def writeBenchmarkFiles(stepBaseDir, solutions, problemSizes, stepName,
filesToCopy):
if not globalParameters["MergeFiles"]:
ensurePath(os.path.join(globalParameters["WorkingPath"], "Solutions"))
ensurePath(os.path.join(globalParameters["WorkingPath"], "Kernels"))
##############################################################################
# Min Naming
##############################################################################
kernels = []
kernelsBetaOnly = []
for solution in solutions:
solutionKernels = solution.getKernels()
for kernel in solutionKernels:
if kernel not in kernels:
kernels.append(kernel)
solutionKernelsBetaOnly = solution.getKernelsBetaOnly()
for kernel in solutionKernelsBetaOnly:
if kernel not in kernelsBetaOnly:
kernelsBetaOnly.append(kernel)
solutionSerialNaming = Solution.getSerialNaming(solutions)
kernelSerialNaming = Solution.getSerialNaming(kernels)
solutionMinNaming = Solution.getMinNaming(solutions)
kernelMinNaming = Solution.getMinNaming(kernels)
solutionWriter = SolutionWriter( \
solutionMinNaming, solutionSerialNaming, \
kernelMinNaming, kernelSerialNaming)
kernelWriterSource = KernelWriterSource( \
kernelMinNaming, kernelSerialNaming)
kernelWriterAssembly = KernelWriterAssembly( \
kernelMinNaming, kernelSerialNaming)
# write solution, kernels and CMake
problemType = solutions[0]["ProblemType"]
writeSolutionsAndKernels( \
globalParameters["WorkingPath"], [problemType], solutions, kernels, kernelsBetaOnly, \
solutionWriter, kernelWriterSource, kernelWriterAssembly )
##############################################################################
# Write CMake
##############################################################################
clientName = "TensileBenchmark_%s" % stepName
writeCMake(globalParameters["WorkingPath"], solutions, kernels, filesToCopy, \
clientName)
forBenchmark = True
writeClientParameters(forBenchmark, solutions, problemSizes, stepName, \
filesToCopy, stepBaseDir)
def writeCMake(outputPath, solutions, kernels, libraryStaticFiles, clientName):
print1("# Writing Custom CMake")
##############################################################################
# Min Naming
##############################################################################
if globalParameters["ShortNames"] and not globalParameters["MergeFiles"]:
solutionSerialNaming = Solution.getSerialNaming(solutions)
kernelSerialNaming = Solution.getSerialNaming(kernels)
else:
solutionSerialNaming = None
kernelSerialNaming = None
solutionMinNaming = Solution.getMinNaming(solutions)
kernelMinNaming = Solution.getMinNaming(kernels)
solutionWriter = SolutionWriter( \
solutionMinNaming, solutionSerialNaming, \
kernelMinNaming, kernelSerialNaming)
kernelWriterSource = KernelWriterSource( \
kernelMinNaming, kernelSerialNaming)
kernelWriterAssembly = KernelWriterAssembly( \
kernelMinNaming, kernelSerialNaming)
generatedFile = open(os.path.join(outputPath, "Generated.cmake"), "w")
generatedFile.write(CMakeHeader)
generatedFile.write("set( TensileClient_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 solution in solutions:
solutionName = solutionWriter.getSolutionName(solution)
generatedFile.write(" ${CMAKE_SOURCE_DIR}/Solutions/%s.h\n" \
% (solutionName) )
generatedFile.write(" ${CMAKE_SOURCE_DIR}/Solutions/%s.cpp\n" \
% (solutionName) )
generatedFile.write(" )\n")
# write kernel names
generatedFile.write("set( TensileClient_KERNELS\n")
if globalParameters["MergeFiles"]:
generatedFile.write(" ${CMAKE_SOURCE_DIR}/Kernels.h\n")
generatedFile.write(" ${CMAKE_SOURCE_DIR}/Kernels.cpp\n")
else:
for kernel in kernels:
kernelName = kernelWriterSource.getKernelName(kernel) if kernel[
"KernelLanguage"] == "Source" else kernelWriterAssembly.getKernelName(
kernel)
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( TensileClient_SOURCE\n")
for fileName in libraryStaticFiles:
# copy file
shutil_copy( os.path.join(globalParameters["SourcePath"], fileName), \
outputPath )
# add file to cmake
generatedFile.write(" ${CMAKE_SOURCE_DIR}/%s\n" % fileName)
generatedFile.write(" )\n\n")
# close generated cmake
generatedFile.close()
def writeClientParameters(forBenchmark, solutions, problemSizes, stepName, \
functionList):
h = ""
##############################################################################
# Min Naming
##############################################################################
if forBenchmark:
kernels = []
for solution in solutions:
solutionKernels = solution.getKernels()
for kernel in solutionKernels:
if kernel not in kernels:
kernels.append(kernel)
solutionSerialNaming = Solution.getSerialNaming(solutions)
kernelSerialNaming = Solution.getSerialNaming(kernels)
solutionMinNaming = Solution.getMinNaming(solutions)
kernelMinNaming = Solution.getMinNaming(kernels)
solutionWriter = SolutionWriter( \
solutionMinNaming, solutionSerialNaming, \
kernelMinNaming, kernelSerialNaming)
if forBenchmark:
if globalParameters["MergeFiles"]:
h += "#include \"Solutions.h\"\n"
else:
for solution in solutions:
solutionName = solutionWriter.getSolutionName(solution)
h += "#include \"" + solutionName + ".h\"\n"
h += "\n"
else:
h += "#include \"Tensile.h\"\n"
h += "typedef enum {\n"
h += " enum_float,\n"
h += " enum_double,\n"
h += " enum_TensileComplexFloat,\n"
h += " enum_TensileComplexDouble\n"
h += "#ifdef Tensile_ENABLE_HALF\n"
h += " ,enum_TensileHalf\n"
h += "#endif\n"
h += "} DataTypeEnum;\n"
h += "\n"
h += "// Debug Params\n"
h += "const bool printTensorA=%s;\n" % toCppBool(
globalParameters["PrintTensorA"])
h += "const bool printTensorB=%s;\n" % toCppBool(
globalParameters["PrintTensorB"])
h += "const bool printTensorC=%s;\n" % toCppBool(
globalParameters["PrintTensorC"])
h += "const bool printWinnersOnly=%s;\n" % toCppBool(
globalParameters["PrintWinnersOnly"])
h += "\n"
h += "const char indexChars[%u] = \"%s" \
% (len(globalParameters["IndexChars"])+1, \
globalParameters["IndexChars"][0])
for i in range(1, len(globalParameters["IndexChars"])):
h += globalParameters["IndexChars"][i]
h += "\";\n"
h += "unsigned int functionIdx;\n"
h += "unsigned int dataTypeIdx;\n"
h += "unsigned int problemTypeIdx;\n"
h += "\n"
##############################################################################
# Problem Types
##############################################################################
#dataTypes = []
#problemTypes = []
#functionSerialToDataTypeAndIdx = []
dataTypes = []
problemTypes = []
problemTypesForDataType = {} # for data type
schedulesForProblemType = {} # for problem type
functionInfo = [
] # dataTypeIdx, problemTypeIdx, idxWithinDataType, idxWithinProblemType
if forBenchmark:
problemType = solutions[0]["ProblemType"]
dataType = problemType["DataType"]
dataTypes.append(dataType)
problemTypes.append(problemType)
problemTypesForDataType[dataType] = [problemType]
schedulesForProblemType[problemType] = solutions
numProblemTypes = 1
for solution in solutions:
functionInfo.append([0, 0, 0, 0, 0, 0])
else:
for functionIdx in range(0, len(functionList)):
function = functionList[functionIdx]
scheduleName = function[0]
problemType = function[1]
dataType = problemType["DataType"]
if dataType not in dataTypes:
dataTypes.append(dataType)
problemTypesForDataType[dataType] = []
if problemType not in problemTypesForDataType[dataType]:
problemTypesForDataType[dataType].append(problemType)
schedulesForProblemType[problemType] = []
schedulesForProblemType[problemType].append(scheduleName)
# sort
dataTypes = sorted(dataTypes)
for dataType in dataTypes:
problemTypesForDataType[dataType] = \
sorted(problemTypesForDataType[dataType])
for problemType in problemTypesForDataType[dataType]:
schedulesForProblemType[problemType] = \
sorted(schedulesForProblemType[problemType])
# assign info
functionIdxSerial = 0
problemTypeIdxSerial = 0
for dataTypeIdxSerial in range(0, len(dataTypes)):
dataType = dataTypes[dataTypeIdxSerial]
functionIdxForDataType = 0
for problemTypeIdxForDataType in range(0, \
len(problemTypesForDataType[dataType])):
problemType = \
problemTypesForDataType[dataType][problemTypeIdxForDataType]
problemTypes.append(problemType)
functionIdxForProblemType = 0
for functionIdxForProblemType in range(0, \
len(schedulesForProblemType[problemType])):
functionInfo.append([ \
dataTypeIdxSerial, \
problemTypeIdxForDataType, \
problemTypeIdxSerial, \
functionIdxSerial,\
functionIdxForDataType,\
functionIdxForProblemType, \
])
functionIdxForProblemType += 1
functionIdxForDataType += 1
functionIdxSerial += 1
problemTypeIdxSerial += 1
numProblemTypes = problemTypeIdxSerial
numFunctions = functionIdxSerial
h += "const unsigned int numFunctions = %u;\n" % numFunctions
##############################################################################
# Data Types
##############################################################################
h += "/* data types */\n"
numDataTypes = len(dataTypes)
h += "const unsigned int numDataTypes = %u;\n" % numDataTypes
h += "const DataTypeEnum dataTypeEnums[numDataTypes] = { enum_%s" \
% dataTypes[0].toCpp()
for dataTypeIdx in range(1, numDataTypes):
h += ", enum_%s" % dataTypes[dataTypeIdx].toCpp()
h += " };\n"
# bytes per elements
h += "const unsigned int bytesPerElement[numDataTypes] = { %u" \
% (dataTypes[0].numBytes())
for dataTypeIdx in range(1, numDataTypes):
dataType = dataTypes[dataTypeIdx]
h += ", %u" % dataType.numBytes()
h += " };\n"
# flops per mac
h += "const unsigned int numFlopsPerMac[numDataTypes] = { %u" \
% (2 if dataTypes[0].isReal() else 8)
for dataTypeIdx in range(1, numDataTypes):
dataType = dataTypes[dataTypeIdx]
h += ", %u" % (2 if dataType.isReal() else 8)
h += " };\n"
for dataTypeIdx in range(0, numDataTypes):
h += "#define Tensile_DATA_TYPE_%s\n" \
% dataTypes[dataTypeIdx].toCpp().upper()
##############################################################################
# Problem Types
##############################################################################
h += "/* problem types */\n"
h += "const unsigned int numProblemTypes = %u;\n" % numProblemTypes
# Num C Indices
h += "const unsigned int numIndicesC[numProblemTypes] = { %u" \
% problemTypes[0]["NumIndicesC"]
for problemTypeIdx in range(1, numProblemTypes):
problemType = problemTypes[problemTypeIdx]
h += ", %u" % problemType["NumIndicesC"]
h += " };\n"
# Num AB Indices
maxNumIndicesAB = len(problemTypes[0]["IndexAssignmentsA"])
h += "const unsigned int numIndicesAB[numProblemTypes] = { %u" \
% len(problemTypes[0]["IndexAssignmentsA"])
for problemTypeIdx in range(1, numProblemTypes):
problemType = problemTypes[problemTypeIdx]
numIndicesAB = len(problemType["IndexAssignmentsA"])
h += ", %u" % numIndicesAB
maxNumIndicesAB = max(numIndicesAB, maxNumIndicesAB)
h += " };\n"
h += "const unsigned int maxNumIndicesAB = %u;\n" % maxNumIndicesAB
# Index Assignments A
h += "const unsigned int indexAssignmentsA[numProblemTypes][maxNumIndicesAB] = {\n"
for problemTypeIdx in range(0, numProblemTypes):
problemType = problemTypes[problemTypeIdx]
indices = problemType["IndexAssignmentsA"]
h += " { %u" % indices[0]
for i in range(1, maxNumIndicesAB):
if i < len(indices):
h += ", %u" % indices[i]
else:
h += ", static_cast<unsigned int>(-1)"
if problemTypeIdx < numProblemTypes - 1:
h += " },\n"
else:
h += " }\n"
h += "};\n"
# Index Assignments B
h += "const unsigned int indexAssignmentsB[numProblemTypes][maxNumIndicesAB] = {\n"
for problemTypeIdx in range(0, numProblemTypes):
problemType = problemTypes[problemTypeIdx]
indices = problemType["IndexAssignmentsB"]
h += " { %u" % indices[0]
for i in range(1, maxNumIndicesAB):
if i < len(indices):
h += ", %u" % indices[i]
else:
h += ", static_cast<unsigned int>(-1)"
if problemTypeIdx < numProblemTypes - 1:
h += " },\n"
else:
h += " }\n"
h += "};\n"
# beta
h += "bool useBeta[numProblemTypes] = { %s" \
% ("true" if problemTypes[0]["UseBeta"] else "false")
for problemTypeIdx in range(1, numProblemTypes):
problemType = problemTypes[problemTypeIdx]
h += ", %s" % ("true" if problemType["UseBeta"] else "false")
h += " };\n"
# Complex Conjugates
h += "const bool complexConjugateA[numProblemTypes] = { %s" \
% ("true" if problemTypes[0]["ComplexConjugateA"] else "false" )
for problemTypeIdx in range(1, numProblemTypes):
problemType = problemTypes[problemTypeIdx]
h += ", %s" % ("true"
if problemTypes[0]["ComplexConjugateA"] else "false")
h += " };\n"
h += "const bool complexConjugateB[numProblemTypes] = { %s" \
% ("true" if problemTypes[0]["ComplexConjugateB"] else "false" )
for problemTypeIdx in range(1, numProblemTypes):
problemType = problemTypes[problemTypeIdx]
h += ", %s" % ("true"
if problemTypes[0]["ComplexConjugateB"] else "false")
h += " };\n"
h += "\n"
if not forBenchmark:
h += "// dataTypeIdxSerial, problemTypeIdxForDataType, problemTypeIdxSerial, functionIdxSerial, functionIdxForDataType, functionIdxForProblemType\n"
first = True
h += "const unsigned int functionInfo[numFunctions][6] = {\n"
for info in functionInfo:
h += "%s{ %u, %u, %u, %u, %u, %u }" % (" " if first else ",\n ", \
info[0], info[1], info[2], info[3], info[4], info[5] )
first = False
h += " };\n"
##############################################################################
# Problem Sizes
##############################################################################
maxNumIndices = problemTypes[0]["TotalIndices"]
if not forBenchmark:
for problemType in problemTypes:
maxNumIndices = max(problemType["TotalIndices"], maxNumIndices)
h += "const unsigned int maxNumIndices = %u;\n" % maxNumIndices
h += "const unsigned int totalIndices[numProblemTypes] = { %u" \
% problemTypes[0]["TotalIndices"]
for problemTypeIdx in range(1, numProblemTypes):
h += ", %u" % problemTypes[problemTypeIdx]["TotalIndices"]
h += " };\n"
if forBenchmark:
h += "const unsigned int numProblems = %u;\n" \
% problemSizes.totalProblemSizes
h += "const unsigned int problemSizes[numProblems][%u] = {\n" \
% problemTypes[0]["TotalIndices"]
for i in range(0, problemSizes.totalProblemSizes):
line = " {%5u" % problemSizes.sizes[i][0]
for j in range(1, problemTypes[0]["TotalIndices"]):
line += ",%5u" % problemSizes.sizes[i][j]
line += " }"
h += line
if i < problemSizes.totalProblemSizes - 1:
h += ","
else:
h += "};"
h += "\n"
h += "const unsigned int minStrides[%u] = {" \
% problemTypes[0]["TotalIndices"]
for i in range(0, len(problemSizes.minStrides)):
if (i != 0):
h += ", "
h += str(problemSizes.minStrides[i])
h += "};\n"
else:
h += "unsigned int userSizes[maxNumIndices];\n"
h += "unsigned int minStrides[%u] = {" \
% maxNumIndices
for i in range(0, maxNumIndices):
if (i != 0):
h += ", "
h += str(0)
# always use 0 for minStrides in benchmark mode
h += "};\n"
if forBenchmark:
h += "/* problem sizes */\n"
"""
h += "const bool indexIsSized[maxNumIndices] = {"
for i in range(0, problemSizes.totalIndices):
h += " %s" % ("true" if problemSizes.indexIsSized[i] else "false")
if i < problemSizes.totalIndices-1:
h += ","
h += " };\n"
h += "const unsigned int numIndicesSized = %u;\n" \
% len(problemSizes.indicesSized)
h += "const unsigned int indicesSized[numIndicesSized][4] = {\n"
h += "// { min, stride, stride_incr, max }\n"
for i in range(0, len(problemSizes.indicesSized)):
r = problemSizes.indicesSized[i]
h += " { %u, %u, %u, %u }" % (r[0], r[1], r[2], r[3])
if i < len(problemSizes.indicesSized)-1:
h += ","
h += "\n"
h += " };\n"
numIndicesMapped = len(problemSizes.indicesMapped)
h += "const unsigned int numIndicesMapped = %u;\n" % numIndicesMapped
if numIndicesMapped > 0:
h += "#define Tensile_INDICES_MAPPED 1\n"
h += "const unsigned int indicesMapped[numIndicesMapped] = {"
for i in range(0, numIndicesMapped):
h += " %u" % problemSizes.indicesMapped[i]
if i < numIndicesMapped-1:
h += ","
h += " };\n"
else:
h += "#define Tensile_INDICES_MAPPED 0\n"
"""
##############################################################################
# Max Problem Sizes
##############################################################################
if forBenchmark:
h += "size_t maxSizeC = %u;\n" % (problemSizes.maxC)
h += "size_t maxSizeA = %u;\n" % (problemSizes.maxA)
h += "size_t maxSizeB = %u;\n" % (problemSizes.maxB)
h += "\n"
else:
h += "size_t maxSizeC;\n"
h += "size_t maxSizeA;\n"
h += "size_t maxSizeB;\n"
h += "\n"
##############################################################################
# Current Problem Size
##############################################################################
h += "/* current problem size */\n"
#h += "unsigned int fullSizes[maxNumIndices];\n"
#h += "unsigned int currentSizedIndexSizes[numIndicesSized];\n"
#h += "unsigned int currentSizedIndexIncrements[numIndicesSized];\n"
h += "\n"
##############################################################################
# Solutions
##############################################################################
if forBenchmark:
h += "/* solutions */\n"
# Problem Type Indices
h += "const unsigned int maxNumSolutions = %u;\n" % len(solutions)
h += "float solutionPerf[numProblems][maxNumSolutions]; // milliseconds\n"
h += "\n"
# Solution Ptrs
h += "typedef TensileStatus (*SolutionFunctionPointer)(\n"
argList = solutionWriter.getArgList(solutions[0]["ProblemType"], True,
True, True)
for i in range(0, len(argList)):
h += " %s %s%s" % (argList[i][0], argList[i][1], \
",\n" if i < len(argList)-1 else ");\n\n")
h += "const SolutionFunctionPointer solutions[maxNumSolutions] = {\n"
for i in range(0, len(solutions)):
solution = solutions[i]
solutionName = solutionWriter.getSolutionName(solution)
h += " %s" % solutionName
if i < len(solutions) - 1:
h += ","
h += "\n"
h += " };\n"
h += "\n"
# Solution Names
h += "const char *solutionNames[maxNumSolutions] = {\n"
for i in range(0, len(solutions)):
solution = solutions[i]
solutionName = solutionWriter.getSolutionName(solution)
h += " \"%s\"" % solutionName
if i < len(solutions) - 1:
h += ","
h += "\n"
h += " };\n"
h += "\n"
else:
# Function Names
functionNames = []
for dataType in dataTypes:
for problemType in problemTypesForDataType[dataType]:
for scheduleName in schedulesForProblemType[problemType]:
#functionNames.append("tensile_%s_%s" % (scheduleName, problemType))
functionNames.append("tensile_%s" % (problemType))
h += "const char *functionNames[numFunctions] = {\n"
for functionIdx in range(0, len(functionNames)):
functionName = functionNames[functionIdx]
h += " \"%s\"%s\n" % (functionName, \
"," if functionIdx < len(functionNames)-1 else "" )
h += " };\n"
##############################################################################
# Runtime Structures
##############################################################################
h += "/* runtime structures */\n"
h += "TensileStatus status;\n"
if globalParameters["RuntimeLanguage"] == "OCL":
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["Device"])
h += "\n"
h += "void *deviceC;\n"
h += "void *deviceA;\n"
h += "void *deviceB;\n"
##############################################################################
# Benchmarking and Validation Parameters
##############################################################################
h += "\n/* benchmarking parameters */\n"
#h += "const bool measureKernelTime = %s;\n" \
# % ("true" if globalParameters["KernelTime"] else "false")
#h += "const unsigned int numEnqueuesPerSync = %u;\n" \
# % (globalParameters["EnqueuesPerSync"])
#h += "const unsigned int numSyncsPerBenchmark = %u;\n" \
# % (globalParameters["SyncsPerBenchmark"])
#h += "unsigned int numElementsToValidate = %s;\n" \
# % (str(globalParameters["NumElementsToValidate"]) \
# if globalParameters["NumElementsToValidate"] >= 0 \
# else "0xFFFFFFFF" )
#h += "unsigned int validationMaxToPrint = %u;\n" \
# % globalParameters["ValidationMaxToPrint"]
#h += "bool validationPrintValids = %s;\n" \
# % ("true" if globalParameters["ValidationPrintValids"] else "false")
h += "size_t validationStride;\n"
if problemType["HighPrecisionAccumulate"]:
h += "static bool useHighPrecisionAccumulate = true;\n"
else:
h += "static bool useHighPrecisionAccumulate = false;\n"
#h += "unsigned int dataInitTypeC = %s;\n" % globalParameters["DataInitTypeC"]
#h += "unsigned int dataInitTypeAB = %s;\n" % globalParameters["DataInitTypeAB"]
h += "\n"
##############################################################################
# Generated Call to Reference
##############################################################################
h += "/* generated call to reference */\n"
h += "template<typename DataType>\n"
h += "TensileStatus generatedCallToReferenceCPU(\n"
h += " const unsigned int *sizes,\n"
h += " const unsigned int *minStrides,\n"
h += " DataType *referenceC,\n"
h += " DataType *initialA,\n"
h += " DataType *initialB,\n"
h += " const unsigned int stride_a,\n"
h += " const unsigned int stride_b,\n"
h += " const unsigned int stride_c,\n"
h += " DataType alpha,\n"
h += " DataType beta,\n"
h += " bool useHighPrecisionAccumulate) {\n"
h += " return tensileReferenceCPU(\n"
h += " referenceC,\n"
h += " initialA,\n"
h += " initialB,\n"
h += " stride_a,\n"
h += " stride_b,\n"
h += " stride_c,\n"
h += " alpha,\n"
h += " beta,\n"
h += " totalIndices[problemTypeIdx],\n"
h += " sizes,\n"
h += " minStrides,\n"
h += " numIndicesC[problemTypeIdx],\n"
h += " numIndicesAB[problemTypeIdx],\n"
h += " indexAssignmentsA[problemTypeIdx],\n"
h += " indexAssignmentsB[problemTypeIdx],\n"
h += " complexConjugateA[problemTypeIdx],\n"
h += " complexConjugateB[problemTypeIdx],\n"
h += " validationStride,\n"
h += " useHighPrecisionAccumulate);\n"
h += "};\n"
h += "\n"
##############################################################################
# Generated Call to Solution
##############################################################################
if forBenchmark:
problemType = solutions[0]["ProblemType"]
h += "/* generated call to solution */\n"
h += "template<typename DataType>\n"
h += "TensileStatus generatedCallToSolution(\n"
h += " unsigned int solutionIdx,\n"
h += " const unsigned int *sizes,\n"
h += " const unsigned int *minStrides,\n"
h += " DataType alpha,\n"
h += " DataType beta, \n"
h += " unsigned int numEvents = 0, \n"
if globalParameters["RuntimeLanguage"] == "OCL":
h += " cl_event *event_wait_list = NULL,\n"
h += " cl_event *outputEvent = NULL ) {\n"
else:
h += " hipEvent_t *startEvent = NULL,\n"
h += " hipEvent_t *stopEvent = NULL ) {\n"
h += " // calculate parameters assuming packed data\n"
# strides
indexChars = globalParameters["IndexChars"]
firstStride = 1
if problemType["UseInitialStrides"]:
firstStride = 0
lastStrideC = problemType["NumIndicesC"]
lastStrideA = len(problemType["IndexAssignmentsA"])
lastStrideB = len(problemType["IndexAssignmentsB"])
# calculate strides
for i in range(0, lastStrideC):
h += " unsigned int strideC%u%s = 1" % (i, indexChars[i])
for j in range(0, i):
h += "* std::max(minStrides[%i], sizes[%i])" % (j, j)
h += ";\n"
for i in range(0, lastStrideA):
h += " unsigned int strideA%u%s = 1" % (i, \
indexChars[problemType["IndexAssignmentsA"][i]])
for j in range(0, i):
h += "* std::max(minStrides[%i], sizes[%i])" % \
(problemType["IndexAssignmentsA"][j],
problemType["IndexAssignmentsA"][j])
h += ";\n"
for i in range(0, lastStrideB):
h += " unsigned int strideB%u%s = 1" % (i, \
indexChars[problemType["IndexAssignmentsB"][i]])
for j in range(0, i):
h += "* std::max(minStrides[%i], sizes[%i])" % \
(problemType["IndexAssignmentsB"][j],
problemType["IndexAssignmentsB"][j])
h += ";\n"
for i in range(0, problemType["TotalIndices"]):
h += " unsigned int size%s = sizes[%u];\n" % (indexChars[i], i)
h += "\n"
# function call
h += " // call solution function\n"
if globalParameters["RuntimeLanguage"] == "OCL":
h += " return solutions[solutionIdx]( static_cast<cl_mem>(deviceC), static_cast<cl_mem>(deviceA), static_cast<cl_mem>(deviceB),\n"
else:
typeName = dataTypes[0].toCpp()
h += " return solutions[solutionIdx]( static_cast<%s *>(deviceC), static_cast<%s *>(deviceA), static_cast<%s *>(deviceB),\n" \
% (typeName, typeName, typeName)
h += " alpha,\n"
if problemType["UseBeta"]:
h += " beta,\n"
h += " 0, 0, 0, // offsets\n"
for i in range(firstStride, lastStrideC):
h += " strideC%u%s,\n" % (i, indexChars[i])
for i in range(firstStride, lastStrideA):
h += " strideA%u%s,\n" % (i, \
indexChars[problemType["IndexAssignmentsA"][i]])
for i in range(firstStride, lastStrideB):
h += " strideB%u%s,\n" % (i, \
indexChars[problemType["IndexAssignmentsB"][i]])
for i in range(0, problemType["TotalIndices"]):
h += " size%s,\n" % indexChars[i]
h += " stream,\n"
if globalParameters["RuntimeLanguage"] == "OCL":
h += " numEvents, event_wait_list, outputEvent ); // events\n"
else:
h += " numEvents, startEvent, stopEvent); // events\n"
h += "};\n"
h += "\n"
else:
############################################################################
# Generated Call to Function
############################################################################
for enqueue in [True, False]:
functionName = "tensile" if enqueue else "tensileGetSolutionName"
returnName = "TensileStatus" if enqueue else "const char *"
h += "/* generated call to function */\n"
h += "template<typename DataType>\n"
h += "%s generatedCallTo_%s(\n" % (returnName, functionName)
h += " unsigned int *sizes,\n"
h += " unsigned int *minStrides,\n"
h += " DataType alpha,\n"
h += " DataType beta, \n"
h += " unsigned int strideA, \n"
h += " unsigned int strideB, \n"
h += " unsigned int strideC, \n"
h += " unsigned int numEvents = 0, \n"
if globalParameters["RuntimeLanguage"] == "OCL":
h += " cl_event *event_wait_list = NULL,\n"
h += " cl_event *outputEvent = NULL );\n\n"
else:
h += " hipEvent_t *startEvent = NULL,\n"
h += " hipEvent_t *stopEvent = NULL );\n\n"
for dataType in dataTypes:
typeName = dataType.toCpp()
functionsForDataType = []
for problemType in problemTypesForDataType[dataType]:
for scheduleName in schedulesForProblemType[problemType]:
functionsForDataType.append(
[scheduleName, problemType])
h += "template<>\n"
h += "inline %s generatedCallTo_%s<%s>(\n" \
% (returnName, functionName, typeName)
h += " unsigned int *sizes,\n"
h += " unsigned int *minStrides,\n"
h += " %s alpha,\n" % typeName
h += " %s beta,\n" % typeName
h += " unsigned int strideA, \n"
h += " unsigned int strideB, \n"
h += " unsigned int strideC, \n"
h += " unsigned int numEvents, \n"
if globalParameters["RuntimeLanguage"] == "OCL":
h += " cl_event *event_wait_list,\n"
h += " cl_event *outputEvent ) {\n\n"
else:
h += " hipEvent_t *startEvent,\n"
h += " hipEvent_t *stopEvent ) {\n\n"
h += " unsigned int functionIdxForDataType = functionInfo[functionIdx][4];\n"
for functionIdx in range(0, len(functionsForDataType)):
function = functionsForDataType[functionIdx]
scheduleName = function[0]
problemType = function[1]
if len(functionsForDataType) > 1:
if functionIdx == 0:
h += " if (functionIdxForDataType == %u) {\n" % functionIdx
elif functionIdx == len(functionsForDataType) - 1:
h += " } else {\n"
else:
h += " } else if (functionIdxForDataType == %u) {\n" \
% functionIdx
# strides
indexChars = globalParameters["IndexChars"]
firstStride = 1
if problemType["UseInitialStrides"]:
firstStride = 0
lastStrideC = problemType["NumIndicesC"]
lastStrideA = len(problemType["IndexAssignmentsA"])
lastStrideB = len(problemType["IndexAssignmentsB"])
# calculate strides
for i in range(0, lastStrideC):
h += " unsigned int strideC%u%s = 1" % (
i, indexChars[i])
for j in range(0, i):
h += "*sizes[%i]" % j
h += ";\n"
h += " if (strideC != std::numeric_limits<unsigned int>::max()) strideC%u%s = strideC;\n" % (
lastStrideC - 1, indexChars[lastStrideC - 1])
for i in range(0, lastStrideA):
h += " unsigned int strideA%u%s = 1" % (i, \
indexChars[problemType["IndexAssignmentsA"][i]])
for j in range(0, i):
h += "*sizes[%i]" % \
problemType["IndexAssignmentsA"][j]
h += ";\n"
h += " if (strideA != std::numeric_limits<unsigned int>::max()) strideA%u%s = strideA;\n" % (
lastStrideA - 1, indexChars[lastStrideA - 1])
for i in range(0, lastStrideB):
h += " unsigned int strideB%u%s = 1" % (i, \
indexChars[problemType["IndexAssignmentsB"][i]])
for j in range(0, i):
h += "*sizes[%i]" % \
problemType["IndexAssignmentsB"][j]
h += ";\n"
h += " if (strideB != std::numeric_limits<unsigned int>::max()) strideB%u%s = strideB;\n" % (
lastStrideB - 1, indexChars[lastStrideB - 1])
for i in range(0, problemType["TotalIndices"]):
h += " unsigned int size%s = sizes[%u];\n" % (
indexChars[i], i)
# function call
h += " // call solution function\n"
h += " return %s_%s(\n" % (functionName, problemType)
if enqueue:
if globalParameters["RuntimeLanguage"] == "OCL":
h += " static_cast<cl_mem>(deviceC),\n"
h += " static_cast<cl_mem>(deviceA),\n"
h += " static_cast<cl_mem>(deviceB),\n"
else:
h += " static_cast<%s *>(deviceC),\n" % typeName
h += " static_cast<%s *>(deviceA),\n" % typeName
h += " static_cast<%s *>(deviceB),\n" % typeName
h += " alpha,\n"
if problemType["UseBeta"]:
h += " beta,\n"
h += " 0, 0, 0, // offsets\n"
for i in range(firstStride, lastStrideC):
h += " strideC%u%s,\n" % (i, indexChars[i])
for i in range(firstStride, lastStrideA):
h += " strideA%u%s,\n" % (i, \
indexChars[problemType["IndexAssignmentsA"][i]])
for i in range(firstStride, lastStrideB):
h += " strideB%u%s,\n" % (i, \
indexChars[problemType["IndexAssignmentsB"][i]])
for i in range(0, problemType["TotalIndices"]):
h += " size%s,\n" % indexChars[i]
h += " stream"
if enqueue:
if globalParameters["RuntimeLanguage"] == "OCL":
h += ",\n numEvents, event_wait_list, outputEvent"
else:
h += ",\n numEvents, startEvent, stopEvent"
h += ");\n"
if len(functionsForDataType) > 1:
h += " }\n" # close last if
h += "};\n" # close callToFunction
##############################################################################
# Results File Name
##############################################################################
if forBenchmark:
h += "/* results file name */\n"
resultsFileName = os.path.join(globalParameters["WorkingPath"], \
"../../Data","%s.csv" % stepName)
resultsFileName = resultsFileName.replace("\\", "\\\\")
h += "const char *resultsFileName = \"%s\";\n" % resultsFileName
##############################################################################
# Write File
##############################################################################
clientParametersFile = open(os.path.join(globalParameters["WorkingPath"], \
"ClientParameters.h"), "w")
clientParametersFile.write(CHeader)
clientParametersFile.write(h)
clientParametersFile.close()
def TensileCreateLibrary():
print1("")
print1(HR)
print1("# Tensile Create Library")
print2(HR)
print2("")
##############################################################################
# Parse Command Line Arguments
##############################################################################
print2("Arguments: %s" % sys.argv)
argParser = argparse.ArgumentParser()
argParser.add_argument("LogicPath",
help="Path to LibraryLogic.yaml files.")
argParser.add_argument("OutputPath", help="Where to write library files?")
argParser.add_argument("RuntimeLanguage", help="Which runtime language?", \
choices=["OCL", "HIP", "HSA"])
argParser.add_argument("--merge-files", dest="MergeFiles", \
action="store_true")
argParser.add_argument("--no-merge-files", dest="MergeFiles", \
action="store_false")
argParser.add_argument("--short-file-names", dest="ShortNames", \
action="store_true")
argParser.add_argument("--no-short-file-names", dest="ShortNames", \
action="store_false")
argParser.add_argument("--library-print-debug", dest="LibraryPrintDebug", \
action="store_true")
argParser.add_argument("--no-library-print-debug", dest="LibraryPrintDebug", \
action="store_false")
args = argParser.parse_args()
logicPath = args.LogicPath
outputPath = args.OutputPath
print2("OutputPath: %s" % outputPath)
ensurePath(outputPath)
arguments = {}
arguments["RuntimeLanguage"] = args.RuntimeLanguage
arguments["MergeFiles"] = args.MergeFiles
arguments["ShortNames"] = args.ShortNames
arguments["LibraryPrintDebug"] = args.LibraryPrintDebug
arguments["CodeFromFiles"] = False
assignGlobalParameters(arguments)
if not os.path.exists(logicPath):
printExit("LogicPath %s doesn't exist" % logicPath)
logicFiles = [os.path.join(logicPath, f) for f in os.listdir(logicPath) \
if (os.path.isfile(os.path.join(logicPath, f)) \
and os.path.splitext(f)[1]==".yaml")]
print1("# LibraryLogicFiles:" % logicFiles)
for logicFile in logicFiles:
print1("# %s" % logicFile)
##############################################################################
# Parse config files
##############################################################################
solutions = []
logicData = {} # keys are problemTypes, values are schedules
for logicFileName in logicFiles:
(scheduleName, deviceNames, problemType, solutionsForSchedule, \
indexOrder, exactLogic, rangeLogic) \
= YAMLIO.readLibraryLogicForSchedule(logicFileName)
if problemType not in logicData:
logicData[problemType] = []
logicData[problemType].append((scheduleName, deviceNames, \
solutionsForSchedule, indexOrder, exactLogic, rangeLogic ))
for solution in solutionsForSchedule:
if solution not in solutions:
solutions.append(solution)
# create solution writer and kernel writer
kernels = []
kernelsBetaOnly = []
for solution in solutions:
solutionKernels = solution.getKernels()
for kernel in solutionKernels:
if kernel not in kernels:
kernels.append(kernel)
solutionKernelsBetaOnly = solution.getKernelsBetaOnly()
for kernel in solutionKernelsBetaOnly:
if kernel not in kernelsBetaOnly:
kernelsBetaOnly.append(kernel)
# if any kernels are assembly, append every ISA supported
if globalParameters["ShortNames"] and not globalParameters["MergeFiles"]:
solutionSerialNaming = Solution.getSerialNaming(solutions)
kernelSerialNaming = Solution.getSerialNaming(kernels)
else:
solutionSerialNaming = None
kernelSerialNaming = None
solutionMinNaming = Solution.getMinNaming(solutions)
kernelMinNaming = Solution.getMinNaming(kernels)
solutionWriter = SolutionWriter( \
solutionMinNaming, solutionSerialNaming, \
kernelMinNaming, kernelSerialNaming)
kernelWriterSource = KernelWriterSource( \
kernelMinNaming, kernelSerialNaming)
kernelWriterAssembly = KernelWriterAssembly( \
kernelMinNaming, kernelSerialNaming)
# write solutions and kernels
writeSolutionsAndKernels(outputPath, solutions, kernels, kernelsBetaOnly, \
solutionWriter, kernelWriterSource, kernelWriterAssembly)
libraryStaticFiles = [
"SolutionMapper.h", "TensileTypes.h", "KernelHeader.h",
"SolutionHelper.cpp", "SolutionHelper.h", "Tools.cpp", "Tools.h"
]
# write cmake
clientName = "LibraryClient"
writeCMake(outputPath, solutions, kernels, libraryStaticFiles, clientName)
# write logic
writeLogic(outputPath, logicData, solutionWriter)
print1("# Tensile Library Writer DONE")
print1(HR)
print1("")
def TensileCreateLibrary():
print1("")
print1(HR)
print1("# Tensile Create Library")
print2(HR)
print2("")
##############################################################################
# Parse Command Line Arguments
##############################################################################
print2("Arguments: %s" % sys.argv)
argParser = argparse.ArgumentParser()
argParser.add_argument("LogicPath",
help="Path to LibraryLogic.yaml files.")
argParser.add_argument("OutputPath", help="Where to write library files?")
argParser.add_argument("RuntimeLanguage", help="Which runtime language?", \
choices=["OCL", "HIP", "HSA"])
argParser.add_argument("--merge-files", dest="MergeFiles", \
action="store_true")
argParser.add_argument("--no-merge-files", dest="MergeFiles", \
action="store_false")
argParser.add_argument("--short-file-names", dest="ShortNames", \
action="store_true")
argParser.add_argument("--no-short-file-names", dest="ShortNames", \
action="store_false")
argParser.add_argument("--library-print-debug", dest="LibraryPrintDebug", \
action="store_true")
argParser.add_argument("--no-library-print-debug", dest="LibraryPrintDebug", \
action="store_false")
argParser.add_argument(
"--isa",
dest="isa",
action="append",
help="which architectures for assembly kernels to target")
args = argParser.parse_args()
logicPath = args.LogicPath
outputPath = args.OutputPath
print2("OutputPath: %s" % outputPath)
ensurePath(outputPath)
arguments = {}
arguments["RuntimeLanguage"] = args.RuntimeLanguage
arguments["MergeFiles"] = args.MergeFiles
arguments["ShortNames"] = args.ShortNames
arguments["LibraryPrintDebug"] = args.LibraryPrintDebug
if args.isa:
newISA = []
for isa in args.isa:
gfxIdx = isa.find("gfx")
if gfxIdx >= 0:
major = int(isa[gfxIdx + 3:gfxIdx + 4])
minor = int(isa[gfxIdx + 4:gfxIdx + 5])
step = int(isa[gfxIdx + 5:gfxIdx + 6])
isaTuple = (major, minor, step)
if isaTuple in globalParameters[
"SupportedISA"] and isaTuple not in newISA:
print1("# User-Specified ISA: gfx%u%u%u" %
(major, minor, step))
newISA.append(isaTuple)
else:
printWarning("isa parameter must be formed as: --isa gfx803")
arguments["SupportedISA"] = newISA
assignGlobalParameters(arguments)
if not os.path.exists(logicPath):
printExit("LogicPath %s doesn't exist" % logicPath)
logicFiles = [os.path.join(logicPath, f) for f in os.listdir(logicPath) \
if (os.path.isfile(os.path.join(logicPath, f)) \
and os.path.splitext(f)[1]==".yaml")]
print1("# LibraryLogicFiles:" % logicFiles)
for logicFile in logicFiles:
print1("# %s" % logicFile)
##############################################################################
# Parse config files
##############################################################################
solutions = []
logicData = {} # keys are problemTypes, values are schedules
for logicFileName in logicFiles:
(scheduleName, deviceNames, problemType, solutionsForSchedule, \
indexOrder, exactLogic, rangeLogic) \
= YAMLIO.readLibraryLogicForSchedule(logicFileName)
if problemType not in logicData:
logicData[problemType] = []
logicData[problemType].append((scheduleName, deviceNames, \
solutionsForSchedule, indexOrder, exactLogic, rangeLogic ))
for solution in solutionsForSchedule:
if solution not in solutions:
solutions.append(solution)
# create solution writer and kernel writer
kernels = []
kernelsBetaOnly = []
for solution in solutions:
solutionKernels = solution.getKernels()
for kernel in solutionKernels:
if kernel not in kernels:
kernels.append(kernel)
solutionKernelsBetaOnly = solution.getKernelsBetaOnly()
for kernel in solutionKernelsBetaOnly:
if kernel not in kernelsBetaOnly:
kernelsBetaOnly.append(kernel)
# if any kernels are assembly, append every ISA supported
if globalParameters["RuntimeLanguage"] == "HIP":
newKernels = []
for kernel in kernels:
if kernel["KernelLanguage"] == "Assembly":
kernel["ISA"] = globalParameters["SupportedISA"][0]
for i in range(1, len(globalParameters["SupportedISA"])):
newKernel = deepcopy(kernel)
newKernel["ISA"] = globalParameters["SupportedISA"][i]
newKernels.append(newKernel)
else:
kernel["ISA"] = (0, 0, 0)
kernels.extend(newKernels)
if globalParameters["ShortNames"] and not globalParameters["MergeFiles"]:
solutionSerialNaming = Solution.getSerialNaming(solutions)
kernelSerialNaming = Solution.getSerialNaming(kernels)
else:
solutionSerialNaming = None
kernelSerialNaming = None
solutionMinNaming = Solution.getMinNaming(solutions)
kernelMinNaming = Solution.getMinNaming(kernels)
solutionWriter = SolutionWriter( \
solutionMinNaming, solutionSerialNaming, \
kernelMinNaming, kernelSerialNaming)
kernelWriterSource = KernelWriterSource( \
kernelMinNaming, kernelSerialNaming)
kernelWriterAssembly = KernelWriterAssembly( \
kernelMinNaming, kernelSerialNaming)
# write solutions and kernels
writeSolutionsAndKernels(outputPath, solutions, kernels, kernelsBetaOnly, \
solutionWriter, kernelWriterSource, kernelWriterAssembly)
libraryStaticFiles = [
"TensileTypes.h", "KernelHeader.h", "SolutionHelper.cpp",
"SolutionHelper.h", "Tools.cpp", "Tools.h"
]
# write cmake
clientName = "LibraryClient"
writeCMake(outputPath, solutions, kernels, libraryStaticFiles, clientName)
# write logic
writeLogic(outputPath, logicData, solutionWriter)
print1("# Tensile Library Writer DONE")
print1(HR)
print1("")
