Mike Tien-Chien Lee

Instruction level power analysis and optimization of software

The increasing popularity of power constrained mobile computers and embedded computing applications drives the need for analyzing and optimizing power in all the components of a system. Software constitutes a major component of todays systems, and its role is projected to grow even further. Thus, an ever increasing portion of the functionality of todays systems is in the form of instructions, as opposed to gates. This motivates the need for analyzing power consumption from the point of view of instructions—something that traditional circuit and gate level power analysis tools are inadequate for. This paper describes an alternative, measurement based instruction level power analysis approach that provides an accurate and practical way of quantifying the power cost of soft-ware. This technique has been applied to three commercial, architecturally different processors. The salient results of these analyses are summarized. Instruction level analysis of a processor helps in the development of models for power consumption of software executing on that processor. The power models for the subject processors are described and interesting observations resulting from the comparison of these models are highlighted. The ability to evaluate software in terms of power consumption makes it feasible to seach fow low power implementations of given programs. In addition, it can guide the development of general tools and techniques for low power software. Several ideas in this regard as motivated by the power analysis of the subject processors are also described.

Using register-transfer paths in code generation for heterogeneous memory-register architectures

In this paper we address the problem of code generation for basic blocks in heterogeneous memory-register DSP processors. We propose a new a technique, based on register-transfer paths, that can be used for efficiently dismantling basic block DAGs (Directed Acyclic Graphs) into expression trees. This approach builds on recent results which report optimal code generation algorithm for expression trees for these architectures. This technique has been implemented and experimentally validated for the TMS320C25, a popular fixed point DSP processor. The results show that good code quality can be obtained using the proposed technique. An analysis of the type of DAGs found in the DSPstone benchmark programs reveals that the majority of basic blocks in this benchmark set are expression trees and leaf DAGs. This leads to our claim that tree based algorithms, like the one described in this paper, should be the technique of choice for basic blocks code generation with heterogeneous memory register architectures.

Power analysis and low-power scheduling techniques for embedded DSP software

Abstract: This paper describes the application of a measurement based power analysis technique for an embedded DSP processor. An instruction-level power model for the processor has been developed using this technique. Significant points of difference have been observed between this model and the ones developed earlier for some general-purpose commercial microprocessors. In particular, the effect of circuit state on the power cost of an instruction stream is more marked in the case of this DSP processor. In addition, the DSP processor has a special architectural feature that allows instructions to be packed into pairs. The energy reduction possible through the use of this feature is studied. The on-chip Booth multiplier on the processor is a major source of energy consumption for DSP programs. A micro-architectural power model for the multiplier is developed and analyzed for further energy minimization. A scheduling algorithm incorporating these new techniques is proposed to reduce the energy consumed by DSP software. Energy reductions varying from 11% to 56% have been observed for several example programs. These energy savings are real and have been verified through physical measurement.

Power analysis of a 32-bit embedded microcontroller

A new approach for power analysis of microprocessors has recently been proposed (Tiwari et al 1994). The idea is to look at the power consumption in a microprocessor from the point of view of the actual software executing on the processor. The basic component of this approach is a measurement based, instruction-level power analysis technique. The technique allows for the development of an instruction-level power model for the given processor, which can be used to evaluate software in terms of the power consumption, and for exploring the optimization of software for lower power. This paper describes the application of this technique for a comprehensive instruction-level power analysis of a commercial 32-bit RISC-based embedded microcontroller. The salient results of the analysis and the basic instruction-level power model are described. Interesting observations and insights based on the results are also presented. Such an instruction-level power analysis can provide cues as to what optimizations in the micro-architecture design of the processor would lead to the most effective power savings in actual software applications. Wherever the results indicate such optimizations, they have been discussed. Furthermore, ideas for low power software design, as suggested by the results, are described in this paper as well.

Power Analysis of a 32-bit Embedded Microcontroller

A new approach for power analysis of microprocessors has recently been proposed [14]. The idea is to look at the power consumption in a microprocessor from the point of view of the actual software executing on the processor. The basic component of this approach is a measurement based, instruction-level power analysis technique. The technique allows for the development of an instruction-level power model for the given processor, which can be used to evaluate software in terms of the power consumption, and for exploring the optimization of software for lower power. This paper describes the application of this technique for a comprehensive instruction-level power analysis of a commercial 32-bit RISC-based embedded microcontroller. The salient results of the analysis and the basic instruction-level power model are described. Interesting observations and insights based on the results are also presented. Such an instruction-level power analysis can provide cues as to what optimizations in the micro-architecture design of the processor would lead to the most effective power savings in actual software applications. Wherever the results indicate such optimizations, they have been discussed. Furthermore, ideas for low power software design, as suggested by the results, are described in this paper as well.

Domain-specific high-level modeling and synthesis for ATM switch design using VHDL

This paper presents our experience on domain-specific high-level modeling and synthesis for Fujitsu ATM switch design. We propose a high-level design methodology using VHDL, where ATM switch architectural features are considered during behavior modeling, and a high-level synthesis compiler, MEBS, is prototyped to synthesize the behavior model down to a gate-level implementation. Since the specific ATM switch architecture is incorporated into both modeling and syntheses phases, a high-quality design is efficiently derived. The synthesis results show that given the design constraints, the proposed high-level design methodology can produce a gate-level implementation by MEBS with about 15% area reduction in shorter design cycle when compared with manual design.

Eliminating false loops caused by sharing in control path

In high-level synthesis, resource sharing may result in a circuit containing false loops that create great difficulty in timing validation during the design sign-off phase. It is hence desirable to avoid generating any false loops in a synthesized circuit. Previous work [Stok 1992; Huang et al. 1995] considered mainly data path sharing for false loop elimination. However, for a complete circuit with both data path and control path, false loops can be created due to control logic sharing. In this article, we present a novel approach to detect and eleminate the false loops caused by control logic sharing. An effective filter is devised to reduce the computational complexity of false loop detection, which is based on checking the level numbers that are propagated from data path operators to imputs and outputs of the control path. Only the input/output pairs of the control path identified by the filter are further investigated by traversing into the data path for false loop detection. A removal algorithm is then applied to eliminate the detected false loops, followed by logic minimization to further optimize the circuit. Experimental results show that for the nine example circuits we tested, the final designs after false loop removal and logic minimization give only slightly larger area than the original ones that contain false loops.

Domain---Specific High---Level Modeling and Synthesis for ATM Switch Prototyping

ATM switch, the core technology of an ATM networking system, is one of the major products in Fujitsu telecommunication business. However, current gate–level design methodology can no longer satisfy its stringent time–to–market requirement. It becomes necessary to exploit high–level methodology to specify and synthesize the design at an abstraction level higher than logic gates. This paper presents our prototyping experience on domain–specific high–level modeling and synthesis for Fujitsu ATM switch design. We propose a high–level design methodology using VHDL, where ATM switch architectural features are considered during behavior modeling, and a high–level synthesis compiler, MEBS, is prototyped to synthesize the behavior model down to a gate–level implementation. Since the specific ATM switch architecture is incorporated into both modeling and synthesis phases, a high–quality design is efficiently derived. The synthesis results shows that given the design constraints, the proposed high–level design methodology can produce a gate–level implementation by MEBS with about 15 percent area reduction in shorter design cycle when compared with manual design.