Merge pull request #509 from Xilinx/bump_to_9190e1c0

[AutoBump] Merge with fixes of 9190e1c (Jan 09) (39) [Only tested MLIR]

Author: jorickert

mlir/lib/Conversion/TosaToLinalg/TosaToLinalg.cpp

@@ -650,30 +650,33 @@ static Value createLinalgBodyCalculationForElementwiseOp(
 static Value expandRank(PatternRewriter &rewriter, Location loc, Value tensor,
                         int64_t rank) {
   // No need to expand if we are already at the desired rank
-  auto shapedType = dyn_cast<ShapedType>(tensor.getType());
-  assert(shapedType && shapedType.hasRank() && "expected a ranked shaped type");
-  int64_t numExtraDims = rank - shapedType.getRank();
+  auto tensorType = dyn_cast<RankedTensorType>(tensor.getType());
+  assert(tensorType && "expected a ranked tensor type");
+  int64_t tensorRank = tensorType.getRank();
+  int64_t numExtraDims = rank - tensorRank;
   assert(numExtraDims >= 0 && "cannot expand tensor to a lower rank");
   if (!numExtraDims)
     return tensor;
 
   // Compute reassociation indices
-  SmallVector<SmallVector<int64_t, 2>> reassociationIndices(
-      shapedType.getRank());
+  SmallVector<ReassociationIndices> reassociationIndices(tensorRank);
   int64_t index = 0;
-  for (index = 0; index <= numExtraDims; index++)
-    reassociationIndices[0].push_back(index);
-  for (size_t position = 1; position < reassociationIndices.size(); position++)
-    reassociationIndices[position].push_back(index++);
+  if (tensorRank != 0) {
+    for (index = 0; index <= numExtraDims; index++)
+      reassociationIndices[0].push_back(index);
+    for (size_t position = 1; position < reassociationIndices.size();
+         position++)
+      reassociationIndices[position].push_back(index++);
+  }
 
   // Compute result type
   SmallVector<int64_t> resultShape;
   for (index = 0; index < numExtraDims; index++)
     resultShape.push_back(1);
-  for (auto size : shapedType.getShape())
+  for (auto size : tensorType.getShape())
     resultShape.push_back(size);
   auto resultType =
-      RankedTensorType::get(resultShape, shapedType.getElementType());
+      RankedTensorType::get(resultShape, tensorType.getElementType());
 
   // Emit 'tensor.expand_shape' op
   return rewriter.create<tensor::ExpandShapeOp>(loc, resultType, tensor,

mlir/test/Conversion/TosaToLinalg/tosa-to-linalg.mlir

@@ -120,6 +120,28 @@ func.func @test_add_uint8(%arg0: tensor<ui8>, %arg1: tensor<ui8>) -> tensor<ui8>
 
 // -----
 
+// CHECK: #[[$MAP0:.+]] = affine_map<(d0, d1) -> (d0, d1)>
+// CHECK: #[[$MAP1:.+]] = affine_map<(d0, d1) -> (0, d1)>
+
+// CHECK-LABEL:   func.func @test_add_0d_broadcast(
+// CHECK-SAME:                                     %[[ARG0:.*]]: tensor<2x1xf32>,
+// CHECK-SAME:                                     %[[ARG1:.*]]: tensor<f32>) -> tensor<2x1xf32> {
+// CHECK:           %[[EXPANDED:.*]] = tensor.expand_shape %[[ARG1]] [] output_shape [1, 1] : tensor<f32> into tensor<1x1xf32>
+// CHECK:           %[[EMPTY_TENSOR:.*]] = tensor.empty() : tensor<2x1xf32>
+// CHECK:           %[[RESULT:.*]] = linalg.generic {indexing_maps = [#[[$MAP0]], #[[$MAP1]], #[[$MAP0]]], iterator_types = ["parallel", "parallel"]} ins(%[[ARG0]], %[[EXPANDED]] : tensor<2x1xf32>, tensor<1x1xf32>) outs(%[[EMPTY_TENSOR]] : tensor<2x1xf32>) {
+// CHECK:           ^bb0(%[[IN0:.*]]: f32, %[[IN1:.*]]: f32, %[[OUT:.*]]: f32):
+// CHECK:             %[[ADD:.*]] = arith.addf %[[IN0]], %[[IN1]] : f32
+// CHECK:             linalg.yield %[[ADD]] : f32
+// CHECK:           } -> tensor<2x1xf32>
+// CHECK:           return %[[RESULT]] : tensor<2x1xf32>
+// CHECK:         }
+func.func @test_add_0d_broadcast(%arg0: tensor<2x1xf32>, %arg1: tensor<f32>) -> tensor<2x1xf32> {
+  %0 = tosa.add %arg0, %arg1 : (tensor<2x1xf32>, tensor<f32>) -> tensor<2x1xf32>
+  return %0 : tensor<2x1xf32>
+}
+
+// -----
+
 // CHECK: #[[$MAP0:.+]] = affine_map<(d0) -> (0)>
 // CHECK: #[[$MAP1:.+]] = affine_map<(d0) -> (d0)>
 // CHECK-LABEL: @test_add_1d_all_dynamic