FP16 support in serialization
* Allow serialization of fp16 data
* Add package to support integrated half data-type (half_float::half), independent of native float: http://half.sourceforge.net/
* Allow passing of accumulate data-type in serialization
Signed-off-by: James Ward <james.ward@arm.com>
Change-Id: I54357f02e3776d81958228f699ea5044f2014f4b
diff --git a/python/serializer/tosa_serializer.py b/python/serializer/tosa_serializer.py
index acec4b7..fb89563 100644
--- a/python/serializer/tosa_serializer.py
+++ b/python/serializer/tosa_serializer.py
@@ -58,6 +58,7 @@
"INT48",
"FLOAT",
"UINT16",
+ "FP16",
]
ByteMask = np.uint64(0xFF)
@@ -145,7 +146,15 @@
def __init__(self):
super().__init__()
- def PoolAttribute(self, kernel, stride, pad, input_zp, output_zp):
+ def PoolAttribute(
+ self,
+ kernel,
+ stride,
+ pad,
+ input_zp,
+ output_zp,
+ accum_dtype,
+ ):
from tosa import PoolAttribute as a, Attribute
self.utype = Attribute.Attribute().PoolAttribute
@@ -156,8 +165,9 @@
self.intvecs.append((a.AddStride, stride))
self.ints.append((a.AddInputZp, input_zp))
self.ints.append((a.AddOutputZp, output_zp))
+ self.ints.append((a.AddAccumDtype, accum_dtype))
- def ConvAttribute(self, pad, stride, dilation, input_zp, weight_zp):
+ def ConvAttribute(self, pad, stride, dilation, input_zp, weight_zp, accum_dtype):
from tosa import ConvAttribute as a, Attribute
self.utype = Attribute.Attribute().ConvAttribute
@@ -168,8 +178,11 @@
self.intvecs.append((a.AddDilation, dilation))
self.ints.append((a.AddInputZp, input_zp))
self.ints.append((a.AddWeightZp, weight_zp))
+ self.ints.append((a.AddAccumDtype, accum_dtype))
- def TransposeConvAttribute(self, outpad, stride, output_shape, input_zp, weight_zp):
+ def TransposeConvAttribute(
+ self, outpad, stride, output_shape, input_zp, weight_zp, accum_dtype
+ ):
from tosa import TransposeConvAttribute as a, Attribute
self.utype = Attribute.Attribute().TransposeConvAttribute
@@ -180,6 +193,7 @@
self.intvecs.append((a.AddOutputShape, output_shape))
self.ints.append((a.AddInputZp, input_zp))
self.ints.append((a.AddWeightZp, weight_zp))
+ self.ints.append((a.AddAccumDtype, accum_dtype))
def PadAttribute(self, padding, pad_const_int, pad_const_fp):
from tosa import PadAttribute as a, Attribute
@@ -316,7 +330,7 @@
self.intvecs.append((a.AddTable, table))
- def MatMulAttribute(self, A_zp, B_zp):
+ def MatMulAttribute(self, A_zp, B_zp, accum_dtype):
from tosa import MatMulAttribute as a, Attribute
self.utype = Attribute.Attribute().MatMulAttribute
@@ -324,8 +338,9 @@
self.ints.append((a.AddAZp, A_zp))
self.ints.append((a.AddBZp, B_zp))
+ self.ints.append((a.AddAccumDtype, accum_dtype))
- def FullyConnectedAttribute(self, input_zp, weight_zp):
+ def FullyConnectedAttribute(self, input_zp, weight_zp, accum_dtype):
from tosa import FullyConnectedAttribute as a, Attribute
self.utype = Attribute.Attribute().FullyConnectedAttribute
@@ -333,6 +348,7 @@
self.ints.append((a.AddInputZp, input_zp))
self.ints.append((a.AddWeightZp, weight_zp))
+ self.ints.append((a.AddAccumDtype, accum_dtype))
def NegateAttribute(self, input1_zp, output_zp):
from tosa import NegateAttribute as a, Attribute
@@ -364,6 +380,8 @@
if dtype == DType.FLOAT:
fntype = np.float32
+ elif dtype == DType.FP16:
+ fntype = np.float16
else:
fntype = int
@@ -445,10 +463,18 @@
b4 = (val_u64 >> np.uint64(32)) & ByteMask
b5 = (val_u64 >> np.uint64(40)) & ByteMask
u8_data.extend([b0, b1, b2, b3, b4, b5])
+ elif self.dtype == DType.FP16:
+ np_arr = np.array(self.data, dtype=np.float16)
+ u8_data.extend(np_arr.view(np.uint8))
elif self.dtype == DType.FLOAT:
for val in self.data:
b = struct.pack("!f", val)
u8_data.extend([b[3], b[2], b[1], b[0]])
+ elif self.dtype == TosaDType.DType:
+ # Serialize DType enum data as uint8 bytes
+ for val in self.data:
+ np_arr = np.array(self.data, dtype=np.uint32)
+ u8_data.extend(np_arr.view(np.uint8))
else:
raise Exception(
"unsupported data type {}".format(DTypeNames[self.dtype])
@@ -873,6 +899,7 @@
)
ConvAttribute.AddInputZp = ConvAttribute.ConvAttributeAddInputZp
ConvAttribute.AddWeightZp = ConvAttribute.ConvAttributeAddWeightZp
+ ConvAttribute.AddAccumDtype = ConvAttribute.ConvAttributeAddAccumDtype
ConvAttribute.End = ConvAttribute.ConvAttributeEnd
from tosa import FullyConnectedAttribute
@@ -886,6 +913,9 @@
FullyConnectedAttribute.AddWeightZp = (
FullyConnectedAttribute.FullyConnectedAttributeAddWeightZp
)
+ FullyConnectedAttribute.AddAccumDtype = (
+ FullyConnectedAttribute.FullyConnectedAttributeAddAccumDtype
+ )
FullyConnectedAttribute.End = (
FullyConnectedAttribute.FullyConnectedAttributeEnd
)
@@ -895,6 +925,7 @@
MatMulAttribute.Start = MatMulAttribute.MatMulAttributeStart
MatMulAttribute.AddAZp = MatMulAttribute.MatMulAttributeAddAZp
MatMulAttribute.AddBZp = MatMulAttribute.MatMulAttributeAddBZp
+ MatMulAttribute.AddAccumDtype = MatMulAttribute.MatMulAttributeAddAccumDtype
MatMulAttribute.End = MatMulAttribute.MatMulAttributeEnd
from tosa import PoolAttribute
@@ -910,6 +941,7 @@
PoolAttribute.StartStrideVector = (
PoolAttribute.PoolAttributeStartStrideVector
)
+ PoolAttribute.AddAccumDtype = PoolAttribute.PoolAttributeAddAccumDtype
PoolAttribute.AddInputZp = PoolAttribute.PoolAttributeAddInputZp
PoolAttribute.AddOutputZp = PoolAttribute.PoolAttributeAddOutputZp
PoolAttribute.End = PoolAttribute.PoolAttributeEnd
@@ -944,6 +976,7 @@
PoolAttribute.StartStrideVector = (
PoolAttribute.PoolAttributeStartStrideVector
)
+ PoolAttribute.AddAccumDtype = PoolAttribute.PoolAttributeAddAccumDtype
PoolAttribute.AddInputZp = PoolAttribute.PoolAttributeAddInputZp
PoolAttribute.AddOutputZp = PoolAttribute.PoolAttributeAddOutputZp
PoolAttribute.End = PoolAttribute.PoolAttributeEnd
@@ -1123,6 +1156,9 @@
TransposeConvAttribute.AddWeightZp = (
TransposeConvAttribute.TransposeConvAttributeAddWeightZp
)
+ TransposeConvAttribute.AddAccumDtype = (
+ TransposeConvAttribute.TransposeConvAttributeAddAccumDtype
+ )
TransposeConvAttribute.End = (
TransposeConvAttribute.TransposeConvAttributeEnd
)