Abstract: With the rapid development of deep learning, neural network and deep learning algorithms play a significant role in various practical applications. Due to the high accuracy and good performance, Convolutional Neural Networks (CNNs) especially have become a research hot spot in the past few years. However, the size of the networks becomes increasingly large scale due to the demands of the practical applications, which poses a significant challenge to construct a high-performance implementation of deep learning neural networks. Meanwhile, many of these application scenarios also have strict requirements on the performance and low-power consumption of hardware devices. Therefore, it is particularly critical to choose a moderate computing platform for hardware acceleration of CNNs. This article aimed to survey the recent advance in Field Programmable Gate Array (FPGA)-based acceleration of CNNs. Various designs and implementations of the accelerator based on FPGA under different devices and network models are overviewed, and the versions of Graphic Processing Units (GPUs), Application Specific Integrated Circuits (ASICs) and Digital Signal Processors (DSPs) are compared to present our own critical analysis and comments. Finally, we give a discussion on different perspectives of these acceleration and optimization methods on FPGA platforms to further explore the opportunities and challenges for future research. More helpfully, we give a prospect for future development of the FPGA-based accelerator.
Abstract: Autonomous driving systems require high reliability to provide people with a safe and comfortable driving experience. However, despite the development of a number of vehicle sensors, it is difficult to always provide high perceived performance in driving environments that vary from time to season. The image segmentation method using deep learning, which has recently evolved rapidly, provides high recognition performance in various road environments stably. However, since the system controls a vehicle in real time, a highly complex deep learning network cannot be used due to time and memory constraints. Moreover, efficient networks are optimized for GPU environments, which degrade performance in embedded processor environments equipped simple hardware accelerators. In this paper, a semantic segmentation network, matrix multiplication accelerator network (MMANet), optimized for matrix multiplication accelerator (MMA) on Texas instrument digital signal processors (TI DSP) is proposed to improve the recognition performance of autonomous driving system. The proposed method is designed to maximize the number of layers that can be performed in a limited time to provide reliable driving environment information in real time. First, the number of channels in the activation map is fixed to fit the structure of MMA. By increasing the number of parallel branches, the lack of information caused by fixing the number of channels is resolved. Second, an efficient convolution is selected depending on the size of the activation. Since MMA is a fixed, it may be more efficient for normal convolution than depthwise separable convolution depending on memory access overhead. Thus, a convolution type is decided according to output stride to increase network depth. In addition, memory access time is minimized by processing operations only in L3 cache. Lastly, reliable contexts are extracted using the extended atrous spatial pyramid pooling (ASPP). The suggested method gets stable features from an extended path by increasing the kernel size and accessing consecutive data. In addition, it consists of two ASPPs to obtain high quality contexts using the restored shape without global average pooling paths since the layer uses MMA as a simple adder. To verify the proposed method, an experiment is conducted using perfsim, a timing simulator, and the Cityscapes validation sets. The proposed network can process an image with 640 x 480 resolution for 6.67 ms, so six cameras can be used to identify the surroundings of the vehicle as 20 frame per second (FPS). In addition, it achieves 73.1% mean intersection over union (mIoU) which is the highest recognition rate among embedded networks on the Cityscapes validation set.
Abstract: ''Cocktail party problem'' is well known as one of the human auditory abilities. We can recognize the specific sound that we want to listen by this ability even if a lot of undesirable sounds or noises are mixed. Blind source separation (BSS) based on independent component analysis (ICA) is one of the methods by which we can separate only a special signal from their mixed signals with simple hypothesis. In this paper, we propose an online approach for blind source separation using the sliding DFT and the time domain independent component analysis. The proposed method can reduce calculation complexity in comparison with conventional methods, and can be applied to parallel processing by using digital signal processors (DSPs) and so on. We evaluate this method and show its availability.
Abstract: Full adders are important components in applications
such as digital signal processors (DSP) architectures and
microprocessors. In addition to its main task, which is adding two
numbers, it participates in many other useful operations such as
subtraction, multiplication, division,, address calculation,..etc. In
most of these systems the adder lies in the critical path that
determines the overall speed of the system. So enhancing the
performance of the 1-bit full adder cell (the building block of the
adder) is a significant goal.Demands for the low power VLSI have
been pushing the development of aggressive design methodologies to
reduce the power consumption drastically. To meet the growing
demand, we propose a new low power adder cell by sacrificing the
MOS Transistor count that reduces the serious threshold loss
problem, considerably increases the speed and decreases the power
when compared to the static energy recovery full (SERF) adder. So a
new improved 14T CMOS l-bit full adder cell is presented in this
paper. Results show 50% improvement in threshold loss problem,
45% improvement in speed and considerable power consumption
over the SERF adder and other different types of adders with
comparable performance.
Abstract: Unlike general-purpose processors, digital signal
processors (DSP processors) are strongly application-dependent. To
meet the needs for diverse applications, a wide variety of DSP
processors based on different architectures ranging from the
traditional to VLIW have been introduced to the market over the
years. The functionality, performance, and cost of these processors
vary over a wide range. In order to select a processor that meets the
design criteria for an application, processor performance is usually
the major concern for digital signal processing (DSP) application
developers. Performance data are also essential for the designers of
DSP processors to improve their design. Consequently, several DSP
performance benchmarks have been proposed over the past decade or
so. However, none of these benchmarks seem to have included recent
new DSP applications.
In this paper, we use a new benchmark that we recently developed
to compare the performance of popular DSP processors from Texas
Instruments and StarCore. The new benchmark is based on the
Selectable Mode Vocoder (SMV), a speech-coding program from the
recent third generation (3G) wireless voice applications. All
benchmark kernels are compiled by the compilers of the respective
DSP processors and run on their simulators. Weighted arithmetic
mean of clock cycles and arithmetic mean of code size are used to
compare the performance of five DSP processors.
In addition, we studied how the performance of a processor is
affected by code structure, features of processor architecture and
optimization of compiler. The extensive experimental data gathered,
analyzed, and presented in this paper should be helpful for DSP
processor and compiler designers to meet their specific design goals.
Abstract: In this paper, low end Digital Signal Processors (DSPs)
are applied to accelerate integer neural networks. The use of DSPs
to accelerate neural networks has been a topic of study for some
time, and has demonstrated significant performance improvements.
Recently, work has been done on integer only neural networks, which
greatly reduces hardware requirements, and thus allows for cheaper
hardware implementation. DSPs with Arithmetic Logic Units (ALUs)
that support floating or fixed point arithmetic are generally more
expensive than their integer only counterparts due to increased circuit
complexity. However if the need for floating or fixed point math
operation can be removed, then simpler, lower cost DSPs can be
used. To achieve this, an integer only neural network is created in
this paper, which is then accelerated by using DSP instructions to
improve performance.