CUDA - How to swap device array pointers? - cuda

Question

I wrote a function swap to conveniently swap device array pointers, but it is not working, I assume I am swapping local array pointers in the swap function and not the ones I am passing to it.
__global__ void device_add_one(float *A, float *B)
{
for (int index = blockIdx.x * blockDim.x + threadIdx.x;
index < N;
index += blockDim.x * gridDim.x)
{
// just for the example
B[index] = A[index] + 1;
{
}
void swap(float *a, float *b)
{
float *temp = a;
a = b;
b = temp;
}
void loop(float *host_array, int size, int loops)
{
cudaMalloc(&A, (size * sizeof(float));
cudaMalloc(&B, (size * sizeof(float));
cudaMemcpy(A, host_array, (size * sizeof(float), cudaMemcpyHostToDevice);
for (int i = 0; i < loops; i++) {
device_add_one<<< 1, 254 >>>(A, B);
// swap pointers like this does not work
swap(A, B);
/* This works:
float *temp = a;
a = b;
b = temp;
*/
}
cudaMemcpy(host_array, A, (size * sizeof(float), cudaMemcpyDeviceToHost);
}

Answer 1

Your function call method of swapping pointers does not work because you are using pass-by-value. This is an ordinary C/C++ programming concept, not unique to CUDA.
When you pass variables (including pointers, in this case) to a function by value:
void swap(float *a, float *b)
the C pass-by-value mechanism creates a local copy of the function arguments, for use within the function body. Changes to those arguments do not show up in the calling context. To work around this, a simple approach would be pass-by-reference (in C++):
void swap(float* &a, float* &b)
Here is a worked example:
$ cat t393.cu
#include <stdio.h>
const int N = 1000;
float *A, *B;
__global__ void device_add_one(float *A, float *B)
{
for (int index = blockIdx.x * blockDim.x + threadIdx.x;
index < N;
index += blockDim.x * gridDim.x)
{
// just for the example
B[index] = A[index] + 1;
}
}
void swap(float* &a, float* &b){
float *temp = a;
a = b;
b = temp;
}
void loop(float *host_array, int size, int loops)
{
cudaMalloc(&A, size * sizeof(float));
cudaMalloc(&B, size * sizeof(float));
cudaMemcpy(A, host_array, (size * sizeof(float)), cudaMemcpyHostToDevice);
for (int i = 0; i < loops; i++) {
device_add_one<<< 1, 254 >>>(A, B);
// swap pointers
swap(A, B);
//float *temp = A;
//A = B;
//B = temp;
}
cudaMemcpy(host_array, A, (size * sizeof(float)), cudaMemcpyDeviceToHost);
}
int main(){
float *data = (float *)malloc(N*sizeof(float));
for (int i = 0; i<N; i++) data[i] = i & 3; // fill with 0 1 2 3 0 1 2 3...
loop(data, N, 100);
for (int i = 0; i<20; i++) printf("%f ", data[i]);
printf("\n");
return 0;
}
$ nvcc -arch=sm_61 -o t393 t393.cu
$ cuda-memcheck ./t393
========= CUDA-MEMCHECK
100.000000 101.000000 102.000000 103.000000 100.000000 101.000000 102.000000 103.000000 100.000000 101.000000 102.000000 103.000000 100.000000 101.000000 102.000000 103.000000 100.000000 101.000000 102.000000 103.000000
========= ERROR SUMMARY: 0 errors
$