add check for qnn tensor dynamic shape

Author: edgchen1

onnxruntime/core/providers/qnn/builder/qnn_def.cc

@@ -394,6 +394,20 @@ const Qnn_QuantizeParams_t& GetQnnTensorQParams(const Qnn_Tensor_t& qnn_tensor)
   ORT_THROW("QNN tensor version not supported, QNN tensor version: ", qnn_tensor.version);
 }
 
+uint8_t* GetQnnTensorIsDynamicDimensions(const Qnn_Tensor_t& qnn_tensor) {
+  if (QNN_TENSOR_VERSION_1 == qnn_tensor.version) {
+    return nullptr;  // not present in v1
+  }
+
+#ifdef QNN_TENSOR_V2_INIT
+  if (QNN_TENSOR_VERSION_2 == qnn_tensor.version) {
+    return qnn_tensor.v2.isDynamicDimensions;
+  }
+#endif  // QNN_TENSOR_V2_INIT
+
+  ORT_THROW("QNN tensor version not supported, QNN tensor version: ", qnn_tensor.version);
+}
+
 Status CompareQnnQuantParams(const Qnn_QuantizeParams_t& qparam0, const Qnn_QuantizeParams_t& qparam1,
                              float& scale_diff, int32_t& offset_diff) {
   scale_diff = 0.0f;

onnxruntime/core/providers/qnn/builder/qnn_def.h

@@ -126,6 +126,7 @@ uint32_t* GetQnnTensorDims(const Qnn_Tensor_t& qnn_tensor);
 const Qnn_ClientBuffer_t& GetQnnTensorClientBuf(const Qnn_Tensor_t& qnn_tensor);
 Qnn_MemHandle_t GetQnnTensorMemHandle(const Qnn_Tensor_t& qnn_tensor);
 const Qnn_QuantizeParams_t& GetQnnTensorQParams(const Qnn_Tensor_t& qnn_tensor);
+uint8_t* GetQnnTensorIsDynamicDimensions(const Qnn_Tensor_t& qnn_tensor);
 
 /**
  * Compares two sets of quantization parameters. Sets the parameters `scale_diff` and `offset_diff`

onnxruntime/core/providers/qnn/builder/qnn_model.cc

@@ -322,6 +322,9 @@ Status QnnModel::SetupTensors(std::vector<QnnTensorInfo>& qnn_tensor_infos,
   qnn_tensor_infos.resize(tensor_count);
 
   for (auto& tensor_wrapper : tensor_wrappers) {
+    ORT_RETURN_IF(utils::QnnTensorHasDynamicShape(tensor_wrapper.GetQnnTensor()),
+                  "QNN tensor (", tensor_wrapper.GetName(), ") has dynamic shape. This is not supported yet.");
+
     const size_t length = utils::GetQnnTensorDataSizeInBytes(tensor_wrapper.GetTensorDims(),
                                                              tensor_wrapper.GetTensorDataType());
     const auto& tensor_name = tensor_wrapper.GetName();

onnxruntime/core/providers/qnn/builder/qnn_utils.cc

@@ -3,6 +3,7 @@
 
 #include "core/providers/qnn/builder/qnn_utils.h"
 
+#include <algorithm>
 #include <functional>
 #include <map>
 #include <numeric>
@@ -71,6 +72,17 @@ size_t GetQnnTensorDataSizeInBytes(gsl::span<const uint32_t> shape, Qnn_DataType
   return std::accumulate(shape.begin(), shape.end(), data_length, std::multiplies<>{});
 }
 
+bool QnnTensorHasDynamicShape(const Qnn_Tensor_t& tensor) {
+  const uint8_t* is_dynamic_dimensions = GetQnnTensorIsDynamicDimensions(tensor);
+  if (is_dynamic_dimensions == nullptr) {
+    return false;
+  }
+
+  const auto rank = GetQnnTensorRank(tensor);
+  return std::any_of(is_dynamic_dimensions, is_dynamic_dimensions + rank,
+                     [](uint8_t is_dynamic_dimension) { return is_dynamic_dimension != 0; });
+}
+
 std::ostream& operator<<(std::ostream& out, const Qnn_Scalar_t& scalar) {
   switch (scalar.dataType) {
     case QNN_DATATYPE_INT_8:

onnxruntime/core/providers/qnn/builder/qnn_utils.h

@@ -29,6 +29,8 @@ size_t GetElementSizeByType(ONNXTensorElementDataType elem_type);
 
 size_t GetQnnTensorDataSizeInBytes(gsl::span<const uint32_t> shape, Qnn_DataType_t element_data_type);
 
+bool QnnTensorHasDynamicShape(const Qnn_Tensor_t& tensor);
+
 // TODO: make these work with Wrappers?
 std::ostream& operator<<(std::ostream& out, const Qnn_Param_t& qnn_param);
 std::ostream& operator<<(std::ostream& out, const Qnn_Tensor_t& tensor);