serialize threads

[tadasv20]

Hi,

My questions sounds a little bit strange. Is it possible to sleep one thread till another one finishes his job. I tried doing something like this:

int current_thread = 0;

__global__ void kernel()

{

  int thread_index = threadIdx.x + blockDim.x * blockIdx.x;

 if (thread_index == 0) {

    /* do some stuff */

    

    current_thread++;

  } else {

    /* here other threads prepare some things for me */

    

    /* kill some time */

    while (current_thread != thread_index)

     ;

    

    /* do some stuff */

    current_thread++;

  }

}

It looks stange, but this while loop just freezes everything, algorithm finishes its job, but it is very slow. I want to do something like semi-parallel (if there is such word): if its not that threads time to work, it just prepares some things and waits for its turn. Is there any way? Thanks.

Tadas

[E.D_Riedijk]

you can use syncthreads to synchronize threads in a block. synchronization between blocks is not generally possible during the runtime of a kernel

[trudger]

It looks like a sequential program. Why do you need to run it in CUDA?

Hi,

My questions sounds a little bit strange. Is it possible to sleep one thread till another one finishes his job. I tried doing something like this:

int current_thread = 0;

__global__ void kernel()

{

 Â int thread_index = threadIdx.x + blockDim.x * blockIdx.x;

 if (thread_index == 0) {

 Â  Â /* do some stuff */

 Â  Â 

 Â  Â current_thread++;

 Â } else {

 Â  Â /* here other threads prepare some things for me */

 Â  Â 

 Â  Â /* kill some time */

 Â  Â while (current_thread != thread_index)

 Â  Â ;

 Â  Â 

 Â  Â /* do some stuff */

 Â  Â current_thread++;

 Â }

}

It looks stange, but this while loop just freezes everything, algorithm finishes its job, but it is very slow. I want to do something like semi-parallel (if there is such word): if its not that threads time to work, it just prepares some things and waits for its turn. Is there any way? Thanks.

Tadas

[snapback]446720[/snapback]

[alex_dubinsky]

No it won’t work. To expand on E.D. Riedijk, all the threads in a grid don’t run at once (only a few blocks at a time run on the card). You can’t wait on threads that haven’t even been loaded into the GPU yet.