help="make xml output pretty")
(options, args) = parser.parse_args()
setupDeferredLogging('preprocessor.log', logging.DEBUG if options.debug else logging.INFO)
if (not options.sourceDir):
logging.error("sourceDirectory option is mandatory. Use '--help' for informations on how to use this module")
sys.exit(1)
#prepare xml output
doc = Document()
callGraphRoot = doc.createElement("callGraph")
doc.appendChild(callGraphRoot)
filesInDir = dirEntries(str(options.sourceDir), True, 'h90')
#first pass: loop through all h90 files (hybrid fortran 90) in the current directory
# and build the basic callgraph based on subprocedures and calls. Also parse @-directives for annotations.
progressIndicatorReset(sys.stderr)
for fileNum, fileInDir in enumerate(filesInDir):
parser = H90XMLCallGraphGenerator(doc)
parser.processFile(fileInDir)
logging.debug("Callgraph generated for " + fileInDir + "")
printProgressIndicator(sys.stderr, fileInDir, fileNum + 1, len(filesInDir), "Callgraph parsing")
#second pass: moved to generateP90Codebase.py since we need symbol analysis already
if (options.pretty):
sys.stdout.write(doc.toprettyxml())
else:
sys.stdout.write(doc.toxml())
if (not options.sourceDir):
logging.error(
"sourceDirectory option is mandatory. Use '--help' for informations on how to use this module"
)
sys.exit(1)
#prepare xml output
doc = Document()
callGraphRoot = doc.createElement("callGraph")
doc.appendChild(callGraphRoot)
filesInDir = dirEntries(str(options.sourceDir), True, 'h90')
#first pass: loop through all h90 files (hybrid fortran 90) in the current directory
# and build the basic callgraph based on subprocedures and calls. Also parse @-directives for annotations.
progressIndicatorReset(sys.stderr)
for fileNum, fileInDir in enumerate(filesInDir):
parser = H90XMLCallGraphGenerator(doc)
parser.processFile(fileInDir)
logging.debug("Callgraph generated for " + fileInDir + "")
printProgressIndicator(sys.stderr, fileInDir, fileNum + 1, len(filesInDir),
"Callgraph parsing")
#second pass: moved to generateP90Codebase.py since we need symbol analysis already
if (options.pretty):
sys.stdout.write(doc.toprettyxml())
else:
sys.stdout.write(doc.toxml())
sys.stdout.flush()
def analyseParallelRegions(doc, appliesTo):
callNodes = doc.getElementsByTagName("call")
routineNodes = doc.getElementsByTagName("routine")
templates = doc.getElementsByTagName("parallelRegionTemplate")
for routineNum, routineNode in enumerate(routineNodes):
filterParallelRegionNodes(doc, routineNode, appliesTo, templates)
printProgressIndicator(
sys.stderr,
"",
routineNum + 1,
len(routineNodes),
"Filtering parallel regions for %s" %(appliesTo) if appliesTo != "" else "Filtering parallel regions"
)
progressIndicatorReset(sys.stderr)
callNodesByCallerName = getCalleesByCallerName(callNodes)
callNodesByCalleeName = getCallersByCalleeName(callNodes)
parallelRegionNodes = doc.getElementsByTagName("parallelRegions")
parallelRegionNodesByRoutineName = {}
sys.stderr.write("Populating parallel region cache\n")
for regionNum, parallelRegionNode in enumerate(parallelRegionNodes):
routine = parallelRegionNode.parentNode
routineName = routine.getAttribute("name")
if appliesTo == "GPU" and parallelRegionNodesByRoutineName.get(routineName) != None:
raise Exception("Multiple GPU parallel regions in subroutine %s" %(routineName))
parallelRegionNodesByRoutineName[routineName] = parallelRegionNode
kernelCallerProblemFound = False
messagesPresentedFor = []
sys.stderr.write("Parallel region analysis\n")
for regionNum, parallelRegionNode in enumerate(parallelRegionNodes):
if not parallelRegionNode.parentNode.tagName == "routine":
raise Exception("Parallel region not within routine.")
parallelRegionNode.parentNode.setAttribute("parallelRegionPosition", "within")
routine = parallelRegionNode.parentNode
routineName = routine.getAttribute("name")
if routineName == None:
raise Exception("Kernel routine without name")
addAttributeToAllCallGraphAncestors(routineNodes, callNodesByCalleeName, routine, "parallelRegionPosition", "inside")
addAttributeToAllCallGraphHeirs(routineNodes, callNodesByCallerName, routine, "parallelRegionPosition", "outside")
#rename this parallelRegion node to 'activeParallelRegion'
children = parallelRegionNode.childNodes
existingActiveParallelRegions = routine.getElementsByTagName("activeParallelRegions")
newRegionNode = None
newRegionNodeNeedsAppending = False
if len(existingActiveParallelRegions) > 0:
newRegionNode = existingActiveParallelRegions[0]
else:
newRegionNode = doc.createElement("activeParallelRegions")
newRegionNodeNeedsAppending = True
for child in children:
newRegionNode.appendChild(child.cloneNode(deep=True))
routine.removeChild(parallelRegionNode)
if newRegionNodeNeedsAppending:
routine.appendChild(newRegionNode)
if appliesTo != "GPU":
continue
#check our kernel callers. Rule: Their direct callees should not be a kernel caller themselves.
kernelCallers = callNodesByCalleeName.get(routineName)
if not kernelCallers:
continue
for kernelCaller in kernelCallers:
kernelCallerName = kernelCaller.getAttribute("caller")
kernelWrapperCall = getFirstKernelCallerInCalleesOf(kernelCallerName, callNodesByCallerName, parallelRegionNodesByRoutineName)
if kernelWrapperCall != None:
kernelWrapperName = kernelWrapperCall.getAttribute("callee")
if not kernelCallerProblemFound:
logging.warning("Subroutine %s calls at least one kernel (%s) and at least one kernel wrapper (%s). \
This may cause device attribute mismatch compiler errors. In this case please wrap all kernels called by %s, such that it does not call a mix of kernel wrappers and kernels.\n"
%(kernelCallerName, routineName, kernelWrapperName, kernelCallerName)
)
kernelCallerProblemFound = True
messagesPresentedFor.append(kernelCallerName)
elif kernelCallerName not in messagesPresentedFor:
messagesPresentedFor.append(kernelCallerName)
logging.warning("...same for %s: calls kernel %s, kernel wrapper %s" %(kernelCallerName, routineName, kernelWrapperName))
