Merge branch 'master' into aniso

10 files changed, 68 insertions, 3 deletions

diff --git a/cuda/2d/sart.cu b/cuda/2d/sart.cu
index e5cb5bb..c8608a3 100644
--- a/cuda/2d/sart.cu
+++ b/cuda/2d/sart.cu
@@ -71,6 +71,8 @@ SART::SART() : ReconAlgo()
 	projectionCount = 0;
 	iteration = 0;
 	customOrder = false;
+
+	fRelaxation = 1.0f;
 }
 
 
@@ -98,6 +100,7 @@ void SART::reset()
 	projectionCount = 0;
 	iteration = 0;
 	customOrder = false;
+	fRelaxation = 1.0f;
 
 	ReconAlgo::reset();
 }
@@ -200,10 +203,10 @@ bool SART::iterate(unsigned int iterations)
 			// BP, mask, and add back
 			// TODO: Try putting the masking directly in the BP
 			zeroVolumeData(D_tmpData, tmpPitch, dims);
-			callBP_SART(D_tmpData, tmpPitch, D_projData, projPitch, angle, 1.0f);
+			callBP_SART(D_tmpData, tmpPitch, D_projData, projPitch, angle, fRelaxation);
 			processVol<opAddMul>(D_volumeData, D_maskData, D_tmpData, volumePitch, dims);
 		} else {
-			callBP_SART(D_volumeData, volumePitch, D_projData, projPitch, angle, 1.0f);
+			callBP_SART(D_volumeData, volumePitch, D_projData, projPitch, angle, fRelaxation);
 		}
 
 		if (useMinConstraint)
diff --git a/cuda/2d/sart.h b/cuda/2d/sart.h
index 7dcd641..eff9ecf 100644
--- a/cuda/2d/sart.h
+++ b/cuda/2d/sart.h
@@ -50,6 +50,8 @@ public:
 
 	virtual float computeDiffNorm();
 
+	void setRelaxation(float r) { fRelaxation = r; }
+
 protected:
 	void reset();
 	bool precomputeWeights();
@@ -78,6 +80,8 @@ protected:
 	// Geometry-specific precomputed data
 	float* D_lineWeight;
 	unsigned int linePitch;
+
+	float fRelaxation;
 };
 
 }
diff --git a/cuda/2d/sirt.cu b/cuda/2d/sirt.cu
index 162ee77..4baaccb 100644
--- a/cuda/2d/sirt.cu
+++ b/cuda/2d/sirt.cu
@@ -50,6 +50,8 @@ SIRT::SIRT() : ReconAlgo()
 	D_minMaskData = 0;
 	D_maxMaskData = 0;
 
+	fRelaxation = 1.0f;
+
 	freeMinMaxMasks = false;
 }
 
@@ -83,6 +85,8 @@ void SIRT::reset()
 	useVolumeMask = false;
 	useSinogramMask = false;
 
+	fRelaxation = 1.0f;
+
 	ReconAlgo::reset();
 }
 
@@ -139,6 +143,9 @@ bool SIRT::precomputeWeights()
 		processVol<opMul>(D_pixelWeight, D_maskData, pixelPitch, dims);
 	}
 
+	// Also fold the relaxation factor into pixel weights
+	processVol<opMul>(D_pixelWeight, fRelaxation, pixelPitch, dims);
+
 	return true;
 }
 
@@ -253,6 +260,7 @@ bool SIRT::iterate(unsigned int iterations)
 
 		callBP(D_tmpData, tmpPitch, D_projData, projPitch, 1.0f);
 
+		// pixel weights also contain the volume mask and relaxation factor
 		processVol<opAddMul>(D_volumeData, D_pixelWeight, D_tmpData, volumePitch, dims);
 
 		if (useMinConstraint)
diff --git a/cuda/2d/sirt.h b/cuda/2d/sirt.h
index 21094a1..bc913f4 100644
--- a/cuda/2d/sirt.h
+++ b/cuda/2d/sirt.h
@@ -53,6 +53,8 @@ public:
 	bool uploadMinMaxMasks(const float* minMaskData, const float* maxMaskData,
 	                       unsigned int pitch);
 
+	void setRelaxation(float r) { fRelaxation = r; }
+
 	virtual bool iterate(unsigned int iterations);
 
 	virtual float computeDiffNorm();
@@ -81,6 +83,8 @@ protected:
 	unsigned int minMaskPitch;
 	float* D_maxMaskData;
 	unsigned int maxMaskPitch;
+
+	float fRelaxation;
 };
 
 bool doSIRT(float* D_volumeData, unsigned int volumePitch,
diff --git a/cuda/3d/astra3d.cu b/cuda/3d/astra3d.cu
index 42cece2..dd6d551 100644
--- a/cuda/3d/astra3d.cu
+++ b/cuda/3d/astra3d.cu
@@ -263,6 +263,7 @@ public:
 	float* angles;
 	float fOriginSourceDistance;
 	float fOriginDetectorDistance;
+	float fRelaxation;
 
 	SConeProjection* projs;
 	SPar3DProjection* parprojs;
@@ -302,6 +303,8 @@ AstraSIRT3d::AstraSIRT3d()
 	pData->projs = 0;
 	pData->parprojs = 0;
 
+	pData->fRelaxation = 1.0f;
+
 	pData->initialized = false;
 	pData->setStartReconstruction = false;
 
@@ -380,6 +383,14 @@ bool AstraSIRT3d::enableSuperSampling(unsigned int iVoxelSuperSampling,
 	return true;
 }
 
+void AstraSIRT3d::setRelaxation(float r)
+{
+	if (pData->initialized)
+		return;
+
+	pData->fRelaxation = r;
+}
+
 bool AstraSIRT3d::enableVolumeMask()
 {
 	if (pData->initialized)
@@ -439,6 +450,8 @@ bool AstraSIRT3d::init()
 	if (!ok)
 		return false;
 
+	pData->sirt.setRelaxation(pData->fRelaxation);
+
 	pData->D_volumeData = allocateVolumeData(pData->dims);
 	ok = pData->D_volumeData.ptr;
 	if (!ok)
diff --git a/cuda/3d/astra3d.h b/cuda/3d/astra3d.h
index 1f1cee2..3345ab8 100644
--- a/cuda/3d/astra3d.h
+++ b/cuda/3d/astra3d.h
@@ -67,6 +67,8 @@ public:
 	bool enableSuperSampling(unsigned int iVoxelSuperSampling,
 	                         unsigned int iDetectorSuperSampling);
 
+	void setRelaxation(float r);
+
 	// Enable volume/sinogram masks
 	//
 	// This may optionally be called before init().
diff --git a/cuda/3d/mem3d.cu b/cuda/3d/mem3d.cu
index d09c203..9a239da 100644
--- a/cuda/3d/mem3d.cu
+++ b/cuda/3d/mem3d.cu
@@ -62,6 +62,25 @@ size_t availableGPUMemory()
 	return free;
 }
 
+int maxBlockDimension()
+{
+	int dev;
+	cudaError_t err = cudaGetDevice(&dev);
+	if (err != cudaSuccess) {
+		ASTRA_WARN("Error querying device");
+		return 0;
+	}
+
+	cudaDeviceProp props;
+	err = cudaGetDeviceProperties(&props, dev);
+	if (err != cudaSuccess) {
+		ASTRA_WARN("Error querying device %d properties", dev);
+		return 0;
+	}
+
+	return std::min(props.maxTexture3D[0], std::min(props.maxTexture3D[1], props.maxTexture3D[2]));
+}
+
 MemHandle3D allocateGPUMemory(unsigned int x, unsigned int y, unsigned int z, Mem3DZeroMode zero)
 {
 	SMemHandle3D_internal hnd;
diff --git a/cuda/3d/mem3d.h b/cuda/3d/mem3d.h
index acb72cb..6fff80b 100644
--- a/cuda/3d/mem3d.h
+++ b/cuda/3d/mem3d.h
@@ -78,6 +78,7 @@ enum Mem3DZeroMode {
 };
 
 size_t availableGPUMemory();
+int maxBlockDimension();
 
 MemHandle3D allocateGPUMemory(unsigned int x, unsigned int y, unsigned int z, Mem3DZeroMode zero);
 
diff --git a/cuda/3d/sirt3d.cu b/cuda/3d/sirt3d.cu
index 4cb35c3..ba6a159 100644
--- a/cuda/3d/sirt3d.cu
+++ b/cuda/3d/sirt3d.cu
@@ -59,6 +59,8 @@ SIRT::SIRT() : ReconAlgo3D()
 
 	useMinConstraint = false;
 	useMaxConstraint = false;
+
+	fRelaxation = 1.0f;
 }
 
 
@@ -89,6 +91,8 @@ void SIRT::reset()
 	useVolumeMask = false;
 	useSinogramMask = false;
 
+	fRelaxation = 1.0f;
+
 	ReconAlgo3D::reset();
 }
 
@@ -196,6 +200,8 @@ bool SIRT::precomputeWeights()
 		// scale pixel weights with mask to zero out masked pixels
 		processVol3D<opMul>(D_pixelWeight, D_maskData, dims);
 	}
+	processVol3D<opMul>(D_pixelWeight, fRelaxation, dims);
+
 
 	return true;
 }
@@ -307,7 +313,7 @@ bool SIRT::iterate(unsigned int iterations)
 	}
 #endif
 
-
+		// pixel weights also contain the volume mask and relaxation factor
 		processVol3D<opAddMul>(D_volumeData, D_tmpData, D_pixelWeight, dims);
 
 		if (useMinConstraint)
diff --git a/cuda/3d/sirt3d.h b/cuda/3d/sirt3d.h
index bb3864a..5e93deb 100644
--- a/cuda/3d/sirt3d.h
+++ b/cuda/3d/sirt3d.h
@@ -48,6 +48,9 @@ public:
 	// 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.
@@ -91,6 +94,8 @@ protected:
 	float fMinConstraint;
 	float fMaxConstraint;
 
+	float fRelaxation;
+
 	cudaPitchedPtr D_maskData;
 	cudaPitchedPtr D_smaskData;