how to find the min value

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi,
In my program, each thread generate many values. I want to find the min value in all the values. Firstly, I want to save all the value into global memory and then compare them, but there are so many values that the global memory cannot contain them. I want to use atomic function atomicMin(), but it only support capability 1.0 and float type. Can you give me some advice?

Thanks!

2

Your best bet is to compute the min per block then have thread #0 of each block use the atomicMin operator to write its block minimum to global memory

The code would go something like this:

compute

store results to a shared array

__syncthreads();

for(i = blockDim.x / 2; i; i >>= 1)

{

if(threadIdx.x < i && array[threadIdx.x] < array[threadIdx.x + i])

      array[threadIdx.x] = array[threadIdx.x + i];

__syncthreads();

}

if(threadIdx.x == 0)

{

 atomicMin(&global_min, array[0]);

}

If you’re not using multiple blocks, just stick to the shared memory.

-Patrick

3

And if you don’t have atomic support (device 1.0) you could write the per-block values out to global memory, then do a short CPU loop over them to find the minimum. Or, make a second kernel to do it on the device in a second pass.