Channels ▼
RSS

Tools

CUDA, Supercomputing for the Masses: Part 11


Texture Memory in CUDA

Graphics processors provide texture memory to accelerate frequently performed operations such as mapping, or deforming, a 2D "skin" onto a 3D polygonal model. Tricks like those in Figure 2 let low-resolution game objects appear as visually richer objects with greater complexity and detail. To remain competitive, graphics card manufacturers added the extra hardware (e.g., texture units) to make these graphics operations occur very fast. CUDA provides mechanisms so C-programmers can exploit some of the added capabilities of the texture unit hardware on CUDA-enabled devices. (For more information on the use of textures in graphics, please see the free online GPU Gems books. A good place to start is Chapter 37 of GPU Gems 2.)

Figure 2: Texture memory in action.

The easiest way to think of texture memory is as an alternative memory access path that the CUDA programmer can bind to regions of the GPU device memory (e.g., global memory). Texture references can be bound to the same, overlapping or different textures in memory. Each on-chip texture unit has some internal memory that buffers data from global memory. For this reason, texture memory can be used as a relaxed mechanism for the thread processors to access global memory because the coalescing requirements discussed in previous articles do not apply to texture memory accesses. This is particularly useful when targeting previous generation CUDA-capable GPUs but may not matter as much with the relaxed coalescing requirements in newer hardware.

Since optimized data access is very important to GPU performance, the use of texture memory can (in the right circumstances) provide a large performance increase. The best performance will be achieved when the threads of a warp read locations that are close together from a spatial locality perspective. CUDA provides 1D, 2D and 3D fetch capabilities using texture memory. Since a texture performs a read operation from global memory only when there is a cache miss it is possible to exceed the maximum theoretical memory bandwidth of the underlying global memory through the judicious use of the texture memory cache. This makes the texture fetch (texfetch) rate the more useful metric for analyzing kernel performance when using textures. For example, Patrick LeGresley notes in slide 32 of High Performance Computing with CUDA that the G80 architecture can provide approximately 18 billion fetches per second.

Especially consider using textures when:

  • There is locality of reference so caching in texture memory can help.
  • Use of the texture cache can reduce the penalty for nearly coalesced accesses (such as a misaligned starting address) that cannot be made fully coalesced.

Each of the previous is problem dependent so some testing is required. Some authors have reported certain texture memory access patterns are many times faster than others (For one example, see Kipton Barro's presentation CUDA Tricks and Computational Physics).

Other performance benefits of texture memory (over global and constant memory) include:

  • Packed data may be broadcast to separate variables in a single operation.
  • 8-bit and 16-bit integer input data may be optionally converted to 32-bit floating-point values in the range [0.0, 1.0] or [-1.0, 1.0] by the texture unit hardware.
  • Linear, bilinear, and tri-linear interpolation using dedicated hardware separate from the thread processors.

Texture memory provides many more capabilities beyond those mentioned in this first introduction. A follow-on article will discuss texture memory further. Until the next article, please look to the NVIDIA_CUDA_SDK projects folder for examples. The Internet also contains many more useful examples as well that you can download and try. Following are two possibilities:

An excellent resource to learn more about the texture cache and other methods for data reuse on GPUs is Mark Silberstein's paper Efficient Computation of Sum-products on GPUs Through Software Managed Cache.

Summary

All of the above (and the previous articles in this series) indicate that CUDA developers must have a firm grasp of CUDA memory spaces. Specifically, be aware that local and global memory are not cached and their access latencies are high. Access latencies as reported by David Kirk and Wen-mei Hwu in The CUDA Memory Model with additional API and Tools info are:

  • Register - dedicated HW - single cycle
  • Shared Memory - dedicated HW - single cycle
  • Local Memory - DRAM, no cache - *slow*
  • Global Memory - DRAM, no cache - *slow*
  • Constant Memory - DRAM, cached, one to hundreds of cycles depending on cache locality
  • Texture Memory - DRAM, cached, hundreds of cycles
  • Instruction Memory (invisible) - DRAM, cached

Consider a typical CUDA template as:

  • Split a task into subtasks
  • Divide input data into chunks that fit into registers and shared memory
  • Load a data chunk from global memory into registers and shared memory
  • Each data chunk is processed by a thread block
  • Copy results back to global memory

However texture and constant memory are cached and can significantly increase application performance -- due to their access characteristics -- depending on the application access patterns:

  • R/O no structure → constant memory
  • R/O array structured → texture memory (CUDA 2.1 and previous)
  • R/W shared within block → shared memory
  • R/W registers spill to local memory and may provide a surprise slowdown
  • R/W inputs/results → global memory

For More Information


Rob Farber is a senior scientist at Pacific Northwest National Laboratory. He has worked in massively parallel computing at several national laboratories and as co-founder of several startups. He can be reached at rmfarber@gmail.com.


Related Reading


More Insights






Currently we allow the following HTML tags in comments:

Single tags

These tags can be used alone and don't need an ending tag.

<br> Defines a single line break

<hr> Defines a horizontal line

Matching tags

These require an ending tag - e.g. <i>italic text</i>

<a> Defines an anchor

<b> Defines bold text

<big> Defines big text

<blockquote> Defines a long quotation

<caption> Defines a table caption

<cite> Defines a citation

<code> Defines computer code text

<em> Defines emphasized text

<fieldset> Defines a border around elements in a form

<h1> This is heading 1

<h2> This is heading 2

<h3> This is heading 3

<h4> This is heading 4

<h5> This is heading 5

<h6> This is heading 6

<i> Defines italic text

<p> Defines a paragraph

<pre> Defines preformatted text

<q> Defines a short quotation

<samp> Defines sample computer code text

<small> Defines small text

<span> Defines a section in a document

<s> Defines strikethrough text

<strike> Defines strikethrough text

<strong> Defines strong text

<sub> Defines subscripted text

<sup> Defines superscripted text

<u> Defines underlined text

Dr. Dobb's encourages readers to engage in spirited, healthy debate, including taking us to task. However, Dr. Dobb's moderates all comments posted to our site, and reserves the right to modify or remove any content that it determines to be derogatory, offensive, inflammatory, vulgar, irrelevant/off-topic, racist or obvious marketing or spam. Dr. Dobb's further reserves the right to disable the profile of any commenter participating in said activities.

 
Disqus Tips To upload an avatar photo, first complete your Disqus profile. | View the list of supported HTML tags you can use to style comments. | Please read our commenting policy.
 

Video