Reference model changes for fp16 support
Change-Id: I72f21fcfa153046274969d327313e3349981dbe6
Signed-off-by: James Ward <james.ward@arm.com>
diff --git a/verif/generator/tosa_test_gen.py b/verif/generator/tosa_test_gen.py
index b76b656..9ff6ec5 100644
--- a/verif/generator/tosa_test_gen.py
+++ b/verif/generator/tosa_test_gen.py
@@ -81,6 +81,8 @@
return np.int64(
self.rng.integers(low=-(1 << 47), high=(1 << 47), size=shape)
)
+ elif dtype == DType.FP16:
+ return np.float16(self.rng.random(size=shape))
elif dtype == DType.FLOAT:
return np.float32(self.rng.random(size=shape))
else:
@@ -128,6 +130,9 @@
def getRandNumberDType(self, dtype):
if dtype == DType.FLOAT:
return self.rng.random()
+ elif dtype == DType.FP16:
+ rand_f32 = self.rng.random()
+ return np.float16(rand_f32)
elif dtype == DType.BOOL:
return self.rng.choice([False, True])
# TOSA specific INT4 weight range from -7 to 7
@@ -178,13 +183,15 @@
return "i32"
elif t == DType.INT48:
return "i48"
+ elif t == DType.FP16:
+ return "f16"
elif t == DType.FLOAT:
return "float"
else:
raise Exception("Unknown dtype, cannot convert to string: {}".format(t))
def typeWidth(self, t):
- """Get the datatype width for integer types"""
+ """Get the datatype width for data types"""
if t == DType.INT4:
return 4
elif t == DType.INT8:
@@ -199,6 +206,8 @@
return 32
elif t == DType.INT48:
return 48
+ elif t == DType.FP16:
+ return 16
elif t == DType.FLOAT:
return 32
elif t == DType.BOOL:
@@ -346,7 +355,7 @@
# Special for multiply:
# Force the result to INT32 for INT types
- if a.dtype != DType.FLOAT:
+ if a.dtype not in (DType.FP16, DType.FLOAT):
result_tens.setDtype(DType.INT32)
if error_name == ErrorIf.WrongOutputType:
all_dtypes = [DType.INT8, DType.INT16, DType.INT48]
@@ -533,6 +542,7 @@
self,
op,
input,
+ accum_dtype,
stride,
pad,
kernel,
@@ -585,17 +595,43 @@
qinfo = [0, 0]
attr = ts.TosaSerializerAttribute()
- attr.PoolAttribute(kernel, stride, pad, qinfo[0], qinfo[1])
+ attr.PoolAttribute(kernel, stride, pad, qinfo[0], qinfo[1], accum_dtype)
self.ser.addOperator(op["op"], input_list, output_list, attr)
return result_tens
+ def build_maxpool2d(
+ self,
+ op,
+ input,
+ stride,
+ pad,
+ kernel,
+ validator_fcns=None,
+ error_name=None,
+ qinfo=None,
+ ):
+ # Same as build_pool2d but manually sets accum_dtype value
+ # (maxpool has no accum_dtype)
+ return self.build_pool2d(
+ op,
+ input,
+ DType.UNKNOWN,
+ stride,
+ pad,
+ kernel,
+ validator_fcns,
+ error_name,
+ qinfo,
+ )
+
def build_conv2d(
self,
op,
ifm,
filter,
bias,
+ accum_dtype,
strides,
padding,
dilations,
@@ -605,7 +641,15 @@
):
assert len(padding) == 4
result_tens = OutputShaper.conv2dOp(
- self.ser, self.rng, ifm, filter, strides, padding, dilations, error_name
+ self.ser,
+ self.rng,
+ ifm,
+ filter,
+ accum_dtype,
+ strides,
+ padding,
+ dilations,
+ error_name,
)
# Ensure new output type has correct qinfo
@@ -648,7 +692,7 @@
return None
attr = ts.TosaSerializerAttribute()
- attr.ConvAttribute(padding, strides, dilations, qinfo[0], qinfo[1])
+ attr.ConvAttribute(padding, strides, dilations, qinfo[0], qinfo[1], accum_dtype)
self.ser.addOperator(op["op"], input_list, output_list, attr)
return result_tens
@@ -659,6 +703,7 @@
ifm,
filter,
bias,
+ accum_dtype,
strides,
padding,
dilations,
@@ -668,7 +713,15 @@
):
assert len(padding) == 6
result_tens = OutputShaper.conv3dOp(
- self.ser, self.rng, ifm, filter, strides, padding, dilations, error_name
+ self.ser,
+ self.rng,
+ ifm,
+ filter,
+ accum_dtype,
+ strides,
+ padding,
+ dilations,
+ error_name,
)
# Ensure new output type has correct qinfo
@@ -711,7 +764,7 @@
return None
attr = ts.TosaSerializerAttribute()
- attr.ConvAttribute(padding, strides, dilations, qinfo[0], qinfo[1])
+ attr.ConvAttribute(padding, strides, dilations, qinfo[0], qinfo[1], accum_dtype)
self.ser.addOperator(op["op"], input_list, output_list, attr)
return result_tens
@@ -722,6 +775,7 @@
ifm,
filter,
bias,
+ accum_dtype,
stride,
out_pad,
output_shape,
@@ -731,7 +785,7 @@
):
assert len(out_pad) == 4
result_tens = OutputShaper.transposeConv2DOp(
- self.ser, self.rng, ifm, output_shape, error_name
+ self.ser, self.rng, ifm, output_shape, accum_dtype, error_name
)
# Ensure new output type has correct qinfo
@@ -773,7 +827,9 @@
return None
attr = ts.TosaSerializerAttribute()
- attr.TransposeConvAttribute(out_pad, stride, output_shape, qinfo[0], qinfo[1])
+ attr.TransposeConvAttribute(
+ out_pad, stride, output_shape, qinfo[0], qinfo[1], accum_dtype
+ )
self.ser.addOperator(op["op"], input_list, output_list, attr)
return result_tens
@@ -784,6 +840,7 @@
ifm,
filter,
bias,
+ accum_dtype,
strides,
padding,
dilations,
@@ -792,7 +849,15 @@
qinfo=None,
):
result_tens = OutputShaper.depthwiseConv2dOp(
- self.ser, self.rng, ifm, filter, strides, padding, dilations, error_name
+ self.ser,
+ self.rng,
+ ifm,
+ filter,
+ accum_dtype,
+ strides,
+ padding,
+ dilations,
+ error_name,
)
# Ensure new output type has correct qinfo
@@ -835,16 +900,24 @@
return None
attr = ts.TosaSerializerAttribute()
- attr.ConvAttribute(padding, strides, dilations, qinfo[0], qinfo[1])
+ attr.ConvAttribute(padding, strides, dilations, qinfo[0], qinfo[1], accum_dtype)
self.ser.addOperator(op["op"], input_list, output_list, attr)
return result_tens
def build_fully_connected(
- self, op, ifm, filter, bias, validator_fcns=None, error_name=None, qinfo=None
+ self,
+ op,
+ ifm,
+ filter,
+ bias,
+ accum_dtype,
+ validator_fcns=None,
+ error_name=None,
+ qinfo=None,
):
result_tens = OutputShaper.fullyConnectedOp(
- self.ser, self.rng, ifm, filter, error_name
+ self.ser, self.rng, ifm, filter, accum_dtype, error_name
)
# Invalidate Input/Output list for error if checks.
@@ -871,17 +944,22 @@
input_list=input_list,
output_list=output_list,
num_operands=num_operands,
+ accum_dtype=accum_dtype,
):
return None
attr = ts.TosaSerializerAttribute()
- attr.FullyConnectedAttribute(qinfo[0], qinfo[1])
+ attr.FullyConnectedAttribute(qinfo[0], qinfo[1], accum_dtype)
self.ser.addOperator(op["op"], input_list, output_list, attr)
return result_tens
- def build_matmul(self, op, a, b, validator_fcns=None, error_name=None, qinfo=None):
- result_tens = OutputShaper.matmulOp(self.ser, self.rng, a, b, error_name)
+ def build_matmul(
+ self, op, a, b, accum_dtype, validator_fcns=None, error_name=None, qinfo=None
+ ):
+ result_tens = OutputShaper.matmulOp(
+ self.ser, self.rng, a, b, accum_dtype, error_name
+ )
# Invalidate Input/Output list for error if checks.
input_list = [a.name, b.name]
@@ -908,11 +986,12 @@
input_list=input_list,
output_list=output_list,
num_operands=num_operands,
+ accum_dtype=accum_dtype,
):
return None
attr = ts.TosaSerializerAttribute()
- attr.MatMulAttribute(qinfo[0], qinfo[1])
+ attr.MatMulAttribute(qinfo[0], qinfo[1], accum_dtype)
self.ser.addOperator(op["op"], input_list, output_list, attr)
return result_tens
@@ -995,7 +1074,7 @@
return None
attr = ts.TosaSerializerAttribute()
- if a.dtype == DType.FLOAT:
+ if a.dtype in (DType.FP16, DType.FLOAT):
attr.ClampAttribute(0, 0, min_val, max_val)
else:
attr.ClampAttribute(min_val, max_val, 0, 0)
@@ -1811,7 +1890,7 @@
op["op"], [cond_tens.name, a.name, b.name], [result_tens.name], attr
)
- if a.dtype in (DType.FLOAT, DType.INT32):
+ if a.dtype in (DType.FLOAT, DType.FP16, DType.INT32):
then_op, else_op = Op.ADD, Op.SUB
elif a.dtype in (DType.INT8, DType.INT16):
then_op, else_op = Op.LOGICAL_RIGHT_SHIFT, Op.LOGICAL_LEFT_SHIFT
@@ -2350,22 +2429,37 @@
# if not specified, defaults to (1, 4)
# 'build_fcn': tuple of the function to (build_operator(), TensorGen function, ArgGen enum)
# 'types': array of datatypes to be tested
- TYPE_FP = [DType.FLOAT]
+ TYPE_FP = [DType.FLOAT, DType.FP16]
TYPE_INT = [DType.INT8, DType.INT16, DType.INT32] # Excludes INT4
- TYPE_INT_FP = [DType.INT8, DType.INT16, DType.INT32, DType.FLOAT] # Excludes INT4
+ TYPE_INT_FP = [
+ DType.INT8,
+ DType.INT16,
+ DType.INT32,
+ DType.FP16,
+ DType.FLOAT,
+ ] # Excludes INT4
TYPE_BOOL = [DType.BOOL]
- TYPE_FI32 = [DType.FLOAT, DType.INT32]
- TYPE_FIB = [DType.FLOAT, DType.INT8, DType.INT16, DType.INT32, DType.BOOL]
+ TYPE_FI32 = [DType.FLOAT, DType.FP16, DType.INT32] # floating-types and INT32
+ TYPE_FIB = [
+ DType.FP16,
+ DType.FLOAT,
+ DType.INT8,
+ DType.INT16,
+ DType.INT32,
+ DType.BOOL,
+ ]
TYPE_FI16 = [DType.FLOAT, DType.INT16]
- TYPE_NARROW_INT_FP = [DType.INT8, DType.INT16, DType.FLOAT]
+ TYPE_NARROW_INT_FP = [DType.INT8, DType.INT16, DType.FP16, DType.FLOAT]
TYPE_CONV = [
[DType.INT8, DType.INT4, DType.INT32],
[DType.INT8, DType.INT8, DType.INT32],
[DType.INT16, DType.INT8, DType.INT48],
+ [DType.FP16, DType.FP16, DType.FP16],
+ [DType.FP16, DType.FP16, DType.FLOAT],
DType.FLOAT,
]
@@ -2524,7 +2618,7 @@
build_fully_connected,
TosaTensorGen.tgFullyConnected,
TosaTensorValuesGen.tvgDefault,
- None,
+ TosaArgGen.agFullyConnected,
),
"qgen": TosaQuantGen.qgConv,
"types": TYPE_CONV,
@@ -2546,7 +2640,7 @@
build_matmul,
TosaTensorGen.tgMatmul,
TosaTensorValuesGen.tvgDefault,
- None,
+ TosaArgGen.agMatMul,
),
"qgen": TosaQuantGen.qgMatmul,
"types": TYPE_NARROW_INT_FP,
@@ -2564,7 +2658,7 @@
"operands": (1, 0),
"rank": (4, 4),
"build_fcn": (
- build_pool2d,
+ build_maxpool2d,
TosaTensorGen.tgNHWC,
TosaTensorValuesGen.tvgDefault,
TosaArgGen.agPooling,
@@ -3384,7 +3478,7 @@
TosaTensorValuesGen.tvgReduceSum,
TosaArgGen.agAxis,
),
- "types": TYPE_FI32,
+ "types": (DType.FP16, DType.FLOAT, DType.INT32),
"error_if_validators": (
TosaErrorValidator.evAxisLargerRank,
TosaErrorValidator.evAxisSmallerZero,
@@ -3571,7 +3665,7 @@
TosaTensorValuesGen.tvgDefault,
None,
),
- "types": TYPE_INT_FP,
+ "types": (DType.INT8, DType.INT16, DType.INT32, DType.FP16, DType.FLOAT),
"error_if_validators": (
TosaErrorValidator.evWrongInputType,
TosaErrorValidator.evWrongOutputType,
@@ -3612,7 +3706,7 @@
TosaTensorValuesGen.tvgDefault,
TosaArgGen.agResize,
),
- "types": [DType.INT8, DType.INT16, DType.FLOAT],
+ "types": (DType.INT8, DType.INT16, DType.FP16, DType.FLOAT),
"invalid_test_validators": (
TosaInvalidValidator.ivWrongDataTypeOrModeResize,
),
@@ -3646,7 +3740,14 @@
TosaTensorValuesGen.tvgDefault,
TosaArgGen.agCast,
),
- "types": [DType.FLOAT, DType.INT8, DType.INT16, DType.INT32, DType.BOOL],
+ "types": (
+ DType.FP16,
+ DType.FLOAT,
+ DType.INT8,
+ DType.INT16,
+ DType.INT32,
+ DType.BOOL,
+ ),
"error_if_validators": (
TosaErrorValidator.evWrongInputType,
TosaErrorValidator.evWrongOutputType,
@@ -3925,7 +4026,9 @@
return ser.addOutput(shape, outputDType)
@staticmethod
- def conv2dOp(ser, rng, ifm, filter, strides, padding, dilations, error_name=None):
+ def conv2dOp(
+ ser, rng, ifm, filter, accum_dtype, strides, padding, dilations, error_name=None
+ ):
# IFM: NHWC
# Filter: OHWI
@@ -3958,26 +4061,26 @@
ofm_shape = [ifm.shape[0], h, w, filter.shape[0]]
- if ifm.dtype == DType.INT8:
- out_dtype = DType.INT32
- elif ifm.dtype == DType.INT16:
- out_dtype = DType.INT48
- elif ifm.dtype == DType.FLOAT:
- out_dtype = DType.FLOAT
- elif error_name == ErrorIf.WrongInputType:
+ if error_name == ErrorIf.WrongInputType:
# Pick some potentially correct output dtype if input type is incorrect
out_dtype = DType.INT32
else:
- raise Exception(f"Unsupported input dtype: {ifm.dtype}")
+ out_dtype = accum_dtype
if error_name == ErrorIf.WrongOutputType:
- wrong_dtypes = list(usableDTypes(excludes=[out_dtype]))
+ if ifm.dtype == DType.FP16:
+ excludes = [DType.FP16, DType.FLOAT]
+ else:
+ excludes = [out_dtype]
+ wrong_dtypes = list(usableDTypes(excludes=excludes))
out_dtype = rng.choice(wrong_dtypes)
return ser.addOutput(ofm_shape, out_dtype)
@staticmethod
- def conv3dOp(ser, rng, ifm, filter, strides, padding, dilations, error_name=None):
+ def conv3dOp(
+ ser, rng, ifm, filter, accum_dtype, strides, padding, dilations, error_name=None
+ ):
# IFM: NDHWC
# Filter: ODHWI
@@ -4020,27 +4123,25 @@
ofm_shape = [ifm.shape[0], d, h, w, filter.shape[0]]
- if ifm.dtype == DType.INT8:
- out_dtype = DType.INT32
- elif ifm.dtype == DType.INT16:
- out_dtype = DType.INT48
- elif ifm.dtype == DType.FLOAT:
- out_dtype = DType.FLOAT
- elif error_name == ErrorIf.WrongInputType:
+ if error_name == ErrorIf.WrongInputType:
# Pick some potentially correct output dtype if input type is incorrect
out_dtype = DType.INT32
else:
- raise Exception(f"Unsupported input dtype: {ifm.dtype}")
+ out_dtype = accum_dtype
if error_name == ErrorIf.WrongOutputType:
- wrong_dtypes = list(usableDTypes(excludes=[out_dtype]))
+ if ifm.dtype == DType.FP16:
+ excludes = [DType.FP16, DType.FLOAT]
+ else:
+ excludes = [out_dtype]
+ wrong_dtypes = list(usableDTypes(excludes=excludes))
out_dtype = rng.choice(wrong_dtypes)
return ser.addOutput(ofm_shape, out_dtype)
@staticmethod
def depthwiseConv2dOp(
- ser, rng, ifm, filter, strides, padding, dilations, error_name=None
+ ser, rng, ifm, filter, accum_dtype, strides, padding, dilations, error_name=None
):
# IFM: NHWC
# Filter: HWCM
@@ -4073,20 +4174,18 @@
ofm_shape = [ifm.shape[0], h, w, filter.shape[2] * filter.shape[3]]
- if ifm.dtype == DType.INT8:
- out_dtype = DType.INT32
- elif ifm.dtype == DType.INT16:
- out_dtype = DType.INT48
- elif ifm.dtype == DType.FLOAT:
- out_dtype = DType.FLOAT
- elif error_name == ErrorIf.WrongInputType:
+ if error_name == ErrorIf.WrongInputType:
# Pick some potentially correct output dtype if input type is incorrect
out_dtype = DType.INT32
else:
- raise Exception(f"Unsupported input dtype: {ifm.dtype}")
+ out_dtype = accum_dtype
if error_name == ErrorIf.WrongOutputType:
- wrong_dtypes = list(usableDTypes(excludes=[out_dtype]))
+ if ifm.dtype == DType.FP16:
+ excludes = [DType.FP16, DType.FLOAT]
+ else:
+ excludes = [out_dtype]
+ wrong_dtypes = list(usableDTypes(excludes=excludes))
out_dtype = rng.choice(wrong_dtypes)
return ser.addOutput(ofm_shape, out_dtype)
@@ -4119,6 +4218,7 @@
DType.INT32,
DType.INT48,
DType.FLOAT,
+ DType.FP16,
]
wrong_dtypes = list(set(all_dtypes) - set([ifm.dtype]))
outputDType = rng.choice(wrong_dtypes)
@@ -4128,55 +4228,20 @@
return ser.addOutput(ofm_shape, outputDType)
@staticmethod
- def fullyConnectedOp(ser, rng, input, filter, error_name=None):
+ def fullyConnectedOp(ser, rng, input, filter, accum_dtype, error_name=None):
# input: N, IC
# filter: OC, IC
# output: N, OC
output_shape = [input.shape[0], filter.shape[0]]
- if error_name == ErrorIf.WrongOutputType:
- if input.dtype == DType.INT8:
- incorrect_types = (
- DType.INT4,
- DType.INT8,
- DType.INT16,
- DType.INT48,
- DType.FLOAT,
- )
- elif input.dtype == DType.INT16:
- incorrect_types = (
- DType.INT4,
- DType.INT8,
- DType.INT16,
- DType.INT32,
- DType.FLOAT,
- )
- elif input.dtype == DType.FLOAT:
- incorrect_types = (
- DType.INT4,
- DType.INT8,
- DType.INT16,
- DType.INT32,
- DType.INT48,
- )
- out_dtype = rng.choice(a=incorrect_types)
- elif input.dtype == DType.INT8:
- out_dtype = DType.INT32
- elif input.dtype == DType.INT16:
- out_dtype = DType.INT48
- elif input.dtype == DType.FLOAT:
- out_dtype = DType.FLOAT
- elif error_name == ErrorIf.WrongInputType:
- # Pick some potentially correct output dtype if input type is incorrect
- out_dtype = DType.INT32
- else:
- raise Exception("Unsupported input dtype: {}".format(input.dtype))
+ # Validated in arg_gen (also invalidated for ErrorIf)
+ out_dtype = accum_dtype
return ser.addOutput(output_shape, out_dtype)
@staticmethod
- def matmulOp(ser, rng, a, b, error_name=None):
+ def matmulOp(ser, rng, a, b, accum_dtype, error_name=None):
# a: N, H, C
# b: N, C, W
# out: N, H, W
@@ -4200,7 +4265,7 @@
DType.INT32,
DType.FLOAT,
)
- elif a.dtype == DType.FLOAT:
+ elif a.dtype == DType.FLOAT or a.dtype == DType.FP16:
incorrect_types = (
DType.INT4,
DType.INT8,
@@ -4209,17 +4274,11 @@
DType.INT48,
)
out_dtype = rng.choice(a=incorrect_types)
- elif a.dtype == DType.INT8:
- out_dtype = DType.INT32
- elif a.dtype == DType.INT16:
- out_dtype = DType.INT48
- elif a.dtype == DType.FLOAT:
- out_dtype = DType.FLOAT
elif error_name == ErrorIf.WrongInputType:
# Pick some potentially correct output dtype if input type is incorrect
out_dtype = DType.INT32
else:
- raise Exception("Unsupported input dtype for matmul: {}".format(a.dtype))
+ out_dtype = accum_dtype # Validated in arg_gen
return ser.addOutput(output_shape, out_dtype)
@@ -4269,10 +4328,6 @@
bad_dim = rng.choice(range(len(output_shape)))
output_shape[bad_dim] -= rng.choice([1, 2])
- # Fix negative output shape if error_if test causes it
- if error_name == ErrorIf.PadSmallerZero and min(output_shape) < 1:
- output_shape = [i if i >= 1 else 1 for i in output_shape]
-
if error_name == ErrorIf.WrongOutputType:
all_dtypes = [
DType.INT8,
@@ -4280,6 +4335,7 @@
DType.INT32,
DType.INT48,
DType.FLOAT,
+ DType.FP16,
]
wrong_dtypes = list(set(all_dtypes) - set([a.dtype]))
outputDType = rng.choice(wrong_dtypes)
@@ -4546,7 +4602,7 @@
return ser.addOutput(val.shape, out_dtype)
@staticmethod
- def transposeConv2DOp(ser, rng, ifm, output_shape, error_name=None):
+ def transposeConv2DOp(ser, rng, ifm, output_shape, accum_dtype, error_name=None):
if error_name == ErrorIf.ConvOutputShapeMismatch:
choices = [1, 2, 3]
change = rng.choice(choices)
@@ -4555,20 +4611,18 @@
if change in [2, 3]:
output_shape[2] = output_shape[2] + rng.choice(choices)
- if ifm.dtype == DType.INT8:
- out_dtype = DType.INT32
- elif ifm.dtype == DType.INT16:
- out_dtype = DType.INT48
- elif ifm.dtype == DType.FLOAT:
- out_dtype = DType.FLOAT
- elif error_name == ErrorIf.WrongInputType:
+ if error_name == ErrorIf.WrongInputType:
# Pick some potentially correct output dtype if input type is incorrect
out_dtype = DType.INT32
else:
- raise Exception(f"Unsupported input dtype: {ifm.dtype}")
+ out_dtype = accum_dtype
if error_name == ErrorIf.WrongOutputType:
- wrong_dtypes = list(usableDTypes(excludes=[out_dtype]))
+ if ifm.dtype == DType.FP16:
+ excludes = [DType.FP16, DType.FLOAT]
+ else:
+ excludes = [out_dtype]
+ wrong_dtypes = list(usableDTypes(excludes=excludes))
out_dtype = rng.choice(wrong_dtypes)
return ser.addOutput(output_shape, out_dtype)