技术背景
在CUDA编程中有可能会遇到一些相对比较隐蔽的报错,但是直接编译运行cu文件是不显现的。那么可以通过添加一个用于检查的宏,来监测CUDA程序运行过程中可能出现的报错。
error.cuh
我们在CUDA头文件中实现这个宏:
cuh
#pragma once
#include <stdio.h>
#define CHECK(call) do{const cudaError_t error_code = call; if (error_code != cudaSuccess){printf("CUDA Error:\n"); printf(" File: %s\n", __FILE__); printf(" Line: %d\n", __LINE__); printf(" Error code: %d\n", error_code); printf(" Error text: %s\n", cudaGetErrorString(error_code)); exit(1);}} while (0)
然后在调用CUDA相关函数或者核函数的时候,就可以使用CHECK操作来监测其中有无相关异常。
调用测试
先用一个简单的测试案例,就是显存分配的场景,如果是一个正常的显存分配:
cu
// nvcc ./test_error.cu -Xcompiler -fPIC -o ./test_error && ./test_error
#include "error.cuh"
#include <stdio.h>
int main(void){
const int N = 100000000;
const int M = sizeof(double) * N;
double *d_x;
CHECK(cudaMalloc((void **)&d_x, M));
CHECK(cudaFree(d_x));
printf("Success!\n");
}
运行结果是没有报错的:
bash
Success!
但是如果我们调大N的值,使其超出显存大小:
cu
// nvcc ./test_error.cu -Xcompiler -fPIC -o ./test_error && ./test_error
#include "error.cuh"
#include <stdio.h>
int main(void){
const int N = 1000000000;
const int M = sizeof(double) * N;
double *d_x;
CHECK(cudaMalloc((void **)&d_x, M));
CHECK(cudaFree(d_x));
printf("Success!\n");
}
再次运行,就会报OOM错误:
bash
./test_error.cu(7): warning #69-D: integer conversion resulted in truncation
const int M = sizeof(double) * N;
^
Remark: The warnings can be suppressed with "-diag-suppress <warning-number>"
./test_error.cu(9): warning #68-D: integer conversion resulted in a change of sign
do{const cudaError_t error_code = cudaMalloc((void **)&d_x, M); if (error_code != cudaSuccess){printf("CUDA Error:\n"); printf(" File: %s\n", "./test_error.cu"); printf(" Line: %d\n", 9); printf(" Error code: %d\n", error_code); printf(" Error text: %s\n", cudaGetErrorString(error_code)); exit(1);}} while (0);
^
./test_error.cu(7): warning #69-D: integer conversion resulted in truncation
const int M = sizeof(double) * N;
^
Remark: The warnings can be suppressed with "-diag-suppress <warning-number>"
./test_error.cu(9): warning #68-D: integer conversion resulted in a change of sign
do{const cudaError_t error_code = cudaMalloc((void **)&d_x, M); if (error_code != cudaSuccess){printf("CUDA Error:\n"); printf(" File: %s\n", "./test_error.cu"); printf(" Line: %d\n", 9); printf(" Error code: %d\n", error_code); printf(" Error text: %s\n", cudaGetErrorString(error_code)); exit(1);}} while (0);
^
./test_error.cu: In function 'int main()':
./test_error.cu:7:31: warning: overflow in conversion from 'long unsigned int' to 'int' changes value from '8000000000' to '-589934592' [-Woverflow]
7 | const int M = sizeof(double) * N;
| ~~~~~~~~~~~~~~~~^~~~
CUDA Error:
File: ./test_error.cu
Line: 9
Error code: 2
Error text: out of memory
当然,中间因为整形溢出,还有一些其他的warnning信息,但是这里主要要展现的是OOM报错问题。
核函数检测
上面的异常检测针对是cudaMalloc这个CUDA操作,其实对于核函数,也是一样可以检测出其异常。我们先演示一个正常的示例:
cu
// nvcc ./test_error.cu -Xcompiler -fPIC -o ./test_error && chmod +x ./test_error && ./test_error
#include "error.cuh"
#include <math.h>
#include <stdio.h>
void __global__ add(const double *x, const double *y, double *z, const int N){
int idx = blockIdx.x * blockDim.x + threadIdx.x;
if (idx < N){
z[idx] = x[idx] + y[idx];
}
}
int main(void){
const int N = 10;
const int M = sizeof(double) * N;
const double a = 1.23;
double *h_x = (double*) malloc(M);
for (int n = 0; n < N; ++n)
{
h_x[n] = a;
}
double *d_x, *d_z;
CHECK(cudaMalloc((void **)&d_x, M));
CHECK(cudaMalloc((void **)&d_z, M));
CHECK(cudaMemcpy(d_x, h_x, M, cudaMemcpyHostToDevice));
const int block_size = 1024;
const int grid_size = (N + block_size - 1) / block_size;
add<<<grid_size, block_size>>>(d_x, d_x, d_z, N);
CHECK(cudaGetLastError());
CHECK(cudaDeviceSynchronize());
CHECK(cudaFree(d_x));
CHECK(cudaFree(d_z));
free(h_x);
printf("Success!\n");
return 0;
}
这个CUDA程序运行的是一个数组加法。运行结果:
bash
$ nvcc ./test_error.cu -Xcompiler -fPIC -o ./test_error && chmod +x ./test_error && ./test_error
Success!
调整一下block_size参数:
cu
// nvcc ./test_error.cu -Xcompiler -fPIC -o ./test_error && chmod +x ./test_error && ./test_error
#include "error.cuh"
#include <math.h>
#include <stdio.h>
void __global__ add(const double *x, const double *y, double *z, const int N){
int idx = blockIdx.x * blockDim.x + threadIdx.x;
if (idx < N){
z[idx] = x[idx] + y[idx];
}
}
int main(void){
const int N = 10;
const int M = sizeof(double) * N;
const double a = 1.23;
double *h_x = (double*) malloc(M);
for (int n = 0; n < N; ++n)
{
h_x[n] = a;
}
double *d_x, *d_z;
CHECK(cudaMalloc((void **)&d_x, M));
CHECK(cudaMalloc((void **)&d_z, M));
CHECK(cudaMemcpy(d_x, h_x, M, cudaMemcpyHostToDevice));
const int block_size = 1025;
const int grid_size = (N + block_size - 1) / block_size;
add<<<grid_size, block_size>>>(d_x, d_x, d_z, N);
CHECK(cudaGetLastError());
CHECK(cudaDeviceSynchronize());
CHECK(cudaFree(d_x));
CHECK(cudaFree(d_z));
free(h_x);
printf("Success!\n");
return 0;
}
由于Block大小在CUDA程序中最大只能是1024,因此如果超出这个数就会出现异常,但是如果没有异常检测函数的话,程序是能够正常执行下去的,这样这个异常就会一直保留在程序中。运行结果:
bash
$ nvcc ./test_error.cu -Xcompiler -fPIC -o ./test_error && chmod +x ./test_error && ./test_error
CUDA Error:
File: ./test_error.cu
Line: 29
Error code: 9
Error text: invalid configuration argument
因为加上了cudaGetLastError()
函数,并使用了异常捕获的宏,所以这里就会提示参数配置异常。
总结概要
本文主要介绍了在CUDA编程的实践中,增加一个异常捕获的宏模块,以保障CUDA项目结果的准确性。主要代码内容参考了樊哲勇所著的《CUDA编程基础与实践》,是一本很好的CUDA编程入门书籍。
版权声明
本文首发链接为:https://www.cnblogs.com/dechinphy/p/cuda_error.html
作者ID:DechinPhy
更多原著文章:https://www.cnblogs.com/dechinphy/
请博主喝咖啡:https://www.cnblogs.com/dechinphy/gallery/image/379634.html
参考内容
- 《CUDA编程基础与实践》------樊哲勇