// includes, system
#include <stdlib.h>
#include <stdio.h>
#include <string.h>
#include <math.h>

// includes, project
#include <cuda.h>

__global__ void stringMatcherComplex(char* subject, int subjectSize, char* pattern, int patternSize, int* numberOfOccurences)
{
	int numberOfThreads = ceil((double)subjectSize / (double)patternSize);
	int threadId = threadIdx.x;
	int i,j;
	if (threadId == 0){
		for (j=0; j<patternSize && pattern[j]==subject[j]; ++j);
			if (j==patternSize)
				(*numberOfOccurences)++;
	}else{
		for (i=((threadId-1)*patternSize+1); i<(threadId*patternSize+1); ++i){
			for (j=0; j<patternSize && pattern[j]==subject[i+j]; ++j);
				if (j==patternSize)
					(*numberOfOccurences)++;
		}
	}
}

__global__ void stringMatcherExhaustive(char* subject, int subjectSize, char* pattern, int patternSize, int* numberOfOccurences)
{
	int threadId = threadIdx.x;
	int j;
	for (j=0; j<patternSize && pattern[j]==subject[j+threadId]; ++j);
		if (j==patternSize)
			(*numberOfOccurences)++;
}

int main(int argc, char** argv)
{
	//Maximum 512 threads
	char* subject = "blablabla";
	int subjectSize = strlen(subject);
	char* subjectd;
	cudaMalloc(&subjectd, subjectSize*sizeof(char));
	cudaMemcpy(subjectd, subject, subjectSize*sizeof(char), cudaMemcpyHostToDevice);

	char* pattern = "bla";
	int patternSize = strlen(pattern);
	char* patternd;
	cudaMalloc(&patternd, patternSize*sizeof(char));
	cudaMemcpy(patternd, pattern, patternSize*sizeof(char), cudaMemcpyHostToDevice);

	int numberOfThreads = ceil((double)subjectSize / (double)patternSize);

	int numberOfOccurences = 0;
	int* numberOfOccurencesd;
	cudaMalloc(&numberOfOccurencesd, sizeof(int));
	cudaMemcpy(numberOfOccurencesd, &numberOfOccurences, sizeof(int), cudaMemcpyHostToDevice);

	stringMatcherComplex<<<1, numberOfThreads>>>(subjectd, subjectSize, patternd, patternSize, numberOfOccurencesd);
	//stringMatcherExhaustive<<<1, subjectSize-patternSize+1>>>(subjectd, subjectSize, patternd, patternSize, numberOfOccurencesd);
	cudaMemcpy(&numberOfOccurences, numberOfOccurencesd, sizeof(int), cudaMemcpyDeviceToHost);

	printf("Number of occurences: %d\n",numberOfOccurences);

	cudaFree(subjectd);
	cudaFree(patternd);
	cudaFree(numberOfOccurencesd);
}