CUDA error when handling large input - c++

Question

So I have a rather strange error that is happening. I have a kernel that is supposed to alter the value of every element in an array. As of right now I only test with launching one thread.
__global__ void kernel(int* data) {
for (int var = 0; var < SIZE; ++var) {
data[var] = data[var] + 1;
}
}
Here is the whole code:
#include "stdint.h"
#include "stdio.h"
#include "kernelLauncher.cuh"
#include <cuda_runtime.h>
#define SIZE 10485760
typedef uint64_t POLY_64;
typedef unsigned char BYTE;
__global__ void kernel(int* data) {
for (int var = 0; var < SIZE; ++var) {
data[var] = data[var] + 1;
}
}
int main() {
int* data = (int*) malloc(sizeof(int) * SIZE);
int* data_d;
for (int var = 0; var < SIZE; ++var) {
data[var] = 1;
}
//allocate device memory for the fingerprinting data
cudaMalloc((void**) &data_d, sizeof(int) * SIZE);
//copy the data to device
CUDA_CHECK_RETURN(
cudaMemcpy(data_d, data, sizeof(int) * SIZE, cudaMemcpyHostToDevice));
kernel<<<1, 1>>>(data_d);
cudaThreadSynchronize();
CUDA_CHECK_RETURN(cudaMemcpy(data, data_d, sizeof(int) * SIZE, cudaMemcpyDeviceToHost));
//try to print the result
for (int var = 0; var < SIZE; ++var) {
printf("%d\n", data[var]);
}
CUDA_CHECK_RETURN(cudaFree(data_d));
return 0;
}
When my SIZE is defined to 1048576, I get my data back just fine. Unfortunatelly when I define it as 10485760 (10 times more). I get:
Error unspecified launch failure at line 40 in file ../src/runTest.cu
Can somebody point me in the right direction. Why is this problem happening ? Thank you in advance
EDIT: So yes.. it is the size definition. I changed my code now so there are no discrepancies between the hard coded loop value in the kernel and the defined constant. However, if I have 10485760 instead of 1048576 it simply does not work.. Why is that. This is not too much allocation at one go.. My card is a Quadro FX 770m with compute capability 1.1

Answer 1

So.. here is what actually seemed to be happening. As some of you suggested, the kernel was indeed taking too long and timing out (although I read from various soruces that this does not happen on Linux systems) So separating the work like this in fact solves the issue and avoids the watchdog killing the kernel :
kernel<<<1, 1>>>(data_d, 0, 1048576);
cudaDeviceSynchronize();
kernel<<<1, 1>>>(data_d, 1048576, 2097152);
cudaDeviceSynchronize();
kernel<<<1, 1>>>(data_d, 2097152, 3145728);
cudaDeviceSynchronize();
kernel<<<1, 1>>>(data_d, 3145728, 4194304);
cudaDeviceSynchronize();
kernel<<<1, 1>>>(data_d, 4194304, 5242880);
cudaDeviceSynchronize();
kernel<<<1, 1>>>(data_d, 5242880, 6291456);
cudaDeviceSynchronize();
kernel<<<1, 1>>>(data_d, 6291456, 7340032);
cudaDeviceSynchronize();
kernel<<<1, 1>>>(data_d, 7340032, 8388608);
cudaDeviceSynchronize();
Now I wonder, what is the way to avoid hitting this threshold. I tried adding
Section "Device"
Identifier "Device0"
Driver "nvidia"
VendorName "NVIDIA Corporation"
Option "Interactive" "0" #<<--- added to avoid kernel time-out
EndSection
into the device section in my Xorg.conf, but this did not really help.