Added a detailed spider class timer that is ran only in TIMING mode (has to be e…

…nabled in CMakeLists)

Added a detailed spider class timer that is ran only in TIMING mode (has to be e…
…nabled in CMakeLists)
Pavel Govyadinov
1 parent 213d109c
Showing 4 changed files with 155 additions and 101 deletions Show diff stats
stim/cuda/cuda_texture.cuh
stim/cuda/spider_cost.cuh
stim/gl/gl_spider.h
stim/image/image.h
@@ -99,8 +99,8 @@ namespace stim
 					&resource,
 					tex,
 					target,
-//					cudaGraphicsMapFlagsReadOnly
-					cudaGraphicsRegisterFlagsNone	
+					cudaGraphicsMapFlagsReadOnly
+//					cudaGraphicsRegisterFlagsNone	
 				)
 			);
@@ -40,11 +40,10 @@ namespace stim{
 	///Returns the value of the template pixel.
 	///@param int x			--location of a pixel.
 	__device__
-	float Template(int x)
+	float Template(int x, int max_x)
 	{
-		if(x < 32/6 || x > 32*5/6 || (x > 32*2/6 && x < 32*4/6)){
-//		if(x < 2 || x > 13 || (x > 5 && x < 10)){
-//	if(x < ceilf(16/6) || x > floorf(16*5/6) || (x > floorf(16*2/6) && x < ceilf(16*4/6))){
+	if(x < max_x/6 || x > max_x*5/6 || (x > max_x*2/6 && x < max_x*4/6))
+		{
 			return 1.0;
 		}else{
 			return 0.0;
@@ -71,7 +70,7 @@ namespace stim{
 		int g_idx = blockIdx.y;
 		float valIn             = tex2D<unsigned char>(texIn, x, y)/255.0;
-		float valTemp           = Template(x);
+		float valTemp           = Template(x, dx);
 //		print[idx]              = abs(valIn);             ///temporary
@@ -149,7 +148,7 @@ namespace stim{
 		for( int i = 0; i<DIM_Y; i++){
 			if(output[i] < mini){
 				mini = output[i];
-				idx = i;                                                                   
+				idx = i;                                          
 			}
 		}
@@ -31,6 +31,11 @@
 #include <iostream>
 #include <fstream>
+#ifdef TIMING
+	#include <ctime>
+	#include <cstdio>	
+#endif
+
 #ifdef TESTING
 	#include <iostream>
 	#include <cstdio>
@@ -48,6 +53,16 @@ class gl_spider : public virtual gl_texture&lt;T&gt;
 	//doesn't even need the texture iD really.
 	private:
+		#ifdef TIMING
+			double branch_time = 0;
+			double direction_time = 0;
+			double position_time = 0;
+			double size_time = 0;
+			double cost_time = 0;
+			double network_time = 0;
+			double hit_time = 0;
+		#endif
+
 		//
 		stim::vec3<float> p;  	//vector designating the position of the spider.
 		stim::vec3<float> d;	//vector designating the orientation of the spider
@@ -118,11 +133,19 @@ class gl_spider : public virtual gl_texture&lt;T&gt;
 		void
 		findOptimalDirection()
 		{
+			#ifdef TIMING
+				gpuStartTimer();
+			#endif
 			setMatrix();			//create the transformation matrix.
 			glCallList(dList);		//move the templates to p, d, m.
+			glFlush();
+			#ifdef TIMING
+				direction_time += gpuStopTimer();
+			#endif
 			#ifdef TESTING
-			test(texbufferID, GL_TEXTURE_2D,2*t_length,numSamples*t_length, "Final_Cost_Direction.bmp");
+//				test(texbufferID, GL_TEXTURE_2D,2*t_length,numSamples*t_length, "Final_Cost_Direction.bmp");
 			#endif
+
 			int best = getCost(texbufferID,numSamples);		//find min cost.
 			stim::vec<float> next(		//find next vector.
 			dV[best][0]*S[0]*R[0],
@@ -143,10 +166,18 @@ class gl_spider : public virtual gl_texture&lt;T&gt;
 		void
 		findOptimalPosition()
 		{
+			#ifdef TIMING
+				gpuStartTimer();
+			#endif
 			setMatrix();			//create the transformation matrix.
 			glCallList(dList+1);		//move the templates to p, d, m.
+			glFlush();
+			#ifdef TIMING
+				position_time += gpuStopTimer();
+			#endif
+
 			#ifdef TESTING
-			test(ptexbufferID, GL_TEXTURE_2D,2*t_length, numSamplesPos*t_length, "Final_Cost_Position.bmp");
+//				test(ptexbufferID, GL_TEXTURE_2D,2*t_length, numSamplesPos*t_length, "Final_Cost_Position.bmp");
 			#endif
 			int best = getCost(ptexbufferID, numSamplesPos);		//find min cost.
 //			std::cerr << best << std::endl;
@@ -169,10 +200,17 @@ class gl_spider : public virtual gl_texture&lt;T&gt;
 		void
 		findOptimalScale()
 		{
+			#ifdef TIMING
+				gpuStartTimer();
+			#endif
 			setMatrix();			//create the transformation.
 			glCallList(dList+2);		//move the templates to p, d, m.
+			glFlush();
+			#ifdef TIMING
+				size_time += gpuStopTimer();
+			#endif
 			#ifdef TESTING
-			test(mtexbufferID, GL_TEXTURE_2D, 2*t_length, numSamplesMag*t_length, "Final_Cost_Position.bmp");
+//				test(mtexbufferID, GL_TEXTURE_2D, 2*t_length, numSamplesMag*t_length, "Final_Cost_Position.bmp");
 			#endif
 			int best = getCost(mtexbufferID, numSamplesMag);		//get best cost.
 			setMagnitude(m[0]*mV[best][0]);	//adjust the magnitude.
@@ -231,6 +269,10 @@ class gl_spider : public virtual gl_texture&lt;T&gt;
 		void
 		branchDetection2(int n = 8, int l_template = 8, int l_square = 8)
 		{
+			#ifdef TIMING
+				gpuStartTimer();
+			#endif
+
 			if(cL.size() < 4){}
 			else{
 			setMatrix(1);
@@ -273,6 +315,9 @@ class gl_spider : public virtual gl_texture&lt;T&gt;
 				}
 			}
 			}
+			#ifdef TIMING
+				branch_time += gpuStopTimer();
+			#endif
 		}
@@ -305,58 +350,6 @@ class gl_spider : public virtual gl_texture&lt;T&gt;
 //---------------------TEMPLATE CREATION METHODS----------------------------//
 //--------------------------------------------------------------------------//
-		void
-		MonteCarloDirectionVectors(int nSamples, float solidAngle = 2*M_PI)
-		{
-///equal area projection.
-//			std::vector<stim::vec<float> > vecsUni;
-//			for(int i = 0; i < nSamples; i++)
-//			{
-//				stim::vec<float> a = uniformRandVector();
-//				a[1] = a[1]*2*M_PI;
-//				a[0] = a[0]*(solidAngle);
-//				a = a.cyl2cart();
-//				stim::vec<float> b(
-//					sqrt(1.-(pow(a[0],2)+(pow(a[1],2)))/4)*a[0],
-//					sqrt(1.-(pow(a[0],2)+(pow(a[1],2)))/4)*a[1],
-//					1.-(pow(a[0],2)+pow(a[1],2))/2);
-//					2.*a[0]/(1.+pow(a[0],2)+pow(a[1],2)),
-//					2.*a[1]/(1.+pow(a[0],2)+pow(a[1],2)),
-//					(-1.0+pow(a[0],2)+pow(a[1],2))/(1.0+pow(a[0],2)+pow(a[1],2)));
-//				b = b.norm();
-//				b[2] = -b[2]; //this doeswn't want to change
-//				vecsUni.push_back(b);
-//			}
-			float PHI[2], Z[2], range;
-			PHI[0] = asin(solidAngle/2);
-			PHI[1] = asin(0);
-			
-			Z[0] = cos(PHI[0]);
-			Z[1] = cos(PHI[1]);
-
-			
-			range = Z[0] - Z[1];
-
-
-			float z, theta, phi;
-			
-			std::vector<stim::vec3<float> > vecsUni;
-			for(int i = 0; i < numSamplesPos; i++)
-			{
-				stim::vec3<float> a(uniformRandom()*0.8, uniformRandom()*0.8, 0.0);
-				a[0] = a[0]-0.4;
-				a[1] = a[1]-0.4;
-				vecsUni.push_back(a);
-			}
-
-			stringstream name;
-			for(int i = 0; i < numSamplesPos; i++)
-				name << vecsUni[i].str() << std::endl;
-			
-			std::ofstream outFile;
-			outFile.open("New_Pos_Vectors.txt");
-			outFile << name.str().c_str();
-		}
 		///@param solidAngle, the size of the arc to sample.
@@ -759,9 +752,6 @@ class gl_spider : public virtual gl_texture&lt;T&gt;
 		{
 			Bind();
 			CHECK_OPENGL_ERROR
-			#ifdef TESTING
-				start = std::clock();
-			#endif
 			findOptimalDirection();
 			findOptimalPosition();
 			findOptimalScale();
@@ -769,12 +759,6 @@ class gl_spider : public virtual gl_texture&lt;T&gt;
 			Bind(btexbufferID, bfboID, 27);
 			branchDetection();
 			Unbind();
-
-			#ifdef TESTING
-				duration_sampling = duration_sampling +
-					 (std::clock() - start) / (double) CLOCKS_PER_SEC;
-				num_sampling = num_sampling + 1.0; 
-			#endif
 			return current_cost;
 		}
@@ -792,17 +776,9 @@ class gl_spider : public virtual gl_texture&lt;T&gt;
 		int
 		getCost()
 		{
-			#ifdef TESTING
-				start = std::clock();
-			#endif
 			stim::vec<int> cost = 
 				stim::cuda::get_cost(texbufferID, GL_TEXTURE_2D, numSamples);
 			cudaDeviceSynchronize();
-			#ifdef TESTING
-				duration_cuda = duration_cuda +
-					 (std::clock() - start) / (double) CLOCKS_PER_SEC;
-				num_cuda = num_cuda + 1.0;
-			#endif
 			current_cost = cost[1];
 			return cost[0];
 		}
@@ -810,16 +786,13 @@ class gl_spider : public virtual gl_texture&lt;T&gt;
 		int
 		getCost(GLuint tID, int n)
 		{
-			#ifdef TESTING
-				start = std::clock();
+			#ifdef TIMING
+				gpuStartTimer();
 			#endif
 			stim::vec<int> cost = 
 				stim::cuda::get_cost(tID, GL_TEXTURE_2D, n, 2*t_length, t_length);
-			cudaDeviceSynchronize();
-			#ifdef TESTING
-				duration_cuda = duration_cuda +
-					 (std::clock() - start) / (double) CLOCKS_PER_SEC;
-				num_cuda = num_cuda + 1.0;
+			#ifdef TIMING
+				cost_time += gpuStopTimer();
 			#endif
 			current_cost = cost[1];
 			return cost[0];
@@ -839,13 +812,9 @@ class gl_spider : public virtual gl_texture&lt;T&gt;
 		//vectical rectangle forming the spider.
 		stim::rect<float> ver;	
-		//Testing and Timing variables.
+		//Timing variables.
 		#ifdef TESTING
 			std::clock_t start;
-			double duration_sampling = 0.0;
-			double duration_cuda = 0.0;
-			double num_sampling = 0.0;
-			double num_cuda = 0.0;
 		#endif
 //--------------------------------------------------------------------------//
@@ -939,10 +908,13 @@ class gl_spider : public virtual gl_texture&lt;T&gt;
 			texID = id;
 			//GenerateFBO(16, numSamples*8);
 			GenerateFBO(t_length*2, numSamples*t_length, texbufferID, fboID); 
+			std::cout << numSamples << std::endl;
 			CHECK_OPENGL_ERROR
 			GenerateFBO(t_length*2, numSamplesPos*t_length, ptexbufferID, pfboID); 
+			std::cout << numSamplesPos << std::endl;
 			CHECK_OPENGL_ERROR
 			GenerateFBO(t_length*2, numSamplesMag*t_length, mtexbufferID, mfboID); 
+			std::cout << numSamplesMag << std::endl;
 			CHECK_OPENGL_ERROR
 			GenerateFBO(16, 216, btexbufferID, bfboID); 
 			CHECK_OPENGL_ERROR
@@ -1204,6 +1176,47 @@ class gl_spider : public virtual gl_texture&lt;T&gt;
 			return seeds;
 		}
+		///sets the number of direction templates.
+		void
+		setNumberDirectionTemplates(int n)
+		{
+			numSamples = n;
+		}
+
+		///sets the number of position templates.
+		void
+		setNumberPositionTemplates(int n)
+		{
+			numSamplesPos = n;
+		}
+
+		///sets the number of position templates.
+		void
+		setNumberSizeTemplates(int n)
+		{
+			numSamplesMag = n;
+		}
+
+		#ifdef TIMING
+		///Returns the timings at the moment the method is called.
+		///In the following order: Branch, Direction, Position, Size, Cost, Network, Hit_detetion.
+		std::vector<double> 
+		getTimings()
+		{
+			std::vector <double> ret;
+			ret.resize(7);
+			ret[0] = branch_time;
+			ret[1] = direction_time;
+			ret[2] = position_time;
+			ret[3] = size_time;
+			ret[4] = cost_time;
+			ret[5] = network_time;
+			ret[6] = hit_time;
+
+			return ret;
+		}
+		#endif
+
 		///returns true if all seed stacks are empty, else false.
 		bool
 		Empty()
@@ -1302,9 +1315,6 @@ class gl_spider : public virtual gl_texture&lt;T&gt;
 		{
 			Bind(texbufferID, fboID, numSamples, t_length);
 			CHECK_OPENGL_ERROR
-			#ifdef TESTING
-				start = std::clock();
-			#endif
 				findOptimalDirection();
 			Unbind();
 			Bind(ptexbufferID, pfboID, numSamplesPos, t_length);
@@ -1315,12 +1325,6 @@ class gl_spider : public virtual gl_texture&lt;T&gt;
 			Unbind();
 			CHECK_OPENGL_ERROR
-			#ifdef TESTING
-				duration_sampling = duration_sampling +
-					 (std::clock() - start) / (double) CLOCKS_PER_SEC;
-				num_sampling = num_sampling + 1.0; 
-			#endif
-			std::cout << current_cost << std::endl;
 			return current_cost;
 		}
@@ -1337,6 +1341,39 @@ class gl_spider : public virtual gl_texture&lt;T&gt;
 //--------------------------------------------------------------------------//
 		void
+		MonteCarloDirectionVectors(int nSamples, float solidAngle = 2*M_PI)
+		{
+			float PHI[2], Z[2], range;
+			PHI[0] = asin(solidAngle/2);
+			PHI[1] = asin(0);
+			
+			Z[0] = cos(PHI[0]);
+			Z[1] = cos(PHI[1]);
+
+			
+			range = Z[0] - Z[1];
+
+
+			float z, theta, phi;
+			
+			std::vector<stim::vec3<float> > vecsUni;
+			for(int i = 0; i < numSamplesPos; i++)
+			{
+				stim::vec3<float> a(uniformRandom()*0.8, uniformRandom()*0.8, 0.0);
+				a[0] = a[0]-0.4;
+				a[1] = a[1]-0.4;
+				vecsUni.push_back(a);
+			}
+
+			stringstream name;
+			for(int i = 0; i < numSamplesPos; i++)
+				name << vecsUni[i].str() << std::endl;
+			
+			std::ofstream outFile;
+			outFile.open("New_Pos_Vectors.txt");
+			outFile << name.str().c_str();
+		}
+		void
 		DrawCylinder()
 		{	 
 			 glNewList(dList+3, GL_COMPILE);
@@ -1442,6 +1479,10 @@ class gl_spider : public virtual gl_texture&lt;T&gt;
 		selectObject(stim::vec3<float> loc, stim::vec3<float> dir, float mag) 
 		{
 		//Define the varibles and turn on Selection Mode
+			
+			#ifdef TIMING
+				gpuStartTimer();
+			#endif
 			float s = 3.0;
 			GLuint selectBuf[2048];
@@ -1501,6 +1542,10 @@ class gl_spider : public virtual gl_texture&lt;T&gt;
 		//	glEnable(GL_CULL_FACE);
 			hits = glRenderMode(GL_RENDER);
 			int found_hits = processHits(hits, selectBuf);
+			#ifdef TIMING
+				hit_time += gpuStopTimer();
+			#endif
+
 			return found_hits;
 		}
@@ -1548,6 +1593,10 @@ class gl_spider : public virtual gl_texture&lt;T&gt;
 		addToNetwork(pair<stim::fiber<float>, int> in, stim::vec3<float> spos, 
 				stim::vec<float> smag, stim::vec3<float> sdir)
 		{
+			#ifdef TIMING
+				 double s = std::clock();
+			#endif
+			
                         std::vector<stim::vec3<float> > ce = in.first.centerline();                
                         std::vector<stim::vec<float> > cm = in.first.centerlinemag();
 			//if the fiber is longer than 2 steps (the number it takes to diverge)
@@ -1581,6 +1630,11 @@ class gl_spider : public virtual gl_texture&lt;T&gt;
 					} else { nt.addEdge(ce,cm, -1, -1);}
 				}
 			}		
+
+			#ifdef TIMING
+				double network_time = (std::clock() - s) / (double) CLOCKS_PER_SEC;
+				network_time += network_time * 1000.0;
+			#endif
 		}
@@ -193,6 +193,7 @@ public:
 			get_interleaved_bgr(buffer);
 		cv::Mat cvImage((int)Y(), (int)X(), cv_type(), buffer);
 		cv::imwrite(filename, cvImage);
+		free(buffer);
 	}
 	void set_interleaved(T* buffer, size_t width, size_t height, size_t channels){