device_enqueue_fix.cl

void krnl1_fn(__global int* A, 
              local int* l)
{
    if (get_local_id(0)==0) {
        *l = A[get_global_id(0)];
    }

    barrier(CLK_LOCAL_MEM_FENCE);

    A[get_global_id(0)] = *l;
}

void kernel krnl1(global int* A)
{    
    // Create a local variable and use it
    local int l;

    krnl1_fn(A, &l);
}
                     

void kernel call(__global int* A)
{
    void (^block)(void) = 
    ^{
        local int l; 
        krnl1_fn(A, &l);
    };
    
    clk_event_t done;
    enqueue_kernel(get_default_queue(),
                   CLK_ENQUEUE_FLAGS_NO_WAIT,
                   ndrange_1D(1024,64), 0, NULL, &done,
                   block);
}