benchmark_ram_to_vram.py

import time
import numpy as np
import pyopencl as cl

def main():
    bytes_size = 10*1024*1024*1024
    num_of_floats = bytes_size // 4

    devices = sum([x.get_devices(device_type=cl.device_type.GPU) for x in cl.get_platforms()], [])
    device = devices[0]
    print(device)
    ctx = cl.Context(devices=[device])
    buf = cl.Buffer(ctx, cl.mem_flags.READ_WRITE, bytes_size)

    arr = np.zeros(num_of_floats, dtype=np.float32)
    arr[:] = 0.1337

    print(f"size of arr: {arr.nbytes / 1024 / 1024 / 1024:.2f} GB")

    time.sleep(1)

    print('start copying', flush=True)
    start = time.perf_counter_ns()
    with cl.CommandQueue(ctx) as queue:
        cl.enqueue_copy(queue, buf, arr)
        cl.enqueue_barrier(queue)
    end = time.perf_counter_ns()
    print('end copying', flush=True)
    print(f"copy from numpy to pyopencl: {end - start} ns")
    b = (bytes_size / 1024 / 1024 / 1024) / ((end - start) / 1e9)
    print(f"bandwith: {b:.2f} GB/s")

main()