Update DML EP to accept broadcasted tensor of size 1 to match CPU (#1…

…9081) ### Description With QDQ enabled for Dml EP we are seeing some models not optimize constant nodes with incorrect tensor size of scale[1] and zeropoint[1] that does not match the input size. CPU accepts this parameter type so updating Dml EP to match CPU behavior. ### Motivation and Context Want to match CPU EP behavior. --------- Co-authored-by: Christian Larson <[email protected]> Co-authored-by: Dwayne Robinson <[email protected]>
microsoft · Jan 11, 2024 · 8a0a972 · 8a0a972
1 parent daa22f9
commit 8a0a972
Show file tree

Hide file tree

Showing 3 changed files with 25 additions and 1 deletion.
diff --git a/onnxruntime/core/providers/dml/DmlExecutionProvider/src/Operators/DmlOperatorElementWise.cpp b/onnxruntime/core/providers/dml/DmlExecutionProvider/src/Operators/DmlOperatorElementWise.cpp
@@ -558,7 +558,11 @@ class DmlOperatorElementwiseQLinear : public DmlOperator
             {
                 ML_CHECK_VALID_ARGUMENT(axis < outputShapeDimCount);
                 uint32_t broadcastAxisLength = outputShape[axis];
-                ML_CHECK_VALID_ARGUMENT(inputTensorShape[0] == broadcastAxisLength);
+                ML_CHECK_VALID_ARGUMENT(
+                    (inputTensorShape[0] == broadcastAxisLength) ||
+                    // Treat as broadcast dimension to match CPU behavior.
+                    (inputTensorShape[0] == 1)
+                );
                 inputTensorShape.insert(inputTensorShape.begin(), axis, 1);
                 inputTensorShape.insert(inputTensorShape.end(), outputShapeDimCount - 1 - axis, 1);
             }

diff --git a/onnxruntime/test/contrib_ops/quantize_ops_test.cc b/onnxruntime/test/contrib_ops/quantize_ops_test.cc
@@ -76,6 +76,16 @@ TEST(DequantizeLinearOpTest, DequantizeLinear_per_tensor_float_int32_cpu) {
   test.Run();
 }
 
+TEST(DequantizeLinearOpTest, DequantizeLinearOpTest_BroadcastTensorOfOne) {
+  OpTester test("DequantizeLinear", 1, onnxruntime::kMSDomain);
+
+  test.AddInput<int32_t>("x", {4}, {-30, -3, 100, 127});
+  test.AddInput<float>("x_scale", {1}, {2.0f}, true);
+  test.AddInput<int32_t>("zero_point", {1}, {0}, true);
+  test.AddOutput<float>("y", {4}, {-60.f, -6.f, 200.f, 254.f});
+  test.Run();
+}
+
 #ifdef USE_CUDA
 TEST(DequantizeLinearOpTest, DequantizeLinear_per_tensor_half_uint8) {
   OpTester test("DequantizeLinear", 1, onnxruntime::kMSDomain);

diff --git a/onnxruntime/test/providers/cpu/tensor/quantize_linear_test.cc b/onnxruntime/test/providers/cpu/tensor/quantize_linear_test.cc
@@ -47,6 +47,16 @@ TEST(DequantizeLinearOpTest, Int32) {
   test.Run();
 }
 
+TEST(DequantizeLinearOpTest_BroadcastTensor, Int32) {
+  OpTester test("DequantizeLinear", 13);
+  test.AddInput<int32_t>("x", {4}, {-30, -3, 100, 127});
+  test.AddAttribute<int64_t>("axis", 0);
+  test.AddInput<float>("x_scale", {1}, {2.0f});
+  test.AddInput<int32_t>("x_zero_point", {1}, {0});
+  test.AddOutput<float>("y", {4}, {-60.f, -6.f, 200.f, 254.f});
+  test.Run();
+}
+
 // 2d inputs
 TEST(DequantizeLinearOpTest, 2D) {
   OpTester test("DequantizeLinear", 10);