I have been looking at the "convolutionSeparable" code sample provided with CUDA 7.5, for the implementation of a 2-D FIR filter for image processing.
The rationale behind the design is discussed in the link.
Has anybody else looked closely at this code and found possible improvements or have any other comments/thoughts?
http://docs.nvidia.com/cuda/samples/3_Imaging/convolutionSeparable/doc/convolutionSeparable.pdf
OK. This is what I have found so far ...
The attached design doc all makes sense (but I would disagree that "convolution measures the amount of overlap between two functions"). However, I found the sample code difficult to follow. So I wrote my own ...
I used the sample code as a performance benchmark and eventually I was able to match it. I found the profiler to be more reliable than timer calls for quantifying execution speed.
Yes, using shared memory improves performance, despite the extra code and (wasted) threads required to populate it.
My code was much more compact than the sample code. Why are loops used to populate the shared memory tile in the sample code? I just used a single thread to populate each element, before proceeding to the convolution (after syncing threads in the block).
My code runs at around 100% GPU occupancy compared with 50% for the sample code, yet the execution time is roughly the same (?).
My row and col calls have approximately equal run times, whereas the sample code run times are very different in each dimension so the total run time will depend on the image aspect ratio!
In my code, I zero out the edge effects around the perimeter of the image.
I'll post my code later, if there is any interest.
Hi Hugh,
It is great to see that you have developed a much better optimized version of the convolution code (CUDA samples).
Can you please share the code along with a small description of how it works? It will benefit us all who are looking into this particular CUDA sample.
I am working on optimizing this code and wish to look into the already optimized versions.
- Badges
- Science topic
- Similar topics
- Engineering
- Electrical Engineering
- Signal Processing
- Image Processing