(low) &&
           isStaticZero(high)) {
         unitDims.insert(dim);
       } else {
         newShape.push_back(size);
+        newResultShape.push_back(outSize);
         newLowPad.push_back(low);
         newHighPad.push_back(high);
       }
@@ -686,8 +689,10 @@ struct DropPadUnitDims : public OpRewritePattern<tensor::PadOp> {
         collapseValue(rewriter, padOp.getLoc(), padOp.getSource(), newShape,
                       reassociationMap, options.rankReductionStrategy);
 
-    auto newPadOp = tensor::PadOp::create(
-        rewriter, padOp.getLoc(), /*result=*/Type(), collapsedSource, newLowPad,
+    auto newResultType = RankedTensorType::get(
+        newResultShape, padOp.getResultType().getElementType());
+    auto newPadOp = rewriter.create<tensor::PadOp>(
+        padOp.getLoc(), /*result=*/newResultType, collapsedSource, newLowPad,
         newHighPad, paddingVal, padOp.getNofold());
 
     Value dest = padOp.getResult();

mlir/test/Dialect/Linalg/drop-unit-extent-dims.mlir

@@ -1076,6 +1076,44 @@ func.func @drop_known_unit_constant_low_high(%arg0: tensor<1x383x128xf32>) -> te
 
 // -----
 
+func.func @drop_unit_dim_mixed_static_dynamic(%arg0: tensor<1x?xf32>) -> tensor<1x16xf32> {
+  %c0 = arith.constant 0 : index
+  %c1 = arith.constant 1 : index
+  %cst = arith.constant 0.000000e+00 : f32
+  %padded = tensor.pad %arg0 low[%c0, %c1] high[%c0, %c0] {
+  ^bb0(%arg1: index, %arg2: index):
+    tensor.yield %cst : f32
+  } : tensor<1x?xf32> to tensor<1x16xf32>
+  return %padded : tensor<1x16xf32>
+}
+// CHECK-LABEL: func @drop_unit_dim_mixed_static_dynamic
+//       CHECK:   %[[CST:.*]] = arith.constant 0.000000e+00 : f32
+//       CHECK:   %[[COLLAPSE:.+]] = tensor.collapse_shape %[[ARGS:.*]] : tensor<1x?xf32> into tensor<?xf32>
+//       CHECK:   %[[PADDED:.*]] = tensor.pad %[[COLLAPSE]] low[1] high[0] {
+//       CHECK:   ^bb0(%[[IDX:.*]]: index):
+//       CHECK:     tensor.yield %[[CST]] : f32
+//       CHECK:   } : tensor<?xf32> to tensor<16xf32>
+//       CHECK:   %[[EXPANDED:.*]] = tensor.expand_shape %[[PADDED]] {{\[\[}}0, 1]] output_shape [1, 16] : tensor<16xf32> into tensor<1x16xf32>
+//       CHECK:   return %[[EXPANDED]] : tensor<1x16xf32>
+
+// -----
+
+#map = affine_map<(d0, d1, d2, d3, d4, d5, d6) -> (d0, d1 + d4, d2 + d5, d6)>
+#map1 = affine_map<(d0, d1, d2, d3, d4, d5, d6) -> (d4, d5, d6, d3)>
+#map2 = affine_map<(d0, d1, d2, d3, d4, d5, d6) -> (d0, d1, d2, d3)>
+module {
+  func.func @drop_unit_dim_corresponding_to_dynamic_dim(%arg0: tensor<1x?x?x1xf32>, %arg1: index) -> tensor<?x1x61x1xf32> {
+    %cst = arith.constant dense<1.000000e+00> : tensor<1x1x1x1xf32>
+    %0 = tensor.empty(%arg1) : tensor<?x1x61x1xf32>
+    %1 = linalg.generic {indexing_maps = [#map, #map1, #map2], iterator_types = ["parallel", "parallel", "parallel", "parallel", "reduction", "reduction", "reduction"]} ins(%arg0, %cst : tensor<1x?x?x1xf32>, tensor<1x1x1x1xf32>) outs(%0 : tensor<?x1x61x1xf32>) {
+    ^bb0(%in: f32, %in_0: f32, %out: f32):
+      %2 = arith.mulf %in, %in_0 : f32
+      %3 = arith.addf %out, %2 : f32
+      linalg.yield %3 : f32
+    } -> tensor<?x1x61x1xf32>
+    return %1 : tensor<?x1x61x1xf32>
+  }
+}
 // CHECK: #[[$MAP1:.+]] = affine_map<(d0) -> (0, d0)>
 // CHECK: #[[$MAP2:.+]] = affine_map<(d0) -> ()>
 
@@ -1097,23 +1135,6 @@ func.func @drop_known_unit_constant_low_high(%arg0: tensor<1x383x128xf32>) -> te
 // CHECK:           return %[[VAL_14]] : tensor<?x1x61x1xf32>
 // CHECK:         }
 
-#map = affine_map<(d0, d1, d2, d3, d4, d5, d6) -> (d0, d1 + d4, d2 + d5, d6)>
-#map1 = affine_map<(d0, d1, d2, d3, d4, d5, d6) -> (d4, d5, d6, d3)>
-#map2 = affine_map<(d0, d1, d2, d3, d4, d5, d6) -> (d0, d1, d2, d3)>
-module {
-  func.func @drop_unit_dim_corresponding_to_dynamic_dim(%arg0: tensor<1x?x?x1xf32>, %arg1: index) -> tensor<?x1x61x1xf32> {
-    %cst = arith.constant dense<1.000000e+00> : tensor<1x1x1x1xf32>
-    %0 = tensor.empty(%arg1) : tensor<?x1x61x1xf32>
-    %1 = linalg.generic {indexing_maps = [#map, #map1, #map2], iterator_types = ["parallel", "parallel", "parallel", "parallel", "reduction", "reduction", "reduction"]} ins(%arg0, %cst : tensor<1x?x?x1xf32>, tensor<1x1x1x1xf32>) outs(%0 : tensor<?x1x61x1xf32>) {
-    ^bb0(%in: f32, %in_0: f32, %out: f32):
-      %2 = arith.mulf %in, %in_0 : f32
-      %3 = arith.addf %out, %2 : f32
-      linalg.yield %3 : f32
-    } -> tensor<?x1x61x1xf32>
-    return %1 : tensor<?x1x61x1xf32>
-  }
-}
-
 // -----
 
 func.func @no_fold_empty_tensor_dim_out_of_bounds(%arg0: tensor<1x?x10xf32>) -> tensor<1x?xf32> {