[QNN EP] Dump QNN json graph

cmake/onnxruntime_providers_qnn.cmake

@@ -30,7 +30,9 @@
 
   source_group(TREE ${ONNXRUNTIME_ROOT}/core FILES ${onnxruntime_providers_qnn_cc_srcs})
   onnxruntime_add_static_library(onnxruntime_providers_qnn ${onnxruntime_providers_qnn_cc_srcs})
-  onnxruntime_add_include_to_target(onnxruntime_providers_qnn onnxruntime_common onnxruntime_framework onnx onnx_proto protobuf::libprotobuf-lite flatbuffers::flatbuffers Boost::mp11)
+  onnxruntime_add_include_to_target(onnxruntime_providers_qnn onnxruntime_common onnxruntime_framework onnx onnx_proto
+	                            protobuf::libprotobuf-lite flatbuffers::flatbuffers Boost::mp11
+				    nlohmann_json::nlohmann_json)
   target_link_libraries(onnxruntime_providers_qnn)
   add_dependencies(onnxruntime_providers_qnn onnx ${onnxruntime_EXTERNAL_DEPENDENCIES})
   set_target_properties(onnxruntime_providers_qnn PROPERTIES CXX_STANDARD_REQUIRED ON)

include/onnxruntime/core/session/onnxruntime_c_api.h

@@ -3662,6 +3662,10 @@ struct OrtApi {
    *   execution provider (typically CPU EP).
    *     - "0": Default. Disabled. QNN EP will handle quantization and dequantization of graph I/O.
    *     - "1": Enabled.
+   *   "enable_qnn_graph_dump": Set to "1" to enable dumping of QNN graphs as JSON files. Each graph partition
+   *      assigned to QNN EP is dumped to a separate file.
+   *   "qnn_graph_dump_dir": Directory in which to dump QNN JSON graphs. If not specified, QNN graphs are dumped in the
+   *      program's current working directory. Ignored if "enable_json_graphs_dump" is not set.
    *
    * SNPE supported keys:
    *   "runtime": SNPE runtime engine, options: "CPU", "CPU_FLOAT32", "GPU", "GPU_FLOAT32_16_HYBRID", "GPU_FLOAT16",

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

@@ -388,6 +388,15 @@ Status CompareQnnQuantParams(const Qnn_QuantizeParams_t& qparam0, const Qnn_Quan
   return Status::OK();
 }
 
+uint32_t CalcQnnTensorNumElems(const Qnn_Tensor_t& qnn_tensor) {
+  uint32_t* qnn_tensor_dims = GetQnnTensorDims(qnn_tensor);
+  uint32_t qnn_tensor_rank = GetQnnTensorRank(qnn_tensor);
+  return std::accumulate(qnn_tensor_dims,
+                         qnn_tensor_dims + qnn_tensor_rank,
+                         1,
+                         std::multiplies<uint32_t>());
+}
+
 bool CreateTensorInQnnGraph(const QNN_INTERFACE_VER_TYPE& qnn_interface,
                             const Qnn_GraphHandle_t& graph,
                             const std::string& node_name,
@@ -422,12 +431,7 @@ bool CreateTensorInQnnGraph(const QNN_INTERFACE_VER_TYPE& qnn_interface,
       return false;
     }
     // verify size expressed by the dims matches the raw tensor size
-    auto qnn_tensor_dims = GetQnnTensorDims(qnn_tensor);
-    auto qnn_tensor_rank = GetQnnTensorRank(qnn_tensor);
-    uint32_t qnn_tensor_size = std::accumulate(qnn_tensor_dims,
-                                               qnn_tensor_dims + qnn_tensor_rank,
-                                               static_cast<uint32_t>(data_size),
-                                               std::multiplies<uint32_t>());
+    uint32_t qnn_tensor_size = CalcQnnTensorNumElems(qnn_tensor) * gsl::narrow_cast<uint32_t>(data_size);
     auto qnn_tensor_buf_size = GetQnnTensorClientBuf(qnn_tensor).dataSize;
     if (qnn_tensor_size != qnn_tensor_buf_size) {
       ss << "Data length mismatch for static tensor. node_name: " << node_name

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

@@ -122,6 +122,7 @@ Qnn_DataType_t GetQnnTensorDataType(const Qnn_Tensor_t& qnn_tensor);
 Qnn_TensorMemType_t GetQnnTensorMemType(const Qnn_Tensor_t& qnn_tensor);
 uint32_t GetQnnTensorRank(const Qnn_Tensor_t& qnn_tensor);
 uint32_t* GetQnnTensorDims(const Qnn_Tensor_t& qnn_tensor);
+uint32_t CalcQnnTensorNumElems(const Qnn_Tensor_t& qnn_tensor);
 const Qnn_ClientBuffer_t& GetQnnTensorClientBuf(const Qnn_Tensor_t& qnn_tensor);
 const Qnn_QuantizeParams_t& GetQnnTensorQParams(const Qnn_Tensor_t& qnn_tensor);
 

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

@@ -97,7 +97,8 @@ Status QnnModel::ComposeGraph(const GraphViewer& graph_viewer,
                               const onnxruntime::Node& fused_node,
                               const qnn::ModelSettings& model_settings,
                               const logging::Logger& logger,
-                              const QnnGraph_Config_t** graph_configs) {
+                              const QnnGraph_Config_t** graph_configs,
+                              const std::string& debug_json_graph_path) {
   LOGS(logger, VERBOSE) << "ComposeGraph Graph name: " << graph_viewer.Name();
 
   // Holder for the NodeUnits in the graph, this will guarantee the NodeUnits is
@@ -140,7 +141,20 @@ Status QnnModel::ComposeGraph(const GraphViewer& graph_viewer,
     }
   }
 
-  ORT_RETURN_IF_NOT(qnn_model_wrapper.ComposeQnnGraph(), "Failed to compose Qnn graph.");
+  const bool build_debug_json_graph = !debug_json_graph_path.empty();
+  ORT_RETURN_IF_NOT(qnn_model_wrapper.ComposeQnnGraph(build_debug_json_graph), "Failed to compose Qnn graph.");
+
+  if (build_debug_json_graph) {
+    const nlohmann::json& json_graph = qnn_model_wrapper.GetQnnJSONGraph();
+    std::ofstream ofs(debug_json_graph_path);
+
+    if (ofs.is_open()) {
+      ofs << json_graph.dump();
+      ofs.close();
+    } else {
+      LOGS(logger_, WARNING) << "Could not open JSON graph file: " << debug_json_graph_path;
+    }
+  }
 
   rt = GetGraphInfoFromModel(qnn_model_wrapper, logger);
   if (!rt) {

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

@@ -37,7 +37,8 @@ class QnnModel {
                       const onnxruntime::Node& fused_node,
                       const qnn::ModelSettings& model_settings,
                       const logging::Logger& logger,
-                      const QnnGraph_Config_t** graph_configs = nullptr);
+                      const QnnGraph_Config_t** graph_configs = nullptr,
+                      const std::string& debug_json_graph_path = "");
 
   Status FinalizeGraphs(const logging::Logger& logger);
 

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

@@ -252,7 +252,7 @@ bool QnnModelWrapper::CreateQnnNode(const std::string& qnn_node_name,
   }
 }
 
-bool QnnModelWrapper::ComposeQnnGraph() {
+bool QnnModelWrapper::ComposeQnnGraph(bool build_debug_json_graph) {
   LOGS(logger_, VERBOSE) << "Compose Qnn Graph.";
   // ORT_RETURN_IF(qnn_op_property_list_.empty(), "Empty Qnn op list, no graph to compose.");
   if (qnn_op_property_list_.empty()) {
@@ -291,6 +291,10 @@ bool QnnModelWrapper::ComposeQnnGraph() {
       LOGS(logger_, ERROR) << error_msg;
       return false;
     }
+
+    if (build_debug_json_graph) {
+      debug_json_graph_.AddOp(op_config_wrapper);
+    }
   }
 
   return true;

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

@@ -9,7 +9,8 @@
 
 #include "core/common/status.h"
 #include "QnnInterface.h"
-#include "qnn_def.h"
+#include "core/providers/qnn/builder/qnn_def.h"
+#include "core/providers/qnn/builder/qnn_utils.h"
 #include "core/common/logging/logging.h"
 #include "core/framework/node_unit.h"
 #include "core/graph/graph_viewer.h"
@@ -91,7 +92,7 @@ class QnnModelWrapper {
                      std::vector<std::string>&& param_tensor_names,
                      bool do_op_validation = false);
 
-  bool ComposeQnnGraph();
+  bool ComposeQnnGraph(bool build_debug_json_graph = false);
 
   Qnn_GraphHandle_t GetQnnGraph() { return graph_; }
 
@@ -127,6 +128,10 @@ class QnnModelWrapper {
     return input_index_map_.find(tensor_name) != input_index_map_.end();
   }
 
+  const nlohmann::json& GetQnnJSONGraph() {
+    return debug_json_graph_.Finalize();
+  }
+
   Qnn_TensorType_t GetTensorType(const std::string& tensor_name) const {
     if (IsInitializerInput(tensor_name)) {
       return QNN_TENSOR_TYPE_STATIC;
@@ -270,6 +275,7 @@ class QnnModelWrapper {
   const Qnn_BackendHandle_t& backend_handle_;
   Qnn_GraphHandle_t graph_ = nullptr;
   std::string graph_name_ = "";
+  utils::QnnJSONGraph debug_json_graph_;
 
   std::vector<std::string> model_input_names_;
   std::vector<std::string> model_output_names_;