def writeIncludes():
print "AutoGemm.py: Generating include files."
if not os.path.exists( Common.getIncludePath() ):
os.makedirs( Common.getIncludePath() )
kernelSourceIncludes = KernelSourceIncludes()
kernelBinaryIncludes = KernelBinaryIncludes()
clKernelIncludes = ClKernelIncludes()
kernelSourceBuildOptions = KernelSourceBuildOptions()
kernelBinaryBuildOptions = KernelBinaryBuildOptions()
cppKernelEnumeration = CppKernelEnumeration()
# for each precision
kernel = KernelParameters.KernelParameters()
for precision in AutoGemmParameters.precisions:
kernel.precision = precision
cppKernelEnumeration.newPrecision(precision)
# valid tiles for this precision
tiles = AutoGemmParameters.getTilesForPrecision(precision)
# add tiles for this precision to Cpp
for tile in tiles:
cppKernelEnumeration.addTile(tile)
# for non tile parameters
for order in AutoGemmParameters.orders:
kernel.order = order
for transA in AutoGemmParameters.transposes[precision]:
kernel.transA = transA
for transB in AutoGemmParameters.transposes[precision]:
kernel.transB = transB
for beta in AutoGemmParameters.betas:
kernel.beta = beta
# add this nonTile combo for this precision to Cpp
cppKernelEnumeration.addNonTile(kernel)
# for tile parameters
for tile in tiles:
kernel.useTile(tile)
kernelSourceIncludes.addKernel(kernel)
kernelBinaryIncludes.addKernel(kernel)
kernelSourceBuildOptions.addKernel(kernel)
kernelBinaryBuildOptions.addKernel(kernel)
clKernelIncludes.addKernel(kernel)
cppKernelEnumeration.addKernel(kernel)
# save written files
kernelSourceIncludes.writeToFile()
kernelBinaryIncludes.writeToFile()
clKernelIncludes.writeToFile()
kernelSourceBuildOptions.writeToFile()
kernelBinaryBuildOptions.writeToFile()
cppKernelEnumeration.writeToFile()
def writeOfflineCompilation(args):
print("AutoGemm.py: Generating list of kernels to pre-compile.")
if not os.path.exists(Common.getIncludePath()):
os.makedirs(Common.getIncludePath())
ocFileName = Common.getIncludePath() + "AutoGemmKernelsToPreCompile.h"
ocFile = open(ocFileName, "w")
ocFile.write(Common.getAutoGemmHeader())
fileStr = "\n/*precision, order, transA, transB, beta, tileNumRows, tileNumCols, unroll*/\n"
fileStr += "\nunsigned int gemmPreCompile[][8] = {\n"
count = 0
for precision in args.precisions:
ocFile.write(fileStr)
fileStr = ""
validTiles = AutoGemmParameters.getTilesForPrecision(precision)
for order in args.orders:
for transpose in args.transposes:
transA = transpose[0]
transB = transpose[1]
if (transA == "C" or transB == "C") and (precision == "s"
or precision == "d"):
# real precision doesn't have conjugate transpose
continue
for beta in args.betas:
for tile in validTiles:
# print combination
kernelStr = " { %1u, %1u, %1u, %1u, %1u, %3u, %3u, %2u },\n" \
% (
Common.precisionInt[precision],
Common.orderInt[order],
Common.transposeInt[transA],
Common.transposeInt[transB],
beta,
tile.macroTileNumRows,
tile.macroTileNumCols,
tile.unroll
)
fileStr += kernelStr
#print kernelStr
count += 1
if count is 0:
fileStr += " { %1u, %1u, %1u, %1u, %1u, %3u, %3u, %2u },\n" \
% ( 0, 0, 0, 0, 0, 0, 0, 0 )
fileStr += "};\n"
fileStr += "unsigned int gemmPreCompileNum = " + str(count) + ";\n"
ocFile.write(fileStr)
ocFile.close()
count *= 4
print("AutoGemm.py: %u kernels will be pre-compiled." % count)
def writeOfflineCompilation(args):
print("AutoGemm.py: Generating list of kernels to pre-compile.")
if not os.path.exists( Common.getIncludePath() ):
os.makedirs( Common.getIncludePath() )
ocFileName = Common.getIncludePath() + "AutoGemmKernelsToPreCompile.h"
ocFile = open(ocFileName, "w")
ocFile.write( Common.getAutoGemmHeader() )
fileStr = "\n/*precision, order, transA, transB, beta, tileNumRows, tileNumCols, unroll*/\n"
fileStr += "\nunsigned int gemmPreCompile[][8] = {\n"
count = 0
for precision in args.precisions:
ocFile.write( fileStr )
fileStr = ""
validTiles = AutoGemmParameters.getTilesForPrecision(precision)
for order in args.orders:
for transpose in args.transposes:
transA = transpose[0]
transB = transpose[1]
if (transA=="C" or transB=="C") and (precision=="s" or precision=="d"):
# real precision doesn't have conjugate transpose
continue
for beta in args.betas:
for tile in validTiles:
# print combination
kernelStr = " { %1u, %1u, %1u, %1u, %1u, %3u, %3u, %2u },\n" \
% (
Common.precisionInt[precision],
Common.orderInt[order],
Common.transposeInt[transA],
Common.transposeInt[transB],
beta,
tile.macroTileNumRows,
tile.macroTileNumCols,
tile.unroll
)
fileStr += kernelStr
#print kernelStr
count+=1
if count is 0:
fileStr += " { %1u, %1u, %1u, %1u, %1u, %3u, %3u, %2u },\n" \
% ( 0, 0, 0, 0, 0, 0, 0, 0 )
fileStr += "};\n"
fileStr += "unsigned int gemmPreCompileNum = " + str(count) + ";\n"
ocFile.write( fileStr )
ocFile.close()
count *= 4
print("AutoGemm.py: %u kernels will be pre-compiled." % count)
def writeKernelSelection():
print("AutoGemm.py: Generating kernel selection.")
if not os.path.exists(Common.getIncludePath()):
os.makedirs(Common.getIncludePath())
########################################
# kernel selection specific
kss = KernelSelectionSpecific()
# for each precision
kernel = KernelParameters.KernelParameters()
for precision in AutoGemmParameters.precisions:
kernel.precision = precision
kss.newPrecision(precision)
# valid tiles for this precision
tiles = AutoGemmParameters.getTilesForPrecision(precision)
# for non tile parameters
for order in AutoGemmParameters.orders:
kernel.order = order
kss.newOrder(order)
for transA in AutoGemmParameters.transposes[precision]:
kernel.transA = transA
for transB in AutoGemmParameters.transposes[precision]:
kernel.transB = transB
kss.newTrans(transA, transB)
for beta in AutoGemmParameters.betas:
kernel.beta = beta
kss.newBeta(beta)
# for tile parameters
for tile in tiles:
kernel.useTile(tile)
kss.newKernel(kernel)
kss.writeToFile()
########################################
# kernel selection
ks = KernelSelection( \
AutoGemmParameters.precisions, \
AutoGemmParameters.orders, \
AutoGemmParameters.transposes, \
AutoGemmParameters.betas, \
AutoGemmParameters.unrolls, \
AutoGemmParameters.kernelSelectionData )
ks.writeToFile()
def writeKernelSelection():
print("AutoGemm.py: Generating kernel selection.")
if not os.path.exists( Common.getIncludePath() ):
os.makedirs( Common.getIncludePath() )
########################################
# kernel selection specific
kss = KernelSelectionSpecific()
# for each precision
kernel = KernelParameters.KernelParameters()
for precision in AutoGemmParameters.precisions:
kernel.precision = precision
kss.newPrecision(precision)
# valid tiles for this precision
tiles = AutoGemmParameters.getTilesForPrecision(precision)
# for non tile parameters
for order in AutoGemmParameters.orders:
kernel.order = order
kss.newOrder(order)
for transA in AutoGemmParameters.transposes[precision]:
kernel.transA = transA
for transB in AutoGemmParameters.transposes[precision]:
kernel.transB = transB
kss.newTrans(transA, transB)
for beta in AutoGemmParameters.betas:
kernel.beta = beta
kss.newBeta(beta)
# for tile parameters
for tile in tiles:
kernel.useTile(tile)
kss.newKernel(kernel)
kss.writeToFile()
########################################
# kernel selection
ks = KernelSelection( \
AutoGemmParameters.precisions, \
AutoGemmParameters.orders, \
AutoGemmParameters.transposes, \
AutoGemmParameters.betas, \
AutoGemmParameters.unrolls, \
AutoGemmParameters.kernelSelectionData )
ks.writeToFile()
def writeOpenCLKernels():
if not os.path.exists( Common.getKernelSourcePath() ):
os.makedirs( Common.getKernelSourcePath() )
if not os.path.exists( Common.getKernelBinaryPath() ):
os.makedirs( Common.getKernelBinaryPath() )
numKernels = 0
# for each precision
kernel = KernelParameters.KernelParameters()
for precision in AutoGemmParameters.precisions:
kernel.precision = precision
# valid tiles for this precision
tiles = AutoGemmParameters.getTilesForPrecision(precision)
# for non tile parameters
for order in AutoGemmParameters.orders:
kernel.order = order
for transA in AutoGemmParameters.transposes[precision]:
kernel.transA = transA
for transB in AutoGemmParameters.transposes[precision]:
kernel.transB = transB
for beta in AutoGemmParameters.betas:
kernel.beta = beta
# for tile parameters
for tile in tiles:
# tile kernel
kernel.useTile(tile)
writeOpenCLKernelToFile(kernel)
# row kernel
rowKernel = copy.copy(kernel)
rowKernel.macroTileNumRows = 1
writeOpenCLKernelToFile(rowKernel)
# col kernel
colKernel = copy.copy(kernel)
colKernel.macroTileNumCols = 1
writeOpenCLKernelToFile(colKernel)
# corner kernel
cornerKernel = copy.copy(kernel)
cornerKernel.macroTileNumRows = 1
cornerKernel.macroTileNumCols = 1
writeOpenCLKernelToFile(cornerKernel)
numKernels += 4
print("AutoGemm.py: generated %d kernels" % numKernels)
def writeOpenCLKernels():
if not os.path.exists(Common.getKernelSourcePath()):
os.makedirs(Common.getKernelSourcePath())
if not os.path.exists(Common.getKernelBinaryPath()):
os.makedirs(Common.getKernelBinaryPath())
numKernels = 0
# for each precision
kernel = KernelParameters.KernelParameters()
for precision in AutoGemmParameters.precisions:
kernel.precision = precision
# valid tiles for this precision
tiles = AutoGemmParameters.getTilesForPrecision(precision)
# for non tile parameters
for order in AutoGemmParameters.orders:
kernel.order = order
for transA in AutoGemmParameters.transposes[precision]:
kernel.transA = transA
for transB in AutoGemmParameters.transposes[precision]:
kernel.transB = transB
for beta in AutoGemmParameters.betas:
kernel.beta = beta
# for tile parameters
for tile in tiles:
# tile kernel
kernel.useTile(tile)
writeOpenCLKernelToFile(kernel)
# row kernel
rowKernel = copy.copy(kernel)
rowKernel.macroTileNumRows = 1
writeOpenCLKernelToFile(rowKernel)
# col kernel
colKernel = copy.copy(kernel)
colKernel.macroTileNumCols = 1
writeOpenCLKernelToFile(colKernel)
# corner kernel
cornerKernel = copy.copy(kernel)
cornerKernel.macroTileNumRows = 1
cornerKernel.macroTileNumCols = 1
writeOpenCLKernelToFile(cornerKernel)
numKernels += 4
print("AutoGemm.py: generated %d kernels" % numKernels)
