Parameter Passing to Device

Accelerated Computing CUDA CUDA Programming and Performance
1

This is my first post. I have a task cut out.

I need to port a function to CUDA. Function reads parametes from file and computes certain other parameters. There are about 13 parameters.

Initially when I wrote the function in a single kernel, all these parameters were hardcoded. I achieved performance and my output was also correct.

Now when I pass these 13 parameters from host to device (kernel function), my performance degrades substanially and there was no resultant output.
If I execute in EMUDebug mode, I am getting the desired output.

What else can be problem. Please advice, this is my first post.

2

Therez some limitation on number of args that you can pass to the CUDA kernel.

Structurize them , copy out to a GPU mem location and pass that GPU location as a pointer to your kernel.

3

Infact I structurized them and copied to GPU mem location.

It was a catastrope. It is taking hell lot of a time, more than what it takes for the CPU equivalent.

Any other suggestions.

4

Can anyone provide me a solution.
A solution is badly needed.

Thanks in advance

Mathew Potter

5

Hi,
Can you post the code where

  1. the kernel is called with multiple arguments - the device kernel function definition (just the top line) and the host variable initialisation, and …
  2. the kernel is called in the struct manner - again, the function definition and host variable init / malloc / memcpy into device struct?
6

In your kernel, are you copying the whole structure from GPU mem to shared mem cache? OR are you using a local array?

It would be dead slow if you had used a local array instead of shared mem array.

It would be slow if all threads in the block copy the structure from global mem to shared mem – something like if you had used “sharedMemStructure = *gMem”. Then all threads will load it from gmem which could make it slow and redundant if you are using more threads per block and more blocks in your kernel.

HTH

7

OK I could fix it. It was internal error due to improper indexing.
Now I am passing arguments through an array.

Thanks a lot for all suggestions.