I am using a constant image of resolution 1024*768 pixels *4 bytes representing Color to be overlapped with several images. The constant image is larger than the GPU constant memory 16K or 64K. If Tiling is used, this means that the constant memory will be used for write several times.
What is the feasibility of using GPU Constant Memory for large images in such scenario?
Like Tahar Maddi said, Constat memory is only used for read operations. One practical way to manage a bidimensional o tridimensional array is to use cuda surface memory, because you can access data with 2 or 3 indexes, respectively and you can do (with surface memory) read and write operations.
Hi Ahmed.
In your case use of constant memory might not be the best idea, as you usually have a thread on the GPU that does the calculations for a unique pixel of the image being processed.
Let me elaborate on the basic concept of constant memory in CUDA:
Constant memory is, apart from being read only as Tahar already mentioned, only feasible in situations where every single thread of a given thread block, executing the same line of code, accesses the exact same address in constant memory. Then, and only then, will you achieve optimum performance for constant memory accesses. If threads of the same block access different addresses in constant memory while executing the same line of code, the accesses to constant memory are serialized for every single thread resulting in extremely slow performance.
If you want to access image data in a read only fashion it’s a better idea to use CUDAs “texture memory” object/reference API. Here accesses need not be synchronised, which is much better suited for image processing where every thread works on a unique pixel of the image.
On the other hand when your data source is read as well as written then you should use CUDAs “surface memory” object/reference API as Daniel already suggested.
Hope that helps you.
Cheers, Sebastian
The usage of loop tiling is correct, but special types of CUDA memories including texture and constant must be written before kernel starts. It means a great overhead for the constant memory due to its size. So, to achieve reasonable efficiency other special types remain: texture, read-only (nearly the same) or shared. The best option is difficult to say since: texture/read-only is "only cached", and byte-oriented access to shared memory is not optimal. Conclusion: try texture or shared memory, measure, and (please) publish the results.
- Badges
- Science topic
- Similar topics
- Data Mining
- Classification