MLBEDSW-4838 Added basic TOSA support.
Added basic TOSA support, enabling Vela to
read and compile a .tosa file corresponding to
CONV2D + Rescale + Clamp, and writing it to an
optimized .tflite file.
The optimized .tflite file, will in this case, hold
a commandstream where the Rescale and Clamp has been
fused into the CONV2D.
The optimized tflite file is not output from Vela.
-Added support to read .tosa file into Vela
internal structure.
- Added tosa_reader.py, tosa_mapper.py and
helper files stored under tosa/
- Support for this limited to ~10 ops
-Added reader_util.py for functions common
for TOSA and TFLite
-Added tosa_graph_optimiser.py
-Added support to fuse Rescale into convolution
-Modified handling for padding
-Added support to fuse Clamp to previous op
-Added graph_optimiser_util.py
-Moved functions common for TOSA/TFLite graph
optimization to this file.
-Renamed graph_optimiser.py to tflite_graph_optmiser.py
-Added separate tosa_supported_operators.py
-Added supported_operator_util.py
-For functions in common for TOSA/TFLite
Signed-off-by: Patrik Gustavsson <patrik.gustavsson@arm.com>
Change-Id: Ic3c540504ec8c5eb4771397fdc6882050ecf33ab
diff --git a/.pre-commit-config.yaml b/.pre-commit-config.yaml
index 8702966..14d098b 100644
--- a/.pre-commit-config.yaml
+++ b/.pre-commit-config.yaml
@@ -1,4 +1,4 @@
-exclude: '^ethosu/vela/(tflite|ethos_u55_regs)/'
+exclude: '^ethosu/vela/(tflite|ethos_u55_regs|tosa)/'
repos:
- repo: https://github.com/asottile/reorder_python_imports
rev: v2.2.0
diff --git a/README.md b/README.md
index 7d8c09e..eecc3db 100644
--- a/README.md
+++ b/README.md
@@ -23,6 +23,10 @@
The tool will also generate performance estimates (EXPERIMENTAL) for the
compiled model.
+The tool has limited functionality for compiling a
+[TOSA](https://git.mlplatform.org/tosa/specification.git/) neural network
+(EXPERIMENTAL).
+
## TensorFlow Support
* Vela 2.1.0 to current supports TensorFlow 2.4
diff --git a/ethosu/vela/architecture_features.py b/ethosu/vela/architecture_features.py
index 19133f5..98d3d8c 100644
--- a/ethosu/vela/architecture_features.py
+++ b/ethosu/vela/architecture_features.py
@@ -38,6 +38,7 @@
from .tensor import MemType
from .tensor import TensorFormat
from .tensor import TensorPurpose
+from .tosa_supported_operators import TosaSupportedOperators
class Block:
@@ -398,6 +399,7 @@
# Setup supported operators and restriction checkers class
self.supported_operators = SupportedOperators()
+ self.tosa_supported_operators = TosaSupportedOperators()
# Returns available number of SHRAM banks depending on activation lookup table
# being used or not
diff --git a/ethosu/vela/compiler_driver.py b/ethosu/vela/compiler_driver.py
index a9e3839..cb47539 100644
--- a/ethosu/vela/compiler_driver.py
+++ b/ethosu/vela/compiler_driver.py
@@ -146,14 +146,14 @@
)
-def compiler_driver(nng, arch, options, scheduler_options):
+def compiler_driver(nng, arch, options, scheduler_options, network_type):
assert verify_graph_health(nng)
# Pre-optimisation operator tracking
for sg in nng.subgraphs:
visit_graph_post_order(sg.output_tensors, arch, [], [_record_operator])
- nng = graph_optimiser.optimise_graph_a(nng, arch, options.verbose_graph)
+ nng = graph_optimiser.optimise_graph(nng, arch, network_type, options.verbose_graph)
assert verify_graph_health(nng)
if options.verbose_quantization:
diff --git a/ethosu/vela/data_type.py b/ethosu/vela/data_type.py
index 3ad642a..07086d6 100644
--- a/ethosu/vela/data_type.py
+++ b/ethosu/vela/data_type.py
@@ -44,9 +44,11 @@
__slots__ = "type", "bits"
+ int4: Any
int8: Any
int16: Any
int32: Any
+ int48: Any
int64: Any
uint8: Any
uint16: Any
@@ -113,9 +115,11 @@
# generate the standard set of data types
+DataType.int4 = DataType(BaseType.SignedInt, 4)
DataType.int8 = DataType(BaseType.SignedInt, 8)
DataType.int16 = DataType(BaseType.SignedInt, 16)
DataType.int32 = DataType(BaseType.SignedInt, 32)
+DataType.int48 = DataType(BaseType.SignedInt, 48)
DataType.int64 = DataType(BaseType.SignedInt, 64)
DataType.uint8 = DataType(BaseType.UnsignedInt, 8)
diff --git a/ethosu/vela/graph_optimiser.py b/ethosu/vela/graph_optimiser.py
index d2598ae..87e3bc8 100644
--- a/ethosu/vela/graph_optimiser.py
+++ b/ethosu/vela/graph_optimiser.py
@@ -1,4 +1,4 @@
-# Copyright (C) 2020-2021 Arm Limited or its affiliates. All rights reserved.
+# Copyright (C) 2021 Arm Limited or its affiliates. All rights reserved.
#
# SPDX-License-Identifier: Apache-2.0
#
@@ -14,1779 +14,32 @@
# See the License for the specific language governing permissions and
# limitations under the License.
# Description:
-# Early optimisation of the network graph, using the rewrite_graph module to do the traversal of the graph. These are
-# split into two parts optimise_graph_a and optimise_graph_b.
-import math
-import uuid
-from typing import Tuple
-
-import numpy as np
-
-from . import fp_math
-from . import lut
+# Early optimisation of the network graph, using the rewrite_graph module to do the traversal of the graph.
from . import rewrite_graph
-from . import scaling
-from .api import NpuRoundingMode
-from .data_type import DataType
-from .debug_database import DebugDatabase
-from .errors import UnsupportedFeatureError
-from .errors import VelaError
-from .ethos_u55_regs.ethos_u55_regs import resampling_mode
-from .numeric_util import clamp_sigmoid
-from .numeric_util import full_shape
-from .numeric_util import round_away_zero
-from .operation import create_activation_function
-from .operation import NpuBlockType
-from .operation import Op
-from .operation import Operation
-from .operation import Padding
-from .operation_util import create_avgpool_nop
-from .operation_util import get_pad_values_from_input
-from .shape4d import Shape4D
-from .softmax import SoftMax
-from .tensor import check_quantized_tens_scaling_equal
-from .tensor import create_const_tensor
-from .tensor import create_equivalence_id
-from .tensor import QuantizationParameters
-from .tensor import Tensor
-from .tensor import TensorPurpose
-from .tflite_mapping import optype_to_builtintype
+from .graph_optimiser_util import check_format_restrictions
+from .graph_optimiser_util import check_reshapes
+from .graph_optimiser_util import record_optimised
+from .nn_graph import NetworkType
+from .tflite_graph_optimiser import tflite_optimise_graph
+from .tosa_graph_optimiser import tosa_optimise_graph
-passthrough_nodes = (Op.Identity,)
-memory_only_ops = (Op.Reshape,)
-
-
-def create_avg_pool_for_concat(concat_op, name, ifm, ifm_shape: Shape4D, write_offset: Shape4D):
- """Creates an average pool for the given concat op/input feature map"""
- ofm = concat_op.ofm
- avgpool_op = create_avgpool_nop(name)
- avgpool_op.inputs = [ifm]
- avgpool_op.outputs = [ofm]
-
- avgpool_op.write_offset = write_offset
- avgpool_op.write_shape = ifm_shape
- ofm.ops.append(avgpool_op)
- DebugDatabase.add_optimised(concat_op, avgpool_op)
- avgpool_op.ifm_shapes.append(ifm_shape)
- avgpool_op.ofm_shapes.append(concat_op.ofm_shapes[0])
- avgpool_op.memory_function = Op.ConcatSliceWrite
- return avgpool_op
-
-
-def remove_passthrough_tensor(tens, arch, nng):
- if len(tens.ops) == 1 and tens.ops[0].type in passthrough_nodes:
- assert len(tens.ops[0].inputs) == 1
- tens = tens.ops[0].inputs[0]
- return tens
-
-
-def rewrite_concat_ops(op, arch):
- if not op.run_on_npu or not op.type.is_concat_op():
- return
-
- axis_4D = 0
- ofm = op.ofm
- ofm.ops = []
- offset = 0
-
- unfuse_activation_function(op)
-
- if op.type == Op.Pack:
- # Pack is also referred to as Stack
- axis = int(op.attrs["axis"])
- if axis < 0: # Convert to positive axis
- axis = len(op.inputs[0].shape) + 1 + axis
-
- desired_shape = op.inputs[0].shape[:axis] + [1] + op.inputs[0].shape[axis:]
-
- axis_4D = axis + (4 - len(desired_shape))
-
- for idx, inp in enumerate(op.inputs):
- op.ifm_shapes[idx] = Shape4D(desired_shape)
- op.type = Op.PackReshaped
-
- inputs, axis = op.get_concat_inputs_axis()
- for idx, inp in enumerate(inputs):
- if op.type != Op.PackReshaped:
- op.ifm_shapes[idx] = Shape4D(inp.shape)
- if axis >= 0:
- axis_4D = axis + (4 - len(inp.shape))
- else:
- axis_4D = axis
- write_offset = [0, 0, 0, 0]
- write_offset[axis_4D] = offset
- concat_end = offset + op.ifm_shapes[idx][axis_4D]
- create_avg_pool_for_concat(
- op, op.name + str(idx) + "_avgpool", inp, op.ifm_shapes[idx], Shape4D.from_list(write_offset)
- )
- offset = concat_end
- assert ofm.shape[axis] == offset
-
- return op
-
-
-def rewrite_split_ops(tens, arch, nng):
-
- if len(tens.ops) == 1 and tens.ops[0].type.is_split_op() and tens.ops[0].type != Op.Unpack:
- split_op = tens.ops[0]
-
- # Not supported so leave it and run on CPU
- if not split_op.run_on_npu:
- return tens
-
- inp, outputs, axis, offset_start, offset_end = split_op.get_split_inputs_axis()
-
- tens.ops = []
- new_op = Operation(Op.SplitSliceRead, split_op.name)
- new_op.inputs = [inp]
- ofm_shape_idx = 0
- read_shape = offset_end
-
- # For Split the offset cannot be extracted from the tensor so it has to
- # be calculated from the index of the output tensor
- if axis is not None:
- # Get the start and end of the split
- offset_start = [0] * 4
- axis_4D_list = split_op.attrs.get("split_axis_4D", None) # Present for UnpackReshaped and some StridedSlice
- for idx, out in enumerate(outputs):
- if axis_4D_list is not None:
- axis_4D = axis_4D_list[idx]
- else:
- split_op.ofm_shapes[idx] = Shape4D(out.shape)
- if axis >= 0:
- axis_4D = axis + (4 - len(out.shape))
- else:
- axis_4D = axis
-
- if out == tens:
- ofm_shape_idx = idx
- read_shape = split_op.ofm_shapes[idx]
- break
-
- offset_start[axis_4D] += split_op.ofm_shapes[idx][axis_4D]
-
- new_op.read_offsets[0] = Shape4D.from_list(offset_start, 0)
- new_op.read_shapes[0] = read_shape
- new_op.run_on_npu = True
- new_op.set_output_tensor(tens)
- new_op.ifm_shapes.append(Shape4D(inp.shape))
- new_op.ofm_shapes.append(split_op.ofm_shapes[ofm_shape_idx])
- DebugDatabase.add_optimised(split_op, new_op)
-
- return tens
-
-
-def remove_SplitSliceRead(op, arch):
-
- if op.type == Op.SplitSliceRead:
- # Check if it is possible to put the SplitSliceRead on the tensor consumer, or if an avgpool need to be inserted
- if (
- len(op.ofm.consumer_list) == 1
- and op.ofm.consumer_list[0] is not None
- and op.ofm.consumer_list[0].run_on_npu
- and op.ofm.consumer_list[0].type != Op.Reshape
- and op.ofm_shapes[0] == Shape4D.from_list(op.ofm.shape)
- ):
- # SplitSliceRead can be performed by tensor consumer
- cons_op = op.ofm.consumer_list[0]
- if cons_op.ifm == op.ofm:
- cons_op.read_offsets[0] = op.read_offsets[0]
- cons_op.read_shapes[0] = op.read_shapes[0]
- cons_op.set_input_tensor(op.ifm, cons_op.type.info.indices.ifms[0])
- cons_op.ifm_shapes[0] = op.ifm_shapes[0]
- elif cons_op.type.is_binary_elementwise_op() and cons_op.ifm2 == op.ofm:
- cons_op.read_offsets[1] = op.read_offsets[0]
- cons_op.read_shapes[1] = op.read_shapes[0]
- cons_op.set_input_tensor(op.ifm, cons_op.type.info.indices.ifms[1])
- cons_op.ifm_shapes[1] = op.ifm_shapes[0]
-
- if "skirt" in cons_op.attrs:
- assert cons_op.attrs["explicit_padding"] == cons_op.attrs["skirt"]
- cons_op.attrs["skirt"] = None
- cons_op.attrs["force_padding"] = True
- op.ofm.consumer_list.remove(cons_op)
- op.ofm.ops = []
- op.ifm.consumer_list.remove(op)
- else:
- avgpool_op = create_avgpool_nop(op.name + "_avgpool")
- avgpool_op.add_input_tensor(op.ifm)
- avgpool_op.outputs = [op.ofm]
- op.ofm.ops.remove(op)
- op.ofm.ops.append(avgpool_op)
- avgpool_op.ifm_shapes.append(op.ifm_shapes[0])
- avgpool_op.ofm_shapes.append(op.ofm_shapes[0])
- avgpool_op.read_offsets[0] = op.read_offsets[0]
- avgpool_op.read_shapes[0] = op.read_shapes[0]
-
- op.ifm.consumer_list.remove(op)
- DebugDatabase.add_optimised(op, avgpool_op)
-
-
-def avoid_nhcwb16_for_concat(tens):
- # If axis corresponds to C-dimension, NHCWB16 can only be used in the output if all the concat_start's are a
- # multiple of 16. This as, it is only then the address offset for the ofm, for all operations, will be 16 byte
- # aligned. For other values of axis the address offsets will be 16 byte aligned, as they are all based on c = 0
- # and those addresses are always 16 byte aligned due to the NHCWB16 format.
- return any(op.write_offset.depth % 16 != 0 for op in tens.ops if op.write_offset is not None)
-
-
-def avoid_nhcwb16_for_split(tens):
- # If read offset is not a multiple of 16 in the C-dimension, NHCWB16 need to be avoided in the input
- for cons_op in tens.consumer_list:
- if cons_op.ifm == tens:
- read_offset = cons_op.read_offsets[0]
- elif cons_op.type.is_binary_elementwise_op() and cons_op.ifm2 == tens:
- read_offset = cons_op.read_offsets[1]
- else:
- assert False
- if read_offset is not None and (read_offset[-1] % 16) != 0:
- return True
- return False
-
-
-def avoid_nhcwb16_for_shapes(tens):
- # check all producers/consumers to see if any op shape is preventing NHCWB16
- for cons_op in tens.consumer_list:
- if cons_op.ifm == tens:
- cons_op_shape = cons_op.ifm_shapes[0]
- elif cons_op.type.is_binary_elementwise_op() and cons_op.ifm2 == tens:
- cons_op_shape = cons_op.ifm_shapes[1]
- else:
- assert False
- if Shape4D(tens.shape) != cons_op_shape:
- return True
-
- for prod_op in tens.ops:
- if Shape4D(tens.shape) != prod_op.ofm_shapes[0]:
- return True
-
- return False
-
-
-# Check if non linear format can be used
-def check_format_restrictions(tens, arch):
- if len(tens.ops) < 1:
- return
- if tens.ops[0].type in (Op.Placeholder, Op.SubgraphInput, Op.Const) or any(
- cons is None for cons in tens.consumer_list
- ):
- return
-
- # Check if any of the producers/consumers is run on CPU
- if not all(cons.run_on_npu for cons in tens.consumer_list):
- return
- if not all(prod.run_on_npu for prod in tens.ops):
- return
-
- # "Concat" ofm exception:
- if avoid_nhcwb16_for_concat(tens):
- return
-
- # "Split" ifm exception:
- if avoid_nhcwb16_for_split(tens):
- return
-
- # Shapes checking: check all producers/consumers are NHCWB16 compatible with tens.shape
- if avoid_nhcwb16_for_shapes(tens):
- return
-
- for op in tens.consumer_list:
- if op.type == Op.ReduceSum and tens.dtype == DataType.int32:
- return
- if op.type == Op.Reshape:
- # Using NHCWB16 format for a no-op reshape is only an option if subsequent
- # consumers do not also need to perform a reshape or if the OFM is going to
- # be processed by CPU operations. No-op reshape consumers with empty lists
- # (those that have no consumers, or null-consumers used as list terminators)
- # must use normal NHWC output.
-
- def incompatible_consumers(oper):
- if oper and oper.type == Op.Reshape:
- for consumer in oper.outputs[0].consumer_list:
- yield from incompatible_consumers(consumer)
- yield not oper or not oper.run_on_npu
-
- if not any(incompatible_consumers(op)):
-
- def get_rewrites(oper):
- if oper and oper.type == Op.Reshape:
- for consumer in oper.outputs[0].consumer_list:
- yield from get_rewrites(consumer)
- yield oper
-
- # Detect no-op reshapes by comparing their full input and output tensor shapes.
- inshape = op.ifm_shapes[0]
- compatible_shape = [(inshape == oper.ofm_shapes[0]) for oper in get_rewrites(op)]
- if not (compatible_shape and all(compatible_shape)):
- return
- else:
- return
-
- tens.needs_linear_format = False
-
-
-def insert_copy_op_after_tens(tens):
- tens_cons_list_copy = tens.consumer_list.copy()
-
- # Create a avg_pool nop op with ifm as input
- copy_tens = tens.clone()
- copy_op = create_avgpool_nop(tens.name + "_avgpool")
- copy_op.add_input_tensor(tens)
- copy_op.set_output_tensor(copy_tens)
- copy_op.set_ifm_ofm_shapes()
- copy_op.run_on_npu = True
-
- # Set copy_ifm consumers
- for tens_cons in tens_cons_list_copy:
- if tens_cons is not None:
- for ifm_idx, cons_inp in enumerate(tens_cons.inputs):
- if cons_inp == tens:
- tens_cons.set_input_tensor(copy_tens, ifm_idx)
-
- DebugDatabase.add_optimised(tens.ops[0], copy_op)
-
-
-def fix_sg_input_output(op, arch, nng):
- if not op.run_on_npu or op.type != Op.Reshape:
- return op
-
- # For the Reshape operators we want to remove, tensors are removed.
- # But in order to to do this, they cannot be outputs of the sg,
- # this need to be fixed prior to the removal.
- # Solution is to add a avgpool NOP, to maintain the original tensor.
- # This is also valid when reshape ifm/ofm is produced respectively
- # consumed by CPU
-
- # Check if operator ifm/ofm are sg ifm/ofm
- ifm_is_sg_ifm = op.ifm.ops[0].type in (Op.Placeholder, Op.SubgraphInput, Op.Const)
- ifm_is_sg_ofm = any(ifm_cons is None for ifm_cons in op.ifm.consumer_list)
- ofm_is_sg_ofm = any(ofm_cons is None for ofm_cons in op.ofm.consumer_list)
- # Check if ifm/ofm is produced repectivly consumed by CPU
- ifm_is_cpu_produced = any(ifm_prod is not None and not ifm_prod.run_on_npu for ifm_prod in op.ifm.ops)
- ofm_is_cpu_consumed = any(ofm_cons is not None and not ofm_cons.run_on_npu for ofm_cons in op.ofm.consumer_list)
-
- if (ifm_is_sg_ofm or ifm_is_sg_ifm or ifm_is_cpu_produced) and (ofm_is_sg_ofm or ofm_is_cpu_consumed):
- # Both ifm and ofm need to persist, but only ifm need a copy, in order to remove the Reshape
- insert_copy_op_after_tens(op.ifm)
-
- return op
-
-
-def needed_total_padding(input_size, stride, filter_size):
- out_size = (input_size + stride - 1) // stride
- needed_input = (out_size - 1) * stride + filter_size
- total_padding = max(0, needed_input - input_size)
- return total_padding
-
-
-def calc_explicit_padding(input_size, stride, filter_size, pad_before, pad_after) -> Tuple[int, int]:
- """
- Based on explicit padding provided in a PAD operation, returns the corresponding hardware padding
- that provides equivalent results.
- """
- total_padding = needed_total_padding(input_size, stride, filter_size)
- # The top/left padding can be taken as is from the PAD
- output_pad_before = pad_before
- # The bottom/right padding might need downward adjustment depending on stride/input size
- output_pad_after = pad_after
- while output_pad_after > 0 and output_pad_after % stride != (total_padding - pad_before) % stride:
- output_pad_after -= 1
- return output_pad_before, output_pad_after
-
-
-def calc_padding_and_skirt(padding_type, kernel, input_shape, explicit_padding):
- k_w, k_h = kernel.dilated_wh()
- s_x, s_y = kernel.stride
- ypad = needed_total_padding(int(input_shape.height), int(s_y), int(k_h))
- xpad = needed_total_padding(int(input_shape.width), int(s_x), int(k_w))
- if padding_type == Padding.SAME:
- left_pad = (xpad + 0) // 2
- right_pad = (xpad + 1) // 2
- top_pad = (ypad + 0) // 2
- bottom_pad = (ypad + 1) // 2
- elif padding_type == Padding.VALID:
- left_pad = 0
- right_pad = 0
- top_pad = 0
- bottom_pad = 0
- elif padding_type == Padding.EXPLICIT:
- # Padding is specified in a PAD operator which has been bypassed.
- top, left, bottom, right = explicit_padding
- top_pad, bottom_pad = calc_explicit_padding(int(input_shape.height), int(s_y), int(k_h), int(top), int(bottom))
- left_pad, right_pad = calc_explicit_padding(int(input_shape.width), int(s_x), int(k_w), int(left), int(right))
- else:
- raise UnsupportedFeatureError(f"Unknown padding")
- padding = (top_pad, left_pad, bottom_pad, right_pad)
- skirt = (top_pad, left_pad, ypad - top_pad, xpad - left_pad)
- return padding, skirt
-
-
-def calc_upscaled_padding_and_skirt(padding_type, kernel_size, stride, input_shape, upscaling_factor):
- kernel_height, kernel_width = kernel_size[0], kernel_size[1]
- if padding_type == Padding.SAME:
- ypad = needed_total_padding(int(input_shape.height) * upscaling_factor, int(stride[1]), int(kernel_height))
- xpad = needed_total_padding(int(input_shape.width) * upscaling_factor, int(stride[2]), int(kernel_width))
- right_pad = max(((xpad + 1) // upscaling_factor) - 1, 0)
- bottom_pad = max(((ypad + 1) // upscaling_factor) - 1, 0)
- left_pad = max(kernel_width - 1 - right_pad, 0)
- top_pad = max(kernel_height - 1 - bottom_pad, 0)
- elif padding_type == Padding.VALID:
- right_pad = max(kernel_width - 2, 0)
- bottom_pad = max(kernel_height - 2, 0)
- left_pad = kernel_width - 1
- top_pad = kernel_height - 1
- else:
- raise UnsupportedFeatureError(f"Unknown padding")
- padding = (top_pad, left_pad, bottom_pad, right_pad)
- skirt = padding
- return padding, skirt
-
-
-def fixup_conv2d_backprop(op, arch, nng):
- if op.type == Op.Conv2DBackpropInput:
- # flip the inputs
- op.inputs[0], op.inputs[2] = op.inputs[2], op.inputs[0]
- op.type = Op.Conv2DBackpropInputSwitchedBias
- op.ifm.resampling_mode = resampling_mode.TRANSPOSE
-
- # Update strides
- op.attrs.update({"stride_w": 1, "stride_h": 1, "strides": (1, 1, 1, 1)})
-
- return op
-
-
-# Convert the op to an elementwise add
-def convert_resizebilinear_1x1_to_add(op):
- op.type = Op.Add
- op.name = op.name + "_add"
- op.attrs["resizebilinear"] = True
- # 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.quantization = QuantizationParameters(0.0, 255.0)
- tens.quantization.scale_f32 = 1.0
- tens.quantization.zero_point = 0
- tens.consumer_list = [op]
- tens_op = op.inputs[1].ops[0]
- tens_op.set_output_tensor(tens)
- # Set the add inputs
- op.inputs[1] = op.inputs[0]
- op.inputs[0] = tens
- op.set_ifm_ofm_shapes()
-
- return op
-
-
-# Convert ResizeBilinear to a number of 2x2 pool ops
-def convert_resizebilinear_to_2x2_pool(op):
- count = 0
- pre_op = op
- outputs = op.outputs
-
- op.attrs.update({"strides": (1, 1, 1, 1), "ksize": (1, 2, 2, 1)})
- if op.attrs["align_corners"]:
- shape_modifier = 1
- op.attrs["padding"] = Padding.VALID
- else:
- shape_modifier = 0
- op.attrs["padding"] = Padding.SAME
- op.inputs[0].resampling_mode = resampling_mode.NEAREST
-
- upscaled_shape = np.array(op.ifm_shapes[0].get_hw_as_list())
- out_shape = np.array(op.ofm_shapes[0].get_hw_as_list())
- if (upscaled_shape == upscaled_shape * 2 - shape_modifier).all():
- return op
-
- while (upscaled_shape < out_shape).all():
- if count == 0:
- scaled_op = pre_op
- else:
- scaled_op = op.clone("_{}".format(count))
- scaled_op.inputs[0] = pre_op.outputs[0]
-
- upscaled_shape = upscaled_shape * 2 - shape_modifier
-
- if (upscaled_shape == out_shape).all():
- scaled_op.outputs = outputs
- scaled_op.outputs[0].ops = [scaled_op]
- else:
- shape = op.ofm_shapes[0].as_list()
- shape[1:3] = upscaled_shape
- out_tens = Tensor(shape, DataType.int16, "{}_{}".format(op.outputs[0].name, count))
- out_tens.quantization = op.outputs[0].quantization.clone()
- out_tens.quantization.quant_min = np.iinfo(np.int16).min
- out_tens.quantization.quant_max = np.iinfo(np.int16).max
- scaled_op.set_output_tensor(out_tens)
- pre_op = scaled_op
- count += 1
-
- # Setup the scale value
- if scaled_op.inputs[0].dtype.bits == 8 and scaled_op.outputs[0].dtype.bits == 16:
- scaled_op.rescale = 128
- elif scaled_op.inputs[0].dtype.bits == 16 and scaled_op.outputs[0].dtype.bits == 8:
- scaled_op.rescale = 1 / 128
- else:
- scaled_op.rescale = None
- scaled_op.set_ifm_ofm_shapes()
-
- return op
-
-
-def fixup_resizebilinear(op, arch, nng):
- if op.type == Op.ResizeBilinear and op.run_on_npu:
- if op.ifm_shapes[0] == op.ofm_shapes[0]:
- # Bypass nop resizebilinear
- op.inputs = op.inputs[:1]
- op.type = Op.Identity
- elif op.ifm_shapes[0].height == 1 and op.ifm_shapes[0].width == 1:
- convert_resizebilinear_1x1_to_add(op)
- else:
- convert_resizebilinear_to_2x2_pool(op)
-
- return op
-
-
-def convert_nop_split_to_identity(op, arch, nng):
- if op.type == Op.Split and op.attrs.get("num_splits") == 1:
- # the list comprehension should return a list with a single tensor
- # if it shouldn't, remove_passthrough_tensor will fail appropriately
- op.inputs = [i for i in op.inputs if i.shape == op.outputs[0].shape]
- op.type = Op.Identity
- return op
-
-
-def rewrite_fully_connected_input(op, arch, nng):
- if op.type == Op.FullyConnected:
- n_in_elems = op.weights.shape[-2]
- elms = op.ifm.elements()
- batch_size = elms // n_in_elems
- assert batch_size * n_in_elems == elms
-
- op.ifm_shapes[0] = Shape4D([batch_size, 1, 1, n_in_elems])
- return op
-
-
-def convert_batched_fc_shape(op, arch, nng):
- if op.type == Op.FullyConnected:
- # Check if the first dimension indicates batching
- if op.ifm_shapes[0].batch > 1:
- batching_split = {4: (2, 2), 8: (2, 4), 16: (4, 4)}
- n = op.ifm_shapes[0].batch
- h, w = batching_split.get(n, (1, n))
- op.ifm_shapes[0] = Shape4D([1, h, w, op.ifm_shapes[0].depth])
-
- # 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))
-
- n = op.ofm_shapes[0].batch
- h, w = batching_split.get(n, (1, n))
- op.ofm_shapes[0] = Shape4D([1, h, w, op.ofm_shapes[0].depth])
- return op
-
-
-def unfuse_activation_function(op):
- if op.type == Op.ConcatTFLite and op.run_on_npu and op.activation is not None:
- act_op = Operation(op.activation.op_type, op.name + op.activation.op_type.name)
- op.activation = None
- out_tens = op.outputs[0]
- intermediate_tens = out_tens.clone("_act_intermediate")
- act_op.set_output_tensor(out_tens)
- act_op.add_input_tensor(intermediate_tens)
- op.set_output_tensor(intermediate_tens)
- act_op.set_ifm_ofm_shapes()
-
-
-def rewrite_stridedslice_output(op, arch, nng):
- if not op.run_on_npu or op.type != Op.StridedSlice:
- return op
-
- new_axis_mask = op.attrs["new_axis_mask"]
- shrink_axis_mask = op.attrs["shrink_axis_mask"]
-
- if shrink_axis_mask == 0 and new_axis_mask == 0:
- return op
-
- axis_4D = [0] * len(op.outputs)
- for idx, out_tens in enumerate(op.outputs):
- output_shape = list(out_tens.shape)
-
- if shrink_axis_mask != 0:
- n = 0
- axis = 0
- while shrink_axis_mask:
- prev_mask = shrink_axis_mask
- n += 1
- shrink_axis_mask &= shrink_axis_mask - 1
- axis = int(math.log2(prev_mask - shrink_axis_mask))
- output_shape = output_shape[:axis] + [1] + output_shape[axis:]
-
- assert len(out_tens.shape) == (len(op.inputs[0].shape) - n)
- op.attrs["shrink_axis_mask"] = 0
- if axis >= 0:
- axis_4D[idx] = axis + (4 - len(output_shape))
- else:
- axis_4D[idx] = axis
- op.ofm_shapes[idx] = Shape4D(output_shape)
-
- elif new_axis_mask != 0:
- n = 0
- axis = 0
- while new_axis_mask:
- prev_mask = new_axis_mask
- n += 1
- new_axis_mask &= new_axis_mask - 1
- axis = int(math.log2(prev_mask - new_axis_mask))
- output_shape = output_shape[:axis] + output_shape[(axis + 1) :]
- new_axis_mask >>= 1
-
- assert len(out_tens.shape) == (len(op.inputs[0].shape) + n)
- op.attrs["new_axis_mask"] = 0
- if axis >= 0:
- axis_4D[idx] = axis + (4 - len(output_shape))
- else:
- axis_4D[idx] = axis
- op.ofm_shapes[idx] = Shape4D(output_shape)
-
- op.attrs["split_axis_4D"] = axis_4D
- return op
-
-
-def rewrite_unpack_output(op, arch, nng):
- tens = op.outputs[0]
- if op.run_on_npu and op.type == Op.Unpack:
- # Unpack is also referred to as Unstack
- axis = int(op.attrs["axis"])
- if axis < 0: # Convert to positive axis
- axis = len(op.inputs[0].shape) + 1 + axis
- op.type = Op.UnpackReshaped
- desired_output_shape = tens.shape[:axis] + [1] + tens.shape[axis:]
-
- axis_4D = axis + (4 - len(desired_output_shape))
- op.attrs["split_axis_4D"] = [axis_4D] * len(op.outputs)
-
- for idx, out_tens in enumerate(op.outputs):
- op.ofm_shapes[idx] = Shape4D(desired_output_shape)
- return op
-
-
-def add_padding_fields(op, arch, nng):
- if op.run_on_npu:
- if "padding" in op.attrs:
- input_shape = op.ifm_shapes[0]
- output_shape = op.ofm_shapes[0]
- if op.type.is_conv2d_op() or op.type.is_depthwise_conv2d_op():
- kernel_size = op.inputs[1].shape[:2]
- elif op.type.is_pool_op() or op.type.npu_block_type == NpuBlockType.ReduceSum:
- kernel_size = op.attrs["ksize"][1:3]
- else:
- raise UnsupportedFeatureError(f"Unknown operation that uses padding: {optype_to_builtintype(op.type)}")
-
- if op.type == Op.Conv2DBackpropInputSwitchedBias:
- upscaling_factor = output_shape.height // input_shape.height
- padding, skirt = calc_upscaled_padding_and_skirt(
- op.attrs["padding"], kernel_size, op.attrs["strides"], input_shape, upscaling_factor
- )
- else:
- padding, skirt = calc_padding_and_skirt(
- op.attrs["padding"], op.kernel, input_shape, op.attrs.get("explicit_padding"),
- )
-
- op.attrs["explicit_padding"] = padding
- op.attrs["skirt"] = skirt
-
- return op
-
-
-def convert_depthwise_to_conv(op, arch, nng):
- # Depthwise is equivalent to a single conv2d if the ifm depth is 1 and
- # the ofm depth equals the depth multipler.
- # If those conditions are true, then we can perform a simple
- # switch of the operator type (and weight order)
-
- if op.type == Op.DepthwiseConv2DBias and (op.attrs["depth_multiplier"] != 1):
- ifm_shape = op.ifm_shapes[0]
- weight_tensor = op.inputs[1]
- ofm_shape = op.ofm_shapes[0]
- if (ifm_shape.depth == 1) and (ofm_shape.depth == op.attrs["depth_multiplier"]):
- # Change op type to Conv2d
- op.type = Op.Conv2DBias
- 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))
- else:
- raise UnsupportedFeatureError(
- f"Unsupported 'DEPTHWISE_CONV_2D' with depth_multiplier = {op.attrs['depth_multiplier']},",
- f" ifm channels = {ifm_shape.depth}, ofm channels = {ofm_shape.depth}",
- )
- DebugDatabase.add_optimised(op, op)
- return op
-
-
-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.weight_transpose_depthwise = True
-
- return op
-
-
-def optimise_strided_conv(op, arch, nng):
- stride_x, stride_y = op.get_kernel_stride()
- ifm_tensor, _, weight_tensor, _ = op.get_ifm_ifm2_weights_ofm()
-
- if (
- op.type == Op.Conv2DBias
- and op.op_index == 0
- and stride_x == 2
- and op.ifm_shapes[0].depth <= 4
- and op.ifm_shapes[0].width % 2 == 0
- and weight_tensor is not None
- and weight_tensor.shape[1] >= 2
- ):
- ifm_shape = op.ifm_shapes[0]
- # IFM
- op.ifm_shapes[0] = Shape4D([ifm_shape.batch, ifm_shape.height, ifm_shape.width // 2, ifm_shape.depth * 2])
-
- # Weights
- weight_shape = weight_tensor.shape
- if weight_shape[1] % 2 != 0:
- weight_shape[1] = weight_shape[1] + 1
- padded_array = np.zeros(weight_shape)
- for i in range(weight_shape[0]):
- padded_array[i] = np.vstack(
- [
- weight_tensor.quant_values[i],
- np.full((1, weight_shape[2], weight_shape[3]), weight_tensor.quantization.zero_point),
- ]
- )
- weight_tensor.quant_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.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
- weight_tensor.value_id = uuid.uuid4()
-
- # Strides
- stride_x = 1
- op.attrs.update({"stride_w": stride_x, "stride_h": stride_y, "strides": (1, stride_y, stride_x, 1)})
-
- return op
-
-
-def convert_conv_to_fc(op, arch, nng):
- # Conv 1x1 can be equivalent to Fully Connected.
- # By representing certain convs as fully connected layers, Vela can better determine wether or not to use
- # caching/double buffering for the weights.
- # (Weights dont need to be reloaded for convs when IFM H and W are 1)
- if op.type == Op.Conv2DBias:
- h = op.ifm_shapes[0].height
- w = op.ifm_shapes[0].width
- kh, kw, _, _ = op.inputs[1].shape
- if h == 1 and w == 1 and kh == 1 and kw == 1:
- # Overwrite this op as a Fully Connected Op
- op.name += "_fc"
- op.type = Op.FullyConnected
- op.attrs = {
- "weights_format": 0,
- }
- # 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))
-
- DebugDatabase.add_optimised(op, op)
- return op
-
-
-def fixup_relus_with_differing_ifm_ofm_scaling(op, arch, nng):
- if op.run_on_npu and op.type.is_relu_op():
- ifm = op.inputs[0]
- ofm = op.outputs[0]
- # Relu with differing IFM and OFM scaling cannot be fused with another primary op
- # and requires its own to be inserted
- if not check_quantized_tens_scaling_equal(ifm, ofm):
- # Override this op with its own primary op (avgpool)
- relu_fused_op = create_avgpool_nop(op.name + "_avgpool")
- # And fuse the original activation function to it
- relu_fused_op.activation = create_activation_function(op.type)
- # Tidy up and assign the ifm and ofm to the new op
- ifm.consumer_list.remove(op)
-
- relu_fused_op.add_input_tensor(ifm)
- relu_fused_op.set_output_tensor(ofm)
- relu_fused_op.set_ifm_ofm_shapes()
- op = relu_fused_op
- return op
-
-
-def fixup_elementwise_with_scalars(op, arch, nng):
- if op.type.is_binary_elementwise_op():
- ifm_tensor, ifm2_tensor, _, _ = op.get_ifm_ifm2_weights_ofm()
- if ifm2_tensor.shape != [] and ifm_tensor.shape != []:
- diff = len(ifm_tensor.shape) - len(ifm2_tensor.shape)
- if diff > 0:
- 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:
- # 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:
- # 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
- return op
-
-
-# Set input/output tensor equivalence to the same id for memory operations
-def set_tensor_equivalence(op, arch, nng):
- if op.type in memory_only_ops:
- eid = op.outputs[0].equivalence_id
- for inp in op.inputs:
- inp.equivalence_id = eid
- return op
-
-
-def set_ifm_ofm_op_shapes(op, arch, nng):
- if op.run_on_npu and op.type.needs_shapes():
- if op.ifm_shapes or op.ofm_shapes:
- # Shapes already set
- return op
- op.set_ifm_ofm_shapes()
- return op
-
-
-def convert_softmax(op, arch, nng):
- if op.type == Op.Softmax and op.run_on_npu:
- softmax = SoftMax(op)
- op = softmax.get_graph()
- return op
-
-
-def convert_mul_max_to_abs_or_lrelu(op, arch, nng):
- r"""Whenever there is a subgraph with this topology:
-
- Input X For X = -1 or X > 0
- | \ / This subgraph can be replaced with either
- | Mul an Abs (if X = -1) or a LeakyReLU (if X > 0)
- | /
- Max
- """
-
- if op.type == Op.Maximum:
- # finds the Mul input(s) to the Max
- muls = [i for i in op.inputs if i.ops[0].type == Op.Mul]
- if len(muls) == 1:
- mul = muls[0].ops[0]
- elif len(muls) == 2:
- # In the case both inputs are Muls, find the one with the same input as the Max
- mul = [m for m in muls if len(set(op.inputs + m.ops[0].inputs)) == 1][0].ops[0]
- else:
- # No Mul inputs
- return op
-
- # make sure the Mul doesn't have any other consumers
- mul_ofm = mul.outputs[0]
- if len(mul_ofm.consumers()) != 1:
- return op
- # make sure the Mul doesn't have a fused activation function
- if mul.activation:
- return op
- ifm, ofm = op.get_ifm_ofm()
- if ifm is None or ofm is None:
- return op
-
- if ifm.dtype not in (DataType.uint8, DataType.int8) or ifm.dtype != ofm.dtype:
- return op
- if not check_quantized_tens_scaling_equal(ifm, ofm) or not check_quantized_tens_scaling_equal(ifm, mul_ofm):
- # rewrite to LeakyRelu currently only makes sense if the quantization is identical
- return op
-
- # finds the branched input that goes to both the Max and the Mul
- shared = set(op.inputs) & set(mul.inputs)
- if len(shared) == 1:
- shared_in = shared.pop()
- # find the constant scalar input to the Mul
- const_tens = (set(mul.inputs) - {shared_in}).pop()
- # check that it is a scalar
- if const_tens.shape != []:
- return op
- const = const_tens.ops[0]
- # check that it is a constant
- if const.type != Op.Const:
- return op
- # Remove the Mul from the shared input's consumers
- shared_in.consumer_list.remove(mul)
- else:
- return op
-
- val = const.outputs[0].values
- if val >= 0:
- new_op = Op.LeakyRelu
- op.attrs["alpha"] = val
- # 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
- 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)
- alpha_scale, alpha_shift = scaling.elementwise_mul_scale(mul_ifm_scale, mul_ifm2_scale, mul_ofm_scale)
- op.attrs["alpha_scaling"] = (alpha_scalar, alpha_scale, alpha_shift)
- elif val == -1:
- new_op = Op.Abs
- else:
- return op
-
- op.type = new_op
- op.name = op.name.replace("Maximum", new_op.name)
- op.outputs[0].name = op.outputs[0].name.replace("Maximum", new_op.name)
- op.inputs = [shared_in]
- op.set_ifm_ofm_shapes()
-
- # Record optimisation in debug database
- DebugDatabase.add_optimised(op, op)
-
- return op
-
-
-def convert_hardswish_to_lut(op, arch, nng):
- if op.type == Op.HardSwish:
- ifm, ofm = op.get_ifm_ofm()
- # Generate the LUT
- ifm_scale = np.double(ifm.quantization.scale_f32)
- ofm_scale = np.double(ofm.quantization.scale_f32)
- zp_in = ifm.quantization.zero_point
- zp_out = ofm.quantization.zero_point
- ifm_scale_hires = (1 / 128) * ifm_scale
- relu_multiplier = np.double(3 / 32768)
- out_scale, out_shift = scaling.quantise_scale(ifm_scale_hires / ofm_scale)
- relu_scale, relu_shift = scaling.quantise_scale(ifm_scale_hires / relu_multiplier)
- # Use 16bit scale
- out_scale_16 = fp_math.downscale_multiplier_int32_to_int16(out_scale)
- relu_scale_16 = fp_math.downscale_multiplier_int32_to_int16(relu_scale)
-
- values = []
- ix = range(256) if ifm.dtype == DataType.uint8 else range(-128, 128)
- quantized_min = min(ix)
- quantized_max = max(ix)
- for x in ix:
- input_value = x - zp_in
- input_value_hires = input_value * 128
- # Compute the input value on essentially the output scale, not shifted yet
- input_value_preshift = fp_math.saturating_rounding_mul16(input_value_hires, out_scale_16)
- # Compute the "relu-ish multiplier". This matches the code in TensorFlow Lite Micro kernel
- relu_value = np.int16(input_value_hires)
- if relu_shift < 31:
- relu_value = fp_math.shift_left16(relu_value, 30 - relu_shift)
-
- relu_value = fp_math.saturating_rounding_mul16(relu_value, relu_scale_16)
-
- if relu_shift < 31:
- relu_value = fp_math.shift_left16(relu_value, 1)
-
- if relu_shift > 31:
- relu_value = fp_math.rounding_divide_by_pot(relu_value, relu_shift - 31)
-
- # Rescaled the value into a 16bit fixedpoint relu_value in [-1, 1]
- # Now convert that to a 16bit fixedpoint value in [0, 1]
- relu_value = (relu_value + (1 << 15)) >> 1
- lut_result = fp_math.saturating_mul16(relu_value, input_value_preshift)
- shift = 31 - out_shift
- shift = -shift if shift < 0 else 0
- # Finally apply the output shift
- lut_result = fp_math.rounding_divide_by_pot(lut_result, shift) + zp_out
- lut_result = min(quantized_max, max(quantized_min, lut_result))
- values.append(lut_result)
- return convert_to_lut(op, values, "hardswish")
- return op
-
-
-def convert_lrelu_to_mul_max(op, arch):
- # Converts LeakyRelu to Max(alpha * IFM, identity * IFM)
- # (the opposite of convert_mul_max_to_abs_or_lrelu)
- ifm, ofm = op.get_ifm_ofm()
- if ifm is None or ofm is None:
- return op
-
- # Add multiplication with alpha
- mul_alpha = Operation(Op.Mul, op.name + "_mul_alpha")
- mul_alpha.add_input_tensor(ifm)
- # Create const tensor containing alpha as scalar
- alpha = op.attrs["alpha"]
- quantization = ifm.quantization.clone()
- quantization.min = 0
- quantization.max = alpha * (quantization.quant_max - quantization.quant_min)
- quantization.zero_point = 0
- if np.isinf(1 / np.float32(alpha)):
- # Handling of alpha near zero
- quantization.scale_f32 = 1
- scalar = 0
- else:
- quantization.scale_f32 = alpha
- scalar = alpha
- alpha_tens = create_const_tensor(
- op.name + "_alpha_scalar", [], ifm.dtype, [scalar], np.float32, quantization=quantization
- )
- alpha_tens.quant_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)
- mul_alpha.set_ifm_ofm_shapes()
- DebugDatabase.add_optimised(op, mul_alpha)
-
- if check_quantized_tens_scaling_equal(ifm, ofm):
- # No identity multiplication is needed
- fm_id = ifm
- else:
- # Add multiplication with identity
- mul_identity = Operation(Op.Mul, op.name + "_mul_identity")
- mul_identity.add_input_tensor(ifm)
- # Create const tensor containing identity as scalar
- quantization = ifm.quantization.clone()
- quantization.min = 0
- quantization.max = quantization.quant_max - quantization.quant_min
- quantization.scale_f32 = 1
- quantization.zero_point = 0
- identity_tens = create_const_tensor(
- op.name + "_id_scalar", [], ifm.dtype, [1], np.uint8, quantization=quantization
- )
- mul_identity.add_input_tensor(identity_tens)
- # Make sure that fm_id is allocated to a different address than fm_alpha
- fm_id = ofm.clone(op.name + "_id", set_unique=True)
- mul_identity.set_output_tensor(fm_id)
- mul_identity.set_ifm_ofm_shapes()
- DebugDatabase.add_optimised(op, mul_identity)
-
- # Convert LeakyRelu to Max, add the results of the multiplication(s) as inputs
- op.type = Op.Maximum
- op.name = op.name.replace("LeakyRelu", "Maximum")
- op.inputs = []
- ifm.consumer_list.remove(op)
- op.add_input_tensor(fm_alpha)
- op.add_input_tensor(fm_id)
- op.set_ifm_ofm_shapes()
-
- DebugDatabase.add_optimised(op, op)
- return op
-
-
-def convert_to_lut(op, lut_values, lut_name):
- # Rewrite the operation by Add with scalar 0 + LUT activation
- ifm = op.inputs[0]
- if ifm is None:
- return op
- assert ifm.dtype.size_in_bytes() == 1
- op.type = Op.Add
- op.name = op.name + "_lut_" + lut_name
- # Mark as no-op to enable potential fusing optimizations
- op.attrs["is_nop"] = True
- # Create an input tensor containing scalar zero
- quantization = QuantizationParameters(0.0, 255.0)
- quantization.scale_f32 = ifm.quantization.scale_f32
- quantization.zero_point = 0
- tens = create_const_tensor(op.inputs[0].name + "_scalar0", [], ifm.dtype, [0], np.uint8, quantization=quantization)
- op.add_input_tensor(tens)
- op.ifm_shapes.append(Shape4D(tens.shape))
-
- # The LUT must be applied without any preceding rescaling (the LUT itself performs the rescale),
- # so even if the OFM has a different scale than the IFM, the generated OFM scale instructions
- # should be the same as the IFM
- op.forced_output_quantization = ifm.quantization
- lut_tensor = lut.create_lut_tensor(op.name + "_values", lut_values, DataType.int8)
- op.set_activation_lut(lut_tensor)
- op.set_ifm_ofm_shapes()
- return op
-
-
-def convert_to_lut8(op, fn, fn_name):
- # Converts op to a no-op + int8/uint8 LUT which is generated with the given function.
- # fn is a function(real) -> real
- ifm, ofm = op.get_ifm_ofm()
- if ifm.dtype not in (DataType.uint8, DataType.int8) or ifm.dtype != ofm.dtype:
- return op
- # Generate the LUT
- ifm_scale = np.double(ifm.quantization.scale_f32)
- ofm_scale = np.double(ofm.quantization.scale_f32)
- zp_in = ifm.quantization.zero_point
- zp_out = ofm.quantization.zero_point
- values = []
- ix = range(256) if ifm.dtype == DataType.uint8 else range(-128, 128)
- quantized_min = min(ix)
- quantized_max = max(ix)
- for x in ix:
- x_real = ifm_scale * (x - zp_in)
- y_real = fn(x_real)
- lut_result = round_away_zero(zp_out + y_real / ofm_scale)
- lut_result = min(quantized_max, max(quantized_min, lut_result))
- values.append(lut_result)
- return convert_to_lut(op, values, fn_name)
-
-
-def convert_lrelu_to_lut(op, arch):
- ifm, ofm = op.get_ifm_ofm()
- # Generate the LUT
- alpha = op.attrs["alpha"]
- ifm_scale = np.double(ifm.quantization.scale_f32)
- ofm_scale = np.double(ofm.quantization.scale_f32)
- zp_in = ifm.quantization.zero_point
- zp_out = ofm.quantization.zero_point
- identity_scale, identity_shift = scaling.elementwise_mul_scale(ifm_scale, 1, ofm_scale)
- alpha_scalar = 1
- alpha_scale, alpha_shift = scaling.elementwise_mul_scale(ifm_scale, alpha, ofm_scale)
- if "alpha_scaling" in op.attrs:
- # The LeakyRelu was the result from convert_mul_max_to_abs_or_lrelu
- alpha_scalar, alpha_scale, alpha_shift = op.attrs["alpha_scaling"]
- values = []
- ix = range(256) if ifm.dtype == DataType.uint8 else range(-128, 128)
- quantized_min = min(ix)
- quantized_max = max(ix)
- for x in ix:
- if x < zp_in:
- lut_result = zp_out + fp_math.multiply_by_quantized_multiplier(
- alpha_scalar * (x - zp_in), alpha_scale, alpha_shift
- )
- else:
- lut_result = zp_out + fp_math.multiply_by_quantized_multiplier(x - zp_in, identity_scale, identity_shift)
- lut_result = min(quantized_max, max(quantized_min, lut_result))
- values.append(lut_result)
- return convert_to_lut(op, values, "lrelu")
-
-
-def convert_lrelu(op, arch, nng):
- # Converts LeakyRelu to a LUT based solution if possible, otherwise a mul + max
- if op.type != Op.LeakyRelu:
- return op
- ifm, ofm = op.get_ifm_ofm()
- if ifm is None or ofm is None:
- return op
- if ifm.dtype in (DataType.uint8, DataType.int8) and ifm.dtype == ofm.dtype:
- # use LUT for int8/uint8
- return convert_lrelu_to_lut(op, arch)
- if check_quantized_tens_scaling_equal(ifm, ofm) and ifm.dtype == ofm.dtype == DataType.int16:
- # use LeakyRelu unmodified for int16 with equal input/output scaling
- return op
- return convert_lrelu_to_mul_max(op, arch)
-
-
-def convert_tanh_sigmoid_to_lut(op, arch, nng):
- # Converts int8/uint8 Sigmoid and Tanh to a LUT based solution
- if op.type == Op.Sigmoid:
- return convert_to_lut8(op, clamp_sigmoid, "sigmoid")
- elif op.type == Op.Tanh:
- return convert_to_lut8(op, math.tanh, "tanh")
- return op
-
-
-def remove_reshapes(op, arch):
- if op.run_on_npu and op.type == Op.Reshape:
- ofm = op.ofm
- ifm = op.ifm
-
- # Check if quantization is the same in the input and output for the reshape ops
- if not check_quantized_tens_scaling_equal(ifm, ofm):
- # TODO Both tensors are needed, since quantisation properties currently are linked to Tensors.
- # In order to remove this reshape either quantization properties need to be moved to Operator,
- # or the reshape need to be replace with a NOP.
- return
-
- # Check if Reshape ifm/ofm are network ifm/ofm
- ifm_is_sg_ifm = ifm.ops[0].type in (Op.Placeholder, Op.SubgraphInput, Op.Const)
- ifm_is_sg_ofm = any(ifm_cons is None for ifm_cons in ifm.consumer_list)
- ofm_is_sg_ofm = any(ofm_cons is None for ofm_cons in ofm.consumer_list)
- # Check if ifm/ofm is produced repectivly consumed by CPU
- ifm_is_cpu_produced = any(ifm_prod is not None and not ifm_prod.run_on_npu for ifm_prod in op.ifm.ops)
- ofm_is_cpu_consumed = any(ofm_cons is not None and not ofm_cons.run_on_npu for ofm_cons in op.ofm.consumer_list)
-
- # This case should be handled prior to this function
- assert not ((ifm_is_sg_ifm or ifm_is_sg_ofm or ifm_is_cpu_produced) and (ofm_is_sg_ofm or ofm_is_cpu_consumed))
-
- if ofm_is_sg_ofm or ofm_is_cpu_consumed:
- # Bypassed by replacing ifm with ofm
- ofm.ops = []
- for prev_op in ifm.ops:
- prev_op.outputs = [ofm]
- ofm.ops.append(prev_op)
-
- # All ifm consumers need to use ofm as input
- for ifm_cons in ifm.consumer_list:
- for ifm_idx, cons_ifm in enumerate(ifm_cons.inputs):
- if cons_ifm == ifm:
- ifm_cons.set_input_tensor(ofm, ifm_idx)
- else:
- # Bypassed Reshape by replacing ofm with ifm
- for cons in ofm.consumer_list:
- for ifm_idx, cons_ifm in enumerate(cons.inputs):
- if cons_ifm == ofm:
- cons.set_input_tensor(ifm, ifm_idx)
-
-
-def check_reshapes(op, arch):
- if op.run_on_npu and op.type == Op.Reshape:
- ofm = op.ofm
-
- if check_quantized_tens_scaling_equal(op.ifm, ofm):
- # Reshape should have been removed
- raise VelaError(f"Reshape op {op} expected to have been removed, still remains")
-
-
-def fuse_activation_function_with_prev(op, arch, nng):
- # if op is a no-op: attempts to move the activation function to the preceding op
- if not op.attrs.get("is_nop", False) or op.activation is None:
- return op
- ifm, ofm = op.get_ifm_ofm()
- if ifm is None or ofm is None:
- return op
- # finds the input(s) to the operation
- prev_op = ifm.ops[0]
- # Note: the below checks on prev_op require that a first optimize pass on the full graph has been performed
- fuse = (
- prev_op.run_on_npu
- and prev_op.type.npu_block_type != NpuBlockType.Default
- and len(ifm.ops) == 1
- and len(prev_op.outputs[0].consumers()) == 1
- and prev_op.activation is None
- )
- if op.activation_lut is not None and arch.shram_reserved_unused_banks == 0:
- # TODO: if SHRAM LUT space is shared with SHRAM ACC (32, 64 MAC),
- # LUT currently only works correctly for elementwise ops
- fuse = False
- if not fuse:
- return op
- # Move the fused activation function + corresponding info to prev_op
- prev_op.activation = op.activation
- prev_op.forced_output_quantization = op.forced_output_quantization
- if op.activation_lut is not None:
- prev_op.set_activation_lut(op.activation_lut)
- # Bypass op
- prev_op.set_output_tensor(ofm)
- DebugDatabase.add_optimised(op, prev_op)
- return op
-
-
-def _leading_pad_ok(leading_pad, stride, kernel_size):
- # If kernel size // 2 > stride, then (left, top) padding must be a multiple of stride,
- # otherwise replacing PAD by hardware padding would iterate the wrong IFM rows/columns
- max_size = kernel_size // 2
- return leading_pad == max_size or max_size <= stride or leading_pad % stride == 0
-
-
-def replace_pad_by_hw_pad(op: Operation, arch, nng):
- """
- Tries to completely remove a PAD operator by using hardware padding.
- E.g. a PAD operation that pads 1, followed by a CONV with VALID padding and kernel size 3
- is rewritten such that the PAD is removed, and the CONV uses SAME padding.
- Converts tens1 -> PAD -> tens2 -> CONV to tens1 -> CONV
- if both operations can be run on the NPU.
- This is the most efficient way to implement PAD, but cannot be done for all pad sizes.
- """
- if (
- (op.type.is_conv2d_op() or op.type.is_depthwise_conv2d_op() or op.type.is_avgpool_op())
- and op.run_on_npu
- and op.attrs["padding"] == Padding.VALID
- ):
- pad_op = op.ifm.ops[0]
- if pad_op.type != Op.Pad or not pad_op.run_on_npu:
- return op
- if pad_op.ifm.dtype != pad_op.ofm.dtype or not check_quantized_tens_scaling_equal(pad_op.ofm, pad_op.ifm):
- return op
- top, left, bottom, right = get_pad_values_from_input(pad_op.inputs[1].values)
- k = op.kernel
- k_w, k_h = k.dilated_wh()
-
- # Check if the PAD operator can be replaced by hardware padding
- if left > k_w // 2 or right > k_w // 2 or top > k_h // 2 or bottom > k_h // 2:
- # Too much padding, it would require hardware padding to actually insert zeros
- return op
- if not _leading_pad_ok(top, k.stride.y, k_h) or not _leading_pad_ok(left, k.stride.x, k_w):
- return op
-
- if op.type.is_avgpool_op():
- # For average pool, hardware padding can only be used if padding is 0 or kernel size / 2
- for pad, k_size in (
- (left, k_w),
- (right, k_w),
- (top, k_h),
- (bottom, k_h),
- ):
- if pad not in (0, k_size // 2):
- return op
- # Average pool is converted to depthwise, because NPU average pool + same padding
- # has a special implementation that is different from PAD followed by average pool with
- # valid padding.
- k_w, k_h = op.kernel.width, op.kernel.height
- ifm = op.ifm
- # Remember other inputs
- other_inputs = op.inputs[1:]
- # Create a weight tensor, all weights are set to 1/(kernel width * kernel height)
- quantization = QuantizationParameters(0.0, 255.0)
- quantization.scale_f32 = 1.0 / (k_w * k_h)
- quantization.zero_point = 0
- shape = [k_h, k_w, 1, op.ofm.shape[-1]]
- weights = np.full(shape, 1)
-
- weight_tens = create_const_tensor(
- op.name + "_weights",
- shape,
- op.ifm.dtype,
- weights,
- np.uint8,
- purpose=TensorPurpose.Weights,
- quantization=quantization,
- )
- weight_tens.quant_values = weights
- op.type = Op.DepthwiseConv2DBias
- op.inputs = []
- op.add_input_tensor(ifm)
- op.add_input_tensor(weight_tens)
- # Add bias tensor, all biases set to 0
- op.inputs.append(None)
- fixup_bias_tensors(op, arch, nng)
- # Add other inputs
- op.inputs.extend(other_inputs)
- op.rounding_mode = NpuRoundingMode.NATURAL
-
- # Bypass the PAD operator
- op.set_input_tensor(pad_op.ifm, 0)
- # Adjust the padding attributes of the convolution operator
- op.attrs["padding"] = Padding.EXPLICIT
- op.attrs["explicit_padding"] = (top, left, bottom, right)
- op.set_ifm_ofm_shapes()
- return op
-
-
-def convert_pad(op: Operation, arch, nng):
- """
- Rewrites PAD operator to an average pool that copies the IFM to the OFM
- + up to 4 average pool operators that fill the OFM with zeros at the borders.
- This is done as fall-back for the PAD operators that remain after replace_pad_by_hw_pad
- """
- if op.type != Op.Pad or not op.run_on_npu:
- return op
- top, left, bottom, right = get_pad_values_from_input(op.inputs[1].values)
-
- ifm = op.ifm
- assert ifm is not None
- ifm_shape = Shape4D(ifm.shape)
- ofm = op.ofm
- assert ofm is not None
- ofm.ops = []
- ofm_shape = op.ofm_shapes[0]
-
- # Average pool op that copies IFM to the right place inside the OFM
- shp0 = Shape4D(0, 0, 0, 0)
- shp_top = shp0.with_height(top)
- avgpool_op = create_avg_pool_for_concat(op, op.name + "_main", ifm, ifm_shape, shp_top.with_width(left))
- avgpool_op.activation = op.activation
- quant = ofm.quantization
- pad_value = quant.zero_point
- # Add operations that fill the borders of the OFM
- if top > 0:
- shape = Shape4D(1, top, ofm_shape.width, ofm_shape.depth)
- zero_tens = create_const_tensor(
- op.name + "_top", shape.as_list(), ofm.dtype, shape.elements() * [pad_value], np.uint8, quantization=quant
- )
- # If top/bottom or left/right are equal, the const tensors can be allocated to the same address
- zero_tens.equivalence_id = create_equivalence_id(tuple(zero_tens.values))
- create_avg_pool_for_concat(op, op.name + "_top", zero_tens, shape, shp0)
- if bottom > 0:
- shape = Shape4D(1, bottom, ofm_shape.width, ofm_shape.depth)
- zero_tens = create_const_tensor(
- op.name + "_bottom",
- shape.as_list(),
- ofm.dtype,
- shape.elements() * [pad_value],
- np.uint8,
- quantization=quant,
- )
- zero_tens.equivalence_id = create_equivalence_id(tuple(zero_tens.values))
- create_avg_pool_for_concat(
- op, op.name + "_bottom", zero_tens, shape, shp0.with_height(ofm_shape.height - bottom)
- )
- if left > 0:
- shape = Shape4D(1, ifm_shape.height, left, ofm_shape.depth)
- zero_tens = create_const_tensor(
- op.name + "_left", shape.as_list(), ofm.dtype, shape.elements() * [pad_value], np.uint8, quantization=quant
- )
- zero_tens.equivalence_id = create_equivalence_id(tuple(zero_tens.values))
- create_avg_pool_for_concat(op, op.name + "_left", zero_tens, shape, shp_top)
- if right > 0:
- shape = Shape4D(1, ifm_shape.height, right, ofm_shape.depth)
- zero_tens = create_const_tensor(
- op.name + "_right", shape.as_list(), ofm.dtype, shape.elements() * [pad_value], np.uint8, quantization=quant
- )
- zero_tens.equivalence_id = create_equivalence_id(tuple(zero_tens.values))
- create_avg_pool_for_concat(
- op, op.name + "_right", zero_tens, shape, shp_top.with_width(ofm_shape.width - right)
- )
-
- op.type = Op.ConcatTFLite
- return avgpool_op
-
-
-def add_attrs_to_resizebilinear(op, arch, nng):
- if op.type == Op.ResizeBilinear and op.run_on_npu:
- input_tensor = op.inputs[0]
- input_shape = op.ifm_shapes[0]
- upscaled_height = input_shape.height * 2
- upscaled_width = input_shape.width * 2
- out_shape = op.ofm_shapes[0]
- if not op.attrs["align_corners"] and out_shape.height == upscaled_height and out_shape.width == upscaled_width:
- # this means the output is supposed to be a x2 upscale,
- # so we need to do SAME padding
- op.attrs["padding"] = Padding.SAME
- elif (
- op.attrs["align_corners"]
- and out_shape.height == (upscaled_height - 1)
- and out_shape.width == (upscaled_width - 1)
- ):
- # here we can just run the avg pool without padding and
- # produce a (M * 2 - 1, N * 2 - 1) sized output
- op.attrs["padding"] = Padding.VALID
- else:
- return op
- input_tensor.resampling_mode = resampling_mode.NEAREST
- op.attrs.update({"strides": (1, 1, 1, 1), "ksize": (1, 2, 2, 1)})
- return op
-
-
-def fixup_bias_tensors(op, arch, nng):
- if op.type.needs_bias() and op.bias is None:
- # Op has no bias, add bias tensor filled with zeros
- 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
-
-
-def convert_mean_to_depthwise_conv_or_avgpool(op, arch, nng):
- if op.type == Op.Mean and op.run_on_npu:
- keep_dims = op.attrs.get("keep_dims", False)
- inp, axis = op.inputs
- shape = inp.shape
- dims = len(shape)
-
- # Height and width axes have different index depending on dimensions
- if axis.shape == [] or axis.shape[0] == 1: # single axis
- axis = int(axis.values) if len(axis.shape) == 0 else int(axis.values[0])
- if dims in (2, 3):
- if axis == 0:
- h, w = shape[axis], 1
- else:
- h, w = 1, shape[axis]
- else:
- if axis == 1:
- h, w = shape[axis], 1
- else:
- h, w = 1, shape[axis]
- else: # multiple axes
- axis = sorted(axis.values)
- h, w = [shape[i] for i in axis]
-
- # Set necessary depthwise attributes
- op.attrs.update(
- {
- "padding": Padding.VALID,
- "stride_h": 1,
- "stride_w": 1,
- "strides": (1, 1, 1, 1),
- "depth_multiplier": 1,
- "channel_multiplier": 1,
- "dilation_h_factor": 1,
- "dilation_w_factor": 1,
- "dilation": (1, 1, 1, 1),
- }
- )
- # Change op type
- op.type = Op.DepthwiseConv2DBias
- # Set IFM/OFM shapes after changing op type
- op.set_ifm_ofm_shapes()
-
- weight_scale, bias = 1, None
- ofmq, ifmq = op.ofm.quantization, inp.quantization
- # Set rounding mode, scaling and zero point based on which reference implementation to match
- if len(shape) == 4 and axis == [1, 2] and keep_dims:
- if inp.dtype == DataType.uint8:
- # This attribute means a different scaling calculation is used in order to match reference
- op.low_precision_scaling = True
- weight_scale = h * w
- # Set zero points to 0 as they will be adjusted for with bias term
- foq = ofmq.clone()
- foq.zero_point = 0
- fiq = ifmq.clone()
- fiq.zero_point = 0
- op.forced_input_quantization = fiq
- bias_term = ofmq.zero_point - int(ifmq.zero_point * ifmq.scale_f32 / ofmq.scale_f32)
- # If the bias term is outside uint8 range, we need an Add op to apply it.
- if bias_term < 0 or bias_term > 255:
- intermediate = op.ofm.clone(suffix="_intermediate", set_unique=True)
- # Bias term has higher bitness (i32) than input/output (u8).
- # 16 bits is enough since the bias is added/subtracted from a u8 value,
- # the bias can only effectively assume values in the range [-255, 255].
- intermediate.dtype = DataType.int16
- intermediate.quantization.zero_point = 0
- add_op = Operation(Op.Add, op.name + "_bias")
- add_op.forced_output_quantization = foq
- add_op.add_input_tensor(intermediate)
- 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,
- )
- add_op.add_input_tensor(bias_term_tens)
- add_op.set_output_tensor(op.ofm)
- add_op.set_ifm_ofm_shapes()
- add_op.activation = op.activation
- op.activation = None
- op.set_output_tensor(intermediate)
- op.set_ifm_ofm_shapes()
- # If not, we can just do it with the OFM zero point.
- else:
- foq.zero_point = bias_term
- op.forced_output_quantization = foq
- else:
- assert inp.dtype == DataType.int8
- # Use a depthwise to calculate the sum,
- # followed by a multiplication with 1/N to get the MEAN
- weight_scale = 1
- intermediate = op.ofm.clone(suffix="_intermediate", set_unique=True)
- intermediate.dtype = DataType.int16
- mul_op = Operation(Op.Mul, op.name + "_mul")
- mul_op.add_input_tensor(intermediate)
- # Create scalar containing 1/N
- quant = QuantizationParameters()
- quant.zero_point = 0
- # The reference rounds negative numbers downwards, e.g. -1.5 is rounded to -2,
- # while rounding mode NATURAL would round this to -1.
- # This can only occur if N is even, and can be emulated by
- # multiplying with a number that is slightly smaller than 1/N.
- # It must be so small that other roundings are not affected;
- # the calculated value is based on worst case,
- # which is sum 256 * N (the maximum sum that can occur with int8)
- n = int(h * w)
- eps = 1 / (256 * (n + 1)) if n % 2 == 0 else 0
- quant.scale_f32 = 1 / (n - eps)
- scalar = create_const_tensor(
- op.name + "_scalar", [1, 1, 1, 1], DataType.uint8, [1], np.uint8, quantization=quant
- )
- mul_op.add_input_tensor(scalar)
- mul_op.set_output_tensor(op.ofm)
- mul_op.set_ifm_ofm_shapes()
- mul_op.rounding_mode = NpuRoundingMode.NATURAL
- mul_op.activation = op.activation
- op.activation = None
- op.set_output_tensor(intermediate)
- op.set_ifm_ofm_shapes()
- elif ifmq.zero_point == ofmq.zero_point and ifmq.scale_f32 == ofmq.scale_f32:
- # Here we can just use a simple AvgPool with truncating rounding,
- # as we're emulating simple integer division.
- op.rounding_mode = NpuRoundingMode.TRUNCATE
- op.type = Op.AvgPool
- op.attrs.update({"ksize": (1, h, w, 1), "filter_height": h, "filter_width": w})
- else:
- op.rounding_mode = NpuRoundingMode.NATURAL
- weight_scale = 1 / (h * w)
- # Input zero point is adjusted after mean calculation, so we emulate that with a bias
- bias = -ifmq.zero_point * h * w
- fiq = ifmq.clone()
- fiq.zero_point = 0
- op.forced_input_quantization = fiq
-
- # Change dimensions to 4
- if dims < 4:
- shape = [1] + shape
- if dims == 2:
- shape += [1]
-
- # If height is greater than max kernel height, reshape to from HxW to 1x(HxW)
- if h > 64:
- shape = [shape[0], 1, h * w, shape[3]]
- op.ifm_shapes[0] = Shape4D(shape)
- if h > 256 and op.type == Op.AvgPool:
- op.attrs.update({"ksize": (1, 1, h * w, 1), "filter_height": 1, "filter_width": h * w})
-
- # If the AvgPool version is used, we don't need to do anything else
- if op.type == Op.AvgPool:
- return op
-
- # Make unit weight tensor quantization
- weight_quant = ifmq.clone()
- weight_quant.min = 0
- weight_quant.max = 255
- weight_quant.scale_f32 = weight_scale
- weight_quant.zero_point = 0
-
- # Set weight shape to [H,W,C,B]
- weight_shape = shape[1:4] + [shape[0]]
- # Add unit weight tensor
- op.set_input_tensor(
- create_const_tensor(
- "weights",
- weight_shape,
- inp.dtype,
- np.ones(weight_shape),
- value_dtype=np.uint8,
- quantization=weight_quant,
- ),
- 1,
- )
- op.weights.quant_values = np.reshape(op.inputs[1].quant_values, weight_shape)
-
- # Add None bias tensor
- op.inputs.append(None)
- # Add bias tensor
- if bias:
- bias_shape = [shape[-1]]
- op.set_input_tensor(
- create_const_tensor(
- "bias",
- bias_shape,
- inp.dtype,
- np.ones(bias_shape) * bias,
- value_dtype=np.int32,
- quant_value_dtype=np.int32,
- quantization=None,
- ),
- 2,
- )
-
- return op
-
-
-def supported_operator_check(op, arch, nng):
- op.run_on_npu = arch.supported_operators.is_operator_supported(op)
- return op
-
-
-def _record_optimised(op, arch):
- if op.type != Op.Const:
- DebugDatabase.add_optimised(op, op)
-
-
-def optimise_graph_a(nng, arch, verbose_graph=False):
+def optimise_graph(nng, arch, network_type, verbose_graph=False):
if verbose_graph:
nng.print_graph("Before Graph Optimization")
- pre_process_list = [
- supported_operator_check,
- set_ifm_ofm_op_shapes,
- # TODO: memory-only Op removal
- ]
-
- for idx, sg in enumerate(nng.subgraphs):
- # rewrite graph pass
- nng.subgraphs[idx] = rewrite_graph.rewrite_graph_pre_order(
- nng, sg, arch, [], pre_process_list, rewrite_unsupported=False,
- )
-
- # Handle Concat Ops
- for idx, sg in enumerate(nng.subgraphs):
- # rewrite graph pass
- rewrite_graph.visit_graph_post_order(sg.output_tensors, arch, [], [rewrite_concat_ops])
- sg.refresh_after_modification()
-
- # Handle Split Ops
- for idx, sg in enumerate(nng.subgraphs):
- # rewrite graph pass
- nng.subgraphs[idx] = rewrite_graph.rewrite_graph_pre_order(
- nng,
- sg,
- arch,
- [],
- [rewrite_unpack_output, rewrite_stridedslice_output, convert_nop_split_to_identity],
- rewrite_unsupported=False,
- )
-
- for idx, sg in enumerate(nng.subgraphs):
- # rewrite graph pass
- nng.subgraphs[idx] = rewrite_graph.rewrite_graph_pre_order(
- nng, sg, arch, [rewrite_split_ops], [], rewrite_unsupported=False,
- )
-
- # Handle sg input output
- for idx, sg in enumerate(nng.subgraphs):
- nng.subgraphs[idx] = rewrite_graph.rewrite_graph_pre_order(
- nng, sg, arch, [], [fix_sg_input_output], rewrite_unsupported=False,
- )
-
- # Removal of reshapes
- for sg in nng.subgraphs:
- rewrite_graph.visit_graph_post_order(sg.output_tensors, arch, [], [remove_reshapes])
- sg.refresh_after_modification()
-
- op_rewrite_list = [
- set_tensor_equivalence,
- convert_mean_to_depthwise_conv_or_avgpool,
- convert_depthwise_to_conv,
- convert_conv_to_fc,
- convert_softmax,
- optimise_strided_conv,
- convert_hardswish_to_lut,
- rewrite_fully_connected_input,
- convert_batched_fc_shape,
- fixup_conv2d_backprop,
- fixup_relus_with_differing_ifm_ofm_scaling,
- fixup_elementwise_with_scalars, # TODO Move to early stage?
- reorder_depthwise_weights,
- fixup_resizebilinear,
- fixup_bias_tensors,
- convert_mul_max_to_abs_or_lrelu,
- convert_lrelu,
- convert_tanh_sigmoid_to_lut,
- replace_pad_by_hw_pad,
- ]
-
- for idx, sg in enumerate(nng.subgraphs):
- # rewrite graph pass
- nng.subgraphs[idx] = rewrite_graph.rewrite_graph_pre_order(
- nng, sg, arch, [], op_rewrite_list, rewrite_unsupported=False,
- )
-
- for idx, sg in enumerate(nng.subgraphs):
- # remove passthrough tensors and attempt further optimizations
- nng.subgraphs[idx] = rewrite_graph.rewrite_graph_pre_order(
- nng,
- sg,
- arch,
- [remove_passthrough_tensor],
- [fuse_activation_function_with_prev, convert_pad, add_padding_fields],
- )
-
- # Removal of SplitSliceRead, need to be done after optimisation has been performed,
- # since ifm/ofm_shapes are of importance to this function
- for sg in nng.subgraphs:
- rewrite_graph.visit_graph_post_order(sg.output_tensors, arch, [], [remove_SplitSliceRead])
- sg.refresh_after_modification()
-
- # Check Tensor Format restrictions
- for sg in nng.subgraphs:
- rewrite_graph.visit_graph_post_order(sg.output_tensors, arch, [check_format_restrictions], [])
- sg.refresh_after_modification()
+ if network_type == NetworkType.TFLite:
+ # TensorFlow Lite graph optimization
+ nng = tflite_optimise_graph(nng, arch)
+ else:
+ # TOSA graph optimization
+ nng = tosa_optimise_graph(nng, arch)
# Post-optimisation operator debug tracing, and checking that no undesired reshapes are left in the graph
for sg in nng.subgraphs:
- rewrite_graph.visit_graph_post_order(sg.output_tensors, arch, [], [check_reshapes, _record_optimised])
+ rewrite_graph.visit_graph_post_order(
+ sg.output_tensors, arch, [check_format_restrictions], [check_reshapes, record_optimised]
+ )
if verbose_graph:
nng.print_graph("After Graph Optimization")
diff --git a/ethosu/vela/graph_optimiser_util.py b/ethosu/vela/graph_optimiser_util.py
new file mode 100644
index 0000000..0b44b8f
--- /dev/null
+++ b/ethosu/vela/graph_optimiser_util.py
@@ -0,0 +1,168 @@
+# Copyright (C) 2021 Arm Limited or its affiliates. All rights reserved.
+#
+# SPDX-License-Identifier: Apache-2.0
+#
+# Licensed under the Apache License, Version 2.0 (the License); you may
+# not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an AS IS BASIS, WITHOUT
+# WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# Description:
+# Common functions and definitions used during the graph optimization.
+from .data_type import DataType
+from .debug_database import DebugDatabase
+from .errors import VelaError
+from .operation import Op
+from .shape4d import Shape4D
+from .tensor import check_quantized_tens_scaling_equal
+
+
+memory_only_ops = (Op.Reshape,)
+
+
+def _avoid_nhcwb16_for_concat(tens):
+ # If axis corresponds to C-dimension, NHCWB16 can only be used in the output if all the concat_start's are a
+ # multiple of 16. This as, it is only then the address offset for the ofm, for all operations, will be 16 byte
+ # aligned. For other values of axis the address offsets will be 16 byte aligned, as they are all based on c = 0
+ # and those addresses are always 16 byte aligned due to the NHCWB16 format.
+ return any(op.write_offset.depth % 16 != 0 for op in tens.ops if op.write_offset is not None)
+
+
+def _avoid_nhcwb16_for_split(tens):
+ # If read offset is not a multiple of 16 in the C-dimension, NHCWB16 need to be avoided in the input
+ for cons_op in tens.consumer_list:
+ if cons_op.ifm == tens:
+ read_offset = cons_op.read_offsets[0]
+ elif cons_op.type.is_binary_elementwise_op() and cons_op.ifm2 == tens:
+ read_offset = cons_op.read_offsets[1]
+ else:
+ assert False
+ if read_offset is not None and (read_offset[-1] % 16) != 0:
+ return True
+ return False
+
+
+def _avoid_nhcwb16_for_shapes(tens):
+ # check all producers/consumers to see if any op shape is preventing NHCWB16
+ for cons_op in tens.consumer_list:
+ if cons_op.ifm == tens:
+ cons_op_shape = cons_op.ifm_shapes[0]
+ elif cons_op.type.is_binary_elementwise_op() and cons_op.ifm2 == tens:
+ cons_op_shape = cons_op.ifm_shapes[1]
+ else:
+ assert False
+ if Shape4D(tens.shape) != cons_op_shape:
+ return True
+
+ for prod_op in tens.ops:
+ if Shape4D(tens.shape) != prod_op.ofm_shapes[0]:
+ return True
+
+ return False
+
+
+# Check if non linear format can be used
+def check_format_restrictions(tens, arch):
+ if len(tens.ops) < 1:
+ return
+ if tens.ops[0].type in (Op.Placeholder, Op.SubgraphInput, Op.Const) or any(
+ cons is None for cons in tens.consumer_list
+ ):
+ return
+
+ # Check if any of the producers/consumers is run on CPU
+ if not all(cons.run_on_npu for cons in tens.consumer_list):
+ return
+ if not all(prod.run_on_npu for prod in tens.ops):
+ return
+
+ # "Concat" ofm exception:
+ if _avoid_nhcwb16_for_concat(tens):
+ return
+
+ # "Split" ifm exception:
+ if _avoid_nhcwb16_for_split(tens):
+ return
+
+ # Shapes checking: check all producers/consumers are NHCWB16 compatible with tens.shape
+ if _avoid_nhcwb16_for_shapes(tens):
+ return
+
+ for op in tens.consumer_list:
+ if op.type == Op.ReduceSum and tens.dtype == DataType.int32:
+ return
+ if op.type == Op.Reshape:
+ # Using NHCWB16 format for a no-op reshape is only an option if subsequent
+ # consumers do not also need to perform a reshape or if the OFM is going to
+ # be processed by CPU operations. No-op reshape consumers with empty lists
+ # (those that have no consumers, or null-consumers used as list terminators)
+ # must use normal NHWC output.
+
+ def incompatible_consumers(oper):
+ if oper and oper.type == Op.Reshape:
+ for consumer in oper.outputs[0].consumer_list:
+ yield from incompatible_consumers(consumer)
+ yield not oper or not oper.run_on_npu
+
+ if not any(incompatible_consumers(op)):
+
+ def get_rewrites(oper):
+ if oper and oper.type == Op.Reshape:
+ for consumer in oper.outputs[0].consumer_list:
+ yield from get_rewrites(consumer)
+ yield oper
+
+ # Detect no-op reshapes by comparing their full input and output tensor shapes.
+ inshape = op.ifm_shapes[0]
+ compatible_shape = [(inshape == oper.ofm_shapes[0]) for oper in get_rewrites(op)]
+ if not (compatible_shape and all(compatible_shape)):
+ return
+ else:
+ return
+
+ tens.needs_linear_format = False
+
+
+def needed_total_padding(input_size, stride, filter_size):
+ out_size = (input_size + stride - 1) // stride
+ needed_input = (out_size - 1) * stride + filter_size
+ total_padding = max(0, needed_input - input_size)
+ return total_padding
+
+
+# Set input/output tensor equivalence to the same id for memory operations
+def set_tensor_equivalence(op, arch, nng):
+ if op.type in memory_only_ops:
+ eid = op.outputs[0].equivalence_id
+ for inp in op.inputs:
+ inp.equivalence_id = eid
+ return op
+
+
+def set_ifm_ofm_op_shapes(op, arch, nng):
+ if op.run_on_npu and op.type.needs_shapes():
+ if op.ifm_shapes or op.ofm_shapes:
+ # Shapes already set
+ return op
+ op.set_ifm_ofm_shapes()
+ return op
+
+
+def check_reshapes(op, arch):
+ if op.run_on_npu and op.type == Op.Reshape:
+ ofm = op.ofm
+
+ if check_quantized_tens_scaling_equal(op.ifm, ofm):
+ # Reshape should have been removed
+ raise VelaError(f"Reshape op {op} expected to have been removed, still remains")
+
+
+def record_optimised(op, arch):
+ if op.type != Op.Const:
+ DebugDatabase.add_optimised(op, op)
diff --git a/ethosu/vela/high_level_command_stream_generator.py b/ethosu/vela/high_level_command_stream_generator.py
index 6fcf80c..b3ea9d4 100644
--- a/ethosu/vela/high_level_command_stream_generator.py
+++ b/ethosu/vela/high_level_command_stream_generator.py
@@ -109,7 +109,9 @@
# Create activation function if needed
for op in ps.ops:
if op.type.is_relu_op() or op.type in (Op.Tanh, Op.Sigmoid):
- ps.primary_op.activation = create_activation_function(op.type)
+ ps.primary_op.activation = create_activation_function(
+ op.type, min=op.attrs.get("min", None), max=op.attrs.get("max", None)
+ )
# Generate commands for the Op that produces this Op's IFM, if applicable
if cascade_info is None or cascade_info.start == sched_op.index:
diff --git a/ethosu/vela/model_reader.py b/ethosu/vela/model_reader.py
index bb49b64..f48645d 100644
--- a/ethosu/vela/model_reader.py
+++ b/ethosu/vela/model_reader.py
@@ -16,7 +16,9 @@
# Description:
# Dispatcher for reading a neural network model.
from . import tflite_reader
+from . import tosa_reader
from .errors import InputFileError
+from .nn_graph import NetworkType
class ModelReaderOptions:
@@ -37,12 +39,33 @@
output_node_names = []
if initialisation_nodes is None:
initialisation_nodes = []
- return tflite_reader.read_tflite(
- fname,
- options.batch_size,
- feed_dict=feed_dict,
- output_node_names=output_node_names,
- initialisation_nodes=initialisation_nodes,
+ return (
+ tflite_reader.read_tflite(
+ fname,
+ options.batch_size,
+ feed_dict=feed_dict,
+ output_node_names=output_node_names,
+ initialisation_nodes=initialisation_nodes,
+ ),
+ NetworkType.TFLite,
+ )
+ elif fname.endswith(".tosa"):
+ if feed_dict is None:
+ feed_dict = {}
+ if output_node_names is None:
+ output_node_names = []
+ if initialisation_nodes is None:
+ initialisation_nodes = []
+
+ return (
+ tosa_reader.read_tosa(
+ fname,
+ options.batch_size,
+ feed_dict=feed_dict,
+ output_node_names=output_node_names,
+ initialisation_nodes=initialisation_nodes,
+ ),
+ NetworkType.TOSA,
)
else:
- raise InputFileError(fname, "Unsupported file extension. Only .tflite files are supported")
+ raise InputFileError(fname, "Unsupported file extension. Only .tflite and .tosa files are supported")
diff --git a/ethosu/vela/nn_graph.py b/ethosu/vela/nn_graph.py
index 7dc2d72..97afde3 100644
--- a/ethosu/vela/nn_graph.py
+++ b/ethosu/vela/nn_graph.py
@@ -44,6 +44,11 @@
return self.name
+class NetworkType(enum.Enum):
+ TFLite = 1
+ TOSA = 2
+
+
class Pass:
def __init__(self, name, placement, is_element_wise, npu_block_type):
self.inputs = []
diff --git a/ethosu/vela/operation.py b/ethosu/vela/operation.py
index 6bd955d..0558e52 100644
--- a/ethosu/vela/operation.py
+++ b/ethosu/vela/operation.py
@@ -238,6 +238,8 @@
Relu = OperatorInfo(indices=IFM_INDICES)
Relu6 = OperatorInfo(indices=IFM_INDICES)
ReluN1To1 = OperatorInfo(indices=IFM_INDICES)
+ ReluN = OperatorInfo(indices=IFM_INDICES) # TOSA specific
+ Rescale = OperatorInfo(indices=IFM_INDICES) # TOSA specific
RescaleAdd = OperatorInfo(block_type=NpuBlockType.ElementWise, indices=IFM_IFM2_INDICES)
Reshape = OperatorInfo(indices=IFM_INDICES)
ResizeBilinear = OperatorInfo(block_type=NpuBlockType.Pooling, indices=IFM_INDICES)
@@ -321,7 +323,7 @@
return self.info.block_type == NpuBlockType.ElementWise and not self.info.is_unary
def is_relu_op(self):
- return self in (Op.Relu, Op.Relu6, Op.ReluN1To1, Op.Clip)
+ return self in (Op.Relu, Op.Relu6, Op.ReluN1To1, Op.ReluN, Op.Clip)
def is_activation_op(self):
return self.is_relu_op() or self in (Op.Tanh, Op.Sigmoid, Op.Softmax, Op.LUT, Op.HardSwish)
@@ -374,7 +376,20 @@
return res
-def create_activation_function(op_type: Op) -> ActivationFunction:
+class ExplicitScaling:
+ """Explicit scaling parameters"""
+
+ def __init__(self, per_channel, shift, multiplier):
+ self.per_channel = per_channel
+ self.shift = shift
+ self.multiplier = multiplier
+
+ def clone(self):
+ res = copy.copy(self)
+ return res
+
+
+def create_activation_function(op_type: Op, min=None, max=None) -> ActivationFunction:
"""Creates activation function with min/max depending on op_type"""
act = ActivationFunction(op_type)
if op_type == Op.Relu:
@@ -393,6 +408,15 @@
act.max = 1.0
elif op_type == Op.HardSwish:
act.min = 0.0
+ if op_type == Op.Clip:
+ assert min is not None and max is not None
+ act.min = min
+ act.max = max
+ elif op_type == Op.ReluN:
+ assert max is not None
+ act.min = 0.0
+ act.max = max
+
return act
@@ -436,6 +460,7 @@
"read_offsets",
"read_shapes",
"rounding_mode",
+ "explicit_scaling",
"low_precision_scaling",
"write_offset",
"write_shape",
@@ -470,6 +495,8 @@
self.read_offsets: List[Shape4D] = [None, None] # offset for [ifm, ifm2]
self.read_shapes: List[Shape4D] = [None, None] # read shape for [ifm, ifm2]
self.rounding_mode: Optional[NpuRoundingMode] = None
+ # Rescale op in TOSA supplies explicit multiplier and shift values
+ self.explicit_scaling: Optional[ExplicitScaling] = None
# The Mean operator (implemented as a depthwise convolution) requires scaling
# to be calculated differently in one case. In that case, this is set to True.
self.low_precision_scaling = False
@@ -498,6 +525,7 @@
res.read_offsets = list(self.read_offsets)
res.read_shapes = list(self.read_shapes)
res.rounding_mode = self.rounding_mode
+ res.explicit_scaling = self.explicit_scaling
res.low_precision_scaling = self.low_precision_scaling
return res
diff --git a/ethosu/vela/reader_util.py b/ethosu/vela/reader_util.py
new file mode 100644
index 0000000..5b454b5
--- /dev/null
+++ b/ethosu/vela/reader_util.py
@@ -0,0 +1,60 @@
+# Copyright (C) 2021 Arm Limited or its affiliates. All rights reserved.
+#
+# SPDX-License-Identifier: Apache-2.0
+#
+# Licensed under the Apache License, Version 2.0 (the License); you may
+# not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an AS IS BASIS, WITHOUT
+# WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# Description:
+# Utlity function for reading .tosa and .tflite files
+from .operation import Op
+from .operation import Operation
+
+
+def decode_str(s):
+ if s is None:
+ return ""
+ return s.decode("utf-8")
+
+
+def clone_and_reshape_tensor(src_tens, reorder, set_unique):
+ tens = src_tens.clone("_reshape", set_unique)
+ tens.shape = [src_tens.shape[idx] for idx in reorder]
+ tens.bandwidth_shape = tens.shape
+ tens.storage_shape = tens.shape
+
+ 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
+
+
+# Fix up tensors without operations. Generate either Placeholder or Constant ops
+def fixup_tensors(input_tensors, tensors):
+ for tens in input_tensors:
+ if len(tens.ops) and tens.ops[0].type == Op.Const:
+ break
+
+ if tens.ops != []:
+ tens.error("This subgraph input tensor has unexpected driving operators.")
+
+ op = Operation(Op.Placeholder, tens.name)
+ op.set_output_tensor(tens)
+
+ for tens in tensors:
+ if not tens.ops:
+ op = Operation(Op.Const, tens.name)
+ op.set_output_tensor(tens)
diff --git a/ethosu/vela/supported_operators.py b/ethosu/vela/supported_operators.py
index dfa2719..c993da1 100644
--- a/ethosu/vela/supported_operators.py
+++ b/ethosu/vela/supported_operators.py
@@ -25,31 +25,19 @@
from .operation import get_slice_offsets
from .operation import Op
from .operation import Padding
+from .supported_operators_util import docstring_format_args
+from .supported_operators_util import list_formatter
from .tensor import check_quantized_tens_scaling_equal
from .tflite_mapping import BUILTIN_OPERATOR_UNKNOWN
from .tflite_mapping import optype_to_builtintype
-# Custom decorator function to allow formatting docstrings containing "{}"
-def docstring_format_args(args):
- def docstring(func):
- func.__doc__ = func.__doc__.format(*args)
- return func
-
- return docstring
-
-
-def _list_formatter(arg):
- # Order and join into a string representation
- return ", ".join(sorted(map(str, arg)))
-
-
def _optype_formatter(op_list):
# Convert internal op types to external names
output = map(optype_to_builtintype, op_list)
# Remove UNKNOWNs
output = (x for x in output if x is not BUILTIN_OPERATOR_UNKNOWN)
- return _list_formatter(output)
+ return list_formatter(output)
class SupportedOperators:
@@ -88,7 +76,8 @@
supported_int32_tensor_ops = (
set((Op.ReduceSum, Op.CLZ,)) | binary_elem_wise_add_mul_sub | binary_elem_wise_shift_ops
)
- relu_ops = Op.op_set(Op.is_relu_op)
+
+ relu_ops = set((Op.Relu, Op.Relu6, Op.ReluN1To1, Op.Clip,))
activation_ops = relu_ops | set((Op.Tanh, Op.Sigmoid, Op.Softmax, Op.HardSwish))
npu_post_ops = (
# activation functions
@@ -354,7 +343,7 @@
return valid, ", ".join(extra)
@classmethod
- @docstring_format_args([_list_formatter(supported_op_dtypes)])
+ @docstring_format_args([list_formatter(supported_op_dtypes)])
def constraint_tens_dtype(cls, op):
"Tensors must be of type: {}"
valid = True
@@ -463,7 +452,7 @@
return res
@classmethod
- @docstring_format_args([_list_formatter(supported_faf_dtypes)])
+ @docstring_format_args([list_formatter(supported_faf_dtypes)])
def constraint_faf_type(cls, op):
"If a fused activation function is present, the Output tensor must be one of type: {}"
if op.activation is None:
@@ -549,7 +538,7 @@
return valid, f"Tensor '{weights.name}' has the sum of weights: {limit}"
@classmethod
- @docstring_format_args([_list_formatter(supported_bias_dtypes)])
+ @docstring_format_args([list_formatter(supported_bias_dtypes)])
def constraint_bias_type(cls, op):
"Optional Bias tensor must be of type: {}"
bias = op.bias
@@ -832,7 +821,7 @@
return valid, f"The pad tensor has the shape: {op.inputs[1].shape}"
@classmethod
- @docstring_format_args([_list_formatter(supported_pad_dtypes)])
+ @docstring_format_args([list_formatter(supported_pad_dtypes)])
def constraint_pad_type(cls, op):
"Pad tensor must be of type: {}"
pad_tensor = op.inputs[1]
diff --git a/ethosu/vela/supported_operators_util.py b/ethosu/vela/supported_operators_util.py
new file mode 100644
index 0000000..24fe72c
--- /dev/null
+++ b/ethosu/vela/supported_operators_util.py
@@ -0,0 +1,31 @@
+# Copyright (C) 2021 Arm Limited or its affiliates. All rights reserved.
+#
+# SPDX-License-Identifier: Apache-2.0
+#
+# Licensed under the Apache License, Version 2.0 (the License); you may
+# not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an AS IS BASIS, WITHOUT
+# WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# Description:
+# Utility functions used in supported operator checking
+
+
+def list_formatter(arg):
+ # Order and join into a string representation
+ return ", ".join(sorted(map(str, arg)))
+
+
+# Custom decorator function to allow formatting docstrings containing "{}"
+def docstring_format_args(args):
+ def docstring(func):
+ func.__doc__ = func.__doc__.format(*args)
+ return func
+
+ return docstring
diff --git a/ethosu/vela/test/test_graph_optimiser.py b/ethosu/vela/test/test_graph_optimiser.py
index 83a3dda..b37bac8 100644
--- a/ethosu/vela/test/test_graph_optimiser.py
+++ b/ethosu/vela/test/test_graph_optimiser.py
@@ -15,17 +15,14 @@
# limitations under the License.
#
# Description:
-# Unit tests for graph_optimiser
+# Unit tests for tflite_graph_optimiser
import numpy as np
import pytest
from ethosu.vela.data_type import DataType
-from ethosu.vela.graph_optimiser import calc_explicit_padding
-from ethosu.vela.graph_optimiser import convert_batched_fc_shape
-from ethosu.vela.graph_optimiser import optimise_graph_a
-from ethosu.vela.graph_optimiser import replace_pad_by_hw_pad
-from ethosu.vela.graph_optimiser import rewrite_fully_connected_input
+from ethosu.vela.graph_optimiser import optimise_graph
from ethosu.vela.nn_graph import Graph
+from ethosu.vela.nn_graph import NetworkType
from ethosu.vela.operation import Op
from ethosu.vela.operation import Padding
from ethosu.vela.rewrite_graph import verify_graph_health
@@ -33,6 +30,10 @@
from ethosu.vela.tensor import Shape4D
from ethosu.vela.tensor import Tensor
from ethosu.vela.test import testutil
+from ethosu.vela.tflite_graph_optimiser import calc_explicit_padding
+from ethosu.vela.tflite_graph_optimiser import convert_batched_fc_shape
+from ethosu.vela.tflite_graph_optimiser import replace_pad_by_hw_pad
+from ethosu.vela.tflite_graph_optimiser import rewrite_fully_connected_input
def test_convert_batched_fc():
@@ -300,7 +301,7 @@
pool_op.run_on_npu = True
nng = testutil.create_graph([pad_op, pool_op])
arch = testutil.create_arch()
- nng = optimise_graph_a(nng, arch)
+ nng = optimise_graph(nng, arch, NetworkType.TFLite)
sg = nng.subgraphs[0]
all_ops = sg.get_all_ops()
print("all_ops: ", all_ops)
@@ -382,7 +383,7 @@
nng, reshape1_op, conv2d_op, reshape2_op = setup_network()
arch = testutil.create_arch()
assert verify_graph_health(nng)
- nng = optimise_graph_a(nng, arch)
+ nng = optimise_graph(nng, arch, NetworkType.TFLite)
assert verify_graph_health(nng)
# Test2 reshape1 with different quantisation, this Reshape op is expected to remain
@@ -393,5 +394,5 @@
quant_zp32.zero_point = 32
reshape1_op.ofm.quantization = quant_zp32
assert verify_graph_health(nng)
- nng = optimise_graph_a(nng, arch)
+ nng = optimise_graph(nng, arch, NetworkType.TFLite)
assert verify_graph_health(nng)
diff --git a/ethosu/vela/tflite_graph_optimiser.py b/ethosu/vela/tflite_graph_optimiser.py
new file mode 100644
index 0000000..3d9eeb8
--- /dev/null
+++ b/ethosu/vela/tflite_graph_optimiser.py
@@ -0,0 +1,1633 @@
+# Copyright (C) 2020-2021 Arm Limited or its affiliates. All rights reserved.
+#
+# SPDX-License-Identifier: Apache-2.0
+#
+# Licensed under the Apache License, Version 2.0 (the License); you may
+# not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an AS IS BASIS, WITHOUT
+# WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# Description:
+# Early optimisation of a TensorFlow Lite based network graph, using the rewrite_graph module
+# to do the traversal of the graph.
+import math
+import uuid
+from typing import Tuple
+
+import numpy as np
+
+from . import fp_math
+from . import lut
+from . import rewrite_graph
+from . import scaling
+from .api import NpuRoundingMode
+from .data_type import DataType
+from .debug_database import DebugDatabase
+from .errors import UnsupportedFeatureError
+from .ethos_u55_regs.ethos_u55_regs import resampling_mode
+from .graph_optimiser_util import needed_total_padding
+from .graph_optimiser_util import set_ifm_ofm_op_shapes
+from .graph_optimiser_util import set_tensor_equivalence
+from .numeric_util import clamp_sigmoid
+from .numeric_util import full_shape
+from .numeric_util import round_away_zero
+from .operation import create_activation_function
+from .operation import NpuBlockType
+from .operation import Op
+from .operation import Operation
+from .operation import Padding
+from .operation_util import create_avgpool_nop
+from .operation_util import get_pad_values_from_input
+from .shape4d import Shape4D
+from .softmax import SoftMax
+from .tensor import check_quantized_tens_scaling_equal
+from .tensor import create_const_tensor
+from .tensor import create_equivalence_id
+from .tensor import QuantizationParameters
+from .tensor import Tensor
+from .tensor import TensorPurpose
+from .tflite_mapping import optype_to_builtintype
+
+passthrough_nodes = (Op.Identity,)
+
+
+def create_avg_pool_for_concat(concat_op, name, ifm, ifm_shape: Shape4D, write_offset: Shape4D):
+ """Creates an average pool for the given concat op/input feature map"""
+ ofm = concat_op.ofm
+ avgpool_op = create_avgpool_nop(name)
+ avgpool_op.inputs = [ifm]
+ avgpool_op.outputs = [ofm]
+
+ avgpool_op.write_offset = write_offset
+ avgpool_op.write_shape = ifm_shape
+ ofm.ops.append(avgpool_op)
+ DebugDatabase.add_optimised(concat_op, avgpool_op)
+ avgpool_op.ifm_shapes.append(ifm_shape)
+ avgpool_op.ofm_shapes.append(concat_op.ofm_shapes[0])
+ avgpool_op.memory_function = Op.ConcatSliceWrite
+ return avgpool_op
+
+
+def remove_passthrough_tensor(tens, arch, nng):
+ if len(tens.ops) == 1 and tens.ops[0].type in passthrough_nodes:
+ assert len(tens.ops[0].inputs) == 1
+ tens = tens.ops[0].inputs[0]
+ return tens
+
+
+def rewrite_concat_ops(op, arch):
+ if not op.run_on_npu or not op.type.is_concat_op():
+ return
+
+ axis_4D = 0
+ ofm = op.ofm
+ ofm.ops = []
+ offset = 0
+
+ unfuse_activation_function(op)
+
+ if op.type == Op.Pack:
+ # Pack is also referred to as Stack
+ axis = int(op.attrs["axis"])
+ if axis < 0: # Convert to positive axis
+ axis = len(op.inputs[0].shape) + 1 + axis
+
+ desired_shape = op.inputs[0].shape[:axis] + [1] + op.inputs[0].shape[axis:]
+
+ axis_4D = axis + (4 - len(desired_shape))
+
+ for idx, inp in enumerate(op.inputs):
+ op.ifm_shapes[idx] = Shape4D(desired_shape)
+ op.type = Op.PackReshaped
+
+ inputs, axis = op.get_concat_inputs_axis()
+ for idx, inp in enumerate(inputs):
+ if op.type != Op.PackReshaped:
+ op.ifm_shapes[idx] = Shape4D(inp.shape)
+ if axis >= 0:
+ axis_4D = axis + (4 - len(inp.shape))
+ else:
+ axis_4D = axis
+ write_offset = [0, 0, 0, 0]
+ write_offset[axis_4D] = offset
+ concat_end = offset + op.ifm_shapes[idx][axis_4D]
+ create_avg_pool_for_concat(
+ op, op.name + str(idx) + "_avgpool", inp, op.ifm_shapes[idx], Shape4D.from_list(write_offset)
+ )
+ offset = concat_end
+ assert ofm.shape[axis] == offset
+
+ return op
+
+
+def rewrite_split_ops(tens, arch, nng):
+
+ if len(tens.ops) == 1 and tens.ops[0].type.is_split_op() and tens.ops[0].type != Op.Unpack:
+ split_op = tens.ops[0]
+
+ # Not supported so leave it and run on CPU
+ if not split_op.run_on_npu:
+ return tens
+
+ inp, outputs, axis, offset_start, offset_end = split_op.get_split_inputs_axis()
+
+ tens.ops = []
+ new_op = Operation(Op.SplitSliceRead, split_op.name)
+ new_op.inputs = [inp]
+ ofm_shape_idx = 0
+ read_shape = offset_end
+
+ # For Split the offset cannot be extracted from the tensor so it has to
+ # be calculated from the index of the output tensor
+ if axis is not None:
+ # Get the start and end of the split
+ offset_start = [0] * 4
+ axis_4D_list = split_op.attrs.get("split_axis_4D", None) # Present for UnpackReshaped and some StridedSlice
+ for idx, out in enumerate(outputs):
+ if axis_4D_list is not None:
+ axis_4D = axis_4D_list[idx]
+ else:
+ split_op.ofm_shapes[idx] = Shape4D(out.shape)
+ if axis >= 0:
+ axis_4D = axis + (4 - len(out.shape))
+ else:
+ axis_4D = axis
+
+ if out == tens:
+ ofm_shape_idx = idx
+ read_shape = split_op.ofm_shapes[idx]
+ break
+
+ offset_start[axis_4D] += split_op.ofm_shapes[idx][axis_4D]
+
+ new_op.read_offsets[0] = Shape4D.from_list(offset_start, 0)
+ new_op.read_shapes[0] = read_shape
+ new_op.run_on_npu = True
+ new_op.set_output_tensor(tens)
+ new_op.ifm_shapes.append(Shape4D(inp.shape))
+ new_op.ofm_shapes.append(split_op.ofm_shapes[ofm_shape_idx])
+ DebugDatabase.add_optimised(split_op, new_op)
+
+ return tens
+
+
+def remove_SplitSliceRead(op, arch):
+
+ if op.type == Op.SplitSliceRead:
+ # Check if it is possible to put the SplitSliceRead on the tensor consumer, or if an avgpool need to be inserted
+ if (
+ len(op.ofm.consumer_list) == 1
+ and op.ofm.consumer_list[0] is not None
+ and op.ofm.consumer_list[0].run_on_npu
+ and op.ofm.consumer_list[0].type != Op.Reshape
+ and op.ofm_shapes[0] == Shape4D.from_list(op.ofm.shape)
+ ):
+ # SplitSliceRead can be performed by tensor consumer
+ cons_op = op.ofm.consumer_list[0]
+ if cons_op.ifm == op.ofm:
+ cons_op.read_offsets[0] = op.read_offsets[0]
+ cons_op.read_shapes[0] = op.read_shapes[0]
+ cons_op.set_input_tensor(op.ifm, cons_op.type.info.indices.ifms[0])
+ cons_op.ifm_shapes[0] = op.ifm_shapes[0]
+ elif cons_op.type.is_binary_elementwise_op() and cons_op.ifm2 == op.ofm:
+ cons_op.read_offsets[1] = op.read_offsets[0]
+ cons_op.read_shapes[1] = op.read_shapes[0]
+ cons_op.set_input_tensor(op.ifm, cons_op.type.info.indices.ifms[1])
+ cons_op.ifm_shapes[1] = op.ifm_shapes[0]
+
+ if "skirt" in cons_op.attrs:
+ assert cons_op.attrs["explicit_padding"] == cons_op.attrs["skirt"]
+ cons_op.attrs["skirt"] = None
+ cons_op.attrs["force_padding"] = True
+ op.ofm.consumer_list.remove(cons_op)
+ op.ofm.ops = []
+ op.ifm.consumer_list.remove(op)
+ else:
+ avgpool_op = create_avgpool_nop(op.name + "_avgpool")
+ avgpool_op.add_input_tensor(op.ifm)
+ avgpool_op.outputs = [op.ofm]
+ op.ofm.ops.remove(op)
+ op.ofm.ops.append(avgpool_op)
+ avgpool_op.ifm_shapes.append(op.ifm_shapes[0])
+ avgpool_op.ofm_shapes.append(op.ofm_shapes[0])
+ avgpool_op.read_offsets[0] = op.read_offsets[0]
+ avgpool_op.read_shapes[0] = op.read_shapes[0]
+
+ op.ifm.consumer_list.remove(op)
+ DebugDatabase.add_optimised(op, avgpool_op)
+
+
+def insert_copy_op_after_tens(tens):
+ tens_cons_list_copy = tens.consumer_list.copy()
+
+ # Create a avg_pool nop op with ifm as input
+ copy_tens = tens.clone()
+ copy_op = create_avgpool_nop(tens.name + "_avgpool")
+ copy_op.add_input_tensor(tens)
+ copy_op.set_output_tensor(copy_tens)
+ copy_op.set_ifm_ofm_shapes()
+ copy_op.run_on_npu = True
+
+ # Set copy_ifm consumers
+ for tens_cons in tens_cons_list_copy:
+ if tens_cons is not None:
+ for ifm_idx, cons_inp in enumerate(tens_cons.inputs):
+ if cons_inp == tens:
+ tens_cons.set_input_tensor(copy_tens, ifm_idx)
+
+ DebugDatabase.add_optimised(tens.ops[0], copy_op)
+
+
+def fix_sg_input_output(op, arch, nng):
+ if not op.run_on_npu or op.type != Op.Reshape:
+ return op
+
+ # For the Reshape operators we want to remove, tensors are removed.
+ # But in order to to do this, they cannot be outputs of the sg,
+ # this need to be fixed prior to the removal.
+ # Solution is to add a avgpool NOP, to maintain the original tensor.
+ # This is also valid when reshape ifm/ofm is produced respectively
+ # consumed by CPU
+
+ # Check if operator ifm/ofm are sg ifm/ofm
+ ifm_is_sg_ifm = op.ifm.ops[0].type in (Op.Placeholder, Op.SubgraphInput, Op.Const)
+ ifm_is_sg_ofm = any(ifm_cons is None for ifm_cons in op.ifm.consumer_list)
+ ofm_is_sg_ofm = any(ofm_cons is None for ofm_cons in op.ofm.consumer_list)
+ # Check if ifm/ofm is produced repectivly consumed by CPU
+ ifm_is_cpu_produced = any(ifm_prod is not None and not ifm_prod.run_on_npu for ifm_prod in op.ifm.ops)
+ ofm_is_cpu_consumed = any(ofm_cons is not None and not ofm_cons.run_on_npu for ofm_cons in op.ofm.consumer_list)
+
+ if (ifm_is_sg_ofm or ifm_is_sg_ifm or ifm_is_cpu_produced) and (ofm_is_sg_ofm or ofm_is_cpu_consumed):
+ # Both ifm and ofm need to persist, but only ifm need a copy, in order to remove the Reshape
+ insert_copy_op_after_tens(op.ifm)
+
+ return op
+
+
+def calc_explicit_padding(input_size, stride, filter_size, pad_before, pad_after) -> Tuple[int, int]:
+ """
+ Based on explicit padding provided in a PAD operation, returns the corresponding hardware padding
+ that provides equivalent results.
+ """
+ total_padding = needed_total_padding(input_size, stride, filter_size)
+ # The top/left padding can be taken as is from the PAD
+ output_pad_before = pad_before
+ # The bottom/right padding might need downward adjustment depending on stride/input size
+ output_pad_after = pad_after
+ while output_pad_after > 0 and output_pad_after % stride != (total_padding - pad_before) % stride:
+ output_pad_after -= 1
+ return output_pad_before, output_pad_after
+
+
+def calc_padding_and_skirt(padding_type, kernel, input_shape, explicit_padding):
+ k_w, k_h = kernel.dilated_wh()
+ s_x, s_y = kernel.stride
+ ypad = needed_total_padding(int(input_shape.height), int(s_y), int(k_h))
+ xpad = needed_total_padding(int(input_shape.width), int(s_x), int(k_w))
+ if padding_type == Padding.SAME:
+ left_pad = (xpad + 0) // 2
+ right_pad = (xpad + 1) // 2
+ top_pad = (ypad + 0) // 2
+ bottom_pad = (ypad + 1) // 2
+ elif padding_type == Padding.VALID:
+ left_pad = 0
+ right_pad = 0
+ top_pad = 0
+ bottom_pad = 0
+ elif padding_type == Padding.EXPLICIT:
+ # Padding is specified in a PAD operator which has been bypassed.
+ top, left, bottom, right = explicit_padding
+ top_pad, bottom_pad = calc_explicit_padding(int(input_shape.height), int(s_y), int(k_h), int(top), int(bottom))
+ left_pad, right_pad = calc_explicit_padding(int(input_shape.width), int(s_x), int(k_w), int(left), int(right))
+ else:
+ raise UnsupportedFeatureError(f"Unknown padding")
+ padding = (top_pad, left_pad, bottom_pad, right_pad)
+ skirt = (top_pad, left_pad, ypad - top_pad, xpad - left_pad)
+ return padding, skirt
+
+
+def calc_upscaled_padding_and_skirt(padding_type, kernel_size, stride, input_shape, upscaling_factor):
+ kernel_height, kernel_width = kernel_size[0], kernel_size[1]
+ if padding_type == Padding.SAME:
+ ypad = needed_total_padding(int(input_shape.height) * upscaling_factor, int(stride[1]), int(kernel_height))
+ xpad = needed_total_padding(int(input_shape.width) * upscaling_factor, int(stride[2]), int(kernel_width))
+ right_pad = max(((xpad + 1) // upscaling_factor) - 1, 0)
+ bottom_pad = max(((ypad + 1) // upscaling_factor) - 1, 0)
+ left_pad = max(kernel_width - 1 - right_pad, 0)
+ top_pad = max(kernel_height - 1 - bottom_pad, 0)
+ elif padding_type == Padding.VALID:
+ right_pad = max(kernel_width - 2, 0)
+ bottom_pad = max(kernel_height - 2, 0)
+ left_pad = kernel_width - 1
+ top_pad = kernel_height - 1
+ else:
+ raise UnsupportedFeatureError(f"Unknown padding")
+ padding = (top_pad, left_pad, bottom_pad, right_pad)
+ skirt = padding
+ return padding, skirt
+
+
+def fixup_conv2d_backprop(op, arch, nng):
+ if op.type == Op.Conv2DBackpropInput:
+ # flip the inputs
+ op.inputs[0], op.inputs[2] = op.inputs[2], op.inputs[0]
+ op.type = Op.Conv2DBackpropInputSwitchedBias
+ op.ifm.resampling_mode = resampling_mode.TRANSPOSE
+
+ # Update strides
+ op.attrs.update({"stride_w": 1, "stride_h": 1, "strides": (1, 1, 1, 1)})
+
+ return op
+
+
+# Convert the op to an elementwise add
+def convert_resizebilinear_1x1_to_add(op):
+ op.type = Op.Add
+ op.name = op.name + "_add"
+ op.attrs["resizebilinear"] = True
+ # 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.quantization = QuantizationParameters(0.0, 255.0)
+ tens.quantization.scale_f32 = 1.0
+ tens.quantization.zero_point = 0
+ tens.consumer_list = [op]
+ tens_op = op.inputs[1].ops[0]
+ tens_op.set_output_tensor(tens)
+ # Set the add inputs
+ op.inputs[1] = op.inputs[0]
+ op.inputs[0] = tens
+ op.set_ifm_ofm_shapes()
+
+ return op
+
+
+# Convert ResizeBilinear to a number of 2x2 pool ops
+def convert_resizebilinear_to_2x2_pool(op):
+ count = 0
+ pre_op = op
+ outputs = op.outputs
+
+ op.attrs.update({"strides": (1, 1, 1, 1), "ksize": (1, 2, 2, 1)})
+ if op.attrs["align_corners"]:
+ shape_modifier = 1
+ op.attrs["padding"] = Padding.VALID
+ else:
+ shape_modifier = 0
+ op.attrs["padding"] = Padding.SAME
+ op.inputs[0].resampling_mode = resampling_mode.NEAREST
+
+ upscaled_shape = np.array(op.ifm_shapes[0].get_hw_as_list())
+ out_shape = np.array(op.ofm_shapes[0].get_hw_as_list())
+ if (upscaled_shape == upscaled_shape * 2 - shape_modifier).all():
+ return op
+
+ while (upscaled_shape < out_shape).all():
+ if count == 0:
+ scaled_op = pre_op
+ else:
+ scaled_op = op.clone("_{}".format(count))
+ scaled_op.inputs[0] = pre_op.outputs[0]
+
+ upscaled_shape = upscaled_shape * 2 - shape_modifier
+
+ if (upscaled_shape == out_shape).all():
+ scaled_op.outputs = outputs
+ scaled_op.outputs[0].ops = [scaled_op]
+ else:
+ shape = op.ofm_shapes[0].as_list()
+ shape[1:3] = upscaled_shape
+ out_tens = Tensor(shape, DataType.int16, "{}_{}".format(op.outputs[0].name, count))
+ out_tens.quantization = op.outputs[0].quantization.clone()
+ out_tens.quantization.quant_min = np.iinfo(np.int16).min
+ out_tens.quantization.quant_max = np.iinfo(np.int16).max
+ scaled_op.set_output_tensor(out_tens)
+ pre_op = scaled_op
+ count += 1
+
+ # Setup the scale value
+ if scaled_op.inputs[0].dtype.bits == 8 and scaled_op.outputs[0].dtype.bits == 16:
+ scaled_op.rescale = 128
+ elif scaled_op.inputs[0].dtype.bits == 16 and scaled_op.outputs[0].dtype.bits == 8:
+ scaled_op.rescale = 1 / 128
+ else:
+ scaled_op.rescale = None
+ scaled_op.set_ifm_ofm_shapes()
+
+ return op
+
+
+def fixup_resizebilinear(op, arch, nng):
+ if op.type == Op.ResizeBilinear and op.run_on_npu:
+ if op.ifm_shapes[0] == op.ofm_shapes[0]:
+ # Bypass nop resizebilinear
+ op.inputs = op.inputs[:1]
+ op.type = Op.Identity
+ elif op.ifm_shapes[0].height == 1 and op.ifm_shapes[0].width == 1:
+ convert_resizebilinear_1x1_to_add(op)
+ else:
+ convert_resizebilinear_to_2x2_pool(op)
+
+ return op
+
+
+def convert_nop_split_to_identity(op, arch, nng):
+ if op.type == Op.Split and op.attrs.get("num_splits") == 1:
+ # the list comprehension should return a list with a single tensor
+ # if it shouldn't, remove_passthrough_tensor will fail appropriately
+ op.inputs = [i for i in op.inputs if i.shape == op.outputs[0].shape]
+ op.type = Op.Identity
+ return op
+
+
+def rewrite_fully_connected_input(op, arch, nng):
+ if op.type == Op.FullyConnected:
+ n_in_elems = op.weights.shape[-2]
+ elms = op.ifm.elements()
+ batch_size = elms // n_in_elems
+ assert batch_size * n_in_elems == elms
+
+ op.ifm_shapes[0] = Shape4D([batch_size, 1, 1, n_in_elems])
+ return op
+
+
+def convert_batched_fc_shape(op, arch, nng):
+ if op.type == Op.FullyConnected:
+ # Check if the first dimension indicates batching
+ if op.ifm_shapes[0].batch > 1:
+ batching_split = {4: (2, 2), 8: (2, 4), 16: (4, 4)}
+ n = op.ifm_shapes[0].batch
+ h, w = batching_split.get(n, (1, n))
+ op.ifm_shapes[0] = Shape4D([1, h, w, op.ifm_shapes[0].depth])
+
+ # 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))
+
+ n = op.ofm_shapes[0].batch
+ h, w = batching_split.get(n, (1, n))
+ op.ofm_shapes[0] = Shape4D([1, h, w, op.ofm_shapes[0].depth])
+ return op
+
+
+def unfuse_activation_function(op):
+ if op.type == Op.ConcatTFLite and op.run_on_npu and op.activation is not None:
+ act_op = Operation(op.activation.op_type, op.name + op.activation.op_type.name)
+ op.activation = None
+ out_tens = op.outputs[0]
+ intermediate_tens = out_tens.clone("_act_intermediate")
+ act_op.set_output_tensor(out_tens)
+ act_op.add_input_tensor(intermediate_tens)
+ op.set_output_tensor(intermediate_tens)
+ act_op.set_ifm_ofm_shapes()
+
+
+def rewrite_stridedslice_output(op, arch, nng):
+ if not op.run_on_npu or op.type != Op.StridedSlice:
+ return op
+
+ new_axis_mask = op.attrs["new_axis_mask"]
+ shrink_axis_mask = op.attrs["shrink_axis_mask"]
+
+ if shrink_axis_mask == 0 and new_axis_mask == 0:
+ return op
+
+ axis_4D = [0] * len(op.outputs)
+ for idx, out_tens in enumerate(op.outputs):
+ output_shape = list(out_tens.shape)
+
+ if shrink_axis_mask != 0:
+ n = 0
+ axis = 0
+ while shrink_axis_mask:
+ prev_mask = shrink_axis_mask
+ n += 1
+ shrink_axis_mask &= shrink_axis_mask - 1
+ axis = int(math.log2(prev_mask - shrink_axis_mask))
+ output_shape = output_shape[:axis] + [1] + output_shape[axis:]
+
+ assert len(out_tens.shape) == (len(op.inputs[0].shape) - n)
+ op.attrs["shrink_axis_mask"] = 0
+ if axis >= 0:
+ axis_4D[idx] = axis + (4 - len(output_shape))
+ else:
+ axis_4D[idx] = axis
+ op.ofm_shapes[idx] = Shape4D(output_shape)
+
+ elif new_axis_mask != 0:
+ n = 0
+ axis = 0
+ while new_axis_mask:
+ prev_mask = new_axis_mask
+ n += 1
+ new_axis_mask &= new_axis_mask - 1
+ axis = int(math.log2(prev_mask - new_axis_mask))
+ output_shape = output_shape[:axis] + output_shape[(axis + 1) :]
+ new_axis_mask >>= 1
+
+ assert len(out_tens.shape) == (len(op.inputs[0].shape) + n)
+ op.attrs["new_axis_mask"] = 0
+ if axis >= 0:
+ axis_4D[idx] = axis + (4 - len(output_shape))
+ else:
+ axis_4D[idx] = axis
+ op.ofm_shapes[idx] = Shape4D(output_shape)
+
+ op.attrs["split_axis_4D"] = axis_4D
+ return op
+
+
+def rewrite_unpack_output(op, arch, nng):
+ tens = op.outputs[0]
+ if op.run_on_npu and op.type == Op.Unpack:
+ # Unpack is also referred to as Unstack
+ axis = int(op.attrs["axis"])
+ if axis < 0: # Convert to positive axis
+ axis = len(op.inputs[0].shape) + 1 + axis
+ op.type = Op.UnpackReshaped
+ desired_output_shape = tens.shape[:axis] + [1] + tens.shape[axis:]
+
+ axis_4D = axis + (4 - len(desired_output_shape))
+ op.attrs["split_axis_4D"] = [axis_4D] * len(op.outputs)
+
+ for idx, out_tens in enumerate(op.outputs):
+ op.ofm_shapes[idx] = Shape4D(desired_output_shape)
+ return op
+
+
+def add_padding_fields(op, arch, nng):
+ if op.run_on_npu:
+ if "padding" in op.attrs:
+ input_shape = op.ifm_shapes[0]
+ output_shape = op.ofm_shapes[0]
+ if op.type.is_conv2d_op() or op.type.is_depthwise_conv2d_op():
+ kernel_size = op.inputs[1].shape[:2]
+ elif op.type.is_pool_op() or op.type.npu_block_type == NpuBlockType.ReduceSum:
+ kernel_size = op.attrs["ksize"][1:3]
+ else:
+ raise UnsupportedFeatureError(f"Unknown operation that uses padding: {optype_to_builtintype(op.type)}")
+
+ if op.type == Op.Conv2DBackpropInputSwitchedBias:
+ upscaling_factor = output_shape.height // input_shape.height
+ padding, skirt = calc_upscaled_padding_and_skirt(
+ op.attrs["padding"], kernel_size, op.attrs["strides"], input_shape, upscaling_factor
+ )
+ else:
+ padding, skirt = calc_padding_and_skirt(
+ op.attrs["padding"], op.kernel, input_shape, op.attrs.get("explicit_padding"),
+ )
+
+ op.attrs["explicit_padding"] = padding
+ op.attrs["skirt"] = skirt
+
+ return op
+
+
+def convert_depthwise_to_conv(op, arch, nng):
+ # Depthwise is equivalent to a single conv2d if the ifm depth is 1 and
+ # the ofm depth equals the depth multipler.
+ # If those conditions are true, then we can perform a simple
+ # switch of the operator type (and weight order)
+
+ if op.type == Op.DepthwiseConv2DBias and (op.attrs["depth_multiplier"] != 1):
+ ifm_shape = op.ifm_shapes[0]
+ weight_tensor = op.inputs[1]
+ ofm_shape = op.ofm_shapes[0]
+ if (ifm_shape.depth == 1) and (ofm_shape.depth == op.attrs["depth_multiplier"]):
+ # Change op type to Conv2d
+ op.type = Op.Conv2DBias
+ 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))
+ else:
+ raise UnsupportedFeatureError(
+ f"Unsupported 'DEPTHWISE_CONV_2D' with depth_multiplier = {op.attrs['depth_multiplier']},",
+ f" ifm channels = {ifm_shape.depth}, ofm channels = {ofm_shape.depth}",
+ )
+ DebugDatabase.add_optimised(op, op)
+ return op
+
+
+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.weight_transpose_depthwise = True
+
+ return op
+
+
+def optimise_strided_conv(op, arch, nng):
+ stride_x, stride_y = op.get_kernel_stride()
+ ifm_tensor, _, weight_tensor, _ = op.get_ifm_ifm2_weights_ofm()
+
+ if (
+ op.type == Op.Conv2DBias
+ and op.op_index == 0
+ and stride_x == 2
+ and op.ifm_shapes[0].depth <= 4
+ and op.ifm_shapes[0].width % 2 == 0
+ and weight_tensor is not None
+ and weight_tensor.shape[1] >= 2
+ ):
+ ifm_shape = op.ifm_shapes[0]
+ # IFM
+ op.ifm_shapes[0] = Shape4D([ifm_shape.batch, ifm_shape.height, ifm_shape.width // 2, ifm_shape.depth * 2])
+
+ # Weights
+ weight_shape = weight_tensor.shape
+ if weight_shape[1] % 2 != 0:
+ weight_shape[1] = weight_shape[1] + 1
+ padded_array = np.zeros(weight_shape)
+ for i in range(weight_shape[0]):
+ padded_array[i] = np.vstack(
+ [
+ weight_tensor.quant_values[i],
+ np.full((1, weight_shape[2], weight_shape[3]), weight_tensor.quantization.zero_point),
+ ]
+ )
+ weight_tensor.quant_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.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
+ weight_tensor.value_id = uuid.uuid4()
+
+ # Strides
+ stride_x = 1
+ op.attrs.update({"stride_w": stride_x, "stride_h": stride_y, "strides": (1, stride_y, stride_x, 1)})
+
+ return op
+
+
+def convert_conv_to_fc(op, arch, nng):
+ # Conv 1x1 can be equivalent to Fully Connected.
+ # By representing certain convs as fully connected layers, Vela can better determine wether or not to use
+ # caching/double buffering for the weights.
+ # (Weights dont need to be reloaded for convs when IFM H and W are 1)
+ if op.type == Op.Conv2DBias:
+ h = op.ifm_shapes[0].height
+ w = op.ifm_shapes[0].width
+ kh, kw, _, _ = op.inputs[1].shape
+ if h == 1 and w == 1 and kh == 1 and kw == 1:
+ # Overwrite this op as a Fully Connected Op
+ op.name += "_fc"
+ op.type = Op.FullyConnected
+ op.attrs = {
+ "weights_format": 0,
+ }
+ # 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))
+
+ DebugDatabase.add_optimised(op, op)
+ return op
+
+
+def fixup_relus_with_differing_ifm_ofm_scaling(op, arch, nng):
+ if op.run_on_npu and op.type.is_relu_op():
+ ifm = op.inputs[0]
+ ofm = op.outputs[0]
+ # Relu with differing IFM and OFM scaling cannot be fused with another primary op
+ # and requires its own to be inserted
+ if not check_quantized_tens_scaling_equal(ifm, ofm):
+ # Override this op with its own primary op (avgpool)
+ relu_fused_op = create_avgpool_nop(op.name + "_avgpool")
+ # And fuse the original activation function to it
+ relu_fused_op.activation = create_activation_function(op.type)
+ # Tidy up and assign the ifm and ofm to the new op
+ ifm.consumer_list.remove(op)
+
+ relu_fused_op.add_input_tensor(ifm)
+ relu_fused_op.set_output_tensor(ofm)
+ relu_fused_op.set_ifm_ofm_shapes()
+ op = relu_fused_op
+ return op
+
+
+def fixup_elementwise_with_scalars(op, arch, nng):
+ if op.type.is_binary_elementwise_op():
+ ifm_tensor, ifm2_tensor, _, _ = op.get_ifm_ifm2_weights_ofm()
+ if ifm2_tensor.shape != [] and ifm_tensor.shape != []:
+ diff = len(ifm_tensor.shape) - len(ifm2_tensor.shape)
+ if diff > 0:
+ 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:
+ # 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:
+ # 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
+ return op
+
+
+def convert_softmax(op, arch, nng):
+ if op.type == Op.Softmax and op.run_on_npu:
+ softmax = SoftMax(op)
+ op = softmax.get_graph()
+ return op
+
+
+def convert_mul_max_to_abs_or_lrelu(op, arch, nng):
+ r"""Whenever there is a subgraph with this topology:
+
+ Input X For X = -1 or X > 0
+ | \ / This subgraph can be replaced with either
+ | Mul an Abs (if X = -1) or a LeakyReLU (if X > 0)
+ | /
+ Max
+ """
+
+ if op.type == Op.Maximum:
+ # finds the Mul input(s) to the Max
+ muls = [i for i in op.inputs if i.ops[0].type == Op.Mul]
+ if len(muls) == 1:
+ mul = muls[0].ops[0]
+ elif len(muls) == 2:
+ # In the case both inputs are Muls, find the one with the same input as the Max
+ mul = [m for m in muls if len(set(op.inputs + m.ops[0].inputs)) == 1][0].ops[0]
+ else:
+ # No Mul inputs
+ return op
+
+ # make sure the Mul doesn't have any other consumers
+ mul_ofm = mul.outputs[0]
+ if len(mul_ofm.consumers()) != 1:
+ return op
+ # make sure the Mul doesn't have a fused activation function
+ if mul.activation:
+ return op
+ ifm, ofm = op.get_ifm_ofm()
+ if ifm is None or ofm is None:
+ return op
+
+ if ifm.dtype not in (DataType.uint8, DataType.int8) or ifm.dtype != ofm.dtype:
+ return op
+ if not check_quantized_tens_scaling_equal(ifm, ofm) or not check_quantized_tens_scaling_equal(ifm, mul_ofm):
+ # rewrite to LeakyRelu currently only makes sense if the quantization is identical
+ return op
+
+ # finds the branched input that goes to both the Max and the Mul
+ shared = set(op.inputs) & set(mul.inputs)
+ if len(shared) == 1:
+ shared_in = shared.pop()
+ # find the constant scalar input to the Mul
+ const_tens = (set(mul.inputs) - {shared_in}).pop()
+ # check that it is a scalar
+ if const_tens.shape != []:
+ return op
+ const = const_tens.ops[0]
+ # check that it is a constant
+ if const.type != Op.Const:
+ return op
+ # Remove the Mul from the shared input's consumers
+ shared_in.consumer_list.remove(mul)
+ else:
+ return op
+
+ val = const.outputs[0].values
+ if val >= 0:
+ new_op = Op.LeakyRelu
+ op.attrs["alpha"] = val
+ # 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
+ 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)
+ alpha_scale, alpha_shift = scaling.elementwise_mul_scale(mul_ifm_scale, mul_ifm2_scale, mul_ofm_scale)
+ op.attrs["alpha_scaling"] = (alpha_scalar, alpha_scale, alpha_shift)
+ elif val == -1:
+ new_op = Op.Abs
+ else:
+ return op
+
+ op.type = new_op
+ op.name = op.name.replace("Maximum", new_op.name)
+ op.outputs[0].name = op.outputs[0].name.replace("Maximum", new_op.name)
+ op.inputs = [shared_in]
+ op.set_ifm_ofm_shapes()
+
+ # Record optimisation in debug database
+ DebugDatabase.add_optimised(op, op)
+
+ return op
+
+
+def convert_hardswish_to_lut(op, arch, nng):
+ if op.type == Op.HardSwish:
+ ifm, ofm = op.get_ifm_ofm()
+ # Generate the LUT
+ ifm_scale = np.double(ifm.quantization.scale_f32)
+ ofm_scale = np.double(ofm.quantization.scale_f32)
+ zp_in = ifm.quantization.zero_point
+ zp_out = ofm.quantization.zero_point
+ ifm_scale_hires = (1 / 128) * ifm_scale
+ relu_multiplier = np.double(3 / 32768)
+ out_scale, out_shift = scaling.quantise_scale(ifm_scale_hires / ofm_scale)
+ relu_scale, relu_shift = scaling.quantise_scale(ifm_scale_hires / relu_multiplier)
+ # Use 16bit scale
+ out_scale_16 = fp_math.downscale_multiplier_int32_to_int16(out_scale)
+ relu_scale_16 = fp_math.downscale_multiplier_int32_to_int16(relu_scale)
+
+ values = []
+ ix = range(256) if ifm.dtype == DataType.uint8 else range(-128, 128)
+ quantized_min = min(ix)
+ quantized_max = max(ix)
+ for x in ix:
+ input_value = x - zp_in
+ input_value_hires = input_value * 128
+ # Compute the input value on essentially the output scale, not shifted yet
+ input_value_preshift = fp_math.saturating_rounding_mul16(input_value_hires, out_scale_16)
+ # Compute the "relu-ish multiplier". This matches the code in TensorFlow Lite Micro kernel
+ relu_value = np.int16(input_value_hires)
+ if relu_shift < 31:
+ relu_value = fp_math.shift_left16(relu_value, 30 - relu_shift)
+
+ relu_value = fp_math.saturating_rounding_mul16(relu_value, relu_scale_16)
+
+ if relu_shift < 31:
+ relu_value = fp_math.shift_left16(relu_value, 1)
+
+ if relu_shift > 31:
+ relu_value = fp_math.rounding_divide_by_pot(relu_value, relu_shift - 31)
+
+ # Rescaled the value into a 16bit fixedpoint relu_value in [-1, 1]
+ # Now convert that to a 16bit fixedpoint value in [0, 1]
+ relu_value = (relu_value + (1 << 15)) >> 1
+ lut_result = fp_math.saturating_mul16(relu_value, input_value_preshift)
+ shift = 31 - out_shift
+ shift = -shift if shift < 0 else 0
+ # Finally apply the output shift
+ lut_result = fp_math.rounding_divide_by_pot(lut_result, shift) + zp_out
+ lut_result = min(quantized_max, max(quantized_min, lut_result))
+ values.append(lut_result)
+ return convert_to_lut(op, values, "hardswish")
+ return op
+
+
+def convert_lrelu_to_mul_max(op, arch):
+ # Converts LeakyRelu to Max(alpha * IFM, identity * IFM)
+ # (the opposite of convert_mul_max_to_abs_or_lrelu)
+ ifm, ofm = op.get_ifm_ofm()
+ if ifm is None or ofm is None:
+ return op
+
+ # Add multiplication with alpha
+ mul_alpha = Operation(Op.Mul, op.name + "_mul_alpha")
+ mul_alpha.add_input_tensor(ifm)
+ # Create const tensor containing alpha as scalar
+ alpha = op.attrs["alpha"]
+ quantization = ifm.quantization.clone()
+ quantization.min = 0
+ quantization.max = alpha * (quantization.quant_max - quantization.quant_min)
+ quantization.zero_point = 0
+ if np.isinf(1 / np.float32(alpha)):
+ # Handling of alpha near zero
+ quantization.scale_f32 = 1
+ scalar = 0
+ else:
+ quantization.scale_f32 = alpha
+ scalar = alpha
+ alpha_tens = create_const_tensor(
+ op.name + "_alpha_scalar", [], ifm.dtype, [scalar], np.float32, quantization=quantization
+ )
+ alpha_tens.quant_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)
+ mul_alpha.set_ifm_ofm_shapes()
+ DebugDatabase.add_optimised(op, mul_alpha)
+
+ if check_quantized_tens_scaling_equal(ifm, ofm):
+ # No identity multiplication is needed
+ fm_id = ifm
+ else:
+ # Add multiplication with identity
+ mul_identity = Operation(Op.Mul, op.name + "_mul_identity")
+ mul_identity.add_input_tensor(ifm)
+ # Create const tensor containing identity as scalar
+ quantization = ifm.quantization.clone()
+ quantization.min = 0
+ quantization.max = quantization.quant_max - quantization.quant_min
+ quantization.scale_f32 = 1
+ quantization.zero_point = 0
+ identity_tens = create_const_tensor(
+ op.name + "_id_scalar", [], ifm.dtype, [1], np.uint8, quantization=quantization
+ )
+ mul_identity.add_input_tensor(identity_tens)
+ # Make sure that fm_id is allocated to a different address than fm_alpha
+ fm_id = ofm.clone(op.name + "_id", set_unique=True)
+ mul_identity.set_output_tensor(fm_id)
+ mul_identity.set_ifm_ofm_shapes()
+ DebugDatabase.add_optimised(op, mul_identity)
+
+ # Convert LeakyRelu to Max, add the results of the multiplication(s) as inputs
+ op.type = Op.Maximum
+ op.name = op.name.replace("LeakyRelu", "Maximum")
+ op.inputs = []
+ ifm.consumer_list.remove(op)
+ op.add_input_tensor(fm_alpha)
+ op.add_input_tensor(fm_id)
+ op.set_ifm_ofm_shapes()
+
+ DebugDatabase.add_optimised(op, op)
+ return op
+
+
+def convert_to_lut(op, lut_values, lut_name):
+ # Rewrite the operation by Add with scalar 0 + LUT activation
+ ifm = op.inputs[0]
+ if ifm is None:
+ return op
+ assert ifm.dtype.size_in_bytes() == 1
+ op.type = Op.Add
+ op.name = op.name + "_lut_" + lut_name
+ # Mark as no-op to enable potential fusing optimizations
+ op.attrs["is_nop"] = True
+ # Create an input tensor containing scalar zero
+ quantization = QuantizationParameters(0.0, 255.0)
+ quantization.scale_f32 = ifm.quantization.scale_f32
+ quantization.zero_point = 0
+ tens = create_const_tensor(op.inputs[0].name + "_scalar0", [], ifm.dtype, [0], np.uint8, quantization=quantization)
+ op.add_input_tensor(tens)
+ op.ifm_shapes.append(Shape4D(tens.shape))
+
+ # The LUT must be applied without any preceding rescaling (the LUT itself performs the rescale),
+ # so even if the OFM has a different scale than the IFM, the generated OFM scale instructions
+ # should be the same as the IFM
+ op.forced_output_quantization = ifm.quantization
+ lut_tensor = lut.create_lut_tensor(op.name + "_values", lut_values, DataType.int8)
+ op.set_activation_lut(lut_tensor)
+ op.set_ifm_ofm_shapes()
+ return op
+
+
+def convert_to_lut8(op, fn, fn_name):
+ # Converts op to a no-op + int8/uint8 LUT which is generated with the given function.
+ # fn is a function(real) -> real
+ ifm, ofm = op.get_ifm_ofm()
+ if ifm.dtype not in (DataType.uint8, DataType.int8) or ifm.dtype != ofm.dtype:
+ return op
+ # Generate the LUT
+ ifm_scale = np.double(ifm.quantization.scale_f32)
+ ofm_scale = np.double(ofm.quantization.scale_f32)
+ zp_in = ifm.quantization.zero_point
+ zp_out = ofm.quantization.zero_point
+ values = []
+ ix = range(256) if ifm.dtype == DataType.uint8 else range(-128, 128)
+ quantized_min = min(ix)
+ quantized_max = max(ix)
+ for x in ix:
+ x_real = ifm_scale * (x - zp_in)
+ y_real = fn(x_real)
+ lut_result = round_away_zero(zp_out + y_real / ofm_scale)
+ lut_result = min(quantized_max, max(quantized_min, lut_result))
+ values.append(lut_result)
+ return convert_to_lut(op, values, fn_name)
+
+
+def convert_lrelu_to_lut(op, arch):
+ ifm, ofm = op.get_ifm_ofm()
+ # Generate the LUT
+ alpha = op.attrs["alpha"]
+ ifm_scale = np.double(ifm.quantization.scale_f32)
+ ofm_scale = np.double(ofm.quantization.scale_f32)
+ zp_in = ifm.quantization.zero_point
+ zp_out = ofm.quantization.zero_point
+ identity_scale, identity_shift = scaling.elementwise_mul_scale(ifm_scale, 1, ofm_scale)
+ alpha_scalar = 1
+ alpha_scale, alpha_shift = scaling.elementwise_mul_scale(ifm_scale, alpha, ofm_scale)
+ if "alpha_scaling" in op.attrs:
+ # The LeakyRelu was the result from convert_mul_max_to_abs_or_lrelu
+ alpha_scalar, alpha_scale, alpha_shift = op.attrs["alpha_scaling"]
+ values = []
+ ix = range(256) if ifm.dtype == DataType.uint8 else range(-128, 128)
+ quantized_min = min(ix)
+ quantized_max = max(ix)
+ for x in ix:
+ if x < zp_in:
+ lut_result = zp_out + fp_math.multiply_by_quantized_multiplier(
+ alpha_scalar * (x - zp_in), alpha_scale, alpha_shift
+ )
+ else:
+ lut_result = zp_out + fp_math.multiply_by_quantized_multiplier(x - zp_in, identity_scale, identity_shift)
+ lut_result = min(quantized_max, max(quantized_min, lut_result))
+ values.append(lut_result)
+ return convert_to_lut(op, values, "lrelu")
+
+
+def convert_lrelu(op, arch, nng):
+ # Converts LeakyRelu to a LUT based solution if possible, otherwise a mul + max
+ if op.type != Op.LeakyRelu:
+ return op
+ ifm, ofm = op.get_ifm_ofm()
+ if ifm is None or ofm is None:
+ return op
+ if ifm.dtype in (DataType.uint8, DataType.int8) and ifm.dtype == ofm.dtype:
+ # use LUT for int8/uint8
+ return convert_lrelu_to_lut(op, arch)
+ if check_quantized_tens_scaling_equal(ifm, ofm) and ifm.dtype == ofm.dtype == DataType.int16:
+ # use LeakyRelu unmodified for int16 with equal input/output scaling
+ return op
+ return convert_lrelu_to_mul_max(op, arch)
+
+
+def convert_tanh_sigmoid_to_lut(op, arch, nng):
+ # Converts int8/uint8 Sigmoid and Tanh to a LUT based solution
+ if op.type == Op.Sigmoid:
+ return convert_to_lut8(op, clamp_sigmoid, "sigmoid")
+ elif op.type == Op.Tanh:
+ return convert_to_lut8(op, math.tanh, "tanh")
+ return op
+
+
+def remove_reshapes(op, arch):
+ if op.run_on_npu and op.type == Op.Reshape:
+ ofm = op.ofm
+ ifm = op.ifm
+
+ # Check if quantization is the same in the input and output for the reshape ops
+ if not check_quantized_tens_scaling_equal(ifm, ofm):
+ # TODO Both tensors are needed, since quantisation properties currently are linked to Tensors.
+ # In order to remove this reshape either quantization properties need to be moved to Operator,
+ # or the reshape need to be replace with a NOP.
+ return
+
+ # Check if Reshape ifm/ofm are network ifm/ofm
+ ifm_is_sg_ifm = ifm.ops[0].type in (Op.Placeholder, Op.SubgraphInput, Op.Const)
+ ifm_is_sg_ofm = any(ifm_cons is None for ifm_cons in ifm.consumer_list)
+ ofm_is_sg_ofm = any(ofm_cons is None for ofm_cons in ofm.consumer_list)
+ # Check if ifm/ofm is produced repectivly consumed by CPU
+ ifm_is_cpu_produced = any(ifm_prod is not None and not ifm_prod.run_on_npu for ifm_prod in op.ifm.ops)
+ ofm_is_cpu_consumed = any(ofm_cons is not None and not ofm_cons.run_on_npu for ofm_cons in op.ofm.consumer_list)
+
+ # This case should be handled prior to this function
+ assert not ((ifm_is_sg_ifm or ifm_is_sg_ofm or ifm_is_cpu_produced) and (ofm_is_sg_ofm or ofm_is_cpu_consumed))
+
+ if ofm_is_sg_ofm or ofm_is_cpu_consumed:
+ # Bypassed by replacing ifm with ofm
+ ofm.ops = []
+ for prev_op in ifm.ops:
+ prev_op.outputs = [ofm]
+ ofm.ops.append(prev_op)
+
+ # All ifm consumers need to use ofm as input
+ for ifm_cons in ifm.consumer_list:
+ for ifm_idx, cons_ifm in enumerate(ifm_cons.inputs):
+ if cons_ifm == ifm:
+ ifm_cons.set_input_tensor(ofm, ifm_idx)
+ else:
+ # Bypassed Reshape by replacing ofm with ifm
+ for cons in ofm.consumer_list:
+ for ifm_idx, cons_ifm in enumerate(cons.inputs):
+ if cons_ifm == ofm:
+ cons.set_input_tensor(ifm, ifm_idx)
+
+
+def fuse_activation_function_with_prev(op, arch, nng):
+ # if op is a no-op: attempts to move the activation function to the preceding op
+ if not op.attrs.get("is_nop", False) or op.activation is None:
+ return op
+ ifm, ofm = op.get_ifm_ofm()
+ if ifm is None or ofm is None:
+ return op
+ # finds the input(s) to the operation
+ prev_op = ifm.ops[0]
+ # Note: the below checks on prev_op require that a first optimize pass on the full graph has been performed
+ fuse = (
+ prev_op.run_on_npu
+ and prev_op.type.npu_block_type != NpuBlockType.Default
+ and len(ifm.ops) == 1
+ and len(prev_op.outputs[0].consumers()) == 1
+ and prev_op.activation is None
+ )
+ if op.activation_lut is not None and arch.shram_reserved_unused_banks == 0:
+ # TODO: if SHRAM LUT space is shared with SHRAM ACC (32, 64 MAC),
+ # LUT currently only works correctly for elementwise ops
+ fuse = False
+ if not fuse:
+ return op
+ # Move the fused activation function + corresponding info to prev_op
+ prev_op.activation = op.activation
+ prev_op.forced_output_quantization = op.forced_output_quantization
+ if op.activation_lut is not None:
+ prev_op.set_activation_lut(op.activation_lut)
+ # Bypass op
+ prev_op.set_output_tensor(ofm)
+ DebugDatabase.add_optimised(op, prev_op)
+ return op
+
+
+def _leading_pad_ok(leading_pad, stride, kernel_size):
+ # If kernel size // 2 > stride, then (left, top) padding must be a multiple of stride,
+ # otherwise replacing PAD by hardware padding would iterate the wrong IFM rows/columns
+ max_size = kernel_size // 2
+ return leading_pad == max_size or max_size <= stride or leading_pad % stride == 0
+
+
+def replace_pad_by_hw_pad(op: Operation, arch, nng):
+ """
+ Tries to completely remove a PAD operator by using hardware padding.
+ E.g. a PAD operation that pads 1, followed by a CONV with VALID padding and kernel size 3
+ is rewritten such that the PAD is removed, and the CONV uses SAME padding.
+ Converts tens1 -> PAD -> tens2 -> CONV to tens1 -> CONV
+ if both operations can be run on the NPU.
+ This is the most efficient way to implement PAD, but cannot be done for all pad sizes.
+ """
+ if (
+ (op.type.is_conv2d_op() or op.type.is_depthwise_conv2d_op() or op.type.is_avgpool_op())
+ and op.run_on_npu
+ and op.attrs["padding"] == Padding.VALID
+ ):
+ pad_op = op.ifm.ops[0]
+ if pad_op.type != Op.Pad or not pad_op.run_on_npu:
+ return op
+ if pad_op.ifm.dtype != pad_op.ofm.dtype or not check_quantized_tens_scaling_equal(pad_op.ofm, pad_op.ifm):
+ return op
+ top, left, bottom, right = get_pad_values_from_input(pad_op.inputs[1].values)
+ k = op.kernel
+ k_w, k_h = k.dilated_wh()
+
+ # Check if the PAD operator can be replaced by hardware padding
+ if left > k_w // 2 or right > k_w // 2 or top > k_h // 2 or bottom > k_h // 2:
+ # Too much padding, it would require hardware padding to actually insert zeros
+ return op
+ if not _leading_pad_ok(top, k.stride.y, k_h) or not _leading_pad_ok(left, k.stride.x, k_w):
+ return op
+
+ if op.type.is_avgpool_op():
+ # For average pool, hardware padding can only be used if padding is 0 or kernel size / 2
+ for pad, k_size in (
+ (left, k_w),
+ (right, k_w),
+ (top, k_h),
+ (bottom, k_h),
+ ):
+ if pad not in (0, k_size // 2):
+ return op
+ # Average pool is converted to depthwise, because NPU average pool + same padding
+ # has a special implementation that is different from PAD followed by average pool with
+ # valid padding.
+ k_w, k_h = op.kernel.width, op.kernel.height
+ ifm = op.ifm
+ # Remember other inputs
+ other_inputs = op.inputs[1:]
+ # Create a weight tensor, all weights are set to 1/(kernel width * kernel height)
+ quantization = QuantizationParameters(0.0, 255.0)
+ quantization.scale_f32 = 1.0 / (k_w * k_h)
+ quantization.zero_point = 0
+ shape = [k_h, k_w, 1, op.ofm.shape[-1]]
+ weights = np.full(shape, 1)
+
+ weight_tens = create_const_tensor(
+ op.name + "_weights",
+ shape,
+ op.ifm.dtype,
+ weights,
+ np.uint8,
+ purpose=TensorPurpose.Weights,
+ quantization=quantization,
+ )
+ weight_tens.quant_values = weights
+ op.type = Op.DepthwiseConv2DBias
+ op.inputs = []
+ op.add_input_tensor(ifm)
+ op.add_input_tensor(weight_tens)
+ # Add bias tensor, all biases set to 0
+ op.inputs.append(None)
+ fixup_bias_tensors(op, arch, nng)
+ # Add other inputs
+ op.inputs.extend(other_inputs)
+ op.rounding_mode = NpuRoundingMode.NATURAL
+
+ # Bypass the PAD operator
+ op.set_input_tensor(pad_op.ifm, 0)
+ # Adjust the padding attributes of the convolution operator
+ op.attrs["padding"] = Padding.EXPLICIT
+ op.attrs["explicit_padding"] = (top, left, bottom, right)
+ op.set_ifm_ofm_shapes()
+ return op
+
+
+def convert_pad(op: Operation, arch, nng):
+ """
+ Rewrites PAD operator to an average pool that copies the IFM to the OFM
+ + up to 4 average pool operators that fill the OFM with zeros at the borders.
+ This is done as fall-back for the PAD operators that remain after replace_pad_by_hw_pad
+ """
+ if op.type != Op.Pad or not op.run_on_npu:
+ return op
+ top, left, bottom, right = get_pad_values_from_input(op.inputs[1].values)
+
+ ifm = op.ifm
+ assert ifm is not None
+ ifm_shape = Shape4D(ifm.shape)
+ ofm = op.ofm
+ assert ofm is not None
+ ofm.ops = []
+ ofm_shape = op.ofm_shapes[0]
+
+ # Average pool op that copies IFM to the right place inside the OFM
+ shp0 = Shape4D(0, 0, 0, 0)
+ shp_top = shp0.with_height(top)
+ avgpool_op = create_avg_pool_for_concat(op, op.name + "_main", ifm, ifm_shape, shp_top.with_width(left))
+ avgpool_op.activation = op.activation
+ quant = ofm.quantization
+ pad_value = quant.zero_point
+ # Add operations that fill the borders of the OFM
+ if top > 0:
+ shape = Shape4D(1, top, ofm_shape.width, ofm_shape.depth)
+ zero_tens = create_const_tensor(
+ op.name + "_top", shape.as_list(), ofm.dtype, shape.elements() * [pad_value], np.uint8, quantization=quant
+ )
+ # If top/bottom or left/right are equal, the const tensors can be allocated to the same address
+ zero_tens.equivalence_id = create_equivalence_id(tuple(zero_tens.values))
+ create_avg_pool_for_concat(op, op.name + "_top", zero_tens, shape, shp0)
+ if bottom > 0:
+ shape = Shape4D(1, bottom, ofm_shape.width, ofm_shape.depth)
+ zero_tens = create_const_tensor(
+ op.name + "_bottom",
+ shape.as_list(),
+ ofm.dtype,
+ shape.elements() * [pad_value],
+ np.uint8,
+ quantization=quant,
+ )
+ zero_tens.equivalence_id = create_equivalence_id(tuple(zero_tens.values))
+ create_avg_pool_for_concat(
+ op, op.name + "_bottom", zero_tens, shape, shp0.with_height(ofm_shape.height - bottom)
+ )
+ if left > 0:
+ shape = Shape4D(1, ifm_shape.height, left, ofm_shape.depth)
+ zero_tens = create_const_tensor(
+ op.name + "_left", shape.as_list(), ofm.dtype, shape.elements() * [pad_value], np.uint8, quantization=quant
+ )
+ zero_tens.equivalence_id = create_equivalence_id(tuple(zero_tens.values))
+ create_avg_pool_for_concat(op, op.name + "_left", zero_tens, shape, shp_top)
+ if right > 0:
+ shape = Shape4D(1, ifm_shape.height, right, ofm_shape.depth)
+ zero_tens = create_const_tensor(
+ op.name + "_right", shape.as_list(), ofm.dtype, shape.elements() * [pad_value], np.uint8, quantization=quant
+ )
+ zero_tens.equivalence_id = create_equivalence_id(tuple(zero_tens.values))
+ create_avg_pool_for_concat(
+ op, op.name + "_right", zero_tens, shape, shp_top.with_width(ofm_shape.width - right)
+ )
+
+ op.type = Op.ConcatTFLite
+ return avgpool_op
+
+
+def add_attrs_to_resizebilinear(op, arch, nng):
+ if op.type == Op.ResizeBilinear and op.run_on_npu:
+ input_tensor = op.inputs[0]
+ input_shape = op.ifm_shapes[0]
+ upscaled_height = input_shape.height * 2
+ upscaled_width = input_shape.width * 2
+ out_shape = op.ofm_shapes[0]
+ if not op.attrs["align_corners"] and out_shape.height == upscaled_height and out_shape.width == upscaled_width:
+ # this means the output is supposed to be a x2 upscale,
+ # so we need to do SAME padding
+ op.attrs["padding"] = Padding.SAME
+ elif (
+ op.attrs["align_corners"]
+ and out_shape.height == (upscaled_height - 1)
+ and out_shape.width == (upscaled_width - 1)
+ ):
+ # here we can just run the avg pool without padding and
+ # produce a (M * 2 - 1, N * 2 - 1) sized output
+ op.attrs["padding"] = Padding.VALID
+ else:
+ return op
+ input_tensor.resampling_mode = resampling_mode.NEAREST
+ op.attrs.update({"strides": (1, 1, 1, 1), "ksize": (1, 2, 2, 1)})
+ return op
+
+
+def fixup_bias_tensors(op, arch, nng):
+ if op.type.needs_bias() and op.bias is None:
+ # Op has no bias, add bias tensor filled with zeros
+ 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
+
+
+def convert_mean_to_depthwise_conv_or_avgpool(op, arch, nng):
+ if op.type == Op.Mean and op.run_on_npu:
+ keep_dims = op.attrs.get("keep_dims", False)
+ inp, axis = op.inputs
+ shape = inp.shape
+ dims = len(shape)
+
+ # Height and width axes have different index depending on dimensions
+ if axis.shape == [] or axis.shape[0] == 1: # single axis
+ axis = int(axis.values) if len(axis.shape) == 0 else int(axis.values[0])
+ if dims in (2, 3):
+ if axis == 0:
+ h, w = shape[axis], 1
+ else:
+ h, w = 1, shape[axis]
+ else:
+ if axis == 1:
+ h, w = shape[axis], 1
+ else:
+ h, w = 1, shape[axis]
+ else: # multiple axes
+ axis = sorted(axis.values)
+ h, w = [shape[i] for i in axis]
+
+ # Set necessary depthwise attributes
+ op.attrs.update(
+ {
+ "padding": Padding.VALID,
+ "stride_h": 1,
+ "stride_w": 1,
+ "strides": (1, 1, 1, 1),
+ "depth_multiplier": 1,
+ "channel_multiplier": 1,
+ "dilation_h_factor": 1,
+ "dilation_w_factor": 1,
+ "dilation": (1, 1, 1, 1),
+ }
+ )
+ # Change op type
+ op.type = Op.DepthwiseConv2DBias
+ # Set IFM/OFM shapes after changing op type
+ op.set_ifm_ofm_shapes()
+
+ weight_scale, bias = 1, None
+ ofmq, ifmq = op.ofm.quantization, inp.quantization
+ # Set rounding mode, scaling and zero point based on which reference implementation to match
+ if len(shape) == 4 and axis == [1, 2] and keep_dims:
+ if inp.dtype == DataType.uint8:
+ # This attribute means a different scaling calculation is used in order to match reference
+ op.low_precision_scaling = True
+ weight_scale = h * w
+ # Set zero points to 0 as they will be adjusted for with bias term
+ foq = ofmq.clone()
+ foq.zero_point = 0
+ fiq = ifmq.clone()
+ fiq.zero_point = 0
+ op.forced_input_quantization = fiq
+ bias_term = ofmq.zero_point - int(ifmq.zero_point * ifmq.scale_f32 / ofmq.scale_f32)
+ # If the bias term is outside uint8 range, we need an Add op to apply it.
+ if bias_term < 0 or bias_term > 255:
+ intermediate = op.ofm.clone(suffix="_intermediate", set_unique=True)
+ # Bias term has higher bitness (i32) than input/output (u8).
+ # 16 bits is enough since the bias is added/subtracted from a u8 value,
+ # the bias can only effectively assume values in the range [-255, 255].
+ intermediate.dtype = DataType.int16
+ intermediate.quantization.zero_point = 0
+ add_op = Operation(Op.Add, op.name + "_bias")
+ add_op.forced_output_quantization = foq
+ add_op.add_input_tensor(intermediate)
+ 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,
+ )
+ add_op.add_input_tensor(bias_term_tens)
+ add_op.set_output_tensor(op.ofm)
+ add_op.set_ifm_ofm_shapes()
+ add_op.activation = op.activation
+ op.activation = None
+ op.set_output_tensor(intermediate)
+ op.set_ifm_ofm_shapes()
+ # If not, we can just do it with the OFM zero point.
+ else:
+ foq.zero_point = bias_term
+ op.forced_output_quantization = foq
+ else:
+ assert inp.dtype == DataType.int8
+ # Use a depthwise to calculate the sum,
+ # followed by a multiplication with 1/N to get the MEAN
+ weight_scale = 1
+ intermediate = op.ofm.clone(suffix="_intermediate", set_unique=True)
+ intermediate.dtype = DataType.int16
+ mul_op = Operation(Op.Mul, op.name + "_mul")
+ mul_op.add_input_tensor(intermediate)
+ # Create scalar containing 1/N
+ quant = QuantizationParameters()
+ quant.zero_point = 0
+ # The reference rounds negative numbers downwards, e.g. -1.5 is rounded to -2,
+ # while rounding mode NATURAL would round this to -1.
+ # This can only occur if N is even, and can be emulated by
+ # multiplying with a number that is slightly smaller than 1/N.
+ # It must be so small that other roundings are not affected;
+ # the calculated value is based on worst case,
+ # which is sum 256 * N (the maximum sum that can occur with int8)
+ n = int(h * w)
+ eps = 1 / (256 * (n + 1)) if n % 2 == 0 else 0
+ quant.scale_f32 = 1 / (n - eps)
+ scalar = create_const_tensor(
+ op.name + "_scalar", [1, 1, 1, 1], DataType.uint8, [1], np.uint8, quantization=quant
+ )
+ mul_op.add_input_tensor(scalar)
+ mul_op.set_output_tensor(op.ofm)
+ mul_op.set_ifm_ofm_shapes()
+ mul_op.rounding_mode = NpuRoundingMode.NATURAL
+ mul_op.activation = op.activation
+ op.activation = None
+ op.set_output_tensor(intermediate)
+ op.set_ifm_ofm_shapes()
+ elif ifmq.zero_point == ofmq.zero_point and ifmq.scale_f32 == ofmq.scale_f32:
+ # Here we can just use a simple AvgPool with truncating rounding,
+ # as we're emulating simple integer division.
+ op.rounding_mode = NpuRoundingMode.TRUNCATE
+ op.type = Op.AvgPool
+ op.attrs.update({"ksize": (1, h, w, 1), "filter_height": h, "filter_width": w})
+ else:
+ op.rounding_mode = NpuRoundingMode.NATURAL
+ weight_scale = 1 / (h * w)
+ # Input zero point is adjusted after mean calculation, so we emulate that with a bias
+ bias = -ifmq.zero_point * h * w
+ fiq = ifmq.clone()
+ fiq.zero_point = 0
+ op.forced_input_quantization = fiq
+
+ # Change dimensions to 4
+ if dims < 4:
+ shape = [1] + shape
+ if dims == 2:
+ shape += [1]
+
+ # If height is greater than max kernel height, reshape to from HxW to 1x(HxW)
+ if h > 64:
+ shape = [shape[0], 1, h * w, shape[3]]
+ op.ifm_shapes[0] = Shape4D(shape)
+ if h > 256 and op.type == Op.AvgPool:
+ op.attrs.update({"ksize": (1, 1, h * w, 1), "filter_height": 1, "filter_width": h * w})
+
+ # If the AvgPool version is used, we don't need to do anything else
+ if op.type == Op.AvgPool:
+ return op
+
+ # Make unit weight tensor quantization
+ weight_quant = ifmq.clone()
+ weight_quant.min = 0
+ weight_quant.max = 255
+ weight_quant.scale_f32 = weight_scale
+ weight_quant.zero_point = 0
+
+ # Set weight shape to [H,W,C,B]
+ weight_shape = shape[1:4] + [shape[0]]
+ # Add unit weight tensor
+ op.set_input_tensor(
+ create_const_tensor(
+ "weights",
+ weight_shape,
+ inp.dtype,
+ np.ones(weight_shape),
+ value_dtype=np.uint8,
+ quantization=weight_quant,
+ ),
+ 1,
+ )
+ op.weights.quant_values = np.reshape(op.inputs[1].quant_values, weight_shape)
+
+ # Add None bias tensor
+ op.inputs.append(None)
+ # Add bias tensor
+ if bias:
+ bias_shape = [shape[-1]]
+ op.set_input_tensor(
+ create_const_tensor(
+ "bias",
+ bias_shape,
+ inp.dtype,
+ np.ones(bias_shape) * bias,
+ value_dtype=np.int32,
+ quant_value_dtype=np.int32,
+ quantization=None,
+ ),
+ 2,
+ )
+
+ return op
+
+
+def supported_operator_check(op, arch, nng):
+ op.run_on_npu = arch.supported_operators.is_operator_supported(op)
+ return op
+
+
+def tflite_optimise_graph(nng, arch):
+ # Pre-processing step
+ pre_process_list = [
+ supported_operator_check,
+ set_ifm_ofm_op_shapes,
+ ]
+
+ for idx, sg in enumerate(nng.subgraphs):
+ nng.subgraphs[idx] = rewrite_graph.rewrite_graph_pre_order(
+ nng, sg, arch, [], pre_process_list, rewrite_unsupported=False,
+ )
+
+ # Handle Concat Ops
+ for idx, sg in enumerate(nng.subgraphs):
+ rewrite_graph.visit_graph_post_order(sg.output_tensors, arch, [], [rewrite_concat_ops])
+ sg.refresh_after_modification()
+
+ # Handle Split Ops
+ for idx, sg in enumerate(nng.subgraphs):
+ nng.subgraphs[idx] = rewrite_graph.rewrite_graph_pre_order(
+ nng,
+ sg,
+ arch,
+ [],
+ [rewrite_unpack_output, rewrite_stridedslice_output, convert_nop_split_to_identity],
+ rewrite_unsupported=False,
+ )
+
+ for idx, sg in enumerate(nng.subgraphs):
+ nng.subgraphs[idx] = rewrite_graph.rewrite_graph_pre_order(
+ nng, sg, arch, [rewrite_split_ops], [], rewrite_unsupported=False,
+ )
+
+ # Handle sg input output
+ for idx, sg in enumerate(nng.subgraphs):
+ nng.subgraphs[idx] = rewrite_graph.rewrite_graph_pre_order(
+ nng, sg, arch, [], [fix_sg_input_output], rewrite_unsupported=False,
+ )
+
+ # Removal of reshapes
+ for sg in nng.subgraphs:
+ rewrite_graph.visit_graph_post_order(sg.output_tensors, arch, [], [remove_reshapes])
+ sg.refresh_after_modification()
+
+ # Rewrite of operators
+ op_rewrite_list = [
+ set_tensor_equivalence,
+ convert_mean_to_depthwise_conv_or_avgpool,
+ convert_depthwise_to_conv,
+ convert_conv_to_fc,
+ convert_softmax,
+ optimise_strided_conv,
+ convert_hardswish_to_lut,
+ rewrite_fully_connected_input,
+ convert_batched_fc_shape,
+ fixup_conv2d_backprop,
+ fixup_relus_with_differing_ifm_ofm_scaling,
+ fixup_elementwise_with_scalars,
+ reorder_depthwise_weights,
+ fixup_resizebilinear,
+ fixup_bias_tensors,
+ convert_mul_max_to_abs_or_lrelu,
+ convert_lrelu,
+ convert_tanh_sigmoid_to_lut,
+ replace_pad_by_hw_pad,
+ ]
+
+ for idx, sg in enumerate(nng.subgraphs):
+ nng.subgraphs[idx] = rewrite_graph.rewrite_graph_pre_order(
+ nng, sg, arch, [], op_rewrite_list, rewrite_unsupported=False,
+ )
+
+ for idx, sg in enumerate(nng.subgraphs):
+ # remove passthrough tensors and attempt further optimizations
+ nng.subgraphs[idx] = rewrite_graph.rewrite_graph_pre_order(
+ nng,
+ sg,
+ arch,
+ [remove_passthrough_tensor],
+ [fuse_activation_function_with_prev, convert_pad, add_padding_fields],
+ )
+
+ # Removal of SplitSliceRead, need to be done after optimisation has been performed,
+ # since ifm/ofm_shapes are of importance to this function
+ for sg in nng.subgraphs:
+ rewrite_graph.visit_graph_post_order(sg.output_tensors, arch, [], [remove_SplitSliceRead])
+ sg.refresh_after_modification()
+
+ return nng
diff --git a/ethosu/vela/tflite_reader.py b/ethosu/vela/tflite_reader.py
index b47177f..1a45a5e 100644
--- a/ethosu/vela/tflite_reader.py
+++ b/ethosu/vela/tflite_reader.py
@@ -27,6 +27,9 @@
from .operation import create_activation_function
from .operation import Op
from .operation import Operation
+from .reader_util import clone_and_reshape_tensor
+from .reader_util import decode_str
+from .reader_util import fixup_tensors
from .tensor import QuantizationParameters
from .tensor import Tensor
from .tflite.BuiltinOperator import BuiltinOperator
@@ -37,29 +40,6 @@
from .tflite_mapping import datatype_map_numpy
-def decode_str(s):
- if s is None:
- return ""
- return s.decode("utf-8")
-
-
-def clone_and_reshape_tensor(src_tens, reorder, set_unique):
- tens = src_tens.clone("_reshape", set_unique)
- tens.shape = [src_tens.shape[idx] for idx in reorder]
- tens.bandwidth_shape = tens.shape
- tens.storage_shape = tens.shape
-
- 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
-
-
class TFLiteSubgraph:
def __init__(self, graph, subgraph):
self.graph = graph
@@ -74,19 +54,7 @@
self.outputs = self.get_tensors_from_indices_remove_duplicates(subgraph.OutputsAsNumpy(), "output")
self.inputs = self.get_tensors_from_indices_remove_duplicates(subgraph.InputsAsNumpy(), "input")
-
- # Fix up tensors without operations. Generate either Placeholder or Constant ops
- for tens in self.inputs:
- if tens.ops != []:
- tens.error("This subgraph input tensor has unexpected driving operators.")
-
- op = Operation(Op.Placeholder, tens.name)
- op.set_output_tensor(tens)
-
- for tens in self.tensors:
- if not tens.ops:
- op = Operation(Op.Const, tens.name)
- op.set_output_tensor(tens)
+ fixup_tensors(self.inputs, self.tensors)
def get_tensors_from_indices_remove_duplicates(self, indices, warning_str):
tensors = []
diff --git a/ethosu/vela/tosa/ArithmeticRightShiftAttribute.py b/ethosu/vela/tosa/ArithmeticRightShiftAttribute.py
new file mode 100644
index 0000000..ad7b9b0
--- /dev/null
+++ b/ethosu/vela/tosa/ArithmeticRightShiftAttribute.py
@@ -0,0 +1,30 @@
+# automatically generated by the FlatBuffers compiler, do not modify
+
+# namespace: tosa
+
+import flatbuffers
+
+class ArithmeticRightShiftAttribute(object):
+ __slots__ = ['_tab']
+
+ @classmethod
+ def GetRootAsArithmeticRightShiftAttribute(cls, buf, offset):
+ n = flatbuffers.encode.Get(flatbuffers.packer.uoffset, buf, offset)
+ x = ArithmeticRightShiftAttribute()
+ x.Init(buf, n + offset)
+ return x
+
+ # ArithmeticRightShiftAttribute
+ def Init(self, buf, pos):
+ self._tab = flatbuffers.table.Table(buf, pos)
+
+ # ArithmeticRightShiftAttribute
+ def Round(self):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(4))
+ if o != 0:
+ return bool(self._tab.Get(flatbuffers.number_types.BoolFlags, o + self._tab.Pos))
+ return False
+
+def ArithmeticRightShiftAttributeStart(builder): builder.StartObject(1)
+def ArithmeticRightShiftAttributeAddRound(builder, round): builder.PrependBoolSlot(0, round, 0)
+def ArithmeticRightShiftAttributeEnd(builder): return builder.EndObject()
diff --git a/ethosu/vela/tosa/Attribute.py b/ethosu/vela/tosa/Attribute.py
new file mode 100644
index 0000000..8a2ccae
--- /dev/null
+++ b/ethosu/vela/tosa/Attribute.py
@@ -0,0 +1,22 @@
+# automatically generated by the FlatBuffers compiler, do not modify
+
+# namespace: tosa
+
+class Attribute(object):
+ NONE = 0
+ Pool2dAttribute = 1
+ Conv2dAttribute = 2
+ TransposeConv2dAttribute = 3
+ ReluNAttribute = 4
+ AxisAttribute = 5
+ ReshapeAttribute = 6
+ SliceAttribute = 7
+ TileAttribute = 8
+ ResizeAttribute = 9
+ ClampAttribute = 10
+ RescaleAttribute = 11
+ MulAttribute = 12
+ ArithmeticRightShiftAttribute = 13
+ CondIfAttribute = 14
+ WhileLoopAttribute = 15
+
diff --git a/ethosu/vela/tosa/AxisAttribute.py b/ethosu/vela/tosa/AxisAttribute.py
new file mode 100644
index 0000000..77f1bca
--- /dev/null
+++ b/ethosu/vela/tosa/AxisAttribute.py
@@ -0,0 +1,30 @@
+# automatically generated by the FlatBuffers compiler, do not modify
+
+# namespace: tosa
+
+import flatbuffers
+
+class AxisAttribute(object):
+ __slots__ = ['_tab']
+
+ @classmethod
+ def GetRootAsAxisAttribute(cls, buf, offset):
+ n = flatbuffers.encode.Get(flatbuffers.packer.uoffset, buf, offset)
+ x = AxisAttribute()
+ x.Init(buf, n + offset)
+ return x
+
+ # AxisAttribute
+ def Init(self, buf, pos):
+ self._tab = flatbuffers.table.Table(buf, pos)
+
+ # AxisAttribute
+ def Axis(self):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(4))
+ if o != 0:
+ return self._tab.Get(flatbuffers.number_types.Int32Flags, o + self._tab.Pos)
+ return 0
+
+def AxisAttributeStart(builder): builder.StartObject(1)
+def AxisAttributeAddAxis(builder, axis): builder.PrependInt32Slot(0, axis, 0)
+def AxisAttributeEnd(builder): return builder.EndObject()
diff --git a/ethosu/vela/tosa/ClampAttribute.py b/ethosu/vela/tosa/ClampAttribute.py
new file mode 100644
index 0000000..9cc8d4d
--- /dev/null
+++ b/ethosu/vela/tosa/ClampAttribute.py
@@ -0,0 +1,54 @@
+# automatically generated by the FlatBuffers compiler, do not modify
+
+# namespace: tosa
+
+import flatbuffers
+
+class ClampAttribute(object):
+ __slots__ = ['_tab']
+
+ @classmethod
+ def GetRootAsClampAttribute(cls, buf, offset):
+ n = flatbuffers.encode.Get(flatbuffers.packer.uoffset, buf, offset)
+ x = ClampAttribute()
+ x.Init(buf, n + offset)
+ return x
+
+ # ClampAttribute
+ def Init(self, buf, pos):
+ self._tab = flatbuffers.table.Table(buf, pos)
+
+ # ClampAttribute
+ def MinInt(self):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(4))
+ if o != 0:
+ return self._tab.Get(flatbuffers.number_types.Int32Flags, o + self._tab.Pos)
+ return 0
+
+ # ClampAttribute
+ def MaxInt(self):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(6))
+ if o != 0:
+ return self._tab.Get(flatbuffers.number_types.Int32Flags, o + self._tab.Pos)
+ return 0
+
+ # ClampAttribute
+ def MinFp(self):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(8))
+ if o != 0:
+ return self._tab.Get(flatbuffers.number_types.Float32Flags, o + self._tab.Pos)
+ return 0.0
+
+ # ClampAttribute
+ def MaxFp(self):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(10))
+ if o != 0:
+ return self._tab.Get(flatbuffers.number_types.Float32Flags, o + self._tab.Pos)
+ return 0.0
+
+def ClampAttributeStart(builder): builder.StartObject(4)
+def ClampAttributeAddMinInt(builder, minInt): builder.PrependInt32Slot(0, minInt, 0)
+def ClampAttributeAddMaxInt(builder, maxInt): builder.PrependInt32Slot(1, maxInt, 0)
+def ClampAttributeAddMinFp(builder, minFp): builder.PrependFloat32Slot(2, minFp, 0.0)
+def ClampAttributeAddMaxFp(builder, maxFp): builder.PrependFloat32Slot(3, maxFp, 0.0)
+def ClampAttributeEnd(builder): return builder.EndObject()
diff --git a/ethosu/vela/tosa/CondIfAttribute.py b/ethosu/vela/tosa/CondIfAttribute.py
new file mode 100644
index 0000000..bc19b95
--- /dev/null
+++ b/ethosu/vela/tosa/CondIfAttribute.py
@@ -0,0 +1,38 @@
+# automatically generated by the FlatBuffers compiler, do not modify
+
+# namespace: tosa
+
+import flatbuffers
+
+class CondIfAttribute(object):
+ __slots__ = ['_tab']
+
+ @classmethod
+ def GetRootAsCondIfAttribute(cls, buf, offset):
+ n = flatbuffers.encode.Get(flatbuffers.packer.uoffset, buf, offset)
+ x = CondIfAttribute()
+ x.Init(buf, n + offset)
+ return x
+
+ # CondIfAttribute
+ def Init(self, buf, pos):
+ self._tab = flatbuffers.table.Table(buf, pos)
+
+ # CondIfAttribute
+ def ThenBranch(self):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(4))
+ if o != 0:
+ return self._tab.String(o + self._tab.Pos)
+ return None
+
+ # CondIfAttribute
+ def ElseBranch(self):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(6))
+ if o != 0:
+ return self._tab.String(o + self._tab.Pos)
+ return None
+
+def CondIfAttributeStart(builder): builder.StartObject(2)
+def CondIfAttributeAddThenBranch(builder, thenBranch): builder.PrependUOffsetTRelativeSlot(0, flatbuffers.number_types.UOffsetTFlags.py_type(thenBranch), 0)
+def CondIfAttributeAddElseBranch(builder, elseBranch): builder.PrependUOffsetTRelativeSlot(1, flatbuffers.number_types.UOffsetTFlags.py_type(elseBranch), 0)
+def CondIfAttributeEnd(builder): return builder.EndObject()
diff --git a/ethosu/vela/tosa/Conv2dAttribute.py b/ethosu/vela/tosa/Conv2dAttribute.py
new file mode 100644
index 0000000..5faecbd
--- /dev/null
+++ b/ethosu/vela/tosa/Conv2dAttribute.py
@@ -0,0 +1,94 @@
+# automatically generated by the FlatBuffers compiler, do not modify
+
+# namespace: tosa
+
+import flatbuffers
+
+class Conv2dAttribute(object):
+ __slots__ = ['_tab']
+
+ @classmethod
+ def GetRootAsConv2dAttribute(cls, buf, offset):
+ n = flatbuffers.encode.Get(flatbuffers.packer.uoffset, buf, offset)
+ x = Conv2dAttribute()
+ x.Init(buf, n + offset)
+ return x
+
+ # Conv2dAttribute
+ def Init(self, buf, pos):
+ self._tab = flatbuffers.table.Table(buf, pos)
+
+ # Conv2dAttribute
+ def Padding(self, j):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(4))
+ if o != 0:
+ a = self._tab.Vector(o)
+ return self._tab.Get(flatbuffers.number_types.Int32Flags, a + flatbuffers.number_types.UOffsetTFlags.py_type(j * 4))
+ return 0
+
+ # Conv2dAttribute
+ def PaddingAsNumpy(self):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(4))
+ if o != 0:
+ return self._tab.GetVectorAsNumpy(flatbuffers.number_types.Int32Flags, o)
+ return 0
+
+ # Conv2dAttribute
+ def PaddingLength(self):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(4))
+ if o != 0:
+ return self._tab.VectorLen(o)
+ return 0
+
+ # Conv2dAttribute
+ def Stride(self, j):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(6))
+ if o != 0:
+ a = self._tab.Vector(o)
+ return self._tab.Get(flatbuffers.number_types.Int32Flags, a + flatbuffers.number_types.UOffsetTFlags.py_type(j * 4))
+ return 0
+
+ # Conv2dAttribute
+ def StrideAsNumpy(self):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(6))
+ if o != 0:
+ return self._tab.GetVectorAsNumpy(flatbuffers.number_types.Int32Flags, o)
+ return 0
+
+ # Conv2dAttribute
+ def StrideLength(self):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(6))
+ if o != 0:
+ return self._tab.VectorLen(o)
+ return 0
+
+ # Conv2dAttribute
+ def Dilation(self, j):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(8))
+ if o != 0:
+ a = self._tab.Vector(o)
+ return self._tab.Get(flatbuffers.number_types.Int32Flags, a + flatbuffers.number_types.UOffsetTFlags.py_type(j * 4))
+ return 0
+
+ # Conv2dAttribute
+ def DilationAsNumpy(self):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(8))
+ if o != 0:
+ return self._tab.GetVectorAsNumpy(flatbuffers.number_types.Int32Flags, o)
+ return 0
+
+ # Conv2dAttribute
+ def DilationLength(self):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(8))
+ if o != 0:
+ return self._tab.VectorLen(o)
+ return 0
+
+def Conv2dAttributeStart(builder): builder.StartObject(3)
+def Conv2dAttributeAddPadding(builder, padding): builder.PrependUOffsetTRelativeSlot(0, flatbuffers.number_types.UOffsetTFlags.py_type(padding), 0)
+def Conv2dAttributeStartPaddingVector(builder, numElems): return builder.StartVector(4, numElems, 4)
+def Conv2dAttributeAddStride(builder, stride): builder.PrependUOffsetTRelativeSlot(1, flatbuffers.number_types.UOffsetTFlags.py_type(stride), 0)
+def Conv2dAttributeStartStrideVector(builder, numElems): return builder.StartVector(4, numElems, 4)
+def Conv2dAttributeAddDilation(builder, dilation): builder.PrependUOffsetTRelativeSlot(2, flatbuffers.number_types.UOffsetTFlags.py_type(dilation), 0)
+def Conv2dAttributeStartDilationVector(builder, numElems): return builder.StartVector(4, numElems, 4)
+def Conv2dAttributeEnd(builder): return builder.EndObject()
diff --git a/ethosu/vela/tosa/ConvQuantInfo.py b/ethosu/vela/tosa/ConvQuantInfo.py
new file mode 100644
index 0000000..9f785b3
--- /dev/null
+++ b/ethosu/vela/tosa/ConvQuantInfo.py
@@ -0,0 +1,38 @@
+# automatically generated by the FlatBuffers compiler, do not modify
+
+# namespace: tosa
+
+import flatbuffers
+
+class ConvQuantInfo(object):
+ __slots__ = ['_tab']
+
+ @classmethod
+ def GetRootAsConvQuantInfo(cls, buf, offset):
+ n = flatbuffers.encode.Get(flatbuffers.packer.uoffset, buf, offset)
+ x = ConvQuantInfo()
+ x.Init(buf, n + offset)
+ return x
+
+ # ConvQuantInfo
+ def Init(self, buf, pos):
+ self._tab = flatbuffers.table.Table(buf, pos)
+
+ # ConvQuantInfo
+ def InputZp(self):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(4))
+ if o != 0:
+ return self._tab.Get(flatbuffers.number_types.Int32Flags, o + self._tab.Pos)
+ return 0
+
+ # ConvQuantInfo
+ def WeightZp(self):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(6))
+ if o != 0:
+ return self._tab.Get(flatbuffers.number_types.Int32Flags, o + self._tab.Pos)
+ return 0
+
+def ConvQuantInfoStart(builder): builder.StartObject(2)
+def ConvQuantInfoAddInputZp(builder, inputZp): builder.PrependInt32Slot(0, inputZp, 0)
+def ConvQuantInfoAddWeightZp(builder, weightZp): builder.PrependInt32Slot(1, weightZp, 0)
+def ConvQuantInfoEnd(builder): return builder.EndObject()
diff --git a/ethosu/vela/tosa/DType.py b/ethosu/vela/tosa/DType.py
new file mode 100644
index 0000000..65432db
--- /dev/null
+++ b/ethosu/vela/tosa/DType.py
@@ -0,0 +1,15 @@
+# automatically generated by the FlatBuffers compiler, do not modify
+
+# namespace: tosa
+
+class DType(object):
+ UNKNOWN = 0
+ BOOL = 1
+ UINT8 = 2
+ INT4 = 3
+ INT8 = 4
+ INT16 = 5
+ INT32 = 6
+ INT48 = 7
+ FLOAT = 8
+
diff --git a/ethosu/vela/tosa/MatMulQuantInfo.py b/ethosu/vela/tosa/MatMulQuantInfo.py
new file mode 100644
index 0000000..6cc3a72
--- /dev/null
+++ b/ethosu/vela/tosa/MatMulQuantInfo.py
@@ -0,0 +1,38 @@
+# automatically generated by the FlatBuffers compiler, do not modify
+
+# namespace: tosa
+
+import flatbuffers
+
+class MatMulQuantInfo(object):
+ __slots__ = ['_tab']
+
+ @classmethod
+ def GetRootAsMatMulQuantInfo(cls, buf, offset):
+ n = flatbuffers.encode.Get(flatbuffers.packer.uoffset, buf, offset)
+ x = MatMulQuantInfo()
+ x.Init(buf, n + offset)
+ return x
+
+ # MatMulQuantInfo
+ def Init(self, buf, pos):
+ self._tab = flatbuffers.table.Table(buf, pos)
+
+ # MatMulQuantInfo
+ def AZp(self):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(4))
+ if o != 0:
+ return self._tab.Get(flatbuffers.number_types.Int32Flags, o + self._tab.Pos)
+ return 0
+
+ # MatMulQuantInfo
+ def BZp(self):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(6))
+ if o != 0:
+ return self._tab.Get(flatbuffers.number_types.Int32Flags, o + self._tab.Pos)
+ return 0
+
+def MatMulQuantInfoStart(builder): builder.StartObject(2)
+def MatMulQuantInfoAddAZp(builder, aZp): builder.PrependInt32Slot(0, aZp, 0)
+def MatMulQuantInfoAddBZp(builder, bZp): builder.PrependInt32Slot(1, bZp, 0)
+def MatMulQuantInfoEnd(builder): return builder.EndObject()
diff --git a/ethosu/vela/tosa/MulAttribute.py b/ethosu/vela/tosa/MulAttribute.py
new file mode 100644
index 0000000..c08f684
--- /dev/null
+++ b/ethosu/vela/tosa/MulAttribute.py
@@ -0,0 +1,30 @@
+# automatically generated by the FlatBuffers compiler, do not modify
+
+# namespace: tosa
+
+import flatbuffers
+
+class MulAttribute(object):
+ __slots__ = ['_tab']
+
+ @classmethod
+ def GetRootAsMulAttribute(cls, buf, offset):
+ n = flatbuffers.encode.Get(flatbuffers.packer.uoffset, buf, offset)
+ x = MulAttribute()
+ x.Init(buf, n + offset)
+ return x
+
+ # MulAttribute
+ def Init(self, buf, pos):
+ self._tab = flatbuffers.table.Table(buf, pos)
+
+ # MulAttribute
+ def Shift(self):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(4))
+ if o != 0:
+ return self._tab.Get(flatbuffers.number_types.Int32Flags, o + self._tab.Pos)
+ return 0
+
+def MulAttributeStart(builder): builder.StartObject(1)
+def MulAttributeAddShift(builder, shift): builder.PrependInt32Slot(0, shift, 0)
+def MulAttributeEnd(builder): return builder.EndObject()
diff --git a/ethosu/vela/tosa/Op.py b/ethosu/vela/tosa/Op.py
new file mode 100644
index 0000000..c71ac44
--- /dev/null
+++ b/ethosu/vela/tosa/Op.py
@@ -0,0 +1,75 @@
+# automatically generated by the FlatBuffers compiler, do not modify
+
+# namespace: tosa
+
+class Op(object):
+ UNKNOWN = 0
+ ARGMAX = 1
+ AVG_POOL2D = 2
+ CONV2D = 3
+ CONV3D = 4
+ DEPTHWISE_CONV2D = 5
+ FULLY_CONNECTED = 6
+ MATMUL = 7
+ MAX_POOL2D = 8
+ TRANSPOSE_CONV2D = 9
+ CLAMP = 10
+ RELUN = 11
+ SIGMOID = 12
+ TANH = 13
+ ADD = 14
+ ARITHMETIC_RIGHT_SHIFT = 15
+ BITWISE_AND = 16
+ BITWISE_OR = 17
+ BITWISE_XOR = 18
+ DIV = 19
+ LOGICAL_AND = 20
+ LOGICAL_LEFT_SHIFT = 21
+ LOGICAL_RIGHT_SHIFT = 22
+ LOGICAL_OR = 23
+ LOGICAL_XOR = 24
+ MAXIMUM = 25
+ MINIMUM = 26
+ MUL = 27
+ POW = 28
+ SUB = 29
+ TABLE = 30
+ ABS = 31
+ BITWISE_NOT = 32
+ CEIL = 33
+ CLZ = 34
+ EXP = 35
+ FLOOR = 36
+ LOG = 37
+ LOGICAL_NOT = 38
+ NEGATE = 39
+ RECIPROCAL = 40
+ RSQRT = 41
+ SELECT = 42
+ EQUAL = 43
+ GREATER = 44
+ GREATER_EQUAL = 45
+ REDUCE_ANY = 46
+ REDUCE_ALL = 47
+ REDUCE_MAX = 48
+ REDUCE_MIN = 49
+ REDUCE_PRODUCT = 50
+ REDUCE_SUM = 51
+ CONCAT = 52
+ PAD = 53
+ RESHAPE = 54
+ REVERSE = 55
+ SLICE = 56
+ TILE = 57
+ TRANSPOSE = 58
+ GATHER = 59
+ SCATTER = 60
+ RESIZE = 61
+ CAST = 62
+ RESCALE = 63
+ CONST = 64
+ IDENTITY = 65
+ CUSTOM = 66
+ COND_IF = 67
+ WHILE_LOOP = 68
+
diff --git a/ethosu/vela/tosa/PadQuantInfo.py b/ethosu/vela/tosa/PadQuantInfo.py
new file mode 100644
index 0000000..825b72c
--- /dev/null
+++ b/ethosu/vela/tosa/PadQuantInfo.py
@@ -0,0 +1,30 @@
+# automatically generated by the FlatBuffers compiler, do not modify
+
+# namespace: tosa
+
+import flatbuffers
+
+class PadQuantInfo(object):
+ __slots__ = ['_tab']
+
+ @classmethod
+ def GetRootAsPadQuantInfo(cls, buf, offset):
+ n = flatbuffers.encode.Get(flatbuffers.packer.uoffset, buf, offset)
+ x = PadQuantInfo()
+ x.Init(buf, n + offset)
+ return x
+
+ # PadQuantInfo
+ def Init(self, buf, pos):
+ self._tab = flatbuffers.table.Table(buf, pos)
+
+ # PadQuantInfo
+ def InputZp(self):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(4))
+ if o != 0:
+ return self._tab.Get(flatbuffers.number_types.Int32Flags, o + self._tab.Pos)
+ return 0
+
+def PadQuantInfoStart(builder): builder.StartObject(1)
+def PadQuantInfoAddInputZp(builder, inputZp): builder.PrependInt32Slot(0, inputZp, 0)
+def PadQuantInfoEnd(builder): return builder.EndObject()
diff --git a/ethosu/vela/tosa/Pool2dAttribute.py b/ethosu/vela/tosa/Pool2dAttribute.py
new file mode 100644
index 0000000..a45c1f1
--- /dev/null
+++ b/ethosu/vela/tosa/Pool2dAttribute.py
@@ -0,0 +1,94 @@
+# automatically generated by the FlatBuffers compiler, do not modify
+
+# namespace: tosa
+
+import flatbuffers
+
+class Pool2dAttribute(object):
+ __slots__ = ['_tab']
+
+ @classmethod
+ def GetRootAsPool2dAttribute(cls, buf, offset):
+ n = flatbuffers.encode.Get(flatbuffers.packer.uoffset, buf, offset)
+ x = Pool2dAttribute()
+ x.Init(buf, n + offset)
+ return x
+
+ # Pool2dAttribute
+ def Init(self, buf, pos):
+ self._tab = flatbuffers.table.Table(buf, pos)
+
+ # Pool2dAttribute
+ def Padding(self, j):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(4))
+ if o != 0:
+ a = self._tab.Vector(o)
+ return self._tab.Get(flatbuffers.number_types.Int32Flags, a + flatbuffers.number_types.UOffsetTFlags.py_type(j * 4))
+ return 0
+
+ # Pool2dAttribute
+ def PaddingAsNumpy(self):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(4))
+ if o != 0:
+ return self._tab.GetVectorAsNumpy(flatbuffers.number_types.Int32Flags, o)
+ return 0
+
+ # Pool2dAttribute
+ def PaddingLength(self):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(4))
+ if o != 0:
+ return self._tab.VectorLen(o)
+ return 0
+
+ # Pool2dAttribute
+ def Kernel(self, j):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(6))
+ if o != 0:
+ a = self._tab.Vector(o)
+ return self._tab.Get(flatbuffers.number_types.Int32Flags, a + flatbuffers.number_types.UOffsetTFlags.py_type(j * 4))
+ return 0
+
+ # Pool2dAttribute
+ def KernelAsNumpy(self):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(6))
+ if o != 0:
+ return self._tab.GetVectorAsNumpy(flatbuffers.number_types.Int32Flags, o)
+ return 0
+
+ # Pool2dAttribute
+ def KernelLength(self):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(6))
+ if o != 0:
+ return self._tab.VectorLen(o)
+ return 0
+
+ # Pool2dAttribute
+ def Stride(self, j):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(8))
+ if o != 0:
+ a = self._tab.Vector(o)
+ return self._tab.Get(flatbuffers.number_types.Int32Flags, a + flatbuffers.number_types.UOffsetTFlags.py_type(j * 4))
+ return 0
+
+ # Pool2dAttribute
+ def StrideAsNumpy(self):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(8))
+ if o != 0:
+ return self._tab.GetVectorAsNumpy(flatbuffers.number_types.Int32Flags, o)
+ return 0
+
+ # Pool2dAttribute
+ def StrideLength(self):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(8))
+ if o != 0:
+ return self._tab.VectorLen(o)
+ return 0
+
+def Pool2dAttributeStart(builder): builder.StartObject(3)
+def Pool2dAttributeAddPadding(builder, padding): builder.PrependUOffsetTRelativeSlot(0, flatbuffers.number_types.UOffsetTFlags.py_type(padding), 0)
+def Pool2dAttributeStartPaddingVector(builder, numElems): return builder.StartVector(4, numElems, 4)
+def Pool2dAttributeAddKernel(builder, kernel): builder.PrependUOffsetTRelativeSlot(1, flatbuffers.number_types.UOffsetTFlags.py_type(kernel), 0)
+def Pool2dAttributeStartKernelVector(builder, numElems): return builder.StartVector(4, numElems, 4)
+def Pool2dAttributeAddStride(builder, stride): builder.PrependUOffsetTRelativeSlot(2, flatbuffers.number_types.UOffsetTFlags.py_type(stride), 0)
+def Pool2dAttributeStartStrideVector(builder, numElems): return builder.StartVector(4, numElems, 4)
+def Pool2dAttributeEnd(builder): return builder.EndObject()
diff --git a/ethosu/vela/tosa/QuantInfo.py b/ethosu/vela/tosa/QuantInfo.py
new file mode 100644
index 0000000..ffdfd32
--- /dev/null
+++ b/ethosu/vela/tosa/QuantInfo.py
@@ -0,0 +1,11 @@
+# automatically generated by the FlatBuffers compiler, do not modify
+
+# namespace: tosa
+
+class QuantInfo(object):
+ NONE = 0
+ UnaryQuantInfo = 1
+ ConvQuantInfo = 2
+ MatMulQuantInfo = 3
+ PadQuantInfo = 4
+
diff --git a/ethosu/vela/tosa/ReluNAttribute.py b/ethosu/vela/tosa/ReluNAttribute.py
new file mode 100644
index 0000000..008b535
--- /dev/null
+++ b/ethosu/vela/tosa/ReluNAttribute.py
@@ -0,0 +1,38 @@
+# automatically generated by the FlatBuffers compiler, do not modify
+
+# namespace: tosa
+
+import flatbuffers
+
+class ReluNAttribute(object):
+ __slots__ = ['_tab']
+
+ @classmethod
+ def GetRootAsReluNAttribute(cls, buf, offset):
+ n = flatbuffers.encode.Get(flatbuffers.packer.uoffset, buf, offset)
+ x = ReluNAttribute()
+ x.Init(buf, n + offset)
+ return x
+
+ # ReluNAttribute
+ def Init(self, buf, pos):
+ self._tab = flatbuffers.table.Table(buf, pos)
+
+ # ReluNAttribute
+ def MaxInt(self):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(4))
+ if o != 0:
+ return self._tab.Get(flatbuffers.number_types.Int32Flags, o + self._tab.Pos)
+ return 0
+
+ # ReluNAttribute
+ def MaxFp(self):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(6))
+ if o != 0:
+ return self._tab.Get(flatbuffers.number_types.Float32Flags, o + self._tab.Pos)
+ return 0.0
+
+def ReluNAttributeStart(builder): builder.StartObject(2)
+def ReluNAttributeAddMaxInt(builder, maxInt): builder.PrependInt32Slot(0, maxInt, 0)
+def ReluNAttributeAddMaxFp(builder, maxFp): builder.PrependFloat32Slot(1, maxFp, 0.0)
+def ReluNAttributeEnd(builder): return builder.EndObject()
diff --git a/ethosu/vela/tosa/RescaleAttribute.py b/ethosu/vela/tosa/RescaleAttribute.py
new file mode 100644
index 0000000..1aa6707
--- /dev/null
+++ b/ethosu/vela/tosa/RescaleAttribute.py
@@ -0,0 +1,110 @@
+# automatically generated by the FlatBuffers compiler, do not modify
+
+# namespace: tosa
+
+import flatbuffers
+
+class RescaleAttribute(object):
+ __slots__ = ['_tab']
+
+ @classmethod
+ def GetRootAsRescaleAttribute(cls, buf, offset):
+ n = flatbuffers.encode.Get(flatbuffers.packer.uoffset, buf, offset)
+ x = RescaleAttribute()
+ x.Init(buf, n + offset)
+ return x
+
+ # RescaleAttribute
+ def Init(self, buf, pos):
+ self._tab = flatbuffers.table.Table(buf, pos)
+
+ # RescaleAttribute
+ def InputZp(self):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(4))
+ if o != 0:
+ return self._tab.Get(flatbuffers.number_types.Int32Flags, o + self._tab.Pos)
+ return 0
+
+ # RescaleAttribute
+ def OutputZp(self):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(6))
+ if o != 0:
+ return self._tab.Get(flatbuffers.number_types.Int32Flags, o + self._tab.Pos)
+ return 0
+
+ # RescaleAttribute
+ def Multiplier(self, j):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(8))
+ if o != 0:
+ a = self._tab.Vector(o)
+ return self._tab.Get(flatbuffers.number_types.Int32Flags, a + flatbuffers.number_types.UOffsetTFlags.py_type(j * 4))
+ return 0
+
+ # RescaleAttribute
+ def MultiplierAsNumpy(self):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(8))
+ if o != 0:
+ return self._tab.GetVectorAsNumpy(flatbuffers.number_types.Int32Flags, o)
+ return 0
+
+ # RescaleAttribute
+ def MultiplierLength(self):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(8))
+ if o != 0:
+ return self._tab.VectorLen(o)
+ return 0
+
+ # RescaleAttribute
+ def Shift(self, j):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(10))
+ if o != 0:
+ a = self._tab.Vector(o)
+ return self._tab.Get(flatbuffers.number_types.Int32Flags, a + flatbuffers.number_types.UOffsetTFlags.py_type(j * 4))
+ return 0
+
+ # RescaleAttribute
+ def ShiftAsNumpy(self):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(10))
+ if o != 0:
+ return self._tab.GetVectorAsNumpy(flatbuffers.number_types.Int32Flags, o)
+ return 0
+
+ # RescaleAttribute
+ def ShiftLength(self):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(10))
+ if o != 0:
+ return self._tab.VectorLen(o)
+ return 0
+
+ # RescaleAttribute
+ def Scale32(self):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(12))
+ if o != 0:
+ return bool(self._tab.Get(flatbuffers.number_types.BoolFlags, o + self._tab.Pos))
+ return False
+
+ # RescaleAttribute
+ def DoubleRound(self):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(14))
+ if o != 0:
+ return bool(self._tab.Get(flatbuffers.number_types.BoolFlags, o + self._tab.Pos))
+ return False
+
+ # RescaleAttribute
+ def PerChannel(self):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(16))
+ if o != 0:
+ return bool(self._tab.Get(flatbuffers.number_types.BoolFlags, o + self._tab.Pos))
+ return False
+
+def RescaleAttributeStart(builder): builder.StartObject(7)
+def RescaleAttributeAddInputZp(builder, inputZp): builder.PrependInt32Slot(0, inputZp, 0)
+def RescaleAttributeAddOutputZp(builder, outputZp): builder.PrependInt32Slot(1, outputZp, 0)
+def RescaleAttributeAddMultiplier(builder, multiplier): builder.PrependUOffsetTRelativeSlot(2, flatbuffers.number_types.UOffsetTFlags.py_type(multiplier), 0)
+def RescaleAttributeStartMultiplierVector(builder, numElems): return builder.StartVector(4, numElems, 4)
+def RescaleAttributeAddShift(builder, shift): builder.PrependUOffsetTRelativeSlot(3, flatbuffers.number_types.UOffsetTFlags.py_type(shift), 0)
+def RescaleAttributeStartShiftVector(builder, numElems): return builder.StartVector(4, numElems, 4)
+def RescaleAttributeAddScale32(builder, scale32): builder.PrependBoolSlot(4, scale32, 0)
+def RescaleAttributeAddDoubleRound(builder, doubleRound): builder.PrependBoolSlot(5, doubleRound, 0)
+def RescaleAttributeAddPerChannel(builder, perChannel): builder.PrependBoolSlot(6, perChannel, 0)
+def RescaleAttributeEnd(builder): return builder.EndObject()
diff --git a/ethosu/vela/tosa/ReshapeAttribute.py b/ethosu/vela/tosa/ReshapeAttribute.py
new file mode 100644
index 0000000..629b6c2
--- /dev/null
+++ b/ethosu/vela/tosa/ReshapeAttribute.py
@@ -0,0 +1,46 @@
+# automatically generated by the FlatBuffers compiler, do not modify
+
+# namespace: tosa
+
+import flatbuffers
+
+class ReshapeAttribute(object):
+ __slots__ = ['_tab']
+
+ @classmethod
+ def GetRootAsReshapeAttribute(cls, buf, offset):
+ n = flatbuffers.encode.Get(flatbuffers.packer.uoffset, buf, offset)
+ x = ReshapeAttribute()
+ x.Init(buf, n + offset)
+ return x
+
+ # ReshapeAttribute
+ def Init(self, buf, pos):
+ self._tab = flatbuffers.table.Table(buf, pos)
+
+ # ReshapeAttribute
+ def Shape(self, j):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(4))
+ if o != 0:
+ a = self._tab.Vector(o)
+ return self._tab.Get(flatbuffers.number_types.Int32Flags, a + flatbuffers.number_types.UOffsetTFlags.py_type(j * 4))
+ return 0
+
+ # ReshapeAttribute
+ def ShapeAsNumpy(self):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(4))
+ if o != 0:
+ return self._tab.GetVectorAsNumpy(flatbuffers.number_types.Int32Flags, o)
+ return 0
+
+ # ReshapeAttribute
+ def ShapeLength(self):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(4))
+ if o != 0:
+ return self._tab.VectorLen(o)
+ return 0
+
+def ReshapeAttributeStart(builder): builder.StartObject(1)
+def ReshapeAttributeAddShape(builder, shape): builder.PrependUOffsetTRelativeSlot(0, flatbuffers.number_types.UOffsetTFlags.py_type(shape), 0)
+def ReshapeAttributeStartShapeVector(builder, numElems): return builder.StartVector(4, numElems, 4)
+def ReshapeAttributeEnd(builder): return builder.EndObject()
diff --git a/ethosu/vela/tosa/ResizeAttribute.py b/ethosu/vela/tosa/ResizeAttribute.py
new file mode 100644
index 0000000..89156d1
--- /dev/null
+++ b/ethosu/vela/tosa/ResizeAttribute.py
@@ -0,0 +1,158 @@
+# automatically generated by the FlatBuffers compiler, do not modify
+
+# namespace: tosa
+
+import flatbuffers
+
+class ResizeAttribute(object):
+ __slots__ = ['_tab']
+
+ @classmethod
+ def GetRootAsResizeAttribute(cls, buf, offset):
+ n = flatbuffers.encode.Get(flatbuffers.packer.uoffset, buf, offset)
+ x = ResizeAttribute()
+ x.Init(buf, n + offset)
+ return x
+
+ # ResizeAttribute
+ def Init(self, buf, pos):
+ self._tab = flatbuffers.table.Table(buf, pos)
+
+ # ResizeAttribute
+ def OutputSize(self, j):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(4))
+ if o != 0:
+ a = self._tab.Vector(o)
+ return self._tab.Get(flatbuffers.number_types.Int32Flags, a + flatbuffers.number_types.UOffsetTFlags.py_type(j * 4))
+ return 0
+
+ # ResizeAttribute
+ def OutputSizeAsNumpy(self):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(4))
+ if o != 0:
+ return self._tab.GetVectorAsNumpy(flatbuffers.number_types.Int32Flags, o)
+ return 0
+
+ # ResizeAttribute
+ def OutputSizeLength(self):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(4))
+ if o != 0:
+ return self._tab.VectorLen(o)
+ return 0
+
+ # ResizeAttribute
+ def Stride(self, j):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(6))
+ if o != 0:
+ a = self._tab.Vector(o)
+ return self._tab.Get(flatbuffers.number_types.Int32Flags, a + flatbuffers.number_types.UOffsetTFlags.py_type(j * 4))
+ return 0
+
+ # ResizeAttribute
+ def StrideAsNumpy(self):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(6))
+ if o != 0:
+ return self._tab.GetVectorAsNumpy(flatbuffers.number_types.Int32Flags, o)
+ return 0
+
+ # ResizeAttribute
+ def StrideLength(self):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(6))
+ if o != 0:
+ return self._tab.VectorLen(o)
+ return 0
+
+ # ResizeAttribute
+ def Offset(self, j):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(8))
+ if o != 0:
+ a = self._tab.Vector(o)
+ return self._tab.Get(flatbuffers.number_types.Int32Flags, a + flatbuffers.number_types.UOffsetTFlags.py_type(j * 4))
+ return 0
+
+ # ResizeAttribute
+ def OffsetAsNumpy(self):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(8))
+ if o != 0:
+ return self._tab.GetVectorAsNumpy(flatbuffers.number_types.Int32Flags, o)
+ return 0
+
+ # ResizeAttribute
+ def OffsetLength(self):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(8))
+ if o != 0:
+ return self._tab.VectorLen(o)
+ return 0
+
+ # ResizeAttribute
+ def Shift(self):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(10))
+ if o != 0:
+ return self._tab.Get(flatbuffers.number_types.Int32Flags, o + self._tab.Pos)
+ return 0
+
+ # ResizeAttribute
+ def StrideFp(self, j):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(12))
+ if o != 0:
+ a = self._tab.Vector(o)
+ return self._tab.Get(flatbuffers.number_types.Float32Flags, a + flatbuffers.number_types.UOffsetTFlags.py_type(j * 4))
+ return 0
+
+ # ResizeAttribute
+ def StrideFpAsNumpy(self):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(12))
+ if o != 0:
+ return self._tab.GetVectorAsNumpy(flatbuffers.number_types.Float32Flags, o)
+ return 0
+
+ # ResizeAttribute
+ def StrideFpLength(self):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(12))
+ if o != 0:
+ return self._tab.VectorLen(o)
+ return 0
+
+ # ResizeAttribute
+ def OffsetFp(self, j):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(14))
+ if o != 0:
+ a = self._tab.Vector(o)
+ return self._tab.Get(flatbuffers.number_types.Float32Flags, a + flatbuffers.number_types.UOffsetTFlags.py_type(j * 4))
+ return 0
+
+ # ResizeAttribute
+ def OffsetFpAsNumpy(self):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(14))
+ if o != 0:
+ return self._tab.GetVectorAsNumpy(flatbuffers.number_types.Float32Flags, o)
+ return 0
+
+ # ResizeAttribute
+ def OffsetFpLength(self):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(14))
+ if o != 0:
+ return self._tab.VectorLen(o)
+ return 0
+
+ # ResizeAttribute
+ def Mode(self):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(16))
+ if o != 0:
+ return self._tab.Get(flatbuffers.number_types.Uint32Flags, o + self._tab.Pos)
+ return 0
+
+def ResizeAttributeStart(builder): builder.StartObject(7)
+def ResizeAttributeAddOutputSize(builder, outputSize): builder.PrependUOffsetTRelativeSlot(0, flatbuffers.number_types.UOffsetTFlags.py_type(outputSize), 0)
+def ResizeAttributeStartOutputSizeVector(builder, numElems): return builder.StartVector(4, numElems, 4)
+def ResizeAttributeAddStride(builder, stride): builder.PrependUOffsetTRelativeSlot(1, flatbuffers.number_types.UOffsetTFlags.py_type(stride), 0)
+def ResizeAttributeStartStrideVector(builder, numElems): return builder.StartVector(4, numElems, 4)
+def ResizeAttributeAddOffset(builder, offset): builder.PrependUOffsetTRelativeSlot(2, flatbuffers.number_types.UOffsetTFlags.py_type(offset), 0)
+def ResizeAttributeStartOffsetVector(builder, numElems): return builder.StartVector(4, numElems, 4)
+def ResizeAttributeAddShift(builder, shift): builder.PrependInt32Slot(3, shift, 0)
+def ResizeAttributeAddStrideFp(builder, strideFp): builder.PrependUOffsetTRelativeSlot(4, flatbuffers.number_types.UOffsetTFlags.py_type(strideFp), 0)
+def ResizeAttributeStartStrideFpVector(builder, numElems): return builder.StartVector(4, numElems, 4)
+def ResizeAttributeAddOffsetFp(builder, offsetFp): builder.PrependUOffsetTRelativeSlot(5, flatbuffers.number_types.UOffsetTFlags.py_type(offsetFp), 0)
+def ResizeAttributeStartOffsetFpVector(builder, numElems): return builder.StartVector(4, numElems, 4)
+def ResizeAttributeAddMode(builder, mode): builder.PrependUint32Slot(6, mode, 0)
+def ResizeAttributeEnd(builder): return builder.EndObject()
diff --git a/ethosu/vela/tosa/ResizeMode.py b/ethosu/vela/tosa/ResizeMode.py
new file mode 100644
index 0000000..65bcd5d
--- /dev/null
+++ b/ethosu/vela/tosa/ResizeMode.py
@@ -0,0 +1,9 @@
+# automatically generated by the FlatBuffers compiler, do not modify
+
+# namespace: tosa
+
+class ResizeMode(object):
+ UNKNOWN = 0
+ NEAREST = 1
+ BILINEAR = 2
+
diff --git a/ethosu/vela/tosa/SliceAttribute.py b/ethosu/vela/tosa/SliceAttribute.py
new file mode 100644
index 0000000..d2f9958
--- /dev/null
+++ b/ethosu/vela/tosa/SliceAttribute.py
@@ -0,0 +1,70 @@
+# automatically generated by the FlatBuffers compiler, do not modify
+
+# namespace: tosa
+
+import flatbuffers
+
+class SliceAttribute(object):
+ __slots__ = ['_tab']
+
+ @classmethod
+ def GetRootAsSliceAttribute(cls, buf, offset):
+ n = flatbuffers.encode.Get(flatbuffers.packer.uoffset, buf, offset)
+ x = SliceAttribute()
+ x.Init(buf, n + offset)
+ return x
+
+ # SliceAttribute
+ def Init(self, buf, pos):
+ self._tab = flatbuffers.table.Table(buf, pos)
+
+ # SliceAttribute
+ def Begin(self, j):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(4))
+ if o != 0:
+ a = self._tab.Vector(o)
+ return self._tab.Get(flatbuffers.number_types.Int32Flags, a + flatbuffers.number_types.UOffsetTFlags.py_type(j * 4))
+ return 0
+
+ # SliceAttribute
+ def BeginAsNumpy(self):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(4))
+ if o != 0:
+ return self._tab.GetVectorAsNumpy(flatbuffers.number_types.Int32Flags, o)
+ return 0
+
+ # SliceAttribute
+ def BeginLength(self):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(4))
+ if o != 0:
+ return self._tab.VectorLen(o)
+ return 0
+
+ # SliceAttribute
+ def Size(self, j):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(6))
+ if o != 0:
+ a = self._tab.Vector(o)
+ return self._tab.Get(flatbuffers.number_types.Int32Flags, a + flatbuffers.number_types.UOffsetTFlags.py_type(j * 4))
+ return 0
+
+ # SliceAttribute
+ def SizeAsNumpy(self):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(6))
+ if o != 0:
+ return self._tab.GetVectorAsNumpy(flatbuffers.number_types.Int32Flags, o)
+ return 0
+
+ # SliceAttribute
+ def SizeLength(self):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(6))
+ if o != 0:
+ return self._tab.VectorLen(o)
+ return 0
+
+def SliceAttributeStart(builder): builder.StartObject(2)
+def SliceAttributeAddBegin(builder, begin): builder.PrependUOffsetTRelativeSlot(0, flatbuffers.number_types.UOffsetTFlags.py_type(begin), 0)
+def SliceAttributeStartBeginVector(builder, numElems): return builder.StartVector(4, numElems, 4)
+def SliceAttributeAddSize(builder, size): builder.PrependUOffsetTRelativeSlot(1, flatbuffers.number_types.UOffsetTFlags.py_type(size), 0)
+def SliceAttributeStartSizeVector(builder, numElems): return builder.StartVector(4, numElems, 4)
+def SliceAttributeEnd(builder): return builder.EndObject()
diff --git a/ethosu/vela/tosa/TileAttribute.py b/ethosu/vela/tosa/TileAttribute.py
new file mode 100644
index 0000000..f1b721b
--- /dev/null
+++ b/ethosu/vela/tosa/TileAttribute.py
@@ -0,0 +1,46 @@
+# automatically generated by the FlatBuffers compiler, do not modify
+
+# namespace: tosa
+
+import flatbuffers
+
+class TileAttribute(object):
+ __slots__ = ['_tab']
+
+ @classmethod
+ def GetRootAsTileAttribute(cls, buf, offset):
+ n = flatbuffers.encode.Get(flatbuffers.packer.uoffset, buf, offset)
+ x = TileAttribute()
+ x.Init(buf, n + offset)
+ return x
+
+ # TileAttribute
+ def Init(self, buf, pos):
+ self._tab = flatbuffers.table.Table(buf, pos)
+
+ # TileAttribute
+ def Multiples(self, j):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(4))
+ if o != 0:
+ a = self._tab.Vector(o)
+ return self._tab.Get(flatbuffers.number_types.Int32Flags, a + flatbuffers.number_types.UOffsetTFlags.py_type(j * 4))
+ return 0
+
+ # TileAttribute
+ def MultiplesAsNumpy(self):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(4))
+ if o != 0:
+ return self._tab.GetVectorAsNumpy(flatbuffers.number_types.Int32Flags, o)
+ return 0
+
+ # TileAttribute
+ def MultiplesLength(self):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(4))
+ if o != 0:
+ return self._tab.VectorLen(o)
+ return 0
+
+def TileAttributeStart(builder): builder.StartObject(1)
+def TileAttributeAddMultiples(builder, multiples): builder.PrependUOffsetTRelativeSlot(0, flatbuffers.number_types.UOffsetTFlags.py_type(multiples), 0)
+def TileAttributeStartMultiplesVector(builder, numElems): return builder.StartVector(4, numElems, 4)
+def TileAttributeEnd(builder): return builder.EndObject()
diff --git a/ethosu/vela/tosa/TosaBasicBlock.py b/ethosu/vela/tosa/TosaBasicBlock.py
new file mode 100644
index 0000000..8f1604a
--- /dev/null
+++ b/ethosu/vela/tosa/TosaBasicBlock.py
@@ -0,0 +1,108 @@
+# automatically generated by the FlatBuffers compiler, do not modify
+
+# namespace: tosa
+
+import flatbuffers
+
+class TosaBasicBlock(object):
+ __slots__ = ['_tab']
+
+ @classmethod
+ def GetRootAsTosaBasicBlock(cls, buf, offset):
+ n = flatbuffers.encode.Get(flatbuffers.packer.uoffset, buf, offset)
+ x = TosaBasicBlock()
+ x.Init(buf, n + offset)
+ return x
+
+ # TosaBasicBlock
+ def Init(self, buf, pos):
+ self._tab = flatbuffers.table.Table(buf, pos)
+
+ # TosaBasicBlock
+ def Name(self):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(4))
+ if o != 0:
+ return self._tab.String(o + self._tab.Pos)
+ return None
+
+ # TosaBasicBlock
+ def Operators(self, j):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(6))
+ if o != 0:
+ x = self._tab.Vector(o)
+ x += flatbuffers.number_types.UOffsetTFlags.py_type(j) * 4
+ x = self._tab.Indirect(x)
+ from .TosaOperator import TosaOperator
+ obj = TosaOperator()
+ obj.Init(self._tab.Bytes, x)
+ return obj
+ return None
+
+ # TosaBasicBlock
+ def OperatorsLength(self):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(6))
+ if o != 0:
+ return self._tab.VectorLen(o)
+ return 0
+
+ # TosaBasicBlock
+ def Tensors(self, j):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(8))
+ if o != 0:
+ x = self._tab.Vector(o)
+ x += flatbuffers.number_types.UOffsetTFlags.py_type(j) * 4
+ x = self._tab.Indirect(x)
+ from .TosaTensor import TosaTensor
+ obj = TosaTensor()
+ obj.Init(self._tab.Bytes, x)
+ return obj
+ return None
+
+ # TosaBasicBlock
+ def TensorsLength(self):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(8))
+ if o != 0:
+ return self._tab.VectorLen(o)
+ return 0
+
+ # TosaBasicBlock
+ def Inputs(self, j):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(10))
+ if o != 0:
+ a = self._tab.Vector(o)
+ return self._tab.String(a + flatbuffers.number_types.UOffsetTFlags.py_type(j * 4))
+ return ""
+
+ # TosaBasicBlock
+ def InputsLength(self):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(10))
+ if o != 0:
+ return self._tab.VectorLen(o)
+ return 0
+
+ # TosaBasicBlock
+ def Outputs(self, j):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(12))
+ if o != 0:
+ a = self._tab.Vector(o)
+ return self._tab.String(a + flatbuffers.number_types.UOffsetTFlags.py_type(j * 4))
+ return ""
+
+ # TosaBasicBlock
+ def OutputsLength(self):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(12))
+ if o != 0:
+ return self._tab.VectorLen(o)
+ return 0
+
+def TosaBasicBlockStart(builder): builder.StartObject(5)
+def TosaBasicBlockAddName(builder, name): builder.PrependUOffsetTRelativeSlot(0, flatbuffers.number_types.UOffsetTFlags.py_type(name), 0)
+def TosaBasicBlockAddOperators(builder, operators): builder.PrependUOffsetTRelativeSlot(1, flatbuffers.number_types.UOffsetTFlags.py_type(operators), 0)
+def TosaBasicBlockStartOperatorsVector(builder, numElems): return builder.StartVector(4, numElems, 4)
+def TosaBasicBlockAddTensors(builder, tensors): builder.PrependUOffsetTRelativeSlot(2, flatbuffers.number_types.UOffsetTFlags.py_type(tensors), 0)
+def TosaBasicBlockStartTensorsVector(builder, numElems): return builder.StartVector(4, numElems, 4)
+def TosaBasicBlockAddInputs(builder, inputs): builder.PrependUOffsetTRelativeSlot(3, flatbuffers.number_types.UOffsetTFlags.py_type(inputs), 0)
+def TosaBasicBlockStartInputsVector(builder, numElems): return builder.StartVector(4, numElems, 4)
+def TosaBasicBlockAddOutputs(builder, outputs): builder.PrependUOffsetTRelativeSlot(4, flatbuffers.number_types.UOffsetTFlags.py_type(outputs), 0)
+def TosaBasicBlockStartOutputsVector(builder, numElems): return builder.StartVector(4, numElems, 4)
+def TosaBasicBlockEnd(builder): return builder.EndObject()
diff --git a/ethosu/vela/tosa/TosaGraph.py b/ethosu/vela/tosa/TosaGraph.py
new file mode 100644
index 0000000..f54a44a
--- /dev/null
+++ b/ethosu/vela/tosa/TosaGraph.py
@@ -0,0 +1,56 @@
+# automatically generated by the FlatBuffers compiler, do not modify
+
+# namespace: tosa
+
+import flatbuffers
+
+class TosaGraph(object):
+ __slots__ = ['_tab']
+
+ @classmethod
+ def GetRootAsTosaGraph(cls, buf, offset):
+ n = flatbuffers.encode.Get(flatbuffers.packer.uoffset, buf, offset)
+ x = TosaGraph()
+ x.Init(buf, n + offset)
+ return x
+
+ # TosaGraph
+ def Init(self, buf, pos):
+ self._tab = flatbuffers.table.Table(buf, pos)
+
+ # TosaGraph
+ def Version(self):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(4))
+ if o != 0:
+ x = self._tab.Indirect(o + self._tab.Pos)
+ from .Version import Version
+ obj = Version()
+ obj.Init(self._tab.Bytes, x)
+ return obj
+ return None
+
+ # TosaGraph
+ def Blocks(self, j):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(6))
+ if o != 0:
+ x = self._tab.Vector(o)
+ x += flatbuffers.number_types.UOffsetTFlags.py_type(j) * 4
+ x = self._tab.Indirect(x)
+ from .TosaBasicBlock import TosaBasicBlock
+ obj = TosaBasicBlock()
+ obj.Init(self._tab.Bytes, x)
+ return obj
+ return None
+
+ # TosaGraph
+ def BlocksLength(self):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(6))
+ if o != 0:
+ return self._tab.VectorLen(o)
+ return 0
+
+def TosaGraphStart(builder): builder.StartObject(2)
+def TosaGraphAddVersion(builder, version): builder.PrependUOffsetTRelativeSlot(0, flatbuffers.number_types.UOffsetTFlags.py_type(version), 0)
+def TosaGraphAddBlocks(builder, blocks): builder.PrependUOffsetTRelativeSlot(1, flatbuffers.number_types.UOffsetTFlags.py_type(blocks), 0)
+def TosaGraphStartBlocksVector(builder, numElems): return builder.StartVector(4, numElems, 4)
+def TosaGraphEnd(builder): return builder.EndObject()
diff --git a/ethosu/vela/tosa/TosaOperator.py b/ethosu/vela/tosa/TosaOperator.py
new file mode 100644
index 0000000..c45efda
--- /dev/null
+++ b/ethosu/vela/tosa/TosaOperator.py
@@ -0,0 +1,102 @@
+# automatically generated by the FlatBuffers compiler, do not modify
+
+# namespace: tosa
+
+import flatbuffers
+
+class TosaOperator(object):
+ __slots__ = ['_tab']
+
+ @classmethod
+ def GetRootAsTosaOperator(cls, buf, offset):
+ n = flatbuffers.encode.Get(flatbuffers.packer.uoffset, buf, offset)
+ x = TosaOperator()
+ x.Init(buf, n + offset)
+ return x
+
+ # TosaOperator
+ def Init(self, buf, pos):
+ self._tab = flatbuffers.table.Table(buf, pos)
+
+ # TosaOperator
+ def Op(self):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(4))
+ if o != 0:
+ return self._tab.Get(flatbuffers.number_types.Uint32Flags, o + self._tab.Pos)
+ return 0
+
+ # TosaOperator
+ def AttributeType(self):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(6))
+ if o != 0:
+ return self._tab.Get(flatbuffers.number_types.Uint8Flags, o + self._tab.Pos)
+ return 0
+
+ # TosaOperator
+ def Attribute(self):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(8))
+ if o != 0:
+ from flatbuffers.table import Table
+ obj = Table(bytearray(), 0)
+ self._tab.Union(obj, o)
+ return obj
+ return None
+
+ # TosaOperator
+ def Inputs(self, j):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(10))
+ if o != 0:
+ a = self._tab.Vector(o)
+ return self._tab.String(a + flatbuffers.number_types.UOffsetTFlags.py_type(j * 4))
+ return ""
+
+ # TosaOperator
+ def InputsLength(self):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(10))
+ if o != 0:
+ return self._tab.VectorLen(o)
+ return 0
+
+ # TosaOperator
+ def Outputs(self, j):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(12))
+ if o != 0:
+ a = self._tab.Vector(o)
+ return self._tab.String(a + flatbuffers.number_types.UOffsetTFlags.py_type(j * 4))
+ return ""
+
+ # TosaOperator
+ def OutputsLength(self):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(12))
+ if o != 0:
+ return self._tab.VectorLen(o)
+ return 0
+
+ # TosaOperator
+ def QuantInfoType(self):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(14))
+ if o != 0:
+ return self._tab.Get(flatbuffers.number_types.Uint8Flags, o + self._tab.Pos)
+ return 0
+
+ # TosaOperator
+ def QuantInfo(self):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(16))
+ if o != 0:
+ from flatbuffers.table import Table
+ obj = Table(bytearray(), 0)
+ self._tab.Union(obj, o)
+ return obj
+ return None
+
+def TosaOperatorStart(builder): builder.StartObject(7)
+def TosaOperatorAddOp(builder, op): builder.PrependUint32Slot(0, op, 0)
+def TosaOperatorAddAttributeType(builder, attributeType): builder.PrependUint8Slot(1, attributeType, 0)
+def TosaOperatorAddAttribute(builder, attribute): builder.PrependUOffsetTRelativeSlot(2, flatbuffers.number_types.UOffsetTFlags.py_type(attribute), 0)
+def TosaOperatorAddInputs(builder, inputs): builder.PrependUOffsetTRelativeSlot(3, flatbuffers.number_types.UOffsetTFlags.py_type(inputs), 0)
+def TosaOperatorStartInputsVector(builder, numElems): return builder.StartVector(4, numElems, 4)
+def TosaOperatorAddOutputs(builder, outputs): builder.PrependUOffsetTRelativeSlot(4, flatbuffers.number_types.UOffsetTFlags.py_type(outputs), 0)
+def TosaOperatorStartOutputsVector(builder, numElems): return builder.StartVector(4, numElems, 4)
+def TosaOperatorAddQuantInfoType(builder, quantInfoType): builder.PrependUint8Slot(5, quantInfoType, 0)
+def TosaOperatorAddQuantInfo(builder, quantInfo): builder.PrependUOffsetTRelativeSlot(6, flatbuffers.number_types.UOffsetTFlags.py_type(quantInfo), 0)
+def TosaOperatorEnd(builder): return builder.EndObject()
diff --git a/ethosu/vela/tosa/TosaTensor.py b/ethosu/vela/tosa/TosaTensor.py
new file mode 100644
index 0000000..2a397db
--- /dev/null
+++ b/ethosu/vela/tosa/TosaTensor.py
@@ -0,0 +1,70 @@
+# automatically generated by the FlatBuffers compiler, do not modify
+
+# namespace: tosa
+
+import flatbuffers
+
+class TosaTensor(object):
+ __slots__ = ['_tab']
+
+ @classmethod
+ def GetRootAsTosaTensor(cls, buf, offset):
+ n = flatbuffers.encode.Get(flatbuffers.packer.uoffset, buf, offset)
+ x = TosaTensor()
+ x.Init(buf, n + offset)
+ return x
+
+ # TosaTensor
+ def Init(self, buf, pos):
+ self._tab = flatbuffers.table.Table(buf, pos)
+
+ # TosaTensor
+ def Name(self):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(4))
+ if o != 0:
+ return self._tab.String(o + self._tab.Pos)
+ return None
+
+ # TosaTensor
+ def Shape(self, j):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(6))
+ if o != 0:
+ a = self._tab.Vector(o)
+ return self._tab.Get(flatbuffers.number_types.Int32Flags, a + flatbuffers.number_types.UOffsetTFlags.py_type(j * 4))
+ return 0
+
+ # TosaTensor
+ def ShapeAsNumpy(self):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(6))
+ if o != 0:
+ return self._tab.GetVectorAsNumpy(flatbuffers.number_types.Int32Flags, o)
+ return 0
+
+ # TosaTensor
+ def ShapeLength(self):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(6))
+ if o != 0:
+ return self._tab.VectorLen(o)
+ return 0
+
+ # TosaTensor
+ def Type(self):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(8))
+ if o != 0:
+ return self._tab.Get(flatbuffers.number_types.Uint32Flags, o + self._tab.Pos)
+ return 0
+
+ # TosaTensor
+ def NpyFilename(self):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(10))
+ if o != 0:
+ return self._tab.String(o + self._tab.Pos)
+ return None
+
+def TosaTensorStart(builder): builder.StartObject(4)
+def TosaTensorAddName(builder, name): builder.PrependUOffsetTRelativeSlot(0, flatbuffers.number_types.UOffsetTFlags.py_type(name), 0)
+def TosaTensorAddShape(builder, shape): builder.PrependUOffsetTRelativeSlot(1, flatbuffers.number_types.UOffsetTFlags.py_type(shape), 0)
+def TosaTensorStartShapeVector(builder, numElems): return builder.StartVector(4, numElems, 4)
+def TosaTensorAddType(builder, type): builder.PrependUint32Slot(2, type, 0)
+def TosaTensorAddNpyFilename(builder, npyFilename): builder.PrependUOffsetTRelativeSlot(3, flatbuffers.number_types.UOffsetTFlags.py_type(npyFilename), 0)
+def TosaTensorEnd(builder): return builder.EndObject()
diff --git a/ethosu/vela/tosa/TransposeConv2dAttribute.py b/ethosu/vela/tosa/TransposeConv2dAttribute.py
new file mode 100644
index 0000000..80f544f
--- /dev/null
+++ b/ethosu/vela/tosa/TransposeConv2dAttribute.py
@@ -0,0 +1,118 @@
+# automatically generated by the FlatBuffers compiler, do not modify
+
+# namespace: tosa
+
+import flatbuffers
+
+class TransposeConv2dAttribute(object):
+ __slots__ = ['_tab']
+
+ @classmethod
+ def GetRootAsTransposeConv2dAttribute(cls, buf, offset):
+ n = flatbuffers.encode.Get(flatbuffers.packer.uoffset, buf, offset)
+ x = TransposeConv2dAttribute()
+ x.Init(buf, n + offset)
+ return x
+
+ # TransposeConv2dAttribute
+ def Init(self, buf, pos):
+ self._tab = flatbuffers.table.Table(buf, pos)
+
+ # TransposeConv2dAttribute
+ def Outpad(self, j):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(4))
+ if o != 0:
+ a = self._tab.Vector(o)
+ return self._tab.Get(flatbuffers.number_types.Int32Flags, a + flatbuffers.number_types.UOffsetTFlags.py_type(j * 4))
+ return 0
+
+ # TransposeConv2dAttribute
+ def OutpadAsNumpy(self):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(4))
+ if o != 0:
+ return self._tab.GetVectorAsNumpy(flatbuffers.number_types.Int32Flags, o)
+ return 0
+
+ # TransposeConv2dAttribute
+ def OutpadLength(self):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(4))
+ if o != 0:
+ return self._tab.VectorLen(o)
+ return 0
+
+ # TransposeConv2dAttribute
+ def Stride(self, j):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(6))
+ if o != 0:
+ a = self._tab.Vector(o)
+ return self._tab.Get(flatbuffers.number_types.Int32Flags, a + flatbuffers.number_types.UOffsetTFlags.py_type(j * 4))
+ return 0
+
+ # TransposeConv2dAttribute
+ def StrideAsNumpy(self):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(6))
+ if o != 0:
+ return self._tab.GetVectorAsNumpy(flatbuffers.number_types.Int32Flags, o)
+ return 0
+
+ # TransposeConv2dAttribute
+ def StrideLength(self):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(6))
+ if o != 0:
+ return self._tab.VectorLen(o)
+ return 0
+
+ # TransposeConv2dAttribute
+ def Dilation(self, j):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(8))
+ if o != 0:
+ a = self._tab.Vector(o)
+ return self._tab.Get(flatbuffers.number_types.Int32Flags, a + flatbuffers.number_types.UOffsetTFlags.py_type(j * 4))
+ return 0
+
+ # TransposeConv2dAttribute
+ def DilationAsNumpy(self):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(8))
+ if o != 0:
+ return self._tab.GetVectorAsNumpy(flatbuffers.number_types.Int32Flags, o)
+ return 0
+
+ # TransposeConv2dAttribute
+ def DilationLength(self):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(8))
+ if o != 0:
+ return self._tab.VectorLen(o)
+ return 0
+
+ # TransposeConv2dAttribute
+ def OutputShape(self, j):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(10))
+ if o != 0:
+ a = self._tab.Vector(o)
+ return self._tab.Get(flatbuffers.number_types.Int32Flags, a + flatbuffers.number_types.UOffsetTFlags.py_type(j * 4))
+ return 0
+
+ # TransposeConv2dAttribute
+ def OutputShapeAsNumpy(self):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(10))
+ if o != 0:
+ return self._tab.GetVectorAsNumpy(flatbuffers.number_types.Int32Flags, o)
+ return 0
+
+ # TransposeConv2dAttribute
+ def OutputShapeLength(self):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(10))
+ if o != 0:
+ return self._tab.VectorLen(o)
+ return 0
+
+def TransposeConv2dAttributeStart(builder): builder.StartObject(4)
+def TransposeConv2dAttributeAddOutpad(builder, outpad): builder.PrependUOffsetTRelativeSlot(0, flatbuffers.number_types.UOffsetTFlags.py_type(outpad), 0)
+def TransposeConv2dAttributeStartOutpadVector(builder, numElems): return builder.StartVector(4, numElems, 4)
+def TransposeConv2dAttributeAddStride(builder, stride): builder.PrependUOffsetTRelativeSlot(1, flatbuffers.number_types.UOffsetTFlags.py_type(stride), 0)
+def TransposeConv2dAttributeStartStrideVector(builder, numElems): return builder.StartVector(4, numElems, 4)
+def TransposeConv2dAttributeAddDilation(builder, dilation): builder.PrependUOffsetTRelativeSlot(2, flatbuffers.number_types.UOffsetTFlags.py_type(dilation), 0)
+def TransposeConv2dAttributeStartDilationVector(builder, numElems): return builder.StartVector(4, numElems, 4)
+def TransposeConv2dAttributeAddOutputShape(builder, outputShape): builder.PrependUOffsetTRelativeSlot(3, flatbuffers.number_types.UOffsetTFlags.py_type(outputShape), 0)
+def TransposeConv2dAttributeStartOutputShapeVector(builder, numElems): return builder.StartVector(4, numElems, 4)
+def TransposeConv2dAttributeEnd(builder): return builder.EndObject()
diff --git a/ethosu/vela/tosa/UnaryQuantInfo.py b/ethosu/vela/tosa/UnaryQuantInfo.py
new file mode 100644
index 0000000..7111c6c
--- /dev/null
+++ b/ethosu/vela/tosa/UnaryQuantInfo.py
@@ -0,0 +1,38 @@
+# automatically generated by the FlatBuffers compiler, do not modify
+
+# namespace: tosa
+
+import flatbuffers
+
+class UnaryQuantInfo(object):
+ __slots__ = ['_tab']
+
+ @classmethod
+ def GetRootAsUnaryQuantInfo(cls, buf, offset):
+ n = flatbuffers.encode.Get(flatbuffers.packer.uoffset, buf, offset)
+ x = UnaryQuantInfo()
+ x.Init(buf, n + offset)
+ return x
+
+ # UnaryQuantInfo
+ def Init(self, buf, pos):
+ self._tab = flatbuffers.table.Table(buf, pos)
+
+ # UnaryQuantInfo
+ def InputZp(self):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(4))
+ if o != 0:
+ return self._tab.Get(flatbuffers.number_types.Int32Flags, o + self._tab.Pos)
+ return 0
+
+ # UnaryQuantInfo
+ def OutputZp(self):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(6))
+ if o != 0:
+ return self._tab.Get(flatbuffers.number_types.Int32Flags, o + self._tab.Pos)
+ return 0
+
+def UnaryQuantInfoStart(builder): builder.StartObject(2)
+def UnaryQuantInfoAddInputZp(builder, inputZp): builder.PrependInt32Slot(0, inputZp, 0)
+def UnaryQuantInfoAddOutputZp(builder, outputZp): builder.PrependInt32Slot(1, outputZp, 0)
+def UnaryQuantInfoEnd(builder): return builder.EndObject()
diff --git a/ethosu/vela/tosa/Version.py b/ethosu/vela/tosa/Version.py
new file mode 100644
index 0000000..403add3
--- /dev/null
+++ b/ethosu/vela/tosa/Version.py
@@ -0,0 +1,54 @@
+# automatically generated by the FlatBuffers compiler, do not modify
+
+# namespace: tosa
+
+import flatbuffers
+
+class Version(object):
+ __slots__ = ['_tab']
+
+ @classmethod
+ def GetRootAsVersion(cls, buf, offset):
+ n = flatbuffers.encode.Get(flatbuffers.packer.uoffset, buf, offset)
+ x = Version()
+ x.Init(buf, n + offset)
+ return x
+
+ # Version
+ def Init(self, buf, pos):
+ self._tab = flatbuffers.table.Table(buf, pos)
+
+ # Version
+ def _major(self):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(4))
+ if o != 0:
+ return self._tab.Get(flatbuffers.number_types.Int32Flags, o + self._tab.Pos)
+ return 0
+
+ # Version
+ def _minor(self):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(6))
+ if o != 0:
+ return self._tab.Get(flatbuffers.number_types.Int32Flags, o + self._tab.Pos)
+ return 22
+
+ # Version
+ def _patch(self):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(8))
+ if o != 0:
+ return self._tab.Get(flatbuffers.number_types.Int32Flags, o + self._tab.Pos)
+ return 0
+
+ # Version
+ def _experimental(self):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(10))
+ if o != 0:
+ return bool(self._tab.Get(flatbuffers.number_types.BoolFlags, o + self._tab.Pos))
+ return False
+
+def VersionStart(builder): builder.StartObject(4)
+def VersionAdd_major(builder, Major): builder.PrependInt32Slot(0, Major, 0)
+def VersionAdd_minor(builder, Minor): builder.PrependInt32Slot(1, Minor, 22)
+def VersionAdd_patch(builder, Patch): builder.PrependInt32Slot(2, Patch, 0)
+def VersionAdd_experimental(builder, Experimental): builder.PrependBoolSlot(3, Experimental, 0)
+def VersionEnd(builder): return builder.EndObject()
diff --git a/ethosu/vela/tosa/WhileLoopAttribute.py b/ethosu/vela/tosa/WhileLoopAttribute.py
new file mode 100644
index 0000000..68655c9
--- /dev/null
+++ b/ethosu/vela/tosa/WhileLoopAttribute.py
@@ -0,0 +1,38 @@
+# automatically generated by the FlatBuffers compiler, do not modify
+
+# namespace: tosa
+
+import flatbuffers
+
+class WhileLoopAttribute(object):
+ __slots__ = ['_tab']
+
+ @classmethod
+ def GetRootAsWhileLoopAttribute(cls, buf, offset):
+ n = flatbuffers.encode.Get(flatbuffers.packer.uoffset, buf, offset)
+ x = WhileLoopAttribute()
+ x.Init(buf, n + offset)
+ return x
+
+ # WhileLoopAttribute
+ def Init(self, buf, pos):
+ self._tab = flatbuffers.table.Table(buf, pos)
+
+ # WhileLoopAttribute
+ def CondBranch(self):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(4))
+ if o != 0:
+ return self._tab.String(o + self._tab.Pos)
+ return None
+
+ # WhileLoopAttribute
+ def BodyBranch(self):
+ o = flatbuffers.number_types.UOffsetTFlags.py_type(self._tab.Offset(6))
+ if o != 0:
+ return self._tab.String(o + self._tab.Pos)
+ return None
+
+def WhileLoopAttributeStart(builder): builder.StartObject(2)
+def WhileLoopAttributeAddCondBranch(builder, condBranch): builder.PrependUOffsetTRelativeSlot(0, flatbuffers.number_types.UOffsetTFlags.py_type(condBranch), 0)
+def WhileLoopAttributeAddBodyBranch(builder, bodyBranch): builder.PrependUOffsetTRelativeSlot(1, flatbuffers.number_types.UOffsetTFlags.py_type(bodyBranch), 0)
+def WhileLoopAttributeEnd(builder): return builder.EndObject()
diff --git a/ethosu/vela/tosa/__init__.py b/ethosu/vela/tosa/__init__.py
new file mode 100644
index 0000000..e69de29
--- /dev/null
+++ b/ethosu/vela/tosa/__init__.py
diff --git a/ethosu/vela/tosa_graph_optimiser.py b/ethosu/vela/tosa_graph_optimiser.py
new file mode 100644
index 0000000..94e6f99
--- /dev/null
+++ b/ethosu/vela/tosa_graph_optimiser.py
@@ -0,0 +1,196 @@
+# Copyright (C) 2021 Arm Limited or its affiliates. All rights reserved.
+#
+# SPDX-License-Identifier: Apache-2.0
+#
+# Licensed under the Apache License, Version 2.0 (the License); you may
+# not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an AS IS BASIS, WITHOUT
+# WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# Description:
+# Early optimisation of the TOSA based network graph, using the rewrite_graph module to do the traversal of the graph.
+from . import rewrite_graph
+from .api import NpuRoundingMode
+from .data_type import DataType
+from .debug_database import DebugDatabase
+from .graph_optimiser_util import needed_total_padding
+from .graph_optimiser_util import set_ifm_ofm_op_shapes
+from .graph_optimiser_util import set_tensor_equivalence
+from .operation import ExplicitScaling
+from .operation import NpuBlockType
+from .operation import Op
+from .operation import Padding
+
+
+def calc_padding_and_skirt(padding_type, kernel, input_shape, explicit_padding):
+ k_w, k_h = kernel.dilated_wh()
+ s_x, s_y = kernel.stride
+ ypad = needed_total_padding(int(input_shape.height), int(s_y), int(k_h))
+ xpad = needed_total_padding(int(input_shape.width), int(s_x), int(k_w))
+ left_pad, right_pad, top_pad, bottom_pad = explicit_padding
+
+ padding = (top_pad, left_pad, bottom_pad, right_pad)
+ skirt = (top_pad, left_pad, ypad - top_pad, xpad - left_pad)
+ return padding, skirt
+
+
+def add_padding_fields(op, arch, nng):
+ if op.run_on_npu:
+ if "padding" in op.attrs:
+ input_shape = op.ifm_shapes[0]
+
+ if op.type == Op.Conv2DBackpropInputSwitchedBias:
+ # TODO not yet supported, but there will be need for separate handling
+ assert False
+ else:
+ padding, skirt = calc_padding_and_skirt(
+ Padding.EXPLICIT, op.kernel, input_shape, op.attrs.get("padding"),
+ )
+
+ op.attrs["explicit_padding"] = padding
+ op.attrs["skirt"] = skirt
+
+ return op
+
+
+def rewrite_activation(op, arch, nng):
+ if not op.type.is_relu_op():
+ return op
+
+ ifm = op.ifm
+ prev_op = ifm.ops[0]
+
+ # Note: the below checks on prev_op require that a first optimize pass on the full graph has been performed
+ fuseable = (
+ prev_op.run_on_npu
+ and prev_op.type.npu_block_type != NpuBlockType.Default
+ and len(ifm.ops) == 1
+ and len(prev_op.outputs[0].consumers()) == 1
+ and prev_op.activation is None
+ )
+ if not fuseable:
+ print("Warning: relu like op will not be possible to fuse, currently not supported")
+ assert False
+
+ zp = ifm.quantization.zero_point if ifm.quantization.zero_point else 0
+ if op.ofm.quantization.zero_point is None:
+ op.ofm.quantization.zero_point = zp
+
+ if op.type == Op.Clip:
+ op.attrs["min"] = op.attrs["min_int"] - zp
+ op.attrs["max"] = op.attrs["max_int"] - zp
+ elif op.type == Op.ReluN:
+ op.attrs["max"] = op.attrs["max_int"] - zp
+ else:
+ print("Warning: Unknown TOSA activation Op")
+ assert False
+
+ return op
+
+
+def rewrite_rescale(op, arch, nng):
+ if op.type == Op.Rescale:
+ ifm = op.ifm
+ ofm = op.ofm
+
+ # some error checking
+ assert len(ifm.ops) == 1
+ prev_op = ifm.ops[0]
+
+ # TODO currently not supported
+ assert prev_op.type not in (Op.Placeholder, Op.SubgraphInput, Op.Const)
+ assert len(ifm.consumer_list) == 1
+
+ input_zp = op.attrs["input_zp"]
+ output_zp = op.attrs["output_zp"]
+ multiplier = op.attrs["multiplier"]
+ shift = op.attrs["shift"]
+ scale32 = op.attrs["scale32"]
+ double_round = op.attrs["double_round"]
+ per_channel = op.attrs["per_channel"]
+
+ assert ifm.dtype in (DataType.uint8, DataType.int8, DataType.int32)
+ assert ifm.dtype in (DataType.uint8, DataType.int8) or input_zp == 0
+ assert ofm.dtype in (DataType.uint8, DataType.int8) or output_zp == 0
+ assert (scale32 and ifm.dtype != DataType.int48) or (not scale32 and not double_round)
+
+ # Check that input tensor has the same zp or no zp
+ ifm_zp = ifm.quantization.zero_point
+ if ifm_zp is not None and ifm_zp != input_zp:
+ print("Error (fuse_rescale): zp of tensors producer/consumer differs unexpectedidly ")
+ assert False
+ ifm.quantization.zero_point = input_zp
+
+ if not scale32:
+ double_round = False
+
+ if prev_op.type.is_depthwise_conv2d_op() or prev_op.type.is_conv2d_op() or prev_op.type == Op.FullyConnected:
+ assert len(multiplier) == len(shift) == len(prev_op.bias.values)
+
+ if ifm.dtype == DataType.int32 and per_channel:
+ for s, m in zip(shift, multiplier):
+ # TODO these are the TOSA limitations
+ assert m >= 0
+ assert 2 <= s <= 62
+ # TODO these are the HW limitations
+ assert 0 <= s < (1 << 6)
+ prev_op.explicit_scaling = ExplicitScaling(per_channel, shift, multiplier)
+ ofm.quantization.zero_point = output_zp
+
+ if double_round:
+ prev_op.rounding_mode = NpuRoundingMode.TFL
+ else:
+ prev_op.rounding_mode = NpuRoundingMode.NATURAL
+
+ # Bypass op
+ prev_op.set_output_tensor(ofm)
+ DebugDatabase.add_optimised(op, prev_op)
+ return op
+ else:
+ print("Warning, unsupported fusing of TOSA Rescale previous operator is of type:", prev_op.type)
+ assert False
+
+ else:
+ print("Warning, unsupported fusing of TOSA Rescale previous operator is of type:", prev_op.type)
+ assert False
+ return op
+
+
+def supported_operator_check(op, arch, nng):
+ op.run_on_npu = arch.tosa_supported_operators.is_operator_supported(op)
+ return op
+
+
+def tosa_optimise_graph(nng, arch):
+ # Pre-processing step
+ pre_process_list = [
+ supported_operator_check,
+ set_ifm_ofm_op_shapes,
+ ]
+
+ for idx, sg in enumerate(nng.subgraphs):
+ nng.subgraphs[idx] = rewrite_graph.rewrite_graph_pre_order(
+ nng, sg, arch, [], pre_process_list, rewrite_unsupported=False,
+ )
+
+ # Rewite Operators step
+ op_rewrite_list = [set_tensor_equivalence, rewrite_rescale]
+
+ for idx, sg in enumerate(nng.subgraphs):
+ nng.subgraphs[idx] = rewrite_graph.rewrite_graph_pre_order(
+ nng, sg, arch, [], op_rewrite_list, rewrite_unsupported=False,
+ )
+
+ # Post-processing step
+ for idx, sg in enumerate(nng.subgraphs):
+ nng.subgraphs[idx] = rewrite_graph.rewrite_graph_pre_order(
+ nng, sg, arch, [], [rewrite_activation, add_padding_fields],
+ )
+
+ return nng
diff --git a/ethosu/vela/tosa_mapping.py b/ethosu/vela/tosa_mapping.py
new file mode 100644
index 0000000..82f61f7
--- /dev/null
+++ b/ethosu/vela/tosa_mapping.py
@@ -0,0 +1,325 @@
+# Copyright (C) 2021 Arm Limited or its affiliates. All rights reserved.
+#
+# SPDX-License-Identifier: Apache-2.0
+#
+# Licensed under the Apache License, Version 2.0 (the License); you may
+# not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an AS IS BASIS, WITHOUT
+# WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# Description:
+# TOSA mapping functions used by reader.
+# Contains a mapping from the various TOSA enums and options structs, generated by the FlatBuffer code
+# generator, to Vela's internal format.
+from .data_type import DataType
+from .operation import Op
+from .operation import TensorIndices
+from .tosa import ArithmeticRightShiftAttribute # noqa: F401
+from .tosa import AxisAttribute # noqa: F401
+from .tosa import ClampAttribute # noqa: F401
+from .tosa import CondIfAttribute # noqa: F401
+from .tosa import Conv2dAttribute # noqa: F401
+from .tosa import ConvQuantInfo # noqa: F401
+from .tosa import MatMulQuantInfo # noqa: F401
+from .tosa import MulAttribute # noqa: F401
+from .tosa import PadQuantInfo # noqa: F401
+from .tosa import Pool2dAttribute # noqa: F401
+from .tosa import ReluNAttribute # noqa: F401
+from .tosa import RescaleAttribute # noqa: F401
+from .tosa import ReshapeAttribute # noqa: F401
+from .tosa import ResizeAttribute # noqa: F401
+from .tosa import SliceAttribute # noqa: F401
+from .tosa import TileAttribute # noqa: F401
+from .tosa import TransposeConv2dAttribute # noqa: F401
+from .tosa import UnaryQuantInfo # noqa: F401
+from .tosa import WhileLoopAttribute # noqa: F401
+from .tosa.DType import DType
+from .tosa.Op import Op as TosaOp
+
+
+datatype_map = {
+ DType.BOOL: DataType.bool,
+ DType.UINT8: DataType.uint8,
+ DType.INT4: DataType.int4,
+ DType.INT8: DataType.int8,
+ DType.INT16: DataType.int16,
+ DType.INT32: DataType.int32,
+ DType.INT48: DataType.int48,
+ DType.FLOAT: DataType.float32,
+}
+
+
+# TODO duplicate of tflite_mapping
+def underscore_to_camel_case(s):
+ return "".join(x.title() for x in s.split("_"))
+
+
+# TODO duplicate of tflite_mapping
+def identity(x):
+ return x
+
+
+class AttrSerializer:
+ def __init__(self, name, members=None):
+ self.name = name
+ self.module = globals()[self.name]
+ self.cls = getattr(self.module, self.name)
+ self.members = []
+ if members is not None:
+ for mem in members:
+ deserialize = identity
+ is_vector = False
+ if isinstance(mem, tuple):
+ if len(mem) == 2:
+ mem, is_vector = mem
+ deserialize = tuple
+ else:
+ assert 0
+ underscore_mem = mem
+ camelcase_mem = underscore_to_camel_case(mem)
+ self.members.append((underscore_mem, camelcase_mem, deserialize, is_vector))
+
+ def deserialize(self, op_data):
+ attr_type = op_data.AttributeType()
+ attr = op_data.Attribute()
+ attrs = {}
+ if attr_type:
+ tosa_attrs = self.cls()
+ tosa_attrs.Init(attr.Bytes, attr.Pos)
+ for underscore_mem, camelcase_mem, deserialize, is_vector in self.members:
+ fun = camelcase_mem
+ if is_vector:
+ fun += "AsNumpy"
+
+ attr = getattr(tosa_attrs, fun)()
+ try:
+ attrs[underscore_mem] = deserialize(attr)
+ except TypeError:
+ print("Warning: {0} could not read attribute '{1}'.".format(self.name, underscore_mem))
+
+ return attrs
+
+
+class QuantSerializer:
+ def __init__(self, name, members=None):
+ self.name = name
+ self.module = globals()[self.name]
+ self.cls = getattr(self.module, self.name)
+ self.members = []
+ if members is not None:
+ for mem in members:
+ deserialize = identity
+ underscore_mem = mem
+ camelcase_mem = underscore_to_camel_case(mem)
+ self.members.append((underscore_mem, camelcase_mem, deserialize))
+
+ def deserialize(self, op_data):
+ quant_info_type = op_data.QuantInfoType()
+ quant_info = op_data.QuantInfo()
+ quant = {}
+ if quant_info_type:
+ tosa_quant = self.cls()
+ tosa_quant.Init(quant_info.Bytes, quant_info.Pos)
+ for underscore_mem, camelcase_mem, deserialize in self.members:
+ attr = getattr(tosa_quant, camelcase_mem)()
+ try:
+ quant[underscore_mem] = deserialize(attr)
+ except TypeError:
+ print("Warning: {0} could not read quant info '{1}'.".format(self.name, underscore_mem))
+
+ return quant
+
+
+is_vec = True
+pool2d_attrs = AttrSerializer("Pool2dAttribute", (("padding", is_vec), ("kernel", is_vec), ("stride", is_vec)))
+conv2d_attrs = AttrSerializer("Conv2dAttribute", (("padding", is_vec), ("stride", is_vec), ("dilation", is_vec)))
+transpose_conv2d_attrs = AttrSerializer(
+ "TransposeConv2dAttribute", (("outpad", is_vec), ("stride", is_vec), ("dilation", is_vec), ("out_shape", is_vec))
+)
+relun_attrs = AttrSerializer("ReluNAttribute", ("max_int"))
+axis_attrs = AttrSerializer("AxisAttribute", ("axis"))
+reshape_attrs = AttrSerializer("ReshapeAttribute", (("shape", is_vec),))
+slice_attrs = AttrSerializer("SliceAttribute", (("begin", is_vec), ("size", is_vec)))
+tile_attrs = AttrSerializer("TileAttribute", (("multiplies", is_vec),))
+resize_attrs = AttrSerializer(
+ "ResizeAttribute", (("output_size", is_vec), ("stride", is_vec), ("offset", is_vec), ("shift"))
+)
+clamp_attrs = AttrSerializer("ClampAttribute", (("min_int"), ("max_int")))
+rescale_attrs = AttrSerializer(
+ "RescaleAttribute",
+ ("input_zp", "output_zp", ("multiplier", is_vec), ("shift", is_vec), "scale32", "double_round", "per_channel"),
+)
+mul_attrs = AttrSerializer("MulAttribute", ("shift"))
+ars_attrs = AttrSerializer("ArithmeticRightShiftAttribute", ("round",))
+condif_attrs = AttrSerializer("CondIfAttribute", (("then_branch"), ("else_branch"))) # TODO these are references
+while_attrs = AttrSerializer("WhileLoopAttribute", (("cond_branch"), ("body_branch"))) # TODO these are references
+
+unary_quant_info = QuantSerializer("UnaryQuantInfo", ("input_zp", "output_zp"))
+conv_quant_info = QuantSerializer("ConvQuantInfo", ("input_zp", "weight_zp"))
+matmul_quant_info = QuantSerializer("MatMulQuantInfo", ("a_zp", "b_zp"))
+pad_quant_info = QuantSerializer("PadQuantInfo", ("input_zp"))
+
+unsupported_tosa_operators = {
+ TosaOp.UNKNOWN,
+ TosaOp.ARGMAX,
+ TosaOp.CONV3D,
+ TosaOp.MATMUL,
+ TosaOp.TRANSPOSE_CONV2D,
+ TosaOp.SIGMOID,
+ TosaOp.TANH,
+ TosaOp.BITWISE_AND,
+ TosaOp.BITWISE_OR,
+ TosaOp.BITWISE_XOR,
+ TosaOp.DIV,
+ TosaOp.LOGICAL_AND,
+ TosaOp.LOGICAL_LEFT_SHIFT,
+ TosaOp.LOGICAL_RIGHT_SHIFT,
+ TosaOp.LOGICAL_OR,
+ TosaOp.LOGICAL_XOR,
+ TosaOp.MAXIMUM,
+ TosaOp.MINIMUM,
+ TosaOp.MUL,
+ TosaOp.POW,
+ TosaOp.TABLE,
+ TosaOp.ABS,
+ TosaOp.BITWISE_NOT,
+ TosaOp.CEIL,
+ TosaOp.CLZ,
+ TosaOp.EXP,
+ TosaOp.FLOOR,
+ TosaOp.LOG,
+ TosaOp.LOGICAL_NOT,
+ TosaOp.NEGATE,
+ TosaOp.RECIPROCAL,
+ TosaOp.RSQRT,
+ TosaOp.SELECT,
+ TosaOp.EQUAL,
+ TosaOp.GREATER,
+ TosaOp.GREATER_EQUAL,
+ TosaOp.REDUCE_ANY,
+ TosaOp.REDUCE_ALL,
+ TosaOp.REDUCE_MAX,
+ TosaOp.REDUCE_MIN,
+ TosaOp.REDUCE_PRODUCT,
+ TosaOp.REDUCE_SUM,
+ TosaOp.CONCAT,
+ TosaOp.PAD,
+ TosaOp.RESHAPE,
+ TosaOp.REVERSE,
+ TosaOp.SLICE,
+ TosaOp.TILE,
+ TosaOp.TRANSPOSE,
+ TosaOp.GATHER,
+ TosaOp.SCATTER,
+ TosaOp.RESIZE,
+ TosaOp.CAST,
+ TosaOp.IDENTITY,
+ TosaOp.CUSTOM,
+ TosaOp.COND_IF,
+ TosaOp.WHILE_LOOP,
+}
+
+
+TOSA_NO_INDICES = TensorIndices([], [], [])
+TOSA_IFM_INDICES = TensorIndices([0], [], [])
+# TOSA_IFM_WEIGHTS_INDICES = TensorIndices([0], [1], [])
+TOSA_IFM_WEIGHTS_BIAS_INDICES = TensorIndices([0], [1], [2])
+TOSA_IFM_IFM2_INDICES = TensorIndices([0, 1], [], [])
+# TOSA_CONV2D_BACKPROP_INDICES = TensorIndices([2], [1], [3])
+# TOSA_TRANSPOSE_CONV_INDICES = TensorIndices([0], [1], [3])
+# TOSA_CONCAT_INDICES = TensorIndices([1, 2], [], [])
+# TOSA_SPLIT_IFM_INDICES = TensorIndices([1], [], [])
+# TOSA_BLOCK_LSTM_INDICES = TensorIndices([3], [4], [])
+
+
+tosa_operator_map = {
+ # TosaOp.UNKNOWN: (),
+ # TODO TosaOp.ARGMAX: (Op.ArgMax, axis_attrs, None),
+ TosaOp.AVG_POOL2D: (Op.AvgPool, pool2d_attrs, unary_quant_info, TOSA_IFM_INDICES),
+ TosaOp.CONV2D: (Op.Conv2DBias, conv2d_attrs, conv_quant_info, TOSA_IFM_WEIGHTS_BIAS_INDICES),
+ # TODO TosaOp.CONV3D:
+ TosaOp.DEPTHWISE_CONV2D: (Op.DepthwiseConv2DBias, conv2d_attrs, conv_quant_info, TOSA_IFM_WEIGHTS_BIAS_INDICES),
+ TosaOp.FULLY_CONNECTED: (Op.FullyConnected, None, conv_quant_info, TOSA_IFM_WEIGHTS_BIAS_INDICES),
+ # TODO TosaOp.MATMUL:
+ TosaOp.MAX_POOL2D: (Op.MaxPool, pool2d_attrs, None, TOSA_IFM_INDICES),
+ # TODO TosaOp.TRANSPOSE_CONV2D: (Op.Conv2DBackpropInput, transpose_conv2d_attrs, conv_quant_info)
+ TosaOp.CLAMP: (Op.Clip, clamp_attrs, None, TOSA_IFM_INDICES),
+ TosaOp.RELUN: (Op.ReluN, relun_attrs, None, TOSA_IFM_INDICES),
+ # TODO TosaOp.SIGMOID
+ # TODO TosaOp.TANH
+ TosaOp.ADD: (Op.Add, None, None, TOSA_IFM_IFM2_INDICES),
+ TosaOp.ARITHMETIC_RIGHT_SHIFT: (Op.SHR, ars_attrs, None, TOSA_IFM_IFM2_INDICES),
+ # TODO TosaOp.BITWISE_AND
+ # TODO TosaOp.BITWISE_OR
+ # TODO TosaOp.BITWISE_XOR
+ # TODO TosaOp.DIV
+ # TODO TosaOp.LOGICAL_AND
+ # TODO TosaOp.LOGICAL_LEFT_SHIFT
+ # TODO TosaOp.LOGICAL_RIGHT_SHIFT
+ # TODO TosaOp.LOGICAL_OR
+ # TODO TosaOp.LOGICAL_XOR
+ # TODO TosaOp.MAXIMUM
+ # TODO TosaOp.MINIMUM
+ # TODO TosaOp.MUL
+ # TODO TosaOp.POW
+ TosaOp.SUB: (Op.Sub, None, None, TOSA_IFM_IFM2_INDICES),
+ # TODO TosaOp.TABLE
+ # TODO TosaOp.ABS
+ # TODO TosaOp.BITWISE_NOT
+ # TODO TosaOp.CEIL
+ # TODO TosaOp.CLZ
+ # TODO TosaOp.EXP
+ # TODO TosaOp.FLOOR
+ # TODO TosaOp.LOG
+ # TODO TosaOp.LOGICAL_NOT
+ # TODO TosaOp.NEGATE
+ # TODO TosaOp.RECIPROCAL
+ # TODO TosaOp.RSQRT
+ # TODO TosaOp.SELECT
+ # TODO TosaOp.EQUAL
+ # TODO TosaOp.GREATER
+ # TODO TosaOp.GREATER_EQUAL
+ # TODO TosaOp.REDUCE_ANY
+ # TODO TosaOp.REDUCE_ALL
+ # TODO TosaOp.REDUCE_MAX
+ # TODO TosaOp.REDUCE_MIN
+ # TODO TosaOp.REDUCE_PRODUCT
+ # TODO TosaOp.REDUCE_SUM
+ # TODO TosaOp.CONCAT
+ # TODO TosaOp.PAD
+ # TODO TosaOp.RESHAPE
+ # TODO TosaOp.REVERSE
+ # TODO TosaOp.SLICE
+ # TODO TosaOp.TILE
+ # TODO TosaOp.TRANSPOSE
+ # TODO TosaOp.GATHER
+ # TODO TosaOp.SCATTER
+ # TODO TosaOp.RESIZE
+ # TODO TosaOp.CAST
+ TosaOp.RESCALE: (Op.Rescale, rescale_attrs, None, TOSA_IFM_INDICES),
+ TosaOp.CONST: (Op.Const, None, None, TOSA_NO_INDICES),
+ # TODO TosaOp.IDENTITY
+ # TODO TosaOp.CUSTOM
+ # TODO TosaOp.COND_IF
+ # TODO TosaOp.WHILE_LOOP
+}
+
+tosa_operator_inv_map = {v[0]: (k, v[1]) for k, v in tosa_operator_map.items()}
+
+
+def tosa_type_name(builtin):
+ return next(k for k, v in vars(TosaOp).items() if v == builtin)
+
+
+# TODO will return UNKNOWN for the once that have not yet been defined in tosa_operator_map
+def optype_to_tosa_op_type(op_type):
+ if op_type in tosa_operator_inv_map:
+ return tosa_type_name(tosa_operator_inv_map[op_type][0])
+ else:
+ return TosaOp.UNKNOWN
diff --git a/ethosu/vela/tosa_reader.py b/ethosu/vela/tosa_reader.py
new file mode 100644
index 0000000..ac0b396
--- /dev/null
+++ b/ethosu/vela/tosa_reader.py
@@ -0,0 +1,259 @@
+# Copyright (C) 2021 Arm Limited or its affiliates. All rights reserved.
+#
+# SPDX-License-Identifier: Apache-2.0
+#
+# Licensed under the Apache License, Version 2.0 (the License); you may
+# not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an AS IS BASIS, WITHOUT
+# WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# Description:
+# Functions used to read from a TOSA format file.
+import os.path
+import struct
+import sys
+
+import numpy as np
+
+from .nn_graph import Graph
+from .nn_graph import Subgraph
+from .operation import Op
+from .operation import Operation
+from .reader_util import clone_and_reshape_tensor
+from .reader_util import decode_str
+from .reader_util import fixup_tensors
+from .tensor import QuantizationParameters
+from .tensor import Tensor
+from .tflite_mapping import DataType
+from .tosa.TosaGraph import TosaGraph as TG
+from .tosa_mapping import datatype_map
+from .tosa_mapping import tosa_operator_map
+from .tosa_mapping import unsupported_tosa_operators
+
+
+class TosaSubgraph:
+ def __init__(self, file_path, graph, block):
+ self.graph = graph
+ self.name = decode_str(block.Name())
+
+ self.tensors = []
+ for idx in range(block.TensorsLength()):
+ self.tensors.append(self.parse_tensor(block.Tensors(idx), file_path))
+
+ for idx in range(block.OperatorsLength()):
+ self.parse_operator(idx, block.Operators(idx))
+
+ # Get the subgraph inputs and outputs
+ self.inputs = self.get_sg_inputs_remove_duplicates(block)
+ self.outputs = self.get_sg_outputs_remove_duplicates(block)
+ fixup_tensors(self.inputs, self.tensors)
+
+ def get_sg_inputs_remove_duplicates(self, block):
+ inputs = []
+ for idx in range(block.InputsLength()):
+ tens_data = block.Inputs(idx)
+ self.add_not_duplicate(tens_data, inputs, "input")
+ return inputs
+
+ def get_sg_outputs_remove_duplicates(self, block):
+ outputs = []
+ for idx in range(block.OutputsLength()):
+ tens_data = block.Outputs(idx)
+ self.add_not_duplicate(tens_data, outputs, "output")
+ return outputs
+
+ def add_not_duplicate(self, tens_data, tensors, warning_str):
+ name = decode_str(tens_data)
+ tensor = self.get_tensor_by_name(name)
+ if tensor not in tensors:
+ tensors.append(tensor)
+ else:
+ print(f"Warning: Subgraph {warning_str} tensor ({tensor}) already seen. Removing the duplicate.")
+
+ def get_tensor_by_name(self, name):
+ for tens in self.tensors:
+ if tens.name == name:
+ return tens
+ return None
+
+ def parse_operator(self, op_index, op_data):
+ op_code = op_data.Op()
+ if op_code in unsupported_tosa_operators:
+ print("Unsupported Operator", op_code)
+ assert False
+
+ op_type, attr_serializer, quant_serializer, indices = tosa_operator_map[op_code]
+ inputs = []
+ outputs = []
+ for idx in range(op_data.InputsLength()):
+ input_tens = self.get_tensor_by_name(decode_str(op_data.Inputs(idx)))
+ inputs.append(input_tens)
+ assert input_tens is not None
+
+ for idx in range(op_data.OutputsLength()):
+ output_tens = self.get_tensor_by_name(decode_str(op_data.Outputs(idx)))
+ outputs.append(output_tens)
+ assert output_tens is not None
+
+ name = "unknown_op_name"
+ if len(outputs):
+ name = outputs[0].name
+ op = Operation(op_type, name)
+ op.type.info.indices = indices
+ op.op_index = op_index
+ op.inputs = inputs
+ op.outputs = outputs
+
+ for out in op.outputs:
+ out.ops = [op]
+
+ # TODO Transpose_conv and conv3d
+ if op.type.is_depthwise_conv2d_op() or op.type.is_conv2d_op() or op.type == Op.FullyConnected:
+ if inputs[1].values is not None:
+ if op.type == Op.FullyConnected:
+ inputs[1] = clone_and_reshape_tensor(inputs[1], (1, 0), False)
+ elif op.type.is_conv2d_op():
+ inputs[1] = clone_and_reshape_tensor(inputs[1], (1, 2, 3, 0), False)
+ elif op.type.is_depthwise_conv2d_op():
+ inputs[1] = clone_and_reshape_tensor(inputs[1], (1, 2, 0, 3), False)
+ if op.type.needs_bias() and len(inputs) <= op_type.info.indices.biases[0]:
+ # No Bias tensor
+ inputs.append(None)
+ if inputs[-1] and inputs[-1].values is not None:
+ # Since bias tensor is used for both bias and scale,
+ # a clone with a unique equivalence_id is needed
+ inputs[-1] = clone_and_reshape_tensor(inputs[-1], (0,), True)
+
+ if attr_serializer is not None:
+ op.attrs = attr_serializer.deserialize(op_data)
+
+ if "dilation" in op.attrs:
+ dilation = op.attrs["dilation"]
+ if len(dilation) == 2:
+ op.attrs["dilation"] = (1, dilation[0], dilation[1], 1)
+ elif len(dilation) == 3:
+ # TODO CONV3D more to be done....
+ op.attrs["dilation"] = (dilation[0], dilation[1], dilation[2], 1)
+ if "kernel" in op.attrs:
+ kernel = op.attrs["kernel"]
+ if len(kernel) == 2:
+ op.attrs["ksize"] = (1, kernel[0], kernel[1], 1)
+ else:
+ # TODO CONV3D more to be done....
+ print("Unsupported kernel dimensions: ", len(kernel))
+ assert False
+
+ if quant_serializer is not None:
+ quant_info = quant_serializer.deserialize(op_data)
+
+ # TODO tensor zero points currently set here
+ # zero points part of Rescale operation, handled in tosa_graph_optimizer
+ if "input_zp" in quant_info:
+ self.set_tensor_zp(op.ifm, quant_info["input_zp"])
+ if "weight_zp" in quant_info:
+ self.set_tensor_zp(op.weights, quant_info["weight_zp"])
+ if "ouput_zp" in quant_info:
+ self.set_tensor_zp(op.ofm, quant_info["output_zp"])
+ if "a_zp" in quant_info:
+ self.set_tensor_zp(op.ifm, quant_info["a_zp"])
+ if "b_zp" in quant_info:
+ self.set_tensor_zp(op.ifm2, quant_info["b_zp"])
+
+ def parse_tensor(self, tens_data, file_path):
+ name = decode_str(tens_data.Name())
+ np_shape = tens_data.ShapeAsNumpy()
+ shape = list(np_shape) if type(np_shape) is np.ndarray else []
+ tens_dtype = tens_data.Type()
+ dtype = datatype_map[tens_dtype]
+
+ tens = Tensor(shape, dtype, name)
+
+ # Initialize quantization parameters
+ tens.quantization = QuantizationParameters()
+
+ tens.quantization.scale_f32 = 1.0
+ if dtype == DataType.uint8:
+ tens.quantization.quant_min = 0
+ tens.quantization.quant_max = (1 << dtype.bits) - 1
+ elif dtype in (DataType.int8, DataType.int16, DataType.int32, DataType.int64):
+ tens.quantization.quant_min = -(1 << (dtype.bits - 1))
+ tens.quantization.quant_max = (1 << (dtype.bits - 1)) - 1
+
+ tens.values = None
+ if tens_data.NpyFilename() is not None:
+ try:
+ 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)
+
+ return tens
+
+ def set_tensor_zp(self, tens, zp):
+ if tens.quantization.zero_point is None:
+ tens.quantization.zero_point = zp
+ elif tens.quantization.zero_point != zp:
+ print(f"Error: Setting tensor zp not possible, tensor already has different zero point")
+ assert False
+
+
+class TosaGraph:
+ def __init__(self, filename, batch_size, feed_dict, output_node_names, initialisation_nodes):
+
+ self.op_times = {}
+ if batch_size is None:
+ batch_size = 1
+ self.batch_size = batch_size
+ self.name = os.path.splitext(os.path.basename(filename))[0]
+ self.initialisation_nodes = initialisation_nodes
+
+ with open(filename, "rb") as f:
+ buf = bytearray(f.read())
+
+ try:
+ parsing_step = "parsing root"
+ tosa_graph = TG.GetRootAsTosaGraph(buf, 0)
+
+ parsing_step = "parsing version"
+ self.check_version(tosa_graph)
+
+ parsing_step = "parsing blocks length"
+ file_path = os.path.dirname(filename)
+ self.subgraphs = []
+ for b_idx in range(tosa_graph.BlocksLength()):
+ parsing_step = f"parsing block {b_idx}"
+ self.subgraphs.append(TosaSubgraph(file_path, self, tosa_graph.Blocks(b_idx)))
+
+ self.nng = Graph(self.name, self.batch_size)
+ for tosa_sg in self.subgraphs:
+ sg = Subgraph(tosa_sg.name)
+ sg.original_inputs = tosa_sg.inputs # Preserve the original input order
+ sg.output_tensors = tosa_sg.outputs
+ self.nng.subgraphs.append(sg)
+
+ except (struct.error, TypeError, RuntimeError) as e:
+ print(f'Error: Invalid .tosa file. Got "{e}" while {parsing_step}.')
+ sys.exit(1)
+
+ def check_version(self, tosa_graph):
+ version = tosa_graph.Version()
+ version_str = f"{version._major()}.{version._minor()}.{version._patch()}"
+ if version_str != "0.22.0":
+ print(f"Unsupported TOSA version: {version_str}")
+ assert False
+
+
+def read_tosa(filename, batch_size, feed_dict, output_node_names, initialisation_nodes):
+ tosa_graph = TosaGraph(filename, batch_size, feed_dict, output_node_names, initialisation_nodes)
+ nng = tosa_graph.nng
+ nng.refresh_after_modification()
+ return nng
diff --git a/ethosu/vela/tosa_supported_operators.py b/ethosu/vela/tosa_supported_operators.py
new file mode 100644
index 0000000..c87d653
--- /dev/null
+++ b/ethosu/vela/tosa_supported_operators.py
@@ -0,0 +1,85 @@
+# Copyright (C) 2021 Arm Limited or its affiliates. All rights reserved.
+#
+# SPDX-License-Identifier: Apache-2.0
+#
+# Licensed under the Apache License, Version 2.0 (the License); you may
+# not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an AS IS BASIS, WITHOUT
+# WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# Description:
+# The TosaSupportedOperators class which is a collection of all supported operators and parameter checks.
+from collections import defaultdict
+
+from .data_type import DataType
+from .operation import Op
+from .supported_operators_util import docstring_format_args
+from .supported_operators_util import list_formatter
+from .tosa_mapping import optype_to_tosa_op_type
+
+
+class TosaSupportedOperators:
+ # TODO currently sparsely populated
+ # Categorised lists of supported operators
+ convolution_ops = set((Op.Conv2DBias,))
+ convolution_like_ops = convolution_ops
+ mac_main_ops = convolution_like_ops
+
+ type_conversion_ops = set((Op.Rescale,))
+ relu_ops = set((Op.Clip, Op.ReluN,))
+ activation_ops = relu_ops
+
+ npu_post_ops = activation_ops
+ supported_operators = mac_main_ops | type_conversion_ops | npu_post_ops
+
+ # Supported data types
+ # TODO will differ compared to TensorFlow Lite, currently set to the same
+ supported_op_dtypes = set((DataType.uint8, DataType.int8, DataType.int16, DataType.int32))
+
+ def __init__(self):
+ # Setup the generic constraints. Note: the order matters
+ self.generic_constraints = []
+ self.generic_constraints.append(TosaSupportedOperators.constraint_tens_dtype)
+
+ # Setup specific constraints. Note: the order matters
+ self.specific_constraints = defaultdict(list)
+
+ def is_operator_supported(self, op):
+ ext_type = optype_to_tosa_op_type(op.type)
+ if op.type not in TosaSupportedOperators.supported_operators:
+ if op.type not in (Op.Placeholder, Op.SubgraphInput, Op.Const):
+ print(f"Info: {ext_type} '{op.name}' is not a NPU op")
+ return False
+
+ for constraint in self.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")
+ print(f" - {constraint.__doc__}")
+ if extra:
+ print(f" {extra}")
+ return False
+
+ return True
+
+ # TODO this function is the same for TensorFlow Lite, but input might differ
+ @classmethod
+ @docstring_format_args([list_formatter(supported_op_dtypes)])
+ def constraint_tens_dtype(cls, op):
+ "Tensors must be of type: {}"
+ valid = True
+ extra = []
+ tensors = [tens for tens in op.get_ifm_ifm2_weights_ofm() if tens]
+ if not tensors:
+ tensors = [tens for tens in op.inputs if tens]
+ for tens in tensors:
+ if tens.dtype not in cls.supported_op_dtypes:
+ valid = False
+ extra.append(f"Tensor '{tens.name}' has data type: {tens.dtype}")
+ return valid, ", ".join(extra)
diff --git a/ethosu/vela/vela.py b/ethosu/vela/vela.py
index f552b21..ecdc7aa 100644
--- a/ethosu/vela/vela.py
+++ b/ethosu/vela/vela.py
@@ -54,7 +54,7 @@
output_basename = os.path.join(compiler_options.output_dir, os.path.splitext(os.path.basename(input_name))[0])
DebugDatabase.show_warnings = enable_debug_db
- nng = model_reader.read_model(input_name, model_reader_options)
+ nng, network_type = model_reader.read_model(input_name, model_reader_options)
if not nng:
raise InputFileError(input_name, "Input file could not be read")
@@ -67,7 +67,7 @@
print("Model reading took %f s" % (stop - start))
start = time.time()
- compiler_driver.compiler_driver(nng, arch, compiler_options, scheduler_options)
+ compiler_driver.compiler_driver(nng, arch, compiler_options, scheduler_options, network_type)
summary_csv_file = "{0}_summary_{1}.csv".format(output_basename, arch.system_config)
stats_writer.write_summary_metrics_csv(nng, summary_csv_file, arch)
diff --git a/ethosu/vela/weight_compressor.py b/ethosu/vela/weight_compressor.py
index 4ba3dee..7e33e93 100644
--- a/ethosu/vela/weight_compressor.py
+++ b/ethosu/vela/weight_compressor.py
@@ -203,7 +203,7 @@
return ohwi[core : ohwi.shape[0] : ncores]
-def _prepare_scale_and_bias(arch, tens, rescale_for_faf):
+def _prepare_scale_and_bias(arch, tens, rescale_for_faf, explicit_scaling):
assert tens.purpose in [TensorPurpose.FeatureMap, TensorPurpose.FSBias]
assert tens.format == TensorFormat.NHWC
# the connected operator should expect a bias input unless it is a FullyConnected
@@ -260,11 +260,15 @@
else:
raise UnsupportedFeatureError(f"Compression of {ifm_dtype} is not implemented; Tensor: '{tens.name}'")
- # quantise all of the weight scales into (scale_factor, shift)
- if ifm_dtype == DataType.int16:
- quantised_scales = [reduced_quantise_scale(scale) for scale in scales]
+ if explicit_scaling:
+ assert len(explicit_scaling.shift) == len(explicit_scaling.multiplier)
+ quantised_scales = [(int(m), int(s)) for s, m in zip(explicit_scaling.shift, explicit_scaling.multiplier)]
else:
- quantised_scales = [quantise_scale(scale) for scale in scales]
+ # quantise all of the weight scales into (scale_factor, shift)
+ if ifm_dtype == DataType.int16:
+ quantised_scales = [reduced_quantise_scale(scale) for scale in scales]
+ else:
+ quantised_scales = [quantise_scale(scale) for scale in scales]
# If only 1 quantised scale is used, repeat that value for the length of the biases
if len(quantised_scales) == 1:
@@ -355,7 +359,7 @@
# Bias & scale
if do_scales:
- quantised_scales, biases = _prepare_scale_and_bias(arch, scale_tens, rescale_for_faf)
+ quantised_scales, biases = _prepare_scale_and_bias(arch, scale_tens, rescale_for_faf, op.explicit_scaling)
scale_tens.element_size_bytes = 10
# Slice the weight stream up depth-ways into bricks and compress