Texas Instruments TMS320 DSP Computer Hardware User Manual


 
www.ti.com
6.13C6xxxSpecificDMARulesandGuidelines
6.13.1CacheCoherencyIssuesforAlgorithmProducers
Y
X = 0ld
Y = new
X
L2 cache External memory
DMA
X
Y
X = 0ld
Y = new
X
L2 cache External memory
Y
DMA
C6xxxSpecificDMARulesandGuidelines
IncertainC6000targets,datathatareinbothexternalmemoryandtheL2cachecancausecoherence
problemswithbackgroundDMAtransfersinseveralways.Thefiguresbelowdepictsomememoryaccess
scenariosthatpotentiallyleadtoproblems.Welaterintroducerulesandguidelinesforbothalgorithmand
frameworkdeveloperstoensurecorrectoperationofC6000algorithms.
InSection6.13.2,CPUaccessofthememorycorrespondingtolocationxbringsitintotheL2cache.
SubsequentwritestoxtakeplaceintheL2cacheuntilthecachelinecontainingxgetswrittenbackto
externalmemory.IfaDMAtransferstartscopyingthedatafromlocationxtoanotherlocation,itmayend
upreadingstalevalueofxinexternalmemorysincecertainDMAcontrollerswillnotdetectpresenceor
flushingofadirtycachelinecontainingx.Toavoidthisproblem,thecachemustbeflushedbeforethe
DMAreadproceeds.
InSection6.13.3,thelocationxhasbeenbroughtintotheL2cache.SupposeaDMAtransferwritesnew
datatolocationx.Inthiscase,theCPUwouldaccesstheoldcacheddatainasubsequentread,unless
thecachedcopyisinvalidated.
Algorithmsmustenforcecoherenceandalignment/sizeconstraintsforinternalbufferstheyrequest
throughtheIALGinterface.Todealwiththesecoherencyproblems,thefollowingnewguidelinesand
ruleshavebeenadded.
DMAGuideline3
Toensurecorrectness,AllC6000algorithmsthatimplementIDMA2needtobesuppliedwiththe
internalmemorytheyrequestfromtheclientapplicationusingalgAlloc().
Thisguidelineappliestotheclientapplication,ratherthantothealgorithm.IfDMAGuideline3isfollowed;
i.e.,ifthetypeofmemoryrequestedisprovided,thealgorithmisguaranteedtooperatecorrectly.
DMARule6
C6000algorithmsmustnotissueanyCPUread/writestobuffersinexternalmemorythatareinvolved
inDMAtransfers.Thisalsoappliestobufferspassedtothealgorithmthroughitsalgorithminterface.
DMARule6isnecessarybecauseitistheonlywayforaneXpressDSP-compliantalgorithmtoavoid
havingtodealwithcachecoherenceoperationssuchascachelinewriteback,cachelineinvalidate,etc.
Theseoperationsarelow-levelandshouldbedealtwithattheclientapplicationlevel.Withthe
introductionofDMARule6,noexternalbuffersinvolvedinDMAtransferswillendupinthecache,and
thereforenoexternalcoherencyproblemswilloccur.
DMARule7
IfaC6000algorithmhasimplementedtheIDMA2interface,theclientmustallocatealltherequired
externalmemoryatacachelineboundary.Thesebuffersmustbeamultipleofcachelinelengthin
size.Theclientmustalsoensurethatthesebuffersarenotincachebeforepassingthemtothe
algorithm.
SPRU352GJune2005RevisedFebruary2007UseoftheDMAResource69
SubmitDocumentationFeedback