Move CUDA include files

14 files changed, 1225 insertions, 0 deletions

diff --git a/include/astra/cuda/3d/algo3d.h b/include/astra/cuda/3d/algo3d.h
new file mode 100644
index 0000000..f5fd207
--- /dev/null
+++ b/include/astra/cuda/3d/algo3d.h
@@ -0,0 +1,71 @@
+/*
+-----------------------------------------------------------------------
+Copyright: 2010-2018, imec Vision Lab, University of Antwerp
+           2014-2018, CWI, Amsterdam
+
+Contact: astra@astra-toolbox.com
+Website: http://www.astra-toolbox.com/
+
+This file is part of the ASTRA Toolbox.
+
+
+The ASTRA Toolbox is free software: you can redistribute it and/or modify
+it under the terms of the GNU General Public License as published by
+the Free Software Foundation, either version 3 of the License, or
+(at your option) any later version.
+
+The ASTRA Toolbox is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
+GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License
+along with the ASTRA Toolbox. If not, see <http://www.gnu.org/licenses/>.
+
+-----------------------------------------------------------------------
+*/
+
+#ifndef _CUDA_ALGO_H
+#define _CUDA_ALGO_H
+
+#include "dims3d.h"
+#include "util3d.h"
+
+namespace astraCUDA3d {
+
+class _AstraExport ReconAlgo3D {
+public:
+	ReconAlgo3D();
+	~ReconAlgo3D();
+
+	bool setConeGeometry(const SDimensions3D& dims, const SConeProjection* projs, const SProjectorParams3D& params);
+	bool setPar3DGeometry(const SDimensions3D& dims, const SPar3DProjection* projs, const SProjectorParams3D& params);
+
+	void signalAbort() { shouldAbort = true; }
+
+protected:
+	void reset();
+
+	bool callFP(cudaPitchedPtr& D_volumeData, 
+	            cudaPitchedPtr& D_projData, 
+	            float outputScale);
+	bool callBP(cudaPitchedPtr& D_volumeData, 
+	            cudaPitchedPtr& D_projData,
+	            float outputScale);
+
+	SDimensions3D dims;
+	SProjectorParams3D params;
+	SConeProjection* coneProjs;
+	SPar3DProjection* par3DProjs;
+
+	float fOutputScale;
+
+	volatile bool shouldAbort;
+
+};
+
+
+}
+
+#endif
+
diff --git a/include/astra/cuda/3d/arith3d.h b/include/astra/cuda/3d/arith3d.h
new file mode 100644
index 0000000..fa9f3a1
--- /dev/null
+++ b/include/astra/cuda/3d/arith3d.h
@@ -0,0 +1,79 @@
+/*
+-----------------------------------------------------------------------
+Copyright: 2010-2018, imec Vision Lab, University of Antwerp
+           2014-2018, CWI, Amsterdam
+
+Contact: astra@astra-toolbox.com
+Website: http://www.astra-toolbox.com/
+
+This file is part of the ASTRA Toolbox.
+
+
+The ASTRA Toolbox is free software: you can redistribute it and/or modify
+it under the terms of the GNU General Public License as published by
+the Free Software Foundation, either version 3 of the License, or
+(at your option) any later version.
+
+The ASTRA Toolbox is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
+GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License
+along with the ASTRA Toolbox. If not, see <http://www.gnu.org/licenses/>.
+
+-----------------------------------------------------------------------
+*/
+
+#ifndef _CUDA_ARITH3D_H
+#define _CUDA_ARITH3D_H
+
+#include <cuda.h>
+
+namespace astraCUDA3d {
+
+struct opAddScaled;
+struct opScaleAndAdd;
+struct opAddMulScaled;
+struct opAddMul;
+struct opAdd;
+struct opMul;
+struct opMul2;
+struct opDividedBy;
+struct opInvert;
+struct opSet;
+struct opClampMin;
+struct opClampMax;
+
+enum VolType {
+  SINO = 0,
+  VOL = 1
+};
+
+
+template<typename op, VolType t> void processVol(CUdeviceptr* out, unsigned int pitch, unsigned int width, unsigned int height);
+template<typename op, VolType t> void processVol(CUdeviceptr* out, float fParam, unsigned int pitch, unsigned int width, unsigned int height);
+template<typename op, VolType t> void processVol(CUdeviceptr* out, const CUdeviceptr* in, unsigned int pitch, unsigned int width, unsigned int height);
+template<typename op, VolType t> void processVol(CUdeviceptr* out, const CUdeviceptr* in, float fParam, unsigned int pitch, unsigned int width, unsigned int height);
+template<typename op, VolType t> void processVol(CUdeviceptr* out, const CUdeviceptr* in1, const CUdeviceptr* in2, float fParam, unsigned int pitch, unsigned int width, unsigned int height);
+template<typename op, VolType t> void processVol(CUdeviceptr* out, const CUdeviceptr* in1, const CUdeviceptr* in2, unsigned int pitch, unsigned int width, unsigned int height);
+
+template<typename op> void processVol3D(cudaPitchedPtr& out, const SDimensions3D& dims);
+template<typename op> void processVol3D(cudaPitchedPtr& out, float fParam, const SDimensions3D& dims);
+template<typename op> void processVol3D(cudaPitchedPtr& out, const cudaPitchedPtr& in, const SDimensions3D& dims);
+template<typename op> void processVol3D(cudaPitchedPtr& out, const cudaPitchedPtr& in, float fParam, const SDimensions3D& dims);
+template<typename op> void processVol3D(cudaPitchedPtr& out, const cudaPitchedPtr& in1, const cudaPitchedPtr& in2, float fParam, const SDimensions3D& dims);
+template<typename op> void processVol3D(cudaPitchedPtr& out, const cudaPitchedPtr& in1, const cudaPitchedPtr& in2, const SDimensions3D& dims);
+
+template<typename op> void processSino3D(cudaPitchedPtr& out, const SDimensions3D& dims);
+template<typename op> void processSino3D(cudaPitchedPtr& out, float fParam, const SDimensions3D& dims);
+template<typename op> void processSino3D(cudaPitchedPtr& out, const cudaPitchedPtr& in, const SDimensions3D& dims);
+template<typename op> void processSino3D(cudaPitchedPtr& out, const cudaPitchedPtr& in, float fParam, const SDimensions3D& dims);
+template<typename op> void processSino3D(cudaPitchedPtr& out, const cudaPitchedPtr& in1, const cudaPitchedPtr& in2, float fParam, const SDimensions3D& dims);
+template<typename op> void processSino3D(cudaPitchedPtr& out, const cudaPitchedPtr& in1, const cudaPitchedPtr& in2, const SDimensions3D& dims);
+
+
+
+}
+
+#endif
diff --git a/include/astra/cuda/3d/astra3d.h b/include/astra/cuda/3d/astra3d.h
new file mode 100644
index 0000000..0ea752d
--- /dev/null
+++ b/include/astra/cuda/3d/astra3d.h
@@ -0,0 +1,314 @@
+/*
+-----------------------------------------------------------------------
+Copyright: 2010-2018, imec Vision Lab, University of Antwerp
+           2014-2018, CWI, Amsterdam
+
+Contact: astra@astra-toolbox.com
+Website: http://www.astra-toolbox.com/
+
+This file is part of the ASTRA Toolbox.
+
+
+The ASTRA Toolbox is free software: you can redistribute it and/or modify
+it under the terms of the GNU General Public License as published by
+the Free Software Foundation, either version 3 of the License, or
+(at your option) any later version.
+
+The ASTRA Toolbox is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
+GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License
+along with the ASTRA Toolbox. If not, see <http://www.gnu.org/licenses/>.
+
+-----------------------------------------------------------------------
+*/
+
+#ifndef _CUDA_ASTRA3D_H
+#define _CUDA_ASTRA3D_H
+
+#include "dims3d.h"
+
+namespace astra {
+
+
+// TODO: Switch to a class hierarchy as with the 2D algorithms
+
+
+class CProjectionGeometry3D;
+class CParallelProjectionGeometry3D;
+class CParallelVecProjectionGeometry3D;
+class CConeProjectionGeometry3D;
+class CConeVecProjectionGeometry3D;
+class CVolumeGeometry3D;
+class AstraSIRT3d_internal;
+
+using astraCUDA3d::Cuda3DProjectionKernel;
+using astraCUDA3d::ker3d_default;
+using astraCUDA3d::ker3d_sum_square_weights;
+
+
+class _AstraExport AstraSIRT3d {
+public:
+
+	AstraSIRT3d();
+	~AstraSIRT3d();
+
+	// Set the volume and projection geometry
+	bool setGeometry(const CVolumeGeometry3D* pVolGeom,
+	                 const CProjectionGeometry3D* pProjGeom);
+
+	// Enable supersampling.
+	//
+	// The number of rays used in FP is the square of iDetectorSuperSampling.
+	// The number of rays used in BP is the cube of iVoxelSuperSampling.
+	bool enableSuperSampling(unsigned int iVoxelSuperSampling,
+	                         unsigned int iDetectorSuperSampling);
+
+	void setRelaxation(float r);
+
+	// Enable volume/sinogram masks
+	//
+	// This may optionally be called before init().
+	// If it is called, setVolumeMask()/setSinogramMask() must be called between
+	// setSinogram() and iterate().
+	bool enableVolumeMask();
+	bool enableSinogramMask();
+
+	// Set GPU index
+	//
+	// This should be called before init(). Note that setting the GPU index
+	// in a thread which has already used the GPU may not work.
+	bool setGPUIndex(int index);
+
+	// Allocate GPU buffers and
+	// precompute geometry-specific data.
+	//
+	// This must be called after calling setReconstructionGeometry() and
+	// setProjectionGeometry() or setFanProjectionGeometry().
+	bool init();
+
+	// Setup input sinogram for a slice.
+	// pfSinogram must be a float array of size XXX
+	// NB: iSinogramPitch is measured in floats, not in bytes.
+	//
+	// This must be called after init(), and before iterate(). It may be
+	// called again after iterate()/getReconstruction() to start a new slice.
+	//
+	// pfSinogram will only be read from during this call.
+	bool setSinogram(const float* pfSinogram, unsigned int iSinogramPitch);
+
+	// Setup volume mask for a slice.
+	// pfMask must be a float array of size XXX
+	// NB: iMaskPitch is measured in floats, not in bytes.
+	//
+	// It may only contain the exact values 0.0f and 1.0f. Only volume pixels
+	// for which pfMask[z] is 1.0f are processed.
+	bool setVolumeMask(const float* pfMask, unsigned int iMaskPitch);
+
+	// Setup sinogram mask for a slice.
+	// pfMask must be a float array of size XXX
+	// NB: iMaskPitch is measured in floats, not in bytes.
+	//
+	// It may only contain the exact values 0.0f and 1.0f. Only sinogram pixels
+	// for which pfMask[z] is 1.0f are processed.
+	bool setSinogramMask(const float* pfMask, unsigned int iMaskPitch);
+
+	// Set the starting reconstruction for SIRT.	
+	// pfReconstruction must be a float array of size XXX
+	// NB: iReconstructionPitch is measured in floats, not in bytes.
+	//
+	// This may be called between setSinogram() and iterate().
+	// If this function is not called before iterate(), SIRT will start
+	// from a zero reconstruction.
+	//
+	// pfReconstruction will only be read from during this call.
+	bool setStartReconstruction(const float* pfReconstruction,
+	                            unsigned int iReconstructionPitch);
+
+	// Enable min/max constraint.
+	//
+	// These may optionally be called between init() and iterate()
+	bool setMinConstraint(float fMin);
+	bool setMaxConstraint(float fMax);
+
+	// Perform a number of (additive) SIRT iterations.
+	// This must be called after setSinogram().
+	//
+	// If called multiple times, without calls to setSinogram() or
+	// setStartReconstruction() in between, iterate() will continue from
+	// the result of the previous call.
+	// Calls to getReconstruction() are allowed between calls to iterate() and
+	// do not change the state.
+	bool iterate(unsigned int iIterations);
+
+	// Get the reconstructed slice.
+	// pfReconstruction must be a float array of size XXX
+	// NB: iReconstructionPitch is measured in floats, not in bytes.
+	//
+	// This may be called after iterate().
+	bool getReconstruction(float* pfReconstruction,
+	                       unsigned int iReconstructionPitch) const;
+
+	// Compute the norm of the difference of the FP of the current
+	// reconstruction and the sinogram. (This performs one FP.)
+	// It can be called after iterate().
+	float computeDiffNorm();
+
+	// Signal the algorithm that it should abort after the current iteration.
+	// This is intended to be called from another thread.
+	void signalAbort();
+
+protected:
+	AstraSIRT3d_internal *pData;
+};
+
+
+class AstraCGLS3d_internal;
+
+
+class _AstraExport AstraCGLS3d {
+public:
+
+	AstraCGLS3d();
+	~AstraCGLS3d();
+
+	// Set the volume and projection geometry
+	bool setGeometry(const CVolumeGeometry3D* pVolGeom,
+	                 const CProjectionGeometry3D* pProjGeom);
+
+	// Enable supersampling.
+	//
+	// The number of rays used in FP is the square of iDetectorSuperSampling.
+	// The number of rays used in BP is the cube of iVoxelSuperSampling.
+	bool enableSuperSampling(unsigned int iVoxelSuperSampling,
+	                         unsigned int iDetectorSuperSampling);
+
+	// Enable volume/sinogram masks
+	//
+	// This may optionally be called before init().
+	// If it is called, setVolumeMask()/setSinogramMask() must be called between
+	// setSinogram() and iterate().
+	bool enableVolumeMask();
+	//bool enableSinogramMask();
+
+	// Set GPU index
+	//
+	// This should be called before init(). Note that setting the GPU index
+	// in a thread which has already used the GPU may not work.
+	bool setGPUIndex(int index);
+
+	// Allocate GPU buffers and
+	// precompute geometry-specific data.
+	//
+	// This must be called after calling setReconstructionGeometry() and
+	// setProjectionGeometry() or setFanProjectionGeometry().
+	bool init();
+
+	// Setup input sinogram for a slice.
+	// pfSinogram must be a float array of size XXX
+	// NB: iSinogramPitch is measured in floats, not in bytes.
+	//
+	// This must be called after init(), and before iterate(). It may be
+	// called again after iterate()/getReconstruction() to start a new slice.
+	//
+	// pfSinogram will only be read from during this call.
+	bool setSinogram(const float* pfSinogram, unsigned int iSinogramPitch);
+
+	// Setup volume mask for a slice.
+	// pfMask must be a float array of size XXX
+	// NB: iMaskPitch is measured in floats, not in bytes.
+	//
+	// It may only contain the exact values 0.0f and 1.0f. Only volume pixels
+	// for which pfMask[z] is 1.0f are processed.
+	bool setVolumeMask(const float* pfMask, unsigned int iMaskPitch);
+
+	// Setup sinogram mask for a slice.
+	// pfMask must be a float array of size XXX
+	// NB: iMaskPitch is measured in floats, not in bytes.
+	//
+	// It may only contain the exact values 0.0f and 1.0f. Only sinogram pixels
+	// for which pfMask[z] is 1.0f are processed.
+	//bool setSinogramMask(const float* pfMask, unsigned int iMaskPitch);
+
+	// Set the starting reconstruction for SIRT.	
+	// pfReconstruction must be a float array of size XXX
+	// NB: iReconstructionPitch is measured in floats, not in bytes.
+	//
+	// This may be called between setSinogram() and iterate().
+	// If this function is not called before iterate(), SIRT will start
+	// from a zero reconstruction.
+	//
+	// pfReconstruction will only be read from during this call.
+	bool setStartReconstruction(const float* pfReconstruction,
+	                            unsigned int iReconstructionPitch);
+
+	// Enable min/max constraint.
+	//
+	// These may optionally be called between init() and iterate()
+	//bool setMinConstraint(float fMin);
+	//bool setMaxConstraint(float fMax);
+
+	// Perform a number of (additive) SIRT iterations.
+	// This must be called after setSinogram().
+	//
+	// If called multiple times, without calls to setSinogram() or
+	// setStartReconstruction() in between, iterate() will continue from
+	// the result of the previous call.
+	// Calls to getReconstruction() are allowed between calls to iterate() and
+	// do not change the state.
+	bool iterate(unsigned int iIterations);
+
+	// Get the reconstructed slice.
+	// pfReconstruction must be a float array of size XXX
+	// NB: iReconstructionPitch is measured in floats, not in bytes.
+	//
+	// This may be called after iterate().
+	bool getReconstruction(float* pfReconstruction,
+	                       unsigned int iReconstructionPitch) const;
+
+	// Compute the norm of the difference of the FP of the current
+	// reconstruction and the sinogram. (This performs one FP.)
+	// It can be called after iterate().
+	float computeDiffNorm();
+
+	// Signal the algorithm that it should abort after the current iteration.
+	// This is intended to be called from another thread.
+	void signalAbort();
+
+protected:
+	AstraCGLS3d_internal *pData;
+};
+
+bool convertAstraGeometry_dims(const CVolumeGeometry3D* pVolGeom,
+                               const CProjectionGeometry3D* pProjGeom,
+                               astraCUDA3d::SDimensions3D& dims);
+
+bool convertAstraGeometry(const CVolumeGeometry3D* pVolGeom,
+                          const CProjectionGeometry3D* pProjGeom,
+                          SPar3DProjection*& pParProjs,
+                          SConeProjection*& pConeProjs,
+                          astraCUDA3d::SProjectorParams3D& params);
+
+_AstraExport bool astraCudaFP(const float* pfVolume, float* pfProjections,
+                      const CVolumeGeometry3D* pVolGeom,
+                      const CProjectionGeometry3D* pProjGeom,
+                      int iGPUIndex, int iDetectorSuperSampling,
+                      Cuda3DProjectionKernel projKernel);
+
+
+_AstraExport bool astraCudaBP(float* pfVolume, const float* pfProjections,
+                      const CVolumeGeometry3D* pVolGeom,
+                      const CProjectionGeometry3D* pProjGeom,
+                      int iGPUIndex, int iVoxelSuperSampling);
+
+_AstraExport bool astraCudaBP_SIRTWeighted(float* pfVolume, const float* pfProjections,
+                      const CVolumeGeometry3D* pVolGeom,
+                      const CProjectionGeometry3D* pProjGeom,
+                      int iGPUIndex, int iVoxelSuperSampling);
+
+}
+
+
+#endif
diff --git a/include/astra/cuda/3d/cgls3d.h b/include/astra/cuda/3d/cgls3d.h
new file mode 100644
index 0000000..b9b9e3a
--- /dev/null
+++ b/include/astra/cuda/3d/cgls3d.h
@@ -0,0 +1,112 @@
+/*
+-----------------------------------------------------------------------
+Copyright: 2010-2018, imec Vision Lab, University of Antwerp
+           2014-2018, CWI, Amsterdam
+
+Contact: astra@astra-toolbox.com
+Website: http://www.astra-toolbox.com/
+
+This file is part of the ASTRA Toolbox.
+
+
+The ASTRA Toolbox is free software: you can redistribute it and/or modify
+it under the terms of the GNU General Public License as published by
+the Free Software Foundation, either version 3 of the License, or
+(at your option) any later version.
+
+The ASTRA Toolbox is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
+GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License
+along with the ASTRA Toolbox. If not, see <http://www.gnu.org/licenses/>.
+
+-----------------------------------------------------------------------
+*/
+
+#ifndef _CUDA_CGLS3D_H
+#define _CUDA_CGLS3D_H
+
+#include "algo3d.h"
+
+namespace astraCUDA3d {
+
+class _AstraExport CGLS : public ReconAlgo3D {
+public:
+	CGLS();
+	~CGLS();
+
+//	bool setConeGeometry(const SDimensions3D& dims, const SConeProjection* projs);
+
+
+	bool enableVolumeMask();
+	bool enableSinogramMask();
+
+	// init should be called after setting all geometry
+	bool init();
+
+	// setVolumeMask should be called after init and before iterate,
+	// but only if enableVolumeMask was called before init.
+	// It may be called again after iterate.
+	bool setVolumeMask(cudaPitchedPtr& D_maskData);
+
+	// setSinogramMask should be called after init and before iterate,
+	// but only if enableSinogramMask was called before init.
+	// It may be called again after iterate.
+	bool setSinogramMask(cudaPitchedPtr& D_smaskData);
+
+
+	// setBuffers should be called after init and before iterate.
+	// It may be called again after iterate.
+	bool setBuffers(cudaPitchedPtr& D_volumeData,
+	                cudaPitchedPtr& D_projData);
+
+
+	// set Min/Max constraints. They may be called at any time, and will affect
+	// any iterate() calls afterwards.
+	bool setMinConstraint(float fMin) { return false; }
+	bool setMaxConstraint(float fMax) { return false; }
+
+	// iterate should be called after init and setBuffers.
+	// It may be called multiple times.
+	bool iterate(unsigned int iterations);
+
+	// Compute the norm of the difference of the FP of the current reconstruction
+	// and the sinogram. (This performs one FP.)
+	// It can be called after iterate.
+	float computeDiffNorm();
+
+protected:
+	void reset();
+
+	bool useVolumeMask;
+	bool useSinogramMask;
+
+	cudaPitchedPtr D_maskData;
+	cudaPitchedPtr D_smaskData;
+
+	// Input/output
+	cudaPitchedPtr D_sinoData;
+	cudaPitchedPtr D_volumeData;
+
+ 	// Temporary buffers
+	cudaPitchedPtr D_r;
+	cudaPitchedPtr D_w;
+	cudaPitchedPtr D_z;
+	cudaPitchedPtr D_p;
+
+	float gamma;
+
+	bool sliceInitialized;
+};
+
+_AstraExport bool doCGLS(cudaPitchedPtr D_volumeData, unsigned int volumePitch,
+            cudaPitchedPtr D_projData, unsigned int projPitch,
+            cudaPitchedPtr D_maskData, unsigned int maskPitch,
+            const SDimensions3D& dims, const SConeProjection* projs,
+            unsigned int iterations);
+
+}
+
+#endif
diff --git a/include/astra/cuda/3d/cone_bp.h b/include/astra/cuda/3d/cone_bp.h
new file mode 100644
index 0000000..6dcadb1
--- /dev/null
+++ b/include/astra/cuda/3d/cone_bp.h
@@ -0,0 +1,45 @@
+/*
+-----------------------------------------------------------------------
+Copyright: 2010-2018, imec Vision Lab, University of Antwerp
+           2014-2018, CWI, Amsterdam
+
+Contact: astra@astra-toolbox.com
+Website: http://www.astra-toolbox.com/
+
+This file is part of the ASTRA Toolbox.
+
+
+The ASTRA Toolbox is free software: you can redistribute it and/or modify
+it under the terms of the GNU General Public License as published by
+the Free Software Foundation, either version 3 of the License, or
+(at your option) any later version.
+
+The ASTRA Toolbox is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
+GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License
+along with the ASTRA Toolbox. If not, see <http://www.gnu.org/licenses/>.
+
+-----------------------------------------------------------------------
+*/
+
+#ifndef _CUDA_CONE_BP_H
+#define _CUDA_CONE_BP_H
+
+namespace astraCUDA3d {
+
+_AstraExport bool ConeBP_Array(cudaPitchedPtr D_volumeData,
+                  cudaArray *D_projArray,
+                  const SDimensions3D& dims, const SConeProjection* angles,
+                  const SProjectorParams3D& params);
+
+_AstraExport bool ConeBP(cudaPitchedPtr D_volumeData,
+            cudaPitchedPtr D_projData,
+            const SDimensions3D& dims, const SConeProjection* angles,
+            const SProjectorParams3D& params);
+         
+}
+
+#endif
diff --git a/include/astra/cuda/3d/cone_fp.h b/include/astra/cuda/3d/cone_fp.h
new file mode 100644
index 0000000..4da72d4
--- /dev/null
+++ b/include/astra/cuda/3d/cone_fp.h
@@ -0,0 +1,45 @@
+/*
+-----------------------------------------------------------------------
+Copyright: 2010-2018, imec Vision Lab, University of Antwerp
+           2014-2018, CWI, Amsterdam
+
+Contact: astra@astra-toolbox.com
+Website: http://www.astra-toolbox.com/
+
+This file is part of the ASTRA Toolbox.
+
+
+The ASTRA Toolbox is free software: you can redistribute it and/or modify
+it under the terms of the GNU General Public License as published by
+the Free Software Foundation, either version 3 of the License, or
+(at your option) any later version.
+
+The ASTRA Toolbox is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
+GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License
+along with the ASTRA Toolbox. If not, see <http://www.gnu.org/licenses/>.
+
+-----------------------------------------------------------------------
+*/
+
+#ifndef _CUDA_CONE_FP_H
+#define _CUDA_CONE_FP_H
+
+namespace astraCUDA3d {
+
+_AstraExport bool ConeFP_Array(cudaArray *D_volArray,
+                  cudaPitchedPtr D_projData,
+                  const SDimensions3D& dims, const SConeProjection* angles,
+                  const SProjectorParams3D& params);
+
+_AstraExport bool ConeFP(cudaPitchedPtr D_volumeData,
+            cudaPitchedPtr D_projData,
+            const SDimensions3D& dims, const SConeProjection* angles,
+            const SProjectorParams3D& params);
+
+}
+
+#endif
diff --git a/include/astra/cuda/3d/darthelper3d.h b/include/astra/cuda/3d/darthelper3d.h
new file mode 100644
index 0000000..385e48e
--- /dev/null
+++ b/include/astra/cuda/3d/darthelper3d.h
@@ -0,0 +1,42 @@
+/*
+-----------------------------------------------------------------------
+Copyright: 2010-2018, imec Vision Lab, University of Antwerp
+           2014-2018, CWI, Amsterdam
+
+Contact: astra@astra-toolbox.com
+Website: http://www.astra-toolbox.com/
+
+This file is part of the ASTRA Toolbox.
+
+
+The ASTRA Toolbox is free software: you can redistribute it and/or modify
+it under the terms of the GNU General Public License as published by
+the Free Software Foundation, either version 3 of the License, or
+(at your option) any later version.
+
+The ASTRA Toolbox is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
+GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License
+along with the ASTRA Toolbox. If not, see <http://www.gnu.org/licenses/>.
+
+-----------------------------------------------------------------------
+*/
+
+#ifndef _CUDA_DARTHELPER3_H
+#define _CUDA_DARTHELPER3_H
+
+#include "dims3d.h"
+
+namespace astraCUDA3d {
+
+	void dartSmoothing(float* out, const float* in, float b, unsigned int radius, SDimensions3D dims);
+	void dartMasking(float* out, const float* in, unsigned int conn, unsigned int radius, unsigned int threshold, SDimensions3D dims);
+
+	bool setGPUIndex(int index);
+
+}
+
+#endif
diff --git a/include/astra/cuda/3d/dims3d.h b/include/astra/cuda/3d/dims3d.h
new file mode 100644
index 0000000..ef4ea78
--- /dev/null
+++ b/include/astra/cuda/3d/dims3d.h
@@ -0,0 +1,77 @@
+/*
+-----------------------------------------------------------------------
+Copyright: 2010-2018, imec Vision Lab, University of Antwerp
+           2014-2018, CWI, Amsterdam
+
+Contact: astra@astra-toolbox.com
+Website: http://www.astra-toolbox.com/
+
+This file is part of the ASTRA Toolbox.
+
+
+The ASTRA Toolbox is free software: you can redistribute it and/or modify
+it under the terms of the GNU General Public License as published by
+the Free Software Foundation, either version 3 of the License, or
+(at your option) any later version.
+
+The ASTRA Toolbox is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
+GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License
+along with the ASTRA Toolbox. If not, see <http://www.gnu.org/licenses/>.
+
+-----------------------------------------------------------------------
+*/
+
+#ifndef _CUDA_CONE_DIMS_H
+#define _CUDA_CONE_DIMS_H
+
+#include "astra/GeometryUtil3D.h"
+
+
+namespace astraCUDA3d {
+
+using astra::SConeProjection;
+using astra::SPar3DProjection;
+
+
+enum Cuda3DProjectionKernel {
+	ker3d_default = 0,
+	ker3d_sum_square_weights
+};
+
+
+struct SDimensions3D {
+	unsigned int iVolX;
+	unsigned int iVolY;
+	unsigned int iVolZ;
+	unsigned int iProjAngles;
+	unsigned int iProjU; // number of detectors in the U direction
+	unsigned int iProjV; // number of detectors in the V direction
+};
+
+struct SProjectorParams3D {
+	SProjectorParams3D() :
+	    iRaysPerDetDim(1), iRaysPerVoxelDim(1),
+	    fOutputScale(1.0f),
+	    fVolScaleX(1.0f), fVolScaleY(1.0f), fVolScaleZ(1.0f),
+	    ker(ker3d_default),
+	    bFDKWeighting(false)
+	{ }
+
+	unsigned int iRaysPerDetDim;
+	unsigned int iRaysPerVoxelDim;
+	float fOutputScale;
+	float fVolScaleX;
+	float fVolScaleY;
+	float fVolScaleZ;
+	Cuda3DProjectionKernel ker;
+	bool bFDKWeighting;
+};
+
+}
+
+#endif
+
diff --git a/include/astra/cuda/3d/fdk.h b/include/astra/cuda/3d/fdk.h
new file mode 100644
index 0000000..6f6e73b
--- /dev/null
+++ b/include/astra/cuda/3d/fdk.h
@@ -0,0 +1,50 @@
+/*
+-----------------------------------------------------------------------
+Copyright: 2010-2018, imec Vision Lab, University of Antwerp
+           2014-2018, CWI, Amsterdam
+
+Contact: astra@astra-toolbox.com
+Website: http://www.astra-toolbox.com/
+
+This file is part of the ASTRA Toolbox.
+
+
+The ASTRA Toolbox is free software: you can redistribute it and/or modify
+it under the terms of the GNU General Public License as published by
+the Free Software Foundation, either version 3 of the License, or
+(at your option) any later version.
+
+The ASTRA Toolbox is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
+GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License
+along with the ASTRA Toolbox. If not, see <http://www.gnu.org/licenses/>.
+
+-----------------------------------------------------------------------
+*/
+
+#ifndef _CUDA_FDK_H
+#define _CUDA_FDK_H
+
+#include "dims3d.h"
+
+namespace astraCUDA3d {
+
+bool FDK_PreWeight(cudaPitchedPtr D_projData,
+                float fSrcOrigin, float fDetOrigin,
+                float fZShift,
+                float fDetUSize, float fDetVSize, float fVoxSize,
+                bool bShortScan,
+                const SDimensions3D& dims, const float* angles);
+
+bool FDK(cudaPitchedPtr D_volumeData,
+         cudaPitchedPtr D_projData,
+         const SConeProjection* angles,
+         const SDimensions3D& dims, SProjectorParams3D params, bool bShortScan,
+         const float* filter);
+
+}
+
+#endif
diff --git a/include/astra/cuda/3d/mem3d.h b/include/astra/cuda/3d/mem3d.h
new file mode 100644
index 0000000..78e1294
--- /dev/null
+++ b/include/astra/cuda/3d/mem3d.h
@@ -0,0 +1,105 @@
+/*
+-----------------------------------------------------------------------
+Copyright: 2010-2018, imec Vision Lab, University of Antwerp
+           2014-2018, CWI, Amsterdam
+
+Contact: astra@astra-toolbox.com
+Website: http://www.astra-toolbox.com/
+
+This file is part of the ASTRA Toolbox.
+
+
+The ASTRA Toolbox is free software: you can redistribute it and/or modify
+it under the terms of the GNU General Public License as published by
+the Free Software Foundation, either version 3 of the License, or
+(at your option) any later version.
+
+The ASTRA Toolbox is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
+GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License
+along with the ASTRA Toolbox. If not, see <http://www.gnu.org/licenses/>.
+
+-----------------------------------------------------------------------
+*/
+
+#ifndef _CUDA_MEM3D_H
+#define _CUDA_MEM3D_H
+
+#include <boost/shared_ptr.hpp>
+
+#include "astra3d.h"
+
+namespace astra {
+class CVolumeGeometry3D;
+class CProjectionGeometry3D;	
+}
+
+namespace astraCUDA3d {
+
+// TODO: Make it possible to delete these handles when they're no longer
+// necessary inside the FP/BP
+//
+// TODO: Add functions for querying capacity
+
+struct SMemHandle3D_internal;
+
+struct MemHandle3D {
+	boost::shared_ptr<SMemHandle3D_internal> d;
+	operator bool() const { return (bool)d; }
+};
+
+struct SSubDimensions3D {
+	unsigned int nx;
+	unsigned int ny;
+	unsigned int nz;
+	unsigned int pitch;
+	unsigned int subnx;
+	unsigned int subny;
+	unsigned int subnz;
+	unsigned int subx;
+	unsigned int suby;
+	unsigned int subz;
+};
+
+/*
+// Useful or not?
+enum Mem3DCopyMode {
+	MODE_SET,
+	MODE_ADD
+};
+*/
+
+enum Mem3DZeroMode {
+	INIT_NO,
+	INIT_ZERO
+};
+
+int maxBlockDimension();
+
+_AstraExport MemHandle3D wrapHandle(float *D_ptr, unsigned int x, unsigned int y, unsigned int z, unsigned int pitch);
+
+MemHandle3D allocateGPUMemory(unsigned int x, unsigned int y, unsigned int z, Mem3DZeroMode zero);
+
+bool copyToGPUMemory(const float *src, MemHandle3D dst, const SSubDimensions3D &pos);
+
+bool copyFromGPUMemory(float *dst, MemHandle3D src, const SSubDimensions3D &pos);
+
+bool freeGPUMemory(MemHandle3D handle);
+
+bool zeroGPUMemory(MemHandle3D handle, unsigned int x, unsigned int y, unsigned int z);
+
+bool setGPUIndex(int index);
+
+
+bool FP(const astra::CProjectionGeometry3D* pProjGeom, MemHandle3D projData, const astra::CVolumeGeometry3D* pVolGeom, MemHandle3D volData, int iDetectorSuperSampling, astra::Cuda3DProjectionKernel projKernel);
+
+bool BP(const astra::CProjectionGeometry3D* pProjGeom, MemHandle3D projData, const astra::CVolumeGeometry3D* pVolGeom, MemHandle3D volData, int iVoxelSuperSampling, bool bFDKWeighting);
+
+bool FDK(const astra::CProjectionGeometry3D* pProjGeom, MemHandle3D projData, const astra::CVolumeGeometry3D* pVolGeom, MemHandle3D volData, bool bShortScan, const float *pfFilter = 0);
+
+}
+
+#endif
diff --git a/include/astra/cuda/3d/par3d_bp.h b/include/astra/cuda/3d/par3d_bp.h
new file mode 100644
index 0000000..b4ec6a3
--- /dev/null
+++ b/include/astra/cuda/3d/par3d_bp.h
@@ -0,0 +1,46 @@
+/*
+-----------------------------------------------------------------------
+Copyright: 2010-2018, imec Vision Lab, University of Antwerp
+           2014-2018, CWI, Amsterdam
+
+Contact: astra@astra-toolbox.com
+Website: http://www.astra-toolbox.com/
+
+This file is part of the ASTRA Toolbox.
+
+
+The ASTRA Toolbox is free software: you can redistribute it and/or modify
+it under the terms of the GNU General Public License as published by
+the Free Software Foundation, either version 3 of the License, or
+(at your option) any later version.
+
+The ASTRA Toolbox is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
+GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License
+along with the ASTRA Toolbox. If not, see <http://www.gnu.org/licenses/>.
+
+-----------------------------------------------------------------------
+*/
+
+#ifndef _CUDA_PAR3D_BP_H
+#define _CUDA_PAR3D_BP_H
+
+namespace astraCUDA3d {
+
+_AstraExport bool Par3DBP_Array(cudaPitchedPtr D_volumeData,
+                   cudaArray *D_projArray,
+                   const SDimensions3D& dims, const SPar3DProjection* angles,
+                   const SProjectorParams3D& params);
+
+_AstraExport bool Par3DBP(cudaPitchedPtr D_volumeData,
+             cudaPitchedPtr D_projData,
+             const SDimensions3D& dims, const SPar3DProjection* angles,
+             const SProjectorParams3D& params);
+         
+
+}
+
+#endif
diff --git a/include/astra/cuda/3d/par3d_fp.h b/include/astra/cuda/3d/par3d_fp.h
new file mode 100644
index 0000000..18f5f2e
--- /dev/null
+++ b/include/astra/cuda/3d/par3d_fp.h
@@ -0,0 +1,50 @@
+/*
+-----------------------------------------------------------------------
+Copyright: 2010-2018, imec Vision Lab, University of Antwerp
+           2014-2018, CWI, Amsterdam
+
+Contact: astra@astra-toolbox.com
+Website: http://www.astra-toolbox.com/
+
+This file is part of the ASTRA Toolbox.
+
+
+The ASTRA Toolbox is free software: you can redistribute it and/or modify
+it under the terms of the GNU General Public License as published by
+the Free Software Foundation, either version 3 of the License, or
+(at your option) any later version.
+
+The ASTRA Toolbox is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
+GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License
+along with the ASTRA Toolbox. If not, see <http://www.gnu.org/licenses/>.
+
+-----------------------------------------------------------------------
+*/
+
+#ifndef _CUDA_PAR3D_FP_H
+#define _CUDA_PAR3D_FP_H
+
+namespace astraCUDA3d {
+
+_AstraExport bool Par3DFP_Array(cudaArray *D_volArray,
+                   cudaPitchedPtr D_projData,
+                   const SDimensions3D& dims, const SPar3DProjection* angles,
+                   const SProjectorParams3D& params);
+
+_AstraExport bool Par3DFP(cudaPitchedPtr D_volumeData,
+            cudaPitchedPtr D_projData,
+            const SDimensions3D& dims, const SPar3DProjection* angles,
+            const SProjectorParams3D& params);
+
+_AstraExport bool Par3DFP_SumSqW(cudaPitchedPtr D_volumeData,
+                    cudaPitchedPtr D_projData,
+                    const SDimensions3D& dims, const SPar3DProjection* angles,
+                    const SProjectorParams3D& params);
+
+}
+
+#endif
diff --git a/include/astra/cuda/3d/sirt3d.h b/include/astra/cuda/3d/sirt3d.h
new file mode 100644
index 0000000..7c9a7c0
--- /dev/null
+++ b/include/astra/cuda/3d/sirt3d.h
@@ -0,0 +1,121 @@
+/*
+-----------------------------------------------------------------------
+Copyright: 2010-2018, imec Vision Lab, University of Antwerp
+           2014-2018, CWI, Amsterdam
+
+Contact: astra@astra-toolbox.com
+Website: http://www.astra-toolbox.com/
+
+This file is part of the ASTRA Toolbox.
+
+
+The ASTRA Toolbox is free software: you can redistribute it and/or modify
+it under the terms of the GNU General Public License as published by
+the Free Software Foundation, either version 3 of the License, or
+(at your option) any later version.
+
+The ASTRA Toolbox is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
+GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License
+along with the ASTRA Toolbox. If not, see <http://www.gnu.org/licenses/>.
+
+-----------------------------------------------------------------------
+*/
+
+#ifndef _CUDA_SIRT3D_H
+#define _CUDA_SIRT3D_H
+
+#include "algo3d.h"
+
+namespace astraCUDA3d {
+
+class _AstraExport SIRT : public ReconAlgo3D {
+public:
+	SIRT();
+	~SIRT();
+
+//	bool setConeGeometry(const SDimensions3D& dims, const SConeProjection* projs);
+
+
+	bool enableVolumeMask();
+	bool enableSinogramMask();
+
+	// init should be called after setting all geometry
+	bool init();
+
+	// Set relaxation factor. This may be called after init and before iterate.
+	void setRelaxation(float r) { fRelaxation = r; }
+
+	// setVolumeMask should be called after init and before iterate,
+	// but only if enableVolumeMask was called before init.
+	// It may be called again after iterate.
+	bool setVolumeMask(cudaPitchedPtr& D_maskData);
+
+	// setSinogramMask should be called after init and before iterate,
+	// but only if enableSinogramMask was called before init.
+	// It may be called again after iterate.
+	bool setSinogramMask(cudaPitchedPtr& D_smaskData);
+
+
+	// setBuffers should be called after init and before iterate.
+	// It may be called again after iterate.
+	bool setBuffers(cudaPitchedPtr& D_volumeData,
+	                cudaPitchedPtr& D_projData);
+
+
+	// set Min/Max constraints. They may be called at any time, and will affect
+	// any iterate() calls afterwards.
+	bool setMinConstraint(float fMin);
+	bool setMaxConstraint(float fMax);
+
+	// iterate should be called after init and setBuffers.
+	// It may be called multiple times.
+	bool iterate(unsigned int iterations);
+
+	// Compute the norm of the difference of the FP of the current reconstruction
+	// and the sinogram. (This performs one FP.)
+	// It can be called after iterate.
+	float computeDiffNorm();
+
+protected:
+	void reset();
+	bool precomputeWeights();
+
+	bool useVolumeMask;
+	bool useSinogramMask;
+
+	bool useMinConstraint;
+	bool useMaxConstraint;
+	float fMinConstraint;
+	float fMaxConstraint;
+
+	float fRelaxation;
+
+	cudaPitchedPtr D_maskData;
+	cudaPitchedPtr D_smaskData;
+
+	// Input/output
+	cudaPitchedPtr D_sinoData;
+	cudaPitchedPtr D_volumeData;
+
+ 	// Temporary buffers
+	cudaPitchedPtr D_projData;
+	cudaPitchedPtr D_tmpData;
+
+	// Geometry-specific precomputed data
+	cudaPitchedPtr D_lineWeight;
+	cudaPitchedPtr D_pixelWeight;
+};
+
+bool doSIRT(cudaPitchedPtr D_volumeData, unsigned int volumePitch,
+            cudaPitchedPtr D_projData, unsigned int projPitch,
+            cudaPitchedPtr D_maskData, unsigned int maskPitch,
+            const SDimensions3D& dims, const SConeProjection* projs,
+            unsigned int iterations);
+
+}
+
+#endif
diff --git a/include/astra/cuda/3d/util3d.h b/include/astra/cuda/3d/util3d.h
new file mode 100644
index 0000000..17eb31e
--- /dev/null
+++ b/include/astra/cuda/3d/util3d.h
@@ -0,0 +1,68 @@
+/*
+-----------------------------------------------------------------------
+Copyright: 2010-2018, imec Vision Lab, University of Antwerp
+           2014-2018, CWI, Amsterdam
+
+Contact: astra@astra-toolbox.com
+Website: http://www.astra-toolbox.com/
+
+This file is part of the ASTRA Toolbox.
+
+
+The ASTRA Toolbox is free software: you can redistribute it and/or modify
+it under the terms of the GNU General Public License as published by
+the Free Software Foundation, either version 3 of the License, or
+(at your option) any later version.
+
+The ASTRA Toolbox is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
+GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License
+along with the ASTRA Toolbox. If not, see <http://www.gnu.org/licenses/>.
+
+-----------------------------------------------------------------------
+*/
+
+#ifndef _CUDA_UTIL3D_H
+#define _CUDA_UTIL3D_H
+
+#include <cuda.h>
+#include "dims3d.h"
+
+#ifndef M_PI
+#define M_PI 3.14159265358979323846
+#endif
+#include "../2d/util.h"
+
+namespace astraCUDA3d {
+
+cudaPitchedPtr allocateVolumeData(const SDimensions3D& dims);
+cudaPitchedPtr allocateProjectionData(const SDimensions3D& dims);
+bool zeroVolumeData(cudaPitchedPtr& D_data, const SDimensions3D& dims);
+bool zeroProjectionData(cudaPitchedPtr& D_data, const SDimensions3D& dims);
+bool copyVolumeToDevice(const float* data, cudaPitchedPtr& D_data, const SDimensions3D& dims, unsigned int pitch = 0);
+bool copyProjectionsToDevice(const float* data, cudaPitchedPtr& D_data, const SDimensions3D& dims, unsigned int pitch = 0);
+bool copyVolumeFromDevice(float* data, const cudaPitchedPtr& D_data, const SDimensions3D& dims, unsigned int pitch = 0);
+bool copyProjectionsFromDevice(float* data, const cudaPitchedPtr& D_data, const SDimensions3D& dims, unsigned int pitch = 0);
+bool duplicateVolumeData(cudaPitchedPtr& D_dest, const cudaPitchedPtr& D_src, const SDimensions3D& dims); 
+bool duplicateProjectionData(cudaPitchedPtr& D_dest, const cudaPitchedPtr& D_src, const SDimensions3D& dims); 
+
+
+bool transferProjectionsToArray(cudaPitchedPtr D_projData, cudaArray* array, const SDimensions3D& dims);
+bool transferVolumeToArray(cudaPitchedPtr D_volumeData, cudaArray* array, const SDimensions3D& dims);
+bool zeroProjectionArray(cudaArray* array, const SDimensions3D& dims);
+bool zeroVolumeArray(cudaArray* array, const SDimensions3D& dims);
+cudaArray* allocateProjectionArray(const SDimensions3D& dims);
+cudaArray* allocateVolumeArray(const SDimensions3D& dims);
+
+bool cudaTextForceKernelsCompletion();
+
+float dotProduct3D(cudaPitchedPtr data, unsigned int x, unsigned int y, unsigned int z);
+
+int calcNextPowerOfTwo(int _iValue);
+
+}
+
+#endif