String matching in CUDA shows different result when increase blocks

You need to synchronize threads of the same block using __syncthreads. For example, local_matches = 0 can theoretically be done concurrently to the atomicAdd of the main loop. Thus, a __syncthreads is needed between both. For the same reason, you also need a __syncthreads before the last if(threadIdx.x == 0). I am not sure this is the only error.

I advise you to use CUDA-GDB to track such bug since the kernel is quite short and relatively simple.

Note that a local_matches do not need to be shared. In fact, it is not efficient to do that. On can perform the reduction in local memory and then perform a final atomicAdd. Additionally, you do not need the innermost conditional. You can simply do: local_matches += i >= pattern_size; (the compiler might already do such optimization).

I finally found a solution for this.

I set a match table with 0 values for every buffer position, set 1 for every position that pattern found, and added the 1's in the CPU.

If you think something better please add an answer.

This is the code:

#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <cuda.h>

__global__ void string_matching(char *buffer, char *pattern, int match_size, int pattern_size, int *match){
    int tid, i;

    for(tid=blockIdx.x*blockDim.x+threadIdx.x; tid<match_size; tid+=blockDim.x){
        for (i = 0; i < pattern_size && pattern[i] == buffer[i + tid]; ++i);
        if(i >= pattern_size){
            match[tid] = 1;
        }
        else{
            match[tid] = 0;
        }
    }

}


int main(int argc, char *argv[]){
    FILE *pFile;
    int i;
    long file_size, match_size, pattern_size;
    char * buffer;
    char * filename, *pattern;
    size_t result;
    int *match, total_matches;

    //CUDA variables
    int blocks, threads_per_block;
    int *match_dev;
    char *buffer_dev, *pattern_dev;

    float total_time, comp_time;
    cudaEvent_t total_start, total_stop, comp_start, comp_stop;
    cudaEventCreate(&total_start);
    cudaEventCreate(&total_stop);
    cudaEventCreate(&comp_start);
    cudaEventCreate(&comp_stop);

    if (argc != 5) {
        printf ("Usage : %s <file_name> <string> <blocks> <threads_per_block>\n", argv[0]);
        return 1;
    }
    filename = argv[1];
    pattern = argv[2];
    blocks = strtol(argv[3], NULL, 10);
    threads_per_block = strtol(argv[4], NULL, 10);
    
    pFile = fopen ( filename , "rb" );
    if (pFile==NULL) {printf ("File error\n"); return 2;}

    // obtain file size:
    fseek (pFile , 0 , SEEK_END);
    file_size = ftell (pFile);
    rewind (pFile);
    printf("file size is %ld\n", file_size);
    
    // allocate memory to contain the file:
    buffer = (char*) malloc (sizeof(char)*file_size);
    if (buffer == NULL) {printf ("Memory error\n"); return 3;}

    // copy the file into the buffer:
    result = fread (buffer,1,file_size,pFile);
    if (result != file_size) {printf ("Reading error\n"); return 4;} 
    
    pattern_size = strlen(pattern);
    match_size = file_size - pattern_size + 1;
    
    match = (int *) malloc (sizeof(int)*match_size);
    if (match == NULL) {printf ("Malloc error\n"); return 5;}

    cudaMalloc((void **)&match_dev, match_size*sizeof(int));
    cudaMalloc((void **)&buffer_dev, file_size*sizeof(char));
    cudaMalloc((void **)&pattern_dev, pattern_size*sizeof(char));

    cudaEventRecord(total_start);

    cudaEventRecord(comp_start);

    cudaMemcpy(buffer_dev, buffer, file_size*sizeof(char), cudaMemcpyHostToDevice);
    cudaMemcpy(pattern_dev, pattern, pattern_size*sizeof(char), cudaMemcpyHostToDevice);

    string_matching<<<blocks, threads_per_block>>>(buffer_dev, pattern_dev, match_size, pattern_size, match_dev);
    cudaThreadSynchronize();

    cudaEventRecord(comp_stop);
    cudaEventSynchronize(comp_stop);
    cudaEventElapsedTime(&comp_time, comp_start, comp_stop);

    cudaMemcpy(match, match_dev, match_size*sizeof(int), cudaMemcpyDeviceToHost);

    cudaEventRecord(total_stop);
    cudaEventSynchronize(total_stop);
    cudaEventElapsedTime(&total_time, total_start, total_stop);

    total_matches = 0;
    for(i=0; i<match_size; i++){
        total_matches += match[i];
    }

    cudaFree(match_dev);
    cudaFree(buffer_dev);
    cudaFree(pattern_dev);

    fclose (pFile);
    free (buffer);

    //Print result
    printf("Total matches: %d\n", total_matches);

    printf("\n\n\nN: %d, Blocks: %d, Threads: %d\n", file_size, blocks, blocks*threads_per_block);
    printf("Total time (ms): %.3f\n", total_time);
    printf("Kernel time (ms): %.3f\n", comp_time);
    printf("Data transfer time(ms): %.3f\n\n\n", total_time-comp_time);

}

I think this is a better solution.

It works only for power of 2 threads per block.

#include <stdio.h>
#include <stdlib.h>
#include <string.h>
#include <cuda.h>

#define MAX_THREADS_PER_BLOCK 100

__global__ void string_matching(char *buffer, char *pattern, int match_size, int pattern_size, int blocks, int slice, int extra, int *gout){
    int tid, i;
    int thread_index = blockIdx.x*blockDim.x + threadIdx.x;
    int start = thread_index*slice;
    int stop = start + slice;
    if(thread_index == blocks*blockDim.x - 1){
        stop += extra;
    }
    if(stop > match_size){
        stop = match_size;
    }
    __shared__ int r[MAX_THREADS_PER_BLOCK];
    int sum = 0;

    for(tid=start; tid<stop; tid++){
        for (i = 0; i < pattern_size && pattern[i] == buffer[i + tid]; ++i);
        if(i >= pattern_size){
            sum++;
        }
    }

    r[threadIdx.x] = sum;

    __syncthreads();

    //works only for power of 2 threads_per_block
    for (int size = blockDim.x/2; size>0; size/=2) { //uniform
        if (threadIdx.x<size)
            r[threadIdx.x] += r[threadIdx.x+size];
        __syncthreads();
    }



    printf("Block: %d, Thread: %d, Global Thread: %d, Start: %d, Stop: %d, Matches: %d, Block Matches: %d\n", blockIdx.x, threadIdx.x, thread_index, start, stop, r[threadIdx.x], r[0]);


    if(threadIdx.x == 0){
        gout[blockIdx.x] = r[0];
    }

    
}


int main(int argc, char *argv[]){
    int i;
    FILE *pFile;
    long file_size, match_size, pattern_size;
    char * buffer;
    char * filename, *pattern;
    size_t result;
    int *results;
    int total_matches;

    //CUDA variables
    int blocks, threads_per_block, total_threads, slice, extra;
    int *results_dev;
    char *buffer_dev, *pattern_dev;

    float total_time, comp_time;
    cudaEvent_t total_start, total_stop, comp_start, comp_stop;
    cudaEventCreate(&total_start);
    cudaEventCreate(&total_stop);
    cudaEventCreate(&comp_start);
    cudaEventCreate(&comp_stop);

    if (argc != 5) {
        printf ("Usage : %s <file_name> <string> <blocks> <threads_per_block>\n", argv[0]);
        return 1;
    }
    filename = argv[1];
    pattern = argv[2];
    blocks = strtol(argv[3], NULL, 10);
    threads_per_block = strtol(argv[4], NULL, 10);
    
    pFile = fopen ( filename , "rb" );
    if (pFile==NULL) {printf ("File error\n"); return 2;}

    // obtain file size:
    fseek (pFile , 0 , SEEK_END);
    file_size = ftell (pFile);
    rewind (pFile);
    printf("file size is %ld\n", file_size);
    
    // allocate memory to contain the file:
    buffer = (char*) malloc (sizeof(char)*file_size);
    if (buffer == NULL) {printf ("Memory error\n"); return 3;}

    // copy the file into the buffer:
    result = fread (buffer,1,file_size,pFile);
    if (result != file_size) {printf ("Reading error\n"); return 4;} 
    
    pattern_size = strlen(pattern);
    match_size = file_size - pattern_size + 1;

    results = (int *)malloc(blocks*sizeof(int));

    cudaMalloc((void **)&results_dev, blocks*sizeof(int));
    cudaMalloc((void **)&buffer_dev, file_size*sizeof(char));
    cudaMalloc((void **)&pattern_dev, pattern_size*sizeof(char));

    cudaEventRecord(total_start);

    cudaEventRecord(comp_start);

    cudaMemcpy(buffer_dev, buffer, file_size*sizeof(char), cudaMemcpyHostToDevice);
    cudaMemcpy(pattern_dev, pattern, pattern_size*sizeof(char), cudaMemcpyHostToDevice);

    total_threads = blocks*threads_per_block;
    slice = match_size/total_threads;
    extra = match_size%total_threads;

    string_matching<<<blocks, threads_per_block>>>(buffer_dev, pattern_dev, match_size, pattern_size, blocks, slice, extra, results_dev);

    cudaEventRecord(comp_stop);
    cudaEventSynchronize(comp_stop);
    cudaEventElapsedTime(&comp_time, comp_start, comp_stop);

    cudaMemcpy(results, results_dev, blocks*sizeof(int), cudaMemcpyDeviceToHost);

    total_matches = 0;
    for(i=0; i<blocks; i++){
        total_matches += results[i];
    }

    cudaEventRecord(total_stop);
    cudaEventSynchronize(total_stop);
    cudaEventElapsedTime(&total_time, total_start, total_stop);

    cudaFree(results_dev);
    cudaFree(buffer_dev);
    cudaFree(pattern_dev);

    fclose (pFile);
    free (buffer);

    //Print result
    printf("Total matches: %d\n", total_matches);

    printf("\n\n\nN: %d, Blocks: %d, Threads: %d\n", file_size, blocks, blocks*threads_per_block);
    printf("Total time (ms): %.3f\n", total_time);
    printf("Kernel time (ms): %.3f\n", comp_time);
    printf("Data transfer time(ms): %.3f\n\n\n", total_time-comp_time);

}

The stride on your loop is not correct for using multiple blocks, i.e. both blocks were doing the full amount of work (and therefore finding two entries each). The correct grid-stride loop looks like the following

for (int tid = blockIdx.x * blockDim.x + threadIdx.x; 
         tid < match_size; 
         tid += blockDim.x * gridDim.x /* <-- fix */) {
    // ...
}