C++ opencldevice::pointer类代码示例

void SeededRegionGrowing::recompileOpenCLCode(Image::pointer input) {
    // Check if there is a need to recompile OpenCL code
    if(input->getDimensions() == mDimensionCLCodeCompiledFor &&
            input->getDataType() == mTypeCLCodeCompiledFor)
        return;

    OpenCLDevice::pointer device = getMainDevice();
    std::string buildOptions = "";
    if(input->getDataType() == TYPE_FLOAT) {
        buildOptions = "-DTYPE_FLOAT";
    } else if(input->getDataType() == TYPE_INT8 || input->getDataType() == TYPE_INT16) {
        buildOptions = "-DTYPE_INT";
    } else {
        buildOptions = "-DTYPE_UINT";
    }
    std::string filename;
    if(input->getDimensions() == 2) {
        filename = "Algorithms/SeededRegionGrowing/SeededRegionGrowing2D.cl";
    } else {
        filename = "Algorithms/SeededRegionGrowing/SeededRegionGrowing3D.cl";
    }
    int programNr = device->createProgramFromSource(std::string(FAST_SOURCE_DIR) + filename, buildOptions);
    mKernel = cl::Kernel(device->getProgram(programNr), "seededRegionGrowing");
    mDimensionCLCodeCompiledFor = input->getDimensions();
    mTypeCLCodeCompiledFor = input->getDataType();
}

void NoneLocalMeans::recompileOpenCLCode(Image::pointer input) {
    // Check if there is a need to recompile OpenCL code
    if(input->getDimensions() == mDimensionCLCodeCompiledFor &&
       input->getDataType() == mTypeCLCodeCompiledFor && !recompile)
        return;
    
    recompile = false;
    OpenCLDevice::pointer device = getMainDevice();
    std::string buildOptions = "";
    if(!device->isWritingTo3DTexturesSupported()) {
        buildOptions = "-DTYPE=" + getCTypeAsString(mOutputType);
    }
    buildOptions += " -D WINDOW=";
    buildOptions += std::to_string((windowSize-1)/2);
    buildOptions += " -D GROUP=";
    buildOptions += std::to_string((groupSize-1)/2);
    buildOptions += " -D KVERSION=";
    buildOptions += std::to_string(k);
    buildOptions += " -D EUCLID=";
    buildOptions += std::to_string(k);
    cl::Program program;
    if(input->getDimensions() == 2) {
        program = getOpenCLProgram(device, "2D", buildOptions);
    } else {
        //createOpenCLProgram(std::string(FAST_SOURCE_DIR) + "Algorithms/NoneLocalMeans/NoneLocalMeans3Dgs.cl", "3D");
        program = getOpenCLProgram(device, "3D", buildOptions);
    }
    mKernel = cl::Kernel(program, "noneLocalMeans");
    mDimensionCLCodeCompiledFor = input->getDimensions();
    mTypeCLCodeCompiledFor = input->getDataType();
}

void BinaryThresholding::execute() {
    if(!mLowerThresholdSet && !mUpperThresholdSet) {
        throw Exception("BinaryThresholding need at least one threshold to be set.");
    }

    Image::pointer input = getStaticInputData<Image>(0);
    Segmentation::pointer output = getStaticOutputData<Segmentation>(0);

    output->createFromImage(input);

    if(getMainDevice()->isHost()) {
        throw Exception("Not implemented yet.");
    } else {
        OpenCLDevice::pointer device = OpenCLDevice::pointer(getMainDevice());
        cl::Program program;
        if(input->getDimensions() == 3) {
            program = getOpenCLProgram(device, "3D");
        } else {
            program = getOpenCLProgram(device, "2D");
        }
        cl::Kernel kernel;
        if(mLowerThresholdSet && mUpperThresholdSet) {
            kernel = cl::Kernel(program, "tresholding");
            kernel.setArg(3, mLowerThreshold);
            kernel.setArg(4, mUpperThreshold);
        } else if(mLowerThresholdSet) {
            kernel = cl::Kernel(program, "thresholdingWithOnlyLower");
            kernel.setArg(3, mLowerThreshold);
        } else {
            kernel = cl::Kernel(program, "thresholdingWithOnlyUpper");
            kernel.setArg(3, mUpperThreshold);
        }
        cl::NDRange globalSize;
        OpenCLImageAccess::pointer access = input->getOpenCLImageAccess(ACCESS_READ, device);
        if(input->getDimensions() == 2) {
            OpenCLImageAccess::pointer access2 = output->getOpenCLImageAccess(ACCESS_READ_WRITE, device);
            kernel.setArg(0, *access->get2DImage());
            kernel.setArg(1, *access2->get2DImage());
            globalSize = cl::NDRange(output->getWidth(), output->getHeight());
        } else {
            // TODO no 3d image write support
            OpenCLImageAccess::pointer access2 = output->getOpenCLImageAccess(ACCESS_READ_WRITE, device);
            kernel.setArg(0, *access->get3DImage());
            kernel.setArg(1, *access2->get3DImage());
            globalSize = cl::NDRange(output->getWidth(), output->getHeight(), output->getDepth());
        }
        kernel.setArg(2, (uchar)mLabel);

        cl::CommandQueue queue = device->getCommandQueue();
        queue.enqueueNDRangeKernel(
                kernel,
                cl::NullRange,
                globalSize,
                cl::NullRange
        );
    }
}

void Dilation::execute() {
    Image::pointer input = getInputData<Image>();
    if(input->getDataType() != TYPE_UINT8) {
        throw Exception("Data type of image given to Dilation must be UINT8");
    }

    Image::pointer output = getOutputData<Image>();
    output->createFromImage(input);
    SceneGraph::setParentNode(output, input);
    output->fill(0);

    OpenCLDevice::pointer device = std::dynamic_pointer_cast<OpenCLDevice>(getMainDevice());
    cl::CommandQueue queue = device->getCommandQueue();
    cl::Program program = getOpenCLProgram(device);
    cl::Kernel dilateKernel(program, "dilate");

    Vector3ui size = input->getSize();

    OpenCLImageAccess::pointer access = input->getOpenCLImageAccess(ACCESS_READ, device);
    dilateKernel.setArg(0, *access->get3DImage());
    dilateKernel.setArg(2, mSize/2);

    if(!device->isWritingTo3DTexturesSupported()) {
        OpenCLBufferAccess::pointer access2 = output->getOpenCLBufferAccess(ACCESS_READ_WRITE, device);
        dilateKernel.setArg(1, *access2->get());

        queue.enqueueNDRangeKernel(
            dilateKernel,
            cl::NullRange,
            cl::NDRange(size.x(), size.y(), size.z()),
            cl::NullRange
        );
    } else {
        OpenCLImageAccess::pointer access2 = output->getOpenCLImageAccess(ACCESS_READ_WRITE, device);
        dilateKernel.setArg(1, *access2->get3DImage());

        queue.enqueueNDRangeKernel(
            dilateKernel,
            cl::NullRange,
            cl::NDRange(size.x(), size.y(), size.z()),
            cl::NullRange
        );
    }

}

void NoneLocalMeans::waitToFinish() {
    if (!getMainDevice()->isHost()) {
        OpenCLDevice::pointer device = getMainDevice();
        device->getCommandQueue().finish();
    }
}

void MeshRenderer::draw2D(
                cl::BufferGL PBO,
                uint width,
                uint height,
                Eigen::Transform<float, 3, Eigen::Affine> pixelToViewportTransform,
                float PBOspacing,
                Vector2f translation
        ) {
    boost::lock_guard<boost::mutex> lock(mMutex);

    OpenCLDevice::pointer device = getMainDevice();
    cl::CommandQueue queue = device->getCommandQueue();
    std::vector<cl::Memory> v;
    v.push_back(PBO);
    queue.enqueueAcquireGLObjects(&v);

    // Map would probably be better here, but doesn't work on NVIDIA, segfault surprise!
    //float* pixels = (float*)queue.enqueueMapBuffer(PBO, CL_TRUE, CL_MAP_WRITE, 0, width*height*sizeof(float)*4);
    boost::shared_array<float> pixels(new float[width*height*sizeof(float)*4]);
    queue.enqueueReadBuffer(PBO, CL_TRUE, 0, width*height*4*sizeof(float), pixels.get());

    boost::unordered_map<uint, Mesh::pointer>::iterator it;
    for(it = mMeshToRender.begin(); it != mMeshToRender.end(); it++) {
    	Mesh::pointer mesh = it->second;
    	if(mesh->getDimensions() != 2) // Mesh must be 2D
    		continue;

		Color color = mDefaultColor;
        ProcessObjectPort port = getInputPort(it->first);
        if(mInputColors.count(port) > 0) {
            color = mInputColors[port];
        }

    	MeshAccess::pointer access = mesh->getMeshAccess(ACCESS_READ);
        std::vector<VectorXui> lines = access->getLines();
        std::vector<MeshVertex> vertices = access->getVertices();

        // Draw each line
        for(int i = 0; i < lines.size(); ++i) {
        	Vector2ui line = lines[i];
        	Vector2f a = vertices[line.x()].getPosition();
        	Vector2f b = vertices[line.y()].getPosition();
        	Vector2f direction = b - a;
        	float lengthInPixels = ceil(direction.norm() / PBOspacing);

        	// Draw the line
        	for(int j = 0; j <= lengthInPixels; ++j) {
        		Vector2f positionInMM = a + direction*((float)j/lengthInPixels);
        		Vector2f positionInPixels = positionInMM / PBOspacing;

        		int x = round(positionInPixels.x());
        		int y = round(positionInPixels.y());
        		y = height - 1 - y;
        		if(x < 0 || y < 0 || x >= width || y >= height)
        			continue;

        		pixels[4*(x + y*width)] = color.getRedValue();
        		pixels[4*(x + y*width) + 1] = color.getGreenValue();
        		pixels[4*(x + y*width) + 2] = color.getBlueValue();
        	}
        }
    }

    //queue.enqueueUnmapMemObject(PBO, pixels);
    queue.enqueueWriteBuffer(PBO, CL_TRUE, 0, width*height*4*sizeof(float), pixels.get());
    queue.enqueueReleaseGLObjects(&v);
}

void BinaryThresholding::waitToFinish() {
    OpenCLDevice::pointer device = OpenCLDevice::pointer(getMainDevice());
    device->getCommandQueue().finish();
}

void DoubleFilter::execute() {
    if(!mInput.isValid()) {
        throw Exception("No input supplied to GaussianSmoothingFilter");
    }

    Image::pointer input = mInput;
    Image::pointer output = mOutput;

    // Initialize output image
    output->createFromImage(input, mDevice);

    if(mDevice->isHost()) {
        // Execution device is Host, use the executeAlgorithmOnHost function with the given data type
        switch(input->getDataType()) {
            // This macro creates a case statement for each data type and sets FAST_TYPE to the correct C++ data type
            fastSwitchTypeMacro(executeAlgorithmOnHost<FAST_TYPE>(input, output));
        }
    } else {
        // Execution device is an OpenCL device
        OpenCLDevice::pointer device = boost::static_pointer_cast<OpenCLDevice>(mDevice);

        // Set build options based on the data type of the data
        std::string buildOptions = "";
        switch(input->getDataType()) {
        case TYPE_FLOAT:
            buildOptions = "-DTYPE=float";
            break;
        case TYPE_INT8:
            buildOptions = "-DTYPE=char";
            break;
        case TYPE_UINT8:
            buildOptions = "-DTYPE=uchar";
            break;
        case TYPE_INT16:
            buildOptions = "-DTYPE=short";
            break;
        case TYPE_UINT16:
            buildOptions = "-DTYPE=ushort";
            break;
        }

        // Compile the code
        int programNr = device->createProgramFromSource(std::string(FAST_SOURCE_DIR) + "Tests/Algorithms/DoubleFilter.cl", buildOptions);
        cl::Kernel kernel = cl::Kernel(device->getProgram(programNr), "doubleFilter");

        // Get global size for the kernel
        cl::NDRange globalSize(input->getWidth()*input->getHeight()*input->getDepth()*input->getNrOfComponents());

        // Set the arguments for the kernel
        OpenCLBufferAccess inputAccess = input->getOpenCLBufferAccess(ACCESS_READ, device);
        OpenCLBufferAccess outputAccess = output->getOpenCLBufferAccess(ACCESS_READ_WRITE, device);
        kernel.setArg(0, *inputAccess.get());
        kernel.setArg(1, *outputAccess.get());

        // Execute the kernel
        device->getCommandQueue().enqueueNDRangeKernel(
                kernel,
                cl::NullRange,
                globalSize,
                cl::NullRange
        );
    }

    // Update timestamp of the output data
    output->updateModifiedTimestamp();
}

void NonLocalMeans::waitToFinish() {
    if (!getMainDevice()->isHost()) {
        OpenCLDevice::pointer device = std::dynamic_pointer_cast<OpenCLDevice>(getMainDevice());
        device->getCommandQueue().finish();
    }
}

void getIntensitySumFromOpenCLImage(OpenCLDevice::pointer device, cl::Image2D image, DataType type, float* sum) {
    // Get power of two size
    unsigned int powerOfTwoSize = getPowerOfTwoSize(std::max(image.getImageInfo<CL_IMAGE_WIDTH>(), image.getImageInfo<CL_IMAGE_HEIGHT>()));

    // Create image levels
    unsigned int size = powerOfTwoSize;
    size /= 2;
    std::vector<cl::Image2D> levels;
    while(size >= 4) {
        cl::Image2D level = cl::Image2D(device->getContext(), CL_MEM_READ_WRITE, getOpenCLImageFormat(device, CL_MEM_OBJECT_IMAGE2D, TYPE_FLOAT, 1), size, size);
        levels.push_back(level);
        size /= 2;
    }

    // Compile OpenCL code
    std::string buildOptions = "";
    switch(type) {
    case TYPE_FLOAT:
        buildOptions = "-DTYPE_FLOAT";
        break;
    case TYPE_UINT8:
        buildOptions = "-DTYPE_UINT8";
        break;
    case TYPE_INT8:
        buildOptions = "-DTYPE_INT8";
        break;
    case TYPE_UINT16:
        buildOptions = "-DTYPE_UINT16";
        break;
    case TYPE_INT16:
        buildOptions = "-DTYPE_INT16";
        break;
    }
    std::string sourceFilename = std::string(FAST_SOURCE_DIR) + "/ImageSum.cl";
    std::string programName = sourceFilename + buildOptions;
    // Only create program if it doesn't exist for this device from before
    if(!device->hasProgram(programName))
        device->createProgramFromSourceWithName(programName, sourceFilename, buildOptions);
    cl::Program program = device->getProgram(programName);
    cl::CommandQueue queue = device->getCommandQueue();

    // Fill first level
    size = powerOfTwoSize/2;
    cl::Kernel firstLevel(program, "createFirstSumImage2DLevel");
    firstLevel.setArg(0, image);
    firstLevel.setArg(1, levels[0]);

    queue.enqueueNDRangeKernel(
            firstLevel,
            cl::NullRange,
            cl::NDRange(size,size),
            cl::NullRange
    );

    // Fill all other levels
    cl::Kernel createLevel(program, "createSumImage2DLevel");
    int i = 0;
    size /= 2;
    while(size >= 4) {
        createLevel.setArg(0, levels[i]);
        createLevel.setArg(1, levels[i+1]);
        queue.enqueueNDRangeKernel(
                createLevel,
                cl::NullRange,
                cl::NDRange(size,size),
                cl::NullRange
        );
        i++;
        size /= 2;
    }

    // Get result from the last level
    unsigned int nrOfElements = 4*4;
    unsigned int nrOfComponents = getOpenCLImageFormat(device, CL_MEM_OBJECT_IMAGE2D, TYPE_FLOAT, 1).image_channel_order == CL_RGBA ? 4 : 1;
    float* result = (float*)allocateDataArray(nrOfElements,TYPE_FLOAT,nrOfComponents);
    queue.enqueueReadImage(levels[levels.size()-1],CL_TRUE,createOrigoRegion(),createRegion(4,4,1),0,0,result);
    *sum = getSumFromOpenCLImageResult<float>(result, nrOfElements, nrOfComponents);
    delete[] result;
}

void getMaxAndMinFromOpenCLBuffer(OpenCLDevice::pointer device, cl::Buffer buffer, unsigned int size, DataType type, float* min, float* max) {
    // Compile OpenCL code
    std::string buildOptions = "";
    switch(type) {
    case TYPE_FLOAT:
        buildOptions = "-DTYPE_FLOAT";
        break;
    case TYPE_UINT8:
        buildOptions = "-DTYPE_UINT8";
        break;
    case TYPE_INT8:
        buildOptions = "-DTYPE_INT8";
        break;
    case TYPE_UINT16:
        buildOptions = "-DTYPE_UINT16";
        break;
    case TYPE_INT16:
        buildOptions = "-DTYPE_INT16";
        break;
    }
    std::string sourceFilename = std::string(FAST_SOURCE_DIR) + "/ImageMinMax.cl";
    std::string programName = sourceFilename + buildOptions;
    // Only create program if it doesn't exist for this device from before
    if(!device->hasProgram(programName))
        device->createProgramFromSourceWithName(programName, sourceFilename, buildOptions);
    cl::Program program = device->getProgram(programName);
    cl::CommandQueue queue = device->getCommandQueue();

    // Nr of work groups must be set so that work-group size does not exceed max work-group size (256 on AMD)
    int length = size;
    cl::Kernel reduce(program, "reduce");

    cl::Buffer current = buffer;
    cl::Buffer clResult;
    int workGroupSize = 256;
    int workGroups = 256;
    int X = ceil((float)length / (workGroups*workGroupSize));

    clResult = cl::Buffer(device->getContext(), CL_MEM_READ_WRITE, getSizeOfDataType(type,1)*workGroups*2);
    reduce.setArg(0, current);
    reduce.setArg(1, workGroupSize * getSizeOfDataType(type,1), NULL);
    reduce.setArg(2, workGroupSize * getSizeOfDataType(type,1), NULL);
    reduce.setArg(3, size);
    reduce.setArg(4, X);
    reduce.setArg(5, clResult);

    queue.enqueueNDRangeKernel(
            reduce,
            cl::NullRange,
            cl::NDRange(workGroups*workGroupSize),
            cl::NDRange(workGroupSize)
    );

    length = workGroups;

    void* result = allocateDataArray(length, type, 2);
    unsigned int nrOfElements = length;
    queue.enqueueReadBuffer(clResult,CL_TRUE,0,getSizeOfDataType(type,1)*workGroups*2,result);
    switch(type) {
    case TYPE_FLOAT:
        getMaxAndMinFromOpenCLImageResult<float>(result, nrOfElements, 2, min, max);
        break;
    case TYPE_INT8:
        getMaxAndMinFromOpenCLImageResult<char>(result, nrOfElements, 2, min, max);
        break;
    case TYPE_UINT8:
        getMaxAndMinFromOpenCLImageResult<uchar>(result, nrOfElements, 2, min, max);
        break;
    case TYPE_INT16:
        getMaxAndMinFromOpenCLImageResult<short>(result, nrOfElements, 2, min, max);
        break;
    case TYPE_UINT16:
        getMaxAndMinFromOpenCLImageResult<ushort>(result, nrOfElements, 2, min, max);
        break;
    }
    deleteArray(result, type);
}

void SeededRegionGrowing::waitToFinish() {
    if(!getMainDevice()->isHost()) {
        OpenCLDevice::pointer device = getMainDevice();
        device->getCommandQueue().finish();
    }
}

void SeededRegionGrowing::execute() {
    if(mSeedPoints.size() == 0)
        throw Exception("No seed points supplied to SeededRegionGrowing");

    Image::pointer input = getStaticInputData<Image>();
    if(input->getNrOfComponents() != 1)
        throw Exception("Seeded region growing currently doesn't support images with several components.");

    Segmentation::pointer output = getStaticOutputData<Segmentation>();

    // Initialize output image
    output->createFromImage(input, getMainDevice());

    if(getMainDevice()->isHost()) {
        ImageAccess::pointer inputAccess = input->getImageAccess(ACCESS_READ);
        void* inputData = inputAccess->get();
        switch(input->getDataType()) {
            fastSwitchTypeMacro(executeOnHost<FAST_TYPE>((FAST_TYPE*)inputData, output));
        }
    } else {
        OpenCLDevice::pointer device = getMainDevice();

        recompileOpenCLCode(input);

        ImageAccess::pointer access = output->getImageAccess(ACCESS_READ_WRITE);
        uchar* outputData = (uchar*)access->get();
        // Initialize to all 0s
        memset(outputData,0,sizeof(uchar)*output->getWidth()*output->getHeight()*output->getDepth());

        // Add sedd points
        for(int i = 0; i < mSeedPoints.size(); i++) {
            Vector3ui pos = mSeedPoints[i];

            // Check if seed point is in bounds
            if(pos.x() < 0 || pos.y() < 0 || pos.z() < 0 ||
                pos.x() >= output->getWidth() || pos.y() >= output->getHeight() || pos.z() >= output->getDepth())
                throw Exception("One of the seed points given to SeededRegionGrowing was out of bounds.");

            outputData[pos.x() + pos.y()*output->getWidth() + pos.z()*output->getWidth()*output->getHeight()] = 2;
        }
        access->release();

        cl::NDRange globalSize;
        if(output->getDimensions() == 2) {
            globalSize = cl::NDRange(input->getWidth(),input->getHeight());
            OpenCLImageAccess2D::pointer inputAccess = input->getOpenCLImageAccess2D(ACCESS_READ, device);
            mKernel.setArg(0, *inputAccess->get());
        } else {
            globalSize = cl::NDRange(input->getWidth(),input->getHeight(), input->getDepth());
            OpenCLImageAccess3D::pointer inputAccess = input->getOpenCLImageAccess3D(ACCESS_READ, device);
            mKernel.setArg(0, *inputAccess->get());
        }

        OpenCLBufferAccess::pointer outputAccess = output->getOpenCLBufferAccess(ACCESS_READ_WRITE, device);
        cl::Buffer stopGrowingBuffer = cl::Buffer(
                device->getContext(),
                CL_MEM_READ_WRITE,
                sizeof(char));
        cl::CommandQueue queue = device->getCommandQueue();
        mKernel.setArg(1, *outputAccess->get());
        mKernel.setArg(2, stopGrowingBuffer);
        mKernel.setArg(3, mMinimumIntensity);
        mKernel.setArg(4, mMaximumIntensity);

        bool stopGrowing = false;
        char stopGrowingInit = 1;
        char * stopGrowingResult = new char;
        int iterations = 0;
        do {
            iterations++;
            queue.enqueueWriteBuffer(stopGrowingBuffer, CL_TRUE, 0, sizeof(char), &stopGrowingInit);

            queue.enqueueNDRangeKernel(
                    mKernel,
                    cl::NullRange,
                    globalSize,
                    cl::NullRange
            );

            queue.enqueueReadBuffer(stopGrowingBuffer, CL_TRUE, 0, sizeof(char), stopGrowingResult);
            if(*stopGrowingResult == 1)
                stopGrowing = true;
        } while(!stopGrowing);
    }

}

void ImageSlicer::orthogonalSlicing(Image::pointer input, Image::pointer output) {
    OpenCLDevice::pointer device = getMainDevice();

    // Determine slice nr and width and height
    unsigned int sliceNr;
    if(mOrthogonalSliceNr < 0) {
        switch(mOrthogonalSlicePlane) {
        case PLANE_X:
            sliceNr = input->getWidth()/2;
            break;
        case PLANE_Y:
            sliceNr = input->getHeight()/2;
            break;
        case PLANE_Z:
            sliceNr = input->getDepth()/2;
            break;
        }
    } else {
        // Check that mSliceNr is valid
        sliceNr = mOrthogonalSliceNr;
        switch(mOrthogonalSlicePlane) {
        case PLANE_X:
            if(sliceNr >= input->getWidth())
                sliceNr = input->getWidth()-1;
            break;
        case PLANE_Y:
            if(sliceNr >= input->getHeight())
                sliceNr = input->getHeight()-1;
            break;
        case PLANE_Z:
            if(sliceNr >= input->getDepth())
                sliceNr = input->getDepth()-1;
            break;
        }
    }
    unsigned int slicePlaneNr, width, height;
    Vector3f spacing(0,0,0);
    switch(mOrthogonalSlicePlane) {
        case PLANE_X:
            slicePlaneNr = 0;
            width = input->getHeight();
            height = input->getDepth();
            spacing.x() = input->getSpacing().y();
            spacing.y() = input->getSpacing().z();
            break;
        case PLANE_Y:
            slicePlaneNr = 1;
            width = input->getWidth();
            height = input->getDepth();
            spacing.x() = input->getSpacing().x();
            spacing.y() = input->getSpacing().z();
            break;
        case PLANE_Z:
            slicePlaneNr = 2;
            width = input->getWidth();
            height = input->getHeight();
            spacing.x() = input->getSpacing().x();
            spacing.y() = input->getSpacing().y();
            break;
    }

    output->create(width, height, input->getDataType(), input->getNrOfComponents());
    output->setSpacing(spacing);

    OpenCLImageAccess::pointer inputAccess = input->getOpenCLImageAccess(ACCESS_READ, device);
    OpenCLImageAccess::pointer outputAccess = output->getOpenCLImageAccess(ACCESS_READ_WRITE, device);

	cl::CommandQueue queue = device->getCommandQueue();
	cl::Program program = getOpenCLProgram(device);
	cl::Kernel kernel(program, "orthogonalSlicing");

    kernel.setArg(0, *inputAccess->get3DImage());
    kernel.setArg(1, *outputAccess->get2DImage());
    kernel.setArg(2, sliceNr);
    kernel.setArg(3, slicePlaneNr);
    queue.enqueueNDRangeKernel(
            kernel,
            cl::NullRange,
            cl::NDRange(width, height),
            cl::NullRange
    );

    // TODO set scene graph transformation
}

void
SegmentationRenderer::draw(Matrix4f perspectiveMatrix, Matrix4f viewingMatrix, float zNear, float zFar, bool mode2D) {
    std::lock_guard<std::mutex> lock(mMutex);
    OpenCLDevice::pointer device = std::dynamic_pointer_cast<OpenCLDevice>(getMainDevice());


    if(mColorsModified) {
        // Transfer colors to device (this doesn't have to happen every render call..)
        std::unique_ptr<float[]> colorData(new float[3*mLabelColors.size()]);
        std::unordered_map<int, Color>::iterator it;
        for(it = mLabelColors.begin(); it != mLabelColors.end(); it++) {
            colorData[it->first*3] = it->second.getRedValue();
            colorData[it->first*3+1] = it->second.getGreenValue();
            colorData[it->first*3+2] = it->second.getBlueValue();
        }

        mColorBuffer = cl::Buffer(
                device->getContext(),
                CL_MEM_READ_ONLY | CL_MEM_COPY_HOST_PTR,
                sizeof(float)*3*mLabelColors.size(),
                colorData.get()
        );
    }

    if(mFillAreaModified) {
        // Transfer colors to device (this doesn't have to happen every render call..)
        std::unique_ptr<char[]> fillAreaData(new char[mLabelColors.size()]);
        std::unordered_map<int, Color>::iterator it;
        for(it = mLabelColors.begin(); it != mLabelColors.end(); it++) {
            if(mLabelFillArea.count(it->first) == 0) {
                // Use default value
                fillAreaData[it->first] = mFillArea;
            } else {
                fillAreaData[it->first] = mLabelFillArea[it->first];
            }
        }

        mFillAreaBuffer = cl::Buffer(
                device->getContext(),
                CL_MEM_READ_ONLY | CL_MEM_COPY_HOST_PTR,
                sizeof(char)*mLabelColors.size(),
                fillAreaData.get()
        );
    }

    mKernel = cl::Kernel(getOpenCLProgram(device), "renderToTexture");
    mKernel.setArg(2, mColorBuffer);
    mKernel.setArg(3, mFillAreaBuffer);
    mKernel.setArg(4, mBorderRadius);
    mKernel.setArg(5, mOpacity);


    for(auto it : mDataToRender) {
        Image::pointer input = std::static_pointer_cast<Image>(it.second);
        uint inputNr = it.first;

        if(input->getDimensions() != 2)
            throw Exception("SegmentationRenderer only supports 2D images. Use ImageSlicer to extract a 2D slice from a 3D image.");

        if(input->getDataType() != TYPE_UINT8)
            throw Exception("SegmentationRenderer only support images with dat type uint8.");

        // Check if a texture has already been created for this image
        if(mTexturesToRender.count(inputNr) > 0 && mImageUsed[inputNr] == input)
            continue; // If it has already been created, skip it

        // If it has not been created, create the texture

        OpenCLImageAccess::pointer access = input->getOpenCLImageAccess(ACCESS_READ, device);
        cl::Image2D *clImage = access->get2DImage();

        // Run kernel to fill the texture
        cl::CommandQueue queue = device->getCommandQueue();

        if (mTexturesToRender.count(inputNr) > 0) {
            // Delete old texture
            glDeleteTextures(1, &mTexturesToRender[inputNr]);
            mTexturesToRender.erase(inputNr);
            glDeleteVertexArrays(1, &mVAO[inputNr]);
            mVAO.erase(inputNr);
        }

        cl::Image2D image;
        cl::ImageGL imageGL;
        std::vector<cl::Memory> v;
        GLuint textureID;
        // TODO The GL-CL interop here is causing glClear to not work on AMD systems and therefore disabled
        /*
        if(DeviceManager::isGLInteropEnabled()) {
            // Create OpenGL texture
            glGenTextures(1, &textureID);
            glBindTexture(GL_TEXTURE_2D, textureID);
            glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_LINEAR);
            glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_LINEAR);
            glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA32F, input->getWidth(), input->getHeight(), 0, GL_RGBA, GL_FLOAT, 0);

            // Create CL-GL image
            imageGL = cl::ImageGL(
                    device->getContext(),
                    CL_MEM_READ_WRITE,
//.........这里部分代码省略.........