High compilation time

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi,

I was just wondering about my compilation time, which increased a lot with the complexity of my program.

There is one kernel which needs a lot of time to compile, its resources are:
Used 39 registers, 5768+5756 bytes smem, 128 bytes cmem[1]

This kernel has several inlined functions. And obviously, so far, it needs a lot of shared memory.

Do anyone of you have similiar experience according to the compilation time?

P.D. I use Netbeans as IDE, which I think shouldn’t affect the compilation time. I use CUDA 2.0, NVCC version V0.2.1221

2

In my experience, compilation time scales with the amount of code that is generated (most likely because of the time the optimizer spends optimizing that code). I had one kernel in the past with a loop unrolled 27 times that called a big inlined device function each time. That one too about ~30 seconds to compile.

3

The compilation of my program takes easily about 10min. But just in device release mode, in emulation mode it is not even 20 seconds.

4

I compile really big kernels, and compilation time has been all over the place (including 10 minutes). It really gets out of hand when you set a limit on the register usage and the compiler tries to match it. Try setting --maxrregcount=128 to give the compiler room to breath.

5

Using the maxrregcount paramenter reduces the compilation time a bit, however not significantly. But as you had similiar experencies, I don’t worry too much about. Then the only thing I can do, is having a coffee while compiling… :)