Merge pull request #73 from guacamoleo/develop

fixes critical bug in LibraryClient for multiple problem types

.gitignore

@@ -6,3 +6,4 @@
 Tensile.egg-info
 build
 dist
+.cache

Tensile/BenchmarkProblems.py

@@ -37,7 +37,6 @@ def benchmarkProblemType( config ):
 
   totalBenchmarkSteps = len(benchmarkProcess)
   winners = WinningParameterDict()
-  determinedParameters = [{}] # winner chosen from benchmark
   print1("NumBenchmarkSteps: %u" % totalBenchmarkSteps)
   print1("")
   print1(HR)
@@ -56,7 +55,7 @@ def benchmarkProblemType( config ):
     numHardcoded = len(benchmarkStep.hardcodedParameters)
     stepName = str(benchmarkStep)
     shortName = benchmarkStep.abbreviation()
-    print1("\n\n")
+    print1("\n")
     print1(HR)
     print1("# %s\n# %s" % (problemTypeName, stepName))
     print1("# NumProblems: %u" % benchmarkStep.problemSizes.totalProblemSizes)
@@ -126,7 +125,6 @@ def benchmarkProblemType( config ):
     # Enumerate Benchmark Permutations
     ############################################################################
     solutions = []
-    currentSolution = {"ProblemType": deepcopy(benchmarkProcess.problemType.state) }
     totalBenchmarkPermutations = 1
     for benchmarkParamName in benchmarkStep.benchmarkParameters:
       totalBenchmarkPermutations *= len(benchmarkStep.benchmarkParameters[benchmarkParamName])
@@ -213,14 +211,12 @@ def benchmarkProblemType( config ):
       for j in range(0, len(solutionsForHardcoded)):
         solution = solutionsForHardcoded[j]
         solutionList.append(solution)
-    solutionsMinNaming = Solution.getMinNaming(solutionList)
     if globalParameters["PrintLevel"] >= 1:
       for i in range(0, len(solutions)):
         solutionsForHardcoded = solutions[i]
         for j in range(0, len(solutionsForHardcoded)):
           solution = solutionsForHardcoded[j]
           print2("#    (%u:%u) %s" % (i, j, \
-              #Solution.getNameMin(solution, solutionsMinNaming) ))
               Solution.getNameFull(solution) ))
       print2(HR)
 
@@ -252,7 +248,9 @@ def benchmarkProblemType( config ):
 
       # run runScript
       process = Popen(runScriptName, cwd=globalParameters["WorkingPath"])
-      status = process.communicate()
+      process.communicate()
+      if process.returncode:
+        printWarning("Benchmark Process exited with code %u" % process.returncode)
       popWorkingPath() # build
 
 
@@ -270,10 +268,6 @@ def benchmarkProblemType( config ):
     YAMLIO.writeSolutions(solutionsFileName, benchmarkStep.problemSizes, \
         solutions )
 
-    #solutionsFromFile = YAMLIO.readSolutions(solutionYAMLFileName)
-    #solutionsMinNaming = Solution.getMinNaming(solutionsFromFile)
-    #for solution in solutionsFromFile:
-
     # End Iteration
     popWorkingPath() # stepName
     print1("%s\n# %s\n# %s: End\n%s\n" \
@@ -318,7 +312,6 @@ def getResults(resultsFileName, solutions):
         printWarning("CSV File %s row %u doesn't have %u elements; ignoring remainer of file." \
             % (resultsFileName, rowIdx, rowLength) )
         break
-      totalFlops = float(row[problemSizeIdx])
       idx = startIdx
       #for i in range(0, len(numBenchmarksPerHardcoded)):
       #  for j in range(0, numBenchmarksPerHardcoded[i]):
@@ -327,9 +320,6 @@ def getResults(resultsFileName, solutions):
         for j in range(0, len(solutionsForHardcoded)):
           solution = solutionsForHardcoded[j]
           gflops = float(row[idx])
-          #time_ms = float(row[idx])
-          #flops = totalFlops / (time_ms / 1000)
-          #gflops = flops / (1000*1000*1000)
           results[i][j].append(gflops)
           idx += 1
   if rowIdx < 2:
@@ -349,8 +339,6 @@ def writeBenchmarkFiles(solutions, problemSizes, stepName, filesToCopy):
   ##############################################################################
   # Min Naming
   ##############################################################################
-  solutionFileNames = []
-  kernelNames = []
   kernels = []
   for solution in solutions:
     solutionKernels = solution.getKernels()
@@ -379,52 +367,6 @@ def writeBenchmarkFiles(solutions, problemSizes, stepName, filesToCopy):
 
   clientName = "TensileBenchmark_%s" % stepName
   writeCMake(globalParameters["WorkingPath"], solutions, filesToCopy, clientName)
-  """
-  generatedFile = open(os.path.join(globalParameters["WorkingPath"], \
-      "Generated.cmake"), "w")
-  generatedFile.write(CMakeHeader)
-  generatedFile.write("set( TensileBenchmark_Solutions\n")
-  # write solution names
-  if globalParameters["MergeFiles"]:
-    generatedFile.write("  ${CMAKE_SOURCE_DIR}/Solutions.h\n")
-    generatedFile.write("  ${CMAKE_SOURCE_DIR}/Solutions.cpp\n")
-  else:
-    for solutionFileName in solutionFileNames:
-      generatedFile.write("  ${CMAKE_SOURCE_DIR}/Solutions/%s.h\n" \
-          % (solutionFileName) )
-      generatedFile.write("  ${CMAKE_SOURCE_DIR}/Solutions/%s.cpp\n" \
-          % (solutionFileName) )
-  generatedFile.write("  )\n")
-
-  # write kernel names
-  generatedFile.write("set( TensileBenchmark_Kernels\n")
-  if globalParameters["MergeFiles"]:
-    generatedFile.write("  ${CMAKE_SOURCE_DIR}/Kernels.h\n")
-    generatedFile.write("  ${CMAKE_SOURCE_DIR}/Kernels.cpp\n")
-  else:
-    for kernelName in kernelNames:
-      generatedFile.write("  ${CMAKE_SOURCE_DIR}/Kernels/%s.h\n" % (kernelName))
-      generatedFile.write("  ${CMAKE_SOURCE_DIR}/Kernels/%s.cpp\n" % kernelName)
-  generatedFile.write("  )\n")
-
-  generatedFile.write("set( TensileBenchmark_Source\n")
-  for fileName in filesToCopy:
-    generatedFile.write("  ${CMAKE_SOURCE_DIR}/%s\n" % fileName)
-  generatedFile.write("  )\n\n")
-
-  # benchmark parameters
-  generatedFile.write("set( ClientName TensileBenchmark_%s)\n" \
-      % (stepName) )
-  generatedFile.write("set( Tensile_BACKEND \"%s\")\n" \
-      % (globalParameters["Backend"]) )
-
-  # build parameters
-  generatedFile.write("set( CMAKE_BUILD_TYPE \"%s\")\n" \
-      % (globalParameters["CMakeBuildType"]) )
-
-  # close generated cmake
-  generatedFile.close()
-  """
 
   forBenchmark = True
   writeClientParameters(forBenchmark, solutions, problemSizes, stepName, filesToCopy)
@@ -492,8 +434,8 @@ def addResults( self, hardcodedParameterList, benchmarkPermutations, \
       if len(matches) != 1:
         printExit("Didn't find exactly 1 match")
       hardcodedParametersKey = matches[0][0]
-      oldWinningParameters = matches[0][1]
-      oldScore = matches[0][2]
+      #oldWinningParameters = matches[0][1]
+      #oldScore = matches[0][2]
       self.winners[hardcodedParametersKey][0].update(winningParameters)
       self.winners[hardcodedParametersKey][1] = winningScore
 
@@ -508,7 +450,6 @@ def __getitem__( self, hardcodedParameters ):
     elif len(matches) == 0:
       return None
     else:
-      happy += 1
       printExit("Didn't find exactly 1 match")
 
   ##########################################################
@@ -536,7 +477,6 @@ def update(self, newHardcodedParameterList ):
           self.winners[FrozenDictionary(newHardcodedParameters)] = \
               [ winningParameters, score ]
         elif len(matches) > 1: # join
-          fastestIdx = -1
           fastestScore = -1
           fastestHardcodedParameters = {}
           fastestWinningParameters = {}
@@ -547,7 +487,6 @@ def update(self, newHardcodedParameterList ):
             score = match[2]
             if score > fastestScore:
               fastestScore = score
-              fastestIdx = matchIdx
               fastestWinningParameters = winningParameters
               fastestHardcodedParameters = hardcodedFrozen.parameters
           newHardcodedParameters.update(fastestHardcodedParameters)

Tensile/BenchmarkStructs.py

@@ -620,7 +620,7 @@ def __str__(self):
     else:
       string += "_Final"
     return string
-  def __repr__():
+  def __repr__(self):
     return self.__str__()
 
 

Tensile/ClientWriter.py

@@ -1,7 +1,6 @@
-from Common import globalParameters, HR, pushWorkingPath, popWorkingPath, print1, CHeader
+from Common import globalParameters, HR, pushWorkingPath, popWorkingPath, print1, CHeader, printExit
 from SolutionStructs import Solution
 from SolutionWriter import SolutionWriter
-from KernelWriter import KernelWriter
 import YAMLIO
 
 import os
@@ -35,7 +34,6 @@ def main( config ):
       ]
 
   for f in filesToCopy:
-    filename = os.path.join(globalParameters["SourcePath"], f)
     shutil_copy(
         os.path.join(globalParameters["SourcePath"], f),
         globalParameters["WorkingPath"] )
@@ -93,7 +91,9 @@ def main( config ):
 
   # run runScript
   process = Popen(runScriptName, cwd=globalParameters["WorkingPath"])
-  status = process.communicate()
+  process.communicate()
+  if process.returncode:
+    printWarning("Benchmark Process exited with code %u" % process.returncode)
   popWorkingPath() # build
 
   popWorkingPath() # LibraryClient
@@ -165,7 +165,6 @@ def writeClientParameters(forBenchmark, solutions, problemSizes, stepName, \
   # Min Naming
   ##############################################################################
   if forBenchmark:
-    kernelNames = []
     kernels = []
     for solution in solutions:
       solutionKernels = solution.getKernels()
@@ -180,20 +179,13 @@ def writeClientParameters(forBenchmark, solutions, problemSizes, stepName, \
     solutionWriter = SolutionWriter( \
         solutionMinNaming, solutionSerialNaming, \
         kernelMinNaming, kernelSerialNaming)
-    kernelWriter = KernelWriter( \
-        kernelMinNaming, kernelSerialNaming)
 
   if forBenchmark:
     if globalParameters["MergeFiles"]:
       h += "#include \"Solutions.h\"\n"
     else:
       for solution in solutions:
         solutionName = solutionWriter.getSolutionName(solution)
-        #if globalParameters["ShortNames"]:
-        #  solutionFileName = \
-        #      Solution.getNameSerial(solution, solutionSerialNaming)
-        #else:
-        #  solutionFileName = Solution.getNameMin(solution, solutionMinNaming)
         h += "#include \"" + solutionName + ".h\"\n"
     h += "\n"
   else:
@@ -279,8 +271,6 @@ def writeClientParameters(forBenchmark, solutions, problemSizes, stepName, \
         functionIdxForProblemType = 0
         for functionIdxForProblemType in range(0, \
             len(schedulesForProblemType[problemType])):
-          schedule = \
-              schedulesForProblemType[problemType][functionIdxForProblemType]
           functionInfo.append([ \
               dataTypeIdxSerial, \
               problemTypeIdxForDataType, \
@@ -292,7 +282,7 @@ def writeClientParameters(forBenchmark, solutions, problemSizes, stepName, \
           functionIdxForProblemType += 1
           functionIdxForDataType += 1
           functionIdxSerial += 1
-      problemTypeIdxSerial += 1
+        problemTypeIdxSerial += 1
     numProblemTypes = problemTypeIdxSerial
     numFunctions = functionIdxSerial
     h += "const unsigned int numFunctions = %u;\n" % numFunctions
@@ -360,7 +350,7 @@ def writeClientParameters(forBenchmark, solutions, problemSizes, stepName, \
       if i < len(indices):
         h += ", %u" % indices[i]
       else:
-        h += ", -1"
+        h += ", static_cast<unsigned int>(-1)"
     if problemTypeIdx < numProblemTypes-1:
       h += " },\n"
     else:
@@ -376,7 +366,7 @@ def writeClientParameters(forBenchmark, solutions, problemSizes, stepName, \
       if i < len(indices):
         h += ", %u" % indices[i]
       else:
-        h += ", -1"
+        h += ", static_cast<unsigned int>(-1)"
     if problemTypeIdx < numProblemTypes-1:
       h += " },\n"
     else:
@@ -557,17 +547,17 @@ def writeClientParameters(forBenchmark, solutions, problemSizes, stepName, \
   h += "TensileStatus status;\n"
   if globalParameters["Backend"] == "OCL":
     h += "unsigned int platformIdx = %u;\n" \
-        % (globalParameters["PlatformIdx"])
+        % (globalParameters["Platform"])
     h += "unsigned int deviceIdx = %u;\n" \
-        % (globalParameters["DeviceIdx"])
+        % (globalParameters["Device"])
     h += "cl_platform_id platform;\n"
     h += "cl_device_id device;\n"
     h += "cl_context context;\n"
     h += "cl_command_queue stream;\n"
   else:
     h += "hipStream_t stream;\n"
     h += "int deviceIdx = %u;\n" \
-        % (globalParameters["DeviceIdx"])
+        % (globalParameters["Device"])
   h += "\n"
   h += "void *deviceC;\n"
   h += "void *deviceA;\n"

Tensile/Common.py

@@ -30,8 +30,8 @@
 globalParameters["LibraryLogicPath"] = "3_LibraryLogic"
 globalParameters["LibraryClientPath"] = "4_LibraryClient"
 # device
-globalParameters["PlatformIdx"] = 0
-globalParameters["DeviceIdx"] = 0
+globalParameters["Platform"] = 0
+globalParameters["Device"] = 0
 # benchmark behavior
 globalParameters["CMakeBuildType"] = "Release" # Debug
 globalParameters["ForceRedoBenchmarkProblems"] = True

Tensile/Configs/rocblas_cgemm.yaml

@@ -1,4 +1,4 @@
-Parameters:
+GlobalParameters:
   Name: Fiji
   PrintLevel: 1
   ForceRedoBenchmarkProblems: False
@@ -13,8 +13,8 @@ Parameters:
   ValidationPrintValids: False
   ShortNames: False
   MergeFiles: True
-  PlatformIdx: 0
-  DeviceIdx: 0
+  Platform: 0
+  Device: 0
   DataInitType: 0
 
 BenchmarkProblems:
@@ -43,11 +43,10 @@ BenchmarkProblems:
       - ThreadTileEdge: [2, 4, 6, 8]
       - NumLoadsCoalescedA: [ 1, 2, 3, 4, 6, 8 ]
       - NumLoadsCoalescedB: [ 1, 2, 3, 4, 6, 8 ]
-    BenchmarkForkParameters:
       - LoopUnroll: [8, 16]
+    BenchmarkForkParameters:
     JoinParameters:
       - MacroTile
-      - DepthU
     BenchmarkJoinParameters:
     BenchmarkFinalParameters:
       - ProblemSizes: [ [32, 32, 32, 4000], [32, 32, 32, 4000], [2], [1536] ]
@@ -75,13 +74,12 @@ BenchmarkProblems:
       - WorkGroupShape: [ 0 ]
       - ThreadTileEdge: [2, 4, 6, 8]
       - ThreadTileShape: [ 0 ]
-    BenchmarkForkParameters:
       - NumLoadsCoalescedA: [ 1, 2, 3, 4, 6, 8 ]
       - NumLoadsCoalescedB: [ 1, 2, 3, 4, 6, 8 ]
       - LoopUnroll: [8, 16]
+    BenchmarkForkParameters:
     JoinParameters:
       - MacroTile
-      - DepthU
     BenchmarkJoinParameters:
     BenchmarkFinalParameters:
       - ProblemSizes: [ [32, 32, 32, 4000], [32, 32, 32, 4000], [2], [1536] ]
@@ -109,13 +107,12 @@ BenchmarkProblems:
       - WorkGroupShape: [ 0 ]
       - ThreadTileEdge: [2, 4, 6, 8]
       - ThreadTileShape: [ 0 ]
-    BenchmarkForkParameters:
       - NumLoadsCoalescedA: [ 1, 2, 3, 4, 6, 8 ]
       - NumLoadsCoalescedB: [ 1, 2, 3, 4, 6, 8 ]
       - LoopUnroll: [8, 16]
+    BenchmarkForkParameters:
     JoinParameters:
       - MacroTile
-      - DepthU
     BenchmarkJoinParameters:
     BenchmarkFinalParameters:
       - ProblemSizes: [ [32, 32, 32, 4000], [32, 32, 32, 4000], [2], [1536] ]
@@ -143,13 +140,12 @@ BenchmarkProblems:
       - WorkGroupShape: [ 0 ]
       - ThreadTileEdge: [2, 4, 6, 8]
       - ThreadTileShape: [ 0 ]
-    BenchmarkForkParameters:
       - NumLoadsCoalescedA: [ 1, 2, 3, 4, 6, 8 ]
       - NumLoadsCoalescedB: [ 1, 2, 3, 4, 6, 8 ]
       - LoopUnroll: [8, 16]
+    BenchmarkForkParameters:
     JoinParameters:
       - MacroTile
-      - DepthU
     BenchmarkJoinParameters:
     BenchmarkFinalParameters:
       - ProblemSizes: [ [32, 32, 32, 4000], [32, 32, 32, 4000], [2], [1536] ]