diff --git a/.gitignore b/.gitignore
index f838c1695130d232ac6a2b888aed0cea31aafaa7..8b3166a44070a4575aac86c445c4504b594cda08 100644
--- a/.gitignore
+++ b/.gitignore
@@ -78,3 +78,6 @@ MANIFEST
# Jenkins cfg dir
/docker/jenkins_home
+
+# SSH key dir mounted into Docker
+/ssh_keys/
diff --git a/docker/Dockerfile.finn_dev b/docker/Dockerfile.finn_dev
index 22e3eb623c7a5da19a5e3ae2284557577898ad23..0e12b504a26ccdb8fd78e162f04cfdeab5a186f1 100644
--- a/docker/Dockerfile.finn_dev
+++ b/docker/Dockerfile.finn_dev
@@ -42,7 +42,7 @@ WORKDIR /workspace
RUN apt-get update
RUN apt-get -y upgrade
RUN apt-get install -y build-essential libglib2.0-0 libsm6 libxext6 libxrender-dev
-RUN apt-get install -y verilator nano zsh
+RUN apt-get install -y verilator nano zsh rsync
RUN apt-get -y install sshpass
RUN echo "StrictHostKeyChecking no" >> /etc/ssh/ssh_config
diff --git a/docker/finn_entrypoint.sh b/docker/finn_entrypoint.sh
index e34b6ce9cc4488c806da8bcf3cc5cc8e500ae806..132d5bdaa286ba3e50bbd06971e9139f5859ef11 100644
--- a/docker/finn_entrypoint.sh
+++ b/docker/finn_entrypoint.sh
@@ -13,9 +13,9 @@ gecho () {
# checkout the correct dependency repo commits
# the repos themselves are cloned in the Dockerfile
-BREVITAS_COMMIT=989cdfdba4700fdd900ba0b25a820591d561c21a
+BREVITAS_COMMIT=f9a27226d4acf1661dd38bc449f71f89e0983cce
CNPY_COMMIT=4e8810b1a8637695171ed346ce68f6984e585ef4
-HLSLIB_COMMIT=13e9b0772a27a3a1efc40c878d8e78ed09efb716
+HLSLIB_COMMIT=8aed899c278c36c977a249558d71795086cf852c
PYVERILATOR_COMMIT=c97a5ba41bbc7c419d6f25c74cdf3bdc3393174f
PYNQSHELL_COMMIT=0c82a61b0ec1a07fa275a14146233824ded7a13d
OMX_COMMIT=1bae737669901e762f581af73348332b5c4b2ada
diff --git a/docs/finn/example_networks.rst b/docs/finn/example_networks.rst
index 9f221871f09bf655db9d81988d6fa83e53473634..86bb2bd11fd805a23a3bdf6da8a8ed686259ecc1 100644
--- a/docs/finn/example_networks.rst
+++ b/docs/finn/example_networks.rst
@@ -20,17 +20,17 @@ version, this is indicated by an x mark in the table.
+-----------------------+------------+----------+----------+----------+----------+----------+----------+
| Export/Import | x | x | x | x | x | x | x |
+-----------------------+------------+----------+----------+----------+----------+----------+----------+
-| Streamlining | x | x | x | x | x | | |
+| Streamlining | x | x | x | x | x | | x |
+-----------------------+------------+----------+----------+----------+----------+----------+----------+
-| Convert to HLS layers | x | x | x | x | x | | |
+| Convert to HLS layers | x | x | x | x | x | | x |
+-----------------------+------------+----------+----------+----------+----------+----------+----------+
-| Stitched IP | x | x | x | x | x | | |
+| Stitched IP | x | x | x | x | x | | x |
+-----------------------+------------+----------+----------+----------+----------+----------+----------+
-| Hardware test | x | x | x | | x | | |
+| Hardware test | x | x | x | | x | | x |
+-----------------------+------------+----------+----------+----------+----------+----------+----------+
-| cppsim | x | x | x | x | x | | |
+| cppsim | x | x | x | x | x | | x |
+-----------------------+------------+----------+----------+----------+----------+----------+----------+
-| rtlsim node-by-node | x | x | x | x | x | | |
+| rtlsim node-by-node | x | x | x | x | x | | x |
+-----------------------+------------+----------+----------+----------+----------+----------+----------+
-| rtlsim stitched IP | x | x | x | x | x | | |
+| rtlsim stitched IP | x | x | x | x | x | | x |
+-----------------------+------------+----------+----------+----------+----------+----------+----------+
diff --git a/docs/finn/getting_started.rst b/docs/finn/getting_started.rst
index 95594bb67a2be3a4c3fbba488c75a704f623c136..f4fa7a13dcbe4fe8ab9667a111df00c605747710 100644
--- a/docs/finn/getting_started.rst
+++ b/docs/finn/getting_started.rst
@@ -18,6 +18,7 @@ Requirements
* A working Vivado 2019.1 installation
* A `VIVADO_PATH` environment variable pointing to the Vivado installation directory (e.g. the directory where settings64.sh is located)
* (optional) A PYNQ board with a network connection
+ * the ``bitstring`` package must be installed on the PYNQ: ``sudo pip3 install bitstring``
Running FINN in Docker
======================
@@ -30,6 +31,7 @@ Getting an interactive shell for development or experimentation
sh run_docker.sh
Simply running sh run-docker.sh without any additional arguments will clone the dependency repos, create a Docker container and give you a terminal with you can use for development for experimentation.
+If you want a new terminal on an already-running container, you can do this with `docker exec -it finn_dev_<username> bash`.
.. warning:: The Docker container is spawned with the `--rm` option, so make sure that any important files you created inside the container are either in the /workspace/finn folder (which is mounted from the host computer) or otherwise backed up.
diff --git a/docs/finn/internals.rst b/docs/finn/internals.rst
index 7a4bc687eeb827320991f7d3f1ef8cc35e97f3da..dee62f09a9253380e05300dac8fa34915c20dab5 100644
--- a/docs/finn/internals.rst
+++ b/docs/finn/internals.rst
@@ -16,6 +16,10 @@ Custom Quantization Annotations
ONNX does not support datatypes smaller than 8-bit integers, whereas in FINN we are interested in smaller integers down to ternary and bipolar. To make this work, FINN uses the quantization_annotation field in ONNX to annotate tensors with their FINN DataType (:py:mod:`finn.core.datatype.DataType`) information. However, all tensors are expected to use single-precision floating point (float32) storage in FINN. This means we store even a 1-bit value as floating point for the purposes of representation. The FINN compiler flow is responsible for eventually producing a packed representation for the target hardware, where the 1-bit is actually stored as 1-bit.
+Note that FINN uses floating point tensors as a carrier data type to represent integers. Floating point arithmetic can introduce rounding errors, e.g. (int_num * float_scale) / float_scale is not always equal to int_num.
+When using the custom ONNX execution flow, FINN will attempt to sanitize any rounding errors for integer tensors. See (:py:mod:`finn.util.basic.sanitize_quant_values`) for more information.
+This behavior can be disabled (not recommended!) by setting the environment variable SANITIZE_QUANT_TENSORS=0.
+
Custom Operations/Nodes
=======================
diff --git a/docs/finn/verification.rst b/docs/finn/verification.rst
index 391c6f999312839daca0d4161336c7c0ae822f89..c52c0840aa40566d930164490b1fd249d7c07757 100644
--- a/docs/finn/verification.rst
+++ b/docs/finn/verification.rst
@@ -28,4 +28,15 @@ This simulation can be used for a model containing several HLS custom operations
Emulation using PyVerilator
===========================
-The emulation using PyVerilator can be used when IP blocks were generated, either node by node or of a whole design. For that purpose PyVerilator gets the generated verilog files.
+The emulation using PyVerilator can be used when IP blocks were generated, either node by node or of a whole (IP-stitched) design. For that purpose PyVerilator gets the generated verilog files.
+
+For debugging purposes, it's possible to generate .vcd trace files that show the value of external & internal signals as the emuation is running. To enable this:
+ - for node-by-node rtlsim, set the `rtlsim_trace` attribute of each node of interest to either a file name for the vcd or `default` to use the node name as the filename.
+ - for IP-stitched rtlsim, set the `rtlsim_trace` metadata_prop for the graph as per above.
+
+To control the tracing depth in the module hierarchy, use the `RTLSIM_TRACE_DEPTH` environment variable (default is 1):
+ - level 1 shows top-level input/output streams
+ - level 2 shows per-layer input/output streams
+ - level 3 shows per full-layer I/O including FIFO count signals
+
+Note that deeper tracing will take longer to execute and may produce very large .vcd files.
diff --git a/run-docker.sh b/run-docker.sh
index e07556716db335421f57a390f1e6a17168ac058b..00ca8f86985a78d8f2af099c51dcd4b80cd2e974 100755
--- a/run-docker.sh
+++ b/run-docker.sh
@@ -65,6 +65,11 @@ DOCKER_INST_NAME="finn_dev_${DOCKER_UNAME}"
# ensure Docker tag and inst. name are all lowercase
DOCKER_TAG=$(echo "$DOCKER_TAG" | tr '[:upper:]' '[:lower:]')
DOCKER_INST_NAME=$(echo "$DOCKER_INST_NAME" | tr '[:upper:]' '[:lower:]')
+# Absolute path to this script, e.g. /home/user/bin/foo.sh
+SCRIPT=$(readlink -f "$0")
+# Absolute path this script is in, thus /home/user/bin
+SCRIPTPATH=$(dirname "$SCRIPT")
+
# the settings below will be taken from environment variables if available,
# otherwise the defaults below will be used
: ${JUPYTER_PORT=8888}
@@ -74,11 +79,7 @@ DOCKER_INST_NAME=$(echo "$DOCKER_INST_NAME" | tr '[:upper:]' '[:lower:]')
: ${PYNQ_BOARD="Pynq-Z1"}
: ${PYNQ_TARGET_DIR="/home/xilinx/$DOCKER_INST_NAME"}
: ${NUM_DEFAULT_WORKERS=1}
-
-# Absolute path to this script, e.g. /home/user/bin/foo.sh
-SCRIPT=$(readlink -f "$0")
-# Absolute path this script is in, thus /home/user/bin
-SCRIPTPATH=$(dirname "$SCRIPT")
+: ${FINN_SSH_KEY_DIR="$SCRIPTPATH/ssh_keys"}
BUILD_LOCAL=/tmp/$DOCKER_INST_NAME
VIVADO_HLS_LOCAL=$VIVADO_PATH
@@ -87,6 +88,7 @@ VIVADO_IP_CACHE=$BUILD_LOCAL/vivado_ip_cache
# ensure build dir exists locally
mkdir -p $BUILD_LOCAL
mkdir -p $VIVADO_IP_CACHE
+mkdir -p $FINN_SSH_KEY_DIR
gecho "Instance is named as $DOCKER_INST_NAME"
gecho "Mounting $BUILD_LOCAL into $BUILD_LOCAL"
@@ -133,6 +135,7 @@ docker run -t --rm --name $DOCKER_INST_NAME $DOCKER_INTERACTIVE --init \
-v $SCRIPTPATH:/workspace/finn \
-v $BUILD_LOCAL:$BUILD_LOCAL \
-v $VIVADO_PATH:$VIVADO_PATH \
+-v $FINN_SSH_KEY_DIR:/home/$DOCKER_UNAME/.ssh \
-e VIVADO_PATH=$VIVADO_PATH \
-e FINN_INST_NAME=$DOCKER_INST_NAME \
-e FINN_ROOT="/workspace/finn" \
diff --git a/src/finn/core/onnx_exec.py b/src/finn/core/onnx_exec.py
index c2f68a35076418e0cf2edb578bdb8d548772fc78..7c3123cd5eb29a54dc5cbfb912225ad3fdb0f219 100644
--- a/src/finn/core/onnx_exec.py
+++ b/src/finn/core/onnx_exec.py
@@ -39,6 +39,7 @@ from finn.core.remote_exec import remote_exec
from finn.core.rtlsim_exec import rtlsim_exec
from finn.custom_op.registry import getCustomOp
import finn.analysis.topology as ta
+from finn.util.basic import sanitize_quant_values, get_sanitize_quant_tensors
def execute_node(node, context, graph):
@@ -102,15 +103,14 @@ def execute_node(node, context, graph):
raise Exception(
"""Output shapes disagree after node execution:
found %s vs expected %s"""
- % (
- str(output_list[list_ind].shape),
- str(context[outp].shape),
- )
+ % (str(output_list[list_ind].shape), str(context[outp].shape))
)
context[outp] = output_list[list_ind]
-def execute_onnx(model, input_dict, return_full_exec_context=False):
+def execute_onnx(
+ model, input_dict, return_full_exec_context=False, start_node=None, end_node=None
+):
"""Executes given ONNX ModelWrapper with given named inputs.
If return_full_exec_context is False, a dict of named outputs is returned
@@ -118,7 +118,12 @@ def execute_onnx(model, input_dict, return_full_exec_context=False):
If return return_full_exec_context is True, the full set of tensors used by
the execution (including inputs, weights, activations and final outputs)
- will be returned as a dict."""
+ will be returned as a dict.
+
+ When start_node and end_node are set to None, the whole graph is executed.
+ If they are set to particular ONNX nodes, only the subgraph between (and
+ including) those nodes is executed.
+ """
if not model.check_all_tensor_shapes_specified():
raise Exception("Found unspecified tensor shapes, try infer_shapes")
@@ -161,8 +166,28 @@ def execute_onnx(model, input_dict, return_full_exec_context=False):
# execute the model node by node
# we can simply walk down the list since the ONNX spec guarantees that it is
# topologically sorted
- for node in graph.node:
+ subgraph = []
+ if start_node is None:
+ start_node = model.graph.node[0]
+ if end_node is None:
+ end_node = model.graph.node[-1]
+ # select the nodes between specified start/end nodes
+ start_ind = model.get_node_index(start_node)
+ end_ind = model.get_node_index(end_node) + 1
+ assert end_ind >= start_ind, "Start/end nodes must define valid subgraph"
+ subgraph = graph.node[start_ind:end_ind]
+ for node in subgraph:
+ if get_sanitize_quant_tensors() != 0:
+ # round input values to match quantization annotation
+ execution_context = sanitize_quant_values(
+ model, node.input, execution_context
+ )
execute_node(node, execution_context, graph)
+ if get_sanitize_quant_tensors() != 0:
+ # round output values to quantization annotation
+ execution_context = sanitize_quant_values(
+ model, node.output, execution_context
+ )
elif model_exec_mode == "remote_pynq":
# use remote exec metadata built into model to execute on a remote PYNQ
remote_exec(model, execution_context)
diff --git a/src/finn/custom_op/fpgadataflow/__init__.py b/src/finn/custom_op/fpgadataflow/__init__.py
index c4800011fe2c944fa877b12d0247795beda4a5e6..71c731f96ca45519c443a5f932ead050770e17de 100644
--- a/src/finn/custom_op/fpgadataflow/__init__.py
+++ b/src/finn/custom_op/fpgadataflow/__init__.py
@@ -173,9 +173,15 @@ class HLSCustomOp(CustomOp):
of the node as a dictionary."""
ret = dict()
ret["BRAM_18K"] = self.bram_estimation()
+ ret["BRAM_efficiency"] = self.bram_efficiency_estimation()
ret["LUT"] = self.lut_estimation()
return ret
+ def bram_efficiency_estimation(self):
+ """Function for BRAM efficiency estimation: actual parameter storage
+ needed divided by the allocated BRAM storage (from estimation)"""
+ return 1
+
def bram_estimation(self):
"""Function for BRAM resource estimation, is member function of
HLSCustomOp class but has to be filled by every node"""
diff --git a/src/finn/custom_op/fpgadataflow/fmpadding.py b/src/finn/custom_op/fpgadataflow/fmpadding_batch.py
similarity index 88%
rename from src/finn/custom_op/fpgadataflow/fmpadding.py
rename to src/finn/custom_op/fpgadataflow/fmpadding_batch.py
index fa321dfa65d14b67fa218fb6a49f602ddab8d57e..d326ae7dfc7830a0081c3b13233d67ef08b12eff 100644
--- a/src/finn/custom_op/fpgadataflow/fmpadding.py
+++ b/src/finn/custom_op/fpgadataflow/fmpadding_batch.py
@@ -21,6 +21,8 @@ class FMPadding_Batch(HLSCustomOp):
"Padding": ("i", True, 2),
# number of channels in input image
"NumChannels": ("i", True, 0),
+ # SIMD Input parallelism
+ "SIMD": ("i", False, 1),
# FINN input datatype
"inputDataType": ("s", True, ""),
# controls distribution of padded pixels
@@ -55,20 +57,22 @@ class FMPadding_Batch(HLSCustomOp):
return oshape
def get_folded_input_shape(self):
- # even though there is no folding in the current hlslib op,
- # insert a time multiplexing axis to remain compatible with the
- # shapes produced by the rest of the dataflow pipeline
- ret = list(self.get_normal_input_shape())
- ret.insert(-1, 1)
- return tuple(ret)
+ normal_ishape = list(self.get_normal_input_shape())
+ ifm_ch = self.get_nodeattr("NumChannels")
+ simd = self.get_nodeattr("SIMD")
+ assert ifm_ch % simd == 0, "SIMD must divide input channels"
+ fold = int(normal_ishape[-1] / simd)
+ folded_ishape = normal_ishape[:-1] + [fold, simd]
+ return tuple(folded_ishape)
def get_folded_output_shape(self):
- # even though there is no folding in the current hlslib op,
- # insert a time multiplexing axis to remain compatible with the
- # shapes produced by the rest of the dataflow pipeline
- ret = list(self.get_normal_output_shape())
- ret.insert(-1, 1)
- return tuple(ret)
+ normal_oshape = list(self.get_normal_output_shape())
+ ifm_ch = self.get_nodeattr("NumChannels")
+ simd = self.get_nodeattr("SIMD")
+ assert ifm_ch % simd == 0, "SIMD must divide input channels"
+ fold = int(normal_oshape[-1] / simd)
+ folded_oshape = normal_oshape[:-1] + [fold, simd]
+ return tuple(folded_oshape)
def make_shape_compatible_op(self, model):
exp_ishape = self.get_normal_input_shape()
@@ -114,15 +118,13 @@ class FMPadding_Batch(HLSCustomOp):
def get_instream_width(self):
ibits = self.get_input_datatype().bitwidth()
- num_ch = self.get_nodeattr("NumChannels")
-
- return ibits * num_ch
+ simd = self.get_nodeattr("SIMD")
+ return ibits * simd
def get_outstream_width(self):
obits = self.get_output_datatype().bitwidth()
- num_ch = self.get_nodeattr("NumChannels")
-
- return obits * num_ch
+ simd = self.get_nodeattr("SIMD")
+ return obits * simd
def get_number_output_values(self):
folded_oshape = self.get_folded_output_shape()
@@ -135,13 +137,15 @@ class FMPadding_Batch(HLSCustomOp):
self.code_gen_dict["$DEFINES$"] = [
"""#define ImgDim1 {}\n#define OutputDim1 {}\n
#define Padding1 {}\n#define NumChannels1 {}\n
- #define PaddingStyle1 {}\n#define numReps {}\n""".format(
+ #define PaddingStyle1 {}\n#define numReps {}
+ #define SIMD1 {}\n""".format(
self.get_nodeattr("ImgDim"),
self.get_padded_odim(),
self.get_nodeattr("Padding"),
self.get_nodeattr("NumChannels"),
self.get_nodeattr("PaddingStyle"),
self.get_nodeattr("numInputVectors"),
+ self.get_nodeattr("SIMD"),
)
]
@@ -176,7 +180,7 @@ class FMPadding_Batch(HLSCustomOp):
in_t = self.get_input_datatype().get_hls_datatype_str()
node = self.onnx_node
self.code_gen_dict["$DOCOMPUTE$"] = [
- """{}<ImgDim1, OutputDim1, Padding1, NumChannels1,
+ """{}<ImgDim1, OutputDim1, Padding1, NumChannels1,SIMD1,
{}, PaddingStyle1> (in0, out, numReps);""".format(
node.op_type, in_t
)
@@ -232,6 +236,7 @@ class FMPadding_Batch(HLSCustomOp):
node = self.onnx_node
exp_ishape = self.get_normal_input_shape()
exp_oshape = self.get_normal_output_shape()
+ folded_ishape = self.get_folded_input_shape()
folded_oshape = self.get_folded_output_shape()
if mode == "cppsim":
@@ -254,10 +259,8 @@ class FMPadding_Batch(HLSCustomOp):
match expected shape (1, ImgDim, ImgDim, NumChannels)."""
export_idt = self.get_input_datatype()
- # no reshaping for input since assuming no folding on input
- # make copy before saving array
- inp = inp.copy()
- np.save(os.path.join(code_gen_dir, "input_0.npy"), inp)
+ reshaped_input = inp.reshape(folded_ishape)
+ np.save(os.path.join(code_gen_dir, "input_0.npy"), reshaped_input)
if mode == "cppsim":
# execute the precompiled model
diff --git a/src/finn/custom_op/fpgadataflow/pool_batch.py b/src/finn/custom_op/fpgadataflow/pool_batch.py
new file mode 100644
index 0000000000000000000000000000000000000000..c7edc24d0e24eef1154293caca2519ab3aa68358
--- /dev/null
+++ b/src/finn/custom_op/fpgadataflow/pool_batch.py
@@ -0,0 +1,395 @@
+# 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.
+
+import os
+import numpy as np
+
+from finn.custom_op.fpgadataflow import HLSCustomOp
+from finn.core.datatype import DataType
+from onnx import TensorProto, helper
+from finn.util.data_packing import npy_to_rtlsim_input, rtlsim_output_to_npy
+
+
+class Pool_Batch(HLSCustomOp):
+ """Class that corresponds to finn-hlslib Pool_batch function.
+ Requires ConvolutionInputGenerator(depthwise == 1) to format its input
+
+ TODO: explain input shape (to reuse im2col code)
+ Input shape (BatchSize,OutImgDim,OutImgDim,KernelSize^2*Channels)
+ Output shape (BatchSize,OutImgDim,OutImgDim,Channels)
+
+ # note: the actual data layout produced by the hlslib kernels is different
+ # for depthwise ops.
+ # * depthwise SWG: (1, OFMDim, OFMDim, IFMChannels/PE, K, K, PE)
+
+ Channels can be folded using PE (SIMD from the input perspective)
+ TODO: doc
+ """
+
+ def get_nodeattr_types(self):
+ my_attrs = {
+ "Channels": ("i", True, 0),
+ "PE": ("i", True, 1),
+ "KernelSize": ("i", True, 0),
+ # Function:
+ # - MaxPool
+ # - AvgPool (not yet supported, but HLSLIB does)
+ # - AccPool (not yet supported, but HLSLIB does)
+ "Function": ("s", True, ""),
+ "OutImgDim": ("i", True, 0),
+ # FINN DataTypes for inputs/outputs
+ "dataType": ("s", True, ""),
+ "BatchSize": ("i", False, 1),
+ }
+
+ my_attrs.update(super().get_nodeattr_types())
+ return my_attrs
+
+ def get_input_datatype(self):
+ """Returns FINN DataType of input."""
+ return DataType[self.get_nodeattr("dataType")]
+
+ def get_output_datatype(self):
+ """Returns FINN DataType of output."""
+ fxn = self.get_nodeattr("Function")
+ if fxn == "MaxPool":
+ # Same as input
+ return DataType[self.get_nodeattr("dataType")]
+ else:
+ raise Exception("Pool_Batch doesn't currently support " + fxn)
+
+ def get_normal_input_shape(self):
+ ifm_ch = self.get_nodeattr("Channels")
+ odim = self.get_nodeattr("OutImgDim")
+ batch_size = self.get_nodeattr("BatchSize")
+ k = self.get_nodeattr("KernelSize")
+ ishape = (batch_size, odim, odim, k * k * ifm_ch)
+ return ishape
+
+ def get_folded_input_shape(self):
+ normal_ishape = list(self.get_normal_input_shape())
+ ifm_ch = self.get_nodeattr("Channels")
+ pe = self.get_nodeattr("PE")
+ assert ifm_ch % pe == 0, "PE must divide input channels"
+ fold = int(normal_ishape[-1] / pe)
+ folded_ishape = normal_ishape[:-1] + [fold, pe]
+ return tuple(folded_ishape)
+
+ def get_normal_output_shape(self):
+ ofm_ch = self.get_nodeattr("Channels")
+ odim = self.get_nodeattr("OutImgDim")
+ batch_size = self.get_nodeattr("BatchSize")
+ oshape = (batch_size, odim, odim, ofm_ch)
+ return oshape
+
+ def get_folded_output_shape(self):
+ normal_oshape = list(self.get_normal_output_shape())
+ ifm_ch = self.get_nodeattr("Channels")
+ pe = self.get_nodeattr("PE")
+ assert ifm_ch % pe == 0, "PE must divide input channels"
+ fold = int(ifm_ch / pe)
+ folded_oshape = normal_oshape[:-1] + [fold, pe]
+ return tuple(folded_oshape)
+
+ def get_number_output_values(self):
+ folded_oshape = self.get_folded_output_shape()
+ return np.prod(folded_oshape[1:-1])
+
+ def get_instream_width(self):
+ dt_bits = self.get_input_datatype().bitwidth()
+ pe = self.get_nodeattr("PE")
+ # ofm_ch = self.get_nodeattr("Channels")
+ # k = self.get_nodeattr("KernelSize")
+ # assert ifm_ch % pe == 0, "PE must divide input channels"
+ # simd = int(ifm_ch/pe)
+ in_width = int(dt_bits * pe)
+ return in_width
+
+ def get_outstream_width(self):
+ fxn = self.get_nodeattr("Function")
+ if fxn == "MaxPool":
+ return self.get_instream_width()
+ else:
+ raise Exception("Pool_Batch doesn't currently support " + fxn)
+
+ def make_shape_compatible_op(self, model):
+ exp_ishape = self.get_normal_input_shape()
+ oshape = self.get_normal_output_shape()
+ ishape = tuple(model.get_tensor_shape(self.onnx_node.input[0]))
+ assert ishape == exp_ishape, "Unexpected input shape for Pool_Batch."
+ # implement tensor with correct shape
+ values = np.random.randn(*oshape).astype(np.float32)
+ return helper.make_node(
+ "Constant",
+ inputs=[],
+ outputs=[self.onnx_node.output[0]],
+ value=helper.make_tensor(
+ name="const_tensor",
+ data_type=TensorProto.FLOAT,
+ dims=values.shape,
+ vals=values.flatten().astype(float),
+ ),
+ )
+
+ def infer_node_datatype(self, model):
+ node = self.onnx_node
+ # data type stays the same
+ dtype = self.get_output_datatype()
+ model.set_tensor_datatype(node.output[0], dtype)
+
+ def verify_node(self):
+ info_messages = []
+
+ # verify that "domain" is set to "finn"
+ domain_value = self.onnx_node.domain
+ if domain_value == "finn":
+ info_messages.append("Attribute domain is set correctly")
+ else:
+ info_messages.append('Attribute domain should be set to "finn"')
+
+ # verify that "backend" is set to "fpgadataflow"
+ backend_value = self.get_nodeattr("backend")
+ if backend_value == "fpgadataflow":
+ info_messages.append("Attribute backend is set correctly")
+ else:
+ info_messages.append('Attribute backend should be set to "fpgadataflow"')
+
+ # verify the number of inputs
+ if len(self.onnx_node.input) == 1:
+ info_messages.append("The number of inputs is correct")
+ else:
+ info_messages.append("""Pool_Batch needs 1 data input""")
+
+ # check supported function
+ fnx = self.get_nodeattr("Function")
+ if fnx == "MaxPool":
+ info_messages.append(
+ "Attribute Function contains a supported pool function"
+ )
+ else:
+ info_messages.append(
+ "Attribute Function contains an unsupported pool function"
+ )
+ return info_messages
+
+ def global_includes(self):
+ self.code_gen_dict["$GLOBALS$"] = ['#include "maxpool.h"']
+ self.code_gen_dict["$GLOBALS$"] += ['#include "pool.hpp"']
+
+ def defines(self, var):
+ self.code_gen_dict["$DEFINES$"] = []
+
+ ifm_ch = self.get_nodeattr("Channels")
+ self.code_gen_dict["$DEFINES$"] += ["#define Channels {}".format(ifm_ch)]
+
+ pe = self.get_nodeattr("PE")
+ self.code_gen_dict["$DEFINES$"] += ["#define PE {}".format(pe)]
+
+ k = self.get_nodeattr("KernelSize")
+ self.code_gen_dict["$DEFINES$"] += ["#define KernelSize {}".format(k)]
+
+ odim = self.get_nodeattr("OutImgDim")
+ self.code_gen_dict["$DEFINES$"] += ["#define OFMDim {}".format(odim)]
+
+ numReps = self.get_nodeattr("BatchSize")
+ self.code_gen_dict["$DEFINES$"] += ["#define numReps {}".format(numReps)]
+
+ def read_npy_data(self):
+ code_gen_dir = self.get_nodeattr("code_gen_dir_cppsim")
+ dtype = self.get_input_datatype()
+ if dtype == DataType.BIPOLAR:
+ # use binary for bipolar storage
+ dtype = DataType.BINARY
+ elem_bits = dtype.bitwidth()
+ packed_bits = self.get_instream_width()
+ packed_hls_type = "ap_uint<%d>" % packed_bits
+ elem_hls_type = dtype.get_hls_datatype_str()
+ npy_type = "float"
+ npy_in = "%s/input_0.npy" % code_gen_dir
+ self.code_gen_dict["$READNPYDATA$"] = []
+ self.code_gen_dict["$READNPYDATA$"].append(
+ 'npy2apintstream<%s, %s, %d, %s>("%s", in0,false);'
+ % (packed_hls_type, elem_hls_type, elem_bits, npy_type, npy_in)
+ )
+
+ def strm_decl(self):
+ self.code_gen_dict["$STREAMDECLARATIONS$"] = []
+ self.code_gen_dict["$STREAMDECLARATIONS$"].append(
+ 'hls::stream<ap_uint<{}>> in0 ("in0");'.format(self.get_instream_width())
+ )
+ self.code_gen_dict["$STREAMDECLARATIONS$"].append(
+ 'hls::stream<ap_uint<{}>> out ("out");'.format(self.get_outstream_width())
+ )
+
+ def docompute(self):
+ idt = self.get_input_datatype()
+ i_hls_dt = idt.get_hls_datatype_str()
+ odt = self.get_output_datatype()
+ o_hls_dt = odt.get_hls_datatype_str()
+
+ self.code_gen_dict["$DOCOMPUTE$"] = []
+
+ fxn = self.get_nodeattr("Function")
+ if fxn == "MaxPool":
+ self.code_gen_dict["$DOCOMPUTE$"] += [
+ "MaxPoolFunction<{},KernelSize> pool_fxn;".format(i_hls_dt)
+ ]
+ else:
+ raise Exception("Pool_Batch doesn't currently support " + fxn)
+
+ self.code_gen_dict["$DOCOMPUTE$"] += [
+ """Pool_batch<Channels, PE, KernelSize,Slice<{} >, Slice< {} > >
+ (in0,out, pool_fxn, OFMDim*OFMDim*numReps);""".format(
+ i_hls_dt, o_hls_dt
+ )
+ ]
+
+ def dataoutstrm(self):
+ code_gen_dir = self.get_nodeattr("code_gen_dir_cppsim")
+ dtype = self.get_output_datatype()
+ if dtype == DataType.BIPOLAR:
+ # use binary for bipolar storage
+ dtype = DataType.BINARY
+ elem_bits = dtype.bitwidth()
+ packed_bits = self.get_outstream_width()
+ packed_hls_type = "ap_uint<%d>" % packed_bits
+ elem_hls_type = dtype.get_hls_datatype_str()
+ npy_type = "float"
+ npy_out = "%s/output.npy" % code_gen_dir
+ oshape = self.get_folded_output_shape()
+ oshape_cpp_str = str(oshape).replace("(", "{").replace(")", "}")
+
+ self.code_gen_dict["$DATAOUTSTREAM$"] = [
+ 'apintstream2npy<%s, %s, %d, %s>(out, %s, "%s",false);'
+ % (
+ packed_hls_type,
+ elem_hls_type,
+ elem_bits,
+ npy_type,
+ oshape_cpp_str,
+ npy_out,
+ )
+ ]
+
+ def save_as_npy(self):
+ self.code_gen_dict["$SAVEASCNPY$"] = []
+
+ def blackboxfunction(self):
+ packed_ibits = self.get_instream_width()
+ packed_in_hls_type = "ap_uint<%d>" % packed_ibits
+
+ packed_obits = self.get_outstream_width()
+ packed_out_hls_type = "ap_uint<%d>" % packed_obits
+ self.code_gen_dict["$BLACKBOXFUNCTION$"] = [
+ "void %s(hls::stream<%s > &in0, hls::stream<%s > &out)"
+ % (self.onnx_node.name, packed_in_hls_type, packed_out_hls_type)
+ ]
+
+ def pragmas(self):
+ self.code_gen_dict["$PRAGMAS$"] = ["#pragma HLS INTERFACE axis port=in0"]
+ self.code_gen_dict["$PRAGMAS$"].append("#pragma HLS INTERFACE axis port=out")
+ self.code_gen_dict["$PRAGMAS$"].append(
+ "#pragma HLS INTERFACE ap_ctrl_none port=return"
+ )
+
+ def execute_node(self, context, graph):
+ mode = self.get_nodeattr("exec_mode")
+ node = self.onnx_node
+ exp_ishape = self.get_normal_input_shape()
+ folded_ishape = self.get_folded_input_shape()
+ exp_oshape = self.get_normal_output_shape()
+ folded_oshape = self.get_folded_output_shape()
+
+ # TODO ensure codegen dir exists
+ if mode == "cppsim":
+ code_gen_dir = self.get_nodeattr("code_gen_dir_cppsim")
+ elif mode == "rtlsim":
+ code_gen_dir = self.get_nodeattr("code_gen_dir_ipgen")
+ else:
+ raise Exception(
+ """Invalid value for attribute exec_mode! Is currently set to: {}
+ has to be set to one of the following value ("cppsim", "rtlsim")""".format(
+ mode
+ )
+ )
+
+ inp = context[node.input[0]]
+
+ assert str(inp.dtype) == "float32", "Input datatype is not float32"
+ assert (
+ inp.shape == exp_ishape
+ ), """Input shape doesn't
+ match expected shape (batch_size,odim,odim,k*k*ifm_ch)."""
+
+ export_idt = self.get_input_datatype()
+ reshaped_input = inp.reshape(folded_ishape)
+
+ np.save(os.path.join(code_gen_dir, "input_0.npy"), reshaped_input)
+
+ if mode == "cppsim":
+ # execute the precompiled model
+ super().exec_precompiled_singlenode_model()
+ # load output npy file
+ super().npy_to_dynamic_output(context)
+ assert (
+ context[node.output[0]].shape == folded_oshape
+ ), "cppsim did not produce expected folded output shape"
+ context[node.output[0]] = context[node.output[0]].reshape(*exp_oshape)
+ elif mode == "rtlsim":
+ sim = self.get_rtlsim()
+ nbits = self.get_instream_width()
+ rtlsim_inp = npy_to_rtlsim_input(
+ "{}/input_0.npy".format(code_gen_dir), export_idt, nbits
+ )
+ super().reset_rtlsim(sim)
+ super().toggle_clk(sim)
+ rtlsim_output = self.rtlsim(sim, rtlsim_inp)
+ odt = export_idt
+ target_bits = odt.bitwidth()
+ packed_bits = self.get_outstream_width()
+ out_npy_path = "{}/output.npy".format(code_gen_dir)
+ out_shape = self.get_folded_output_shape()
+ rtlsim_output_to_npy(
+ rtlsim_output, out_npy_path, odt, out_shape, packed_bits, target_bits
+ )
+ # load and reshape output
+ output = np.load(out_npy_path)
+ output = np.asarray([output], dtype=np.float32).reshape(*exp_oshape)
+ context[node.output[0]] = output
+ else:
+ raise Exception(
+ """Invalid value for attribute exec_mode! Is currently set to: {}
+ has to be set to one of the following value ("cppsim", "rtlsim")""".format(
+ mode
+ )
+ )
+
+ assert (
+ context[node.output[0]].shape == exp_oshape
+ ), """Output
+ shape doesn't match expected shape (1, ofm_dim, ofm_dim, k*k*ifm_ch)."""
diff --git a/src/finn/custom_op/fpgadataflow/streamingfclayer_batch.py b/src/finn/custom_op/fpgadataflow/streamingfclayer_batch.py
index 9b73ba1e100aa83fd19aa8799195c99891fca3fd..a7ebff68749120868cae9ce5ac18d2856fe2cb8a 100644
--- a/src/finn/custom_op/fpgadataflow/streamingfclayer_batch.py
+++ b/src/finn/custom_op/fpgadataflow/streamingfclayer_batch.py
@@ -240,11 +240,21 @@ class StreamingFCLayer_Batch(HLSCustomOp):
Q = self.get_nodeattr("SIMD")
wdt = self.get_weight_datatype()
W = wdt.bitwidth()
- D_in = self.get_instream_width()
- D_out = self.get_outstream_width()
+ D_in = self.get_nodeattr("MW")
+ D_out = self.get_nodeattr("MH")
omega = (D_in * D_out) / (Q * P)
return P * (math.ceil(omega / 512)) * (math.ceil((Q * W) / 36))
+ def bram_efficiency_estimation(self):
+ wdt = self.get_weight_datatype()
+ W = wdt.bitwidth()
+ D_in = self.get_nodeattr("MW")
+ D_out = self.get_nodeattr("MH")
+ bram16_est = self.bram_estimation()
+ wbits = W * D_in * D_out
+ bram16_est_capacity = bram16_est * 36 * 512
+ return wbits / bram16_est_capacity
+
def lut_estimation(self):
"""Calculates resource estimations for LUTs based on:
- FINN-R: An End-to-End Deep-Learning Framework for Fast
@@ -290,12 +300,15 @@ class StreamingFCLayer_Batch(HLSCustomOp):
return out_width
def get_weightstream_width(self):
- """Returns weight stream width. Used in decoupled mode."""
- pe = self.get_nodeattr("PE")
- simd = self.get_nodeattr("SIMD")
- wp = self.get_weight_datatype().bitwidth()
- w_width = pe * simd * wp
- return w_width
+ """Returns weight stream width. Used only in decoupled mode."""
+ if self.get_nodeattr("mem_mode") == "decoupled":
+ pe = self.get_nodeattr("PE")
+ simd = self.get_nodeattr("SIMD")
+ wp = self.get_weight_datatype().bitwidth()
+ w_width = pe * simd * wp
+ return w_width
+ else:
+ return 0
def get_weightstream_width_padded(self):
"""Returns weight stream width padded to a multiple of 8. This is required
diff --git a/src/finn/custom_op/fpgadataflow/templates.py b/src/finn/custom_op/fpgadataflow/templates.py
index 1a8216f64bf71b7fb9f1f8becf4732970b5bf451..1da60a5124fa86b4336bae8fd1a587672f2f2e6f 100644
--- a/src/finn/custom_op/fpgadataflow/templates.py
+++ b/src/finn/custom_op/fpgadataflow/templates.py
@@ -99,6 +99,7 @@ set_top $config_toplevelfxn
open_solution sol1
set_part $config_proj_part
+config_compile -ignore_long_run_time -disable_unroll_code_size_check
config_interface -m_axi_addr64
config_rtl -auto_prefix
$EXTRA_DIRECTIVES$
diff --git a/src/finn/custom_op/fpgadataflow/tlastmarker.py b/src/finn/custom_op/fpgadataflow/tlastmarker.py
index 25ea05e3607a52731ae1b64de421837bf137ee2b..17ba44b959577faf573d77ae222f7b2a3be6669d 100644
--- a/src/finn/custom_op/fpgadataflow/tlastmarker.py
+++ b/src/finn/custom_op/fpgadataflow/tlastmarker.py
@@ -30,20 +30,30 @@ from finn.custom_op.fpgadataflow import HLSCustomOp
class TLastMarker(HLSCustomOp):
- """Class that corresponds to the TLastMarker node that needs to be
- inserted at the end of the model for rtlsim with stitched IP.
- It marks the end of the current image/input sample."""
+ """Node that adds/removes AXI stream TLAST signals where needed. Its behavior
+ is transparent in node-by-node execution, only visible in IP-stitched rtlsim or
+ actual hardware.
+ This node may be needed at the end of the network to signal a DMA write (needed by the
+ FINN PYNQ shell) or at the beginning to remove the end-of-burst from DMA read."""
def __init__(self, onnx_node):
super().__init__(onnx_node)
def get_nodeattr_types(self):
my_attrs = {
+ # number of (static) iterations until TLAST=1 is generated for Direction=out
"NumIters": ("i", True, 0),
+ # whether static or dynamic (from AXI lite) number of iterations are used
+ "DynIters": ("i", False, 1),
+ # direction: whether to insert or remove TLAST
+ "Direction": ("s", False, "out"),
# width of input-output data streams, in bits
"StreamWidth": ("i", True, 0),
# width of individual element in stream, in bits
"ElemWidth": ("i", True, 0),
+ # Protocol: external or internal
+ # Vitis docs recommend using qdma_axis for external, ap_axiu for internal
+ "Protocol": ("s", False, "external"),
}
my_attrs.update(super().get_nodeattr_types())
return my_attrs
@@ -76,12 +86,33 @@ class TLastMarker(HLSCustomOp):
def defines(self, var):
stream_width = self.get_nodeattr("StreamWidth")
+ direction = self.get_nodeattr("Direction")
+ protocol = self.get_nodeattr("Protocol")
# output stream must have TLAST, so we use this stream data type:
# qdma_axis<stream_data_width,0,0,0 >
- out_stream_dtype = "qdma_axis<%d,0,0,0>" % stream_width
+ if direction == "out":
+ if protocol == "external":
+ out_stream_dtype = "qdma_axis<%d,0,0,0>" % stream_width
+ elif protocol == "internal":
+ out_stream_dtype = "ap_axiu<%d,0,0,0>" % stream_width
+ else:
+ raise Exception("Unrecognized Protocol in TLastMarker")
+ in_stream_dtype = "ap_uint<%d>" % stream_width
+ elif direction == "in":
+ out_stream_dtype = "ap_uint<%d>" % stream_width
+ if protocol == "external":
+ in_stream_dtype = "qdma_axis<%d,0,0,0>" % stream_width
+ elif protocol == "internal":
+ in_stream_dtype = "ap_axiu<%d,0,0,0>" % stream_width
+ else:
+ raise Exception("Unrecognized Protocol in TLastMarker")
+ else:
+ raise Exception("Unrecognized Direction in TLastMarker")
+
self.code_gen_dict["$DEFINES$"] = [
"#define StreamWidth %d" % stream_width,
"#define OutDType %s" % out_stream_dtype,
+ "#define InDType %s" % in_stream_dtype,
"#define NumItersPerImg %d" % self.get_nodeattr("NumIters"),
]
@@ -89,27 +120,60 @@ class TLastMarker(HLSCustomOp):
self.code_gen_dict["$READNPYDATA$"] = []
def docompute(self):
- self.code_gen_dict["$DOCOMPUTE$"] = [
- "unsigned int n = 1;",
- "OutDType t;",
- "t.set_keep(-1);",
- "io_section: { // start of cycle accurate region",
- "#pragma HLS protocol fixed",
- "// do a first read from stream before we decide on numIters",
- "// giving software a chance to set up the numIters prior to startup",
- "t.set_data(in0.read());",
- "n = (numIters == 0 ? NumItersPerImg : numIters);",
- "t.set_last(n==1);",
- "out.write(t);",
- "} // end of cycle accurate region",
- "// do one less iteration than spec since we already did one",
- "for(unsigned int i=1; i<n; i++) {",
- "#pragma HLS PIPELINE II=1",
- "t.set_data(in0.read());",
- "t.set_last(i==(n-1));",
- "out.write(t);",
- "}",
- ]
+ dyn_iters = self.get_nodeattr("DynIters")
+ direction = self.get_nodeattr("Direction")
+ use_qdma_axis = self.get_nodeattr("Protocol") == "external"
+ if direction == "in":
+ # read from input and just pass data along; ignore tlast
+ # no dyn iters on input, it doesnt make sense
+ self.code_gen_dict["$DOCOMPUTE$"] = [
+ "for(unsigned int i=0; i<NumItersPerImg; i++) {",
+ "#pragma HLS PIPELINE II=1",
+ "out.write(in0.read().get_data());"
+ if use_qdma_axis
+ else "out.write(in0.read().data);",
+ "}",
+ ]
+
+ elif dyn_iters == 1:
+ # output, with dynamic iteration counts
+ self.code_gen_dict["$DOCOMPUTE$"] = [
+ "unsigned int n = 1;",
+ "OutDType t;",
+ "t.set_keep(-1);" if use_qdma_axis else "t.keep = -1;",
+ "io_section: { // start of cycle accurate region",
+ "#pragma HLS protocol fixed",
+ "// do a first read from stream before we decide on numIters",
+ "// giving software a chance to set up the numIters prior to startup",
+ "t.set_data(in0.read());" if use_qdma_axis else "t.data = in0.read();",
+ "n = (numIters == 0 ? NumItersPerImg : numIters);",
+ "t.set_last(n==1);" if use_qdma_axis else "t.last = (n==1);",
+ "out.write(t);",
+ "} // end of cycle accurate region",
+ "// do one less iteration than spec since we already did one",
+ "for(unsigned int i=1; i<n; i++) {",
+ "#pragma HLS PIPELINE II=1",
+ "t.set_data(in0.read());" if use_qdma_axis else "t.data = in0.read();",
+ "t.set_last(i==(n-1));" if use_qdma_axis else "t.last = (i==(n-1));",
+ "out.write(t);",
+ "}",
+ ]
+
+ else:
+ # output, with static iteration counts
+ self.code_gen_dict["$DOCOMPUTE$"] = [
+ "unsigned int n = 1;",
+ "OutDType t;",
+ "t.set_keep(-1);" if use_qdma_axis else "t.keep = -1;",
+ "for(unsigned int i=0; i<NumItersPerImg; i++) {",
+ "#pragma HLS PIPELINE II=1",
+ "t.set_data(in0.read());" if use_qdma_axis else "t.data = in0.read();",
+ "t.set_last(i==(NumItersPerImg-1));"
+ if use_qdma_axis
+ else "t.last = (i==(NumItersPerImg-1));",
+ "out.write(t);",
+ "}",
+ ]
def dataoutstrm(self):
self.code_gen_dict["$DATAOUTSTREAM$"] = []
@@ -118,18 +182,30 @@ class TLastMarker(HLSCustomOp):
self.code_gen_dict["$SAVEASCNPY$"] = []
def blackboxfunction(self):
- self.code_gen_dict["$BLACKBOXFUNCTION$"] = [
- """void %s(hls::stream<ap_uint<StreamWidth> > &in0,
- hls::stream<OutDType> &out, unsigned int numIters)"""
- % self.onnx_node.name
- ]
+ dyn_iters = self.get_nodeattr("DynIters")
+
+ if dyn_iters == 1:
+ self.code_gen_dict["$BLACKBOXFUNCTION$"] = [
+ """void %s(hls::stream<InDType> &in0,
+ hls::stream<OutDType> &out, unsigned int numIters)"""
+ % self.onnx_node.name
+ ]
+ else:
+ self.code_gen_dict["$BLACKBOXFUNCTION$"] = [
+ """void %s(hls::stream<InDType> &in0, hls::stream<OutDType> &out)"""
+ % self.onnx_node.name
+ ]
def pragmas(self):
self.code_gen_dict["$PRAGMAS$"] = ["#pragma HLS INTERFACE axis port=in0"]
self.code_gen_dict["$PRAGMAS$"].append("#pragma HLS INTERFACE axis port=out")
- self.code_gen_dict["$PRAGMAS$"].append(
- "#pragma HLS INTERFACE s_axilite port=numIters bundle=control"
- )
+
+ dyn_iters = self.get_nodeattr("DynIters")
+ if dyn_iters == 1:
+ self.code_gen_dict["$PRAGMAS$"].append(
+ "#pragma HLS INTERFACE s_axilite port=numIters bundle=control"
+ )
+
self.code_gen_dict["$PRAGMAS$"].append(
"#pragma HLS INTERFACE ap_ctrl_none port=return"
)
@@ -158,7 +234,7 @@ class TLastMarker(HLSCustomOp):
def strm_decl(self):
self.code_gen_dict["$STREAMDECLARATIONS$"] = []
self.code_gen_dict["$STREAMDECLARATIONS$"].append(
- 'hls::stream<ap_uint<{}>> in0 ("in0");'.format(self.get_instream_width())
+ 'hls::stream<InDType> in0 ("in0");'
)
self.code_gen_dict["$STREAMDECLARATIONS$"].append(
'hls::stream<OutDType> out ("out");'
diff --git a/src/finn/custom_op/im2col.py b/src/finn/custom_op/im2col.py
index 82a6b140f7af1be4e5c0f429d077b99c7865383e..8ed0041704d421dab587f08bcbcd9e739e8434e9 100644
--- a/src/finn/custom_op/im2col.py
+++ b/src/finn/custom_op/im2col.py
@@ -80,6 +80,8 @@ class Im2Col(CustomOp):
"input_shape": ("s", True, ""),
"pad_amount": ("i", False, 0),
"pad_value": ("i", False, 0),
+ # depthwise: if != 0, infer ConvolutionInputGenerator with depthwise == 1
+ "depthwise": ("i", False, 0),
}
def make_shape_compatible_op(self, model):
diff --git a/src/finn/custom_op/quantavgpool2d.py b/src/finn/custom_op/quantavgpool2d.py
new file mode 100644
index 0000000000000000000000000000000000000000..fb5c78bc0c8419ba519c5c3113d9b0c7ae2dd3b7
--- /dev/null
+++ b/src/finn/custom_op/quantavgpool2d.py
@@ -0,0 +1,128 @@
+import numpy as np
+from onnx import TensorProto, helper
+import onnxruntime as rt
+
+from finn.custom_op import CustomOp
+from finn.core.datatype import DataType
+from finn.custom_op.maxpoolnhwc import compute_pool_output_dim
+
+
+class QuantAvgPool2d(CustomOp):
+ """Class that corresponds to the quantized average pooling
+ layer from brevitas"""
+
+ def get_nodeattr_types(self):
+ return {
+ "stride": ("i", True, 1),
+ "kernel": ("i", True, 1),
+ "ibits": ("i", True, 1),
+ "obits": ("i", True, 1),
+ # determines if values are signed (set to "1") or unsigned ("0")
+ "signed": ("i", True, 0),
+ # data layout attribute can be set to "NCHW" or "NHWC"
+ "data_layout": ("s", False, "NCHW"),
+ }
+
+ def make_shape_compatible_op(self, model):
+ node = self.onnx_node
+ k = self.get_nodeattr("kernel")
+ s = self.get_nodeattr("stride")
+ data_layout = self.get_nodeattr("data_layout")
+ if data_layout == "NCHW":
+ return helper.make_node(
+ "AveragePool",
+ inputs=[node.input[0]],
+ outputs=[node.output[0]],
+ kernel_shape=[k, k],
+ strides=[s, s],
+ )
+ elif data_layout == "NHWC":
+ iname = node.input[0]
+ ishape = model.get_tensor_shape(iname)
+ (n, hi, wi, c) = ishape
+ ho = compute_pool_output_dim(hi, k, s)
+ wo = compute_pool_output_dim(wi, k, s)
+ oshape = (n, ho, wo, c)
+ # implement tensor with correct shape
+ values = np.random.randn(*oshape).astype(np.float32)
+ return helper.make_node(
+ "Constant",
+ inputs=[],
+ outputs=[node.output[0]],
+ value=helper.make_tensor(
+ name="const_tensor",
+ data_type=TensorProto.FLOAT,
+ dims=values.shape,
+ vals=values.flatten().astype(float),
+ ),
+ )
+
+ else:
+ raise Exception(
+ """Datalayout for QuantAvgPool2d is set to an invalid value.
+ Has to be set to "NCHW" or "NHWC"."""
+ )
+
+ def infer_node_datatype(self, model):
+ node = self.onnx_node
+ bw = self.get_nodeattr("obits")
+ if bw in [2, 4, 8, 16, 32]:
+ if self.get_nodeattr("signed") == 0:
+ dtype = DataType["UINT%d" % bw]
+ else:
+ dtype = DataType["INT%d" % bw]
+ else:
+ raise Exception("Unsupported output datatype for QuantAvgPool2d")
+ model.set_tensor_datatype(node.output[0], dtype)
+
+ def execute_node(self, context, graph):
+ # create a standard average pooling node to help calculate the result
+ node = self.onnx_node
+ k = self.get_nodeattr("kernel")
+ s = self.get_nodeattr("stride")
+ inp_values = context[node.input[0]]
+ oshape = context[node.output[0]].shape
+ if self.get_nodeattr("data_layout") == "NHWC":
+ inp_values = inp_values.transpose(0, 3, 1, 2)
+ oshape = (context[node.output[0]]).transpose(0, 3, 1, 2).shape
+ ishape = inp_values.shape
+ inp = helper.make_tensor_value_info(node.input[0], TensorProto.FLOAT, ishape)
+ outp = helper.make_tensor_value_info(node.output[0], TensorProto.FLOAT, oshape)
+ node_avgpool = helper.make_node(
+ "AveragePool",
+ inputs=[node.input[0]],
+ outputs=[node.output[0]],
+ kernel_shape=[k, k],
+ strides=[s, s],
+ )
+ graph_avgpool = helper.make_graph(
+ nodes=[node_avgpool],
+ name="single-avgpool-exec",
+ inputs=[inp],
+ outputs=[outp],
+ )
+ model_avgpool = helper.make_model(graph_avgpool)
+ idict = {node.input[0]: inp_values}
+ sess = rt.InferenceSession(model_avgpool.SerializeToString())
+ result_temp = sess.run(None, idict)
+ # remove scaling introduced by average
+ result_temp = result_temp[0] * (k * k)
+ ibits = self.get_nodeattr("ibits")
+ max_value = 2 ** ibits - 1
+ max_value = max_value * k * k
+ max_bit_width = int(max_value).bit_length()
+ shift_bits = max_bit_width - self.get_nodeattr("obits")
+ result = np.right_shift(result_temp.astype(int), shift_bits)
+ if self.get_nodeattr("data_layout") == "NHWC":
+ result = result.transpose(0, 2, 3, 1)
+ context[node.output[0]] = result.astype(np.float32)
+
+ def verify_node(self):
+ info_messages = []
+ # verify that "domain" is set to "finn"
+ domain_value = self.onnx_node.domain
+ if domain_value == "finn":
+ info_messages.append("Attribute domain is set correctly")
+ else:
+ info_messages.append('Attribute domain should be set to "finn"')
+ return info_messages
diff --git a/src/finn/custom_op/registry.py b/src/finn/custom_op/registry.py
index 614a3d7ffd70d0b102bad2b76177a2d3b32765c7..0060e5d400f30055d532671c8cf1680f0668442a 100644
--- a/src/finn/custom_op/registry.py
+++ b/src/finn/custom_op/registry.py
@@ -44,10 +44,12 @@ from finn.custom_op.fpgadataflow.streamingdatawidthconverter_batch import (
StreamingDataWidthConverter_Batch,
)
from finn.custom_op.fpgadataflow.globalaccpool_batch import GlobalAccPool_Batch
-from finn.custom_op.fpgadataflow.fmpadding import FMPadding_Batch
+from finn.custom_op.fpgadataflow.pool_batch import Pool_Batch
+from finn.custom_op.fpgadataflow.fmpadding_batch import FMPadding_Batch
from finn.custom_op.fpgadataflow.thresholding_batch import Thresholding_Batch
from finn.custom_op.fpgadataflow.addstreams_batch import AddStreams_Batch
from finn.custom_op.fpgadataflow.labelselect_batch import LabelSelect_Batch
+from finn.custom_op.quantavgpool2d import QuantAvgPool2d
from finn.custom_op.fpgadataflow.duplicatestreams_batch import DuplicateStreams_Batch
# create a mapping of all known CustomOp names and classes
@@ -65,10 +67,12 @@ custom_op["MaxPoolNHWC"] = MaxPoolNHWC
custom_op["StreamingDataWidthConverter_Batch"] = StreamingDataWidthConverter_Batch
custom_op["StreamingFIFO"] = StreamingFIFO
custom_op["GlobalAccPool_Batch"] = GlobalAccPool_Batch
+custom_op["Pool_Batch"] = Pool_Batch
custom_op["FMPadding_Batch"] = FMPadding_Batch
custom_op["Thresholding_Batch"] = Thresholding_Batch
custom_op["AddStreams_Batch"] = AddStreams_Batch
custom_op["LabelSelect_Batch"] = LabelSelect_Batch
+custom_op["QuantAvgPool2d"] = QuantAvgPool2d
custom_op["DuplicateStreams_Batch"] = DuplicateStreams_Batch
diff --git a/src/finn/transformation/change_datalayout.py b/src/finn/transformation/change_datalayout.py
new file mode 100644
index 0000000000000000000000000000000000000000..d5b393a25e57122b059a44f70904a6dbe5bbaa3f
--- /dev/null
+++ b/src/finn/transformation/change_datalayout.py
@@ -0,0 +1,110 @@
+# 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 onnx import helper, TensorProto
+
+from finn.transformation import Transformation
+from finn.transformation.infer_shapes import InferShapes
+from finn.util.basic import get_by_name
+
+
+class ChangeDataLayoutQuantAvgPool2d(Transformation):
+ """Replace QuantAvgPool2d with datalayout (N,C,H,W) with Transpose nodes
+ and QuantAvgPool2dNHWC with datalayout (N,H,W,C)"""
+
+ def apply(self, model):
+ graph = model.graph
+ node_ind = 0
+ graph_modified = False
+ for n in graph.node:
+ node_ind += 1
+ if n.op_type == "QuantAvgPool2d" and (
+ get_by_name(n.attribute, "data_layout") is None
+ or get_by_name(n.attribute, "data_layout").s.decode("UTF-8") == "NCHW"
+ ):
+ graph_modified = True
+ node_input = n.input[0]
+ node_output = n.output[0]
+ s = get_by_name(n.attribute, "stride").i
+ k = get_by_name(n.attribute, "kernel").i
+ ibits = get_by_name(n.attribute, "ibits").i
+ obits = get_by_name(n.attribute, "obits").i
+ signed = get_by_name(n.attribute, "signed").i
+ batchsize = model.get_tensor_shape(n.input[0])[0] # assume NCHW
+ channels = model.get_tensor_shape(n.input[0])[1] # assume NCHW
+ idim = model.get_tensor_shape(n.input[0])[-1] # assume NCHW
+ odim = model.get_tensor_shape(n.output[0])[-1] # assume NCHW
+
+ # create new nodes
+ # NCHW -> NHWC
+ # create new intermediate values
+ inp_trans_out = helper.make_tensor_value_info(
+ model.make_new_valueinfo_name(),
+ TensorProto.FLOAT,
+ (batchsize, idim, idim, channels), # NHWC
+ )
+ graph.value_info.append(inp_trans_out)
+ inp_trans_out = inp_trans_out.name
+ quantavg_out = helper.make_tensor_value_info(
+ model.make_new_valueinfo_name(),
+ TensorProto.FLOAT,
+ (batchsize, odim, odim, channels),
+ )
+ graph.value_info.append(quantavg_out)
+ quantavg_out = quantavg_out.name
+ inp_trans_node = helper.make_node(
+ "Transpose", [node_input], [inp_trans_out], perm=[0, 2, 3, 1]
+ )
+ quantavg_node = helper.make_node(
+ "QuantAvgPool2d",
+ [inp_trans_out],
+ [quantavg_out],
+ domain="finn",
+ stride=s,
+ kernel=k,
+ ibits=ibits,
+ obits=obits,
+ signed=signed,
+ data_layout="NHWC",
+ )
+ # NHWC -> NCHW
+ out_trans_node = helper.make_node(
+ "Transpose", [quantavg_out], [node_output], perm=[0, 3, 1, 2]
+ )
+ # insert nodes
+ graph.node.insert(node_ind, inp_trans_node)
+ graph.node.insert(node_ind + 1, quantavg_node)
+ graph.node.insert(node_ind + 2, out_trans_node)
+ # remove old nodes
+ graph.node.remove(n)
+
+ # set shapes
+ model.set_tensor_shape(inp_trans_out, (batchsize, idim, idim, channels))
+ model.set_tensor_shape(quantavg_out, (batchsize, odim, odim, channels))
+ model = model.transform(InferShapes())
+ return (model, graph_modified)
diff --git a/src/finn/transformation/fpgadataflow/annotate_resources.py b/src/finn/transformation/fpgadataflow/annotate_resources.py
index 207075b00de1871da19ea78472125d435449ed6e..62ee92df54eee2b63d84657515d7fbc3a8808b81 100644
--- a/src/finn/transformation/fpgadataflow/annotate_resources.py
+++ b/src/finn/transformation/fpgadataflow/annotate_resources.py
@@ -69,6 +69,9 @@ class AnnotateResources(Transformation):
total_dict[r_type] += r_amount
else:
total_dict[r_type] = r_amount
+ for k in total_dict.keys():
+ if "efficiency" in k:
+ total_dict[k] = total_dict[k] / len(graph.node)
model.set_metadata_prop("res_total_" + self.mode, str(total_dict))
for node in graph.node:
if _is_fpgadataflow_node(node) and node.name in res_dict.keys():
diff --git a/src/finn/transformation/fpgadataflow/convert_to_hls_layers.py b/src/finn/transformation/fpgadataflow/convert_to_hls_layers.py
index d421a5f3ef8ca980b399087de1482b2ae913da1b..b70b126680d650547cf376dd601c048c73a1cfd4 100644
--- a/src/finn/transformation/fpgadataflow/convert_to_hls_layers.py
+++ b/src/finn/transformation/fpgadataflow/convert_to_hls_layers.py
@@ -34,6 +34,7 @@ from finn.custom_op.registry import getCustomOp
from finn.transformation.infer_shapes import InferShapes
from finn.transformation.infer_datatypes import InferDataTypes
import finn.core.data_layout as DataLayout
+from finn.util.basic import get_by_name
class InferConvInpGen(Transformation):
@@ -56,6 +57,7 @@ class InferConvInpGen(Transformation):
k = i2c_inst.get_nodeattr("kernel_size")
pad = i2c_inst.get_nodeattr("pad_amount")
pad_val = i2c_inst.get_nodeattr("pad_value")
+ depthwise = i2c_inst.get_nodeattr("depthwise")
ifm_ch = i2c_in_shape[-1]
ifm_dim = i2c_in_shape[1]
ofm_dim = i2c_out_shape[1]
@@ -67,7 +69,11 @@ class InferConvInpGen(Transformation):
if pad > 0:
# if padding enabled, ensure pad_val supported by DataType
- assert dt.allowed(pad_val), "Im2Col DataType must support pad_val"
+ # assert dt.allowed(pad_val),"""FMPadding_Batch DataType
+ # must support pad_val"""
+ assert (
+ pad_val == 0
+ ), "FMPadding_Batch doesn't currently support pad_val!= 0"
odim_padding = ifm_dim + 2 * pad
@@ -112,6 +118,7 @@ class InferConvInpGen(Transformation):
Stride=stride,
inputDataType=dt.name,
outputDataType=dt.name,
+ depthwise=depthwise,
)
graph.node.insert(ConvInpGen_node_idx, ConvInpGen_node)
# remove old nodes
@@ -169,6 +176,137 @@ class InferStreamingMaxPool(Transformation):
return (model, graph_modified)
+class InferPool_Batch(Transformation):
+ """If kernel_shape > strides, replace Pool layer with with of Im2col
+ + pool(with kernel_shape == strides), plus Transpose layers to keep the original
+ data layout."""
+
+ def apply(self, model):
+ graph = model.graph
+ node_ind = 0
+ graph_modified = False
+ for n in graph.node:
+ node_ind += 1
+ if n.op_type in ["MaxPool"]:
+ # extract pool parameters
+ k = get_by_name(n.attribute, "kernel_shape").ints[-1]
+ stride = get_by_name(n.attribute, "strides").ints[-1]
+
+ if k <= stride:
+ continue
+
+ try:
+ pad = get_by_name(n.attribute, "pads").ints[-1]
+ except AttributeError:
+ pad = 0
+
+ node_input = n.input[0]
+ node_output = n.output[0]
+ idt = model.get_tensor_datatype(node_input)
+ if not idt.is_integer():
+ continue
+
+ # odt = model.get_tensor_datatype(node_output)
+
+ ifm_ch = model.get_tensor_shape(n.input[0])[1] # assume NCHW
+ ofm_ch = ifm_ch
+ ifm_dim = model.get_tensor_shape(n.input[0])[-1] # assume NCHW
+ ofm_dim = model.get_tensor_shape(n.output[0])[-1] # assume NCHW
+ # create new intermediate values
+ inp_trans_out = helper.make_tensor_value_info(
+ model.make_new_valueinfo_name(),
+ TensorProto.FLOAT,
+ (1, ifm_dim, ifm_dim, ifm_ch), # NHWC
+ )
+ graph.value_info.append(inp_trans_out)
+ inp_trans_out = inp_trans_out.name
+ model.set_tensor_datatype(inp_trans_out, idt)
+
+ im2col_out = helper.make_tensor_value_info(
+ model.make_new_valueinfo_name(),
+ TensorProto.FLOAT,
+ (1, ofm_dim, ofm_dim, ifm_ch * k * k),
+ )
+ graph.value_info.append(im2col_out)
+ im2col_out = im2col_out.name
+ model.set_tensor_datatype(im2col_out, idt)
+
+ pool_output = helper.make_tensor_value_info(
+ model.make_new_valueinfo_name(),
+ TensorProto.FLOAT,
+ (1, ofm_dim, ofm_dim, ofm_ch),
+ )
+ graph.value_info.append(pool_output)
+ pool_output = pool_output.name
+ # model.set_tensor_datatype(pool_output, odt)
+
+ # create new nodes
+ # NCHW -> NHWC
+ inp_trans_node = helper.make_node(
+ "Transpose", [node_input], [inp_trans_out], perm=[0, 2, 3, 1]
+ )
+
+ if n.op_type == "MaxPool":
+ pool_fxn = "MaxPool"
+ pad_value = idt.min()
+ else:
+ raise Exception(
+ "pad_value and pool_fxn not configured for {}".format(n.op_type)
+ )
+
+ # format input tensor
+ im2col_node = helper.make_node(
+ "Im2Col",
+ [inp_trans_out],
+ [im2col_out],
+ domain="finn",
+ stride=stride,
+ kernel_size=k,
+ pad_amount=pad,
+ pad_value=pad_value,
+ depthwise=1,
+ input_shape="(1,{},{},{})".format(ifm_dim, ifm_dim, ifm_ch),
+ )
+
+ # Warning PE has to be equal to ifm_ch until Im2Col is replaced by
+ # ConvolutionInputGenerator with depthwise=1.
+ # For other settings the output will be incorrect due to incorrect input
+ # data layout
+ pool_node = helper.make_node(
+ "Pool_Batch",
+ [im2col_out],
+ [pool_output],
+ domain="finn",
+ backend="fpgadataflow",
+ dataType=idt.name,
+ Channels=ifm_ch,
+ PE=ifm_ch,
+ KernelSize=k,
+ Function=pool_fxn,
+ OutImgDim=ofm_dim,
+ BatchSize=1,
+ )
+
+ # NHWC -> NCHW
+ out_trans_node = helper.make_node(
+ "Transpose", [pool_output], [node_output], perm=[0, 3, 1, 2]
+ )
+
+ # insert nodes where the conv is to preserve topological ordering
+ graph.node.insert(node_ind, inp_trans_node)
+ graph.node.insert(node_ind + 1, im2col_node)
+ graph.node.insert(node_ind + 2, pool_node)
+ graph.node.insert(node_ind + 3, out_trans_node)
+ # remove old node
+ graph.node.remove(n)
+ graph_modified = True
+
+ if graph_modified:
+ model = model.transform(InferShapes())
+ model = model.transform(InferDataTypes())
+ return (model, graph_modified)
+
+
class InferBinaryStreamingFCLayer(Transformation):
"""Convert XnorPopcountMatMul layers to
StreamingFCLayer_Batch layers. Any immediately following MultiThreshold
diff --git a/src/finn/transformation/fpgadataflow/insert_tlastmarker.py b/src/finn/transformation/fpgadataflow/insert_tlastmarker.py
index 32f32ece585a93465ba32fede45d5eb606a2b0a3..04dd437af27b9fbe18b2255c20a8e4acda03b3d0 100644
--- a/src/finn/transformation/fpgadataflow/insert_tlastmarker.py
+++ b/src/finn/transformation/fpgadataflow/insert_tlastmarker.py
@@ -31,23 +31,34 @@ from onnx import helper as oh
from finn.custom_op.registry import getCustomOp
from finn.transformation import Transformation
+from finn.util.basic import get_by_name
+
+import numpy as np
class InsertTLastMarker(Transformation):
- """Ensure that the graph is terminated with a TLastMarker node, inserting
- one if necessary."""
+ """Ensure that the graph is started/terminated with a TLastMarker node, inserting
+ one if necessary. Use constructor args to determine type of TLastMarker to be inserted.
+ More information available on the TLastMarker documentation.
+ """
- def __init__(self):
+ def __init__(self, both=False, external=True, dynamic=True):
super().__init__()
+ self.dyniters = dynamic
+ self.external = external
+ self.both = both
def apply(self, model):
# TODO only makes sense for a pure fpgadataflow graph -- check!
graph_out_name = model.graph.output[0].name
final_node = model.find_producer(graph_out_name)
- if final_node.op_type == "TLastMarker":
- # TODO maybe check the correctness of properties
- return (model, False)
- else:
+ graph_modified = False
+ if final_node.op_type != "TLastMarker" and not (
+ final_node.op_type == "IODMA"
+ and get_by_name(final_node.attribute, "direction").s.decode("UTF-8")
+ == "out"
+ ):
+
custom_op = getCustomOp(final_node)
num_iters = int(custom_op.get_number_output_values())
stream_width = int(custom_op.get_outstream_width())
@@ -69,8 +80,51 @@ class InsertTLastMarker(Transformation):
NumIters=num_iters,
StreamWidth=stream_width,
ElemWidth=elem_width,
+ DynIters=(1 if self.dyniters else 0),
+ Direction="out",
+ Protocol=("external" if self.external else "internal"),
domain="finn",
backend="fpgadataflow",
)
model.graph.node.append(tlast_node)
- return (model, True)
+ graph_modified = True
+ # if both is True, also insert marker on input
+ if self.both:
+ graph_in_name = model.graph.input[0].name
+ first_node = model.find_consumer(graph_in_name)
+ if first_node.op_type != "TLastMarker" and not (
+ first_node.op_type == "IODMA"
+ and get_by_name(first_node.attribute, "direction").s.decode("UTF-8")
+ == "in"
+ ):
+
+ custom_op = getCustomOp(first_node)
+ num_iters = np.prod(custom_op.get_folded_input_shape()[1:-1])
+ stream_width = int(custom_op.get_instream_width())
+ in_shape = model.get_tensor_shape(graph_in_name)
+ in_dtype = model.get_tensor_datatype(graph_in_name)
+ elem_width = in_dtype.bitwidth()
+ # make new buffer
+ first_node_in = oh.make_tensor_value_info(
+ model.make_new_valueinfo_name(), TensorProto.FLOAT, in_shape
+ )
+ model.graph.value_info.append(first_node_in)
+ model.set_tensor_datatype(first_node_in.name, in_dtype)
+ # reroute final node output to first_node_in_name
+ first_node.input[0] = first_node_in.name
+ tlast_node = oh.make_node(
+ "TLastMarker",
+ [graph_in_name],
+ [first_node_in.name],
+ NumIters=num_iters,
+ StreamWidth=stream_width,
+ ElemWidth=elem_width,
+ DynIters=(1 if self.dyniters else 0),
+ Direction="in",
+ Protocol=("external" if self.external else "internal"),
+ domain="finn",
+ backend="fpgadataflow",
+ )
+ model.graph.node.insert(0, tlast_node)
+ graph_modified = True
+ return (model, graph_modified)
diff --git a/src/finn/transformation/fpgadataflow/prepare_cppsim.py b/src/finn/transformation/fpgadataflow/prepare_cppsim.py
index a1524322ec03a4e96ef41f999144e3eed349c5af..6eae560e1191642cfaf85d92c6d0fcf644630973 100644
--- a/src/finn/transformation/fpgadataflow/prepare_cppsim.py
+++ b/src/finn/transformation/fpgadataflow/prepare_cppsim.py
@@ -29,9 +29,12 @@
import os
import finn.custom_op.registry as registry
-from finn.transformation import Transformation
from finn.util.basic import make_build_dir
from finn.util.fpgadataflow import is_fpgadataflow_node
+from finn.transformation import Transformation
+from finn.util.basic import get_num_default_workers
+import multiprocessing as mp
+import copy
def _codegen_single_node(node, model):
@@ -66,8 +69,39 @@ class PrepareCppSim(Transformation):
that contains generated C++ code that can be used to simulate node using cppsim.
The subsequent transformation is CompileCppSim"""
+ def __init__(self, num_workers=None):
+ super().__init__()
+ if num_workers is None:
+ self._num_workers = get_num_default_workers()
+ else:
+ self._num_workers = num_workers
+ assert self._num_workers >= 0, "Number of workers must be nonnegative."
+ if self._num_workers == 0:
+ self._num_workers = mp.cpu_count()
+
+ def prepareCppSim_node(self, node):
+ if is_fpgadataflow_node(node) is True:
+ _codegen_single_node(node, self.model)
+ return (node, False)
+
def apply(self, model):
- for node in model.graph.node:
- if is_fpgadataflow_node(node) is True:
- _codegen_single_node(node, model)
- return (model, False)
+ # Remove old nodes from the current model
+ self.model = copy.deepcopy(model)
+ old_nodes = []
+ for i in range(len(model.graph.node)):
+ old_nodes.append(model.graph.node.pop())
+
+ # Execute transformation in parallel
+ with mp.Pool(self._num_workers) as p:
+ new_nodes_and_bool = p.map(self.prepareCppSim_node, old_nodes, chunksize=1)
+
+ # extract nodes and check if the transformation needs to run again
+ # Note: .pop() had initially reversed the node order
+ run_again = False
+ for node, run in reversed(new_nodes_and_bool):
+ # Reattach new nodes to old model
+ model.graph.node.append(node)
+ if run is True:
+ run_again = True
+
+ return (model, run_again)
diff --git a/src/finn/transformation/infer_data_layouts.py b/src/finn/transformation/infer_data_layouts.py
index 9ac75578ffb911cc44cfddc2b2119b55e6abf2dd..e7a6b88239a1735d5379e165333f8356ae6f88a1 100644
--- a/src/finn/transformation/infer_data_layouts.py
+++ b/src/finn/transformation/infer_data_layouts.py
@@ -38,7 +38,7 @@ def _dims_to_layout(model, node, ndims):
return DataLayout.NC
else:
if node.domain == "finn":
- if node.op_type == "MultiThreshold":
+ if node.op_type == "MultiThreshold" or node.op_type == "QuantAvgPool2d":
mt_inst = registry.getCustomOp(node)
layout = mt_inst.get_nodeattr("data_layout")
if layout == "NHWC" and ndims == 4:
diff --git a/src/finn/transformation/infer_datatypes.py b/src/finn/transformation/infer_datatypes.py
index 1acd4e3abe2d77248810cf15c15475e806a3bd32..39b7a787be8c725e7b6d474757dd96fc4848dfe0 100644
--- a/src/finn/transformation/infer_datatypes.py
+++ b/src/finn/transformation/infer_datatypes.py
@@ -71,7 +71,13 @@ def _infer_node_datatype(model, node):
else:
# unknown, assume node produces float32 outputs
for o in node.output:
- model.set_tensor_datatype(o, DataType.FLOAT32)
+ # check if output datatype is already set to a value != FLOAT32
+ odtype = model.get_tensor_datatype(o)
+ if odtype is not None and odtype != DataType.FLOAT32:
+ # don't change data type
+ model.set_tensor_datatype(o, odtype)
+ else:
+ model.set_tensor_datatype(o, DataType.FLOAT32)
# compare old and new output dtypes to see if anything changed
new_odtypes = list(map(lambda x: model.get_tensor_datatype(x), node.output))
graph_modified = new_odtypes != odtypes
diff --git a/src/finn/transformation/lower_convs_to_matmul.py b/src/finn/transformation/lower_convs_to_matmul.py
index 3da785d8dd21b2c6701bffc8ce3869fb14b237a9..aa231a43a3865a161a501b4997ff2f538800554f 100644
--- a/src/finn/transformation/lower_convs_to_matmul.py
+++ b/src/finn/transformation/lower_convs_to_matmul.py
@@ -80,14 +80,19 @@ class LowerConvsToMatMul(Transformation):
inp_trans_out = inp_trans_out.name
model.set_tensor_datatype(inp_trans_out, idt)
- im2col_out = helper.make_tensor_value_info(
- model.make_new_valueinfo_name(),
- TensorProto.FLOAT,
- (1, ofm_dim, ofm_dim, ifm_ch * k * k),
- )
- graph.value_info.append(im2col_out)
- im2col_out = im2col_out.name
- model.set_tensor_datatype(im2col_out, idt)
+ need_im2col = True
+ if k == 1 and pad == 0 and stride == 1:
+ need_im2col = False
+
+ if need_im2col:
+ im2col_out = helper.make_tensor_value_info(
+ model.make_new_valueinfo_name(),
+ TensorProto.FLOAT,
+ (1, ofm_dim, ofm_dim, ifm_ch * k * k),
+ )
+ graph.value_info.append(im2col_out)
+ im2col_out = im2col_out.name
+ model.set_tensor_datatype(im2col_out, idt)
matmul_out = helper.make_tensor_value_info(
model.make_new_valueinfo_name(),
@@ -104,19 +109,23 @@ class LowerConvsToMatMul(Transformation):
"Transpose", [cnv_input], [inp_trans_out], perm=[0, 2, 3, 1]
)
# lower input tensor
- im2col_node = helper.make_node(
- "Im2Col",
- [inp_trans_out],
- [im2col_out],
- domain="finn",
- stride=stride,
- kernel_size=k,
- pad_amount=pad,
- input_shape="(1,{},{},{})".format(ifm_dim, ifm_dim, ifm_ch),
- )
+ matmul_input = inp_trans_out
+ if need_im2col:
+ matmul_input = im2col_out
+ im2col_node = helper.make_node(
+ "Im2Col",
+ [inp_trans_out],
+ [im2col_out],
+ domain="finn",
+ stride=stride,
+ kernel_size=k,
+ pad_amount=pad,
+ input_shape="(1,{},{},{})".format(ifm_dim, ifm_dim, ifm_ch),
+ )
+
# do matmul
matmul_node = helper.make_node(
- "MatMul", [im2col_out, weight_name], [matmul_out]
+ "MatMul", [matmul_input, weight_name], [matmul_out]
)
# NHWC -> NCHW
out_trans_node = helper.make_node(
@@ -124,9 +133,13 @@ class LowerConvsToMatMul(Transformation):
)
# insert nodes where the conv is to preserve topological ordering
graph.node.insert(node_ind, inp_trans_node)
- graph.node.insert(node_ind + 1, im2col_node)
- graph.node.insert(node_ind + 2, matmul_node)
- graph.node.insert(node_ind + 3, out_trans_node)
+ if need_im2col:
+ graph.node.insert(node_ind + 1, im2col_node)
+ graph.node.insert(node_ind + 2, matmul_node)
+ graph.node.insert(node_ind + 3, out_trans_node)
+ else:
+ graph.node.insert(node_ind + 1, matmul_node)
+ graph.node.insert(node_ind + 2, out_trans_node)
# remove old nodes
graph.node.remove(n)
model = model.transform(InferShapes())
diff --git a/src/finn/transformation/remove_identity.py b/src/finn/transformation/remove_identity.py
new file mode 100644
index 0000000000000000000000000000000000000000..d7a58d59c1bb8ff643e691442e7eda3c0516aa5c
--- /dev/null
+++ b/src/finn/transformation/remove_identity.py
@@ -0,0 +1,62 @@
+# 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.transformation import Transformation
+
+
+def _is_identity(node, model):
+ if node.op_type == "Mul":
+ scale = model.get_initializer(node.input[1])
+ if scale is not None:
+ return (scale == 1).all()
+ elif node.op_type == "Add":
+ bias = model.get_initializer(node.input[1])
+ if bias is not None:
+ return (bias == 0).all()
+ return False
+
+
+class RemoveIdentity(Transformation):
+ """Remove nodes that apply identity ops from the graph, including:
+ * Multiply by 1
+ * Add 0
+ ."""
+
+ def apply(self, model):
+ graph = model.graph
+ node_ind = 0
+ graph_modified = False
+ for node in graph.node:
+ node_ind += 1
+ if _is_identity(node, model):
+ node_src = node.input[0]
+ node_dst = node.output[0]
+ graph.node.remove(node)
+ model.rename_tensor(node_dst, node_src)
+ graph_modified = True
+ return (model, graph_modified)
diff --git a/src/finn/transformation/streamline/__init__.py b/src/finn/transformation/streamline/__init__.py
index c9c73fa4c8303ee28bc1cc6aee879d633740e01e..d9c12a20975084705b801c0ff027d4b99aff9490 100644
--- a/src/finn/transformation/streamline/__init__.py
+++ b/src/finn/transformation/streamline/__init__.py
@@ -41,6 +41,7 @@ from finn.transformation.streamline.absorb import (
FactorOutMulSignMagnitude,
Absorb1BitMulIntoMatMul,
Absorb1BitMulIntoConv,
+ AbsorbSignBiasIntoMultiThreshold,
)
from finn.transformation.streamline.collapse_repeated import (
@@ -59,6 +60,7 @@ from finn.transformation.streamline.reorder import (
from finn.transformation.streamline.round_thresholds import RoundAndClipThresholds
from finn.transformation.streamline.sign_to_thres import ConvertSignToThres
from finn.transformation.batchnorm_to_affine import BatchNormToAffine
+from finn.transformation.streamline.remove import RemoveIdentityOps
class Streamline(Transformation):
@@ -70,6 +72,7 @@ class Streamline(Transformation):
ConvertDivToMul(),
BatchNormToAffine(),
ConvertSignToThres(),
+ AbsorbSignBiasIntoMultiThreshold(),
MoveAddPastMul(),
MoveScalarAddPastMatMul(),
MoveScalarAddPastConv(),
@@ -87,6 +90,7 @@ class Streamline(Transformation):
]
for trn in streamline_transformations:
model = model.transform(trn)
+ model = model.transform(RemoveIdentityOps())
model = model.transform(GiveUniqueNodeNames())
model = model.transform(GiveReadableTensorNames())
model = model.transform(InferDataTypes())
diff --git a/src/finn/transformation/streamline/absorb.py b/src/finn/transformation/streamline/absorb.py
index dbcf97361017144174f9fbfca35a84361b5abd26..dc01eea411fc1f640e481c9be02a92acdd59533f 100644
--- a/src/finn/transformation/streamline/absorb.py
+++ b/src/finn/transformation/streamline/absorb.py
@@ -28,14 +28,80 @@
import numpy as np
from onnx import helper as oh
+import warnings
from finn.core.datatype import DataType
from finn.transformation import Transformation
from finn.util.basic import get_by_name
from finn.custom_op.registry import getCustomOp
+from finn.transformation.infer_shapes import InferShapes
from finn.transformation.infer_datatypes import InferDataTypes
+class AbsorbSignBiasIntoMultiThreshold(Transformation):
+ """Absorb scalar bias originating from signed int export back into
+ MultiThreshold and re-evaluate the output datatype."""
+
+ def apply(self, model):
+ graph = model.graph
+ node_ind = 0
+ graph_modified = False
+ for n in graph.node:
+ # search for (MultiThreshold, Add) pair
+ node_ind += 1
+ if (
+ n.op_type == "MultiThreshold"
+ and not model.is_fork_node(n)
+ and not model.is_join_node(n)
+ ):
+ consumer = model.find_consumer(n.output[0])
+ if consumer is not None and consumer.op_type == "Add":
+ mt_node = n
+ add_node = consumer
+ threshold_name = mt_node.input[1]
+ add_weight_name = add_node.input[1]
+ T = model.get_initializer(threshold_name)
+ A = model.get_initializer(add_weight_name)
+ if (A is None) or (T is None):
+ warnings.warn("Threshold or add bias not constant, skipping")
+ continue
+ end_name = add_node.output[0]
+ # we can only absorb scalar adds
+ is_scalar = A.ndim == 0 or all(x == 1 for x in A.shape)
+ if not is_scalar:
+ continue
+ bias = A.flatten()[0]
+ # set MultiThreshold bias property
+ mt_inst = getCustomOp(mt_node)
+ bias += mt_inst.get_nodeattr("out_bias")
+ mt_inst.set_nodeattr("out_bias", bias)
+ graph_modified = True
+ # compute new DataType for MultiThreshold output
+ steps = T.shape[-1]
+ new_min = bias
+ new_max = steps + bias
+ odt = DataType.get_smallest_possible(steps).name.replace(
+ "UINT", "INT"
+ )
+ odt = DataType[odt]
+ assert odt.allowed(new_max) and odt.allowed(
+ new_min
+ ), """Could
+ not compute new MultiThreshold DataType (min = %d max = %d)""" % (
+ new_min,
+ new_max,
+ )
+ mt_inst.set_nodeattr("out_dtype", odt.name)
+ # remove Add node, rewire MultiThreshold
+ graph.node.remove(add_node)
+ mt_node.output[0] = end_name
+ # set datatype
+ model.set_tensor_datatype(end_name, odt)
+ if graph_modified:
+ model = model.transform(InferDataTypes())
+ return (model, graph_modified)
+
+
class AbsorbAddIntoMultiThreshold(Transformation):
"""Absorb preceding Add ops into MultiThreshold by updating the threshold
values. Only scalar/1D add vectors can be absorbed."""
@@ -290,3 +356,38 @@ class AbsorbTransposeIntoMultiThreshold(Transformation):
if graph_modified:
model = model.transform(InferDataTypes())
return (model, graph_modified)
+
+
+class AbsorbScalarMulIntoTopK(Transformation):
+ """Absorb a mul node into a suceeding topk node if the mul is scalar."""
+
+ def apply(self, model):
+ graph = model.graph
+ node_ind = 0
+ graph_modified = False
+ for n in graph.node:
+ node_ind += 1
+ if n.op_type == "TopK":
+ prod = model.find_producer(n.input[0])
+ if prod is not None and prod.op_type == "Mul":
+ prod_input = prod.input[0]
+ param_name = prod.input[1]
+ A = model.get_initializer(param_name)
+ if A is None:
+ warnings.warn("Param is not constant, skipping")
+ continue
+ if all(x == 1 for x in A.shape) and A > 0:
+ # if the mul is scalar and positive, we can just delete the
+ # mul node and rewire the top k node. Because the top k node
+ # works with probabilities and their relation to each other
+ # the relation doesn't change if every value is multiplied
+ # with a scalar
+ graph.node.remove(prod)
+ n.input[0] = prod_input
+ # to avoid error the dataype is set to float32
+ model.set_tensor_datatype(n.input[0], DataType.FLOAT32)
+ graph_modified = True
+ if graph_modified:
+ model = model.transform(InferShapes())
+ model = model.transform(InferDataTypes())
+ return (model, graph_modified)
diff --git a/src/finn/transformation/streamline/collapse_repeated.py b/src/finn/transformation/streamline/collapse_repeated.py
index 67824ad4f633983b93e3178d03118927a1ddd85b..769bed841ce07c1c9c62f762de4b2c0937a6d68f 100644
--- a/src/finn/transformation/streamline/collapse_repeated.py
+++ b/src/finn/transformation/streamline/collapse_repeated.py
@@ -30,6 +30,7 @@ from onnx import helper as oh
from finn.transformation import Transformation
from finn.transformation.infer_shapes import InferShapes
+from finn.core.datatype import DataType
class CollapseRepeatedOp(Transformation):
@@ -83,6 +84,9 @@ class CollapseRepeatedOp(Transformation):
graph.node.insert(node_ind, new_node)
# replace parameter value
model.set_initializer(new_node_param_name, new_param)
+ # be conservative with param/output DataTypes
+ model.set_tensor_datatype(new_node_param_name, DataType.FLOAT32)
+ model.set_tensor_datatype(end_name, DataType.FLOAT32)
# remove old nodes
graph.node.remove(n)
graph.node.remove(consumer)
diff --git a/src/finn/transformation/streamline/remove.py b/src/finn/transformation/streamline/remove.py
new file mode 100644
index 0000000000000000000000000000000000000000..ddc4233ddafbc70c4d20d316ea72ea6bba1b82a8
--- /dev/null
+++ b/src/finn/transformation/streamline/remove.py
@@ -0,0 +1,69 @@
+# 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.transformation import Transformation
+from finn.transformation.infer_shapes import InferShapes
+import numpy as np
+
+class RemoveIdentityOps(Transformation):
+ """Remove identity ops like Add/Sub with zero or Mul/Div with one"""
+
+ def apply(self, model):
+ graph = model.graph
+ node_ind = 0
+ graph_modified = False
+ for n in graph.node:
+ node_ind += 1
+ if (
+ n.op_type in ["Add", "Sub"]
+ and not model.is_fork_node(n)
+ and not model.is_join_node(n)
+ ):
+ A = model.get_initializer(n.input[1])
+ if A is not None and (A == np.zeros_like(A)).all():
+ producer = model.find_producer(n.input[0])
+ # remove node and wire output tensor to
+ # output of producer node
+ producer.output[0] = n.output[0]
+ graph.node.remove(n)
+
+ elif (
+ n.op_type in ["Mul", "Div"]
+ and not model.is_fork_node(n)
+ and not model.is_join_node(n)
+ ):
+ A = model.get_initializer(n.input[1])
+ if A is not None and (A == np.ones_like(A)).all():
+ producer = model.find_producer(n.input[0])
+ # remove node and wire output tensor to
+ # output of producer node
+ producer.output[0] = n.output[0]
+ graph.node.remove(n)
+ model = model.transform(InferShapes())
+ return (model, graph_modified)
diff --git a/src/finn/transformation/streamline/reorder.py b/src/finn/transformation/streamline/reorder.py
index 0b6259a61d3eb67b7b38d4c6939019ce2893a875..a1bd16f6d0b70193122d5d067ccdee395260c7b1 100644
--- a/src/finn/transformation/streamline/reorder.py
+++ b/src/finn/transformation/streamline/reorder.py
@@ -27,12 +27,15 @@
# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
import numpy as np
+import warnings
from onnx import helper as oh
from finn.transformation import Transformation
from finn.transformation.infer_shapes import InferShapes
+from finn.core.datatype import DataType
from finn.core.onnx_exec import execute_node
from finn.util.basic import get_by_name
+from finn.custom_op.registry import getCustomOp
class MoveAddPastMul(Transformation):
@@ -336,6 +339,71 @@ class MoveScalarMulPastConv(Transformation):
return (model, graph_modified)
+class MoveMulPastDWConv(Transformation):
+ """Move channelwise mul operations past depthwise conv operations. We want to have muls
+ next to each other such that they can be collapsed into a single mul."""
+
+ def apply(self, model):
+ graph = model.graph
+ node_ind = 0
+ graph_modified = False
+ for n in graph.node:
+ node_ind += 1
+ if (
+ n.op_type == "Mul"
+ and not model.is_fork_node(n)
+ and not model.is_join_node(n)
+ ):
+ consumer = model.find_consumer(n.output[0])
+ if (
+ consumer is not None
+ and consumer.op_type == "Conv"
+ and not model.is_join_node(consumer)
+ ):
+ mul_weight_name = n.input[1]
+ A = model.get_initializer(mul_weight_name)
+ if A is None:
+ warnings.warn(
+ """Mul weight tensor is not set. If it is a constant,
+ please use set_initializer to set the tensor."""
+ )
+ continue
+ conv_node = consumer
+ mul_node = n
+ start_name = mul_node.input[0]
+ conv_in_name = conv_node.input[0]
+ conv_in_shape = model.get_tensor_shape(conv_in_name)
+ ifm_ch = conv_in_shape[1]
+ group_attribute = get_by_name(consumer.attribute, "group")
+ if group_attribute is None:
+ continue
+ group_attribute = group_attribute.i
+ conv_out_name = conv_node.output[0]
+ conv_out_shape = model.get_tensor_shape(conv_out_name)
+ if A.shape == (1, ifm_ch, 1, 1) and ifm_ch == group_attribute:
+ # if the mul is channelwise and conv is depthwise,
+ # we can simply swap the order of ops
+ # rewire mul input to be conv input
+ conv_node.input[0] = start_name
+ model.set_tensor_shape(start_name, conv_in_shape)
+ model.set_tensor_datatype(start_name, DataType.FLOAT32)
+ # use old conv input tensor as conv output
+ conv_node.output[0] = conv_in_name
+ model.set_tensor_shape(conv_in_name, conv_out_shape)
+ model.set_tensor_datatype(conv_in_name, DataType.FLOAT32)
+ # use new conv output as new mul node input
+ mul_node.input[0] = conv_in_name
+ # use old conv output as new mul node output
+ mul_node.output[0] = conv_out_name
+ model.set_tensor_datatype(conv_out_name, DataType.FLOAT32)
+ # move mul node past conv node
+ graph.node.remove(mul_node)
+ graph.node.insert(node_ind, mul_node)
+ graph_modified = True
+ model = model.transform(InferShapes())
+ return (model, graph_modified)
+
+
class MoveLinearPastEltwiseAdd(Transformation):
"""Move linear operations (mul, add) past elementwise add operations where possible.
Specifically,matches and transforms the following patterns:
@@ -531,3 +599,67 @@ class MoveMulPastFork(MoveOpPastFork):
class MoveLinearPastFork(MoveOpPastFork):
def __init__(self):
super().__init__(["Add", "Mul"])
+
+
+class MoveMaxPoolPastMultiThreshold(Transformation):
+ """Move MaxPool nodes past MultiThreshold nodes on linear segments of the graph."""
+
+ def apply(self, model):
+ graph = model.graph
+ node_ind = 0
+ graph_modified = False
+ nodes = [n for n in graph.node]
+ for n in nodes:
+ node_ind += 1
+ if n.op_type == "MaxPool" and not model.is_fork_node(n):
+ consumer = model.find_consumer(n.output[0])
+ pads = get_by_name(n.attribute, "pads")
+ has_padding = False
+ if pads is not None:
+ pads = list(pads.ints)
+ has_padding = np.prod(pads) != 0
+ if consumer is not None and consumer.op_type == "MultiThreshold":
+ mt_out = consumer.output[0]
+ mt_odt = model.get_tensor_datatype(mt_out)
+ if mt_odt.signed() and has_padding:
+ warnings.warn(
+ "Skipping padded MaxPool + signed-output MultiThreshold"
+ )
+ continue
+ # check for non-decreasing thresholds and nonnegative
+ # scale factor in MultiThreshold
+ # otherwise we cannot do the reordering
+ T = model.get_initializer(consumer.input[1])
+ T_sorted = np.sort(T, axis=1)
+ assert (
+ T == T_sorted
+ ).all(), "MultiThreshold must have non-decreasing thresholds"
+ mt_inst = getCustomOp(consumer)
+ if mt_inst.get_nodeattr("out_scale") < 0:
+ warnings.warn("Skipping MultiThreshold with negative out_scale")
+ continue
+
+ # remove old nodes
+ graph.node.remove(n)
+ graph.node.remove(consumer)
+
+ # swap conections
+ group_in = n.input[0]
+ # new tensor because dims change
+ group_middle = model.make_new_valueinfo_name()
+ group_out = consumer.output[0]
+
+ consumer.input[0] = group_in
+ consumer.output[0] = group_middle
+
+ n.input[0] = group_middle
+ n.output[0] = group_out
+
+ # insert them back in
+ graph.node.insert(node_ind - 1, consumer)
+ graph.node.insert(node_ind, n)
+
+ graph_modified = True
+
+ model = model.transform(InferShapes())
+ return (model, graph_modified)
diff --git a/src/finn/util/basic.py b/src/finn/util/basic.py
index eb3d46bcd66e3dc307a679e6b8dfbb9913398d36..4a8277e08d3fc21e0b20668edf2ecad947b36647 100644
--- a/src/finn/util/basic.py
+++ b/src/finn/util/basic.py
@@ -31,6 +31,7 @@ import random
import string
import subprocess
import tempfile
+import warnings
import numpy as np
@@ -105,6 +106,25 @@ def get_finn_root():
)
+def get_execution_error_thresh():
+ "Return the max error that is allowed for rounding in FINN execution."
+ try:
+ return float(os.environ["ERROR_THRESH"])
+ except KeyError:
+ return 1e-2
+
+
+def get_sanitize_quant_tensors():
+ """Return whether tensors with quantization annotations should be sanitized.
+ Enabled by default, disabling will yield faster ONNX execution but may give
+ incorrect results. Use with caution."""
+ try:
+ return int(os.environ["SANITIZE_QUANT_TENSORS"])
+ except KeyError:
+ # enabled by default
+ return 1
+
+
def make_build_dir(prefix=""):
"""Creates a temporary folder with given prefix to be used as a build dir.
Use this function instead of tempfile.mkdtemp to ensure any generated files
@@ -264,6 +284,69 @@ def calculate_signed_dot_prod_range(dt_a, dt_b, len):
return (min_prod, max_prod)
+def sanitize_quant_values(model, node_tensors, execution_context, check_values=False):
+ """ Sanitize given list of tensors in execution_context by rounding values
+ that are supposed to be integers (as indicated by their quantization
+ annotation). Will raise an assertion if the amount of rounding is too large.
+ Returns the sanitized execution context.
+
+ If check_values is specified, an extra DataType.allowed() check will be
+ performed on any rounded tensors.
+
+ Background:
+ FINN uses floating point tensors as a carrier data type to represent
+ integers. Floating point arithmetic can introduce rounding errors, e.g.
+ (int_num * float_scale) / float_scale is not always equal to int_num.
+ We use this function to ensure that the values that are supposed to be
+ integers are indeed integers.
+ """
+
+ for tensor in node_tensors:
+ dtype = model.get_tensor_datatype(tensor)
+ # floats don't need sanitization, skip to next
+ # introduces less quicker runtime
+ if dtype == DataType.FLOAT32:
+ continue
+ current_values = execution_context[tensor]
+ updated_values = current_values
+ has_to_be_rounded = False
+ # TODO: vectorize with numpy
+ for value in np.nditer(current_values):
+ if not dtype.allowed(value):
+ has_to_be_rounded = True
+ break
+ if has_to_be_rounded:
+ updated_values = np.round(current_values)
+ warnings.warn(
+ "The values of tensor {} can't be represented "
+ "with the set FINN datatype ({}), they will be rounded to match the "
+ "FINN datatype.".format(tensor, dtype)
+ )
+ # check if rounded values are not too far from original values
+ max_error = max(np.abs(current_values - updated_values).flatten())
+ if max_error <= get_execution_error_thresh():
+ if check_values is True:
+ # check again if values can now be represented with set finn datatype
+ # TODO: vectorize with numpy
+ for value in np.nditer(updated_values):
+ if not dtype.allowed(value):
+ raise Exception(
+ """Values can't be represented with set
+ finn datatype ({}) for input {}""".format(
+ dtype, tensor
+ )
+ )
+ execution_context[tensor] = updated_values
+ else:
+ raise Exception(
+ """Rounding error is too high to match set FINN
+ datatype ({}) for input {}""".format(
+ dtype, tensor
+ )
+ )
+ return execution_context
+
+
class CppBuilder:
"""Builds the g++ compiler command to produces the executable of the c++ code
in code_gen_dir which is passed to the function build() of this class."""
diff --git a/src/finn/util/vivado.py b/src/finn/util/vivado.py
index 0f82c52cb2c1fc5ee4ed5a1927f46e222e0ab9b5..6b6df3940cfeeed292345382471719c49f725de6 100644
--- a/src/finn/util/vivado.py
+++ b/src/finn/util/vivado.py
@@ -28,6 +28,7 @@
import os
import subprocess
+import stat
from finn.util.basic import get_remote_vivado
@@ -91,6 +92,7 @@ def out_of_context_synth(
vivado_proj_folder = "%s/results_%s" % (verilog_dir, top_name)
res_counts_path = vivado_proj_folder + "/res.txt"
if remote_server is not None:
+ print("Using remote Vivado OOC synth, remote server %s" % remote_server)
run_synth = """
#!/bin/bash
which vivado;
@@ -105,14 +107,17 @@ cat %s
)
with open(vivado_proj_folder + "/run.sh", "w") as f:
f.write(run_synth)
+ st = os.stat(vivado_proj_folder + "/run.sh")
+ os.chmod(vivado_proj_folder + "/run.sh", st.st_mode | stat.S_IEXEC)
# note that this assumes the same temp folder can be created on the
# remote server
- remote_server_uri = remote_server + ":" + verilog_dir
- copy_files = "rsync -avz %s %s" % (verilog_dir + "/", remote_server_uri + "/")
+ # note we set target path as / due to use of -R (relative)
+ remote_server_uri = remote_server + ":/"
+ copy_files = "rsync -avzR %s %s" % (verilog_dir + "/", remote_server_uri)
copy_files = copy_files.split()
proc = subprocess.Popen(copy_files, cwd=verilog_dir, env=os.environ)
proc.communicate()
- vivado_cmd = "bash %s/run.sh" % vivado_proj_folder
+ vivado_cmd = "bash -ic %s/run.sh" % vivado_proj_folder
run_vivado = ["ssh", "-t", remote_server, vivado_cmd]
proc = subprocess.Popen(run_vivado, cwd=verilog_dir, env=os.environ)
proc.communicate()
diff --git a/tests/brevitas/test_brevitas_avg_pool_export.py b/tests/brevitas/test_brevitas_avg_pool_export.py
new file mode 100644
index 0000000000000000000000000000000000000000..24854a2153df9af78feb8352ca119e831a9ac9eb
--- /dev/null
+++ b/tests/brevitas/test_brevitas_avg_pool_export.py
@@ -0,0 +1,103 @@
+import os
+
+import onnx # noqa
+import torch
+import numpy as np
+import brevitas.onnx as bo
+from brevitas.nn import QuantAvgPool2d
+from brevitas.quant_tensor import pack_quant_tensor
+from brevitas.core.quant import QuantType
+from finn.core.modelwrapper import ModelWrapper
+from finn.core.datatype import DataType
+from finn.transformation.infer_shapes import InferShapes
+from finn.transformation.infer_datatypes import InferDataTypes
+from finn.util.basic import gen_finn_dt_tensor
+import finn.core.onnx_exec as oxe
+
+import pytest
+
+export_onnx_path = "test_avg_pool.onnx"
+
+
+@pytest.mark.parametrize("kernel_size", [2, 3])
+@pytest.mark.parametrize("stride", [1, 2])
+@pytest.mark.parametrize("signed", [False, True])
+@pytest.mark.parametrize("bit_width", [2, 4])
+@pytest.mark.parametrize("input_bit_width", [4, 8, 32])
+@pytest.mark.parametrize("channels", [2, 4])
+@pytest.mark.parametrize("idim", [7, 8])
+def test_brevitas_avg_pool_export(
+ kernel_size, stride, signed, bit_width, input_bit_width, channels, idim
+):
+ ishape = (1, channels, idim, idim)
+ ibw_tensor = torch.Tensor([input_bit_width])
+
+ b_avgpool = QuantAvgPool2d(
+ kernel_size=kernel_size,
+ stride=stride,
+ signed=signed,
+ min_overall_bit_width=bit_width,
+ max_overall_bit_width=bit_width,
+ quant_type=QuantType.INT,
+ )
+ # call forward pass manually once to cache scale factor and bitwidth
+ input_tensor = torch.from_numpy(np.zeros(ishape)).float()
+ scale = np.ones((1, channels, 1, 1))
+ output_scale = torch.from_numpy(scale).float()
+ input_quant_tensor = pack_quant_tensor(
+ tensor=input_tensor, scale=output_scale, bit_width=ibw_tensor
+ )
+ bo.export_finn_onnx(b_avgpool, ishape, export_onnx_path, input_t=input_quant_tensor)
+ model = ModelWrapper(export_onnx_path)
+
+ # determine input FINN datatype
+ if signed is True:
+ prefix = "INT"
+ else:
+ prefix = "UINT"
+ dt_name = prefix + str(input_bit_width // 2)
+ dtype = DataType[dt_name]
+ model = model.transform(InferShapes())
+ model = model.transform(InferDataTypes())
+
+ # execution with input tensor using integers and scale = 1
+ # calculate golden output
+ inp = gen_finn_dt_tensor(dtype, ishape)
+ input_tensor = torch.from_numpy(inp).float()
+ input_quant_tensor = pack_quant_tensor(
+ tensor=input_tensor, scale=output_scale, bit_width=ibw_tensor
+ )
+ b_avgpool.eval()
+ expected = b_avgpool.forward(input_quant_tensor).tensor.detach().numpy()
+
+ # finn execution
+ idict = {model.graph.input[0].name: inp}
+ odict = oxe.execute_onnx(model, idict, True)
+ produced = odict[model.graph.output[0].name]
+ assert (expected == produced).all()
+
+ # execution with input tensor using float and scale != 1
+ scale = np.random.uniform(low=0, high=1, size=(1, channels, 1, 1)).astype(
+ np.float32
+ )
+ inp_tensor = inp * scale
+ input_tensor = torch.from_numpy(inp_tensor).float()
+ input_scale = torch.from_numpy(scale).float()
+ input_quant_tensor = pack_quant_tensor(
+ tensor=input_tensor, scale=input_scale, bit_width=ibw_tensor
+ )
+ # export again to set the scale values correctly
+ bo.export_finn_onnx(b_avgpool, ishape, export_onnx_path, input_t=input_quant_tensor)
+ model = ModelWrapper(export_onnx_path)
+ model = model.transform(InferShapes())
+ model = model.transform(InferDataTypes())
+ b_avgpool.eval()
+ expected = b_avgpool.forward(input_quant_tensor).tensor.detach().numpy()
+ # finn execution
+ idict = {model.graph.input[0].name: inp_tensor}
+ odict = oxe.execute_onnx(model, idict, True)
+ produced = odict[model.graph.output[0].name]
+
+ assert np.isclose(expected, produced).all()
+
+ os.remove(export_onnx_path)
diff --git a/tests/brevitas/test_brevitas_relu_act_export.py b/tests/brevitas/test_brevitas_relu_act_export.py
index c9d8f2d812bc7bea1a2fd2598a7711099ad421e6..c5ddad12ca3e8d353682fbb20449d44358485f69 100644
--- a/tests/brevitas/test_brevitas_relu_act_export.py
+++ b/tests/brevitas/test_brevitas_relu_act_export.py
@@ -23,6 +23,7 @@ export_onnx_path = "test_act.onnx"
def test_brevitas_act_export_relu(abits, max_val, scaling_impl_type):
min_val = -1.0
ishape = (1, 15)
+
b_act = QuantReLU(
bit_width=abits,
max_val=max_val,
@@ -67,3 +68,60 @@ scaling_impl.learned_value": torch.tensor(
assert np.isclose(produced, expected, atol=1e-3).all()
os.remove(export_onnx_path)
+
+
+@pytest.mark.parametrize("abits", [1, 2, 4, 8])
+@pytest.mark.parametrize("max_val", [1.0, 1.5, 1 - 2 ** (-7)])
+@pytest.mark.parametrize("scaling_per_channel", [True, False])
+def test_brevitas_act_export_relu_imagenet(abits, max_val, scaling_per_channel):
+ out_channels = 32
+ ishape = (1, out_channels, 1, 1)
+ min_val = -1.0
+ b_act = QuantReLU(
+ bit_width=abits,
+ quant_type=QuantType.INT,
+ scaling_impl_type=ScalingImplType.PARAMETER,
+ scaling_per_channel=scaling_per_channel,
+ restrict_scaling_type=RestrictValueType.LOG_FP,
+ scaling_min_val=2e-16,
+ max_val=6.0,
+ return_quant_tensor=True,
+ per_channel_broadcastable_shape=(1, out_channels, 1, 1),
+ )
+ if scaling_per_channel is True:
+ rand_tensor = (2) * torch.rand((1, out_channels, 1, 1))
+ else:
+ rand_tensor = torch.tensor(1.2398)
+ checkpoint = {
+ "act_quant_proxy.fused_activation_quant_proxy.tensor_quant.\
+scaling_impl.learned_value": rand_tensor.type(
+ torch.FloatTensor
+ )
+ }
+ b_act.load_state_dict(checkpoint)
+ bo.export_finn_onnx(b_act, ishape, export_onnx_path)
+ model = ModelWrapper(export_onnx_path)
+ model = model.transform(InferShapes())
+ inp_tensor = np.random.uniform(low=min_val, high=max_val, size=ishape).astype(
+ np.float32
+ )
+ idict = {model.graph.input[0].name: inp_tensor}
+ odict = oxe.execute_onnx(model, idict, True)
+ produced = odict[model.graph.output[0].name]
+ inp_tensor = torch.from_numpy(inp_tensor).float()
+ b_act.eval()
+ expected = b_act.forward(inp_tensor).tensor.detach().numpy()
+ if not np.isclose(produced, expected, atol=1e-3).all():
+ print(abits, max_val)
+ print("scale: ", b_act.quant_act_scale().type(torch.FloatTensor).detach())
+ if abits < 5:
+ print(
+ "thres:",
+ ", ".join(["{:8.4f}".format(x) for x in b_act.export_thres[0]]),
+ )
+ print("input:", ", ".join(["{:8.4f}".format(x) for x in inp_tensor[0]]))
+ print("prod :", ", ".join(["{:8.4f}".format(x) for x in produced[0]]))
+ print("expec:", ", ".join(["{:8.4f}".format(x) for x in expected[0]]))
+
+ assert np.isclose(produced, expected, atol=1e-3).all()
+ os.remove(export_onnx_path)
diff --git a/tests/core/test_basic_onnx_exec.py b/tests/core/test_basic_onnx_exec.py
index a7b6da9965aa5912870812a8c1f8d6da2ee0d181..ddb2cbfc40c7647970f0c51ecb95340e7d1dddae 100644
--- a/tests/core/test_basic_onnx_exec.py
+++ b/tests/core/test_basic_onnx_exec.py
@@ -35,6 +35,8 @@ import onnx.numpy_helper as np_helper
import finn.core.onnx_exec as oxe
from finn.core.modelwrapper import ModelWrapper
from finn.transformation.infer_shapes import InferShapes
+from finn.core.datatype import DataType
+from finn.util.basic import gen_finn_dt_tensor
def test_mnist_onnx_download_extract_run():
@@ -47,9 +49,50 @@ def test_mnist_onnx_download_extract_run():
raw_o = get_data("finn", "data/onnx/mnist-conv/test_data_set_0/output_0.pb")
input_tensor = onnx.load_tensor_from_string(raw_i)
output_tensor = onnx.load_tensor_from_string(raw_o)
- # run using FINN-based execution
+ # run using FINN-based execution (full graph)
input_dict = {"Input3": np_helper.to_array(input_tensor)}
- output_dict = oxe.execute_onnx(model, input_dict)
+ output_dict = oxe.execute_onnx(model, input_dict, return_full_exec_context=True)
assert np.isclose(
np_helper.to_array(output_tensor), output_dict["Plus214_Output_0"], atol=1e-3
).all()
+ # test subgraph execution
+ start_node = model.graph.node[1]
+ end_node = model.graph.node[3]
+ subgraph_i_dict = {start_node.input[0]: output_dict[start_node.input[0]]}
+ subgraph_o_dict = oxe.execute_onnx(
+ model,
+ subgraph_i_dict,
+ return_full_exec_context=True,
+ start_node=start_node,
+ end_node=end_node,
+ )
+ assert np.isclose(
+ subgraph_o_dict[end_node.output[0]], output_dict[end_node.output[0]], atol=1e-3
+ ).all()
+
+
+def test_onnx_exec_internal_rounding():
+ inp0 = onnx.helper.make_tensor_value_info("inp0", onnx.TensorProto.FLOAT, [2, 2])
+ inp1 = onnx.helper.make_tensor_value_info("inp1", onnx.TensorProto.FLOAT, [1])
+ outp = onnx.helper.make_tensor_value_info("outp", onnx.TensorProto.FLOAT, [2, 2])
+ mul_node = onnx.helper.make_node("Mul", inputs=["inp0", "inp1"], outputs=["outp"])
+ graph = onnx.helper.make_graph(
+ nodes=[mul_node], name="mul_graph", inputs=[inp0, inp1], outputs=[outp]
+ )
+
+ model = onnx.helper.make_model(graph, producer_name="mul-model")
+ model = ModelWrapper(model)
+ idt = DataType.INT2
+ model.set_tensor_datatype("inp0", idt)
+ model.set_tensor_datatype("inp1", idt)
+ model.transform(InferShapes())
+
+ mul_value = np.asarray([-1], dtype=np.float32)
+ inp_int = gen_finn_dt_tensor(idt, [2, 2])
+ scale = np.random.uniform(low=0, high=1, size=(2, 2)).astype(np.float32)
+ inp_rounded = (inp_int * scale) / (scale + 1e-7)
+ input_dict = {"inp0": inp_rounded, "inp1": mul_value}
+ output_dict = oxe.execute_onnx(model, input_dict)
+ produced = output_dict["outp"]
+ expected = np.multiply(inp_int, mul_value)
+ assert (produced == expected).all()
diff --git a/tests/end2end/test_end2end_cnv_w2a2.py b/tests/end2end/test_end2end_cnv_w2a2.py
new file mode 100644
index 0000000000000000000000000000000000000000..31ccebd4c175ad2badef17499bf113d978b637f7
--- /dev/null
+++ b/tests/end2end/test_end2end_cnv_w2a2.py
@@ -0,0 +1,377 @@
+# 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.
+
+import os
+
+import numpy as np
+
+# as of Feb'20 there is a bug that segfaults ONNX shape inference if we
+# import pytorch before onnx, so we make sure to import onnx first
+import onnx # NOQA
+
+import pytest
+import pkg_resources as pk
+from finn.custom_op.registry import getCustomOp
+from finn.core.onnx_exec import execute_onnx
+from finn.transformation.double_to_single_float import DoubleToSingleFloat
+from finn.transformation.infer_shapes import InferShapes
+from finn.transformation.move_reshape import RemoveCNVtoFCFlatten
+from finn.transformation.fold_constants import FoldConstants
+from finn.transformation.general import GiveReadableTensorNames, GiveUniqueNodeNames
+from finn.transformation.streamline import Streamline
+from finn.transformation.lower_convs_to_matmul import LowerConvsToMatMul
+import finn.transformation.streamline.absorb as absorb
+from finn.transformation.streamline.reorder import MakeMaxPoolNHWC
+import finn.transformation.fpgadataflow.convert_to_hls_layers as to_hls
+from finn.transformation.fpgadataflow.create_dataflow_partition import (
+ CreateDataflowPartition,
+)
+from finn.transformation.fpgadataflow.insert_dwc import InsertDWC
+from finn.transformation.fpgadataflow.insert_tlastmarker import InsertTLastMarker
+from finn.transformation.fpgadataflow.prepare_ip import PrepareIP
+from finn.transformation.fpgadataflow.hlssynth_ip import HLSSynthIP
+from finn.transformation.fpgadataflow.replace_verilog_relpaths import (
+ ReplaceVerilogRelPaths,
+)
+from finn.transformation.fpgadataflow.create_stitched_ip import CreateStitchedIP
+from finn.transformation.fpgadataflow.set_exec_mode import SetExecMode
+from finn.transformation.fpgadataflow.prepare_cppsim import PrepareCppSim
+from finn.transformation.fpgadataflow.compile_cppsim import CompileCppSim
+from finn.transformation.fpgadataflow.make_pynq_driver import MakePYNQDriver
+from finn.transformation.fpgadataflow.make_pynq_proj import MakePYNQProject
+from finn.transformation.fpgadataflow.synth_pynq_proj import SynthPYNQProject
+from finn.transformation.fpgadataflow.make_deployment import DeployToPYNQ
+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
+from finn.transformation.fpgadataflow.insert_fifo import InsertFIFO
+from finn.core.throughput_test import throughput_test_rtlsim
+
+build_dir = "/tmp/" + os.environ["FINN_INST_NAME"]
+test_pynq_board = os.getenv("PYNQ_BOARD", default="Pynq-Z1")
+test_fpga_part = pynq_part_map[test_pynq_board]
+target_clk_ns = 10
+mem_mode = "decoupled"
+
+
+def test_end2end_cnv_w2a2_export():
+ import brevitas.onnx as bo
+
+ cnv = get_test_model_trained("CNV", 2, 2)
+ bo.export_finn_onnx(
+ cnv, (1, 3, 32, 32), build_dir + "/end2end_cnv_w2a2_export.onnx"
+ )
+
+
+def test_end2end_cnv_w2a2_import_and_tidy():
+ model = load_test_checkpoint_or_skip(build_dir + "/end2end_cnv_w2a2_export.onnx")
+ model = model.transform(DoubleToSingleFloat())
+ model = model.transform(InferShapes())
+ model = model.transform(FoldConstants())
+ model = model.transform(GiveUniqueNodeNames())
+ model = model.transform(GiveReadableTensorNames())
+ model.save(build_dir + "/end2end_cnv_w2a2_tidy.onnx")
+
+
+def test_end2end_cnv_w2a2_streamline():
+ model = load_test_checkpoint_or_skip(build_dir + "/end2end_cnv_w2a2_tidy.onnx")
+ model = model.transform(Streamline())
+ model = model.transform(LowerConvsToMatMul())
+ model = model.transform(MakeMaxPoolNHWC())
+ model = model.transform(absorb.AbsorbTransposeIntoMultiThreshold())
+ model = model.transform(Streamline())
+ model.save(build_dir + "/end2end_cnv_w2a2_streamlined.onnx")
+
+
+def test_end2end_cnv_w2a2_convert_to_hls_layers():
+ model = load_test_checkpoint_or_skip(
+ build_dir + "/end2end_cnv_w2a2_streamlined.onnx"
+ )
+ model = model.transform(to_hls.InferQuantizedStreamingFCLayer(mem_mode))
+ model = model.transform(to_hls.InferConvInpGen())
+ model = model.transform(to_hls.InferStreamingMaxPool())
+ model = model.transform(RemoveCNVtoFCFlatten())
+ model.save(build_dir + "/end2end_cnv_w2a2_hls_layers.onnx")
+
+
+def test_end2end_cnv_w2a2_create_dataflow_partition():
+ model = load_test_checkpoint_or_skip(
+ build_dir + "/end2end_cnv_w2a2_hls_layers.onnx"
+ )
+ parent_model = model.transform(CreateDataflowPartition())
+ parent_model.save(build_dir + "/end2end_cnv_w2a2_dataflow_parent.onnx")
+ sdp_node = parent_model.get_nodes_by_op_type("StreamingDataflowPartition")[0]
+ sdp_node = getCustomOp(sdp_node)
+ dataflow_model_filename = sdp_node.get_nodeattr("model")
+ dataflow_model = load_test_checkpoint_or_skip(dataflow_model_filename)
+ dataflow_model.save(build_dir + "/end2end_cnv_w2a2_dataflow_model.onnx")
+
+
+def test_end2end_cnv_w2a2_fold_and_tlastmarker():
+ model = load_test_checkpoint_or_skip(
+ build_dir + "/end2end_cnv_w2a2_dataflow_model.onnx"
+ )
+ fc_layers = model.get_nodes_by_op_type("StreamingFCLayer_Batch")
+ # each tuple is (PE, SIMD, in_fifo_depth) for a layer
+ folding = [
+ (8, 3, 256, "auto"),
+ (16, 16, 256, "auto"),
+ (8, 16, 256, "auto"),
+ (8, 16, 256, "block"),
+ (4, 8, 214, "auto"),
+ (1, 8, 2, "auto"),
+ (1, 2, 126, "distributed"),
+ (2, 2, 62, "block"),
+ (5, 1, 6, "distributed"),
+ ]
+ for fcl, (pe, simd, ififodepth, ramstyle) in zip(fc_layers, folding):
+ fcl_inst = getCustomOp(fcl)
+ fcl_inst.set_nodeattr("PE", pe)
+ fcl_inst.set_nodeattr("SIMD", simd)
+ fcl_inst.set_nodeattr("inFIFODepth", ififodepth)
+ fcl_inst.set_nodeattr("ram_style", ramstyle)
+
+ swg_layers = model.get_nodes_by_op_type("ConvolutionInputGenerator")
+ swg_idepth = [2, 51, 9, 106, 2, 2]
+ for i in range(len(swg_layers)):
+ swg_inst = getCustomOp(swg_layers[i])
+ simd = folding[i][1]
+ swg_inst.set_nodeattr("SIMD", simd)
+ swg_inst.set_nodeattr("inFIFODepth", swg_idepth[i])
+
+ model = model.transform(InsertDWC())
+ model = model.transform(InsertFIFO())
+ model = model.transform(InsertTLastMarker())
+ model = model.transform(GiveUniqueNodeNames())
+ model = model.transform(AnnotateResources("estimate"))
+ model.save(build_dir + "/end2end_cnv_w2a2_folded.onnx")
+
+
+@pytest.mark.slow
+@pytest.mark.vivado
+def test_end2end_cnv_w2a2_gen_hls_ip():
+ model = load_test_checkpoint_or_skip(build_dir + "/end2end_cnv_w2a2_folded.onnx")
+ model = model.transform(PrepareIP(test_fpga_part, target_clk_ns))
+ model = model.transform(HLSSynthIP())
+ model = model.transform(AnnotateResources("hls"))
+ model.save(build_dir + "/end2end_cnv_w2a2_ipgen.onnx")
+
+
+@pytest.mark.vivado
+def test_end2end_cnv_w2a2_ip_stitch():
+ model = load_test_checkpoint_or_skip(build_dir + "/end2end_cnv_w2a2_ipgen.onnx")
+ model = model.transform(ReplaceVerilogRelPaths())
+ model = model.transform(CreateStitchedIP(test_fpga_part, target_clk_ns))
+ model.save(build_dir + "/end2end_cnv_w2a2_ipstitch.onnx")
+
+
+@pytest.mark.vivado
+def test_end2end_cnv_w2a2_verify_dataflow_part():
+ model = load_test_checkpoint_or_skip(build_dir + "/end2end_cnv_w2a2_ipstitch.onnx")
+ x = np.zeros((1, 32, 32, 3), dtype=np.float32)
+ inp_name = model.graph.input[0].name
+ out_name = model.graph.output[0].name
+ inp_dict = {inp_name: x}
+ # cppsim
+ model = model.transform(PrepareCppSim())
+ model = model.transform(CompileCppSim())
+ model = model.transform(SetExecMode("cppsim"))
+ model.save(build_dir + "/end2end_cnv_w2a2_ipgen_cppsim.onnx")
+ ret_cppsim = execute_onnx(model, inp_dict, True)
+ res_cppsim = ret_cppsim[out_name]
+ # node-by-node rtlsim
+ model = model.transform(SetExecMode("rtlsim"))
+ model = model.transform(PrepareRTLSim())
+ model.save(build_dir + "/end2end_cnv_w2a2_ipgen_nodebynode_rtlsim.onnx")
+ ret_rtlsim_nodebynode = execute_onnx(model, inp_dict, True)
+ res_rtlsim_nodebynode = ret_rtlsim_nodebynode[out_name]
+ # whole-network (ip-stitched) rtlsim
+ model.set_metadata_prop("exec_mode", "rtlsim")
+ model.save(build_dir + "/end2end_cnv_w2a2_ipstitch_whole_rtlsim.onnx")
+ # this is a particularly long-running test, set liveness thr. to unlimited
+ os.environ["LIVENESS_THRESHOLD"] = "-1"
+ ret_rtlsim_whole = execute_onnx(model, inp_dict, True)
+ res_rtlsim_whole = ret_rtlsim_whole[out_name]
+ assert np.isclose(res_cppsim, res_rtlsim_nodebynode).all()
+ assert np.isclose(res_cppsim, res_rtlsim_whole).all()
+
+
+@pytest.mark.vivado
+def test_end2end_cnv_w2a2_throughput_test_rtlsim():
+ model = load_test_checkpoint_or_skip(
+ build_dir + "/end2end_cnv_w2a2_ipstitch_whole_rtlsim.onnx"
+ )
+ model.set_metadata_prop("rtlsim_trace", "rtlsim_trace.vcd")
+ # os.environ["RTLSIM_TRACE_DEPTH"] = "4"
+ # run through IP-stitched rtlsim with increasing batch sizes and
+ # check the number of cycles it takes to execute
+ ret = throughput_test_rtlsim(model, 10)
+ # TODO check for expected performance
+ assert ret["cycles"] > 0
+
+
+@pytest.mark.vivado
+def test_end2end_cnv_w2a2_verify_all():
+ # use the streamlined model as the "golden" model for right answers
+ golden = load_test_checkpoint_or_skip(
+ build_dir + "/end2end_cnv_w2a2_streamlined.onnx"
+ )
+ iname = golden.graph.input[0].name
+ oname = golden.graph.output[0].name
+ # load one of the test vectors
+ fn = pk.resource_filename("finn", "data/cifar10/cifar10-test-data-class3.npz")
+ input_tensor = np.load(fn)["arr_0"].astype(np.float32)
+ input_tensor = input_tensor / 255
+ assert input_tensor.shape == (1, 3, 32, 32)
+ x = input_tensor
+ # x = np.zeros(ishape, dtype=np.float32)
+ ret_golden = execute_onnx(golden, {iname: x}, True)
+ y_golden = ret_golden[oname]
+ # set up parent+child graph to test
+ # we'll use models from the previous step as the child model
+ parent_model = load_test_checkpoint_or_skip(
+ build_dir + "/end2end_cnv_w2a2_dataflow_parent.onnx"
+ )
+ iname = parent_model.graph.input[0].name
+ oname = parent_model.graph.output[0].name
+ # produce results with cppsim
+ sdp_node = parent_model.get_nodes_by_op_type("StreamingDataflowPartition")[0]
+ sdp_node = getCustomOp(sdp_node)
+ load_test_checkpoint_or_skip(build_dir + "/end2end_cnv_w2a2_ipgen_cppsim.onnx")
+ sdp_node.set_nodeattr("model", build_dir + "/end2end_cnv_w2a2_ipgen_cppsim.onnx")
+ ret_cppsim = execute_onnx(parent_model, {iname: x}, True)
+ y_cppsim = ret_cppsim[oname]
+ # produce results with node-by-node rtlsim
+ load_test_checkpoint_or_skip(
+ build_dir + "/end2end_cnv_w2a2_ipgen_nodebynode_rtlsim.onnx"
+ )
+ sdp_node.set_nodeattr(
+ "model", build_dir + "/end2end_cnv_w2a2_ipgen_nodebynode_rtlsim.onnx"
+ )
+ ret_nodebynode_rtlsim = execute_onnx(parent_model, {iname: x}, True)
+ y_nodebynode_rtlsim = ret_nodebynode_rtlsim[oname]
+ # produce results with whole-network (stitched ip) rtlsim
+ load_test_checkpoint_or_skip(
+ build_dir + "/end2end_cnv_w2a2_ipstitch_whole_rtlsim.onnx"
+ )
+ sdp_node.set_nodeattr(
+ "model", build_dir + "/end2end_cnv_w2a2_ipstitch_whole_rtlsim.onnx"
+ )
+ # this is a particularly long-running test, set liveness thr. to unlimited
+ os.environ["LIVENESS_THRESHOLD"] = "-1"
+ ret_whole_rtlsim = execute_onnx(parent_model, {iname: x}, True)
+ y_whole_rtlsim = ret_whole_rtlsim[oname]
+ assert np.isclose(y_golden, y_cppsim).all()
+ assert np.isclose(y_golden, y_nodebynode_rtlsim).all()
+ assert np.isclose(y_golden, y_whole_rtlsim).all()
+ assert np.argmax(y_golden) == 3
+
+
+@pytest.mark.vivado
+def test_end2end_cnv_w2a2_make_pynq_proj():
+ model = load_test_checkpoint_or_skip(build_dir + "/end2end_cnv_w2a2_ipstitch.onnx")
+ model = model.transform(MakePYNQProject(test_pynq_board))
+ model.save(build_dir + "/end2end_cnv_w2a2_pynq_project.onnx")
+
+
+@pytest.mark.slow
+@pytest.mark.vivado
+def test_end2end_cnv_w2a2_synth_pynq_project():
+ model = load_test_checkpoint_or_skip(
+ build_dir + "/end2end_cnv_w2a2_pynq_project.onnx"
+ )
+ model = model.transform(SynthPYNQProject())
+ model = model.transform(AnnotateResources("synth"))
+ model.save(build_dir + "/end2end_cnv_w2a2_synth.onnx")
+
+
+def test_end2end_cnv_w2a2_make_driver():
+ model = load_test_checkpoint_or_skip(build_dir + "/end2end_cnv_w2a2_synth.onnx")
+ model = model.transform(MakePYNQDriver())
+ model.save(build_dir + "/end2end_cnv_w2a2_pynq_driver.onnx")
+
+
+def test_end2end_cnv_w2a2_deploy_on_pynq():
+ model = load_test_checkpoint_or_skip(
+ build_dir + "/end2end_cnv_w2a2_pynq_driver.onnx"
+ )
+ try:
+ ip = os.environ["PYNQ_IP"] # no fault for this one; skip if not defined
+ if ip == "":
+ pytest.skip("PYNQ board IP address not specified")
+ username = os.getenv("PYNQ_USERNAME", "xilinx")
+ password = os.getenv("PYNQ_PASSWORD", "xilinx")
+ port = os.getenv("PYNQ_PORT", 22)
+ target_dir = os.getenv("PYNQ_TARGET_DIR", "/home/xilinx/finn")
+ model = model.transform(DeployToPYNQ(ip, port, username, password, target_dir))
+ # save the model to be able to link it to the parent
+ model.save(build_dir + "/end2end_cnv_w2a2_pynq_deploy.onnx")
+ except KeyError:
+ pytest.skip("PYNQ board IP address not specified")
+
+
+def test_end2end_cnv_w2a2_run_on_pynq():
+ # use the streamlined model as the "golden" model for right answers
+ golden = load_test_checkpoint_or_skip(
+ build_dir + "/end2end_cnv_w2a2_streamlined.onnx"
+ )
+ iname = golden.graph.input[0].name
+ oname = golden.graph.output[0].name
+ # load one of the test vectors
+ fn = pk.resource_filename("finn", "data/cifar10/cifar10-test-data-class3.npz")
+ input_tensor = np.load(fn)["arr_0"].astype(np.float32)
+ input_tensor = input_tensor / 255
+ assert input_tensor.shape == (1, 3, 32, 32)
+ x = input_tensor
+ # run using FINN-based execution
+ ret_golden = execute_onnx(golden, {iname: x}, True)
+ y_golden = ret_golden[oname]
+ # set up parent+child graph to test
+ # we'll use models from the previous step as the child model
+ parent_model = load_test_checkpoint_or_skip(
+ build_dir + "/end2end_cnv_w2a2_dataflow_parent.onnx"
+ )
+ iname = parent_model.graph.input[0].name
+ oname = parent_model.graph.output[0].name
+ try:
+ ip = os.environ["PYNQ_IP"] # NOQA
+ if ip == "":
+ pytest.skip("PYNQ board IP address not specified")
+ # produce results with cppsim
+ sdp_node = parent_model.get_nodes_by_op_type("StreamingDataflowPartition")[0]
+ sdp_node = getCustomOp(sdp_node)
+ load_test_checkpoint_or_skip(build_dir + "/end2end_cnv_w2a2_pynq_deploy.onnx")
+ sdp_node.set_nodeattr("model", build_dir + "/end2end_cnv_w2a2_pynq_deploy.onnx")
+ ret = execute_onnx(parent_model, {iname: x}, True)
+ y = ret[oname]
+ assert np.isclose(y, y_golden).all()
+ assert np.argmax(y) == 3
+
+ except KeyError:
+ pytest.skip("PYNQ board IP address not specified")
diff --git a/tests/fpgadataflow/test_convert_to_hls_pool_batch.py b/tests/fpgadataflow/test_convert_to_hls_pool_batch.py
new file mode 100644
index 0000000000000000000000000000000000000000..c9f78dcea1a1ce364d0657ad64de7d440d41b822
--- /dev/null
+++ b/tests/fpgadataflow/test_convert_to_hls_pool_batch.py
@@ -0,0 +1,160 @@
+# 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.
+
+import pytest
+
+from onnx import TensorProto, helper
+import numpy as np
+import finn.core.onnx_exec as oxe
+from finn.core.datatype import DataType
+from finn.core.modelwrapper import ModelWrapper
+from finn.transformation.fpgadataflow.prepare_ip import PrepareIP
+from finn.transformation.fpgadataflow.prepare_cppsim import PrepareCppSim
+from finn.transformation.fpgadataflow.compile_cppsim import CompileCppSim
+from finn.transformation.fpgadataflow.hlssynth_ip import HLSSynthIP
+from finn.transformation.fpgadataflow.set_exec_mode import SetExecMode
+from finn.transformation.fpgadataflow.prepare_rtlsim import PrepareRTLSim
+import finn.transformation.fpgadataflow.convert_to_hls_layers as to_hls
+from finn.transformation.general import GiveUniqueNodeNames
+from finn.custom_op.registry import getCustomOp
+from finn.util.basic import gen_finn_dt_tensor
+from finn.transformation.infer_shapes import InferShapes
+
+
+def make_single_maxpool_modelwrapper(k, stride, pad, ifm_ch, ifm_dim, ofm_dim, idt):
+ odt = idt
+ inp = helper.make_tensor_value_info(
+ "inp", TensorProto.FLOAT, [1, ifm_ch, ifm_dim, ifm_dim]
+ )
+ outp = helper.make_tensor_value_info(
+ "outp", TensorProto.FLOAT, [1, ifm_ch, ofm_dim, ofm_dim]
+ )
+
+ mp_node = helper.make_node(
+ "MaxPool",
+ ["inp"],
+ ["outp"],
+ kernel_shape=[k, k],
+ pads=[pad, pad, pad, pad],
+ strides=[stride, stride],
+ )
+ graph = helper.make_graph(
+ nodes=[mp_node], name="mp_graph", inputs=[inp], outputs=[outp]
+ )
+
+ model = helper.make_model(graph, producer_name="mp-model")
+ model = ModelWrapper(model)
+
+ model.set_tensor_datatype("inp", idt)
+ model.set_tensor_datatype("outp", odt)
+ model = model.transform(InferShapes())
+
+ return model
+
+
+def prepare_inputs(input_tensor):
+ return {"inp": input_tensor}
+
+
+# input datatype
+@pytest.mark.parametrize("idt", [DataType.UINT4, DataType.INT4])
+# pool configuration: ( k,stride, pad, ifm_dim )
+@pytest.mark.parametrize(
+ "pool_config", [(3, 2, 0, 5), (3, 2, 1, 5), (2, 2, 0, 8), (5, 2, 2, 7)]
+)
+# input channels
+@pytest.mark.parametrize("ifm_ch", [1, 4, 20])
+# number of out channel computed in parallel
+@pytest.mark.parametrize("pe", [1, 4, 20])
+# execution mode
+@pytest.mark.parametrize("exec_mode", ["cppsim", "rtlsim"])
+# pool type
+@pytest.mark.parametrize("op_type", ["MaxPool"])
+@pytest.mark.slow
+@pytest.mark.vivado
+def test_convert_to_hls_pool_batch(idt, pool_config, ifm_ch, pe, exec_mode, op_type):
+ k, stride, pad, ifm_dim = pool_config
+
+ if ifm_ch % pe != 0:
+ pytest.skip("ifm_ch%pe != 0. Skipping")
+
+ if pad != 0 and idt.signed():
+ pytest.skip("No support for pal_val != 0. Skipping")
+
+ np.random.seed(0)
+ ofm_dim = int(((ifm_dim + 2 * pad - k) / stride) + 1)
+
+ x = gen_finn_dt_tensor(idt, (1, ifm_ch, ifm_dim, ifm_dim))
+ # prepare input data
+ input_dict = prepare_inputs(x)
+ if op_type == "MaxPool":
+ model = make_single_maxpool_modelwrapper(
+ k, stride, pad, ifm_ch, ifm_dim, ofm_dim, idt
+ )
+ else:
+ assert False, "{} is not a supported op_type".format(op_type)
+
+ y_expected = oxe.execute_onnx(model, input_dict)["outp"]
+
+ new_model = model.transform(to_hls.InferPool_Batch())
+ new_model = new_model.transform(GiveUniqueNodeNames())
+
+ if ifm_ch != pe:
+ new_model = new_model.transform(to_hls.InferConvInpGen())
+ # Folding
+ for n in new_model.graph.node:
+ if n.op_type == "ConvolutionInputGenerator":
+ inst = getCustomOp(n)
+ inst.set_nodeattr("SIMD", pe)
+ elif n.op_type == "Pool_Batch":
+ inst = getCustomOp(n)
+ inst.set_nodeattr("PE", pe)
+
+ if exec_mode == "cppsim":
+ new_model = new_model.transform(SetExecMode("cppsim"))
+ new_model = new_model.transform(PrepareCppSim())
+ new_model = new_model.transform(CompileCppSim())
+ elif exec_mode == "rtlsim":
+ new_model = new_model.transform(SetExecMode("rtlsim"))
+ new_model = new_model.transform(GiveUniqueNodeNames())
+ new_model = new_model.transform(PrepareIP("xc7z020clg400-1", 5))
+ new_model = new_model.transform(HLSSynthIP())
+ new_model = new_model.transform(PrepareRTLSim())
+ else:
+ raise Exception("Unknown exec_mode")
+
+ # execute new_model
+ y_produced = oxe.execute_onnx(new_model, input_dict)["outp"]
+ assert (y_produced == y_expected).all()
+ if stride != k:
+ if pad == 0 or ifm_ch == pe:
+ assert len(new_model.graph.node) == 4
+ else:
+ assert len(new_model.graph.node) == 5
+ else:
+ assert len(new_model.graph.node) == 1
diff --git a/tests/fpgadataflow/test_fpgadataflow_fmpadding.py b/tests/fpgadataflow/test_fpgadataflow_fmpadding.py
index 9d6390b2673e5d2c0e72748183ac04ed222d078e..5ff3da87228a2a32a41226bb46e0b16b1a44df50 100644
--- a/tests/fpgadataflow/test_fpgadataflow_fmpadding.py
+++ b/tests/fpgadataflow/test_fpgadataflow_fmpadding.py
@@ -23,7 +23,7 @@ test_fpga_part = pynq_part_map[test_pynq_board]
target_clk_ns = 10
-def make_single_fmpadding_modelwrapper(idim, padding, num_ch, idt, pad_style):
+def make_single_fmpadding_modelwrapper(idim, padding, num_ch, simd, idt, pad_style):
assert pad_style == 2, "only pad_style == 2 supported in hlslib"
assert padding > 0, "Output dim should be greater than input dim"
odim = idim + padding
@@ -47,6 +47,7 @@ def make_single_fmpadding_modelwrapper(idim, padding, num_ch, idt, pad_style):
inputDataType=str(idt.name),
PaddingStyle=pad_style,
numInputVectors=1,
+ SIMD=simd,
)
graph = helper.make_graph(
@@ -63,11 +64,13 @@ def make_single_fmpadding_modelwrapper(idim, padding, num_ch, idt, pad_style):
# input image dimension
-@pytest.mark.parametrize("idim", [8, 16])
+@pytest.mark.parametrize("idim", [8])
# number of rows and number of cols to add
@pytest.mark.parametrize("pad", [2, 3])
# number of channels
-@pytest.mark.parametrize("num_ch", [1, 2])
+@pytest.mark.parametrize("num_ch", [2, 4])
+# Input parallelism
+@pytest.mark.parametrize("simd", [1, 2])
# PaddingStyle: selects behavior when (odim-idim)%2 != 0
@pytest.mark.parametrize("pad_style", [2])
# FINN input datatype
@@ -76,14 +79,15 @@ def make_single_fmpadding_modelwrapper(idim, padding, num_ch, idt, pad_style):
@pytest.mark.parametrize("mode", ["cppsim", "rtlsim"])
@pytest.mark.slow
@pytest.mark.vivado
-def test_fpgadataflow_fmpadding(idim, pad, num_ch, pad_style, idt, mode):
-
+def test_fpgadataflow_fmpadding(idim, pad, num_ch, simd, pad_style, idt, mode):
+ if num_ch % simd != 0:
+ pytest.skip(" num_ch % simd != 0, skipping")
# generate input data
x = gen_finn_dt_tensor(idt, [1, idim, idim, num_ch])
input_dict = {"inp": x}
odim = idim + pad
- model = make_single_fmpadding_modelwrapper(idim, pad, num_ch, idt, pad_style)
+ model = make_single_fmpadding_modelwrapper(idim, pad, num_ch, simd, idt, pad_style)
model = model.transform(InferShapes())
model = model.transform(SetExecMode(mode))
model = model.transform(GiveUniqueNodeNames())
diff --git a/tests/fpgadataflow/test_fpgadataflow_ip_stitch.py b/tests/fpgadataflow/test_fpgadataflow_ipstitch.py
similarity index 99%
rename from tests/fpgadataflow/test_fpgadataflow_ip_stitch.py
rename to tests/fpgadataflow/test_fpgadataflow_ipstitch.py
index 61dd81b728aafcd8ccc812cf0cb4c27eff00f471..b830693c32afe629dd6fc70868d0bddacac4c887 100644
--- a/tests/fpgadataflow/test_fpgadataflow_ip_stitch.py
+++ b/tests/fpgadataflow/test_fpgadataflow_ipstitch.py
@@ -300,7 +300,7 @@ def test_fpgadataflow_ipstitch_synth_ooc():
assert ret["FF"] > 0
assert ret["DSP"] == 0
assert ret["BRAM"] == 0
- assert ret["fmax_mz"] > 100
+ assert ret["fmax_mhz"] > 100
@pytest.mark.vivado
diff --git a/tests/fpgadataflow/test_fpgadataflow_res_estimate.py b/tests/fpgadataflow/test_fpgadataflow_res_estimate.py
index 38f792ed3cdd52044b28b4c19ac0603da4e502e6..398a17132a2ef6c92e600102ff5c0b71a1f65aaa 100644
--- a/tests/fpgadataflow/test_fpgadataflow_res_estimate.py
+++ b/tests/fpgadataflow/test_fpgadataflow_res_estimate.py
@@ -92,7 +92,7 @@ def test_res_estimate():
model = model.transform(GiveUniqueNodeNames())
prod_resource_estimation = model.analysis(res_estimation)
expect_resource_estimation = {
- "StreamingFCLayer_Batch_0": {"BRAM_18K": 1, "LUT": 304.4}
+ "StreamingFCLayer_Batch_0": {"BRAM_18K": 1, 'BRAM_efficiency': 0.001736111111111111, "LUT": 304.4}
}
assert check_two_dict_for_equality(
diff --git a/tests/pynq/test_pynq_performance_end2end.py b/tests/pynq/test_pynq_performance_end2end.py
index 66a93a190061e0142637be19bb2ea841d192745a..3b6ea86741b8adefce4faaa65b791f1d213cf3ae 100644
--- a/tests/pynq/test_pynq_performance_end2end.py
+++ b/tests/pynq/test_pynq_performance_end2end.py
@@ -10,7 +10,7 @@ from finn.core.throughput_test import throughput_test
build_dir = "/tmp/" + os.environ["FINN_INST_NAME"]
-@pytest.mark.parametrize("end2end_example", ["tfc_w1a1", "cnv_w1a1"])
+@pytest.mark.parametrize("end2end_example", ["tfc_w1a1", "cnv_w1a1", "cnv_w2a2"])
@pytest.mark.slow
def test_pynq_performance_end2end(end2end_example):
model = load_test_checkpoint_or_skip(
diff --git a/tests/transformation/streamline/test_streamline_cnv.py b/tests/transformation/streamline/test_streamline_cnv.py
index 56dcd26076ec0a5fba6e9be6acac7f5e13572c3d..103967dfb6b86cc6e2ce2bc9ab78249d8945d47d 100644
--- a/tests/transformation/streamline/test_streamline_cnv.py
+++ b/tests/transformation/streamline/test_streamline_cnv.py
@@ -44,9 +44,9 @@ from finn.transformation.double_to_single_float import DoubleToSingleFloat
export_onnx_path = make_build_dir("test_streamline_cnv_")
# act bits
-@pytest.mark.parametrize("abits", [1])
+@pytest.mark.parametrize("abits", [1, 2])
# weight bits
-@pytest.mark.parametrize("wbits", [1])
+@pytest.mark.parametrize("wbits", [1, 2])
# network topology / size
@pytest.mark.parametrize("size", ["CNV"])
def test_streamline_cnv(size, wbits, abits):
@@ -74,6 +74,7 @@ def test_streamline_cnv(size, wbits, abits):
# model.save("orig_cnv.onnx")
model = model.transform(Streamline())
# model.save("streamlined_cnv.onnx")
+ assert len(model.graph.node) == 23
produced_ctx = oxe.execute_onnx(model, input_dict, True)
produced = produced_ctx[model.graph.output[0].name]
assert np.isclose(expected, produced, atol=1e-3).all()
diff --git a/tests/transformation/test_absorb_mul_into_topk.py b/tests/transformation/test_absorb_mul_into_topk.py
new file mode 100644
index 0000000000000000000000000000000000000000..1394220f7c336ccea8fe9c494734c4175bf2e847
--- /dev/null
+++ b/tests/transformation/test_absorb_mul_into_topk.py
@@ -0,0 +1,108 @@
+# 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.
+import pytest
+
+import numpy as np
+from onnx import TensorProto, helper
+
+from finn.core.modelwrapper import ModelWrapper
+from finn.transformation.infer_shapes import InferShapes
+from finn.transformation.infer_datatypes import InferDataTypes
+from finn.transformation.general import GiveUniqueNodeNames, GiveReadableTensorNames
+from finn.transformation.insert_topk import InsertTopK
+from finn.transformation.streamline.absorb import AbsorbScalarMulIntoTopK
+import finn.core.onnx_exec as oxe
+
+# parameter to indicate if mul parameter is negative or positive
+@pytest.mark.parametrize("mul_positive", [True, False])
+# parameter to indicate if mul parameter is scalar or not
+@pytest.mark.parametrize("scalar", [True, False])
+def test_absorb_mul_into_topk(mul_positive, scalar):
+ if scalar is True:
+ shape = [1]
+ else:
+ shape = [1, 1, 1, 1000]
+ inp = helper.make_tensor_value_info("inp", TensorProto.FLOAT, [1, 1, 1, 1000])
+ a0 = helper.make_tensor_value_info("a0", TensorProto.FLOAT, shape)
+ outp = helper.make_tensor_value_info("outp", TensorProto.FLOAT, [1, 1, 1, 1000])
+
+ mul_node = helper.make_node("Mul", ["inp", "a0"], ["outp"])
+ mul_graph = helper.make_graph(
+ nodes=[mul_node],
+ name="mul-graph",
+ inputs=[inp],
+ outputs=[outp],
+ value_info=[a0],
+ )
+
+ model = helper.make_model(mul_graph, producer_name="mul_model")
+ model = ModelWrapper(model)
+ # initialize values
+ if mul_positive is True:
+ a0_values = np.random.uniform(low=0.1, high=1, size=tuple(shape)).astype(
+ np.float32
+ )
+ else:
+ a0_values = np.random.uniform(low=-1, high=-0.1, size=tuple(shape)).astype(
+ np.float32
+ )
+ model.set_initializer("a0", a0_values)
+ model = model.transform(InsertTopK())
+ model = model.transform(InferShapes())
+ model = model.transform(InferDataTypes())
+ model = model.transform(GiveUniqueNodeNames())
+ model = model.transform(GiveReadableTensorNames())
+ model_transformed = model.transform(AbsorbScalarMulIntoTopK())
+
+ # compare execution results
+ inp_values = np.random.uniform(low=-10, high=10, size=(1, 1, 1, 1000)).astype(
+ np.float32
+ )
+ idict = {"global_in": inp_values}
+ odict = oxe.execute_onnx(model, idict, True)
+ y_indices = odict["global_out"]
+ y_values = odict["TopK_0_out0"]
+ odict = oxe.execute_onnx(model_transformed, idict, True)
+ y_tr_indices = odict["global_out"]
+ y_tr_values = odict["TopK_0_out0"]
+
+ # the indices stay the same, if the model is transformed or not
+ assert (y_indices == y_tr_indices).all()
+
+ if scalar is True and mul_positive is True:
+ # the values change if the model was transformed
+ assert (y_values != y_tr_values).all()
+
+ # check for new order
+ assert model.graph != model_transformed.graph
+ assert len(model.graph.node) - 1 == len(model_transformed.graph.node)
+ assert model_transformed.graph.node[0].op_type == "TopK"
+
+ else:
+ assert (y_values == y_tr_values).all()
+ assert model.graph == model_transformed.graph
diff --git a/tests/transformation/test_change_datalayout.py b/tests/transformation/test_change_datalayout.py
new file mode 100644
index 0000000000000000000000000000000000000000..66459d574957575e61ec1bec631fb7030a27cca1
--- /dev/null
+++ b/tests/transformation/test_change_datalayout.py
@@ -0,0 +1,112 @@
+# 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.
+import pytest
+from onnx import helper, TensorProto
+
+from finn.custom_op.maxpoolnhwc import compute_pool_output_dim
+from finn.core.modelwrapper import ModelWrapper
+from finn.core.datatype import DataType
+import finn.core.data_layout as DataLayout
+from finn.transformation.change_datalayout import ChangeDataLayoutQuantAvgPool2d
+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 GiveUniqueNodeNames, GiveReadableTensorNames
+from finn.util.basic import gen_finn_dt_tensor
+from finn.util.basic import get_by_name
+import finn.core.onnx_exec as oxe
+
+# stride
+@pytest.mark.parametrize("s", [1, 2])
+# kernel
+@pytest.mark.parametrize("k", [3, 4])
+# ibits
+@pytest.mark.parametrize("ibits", [4, 8])
+# obits
+@pytest.mark.parametrize("obits", [2, 4])
+# signed
+@pytest.mark.parametrize("signed", [False, True])
+# channels
+@pytest.mark.parametrize("c", [2, 3])
+# input dimension
+@pytest.mark.parametrize("idim", [6, 7])
+def test_change_datalayout_quantavgpool(s, k, ibits, obits, signed, c, idim):
+ n = 1
+ odim = compute_pool_output_dim(idim, k, s)
+ # determine input FINN datatype
+ if signed is True:
+ prefix = "INT"
+ else:
+ prefix = "UINT"
+ dt_name = prefix + str(ibits)
+ dtype = DataType[dt_name]
+
+ inp = helper.make_tensor_value_info("inp", TensorProto.FLOAT, [n, c, idim, idim])
+ outp = helper.make_tensor_value_info("outp", TensorProto.FLOAT, [n, c, odim, odim])
+
+ node = helper.make_node(
+ "QuantAvgPool2d",
+ ["inp"],
+ ["outp"],
+ domain="finn",
+ stride=s,
+ kernel=k,
+ ibits=ibits,
+ obits=obits,
+ signed=signed,
+ data_layout="NCHW",
+ )
+ graph = helper.make_graph(
+ nodes=[node], name="single-quantavgpool", inputs=[inp], outputs=[outp]
+ )
+
+ model = helper.make_model(graph)
+ model = ModelWrapper(model)
+ model = model.transform(InferShapes())
+ model = model.transform(InferDataTypes())
+ model = model.transform(InferDataLayouts())
+ model = model.transform(GiveUniqueNodeNames())
+ model = model.transform(GiveReadableTensorNames())
+ model_transformed = model.transform(ChangeDataLayoutQuantAvgPool2d())
+ model_transformed = model_transformed.transform(InferShapes())
+ model_transformed = model_transformed.transform(InferDataTypes())
+ model_transformed = model_transformed.transform(InferDataLayouts())
+ model_transformed = model_transformed.transform(GiveUniqueNodeNames())
+ model_transformed = model_transformed.transform(GiveReadableTensorNames())
+ inp_values = gen_finn_dt_tensor(dtype, [n, c, idim, idim])
+ idict = {"inp": inp_values}
+ assert oxe.compare_execution(model, model_transformed, idict)
+ assert len(model.graph.node) + 2 == len(model_transformed.graph.node)
+ assert model_transformed.graph.node[-1].op_type == "Transpose"
+ assert model_transformed.graph.node[0].op_type == "Transpose"
+ # check if QuantAvgPool2d node has datalayout set correctly
+ node = model_transformed.graph.node[1]
+ d_layout = get_by_name(node.attribute, "data_layout").s.decode("UTF-8")
+ assert d_layout == "NHWC"
+ assert model_transformed.get_tensor_layout(node.input[0]) == DataLayout.NHWC
+ assert model_transformed.get_tensor_layout(node.output[0]) == DataLayout.NHWC
diff --git a/tests/transformation/test_conv_lowering.py b/tests/transformation/test_conv_lowering.py
index 2cbc8e558940517168678b05c3bb46af8170abce..73891ded1b9691c7c48a2075ad6ca4668fcf6bfe 100644
--- a/tests/transformation/test_conv_lowering.py
+++ b/tests/transformation/test_conv_lowering.py
@@ -26,12 +26,13 @@
# 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.
+import onnx.helper as oh
+from onnx import TensorProto
import os
import pkg_resources as pk
import brevitas.onnx as bo
import numpy as np
-
from finn.core.modelwrapper import ModelWrapper
from finn.transformation.fold_constants import FoldConstants
from finn.transformation.infer_shapes import InferShapes
@@ -65,3 +66,51 @@ def test_conv_lowering_cnv_w1a1():
assert np.isclose(produced, expected).all()
assert np.argmax(produced) == 3
os.remove(export_onnx_path)
+
+
+def test_conv_lowering_conv_1x1():
+ np.random.seed(0)
+
+ in_feature_dim = 7
+ in_chn = 3
+ kernel_size = 1
+ out_feature_dim = in_feature_dim
+
+ input_shape = [1, in_chn, in_feature_dim, in_feature_dim]
+ output_shape = [1, in_chn, out_feature_dim, out_feature_dim]
+
+ conv_param_shape = [in_chn, in_chn, kernel_size, kernel_size]
+
+ conv_config = {}
+ conv_config["dilations"] = [1, 1]
+ conv_config["group"] = 1
+ conv_config["kernel_shape"] = [kernel_size, kernel_size]
+ conv_config["pads"] = [0, 0, 0, 0]
+ conv_config["strides"] = [1, 1]
+
+ top_in = oh.make_tensor_value_info("top_in", TensorProto.FLOAT, input_shape)
+ top_out = oh.make_tensor_value_info("top_out", TensorProto.FLOAT, output_shape)
+
+ value_info = [oh.make_tensor_value_info("p1", TensorProto.FLOAT, conv_param_shape)]
+
+ modelproto = oh.make_model(
+ oh.make_graph(
+ name="test",
+ inputs=[top_in],
+ outputs=[top_out],
+ value_info=value_info,
+ nodes=[oh.make_node("Conv", ["top_in", "p1"], ["top_out"], **conv_config)],
+ )
+ )
+ model = ModelWrapper(modelproto)
+ model = model.transform(InferShapes())
+ model.set_initializer("p1", np.random.rand(*conv_param_shape).astype(np.float32))
+
+ new_model = model.transform(LowerConvsToMatMul())
+ inp_dict = {"top_in": np.random.rand(*input_shape).astype(np.float32)}
+
+ assert oxe.compare_execution(model, new_model, inp_dict)
+ assert new_model.graph.node[0].op_type == "Transpose"
+ assert new_model.graph.node[1].op_type == "MatMul"
+ assert new_model.graph.node[2].op_type == "Transpose"
+ assert len(new_model.graph.node) == 3
diff --git a/tests/transformation/test_move_maxpool_past_multithreshold.py b/tests/transformation/test_move_maxpool_past_multithreshold.py
new file mode 100644
index 0000000000000000000000000000000000000000..2fc19debf8d6fc89d15e3d731f1e54daa491c321
--- /dev/null
+++ b/tests/transformation/test_move_maxpool_past_multithreshold.py
@@ -0,0 +1,100 @@
+from onnx import TensorProto, helper
+import numpy as np
+
+import finn.core.onnx_exec as oxe
+from finn.core.modelwrapper import ModelWrapper
+from finn.transformation.streamline.reorder import MoveMaxPoolPastMultiThreshold
+from finn.transformation.infer_shapes import InferShapes
+from finn.transformation.infer_datatypes import InferDataTypes
+
+
+def get_multithreshold_rand_params(channels, num_of_thres, seed=None):
+ if seed is not None:
+ np.random.seed(seed)
+ steps = np.random.rand(channels, 1) * 2
+ bias = np.random.rand(channels, 1) * 10
+ thres = [np.arange(num_of_thres) for chn in range(channels)]
+ thres = ((thres - bias) * steps).astype(np.float32)
+ return thres
+
+
+def test_move_maxpool_past_multithreshold():
+ # generate test vectors of correct shape
+ ch = 64
+ ifmdim = 16
+ ofmdim = 16 // 4
+ input_shape = (1, ch, ifmdim, ifmdim)
+ output_shape = (1, ch, ofmdim, ofmdim)
+
+ top_in = helper.make_tensor_value_info("top_in", TensorProto.FLOAT, input_shape)
+ top_out = helper.make_tensor_value_info("top_out", TensorProto.FLOAT, output_shape)
+
+ maxpool_config = {}
+ maxpool_config["pads"] = [1, 1, 1, 1]
+ maxpool_config["kernel_shape"] = [3, 3]
+ maxpool_config["strides"] = [2, 2]
+
+ value_info = []
+ thres1_shape = [1, 1]
+ value_info += [
+ helper.make_tensor_value_info("thres1", TensorProto.FLOAT, thres1_shape)
+ ]
+
+ thres2_shape = [ch, 14]
+ value_info += [
+ helper.make_tensor_value_info("thres2", TensorProto.FLOAT, thres2_shape)
+ ]
+
+ nodes = []
+ nodes += [helper.make_node("MaxPool", ["top_in"], ["t1"], **maxpool_config)]
+ nodes += [
+ helper.make_node(
+ "MultiThreshold",
+ ["t1", "thres1"],
+ ["t2"],
+ domain="finn",
+ out_dtype="BIPOLAR",
+ out_bias=-1.0,
+ out_scale=1.0,
+ )
+ ]
+ nodes += [helper.make_node("MaxPool", ["t2"], ["t3"], **maxpool_config)]
+ nodes += [
+ helper.make_node(
+ "MultiThreshold",
+ ["t3", "thres2"],
+ ["top_out"],
+ domain="finn",
+ out_dtype="UINT4",
+ )
+ ]
+
+ modelproto = helper.make_model(
+ helper.make_graph(
+ name="test",
+ inputs=[top_in],
+ outputs=[top_out],
+ value_info=value_info,
+ nodes=nodes,
+ )
+ )
+ model = ModelWrapper(modelproto)
+ model = model.transform(InferShapes())
+ model = model.transform(InferDataTypes())
+
+ model.set_initializer("thres1", np.array([[0]]))
+ model.set_initializer(
+ "thres2", get_multithreshold_rand_params(*thres2_shape, seed=0)
+ )
+
+ # Transform
+ new_model = model.transform(MoveMaxPoolPastMultiThreshold())
+ inp_dict = {"top_in": np.random.rand(*input_shape).astype(np.float32)}
+
+ # Test
+ assert oxe.compare_execution(model, new_model, inp_dict)
+ assert new_model.graph.node[0].op_type == "MaxPool"
+ assert new_model.graph.node[1].op_type == "MultiThreshold"
+ assert new_model.graph.node[2].op_type == "MultiThreshold"
+ assert new_model.graph.node[3].op_type == "MaxPool"
+ assert len(new_model.graph.node) == 4
diff --git a/tests/transformation/test_move_mul_past_dw_conv.py b/tests/transformation/test_move_mul_past_dw_conv.py
new file mode 100644
index 0000000000000000000000000000000000000000..1ae8fbfe89986d58d3d71f5f8735a98469d9d1e3
--- /dev/null
+++ b/tests/transformation/test_move_mul_past_dw_conv.py
@@ -0,0 +1,93 @@
+import pytest
+
+from onnx import helper, TensorProto
+from finn.custom_op.im2col import compute_conv_output_dim
+import finn.core.onnx_exec as oxe
+from finn.core.datatype import DataType
+from finn.core.modelwrapper import ModelWrapper
+from finn.transformation.infer_datatypes import InferDataTypes
+from finn.transformation.infer_shapes import InferShapes
+from finn.util.basic import gen_finn_dt_tensor
+from finn.transformation.streamline.reorder import MoveMulPastDWConv
+
+
+# input dimension
+@pytest.mark.parametrize("ifm_dim", [4, 7])
+# input channels
+@pytest.mark.parametrize("ifm_ch", [2, 3])
+# kernel size
+@pytest.mark.parametrize("k", [2, 3])
+# stride
+@pytest.mark.parametrize("stride", [1, 2])
+# padding
+@pytest.mark.parametrize("pad_amt", [0, 1])
+# depthwise
+@pytest.mark.parametrize("dw", [0, 1])
+def test_move_mul_past_dw_conv(ifm_dim, ifm_ch, k, stride, pad_amt, dw):
+ if dw == 1:
+ ofm_ch = ifm_ch
+ groups = ifm_ch
+ W_shape = [ofm_ch, 1, k, k]
+ else:
+ ofm_ch = ifm_ch + 2
+ groups = 1
+ W_shape = [ofm_ch, ifm_ch, k, k]
+
+ ofm_dim = compute_conv_output_dim(ifm_dim, k, stride, pad_amt)
+
+ # set up onnx model
+ inp = helper.make_tensor_value_info(
+ "inp", TensorProto.FLOAT, [1, ifm_ch, ifm_dim, ifm_dim]
+ )
+ mul = helper.make_tensor_value_info("mul", TensorProto.FLOAT, [1, ifm_ch, 1, 1])
+ W = helper.make_tensor_value_info("W", TensorProto.FLOAT, W_shape)
+ outp = helper.make_tensor_value_info(
+ "outp", TensorProto.FLOAT, [1, ofm_ch, ofm_dim, ofm_dim]
+ )
+
+ Mul_node = helper.make_node("Mul", ["inp", "mul"], ["mul_out"])
+
+ Conv_node = helper.make_node(
+ "Conv",
+ ["mul_out", "W"],
+ ["outp"],
+ group=groups,
+ kernel_shape=[k, k],
+ pads=[pad_amt, pad_amt, pad_amt, pad_amt],
+ strides=[stride, stride],
+ )
+
+ graph = helper.make_graph(
+ nodes=[Mul_node, Conv_node],
+ name="mulpastconv_graph",
+ inputs=[inp],
+ outputs=[outp],
+ value_info=[mul, W],
+ )
+
+ model = helper.make_model(graph, producer_name="mulpastconv-model")
+ model = ModelWrapper(model)
+ inp_values = gen_finn_dt_tensor(DataType.INT2, [1, ifm_ch, ifm_dim, ifm_dim])
+ mul_values = gen_finn_dt_tensor(DataType.INT2, [1, ifm_ch, 1, 1])
+ W_values = gen_finn_dt_tensor(DataType.INT2, W_shape)
+ model.set_initializer("W", W_values)
+ model.set_initializer("mul", mul_values)
+ model = model.transform(InferShapes())
+ model = model.transform(InferDataTypes())
+ idict = {"inp": inp_values}
+ odict = oxe.execute_onnx(model, idict, True)
+ out_before = odict["outp"]
+
+ # move channelwise multiplication past depthwise conv
+ model_transformed = model.transform(MoveMulPastDWConv())
+ odict = oxe.execute_onnx(model_transformed, idict, True)
+ out_after = odict["outp"]
+
+ assert (out_before == out_after).all()
+
+ if dw == 0:
+ assert model.graph.node[0].op_type == model_transformed.graph.node[0].op_type
+ assert model.graph.node[1].op_type == model_transformed.graph.node[1].op_type
+ else:
+ assert model.graph.node[0].op_type == model_transformed.graph.node[1].op_type
+ assert model.graph.node[1].op_type == model_transformed.graph.node[0].op_type
diff --git a/tests/transformation/test_remove_identity_ops.py b/tests/transformation/test_remove_identity_ops.py
new file mode 100644
index 0000000000000000000000000000000000000000..536c1ab0b48fa44388da23f45b528da3c5f3b2f2
--- /dev/null
+++ b/tests/transformation/test_remove_identity_ops.py
@@ -0,0 +1,81 @@
+import pytest
+
+import numpy as np
+from onnx import helper, TensorProto
+import finn.core.onnx_exec as oxe
+from finn.core.datatype import DataType
+from finn.core.modelwrapper import ModelWrapper
+from finn.transformation.infer_datatypes import InferDataTypes
+from finn.transformation.infer_shapes import InferShapes
+from finn.transformation.streamline.remove import RemoveIdentityOps
+from finn.util.basic import gen_finn_dt_tensor
+
+
+def insert_identity_op(model, op):
+ if op in ["Add", "Sub"]:
+ val = np.asarray([0.0], dtype=np.float32)
+ elif op in ["Mul", "Div"]:
+ val = np.asarray([1.0], dtype=np.float32)
+ else:
+ return
+
+ identity_node = helper.make_node(op, ["div_out", "value"], ["ident_out"])
+ graph = model.graph
+ graph.node.insert(3, identity_node)
+ graph.node[-1].input[0] = "ident_out"
+ model.set_initializer("value", val)
+
+ return model
+
+
+# identity operations to be inserted
+@pytest.mark.parametrize("op", ["Add", "Sub", "Mul", "Div"])
+def test_remove_identity_ops(op):
+
+ # set up onnx model
+ inp = helper.make_tensor_value_info("inp", TensorProto.FLOAT, [1, 4, 1, 1])
+ mul = helper.make_tensor_value_info("mul", TensorProto.FLOAT, [])
+ shape = helper.make_tensor_value_info("shape", TensorProto.FLOAT, [2])
+ div = helper.make_tensor_value_info("div", TensorProto.FLOAT, [])
+ matmul = helper.make_tensor_value_info("matmul", TensorProto.FLOAT, [4, 2])
+ outp = helper.make_tensor_value_info("outp", TensorProto.FLOAT, [1, 2])
+
+ mul_node = helper.make_node("Mul", ["inp", "mul"], ["mul_out"])
+ reshape_node = helper.make_node("Reshape", ["mul_out", "shape"], ["reshape_out"])
+ div_node = helper.make_node("Div", ["reshape_out", "div"], ["div_out"])
+ matmul_node = helper.make_node("MatMul", ["div_out", "matmul"], ["outp"])
+
+ graph = helper.make_graph(
+ nodes=[mul_node, reshape_node, div_node, matmul_node],
+ name="identity-graph",
+ inputs=[inp],
+ outputs=[outp],
+ value_info=[mul, shape, div, matmul],
+ )
+
+ model = helper.make_model(graph, producer_name="mulpastconv-model")
+ model = ModelWrapper(model)
+ inp_values = gen_finn_dt_tensor(DataType.INT2, [1, 4, 1, 1])
+ mul_values = np.random.uniform(low=0.1, high=0.99, size=(1)).astype(np.float32)
+ shape_values = np.asarray([1, -1], dtype=np.int64)
+ div_values = np.random.uniform(low=0.1, high=0.99, size=(1)).astype(np.float32)
+ matmul_values = gen_finn_dt_tensor(DataType.INT2, [4, 2])
+ model.set_initializer("mul", mul_values)
+ model.set_initializer("shape", shape_values)
+ model.set_initializer("div", div_values)
+ model.set_initializer("matmul", matmul_values)
+ insert_identity_op(model, op)
+ model = model.transform(InferShapes())
+ model = model.transform(InferDataTypes())
+ idict = {"inp": inp_values}
+ odict = oxe.execute_onnx(model, idict)
+ out_before = odict["outp"]
+ num_of_nodes_before = len(model.graph.node)
+
+ model = model.transform(RemoveIdentityOps())
+ num_of_nodes_after = len(model.graph.node)
+ assert num_of_nodes_before - 1 == num_of_nodes_after
+
+ odict = oxe.execute_onnx(model, idict)
+ out_after = odict["outp"]
+ assert (out_before == out_after).all()