Applying the fundamentals of parallel programming to multiprocessor designs: Part 2To see how you might apply the methods outlined in Part 1 to a practical computing problem, consider the error diffusion algorithm that is used in many computer graphics and image processing programs.
Originally proposed by Floyd and Steinberg, error diffusion is a technique for displaying continuous-tone digital images on devices that have limited color (tone) range.
Printing an 8-bit grayscale image to a black-and-white printer is problematic. The printer, being a bi-level device, cannot print the 8-bit image natively. It must simulate multiple shades of gray by using an approximation technique.An example of an image before and after the error diffusion process is shown in Figure 3.2 below. The original image, composed of 8-bit grayscale pixels, is shown on the left, and the result of the image that has been processed using the error diffusion algorithm is shown on the right.
The output image is composed of pixels of only two colors: black and white. Original 8-bit image on the left, resultant 2-bit image on the right. At the resolution of this printing, they look similar.
The same images as above but zoomed to 400 percent and cropped to 25 percent to show pixel detail. Now you can clearly see the 2-bit black-white rendering on the right and 8-bit gray-scale on the left.
|Figure 3.2 Error Diffusion Algorithm Output|
The basic error diffusion algorithm does its work in a simple three-step process:
Step #1. Determine the output value given the input value of the current pixel. This step often uses quantization, or in the binary case, thresholding. For an 8-bit grayscale image that is displayed on a 1-bit output device, all input values in the range [0, 127] are to be displayed as a 0 and all input values between [128, 255] are to be displayed as a 1 on the output device.
Step #2. Once the output value is determined, the code computes the error between what should be displayed on the output device and what is actually displayed. As an example, assume that the current input pixel value is 168. Given that it is greater than our threshold value (128), we determine that the output value will be a 1.
This value is stored in the output array. To compute the error, the program must normalize output first, so it is in the same scale as the input value. That is, for the purposes of computing the display error, the output pixel must be 0 if the output pixel is 0 or 255 if the output pixel is 1. In this case, the display error is the difference between the actual value that should have been displayed (168) and the output value (255), which is -87.
Step #3. Finally, the error value is distributed on a fractional basis to the neighboring pixels in the region, as shown in Figure 3.3 below.
|Figure 3.3. Distributing Error Values to Neighboring Pixels|
This example uses the Floyd-Steinberg error weights to propagate errors to neighboring pixels. 7/16ths of the error is computed and added to the pixel to the right of the current pixel that is being processed. 5/16ths of the error is added to the pixel in the next row, directly below the current pixel.
The remaining errors propagate in a similar fashion. While you can use other error weighting schemes, all error diffusion algorithms follow this general method.
The three-step process is applied to all pixels in the image. Listing 3.1 below shows a simple C implementation of the error diffusion algorithm, using Floyd-Steinberg error weights.
|Listing 3.1. C-language Implementation of the Error Diffusion Algorithm|
Analysis of the Error Diffusion
At first glance, one might think that the error diffusion algorithm is an inherently serial process. The conventional approach distributes errors to neighboring pixels as they are computed.
As a result, the previous pixel's error must be known in order to compute the value of the next pixel. This interdependency implies that the code can only process one pixel at a time. It's not that difficult, however, to approach this problem in a way that is more suitable to a multithreaded approach.
An Alternate Approach: Parallel
To transform the conventional error diffusion algorithm into an approach that is more conducive to a parallel solution, consider the different decomposition that were covered previously in this chapter.
Which would be appropriate in this case? As a hint, consider Figure 3.4 below, which revisits the error distribution illustrated in Figure 3.3, from a slightly different perspective.
|Figure 3.4. Error-Diffusion Error Computation from the Receiving Pixel's Perspective|
Given that a pixel may not be processed until its spatial predecessors have been processed, the problem appears to lend itself to an approach where we have a producer - or in this case, multiple producers - producing data (error values) which a consumer (the current pixel) will use to compute the proper output pixel.
The flow of error data to the current pixel is critical. Therefore, the problem seems to break down into a data-flow decomposition.
Now that we identified the approach, the next step is to determine the best pattern that can be applied to this particular problem. Each independent thread of execution should process an equal amount of work (load balancing).
How should the work be partitioned? One way, based on the algorithm presented in the previous section, would be to have a thread that processed the even pixels in a given row, and another thread that processed the odd pixels in the same row. This approach is ineffective however; each thread will be blocked waiting for the other to complete, and the performance could be worse than in the sequential case.
To effectively subdivide the work among threads, we need a way to reduce (or ideally eliminate) the dependency between pixels.
Figure 3.4 above illustrates an important point that's not obvious in Figure 3.3 earlier - that in order for a pixel to be able to be processed, it must have three error values - labeled eA, eB, and eC1 in Figure 3.3 - from the previous row, and one error value from the pixel immediately to the left on the current row. [We assume eA = eD = 0 at the left edge of the page (for pixels in column 0); and that eC = 0 at the right edge of the page (for pixels in column W-1, where W = the number of pixels in the image)].
Thus, once these pixels are processed, the current pixel may complete its processing. This ordering suggests an implementation where each thread processes a row of data. Once a row has completed processing of the first few pixels, the thread responsible for the next row may begin its processing. Figure 3.5 below shows this sequence.
|Figure 3.5 Parallel Error Diffusion for Multi-thread, Multi-row Situation|
Notice that a small latency occurs at the start of each row. This latency is due to the fact that the previous row's error data must be calculated before the current row can be processed. These types of latency are generally unavoidable in producer-consumer implementations; however, you can minimize the impact of the latency as illustrated here.
The trick is to derive the proper workload partitioning so that each
thread of execution works as efficiently as possible. In this case, you
incur a two-pixel latency before processing of the next thread can
In the previous section, we proposed a method of error diffusion where each thread processed a row of data at a time. However, one might consider subdividing the work at a higher level of granularity.
Instinctively, when partitioning work between threads, one tends to look for independent tasks. The simplest way of parallelizing this problem would be to process each page separately. Generally speaking, each page would be an independent data set, and thus, it would not have any interdependencies.
So why did we propose a row-based solution instead of processing individual pages? The three key reasons are:
1) An image may span multiple pages. This implementation would impose a restriction of one image per page, which might or might not be suitable for the given application.
2) Increased memory usage. An 8.5 x 11-inch page at 1,200 dpi consumes 131 megabytes of RAM. Intermediate results must be saved; therefore, this approach would be less memory efficient.
3) An application might, in a common use-case, print only a single page at a time. Subdividing the problem at the page level would offer no performance improvement from the sequential case.
A hybrid approach would be to subdivide the pages and process regions of a page in a thread, as illustrated in Figure 3.6 below.
|Figure 3.6. Parallel Error Diffusion for Multi-thread, Multi-page Situation|
Note that each thread must work on sections from different page. This increases the startup latency involved before the threads can begin work. In Figure 3.6 above, Thread 2 incurs a 1/3 page startup latency before it can begin to process data, while Thread 3 incurs a 2/3 page startup latency.
While somewhat improved, the hybrid approach suffers from similar limitations as the page-based partitioning scheme described above. To avoid these limitations, you should focus on the row-based error diffusion implementation illustrated in Figure 3.5.
The key points to keep in mind when developing solutions for parallel computing architectures are:
fall into one of three categories: task, data, and data flow.
2) Task-level parallelism partitions the work between threads based on tasks.
3) Data decomposition breaks down tasks based on the data that the threads work on.
4) Data flow decomposition breaks down the problem in terms of how data flows between the tasks.
5) Most parallel programming problems fall into one of several well known patterns.
6) The constraints of synchronization, communication, load balancing, and scalability must be dealt with to get the most benefit out of a parallel program.
Many problems that appear to be serial may, through a simple transformation, be adapted to a parallel implementation.
To read Part 1, go to Breaking
up tasks for multicore multithreading
This series of two articles was excerpted from Multi-Core Programming by Shameem Akhter and Jason Roberts, published by Intel Press. Copyright © 2006 Intel Corporation. All rights reserved. This book can be purchased on line.
Shameem Akhter is a platform
architect at Intel Corp., focusing on single socket multicore
architecture and performance analysis. Jason Roberts is a senior
software engineer at Intel, working on a number of different
multithreaded software products ranging from desktop, handheld and
embedded DSP platforms.
To learn more about designing
go to More
About Multicores and Multiprocessing.