// Consulting cap-lore.com/MathPhys/fft/MPfft/ft.c
// Eight threads contend (with __sync_fetch_and_add) for indexes
// within array t at which to store their thread number.
#include <stdio.h>
#include <pthread.h>
#include <stdlib.h>
#define cpus 8
typedef void * thnk(void *);
static pthread_t mt(void * msg, thnk rtn){
   pthread_attr_t attr;
   pthread_attr_init(&attr);
   pthread_attr_setdetachstate(&attr, PTHREAD_CREATE_JOINABLE);
   {
      pthread_t zz;
      int rc = pthread_create(&zz, &attr, rtn, msg);
      if(rc) {printf(
            "ERROR; return code from pthread_create() is %d\n", rc);
         exit(-1);}
       pthread_attr_destroy(&attr);
       return zz;}}
typedef struct{int first; int * ret;} blkmsg;
// unneeded type, but usually a way to send binary messages to each thread.
#define sz 10000
int cur = 0;
char t[sz]; int cnts[cpus];
void* fab(void* w){int me=((blkmsg*)w)->first; int p=0;
  while(1){int k = __sync_fetch_and_add(&cur, 1);
    if(k>=sz) break; t[k] = me+1; ++p;}
  printf("Bye %d %d\n", me+1, p); cnts[me] = p; return w;}
int main(){for(int j=0; j<sz; ++j) t[j] = 0;
  blkmsg bm[cpus];
  pthread_t thrds[cpus];
  for(int n = 0; n < cpus; ++n) {bm[n] = (blkmsg){n, cnts+n};      
         thrds[n] = mt(&bm[n], fab);}
   for(int n = 0; n < cpus; ++n) {int rc = pthread_join(thrds[n], NULL);
          if(rc) {printf("ERROR; return code from pthread_join() is %d\n", rc);
             exit(-1);}}
{int H[cpus]; for(int j=0; j<cpus; ++j) H[j]=0;
 for(int j=0; j<sz; ++j) ++H[t[j]-1];
 for(int j=0; j<cpus; ++j) if(cnts[j] != H[j]) exit(printf("Oops %d\n", j));}
for(int j=0; j<sz; ++j) printf("%d", t[j]); printf("\n");
for(int j=0; j<cpus; ++j) printf("%d %5d\n", j, cnts[j]);
return 0;}