ExecutionStatus execute(NodeSocketReader& reader, NodeSocketWriter& writer) override
{
const clw::Image2D& input = reader.readSocket(0).getDeviceImageMono();
clw::Image2D& output = writer.acquireSocket(0).getDeviceImageMono();
int imageWidth = input.width();
int imageHeight = input.height();
if(imageWidth == 0 || imageHeight == 0)
return ExecutionStatus(EStatus::Ok);
clw::Kernel kernelConvertBayer2Gray;
switch(_BayerCode.cast<Enum>().cast<cvu::EBayerCode>())
{
case cvu::EBayerCode::RG:
kernelConvertBayer2Gray = _gpuComputeModule->acquireKernel(_kidConvertRG2Gray);
break;
case cvu::EBayerCode::GB:
kernelConvertBayer2Gray = _gpuComputeModule->acquireKernel(_kidConvertGB2Gray);
break;
case cvu::EBayerCode::GR:
kernelConvertBayer2Gray = _gpuComputeModule->acquireKernel(_kidConvertGR2Gray);
break;
case cvu::EBayerCode::BG:
kernelConvertBayer2Gray = _gpuComputeModule->acquireKernel(_kidConvertBG2Gray);
break;
}
if(kernelConvertBayer2Gray.isNull())
return ExecutionStatus(EStatus::Error, "Bad bayer code");
// Ensure output image size is enough
if(output.isNull() || output.width() != imageWidth || output.height() != imageHeight)
{
output = _gpuComputeModule->context().createImage2D(
clw::EAccess::ReadWrite, clw::EMemoryLocation::Device,
clw::ImageFormat(clw::EChannelOrder::R, clw::EChannelType::Normalized_UInt8),
imageWidth, imageHeight);
}
cl_float3 gains = { (float) _redGain, (float) _greenGain, (float) _blueGain };
int sharedWidth = 16 + 2;
int sharedHeight = 16 + 2;
kernelConvertBayer2Gray.setLocalWorkSize(16, 16);
kernelConvertBayer2Gray.setRoundedGlobalWorkSize(imageWidth, imageHeight);
kernelConvertBayer2Gray.setArg(0, input);
kernelConvertBayer2Gray.setArg(1, output);
kernelConvertBayer2Gray.setArg(2, gains);
kernelConvertBayer2Gray.setArg(3, clw::LocalMemorySize(sharedWidth*sharedHeight*sizeof(float)));
kernelConvertBayer2Gray.setArg(4, sharedWidth);
kernelConvertBayer2Gray.setArg(5, sharedHeight);
_gpuComputeModule->queue().asyncRunKernel(kernelConvertBayer2Gray);
_gpuComputeModule->queue().finish();
return ExecutionStatus(EStatus::Ok);
}
ExecutionStatus execute(NodeSocketReader& reader, NodeSocketWriter& writer) override
{
// Read input sockets
const cv::Mat& src = reader.readSocket(0).getImageMono();
// ouputs
KeyPoints& kp = writer.acquireSocket(0).getKeypoints();
cv::Mat& descriptors = writer.acquireSocket(1).getArray();
// Validate inputs
if(src.empty())
return ExecutionStatus(EStatus::Ok);
(*_brisk)(src, cv::noArray(), kp.kpoints, descriptors);
kp.image = src;
return ExecutionStatus(EStatus::Ok,
string_format("Keypoints detected: %d", (int) kp.kpoints.size()));
}
ExecutionStatus execute(NodeSocketReader& reader, NodeSocketWriter& writer) override
{
// Read input sockets
const KeyPoints& kp = reader.readSocket(0).getKeypoints();
// Validate inputs
if(kp.kpoints.empty() || kp.image.empty())
return ExecutionStatus(EStatus::Ok);
// Acquire output sockets
KeyPoints& outKp = writer.acquireSocket(0).getKeypoints();
cv::Mat& outDescriptors = writer.acquireSocket(1).getArray();
outKp = kp;
_brisk->compute(kp.image, outKp.kpoints, outDescriptors);
return ExecutionStatus(EStatus::Ok);
}
ExecutionStatus execute(NodeSocketReader& reader, NodeSocketWriter& writer) override
{
// Read input sockets
const cv::Mat& src = reader.readSocket(0).getImageMono();
// Acquire output sockets
KeyPoints& kp = writer.acquireSocket(0).getKeypoints();
cv::Mat& descriptors = writer.acquireSocket(1).getArray();
// Validate inputs
if(src.empty())
return ExecutionStatus(EStatus::Ok);
// Do stuff
cv::SIFT sift(_nFeatures, _nOctaveLayers,
_contrastThreshold, _edgeThreshold, _sigma);
sift(src, cv::noArray(), kp.kpoints, descriptors);
kp.image = src;
return ExecutionStatus(EStatus::Ok,
string_format("Keypoints detected: %d", (int) kp.kpoints.size()));
}
ExecutionStatus execute(NodeSocketReader& reader, NodeSocketWriter& writer) override
{
// Read input sockets
const cv::Mat& src = reader.readSocket(0).getImageMono();
// Acquire output sockets
KeyPoints& kp = writer.acquireSocket(0).getKeypoints();
// Validate inputs
if(src.empty())
return ExecutionStatus(EStatus::Ok);
_brisk->detect(src, kp.kpoints);
kp.image = src;
return ExecutionStatus(EStatus::Ok,
string_format("Keypoints detected: %d", (int) kp.kpoints.size()));
}
ExecutionStatus execute(NodeSocketReader& reader, NodeSocketWriter& writer) override
{
const clw::Image2D& deviceImage = reader.readSocket(0).getDeviceImageMono();
KeyPoints& kp = writer.acquireSocket(0).getKeypoints();
cv::Mat& descriptors = writer.acquireSocket(1).getArray();
DeviceArray& descriptors_dev = writer.acquireSocket(2).getDeviceArray();
int imageWidth = deviceImage.width();
int imageHeight = deviceImage.height();
if(imageWidth == 0 || imageHeight == 0)
return ExecutionStatus(EStatus::Ok);
GpuPerformanceMarker marker(_gpuComputeModule->activityLogger(), "SURF");
if(!_constantsUploaded)
uploadSurfConstants();
ensureKeypointsBufferIsEnough();
// Zero keypoints counter (use pinned memory)
// Another way would be to map pinned buffer (allocated on a host)
// and read the device buffer using just obtained pointer. If data transfer is single-direction
// we don't need to map and unmap every time - just after creation and before destruction of a buffer
// From AMD APP OpenCL Programming Guide:
// pinnedBuffer = clCreateBuffer(CL_MEM_ALLOC_HOST_PTR or CL_MEM_USE_HOST_PTR)
// deviceBuffer = clCreateBuffer()
// 1)
// void* pinnedMemory = clEnqueueMapBuffer(pinnedBuffer) -> can be done only once
// clEnqueueRead/WriteBuffer(deviceBuffer, pinnedMemory)
// clEnqueueUnmapMemObject(pinnedBuffer, pinnedMemory) -> can be done only once
//
// 2)
// void* pinnedMemory = clEnqueueMapBuffer(pinnedBuffer)
// [Application writes or modifies memory (host memory bandwith)]
// clEnqueueUnmapMemObject(pinnedBuffer, pinnedMemory)
// clEnqueueCopyBuffer(pinnedBuffer, deviceBuffer)
// - or -
// clEnqueueCopyBuffer(deviceBuffer, pinnedBuffer)
// void* pinnedMemory = clEnqueueMapBuffer(pinnedBuffer)
// [Application reads memory (host memory bandwith)]
// clEnqueueUnmapMemObject(pinnedBuffer, pinnedMemory)
// On AMD these are the same solutions
int* intPtr = (int*) _gpuComputeModule->queue().mapBuffer(_pinnedKeypointsCount_cl, clw::EMapAccess::Write);
if(!intPtr)
return ExecutionStatus(EStatus::Error, "Couldn't mapped keypoints counter buffer to host address space");
intPtr[0] = 0;
_gpuComputeModule->queue().asyncUnmap(_pinnedKeypointsCount_cl, intPtr);
_gpuComputeModule->queue().asyncCopyBuffer(_pinnedKeypointsCount_cl, _keypointsCount_cl);
convertImageToIntegral(deviceImage, imageWidth, imageHeight);
prepareScaleSpaceLayers(imageWidth, imageHeight);
buildScaleSpace(imageWidth, imageHeight);
findScaleSpaceMaxima();
kp.image.create(imageHeight, imageWidth, CV_8UC1);
_gpuComputeModule->dataQueue().asyncReadImage2D(deviceImage, kp.image.data, (int) kp.image.step);
_gpuComputeModule->dataQueue().flush();
_gpuComputeModule->activityLogger().beginPerfMarker("Read number of keypoints", "SURF");
// Read keypoints counter (use pinned memory)
_gpuComputeModule->queue().asyncCopyBuffer(_keypointsCount_cl, _pinnedKeypointsCount_cl);
intPtr = (int*) _gpuComputeModule->queue().mapBuffer(_pinnedKeypointsCount_cl, clw::EMapAccess::Read);
if(!intPtr)
return ExecutionStatus(EStatus::Error, "Couldn't mapped keypoints counter buffer to host address space");
int keypointsCount = min(intPtr[0], kKeypointsMax);
_gpuComputeModule->queue().asyncUnmap(_pinnedKeypointsCount_cl, intPtr);
_gpuComputeModule->activityLogger().endPerfMarker();
if(keypointsCount > 0)
{
if(!_upright)
findKeypointOrientation(keypointsCount);
else
uprightKeypointOrientation(keypointsCount);
calculateDescriptors(keypointsCount);
GpuPerformanceMarker marker(_gpuComputeModule->activityLogger(), "Read results", "SURF");
// Start copying descriptors to pinned buffer
if(_downloadDescriptors)
_gpuComputeModule->queue().asyncCopyBuffer(_descriptors_cl, _pinnedDescriptors_cl);
vector<KeyPoint> kps = downloadKeypoints(keypointsCount);
kp.kpoints = transformKeyPoint(kps);
descriptors_dev = DeviceArray::createFromBuffer(_descriptors_cl, 64, keypointsCount, EDataType::Float);
if(_downloadDescriptors)
{
descriptors.create(keypointsCount, 64, CV_32F);
float* floatPtr = (float*) _gpuComputeModule->queue().mapBuffer(_pinnedDescriptors_cl, clw::EMapAccess::Read);
if(!floatPtr)
return ExecutionStatus(EStatus::Error, "Couldn't mapped descriptors buffer to host address space");
if(descriptors.step == 64*sizeof(float))
{
//copy(floatPtr, floatPtr + 64*keypointsCount, descriptors.ptr<float>());
memcpy(descriptors.ptr<float>(), floatPtr, sizeof(float)*64 * keypointsCount);
}
else
{
for(int row = 0; row < keypointsCount; ++row)
//copy(floatPtr + 64*row, floatPtr + 64*row + 64, descriptors.ptr<float>(row));
memcpy(descriptors.ptr<float>(row), floatPtr + 64*row, sizeof(float)*64);
}
_gpuComputeModule->queue().asyncUnmap(_pinnedDescriptors_cl, floatPtr);
}
// Finish downloading input image
_gpuComputeModule->dataQueue().finish();
}
else
{
// Finish downloading input image
_gpuComputeModule->dataQueue().finish();
kp.kpoints = vector<cv::KeyPoint>();
descriptors = cv::Mat();
descriptors_dev = DeviceArray();
}
return ExecutionStatus(EStatus::Ok,
string_format("Keypoints detected: %d", (int) kp.kpoints.size()));
}
ExecutionStatus execute(NodeSocketReader& reader, NodeSocketWriter& writer) override
{
const clw::Image2D& deviceImage = reader.readSocket(0).getDeviceImageMono();
clw::Image2D& deviceDest = writer.acquireSocket(0).getDeviceImageMono();
int srcWidth = deviceImage.width();
int srcHeight = deviceImage.height();
if(srcWidth == 0 || srcHeight == 0)
return ExecutionStatus(EStatus::Ok);
clw::Kernel kernelGaussMix = _gpuComputeModule->acquireKernel(_kidGaussMix);
/*
Create mixture data buffer
*/
resetMixturesState(srcWidth * srcHeight);
if(deviceDest.isNull()
|| deviceDest.width() != srcWidth
|| deviceDest.height() != srcHeight)
{
// Obraz (w zasadzie maska) pierwszego planu
deviceDest = _gpuComputeModule->context().createImage2D(
clw::EAccess::ReadWrite, clw::EMemoryLocation::Device,
clw::ImageFormat(clw::EChannelOrder::R, clw::EChannelType::Normalized_UInt8),
srcWidth, srcHeight);
}
// Calculate dynamic learning rate (if necessary)
++_nframe;
float alpha = _learningRate >= 0 && _nframe > 1
? _learningRate
: 1.0f/std::min(_nframe, _history.cast<int>());
kernelGaussMix.setLocalWorkSize(16, 16);
kernelGaussMix.setRoundedGlobalWorkSize(srcWidth, srcHeight);
kernelGaussMix.setArg(0, deviceImage);
kernelGaussMix.setArg(1, deviceDest);
kernelGaussMix.setArg(2, _mixtureDataBuffer);
kernelGaussMix.setArg(3, _mixtureParamsBuffer);
kernelGaussMix.setArg(4, alpha);
_gpuComputeModule->queue().asyncRunKernel(kernelGaussMix);
if(_showBackground)
{
clw::Image2D& deviceDestBackground = writer.acquireSocket(1).getDeviceImageMono();
if(deviceDestBackground.isNull()
|| deviceDestBackground.width() != srcWidth
|| deviceDestBackground.height() != srcHeight)
{
// Obraz (w zasadzie maska) pierwszego planu
deviceDestBackground = _gpuComputeModule->context().createImage2D(
clw::EAccess::ReadWrite, clw::EMemoryLocation::Device,
clw::ImageFormat(clw::EChannelOrder::R, clw::EChannelType::Normalized_UInt8),
srcWidth, srcHeight);
}
clw::Kernel kernelBackground = _gpuComputeModule->acquireKernel(_kidGaussBackground);
kernelBackground.setLocalWorkSize(16, 16);
kernelBackground.setRoundedGlobalWorkSize(srcWidth, srcHeight);
kernelBackground.setArg(0, deviceDestBackground);
kernelBackground.setArg(1, _mixtureDataBuffer);
kernelBackground.setArg(2, _mixtureParamsBuffer);
_gpuComputeModule->queue().asyncRunKernel(kernelBackground);
}
_gpuComputeModule->queue().finish();
return ExecutionStatus(EStatus::Ok);
}
