call-cuda-10000-times-with-memory-optimization.py

import numpy as np
from numba import cuda

@cuda.jit
def add_2(array):
    pos = cuda.grid(1)
    array[pos] = array[pos] + 2

# criamos um vetor ordenado até 100000, [0, 1, 2, .. 999999] dentro da GPU
an_array = cuda.to_device(np.arange(100000))

threadsperblock = 32
blockspergrid = (an_array.size + (threadsperblock - 1)) // threadsperblock

for x in range(1000):
    add_2[blockspergrid, threadsperblock](an_array)


print(an_array.copy_to_host())