CUDA Timing Woes

Sep 25, 2010 · 3 minute read · Comments
CCUDA

I have come to realize that programming using Nvidia’s CUDA can be a major PITA pit when a problem crops up. Generally there are few resources besides the CUDA Programming Guide that are helpful. Most of the answers that one would find in their searches are on the forums at Nvidia, and even then many searches turn up unanswered questions.

My particular problem was in the realm of timing the execution of a kernel on the device. In the example code that was provided to me in my class, they used several undocumented functions from the cutil.h (or, in my case, the cutil_inline.h) header. These functions included:

• cutCreateTimer
• cutStartTimer
• cutStopTimer
• cutGetTimerValue
• cutDeleteTimer

All of these functions are not officially supported by Nvidia. The exact error that I was getting was this:

ld: cannot find -lcutil

Basically, the linker could not find the static cutil library to link against. I used the command line

nvcc example.cu -Iinclude -lcutil -L/opt/cuda-sdk/lib

where the path /opt/cuda-sdk/lib contained the file libcutil.a, which is the shared library version of cutil. I searched high and low and couldn’t find a libcutil.so anywhere inside the SDK file structure. As it turns out, it doesn’t exist, at least not in the SDK version that I am using.

The way you are supposed to time something is by using events. I will get to events in a minute, but first I want to describe the journey I took to get there.

Google, thou hast failed me

I spent a total of probably 4 hours of Google searching, interspersed between other classes and homework, which got me basically nowhere. I found various sources all telling me something different. I tried many of these different approaches, including:

• looking in cutil.h for clues on what the function declarations looked like to see if maybe I could add a function prototype to help out the linker (wrong direction… I think. No effect anyway)
• adding the phrase ‘extern “C”’ before the prototype declaration (no effect)
• decompiling the shared library and recompiling it into a static library (which didn’t work anyway)

The above are just a few of the many different things I tried while searching. Most of the suggestions I tried were from Nvidia forums. In the forums also were about 20 threads that I came across that said to just reference the library using -lcutil and -L/path/to/library, which didn’t work in the first place.

Nvidia to the rescue

In the end I gave up and started looking through the Nvidia CUDA Programming Guide(warning: large PDF. source) page by page. I came across an example in the middle (Page 39, Section 3.2.7.6) that focuses on events and how they can be used to time execution. Here is the basic structure for timing something in your program (in C):

cudaEvent_t start, stop;
cudaEventCreate(&start);
cudaEventCreate(&stop);
cudaEventRecord(start, 0);

// Whatever needs to be timed goes here

cudaEventRecord(stop, 0);
cudaEventSynchronize(stop);
float elapsedTime;
cudaEventElapsedTime(&elapsedTime, start, stop);

So, there you have it

If you are trying to time something using any of the cutXxXxTimer functions, stop using those and start using events. The functions are officially supported, aren’t any harder to use, and, best of all, they actually link correctly when you go to compile your program.