updated mark_nvidia.py script

pull/13/merge
Vladislav Vinogradov 13 years ago
parent 2634dc6cce
commit 00546f24e0
  1. 359
      modules/gpu/misc/mark_nvidia.py

@ -1,255 +1,234 @@
import sys, re
spaces = '[\s]*'
symbols = '[\s\w\d,.=:|]*'
symbols = '[\s\w\d,.:|]*'
def pattern1(prefix, test):
return re.compile(spaces + 'perf::' + prefix + '/' + test + '::' + '\(' + symbols + '\)' + spaces)
return re.compile(spaces + prefix + '_' + test + '::' + symbols + '::' + '\(' + symbols + '\)' + spaces)
def pattern2(prefix, test, cvtype):
return re.compile(spaces + 'perf::' + prefix + '/' + test + '::' + '\(' + symbols + cvtype + symbols + '\)' + spaces)
def pattern2(prefix, test, param1):
return re.compile(spaces + prefix + '_' + test + '::' + symbols + '::' + '\(' + symbols + param1 + symbols + '\)' + spaces)
def pattern3(prefix, test, cvtype, param1):
return re.compile(spaces + 'perf::' + prefix + '/' + test + '::' + '\(' + symbols + cvtype + symbols + param1 + symbols + '\)' + spaces)
def pattern3(prefix, test, param1, param2):
return re.compile(spaces + prefix + '_' + test + '::' + symbols + '::' + '\(' + symbols + param1 + symbols + param2 + symbols + '\)' + spaces)
def pattern4(prefix, test, cvtype, param1, param2):
return re.compile(spaces + 'perf::' + prefix + '/' + test + '::' + '\(' + symbols + cvtype + symbols + param1 + symbols + param2 + symbols + '\)' + spaces)
def pattern4(prefix, test, param1, param2, param3):
return re.compile(spaces + prefix + '_' + test + '::' + symbols + '::' + '\(' + symbols + param1 + symbols + param2 + symbols + param3 + symbols + '\)' + spaces)
def pattern5(prefix, test, param1, param2, param3, param5):
return re.compile(spaces + prefix + '_' + test + '::' + symbols + '::' + '\(' + symbols + param1 + symbols + param2 + symbols + param3 + symbols + param4 + symbols + '\)' + spaces)
npp_patterns = [
##############################################################
# Core
# Core/Add_Mat (CV_8U | CV_16U | CV_32F)
pattern2('Core', 'Add_Mat', '8U'),
pattern2('Core', 'Add_Mat', '16U'),
pattern2('Core', 'Add_Mat', '32F'),
# Core/Add_Scalar (CV_8U | CV_16U | CV_32F)
pattern2('Core', 'Add_Scalar', '8U'),
pattern2('Core', 'Add_Scalar', '16U'),
pattern2('Core', 'Add_Scalar', '32F'),
# Core/Subtract_Mat (CV_8U | CV_16U | CV_32F)
pattern2('Core', 'Subtract_Mat', '8U'),
pattern2('Core', 'Subtract_Mat', '16U'),
pattern2('Core', 'Subtract_Mat', '32F'),
# Core/Subtract_Scalar (CV_8U | CV_16U | CV_32F)
pattern2('Core', 'Subtract_Scalar', '8U'),
pattern2('Core', 'Subtract_Scalar', '16U'),
pattern2('Core', 'Subtract_Scalar', '32F'),
# Core/Multiply_Mat (CV_8U | CV_16U | CV_32F)
pattern2('Core', 'Multiply_Mat', '8U'),
pattern2('Core', 'Multiply_Mat', '16U'),
pattern2('Core', 'Multiply_Mat', '32F'),
# Core/Multiply_Scalar (CV_8U | CV_16U | CV_32F)
pattern2('Core', 'Multiply_Scalar', '8U'),
pattern2('Core', 'Multiply_Scalar', '16U'),
pattern2('Core', 'Multiply_Scalar', '32F'),
# Core/Divide_Mat (CV_8U | CV_16U | CV_32F)
pattern2('Core', 'Divide_Mat', '8U'),
pattern2('Core', 'Divide_Mat', '16U'),
pattern2('Core', 'Divide_Mat', '32F'),
# Core/Divide_Scalar (CV_8U | CV_16U | CV_32F)
pattern2('Core', 'Divide_Scalar', '8U'),
pattern2('Core', 'Divide_Scalar', '16U'),
pattern2('Core', 'Divide_Scalar', '32F'),
# Core/AbsDiff_Mat (CV_8U | CV_16U | CV_32F)
pattern2('Core', 'AbsDiff_Mat', '8U'),
pattern2('Core', 'AbsDiff_Mat', '16U'),
pattern2('Core', 'AbsDiff_Mat', '32F'),
# Core/AbsDiff_Scalar (CV_8U | CV_16U | CV_32F)
pattern2('Core', 'AbsDiff_Scalar', '8U'),
pattern2('Core', 'AbsDiff_Scalar', '16U'),
pattern2('Core', 'AbsDiff_Scalar', '32F'),
# Core/Abs
# Core_AddMat (CV_8U | CV_16U | CV_32F)
pattern2('Core', 'AddMat', '8U'),
pattern2('Core', 'AddMat', '16U'),
pattern2('Core', 'AddMat', '32F'),
# Core_AddScalar (CV_8U | CV_16U | CV_32F)
pattern2('Core', 'AddScalar', '8U'),
pattern2('Core', 'AddScalar', '16U'),
pattern2('Core', 'AddScalar', '32F'),
# Core_SubtractMat (CV_8U | CV_16U | CV_32F)
pattern2('Core', 'SubtractMat', '8U'),
pattern2('Core', 'SubtractMat', '16U'),
pattern2('Core', 'SubtractMat', '32F'),
# Core_SubtractScalar (CV_8U | CV_16U | CV_32F)
pattern2('Core', 'SubtractScalar', '8U'),
pattern2('Core', 'SubtractScalar', '16U'),
pattern2('Core', 'SubtractScalar', '32F'),
# Core_MultiplyMat (CV_8U | CV_16U | CV_32F)
pattern2('Core', 'MultiplyMat', '8U'),
pattern2('Core', 'MultiplyMat', '16U'),
pattern2('Core', 'MultiplyMat', '32F'),
# Core_MultiplyScalar (CV_8U | CV_16U | CV_32F)
pattern2('Core', 'MultiplyScalar', '8U'),
pattern2('Core', 'MultiplyScalar', '16U'),
pattern2('Core', 'MultiplyScalar', '32F'),
# Core_DivideMat (CV_8U | CV_16U | CV_32F)
pattern2('Core', 'DivideMat', '8U'),
pattern2('Core', 'DivideMat', '16U'),
pattern2('Core', 'DivideMat', '32F'),
# Core_Divide_Scalar (CV_8U | CV_16U | CV_32F)
pattern2('Core', 'DivideScalar', '8U'),
pattern2('Core', 'DivideScalar', '16U'),
pattern2('Core', 'DivideScalar', '32F'),
# Core_AbsDiff_Mat (CV_8U | CV_16U | CV_32F)
pattern2('Core', 'AbsDiffMat', '8U'),
pattern2('Core', 'AbsDiffMat', '16U'),
pattern2('Core', 'AbsDiffMat', '32F'),
# Core_AbsDiffScalar (CV_8U | CV_16U | CV_32F)
pattern2('Core', 'AbsDiffScalar', '8U'),
pattern2('Core', 'AbsDiffScalar', '16U'),
pattern2('Core', 'AbsDiffScalar', '32F'),
# Core_Abs
pattern1('Core', 'Abs'),
# Core/Sqr
# Core_Sqr
pattern1('Core', 'Sqr'),
# Core/Sqrt
# Core_Sqrt
pattern1('Core', 'Sqrt'),
# Core/Log
# Core_Log
pattern1('Core', 'Log'),
# Core/Exp
# Core_Exp
pattern1('Core', 'Exp'),
# Core/Bitwise_And_Scalar
pattern1('Core', 'Bitwise_And_Scalar'),
# Core_BitwiseAndScalar
pattern1('Core', 'BitwiseAndScalar'),
# Core/Bitwise_Or_Scalar
pattern1('Core', 'Bitwise_Or_Scalar'),
# Core_BitwiseOrScalar
pattern1('Core', 'BitwiseOrScalar'),
# Core/Bitwise_Xor_Scalar
pattern1('Core', 'Bitwise_Xor_Scalar'),
# Core_BitwiseXorScalar
pattern1('Core', 'BitwiseXorScalar'),
# Core/RShift
# Core_RShift
pattern1('Core', 'RShift'),
# Core/LShift
# Core_LShift
pattern1('Core', 'LShift'),
# Core/Transpose
# Core_Transpose
pattern1('Core', 'Transpose'),
# Core/Flip
# Core_Flip
pattern1('Core', 'Flip'),
# Core/LUT_OneChannel
pattern1('Core', 'LUT_OneChannel'),
# Core_LutOneChannel
pattern1('Core', 'LutOneChannel'),
# Core/LUT_MultiChannel
pattern1('Core', 'LUT_MultiChannel'),
# Core_LutMultiChannel
pattern1('Core', 'LutMultiChannel'),
# Core/Magnitude_Complex
pattern1('Core', 'Magnitude_Complex'),
# Core_MagnitudeComplex
pattern1('Core', 'MagnitudeComplex'),
# Core/Magnitude_Sqr_Complex
pattern1('Core', 'Magnitude_Sqr_Complex'),
# Core_MagnitudeSqrComplex
pattern1('Core', 'MagnitudeSqrComplex'),
# Core/MeanStdDev
# Core_MeanStdDev
pattern1('Core', 'MeanStdDev'),
# Core/NormDiff
# Core_NormDiff
pattern1('Core', 'NormDiff'),
##############################################################
# Filters
# Filters/Blur
# Filters_Blur
pattern1('Filters', 'Blur'),
# Filters/Erode
# Filters_Erode
pattern1('Filters', 'Erode'),
# Filters/Dilate
# Filters_Dilate
pattern1('Filters', 'Dilate'),
# Filters/MorphologyEx
# Filters_MorphologyEx
pattern1('Filters', 'MorphologyEx'),
##############################################################
# ImgProc
# ImgProc/Resize (8UC1 | 8UC4, INTER_NEAREST | INTER_LINEAR)
pattern3('ImgProc', 'Resize', '8UC1', 'INTER_NEAREST'),
pattern3('ImgProc', 'Resize', '8UC4', 'INTER_NEAREST'),
pattern3('ImgProc', 'Resize', '8UC1', 'INTER_LINEAR'),
pattern3('ImgProc', 'Resize', '8UC4', 'INTER_LINEAR'),
# ImgProc/Resize (8UC4, INTER_CUBIC)
pattern3('ImgProc', 'Resize', '8UC4', 'INTER_CUBIC'),
# ImgProc/WarpAffine (8UC1 | 8UC3 | 8UC4 | 32FC1 | 32FC3 | 32FC4, INTER_NEAREST | INTER_LINEAR | INTER_CUBIC, BORDER_CONSTANT)
pattern4('ImgProc', 'WarpAffine', '8UC1', 'INTER_NEAREST', 'BORDER_CONSTANT'),
pattern4('ImgProc', 'WarpAffine', '8UC1', 'INTER_LINEAR', 'BORDER_CONSTANT'),
pattern4('ImgProc', 'WarpAffine', '8UC1', 'INTER_CUBIC', 'BORDER_CONSTANT'),
pattern4('ImgProc', 'WarpAffine', '8UC3', 'INTER_NEAREST', 'BORDER_CONSTANT'),
pattern4('ImgProc', 'WarpAffine', '8UC3', 'INTER_LINEAR', 'BORDER_CONSTANT'),
pattern4('ImgProc', 'WarpAffine', '8UC3', 'INTER_CUBIC', 'BORDER_CONSTANT'),
pattern4('ImgProc', 'WarpAffine', '8UC4', 'INTER_NEAREST', 'BORDER_CONSTANT'),
pattern4('ImgProc', 'WarpAffine', '8UC4', 'INTER_LINEAR', 'BORDER_CONSTANT'),
pattern4('ImgProc', 'WarpAffine', '8UC4', 'INTER_CUBIC', 'BORDER_CONSTANT'),
pattern4('ImgProc', 'WarpAffine', '32FC1', 'INTER_NEAREST', 'BORDER_CONSTANT'),
pattern4('ImgProc', 'WarpAffine', '32FC1', 'INTER_LINEAR', 'BORDER_CONSTANT'),
pattern4('ImgProc', 'WarpAffine', '32FC1', 'INTER_CUBIC', 'BORDER_CONSTANT'),
pattern4('ImgProc', 'WarpAffine', '32FC3', 'INTER_NEAREST', 'BORDER_CONSTANT'),
pattern4('ImgProc', 'WarpAffine', '32FC3', 'INTER_LINEAR', 'BORDER_CONSTANT'),
pattern4('ImgProc', 'WarpAffine', '32FC3', 'INTER_CUBIC', 'BORDER_CONSTANT'),
pattern4('ImgProc', 'WarpAffine', '32FC4', 'INTER_NEAREST', 'BORDER_CONSTANT'),
pattern4('ImgProc', 'WarpAffine', '32FC4', 'INTER_LINEAR', 'BORDER_CONSTANT'),
pattern4('ImgProc', 'WarpAffine', '32FC4', 'INTER_CUBIC', 'BORDER_CONSTANT'),
# ImgProc/WarpPerspective (8UC1 | 8UC3 | 8UC4 | 32FC1 | 32FC3 | 32FC4, INTER_NEAREST | INTER_LINEAR | INTER_CUBIC, BORDER_CONSTANT)
pattern4('ImgProc', 'WarpPerspective', '8UC1', 'INTER_NEAREST', 'BORDER_CONSTANT'),
pattern4('ImgProc', 'WarpPerspective', '8UC1', 'INTER_LINEAR', 'BORDER_CONSTANT'),
pattern4('ImgProc', 'WarpPerspective', '8UC1', 'INTER_CUBIC', 'BORDER_CONSTANT'),
pattern4('ImgProc', 'WarpPerspective', '8UC3', 'INTER_NEAREST', 'BORDER_CONSTANT'),
pattern4('ImgProc', 'WarpPerspective', '8UC3', 'INTER_LINEAR', 'BORDER_CONSTANT'),
pattern4('ImgProc', 'WarpPerspective', '8UC3', 'INTER_CUBIC', 'BORDER_CONSTANT'),
pattern4('ImgProc', 'WarpPerspective', '8UC4', 'INTER_NEAREST', 'BORDER_CONSTANT'),
pattern4('ImgProc', 'WarpPerspective', '8UC4', 'INTER_LINEAR', 'BORDER_CONSTANT'),
pattern4('ImgProc', 'WarpPerspective', '8UC4', 'INTER_CUBIC', 'BORDER_CONSTANT'),
pattern4('ImgProc', 'WarpPerspective', '32FC1', 'INTER_NEAREST', 'BORDER_CONSTANT'),
pattern4('ImgProc', 'WarpPerspective', '32FC1', 'INTER_LINEAR', 'BORDER_CONSTANT'),
pattern4('ImgProc', 'WarpPerspective', '32FC1', 'INTER_CUBIC', 'BORDER_CONSTANT'),
pattern4('ImgProc', 'WarpPerspective', '32FC3', 'INTER_NEAREST', 'BORDER_CONSTANT'),
pattern4('ImgProc', 'WarpPerspective', '32FC3', 'INTER_LINEAR', 'BORDER_CONSTANT'),
pattern4('ImgProc', 'WarpPerspective', '32FC3', 'INTER_CUBIC', 'BORDER_CONSTANT'),
pattern4('ImgProc', 'WarpPerspective', '32FC4', 'INTER_NEAREST', 'BORDER_CONSTANT'),
pattern4('ImgProc', 'WarpPerspective', '32FC4', 'INTER_LINEAR', 'BORDER_CONSTANT'),
pattern4('ImgProc', 'WarpPerspective', '32FC4', 'INTER_CUBIC', 'BORDER_CONSTANT'),
# ImgProc/CopyMakeBorder (8UC1 | 8UC4 | 32SC1 | 32FC1, BORDER_CONSTANT)
pattern3('ImgProc', 'CopyMakeBorder', '8UC1', 'BORDER_CONSTANT'),
pattern3('ImgProc', 'CopyMakeBorder', '8UC4', 'BORDER_CONSTANT'),
pattern3('ImgProc', 'CopyMakeBorder', '32SC1', 'BORDER_CONSTANT'),
pattern3('ImgProc', 'CopyMakeBorder', '32FC1', 'BORDER_CONSTANT'),
# ImgProc/Threshold (32F, THRESH_TRUNC)
# ImgProc_Resize (8U, 1 | 4, INTER_NEAREST | INTER_LINEAR)
pattern4('ImgProc', 'Resize', '8U', '1', 'INTER_NEAREST'),
pattern4('ImgProc', 'Resize', '8U', '4', 'INTER_NEAREST'),
pattern4('ImgProc', 'Resize', '8U', '1', 'INTER_LINEAR'),
pattern4('ImgProc', 'Resize', '8U', '4', 'INTER_LINEAR'),
# ImgProc_Resize (8U, 4, INTER_CUBIC)
pattern4('ImgProc', 'Resize', '8U', '4', 'INTER_CUBIC'),
# ImgProc_WarpAffine (8U | 32F, INTER_NEAREST | INTER_LINEAR | INTER_CUBIC, BORDER_CONSTANT)
pattern4('ImgProc', 'WarpAffine', '8U' , 'INTER_NEAREST', 'BORDER_CONSTANT'),
pattern4('ImgProc', 'WarpAffine', '8U' , 'INTER_LINEAR', 'BORDER_CONSTANT'),
pattern4('ImgProc', 'WarpAffine', '8U' , 'INTER_CUBIC', 'BORDER_CONSTANT'),
pattern4('ImgProc', 'WarpAffine', '32F', 'INTER_NEAREST', 'BORDER_CONSTANT'),
pattern4('ImgProc', 'WarpAffine', '32F', 'INTER_LINEAR', 'BORDER_CONSTANT'),
pattern4('ImgProc', 'WarpAffine', '32F', 'INTER_CUBIC', 'BORDER_CONSTANT'),
# ImgProc_WarpPerspective (8U | 32F, INTER_NEAREST | INTER_LINEAR | INTER_CUBIC, BORDER_CONSTANT)
pattern4('ImgProc', 'WarpPerspective', '8U' , 'INTER_NEAREST', 'BORDER_CONSTANT'),
pattern4('ImgProc', 'WarpPerspective', '8U' , 'INTER_LINEAR', 'BORDER_CONSTANT'),
pattern4('ImgProc', 'WarpPerspective', '8U' , 'INTER_CUBIC', 'BORDER_CONSTANT'),
pattern4('ImgProc', 'WarpPerspective', '32F', 'INTER_NEAREST', 'BORDER_CONSTANT'),
pattern4('ImgProc', 'WarpPerspective', '32F', 'INTER_LINEAR', 'BORDER_CONSTANT'),
pattern4('ImgProc', 'WarpPerspective', '32F', 'INTER_CUBIC', 'BORDER_CONSTANT'),
# ImgProc_CopyMakeBorder (8UC1 | 8UC4 | 32SC1 | 32FC1, BORDER_CONSTANT)
pattern4('ImgProc', 'CopyMakeBorder', '8U' , '1', 'BORDER_CONSTANT'),
pattern4('ImgProc', 'CopyMakeBorder', '8U' , '4', 'BORDER_CONSTANT'),
pattern4('ImgProc', 'CopyMakeBorder', '32S', '1', 'BORDER_CONSTANT'),
pattern4('ImgProc', 'CopyMakeBorder', '32F', '1', 'BORDER_CONSTANT'),
# ImgProc_Threshold (32F, THRESH_TRUNC)
pattern3('ImgProc', 'Threshold', '32F', 'THRESH_TRUNC'),
# ImgProc/Integral_Sqr
pattern1('ImgProc', 'Integral_Sqr'),
# ImgProc_IntegralSqr
pattern1('ImgProc', 'IntegralSqr'),
# ImgProc/HistEven_OneChannel
pattern1('ImgProc', 'HistEven_OneChannel'),
# ImgProc_HistEven_OneChannel
pattern1('ImgProc', 'HistEvenOneChannel'),
# ImgProc/HistEven_FourChannel
pattern1('ImgProc', 'HistEven_FourChannel'),
# ImgProc_HistEven_FourChannel
pattern1('ImgProc', 'HistEvenFourChannel'),
# ImgProc/Rotate
# ImgProc_Rotate
pattern1('ImgProc', 'Rotate'),
# ImgProc/SwapChannels
# ImgProc_SwapChannels
pattern1('ImgProc', 'SwapChannels'),
# ImgProc/AlphaComp
# ImgProc_AlphaComp
pattern1('ImgProc', 'AlphaComp'),
# ImgProc/ImagePyramid_build
pattern1('ImgProc', 'ImagePyramid_build'),
# ImgProc_ImagePyramidBuild
pattern1('ImgProc', 'ImagePyramidBuild'),
# ImgProc_ImagePyramid_getLayer
pattern1('ImgProc', 'ImagePyramidGetLayer'),
# ImgProc/ImagePyramid_getLayer
pattern1('ImgProc', 'ImagePyramid_getLayer'),
##############################################################
# MatOp
# MatOp/SetTo (8UC4 | 16UC1 | 16UC4 | 32FC1 | 32FC4)
pattern2('MatOp', 'SetTo', '8UC4'),
pattern2('MatOp', 'SetTo', '16UC1'),
pattern2('MatOp', 'SetTo', '16UC4'),
pattern2('MatOp', 'SetTo', '32FC1'),
pattern2('MatOp', 'SetTo', '32FC4'),
# MatOp/SetToMasked (8UC4 | 16UC1 | 16UC4 | 32FC1 | 32FC4)
pattern2('MatOp', 'SetToMasked', '8UC4'),
pattern2('MatOp', 'SetToMasked', '16UC1'),
pattern2('MatOp', 'SetToMasked', '16UC4'),
pattern2('MatOp', 'SetToMasked', '32FC1'),
pattern2('MatOp', 'SetToMasked', '32FC4'),
# MatOp/CopyToMasked (8UC1 | 8UC3 |8UC4 | 16UC1 | 16UC3 | 16UC4 | 32FC1 | 32FC3 | 32FC4)
pattern2('MatOp', 'CopyToMasked', '8UC1'),
pattern2('MatOp', 'CopyToMasked', '8UC3'),
pattern2('MatOp', 'CopyToMasked', '8UC4'),
pattern2('MatOp', 'CopyToMasked', '16UC1'),
pattern2('MatOp', 'CopyToMasked', '16UC3'),
pattern2('MatOp', 'CopyToMasked', '16UC4'),
pattern2('MatOp', 'CopyToMasked', '32FC1'),
pattern2('MatOp', 'CopyToMasked', '32FC3'),
pattern2('MatOp', 'CopyToMasked', '32FC4'),
# MatOp_SetTo (8UC4 | 16UC1 | 16UC4 | 32FC1 | 32FC4)
pattern3('MatOp', 'SetTo', '8U' , '4'),
pattern3('MatOp', 'SetTo', '16U', '1'),
pattern3('MatOp', 'SetTo', '16U', '4'),
pattern3('MatOp', 'SetTo', '32F', '1'),
pattern3('MatOp', 'SetTo', '32F', '4'),
# MatOp_SetToMasked (8UC4 | 16UC1 | 16UC4 | 32FC1 | 32FC4)
pattern3('MatOp', 'SetToMasked', '8U' , '4'),
pattern3('MatOp', 'SetToMasked', '16U', '1'),
pattern3('MatOp', 'SetToMasked', '16U', '4'),
pattern3('MatOp', 'SetToMasked', '32F', '1'),
pattern3('MatOp', 'SetToMasked', '32F', '4'),
# MatOp_CopyToMasked (8UC1 | 8UC3 |8UC4 | 16UC1 | 16UC3 | 16UC4 | 32FC1 | 32FC3 | 32FC4)
pattern3('MatOp', 'CopyToMasked', '8U' , '1'),
pattern3('MatOp', 'CopyToMasked', '8U' , '3'),
pattern3('MatOp', 'CopyToMasked', '8U' , '4'),
pattern3('MatOp', 'CopyToMasked', '16U', '1'),
pattern3('MatOp', 'CopyToMasked', '16U', '3'),
pattern3('MatOp', 'CopyToMasked', '16U', '4'),
pattern3('MatOp', 'CopyToMasked', '32F', '1'),
pattern3('MatOp', 'CopyToMasked', '32F', '3'),
pattern3('MatOp', 'CopyToMasked', '32F', '4'),
]
cublasPattern = pattern1('Core', 'GEMM')
@ -260,7 +239,7 @@ if __name__ == "__main__":
inputFile = open(sys.argv[1], 'r')
lines = inputFile.readlines()
inputFile.close()
for i in range(len(lines)):
if cublasPattern.match(lines[i]):

Loading…
Cancel
Save