Optimization of Audio Processing algorithms (Reverb) on ARMv6 family of processors
Audio processing algorithms are increasingly used in cell phones and today’s customers are placing more demands on cell phones. Feature phones, once the advent of mobile phone technology, nowadays do more than just providing the user with MP3 play back or advanced audio effects. These features have become an integral part of medium as well as low-end phones. On the other hand, there is also an endeavor to include as improved quality as possible into products to compete in market and satisfy users’ needs. Tackling the above requirements has been partly satisfied by the advance in hardware design and manufacturing technology. However, as new hardware emerges into market the need for competence to write efficient software and exploit the new features thoroughly and effectively arises. Even though compilers are also keeping up with the new tide space for hand optimized code still exist. Wrapped in the above goal, an effort was made in this thesis to partly cover the competence requirement at Multimedia Section (part of Ericsson Mobile Platforms) to develope optimized code for new processors. Forging persistently ahead with new products, EMP has always incorporated the latest technology into its products among which ARMv6 family of processors has the main central processing role in a number of upcoming products. To fully exploit latest features provided by ARMv6, it was required to probe its new instruction set among which new media processing instructions are of outmost importance. In order to execute DSP-intensive algorithms (e.g. Audio Processing algorithms) efficiently, the implementation should be done in low-level code applying available instruction set. Meanwhile, ARMv6 comes with a number of new features in comparison with its predecessors. SIMD (Single Instruction Multiple Data) and VFP (Vector Floating Point) are the most prominent media processing improvements in ARMv6. Aligned with thesis goals and guidelines, Reverb algorithm which is among one of the most complicated audio features on a hand-held devices was probed. Consequently, its kernel parts were identified and implementation was done both in fixed-point and floating-point using the available resources on hardware. Besides execution time and amount of code memory for each part were measured and provided in tables and charts for comparison purposes. Conclusions were finally drawn based on developed code’s efficiency over ARM compiler’s as well as existing code already developed and tailored to ARMv5 processors. The main criteria for optimization was the execution time. Moreover, quantization effect due to limited precision fixed-point arithmetic was formulated and its effect on quality was elaborated. The outcomes, clearly indicate that hand optimization of kernel parts are superior to Compiler optimized alternative both from the point of code memory as well as execution time. The results also confirmed the presumption that hand optimized code using new instruction set can improve efficiency by an average 25%-50% depending on the algorithm structure and its interaction with other parts of audio effect. Despite its many draw backs, fixed-point implementation remains yet to be the dominant implementation for majority of DSP algorithms on low-power devices.
Automatic Parallel Memory Address Generation for Parallel DSP Computing
The concept of Parallel Vector (scratch pad) Memories (PVM) was introduced as one solution for Parallel Computing in DSP, which can provides parallel memory addressing efficiently with minimum latency. The parallel programming more efficient by using the parallel addressing generator for parallel vector memory (PVM) proposed in this thesis. However, without hiding complexities by cache, the cost of programming is high. To minimize the programming cost, automatic parallel memory address generation is needed to hide the complexities of memory access. This thesis investigates methods for implementing conflict-free vector addressing algorithms on a parallel hardware structure. In particular, match vector addressing requirements extracted from the behaviour model to a prepared parallel memory addressing template, in order to supply data in parallel from the main memory to the on-chip vector memory. According to the template and usage of the main and on-chip parallel vector memory, models for data pre-allocation and permutation in scratch pad memories of ASIP can be decided and configured. By exposing the parallel memory access of source code, the memory access flow graph (MFG) will be generated. Then MFG will be used combined with hardware information to match templates in the template library. When it is matched with one template, suited permutation equation will be gained, and the permutation table that include target addresses for data pre-allocation and permutation is created. Thus it is possible to automatically generate memory address for parallel memory accesses. A tool for achieving the goal mentioned above is created, Permutator, which is implemented in C++ combined with XML. Memory access coding template is selected, as a result that permutation formulas are specified. And then PVM address table could be generated to make the data pre-allocation, so that efficient parallel memory access is possible. The result shows that the memory access complexities is hiden by using Permutator, so that the programming cost is reduced.It works well in the context that each algorithm with its related hardware information is corresponding to a template case, so that extra memory cost is eliminated.
Voice Codec for Floating Point Processor
As part of an ongoing project at the department of electrical engineering, ISY, at Linköping University, a voice decoder using floating point formats has been the focus of this master thesis. Previous work has been done developing an mp3-decoder using the floating point formats. All is expected to be implemented on a single DSP.The ever present desire to make things smaller, more efficient and less power consuming are the main reasons for this master thesis regarding the use of a floating point format instead of the traditional integer format in a GSM codec. The idea with the low precision floating point format is to be able to reduce the size of the memory. This in turn reduces the size of the total chip area needed and also decreases the power consumption.One main question is if this can be done with the floating point format without losing too much sound quality of the speech. When using the integer format, one can represent every value in the range depending on how many bits are being used. When using a floating point format you can represent larger values using fewer bits compared to the integer format but you lose representation of some values and have to round the values off.From the tests that have been made with the decoder during this thesis, it has been found that the audible difference between the two formats is very small and can hardly be heard, if at all. The rounding seems to have very little effect on the quality of the sound and the implementation of the codec has succeeded in reproducing similar sound quality to the GSM standard decoder.
Evaluation of Image Warping Algorithms for Implementation in FPGA
The target of this master thesis is to evaluate the Image Warping technique and propose a possible design for an implementation in FPGA. The Image Warping is widely used in the image processing for image correction and rectification. A DSP is a usual choice for implantation of the image processing algorithms, but to decrease a cost of the target system it was proposed to use an FPGA for implementation. In this work a different Image Warping methods was evaluated in terms of performance, produced image quality, complexity and design size. Also, considering that it is not only Image Warping algorithm which will be implemented on the target system, it was important to estimate a possible memory bandwidth used by the proposed design. The evaluation was done by implemented a C-model of the proposed design with a finite datapath to simulate hardware implementation as close as possible.
Implementation of Algorithms on FPGAs
This thesis describes how an algorithm is transferred from a digital signal processor to an embedded microprocessor in an FPGA using C to hardware program from Altera. Saab Avitronics develops the secondary high lift control system for the Boeing 787 aircraft. The high lift system consists of electric motors controlling the trailing edge wing flaps and the leading edge wing slats. The high lift motors manage to control the Boeing 787 aircraft with full power even if half of each motor’s stators are damaged. The motor is a PMDC brushless motor which is controlled by an advanced algorithm. The algorithm needs to be calculated by a fast special digital signal processor. In this thesis I have tested if the algorithm can be transferred to an FPGA and still manage the time and safety demands. This was done by transferring an already working algorithm from the digital signal processor to an FPGA. The idea was to put the algorithm in an embedded NIOS II microprocessor and speed up the bottlenecks with Altera’s C to hardware program. The study shows that the C-code needs to be optimized for C to hardware to manage the up speeding part, as the tests showed that the calculation time for the algorithm actually became longer with C to hardware. This thesis also shows that it is highly probable to use an FPGA equipped with Altera’s NIOS II safety critical microprocessor instead of a digital signal processor to control the electrical high lift motors in the Boeing 787 aircraft.
DSP Platform Benchmarking
●1 commentBenchmarking of DSP kernel algorithms was conducted in the thesis on a DSP processor for teaching in the course TESA26 in the department of Electrical Engineering. It includes benchmarking on cycle count and memory usage. The goal of the thesis is to evaluate the quality of a single MAC DSP instruction set and provide suggestions for further improvement in instruction set architecture accordingly. The scope of the thesis is limited to benchmark the processor only based on assembly coding. The quality check of compiler is not included. The method of the benchmarking was proposed by BDTI, Berkeley Design Technology Incorporations, which is the general methodology used in world wide DSP industry. Proposals on assembly instruction set improvements include the enhancement of FFT and DCT. The cycle cost of the new FFT benchmark based on the proposal was XX% lower, showing that the proposal was right and qualified. Results also show that the proposal promotes the cycle cost score for matrix computing, especially matrix multiplication. The benchmark results were compared with general scores of single MAC DSP processors offered by BDTI.
Efficient arithmetic for high speed DSP implementation on FPGAs
The author was sponsored by EnTegra Ltd, a company who develop hardware and software products and services for the real time implementation of DSP and RF systems. The field programmable gate array (FPGA) is being used increasingly in the field of DSP. This is due to the fact that the parallel computing power of such devices is ideal for today’s truly demanding DSP algorithms. Algorithms such as the QR-RLS update are computationally intensive and must be carried out at extremely high speeds (MHz). This means that the DSP processor is simply not an option. ASICs can be used but the expense of developing custom logic is prohibitive. The increased use of the FPGA in DSP means that there is a significant requirement for efficient arithmetic cores that utilises the resources on such devices. This thesis presents the research and development effort that was carried out to produce fixed point division and square root cores for use in a new Electronic Design Automation (EDA) tool for EnTegra, which is targeted at FPGA implementation of DSP systems. Further to this, a new technique for predicting the accuracy of CORDIC systems computing vector magnitudes and cosines/sines is presented. This work allows the most efficient CORDIC design for a specified level of accuracy to be found quickly and easily without the need to run lengthy simulations, as was the case before. The CORDIC algorithm is a technique using mainly shifts and additions to compute many arithmetic functions and is thus ideal for FPGA implementation.
Algorithm Adaptation and Optimization of a Novel DSP Vector Co-processor
The Division of Computer Engineering at Linköping's university is currently researching the possibility to create a highly parallel DSP platform, that can keep up with the computational needs of upcoming standards for various applications, at low cost and low power consumption. The architecture is called ePUMA and it combines a general RISC DSP master processor with eight SIMD co-processors on a single chip. The master processor will act as the main processor for general tasks and execution control, while the co-processors will accelerate computing intensive and parallel DSP kernels.This thesis investigates the performance potential of the co-processors by implementing matrix algebra kernels for QR decomposition, LU decomposition, matrix determinant and matrix inverse, that run on a single co-processor. The kernels will then be evaluated to find possible problems with the co-processors' microarchitecture and suggest solutions to the problems that might exist. The evaluation shows that the performance potential is very good, but a few problems have been identified, that causes significant overhead in the kernels. Pipeline mismatches, that occurs due to different pipeline lengths for different instructions, causes pipeline hazards and the current solution to this, doesn't allow effective use of the pipeline. In some cases, the single port memories will cause bottlenecks, but the thesis suggests that the situation could be greatly improved by using buffered memory write-back. Also, the lack of register forwarding makes kernels with many data dependencies run unnecessarily slow.
Implementation of Elementary Functions for a Fixed Point SIMD DSP Coprocessor
This thesis is about implementing the functions for reciprocal, square root, inverse square root and logarithms on a DSP platform. A multi-core DSP platform that consists of one master processor core and several SIMD coprocessor cores is currently being designed by a team at the Computer Engineering Department of Linköping University. The SIMD coprocessors’ arithmetic logic unit (ALU) has 16 multipliers to support vector multiplication instructions. By efficiently using the 16 multipliers, it is possible to evaluate polynomials very fast. The ALU does not have (hardware) support for floating point arithmetic, so the challenge is to get good precision by using fixed point arithmetic. Precise and fast solutions to implement the mathematical functions are found by converting the fixed point input to a soft floating point format before polynomial approximation, choosing a polynomial based on an error analysis of the polynomial approximation, and using Newton-Raphson or Goldschmidt iterations to improve the precision of the polynomial approximations. Finally, suggestions are made of changes and additions to the instruction set architecture, in order to make the implementations faster, by efficiently using the currently existing hardware.
Benchmarking a DSP processor
This Master thesis describes the benchmarking of a DSP processor. Benchmarking means measuring the performance in some way. In this report, we have focused on the number of instruction cycles needed to execute certain algorithms. The algorithms we have used in the benchmark are all very common in signal processing today. The results we have reached in this thesis have been compared to benchmarks for other processors, performed by Berkeley Design Technology, Inc. The algorithms were programmed in assembly code and then executed on the instruction set simulator. After that, we proposed changes to the instruction set, with the aim to reduce the execution time for the algorithms. The results from the benchmark show that our processor is at the same level as the ones tested by BDTI. Probably would a more experienced programmer be able to reduce the cycle count even more, especially for some of the more complex benchmarks.
Real-time Motion Picture Restoration
Through age or misuse, motion picture films can develop damage in the form of dirt or scratches which detract from the quality of the film. Removal of these artifacts is a worthwhile process as it makes the films more visually attractive and extends the life of the material. In this thesis, various methods for detecting and concealing the effects of film damage are described. Appropriate algorithms are selected for implementation of a system, based on a TMS320C80 video processor, which can remove the effects of film defects using digital processing. The restoration process operates in real-time at video frame rates (30 frames per second). Details of the software implementation of this system are presented along with results from processing damaged film material. The effects of damage are significantly reduced after processing.
Energy Profiling of DSP Applications, A Case Study of an Intelligent ECG Monitor
●7 commentsProper balance of power and performance for optimum system organization requires precise profiling of the power consumption of different hardware subsystems as well as software functions. Moreover, power consumption of mobile systems is even more important, since the battery is a large portion of the overall size and weight of the system. Average power consumption is only a crude estimate of power requirements and battery life; a much better estimate can be made using dynamic power consumption. Dynamic power consumption is a function of the execution profile of the given application running on specific hardware platform. In this paper we introduce a new environment for energy profiling of DSP applications. The environment consists of a JTAG emulator, a high-resolution HP 3583A multimeter and a workstation that controls devices and stores the traces. We use Texas Instruments’ Real Time Data Exchange mechanism (RTDXÔ) to generate an execution profile and custom procedures for energy profile data acquisition using GPIB interface. We developed custom procedures to correlate and analyze both energy and execution profiles. The environment allows us to improve the system power consumption through changes in software organization and to measure real battery life for the given hardware, software and battery configuration. As a case study, we present the analysis of a real-time portable ECG monitor implemented using a Texas Instruments TMS320C5410-100 processor board, and a Del Mar PWA ECG Amplifier.
Teaching MODEM Concepts and Design Procedure with MATLAB Simulations
●4 commentsMATLAB simulation is used as the primary tool to illustrate concepts, to validate MODEM designs, and to vent' operation of the subsystems employed in DSP based transmitters and receivers presented in a pair of classes on MODEM Design and Digital Receiver Design. The whole gamut of subsystems found in conventional and experimental modem designs are simulated and assembled to form a full end-to-end simulation of an operating MODEM. This paper describes the philosophy used to guide class involvement and assess the experience and the learning value to student participants.
Cascaded Integrator-Comb (CIC) Filter Introduction
●3 commentsIn the classic paper, "An Economical Class of Digital Filters for Decimation and Interpolation", Hogenauer introduced an important class of digital filters called "Cascaded Integrator-Comb", or "CIC" for short (also sometimes called "Hogenauer filters"). Here, Matthew Donadio provides a more gentle introduction to the subject of CIC filters, geared specifically to the needs of practicing DSP designers.
Using the DFT as a Filter: Correcting a Misconception
●1 commentI have read, in some of the literature of DSP, that when the discrete Fourier transform (DFT) is used as a filter the process of performing a DFT causes an input signal's spectrum to be frequency translated down to zero Hz (DC). I can understand why someone might say that, but I challenge that statement as being incorrect. Here are my thoughts.
An Introduction To Compressive Sampling
This article surveys the theory of compressive sensing, also known as compressed sensing or CS, a novel sensing/sampling paradigm that goes against the common wisdom in data acquisition.
Region based Active Contour Segmentation
In this paper, we propose a natural framework that allows any region-based segmentation energy to be re-formulated in a local way. We consider local rather than global image statistics and evolve a contour based on local information. Localized contours are capable of segmenting objects with heterogeneous feature profiles that would be difficult to capture correctly using a standard global method. The presented technique is versatile enough to be used with any global region-based active contour energy and instill in it the benefits of localization. We describe this framework and demonstrate the localization of three well-known energies in order to illustrate how our framework can be applied to any energy. We then compare each localized energy to its global counterpart to show the improvements that can be achieved. Next, an in-depth study of the behaviors of these energies in response to the degree of localization is given. Finally, we show results on challenging images to illustrate the robust and accurate segmentations that are possible with this new class of active contour models.
LOW-RESOURCE DELAYLESS SUBBAND ADAPTIVE FILTER USING WEIGHTED OVERLAP-ADD
●2 commentsA delayless structure targeted for low-resource implementation is proposed to eliminate filterbank processing delays in subband adaptive filters (SAFs). Rather than using direct IFFT or polyphase filterbanks to transform the SAFs back into the time-domain, the proposed method utilizes a weighted overlap-add (WOLA) synthesis. Low-resource real-time implementations are targeted and as such do not involve long (as long as the echo plant) FFT or IFFT operations. Also, the proposed approach facilitates time distribution of the adaptive filter reconstruction calculations crucial for efficient real-time and hardware implementation. The method is implemented on an oversampled WOLA filterbank employed as part of an echo cancellation application. Evaluation results demonstrate that the proposed implementation outperforms conventional SAF systems since the signals used in actual adaptive filtering are not distorted by filterbank aliasing. The method is a good match for partial update adaptive algorithms since segments of the time-domain adaptive filter are sequentially reconstructed and updated.
Closing the gap: CPU and FPGA Trends in sustainable floating-point BLAS performance
Field programmable gate arrays (FPGAs) have long been an attractive alternative to microprocessors for computing tasks — as long as floating-point arithmetic is not required. Fueled by the advance of Moore’s Law, FPGAs are rapidly reaching sufficient densities to enhance peak floating-point performance as well. The question, however, is how much of this peak performance can be sustained. This paper examines three of the basic linear algebra subroutine (BLAS) functions: vector dot product, matrix-vector multiply, and matrix multiply. A comparison of microprocessors, FPGAs, and Reconfigurable Computing platforms is performed for each operation. The analysis highlights the amount of memory bandwidth and internal storage needed to sustain peak performance with FPGAs. This analysis considers the historical context of the last six years and is extrapolated for the next six years.
Gauss-Newton Based Learning for Fully Recurrent Neural Networks
●4 commentsThe thesis discusses a novel off-line and on-line learning approach for Fully Recurrent Neural Networks (FRNNs). The most popular algorithm for training FRNNs, the Real Time Recurrent Learning (RTRL) algorithm, employs the gradient descent technique for finding the optimum weight vectors in the recurrent neural network. Within the framework of the research presented, a new off-line and on-line variation of RTRL is presented, that is based on the Gauss-Newton method. The method itself is an approximate Newton’s method tailored to the specific optimization problem, (non-linear least squares), which aims to speed up the process of FRNN training. The new approach stands as a robust and effective compromise between the original gradient-based RTRL (low computational complexity, slow convergence) and Newton-based variants of RTRL (high computational complexity, fast convergence). By gathering information over time in order to form Gauss-Newton search vectors, the new learning algorithm, GN-RTRL, is capable of converging faster to a better quality solution than the original algorithm. Experimental results reflect these qualities of GN-RTRL, as well as the fact that GN-RTRL may have in practice lower computational cost in comparison, again, to the original RTRL.
