def test_pad_4x3():
z = {} # problemType definition
conv = Convolution(z,
'ConvolutionForward',
config={
'TensorAFormat': 'NCHW',
'Filter': '4x3',
'PadStart': 'NxN',
'PadEnd': 'NxN',
})
log.debug(conv.printUsage(z))
e = {
'n': 1,
'c': 4,
'h': 12,
'w': 8,
'k': 1,
'p': 1,
'p_': 2,
'q': 3,
'q_': 4
}
ec = ConvProblem(e, conv)
assert (ec.zeroPadA[0] == [0, 5, 3, 4])
assert (ec.zeroPadA[1] == [1, 4, 8, 16])
assert (len(ec.zeroPadA) == 2)
def ConvolutionContraction(cls, conv, problemType, solution, dataType, \
problemFunc, generateConvFormat=True, problemLevel=1):
"""
Generates a YamlBuilder object that will run a convolution, in normal
contraction mode.
"""
availableArchs = cls.findAvailableArchs()
benchmarkParams = solution()
for item in benchmarkParams["ForkParameters"]:
if 'MatrixInstruction' in item and 'gfx908' not in availableArchs:
pytest.skip()
break
doc = cls.Header(debug=False)
if generateConvFormat:
tensileProblemType = {
"OperationType":
conv.convolutionType,
"ConvolutionConfig": [{
key: val
} for (key, val) in conv.config.items()],
"DataType":
dataType
}
else:
tensileProblemType = {
"OperationType": "TensorContraction",
"DataType": dataType
}
tensileProblemType.update(problemType)
for (key, value) in conv.solutionParms.items():
benchmarkParams["ForkParameters"].append({key: [value]})
problems = problemFunc(conv, problemType, problemLevel)
#print("problems:", problems)
if generateConvFormat:
convs = [{"Conv": e} for e in problems]
benchmarkParams["BenchmarkFinalParameters"] = [{
"ProblemSizes":
convs
}]
else:
exacts = [{
"Exact": dict(ConvProblem(p, conv).toExactDict())
} for p in problems]
benchmarkParams["BenchmarkFinalParameters"] = [{
"ProblemSizes":
exacts
}]
doc["BenchmarkProblems"] = [[tensileProblemType, benchmarkParams]]
#print (doc)
return cls(doc)
def test_bad_invalid_list():
z = {} # problemType definition
conv = Convolution(z,
'ConvolutionForward',
config={
'TensorAFormat': 'NCHW',
'Stride': '2x3',
})
log.debug(conv.printUsage(z))
e = [1, 2, 3, 4]
with pytest.raises(Exception, match="ConvProblem must be a dictionary"):
ConvProblem(e, conv)
def test_stride2x3_defaults():
z = {} # problemType definition
conv = Convolution(z,
'ConvolutionForward',
config={
'TensorAFormat': 'NCHW',
'Stride': '2x3',
})
log.debug(conv.printUsage(z))
e = {'n': 64, 'c': 256, 'h': 20, 'w': 14, 'k': 1024}
ec = ConvProblem(e, conv)
assert (ec.sizes == (4, 10, e['k'], e['n'], e['c']))
assert (ec.stridesA == (3, 28, -1, -1))
def test_pad_4x1():
z = {} # problemType definition
conv = Convolution(z,
'ConvolutionForward',
config={
'TensorAFormat': 'NCHW',
'Filter': '4x1',
'PadStart': 'Nx0',
'PadEnd': 'Nx0',
})
log.debug(conv.printUsage(z))
e = {'n': 1, 'c': 4, 'h': 12, 'w': 8, 'k': 1, 'p': 2, 'p_': 3}
ec = ConvProblem(e, conv)
assert (ec.zeroPadA == [[1, 4, 2, 3]])
def test_bad_missing_field():
z = {} # problemType definition
conv = Convolution(z,
'ConvolutionForward',
config={
'TensorAFormat': 'NCHW',
'Stride': '2x3',
})
log.debug(conv.printUsage(z))
e = {'c': 256, 'h': 20, 'w': 14, 'k': 1024, 'x': 1, 'y': 1, 'u': 1, 'v': 1}
with pytest.raises(
Exception,
match="required ConvProblem field 'n' not present in ConvProblem"):
ConvProblem(e, conv)
def test_bad_invalid_field():
z = {} # problemType definition
conv = Convolution(z,
'ConvolutionForward',
config={
'TensorAFormat': 'NCHW',
'Stride': '2x3',
})
log.debug(conv.printUsage(z))
e = {
'A': 123,
'n': 64,
'c': 256,
'h': 20,
'w': 14,
'k': 1024,
'x': 1,
'y': 1,
'u': 1,
'v': 1
}
with pytest.raises(Exception, match="unknown ConvProblem field 'A'"):
ConvProblem(e, conv)
def test_bad_stride():
z = {} # problemType definition
conv = Convolution(z,
'ConvolutionForward',
config={
'TensorAFormat': 'NCHW',
'Stride': '2x3',
})
log.debug(conv.printUsage(z))
e = {
'n': 64,
'c': 256,
'h': 20,
'w': 14,
'k': 1024,
'x': 1,
'y': 1,
'u': 7,
'v': 2
}
with pytest.raises(RuntimeError, \
match="Mismatch between ConvolutionConfig value.*stride"):
ConvProblem(e, conv)
def test_bad_filter():
z = {} # problemType definition
conv = Convolution(z,
'ConvolutionForward',
config={
'TensorAFormat': 'NCHW',
})
log.debug(conv.printUsage(z))
e = {
'n': 64,
'c': 256,
'h': 20,
'w': 14,
'k': 1024,
'x': 2,
'y': 1,
'u': 1,
'v': 1
}
with pytest.raises(RuntimeError, \
#match="Mismatch between ConvolutionConfig value \(1\) and specified value \(2\) for filter[0]."):
match="Mismatch between ConvolutionConfig value"):
ConvProblem(e, conv)