MLBEDSW-4075 PACK axis 0 + tanh fails with output diff
The test failed since the tanh had batch size > 1.
Added checks for batch size for all supported operators.
Signed-off-by: Fredrik Svedberg <fredrik.svedberg@arm.com>
Change-Id: I3570352740c40eb96bd9db965dfa3c91c81ff2ad
diff --git a/SUPPORTED_OPS.md b/SUPPORTED_OPS.md
index 3ccf3ab..9488a78 100644
--- a/SUPPORTED_OPS.md
+++ b/SUPPORTED_OPS.md
@@ -1,7 +1,7 @@
# Supported Ops
This file was automatically generated by Vela using the `--supported-ops-report` parameter.
-Vela version: `3.6.0rc1.dev11+gac5e33e`
+Vela version: `3.5.1.dev10+gf616c9d6.d20220915`
This file complies with
[**Gitiles Markdown syntax**](https://github.com/google/gitiles/blob/master/Documentation/markdown.md)
@@ -61,19 +61,20 @@
This is a list of constraints most NPU operators must satisfy in order to be scheduled on the NPU.
(Operators excluded from certain constraints are shown in brackets [ ] )
-- Input(s) and Output tensors must not be dynamic - [Quantize]
-- Input(s) and Output tensors must have a defined shape
-- Output tensors cannot be scalar - [Quantize]
-- Scalar Input tensors are only valid for op type: ADD, EXPAND_DIMS, MAXIMUM, MEAN, MINIMUM, MUL, QUANTIZE, SPLIT, SPLIT_V, SUB
-- Input(s) and Output tensors must not be greater than 4D
-- Input(s), Output and Weight tensors must have quantization parameters - [Shape]
-- Input(s), Output and Weight tensors with quantization scales must be finite
-- Input and Output tensors must have quantization scales that fit within float32 precision
-- Constant tensors should not have NoneType-values
+- Input(s) and Output tensors must not be dynamic - [QUANTIZE]
+- Input(s) and Output tensors must have a defined shape
+- Output tensors cannot be scalar - [QUANTIZE]
+- Scalar Input tensors are only valid for op type: ADD, EXPAND_DIMS, MAXIMUM, MEAN, MINIMUM, MUL, QUANTIZE, SPLIT, SPLIT_V, SUB
+- Input(s) and Output tensors must not be greater than 4D
+- Input(s), Output and Weight tensors must have quantization parameters - [SHAPE]
+- Input(s), Output and Weight tensors with quantization scales must be finite
+- Input and Output tensors must have quantization scales that fit within float32 precision
+- Constant tensors should not have NoneType-values
- Tensors must be of type: int16, int32, int8, uint8
- Tensors which are int32 are only valid when op type is: ADD, MUL, SHAPE, SUB
- Tensor dimensions must be in the range [1, 65535]
- Per-axis quantization is only supported for the following op types: CONV_2D, DEPTHWISE_CONV_2D, TRANSPOSE_CONV
+- IFM Tensor batch size must be 1 - [FULLY_CONNECTED, RESHAPE, SHAPE, SLICE, SOFTMAX, SPLIT, SPLIT_V, SQUEEZE, STRIDED_SLICE, UNPACK]
- The fused activation function (if present) must be one of type: LOGISTIC, RELU, RELU6, RELU_N1_TO_1, TANH
- If a fused activation function is present, the Output tensor must be one of type: int16, int8, uint8
@@ -83,7 +84,6 @@
- At least one Input's shape must match the OFM's shape
- IFM and OFM data types must match
-- Batch size must be 1 for Input tensors with more than 2 dimensions
### TFLite ADD Constraints
@@ -93,7 +93,6 @@
- Both Input data types must match
- For IFM that are signed, OFM must also be signed
- For IFM that are unsigned, OFM must either be the same type or int32
-- Batch size must be 1 for Input tensors with more than 2 dimensions
- Broadcasting is only allowed for rank indices with dimension 1, from either IFM1 or IFM2
### TFLite AVERAGE_POOL_2D Constraints
@@ -103,7 +102,6 @@
- Stride values for both width and height must be integer types
- IFM and OFM data types must match
- Kernel filter values for both width and height must be integer types
-- IFM Tensor batch size must be 1
- Stride values for both width and height must be in the range [1, 3]
- Kernel filter values for both width and height must be in the range [1, 8]
- VALID padding: Kernel filter height must be in the range [1, 256]
@@ -134,7 +132,6 @@
- The sum of the weights cannot exceed 8323072
- Optional Bias tensor must be of type: int32, int64
- Optional Bias tensor values must fit within 40-bits
-- IFM Tensor batch size must be 1
### TFLite DEPTHWISE_CONV_2D Constraints
@@ -151,7 +148,6 @@
- The sum of the weights cannot exceed 8323072
- Optional Bias tensor must be of type: int32, int64
- Optional Bias tensor values must fit within 40-bits
-- IFM Tensor batch size must be 1
- For depth multipliers > 1, IFM channels must be 1 and OFM channels must be equal to the depth multiplier
### TFLite EXPAND_DIMS Constraints
@@ -184,7 +180,6 @@
- At least one Input's shape must match the OFM's shape
- IFM and OFM data types must match
-- Batch size must be 1 for Input tensors with more than 2 dimensions
### TFLite MAXIMUM Constraints
@@ -192,7 +187,6 @@
- At least one Input's shape must match the OFM's shape
- IFM and OFM data types must match
-- Batch size must be 1 for Input tensors with more than 2 dimensions
- Both Input quantization parameters must match OFM quantization parameters
- Broadcasting is only allowed for rank indices with dimension 1, from either IFM1 or IFM2
@@ -203,7 +197,6 @@
- Stride values for both width and height must be integer types
- IFM and OFM data types must match
- Kernel filter values for both width and height must be integer types
-- IFM Tensor batch size must be 1
- Stride values for both width and height must be in the range [1, 3]
- Kernel filter height must be in the range [1, 256]
- Product of kernel filter width and height must be in the range [1, 65536]
@@ -215,7 +208,6 @@
- IFM must be int8 or uint8
- Input tensor must be at least 2D
- Axis indices must correspond to height and width axes
-- IFM Tensor batch size must be 1
- Product of height and width must be no greater than 65536
- Product of height and width must be no greater than 4096 when:
IFM and OFM have different scale or zero point; or
@@ -236,7 +228,6 @@
- At least one Input's shape must match the OFM's shape
- IFM and OFM data types must match
-- Batch size must be 1 for Input tensors with more than 2 dimensions
- Both Input quantization parameters must match OFM quantization parameters
- Broadcasting is only allowed for rank indices with dimension 1, from either IFM1 or IFM2
@@ -248,7 +239,6 @@
- Both Input data types must match
- For IFM that are signed, OFM must also be signed
- For IFM that are unsigned, OFM must either be the same type or int32
-- Batch size must be 1 for Input tensors with more than 2 dimensions
- Broadcasting is only allowed for rank indices with dimension 1, from either IFM1 or IFM2
### TFLite PAD Constraints
@@ -335,7 +325,6 @@
- Both Input data types must match
- For IFM that are signed, OFM must also be signed
- For IFM that are unsigned, OFM must either be the same type or int32
-- Batch size must be 1 for Input tensors with more than 2 dimensions
- Broadcasting is only allowed for rank indices with dimension 1, from either IFM1 or IFM2
### TFLite TRANSPOSE_CONV Constraints
@@ -353,7 +342,6 @@
- The sum of the weights cannot exceed 8323072
- Optional Bias tensor must be of type: int32, int64
- Optional Bias tensor values must fit within 40-bits
-- IFM Tensor batch size must be 1
- Stride values for both width and height must be 2
- SAME padding: OFM dimensions must equal IFM dimensions multiplied by stride
- VALID padding: OFM dimensions must equal IFM dimensions multiplied by stride,
diff --git a/ethosu/vela/test/test_tflite_supported_operators.py b/ethosu/vela/test/test_tflite_supported_operators.py
index 3872bdc..35fc1a6 100644
--- a/ethosu/vela/test/test_tflite_supported_operators.py
+++ b/ethosu/vela/test/test_tflite_supported_operators.py
@@ -550,25 +550,25 @@
def test_constraint_elemwise_batch_size():
# BINARY CASE
- # Batch can be >1 if dims is <=2D
- op = testutil.create_elemwise_op(Op.Add, "op", [2, 2], [2, 2], [2, 2])
+ # Batch can be >1 if dims is <=3D
+ op = testutil.create_elemwise_op(Op.Add, "op", [2, 2, 2], [2, 2, 2], [2, 2, 2])
assert support.is_operator_supported(op)
- # For dims >2D, batch must be 1
- op = testutil.create_elemwise_op(Op.Add, "op", [1, 2, 2], [1, 2, 2], [1, 2, 2])
+ # For dims >3D, batch must be 1
+ op = testutil.create_elemwise_op(Op.Add, "op", [1, 2, 2, 2], [1, 2, 2, 2], [1, 2, 2, 2])
assert support.is_operator_supported(op)
# invalid case
- op = testutil.create_elemwise_op(Op.Add, "op", [2, 2, 2], [2, 2, 2], [2, 2, 2])
+ op = testutil.create_elemwise_op(Op.Add, "op", [2, 2, 2, 2], [2, 2, 2, 2], [2, 2, 2, 2])
assert not support.is_operator_supported(op)
# UNARY CASE
- # Batch can be >1 if dims is <=2D
- op = testutil.create_elemwise_op(Op.CLZ, "op", [2, 2], None, [2, 2], datatype=DataType.int32)
+ # Batch can be >1 if dims is <=3D
+ op = testutil.create_elemwise_op(Op.CLZ, "op", [2, 2, 2], None, [2, 2, 2], datatype=DataType.int32)
assert support.is_operator_supported(op)
- # For dims >2D, batch must be 1
- op = testutil.create_elemwise_op(Op.CLZ, "op", [1, 2, 2], None, [1, 2, 2], datatype=DataType.int32)
+ # For dims >3D, batch must be 1
+ op = testutil.create_elemwise_op(Op.CLZ, "op", [1, 2, 2, 2], None, [1, 2, 2, 2], datatype=DataType.int32)
assert support.is_operator_supported(op)
# invalid case
- op = testutil.create_elemwise_op(Op.CLZ, "op", [2, 2, 2], None, [2, 2, 2], datatype=DataType.int32)
+ op = testutil.create_elemwise_op(Op.CLZ, "op", [2, 2, 2, 2], None, [2, 2, 2, 2], datatype=DataType.int32)
assert not support.is_operator_supported(op)
diff --git a/ethosu/vela/tflite_supported_operators.py b/ethosu/vela/tflite_supported_operators.py
index b6f9796..d42caf5 100644
--- a/ethosu/vela/tflite_supported_operators.py
+++ b/ethosu/vela/tflite_supported_operators.py
@@ -20,6 +20,7 @@
import numpy as np
from .data_type import DataType
+from .numeric_util import full_shape
from .operation import Op
from .operation import Padding
from .supported_operators_util import docstring_format_args
@@ -206,9 +207,20 @@
self.generic_constraints.append(TFLiteSupportedOperators.constraint_tens_int32_ops)
self.generic_constraints.append(TFLiteSupportedOperators.constraint_tens_dimension)
self.generic_constraints.append(TFLiteSupportedOperators.constraint_tens_quant_per_axis)
+ self.generic_constraints.append(TFLiteSupportedOperators.constraint_batch_size)
self.generic_constraints.append(TFLiteSupportedOperators.constraint_faf)
self.generic_constraints.append(TFLiteSupportedOperators.constraint_faf_type)
+ # Setup generic constraint exceptions
+ self.generic_constraints_exceptions = defaultdict(list)
+ self.generic_constraints_exceptions[Op.FullyConnected].append(TFLiteSupportedOperators.constraint_batch_size)
+ self.generic_constraints_exceptions[Op.Softmax].append(TFLiteSupportedOperators.constraint_batch_size)
+ self.generic_constraints_exceptions[Op.Reshape].append(TFLiteSupportedOperators.constraint_batch_size)
+ self.generic_constraints_exceptions[Op.Shape].append(TFLiteSupportedOperators.constraint_batch_size)
+ self.generic_constraints_exceptions[Op.Squeeze].append(TFLiteSupportedOperators.constraint_batch_size)
+ for op_type in TFLiteSupportedOperators.split_ops - set((Op.UnpackReshaped,)):
+ self.generic_constraints_exceptions[op_type].append(TFLiteSupportedOperators.constraint_batch_size)
+
# Setup specific constraints. Note: the order matters
self.specific_constraints = defaultdict(list)
@@ -223,7 +235,6 @@
self.specific_constraints[op_type].append(TFLiteSupportedOperators.constraint_weights_limit)
self.specific_constraints[op_type].append(TFLiteSupportedOperators.constraint_bias_type)
self.specific_constraints[op_type].append(TFLiteSupportedOperators.constraint_bias_40bit)
- self.specific_constraints[op_type].append(TFLiteSupportedOperators.constraint_batch_size)
# Depthwise Conv specific checks:
for op_type in TFLiteSupportedOperators.depthwise_convolution_ops:
self.specific_constraints[op_type].append(TFLiteSupportedOperators.constraint_depth_multiplier)
@@ -235,7 +246,6 @@
# Pooling checks:
for op_type in TFLiteSupportedOperators.pooling_ops:
- self.specific_constraints[op_type].append(TFLiteSupportedOperators.constraint_batch_size)
self.specific_constraints[op_type].append(TFLiteSupportedOperators.constraint_stride_range)
# AVG pooling specific checks:
for op_type in TFLiteSupportedOperators.avg_pooling_ops:
@@ -268,9 +278,7 @@
self.specific_constraints[op_type].append(TFLiteSupportedOperators.constraint_bias_type)
self.specific_constraints[op_type].append(TFLiteSupportedOperators.constraint_bias_40bit)
- # Element-wise checks:
- for op_type in TFLiteSupportedOperators.elem_wise_main_ops:
- self.specific_constraints[op_type].append(TFLiteSupportedOperators.constraint_elemwise_batch_size)
+ # Element-wise checks
# Binary Min/Max specific checks:
for op_type in TFLiteSupportedOperators.binary_elem_wise_min_max_ops:
self.specific_constraints[op_type].append(
@@ -302,7 +310,6 @@
self.specific_constraints[Op.Pad].append(TFLiteSupportedOperators.constraint_pad_type)
# Mean specific checks:
- self.specific_constraints[Op.Mean].append(TFLiteSupportedOperators.constraint_batch_size)
self.specific_constraints[Op.Mean].append(TFLiteSupportedOperators.constraint_mean_height_width_product_avgpool)
self.specific_constraints[Op.Mean].append(TFLiteSupportedOperators.constraint_mean_height_width_product)
self.specific_constraints[Op.Mean].append(TFLiteSupportedOperators.constraint_mean_height_width_product_int8)
@@ -319,7 +326,10 @@
print(f"Info: {ext_type} '{op.name}' is a CPU only op")
return False
- for constraint in self.generic_constraints + self.specific_constraints[op.type]:
+ op_exceptions = self.generic_constraints_exceptions[op.type]
+ generic_constraints = [constraint for constraint in self.generic_constraints if constraint not in op_exceptions]
+
+ for constraint in generic_constraints + self.specific_constraints[op.type]:
valid, extra = constraint(op)
if not valid:
print(f"Warning: {ext_type} '{op.name}' is not supported on the NPU. Placing on CPU instead")
@@ -497,9 +507,16 @@
@staticmethod
def constraint_batch_size(op):
"IFM Tensor batch size must be 1"
- ifm = op.ifm
- valid = ifm.shape[0] == 1
- return valid, f"Tensor '{ifm.name}' has batch size: {ifm.shape[0]}"
+ valid = True
+ extra = []
+ for tens in (op.ifm, op.ifm2):
+ if tens is not None:
+ batch_size = full_shape(4, tens.shape, 1)[0]
+ if batch_size != 1:
+ valid = False
+ extra.append(f"Tensor '{tens.name}' has batch size: {batch_size}")
+ extra = "\n ".join(extra)
+ return valid, extra
@staticmethod
def constraint_depth_multiplier(op):
@@ -753,20 +770,6 @@
return valid, f"Op has tensors with different quantization parameters to the OFM '{op.ofm.name}': {extra}"
@staticmethod
- def constraint_elemwise_batch_size(op):
- "Batch size must be 1 for Input tensors with more than 2 dimensions"
- valid = True
- extra = []
- for tens in (op.ifm, op.ifm2):
- # Unary ops have ifm2 as None
- if tens is not None:
- if (len(tens.shape) > 2) and (tens.shape[0] != 1):
- valid = False
- extra.append(tens.name)
- extra = ", ".join(extra)
- return valid, f"Op has invalid input tensors: {extra}"
-
- @staticmethod
def constraint_broadcast_shapes(op):
"Broadcasting is only allowed for rank indices with dimension 1, from either IFM1 or IFM2"
ifm_shape = op.ifm.shape
diff --git a/ethosu/vela/tosa_supported_operators.py b/ethosu/vela/tosa_supported_operators.py
index 15e1569..192862e 100644
--- a/ethosu/vela/tosa_supported_operators.py
+++ b/ethosu/vela/tosa_supported_operators.py
@@ -94,6 +94,9 @@
self.generic_constraints.append(TosaSupportedOperators.constraint_rank) # TODO not supported for all ops yet
self.generic_constraints.append(TosaSupportedOperators.constraint_batch) # TODO not supported for all ops yet
+ # Setup generic constraint exceptions
+ self.generic_constraints_exceptions = defaultdict(list)
+
# Setup specific constraints. Note: the order matters
self.specific_constraints = defaultdict(list)
diff --git a/ethosu/vela/vela.py b/ethosu/vela/vela.py
index a42b218..7740711 100644
--- a/ethosu/vela/vela.py
+++ b/ethosu/vela/vela.py
@@ -45,6 +45,7 @@
from .tflite.Model import Model
from .tflite_mapping import builtin_operator_map
from .tflite_mapping import builtin_operator_name_map
+from .tflite_mapping import optype_to_builtintype
from .tflite_model_semantic import TFLiteSemantic
from .tflite_supported_operators import TFLiteSupportedOperators
from .tosa_model_semantic import TosaSemantic
@@ -178,6 +179,12 @@
# To easily exclude NetworkType from generated documentation.
exclude_generation_network_type_value = [NetworkType.TOSA.value]
+ def _exclude_list_names(constraint, exclude_list):
+ constraints_excluded_names = [
+ optype_to_builtintype(op) for op, exclude_constraint in exclude_list if constraint in exclude_constraint
+ ]
+ return f" - [{', '.join(sorted(constraints_excluded_names))}]" if constraints_excluded_names else ""
+
lines = [
"# Supported Ops",
"",
@@ -256,20 +263,13 @@
for constraint in semantic_checker.generic_constraints:
# Markdown needs two spaces at the end of a line to render it as a separate line
reason = constraint.__doc__.replace("\n", " \n")
-
exclude_list = TFLiteSemantic.get_generic_constraint_exclude_list().items()
- constraints_excluded_names = [
- op.name for op, exclude_constraint in exclude_list if constraint in exclude_constraint
- ]
- excluded_constraint_text = ""
- if constraints_excluded_names:
- excluded_constraint_text = f"- [{', '.join(constraints_excluded_names)}]"
-
- lines.append(f"- {reason} {excluded_constraint_text}")
+ lines.append(f"- {reason}{_exclude_list_names(constraint, exclude_list)}")
for constraint in supported.generic_constraints:
# Markdown needs two spaces at the end of a line to render it as a separate line
reason = constraint.__doc__.replace("\n", " \n")
- lines.append(f"- {reason}")
+ exclude_list = supported.generic_constraints_exceptions.items()
+ lines.append(f"- {reason}{_exclude_list_names(constraint, exclude_list)}")
for op, name in op_constraint_links:
lines += [
"",