Add QNN UTs for QNN Pad Op with FP16 data on HTP backend (microsoft#21142)

### Description
1. Add QNN UTs for QNN Pad Op with FP16 data on HTP backend
2. Improve Pad op builder to handle invalid optional input
3. Add UT for ReduceSum for FP16 precision with 5D for issue reproduce

Author: HectorSVC

onnxruntime/core/providers/qnn/builder/opbuilder/pad_op_builder.cc

@@ -51,7 +51,7 @@ Status PadOpBuilder::ProcessInputs(QnnModelWrapper& qnn_model_wrapper,
     auto& pads_input_name = inputs[1].node_arg.Name();
     ORT_RETURN_IF_NOT(qnn_model_wrapper.IsInitializerInput(pads_input_name),
                       "Qnn doesn't support dynamic pad input");
-    if (node_unit.Inputs().size() > 2) {
+    if (inputs.size() > 2 && inputs[2].node_arg.Exists()) {
       auto& constant_value_input_name = inputs[2].node_arg.Name();
       ORT_RETURN_IF_NOT(qnn_model_wrapper.IsInitializerInput(constant_value_input_name),
                         "Qnn doesn't support dynamic constant_value input");
@@ -227,13 +227,13 @@ Status PadOpBuilder::ProcessAttributesAndOutputs(QnnModelWrapper& qnn_model_wrap
   param_tensor_names.push_back(mode_param.GetParamTensorName());
   qnn_model_wrapper.AddParamWrapper(std::move(mode_param));
 
-  QnnParamWrapper multiples_param(node_unit.Index(), node_unit.Name(), QNN_OP_PAD_PARAM_PAD_AMOUNT,
-                                  std::move(pad_amount_dim), std::move(pad_amount));
-  param_tensor_names.push_back(multiples_param.GetParamTensorName());
-  qnn_model_wrapper.AddParamWrapper(std::move(multiples_param));
+  QnnParamWrapper pad_amount_param(node_unit.Index(), node_unit.Name(), QNN_OP_PAD_PARAM_PAD_AMOUNT,
+                                   std::move(pad_amount_dim), std::move(pad_amount));
+  param_tensor_names.push_back(pad_amount_param.GetParamTensorName());
+  qnn_model_wrapper.AddParamWrapper(std::move(pad_amount_param));
 
   // Process optional input constant_value
-  if (node_unit.Inputs().size() > 2) {
+  if (inputs.size() > 2 && inputs[2].node_arg.Exists()) {
     ORT_RETURN_IF_ERROR(ProcessConstantValue(qnn_model_wrapper, param_tensor_names, node_unit, inputs[2]));
   }  // constant_value
 

onnxruntime/test/providers/qnn/pad_op_test.cpp

@@ -98,18 +98,33 @@ static void RunPadOpTest(const TestInputDef<float>& data_def,
                          const std::vector<ONNX_NAMESPACE::AttributeProto>& attrs,
                          ExpectedEPNodeAssignment expected_ep_assignment,
                          bool has_constant_value = true,
-                         int opset = 18) {
+                         int opset = 18,
+                         bool use_htp = false,
+                         bool enable_fp16_precision = false,
+                         float f32_abs_err = 1e-5f) {
   ProviderOptions provider_options;
+  if (use_htp) {
 #if defined(_WIN32)
-  provider_options["backend_path"] = "QnnCpu.dll";
+    provider_options["backend_path"] = "QnnHtp.dll";
 #else
-  provider_options["backend_path"] = "libQnnCpu.so";
+    provider_options["backend_path"] = "libQnnHtp.so";
 #endif
+  } else {
+#if defined(_WIN32)
+    provider_options["backend_path"] = "QnnCpu.dll";
+#else
+    provider_options["backend_path"] = "libQnnCpu.so";
+#endif
+  }
+
+  if (enable_fp16_precision) {
+    provider_options["enable_htp_fp16_precision"] = "1";
+  }
 
   RunQnnModelTest(BuildPadTestCase(data_def, pads_def, constant_value_def, attrs, has_constant_value),
                   provider_options,
                   opset,
-                  expected_ep_assignment);
+                  expected_ep_assignment, f32_abs_err);
 }
 
 // Runs a QDQ Pad model on the QNN HTP backend. Checks the graph node assignment, and that inference
@@ -229,6 +244,60 @@ TEST_F(QnnCPUBackendTests, Pad6d) {
 #if defined(__aarch64__) || defined(_M_ARM64) || defined(__linux__)
 //
 // HTP tests:
+TEST_F(QnnHTPBackendTests, PadNoConstantValue_fp16_test) {
+  bool has_constant_value_input = false;
+  bool use_htp = true;
+  bool enable_fp16_precision = true;
+  RunPadOpTest(TestInputDef<float>({3, 2}, false, {1.0f, 1.2f, 2.3f, 3.4f, 4.5f, 5.6f}),
+               TestInputDef<int64_t>({4}, true, {0, 2, 0, 0}),
+               TestInputDef<float>({1}, true, {0.0f}),
+               {utils::MakeAttribute("mode", "constant")},
+               ExpectedEPNodeAssignment::All,
+               has_constant_value_input,
+               18,  // opset
+               use_htp,
+               enable_fp16_precision,
+               2e-3f);
+}
+
+TEST_F(QnnHTPBackendTests, PadReflectMode_fp16) {
+  bool has_constant_value_input = false;
+  bool use_htp = true;
+  bool enable_fp16_precision = true;
+  RunPadOpTest(TestInputDef<float>({3, 2}, false, {1.0f, 1.2f, 2.3f, 3.4f, 4.5f, 5.6f}),
+               TestInputDef<int64_t>({4}, true, {0, 1, 0, 0}),
+               TestInputDef<float>({1}, true, {0.0f}),
+               {utils::MakeAttribute("mode", "reflect")},
+               ExpectedEPNodeAssignment::All,
+               has_constant_value_input,
+               18,  // opset
+               use_htp,
+               enable_fp16_precision,
+               2e-3f);
+}
+
+// HTP\HTP\src\hexagon\prepare\graph_prepare.cc:203:ERROR:could not create op: q::flat_from_vtcm
+// HTP\HTP\src\hexagon\prepare\graph_prepare.cc:1238:ERROR:Op 0x104100000011 preparation failed with err:-1
+// Completed stage: Graph Transformations and Optimizations (13372 us)
+// QnnDsp <E> "node" generated: could not create op
+// QnnDsp <E> RouterWindows graph prepare failed 12
+// QnnDsp <E> Failed to finalize graph (id: 1) with err 1002
+TEST_F(QnnHTPBackendTests, DISABLED_PadReflectMode_FP16_big_data) {
+  bool has_constant_value_input = false;
+  bool use_htp = true;
+  bool enable_fp16_precision = true;
+  RunPadOpTest(TestInputDef<float>({1, 4, 512, 512}, false, GetFloatDataInRange(1.0f, 10.0f, 4 * 512 * 512)),
+               TestInputDef<int64_t>({8}, true, {0, 0, 3, 3, 0, 0, 3, 3}),
+               TestInputDef<float>({1}, true, {0.0f}),
+               {utils::MakeAttribute("mode", "reflect")},
+               ExpectedEPNodeAssignment::All,
+               has_constant_value_input,
+               18,  // opset
+               use_htp,
+               enable_fp16_precision,
+               2e-3f);
+}
+
 //
 // QDQ Pad
 TEST_F(QnnHTPBackendTests, PadNoConstantValue) {