如果模型中有些算子不被ONNX算子库支持,我们就需要利用ONNXRuntime提供的API手动添加新算子。在官方文档中已经对如何添加定制算子进行了介绍(https://onnxruntime.ai/docs/reference/operators/add-custom-op.html ),这里我们主要把源码中对应的流程给捋清楚。
添加定制算子(Custom Operators)主要分为三步:
首先看看源码中给出的定制算子样例:
// file path: onnxruntime/test/shared_lib/custom_op_utils.h// 首先定义定制算子的核struct MyCustomKernel { MyCustomKernel(Ort::CustomOpApi ort, const OrtKernelInfo* /*info*/, void* compute_stream) : ort_(ort), compute_stream_(compute_stream) { } void Compute(OrtKernelContext* context); private: Ort::CustomOpApi ort_; void* compute_stream_;};// 然后定义定制算子的各个操作,各个成员函数均已实现,其中 CreateKernel 会返回前面定义的算子核对象struct MyCustomOp : Ort::CustomOpBase<MyCustomOp, MyCustomKernel> { explicit MyCustomOp(const char* provider, void* compute_stream) : provider_(provider), compute_stream_(compute_stream) {} void* CreateKernel(Ort::CustomOpApi api, const OrtKernelInfo* info) const { return new MyCustomKernel(api, info, compute_stream_); }; const char* GetName() const { return "Foo"; }; const char* GetExecutionProviderType() const { return provider_; }; size_t GetInputTypeCount() const { return 2; }; ONNXTensorElementDataType GetInputType(size_t /*index*/) const { // Both the inputs need to be necessarily of float type return ONNX_TENSOR_ELEMENT_DATA_TYPE_FLOAT; }; size_t GetOutputTypeCount() const { return 1; }; ONNXTensorElementDataType GetOutputType(size_t /*index*/) const { return ONNX_TENSOR_ELEMENT_DATA_TYPE_FLOAT; }; private: const char* provider_; void* compute_stream_;};在上面代码中,我们看到定制算子继承自 Ort::CustomOpBase<MyCustomOp, MyCustomKernel>,这种扩展类作为模板基类的模板参数的方式又被称为CRTP,接着深入到这个模板类内部:
// file path: include/onnxruntime/core/session/onnxruntime_cxx_api.htemplate <typename TOp, typename TKernel>struct CustomOpBase : OrtCustomOp { CustomOpBase() { OrtCustomOp::version = ORT_API_VERSION; OrtCustomOp::CreateKernel = [](const OrtCustomOp* this_, const OrtApi* api, const OrtKernelInfo* info) { return static_cast<const TOp*>(this_)->CreateKernel(*api, info); }; OrtCustomOp::GetName = [](const OrtCustomOp* this_) { return static_cast<const TOp*>(this_)->GetName(); }; OrtCustomOp::GetExecutionProviderType = [](const OrtCustomOp* this_) { return static_cast<const TOp*>(this_)->GetExecutionProviderType(); }; OrtCustomOp::GetInputTypeCount = [](const OrtCustomOp* this_) { return static_cast<const TOp*>(this_)->GetInputTypeCount(); }; OrtCustomOp::GetInputType = [](const OrtCustomOp* this_, size_t index) { return static_cast<const TOp*>(this_)->GetInputType(index); }; OrtCustomOp::GetOutputTypeCount = [](const OrtCustomOp* this_) { return static_cast<const TOp*>(this_)->GetOutputTypeCount(); }; OrtCustomOp::GetOutputType = [](const OrtCustomOp* this_, size_t index) { return static_cast<const TOp*>(this_)->GetOutputType(index); }; OrtCustomOp::KernelCompute = [](void* op_kernel, OrtKernelContext* context) { static_cast<TKernel*>(op_kernel)->Compute(context); }; OrtCustomOp::KernelDestroy = [](void* op_kernel) { delete static_cast<TKernel*>(op_kernel); }; OrtCustomOp::GetInputCharacteristic = [](const OrtCustomOp* this_, size_t index) { return static_cast<const TOp*>(this_)->GetInputCharacteristic(index); }; OrtCustomOp::GetOutputCharacteristic = [](const OrtCustomOp* this_, size_t index) { return static_cast<const TOp*>(this_)->GetOutputCharacteristic(index); }; } // Default implementation of GetExecutionProviderType that returns nullptr to default to the CPU provider const char* GetExecutionProviderType() const { return nullptr; } // Default implementations of GetInputCharacteristic() and GetOutputCharacteristic() below // (inputs and outputs are required by default) OrtCustomOpInputOutputCharacteristic GetInputCharacteristic(size_t /*index*/) const { return OrtCustomOpInputOutputCharacteristic::INPUT_OUTPUT_REQUIRED; } OrtCustomOpInputOutputCharacteristic GetOutputCharacteristic(size_t /*index*/) const { return OrtCustomOpInputOutputCharacteristic::INPUT_OUTPUT_REQUIRED; }};这里的 CustomOpBase 又继承自 OrtCustomOp:
// include/onnxruntime/core/session/onnxruntime_c_api.hstruct OrtCustomOp;typedef struct OrtCustomOp OrtCustomOp;struct OrtCustomOp { uint32_t version; // Must be initialized to ORT_API_VERSION // This callback creates the kernel, which is a user defined parameter that is passed to the Kernel* callbacks below. void*(ORT_API_CALL* CreateKernel)(_In_ const struct OrtCustomOp* op, _In_ const OrtApi* api, _In_ const OrtKernelInfo* info); // Returns the name of the op const char*(ORT_API_CALL* GetName)(_In_ const struct OrtCustomOp* op); // Returns the type of the execution provider, return nullptr to use CPU execution provider const char*(ORT_API_CALL* GetExecutionProviderType)(_In_ const struct OrtCustomOp* op); // Returns the count and types of the input & output tensors ONNXTensorElementDataType(ORT_API_CALL* GetInputType)(_In_ const struct OrtCustomOp* op, _In_ size_t index); size_t(ORT_API_CALL* GetInputTypeCount)(_In_ const struct OrtCustomOp* op); ONNXTensorElementDataType(ORT_API_CALL* GetOutputType)(_In_ const struct OrtCustomOp* op, _In_ size_t index); size_t(ORT_API_CALL* GetOutputTypeCount)(_In_ const struct OrtCustomOp* op); // Op kernel callbacks void(ORT_API_CALL* KernelCompute)(_In_ void* op_kernel, _In_ OrtKernelContext* context); void(ORT_API_CALL* KernelDestroy)(_In_ void* op_kernel); // Returns the characteristics of the input & output tensors OrtCustomOpInputOutputCharacteristic(ORT_API_CALL* GetInputCharacteristic)(_In_ const struct OrtCustomOp* op, _In_ size_t index); OrtCustomOpInputOutputCharacteristic(ORT_API_CALL* GetOutputCharacteristic)(_In_ const struct OrtCustomOp* op, _In_ size_t index);};可以发现,OrtCustomOp 中定义了定制算子应该实现的模式,其中的一系列回调函数由其派生类一一实现,比如上文提到的 CustomOpBase 在其构造函数中,以 lambda 函数的方式实现各个回调函数。
至此,我们已经完整地梳理了定义定制算子在源码内部是如何实现的,接下来介绍如何将定义好的定制算子使用起来。
从如下官方测试代码开始分析:
// file path: onnxruntime/test/shared_lib/test_inference.ccTEST(CApiTest, custom_op_handler) { std::cout << "Running custom op inference" << std::endl; std::vector<Input> inputs(1); Input& input = inputs[0]; input.name = "X"; input.dims = {3, 2}; input.values = {1.0f, 2.0f, 3.0f, 4.0f, 5.0f, 6.0f}; // prepare expected inputs and outputs std::vector<int64_t> expected_dims_y = {3, 2}; std::vector<float> expected_values_y = {2.0f, 4.0f, 6.0f, 8.0f, 10.0f, 12.0f}; // 创建定制算子(MyCustomOp)#ifdef USE_CUDA cudaStream_t compute_stream = nullptr; // 声明一个 cuda stream cudaStreamCreateWithFlags(&compute_stream, cudaStreamNonBlocking); // 创建一个 cuda stream MyCustomOp custom_op{onnxruntime::kCudaExecutionProvider, compute_stream};#else MyCustomOp custom_op{onnxruntime::kCpuExecutionProvider, nullptr};#endif // 创建定制算子域(CustomOpDomain) Ort::CustomOpDomain custom_op_domain(""); // 在定制算子域中添加定制算子 custom_op_domain.Add(&custom_op); // 进入 TestInference#ifdef USE_CUDA TestInference<float>(*ort_env, CUSTOM_OP_MODEL_URI, inputs, "Y", expected_dims_y, expected_values_y, 1, custom_op_domain, nullptr, nullptr, false, compute_stream); cudaStreamDestroy(compute_stream);#else TestInference<float>(*ort_env, CUSTOM_OP_MODEL_URI, inputs, "Y", expected_dims_y, expected_values_y, 0, custom_op_domain, nullptr);#endif}以上代码需要特别注意的是,需要根据宏(USE_CUDA)用来判断是否使用CUDA。如果使用 CUDA:
这里创建 cuda stream 的方式是 cudaStreamCreateWithFlags,该函数和 cudaStreamCreate 不同,后者在多次调用时是串行方式执行,而前者可同步执行。如果将参数 cudaStreamNonBlocking 替换为 cudaStreamDefault,则 cudaStreamCreateWithFlags 的行为将和 cudaStreamCreate 相同。【参考内容:CUDA 5.0 中cudaStreamCreateWithFlags 的用法】
无论是否使用CDUA,我们都需要创建定制算子(MyCustomOp)。
进入 TestInference 函数内部:
// file path: onnxruntime/test/shared_lib/test_inference.cctemplate <typename OutT>static void TestInference(Ort::Env& env, const std::basic_string<ORTCHAR_T>& model_uri, const std::vector<Input>& inputs, const char* output_name, const std::vector<int64_t>& expected_dims_y, const std::vector<OutT>& expected_values_y, int provider_type, OrtCustomOpDomain* custom_op_domain_ptr, const char* custom_op_library_filename, void** library_handle = nullptr, bool test_session_creation_only = false, void* cuda_compute_stream = nullptr) { Ort::SessionOptions session_options; if (provider_type == 1) {#ifdef USE_CUDA std::cout << "Running simple inference with cuda provider" << std::endl; auto cuda_options = CreateDefaultOrtCudaProviderOptionsWithCustomStream(cuda_compute_stream); session_options.AppendExecutionProvider_CUDA(cuda_options);#else ORT_UNUSED_PARAMETER(cuda_compute_stream); return;#endif } else if (provider_type == 2) {#ifdef USE_DNNL Ort::ThrowOnError(OrtSessionOptionsAppendExecutionProvider_Dnnl(session_options, 1)); std::cout << "Running simple inference with dnnl provider" << std::endl;#else return;#endif } else if (provider_type == 3) {#ifdef USE_NUPHAR Ort::ThrowOnError(OrtSessionOptionsAppendExecutionProvider_Nuphar(session_options, /*allow_unaligned_buffers*/ 1, "")); std::cout << "Running simple inference with nuphar provider" << std::endl;#else return;#endif } else { std::cout << "Running simple inference with default provider" << std::endl; } if (custom_op_domain_ptr) { session_options.Add(custom_op_domain_ptr); } if (custom_op_library_filename) { Ort::ThrowOnError(Ort::GetApi().RegisterCustomOpsLibrary(session_options, custom_op_library_filename, library_handle)); } // if session creation passes, model loads fine Ort::Session session(env, model_uri.c_str(), session_options); // caller wants to test running the model (not just loading the model) if (!test_session_creation_only) { // Now run auto default_allocator = std::make_unique<MockedOrtAllocator>(); //without preallocated output tensor RunSession<OutT>(default_allocator.get(), session, inputs, output_name, expected_dims_y, expected_values_y, nullptr); //with preallocated output tensor Ort::Value value_y = Ort::Value::CreateTensor<float>(default_allocator.get(), expected_dims_y.data(), expected_dims_y.size()); //test it twice for (int i = 0; i != 2; ++i) RunSession<OutT>(default_allocator.get(), session, inputs, output_name, expected_dims_y, expected_values_y, &value_y); }}前文提到,如果对应EP是CUDA,需要确保 ORT 的 CUDA kernels 和定制 CUDA kernels 之间的同步。为了实现这一目标,首先通过 CreateDefaultOrtCudaProviderOptionsWithCustomStream 函数将新创建的 CUDA 计算流以 OrtCudaProviderOptions 的形式传递给 SessionOptions:
OrtCUDAProviderOptions cuda_options = CreateDefaultOrtCudaProviderOptionsWithCustomStream(cuda_compute_stream);session_options.AppendExecutionProvider_CUDA(cuda_options)之后,将定制算子域也添加到 SessionOptions 中:
if (custom_op_domain_ptr) { session_options.Add(custom_op_domain_ptr);}至此,SessionOptions 已经构建完成,下面创建 Session 并通过 model_uri 加载模型:
Ort::Session session(env, model_uri.c_str(), session_options);这里的(1)Ort::Session 是在 onnxruntime_cxx_api.h 文件中声明的类,(2)对应的构造函数在 onnxruntime_cxx_inline.h 中实现,(3)实现方式是进一步调用 onnxruntime_c_api.h 中定义的 API,该 API 也仅仅是声明,(4)最终对应的实现在 onnxruntime_c_api.cc 文件中:
// (1) include/onnxruntime/core/session/onnxruntime_cxx_api.hstruct Session : Base<OrtSession> { explicit Session(std::nullptr_t) {} Session(Env& env, const ORTCHAR_T* model_path, const SessionOptions& options);}// (2) include/onnxruntime/core/session/onnxruntime_cxx_inline.hinline Session::Session(Env& env, const ORTCHAR_T* model_path, const SessionOptions& options) { ThrowOnError(GetApi().CreateSession(env, model_path, options, &p_));}// (3) include/onnxruntime/core/session/onnxruntime_c_api.hORT_API2_STATUS(CreateSession, _In_ const OrtEnv* env, _In_ const ORTCHAR_T* model_path, _In_ const OrtSessionOptions* options, _Outptr_ OrtSession** out);// (4) onnxruntime/core/session/onnxruntime_c_api.ccORT_API_STATUS_IMPL(OrtApis::CreateSession, _In_ const OrtEnv* env, _In_ const ORTCHAR_T* model_path, _In_ const OrtSessionOptions* options, _Outptr_ OrtSession** out) { API_IMPL_BEGIN std::unique_ptr<onnxruntime::InferenceSession> sess; OrtStatus* status = nullptr; *out = nullptr; ORT_TRY { ORT_API_RETURN_IF_ERROR(CreateSessionAndLoadModel(options, env, model_path, nullptr, 0, sess)); ORT_API_RETURN_IF_ERROR(InitializeSession(options, sess)); *out = reinterpret_cast<OrtSession*>(sess.release()); } ORT_CATCH(const std::exception& e) { ORT_HANDLE_EXCEPTION([&]() { status = OrtApis::CreateStatus(ORT_FAIL, e.what()); }); } return status; API_IMPL_END}可以发现,Ort::Session 内部还是调用了 onnxruntime::InferenceSession。
扯远了,下面回归主题。
创建 Session 完成之后,便开始运行,进入 RunSession 函数内部:
// file path: onnxruntime/test/shared_lib/test_inference.cctemplate <typename OutT>void RunSession(OrtAllocator* allocator, Ort::Session& session_object, const std::vector<Input>& inputs, const char* output_name, const std::vector<int64_t>& dims_y, const std::vector<OutT>& values_y, Ort::Value* output_tensor) { // 构建模型输入 std::vector<Ort::Value> ort_inputs; std::vector<const char*> input_names; for (size_t i = 0; i < inputs.size(); i++) { input_names.emplace_back(inputs[i].name); ort_inputs.emplace_back( Ort::Value::CreateTensor<float>(allocator->Info(allocator), const_cast<float*>(inputs[i].values.data()), inputs[i].values.size(), inputs[i].dims.data(), inputs[i].dims.size())); } // 运行 RUN std::vector<Ort::Value> ort_outputs; if (output_tensor) session_object.Run(Ort::RunOptions{nullptr}, input_names.data(), ort_inputs.data(), ort_inputs.size(), &output_name, output_tensor, 1); else { ort_outputs = session_object.Run(Ort::RunOptions{}, input_names.data(), ort_inputs.data(), ort_inputs.size(), &output_name, 1); ASSERT_EQ(ort_outputs.size(), 1u); output_tensor = &ort_outputs[0]; } auto type_info = output_tensor->GetTensorTypeAndShapeInfo(); ASSERT_EQ(type_info.GetShape(), dims_y); size_t total_len = type_info.GetElementCount(); ASSERT_EQ(values_y.size(), total_len); OutT* f = output_tensor->GetTensorMutableData<OutT>(); for (size_t i = 0; i != total_len; ++i) { ASSERT_EQ(values_y[i], f[i]); }}这里使用了一些GTest中的断言来判定运行结果是否符合预期。
至此,我们已经完整地分析了定制算子从定义到使用的全部流程。
文档中还提到了 Contrib ops,这类算子归属于 contrib ops domain,是嵌入到 runtime 内部的,对于一些使用低频的算子最好不要加入这个域中,否则会导致运行时库(runtime library)过大。
官方文档中给出了添加算子到这个域中的方法,这里就不再进行介绍了,以后用到了再说吧。