麻巴有多个参数时，CUDA内核速度变慢

问题描述

我写一些测试代码并打印运行时间。代码和时间如下：

from numba import cuda
import time
import numpy as np

@cuda.jit()
def test1(a,b,c,d,e,f,g,h,i):
    pass

@cuda.jit()
def test2():
    pass

ca=cuda.to_device(np.full(shape=10000,dtype=np.int32,fill_value=-1))
cb=cuda.to_device(np.full(shape=10000,fill_value=-1))
cc=cuda.to_device(np.full(shape=10000,fill_value=-1))
cd=cuda.to_device(np.full(shape=10000,fill_value=-1))
ce=cuda.to_device(np.full(shape=10000,fill_value=-1))
cf=cuda.to_device(np.full(shape=10000,fill_value=-1))
cg=cuda.to_device(np.full(shape=10000,fill_value=-1))
ch=cuda.to_device(np.full(shape=10000,fill_value=-1))
ci=cuda.to_device(np.full(shape=10000,fill_value=-1))

for i in range(4):
    stime=time.time()
    test1[1,1](ca,cb,cc,cd,ce,cf,cg,ch,ci)
    cuda.synchronize()
    etime=time.time()
    print(etime-stime)
    stime = time.time()
    test2[1,1]()
    cuda.synchronize()
    etime = time.time()
    print(etime - stime)
    print()

打印时间

0.21714425086975098
0.1640312671661377

0.0014150142669677734
0.0002522468566894531

0.0009179115295410156
0.00022268295288085938

0.0008695125579833984
0.000217437744140625

这些变量已放置在GPU上。为什么将参数传递给内核时时间仍然增加

解决方法

暂无找到可以解决该程序问题的有效方法，小编努力寻找整理中！

如果你已经找到好的解决方法，欢迎将解决方案带上本链接一起发送给小编。

小编邮箱:dio#foxmail.com (将#修改为@）