from __future__ import print_function, division, absolute_import import os import six __all__ = ['gen', 'DATA_SIZES', 'MAX_VALUES', 'MAX', 'MIN', 'BMIN', 'BMAX', 'SMIN', 'SMAX', 'UMIN', 'UMAX', 'TYPE', 'T', 'U', 'B'] _NUMERIC_TYPES = tuple(list(six.integer_types) + [float]) DATA_SIZES = { 'char': 8, 'uchar': 8, 'short': 16, 'ushort': 16, 'int': 32, 'uint': 32, 'long': 64, 'ulong': 64 } # By default, just test what is part of the CL1.1 spec, leave vec3 for later # VEC_WIDTHS = [2, 3, 4, 8, 16] VEC_WIDTHS = [2, 4, 8, 16] # ALL_WIDTHS = [1, 2, 3, 4, 8, 16] ALL_WIDTHS = [1, 2, 4, 8, 16] MIN_VALUES = { 'char': -128, 'uchar': 0, 'short': -32768, 'ushort': 0, 'int': -2147483648, 'uint': 0, 'long': -9223372036854775808, 'ulong': 0 } MAX_VALUES = { 'char': 127, 'uchar': 255, 'short': 32767, 'ushort': 65535, 'int': 2147483647, 'uint': 4294967295, 'long': 9223372036854775807, 'ulong': 18446744073709551615 } # Define placeholders to reduce magic number usage MAX = 'MAX_VAL' MIN = 'MIN_VAL' BMIN = 'min_for_larger_type' BMAX = 'max_for_larger_type' SMIN = 'signed_min_for_type' SMAX = 'signed_max_for_type' UMIN = 'unsigned_min_for_type' UMAX = 'unsigned_max_for_type' TYPE = 'TYPE' SIZE = 'SIZE' TRUE = 'true_value_for_type' #1 for scalar, -1 for vector NEGNAN = 'Negative NAN as a string, because float("-nan") just produces nan' # Identity type list T = { 'char': 'char', 'uchar': 'uchar', 'short': 'short', 'ushort': 'ushort', 'int': 'int', 'uint': 'uint', 'long': 'long', 'ulong': 'ulong' } # Signed type for each type SIGNED = { 'char': 'char', 'uchar': 'char', 'short': 'short', 'ushort': 'short', 'int': 'int', 'uint': 'int', 'long': 'long', 'ulong': 'long' } # Unsigned type for each source type U = { 'char': 'uchar', 'uchar': 'uchar', 'short': 'ushort', 'ushort': 'ushort', 'int': 'uint', 'uint': 'uint', 'long': 'ulong', 'ulong': 'ulong' } # Next larger type with same signedness B = { 'char': 'short', 'uchar': 'ushort', 'short': 'int', 'ushort': 'uint', 'int': 'long', 'uint': 'ulong', } # vecSizes has the layout [in0width, ..., inNwidth] where outType width is # assumed to match the width of the first input def gen_kernel(f, fnName, inTypes, outType, vecSizes, typePrefix): f.write('kernel void test_' + typePrefix + str(vecSizes[0]) + '_' + fnName + '_' + inTypes[0]+'(global '+outType+'* out') for arg in range(0, len(inTypes)): f.write(', global '+inTypes[arg]+'* in'+str(arg)) f.write('){\n') suffix = ';' if (vecSizes[0] == 1): f.write(' out[get_global_id(0)] = ') else: f.write(' vstore'+str(vecSizes[0])+'(') suffix = ', get_global_id(0), out)' + suffix f.write(fnName+'(') suffix = ')' + suffix for arg in range(0, len(inTypes)): if (arg > 0): f.write(', ') # if scalar, don't print vload/vstore if (vecSizes[arg] == 1): f.write('in'+str(arg)+'[get_global_id(0)]') else: f.write('vload'+str(vecSizes[arg])+'(get_global_id(0), in'+str(arg)+')') f.write(suffix+'\n}\n\n') def gen_kernel_1_arg(f, fnName, inType, outType): for vecSize in ALL_WIDTHS: gen_kernel(f, fnName, [inType], outType, [vecSize], '') # 2 argument kernel with input types that match their vector size def gen_kernel_2_arg_same_size(f, fnName, inTypes, outType): for vecSize in ALL_WIDTHS: gen_kernel(f, fnName, inTypes, outType, [vecSize, vecSize], '') # 2 argument kernel with 1 vector and one scalar input argument def gen_kernel_2_arg_mixed_size(f, fnName, inTypes, outType): for vecSize in VEC_WIDTHS: gen_kernel(f, fnName, inTypes, outType, [vecSize, 1], 'tss_') # 2 argument kernel with 1 vector and one scalar input argument with multiple # input data types def gen_kernel_2_arg_mixed_sign(f, fnName, inTypes, outType): for vecSize in ALL_WIDTHS: gen_kernel(f, fnName, inTypes, outType, [vecSize, vecSize], '') # 3-argument built-in functions def gen_kernel_3_arg_same_type(f, fnName, inTypes, outType): for vecSize in ALL_WIDTHS: gen_kernel(f, fnName, inTypes, outType, [vecSize, vecSize, vecSize], '' ) def gen_kernel_3_arg_mixed_size_tss(f, fnName, inTypes, outType): for vecSize in VEC_WIDTHS: gen_kernel(f, fnName, inTypes, outType, [vecSize, 1, 1], 'tss_') def gen_kernel_3_arg_mixed_size_tts(f, fnName, inTypes, outType): for vecSize in VEC_WIDTHS: gen_kernel(f, fnName, inTypes, outType, [vecSize, vecSize, 1], 'tts_') def generate_kernels(f, dataType, fnName, fnDef): argTypes = getArgTypes(dataType, fnDef['arg_types']) # For len(argTypes), remember that this includes the output arg if (len(argTypes) == 2): gen_kernel_1_arg(f, fnName, argTypes[1], argTypes[0]) return if (len(argTypes) == 3 and not fnName is 'upsample'): gen_kernel_2_arg_same_size(f, fnName, [argTypes[1], argTypes[2]], argTypes[0]) if (fnDef['function_type'] is 'tss'): gen_kernel_2_arg_mixed_size(f, fnName, [argTypes[1], argTypes[2]], argTypes[0]) return if (len(argTypes) == 4): gen_kernel_3_arg_same_type(f, fnName, [argTypes[1], argTypes[2], argTypes[3]], argTypes[0]) if (fnDef['function_type'] is 'tss'): gen_kernel_3_arg_mixed_size_tss(f, fnName, [argTypes[1], argTypes[2], argTypes[3]], argTypes[0]) if (fnDef['function_type'] is 'tts'): gen_kernel_3_arg_mixed_size_tts(f, fnName, [argTypes[1], argTypes[2], argTypes[3]], argTypes[0]) return if (fnName is 'upsample'): gen_kernel_2_arg_mixed_sign(f, fnName, [argTypes[1], argTypes[2]], argTypes[0]) return def getValue(type, val, isVector): # Check if val is a str, list, or value if (isinstance(val, str)): if (val == MIN): return MIN_VALUES[type] elif (val == MAX): return MAX_VALUES[type] elif (val == BMIN): return MIN_VALUES[B[type]] elif (val == BMAX): return MAX_VALUES[B[type]] elif (val == SMIN): return MIN_VALUES[SIGNED[type]] elif (val == SMAX): return MAX_VALUES[SIGNED[type]] elif (val == UMIN): return MIN_VALUES[U[type]] elif (val == UMAX): return MAX_VALUES[U[type]] elif (val == TYPE): return type elif (val == SIZE): return DATA_SIZES[type] elif (val == TRUE): if (isVector): return -1 else: return 1 elif (val == NEGNAN): return '-nan' #cl-program-tester translates this for us else: print('Unknown string value: ' + val + '\n') elif (isinstance(val, list)): # The list should be of the format: [op, arg1, ... argN] where op is a # Fn ref and arg[1-n] are either MIN/MAX or numbers (They could be # nested lists). The exception for arg1 is TYPE, which means to # substitute the data type # Evaluate the value of the requested function and arguments # TODO: Change to varargs calls after unshifting the first list element if (callable(val[0])): if (len(val) == 2): return (val[0])(getValue(type, val[1], isVector)) elif (len(val) == 3): return (val[0])(getValue(type, val[1], isVector), getValue(type, val[2], isVector)) elif (len(val) == 4): return (val[0])(getValue(type, val[1], isVector), getValue(type, val[2], isVector), getValue(type, val[3], isVector)) else: return (val[0])(getValue(type, val[1], isVector), getValue(type, val[2], isVector), getValue(type, val[3], isVector), getValue(type, val[4], isVector)) else: return map(lambda x: getValue(type, x, isVector), val); # At this point, we should have been passed a number if (isinstance(val, _NUMERIC_TYPES)): return val print('Invalid value '+repr(val)+' encountered in getValue\n') def getStrVal(type, val, isVector): return " ".join(map(str, getValue(type, val, isVector))) def getArgType(baseType, argType): # If the argType is a string, it's a literal data type... return it if (isinstance(argType, str)): return argType # otherwise it's a list to pick from return argType[baseType] def getArgTypes(baseType, argTypes): ret = [] for argType in argTypes: ret.append(getArgType(baseType, argType)) return ret def isFloatType(t): return t not in U # Print a test with all-vector inputs/outputs and/or mixed vector/scalar args def print_test(f, fnName, argType, functionDef, tests, numTests, vecSize, fntype): # If the test allows mixed vector/scalar arguments, handle the case with # only vector arguments through a recursive call. if (fntype is 'tss' or fntype is 'tts'): print_test(f, fnName, argType, functionDef, tests, numTests, vecSize, 'ttt') # The tss && vecSize==1 case is handled in the non-tss case. if ((not fntype is 'ttt') and vecSize == 1): return # If we're handling mixed vector/scalar input widths, the kernels have # different names than when the vector widths match tssStr = fntype + '_' if (not fntype is 'ttt') else '' argTypes = getArgTypes(argType, functionDef['arg_types']) argCount = len(argTypes) tolerance = functionDef['tolerance'] if 'tolerance' in functionDef else 0 # Write the test header f.write('[test]\n' + 'name: ' + tssStr + fnName + ' ' + argType + str(vecSize) + '\n' + 'kernel_name: test_' + tssStr + str(vecSize) + '_' + fnName + '_' + argType + '\n' + 'global_size: ' + str(numTests) + ' 0 0\n\n' ) # For each argument, write a line containing its type, index, and values for arg in range(0, argCount): argInOut = '' argVal = getStrVal(argType, tests[arg], (vecSize > 1)) if arg == 0: argInOut = 'arg_out: ' else: argInOut = 'arg_in: ' # The output argument and first tss argument are vectors, any that # follow are scalar. If !tss, then everything has a matching vector # width if (fntype is 'ttt' or (arg < 2 and fntype is 'tss') or (arg < 3 and fntype is 'tts')): f.write(argInOut + str(arg) + ' buffer ' + argTypes[arg] + '[' + str(numTests * vecSize) + '] ' + ''.join(map(lambda x: (x + ' ') * vecSize, argVal.split())) ) if arg == 0: f.write(' tolerance {0} '.format(tolerance)) # Use ulp tolerance for float types if isFloatType(argTypes[arg]): f.write('ulp') f.write('\n') else: argInOut = 'arg_in: ' f.write(argInOut + str(arg) + ' buffer ' + argTypes[arg] + '[' + str(numTests) + '] ' + argVal + '\n' ) # Blank line between tests for formatting reasons f.write('\n') def gen(types, minVersions, functions, testDefs, dirName): # Create the output directory if required if not os.path.exists(dirName): try: os.makedirs(dirName) except OSError as e: if e.errno == 17: # file exists pass raise # Loop over all data types being tested. Create one output file per data # type for dataType in types: for fnName in functions: # Merge all of the generic/signed/unsigned/custom test definitions if (dataType, fnName) not in testDefs: continue functionDef = testDefs[(dataType, fnName)] # Check if the function actually exists for this data type if (not functionDef.keys()): continue clcVersionMin = minVersions[fnName] fileName = 'builtin-' + dataType + '-' + fnName + '-' + \ str(float(clcVersionMin)/10)+'.generated.cl' fileName = os.path.join(dirName, fileName) f = open(fileName, 'w') print(fileName) # Write the file header f.write('/*!\n' + '[config]\n' + 'name: Test '+dataType+' '+fnName+' built-in on CL 1.1\n' + 'clc_version_min: '+str(clcVersionMin)+'\n' + 'dimensions: 1\n' ) if (dataType == 'double'): f.write('require_device_extensions: cl_khr_fp64\n') # Blank line to provide separation between config header and tests f.write('\n') # Write all tests for the built-in function tests = functionDef['values'] argCount = len(functionDef['arg_types']) fnType = functionDef['function_type'] outputValues = tests[0] numTests = len(outputValues) # Handle all available scalar/vector widths sizes = sorted(VEC_WIDTHS) sizes.insert(0, 1) # Add 1-wide scalar to the vector widths for vecSize in sizes: print_test(f, fnName, dataType, functionDef, tests, numTests, vecSize, fnType) # Terminate the header section f.write('!*/\n\n') if (dataType == 'double'): f.write('#pragma OPENCL EXTENSION cl_khr_fp64 : enable\n\n') # Generate the actual kernels generate_kernels(f, dataType, fnName, functionDef) f.close()