Merge pull request #33 from Xilinx/christopher.FXML-1991_update_add_and_sub_folding

[FXML-1991] Update Add and Sub Constant Folding

Author: cmcgirr-amd

mlir/lib/Dialect/Tosa/IR/TosaCanonicalizations.cpp

@@ -474,8 +474,6 @@ OpFoldResult AddOp::fold(ArrayRef<Attribute> operands) {
   auto resultTy = getType().dyn_cast<RankedTensorType>();
   if (!lhsTy || !rhsTy || !resultTy)
     return {};
-  if (lhsTy != rhsTy)
-    return {};
 
   auto resultETy = resultTy.getElementType();
   auto lhsAttr = operands[0].dyn_cast_or_null<DenseElementsAttr>();
@@ -504,6 +502,9 @@ OpFoldResult AddOp::fold(ArrayRef<Attribute> operands) {
   if (!lhsAttr || !rhsAttr)
     return {};
 
+  if (lhsTy != rhsTy)
+    return {};
+
   return binaryFolder<std::plus<APInt>, std::plus<APFloat>>(lhsAttr, rhsAttr,
                                                             lhsTy);
 }
@@ -635,8 +636,6 @@ OpFoldResult SubOp::fold(ArrayRef<Attribute> operands) {
   auto resultTy = getType().dyn_cast<RankedTensorType>();
   if (!lhsTy || !rhsTy || !resultTy)
     return {};
-  if (lhsTy != rhsTy)
-    return {};
 
   auto resultETy = resultTy.getElementType();
   auto lhsAttr = operands[0].dyn_cast_or_null<DenseElementsAttr>();
@@ -655,6 +654,9 @@ OpFoldResult SubOp::fold(ArrayRef<Attribute> operands) {
   if (!lhsAttr || !rhsAttr)
     return {};
 
+  if (lhsTy != rhsTy)
+    return {};
+
   return binaryFolder<std::minus<APInt>, std::minus<APFloat>>(lhsAttr, rhsAttr,
                                                               lhsTy);
 }

mlir/test/Dialect/Tosa/constant-op-fold.mlir

@@ -164,6 +164,28 @@ func.func @fold_add_splat_f32() -> tensor<10xf32> {
 
 // -----
 
+// CHECK-LABEL: @fold_add_zero_splat_different_shape_f32
+func.func @fold_add_zero_splat_different_shape_f32(%arg0: tensor<1x10xf32>) -> tensor<1x10xf32> {
+  %zero = "tosa.const"() {value = dense<0.0> : tensor<1x1xf32>} : () -> tensor<1x1xf32>
+  %add = "tosa.add"(%arg0, %zero) : (tensor<1x10xf32>, tensor<1x1xf32>) -> tensor<1x10xf32>
+  // CHECK: return %arg0
+  return %add : tensor<1x10xf32>
+}
+
+// -----
+
+// CHECK-LABEL: @fold_add_zero_broadcast_arg_f32
+func.func @fold_add_zero_broadcast_arg_f32(%arg0: tensor<1x10xf32>) -> tensor<4x10xf32> {
+  %zero = "tosa.const"() {value = dense<0.0> : tensor<1x1xf32>} : () -> tensor<4x10xf32>
+  %add = "tosa.add"(%arg0, %zero) : (tensor<1x10xf32>, tensor<4x10xf32>) -> tensor<4x10xf32>
+  // CHECK: %[[ZERO:.+]] = "tosa.const"() {value = dense<0.000000e+00> : tensor<1x1xf32>} : () -> tensor<4x10xf32>
+  // CHECK: %[[ADD:.+]] = "tosa.add"(%arg0, %[[ZERO]]) : (tensor<1x10xf32>, tensor<4x10xf32>) -> tensor<4x10xf32>
+  // CHECK: return %[[ADD]] : tensor<4x10xf32>
+  return %add : tensor<4x10xf32>
+}
+
+// -----
+
 // CHECK-LABEL: @fold_div_zero_lhs_i32
 func.func @fold_div_zero_lhs_i32(%arg0: tensor<i32>) -> tensor<i32> {
   %zero = "tosa.const"() {value = dense<0> : tensor<i32>} : () -> tensor<i32>
@@ -350,6 +372,16 @@ func.func @fold_sub_splat_f32() -> tensor<10xf32> {
 
 // -----
 
+// CHECK-LABEL: @fold_sub_zero_splat_different_shape_f32
+func.func @fold_sub_zero_splat_different_shape_f32(%arg0: tensor<1x10xf32>) -> tensor<1x10xf32> {
+  %zero = "tosa.const"() {value = dense<0.0> : tensor<1x1xf32>} : () -> tensor<1x1xf32>
+  %sub = "tosa.sub"(%arg0, %zero) : (tensor<1x10xf32>, tensor<1x1xf32>) -> tensor<1x10xf32>
+  // CHECK: return %arg0
+  return %sub : tensor<1x10xf32>
+}
+
+// -----
+
 // CHECK-LABEL: @fold_greater_splat_f32
 func.func @fold_greater_splat_f32() -> (tensor<10xi1>, tensor<10xi1>) {
   %0 = "tosa.const"() {value = dense<4.0> : tensor<10xf32>} : () -> tensor<10xf32>