为什么在Python中同样的矩阵乘法，Pytorch和Cupy比Numpy慢？
import cupy as cp
import numpy as np
x_cpu = np.random.randn(100, 100).astype(np.float32)
x_gpu = cp.asarray(x_cpu)
%%timeit
x_cpu*x_cpu
2.41 µs ± 19.7 ns per loop (mean ± std. dev. of 7 runs, 100000 loops each)
%%timeit
x_gpu*x_gpu
14.3 µs ± 53.4 ns per loop (mean ± std. dev. of 7 runs, 100000 loops each)
import torch
x_tensor = torch.from_numpy(x_cpu)
%%timeit
x_tensor*x_tensor
3.01 µs ± 33.5 ns per loop (mean ± std. dev. of 7 runs, 100000 loops each)
x_gpu_tensor = x_tensor.cuda()
%%timeit
x_gpu_tensor*x_gpu_tensor
7.85 µs ± 13.9 ns per loop (mean ± std. dev. of 7 runs, 100000 loops each)
为什么在Python中同样的矩阵乘法，Pytorch和Cupy比Numpy慢？
11 回复
给个数量级参照：
def add(a, b):
return a+b
def add2(a, b):
return add(a, b)
x = 100.0
%%timeit
add(x, x)
180 ns ± 6.67 ns per loop (mean ± std. dev. of 7 runs, 10000000 loops each)
%%timeit
add2(x, x)
281 ns ± 3.59 ns per loop (mean ± std. dev. of 7 runs, 1000000 loops each)
函数调用的开销都只比你低一个数量级了
