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 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 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
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,
GL_TEXTURE_2D,
0,
textureID
);
glBindTexture(GL_TEXTURE_2D, 0);
glFinish();
mKernel.setArg(1, imageGL);
v.push_back(imageGL);
queue.enqueueAcquireGLObjects(&v);
} else {
*/
image = cl::Image2D(
device->getContext(),
CL_MEM_READ_WRITE,
cl::ImageFormat(CL_RGBA, CL_FLOAT),
input->getWidth(), input->getHeight()
);
mKernel.setArg(1, image);
//}
mKernel.setArg(0, *clImage);
queue.enqueueNDRangeKernel(
mKernel,
cl::NullRange,
cl::NDRange(input->getWidth(), input->getHeight()),
cl::NullRange
);
/*if(DeviceManager::isGLInteropEnabled()) {
queue.enqueueReleaseGLObjects(&v);
} else {*/
// Copy data from CL image to CPU
auto data = make_uninitialized_unique<float[]>(input->getWidth() * input->getHeight() * 4);
queue.enqueueReadImage(
image,
CL_TRUE,
createOrigoRegion(),
createRegion(input->getWidth(), input->getHeight(), 1),
0, 0,
data.get()
);
// Copy data from CPU to GL 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, data.get());
glBindTexture(GL_TEXTURE_2D, 0);
glFinish();
//}
mTexturesToRender[inputNr] = textureID;
mImageUsed[inputNr] = input;
queue.finish();
}
glEnable(GL_BLEND);
glBlendFunc(GL_SRC_ALPHA, GL_ONE_MINUS_SRC_ALPHA);
drawTextures(perspectiveMatrix, viewingMatrix, mode2D);
glDisable(GL_BLEND);
}
