#include <stdio.h>
#include <string.h>
#include <stdlib.h>
#include <unistd.h>
#include <sys/times.h>
#include <cilk/cilk.h>
#include <cilk/cilk_api.h>

int main(argc,argv) int argc; char *argv[]; {
    unsigned *buf[128];
    unsigned ncheck[129],jcheck[128];
    unsigned bsz,stp;
    clock_t rstart,rend,cstart,cend,istart,iend;
    struct tms tm;
    char *pname;
    double srun,sela,crun,cela,cper,scre;
    long clk_tck=sysconf(_SC_CLK_TCK);

    istart=times(&tm);
    if(pname=rindex(argv[0],'/')) pname++;
    else pname=argv[0];
    printf("%s - Eric's memory/cache benchmark v2.0-cilk\n",pname);
    if(argc!=3){
        fprintf(stderr,"Usage: %s size steps\n\n"
"where size is the size of the memory buffer to allocate and\n"
"steps is the number of memory accesses to make.\n",pname);
        exit(1);
    }
    bsz=atoi(argv[1]);
    stp=atoi(argv[2]);
    int ncpu = __cilkrts_get_nworkers();
    printf("CILK_NWORKERS=%d\nsizeof(unsigned)=%u\n",
        ncpu,(unsigned)sizeof(unsigned));
    printf(
"Performing test: size=%d, steps=%d, sizeof(int)=%d...",
        bsz,stp,sizeof(int));
    for(int cpu=0;cpu<ncpu;cpu++){
        if(!(buf[cpu]=calloc(bsz,sizeof(int)))){
            fprintf(stderr,"Memory allocation of integer array failed!\n");
            exit(2);
        }
    }
    for(int cpu=0;cpu<ncpu;cpu++){
        for(unsigned i=0;i<bsz;i++) buf[cpu][i]=i;
    }
    iend=times(&tm);

    rstart=times(&tm);
    cstart=tm.tms_utime+tm.tms_stime;
    cilk_for(int cpu=0;cpu<ncpu;cpu++){
        unsigned i,j,n;
        unsigned *mybuf=buf[cpu];
        for(j=i=0;i<stp;i++) {
            n=(j*mybuf[j]+i)%bsz;
            mybuf[n]+=j;
            j=(n*mybuf[n]+i+3)%bsz;
            mybuf[j]+=n+3;
        }
        ncheck[cpu]=n;
        jcheck[cpu]=j;
    }
    rend=times(&tm);
    cend=tm.tms_utime+tm.tms_stime;

    scre=(double)(iend-istart)/(double)clk_tck;
    srun=(double)(cend-cstart)/(double)clk_tck;
    sela=(double)(rend-rstart)/(double)clk_tck;
    if(srun==0.0) {
        crun=0.0;
    } else {
        crun=(double)stp*ncpu/srun;
    }
    if(sela==0.0) {
        cper=100.0;
        cela=0.0;
    } else {
        cper=100.0*srun/sela;
        cela=(double)stp*ncpu/sela;
    }

    unsigned nch0=buf[0][ncheck[0]],jch0=buf[0][jcheck[0]];
    for(int cpu=1;cpu<ncpu;cpu++) {
        if(buf[cpu][ncheck[cpu]]!=nch0
         ||buf[cpu][jcheck[cpu]]!=jch0){
            fprintf(stderr,"Inconsistent check values!\n");
        }
    }
    printf("Done.\n\n"
           "Startup time: %g sec\n"
           "Running time: %g sec -- %g/sec\n"
           "Elapsed time: %g sec -- %g/sec given %g %%CPU\n"
           "Check values: %x -- %x\n",
            scre,srun,crun,sela,cela,cper,
            nch0,jch0);
    exit(0);
}