The SIMD (single instruction, multiple data) control style achieves a very cost effective processor element (PE) control mechanism. In this paper, two efficient data transfer operations for a SIMD PE array processor are proposed to address the inefficiency of most SIMD processor arrays in performing irregular memory access per PE. First, the SIMD random access provides simultaneous data transfer of randomly, from each PE independently addressed elements. Second, the SIMD ROI (region of interest) access enables concurrent data transfer of ROI areas with optional different ROI parameters for each PE. Speed-up values for both new transfer operations show an at least six times faster execution and the synthesis output reports an area increase of merely 1% for a 32 PE SIMD array processor configuration.

This paper describes existing designs and future design trends of in-vehicle vision processors for driver assistance systems. First, requirements of vision processors for driver assistance systems are summarized. Next, the characteristics of vision tasks for safety are described. Then several in-vehicle vision processor LSI implementations are reviewed, and the design approach of one of them, the IMAPCAR highly parallel processor, is further described in detail. Finally, future trends of in-vehicle vision processors focusing on their architectures and application coverage expansion such as integration of vision for safety, Digital TV codec, and 3D graphics functions of future car navigation, are discussed.

A scalable SIMD/MIMD mixed-mode parallel processor architecture called XC core is proposed to meet the high and diverse performance requirements of embedded multimedia applications. XC core supports both the SIMD and MIMD computing models at low hardware cost by dynamically reconfiguring itself into datapath circuits or control circuits, i.e., trading off between performance and flexibility. A control processor is used to broadcast instructions to a whole SIMD PE (Processing Element) array or to a part of it while assigning a separate program to each PU (Processing Unit), that is mainly composed of the hardware resources of several PEs. RTL synthesis results show that area overhead for reconfiguration is merely 10% of the total area. Benchmark results show that the SIMD mode is effectively achieving high performance towards the regular and massive data parallelism portions of applications, while the MIMD mode enables acceleration of the remaining part of applications whose implementation using a pure highly parallel SIMD architecture would otherwise be impossible. The results show that the XC core design is competitive against more complex processors, with respect to both its cost efficiency as a highly parallel SIMD processor and its flexibility as a multicore MIMD processor, against a wide range of applications.

Embedded processors for video image recognition in most cases not only need to address the conventional cost (die size and power) versus real-time performance issue, but must also maintain high flexibility due to the immense diversity of recognition targets, situations, and applications. This paper describes IMAP, a highly parallel SIMD linear processor and memory array architecture that addresses these trade-off requirements. By using parallel and systolic algorithmic techniques, but based on a simple linear array architecture, IMAP successfully exploits not only the straightforward per-image row data level parallelism (DLP), but also the inherent DLP of other memory access patterns frequently found in various image recognition tasks, while allowing programming to be done using an explicit parallel C language (1DC). We describe and evaluate IMAP-CE, one of the latest IMAP processors, integrating 128 100 MHz 8 bit 4-way VLIW PEs, 128 2 KByte RAMs, and one 16 bit RISC control processor onto a single chip. The PE instruction set is enhanced to support 1DC code. The die size of IMAP-CE is 11 times11 mm2integrating 32.7 M transistors, while the power consumption is, on average, approximately 2 watts. IMAP-CE is evaluated mainly by comparing its performance while running 1DC code with that of a 2.4 GHz Intel P4 running optimized C code. Based on the use of parallelizing techniques, benchmark results show a speed increase of up to 20 times for image filter kernels and of 4 times for a full image recognition application

Embedded processors for video image recognition require to address both the cost (die size and power) versus real-time performance issue, and also to achieve high flexibility due to the immense diversity of recognition targets, situations, and applications. This paper describes IMAP, a highly parallel SIMD linear processor and memory array architecture that addresses these trading-off requirements. By using parallel and systolic algorithmic techniques, despite of its simple architecture IMAP achieves to exploit not only the straightforward per image row data level parallelism (DLP), but also the inherent DLP of other memory access patterns frequently found in various image recognition tasks, under the use of an explicit parallel C language (IDC). We describe and evaluate IMAP-CE, a latest IMAP processor, which integrates 128 of 100MHz 8 bit 4-way VLIW PEs, 128 of 2KByte RAMs, and one 16 bit RISC control processor, into a single chip. The PE instruction set is enhanced for supporting IDC codes. IMAP-CE is evaluated mainly by comparing its performance running IDC codes with that of a 2.4GHz Intel P4 running optimized C codes. Based on the use of parallelizing techniques, benchmark results show a speedup of up to 20 for image filter kernels, and of 4 for a full image recognition application.

IMAP-CE is the fourth generation of a series of SIMD linear processor arrays based on the IMAP (integrated memory array processor) architecture. The aim of IMAP-CE is to provide a compact, cost effective and yet high performance solution for various embedded real-time vision applications, especially for vision based driving assistance applications in the ITS (intelligent transportation system) fields. IMAP-CE integrates 128 VLIW processing elements, and a RISC control processor which provides the single instruction stream for the processor array. The peak performance of IMAP-CE is up to 51.2 GOPS operating under 100 MHz. This paper describes the design features of IMAP-CE, its enhanced instruction set for image processing, and the estimated performance.

A newly developed one-board real-time image processing system with 10.24-GIPS peak performance is described. Such performance is possible due to a linear processor array of 256 SIMD processing elements. The IMAP-VISION chip, which is a key device of the system, integrates 32 8-bit processing elements and 32 l-Kbyte/PE memories. It also has a synchronous DRAM (SDRAM) interface with 160-MB/s bandwidth which enables a large capacity memory to be used for real-time vision applications. As the chip is designed to be able to overlap data transfer between the on-chip memories and the SDRAMs with computations, the system is capable of executing complicated algorithms flexibly without memory access overhead or shortage of memory capacity. In addition, a bit-pack function is used to improve binary image processing performance. The efficiency of this function is illustrated by a region labeling algorithm which is often used in object recognition. The system manages to process real-time region labeling and Hough transform, and shows high performance for most low-level image processing tasks.

This paper describes the real-time vision system (RVS-2) which shows quite high performance for low-level image processing while it is implemented in a one-board type compact size format with small power consumption. The RVS-2 consists of an IMAP board, a video board and a host workstation. The IMAP board consists of eight highly-integrated IMAP LSIs and a dedicated control LSI (RVSC). The IMAP chip integrates 2 Mb image memory and 64 processing elements that operate in the SIMD mode. The RVSC chip performs global data operations efficiently without interactions with the host workstation, as well providing an instruction stream to the IMAP chips. The peak performance of the RVS-2 is 30 GIPS and most of the basic image processing tasks are carried out within about 0.1-0.7 ms, which is about 50-300 times faster than the video frame rate.>

In general, datafiow graph for which it is difficult to optimize quality are those used for branches (conditionals) and loops. Although a number of dataflow graph transformation schemes have taken these into accounts, the quality of graphs been produced was not compared with that of hand-written optimized graphs.