5-ProgramDesign

8/10/2019 5-ProgramDesign

1/78


2/78


3/78


4/78


5/78


6/78


7/78


8/78


9/78


10/78


11/78


12/78


13/78


14/78


15/78


16/78


17/78


18/78


19/78


20/78


21/78


22/78


23/78


24/78


25/78


26/78


27/78


28/78


29/78


30/78


31/78


32/78


33/78


34/78


35/78


36/78


37/78


38/78


39/78


40/78


41/78


42/78


43/78


44/78


45/78


46/78


47/78


48/78


49/78


50/78


51/78


52/78


53/78


54/78


55/78


56/78


57/78


58/78


59/78


60/78


61/78


62/78


63/78


64/78


65/78


66/78


67/78


68/78

stemas Embebidos 2013/14

ogram Design and Analysis 68

d cc ][

Performance profiling

Count the nr of times procedures or basicblocks are called (not time)Use a counter on that block/procedureAnalyse PC during executionHLL may hide long sequence ofinstructions

SE 2013/14 - Program design and analysis - pbrandao (see at beginning)

135

d cc ][

Performance optimization hints

Use registers efficiently.Use page mode memory accesses.

Re-arrange variables to reference themcontiguously

Analyse cache behaviour:instruction conflicts can be handled byrewriting code, rescheduling;

conflicting scalar data can easily be moved;

conflicting array data can be moved,padded.

136



69/78



d cc ][

HLL Performance optimization

Use a better algorithm or data structureEnable compiler optimizations

The ARM Infocenter as a coding practice

Don't optimize what doesn't matter


137

From The Practice ofProgramming Kernighanand Pike

d cc ][

HLL Performance optimization:tuning the code

Collect common sub expressions

Replace expensive operations by cheap ones

Unroll or eliminate loops

Cache frequently-used values.

Write a special-purpose allocator

Buffer input and output

Handle special cases separately

Precompute results

Use approximate values


138

http://infocenter.arm.com/help/index.jsp?topic=/com.arm.doc.dui0205j/CJAFJCFG.htmlhttp://infocenter.arm.com/help/index.jsp?topic=/com.arm.doc.dui0205j/CJAFJCFG.html


70/78



Program level Energy andPower Analysis andOptimization


139

d cc ][

Energy/power optimization

Energy : ability to do work.Most important in battery-powered systems.

Power : energy per unit time.Important even in wall-plug systems---powerbecomes heat.

140



71/78



d cc ][

Measuring energy consumption

Execute a small loop, measure current:


141

while (TRUE){testCode();

}

IA

CPU

d cc ][

Sources of energy consumption

Relative energy per operation (Catthoor et al):

memory transfer: 33

external I/O: 10

SRAM write: 9

SRAM read: 4.4multiply: 3.6

add: 1


142


72/78



d cc ][

Cache behaviour is important

Energy consumption has a sweet spot ascache size changes:

cache too small : program thrashes, burningenergy on external memory accesses;

cache too large : cache itself burns too muchpower.

143


d cc ][

Cache sweet spot

[Li98] 1998 IEEE

144



73/78



d cc ][

Optimizing for energy

First-order optimization:high performance = low energy .

Not many instructions trade speed forenergy.

145


d cc ][

Optimizing for energy, cont d.

Use registers efficiently.Identify and eliminate cache conflicts.Moderate loop unrolling eliminates someloop overhead instructions.Eliminate pipeline stalls.Inlining procedures may help: reduceslinkage, but may increase cachethrashing.

146



74/78



d cc ][

Efficient loops

General rules:Don t use function calls.

Keep loop body small to enable local repeat(only forward branches).

Use unsigned integer for loop counter.

Use


75/78



Analysis and Optimizationof Program Size


149

d cc ][

Optimizing for program size

Goal:reduce hardware cost of memory;

reduce power consumption of memory units.

Two opportunities:data;

instructions.

150


This big


76/78



d cc ][

Data size minimization

Reuse constants, variables, data buffers indifferent parts of code.

Requires careful verification of correctness.

Generate data using instructions.

151


d cc ][

Reducing code size

Avoid function inlining.Choose CPU with compact instructions.

ARM has Thumb and ARM instruction sets

Thumb are 16 bit instructions (see diff )

Use specialized instructions where possible.

152

http://infocenter.arm.com/help/index.jsp?topic=/com.arm.doc.dui0068b/ch02s02s09.htmlhttp://infocenter.arm.com/help/index.jsp?topic=/com.arm.doc.dui0068b/ch02s02s09.html


77/78



d cc ][

Size Hints

Use the smallest possible data typeDon't store what you can easilyrecomputed

This may have a cost in run-time


153


Program design andanalysisThe End

154



78/78


d cc ][

Summary

Components for Embedded ProgramsModels of ProgramsAssembly, Linking, and LoadingBasic Compilation TechniquesProgram OptimizationProgram-level Performance Analysis

Software Performance Optimization

155


5-ProgramDesign

Documents