Does enqueue_nd_range_kernel() runs asynchronously? (double buffering)

[Nyanraltotlapun]

Hi. I am trying to do some batched data processing (success), and now looking into implementing double buffer pattern in order to separate kernel execution and data copy operations.
But cannot find info about kernel invocation in pyopencl regardless its synchronization properties.

Is what I am doing is sane?

My simplified code:

data_size = 1000000
batch_size = 1000
for b_n, s_i in enumerate(range(0, data_size, batch_size)):
    r_end = s_i + batch_size

    if b_n % 2 == 0:
        c_k_ev = cl.enqueue_copy(cl_queue, k_buff1, k_array_n[s_i:r_end], wait_for=[knl_ev], is_blocking=False)
        knl_ev = mykernel(cl_queue, global_size, local_size,k_buff1, res_x1,wait_for=[c_k_ev, c_x_ev1])
        c_x_ev1 = cl.enqueue_copy(cl_queue, x_array_n[s_i:r_end], res_x1, wait_for=[knl_ev], is_blocking=False)

    else:
        c_k_ev = cl.enqueue_copy(cl_queue, k_buff2, k_array_n[s_i:r_end], wait_for=[knl_ev], is_blocking=False)
        nl_ev = mykernel(cl_queue, global_size, local_size,k_buff2, res_x2,wait_for=[c_k_ev, c_x_ev2])
        c_x_ev2 = cl.enqueue_copy(cl_queue, x_array_n[s_i:r_end], res_x2, wait_for=[knl_ev], is_blocking=False)

I don't see any improvement from single buffer code (not necessary they are exist)

[inducer]

Does enqueue_nd_range_kernel() run[s] asynchronously?

Please look at the OpenCL spec to understand the expected synchronization behavior. Short answer: yes, it doesn't wait for kernel completion.