Study Note: Schedule Optimisation and math_intrinsic in CUDA Programming

原创于 2016-02-29 09:41:27 发布 · 695 阅读

0 ·

CC 4.0 BY-SA版权

文章标签：

#math #gpu #cuda #schedule

Program Optimisation 同时被 2 个专栏收录

9 篇文章

订阅专栏

GPU/CUDA

5 篇文章

订阅专栏

本文介绍了如何通过调整线程块大小、寄存器使用及共享内存来提高GPU占用率的方法。此外，还讨论了在牺牲一定精度的情况下使用fast_math选项或CUDA内在数学函数来进一步优化代码性能。

摘要生成于 C知道，由 DeepSeek-R1 满血版支持，前往体验 >

Let us introduce a new term first[1].

It is the ratio of active warps / maximum number(32) of warps.

It depends on three parameters:

1) threads/block (set in <<<>>>)

2) registers/thread (can see in the ptx file or use --ptxas-option=-v to see after finish compiling)

3) shared memory/block(also can see the ptx file and use--ptxas-option=-v to see after finish compiling). However, if our shared memory variable is set extern (use 'extern' to define the shared memory). We get this variable from the runtime.

We can use these charts to see how can we improve the occupancy.(By keeping other two variables the same, changing one variable.)[1]

Also, at the expense of accuracy, we can -use fast_math or replace some math function with CUDA math intrinsic function in the code. [1]

Reference:

[1] 18645 CMU How to write fast code Jike Chong and Ian Lane