CUDA双矩阵溢出

Question

I wrote a program that double the element for a given matrix, if I change the matrix size to be 500, it will "stopped working" due to overflow, can people help me understand why? (it works fine for 100)

#include "cuda_runtime.h"
#include "device_launch_parameters.h"

#include <stdio.h>
#include <stdlib.h>
__global__ void kernel_double(int *c, int *a)
{
    int i = blockIdx.x * blockDim.x + threadIdx.x;
    c[i] = a[i] * 2;
}
int main()
{
    const int size = 100; 
    // failed when size = 500, Unhandled exception at 0x00123979 in
    // doublify.exe: 0xC00000FD: 
    // Stack overflow (parameters: 0x00000000, 0x00602000).
    int a[size][size], c[size][size];
    int sum_a = 0;
    int sum_c = 0;

    for (int i = 0; i < size; i++) {
        for (int j = 0; j < size; j++) {
            a[i][j] = rand() % 10;
            sum_a += a[i][j];
        }
    }
    printf("sum of matrix a is %d \n", sum_a);

    int *dev_a = 0;
    int *dev_c = 0;
    cudaMalloc((void**)&dev_c, size * size * sizeof(int));
    cudaMalloc((void**)&dev_a, size * size * sizeof(int));
    cudaMemcpy(dev_a, a, size * size * sizeof(int), cudaMemcpyHostToDevice);
    printf("grid size %d \n", int(size * size / 1024) + 1);
    kernel_double << <int(size * size / 1024) + 1, 1024  >> >(dev_c, dev_a);
    cudaDeviceSynchronize();
    cudaMemcpy(c, dev_c, size * size * sizeof(int), cudaMemcpyDeviceToHost);
    cudaFree(dev_c);
    cudaFree(dev_a);
    for (int i = 0; i < size; i++) {
        for (int j = 0; j < size; j++) {
            sum_c += c[i][j];
        }
    }
    printf("sum of matrix c is %d \n", sum_c);
    return 0;
} 

And here is the output when size equals to 100:

sum of matrix a is 44949
grid size 10
sum of matrix c is 89898
Press any key to continue . . .

My development environment is MSVS2015 V14, CUDA8.0 and GTX1050Ti

Answer 1

You're getting a stack overflow with a size of 500 because you declare 2 local variable arrays with 250,000 elements each. This works out to about 2MB of stack space.

You may be able to supply a linker option to increase the initial stack size, but a better solution would be dynamically allocate the space for your arrays. (You could create a class with the arrays in them, then just allocate an instance of that class.)

For example, before your main function add a new struct:

struct mats {
    int a[size][size];
    int c[size][size];
};

Then, in your main , remove the a and c arrays, and replace it with

auto ary = std::make_unique<mats>();

everywhere you reference a or c , use ary->a and ary->c instead. (The unique_ptr will automatically delete the allocated memory when ary goes out of scope.)