ai-box/cxx/mnn/node.cc

#include <iostream>
#include <vector>
#include <map>
#include <cstring>
#include <MNN/Interpreter.hpp>
#include <MNN/ImageProcess.hpp>
#include "common/tensor.h"
#include "node.h"

using namespace Napi;

#define SESSION_INSTANCE_METHOD(method) InstanceMethod<&MNNSession::method>(#method, static_cast<napi_property_attributes>(napi_writable | napi_configurable))

static const std::map<TensorDataType, halide_type_t> DATA_TYPE_MAP = {
	{TensorDataType::Float32, halide_type_of<float>()},
	{TensorDataType::Float64, halide_type_of<double>()},
	{TensorDataType::Int32, halide_type_of<int32_t>()},
	{TensorDataType::Uint32, halide_type_of<uint32_t>()},
	{TensorDataType::Int16, halide_type_of<int16_t>()},
	{TensorDataType::Uint16, halide_type_of<uint16_t>()},
	{TensorDataType::Int8, halide_type_of<int8_t>()},
	{TensorDataType::Uint8, halide_type_of<uint8_t>()},
	{TensorDataType::Int64, halide_type_of<int64_t>()},
	{TensorDataType::Uint64, halide_type_of<uint64_t>()},
};

static size_t getShapeSize(const std::vector<int> &shape)
{
	if (!shape.size()) return 0;
	size_t sum = 1;
	for (auto i : shape) {
		if (i > 1) sum *= i;
	};
	return sum;
}

static Napi::Value mnnSizeToJavascript(Napi::Env env, const std::vector<int> &shape)
{
	auto result = Napi::Array::New(env, shape.size());
	for (int i = 0; i < shape.size(); ++i) result.Set(i, Napi::Number::New(env, shape[i]));
	return result;
}

class MNNSessionRunWorker : public AsyncWorker {
  public:
	MNNSessionRunWorker(const Napi::Function &callback, MNN::Interpreter *interpreter, MNN::Session *session)
		: AsyncWorker(callback), interpreter_(interpreter), session_(session) {}

	~MNNSessionRunWorker()
	{
		interpreter_->releaseSession(session_);
	}

	void Execute()
	{
		if (MNN::ErrorCode::NO_ERROR != interpreter_->runSession(session_)) {
			SetError(std::string("Run session failed"));
		}
	}

	void OnOK()
	{
		if (HasError()) {
			Callback().Call({Error::New(Env(), errorMessage_.c_str()).Value(), Env().Undefined()});
		}
		else {
			auto result = Object::New(Env());
			for (auto it : interpreter_->getSessionOutputAll(session_)) {
				auto tensor = it.second;
				auto item = Object::New(Env());
				auto buffer = ArrayBuffer::New(Env(), tensor->size());
				memcpy(buffer.Data(), tensor->host<float>(), tensor->size());
				item.Set("data", buffer);
				item.Set("shape", mnnSizeToJavascript(Env(), tensor->shape()));
				result.Set(it.first, item);
			}
			Callback().Call({Env().Undefined(), result});
		}
	}

	void SetInput(const std::string &name, TensorDataType dataType, const std::vector<int> &shape, void *data, size_t dataBytes)
	{
		auto tensor = interpreter_->getSessionInput(session_, name.c_str());
		if (!tensor) {
			SetError(std::string("input name #" + name + " not exists"));
			return;
		}

		halide_type_t type = tensor->getType();
		if (dataType != TensorDataType::Unknown) {
			auto it = DATA_TYPE_MAP.find(dataType);
			if (it != DATA_TYPE_MAP.end()) type = it->second;
		}

		if (shape.size()) {
			interpreter_->resizeTensor(tensor, shape);
			interpreter_->resizeSession(session_);
		}
		auto tensorBytes = getShapeSize(tensor->shape()) * type.bits / 8;
		if (tensorBytes != dataBytes) {
			SetError(std::string("input name #" + name + " data size not matched"));
			return;
		}

		auto hostTensor = MNN::Tensor::create(tensor->shape(), type, data, MNN::Tensor::CAFFE);
		tensor->copyFromHostTensor(hostTensor);
		delete hostTensor;
	}

	inline void SetError(const std::string &what) { errorMessage_ = what; }
	inline bool HasError() { return errorMessage_.size() > 0; }

  private:
	MNN::Interpreter *interpreter_;
	MNN::Session *session_;
	std::string errorMessage_;
};

class MNNSession : public ObjectWrap<MNNSession> {
  public:
	static Napi::Object Init(Napi::Env env, Napi::Object exports)
	{
		Function func = DefineClass(env, "MNNSession", {
														   SESSION_INSTANCE_METHOD(GetInputsInfo),
														   SESSION_INSTANCE_METHOD(GetOutputsInfo),
														   SESSION_INSTANCE_METHOD(Run),
													   });
		FunctionReference *constructor = new FunctionReference();
		*constructor = Napi::Persistent(func);
		exports.Set("MNNSession", func);
		env.SetInstanceData<FunctionReference>(constructor);
		return exports;
	}

	MNNSession(const CallbackInfo &info)
		: ObjectWrap(info)
	{
		try {
			if (info[0].IsString()) {
				interpreter_ = MNN::Interpreter::createFromFile(info[0].As<String>().Utf8Value().c_str());
			}
			else if (info[0].IsTypedArray()) {
				size_t bufferBytes;
				auto buffer = dataFromTypedArray(info[0], bufferBytes);
				interpreter_ = MNN::Interpreter::createFromBuffer(buffer, bufferBytes);
			}
			else interpreter_ = nullptr;

			if (interpreter_) {
				backendConfig_.precision = MNN::BackendConfig::Precision_High;
				backendConfig_.power = MNN::BackendConfig::Power_High;
				scheduleConfig_.type = MNN_FORWARD_CPU;
				scheduleConfig_.numThread = 1;
				scheduleConfig_.backendConfig = &backendConfig_;
				session_ = interpreter_->createSession(scheduleConfig_);
			}
			else session_ = nullptr;
		}
		catch (std::exception &e) {
			Error::New(info.Env(), e.what()).ThrowAsJavaScriptException();
		}
	}

	~MNNSession() {}

	Napi::Value GetInputsInfo(const Napi::CallbackInfo &info) { return BuildInputOutputInfo(info.Env(), interpreter_->getSessionInputAll(session_)); }

	Napi::Value GetOutputsInfo(const Napi::CallbackInfo &info) { return BuildInputOutputInfo(info.Env(), interpreter_->getSessionOutputAll(session_)); }

	Napi::Value Run(const Napi::CallbackInfo &info)
	{
		auto worker = new MNNSessionRunWorker(info[1].As<Function>(), interpreter_, interpreter_->createSession(scheduleConfig_));
		auto inputArgument = info[0].As<Object>();
		for (auto it = inputArgument.begin(); it != inputArgument.end(); ++it) {
			auto name = (*it).first.As<String>().Utf8Value();
			auto inputOption = static_cast<Napi::Value>((*it).second).As<Object>();
			auto type = inputOption.Has("type") ? static_cast<TensorDataType>(inputOption.Get("type").As<Number>().Int32Value()) : TensorDataType::Unknown;
			size_t dataByteLen;
			void *data = dataFromTypedArray(inputOption.Get("data"), dataByteLen);
			auto shape = inputOption.Has("shape") ? GetShapeFromJavascript(inputOption.Get("shape").As<Array>()) : std::vector<int>();
			worker->SetInput(name, type, shape, data, dataByteLen);
		}
		worker->Queue();
		return info.Env().Undefined();
	}

  private:
	Napi::Object BuildInputOutputInfo(Napi::Env env, const std::map<std::string, MNN::Tensor *> &tensors)
	{
		auto result = Object::New(env);
		for (auto it : tensors) {
			auto item = Object::New(env);
			auto name = it.first;
			auto shape = it.second->shape();
			auto type = it.second->getType();
			TensorDataType dataType = TensorDataType::Unknown;
			for (auto dt : DATA_TYPE_MAP) {
				if (dt.second == type) {
					dataType = dt.first;
					break;
				}
			}
			auto shapeArr = Array::New(env, shape.size());
			for (size_t i = 0; i < shape.size(); i++) {
				shapeArr.Set(i, Number::New(env, shape[i]));
			}
			item.Set("name", String::New(env, name));
			item.Set("shape", shapeArr);
			item.Set("type", Number::New(env, static_cast<int>(dataType)));
			result.Set(name, item);
		}
		return result;
	}

	std::vector<int> GetShapeFromJavascript(const Napi::Array &shape)
	{
		std::vector<int> result;
		for (size_t i = 0; i < shape.Length(); i++) {
			result.push_back(shape.Get(i).As<Number>().Int32Value());
		}
		return result;
	}

  private:
	MNN::Interpreter *interpreter_;
	MNN::Session *session_;
	MNN::BackendConfig backendConfig_;
	MNN::ScheduleConfig scheduleConfig_;
};

void InstallMNNAPI(Napi::Env env, Napi::Object exports)
{
	MNNSession::Init(env, exports);
}


#if defined(USE_MNN) && !defined(BUILD_MAIN_WORD)
static Object Init(Env env, Object exports)
{
#ifdef RELEASE_VERSION
	exports.Set("__release__", Napi::String::New(env, RELEASE_VERSION));
#endif
	InstallMNNAPI(env, exports);
	return exports;
}
NODE_API_MODULE(addon, Init)
#endif