simple CUDA invocation question (newbie error) simple CUDA invocation question

I’ve been playing with CUDA running on a RHEL box for the past day or so. After getting 2.2 installed, I tried to implement the first snippet described by the NVIDIA programming manual, a simple 1-d vector addition. To make sure the answer makes sense, I wrote the routine the “old way” and then tried the implementation below. As the output shows, the CUDA call doesn’t work (and I think I lifted it directly from the user manual.

Any suggestions on what I screwed up?

Code:

[codebox]

#include <stdio.h>

#include <assert.h>

#include <cuda.h>

// Kernel definition

global void

VecAdd (float *A, float *B, float *C)

{

int i = threadIdx.x;

C[i] = A[i] + B[i];

}

void add_vectors (int N, float A, float B, float C);

void

add_vectors (int N, float A, float B, float C)

{

int j;

for (j = 0; j < N; j++)

{

  C[j] = A[j] + B[j];

}

return;

}

int

main (void)

{

int i, N;

N = 10;

float A[N], B[N], C[N];

for (i = 0; i < N; i++)

{

  A[i] = 1.0 * i;

  B[i] = 0.01 * i;

}

add_vectors (N, A, B, C);

printf (“adding the old-fashioned way:\n”);

for (i = 0; i < N; i++)

{

  printf ("%7.2f +%7.2f =%7.2f\n", A[i], B[i], C[i]);

}

for (i = 0; i < N; i++)

{

  C[i] = acos (-1.0);

}

// Kernel invocation

VecAdd <<< 1, N >>> (A, B, C);

printf ("\nAdding with CUDA\n");

for (i = 0; i < N; i++)

{

  printf ("%7.2f +%7.2f =%7.2f\n", A[i], B[i], C[i]);

}

return 1;

}

[/codebox]

Output:

[codebox]

[nmoore@buff cuda_play]$ nvcc dot_prod.cu

[nmoore@buff cuda_play]$ ./a.out

adding the old-fashioned way:

0.00 + 0.00 = 0.00

1.00 + 0.01 = 1.01

2.00 + 0.02 = 2.02

3.00 + 0.03 = 3.03

4.00 + 0.04 = 4.04

5.00 + 0.05 = 5.05

6.00 + 0.06 = 6.06

7.00 + 0.07 = 7.07

8.00 + 0.08 = 8.08

9.00 + 0.09 = 9.09

Adding with CUDA

0.00 + 0.00 = 3.14

1.00 + 0.01 = 3.14

2.00 + 0.02 = 3.14

3.00 + 0.03 = 3.14

4.00 + 0.04 = 3.14

5.00 + 0.05 = 3.14

6.00 + 0.06 = 3.14

7.00 + 0.07 = 3.14

8.00 + 0.08 = 3.14

9.00 + 0.09 = 3.14

[/codebox]

CUDA kernels can only operate on GPU memory, so you’ll need to allocate memory on the GPU (cudaMalloc) and copy your buffers (cudaMemcpy) before the kernel does anything useful. as it stands now, the kernel is just generated an unspecified launch failure, which on the CPU side would mean a segfault.

GPU has its own memory. You have to allocate memory on the device (GPU) and copy data from host (RAM) into there, before performing any computations.
At the end, you have to copy the result back from device to host, and only later print it out.
This is where cudaMalloc and cudaMemcpy come into play - check those out!

With SDK you should have some examples there already (some fairly easy to understand). Check those out too.