#ifndef STIM_CUDA_UPDATE_DIR_GLOBALD_H
#define STIM_CUDA_UPDATE_DIR_GLOBAL_H

# include <iostream>
# include <cuda.h>
#include <stim/cuda/cudatools.h>
#include <stim/cuda/sharedmem.cuh>
#include "cpyToshare.cuh" 

#define RMAX_TEST	8

namespace stim{
	namespace cuda{
	
		// this kernel calculates the voting direction for the next iteration based on the angle between the location of this voter and the maximum vote value in its voting area.
		template<typename T>
		__global__ void cuda_update_dir(T* gpuDir, T* gpuVote, T* gpuGrad, T* gpuTable, T phi, int rmax,  int x,  int y){

			
			//calculate the start point for this block
			int bxi = blockIdx.x * blockDim.x;
			
			// calculate the 2D coordinates for this current thread.
			int xi = bxi + threadIdx.x;
			if(xi >= x) return;													//if the index is outside of the image, terminate the kernel
			int yi = blockIdx.y * blockDim.y + threadIdx.y;			
			int i = yi * x + xi;												// convert 2D coordinates to 1D
			
			float theta = gpuGrad[2*i];											// calculate the voting direction based on the grtadient direction - global memory fetch			
			gpuDir[i] = 0;														//initialize the vote direction to zero			
			float max = 0;														// define a local variable to maximum value of the vote image in the voting area for this voter
			int id_x = 0;														// define two local variables for the x and y position of the maximum
			int id_y = 0;
			
			int x_table = 2*rmax +1;											// compute the size of window which will be checked for finding the voting area for this voter
			int rmax_sq = rmax * rmax;
			int tx_rmax = threadIdx.x + rmax;
			float atan_angle;
			float vote_c;
			for(int yr = -RMAX_TEST; yr <= RMAX_TEST; yr++){
				if (yi+yr >= 0 && yi + yr < y){
					for(int xr = -RMAX_TEST; xr <= RMAX_TEST; xr++){

						unsigned int ind_t = (RMAX_TEST - yr) * x_table + RMAX_TEST - xr;

						// calculate the angle between the voter and the current pixel in x and y directions
						atan_angle = gpuTable[ind_t];
						
						// find the vote value for the current counter
						vote_c = gpuVote[(yi+yr)*x + (xi+xr)];
						
						// check if the current pixel is located in the voting area of this voter.
						if (((xr * xr + yr *yr)< rmax_sq) && (abs(atan_angle - theta) <phi)){

						// compare the vote value of this pixel with the max value to find the maxima and its index.
							if  (vote_c>max) {

								max = vote_c;
								id_x =  xr;
								id_y =  yr;
							}
						}
					}
				}
			}
							
			unsigned int ind_m = (rmax - id_y) * x_table + (rmax - id_x);
			float new_angle = gpuTable[ind_m];

			if(xi < x && yi < y)
				gpuDir[i] = new_angle;
		}										//end kernel

		// this kernel updates the gradient direction by the calculated voting direction.
		template<typename T>
		__global__ void cuda_update_grad(T* gpuGrad, T* gpuDir, int x, int y){

			// calculate the 2D coordinates for this current thread.
			int xi = blockIdx.x * blockDim.x + threadIdx.x;
			int yi = blockIdx.y * blockDim.y + threadIdx.y;
		
			// convert 2D coordinates to 1D
			int i = yi * x + xi;
			
			//update the gradient image with the vote direction
			gpuGrad[2*i] = gpuDir[i];
		}
		
		template<typename T>
		void gpu_update_dir(T* gpuVote, T* gpuGrad, T* gpuTable, T phi, unsigned int rmax, unsigned int x, unsigned int y){

			//calculate the number of bytes in the array
			unsigned int bytes = x * y * sizeof(T);

			unsigned int max_threads = stim::maxThreadsPerBlock();
			dim3 threads(max_threads, 1);
			dim3 blocks(x/threads.x + (x %threads.x == 0 ? 0:1) , y);
			
			// allocate space on the GPU for the updated vote direction
			T* gpuDir;
			cudaMalloc(&gpuDir, bytes);	

			//call the kernel to calculate the new voting direction
			cuda_update_dir <<< blocks, threads>>>(gpuDir, gpuVote, gpuGrad, gpuTable, phi, rmax, x , y);

			//call the kernel to update the gradient direction
			cuda_update_grad <<< blocks, threads >>>(gpuGrad, gpuDir, x , y);
			
			//free allocated memory
			cudaFree(gpuDir);

		}
		
		template<typename T>
		void cpu_update_dir(T* cpuVote, T* cpuGrad,T* cpuTable, T phi, unsigned int rmax, unsigned int x, unsigned int y){

			//calculate the number of bytes in the array
			unsigned int bytes = x * y * sizeof(T);

			//calculate the number of bytes in the atan2 table
			unsigned int bytes_table = (2*rmax+1) * (2*rmax+1) * sizeof(T);

			//allocate space on the GPU for the Vote Image
			T* gpuVote;
			cudaMalloc(&gpuVote, bytes);

			//copy the input vote image to the GPU
			HANDLE_ERROR(cudaMemcpy(gpuVote, cpuVote, bytes, cudaMemcpyHostToDevice));	

			//allocate space on the GPU for the input Gradient image
			T* gpuGrad;
			HANDLE_ERROR(cudaMalloc(&gpuGrad, bytes*2));

			//copy the Gradient data to the GPU
			HANDLE_ERROR(cudaMemcpy(gpuGrad, cpuGrad, bytes*2, cudaMemcpyHostToDevice));

			//allocate space on the GPU for the atan2 table
			T* gpuTable;
			HANDLE_ERROR(cudaMalloc(&gpuTable, bytes_table));

			//copy the atan2 values to the GPU
			HANDLE_ERROR(cudaMemcpy(gpuTable, cpuTable, bytes_table, cudaMemcpyHostToDevice));
						
			//call the GPU version of the update direction function
			gpu_update_dir<T>(gpuVote, gpuGrad, gpuTable, phi, rmax, x , y);
							
			//copy the new gradient image back to the CPU
			cudaMemcpy(cpuGrad, gpuGrad, bytes*2, cudaMemcpyDeviceToHost) ;

			//free allocated memory
			cudaFree(gpuTable);
			cudaFree(gpuVote);
			cudaFree(gpuGrad);
		}
		
	}
}

#endif