diff --git a/docker/Dockerfile.finn_dev b/docker/Dockerfile.finn_dev
index f8e15f34fb4da3dc4ee353a29d26866b68879144..db49dceb2d06670dfc43059d3a4fa6160a8ded58 100644
--- a/docker/Dockerfile.finn_dev
+++ b/docker/Dockerfile.finn_dev
@@ -49,13 +49,14 @@ RUN echo "StrictHostKeyChecking no" >> /etc/ssh/ssh_config
COPY requirements.txt .
RUN pip install -r requirements.txt
RUN rm requirements.txt
-RUN pip install jupyter
-RUN pip install matplotlib
-RUN pip install pytest-dependency
-RUN pip install sphinx
-RUN pip install sphinx_rtd_theme
-RUN pip install pytest-xdist
-RUN pip install pytest-parallel
+RUN pip install jupyter==1.0.0
+RUN pip install matplotlib==3.3.1 --ignore-installed certifi
+RUN pip install pytest-dependency==0.5.1
+RUN pip install sphinx==3.1.2
+RUN pip install sphinx_rtd_theme==0.5.0
+RUN pip install pytest-xdist==2.0.0
+RUN pip install pytest-parallel==0.1.0
+RUN pip install netron==4.4.7
# switch user
RUN groupadd -g $GID $GNAME
@@ -80,19 +81,6 @@ RUN git clone https://github.com/maltanar/pyverilator /workspace/pyverilator
RUN git clone https://github.com/maltanar/PYNQ-HelloWorld.git /workspace/PYNQ-HelloWorld
# oh-my-xilinx
RUN git clone https://bitbucket.org/maltanar/oh-my-xilinx.git /workspace/oh-my-xilinx
-# netron
-RUN git clone https://github.com/lutzroeder/netron.git /workspace/netron
-
-# build and install netron
-USER root
-RUN curl -sL https://deb.nodesource.com/setup_12.x | bash -
-RUN apt-get install -y nodejs
-WORKDIR /workspace/netron
-RUN git checkout 376e9d33733a3eacfe3c432808fd46e6cd1460cb
-RUN npm install
-RUN python setup.py build
-RUN pip install /workspace/netron
-USER $UNAME
# for this developer-oriented Docker container we assume the FINN repo is cloned and mounted from the host
# at /workspace/finn -- see run-docker.sh for an example of how to do this.
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/requirements.txt b/requirements.txt
index b15d86ed89f7b0e76b772ce42aba6481937310b0..4aa1cbe3484a3447851879d7da9ce9d48b066592 100644
--- a/requirements.txt
+++ b/requirements.txt
@@ -1,13 +1,11 @@
-bitstring
-docrep
-future
+bitstring==3.1.7
+docrep==0.2.7
+future==0.18.2
numpy==1.18.0
onnx==1.6.0
onnxruntime==1.2.0
-pre-commit
-pyverilator
-scipy
-sphinx
-toposort
-vcdvcd
-wget
+pre-commit==2.6.0
+scipy==1.5.2
+toposort==1.5
+vcdvcd==1.0.5
+wget==3.2
diff --git a/src/finn/core/datatype.py b/src/finn/core/datatype.py
index 222d11a8872f9be757fd60fbfa5f8abea683311a..df895a1ad446d6b2cc3ebb24f1179944f4cfe9ab 100644
--- a/src/finn/core/datatype.py
+++ b/src/finn/core/datatype.py
@@ -50,17 +50,69 @@ class DataType(Enum):
UINT2 = auto()
UINT3 = auto()
UINT4 = auto()
+ UINT5 = auto()
+ UINT6 = auto()
+ UINT7 = auto()
UINT8 = auto()
+ UINT9 = auto()
+ UINT10 = auto()
+ UINT11 = auto()
+ UINT12 = auto()
+ UINT13 = auto()
+ UINT14 = auto()
+ UINT15 = auto()
UINT16 = auto()
+ UINT17 = auto()
+ UINT18 = auto()
+ UINT19 = auto()
+ UINT20 = auto()
+ UINT21 = auto()
+ UINT22 = auto()
+ UINT23 = auto()
+ UINT24 = auto()
+ UINT25 = auto()
+ UINT26 = auto()
+ UINT27 = auto()
+ UINT28 = auto()
+ UINT29 = auto()
+ UINT30 = auto()
+ UINT31 = auto()
UINT32 = auto()
+ UINT64 = auto()
BIPOLAR = auto()
TERNARY = auto()
INT2 = auto()
INT3 = auto()
INT4 = auto()
+ INT5 = auto()
+ INT6 = auto()
+ INT7 = auto()
INT8 = auto()
+ INT9 = auto()
+ INT10 = auto()
+ INT11 = auto()
+ INT12 = auto()
+ INT13 = auto()
+ INT14 = auto()
+ INT15 = auto()
INT16 = auto()
+ INT17 = auto()
+ INT18 = auto()
+ INT19 = auto()
+ INT20 = auto()
+ INT21 = auto()
+ INT22 = auto()
+ INT23 = auto()
+ INT24 = auto()
+ INT25 = auto()
+ INT26 = auto()
+ INT27 = auto()
+ INT28 = auto()
+ INT29 = auto()
+ INT30 = auto()
+ INT31 = auto()
INT32 = auto()
+ INT64 = auto()
FLOAT32 = auto()
def bitwidth(self):
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 87f52eeea591ba42bf5374df3c93bcc3e4f8e944..ea6922123a1334a7ea0d0568e09c043e06490f38 100644
--- a/src/finn/custom_op/fpgadataflow/streamingfclayer_batch.py
+++ b/src/finn/custom_op/fpgadataflow/streamingfclayer_batch.py
@@ -36,6 +36,7 @@ from finn.custom_op.fpgadataflow import HLSCustomOp
from finn.util.basic import (
interleave_matrix_outer_dim_from_partitions,
roundup_to_integer_multiple,
+ calculate_matvec_accumulator_range,
)
from finn.util.data_packing import (
npy_to_rtlsim_input,
@@ -72,6 +73,8 @@ class StreamingFCLayer_Batch(HLSCustomOp):
"inputDataType": ("s", True, ""),
"weightDataType": ("s", True, ""),
"outputDataType": ("s", True, ""),
+ # FINN DataType for accumulator -- auto-computed and updated
+ "accDataType": ("s", False, "INT32"),
# use xnor-popcount for binary weights/inputs, thus treating them
# as bipolar
"binaryXnorMode": ("i", False, 0),
@@ -424,6 +427,51 @@ class StreamingFCLayer_Batch(HLSCustomOp):
ret = np.flip(ret, axis=-1)
return ret
+ def minimize_accumulator_width(self, model):
+ weights = model.get_initializer(self.onnx_node.input[1])
+ if len(self.onnx_node.input) > 2:
+ thresholds = model.get_initializer(self.onnx_node.input[2])
+ else:
+ thresholds = None
+ idt = self.get_input_datatype()
+ # calculate minimum and maximum values of accumulator
+ (acc_min, acc_max) = calculate_matvec_accumulator_range(weights, idt)
+ if thresholds is not None:
+ threshold_tensor = self.get_hls_compatible_threshold_tensor(thresholds)
+ # set threshold datatype (and accumulator datatype implicitly)
+ min_threshold = thresholds.min()
+ max_threshold = thresholds.max()
+ # get range required by threshold values
+ tdt_min = min(acc_min, min_threshold)
+ tdt_max = max(acc_max, max_threshold)
+ if tdt_min < 0:
+ if abs(tdt_min) > tdt_max:
+ tdt = DataType.get_smallest_possible(tdt_min)
+ else:
+ tdt = DataType.get_smallest_possible(0 - tdt_max)
+ else:
+ tdt = DataType.get_smallest_possible(tdt_max)
+ assert np.vectorize(tdt.allowed)(
+ threshold_tensor
+ ).all(), "Thresholds can't be expressed with type %s" % str(tdt)
+ self.set_nodeattr("accDataType", tdt.name)
+ else:
+ if acc_min < 0:
+ if abs(acc_min) > acc_max:
+ adt = DataType.get_smallest_possible(acc_min)
+ else:
+ 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)
+ return DataType[self.get_nodeattr("accDataType")]
+
def get_hls_compatible_threshold_tensor(self, orig_thres_matrix):
"""Convert the original numpy weight matrix orig_weight_matrix into
a form suitable for passing to the hlslib call:
@@ -573,7 +621,6 @@ class StreamingFCLayer_Batch(HLSCustomOp):
thresholds = model.get_initializer(self.onnx_node.input[2])
if thresholds is not None:
threshold_tensor = self.get_hls_compatible_threshold_tensor(thresholds)
- tdt = DataType.INT32
# use UINT32 threshold export for bipolar times bipolar
inp_is_bipolar = self.get_input_datatype() == DataType.BIPOLAR
wt_is_bipolar = self.get_weight_datatype() == DataType.BIPOLAR
@@ -583,11 +630,12 @@ class StreamingFCLayer_Batch(HLSCustomOp):
bin_xnor_mode = self.get_nodeattr("binaryXnorMode") == 1
inp_is_bipolar = inp_is_bipolar or (inp_is_binary and bin_xnor_mode)
wt_is_bipolar = wt_is_bipolar or (wt_is_binary and bin_xnor_mode)
- if inp_is_bipolar and wt_is_bipolar:
- tdt = DataType.UINT32
+ # get computed threshold datatype from attribute
+ tdt = DataType[self.get_nodeattr("accDataType")]
+
assert np.vectorize(tdt.allowed)(
threshold_tensor
- ).all(), "Thresholds are not int"
+ ).all(), "Thresholds can't be expressed with type %s" % str(tdt)
thresholds_hls_code = numpy_to_hls_code(
threshold_tensor, tdt, "thresholds", False, True
)
diff --git a/src/finn/custom_op/fpgadataflow/thresholding_batch.py b/src/finn/custom_op/fpgadataflow/thresholding_batch.py
index 379ebd92d86d54c6bc621c7f89b01eacba2b5d3f..562bab0f18990096f7364b3a4e2bcbbbf4ce2b58 100644
--- a/src/finn/custom_op/fpgadataflow/thresholding_batch.py
+++ b/src/finn/custom_op/fpgadataflow/thresholding_batch.py
@@ -283,10 +283,25 @@ class Thresholding_Batch(HLSCustomOp):
thresholds = model.get_initializer(self.onnx_node.input[1])
threshold_tensor = self.get_hls_compatible_threshold_tensor(thresholds)
- tdt = DataType.INT32
+
+ min_threshold = thresholds.min()
+ max_threshold = thresholds.max()
+ min_input = self.get_input_datatype().min()
+ max_input = self.get_input_datatype().max()
+ # get range required by threshold values
+ tdt_min = min(min_input, min_threshold)
+ tdt_max = max(max_input, max_threshold)
+ if tdt_min < 0:
+ if abs(tdt_min) > tdt_max:
+ tdt = DataType.get_smallest_possible(tdt_min)
+ else:
+ tdt = DataType.get_smallest_possible(0 - tdt_max - 1)
+ else:
+ tdt = DataType.get_smallest_possible(tdt_max)
assert np.vectorize(tdt.allowed)(
threshold_tensor
- ).all(), "Thresholds are not int"
+ ).all(), "Thresholds can't be expressed with type %s" % str(tdt)
+
thresholds_hls_code = numpy_to_hls_code(
threshold_tensor, tdt, "thresholds", False, True
)
diff --git a/src/finn/transformation/fpgadataflow/convert_to_hls_layers.py b/src/finn/transformation/fpgadataflow/convert_to_hls_layers.py
index e6dca0e4b05f943c971bc0f97af03f5038fd0dab..88f5fa926f73d5cb1919a02c83153cb8d1894711 100644
--- a/src/finn/transformation/fpgadataflow/convert_to_hls_layers.py
+++ b/src/finn/transformation/fpgadataflow/convert_to_hls_layers.py
@@ -40,6 +40,9 @@ from finn.transformation.general import SortGraph
import finn.core.data_layout as DataLayout
from finn.util.onnx import nchw_to_nhwc
from finn.util.basic import get_by_name
+from finn.transformation.fpgadataflow.minimize_accumulator_width import (
+ MinimizeAccumulatorWidth,
+)
class InferConvInpGen(Transformation):
@@ -489,6 +492,7 @@ class InferBinaryStreamingFCLayer(Transformation):
graph.node.remove(n)
graph_modified = True
if graph_modified:
+ model = model.transform(MinimizeAccumulatorWidth())
model = model.transform(InferShapes())
model = model.transform(InferDataTypes())
return (model, graph_modified)
@@ -623,6 +627,7 @@ class InferQuantizedStreamingFCLayer(Transformation):
graph.node.remove(n)
graph_modified = True
if graph_modified:
+ model = model.transform(MinimizeAccumulatorWidth())
model = model.transform(InferShapes())
model = model.transform(InferDataTypes())
return (model, graph_modified)
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/fpgadataflow/make_zynq_proj.py b/src/finn/transformation/fpgadataflow/make_zynq_proj.py
index 1558d7399fe5399053c3f347cd06c4e0d76753e7..095327be0d3c36f201bcf343d8aea61aa069b8e1 100644
--- a/src/finn/transformation/fpgadataflow/make_zynq_proj.py
+++ b/src/finn/transformation/fpgadataflow/make_zynq_proj.py
@@ -294,11 +294,12 @@ class ZynqBuild(Transformation):
# Build each kernel individually
sdp_nodes = model.get_nodes_by_op_type("StreamingDataflowPartition")
for sdp_node in sdp_nodes:
+ prefix = sdp_node.name + "_"
sdp_node = getCustomOp(sdp_node)
dataflow_model_filename = sdp_node.get_nodeattr("model")
kernel_model = ModelWrapper(dataflow_model_filename)
kernel_model = kernel_model.transform(InsertFIFO())
- kernel_model = kernel_model.transform(GiveUniqueNodeNames())
+ kernel_model = kernel_model.transform(GiveUniqueNodeNames(prefix))
kernel_model.save(dataflow_model_filename)
kernel_model = kernel_model.transform(
PrepareIP(self.fpga_part, self.period_ns)
diff --git a/src/finn/transformation/fpgadataflow/minimize_accumulator_width.py b/src/finn/transformation/fpgadataflow/minimize_accumulator_width.py
new file mode 100644
index 0000000000000000000000000000000000000000..2c54a5efbd3b28f0fbfd074b512929edab234e78
--- /dev/null
+++ b/src/finn/transformation/fpgadataflow/minimize_accumulator_width.py
@@ -0,0 +1,48 @@
+# Copyright (c) 2020, Xilinx
+# All rights reserved.
+#
+# Redistribution and use in source and binary forms, with or without
+# modification, are permitted provided that the following conditions are met:
+#
+# * Redistributions of source code must retain the above copyright notice, this
+# list of conditions and the following disclaimer.
+#
+# * Redistributions in binary form must reproduce the above copyright notice,
+# this list of conditions and the following disclaimer in the documentation
+# and/or other materials provided with the distribution.
+#
+# * Neither the name of FINN nor the names of its
+# contributors may be used to endorse or promote products derived from
+# this software without specific prior written permission.
+#
+# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+# AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+# IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+# DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
+# FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+# DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+# SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+# CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+# OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+from finn.custom_op.registry import getCustomOp
+from finn.transformation import Transformation
+from finn.util.fpgadataflow import is_fpgadataflow_node
+
+
+class MinimizeAccumulatorWidth(Transformation):
+ """For relevant nodes, call the accumulator width minimization
+ functions to save on resources. May alter tensor DataType for
+ certain nodes if they produce an accumulator as result."""
+
+ def __init__(self):
+ super().__init__()
+
+ def apply(self, model):
+ for node in model.graph.node:
+ if is_fpgadataflow_node(node) is True:
+ inst = getCustomOp(node)
+ if hasattr(inst, "minimize_accumulator_width"):
+ inst.minimize_accumulator_width(model)
+ return (model, False)
diff --git a/src/finn/transformation/general.py b/src/finn/transformation/general.py
index 4303eb17f39a9949f5729e895e449bbb6a633033..8ad59d2baf3015cfebffeff88a059f48d9428371 100644
--- a/src/finn/transformation/general.py
+++ b/src/finn/transformation/general.py
@@ -81,14 +81,19 @@ class RemoveStaticGraphInputs(Transformation):
class GiveUniqueNodeNames(Transformation):
- """Give unique names to each node in the graph using enumeration."""
+ """Give unique names to each node in the graph using enumeration, starting
+ with given prefix (if specified in the constructor)."""
+
+ def __init__(self, prefix=""):
+ super().__init__()
+ self.prefix = prefix
def apply(self, model):
optype_count = {}
for n in model.graph.node:
if n.op_type not in optype_count.keys():
optype_count[n.op_type] = 0
- n.name = "%s_%d" % (n.op_type, optype_count[n.op_type])
+ n.name = "%s%s_%d" % (self.prefix, n.op_type, optype_count[n.op_type])
optype_count[n.op_type] += 1
# return model_was_changed = False as single iteration is always enough
return (model, False)
@@ -189,6 +194,9 @@ 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 +222,7 @@ 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/transformation/streamline/round_thresholds.py b/src/finn/transformation/streamline/round_thresholds.py
index c33281d85449c173a4631297fd1d67ac0aed8c81..8626ef40619b067c6672c9017ddcb747998c3f2c 100644
--- a/src/finn/transformation/streamline/round_thresholds.py
+++ b/src/finn/transformation/streamline/round_thresholds.py
@@ -51,10 +51,20 @@ class RoundAndClipThresholds(Transformation):
model.set_tensor_datatype(n.input[1], idtype)
graph_modified = True
if idtype.is_integer() and not idtype.signed() and (Tnew < 0).any():
- # clip any negative thresholds
+ # clip any negative thresholds if input is unsigned
Tnew = np.clip(Tnew, 0, None)
model.set_initializer(n.input[1], Tnew)
# use same datatype as inputs for thresholds
model.set_tensor_datatype(n.input[1], idtype)
graph_modified = True
+ if idtype.is_integer() and (
+ (Tnew < (idtype.min() - 1)).any()
+ or (Tnew > (idtype.max() + 1)).any()
+ ):
+ # clip any large thresholds to input range + 1
+ Tnew = np.clip(Tnew, idtype.min() - 1, idtype.max() + 1)
+ model.set_initializer(n.input[1], Tnew)
+ # use same datatype as inputs for thresholds
+ model.set_tensor_datatype(n.input[1], idtype)
+ graph_modified = True
return (model, graph_modified)
diff --git a/src/finn/util/basic.py b/src/finn/util/basic.py
index 6c92e9b2765b1c2be6f95ee148964bccfb3cd7be..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"):
@@ -259,6 +265,33 @@ def pad_tensor_to_multiple_of(ndarray, pad_to_dims, val=0, distr_pad=False):
return ret
+def calculate_matvec_accumulator_range(matrix, vec_dt):
+ """Calculate the minimum and maximum possible result (accumulator) values
+ for a dot product x * A, given matrix A of dims (MW, MH), and vector (1, MW)
+ with datatype vec_dt. Returns (acc_min, acc_max).
+ """
+ min_weight = matrix.min()
+ max_weight = matrix.max()
+ perceptive_field_elems = matrix.shape[0]
+ min_input = vec_dt.min()
+ max_input = vec_dt.max()
+ # calculate minimum and maximum values of accumulator
+ # assume inputs span the whole range of the input datatype
+ acc_min = perceptive_field_elems * min(
+ min_weight * max_input,
+ min_weight * min_input,
+ max_weight * max_input,
+ max_weight * min_input,
+ )
+ acc_max = perceptive_field_elems * max(
+ min_weight * max_input,
+ min_weight * min_input,
+ max_weight * max_input,
+ max_weight * min_input,
+ )
+ return (acc_min, acc_max)
+
+
def gen_finn_dt_tensor(finn_dt, tensor_shape):
"""Generates random tensor in given shape and with given FINN DataType."""
if type(tensor_shape) == list:
diff --git a/tests/end2end/test_end2end_cnv_w1a1.py b/tests/end2end/test_end2end_cnv_w1a1.py
index ebca224389550929cebd542cf4201cf62481a169..f931f91c89f738899ff9e6584be81a3b2d542227 100644
--- a/tests/end2end/test_end2end_cnv_w1a1.py
+++ b/tests/end2end/test_end2end_cnv_w1a1.py
@@ -78,6 +78,7 @@ from finn.transformation.fpgadataflow.annotate_resources import AnnotateResource
from finn.transformation.fpgadataflow.prepare_rtlsim import PrepareRTLSim
from finn.transformation.fpgadataflow.insert_fifo import InsertFIFO
from finn.core.throughput_test import throughput_test_rtlsim
+import warnings
build_dir = "/tmp/" + os.environ["FINN_INST_NAME"]
test_pynq_board = os.getenv("PYNQ_BOARD", default="Pynq-Z1")
@@ -317,6 +318,10 @@ def test_end2end_cnv_w1a1_synth_pynq_project():
)
model = model.transform(SynthPYNQProject())
model = model.transform(AnnotateResources("synth"))
+ warnings.warn(
+ "Post-synthesis resources (excluding shell): "
+ + model.get_metadata_prop("res_total_synth")
+ )
model.save(build_dir + "/end2end_cnv_w1a1_synth.onnx")
diff --git a/tests/end2end/test_end2end_cnv_w2a2.py b/tests/end2end/test_end2end_cnv_w2a2.py
index 2e34990007677ce1b8e0a9ae4a1781d4527ee040..239094a3c931c16b3afe8d1874345e4dc90334ef 100644
--- a/tests/end2end/test_end2end_cnv_w2a2.py
+++ b/tests/end2end/test_end2end_cnv_w2a2.py
@@ -77,6 +77,7 @@ from finn.transformation.fpgadataflow.annotate_resources import AnnotateResource
from finn.transformation.fpgadataflow.prepare_rtlsim import PrepareRTLSim
from finn.transformation.fpgadataflow.insert_fifo import InsertFIFO
from finn.core.throughput_test import throughput_test_rtlsim
+import warnings
build_dir = "/tmp/" + os.environ["FINN_INST_NAME"]
test_pynq_board = os.getenv("PYNQ_BOARD", default="Pynq-Z1")
@@ -315,6 +316,10 @@ def test_end2end_cnv_w2a2_synth_pynq_project():
)
model = model.transform(SynthPYNQProject())
model = model.transform(AnnotateResources("synth"))
+ warnings.warn(
+ "Post-synthesis resources (excluding shell): "
+ + model.get_metadata_prop("res_total_synth")
+ )
model.save(build_dir + "/end2end_cnv_w2a2_synth.onnx")
diff --git a/tests/end2end/test_end2end_tfc_w1a1.py b/tests/end2end/test_end2end_tfc_w1a1.py
index b827cbb1c31cc84de9fa5d4df4d6b23e02a02a5f..1a3cc4f1bb9232809e864bb0c784498534f63631 100644
--- a/tests/end2end/test_end2end_tfc_w1a1.py
+++ b/tests/end2end/test_end2end_tfc_w1a1.py
@@ -79,6 +79,7 @@ from finn.transformation.fpgadataflow.annotate_resources import AnnotateResource
from finn.transformation.fpgadataflow.prepare_rtlsim import PrepareRTLSim
from finn.core.throughput_test import throughput_test_rtlsim
import finn.util.vcd as vcd
+import warnings
build_dir = "/tmp/" + os.environ["FINN_INST_NAME"]
test_pynq_board = os.getenv("PYNQ_BOARD", default="Pynq-Z1")
@@ -241,11 +242,11 @@ def test_end2end_tfc_w1a1_throughput_test_rtlsim():
# run through IP-stitched rtlsim with increasing batch sizes and
# check the number of cycles it takes to execute
ret = throughput_test_rtlsim(model, 1)
- assert ret["cycles"] == 205
+ assert np.isclose(ret["cycles"], 205, atol=5)
ret = throughput_test_rtlsim(model, 10)
- assert ret["cycles"] == 844
+ assert np.isclose(ret["cycles"], 844, atol=10)
ret = throughput_test_rtlsim(model, 100)
- assert ret["cycles"] == 7234
+ assert np.isclose(ret["cycles"], 7234, atol=100)
@pytest.mark.vivado
@@ -314,6 +315,10 @@ def test_end2end_tfc_w1a1_synth_pynq_project():
)
model = model.transform(SynthPYNQProject())
model = model.transform(AnnotateResources("synth"))
+ warnings.warn(
+ "Post-synthesis resources (excluding shell): "
+ + model.get_metadata_prop("res_total_synth")
+ )
model.save(build_dir + "/end2end_tfc_w1a1_synth.onnx")
diff --git a/tests/end2end/test_end2end_tfc_w1a2.py b/tests/end2end/test_end2end_tfc_w1a2.py
index 755650e3d4da6947a93495fd5bbe0464cf485193..0f066cb06c53ce118d0a357fce0999299d7f3305 100644
--- a/tests/end2end/test_end2end_tfc_w1a2.py
+++ b/tests/end2end/test_end2end_tfc_w1a2.py
@@ -74,6 +74,7 @@ from finn.util.basic import pynq_part_map
from finn.util.test import get_test_model_trained, load_test_checkpoint_or_skip
from finn.transformation.fpgadataflow.annotate_resources import AnnotateResources
from finn.transformation.fpgadataflow.prepare_rtlsim import PrepareRTLSim
+import warnings
build_dir = "/tmp/" + os.environ["FINN_INST_NAME"]
test_pynq_board = os.getenv("PYNQ_BOARD", default="Pynq-Z1")
@@ -270,6 +271,10 @@ def test_end2end_tfc_w1a2_synth_pynq_project():
)
model = model.transform(SynthPYNQProject())
model = model.transform(AnnotateResources("synth"))
+ warnings.warn(
+ "Post-synthesis resources (excluding shell): "
+ + model.get_metadata_prop("res_total_synth")
+ )
model.save(build_dir + "/end2end_tfc_w1a2_synth.onnx")
diff --git a/tests/end2end/test_end2end_tfc_w2a2.py b/tests/end2end/test_end2end_tfc_w2a2.py
index 4b2dd9ef01850897d95ede1214f87e9aa5b79f63..6eb613fc877b6e6801140f2a03c3a9509c08c0cb 100644
--- a/tests/end2end/test_end2end_tfc_w2a2.py
+++ b/tests/end2end/test_end2end_tfc_w2a2.py
@@ -74,6 +74,7 @@ from finn.util.basic import pynq_part_map
from finn.util.test import get_test_model_trained, load_test_checkpoint_or_skip
from finn.transformation.fpgadataflow.annotate_resources import AnnotateResources
from finn.transformation.fpgadataflow.prepare_rtlsim import PrepareRTLSim
+import warnings
build_dir = "/tmp/" + os.environ["FINN_INST_NAME"]
test_pynq_board = os.getenv("PYNQ_BOARD", default="Pynq-Z1")
@@ -270,6 +271,10 @@ def test_end2end_tfc_w2a2_synth_pynq_project():
)
model = model.transform(SynthPYNQProject())
model = model.transform(AnnotateResources("synth"))
+ warnings.warn(
+ "Post-synthesis resources (excluding shell): "
+ + model.get_metadata_prop("res_total_synth")
+ )
model.save(build_dir + "/end2end_tfc_w2a2_synth.onnx")
diff --git a/tests/end2end/test_zynqbuild_end2end_cnv_w1a1.py b/tests/end2end/test_zynqbuild_end2end_cnv_w1a1.py
index 25cafcfd4c552fb368cbaca2d1d2714cf2d14011..a272fadc12f095034693e555e4d791e9e73262ab 100644
--- a/tests/end2end/test_zynqbuild_end2end_cnv_w1a1.py
+++ b/tests/end2end/test_zynqbuild_end2end_cnv_w1a1.py
@@ -63,6 +63,7 @@ from finn.transformation.lower_convs_to_matmul import LowerConvsToMatMul
from finn.transformation.streamline.reorder import MakeMaxPoolNHWC
from finn.transformation.infer_data_layouts import InferDataLayouts
from finn.transformation.fpgadataflow.annotate_cycles import AnnotateCycles
+import warnings
build_dir = "/tmp/" + os.environ["FINN_INST_NAME"]
@@ -178,6 +179,10 @@ def test_end2end_zynqbuild_cnv_w1a1_build():
)
model = model.transform(ZynqBuild(test_pynq_board, target_clk_ns))
model = model.transform(AnnotateResources("synth"))
+ warnings.warn(
+ "Post-synthesis resources (excluding shell): "
+ + model.get_metadata_prop("res_total_synth")
+ )
model.save(build_dir + "/end2end_zynqbuild_cnv_w1a1_build.onnx")
diff --git a/tests/end2end/test_zynqbuild_end2end_tfc_w1a1.py b/tests/end2end/test_zynqbuild_end2end_tfc_w1a1.py
index ff2af70731d9248dd2593db5be9e465fa86157dd..8b298d5644d6d6cda038e8ca1757be7538ba9804 100644
--- a/tests/end2end/test_zynqbuild_end2end_tfc_w1a1.py
+++ b/tests/end2end/test_zynqbuild_end2end_tfc_w1a1.py
@@ -64,6 +64,7 @@ from finn.util.test import get_test_model_trained, load_test_checkpoint_or_skip
from finn.transformation.fpgadataflow.annotate_resources import AnnotateResources
from finn.transformation.infer_data_layouts import InferDataLayouts
from finn.transformation.fpgadataflow.make_zynq_proj import ZynqBuild
+import warnings
build_dir = "/tmp/" + os.environ["FINN_INST_NAME"]
test_pynq_board = os.getenv("PYNQ_BOARD", default="Pynq-Z1")
@@ -160,6 +161,10 @@ def test_end2end_zynqbuild_tfc_w1a1_build():
)
model = model.transform(ZynqBuild(test_pynq_board, target_clk_ns))
model = model.transform(AnnotateResources("synth"))
+ warnings.warn(
+ "Post-synthesis resources (excluding shell): "
+ + model.get_metadata_prop("res_total_synth")
+ )
model.save(build_dir + "/end2end_zynqbuild_tfc_w1a1_build.onnx")
diff --git a/tests/end2end/test_zynqbuild_end2end_tfc_w2a2.py b/tests/end2end/test_zynqbuild_end2end_tfc_w2a2.py
index 7b28090854adbbcb6f400f73c2b6f6557f540e5e..bdb24d82dd639abe52aac9688b0b98430f72cabd 100644
--- a/tests/end2end/test_zynqbuild_end2end_tfc_w2a2.py
+++ b/tests/end2end/test_zynqbuild_end2end_tfc_w2a2.py
@@ -58,6 +58,7 @@ from finn.util.basic import pynq_part_map
from finn.util.test import get_test_model_trained, load_test_checkpoint_or_skip
from finn.transformation.fpgadataflow.annotate_resources import AnnotateResources
from finn.transformation.fpgadataflow.make_zynq_proj import ZynqBuild
+import warnings
build_dir = "/tmp/" + os.environ["FINN_INST_NAME"]
test_pynq_board = os.getenv("PYNQ_BOARD", default="Pynq-Z1")
@@ -149,6 +150,10 @@ def test_end2end_zynqbuild_tfc_w2a2_build():
)
model = model.transform(ZynqBuild(test_pynq_board, target_clk_ns))
model = model.transform(AnnotateResources("synth"))
+ warnings.warn(
+ "Post-synthesis resources (excluding shell): "
+ + model.get_metadata_prop("res_total_synth")
+ )
model.save(build_dir + "/end2end_zynqbuild_tfc_w2a2_build.onnx")
diff --git a/tests/fpgadataflow/test_convert_to_hls_layers_fc.py b/tests/fpgadataflow/test_convert_to_hls_layers_fc.py
index d77065ad9396d0cc8dd57a39ed823fffcb30ee47..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)
@@ -123,6 +122,7 @@ def test_convert_to_hls_layers_tfc_w1a1():
# do forward pass in PyTorch/Brevitas
expected = tfc.forward(input_tensor).detach().numpy()
assert np.isclose(produced, expected, atol=1e-3).all()
+ os.remove(export_onnx_path)
@pytest.mark.vivado