#%%
import torch
import numpy as np

r = torch.randn(10000, 1000)
i = torch.zeros_like(r)
c = torch.stack((r, i), 2)
nc = c.numpy()
g = c.cuda()
print('c.device : {}'.format(c.device))
print('g.device : {}'.format(g.device))

#%%
# timeit c.cuda()

#%%
# timeit g.cpu()

#%%
import time
N = 10

gpu_time = 0


print('N = {:d}\n'.format(N))

start_time = time.time()
for i in range(0, N) :
    g = c.cuda()
    torch.cuda.empty_cache()
elapsed_time = time.time() - start_time
gpu_time += elapsed_time
print('{:<10} : {:2.3f} ms/loop'.format('to GPU', elapsed_time/N*1000))

start_time = time.time()
for i in range(0, N) :
    a=torch.ifft(torch.fft(g, 2),2)
    torch.cuda.empty_cache()
elapsed_time = time.time() - start_time
gpu_time += elapsed_time
print('{:<10} : {:2.3f} ms/loop'.format('GPU cal.', elapsed_time/N*1000))

b=a.cpu()
start_time = time.time()
for i in range(0, N) :
    b=a.cpu()
    torch.cuda.empty_cache()
elapsed_time = time.time() - start_time
gpu_time += elapsed_time
print('{:<10} : {:2.3f} ms/loop'.format('to Mem.', elapsed_time/N*1000))
print('{:<10} : {:2.3f} ms/loop'.format('GPU tot.', gpu_time/N*1000))

print('\n')

for nthread in [1] :
    torch.set_num_threads(nthread)
    start_time = time.time()
    for i in range(0, N) :
        torch.irfft(torch.rfft(c, 2),2).numpy()
        torch.cuda.empty_cache()
    elapsed_time = time.time() - start_time
    print('{:<10} : {:d} threads : {:2.3f} ms/loop/thread'.format('CPU tot.', nthread, elapsed_time*nthread/N*1000))

#%%
