Function reinit_threads

numexpr3/module.cpp:164–206

Initialize threads and allocate space for arrays in global_state

        
// Initialize threads and allocate space for arrays in global_state
int reinit_threads(int n_thread_old) {
    int tid, rc;

    gs.count_threads = 0;      // Reset threads counter
    gs.barrier_passed = BARRIER_HALT;

    // Allocate memory
    gs.threads = (pthread_t*)realloc( gs.threads, gs.n_thread * sizeof(pthread_t) );
    gs.tids = (int*)realloc( gs.tids, gs.n_thread * sizeof(int*) );

    gs.params = (NumExprObject*)realloc( gs.params, gs.n_thread  * sizeof(NumExprObject) );
    // params[R].program is shared.

    // Allocate an arena for the registers
    // Since NPY_MAXARGS=32, we can just allocate a maximum space and only 
    // consume a kB or so of RAM.
    gs.registerArena = (char *)realloc( gs.registerArena, gs.n_thread * NPY_MAXARGS * sizeof(NumExprReg) );
    for( int I=0; I < gs.n_thread; I++ ) {
        gs.params[I].registers = (NumExprReg*)(gs.registerArena + I*NPY_MAXARGS*sizeof(NumExprReg) );
    }

    gs.stridesArray = (npy_intp*)realloc( gs.stridesArray, gs.n_thread * sizeof(npy_intp) );
    gs.iter = (NpyIter**)realloc( gs.iter, gs.n_thread  * sizeof(NpyIter*) );
    gs.reduce_iter = (NpyIter**)realloc( gs.reduce_iter, gs.n_thread * sizeof(NpyIter*) );

    // Finally, create the threads
    for (tid = 0; tid < gs.n_thread; tid++) {
        gs.tids[tid] = tid;
        rc = pthread_create(&gs.threads[tid], NULL, th_worker, (void *)&gs.tids[tid]);

        if (rc) {
            fprintf(stderr,
                    "ERROR; return code from pthread_create() is %d\n", rc);
            fprintf(stderr, "\tError detail: %s\n", strerror(rc));
            exit(-1);
        }
    }

    gs.init_threads_done = 1;         // Initialization done!
    gs.pid = (int)getpid();           // save the PID for this process

    return(0);
}
        

int numexpr_set_nthreads(int n_thread_new) {