diff --git a/cpp/CMakeLists.txt b/cpp/CMakeLists.txt
new file mode 100644
index 0000000..665884a
--- /dev/null
+++ b/cpp/CMakeLists.txt
@@ -0,0 +1,80 @@
+cmake_minimum_required(VERSION 3.11 FATAL_ERROR)
+
+include(GNUInstallDirs)
+
+set(CMAKE_CXX_STANDARD 14)
+set(CMAKE_CXX_STANDARD_REQUIRED ON)
+set(CMAKE_CXX_EXTENSIONS OFF)
+
+SET(PROJECT_NAME pynvjpeg)
+PROJECT(${PROJECT_NAME} VERSION 0.1 LANGUAGES CUDA CXX)
+
+if(NOT DEFINED CMAKE_CUDA_STANDARD)
+    set(CMAKE_CUDA_STANDARD 11)
+    set(CMAKE_CUDA_STANDARD_REQUIRED ON)    
+endif()
+
+include_directories(${CMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES}) 
+link_directories(${CMAKE_CUDA_INCLUDE_DIRS})
+
+find_library(CUDART_LIBRARY cudart ${CMAKE_CUDA_IMPLICIT_LINK_DIRECTORIES})
+find_library(NVJPEG_LIBRARY nvjpeg ${CMAKE_CUDA_IMPLICIT_LINK_DIRECTORIES})
+
+include_directories(
+  ${PROJECT_SOURCE_DIR}
+  )
+
+include(FetchContent)
+FetchContent_Declare(
+  pybind11_sources
+  GIT_REPOSITORY https://github.com/pybind/pybind11.git
+  GIT_TAG v2.9
+  )
+FetchContent_GetProperties(pybind11_sources)
+if(NOT pybind11_sources_POPULATED)
+  FetchContent_Populate(pybind11_sources)
+  add_subdirectory(
+    ${pybind11_sources_SOURCE_DIR}
+    ${pybind11_sources_BINARY_DIR}
+    )
+endif()
+
+if(CMAKE_INSTALL_PREFIX_INITIALIZED_TO_DEFAULT)
+    set(CMAKE_INSTALL_PREFIX ${CMAKE_BINARY_DIR} CACHE PATH "" FORCE)
+endif(CMAKE_INSTALL_PREFIX_INITIALIZED_TO_DEFAULT)
+
+add_library(${PROJECT_NAME}
+  SHARED
+  nvjpeg_decoder.cpp
+  jpeg_image.cpp
+  cuda_util.cpp
+  python_nvjpeg_decoder.cpp
+  )
+
+target_link_libraries(${PROJECT_NAME}
+  PUBLIC
+  ${CUDART_LIBRARY}
+  ${NVJPEG_LIBRARY}
+  pybind11::module)
+
+set_target_properties(${PROJECT_NAME}
+  PROPERTIES
+  PREFIX "${PYTHON_MODULE_PREFIX}"
+  SUFFIX "${PYTHON_MODULE_EXTENSION}"
+  )
+
+add_executable(test_decode test.cpp)
+target_link_libraries(test_decode ${PROJECT_NAME} pybind11::embed)
+
+include(CTest)
+enable_testing()
+
+add_test(
+  NAME test_decode
+  COMMAND $<TARGET_FILE:test_decode> ${CMAKE_CURRENT_SOURCE_DIR}/cat.jpg
+  )
+
+add_test(
+  NAME python_test_short
+  COMMAND ${PYTHON_EXECUTABLE} ${CMAKE_CURRENT_SOURCE_DIR}/test.py $<TARGET_FILE:${PROJECT_NAME}> ${CMAKE_CURRENT_SOURCE_DIR}/cat.jpg
+  )
diff --git a/cpp/cat.jpg b/cpp/cat.jpg
new file mode 100644
index 0000000..2e24329
Binary files /dev/null and b/cpp/cat.jpg differ
diff --git a/cpp/cuda_util.cpp b/cpp/cuda_util.cpp
new file mode 100644
index 0000000..ad69a75
--- /dev/null
+++ b/cpp/cuda_util.cpp
@@ -0,0 +1,12 @@
+
+#include "cuda_util.h"
+
+
+namespace NVJpegDecoder {
+
+int DevMalloc(void **p, size_t s) { return (int)cudaMalloc(p, s); }
+int DevFree(void *p) { return (int)cudaFree(p); }
+int HostMalloc(void** p, size_t s, unsigned int f) { return (int)cudaHostAlloc(p, s, f); }
+int HostFree(void* p) { return (int)cudaFreeHost(p); }
+
+} // namespace NVJpegDecoder
diff --git a/cpp/cuda_util.h b/cpp/cuda_util.h
new file mode 100644
index 0000000..3b946ac
--- /dev/null
+++ b/cpp/cuda_util.h
@@ -0,0 +1,98 @@
+#pragma once
+
+#include <string>
+#include <iostream>
+#include <vector>
+#include <cuda_runtime_api.h>
+#include <nvjpeg.h>
+
+namespace NVJpegDecoder {
+
+
+#define CHECK_CUDA(call)                                                \
+  do {                                                                  \
+    CudaStatus s(call);                                                 \
+    if (!s.IsOk()) {                                                    \
+      std::cout << "CUDA Runtime failure: '#" << s.Msg() << "' at " <<  __FILE__ << ":" << __LINE__ << std::endl; \
+      return false;                                                     \
+    }                                                                   \
+  } while (false)                                                       \
+
+#define CHECK_NVJPEG(call)                                              \
+  do {                                                                  \
+    NvJpegStatus s(call);                                               \
+    if (!s.IsOk()) {                                                    \
+      std::cout << "NVJPEG failure: '#" << s.Msg() << "' at " <<  __FILE__ << ":" << __LINE__ << std::endl; \
+      return false;                                                     \
+    }                                                                   \
+  } while (false)
+
+
+int DevMalloc(void **p, size_t s);
+int DevFree(void *p);
+int HostMalloc(void** p, size_t s, unsigned int f);
+int HostFree(void* p);
+
+
+class CudaStatus {
+public:
+   explicit CudaStatus(cudaError_t error) : mCode(error) {}
+   
+   CudaStatus(CudaStatus&) = default;
+   CudaStatus& operator=(CudaStatus&) = default;
+   
+   bool IsOk() {
+     return mCode == cudaSuccess;
+   }
+   
+   std::string Msg() {
+     return cudaGetErrorString(mCode);
+   }
+
+private:
+   cudaError_t mCode;
+};
+
+
+class NvJpegStatus {
+public:
+   explicit NvJpegStatus(nvjpegStatus_t error) : mCode(error) {}
+
+   bool IsOk() {
+     return mCode == NVJPEG_STATUS_SUCCESS;
+   }
+
+   inline nvjpegStatus_t Code() {return mCode;}
+
+   std::string Msg() {
+     switch (mCode) {
+     case NVJPEG_STATUS_NOT_INITIALIZED:
+       return "NVJPEG_STATUS_NOT_INITIALIZED";
+     case NVJPEG_STATUS_INVALID_PARAMETER:
+       return "NVJPEG_STATUS_INVALID_PARAMETER";
+     case NVJPEG_STATUS_BAD_JPEG:
+       return "NVJPEG_STATUS_BAD_JPEG";
+     case NVJPEG_STATUS_JPEG_NOT_SUPPORTED:
+       return "NVJPEG_STATUS_JPEG_NOT_SUPPORTED";
+     case NVJPEG_STATUS_ALLOCATOR_FAILURE:
+       return "NVJPEG_STATUS_ALLOCATOR_FAILURE";
+     case NVJPEG_STATUS_EXECUTION_FAILED:
+       return "NVJPEG_STATUS_EXECUTION_FAILED";
+     case NVJPEG_STATUS_ARCH_MISMATCH:
+       return "NVJPEG_STATUS_ARCH_MISMATCH";
+     case NVJPEG_STATUS_INTERNAL_ERROR:
+       return "NVJPEG_STATUS_INTERNAL_ERROR";
+     case NVJPEG_STATUS_IMPLEMENTATION_NOT_SUPPORTED:
+       return "NVJPEG_STATUS_IMPLEMENTATION_NOT_SUPPORTED";
+     case NVJPEG_STATUS_INCOMPLETE_BITSTREAM:
+       return "NVJPEG_STATUS_INCOMPLETE_BITSTREAM";
+     default:
+       return "UNKNOWN NVJPEG ERROR";
+     }
+   }
+
+private:
+   nvjpegStatus_t mCode;
+};
+
+} // namespace NVJpegDecoder
diff --git a/cpp/jpeg_image.cpp b/cpp/jpeg_image.cpp
new file mode 100644
index 0000000..14d2117
--- /dev/null
+++ b/cpp/jpeg_image.cpp
@@ -0,0 +1,59 @@
+#include "jpeg_image.h"
+#include "cuda_util.h"
+
+
+namespace NVJpegDecoder {
+
+bool JpegImage::Init(int width, int height, int channels) {
+  mNvImage = std::make_unique<nvjpegImage_t>();
+  unsigned char * pBuffer = nullptr;
+  CHECK_CUDA(cudaMalloc((void **)&pBuffer, height * width * channels));
+  for(int i = 0; i < channels; i++) {
+    mNvImage->channel[i] = pBuffer + (height *width * i);
+    mNvImage->pitch[i] = (unsigned int)width;
+  }
+
+  mNvImage->pitch[0] = (unsigned int)width * channels;
+  mWidth = width;
+  mHeight = height;
+  mChannels = channels;
+  return true;
+}
+
+JpegImage::JpegImage(JpegImage&& rhs) {
+  mWidth = rhs.mWidth;
+  mHeight = rhs.mHeight;
+  mChannels = rhs.mChannels;  
+  mSubsampling = rhs.mSubsampling;
+  mNvImage = std::move(rhs.mNvImage);
+}
+
+JpegImage& JpegImage::operator=(JpegImage&& rhs) {
+  mWidth = rhs.mWidth;
+  mHeight = rhs.mHeight;
+  mChannels = rhs.mChannels;
+  mSubsampling = rhs.mSubsampling;
+  mNvImage = std::move(rhs.mNvImage);
+  return *this;
+}
+
+unsigned char* JpegImage::Cpu() {
+  size_t size = mHeight * mWidth * mChannels;
+  auto buffer = std::make_unique<unsigned char[]>(size) ;
+  CudaStatus s(cudaMemcpy(buffer.get(), mNvImage->channel[0], size, cudaMemcpyDeviceToHost));
+
+  if (!s.IsOk()) {
+    std::cout << "Copy image from GPU to CPU failed: " << s.Msg() << std::endl;
+    return nullptr;
+  }
+  return buffer.release();
+}
+
+JpegImage::~JpegImage(){
+  if (mNvImage != nullptr) {
+    cudaFree(mNvImage->channel[0]);
+  }
+}
+
+
+} // namespace NVJpegDecoder
diff --git a/cpp/jpeg_image.h b/cpp/jpeg_image.h
new file mode 100644
index 0000000..1dc9a3a
--- /dev/null
+++ b/cpp/jpeg_image.h
@@ -0,0 +1,39 @@
+#pragma once
+#include <vector>
+#include <memory>
+#include <nvjpeg.h>
+
+namespace NVJpegDecoder {
+
+class JpegImage {
+
+public:
+   JpegImage() noexcept : mNvImage(nullptr) {}
+   virtual ~JpegImage();
+
+   JpegImage(const JpegImage&) = delete;
+   JpegImage& operator=(const JpegImage&) = delete;
+
+   JpegImage(JpegImage&& rhs);
+   JpegImage& operator=(JpegImage&& rhs);
+
+   bool Init(int width, int height, int channels);
+
+   nvjpegImage_t* GetImagePoint() {
+     return mNvImage.get();
+   }
+
+   const std::vector<int64_t> Dims() {
+     return std::vector<int64_t>{mHeight, mWidth, mChannels};
+   }
+
+   unsigned char* Cpu();
+
+private:
+   int mWidth = 0;
+   int mHeight = 0;
+   int mChannels = 0;
+   nvjpegChromaSubsampling_t mSubsampling;
+   std::unique_ptr<nvjpegImage_t> mNvImage;
+};
+} // namespace NVJpegDecoder
diff --git a/cpp/my_test.cpp b/cpp/my_test.cpp
new file mode 100644
index 0000000..596b0c6
--- /dev/null
+++ b/cpp/my_test.cpp
@@ -0,0 +1,23 @@
+#include <pybind11/pybind11.h>
+#include <pybind11/embed.h>
+#include <iostream>
+
+namespace py = pybind11;
+
+
+void print(py::list l) {
+  for (auto item: l) {
+    std::cout << item.attr("__str__")().cast<std::string>() << std::endl;
+  }
+}
+
+int main() {
+  py::scoped_interpreter guard{};
+  py::print("Hello, World!");
+  py::list data;
+  for (int i = 0; i < 10; i++) {
+    data.append(i);
+  }  
+  print(data);
+  return 0;
+}
diff --git a/cpp/nvjpeg_decoder.cpp b/cpp/nvjpeg_decoder.cpp
new file mode 100644
index 0000000..b922928
--- /dev/null
+++ b/cpp/nvjpeg_decoder.cpp
@@ -0,0 +1,93 @@
+#include <string>
+#include <iostream>
+#include <fstream>
+#include <cassert>
+#include "cuda_util.h"
+#include "nvjpeg_decoder.h"
+
+
+namespace NVJpegDecoder {
+
+bool Decoder::BindDevice(int device_id) {
+  if (device_id == mDeviceId) {
+    return true;
+  }
+
+  if (device_id < 0) {
+    std::cout<< "Device id must >= 0, the input is "
+             << device_id << std::endl;
+    return false;
+  }
+
+  Destroy();
+  mDeviceId = device_id;
+  CHECK_CUDA(cudaSetDevice(mDeviceId));
+  CHECK_NVJPEG(nvjpegCreateSimple(&mHandle));
+  CHECK_NVJPEG(nvjpegJpegStateCreate(mHandle, &mState));
+  CHECK_CUDA(cudaStreamCreateWithFlags(&mStream, cudaStreamNonBlocking));
+  return true;
+}
+
+bool Decoder::Read(const char* filename, JpegImage& image) {
+  std::ifstream input(filename);
+  if (!(input.is_open())) {
+    std::cout << "Open file " << filename << " failed" << std::endl;
+    return false;
+  }
+  std::string imagedata((std::istreambuf_iterator<char>(input)), std::istreambuf_iterator<char>());
+  if (!Decode(imagedata, image)) {
+    return false;
+  }
+  return true;
+}
+
+bool Decoder::Decode(std::string& imagedata , JpegImage& image) {
+  if (!PrepareJpegImage(imagedata, image)) {
+    return false;
+  }
+  CHECK_NVJPEG(nvjpegDecode(
+    mHandle,
+    mState,
+    (const unsigned char *)imagedata.data(),
+    imagedata.size(),
+    NVJPEG_OUTPUT_RGBI,
+    image.GetImagePoint(),
+    mStream));
+  return true;
+}
+
+bool Decoder::PrepareJpegImage(const std::string& image, JpegImage& output) {
+  int widths[NVJPEG_MAX_COMPONENT];
+  int heights[NVJPEG_MAX_COMPONENT];
+  int channels;
+  nvjpegChromaSubsampling_t subsampling;
+
+  CHECK_NVJPEG(nvjpegGetImageInfo(
+                 mHandle, (unsigned char *)image.data(), image.size(),
+                 &channels, &subsampling, widths, heights));
+  
+  if (NVJPEG_CSS_UNKNOWN == subsampling) {
+    std::cout << "Unknown chroma subsampling" << std::endl;
+    return false;
+  }
+
+  if (!output.Init(widths[0], heights[0], channels)) {
+    return false;
+  }
+  return true;
+}
+
+void Decoder::Destroy() {
+  if (mDeviceId >= 0) {
+    cudaStreamDestroy(mStream);
+    nvjpegJpegStateDestroy(mState);
+    nvjpegDestroy(mHandle);
+    mDeviceId = -1;
+  }
+}
+
+Decoder::~Decoder() {
+  Destroy();
+}
+
+} // namespace NVJpegDecoder
diff --git a/cpp/nvjpeg_decoder.h b/cpp/nvjpeg_decoder.h
new file mode 100644
index 0000000..5c0c013
--- /dev/null
+++ b/cpp/nvjpeg_decoder.h
@@ -0,0 +1,39 @@
+#pragma once
+
+#include <vector>
+#include <string>
+#include <cuda_runtime_api.h>
+#include <nvjpeg.h>
+
+#include "jpeg_image.h"
+
+namespace NVJpegDecoder {
+
+typedef std::vector<std::string> OriginJpegImages;
+typedef std::vector<JpegImage> JpegImages;
+
+class Decoder {
+public:
+   Decoder():mDeviceId(-1) {}
+   virtual ~Decoder();
+
+   Decoder(Decoder&) = delete;
+   Decoder& operator=(Decoder&) = delete;
+   
+   bool BindDevice(int device_id=0);
+   bool Read(const char* filename, JpegImage& image);
+   bool Decode(std::string& imagedata , JpegImage& image);
+   
+private:
+   bool PrepareJpegImage(const std::string& image, JpegImage& output);
+   void Destroy();
+   
+private:
+   nvjpegHandle_t mHandle;
+   nvjpegJpegState_t mState;
+
+   int mDeviceId;
+   cudaStream_t mStream;
+};
+
+} // namespace NVJpegDecoder
diff --git a/cpp/python_nvjpeg_decoder.cpp b/cpp/python_nvjpeg_decoder.cpp
new file mode 100644
index 0000000..b837ff8
--- /dev/null
+++ b/cpp/python_nvjpeg_decoder.cpp
@@ -0,0 +1,52 @@
+#include <string>
+#include <iostream>
+#include <pybind11/numpy.h>
+#include <pybind11/stl_bind.h>
+#include "python_nvjpeg_decoder.h"
+
+
+namespace NVJpegDecoder {
+
+bool PythonDecoder::BindDevice(int device_id) {
+  py::gil_scoped_release release;
+  return mDecoder.BindDevice(device_id);
+}
+
+py::object PythonDecoder::Read(std::string& filename) {
+  unsigned char* data = nullptr;
+  JpegImage image;
+  {
+    py::gil_scoped_release release;
+    if (mDecoder.Read(filename.c_str(), image)) {
+      data = image.Cpu();
+    }
+  } // gets gil
+  
+  std::unique_ptr<unsigned char> ret(data);
+  if (ret != nullptr) {
+    return py::array(py::dtype(py::format_descriptor<uint8_t>::format()), image.Dims(), (void*)ret.get());
+  }
+  return py::none();
+}
+
+py::object PythonDecoder::Decode(std::string& image_bytes) {
+  JpegImage image;
+  if (!mDecoder.Decode(image_bytes, image)) {
+    return py::none();
+  }
+  std::unique_ptr<unsigned char> data(image.Cpu());
+  if (nullptr == data) {
+    return py::none();
+  }
+  return py::array(py::dtype(py::format_descriptor<uint8_t>::format()), image.Dims(), (void*)data.get());
+}
+
+PYBIND11_MODULE(pynvjpeg, m) {
+  py::class_<PythonDecoder, std::shared_ptr<PythonDecoder>>(m, "Decoder")
+     .def(py::init())
+     .def("bind_device", &PythonDecoder::BindDevice)
+     .def("imread", &PythonDecoder::Read)
+     .def("imdecode", &PythonDecoder::Decode);  
+}
+
+} // namespace NVJpegDecoder
diff --git a/cpp/python_nvjpeg_decoder.h b/cpp/python_nvjpeg_decoder.h
new file mode 100644
index 0000000..19bf1ca
--- /dev/null
+++ b/cpp/python_nvjpeg_decoder.h
@@ -0,0 +1,26 @@
+#pragma once
+
+#include <pybind11/pybind11.h>
+#include <pybind11/stl.h>
+#include "nvjpeg_decoder.h"
+
+namespace py = pybind11;
+
+namespace NVJpegDecoder {
+
+class PythonDecoder {
+public:
+   PythonDecoder() = default;
+
+   PythonDecoder(PythonDecoder&) = delete;
+   PythonDecoder& operator=(PythonDecoder&) = delete;
+
+   bool BindDevice(int device_id=0);
+   py::object Read(std::string&);
+   py::object Decode(std::string&);
+
+private:
+   Decoder mDecoder;
+};
+
+} // namespace NVJpegDecoder
diff --git a/cpp/test.cpp b/cpp/test.cpp
new file mode 100644
index 0000000..1ad4cf7
--- /dev/null
+++ b/cpp/test.cpp
@@ -0,0 +1,90 @@
+#include <string>
+#include <sys/timeb.h>
+#include <sstream>
+#include <fstream>
+#include <iostream>
+#include "nvjpeg_decoder.h"
+
+
+bool ReadImage(const char* filename, std::string& imagedata) {
+  std::ifstream input(filename);
+  if (!(input.is_open())) {
+    std::cout << "Open file " << filename << " failed" << std::endl;
+    return false;
+  }
+
+  imagedata = std::string((std::istreambuf_iterator<char>(input)), std::istreambuf_iterator<char>());
+  return true;
+}
+
+int main(int argc, char *argv[]) {
+  const char* image_file = argv[1];
+  NVJpegDecoder::Decoder decoder;
+  if (!decoder.BindDevice(0)) {
+    std::cout << "Init Failed" << std::endl;
+    return -1;
+  }
+
+  // test deocde from bytes
+  {
+    std::string image_btyes;
+    if (!ReadImage(image_file, image_btyes)) {
+      return -1;
+    }
+    int count = 0;
+    timeb t1, t2;
+    ftime(&t1);
+
+    while (count < 1000) {
+      NVJpegDecoder::JpegImage image;
+      if (!decoder.Decode(image_btyes, image)) {
+        return -1;
+      }
+      count++;
+    }
+    ftime(&t2);
+    std::cout << "--------- " << t2.time * 1000 + t2.millitm - t1.time * 1000 - t1.millitm << std::endl;
+  }
+    
+  // test decode from file
+  {
+    timeb t1, t2;
+    ftime(&t1);    
+    {
+      int count = 0;
+      while (count < 1000) {
+        NVJpegDecoder::JpegImage image;
+        if (!decoder.Read(image_file, image)) {
+          return -1;
+        }
+        unsigned char* d = image.Cpu();
+        delete[] d;
+        count++;
+      }
+    }
+    ftime(&t2);
+    std::cout << "--------- " << t2.time * 1000 + t2.millitm - t1.time * 1000 - t1.millitm << std::endl;        
+  }
+
+  decoder.BindDevice(1);  
+  {
+    std::string image_btyes;
+    if (!ReadImage(image_file, image_btyes)) {
+      return -1;
+    }
+    int count = 0;
+    timeb t1, t2;
+    ftime(&t1);
+
+    while (count < 1000) {
+      NVJpegDecoder::JpegImage image;
+      if (!decoder.Decode(image_btyes, image)) {
+        return -1;
+      }
+      count++;
+    }
+    ftime(&t2);
+    std::cout << "--------- " << t2.time * 1000 + t2.millitm - t1.time * 1000 - t1.millitm << std::endl;
+  }
+  return 0;
+}
diff --git a/cpp/test.py b/cpp/test.py
new file mode 100644
index 0000000..9e6ddf1
--- /dev/null
+++ b/cpp/test.py
@@ -0,0 +1,26 @@
+import sys
+import numpy as np
+import cv2
+import time
+sys.path.insert(0, '/'.join(sys.argv[1].split('/')[:-1]))
+import pynvjpeg
+
+
+if __name__ == '__main__':
+    decoder = pynvjpeg.Decoder()
+    assert(decoder.bind_device(0) is True)
+    filename = sys.argv[2]
+    image0 = decoder.imread(filename)
+    image0 = image0.astype(np.int32)
+    h, w, c = image0.shape
+    with open(filename, 'rb') as f:
+        data = f.read()
+
+    
+    image1 = cv2.imread(filename)
+    image1 = cv2.cvtColor(image1, cv2.COLOR_BGR2RGB)
+    image1 = image1.astype(np.int32)
+    if np.sum((image1 - image0)) > h * w * c:
+        exit(-1)
+    exit(0)
+    
diff --git a/cpp/test_pydecoder.cpp b/cpp/test_pydecoder.cpp
new file mode 100644
index 0000000..c1d1b15
--- /dev/null
+++ b/cpp/test_pydecoder.cpp
@@ -0,0 +1,22 @@
+#include <iostream>
+#include <pybind11/embed.h>
+#include "python_nvjpeg_decoder.h"
+
+
+int main() {
+  py::scoped_interpreter guard{};  
+  NVJpegDecoder::PythonDecoder decoder;
+  if (!decoder.Init()) {
+    std::cout << "Init Failed" << std::endl;
+    return -1;
+  }
+  std::cout << 1 << std::endl;;
+  py::list data;
+  std::cout << 2 << std::endl;;  
+  for (int i = 0; i < 10; i++) {
+    std::cout << i << ' ';
+    data.append(i);
+  }
+  decoder.BatchDecode(data);
+  return 0;
+}
diff --git a/image_decoder.py b/image_decoder.py
new file mode 100644
index 0000000..57c14fa
--- /dev/null
+++ b/image_decoder.py
@@ -0,0 +1,96 @@
+# encoding=utf-8
+import os
+import sys
+import logging
+import platform
+
+import cv2
+import requests
+import numpy as np
+from towhee.types import Image
+from towhee.operator import PyOperator, SharedType
+
+logger = logging.getLogger()
+
+
+def create_pynvjpeg():
+    if platform.system().lower() != 'linux':
+        logger.warning('The GPU deocder only support linux')
+        return None
+
+    py_version = platform.python_version()
+    version = ''.join(py_version.split('.')[0:2])
+    cur_dir = os.path.dirname(os.path.abspath(__file__))
+    lib_path = cur_dir + '/' + 'py' + version + 'nvjpeg'
+    sys.path.insert(0, lib_path)
+    try:
+        import pynvjpeg
+        decoder = pynvjpeg.Decoder()
+        if not decoder.bind_device(0):
+            logger.info('Init GPU deocder failed, use CPU decoder')
+            return None
+        else:
+            return decoder
+    except Exception as e:
+        logger.error('Create nvjpeg failed, use opencv decoder, errors: ', str(e))
+        return None
+
+
+class ImageDecoder(PyOperator):
+    def __init__(self):
+        self._nvjpeg = create_pynvjpeg()
+
+    def to_device(self, device):
+        if self._nvjpeg is not None:
+            if not self._nvjpeg.bind_device(device):
+                self._nvjpeg = None
+
+    def _download_image(self, image_url):
+        image_type = image_url.split('?')[0].split('.')[-1].lower()
+        is_jpeg = True if image_type == 'jpeg' or image_type == 'jpg' else False
+        r = requests.get(image_url, timeout=(20, 20))
+        if r.status_code // 100 != 2:
+            logging.error('Download image from %s failed, error msg: %s, request code: %s ' % (image_url,
+                                                                                               r.text,
+                                                                                               r.status_code))
+            return None
+        return is_jpeg, r.content
+
+    def _read_image(self, image_path):
+        is_jpeg = image_path.lower().endswith('jpg') or image_path.lower().endswith('jpeg')
+        with open(image_path, 'rb') as f:
+            return is_jpeg, f.read()
+
+    def load_image(self, image_path):
+        if image_path.startswith('http'):
+            return self._download_image(image_path)
+        else:
+            return self._read_image(image_path)
+
+    def __call__(self, image_path: str):
+        is_jpeg, image_content = self.load_image(image_path)
+        if is_jpeg and self._nvjpeg is not None:
+            image = self._nvjpeg.imdecode(image_content)
+        else:
+            arr = np.asarray(bytearray(image_content), dtype=np.uint8)
+            image = cv2.imdecode(arr, -1)
+            if image is not None:
+                image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
+        if image is None:
+            raise RuntimeError('Decode image %s failed' % image_path)
+        return Image(image, 'RGB')
+
+    def input_schema(self):
+        return [(str, (1,))]
+
+    def output_schema(self):
+        return [(Image, (-1, -1, 3))]
+
+    @property
+    def shared_type(self):
+        return SharedType.NotShareable
+
+if __name__ == '__main__':
+    d = ImageDecoder()
+    print(d('/home/junjie.jiangjjj/images/1.png'))
+    print(d('/home/junjie.jiangjjj/images/towhee.jpeg'))    
diff --git a/py38nvjpeg/__init__.py b/py38nvjpeg/__init__.py
new file mode 100644
index 0000000..e69de29
diff --git a/py38nvjpeg/pynvjpeg.cpython-38-x86_64-linux-gnu.so b/py38nvjpeg/pynvjpeg.cpython-38-x86_64-linux-gnu.so
new file mode 100755
index 0000000..af632c0
Binary files /dev/null and b/py38nvjpeg/pynvjpeg.cpython-38-x86_64-linux-gnu.so differ