threads inside threads problem

[Ivan Mili]

So I'm working on multipling 2 matrices in C++ using threads.
This is the code:

#include<iostream>#include<cstdlib>
#include<pthread.h>
using namespace std;


const int m=5;
const int n=4;
const int s=4;
const int r=6;
const int num_thrd_row=2;
const int num_thrd_col=3;


class data
{
public:
  int* A;
  int* B;
  int* C;
  int start;
  int row;
  int column;
  data(int* a,int* b,int* c): A(a),B(b),C(c),start(0) {}
};


void* multiply_columns(void* dat)
{
  data* P=(data*)dat;
  int i=P->row;


  for(int j=P->column;j<r;j+=num_thrd_col)//***problem***
  {
    float sum=0;


    for(int k=0;k<s;k++)
      sum+=P->A[i*n+k]*P->B[k*r+j];


    P->C[i*r+j]=sum;
  }
}


void* multiply_rows(void* dat)
{
  data* P=(data*)dat;
  pthread_t tID[m*num_thrd_col];


  for(int i=P->start;i<m;i+=num_thrd_row)
  {
    P->row=i;
    P->column=0;


    for(int t=0;t<num_thrd_col;t++)
    {
      pthread_create(&tID[i*num_thrd_col+t],NULL,multiply_columns,(void*)P);
      P->column++;//***problem***
    }


    for(int t=0;t<num_thrd_col;t++)
      pthread_join(tID[i*num_thrd_col+t],NULL);
  }
}


int main()
{
  srand(time(NULL));


  int* A=new int[m*n];
  int* B=new int[s*r];


  for(int i=0;i<m*n;i++)
    A[i]=rand()%10+1;


  for(int i=0;i<s*r;i++)
    B[i]=rand()%10+1;


  if(n==s)
  {
    int* C=new int[m*r];
    pthread_t tID[num_thrd_row];


    data* P=new data(A,B,C);


    for(int i=0;i<num_thrd_row;i++)
    {
      pthread_create(&tID[i],NULL,multiply_rows,(void*)P);
      P->start++;
    }


    for(int i=0;i<num_thrd_row;i++)
      pthread_join(tID[i],NULL);


    cout << "A [" << m << "*" << n << "]=" << endl;
    for(int i=0;i<m*n;i++)
    {
      cout << A[i] << "\t";
      if((i+1)%n==0) cout << endl;
    }


    cout << endl << "B [" << s << "*" << r << "]=" << endl;
    for(int i=0;i<s*r;i++)
    {
      cout << B[i] << "\t";
      if((i+1)%r==0) cout << endl;
    }


    cout << endl << "AB [" << m << "*" << r << "]=" << endl;
    for(int i=0;i<m*r;i++)
    {
      cout << C[i] << "\t";
      if((i+1)%r==0) cout << endl;
    }


    delete[] C;
  }
  else
    cout << "Matrices cannot be multiplied!" << endl;


  delete[] A;
  delete[] B;


  return 1;
}

Problem is, "j=P->column", because j is always equal to 3.
How can that be possible? In main thread is done the same thing (with P->start, and "i=P->start" works just fine. Thanks for your help.

[grumpy]

You need to read up on threads properly, because you are completely misinterpreting how they work.

If multiple thread threads act on the same data concurrently, the result is a race condition (anything can happen, depending on which thread affects the data when, whether the operating system causes threads to preempt each other). Threads do not self-synchronise amongst themselves.

In your calls of pthread_create(), you are providing the same pointer (P) to every thread. That means all threads act on the same data concurrently.

Either your threads need to be passed different data (e.g. each thread is passed its own P, which is not shared with other threads) or they need to synchronise access to it (e.g. each thread cooperates with every other thread, and waits until it has exclusive access to the data) using synchronisation primitives (mutexes, critical sections, etc).

[Ivan Mili]

I tried this before, and it still didnt work, and I also tried mutex to try to see what is the problem, and even with mutex, j=P->column was always 3.

void* multiply_rows(void* dat)
{
  data* P=(data*)dat;
  data* P_new=new data(P->A,P->B,P->C);
  pthread_t tID[m*num_thrd_col];


  for(int i=P->start;i<m;i+=num_thrd_row)
  {
    P_new->row=i;
    P_new->column=0;


    for(int t=0;t<num_thrd_col;t++)
    {
      pthread_create(&tID[i*num_thrd_col+t],NULL,multiply_columns,(void*)P_new);
      P_new->column++;
    }


    for(int t=0;t<num_thrd_col;t++)
      pthread_join(tID[i*num_thrd_col+t],NULL);
  }
}

[grumpy]

There is no guarantee that a created thread executes (calls the supplied function) before pthread_create() returns. My previous comments still stand.

[MutantJohn]

Also, C++11 changed how we do threads and tbh, it's entirely for the better.

Instead, using std::thread and lambda functions (Lambda functions (since C++11) - cppreference.com).

A good thing to do is to create a vector like so :

std::vector<std::thread> threads;
threads.emplace_back([/* captures (what your functions need) */](parameters)->return_type
{
/* ... */
})

for (auto &t : threads)
    t.join()

I'm not sure if that's confusing or not but to me, it was always simplest to just see exactly what code each thread would be running. Note that you can capture pretty much everything into your lambdas and you can pretty much call anything with them too. I think they simplify the threading model greatly.

I also believe valgrind comes with multithreaded support now or at least it comes with a data race detector : Valgrind