Merge pull request #11 from kilikkuo/master

Merge from kilikkuo's fork

Author: Kilik Kuo

oclConfigurar.py

@@ -13,6 +13,7 @@ def __init__(self):
         self.dicPlatform2Devices = {}
 
         self.context = None
+        self.curDevice = None
         self.queue = None
         self.program = None
         self.mem_pool = None
@@ -42,10 +43,10 @@ def setupProgramAndDataStructure(self, program, lstIPath=[], dicName2DS={}):
         #                                                  ('open', numpy.float32)]}
         assert (self.context != None), "Setup context seems incorrectly !!"
         assert (len(self.context.devices) > 0), "Error, No device for context !!"
-        device = self.context.devices[0]
+        self.curDevice = self.context.devices[0]
         dicReturnStruct = {}
         for k, v in dicName2DS.iteritems():
-            kObj, k_c_decl = cl.tools.match_dtype_to_c_struct(device, k, v)
+            kObj, k_c_decl = cl.tools.match_dtype_to_c_struct(self.curDevice, k, v)
             retV = cl.tools.get_or_register_dtype(k, kObj)
 
             dicReturnStruct[k] = retV
@@ -64,8 +65,15 @@ def setupProgramAndDataStructure(self, program, lstIPath=[], dicName2DS={}):
     def callFuncFromProgram(self, strMethodName, *args, **argd):
         methodCall = getattr(self.program, strMethodName)
         if methodCall:
+            if len(args) >= 2 and type(args[1])==tuple and (not args[1]) != True:
+                wgs = cl.Kernel(self.program, strMethodName).get_work_group_info(
+                    cl.kernel_work_group_info.WORK_GROUP_SIZE, self.curDevice)
+                local_worksize = reduce(lambda x,y: x*y, args[1])
+                print 'local size : ', local_worksize
+                assert wgs >= local_worksize, 'Out of capability, please reduce the local work size for %s()'%(strMethodName)
             evt = methodCall(self.queue, *args)
             return evt
+        return None
 
     def getContext(self, device=PREFERRED_GPU):
         assert len(self.dicIdx2Platform) > 0, 'No platform for OCL operation'
@@ -104,8 +112,8 @@ def createOCLArrayEmpty(self, stDType, size):
         assert size > 0, "Can NOT create array size <= 0"
         assert (self.queue != None), " Make sure setup correctly"
         # Creat a list which contains element initialized with structure stDType
-        npArrData = np.zeros(size, dtype=stDType, allocator=self.mem_pool)
-        clArrData = cl.array.to_device(self.queue, npArrData)
+        npArrData = np.zeros(size, dtype=stDType)
+        clArrData = cl.array.to_device(self.queue, npArrData, allocator=self.mem_pool)
         return clArrData
 
     def createOCLArrayForInput(self, stDType, lstData):
@@ -114,6 +122,6 @@ def createOCLArrayForInput(self, stDType, lstData):
         assert len(lstData) > 0, "Size of input data list = 0"
         assert (self.queue != None), " Make sure setup correctly"
 
-        arrayData = np.array(lstData, dtype=stDType, allocator=self.mem_pool)
-        clArrayData = cl.array.to_device(self.queue, arrayData)
+        arrayData = np.array(lstData, dtype=stDType)
+        clArrayData = cl.array.to_device(self.queue, arrayData, allocator=self.mem_pool)
         return clArrayData

oclPyObjGenerator.py

@@ -46,7 +46,8 @@ def getMethodDef(strMethodName, dicArgd={}, nIndent=0):
     return head
 
 def prepareImport(nIndent=0):
-    lstImport = ['import os', 'import numpy', 'from oclConfigurar import OCLConfigurar, PREFERRED_GPU']
+    lstImport = ['import os', 'import numpy', 'from oclConfigurar import OCLConfigurar, PREFERRED_GPU',
+                 'from threading import Lock']
     strResult = ''
     for item in lstImport:
         strResult += item + os.linesep
@@ -61,6 +62,11 @@ def prepareNumpyDS(dicDS={}, nIndent=0):
         strTemp += os.linesep
     return strTemp
 
+def prepareLock(nIndent=0):
+    strTemp = ''
+    strTemp += tabSpace * nIndent + 'lk = Lock()' + os.linesep
+    return strTemp
+
 def prepareOCLConfigurar(nIndent=0):
     return tabSpace*nIndent + "self.oclConfigurar = OCLConfigurar()" + os.linesep
 
@@ -72,7 +78,11 @@ def prepareOCLSetup(clsName, dicNumpyDS={}, nIndent=0):
     test = space + 'self.oclConfigurar.setupContextAndQueue(PREFERRED_GPU)' + os.linesep
     test += space + 'dirPath = os.path.dirname(__file__)' + os.linesep
     test += space + 'path = os.path.join(dirPath, %s)'%('\''+str(getKernelFileName(clsName))+'\'') + os.linesep
-    test += space + 'dicRetDS = self.oclConfigurar.setupProgramAndDataStructure(path, [dirPath], {%s})'%(strDS) + os.linesep
+    test += space + '%s.lk.acquire()'%(clsName) + os.linesep
+    test += space + 'try:' + os.linesep
+    test += space + tabSpace + 'dicRetDS = self.oclConfigurar.setupProgramAndDataStructure(path, [dirPath], {%s})'%(strDS) + os.linesep
+    test += space + 'finally:' + os.linesep
+    test += space + tabSpace + '%s.lk.release()'%(clsName) + os.linesep
     for k in dicNumpyDS.iterkeys():
         test += space + 'self.%s = dicRetDS.get(%s, None)'%(k, '\''+k+'\'') + os.linesep
     return test + os.linesep
@@ -93,7 +103,8 @@ def prepareOCLFuncImpl(strMethodName, dicArgd={}, nIndent=0):
     if strRepresArgs == '':
         return '' + os.linesep
     strRepresArgs = strRepresArgs[:-2]
-    body = tabSpace*nIndent + 'self.oclConfigurar.callFuncFromProgram(%s, %s)'%('\''+strMethodName+'\'', strRepresArgs) + os.linesep
+    body = tabSpace*nIndent + 'evt = self.oclConfigurar.callFuncFromProgram(%s, %s)'%('\''+strMethodName+'\'', strRepresArgs) + os.linesep
+    body += tabSpace*nIndent + 'return evt' + os.linesep
     return body
 
 def prepareOCLBufferByDS(nIndent=0):
@@ -156,6 +167,7 @@ def generateOCLPyObj(self):
 
             # class
             fPyObj.write(getClassDef(self.className))
+            fPyObj.write(prepareLock(1))
             fPyObj.write(prepareNumpyDS(self.dicNumpyDS, 1))
 
             # __init__