Multidimensional digital signal processing (MDSP) refers to the extension of digital signal processing (DSP) techniques to signals that vary in more than one dimension. While conventional DSP typically deals with one-dimensional data, such as time-varying audio signals, MDSP involves processing signals in two or more dimensions. Many of the principles from one-dimensional DSP, such as Fourier transforms and filter design, have analogous counterparts in multidimensional signal processing. Modern general-purpose computing on graphics processing units (GPGPUs) have an excellent throughput on vector operations and numeric manipulations through a high degree of parallel computations. Processing digital signals, particularly multidimensional signals, often involves a series of vector operations on massive numbers of independent data samples, GPGPUs are now widely employed to accelerate multidimensional DSP, such as image processing, video codecs, radar signal analysis, sonar signal processing, and ultrasound scanning. Conceptually, GPGPUs dramatically reduce the computation complexity when compared with central processing units (CPUs), digital signal processors (DSPs), or other FPGA accelerators.
Motivation Processing multidimensional signals is a common problem in scientific research and/or engineering computations. Typically, a DSP problem's computation complexity grows exponentially with the number of dimensions. Notwithstanding, with a high degree of time and storage complexity, it is extremely difficult to process multidimensional signals in real-time. Although many fast algorithms (e.g. FFT) have been proposed for 1-D DSP problems, they are still not efficient enough to be adapted in high dimensional DSP problems. Therefore, it is still hard to obtain the desired computation results with digital signal processors. Hence, better algorithms and hardware architecture are needed to accelerate multidimensional DSP computations.
Existing approaches Practically, to accelerate multidimensional DSP, some common approaches have been proposed and developed in the past decades.
Lower sampling rate A makeshift to achieve a real-time requirement in multidimensional DSP applications is to use a lower sampling rate, which can efficiently reduce the number of samples to be processed at one time and thereby decrease the total processing time. However, this can lead to the aliasing problem due to the sampling theorem and poor-quality outputs. In some applications, such as military radars and medical images, we are eager to have highly precise and accurate results. In such cases, using a lower sampling rate to reduce the amount of computation in the multidimensional DSP domain is not always allowable.
Digital signal processors Digital signal processors are designed specifically to process vector operations. They have been widely used in DSP computations for decades. However, most digital signal processors are only capable of manipulating a few operations in parallel. This kind of design is sufficient to accelerate audio processing (1-D signals) and image processing (2-D signals). However, with a large number of data samples of multidimensional signals, this is still not powerful enough to retrieve computation results in real-time.
Supercomputer assistance In order to accelerate multidimensional DSP computations, using dedicated supercomputers or cluster computers is required in some circumstances, e.g., weather forecasting and military radars. Nevertheless, using supercomputers designated to simply perform DSP operations takes considerable money cost and energy consumption. Also, it is not practical and suitable for all multidimensional DSP applications.
GPU acceleration GPUs are originally devised to accelerate image processing and video stream rendering. Moreover, since modern GPUs have good ability to perform numeric computations in parallel with a relatively low cost and better energy efficiency, GPUs are becoming a popular alternative to replace supercomputers performing multidimensional DSP.
GPGPU computations
Modern GPU designs are mainly based on the SIMD (Single Instruction Multiple Data) computation paradigm. This type of GPU devices is so-called general-purpose GPUs (GPGPUs). GPGPUs are able to perform an operation on multiple independent data concurrently with their vector or SIMD functional units. A modern GPGPU can spawn thousands of concurrent threads and process all threads in a batch manner. With this nature, GPGPUs can be employed as DSP accelerators easily while many DSP problems can be solved by divide-and-conquer algorithms. A large scale and complex DSP problem can be divided into a bunch of small numeric problems and be processed altogether at one time so that the overall time complexity can be reduced significantly. For example, multiplying two M × M matrices can be processed by M × M concurrent threads on a GPGPU device without any output data dependency. Therefore, theoretically, by means of GPGPU acceleration, we can gain up to M × M speedup compared with a traditional CPU or digital signal processor.
GPU programming languages Currently, there are several existing programming languages or interfaces which support GPGPU programming.
CUDA CUDA is the standard interface to program Nvidia GPUs. Nvidia also provides many CUDA libraries to support DSP acceleration on Nvidia GPU devices.
OpenCL OpenCL is an industrial standard which was originally proposed by Apple Inc. and is maintained and developed by the Khronos Group now. OpenCL provides C++ like APIs for programming different devices universally, including GPGPUs.
The following figure illustrates the execution flow of launching an OpenCL program on a GPU device. The CPU first detects OpenCL devices (GPU in this case) and then invokes a just-in-time compiler to translate the OpenCL source code into target binary. CPU then sends data to GPU to perform computations. When the GPU is processing data, CPU is free to process its own tasks.
C++ AMP C++ AMP is a programming model proposed by Microsoft. C++ AMP is a C++ based library designed for programming SIMD processors
OpenACC OpenACC is a programming standard for parallel computing developed by Cray, CAPS, NVIDIA and PGI. OpenAcc targets programming for CPU and GPU heterogeneous systems with C, C++, and Fortran extensions.
Examples of GPU programming for multidimensional DSP
m × m matrix multiplication Suppose A and B are two m × m matrices and we would like to compute C = A × B.
… excerpt ends here. Continue reading the full article.
