Tips of Day1 practice in 2022cuda summer training camp

Previous review ：

2022CUDA Summer training camp Day1 practice  https://bbs.huaweicloud.com/blogs/364478

TIPS1： solve jupyter lab in sudo When you need a password  

Jupyter Lab It's a good thing . It makes programming visible .

However , If you carry out such instructions ：

!sudo /usr/local/cuda/bin/nvprof --print-api-trace ./hello_cuda

That's a little bad ：

Because it will ask you for the password , But it doesn't give you a place to enter your password .

So you have to wait for the asterisk on the left to light up for a while ... Not like others cell After that, it becomes a number .

However , There is no solution . You can try this ：

!echo nano|sudo -S /usr/local/cuda/bin/nvprof --print-api-trace ./hello_cuda

Take the password as the input behind the pipeline .

You'll find that , Unexpectedly, he can run ！

TIPS2： A mistake deepened the right CUDA Understanding of asynchronous execution  

Zhang Xiaobai Jupyter Lab In the implementation of cuda Of helloworld（ I'm sorry , The third time helloworld 了 ） I found a strange phenomenon , occasionally , After compilation and execution, there will be output , occasionally , No output after compilation and execution . But to perform nprof There is always output when .

So Zhang Xiaobai consulted the teacher ：

The teacher pointed out that , A synchronization process is missing from my code . Less cudaDeviceSynchronize();

because CUDA The code itself is executed asynchronously . When the program instruction is issued, it is over . This is the time , without perform cudaDeviceSynchronize(), Wait for synchronization , Then it is possible that the result will come out before the program is executed , It is possible that the result will not come out until the program is executed . At that time , The caller's program has long exited . This also led to the strange phenomenon found by Zhang Xiaobai —— Sometimes there are results , Sometimes there is no result .

After synchronization, the code is as follows ：

#include <stdio.h>__global__ void hello_from_gpu(){    printf("Hello World from the GPU!\n");}int main(void){    hello_from_gpu<<<2,2>>>();    cudaDeviceSynchronize();    return 0;}

The results are as follows ：

This time , Whether it's nvcc When compiling, add -run Parameters , Or run it separately after compilation , Or in nvprof Run in , There are printed results .

So it is better to , Make a deep impression . In the future, there will be no error of forgetting to add synchronization in the code ...（ Maybe there will be ！）

（ The full text after , Thank you for reading ）