General-purpose computing on graphics processing units (GPGPU, also referred to as GPGP and less often GP²U) is the technique of using a GPU, which typically handles computation only for computer graphics, to perform computation in applications traditionally handled by the CPU. It is made possible by the addition of programmable stages and higher precision arithmetic to the rendering pipelines, which allows programmers to use stream processing on non-graphics data[1][2][3]. Additionally, the use of multiple graphics cards in a single computer, or large numbers of graphics chips, further parallelizes the already parallel nature of graphics processing[4]
Contents |
GPU functionality has, traditionally, been very limited. In fact, for many years the GPU was only used to accelerate certain parts of the graphics pipeline. Some improvements were needed before GPGPU became feasible.
Programmable vertex and fragment shaders were added to the graphics pipeline to enable game programmers to generate even more realistic effects. Vertex shaders allow the programmer to alter per-vertex attributes, such as position, color, texture coordinates, and normal vector. Fragment shaders are used to calculate the color of a fragment, or per-pixel. Programmable fragment shaders allow the programmer to substitute, for example, a lighting model other than those provided by default by the graphics card, typically simple Gouraud shading. Shaders have enabled graphics programmers to create lens effects, displacement mapping, and depth of field.
The programmability of the pipelines have trended according to Microsoft’s DirectX specification , with DirectX 8 introducing Shader Model 1.1, DirectX 8.1 Pixel Shader Models 1.2, 1.3 and 1.4, and DirectX 9 defining Shader Model 2.x and 3.0. Each shader model increased the programming model flexibilities and capabilities, ensuring the conforming hardware follows suit. The DirectX 10 specification introduces Shader Model 4.0 which unifies the programming specification for vertex, geometry (“Geometry Shaders” are new to DirectX 10) and fragment processing allowing for a better fit for unified shader hardware, thus providing a single computational pool of programmable resource.
Pre-DirectX 9 graphics cards only supported paletted or integer color types. Various formats are available, each containing a red element, a green element, and a blue element. Sometimes an additional alpha value is added, to be used for transparency. Common formats are:
For early fixed-function or limited programmability graphics (i.e. up to and including DirectX 8.1-compliant GPUs) this was sufficient because this is also the representation used in displays. This representation does have certain limitations, however. Given sufficient graphics processing power even graphics programmers would like to use better formats, such as floating point data formats, in order to obtain effects such as high dynamic range imaging. Many GPGPU applications require floating point accuracy, which came with graphics cards conforming to the DirectX 9 specification.
DirectX 9 Shader Model 2.x suggested the support of two precision types: full and partial precision. Full precision support could either be FP32 or FP24 (floating point 32- or 24-bit per component) or greater, while partial precision was FP16. ATI’s R300 series of GPUs supported FP24 precision only in the programmable fragment pipeline (although FP32 was supported in the vertex processors) while Nvidia’s NV30 series supported both FP16 and FP32; other vendors such as S3 Graphics and XGI supported a mixture of formats up to FP24.
Shader Model 3.0 altered the specification, increasing full precision requirements to a minimum of FP32 support in the fragment pipeline. ATI’s Shader Model 3.0 compliant R5xx generation (Radeon X1000 series) supports just FP32 throughout the pipeline while Nvidia’s NV4x and G7x series continued to support both FP32 full precision and FP16 partial precisions. Although not stipulated by Shader Model 3.0, both ATI and Nvidia’s Shader Model 3.0 GPUs introduced support for blendable FP16 render targets, more easily facilitating the support for High Dynamic Range Rendering.
The implementations of floating point on Nvidia GPUs are mostly IEEE compliant; however, this is not true across all vendors.[5] This has implications for correctness which are considered important to some scientific applications. While 64-bit floating point values (double precision float) are commonly available on CPUs, these are not universally supported on GPUs; some GPU architectures sacrifice IEEE compliance while others lack double-precision altogether. There have been efforts to emulate double-precision floating point values on GPUs; however, the speed tradeoff negates any benefit to offloading the computation onto the GPU in the first place.[6]
Most operations on the GPU operate in a vectorized fashion: a single operation can be performed on up to four values at once. For instance, if one color <R1, G1, B1> is to be modulated by another color <R2, G2, B2>, the GPU can produce the resulting color <R1*R2, G1*G2, B1*B2> in a single operation. This functionality is useful in graphics because almost every basic data type is a vector (either 2-, 3-, or 4-dimensional). Examples include vertices, colors, normal vectors, and texture coordinates. Many other applications can put this to good use, and because of their higher performance, vector instructions (SIMD) have long been available on CPUs.
In 2002 Fung etal developed OpenVIDIA at University of Toronto, and demonstrated this work, which was later published in 2003, 2004, and 2005[7], in conjunction with a collaboration between University of Toronto and nVIDIA. In November 2006 Nvidia launched CUDA, an SDK and API that allows a programmer to use the C programming language to code algorithms for execution on Geforce 8 series GPUs. . OpenCL, an open standard defined by the Khronos Group[8] provides a cross platform GPGPU platform that additionally supports data parallel compute on CPUs. OpenCL is actively supported on Intel, AMD, Nvidia and Arm platforms. GPGPU compared, for example, to traditional floating point accelerators such as the 64-bit CSX700 boards from ClearSpeed that are used in today's supercomputers, current top-end GPUs from AMD and Nvidia emphasize single-precision (32-bit) computation; double-precision (64-bit) computation executes more slowly.
GPUs are designed specifically for graphics and thus are very restrictive in terms of operations and programming. Because of their nature, GPUs are only effective at tackling problems that can be solved using stream processing and the hardware can only be used in certain ways.
GPUs can only process independent vertices and fragments, but can process many of them in parallel. This is especially effective when the programmer wants to process many vertices or fragments in the same way. In this sense, GPUs are stream processors – processors that can operate in parallel by running a single kernel on many records in a stream at once.
A stream is simply a set of records that require similar computation. Streams provide data parallelism. Kernels are the functions that are applied to each element in the stream. In the GPUs, vertices and fragments are the elements in streams and vertex and fragment shaders are the kernels to be run on them. Since GPUs process elements independently there is no way to have shared or static data. For each element we can only read from the input, perform operations on it, and write to the output. It is permissible to have multiple inputs and multiple outputs, but never a piece of memory that is both readable and writable.
Arithmetic intensity is defined as the number of operations performed per word of memory transferred. It is important for GPGPU applications to have high arithmetic intensity else the memory access latency will limit computational speedup.[9]
Ideal GPGPU applications have large data sets, high parallelism, and minimal dependency between data elements.
There are a variety of computational resources available on the GPU:
In fact, the programmer can substitute a write only texture for output instead of the framebuffer. This is accomplished either through Render to Texture (RTT), Render-To-Backbuffer-Copy-To-Texture (RTBCTT), or the more recent stream-out.
The most common form for a stream to take in GPGPU is a 2D grid because this fits naturally with the rendering model built into GPUs. Many computations naturally map into grids: matrix algebra, image processing, physically based simulation, and so on.
Since textures are used as memory, texture lookups are then used as memory reads. Certain operations can be done automatically by the GPU because of this.
Kernels can be thought of as the body of loops. For example, if the programmer were operating on a grid on the CPU they might have code that looked like this:
// Input and output grids have 10000 x 10000 or 100 million elements. void transform_10k_by_10k_grid(float in[10000][10000], float out[10000][10000]) { for(int x = 0; x < 10000; x++) { for(int y = 0; y < 10000; y++) { // The next line is executed 100 million times out[x][y] = do_some_hard_work(in[x][y]); } } }
On the GPU, the programmer only specifies the body of the loop as the kernel and what data to loop over by invoking geometry processing.
In sequential code it is possible to control the flow of the program using if-then-else statements and various forms of loops. Such flow control structures have only recently been added to GPUs.[10] Conditional writes could be accomplished using a properly crafted series of arithmetic/bit operations, but looping and conditional branching were not possible.
Recent GPUs allow branching, but usually with a performance penalty. Branching should generally be avoided in inner loops, whether in CPU or GPU code, and various techniques, such as static branch resolution, pre-computation, predication, loop splitting[11], and Z-cull[12] can be used to achieve branching when hardware support does not exist.
The map operation simply applies the given function (the kernel) to every element in the stream. A simple example is multiplying each value in the stream by a constant (increasing the brightness of an image). The map operation is simple to implement on the GPU. The programmer generates a fragment for each pixel on screen and applies a fragment program to each one. The result stream of the same size is stored in the output buffer.
Some computations require calculating a smaller stream (possibly a stream of only 1 element) from a larger stream. This is called a reduction of the stream. Generally a reduction can be accomplished in multiple steps. The results from the previous step are used as the input for the current step and the range over which the operation is applied is reduced until only one stream element remains.
Stream filtering is essentially a non-uniform reduction. Filtering involves removing items from the stream based on some criteria.
The scatter operation is most naturally defined on the vertex processor. The vertex processor is able to adjust the position of the vertex, which allows the programmer to control where information is deposited on the grid. Other extensions are also possible, such as controlling how large an area the vertex affects.
The fragment processor cannot perform a direct scatter operation because the location of each fragment on the grid is fixed at the time of the fragment's creation and cannot be altered by the programmer. However, a logical scatter operation may sometimes be recast or implemented with an additional gather step. A scatter implementation would first emit both an output value and an output address. An immediately following gather operation uses address comparisons to see whether the output value maps to the current output slot.
The fragment processor is able to read textures in a random access fashion, so it can gather information from any grid cell, or multiple grid cells, as desired.
The sort operation transforms an unordered set of elements into an ordered set of elements. The most common implementation on GPUs is using sorting networks.[12]
The search operation allows the programmer to find a particular element within the stream, or possibly find neighbors of a specified element. The GPU is not used to speed up the search for an individual element, but instead is used to run multiple searches in parallel.
A variety of data structures can be represented on the GPU:
The following are some of the areas where GPUs have been used for general purpose computing:
|
||||||||||||||||||||||||||||
|
||||||||||||||||||||||||||||||||||||||