pytorch_cuda_pow_test

gistfile1.txt

import torch
import torch.nn as nn

import time

torch.backends.cudnn.benachmark = True

a = torch.randn(1024, 1024, 10).cuda()
b = torch.randn(1024, 1024, 10).cuda()

nb_iters = 1000

for _ in range(10):
    c = (a**2 + b**2) **0.5

torch.cuda.synchronize()
t0 = time.time()
for _ in range(nb_iters):
    c = (a**2 + b**2)** 0.5
torch.cuda.synchronize()
t1 = time.time()
print('slow took {}s'.format((t1 - t0)/nb_iters))


for _ in range(10):
    c = (a*a + b*b) **0.5

torch.cuda.synchronize()
t0 = time.time()
for _ in range(nb_iters):
    c = (a*a + b*b)**0.5
torch.cuda.synchronize()
t1 = time.time()
print('fast took {}s'.format((t1 - t0)/nb_iters))