Embedded Systems Design - July 2008 - (Page 20)

cover feature which parts of a video frame may be One of the design goals of these the source code, but in the end, there is safely processed in parallel by a tools was the ability to retain the no silver bullet: either the designer or video encoder application. structure of the original source code the tool (or both) must sufﬁciently “unTo achieve this goal, a subset of as much as possible when performderstand” the source code to be able to the CleanC rules detail how to use ing transformations. This makes the transform it correctly and efﬁciently. constructs that reveal the intention transformed code easier to recogHence, it’s crucial that we identify those of the designer more clearly. For innize and understand, which is useful constructs in the source code that hamstance, when processing a video when the designer must edit the per this process. These can be code conframe, it’s easier to analyze the code transformed code. structs at the level of programming lanwhen the designer uses an explicit Undisciplined use of header ﬁles guage idioms or even at the level of code construct to represent the and C preprocessor statements design patterns for parallel programframe as a two-dimensional array makes this design goal difﬁcult to ming. Once we’ve identiﬁed these conwith explicit indexing to access the achieve. A subset of CleanC rules structs, we may elaborate rules that pixel values, and when the designer detail how to properly use header specify in which circumstances these uses canonical for loops with ﬁles and preprocessor statements. constructs should be avoided. Source-code transformations may This is the rationale of IMEC’s soknown bounds to enumerate the also introduce conﬂicts between called CleanC rules that have been elabpixels. (The IMEC mapping tools variable names; hence, CleanC sugorated during the development of an understand a controlled form of dygests using different names for difMPSoC application-mapping tool suite. namic multidimensional memory ferent things. Other constructs that These rules target ANSI-C language idallocation.) Equivalent code with make source code transformations ioms that make the process of mapping data and function pointers and with more difﬁcult include irregular consequential C programs onto an MPSoC pointer arithmetic often obscures trol ﬂow and expressions with side platform more difﬁcult. They consist of the intentions and makes the analyeffects. a few tens of guidelines and restrictions. sis more difﬁcult. Code restrictions are conSome exceptions to structs that aren’t allowed to this intention-revealUndisciplined use of header ﬁles and be present in the input C code ing principle exist, C preprocessor statements makes for the tools to work (on that however. For instance, part of the code). while recursive functhis design goal difﬁcult to achieve. For instance, undiscition deﬁnitions often plined use of preprocessing express the intention information in C can make life very hard more clearly than their nonrecursive for any kind of source-code transformacounterparts, they are typically Category 2—Rules to enable and tion tool. Code guidelines describe how harder to analyze by tools. improve data-dependency analycode should be written to give the tools sis. A major challenge in parallelizCategory 3—Rules to minimize maximum accuracy and transformation ing code and optimizing usage of data dependencies. Needless data freedom. The majority of these rules the memory hierarchy consists of dependencies complicate the analytranscends the speciﬁc functionality of dealing with data dependencies. sis process and may lower the qualithe IMEC mapping tools and may form Knowing which data are needed by ty of the analysis results. This, in the basis for a standard set of guidelines which parts of the code and underturn, affects the ability to optimize for programming MPSoC platforms in a standing the exact dependencies beand parallelize the source code. “clean” code called CleanC. tween the program parts that upCleanC not only details to avoid date the data and the parts that read global variables and duplicated HOW CLEANC DIFFERS FROM C those data are crucial instruments functionality, but it also imposes We can roughly classify the CleanC rules for parallelizing and optimizing the stricter constraints on the usage of in four different groups. These classiﬁcacode. Brute-force approaches for types: types with different names are tions are not completely independent. synchronizing accesses to shared basically considered different. For instance, applying a rule from one data resources may result in worstUsing code constructs like globcategory may strengthen the effects of case assumptions about the proal variables also has more direct rules in other categories. grams behavior. Hence they are typnegative effects, for instance by ically bound to be suboptimal at needlessly increasing the lifecycle of best and highly inefﬁcient at worst. large data structures. Category 1—Rules to enable and Thus, we must be able to analyze improve source-code transformaCategory 4—Rules to avoid the data dependencies in more precise tions. The IMEC mapping tools are dark corners of the C standards. and ﬁne-grained ways: for instance, source-code transformation tools. The C standards contain many ele- • • • • 20 JULY 2008 | embedded systems design | www.embedded.com http://www.embedded.com

Table of Contents Feed for the Digital Edition of Embedded Systems Design - July 2008

Embedded Systems Design - July 2008 Contents #Include Parity Bit Programming Pointers Interactive C-code Cleaning Tool Supports Multiprocessor SoC Design Building a Power Supply for Discontinuous Transmission Wireless Networks An Exception Primer Advertising Index Break Points Marketplace

Embedded Systems Design - July 2008

For optimal viewing of this digital publication, please enable JavaScript and then refresh the page. If you would like to try to load the digital publication without using Flash Player detection, please click here.