All, I have a 3D array in CPU declared like ARRAY[128][128][128]. Since the accesses can benefit from 3D Spacial Locality, I decided to move it to a 3D texture… So, I allocated a 3D array using cudaMalloc3DArray() API, did a cudaMemcpy3D() (which is not so straightforward) and acessed it from kern…

Oh… Is it? Thanks! Documentation always helps… Avidday, Have a good day! Thanks,

I have attached a small CU file in my first post that reproduces the issue. FYI. @NV, Kindly throw some light on this and probably update the 3D texture documentation a bit - so that it covers important cases like this.

IMO, this is normal behaviour which also applies to 2D textures. It just depends on how you store the texture data. e.g. it is specified somewhere in the documentation that the data for a 3D texture should be a concatenation of the various slices, where each slice is a concatenation of rows an eac…

@Nico , Thanks! The Progrmaming guide explains how to access 3D arrays in the section under “device memory - section 3.2.1”. There in, the Z,Y,X relation shows up correctly… I think it is left to the reader to assume that “cudaMalloc3Darray()” works similarly and texture acceses also need to be done…

While the documentation probably could be clearer, the real problem is your lack of understanding of concepts like row-major ordering. The CPU and the GPU are using the same ordering, which is the one you decide on. The texture function parameter names imply that they expect row-major ordering in…

Very difficult concepts… Isnt it? Yeah they use the same ordering… CPU calls it as X,Y,Z and GPU prefers to call it as Z,Y,X. This is just how they prefer to call it. If you read my previous post, I have mentioned the place in the programmers guide where this is quite clear (Section 3.2.1). So,…

This seems to be the source of your confusion. The CPU and the GPU “prefer to call it” the same thing. Consider this array of data: 0 1 2 3 The element at (x = 0, y = 1) is 2. However you arrange these four numbers in memory, the conceptual layout should still be the same: a 2 x 2 matrix of…

Stickguy, Thanks for your time and a good example. Let me take your example. 0 1 2 3 You prefer to call [0,1] as 2 because of “graphics” knowledge… i.e. If you look @ your screen, the top-left most is 0,0 and as you descend down “Y” increases… This is where the confusion starts… This ordering…

You’re absolutely right. It’s mostly due to my OpenGL experience that I don’t have a problem understanding this coordinate mapping and I can imagine that it can be difficult to interpret without graphics knowledge. The “upper left corner” you’re referring to is mentioned in the CUDA reference manua…

3D Textures -- xyz@CPU becomes zyx@GPU Is this the normal behaviour?

Accelerated Computing CUDA CUDA Programming and Performance

avidday December 29, 2009, 9:21am 2

Probably row major order versus column major order. I have only ever used 3D textures with a code with column major ordered storage, and it just worked, but I have no proof that it is really ordering that is in play here. The 3D texture documentation is rather, err scant…

Topic		Replies	Views
1D access vs 3D access CUDA Programming and Performance	5	2784	March 23, 2009
tex3D axis order problem CUDA Programming and Performance	0	2385	February 11, 2011
Cuda array memory layout question CUDA Programming and Performance cuda , kernel	4	239	June 4, 2024
cudaMemcpy3D for allocating arrays for textures only? CUDA Programming and Performance	3	1213	March 2, 2011
c++array, cudaarray, texture CUDA Programming and Performance	1	4558	October 18, 2008
3D texture CUDA Programming and Performance	0	999	October 17, 2008
Fast texture fetching An algorithm for fast texture fetching CUDA Programming and Performance	12	15180	December 19, 2013
3D Texturing CUDA Programming and Performance	7	6252	August 8, 2008
Writing in to cudaArray CUDA Programming and Performance	7	5779	July 2, 2010
3D texture problem CUDA Programming and Performance	0	2587	May 13, 2011

3D Textures -- xyz@CPU becomes zyx@GPU Is this the normal behaviour?

Related topics