diff --git a/docker/quicktest.sh b/docker/quicktest.sh
index 02e014cd3cc7bb88eebd02f03ff599913079152b..b06feccdc578a59c8ef00531871e1211c2a407e5 100755
--- a/docker/quicktest.sh
+++ b/docker/quicktest.sh
@@ -16,6 +16,11 @@ elif [ $1 = "rtlsim" ]; then
elif [ $1 = "end2end" ]; then
echo "Running end2end test suite with no parallelism"
python setup.py test --addopts "-k end2end"
+elif [ $1 = "full" ]; then
+ echo "Running full test suite, each step with appropriate parallelism"
+ $0 main;
+ $0 rtlsim;
+ $0 end2end;
else
echo "Unrecognized argument to quicktest.sh"
fi
diff --git a/src/finn/core/modelwrapper.py b/src/finn/core/modelwrapper.py
index 646add188c5d475cf37ccd33cf24d29d61754ae1..98b234592ebe0c704fafd1eed980325d8566e7e2 100644
--- a/src/finn/core/modelwrapper.py
+++ b/src/finn/core/modelwrapper.py
@@ -36,6 +36,11 @@ from onnx import TensorProto
import finn.util.basic as util
import finn.util.onnx as onnxutil
from finn.core.datatype import DataType
+from finn.transformation.general import (
+ RemoveUnusedTensors,
+ RemoveStaticGraphInputs,
+ SortGraph,
+)
class ModelWrapper:
@@ -87,7 +92,7 @@ class ModelWrapper:
"""Runs given anaylsis_fxn on this model and return resulting dict."""
return analysis_fxn(self)
- def transform(self, transformation, make_deepcopy=True):
+ def transform(self, transformation, make_deepcopy=True, cleanup=True):
"""Applies given Transformation repeatedly until no more changes can be made
and returns a transformed ModelWrapper instance.
@@ -101,6 +106,22 @@ class ModelWrapper:
(transformed_model, model_was_changed) = transformation.apply(
transformed_model
)
+ if cleanup:
+ transformed_model.cleanup()
+ return transformed_model
+
+ def cleanup(self):
+ "Run cleanup transformations on the model."
+ transformed_model = self
+ cleanup_transforms = [
+ RemoveUnusedTensors(),
+ RemoveStaticGraphInputs(),
+ SortGraph(),
+ ]
+ for trn in cleanup_transforms:
+ transformed_model = transformed_model.transform(
+ trn, cleanup=False, make_deepcopy=False
+ )
return transformed_model
def check_compatibility(self):
diff --git a/src/finn/custom_op/fpgadataflow/streamingfclayer_batch.py b/src/finn/custom_op/fpgadataflow/streamingfclayer_batch.py
index bc266e4934c41d6f5f1261e0a30e90cb72ba83a8..181e04f7142053708cc5b2338a8078f6c9fc8303 100644
--- a/src/finn/custom_op/fpgadataflow/streamingfclayer_batch.py
+++ b/src/finn/custom_op/fpgadataflow/streamingfclayer_batch.py
@@ -77,7 +77,7 @@ class StreamingFCLayer_Batch(HLSCustomOp):
"weightDataType": ("s", True, ""),
"outputDataType": ("s", True, ""),
# FINN DataType for accumulator -- auto-computed and updated
- "accDataType": ("s", False, "DataType.INT32"),
+ "accDataType": ("s", False, "INT32"),
# use xnor-popcount for binary weights/inputs, thus treating them
# as bipolar
"binaryXnorMode": ("i", False, 0),
@@ -483,6 +483,10 @@ class StreamingFCLayer_Batch(HLSCustomOp):
adt = DataType.get_smallest_possible(0 - acc_max)
else:
adt = DataType.get_smallest_possible(acc_max)
+ # ensure a datatype divisible by 8-bits in case this is the last node
+ bw = roundup_to_integer_multiple(adt.bitwidth(), 8)
+ new_adt_name = adt.name.replace(str(adt.bitwidth()), str(bw))
+ adt = DataType[new_adt_name]
self.set_nodeattr("accDataType", adt.name)
# for no-activation nodes, output dt = acc dt
self.set_nodeattr("outputDataType", adt.name)
diff --git a/src/finn/transformation/fpgadataflow/create_dataflow_partition.py b/src/finn/transformation/fpgadataflow/create_dataflow_partition.py
index 5ec4ab14d65d63523856a6bb107bf75c1ca5a261..fb8b4358abd772d13c355f797649dc3b51975b4d 100644
--- a/src/finn/transformation/fpgadataflow/create_dataflow_partition.py
+++ b/src/finn/transformation/fpgadataflow/create_dataflow_partition.py
@@ -112,6 +112,7 @@ class CreateDataflowPartition(Transformation):
"dataflow_partition" + str(target_partition_id) + "_"
)
df_model_filename = df_model_dir + "/df_model.onnx"
+ df_model.cleanup()
df_model.save(df_model_filename)
# remove all dataflow nodes from the non-dataflow model
# keep track of where the dataflow part starts
diff --git a/src/finn/transformation/general.py b/src/finn/transformation/general.py
index 4303eb17f39a9949f5729e895e449bbb6a633033..690364a7d827086217ffdf9d246c2f4c5f395772 100644
--- a/src/finn/transformation/general.py
+++ b/src/finn/transformation/general.py
@@ -189,6 +189,9 @@ class SortGraph(Transformation):
# Probably this is faster than copying initializers and more robust in general
def apply(self, model):
+ if len(model.graph.node) == 1:
+ # single-node graph, nothing to sort
+ return (model, False)
# Gather graph structure
graph_dependencies = {}
node_list = [
@@ -214,7 +217,7 @@ class SortGraph(Transformation):
for new_idx, sorted_idx in enumerate(sorted_node_indexes):
model.graph.node.insert(new_idx, node_list[sorted_idx])
- return model, False
+ return (model, False)
class ConvertSubToAdd(Transformation):
diff --git a/src/finn/transformation/merge_onnx_models.py b/src/finn/transformation/merge_onnx_models.py
index 5dc6127ed189311c72a119932394aca4745e3608..ceacab197150fe6d32e3a9eda268aed186b1a8bc 100644
--- a/src/finn/transformation/merge_onnx_models.py
+++ b/src/finn/transformation/merge_onnx_models.py
@@ -31,12 +31,12 @@ from onnx import helper
from finn.transformation import Transformation
from finn.core.modelwrapper import ModelWrapper
-import finn.util.basic as util
from finn.transformation.infer_shapes import InferShapes
from finn.transformation.infer_datatypes import InferDataTypes
from finn.transformation.infer_data_layouts import InferDataLayouts
from finn.transformation.general import (
GiveReadableTensorNames,
+ GiveRandomTensorNames,
GiveUniqueNodeNames,
GiveUniqueParameterTensors,
)
@@ -59,6 +59,9 @@ class MergeONNXModels(Transformation):
graph_modified = False
pre_model = self.pre_model
post_model = copy.deepcopy(model)
+ # to avoid mix-ups, start by giving all tensors random names
+ pre_model = pre_model.transform(GiveRandomTensorNames())
+ post_model = post_model.transform(GiveRandomTensorNames())
# check for dynamic outputs of pre model
dyn_outp = []
@@ -94,27 +97,6 @@ class MergeONNXModels(Transformation):
for n in post_model.graph.node:
n.name = ""
- # randomize all tensor names
- names1 = pre_model.get_all_tensor_names()
- names2 = post_model.get_all_tensor_names()
- used_names = names1 + names2
-
- # pre_model
- for tensor_name in names1:
- new_name = util.random_string()
- while new_name in used_names:
- new_name = util.random_string()
- pre_model.rename_tensor(tensor_name, new_name)
- used_names.append(new_name)
-
- # post_model
- for tensor in names2:
- new_name = util.random_string()
- while new_name in used_names:
- new_name = util.random_string()
- post_model.rename_tensor(tensor_name, new_name)
- used_names.append(new_name)
-
# check if models can be merged
output_model_a = dyn_outp[0].name
input_model_b = dyn_inp[0].name
@@ -124,6 +106,9 @@ class MergeONNXModels(Transformation):
output_a_shape == input_b_shape
), "Models can't be merged! Shapes don't match."
+ pre_model.save("pre.onnx")
+ post_model.save("post.onnx")
+
# connect output of one model to input of the other
for n in pre_model.graph.node:
if output_model_a == n.output[0]:
@@ -132,83 +117,43 @@ class MergeONNXModels(Transformation):
# extract information for new model
# nodes
- node_list_a = pre_model.graph.node
- node_list_b = post_model.graph.node
-
- node_list = node_list_a
- for node in node_list_b:
- node_list.append(node)
+ node_pre = [node for node in pre_model.graph.node]
+ node_post = [node for node in post_model.graph.node]
+ node_new = node_pre + node_post
# in and output
inp = pre_model.graph.input[0]
outp = post_model.graph.output[0]
+ vi_pre = [x for x in pre_model.graph.value_info]
+ out_pre = [x for x in pre_model.graph.output]
+ qa_pre = [x for x in pre_model.graph.quantization_annotation]
+ init_pre = [x for x in pre_model.graph.initializer]
+
+ vi_post = [x for x in post_model.graph.value_info]
+ qa_post = [x for x in post_model.graph.quantization_annotation]
+ init_post = [x for x in post_model.graph.initializer]
+
+ vi_new = vi_pre + vi_post + out_pre
+ qa_new = qa_pre + qa_post
+ init_new = init_pre + init_post
+
# create new graph and model
new_graph = helper.make_graph(
- nodes=node_list,
+ nodes=node_new,
name="fuse-graph",
inputs=[inp],
outputs=[outp],
- value_info=[],
+ value_info=vi_new,
)
new_model = helper.make_model(new_graph, producer_name="fuse_model")
new_model = ModelWrapper(new_model)
- # add value info from both models to new model
- # pre model
- vi_pre = [x for x in pre_model.graph.input]
- vi_pre += [x for x in pre_model.graph.output]
- vi_pre += [x for x in pre_model.graph.value_info]
- for vi in vi_pre:
- # preserve intializers, quantization/sparsity annotation, etc.
- # initializer
- init_val = pre_model.get_initializer(vi.name)
- if init_val is not None:
- new_model.set_initializer(vi.name, init_val)
- # FINN datatype
- dtype = pre_model.get_tensor_datatype(vi.name)
- new_model.set_tensor_datatype(vi.name, dtype)
- # data layout
- data_layout = pre_model.get_tensor_layout(vi.name)
- if data_layout is not None:
- new_model.set_tensor_layout(vi.name, data_layout)
- # sparsity
- sparsity = pre_model.get_tensor_sparsity(vi.name)
- if sparsity is not None:
- new_model.set_tensor_sparsity(vi.name, sparsity)
- # graph input should not be part of graph.value_info, so don't insert
- # if current vi == inp, but the quantization annotation is preserved
- if vi == inp:
- continue
- new_model.graph.value_info.append(vi)
-
- # post model
- vi_model = [x for x in post_model.graph.input]
- vi_model += [x for x in post_model.graph.output]
- vi_model += [x for x in post_model.graph.value_info]
- for vi in vi_model:
- # preserve intializers, quantization/sparsity annotation, etc.
- # initializer
- init_val = post_model.get_initializer(vi.name)
- if init_val is not None:
- new_model.set_initializer(vi.name, init_val)
- # FINN datatype
- dtype = post_model.get_tensor_datatype(vi.name)
- new_model.set_tensor_datatype(vi.name, dtype)
- # data layout
- data_layout = post_model.get_tensor_layout(vi.name)
- if data_layout is not None:
- new_model.set_tensor_layout(vi.name, data_layout)
- # sparsity
- sparsity = post_model.get_tensor_sparsity(vi.name)
- if sparsity is not None:
- new_model.set_tensor_sparsity(vi.name, sparsity)
- # graph output should not be part of graph.value_info, so don't insert
- # if current vi == outp, but the quantization annotation is preserved
- if vi == outp:
- continue
- new_model.graph.value_info.append(vi)
+ for i in init_new:
+ new_model.graph.initializer.append(i)
+ for qa in qa_new:
+ new_model.graph.quantization_annotation.append(qa)
# tidy-up new model
model = new_model
diff --git a/src/finn/util/basic.py b/src/finn/util/basic.py
index 62d5947e3b7e06375cc9d48a2cf32b4f685e7861..cc759bebb1b856a84e25978d442e460332092d23 100644
--- a/src/finn/util/basic.py
+++ b/src/finn/util/basic.py
@@ -156,13 +156,19 @@ def make_build_dir(prefix=""):
def get_by_name(container, name, name_field="name"):
- """Return item from container by .name field if it exists, None otherwise"""
+ """Return item from container by .name field if it exists, None otherwise.
+ Will throw an Exception if multiple items are found, since this violates the
+ ONNX standard."""
names = [getattr(x, name_field) for x in container]
- try:
- ind = names.index(name)
- return container[ind]
- except ValueError:
+
+ inds = [i for i, e in enumerate(names) if e == name]
+ if len(inds) > 1:
+ raise Exception("Found multiple get_by_name matches, undefined behavior")
+ elif len(inds) == 0:
return None
+ else:
+ ind = inds[0]
+ return container[ind]
def remove_by_name(container, name, name_field="name"):
diff --git a/tests/fpgadataflow/test_convert_to_hls_layers_fc.py b/tests/fpgadataflow/test_convert_to_hls_layers_fc.py
index 30d5ae64cfec84e089426d389a8cb607cd71c12f..bd600c6c57d00d5fc03152f75b9f2f8c6beeeb2c 100644
--- a/tests/fpgadataflow/test_convert_to_hls_layers_fc.py
+++ b/tests/fpgadataflow/test_convert_to_hls_layers_fc.py
@@ -89,7 +89,6 @@ def test_convert_to_hls_layers_tfc_w1a1():
assert fc3.op_type == "StreamingFCLayer_Batch"
assert model.get_tensor_shape(fc3.input[0]) == [1, 64]
assert model.get_tensor_shape(fc3.input[1]) == [64, 10]
- os.remove(export_onnx_path)
fc0w = getCustomOp(fc0)
fc0w.set_nodeattr("SIMD", 784)