Fastest way to swap floats and integers? also looking for conditional swaps

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi there,

I am currently doing some experiments with a sort algorithm and I was wondering if there is a really quick way of swapping the contents of two variables in shared memory.

// array to sort, assume pointer is initialized with a location in shared mem
float *var;

// this swap method is a classic, using a temp register
float tmp = var[0];
var[1]=var[0];
var[0]=tmp;

For integers, there are swap methods using xor that can work without a temp register, but this does not map well to floats… unless one treats the float as an integer representation. And these hacks also need 3 instructions to complete.

For applicability in sorting I need to compare the contents of both variables and swap conditionally, i.e.

// compiler uses predication hopefully instead of branch divergence
if (var[0] > var[1])
{
float tmp = var[0];
var[1]=var[0];
var[0]=tmp;
}

But now the question to CUDA experts. Are there any PTX instructions that could be used to accelerate this procedure? I would like the entire swap operation to consume as few clock cycles as possible. Are there any built-in swap or conditional swap primitives in the GPU? If not… this would go on my wish list!

Christian

2

Atomic CAS (Compare And Swap) instruction can do that look at PTX pdf. You’ll need SM11 compute capability for swa[ing in global memory and SM12 for swapping in shared mem. SM13 can swap 64bits float too.

3

I have compute capability 1.1 (G92 chip), and atomic swapping in global memory might be a bit slow… I wonder if Atomic CAS allows for any read/write coalescing when accessing global memory?

Christian