CUDA texture memory size

Question

When I bind an array to a texture in CUDA,

• Is copying an array into texture space? or,
• is an array reference like a texture?

If the answer is 1., then I can bind the texture and security data from the texture memory space when I write the result to an array that is allocated in global memory.

If the answer is 2., then is texture memory a global memory space in which data is cached and spatially fetched?

I would like to know about this topic, since I saw some question related to this topic, and I do not know what is clear now.

Thanks in advance.

Answer 1

The answer is the second option, but from there it gets a little more complicated. There is not such a thing as “texture memory”, but only global memory, which is accessed through dedicated equipment, which includes cache reading on the GPU (6-8 kbytes per MP depending on the card, see table F-2 in Appendix F of the Cuda Programming Guide) and a series of hardware accelerated filtering / interpolation actions. There are two ways to use texture equipment in CUDA:

• Associate linear memory with a texture and read it in the kernel using the 1D fetch API. In this case, the texture hardware really just acts as a pass-through cache, and (IIRC) there are no filtering actions available.
• Create a CUDA array, copy the contents of the linear memory into this array and bind it to the texture. The resulting CUDA array contains a spatially ordered version of a linear source stored in global memory in the form of an (undocumented) gap filling curve . Texture hardware provides cached access to this array, includes simultaneous memory reading with hardware accelerated filtering.

You can find a review of the GT200 architecture written by David Kanter to read to better understand how the real architecture implements the memory hierarchy of APIs.