Skip to content
GitLab
Commit ace48281 authored by Declara Denis's avatar Declara Denis Committed by Christian Schulte zu Berge
Browse files

Added time driven Monitor to CM Solver

parent 57fcba2d
Hide whitespace changes
Inline Side-by-side
modules/cudaconfidencemaps/core/cudaconfidencemaps_cuda.cu
View file @ ace48281
#include "cudaconfidencemaps_cuda.h"
#include "cudautils.h"
#include "cuspmonitors.h"
#include <cusp/blas.h>
#include <cusp/dia_matrix.h>
......@@ -69,32 +70,6 @@ namespace cuda {
float systemSolveTime;
};
template <typename ValueType>
class iteration_monitor : public cusp::default_monitor<ValueType>
{
typedef typename cusp::norm_type<ValueType>::type Real;
typedef cusp::default_monitor<ValueType> super;
public:
template <typename Vector>
iteration_monitor(const Vector& b, size_t iteration_limit = 500)
: super(b, iteration_limit, 0.0f, 0.0f)
{ }
template <typename Vector>
bool finished(const Vector& r)
{
// Only if the maximum iteration count has been reached, actually go ahead and
// compute the error
if (super::iteration_count() >= super::iteration_limit()) {
super::r_norm = cusp::blas::nrm2(r);
return true;
}
return false;
}
};
CudaConfidenceMapsSystemSolver::CudaConfidenceMapsSystemSolver()
: _gpuData(new CudaConfidenceMapsSystemGPUData())
{
......@@ -173,12 +148,12 @@ namespace cuda {
}
// FIXME: Remove errorTolerance parameter
void CudaConfidenceMapsSystemSolver::solve(int maximumIterations, float errorTolerance) {
void CudaConfidenceMapsSystemSolver::solve(float millisecondBudget) {
// Measure execution time and record it in the _gpuData datastructure
CUDAClock clock; clock.start();
// The solution is computed using Conjugate Gradient with a Diagonal (Jacobi) preconditioner
iteration_monitor<float> monitor(_gpuData->b_d, maximumIterations);
deadline_monitor<float> monitor(_gpuData->b_d, millisecondBudget);
cusp::precond::diagonal<float, cusp::device_memory> M(_gpuData->L_d);
cusp::krylov::cg(_gpuData->L_d, _gpuData->x_d, _gpuData->b_d, monitor, M);
_gpuData->solutionResidualNorm = monitor.residual_norm();
......
modules/cudaconfidencemaps/core/cudaconfidencemaps_cuda.h
View file @ ace48281
......@@ -58,10 +58,9 @@ namespace cuda {
/**
* After calling \see uploadImage(), this functions launches a solver on the GPU that will solve
* the diffusion problem.
* \param maximumIterations maximum number of iterations the solver will preform
* \param errorTolerance if the solution error sinks below this value, the solver stops early
* \param millisecondBudget the time budget the solver has to come up with a solution.
*/
void solve(int maximumIterations, float errorTolerance);
void solve(float millisecondBudget);
/**
* Returns a host buffer of the last solution computed by the solver. The pointer is guaranteed to
......
modules/cudaconfidencemaps/core/cuspmonitors.h 0 → 100644
View file @ ace48281
#ifndef CUSPMONITORS_H__
#define CUSPMONITORS_H__
#include <cusp/monitor.h>
#include <tbb/tick_count.h>
namespace campvis {
namespace cuda {
/**
* This class allows to execute a fixed number of conjugate gradient iterations
* in CUSP. Unlike the default_monitor, this class only calculates the residual
* norm when the iteration count is reached.
*/
template <typename ValueType>
class iteration_monitor : public cusp::default_monitor<ValueType>
{
typedef typename cusp::norm_type<ValueType>::type Real;
typedef cusp::default_monitor<ValueType> super;
public:
template <typename Vector>
iteration_monitor(const Vector& b, size_t iteration_limit = 500)
: super(b, iteration_limit, 0.0f, 0.0f)
{ }
template <typename Vector>
bool finished(const Vector& r)
{
// Only if the maximum iteration count has been reached, actually go ahead and
// compute the error
if (super::iteration_count() >= super::iteration_limit()) {
super::r_norm = cusp::blas::nrm2(r);
return true;
}
return false;
}
};
/**
* This monitor allows to set a deadline, after which the computation has to stop.
*/
template <typename ValueType>
class deadline_monitor : public cusp::default_monitor<ValueType>
{
typedef typename cusp::norm_type<ValueType>::type Real;
typedef cusp::default_monitor<ValueType> super;
public:
template <typename Vector>
deadline_monitor(const Vector& b, float milliseconds)
: super(b, 0, 0.0f, 0.0f), _startTime(tbb::tick_count::now()), _seconds(milliseconds / 1000.0f)
{ }
template <typename Vector>
bool finished(const Vector& r)
{
// Only if the deadline is reached, stop and compute the error
if ((tbb::tick_count::now() - _startTime).seconds() > _seconds) {
super::r_norm = cusp::blas::nrm2(r);
return true;
}
return false;
}
private:
tbb::tick_count _startTime;
float _seconds;
};
}
}
#endif // CUSPMONITORS_H__
\ No newline at end of file
modules/cudaconfidencemaps/cudaconfidencemaps.cmake
View file @ ace48281
......@@ -47,7 +47,10 @@ IF(ModuleEnabled)
ENDIF()
# CUSP Include directory
CUDA_INCLUDE_DIRECTORIES(${ThisModDir}/ext/cusplibrary-0.4.0)
CUDA_INCLUDE_DIRECTORIES(
${ThisModDir}/ext/cusplibrary-0.4.0
${TBB_INCLUDE_DIR}
)
# Build CUDA portion of the code (STATICALLY!?)
FILE(GLOB cuda_SOURCES modules/cudaconfidencemaps/core/*.cu)
......@@ -55,10 +58,6 @@ IF(ModuleEnabled)
${cuda_SOURCES}
)
# Make sure code can find the CUSP include files included with this module
#SET(CUDA_NVCC_FLAGS ${CUDA_NVCC_FLAGS};"-I ${ThisModDir}/ext/cusplibrary-0.4.0")
set(CUDA_NVCC_FLAGS "ajkladjfl" CACHE STRING "adsf")
# Link CUDA code to module
LIST(APPEND ThisModExternalLibs cudaconfidencemaps-cuda)
ELSE()
......
modules/cudaconfidencemaps/pipelines/cudaconfidencemapsdemo.cpp
View file @ ace48281
......@@ -30,18 +30,17 @@
namespace campvis {
CudaConfidenceMapsDemo::CudaConfidenceMapsDemo(DataContainer* dc)
: AutoEvaluationPipeline(dc)
, _usIgtlReader()
, _usCropFilter(&_canvasSize)
CudaConfidenceMapsDemo::CudaConfidenceMapsDemo(DataContainer* dc)
: AutoEvaluationPipeline(dc)
, _usIgtlReader()
, _usCropFilter(&_canvasSize)
, _usBlurFilter(&_canvasSize)
, _usResampler(&_canvasSize)
, _usMapsSolver()
, _usFusion(&_canvasSize)
, _usFanRenderer(&_canvasSize)
, p_iterations("Iterations", "Number of CG Iterations", 150, 1, 500)
, p_autoIterationCount("AutoIterationCount", "Estimate iteration count based on time slot", true)
, p_timeSlot("TimeSlot", "Milliseconds per frame", 32.0f, 10.0f, 250.0f)
, p_millisecondBudget("MillisecondBudget", "Milliseconds per frame", 32.0f, 10.0f, 1000.0f)
, p_connectToIGTLinkServer("ConnectToIGTLink", "Connect/Disconnect to IGTLink")
, p_gaussianFilterSize("GaussianSigma", "Blur amount", 2.5f, 1.0f, 10.0f)
, p_resamplingScale("ResampleScale", "Resample Scale", 0.25f, 0.01f, 1.0f)
......@@ -52,21 +51,18 @@ namespace campvis {
, p_useAlphaBetaFilter("UseAlphaBetaFilter", "Alpha-Beta-Filter", true)
, p_fanHalfAngle("FanHalfAngle", "Fan Half Angle", 37.0f, 1.0f, 90.0f)
, p_fanInnerRadius("FanInnerRadius", "Fan Inner Radius", 0.222f, 0.001f, 0.999f)
, _cgIterationsPerMsRunningAverage(1.0f)
, _cgTimeslotRunningAverage(1.0f)
, _statisticsLastUpdateTime()
{
addProcessor(&_usIgtlReader);
addProcessor(&_usCropFilter);
addProcessor(&_usIgtlReader);
addProcessor(&_usCropFilter);
addProcessor(&_usBlurFilter);
addProcessor(&_usResampler);
addProcessor(&_usMapsSolver);
addProcessor(&_usFusion);
addProcessor(&_usFanRenderer);
addProperty(p_iterations);
addProperty(p_autoIterationCount);
addProperty(p_timeSlot);
addProperty(p_millisecondBudget);
addProperty(p_connectToIGTLinkServer);
addProperty(p_gaussianFilterSize);
addProperty(p_resamplingScale);
......@@ -98,8 +94,8 @@ namespace campvis {
// Create connectors
_usIgtlReader.p_targetImagePrefix.setValue("us.igtl.");
_usCropFilter.p_inputImage.setValue("us.igtl.CAMPUS");
_usCropFilter.p_outputImage.setValue("us");
_usCropFilter.p_inputImage.setValue("us.igtl.CAMPUS");
_usCropFilter.p_outputImage.setValue("us");
_usBlurFilter.p_inputImage.setValue("us");
_usBlurFilter.p_outputImage.setValue("us.blurred");
......@@ -125,9 +121,6 @@ namespace campvis {
_renderTargetID.setValue("us.fused_fan");
// Bind pipeline proeprties to processor properties
p_iterations.addSharedProperty(&_usMapsSolver.p_iterations);
//p_autoIterationCount.addSharedProperty();
//p_timeSlot.addSharedProperty();
p_connectToIGTLinkServer.addSharedProperty(&_usIgtlReader.p_connect);
p_gaussianFilterSize.addSharedProperty(&_usBlurFilter.p_sigma);
p_resamplingScale.addSharedProperty(&_usResampler.p_resampleScale);
......@@ -151,6 +144,7 @@ namespace campvis {
else {
// Only launch the pipeline if the IgtlReader has recieved new data
if (!_usIgtlReader.isValid()) {
float millisecondBudget = p_millisecondBudget.getValue();
auto startTime = tbb::tick_count::now();
// Make sure that the whole pipeline gets invalidated
......@@ -160,12 +154,13 @@ namespace campvis {
_usMapsSolver.invalidate(AbstractProcessor::INVALID_RESULT);
_usFusion.invalidate(AbstractProcessor::INVALID_RESULT);
executeProcessorAndCheckOpenGLState(&_usIgtlReader);
executeProcessorAndCheckOpenGLState(&_usCropFilter);
executeProcessorAndCheckOpenGLState(&_usBlurFilter);
executeProcessorAndCheckOpenGLState(&_usIgtlReader);
executeProcessorAndCheckOpenGLState(&_usCropFilter);
executeProcessorAndCheckOpenGLState(&_usBlurFilter) ;
executeProcessorAndCheckOpenGLState(&_usResampler);
auto solverStartTime = tbb::tick_count::now();
_usMapsSolver.p_millisecondBudget.setValue(millisecondBudget);
executeProcessorAndCheckOpenGLState(&_usMapsSolver);
auto solverEndTime = tbb::tick_count::now();
......@@ -186,24 +181,6 @@ namespace campvis {
string << "Error: " << _usMapsSolver.getResidualNorm() << std::endl;
_usFanRenderer.p_text.setValue(string.str());
}
auto ms = (solverEndTime - solverStartTime).seconds() * 1000.0f;
auto iterationsPerMs = _usMapsSolver.getActualConjugentGradientIterations() / ms;
// Factor out the time needed for pre- and postprocessing the image from the time slot
float timeSlot = (p_timeSlot.getValue() -
(solverStartTime - startTime).seconds() * 1000.0f -
(endTime - solverEndTime).seconds() * 1000.0f);
// Compute the running average using an exponential filter
float expAlpha = 0.05f;
_cgIterationsPerMsRunningAverage = _cgIterationsPerMsRunningAverage * (1.0f - expAlpha) + iterationsPerMs * expAlpha;
_cgTimeslotRunningAverage = _cgTimeslotRunningAverage * (1.0f - expAlpha) + timeSlot * expAlpha;
int iterations = static_cast<int>(_cgTimeslotRunningAverage * _cgIterationsPerMsRunningAverage);
p_iterations.setValue(iterations);
}
}
}
......
modules/cudaconfidencemaps/pipelines/cudaconfidencemapsdemo.h
View file @ ace48281
......@@ -72,7 +72,7 @@ namespace campvis {
protected:
OpenIGTLinkClient _usIgtlReader;
GlImageCrop _usCropFilter;
GlImageCrop _usCropFilter;
GlGaussianFilter _usBlurFilter;
GlImageResampler _usResampler;
CudaConfidenceMapsSolver _usMapsSolver;
......@@ -80,9 +80,8 @@ namespace campvis {
UsFanRenderer _usFanRenderer;
NumericProperty<int> p_iterations;
BoolProperty p_autoIterationCount;
FloatProperty p_timeSlot;
FloatProperty p_millisecondBudget;
ButtonProperty p_connectToIGTLinkServer;
......@@ -98,9 +97,6 @@ namespace campvis {
FloatProperty p_fanHalfAngle;
FloatProperty p_fanInnerRadius;
float _cgIterationsPerMsRunningAverage;
float _cgTimeslotRunningAverage;
tbb::tick_count _statisticsLastUpdateTime;
};
......
modules/cudaconfidencemaps/processors/cudaconfidencemapssolver.cpp
View file @ ace48281
......@@ -41,7 +41,7 @@ namespace campvis {
, p_outputConfidenceMap("OutputConfidenceMap", "Output Confidence Map", "us.confidence", DataNameProperty::WRITE)
, p_resetResult("ResetSolution", "Reset solution vector")
, p_use8Neighbourhood("Use8Neighbourhood", "Use 8 Neighbourhood (otherwise 4)", true)
, p_iterations("IterationCount", "Conjugate Gradient Iterations", 200, 1, 500)
, p_millisecondBudget("IterationCount", "Conjugate Gradient Iterations", 25.0f, 1.0f, 1000.0f)
, p_gradientScaling("GradientScaling", "Scaling factor for gradients", 2.0f, 0.001f, 10.0f)
, p_paramAlpha("Alpha", "Alpha (TGC)", 2.0f, 0.001f, 10.0f)
, p_paramBeta("Beta", "Beta (Weight mapping)", 20.0f, 0.001f, 200.0f)
......@@ -57,7 +57,7 @@ namespace campvis {
addProperty(p_resetResult);
addProperty(p_use8Neighbourhood);
addProperty(p_iterations);
addProperty(p_millisecondBudget);
addProperty(p_gradientScaling);
addProperty(p_paramAlpha);
addProperty(p_paramBeta);
......@@ -84,7 +84,7 @@ namespace campvis {
ImageRepresentationLocal::ScopedRepresentation img(data, p_inputImage.getValue());
if (img != 0) {
bool use8Neighbourhood = p_use8Neighbourhood.getValue();
int iterations = p_iterations.getValue();
float millisecondBudget = p_millisecondBudget.getValue();
float gradientScaling = p_gradientScaling.getValue();
float alpha = p_paramAlpha.getValue();
float beta = p_paramBeta.getValue();
......@@ -99,7 +99,7 @@ namespace campvis {
auto image = (unsigned char*)img->getWeaklyTypedPointer()._pointer;
_solver.uploadImage(image, size.x, size.y, gradientScaling, alpha, beta, gamma, use8Neighbourhood);
_solver.solve(iterations, 1e-10f);
_solver.solve(millisecondBudget);
const float *solution = _solver.getSolution(size.x, size.y);
......
modules/cudaconfidencemaps/processors/cudaconfidencemapssolver.h
View file @ ace48281
......@@ -90,7 +90,7 @@ namespace campvis {
BoolProperty p_use8Neighbourhood; ///< Wether to use 8- or 4-neighbourhood
NumericProperty<int> p_iterations; ///< Number of CG-Iterations to do
FloatProperty p_millisecondBudget; ///< Maximum number of ms the solver can run
FloatProperty p_gradientScaling;
FloatProperty p_paramAlpha;
......
modules/cudaconfidencemaps/processors/usfanrenderer.cpp
View file @ ace48281
......@@ -71,7 +71,7 @@ namespace campvis {
// Creates the grid, with the origin at the center of the top edge, with the +y axis representing depth
_grid = GeometryDataFactory::createGrid(cgt::vec3(-0.5f, 1.0f, 0.0f), cgt::vec3(0.5f, 0.0f, 0.0f),
cgt::vec3(0.0f, 1.0f, 0.0f), cgt::vec3(1.0f, 0.0f, 0.0f),
16, 4);
32, 32);
// Initialize font rendering
updateFontAtlas();
......
modules/openigtlink/processors/openigtlinkclient.cpp
View file @ ace48281
......@@ -160,7 +160,7 @@ namespace campvis {
// If the voxel size boundled with the packet is practically 0.0f, make it 1.0f
// this makes sure we don't get invalid mapping informations (non invertable matrix)
if (minElem(voxelSize) <= 1e-10f) {
voxelSize = 1.0f;
voxelSize = cgt::vec3(1.0f);
}
size_t dimensionality = (size_i[2] == 1) ? ((size_i[1] == 1) ? 1 : 2) : 3;
......
modules/preprocessing/glsl/glimageresampler.frag
View file @ ace48281
......@@ -35,7 +35,7 @@ uniform sampler2D _texture;
#endif
void main() {
#ifdef GLRESAMPLER_3D
#ifdef GLRESAMPLER_3D
vec4 sample = texture(_texture, vec3(ex_TexCoord.xy, _zTexCoord));
#endif
......
Supports Markdown
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment

LRZ Homepage | Datenschutz | Dokumentation und Betriebsbedingungen | Impressum