MLBEDSW-4892: Fix crash affecting biases without quantization.
Remove quant_values attribute from Tensor class.
It only needs a single values attribute, holding either
quantized or unquantized values as appropriate.
Change-Id: Ie96f80ac58061b6077e0f7048dc60209fdfbcafa
Signed-off-by: James Peet <james.peet@arm.com>
diff --git a/ethosu/vela/data_type.py b/ethosu/vela/data_type.py
index 07086d6..470504d 100644
--- a/ethosu/vela/data_type.py
+++ b/ethosu/vela/data_type.py
@@ -18,6 +18,8 @@
import enum
from typing import Any
+import numpy as np
+
from .numeric_util import round_up_divide
@@ -99,6 +101,16 @@
__repr__ = __str__
+ def as_numpy_type(self):
+ numpy_dtype_code = {
+ BaseType.UnsignedInt: "u",
+ BaseType.SignedInt: "i",
+ BaseType.Float: "f",
+ BaseType.Complex: "c",
+ }
+ assert self.type in numpy_dtype_code, f"Failed to interpret {self} as a numpy dtype"
+ return np.dtype(numpy_dtype_code[self.type] + str(self.size_in_bytes()))
+
stem_name = {
BaseType.UnsignedInt: ("uint%s", True),
BaseType.SignedInt: ("int%s", True),
diff --git a/ethosu/vela/high_level_command_to_npu_op.py b/ethosu/vela/high_level_command_to_npu_op.py
index 80d0e47..9b76ec1 100644
--- a/ethosu/vela/high_level_command_to_npu_op.py
+++ b/ethosu/vela/high_level_command_to_npu_op.py
@@ -437,8 +437,7 @@
npu_op.ifm2.quantization = get_ifm_or_ifm2_quantization(ps, cmd.ifm2_tensor)
if cmd.ifm2_tensor.shape == []:
# scalar
- assert cmd.ifm2_tensor.quant_values.size == 1
- npu_op.ifm2_scalar = cmd.ifm2_tensor.values.item(0)
+ npu_op.ifm2_scalar = cmd.ifm2_tensor.get_scalar()
npu_op.ifm2.shape = NpuShape3D(height=0, width=0, depth=0)
else:
ifm2_blk = cmd.ifm2_box.get_block()
diff --git a/ethosu/vela/npu_serialisation.py b/ethosu/vela/npu_serialisation.py
index 06ea61d..ea35ac6 100644
--- a/ethosu/vela/npu_serialisation.py
+++ b/ethosu/vela/npu_serialisation.py
@@ -48,7 +48,7 @@
def copy_ifm_values_to_memory_tensor(memory_tensor, src_tensor):
start_addr = src_tensor.address
- values = src_tensor.quant_values.flatten() if src_tensor.quant_values is not None else src_tensor.values.flatten()
+ values = src_tensor.values.flatten()
if src_tensor.dtype.size_in_bytes() > 1:
values = np.frombuffer(values.tobytes(), dtype=np.uint8)
end_addr = start_addr + values.size
diff --git a/ethosu/vela/reader_util.py b/ethosu/vela/reader_util.py
index 233286c..476b70a 100644
--- a/ethosu/vela/reader_util.py
+++ b/ethosu/vela/reader_util.py
@@ -34,9 +34,6 @@
if tens.values is not None:
tens.values = tens.values.transpose(reorder)
- if tens.quant_values is not None:
- tens.quant_values = tens.quant_values.transpose(reorder)
-
op = Operation(Op.Const, tens.name)
op.set_output_tensor(tens)
return tens
diff --git a/ethosu/vela/supported_operators.py b/ethosu/vela/supported_operators.py
index c993da1..663c78f 100644
--- a/ethosu/vela/supported_operators.py
+++ b/ethosu/vela/supported_operators.py
@@ -532,7 +532,7 @@
def constraint_weights_limit(cls, op):
"The sum of the weights cannot exceed {}"
weights = op.weights
- values = weights.quant_values.astype(np.int64) - weights.quantization.zero_point
+ values = weights.values.astype(np.int64) - weights.quantization.zero_point
limit = np.amax(np.sum(np.absolute(values), axis=(0, 1, 2)))
valid = limit <= cls.weights_limit
return valid, f"Tensor '{weights.name}' has the sum of weights: {limit}"
@@ -551,8 +551,8 @@
def constraint_bias_40bit(op):
"Optional Bias tensor values must fit within 40-bits"
bias = op.bias
- if bias and bias.dtype == DataType.int64 and bias.quant_values is not None:
- valid = all(len(bin(quant_value)[2:]) <= 40 for quant_value in bias.quant_values)
+ if bias and bias.dtype == DataType.int64 and bias.values is not None:
+ valid = all(len(bin(quant_value)[2:]) <= 40 for quant_value in bias.values)
return valid, f"Tensor '{bias.name}' has values larger than 40-bits"
return True, "Op has no bias tensor, or it fits in 40-bit"
diff --git a/ethosu/vela/tensor.py b/ethosu/vela/tensor.py
index 7dbdcdd..677757c 100644
--- a/ethosu/vela/tensor.py
+++ b/ethosu/vela/tensor.py
@@ -254,20 +254,8 @@
res.quant_max = self.quant_max
return res
- def dequantize(self, values):
- if self.zero_point.size == 1 and self.scale_f32.size == 1:
- # same scale is used for all values
- res = (values.astype(np.float64) - self.zero_point) * self.scale_f32
- else:
- # a different scale is used for different sets of values
- values_as_float = values.astype(np.float64)
-
- # this is not compatible with the format of depthwise weights,
- # where input is at index 3 (Output, Kh, Kw, Input)
- # return the quantized values
- return np.ndarray((values_as_float.shape))
-
- return res
+ def dequantize(self, values) -> np.ndarray:
+ return np.subtract(values, self.zero_point) * self.scale_f32
def is_scaling_equal(self, other: Optional["QuantizationParameters"]) -> bool:
# quantisation parameter scaling is not equal if 'other' is None because
@@ -300,16 +288,12 @@
value_dtype: np.dtype = None,
purpose: TensorPurpose = TensorPurpose.Unknown,
quantization: QuantizationParameters = None,
- quant_value_dtype: np.dtype = None,
):
# Tensor
const_tensor = Tensor(shape, dtype, name + "_0")
const_tensor.purpose = purpose
const_tensor.quantization = quantization
const_tensor.values = np.array(values, dtype=value_dtype)
- const_tensor.quant_values = np.frombuffer(
- const_tensor.values.tobytes(), dtype=np.uint8 if not quant_value_dtype else quant_value_dtype
- )
# Operator
const_op = Operation(Op.Const, name)
const_op.set_output_tensor(const_tensor)
@@ -349,7 +333,6 @@
"ops",
"consumer_list",
"values",
- "quant_values",
"compressed_values",
"compressed_values_substream_offsets",
"mem_area",
@@ -391,8 +374,7 @@
self.ops: List[Operation] = []
self.consumer_list: List[Operation] = []
- self.values: Optional[np.ndarray] = None
- self.quant_values: Optional[np.ndarray] = None
+ self.values: Optional[np.ndarray] = None # elements are of type self.dtype
self.compressed_values: Optional[np.ndarray] = None
self.compressed_values_substream_offsets: Optional[List] = None
self.mem_area: MemArea = MemArea.Unknown
@@ -816,6 +798,17 @@
return (self.dtype.type & BaseType.Int) != 0 and self.quantization.is_valid()
+ def get_scalar(self):
+ """
+ return: Unquantized or dequantized scalar value
+ rtype: self.dtype (if unquantized) or float (if dequantized)
+ """
+ assert self.values.size == 1, "get_scalar called on non-scalar tensor"
+ if self.is_quantized():
+ return self.quantization.dequantize(self.values).item(0)
+ else:
+ return self.values.item(0)
+
def __lt__(self, other: "Tensor") -> bool:
return self.equivalence_id < other.equivalence_id
diff --git a/ethosu/vela/test/test_graph_optimiser.py b/ethosu/vela/test/test_graph_optimiser.py
index b37bac8..e0eedd6 100644
--- a/ethosu/vela/test/test_graph_optimiser.py
+++ b/ethosu/vela/test/test_graph_optimiser.py
@@ -139,8 +139,7 @@
conv_out_tens = Tensor(in_shape, in_dtype, "output")
conv_out_tens.quantization = qp.clone()
weight_tens = Tensor([kernel_size, kernel_size, in_shape[-1], out_shape[-1]], in_dtype, "weights")
- weight_tens.values = np.zeros(weight_tens.shape)
- weight_tens.quant_values = np.zeros(weight_tens.shape, np.int8)
+ weight_tens.values = np.zeros(weight_tens.shape, in_dtype.as_numpy_type())
weight_tens.quantization = qp.clone()
bias_tens = Tensor(out_shape, pad_dtype, "biases")
attrs = {"padding": pad_setting, "stride_w": 2, "stride_h": 2, "dilation_w_factor": 1, "dilation_h_factor": 1}
@@ -349,8 +348,7 @@
conv_ofm = Tensor([1, 8, 8, 16], DataType.uint8, "output")
conv_ofm.quantization = quant.clone()
weight_tens = Tensor([1, 1, 16, 16], DataType.uint8, "weights")
- weight_tens.values = np.zeros(weight_tens.shape)
- weight_tens.quant_values = np.zeros(weight_tens.shape, np.uint8)
+ weight_tens.values = np.zeros(weight_tens.shape, np.uint8)
weight_tens.quantization = quant.clone()
bias_tens = Tensor([16], DataType.int32, "biases")
diff --git a/ethosu/vela/test/test_supported_operators.py b/ethosu/vela/test/test_supported_operators.py
index 666a5ec..3830815 100644
--- a/ethosu/vela/test/test_supported_operators.py
+++ b/ethosu/vela/test/test_supported_operators.py
@@ -246,7 +246,7 @@
op = testutil.create_op_with_quant_tensors(Op.Conv2DBias, [1, 1, 1, 1], [1, 1, 1, 1], weights_shape=[1, 1, 1, 1])
op.attrs = {"stride_w": 1, "stride_h": 1}
bias = Tensor([1, 1, 1, 1], DataType.int64, "bias")
- bias.quant_values = np.array([0x01FF_FFFF_FFFF])
+ bias.values = np.array([0x01FF_FFFF_FFFF])
op.add_input_tensor(bias)
assert not support.is_operator_supported(op)
diff --git a/ethosu/vela/tflite_graph_optimiser.py b/ethosu/vela/tflite_graph_optimiser.py
index 3d9eeb8..9fdff8f 100644
--- a/ethosu/vela/tflite_graph_optimiser.py
+++ b/ethosu/vela/tflite_graph_optimiser.py
@@ -354,8 +354,7 @@
# Create an input tensor filled with zeros
shape = op.ofm_shapes[0].as_list()
tens = Tensor(shape, op.inputs[0].dtype, op.inputs[1].name + "_add")
- tens.values = np.zeros(shape)
- tens.quant_values = np.zeros(shape, np.uint8)
+ tens.values = np.zeros(shape, tens.dtype.as_numpy_type())
tens.quantization = QuantizationParameters(0.0, 255.0)
tens.quantization.scale_f32 = 1.0
tens.quantization.zero_point = 0
@@ -470,8 +469,8 @@
# Reshape Weights to be 4D. IO becomes HWIO
weight_tensor = op.inputs[1]
- weight_tensor.quant_values = np.expand_dims(np.expand_dims(weight_tensor.quant_values, axis=0), axis=0)
- weight_tensor.set_all_shapes(list(weight_tensor.quant_values.shape))
+ weight_tensor.values = np.expand_dims(np.expand_dims(weight_tensor.values, axis=0), axis=0)
+ weight_tensor.set_all_shapes(list(weight_tensor.values.shape))
n = op.ofm_shapes[0].batch
h, w = batching_split.get(n, (1, n))
@@ -608,8 +607,8 @@
del op.attrs["channel_multiplier"]
del op.attrs["depth_multiplier"]
- weight_tensor.quant_values = np.transpose(weight_tensor.quant_values, (0, 1, 3, 2))
- weight_tensor.set_all_shapes(list(weight_tensor.quant_values.shape))
+ weight_tensor.values = np.transpose(weight_tensor.values, (0, 1, 3, 2))
+ weight_tensor.set_all_shapes(list(weight_tensor.values.shape))
else:
raise UnsupportedFeatureError(
f"Unsupported 'DEPTHWISE_CONV_2D' with depth_multiplier = {op.attrs['depth_multiplier']},",
@@ -622,8 +621,8 @@
def reorder_depthwise_weights(op, arch, nng):
if op.type.is_depthwise_conv2d_op():
weight_tensor = op.inputs[1]
- weight_tensor.quant_values = np.transpose(weight_tensor.quant_values, (0, 1, 3, 2))
- weight_tensor.set_all_shapes(list(weight_tensor.quant_values.shape))
+ weight_tensor.values = np.transpose(weight_tensor.values, (0, 1, 3, 2))
+ weight_tensor.set_all_shapes(list(weight_tensor.values.shape))
weight_tensor.weight_transpose_depthwise = True
return op
@@ -654,14 +653,14 @@
for i in range(weight_shape[0]):
padded_array[i] = np.vstack(
[
- weight_tensor.quant_values[i],
+ weight_tensor.values[i],
np.full((1, weight_shape[2], weight_shape[3]), weight_tensor.quantization.zero_point),
]
)
- weight_tensor.quant_values = padded_array
+ weight_tensor.values = padded_array
weight_shape[1] //= 2
weight_shape[2] *= 2
- weight_tensor.quant_values = np.reshape(weight_tensor.quant_values, weight_shape)
+ weight_tensor.values = np.reshape(weight_tensor.values, weight_shape)
weight_tensor.set_all_shapes(weight_shape)
# If multiple copies of the weights are used, we could avoid
# them having the same address by changing the value_id
@@ -692,8 +691,8 @@
}
# Reshape Weights to be 2D. HWIO becomes just IO (as H and W are 1, they can just be dropped)
weight_tensor = op.inputs[1]
- weight_tensor.quant_values = weight_tensor.quant_values.squeeze(axis=(0, 1))
- weight_tensor.set_all_shapes(list(weight_tensor.quant_values.shape))
+ weight_tensor.values = weight_tensor.values.squeeze(axis=(0, 1))
+ weight_tensor.set_all_shapes(list(weight_tensor.values.shape))
DebugDatabase.add_optimised(op, op)
return op
@@ -729,11 +728,11 @@
ifm2_tensor.shape = full_shape(len(ifm_tensor.shape), ifm2_tensor.shape, 1)
elif diff < 0:
ifm_tensor.shape = full_shape(len(ifm2_tensor.shape), ifm_tensor.shape, 1)
- elif ifm_tensor.shape == [] and ifm_tensor.quant_values is None:
+ elif ifm_tensor.shape == [] and ifm_tensor.values is None:
# IFM is marked as a scalar, but is a result of an operation; change it to a shape of size 1
ifm_tensor.shape = len(ifm2_tensor.shape) * [1]
ifm_tensor.storage_shape = ifm_tensor.shape
- elif ifm2_tensor.shape == [] and ifm2_tensor.quant_values is None:
+ elif ifm2_tensor.shape == [] and ifm2_tensor.values is None:
# IFM2 is marked as a scalar, but is a result of an operation; change it to a shape of size 1
ifm2_tensor.shape = len(ifm_tensor.shape) * [1]
ifm2_tensor.storage_shape = ifm2_tensor.shape
@@ -811,7 +810,7 @@
# to produce bit exact results, the alpha is not enough;
# save additional scaling info in attr "alpha_scale", to be used as input
# to the LUT construction
- alpha_scalar = const_tens.quant_values - const_tens.quantization.zero_point
+ alpha_scalar = const_tens.values - const_tens.quantization.zero_point
mul_ifm_scale = np.double(ifm.quantization.scale_f32)
mul_ifm2_scale = np.double(const_tens.quantization.scale_f32)
mul_ofm_scale = np.double(mul_ofm.quantization.scale_f32)
@@ -912,7 +911,7 @@
alpha_tens = create_const_tensor(
op.name + "_alpha_scalar", [], ifm.dtype, [scalar], np.float32, quantization=quantization
)
- alpha_tens.quant_values = np.array([1])
+ alpha_tens.values = np.array([1])
mul_alpha.add_input_tensor(alpha_tens)
fm_alpha = ofm.clone(op.name + "_alpha", set_unique=True)
mul_alpha.set_output_tensor(fm_alpha)
@@ -1209,7 +1208,7 @@
purpose=TensorPurpose.Weights,
quantization=quantization,
)
- weight_tens.quant_values = weights
+ weight_tens.values = weights
op.type = Op.DepthwiseConv2DBias
op.inputs = []
op.add_input_tensor(ifm)
@@ -1331,7 +1330,6 @@
nr_biases = op.inputs[1].shape[-1]
bias_values = [0] * nr_biases
bias_tensor = create_const_tensor(op.name + "_bias", [nr_biases], DataType.int32, bias_values)
- bias_tensor.quant_values = bias_tensor.values
op.set_input_tensor(bias_tensor, op.type.info.indices.biases[0])
return op
@@ -1409,13 +1407,7 @@
quant = QuantizationParameters()
quant.zero_point = 0
bias_term_tens = create_const_tensor(
- op.name + "_bias",
- [1, 1, 1, 1],
- DataType.int16,
- [bias_term],
- np.int16,
- quantization=quant,
- quant_value_dtype=np.int16,
+ op.name + "_bias", [1, 1, 1, 1], DataType.int16, [bias_term], np.int16, quantization=quant,
)
add_op.add_input_tensor(bias_term_tens)
add_op.set_output_tensor(op.ofm)
@@ -1514,7 +1506,7 @@
),
1,
)
- op.weights.quant_values = np.reshape(op.inputs[1].quant_values, weight_shape)
+ op.weights.values = np.reshape(op.inputs[1].values, weight_shape)
# Add None bias tensor
op.inputs.append(None)
diff --git a/ethosu/vela/tflite_reader.py b/ethosu/vela/tflite_reader.py
index 30bf32a..fbee793 100644
--- a/ethosu/vela/tflite_reader.py
+++ b/ethosu/vela/tflite_reader.py
@@ -107,9 +107,6 @@
tens.values = np.array(buf.view(np_dtype))
else:
tens.values = np.array(buf.view(np_dtype).reshape(shape))
- if tens.quantization is not None:
- tens.quant_values = tens.values
- tens.values = tens.quantization.dequantize(tens.quant_values)
return tens
def parse_operator(self, op_index, op_data):
diff --git a/ethosu/vela/tflite_writer.py b/ethosu/vela/tflite_writer.py
index fd3bf42..e6dd85b 100644
--- a/ethosu/vela/tflite_writer.py
+++ b/ethosu/vela/tflite_writer.py
@@ -243,9 +243,7 @@
def serialise_tensor(self, tens):
builder = self.builder
tens_shape = tens.shape
- values = tens.quant_values
- if values is None:
- values = tens.values
+ values = tens.values
if values is None:
values = np.empty(shape=(0), dtype=np.uint8)
diff --git a/ethosu/vela/tosa_reader.py b/ethosu/vela/tosa_reader.py
index e51ead1..364d9a6 100644
--- a/ethosu/vela/tosa_reader.py
+++ b/ethosu/vela/tosa_reader.py
@@ -192,7 +192,6 @@
fname = decode_str(tens_data.NpyFilename())
tens.values = np.load(os.path.join(file_path, fname))
assert list(tens.values.shape) == tens.shape
- tens.quant_values = tens.values
except (struct.error, TypeError, RuntimeError) as e:
print(f'Error: Invalid npy file. Got "{e}" ')
sys.exit(1)
diff --git a/ethosu/vela/weight_compressor.py b/ethosu/vela/weight_compressor.py
index 7e33e93..6536143 100644
--- a/ethosu/vela/weight_compressor.py
+++ b/ethosu/vela/weight_compressor.py
@@ -100,7 +100,7 @@
def create_weight_compression_config(weight_tens, npu_block_type, ofm_block_depth, ofm_depth_step, dilation):
# Note: for an ofm block only its depth is used in weight compression.
# And block depth > ofm depth gives same result as block depth == ofm depth
- block_depth = min(ofm_block_depth, weight_tens.quant_values.shape[-1])
+ block_depth = min(ofm_block_depth, weight_tens.values.shape[-1])
return WeightCompressionConfig(npu_block_type, block_depth, ofm_depth_step, dilation, weight_tens.value_id)
@@ -214,7 +214,7 @@
# the operator should only have a single output
assert len(tens.consumer_list[0].outputs) == 1
- biases = tens.quant_values
+ biases = tens.values
first_consumer_op = tens.consumer_list[0]
ifm_dtype = first_consumer_op.inputs[0].dtype
@@ -318,7 +318,7 @@
assert weight_tens.quantization.zero_point is not None
# Early zero-point correction
- quant_buf = weight_tens.quant_values.astype(np.int16)
+ quant_buf = weight_tens.values.astype(np.int16)
# the zero point can be either a native or numpy type
if isinstance(weight_tens.quantization.zero_point, (int, float)):
zero_point = np.int16(weight_tens.quantization.zero_point)
@@ -363,7 +363,7 @@
scale_tens.element_size_bytes = 10
# Slice the weight stream up depth-ways into bricks and compress
- full_ofm_depth = weight_tens.quant_values.shape[-1]
+ full_ofm_depth = weight_tens.values.shape[-1]
ofm_block_depth = block_config.ofm_block.depth
weight_range_index = 0