diff --git a/.github/workflows/pre-commit.yml b/.github/workflows/pre-commit.yml
index 2fbb9265beb49644f08a2c6e916ab9c23d4bd339..20f5b48f7acc65ab18702ef2509e9791f919b825 100644
--- a/.github/workflows/pre-commit.yml
+++ b/.github/workflows/pre-commit.yml
@@ -13,10 +13,10 @@ jobs:
steps:
- name: Checkout
- uses: actions/checkout@v2
+ uses: actions/checkout@v3
- name: Setup Python
- uses: actions/setup-python@v2
+ uses: actions/setup-python@v3
- name: Run Lint
- uses: pre-commit/action@v2.0.0
+ uses: pre-commit/action@v3.0.0
diff --git a/.github/workflows/quicktest-dev-pr.yml b/.github/workflows/quicktest-dev-pr.yml
index d188007465cd27662ffadfb3ece0d8bf2e8e28be..ec92c84665d868b8a4376c82ecdf72395f1367a8 100644
--- a/.github/workflows/quicktest-dev-pr.yml
+++ b/.github/workflows/quicktest-dev-pr.yml
@@ -17,37 +17,9 @@ jobs:
- name: checkout
uses: actions/checkout@v2
- - name: set up Docker Buildx
- uses: docker/setup-buildx-action@v1
-
- - name: cache Docker layers
- uses: actions/cache@v2
- with:
- path: /tmp/.buildx-cache
- key: ${{ runner.os }}-buildx-${{ github.sha }}
- restore-keys: |
- ${{ runner.os }}-buildx-
-
- - name: Build and push
- uses: docker/build-push-action@v2
- with:
- file: docker/Dockerfile.finn
- context: .
- push: false
- load: true
- tags: finn_gha
- cache-from: type=local,src=/tmp/.buildx-cache
- cache-to: type=local,dest=/tmp/.buildx-cache-new
- -
- # Temp fix
- # https://github.com/docker/build-push-action/issues/252
- # https://github.com/moby/buildkit/issues/1896
- name: Move cache
- run: |
- rm -rf /tmp/.buildx-cache
- mv /tmp/.buildx-cache-new /tmp/.buildx-cache
-
-
- name: DockerRunQuicktest
run: |
- docker run --init --hostname finn_gha -w $(pwd) -v $(pwd):$(pwd) -e FINN_BUILD_DIR=/tmp/finn_gha -e FINN_INST_NAME=finn_gha finn_gha quicktest.sh
+ export FINN_ROOT=$(pwd)
+ export FINN_BUILD_DIR=/tmp/finn_gha
+ export FINN_INST_NAME=finn_gha
+ ./run-docker.sh quicktest
diff --git a/AUTHORS.rst b/AUTHORS.rst
index 1d42d35a3b269176fcab79d8239b84ac8442fa43..d011ce3d7ad74125b7013b7a7e987eb22e70a9f3 100644
--- a/AUTHORS.rst
+++ b/AUTHORS.rst
@@ -2,8 +2,9 @@
Contributors
============
-* Yaman Umuroglu (@maltanar) (maintainer)
-* Jakoba Petri-Koenig (@auphelia)
+* Jakoba Petri-Koenig (@auphelia) (maintainer)
+* Thomas Preusser (@preusser)
+* Yaman Umuroglu (@maltanar)
* Andrea Rigoni (@AndreaRigoni)
* Hendrik Borras (@HenniOVP)
* Lucian Petrica (@quetric)
@@ -22,3 +23,6 @@ Contributors
* Javier Duarte (@jmduarte)
* Uma Maheshwari (@umav1511)
* José Rosa (@pinxau1000)
+* Aziz Bahri (@azizb-xlnx)
+* Fionn O'Donohoe (@fionnodonohoe-xlnx)
+* Matthias Gehre (@mgehre-amd)
diff --git a/CONTRIBUTING.md b/CONTRIBUTING.md
index deed89651db34d3821df35c8a1eb0f85b72f23a5..d376a1b42b0f1f3856f40b3993533785fb254a9b 100644
--- a/CONTRIBUTING.md
+++ b/CONTRIBUTING.md
@@ -2,7 +2,7 @@ We welcome contributions to FINN.
Please follow the steps below and be sure that your contribution complies with our guidelines.
-1. Share your proposal via <a href="https://github.com/Xilinx/finn/issues" target="_blank">Github issues</a>. If you are looking for some issues to get started with, we have a list of <a href="https://github.com/Xilinx/finn/labels/good%20first%20issue">good first issues</a> in the issue tracker. Feel free to ask questions on the <a href="https://gitter.im/xilinx-finn/community?utm_source=badge&utm_medium=badge&utm_campaign=pr-badge">FINN gitter channel as well</a>.
+1. Share your proposal via <a href="https://github.com/Xilinx/finn/issues" target="_blank">Github issues</a>. If you are looking for some issues to get started with, we have a list of <a href="https://github.com/Xilinx/finn/labels/good%20first%20issue">good first issues</a> in the issue tracker. Feel free to ask questions in the <a href="https://github.com/Xilinx/finn/discussions">FINN GitHub discussions</a> as well.
We welcome submissions to:
@@ -31,4 +31,4 @@ Please follow the steps below and be sure that your contribution complies with o
3. We will review your contribution and, if any additional fixes or modifications are
necessary, may provide feedback to guide you. When accepted, your pull request will
-be merged to the repository. If you have more questions please contact us via the <a href="https://gitter.im/xilinx-finn/community" target="_blank">FINN gitter channel</a>.
+be merged to the repository. If you have more questions please contact us.
diff --git a/README.md b/README.md
index 4cc995fc8c991ccc851e95fd30897aeea8ca266a..1b8efc8f19d0b664a17320585f5ea60acbe03eb4 100644
--- a/README.md
+++ b/README.md
@@ -24,9 +24,7 @@ Please see the [Getting Started](https://finn.readthedocs.io/en/latest/getting_s
## What's New in FINN?
-* **2021-11-05:** v0.7 is released, introducing QONNX support, three new example networks and many other improvements. Read more on the [v0.7 release blog post](https://xilinx.github.io/finn//2021/11/05/finn-v07-is-released.html).
-* **2021-06-15:** v0.6 is released, with ResNet-50 on U250 and ZCU104 MobileNet-v1 in finn-examples showcasing new features plus a lot more. Read more on the [v0.6 release blog post](https://xilinx.github.io/finn//2021/06/15/finn-v06-is-released.html).
-* **2020-12-17:** v0.5b (beta) is released, with a new [examples repo](https://github.com/Xilinx/finn-examples) including MobileNet-v1. Read more on the <a href="https://xilinx.github.io/finn/2020/12/17/finn-v05b-beta-is-released.html">release blog post</a>.
+* Please find all news under [GitHub discussions Announcements](https://github.com/Xilinx/finn/discussions/categories/announcements).
## Documentation
diff --git a/docker/Dockerfile.finn b/docker/Dockerfile.finn
index 1a1ae4b4f4d55dff231ccda6c3885ecef436b0b8..da88c7fbfa5fd3454cd7945dd5febd31736db0ce 100644
--- a/docker/Dockerfile.finn
+++ b/docker/Dockerfile.finn
@@ -28,6 +28,9 @@
FROM pytorch/pytorch:1.7.1-cuda11.0-cudnn8-runtime
LABEL maintainer="Yaman Umuroglu <yamanu@xilinx.com>"
+
+ARG XRT_DEB_VERSION="xrt_202210.2.13.466_18.04-amd64-xrt"
+
WORKDIR /workspace
# some Vitis deps require a timezone to be specified, which hangs in Docker
@@ -49,7 +52,6 @@ RUN apt-get update && \
libsm6 \
libxext6 \
libxrender-dev \
- verilator \
nano \
zsh \
rsync \
@@ -65,6 +67,16 @@ RUN apt-get update && \
RUN echo "StrictHostKeyChecking no" >> /etc/ssh/ssh_config
RUN locale-gen "en_US.UTF-8"
+# install Verilator from source to get the right version
+RUN apt-get install -y git perl python3 make autoconf g++ flex bison ccache libgoogle-perftools-dev numactl perl-doc libfl2 libfl-dev zlibc zlib1g zlib1g-dev
+RUN git clone https://github.com/verilator/verilator
+RUN cd verilator && \
+ git checkout v4.012 && \
+ autoconf && \
+ ./configure && \
+ make -j4 && \
+ make install
+
# install XRT
RUN wget https://www.xilinx.com/bin/public/openDownload?filename=$XRT_DEB_VERSION.deb -O /tmp/$XRT_DEB_VERSION.deb
RUN apt install -y /tmp/$XRT_DEB_VERSION.deb
diff --git a/docker/finn_entrypoint.sh b/docker/finn_entrypoint.sh
index 556e6d040db3140916d75632b9bdead3c1d38747..b5c702111ac706df8d29947ce8768e499ee3627a 100644
--- a/docker/finn_entrypoint.sh
+++ b/docker/finn_entrypoint.sh
@@ -113,5 +113,6 @@ else
yecho "If you need Vitis HLS, ensure HLS_PATH is set correctly and mounted into the Docker container."
fi
+export PATH=$PATH:$HOME/.local/bin
# execute the provided command(s) as root
exec "$@"
diff --git a/docker/jenkins/Jenkinsfile b/docker/jenkins/Jenkinsfile
index ad533efa5d8bbab68837e6092f91c4767cde60f7..e3e5b5f7f93c312269f2c96942e44318875903e0 100644
--- a/docker/jenkins/Jenkinsfile
+++ b/docker/jenkins/Jenkinsfile
@@ -9,7 +9,7 @@ node {
"FINN_XILINX_VERSION=2022.1",
"FINN_DOCKER_TAG=xilinx/finn:jenkins",
"FINN_HOST_BUILD_DIR=/scratch/users/finn_ci",
- "PLATFORM_REPO_PATHS=/opt/xilinx/dsa"
+ "PLATFORM_REPO_PATHS=/opt/xilinx/platforms"
]){
parallel firstBranch: {
stage('Brevitas export') {
diff --git a/docker/quicktest.sh b/docker/quicktest.sh
index f625f2b1ef722f386180a8409a9eb9e759a2f3b6..b4ad37232fa69754a86e9064d7592d7474e8617e 100755
--- a/docker/quicktest.sh
+++ b/docker/quicktest.sh
@@ -2,7 +2,7 @@
: ${PYTEST_PARALLEL=auto}
-cd $FINN_ROOT/finn
+cd $FINN_ROOT
# check if command line argument is empty or not present
if [ -z $1 ]; then
echo "Running quicktest: not (vivado or slow or board) with pytest-xdist"
diff --git a/fetch-repos.sh b/fetch-repos.sh
index 88f0a3822a36df7d5ff3a86df31f5f3e9bb2181c..fb00faccea87abf56d6e2fce6d5c5c92af57199a 100755
--- a/fetch-repos.sh
+++ b/fetch-repos.sh
@@ -27,13 +27,13 @@
# OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
-QONNX_COMMIT="4a4826641db8d34619d31eac155fe95af11692eb"
+QONNX_COMMIT="92184fea2dd417bc7a53c82811fef271e4833c4c"
FINN_EXP_COMMIT="9cbd2787b5160e2b44e0e8164a0df1457dbd5366"
BREVITAS_COMMIT="a5b71d6de1389d3e7db898fef72e014842670f03"
PYVERILATOR_COMMIT="64b8294ff1afebb47be76fcad6ae87027e0402c2"
CNPY_COMMIT="4e8810b1a8637695171ed346ce68f6984e585ef4"
-HLSLIB_COMMIT="e9946e5e56acd85837e8e79224d2bb60764bed69"
-OMX_COMMIT="a97f0bf145a2f7e57ca416ea76c9e45df4e9aa37"
+HLSLIB_COMMIT="79d7c61fbe318bfcd56e3c35bbfb774995a7870c"
+OMX_COMMIT="d1065a788219ca0eb54d5e57600b1f9d7f67d4cc"
AVNET_BDF_COMMIT="2d49cfc25766f07792c0b314489f21fe916b639b"
XIL_BDF_COMMIT="8cf4bb674a919ac34e3d99d8d71a9e60af93d14e"
EXP_BOARD_FILES_MD5="30eecc497c31050bd46d10ea20eba232"
diff --git a/requirements.txt b/requirements.txt
index e1c65bbf923a3561832d2d5296d9894ff9419855..970acc342bb7984e69929d1ef5eaa027b765ced0 100644
--- a/requirements.txt
+++ b/requirements.txt
@@ -4,7 +4,7 @@ dataclasses-json==0.5.7
docrep==0.2.7
future==0.18.2
gspread==3.6.0
-numpy==1.18.0
+numpy==1.22.0
onnx==1.11.0
onnxoptimizer
onnxruntime==1.11.1
@@ -13,6 +13,7 @@ protobuf==3.20.1
pyscaffold==3.2.1
scipy==1.5.2
setupext-janitor>=1.1.2
+sigtools==2.0.3
toposort==1.5
vcdvcd==1.0.5
wget==3.2
diff --git a/run-docker.sh b/run-docker.sh
index 95c023ee8795797000425ab421798c1a969e6f38..aab0df1f1c60d894cc26ddde5facfdd93ee18c6a 100755
--- a/run-docker.sh
+++ b/run-docker.sh
@@ -88,7 +88,7 @@ SCRIPTPATH=$(dirname "$SCRIPT")
: ${PLATFORM_REPO_PATHS="/opt/xilinx/platforms"}
: ${XRT_DEB_VERSION="xrt_202210.2.13.466_18.04-amd64-xrt"}
: ${FINN_HOST_BUILD_DIR="/tmp/$DOCKER_INST_NAME"}
-: ${FINN_DOCKER_TAG="xilinx/finn:$(git describe --tags --dirty).$XRT_DEB_VERSION"}
+: ${FINN_DOCKER_TAG="xilinx/finn:$(git describe --always --tags --dirty).$XRT_DEB_VERSION"}
: ${FINN_DOCKER_PREBUILT="0"}
: ${FINN_DOCKER_RUN_AS_ROOT="0"}
: ${FINN_DOCKER_GPU="$(docker info | grep nvidia | wc -m)"}
diff --git a/src/finn/builder/build_dataflow.py b/src/finn/builder/build_dataflow.py
index 238083f653d410772a81115ff12dd987835d1f32..d6864994a70a0ea4c24567155ff7c0599bc0fb6f 100644
--- a/src/finn/builder/build_dataflow.py
+++ b/src/finn/builder/build_dataflow.py
@@ -155,12 +155,14 @@ def build_dataflow_cfg(model_filename, cfg: DataflowBuildConfig):
% (step_name, step_num, len(build_dataflow_steps))
)
# redirect output to logfile
- sys.stdout = stdout_logger
- sys.stderr = stderr_logger
- print(
- "Running step: %s [%d/%d]"
- % (step_name, step_num, len(build_dataflow_steps))
- )
+ if not cfg.verbose:
+ sys.stdout = stdout_logger
+ sys.stderr = stderr_logger
+ # also log current step name to logfile
+ print(
+ "Running step: %s [%d/%d]"
+ % (step_name, step_num, len(build_dataflow_steps))
+ )
# run the step
step_start = time.time()
model = transform_step(model, cfg)
diff --git a/src/finn/builder/build_dataflow_config.py b/src/finn/builder/build_dataflow_config.py
index 09e9ec3a564dc2b459cd1ea3205e541f922b1af0..92263bd82ce291833c6868847876ac7e3b68e6f8 100644
--- a/src/finn/builder/build_dataflow_config.py
+++ b/src/finn/builder/build_dataflow_config.py
@@ -285,6 +285,10 @@ class DataflowBuildConfig:
#: Whether pdb postmortem debuggig will be launched when the build fails
enable_build_pdb_debug: Optional[bool] = True
+ #: When True, all warnings and compiler output will be printed in stdout.
+ #: Otherwise, these will be suppressed and only appear in the build log.
+ verbose: Optional[bool] = False
+
#: If given, only run the steps in the list. If not, run default steps.
#: See `default_build_dataflow_steps` for the default list of steps.
#: When specified:
diff --git a/src/finn/core/throughput_test.py b/src/finn/core/throughput_test.py
index 07eda6aa1d82df0a9f9a01d4f17f7880a8cf8b26..3533fd13399a4ba4392d66af785979afc32cab29 100644
--- a/src/finn/core/throughput_test.py
+++ b/src/finn/core/throughput_test.py
@@ -157,8 +157,8 @@ def throughput_test_rtlsim(model, batchsize=100):
res["cycles"] = cycles
res["runtime[ms]"] = runtime_s * 1000
res["throughput[images/s]"] = batchsize / runtime_s
- res["DRAM_in_bandwidth[Mb/s]"] = i_bytes * 0.000001 / runtime_s
- res["DRAM_out_bandwidth[Mb/s]"] = o_bytes * 0.000001 / runtime_s
+ res["DRAM_in_bandwidth[MB/s]"] = i_bytes * 0.000001 / runtime_s
+ res["DRAM_out_bandwidth[MB/s]"] = o_bytes * 0.000001 / runtime_s
res["fclk[mhz]"] = fclk_mhz
res["N"] = batchsize
diff --git a/src/finn/custom_op/fpgadataflow/hlscustomop.py b/src/finn/custom_op/fpgadataflow/hlscustomop.py
index 9978ab0c7138aa6846a1427cd346c5257e4f8728..b202e95a28a26de3dabc098c2030bafcf840d164 100644
--- a/src/finn/custom_op/fpgadataflow/hlscustomop.py
+++ b/src/finn/custom_op/fpgadataflow/hlscustomop.py
@@ -397,18 +397,20 @@ class HLSCustomOp(CustomOp):
builder.build(code_gen_dir)
self.set_nodeattr("executable_path", builder.executable_path)
- def dynamic_input_to_npy(self, context, count):
+ def dynamic_input_to_npy(self, context, count, target_dir=""):
"""Saves input (given context) into .npy files.
Count indicates the number of inputs that have to be saved."""
node = self.onnx_node
- code_gen_dir = self.get_nodeattr("code_gen_dir_cppsim")
- if code_gen_dir == "":
- raise Exception(
+ if target_dir == "":
+ code_gen_dir = self.get_nodeattr("code_gen_dir_cppsim")
+ if code_gen_dir == "":
+ raise Exception(
+ """
+ Found no codegen dir for this node, did you run the prepare_cppsim transformation?
"""
-Found no codegen dir for this node, did you run the prepare_cppsim transformation?
- """
- )
+ )
+ target_dir = code_gen_dir
# create a npy file for each input of the node (in_ind is input index)
# assuming dynamic inputs start from 0
for in_ind in range(count):
@@ -427,7 +429,7 @@ Found no codegen dir for this node, did you run the prepare_cppsim transformatio
# make copy before saving the array
reshaped_input = reshaped_input.copy()
np.save(
- os.path.join(code_gen_dir, "input_{}.npy".format(in_ind)),
+ os.path.join(target_dir, "input_{}.npy".format(in_ind)),
reshaped_input,
)
diff --git a/src/finn/custom_op/fpgadataflow/streamingmaxpool_batch.py b/src/finn/custom_op/fpgadataflow/streamingmaxpool_batch.py
index d9ffea4d9cd8895fdf55a497e8c7d0e49808ac95..882b40a0aaf542e6dcaf427ca3567ae78394ede5 100755
--- a/src/finn/custom_op/fpgadataflow/streamingmaxpool_batch.py
+++ b/src/finn/custom_op/fpgadataflow/streamingmaxpool_batch.py
@@ -138,14 +138,22 @@ class StreamingMaxPool_Batch(HLSCustomOp):
def get_exp_cycles(self):
# derived from StreamingMaxPool_Batch loop nest
ifm_dim, k, ifm_ch = self.get_1d_attrs_normalized()
- _, _, ofm_dim_w, nf, _ = self.get_folded_output_shape()
+ warnings.warn(
+ """Estimated latency for layer {} can be lower than
+ actual latency!""".format(
+ self.onnx_node.name
+ )
+ )
if self.is_1d():
- exp_cycles = ofm_dim_w * nf * (k[1] + 1)
+ _, _, _, nf, _ = self.get_folded_output_shape()
+ ceil_mode = self.get_nodeattr("CeilMode")
+ ofm_dim = compute_pool_output_dim(ifm_dim[1], k[1], k[1], 0, ceil_mode)
+ exp_cycles = ofm_dim * nf * (k[1] + 1)
return int(exp_cycles)
else:
# TODO: adjust inaccurate formula
- return int(ifm_dim[1] * (ifm_dim[1] + (ifm_dim[1] / k[1])))
+ return int(ifm_dim[1] * ifm_dim[1] * (1 + 1 / (k[1] * k[1])))
def get_instream_width(self):
dt_bits = self.get_input_datatype().bitwidth()
diff --git a/src/finn/custom_op/fpgadataflow/upsampler.py b/src/finn/custom_op/fpgadataflow/upsampler.py
index b62e4f2f6784e8964232efcc9971f0b8bc35ac5d..eb51fe39fc6e7ec84204f9d541a0e47c333bbf43 100644
--- a/src/finn/custom_op/fpgadataflow/upsampler.py
+++ b/src/finn/custom_op/fpgadataflow/upsampler.py
@@ -27,7 +27,6 @@
# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
import numpy as np
-import os
import warnings
from qonnx.core.datatype import DataType
@@ -57,6 +56,8 @@ class UpsampleNearestNeighbour_Batch(HLSCustomOp):
"inputDataType": ("s", True, ""),
# Batch size
"numInputVectors": ("i", False, 1),
+ # Dimensionality mode: 0 = 2D square, 1 = 1D in H dim
+ "DimMode": ("i", False, 0),
}
my_attrs.update(super().get_nodeattr_types())
return my_attrs
@@ -64,21 +65,34 @@ class UpsampleNearestNeighbour_Batch(HLSCustomOp):
def get_exp_cycles(self):
OFMDim = self.get_nodeattr("OFMDim")
batch_size = self.get_nodeattr("numInputVectors")
- exp_cycles = OFMDim * OFMDim * batch_size
+ is_2d = self.get_nodeattr("DimMode") == 0
+ reps = 1
+ if is_2d:
+ OFMDim = OFMDim * OFMDim
+ reps = batch_size
+ exp_cycles = OFMDim * reps
return int(exp_cycles)
def get_normal_input_shape(self):
IFMDim = self.get_nodeattr("IFMDim")
num_ch = self.get_nodeattr("NumChannels")
batch = self.get_nodeattr("numInputVectors")
- ishape = (batch, IFMDim, IFMDim, num_ch)
+ is_2d = self.get_nodeattr("DimMode") == 0
+ if is_2d:
+ ishape = (batch, IFMDim, IFMDim, num_ch)
+ else:
+ ishape = (batch, IFMDim, 1, num_ch)
return ishape
def get_normal_output_shape(self):
OFMDim = self.get_nodeattr("OFMDim")
num_ch = self.get_nodeattr("NumChannels")
batch = self.get_nodeattr("numInputVectors")
- oshape = (batch, OFMDim, OFMDim, num_ch)
+ is_2d = self.get_nodeattr("DimMode") == 0
+ if is_2d:
+ oshape = (batch, OFMDim, OFMDim, num_ch)
+ else:
+ oshape = (batch, OFMDim, 1, num_ch)
return oshape
def get_folded_input_shape(self):
@@ -187,10 +201,19 @@ class UpsampleNearestNeighbour_Batch(HLSCustomOp):
)
def docompute(self):
- self.code_gen_dict["$DOCOMPUTE$"] = [
- """UpsampleNearestNeighbour_Batch<OFMDim, IFMDim, IFMChannels,
- ap_uint<Input_precision> > (in0, out, numReps);"""
- ]
+ is_2d = self.get_nodeattr("DimMode") == 0
+ batch = self.get_nodeattr("numInputVectors")
+ if is_2d:
+ self.code_gen_dict["$DOCOMPUTE$"] = [
+ """UpsampleNearestNeighbour_Batch<OFMDim, IFMDim, IFMChannels,
+ ap_uint<Input_precision> > (in0, out, numReps);"""
+ ]
+ else:
+ assert batch == 1, "1D upsampler currently needs numReps=1"
+ self.code_gen_dict["$DOCOMPUTE$"] = [
+ """UpsampleNearestNeighbour_1D<OFMDim, IFMDim, IFMChannels,
+ ap_uint<Input_precision> > (in0, out);"""
+ ]
def dataoutstrm(self):
code_gen_dir = self.get_nodeattr("code_gen_dir_cppsim")
@@ -246,7 +269,6 @@ class UpsampleNearestNeighbour_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":
@@ -268,9 +290,7 @@ class UpsampleNearestNeighbour_Batch(HLSCustomOp):
), """Input shape doesn't
match expected shape (numInputVectors, ImgDim, ImgDim, NumChannels)."""
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)
+ self.dynamic_input_to_npy(context, 1, target_dir=code_gen_dir)
if mode == "cppsim":
# execute the precompiled model
diff --git a/src/finn/qnn-data/templates/driver/driver_base.py b/src/finn/qnn-data/templates/driver/driver_base.py
index 497477da9d4cff736dc32eb27532e658890d5cc7..2096760580b4f33ba1ab09564ebba1601c4dc23c 100644
--- a/src/finn/qnn-data/templates/driver/driver_base.py
+++ b/src/finn/qnn-data/templates/driver/driver_base.py
@@ -439,13 +439,13 @@ class FINNExampleOverlay(Overlay):
total_in = 0
for i in range(self.num_inputs):
total_in += np.prod(self.ishape_packed(i))
- res["DRAM_in_bandwidth[Mb/s]"] = total_in * 0.000001 / runtime
+ res["DRAM_in_bandwidth[MB/s]"] = total_in * 0.000001 / runtime
total_out = 0
for o in range(self.num_outputs):
total_out += np.prod(self.oshape_packed(o))
- res["DRAM_out_bandwidth[Mb/s]"] = total_out * 0.000001 / runtime
+ res["DRAM_out_bandwidth[MB/s]"] = total_out * 0.000001 / runtime
for iwdma, iwbuf, iwdma_name in self.external_weights:
- res["DRAM_extw_%s_bandwidth[Mb/s]" % iwdma_name] = (
+ res["DRAM_extw_%s_bandwidth[MB/s]" % iwdma_name] = (
self.batch_size * np.prod(iwbuf.shape) * 0.000001 / runtime
)
if self.platform == "zynq-iodma":
diff --git a/src/finn/transformation/fpgadataflow/convert_to_hls_layers.py b/src/finn/transformation/fpgadataflow/convert_to_hls_layers.py
index f0bd5fbd0670e5088372383b16690ab67878334d..429bc34ffc59b5d98bb559f36ac557de4dbba92f 100644
--- a/src/finn/transformation/fpgadataflow/convert_to_hls_layers.py
+++ b/src/finn/transformation/fpgadataflow/convert_to_hls_layers.py
@@ -285,20 +285,25 @@ class InferUpsample(Transformation):
)
# Assumes nhwc layout for scales and input
- assert scales[1] == scales[2], (
- "%s: Upsampling is only supported for quadratic scales." % n.name
+ is_scale_square_2d = scales[1] == scales[2]
+ is_scale_1d = scales[1] > 1 and scales[2] == 1
+ assert is_scale_square_2d or is_scale_1d, (
+ "%s: Upsampling only supported for 1D H, or 2D square scaling"
+ % n.name
)
assert scales[0] == scales[3] == 1, (
n.name + ": Upsampling is only supported for scales with "
- "the first and last dimensions being 1."
+ "the first and last dimensions being 1 in NHWC."
)
spatial_scale = scales[1]
assert spatial_scale == int(spatial_scale), (
"%s: Upsampling is only supported for integer scales." % n.name
)
+ is_shape_square_2d = in_shape[1] == in_shape[2]
+ is_shape_1d = in_shape[1] > 1 and in_shape[2] == 1
- assert in_shape[1] == in_shape[2], (
- "%s: Upsampling is only supported for quadratic input shapes."
+ assert is_shape_square_2d or is_shape_1d, (
+ "%s: Upsampling is only supported for 1D H or 2D square inputs."
% n.name
)
@@ -308,6 +313,7 @@ class InferUpsample(Transformation):
NumChannels = in_shape[-1]
numInputVectors = in_shape[0]
inputDataType = dt.name
+ dim_mode = 0 if is_shape_square_2d else 1
# Insert the HLSCustomOp node
Upsample_HLS_node = helper.make_node(
@@ -321,6 +327,7 @@ class InferUpsample(Transformation):
NumChannels=NumChannels,
inputDataType=inputDataType,
numInputVectors=numInputVectors,
+ DimMode=dim_mode,
name="UpsampleNearestNeighbour_Batch_" + n.name,
)
diff --git a/src/finn/transformation/fpgadataflow/create_stitched_ip.py b/src/finn/transformation/fpgadataflow/create_stitched_ip.py
index 7c978cf61a465cacb4d562634d950311ed992021..892ab09fdf41947f86e2bf122e057e94585dfa8c 100644
--- a/src/finn/transformation/fpgadataflow/create_stitched_ip.py
+++ b/src/finn/transformation/fpgadataflow/create_stitched_ip.py
@@ -534,8 +534,9 @@ class CreateStitchedIP(Transformation):
tcl.append("ipx::save_core [ipx::find_open_core %s]" % block_vlnv)
# export list of used Verilog files (for rtlsim later on)
tcl.append(
- "set all_v_files [get_files -filter {FILE_TYPE == Verilog "
- + "&& USED_IN_SYNTHESIS == 1} ]"
+ "set all_v_files [get_files -filter {USED_IN_SYNTHESIS == 1 "
+ + "&& (FILE_TYPE == Verilog || FILE_TYPE == SystemVerilog "
+ + '|| FILE_TYPE =="Verilog Header")}]'
)
v_file_list = "%s/all_verilog_srcs.txt" % vivado_stitch_proj_dir
tcl.append("set fp [open %s w]" % v_file_list)
diff --git a/src/finn/transformation/fpgadataflow/make_pynq_driver.py b/src/finn/transformation/fpgadataflow/make_pynq_driver.py
index 863523605580ef77559b65a1abd72802daff187d..dce98e54a3d62d72b83ebed21aa0604f0f6fa8ce 100644
--- a/src/finn/transformation/fpgadataflow/make_pynq_driver.py
+++ b/src/finn/transformation/fpgadataflow/make_pynq_driver.py
@@ -118,12 +118,21 @@ class MakePYNQDriver(Transformation):
files_to_copy.append(
(qonnx_path + "/util/basic.py", qonnx_target_path + "/util/basic.py")
)
+ files_to_copy.append(
+ (qonnx_path + "/util/__init__.py", qonnx_target_path + "/util/__init__.py")
+ )
files_to_copy.append(
(
finn_util_path + "/data_packing.py",
finn_target_path + "/util/data_packing.py",
)
)
+ files_to_copy.append(
+ (
+ finn_util_path + "/__init__.py",
+ finn_target_path + "/util/__init__.py",
+ )
+ )
for (src_file, target_file) in files_to_copy:
shutil.copy(src_file, target_file)
# extract input-output shapes from the graph
diff --git a/src/finn/transformation/fpgadataflow/synth_ooc.py b/src/finn/transformation/fpgadataflow/synth_ooc.py
index 8d4aec259c440e311f6e3a6fb4d0359d55d738ca..6070cce636f50473545ab8a33c7867b7e1eb7f9c 100644
--- a/src/finn/transformation/fpgadataflow/synth_ooc.py
+++ b/src/finn/transformation/fpgadataflow/synth_ooc.py
@@ -52,7 +52,7 @@ class SynthOutOfContext(Transformation):
top_module_name = model.get_metadata_prop("wrapper_filename")
top_module_name = file_to_basename(top_module_name).strip(".v")
build_dir = make_build_dir("synth_out_of_context_")
- verilog_extensions = [".v", ".vh"]
+ verilog_extensions = [".v", ".sv", ".vh"]
with open(vivado_stitch_proj_dir + "/all_verilog_srcs.txt", "r") as f:
all_verilog_srcs = f.read().split()
for file in all_verilog_srcs:
diff --git a/src/finn/transformation/qonnx/fold_quant_weights.py b/src/finn/transformation/qonnx/fold_quant_weights.py
index 80b6042d03ea11a45493011288133ed3a6f57c8d..e8339ae24472fa238e5c5da176b1316611218a54 100644
--- a/src/finn/transformation/qonnx/fold_quant_weights.py
+++ b/src/finn/transformation/qonnx/fold_quant_weights.py
@@ -126,10 +126,20 @@ class FoldQuantWeights(Transformation):
model.set_tensor_datatype(node_out, new_dtype)
# Reshape scale for Conv if required
+ target_output_shape = model.get_tensor_shape(
+ target_node.output[0]
+ )
if target_node.op_type == "Conv" and len(scale.shape) > 0:
- bias_shape = [1] * len(scale.shape)
- bias_shape[1] = -1
- scale = scale.reshape(bias_shape)
+ conv_out_shape = [1] * len(target_output_shape)
+ # only support per-output channel scaling
+ # (i.e. all scale shape elems besides 0th must be 1s)
+ if len(scale.shape) > 1:
+ assert (
+ np.prod(scale.shape[1:]) == 1
+ ), "Can't fold scale beyond per-out-channel granularity"
+ # collect all scaling in channels dim (since we constrain)
+ conv_out_shape[1] = -1
+ scale = scale.reshape(conv_out_shape)
if scale.shape == (1,):
scale = scale[0]
diff --git a/src/finn/transformation/qonnx/quant_act_to_multithreshold.py b/src/finn/transformation/qonnx/quant_act_to_multithreshold.py
index c52d69b0f09d306c5b076bb6ef1775f38977241a..77025ecdf57d5a422992d4163d05c740454986bb 100644
--- a/src/finn/transformation/qonnx/quant_act_to_multithreshold.py
+++ b/src/finn/transformation/qonnx/quant_act_to_multithreshold.py
@@ -110,11 +110,6 @@ class ConvertQuantActToMultiThreshold(Transformation):
predecessor_op_type = predecessor[0].op_type
else:
predecessor_op_type = predecessor
- if model.is_fork_node(n):
- raise ValueError(
- "Forking Quant/BipolarQuant nodes are currently "
- "not supported by FINN."
- )
if n.op_type == "Quant" and not model.get_initializer(n.input[2]) == 0:
raise ValueError(
"Only Quant nodes with zero-point == 0 are currently supported."
diff --git a/src/finn/transformation/streamline/absorb.py b/src/finn/transformation/streamline/absorb.py
index 0299c4f4d89d1fdd94434db77c77a0e529c86d26..a983e67750a0a860eeeb4b429f7d6b181fc84fe3 100644
--- a/src/finn/transformation/streamline/absorb.py
+++ b/src/finn/transformation/streamline/absorb.py
@@ -473,7 +473,7 @@ class AbsorbConsecutiveTransposes(Transformation):
"""Remove (Transpose -> Transpose) patterns when the input and output
of the pattern have the same layout."""
- def Are_opposite_permutations(self, perms1, perms2):
+ def are_opposite_permutations(self, perms1, perms2):
if len(perms1) != len(perms2):
return False
assert 0 <= max(perms2) < len(perms2), "invalid permutation"
@@ -488,72 +488,40 @@ class AbsorbConsecutiveTransposes(Transformation):
def apply(self, model):
graph = model.graph
graph_modified = False
- for n in graph.node:
- if n.op_type == "Transpose":
- if model.is_fork_node(n):
- next_nodes = model.find_direct_successors(n)
- perms1 = list(get_by_name(n.attribute, "perm").ints)
-
- # check if all nodes after fork are opposite transposes
- all_opposite_transposes = True
- for next_node in next_nodes:
- if next_node is not None and next_node.op_type == "Transpose":
- perms2 = list(get_by_name(next_node.attribute, "perm").ints)
- if not self.Are_opposite_permutations(perms1, perms2):
- all_opposite_transposes = False
- break
- else:
- all_opposite_transposes = False
- break
-
- if not all_opposite_transposes:
- continue
-
- prod = model.find_producer(n.input[0])
- for next_node in next_nodes:
- # connect next_node's consumer input to n's producer output
- # TODO implement this to allow for forks as producers and
- # joins as consumers
- cons = model.find_consumer(next_node.output[0])
- cons.input[0] = prod.output[0]
-
- # remove consumer transpose
- graph.node.remove(next_node)
-
- # remove producer transpose
- graph.node.remove(n)
- graph_modified = True
-
- else:
- next_node = model.find_consumer(n.output[0])
+ for node in graph.node:
+ if node.op_type == "Transpose":
+ next_nodes = model.find_consumers(node.output[0])
+ perms1 = list(get_by_name(node.attribute, "perm").ints)
+ # check if all nodes after fork are opposite transposes
+ all_opposite_transposes = True
+ for next_node in next_nodes:
if next_node is not None and next_node.op_type == "Transpose":
- perms1 = list(get_by_name(n.attribute, "perm").ints)
perms2 = list(get_by_name(next_node.attribute, "perm").ints)
- if self.Are_opposite_permutations(perms1, perms2):
-
- # connect next_node's consumer input to n's producer output
- # TODO implement this to allow for forks as producers
- consumers = model.find_direct_successors(next_node)
- prod = model.find_producer(n.input[0])
- if prod is not None:
- for cons in consumers:
- for cons_in in cons.input:
- if cons_in == next_node.output[0]:
- prod.output[0] = cons_in
- break
- else:
- # n.input[0] is top-level graph input
- # wire consumers directly to that
- for cons in consumers:
- for i, iname in enumerate(cons.input):
- if iname == next_node.output[0]:
- cons.input[i] = n.input[0]
-
- # remove both transposes
- graph.node.remove(n)
- graph.node.remove(next_node)
+ if not self.are_opposite_permutations(perms1, perms2):
+ all_opposite_transposes = False
+ break
+ else:
+ all_opposite_transposes = False
+ break
+ if not all_opposite_transposes:
+ continue
+ source_tensor = node.input[0]
+ for next_node in next_nodes:
+ # connect next_node's consumers' appropriate input to n's input
+ # TODO how to handle top-level outputs if any?
+ nextnode_out = next_node.output[0]
+ assert nextnode_out not in [x.name for x in model.graph.output]
+ consumers = model.find_consumers(nextnode_out)
+ for cons in consumers:
+ for i, iname in enumerate(cons.input):
+ if iname == nextnode_out:
+ cons.input[i] = source_tensor
+ # remove consumer transpose
+ graph.node.remove(next_node)
+ # remove producer transpose
+ graph.node.remove(node)
+ graph_modified = True
- graph_modified = True
if graph_modified:
model = model.transform(InferDataTypes())
return (model, graph_modified)
diff --git a/src/finn/transformation/streamline/reorder.py b/src/finn/transformation/streamline/reorder.py
index 9ff8a2173ce81e2a19c56bbd20a326759c3b9df2..3e815c1537353cc2be970a2068d4ded30cc48bc8 100644
--- a/src/finn/transformation/streamline/reorder.py
+++ b/src/finn/transformation/streamline/reorder.py
@@ -553,6 +553,8 @@ class MoveLinearPastEltwiseAdd(Transformation):
# Other transform should handle that
if prod0 is None or prod1 is None or (prod0 == prod1):
continue
+ if len(prod0.input) < 2 or len(prod1.input) < 2:
+ continue
init0 = model.get_initializer(prod0.input[1])
init1 = model.get_initializer(prod1.input[1])
# if either initializer is None, skip
@@ -728,9 +730,10 @@ class MoveOpPastFork(Transformation):
can be merged with nodes in the branches
"""
- def __init__(self, op_name_list):
+ def __init__(self, op_name_list, get_attrs_fxn=lambda x: {}):
super().__init__()
self.ops_to_move = op_name_list
+ self.get_attrs_fxn = get_attrs_fxn
def apply(self, model):
graph = model.graph
@@ -747,9 +750,10 @@ class MoveOpPastFork(Transformation):
# Restrict this transform to operations with constant parameters
# Assuming parameters is in input 1
- op_init_param = model.get_initializer(n.input[1])
- if op_init_param is None:
- continue
+ if len(n.input) > 1:
+ op_init_param = model.get_initializer(n.input[1])
+ else:
+ op_init_param = None
# Check case when branches are empty and go
# to the same node
@@ -766,16 +770,20 @@ class MoveOpPastFork(Transformation):
for consumer_node in consumers[1:]:
# create new node
- new_param_name = model.make_new_valueinfo_name()
new_output_tensor_name = model.make_new_valueinfo_name()
+ if op_init_param is None:
+ new_inp_list = [n.input[0]]
+ else:
+ new_param_name = model.make_new_valueinfo_name()
+ new_inp_list = [n.input[0], new_param_name]
+ model.set_initializer(new_param_name, op_init_param)
+ attrs = self.get_attrs_fxn(n)
+ # TODO use copy of original node instead to get attrs?
new_node = oh.make_node(
- n.op_type,
- [n.input[0], new_param_name],
- [new_output_tensor_name],
+ n.op_type, new_inp_list, [new_output_tensor_name], **attrs
)
graph.node.insert(node_ind, new_node)
node_ind += 1
- model.set_initializer(new_param_name, op_init_param)
# change consumer input tensor
graph.node.remove(consumer_node)
@@ -811,6 +819,13 @@ class MoveLinearPastFork(MoveOpPastFork):
super().__init__(["Add", "Mul"])
+class MoveTransposePastFork(MoveOpPastFork):
+ def __init__(self):
+ super().__init__(
+ ["Transpose"], lambda x: {"perm": get_by_name(x.attribute, "perm").ints}
+ )
+
+
class MoveMaxPoolPastMultiThreshold(Transformation):
"""Move MaxPool nodes past MultiThreshold nodes on linear segments of the graph."""
diff --git a/src/finn/util/pyverilator.py b/src/finn/util/pyverilator.py
index 3396561e06f553785e842ec0b6626bc405d262c5..f6a51da8e44ea60ae5693cdd033b39bdf51376ac 100644
--- a/src/finn/util/pyverilator.py
+++ b/src/finn/util/pyverilator.py
@@ -74,7 +74,9 @@ def pyverilate_stitched_ip(
# are identical but in multiple directories (regslice_core.v)
# remove duplicates from list by doing list -> set -> list
- all_verilog_files = list(set(filter(lambda x: x.endswith(".v"), all_verilog_srcs)))
+ all_verilog_files = list(
+ set(filter(lambda x: x.endswith(".v") or x.endswith(".sv"), all_verilog_srcs))
+ )
# remove all but one instances of regslice_core.v
filtered_verilog_files = []
diff --git a/tests/brevitas/test_brevitas_relu_act_export.py b/tests/brevitas/test_brevitas_relu_act_export.py
index b0c3d6088c27291f1f49dd2f1ee746b65ca0a737..3dc46ec31e49d7115b19b3373d54be6ddc29bb80 100644
--- a/tests/brevitas/test_brevitas_relu_act_export.py
+++ b/tests/brevitas/test_brevitas_relu_act_export.py
@@ -41,6 +41,7 @@ from brevitas.nn import QuantReLU
from qonnx.core.modelwrapper import ModelWrapper
from qonnx.transformation.infer_shapes import InferShapes
from qonnx.util.cleanup import cleanup as qonnx_cleanup
+from torch import nn
import finn.core.onnx_exec as oxe
from finn.transformation.qonnx.convert_qonnx_to_finn import ConvertQONNXtoFINN
@@ -179,3 +180,83 @@ scaling_impl.learned_value": rand_tensor.type(
assert np.isclose(produced, expected, atol=1e-3).all()
os.remove(export_onnx_path)
+
+
+class PyTorchTestModel(nn.Module):
+ def __init__(self, abits):
+ super(PyTorchTestModel, self).__init__()
+ out_channels = 32
+ self.b_act = QuantReLU(
+ bit_width=abits,
+ quant_type=QuantType.INT,
+ scaling_impl_type=ScalingImplType.PARAMETER,
+ scaling_per_channel=True,
+ restrict_scaling_type=RestrictValueType.LOG_FP,
+ scaling_min_val=2e-16,
+ max_val=6.0,
+ return_quant_tensor=False,
+ per_channel_broadcastable_shape=(1, out_channels, 1, 1),
+ )
+
+ def forward(self, x):
+ act_out = self.b_act(x)
+ y0 = act_out * 2.0
+ y1 = act_out * -1.0
+ y = y0 + y1
+ return y
+
+
+@pytest.mark.brevitas_export
+@pytest.mark.parametrize("abits", [2, 4, 8])
+@pytest.mark.parametrize("max_val", [1.0, 1.5, 1 - 2 ** (-7)])
+@pytest.mark.parametrize("scaling_per_channel", [True])
+@pytest.mark.parametrize("QONNX_export", [True])
+def test_brevitas_act_export_relu_forking(
+ abits, max_val, scaling_per_channel, QONNX_export
+):
+ out_channels = 32
+ ishape = (1, out_channels, 1, 1)
+ min_val = -1.0
+ model_pyt = PyTorchTestModel(abits)
+
+ rand_tensor = (2) * torch.rand((1, out_channels, 1, 1))
+
+ checkpoint = {
+ "b_act.act_quant_proxy.fused_activation_quant_proxy."
+ "tensor_quant.scaling_impl.learned_value": rand_tensor.type(torch.FloatTensor)
+ }
+ model_pyt.load_state_dict(checkpoint)
+
+ if QONNX_export:
+ m_path = export_onnx_path
+ BrevitasONNXManager.export(model_pyt, ishape, m_path)
+ qonnx_cleanup(m_path, out_file=m_path)
+ model = ModelWrapper(m_path)
+ model = model.transform(ConvertQONNXtoFINN())
+ model.save(m_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()
+ model_pyt.eval()
+ expected = model_pyt.forward(inp_tensor).detach().numpy()
+ if not np.isclose(produced, expected, atol=1e-3).all():
+ print(abits, max_val)
+ print("scale: ", model_pyt.quant_act_scale().type(torch.FloatTensor).detach())
+ if abits < 5:
+ print(
+ "thres:",
+ ", ".join(["{:8.4f}".format(x) for x in model_pyt.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/end2end/test_end2end_bnn_pynq.py b/tests/end2end/test_end2end_bnn_pynq.py
index 4139c74de677e7e7108211b35ec7fdf01bd138d9..103f18b514c23c4e1ad35a85d020dc0481aa9c47 100644
--- a/tests/end2end/test_end2end_bnn_pynq.py
+++ b/tests/end2end/test_end2end_bnn_pynq.py
@@ -788,7 +788,7 @@ class TestEnd2End:
ret_str += "\n" + "Raw data:"
ret_str += "\n" + "{:<8} {:<16} {:<16} {:<16} {:<16} {:<16}".format(
- "N", "runtime[ms]", "fclk[mhz]", "fps", "DRAM rd[Mb/s]", "DRAM wr[Mb/s]"
+ "N", "runtime[ms]", "fclk[mhz]", "fps", "DRAM rd[MB/s]", "DRAM wr[MB/s]"
)
for k in bsize_range:
v = ret[k]
@@ -797,8 +797,8 @@ class TestEnd2End:
np.round(v["runtime[ms]"], 4),
v["fclk[mhz]"],
np.round(v["throughput[images/s]"], 2),
- np.round(v["DRAM_in_bandwidth[Mb/s]"], 2),
- np.round(v["DRAM_out_bandwidth[Mb/s]"], 2),
+ np.round(v["DRAM_in_bandwidth[MB/s]"], 2),
+ np.round(v["DRAM_out_bandwidth[MB/s]"], 2),
)
ret_str += "\n" + "-----------------------------"
warnings.warn(ret_str)
diff --git a/tests/fpgadataflow/test_fpgadataflow_checksum.py b/tests/fpgadataflow/test_fpgadataflow_checksum.py
index 5e79ea2dad2aa4200f998fd8953672b9f49b2b86..495fcd10b6a977c6b0917ac37b58ec5595185c25 100644
--- a/tests/fpgadataflow/test_fpgadataflow_checksum.py
+++ b/tests/fpgadataflow/test_fpgadataflow_checksum.py
@@ -133,6 +133,7 @@ def create_two_fc_model():
return model
+@pytest.mark.vivado
@pytest.mark.fpgadataflow
def test_fpgadataflow_checksum():
# use a graph consisting of two fc layers to test
diff --git a/tests/fpgadataflow/test_fpgadataflow_concat.py b/tests/fpgadataflow/test_fpgadataflow_concat.py
index dddc470ec2ed88faf078f19bd0d2a7a4a6b5b6cd..8488a34dff52d39c28fbea25275c9a4b59c37f80 100644
--- a/tests/fpgadataflow/test_fpgadataflow_concat.py
+++ b/tests/fpgadataflow/test_fpgadataflow_concat.py
@@ -144,6 +144,5 @@ def test_fpgadataflow_concat_stitchedip():
)
model.set_metadata_prop("exec_mode", "rtlsim")
model.set_metadata_prop("rtlsim_trace", "trace.vcd")
- model.save("dbg.onnx")
ret_sim = execute_onnx(model, inp_dict)
assert (exp_out == ret_sim[oname]).all()
diff --git a/tests/fpgadataflow/test_layer_streaming_maxpool_batch.py b/tests/fpgadataflow/test_fpgadataflow_streamingmaxpool.py
similarity index 84%
rename from tests/fpgadataflow/test_layer_streaming_maxpool_batch.py
rename to tests/fpgadataflow/test_fpgadataflow_streamingmaxpool.py
index 55c90644dfbb23fbc2da10cf969461abe6d38bf3..a3968cf79704092ffb5ec53c887842372b625f4d 100644
--- a/tests/fpgadataflow/test_layer_streaming_maxpool_batch.py
+++ b/tests/fpgadataflow/test_fpgadataflow_streamingmaxpool.py
@@ -32,6 +32,7 @@ from onnx import TensorProto, helper
from qonnx.core.datatype import DataType
from qonnx.core.modelwrapper import ModelWrapper
from qonnx.custom_op.general.maxpoolnhwc import compute_pool_output_dim
+from qonnx.custom_op.registry import getCustomOp
from qonnx.transformation.general import GiveUniqueNodeNames
from qonnx.transformation.infer_shapes import InferShapes
from qonnx.util.basic import gen_finn_dt_tensor
@@ -82,46 +83,6 @@ def make_single_maxpoolnhwc_modelwrapper(k, ifm_ch, ifm_dim, ofm_dim, idt, ceil_
return model
-def make_single_streamingmaxpool_modelwrapper(
- k, ifm_ch, pe, ifm_dim, ofm_dim, idt, ceil_mode
-):
- k_h, k_w = k
- ifm_dim_h, ifm_dim_w = ifm_dim
- ofm_dim_h, ofm_dim_w = ofm_dim
- odt = idt
- inp = helper.make_tensor_value_info(
- "inp", TensorProto.FLOAT, [1, ifm_dim_h, ifm_dim_w, ifm_ch]
- )
- outp = helper.make_tensor_value_info(
- "outp", TensorProto.FLOAT, [1, ofm_dim_h, ofm_dim_w, ifm_ch]
- )
-
- smp_node = helper.make_node(
- "StreamingMaxPool_Batch",
- ["inp"],
- ["outp"],
- domain="finn.custom_op.fpgadataflow",
- backend="fpgadataflow",
- PoolDim=[k_h, k_w],
- NumChannels=ifm_ch,
- PE=pe,
- ImgDim=[ifm_dim_h, ifm_dim_w],
- CeilMode=ceil_mode,
- dataType=idt.name,
- )
- graph = helper.make_graph(
- nodes=[smp_node], name="smp_graph", inputs=[inp], outputs=[outp]
- )
-
- model = helper.make_model(graph, producer_name="smp-model")
- model = ModelWrapper(model)
-
- model.set_tensor_datatype("inp", idt)
- model.set_tensor_datatype("outp", odt)
-
- return model
-
-
def prepare_inputs(input_tensor):
return {"inp": input_tensor}
@@ -187,6 +148,10 @@ def test_fpgadataflow_streamingmaxpool(
assert model.graph.node[0].op_type == "StreamingMaxPool_Batch"
+ # Ensure PE value is set
+ streamingmaxpool_node = model.get_nodes_by_op_type("StreamingMaxPool_Batch")[0]
+ getCustomOp(streamingmaxpool_node).set_nodeattr("PE", pe)
+
if exec_mode == "cppsim":
model = model.transform(SetExecMode("cppsim"))
model = model.transform(PrepareCppSim())
@@ -198,7 +163,7 @@ def test_fpgadataflow_streamingmaxpool(
model = model.transform(HLSSynthIP())
model = model.transform(PrepareRTLSim())
else:
- raise Exception("Unknown exec_mode in test_layer_streaming_maxpool_batch")
+ raise Exception("Unknown exec_mode in test_fpgadataflow_streamingmaxpool")
# execute model
y_produced = oxe.execute_onnx(model, input_dict)["outp"]
@@ -211,6 +176,7 @@ def test_fpgadataflow_streamingmaxpool(
exp_cycles_dict = model.analysis(exp_cycles_per_layer)
exp_cycles = exp_cycles_dict[node.name]
# FIXME: maxpool cycles prediction needs a fix
- # mostl likely due to some loops not flattening
+ # most likely due to inaccurate cycle prediction of
+ # nested for-loops
# assert np.isclose(exp_cycles, cycles_rtlsim, atol=15)
assert exp_cycles != 0
diff --git a/tests/fpgadataflow/test_fpgadataflow_upsampler.py b/tests/fpgadataflow/test_fpgadataflow_upsampler.py
index d1ef0b890a66524b7cbd055a413561961ebcb4a7..a08d31f7b05184a4d5c84ef927a05fe1fd6e43c3 100644
--- a/tests/fpgadataflow/test_fpgadataflow_upsampler.py
+++ b/tests/fpgadataflow/test_fpgadataflow_upsampler.py
@@ -30,6 +30,7 @@ import pytest
import numpy as np
import os
+import shutil
import torch
from brevitas.export import FINNManager
from qonnx.core.datatype import DataType
@@ -51,6 +52,7 @@ from finn.transformation.fpgadataflow.prepare_cppsim import PrepareCppSim
from finn.transformation.fpgadataflow.prepare_ip import PrepareIP
from finn.transformation.fpgadataflow.prepare_rtlsim import PrepareRTLSim
from finn.transformation.fpgadataflow.set_exec_mode import SetExecMode
+from finn.util.basic import make_build_dir
tmpdir = os.environ["FINN_BUILD_DIR"]
@@ -117,7 +119,7 @@ class PyTorchTestModel(nn.Module):
# param datatype
@pytest.mark.parametrize("dt", [DataType["INT8"]])
-# Width/height of square input feature map
+# spatial dim input feature map
@pytest.mark.parametrize("IFMDim", [3, 5])
# upscaling factor
@pytest.mark.parametrize("scale", [2, 3])
@@ -125,14 +127,22 @@ class PyTorchTestModel(nn.Module):
@pytest.mark.parametrize("NumChannels", [4])
# execution mode
@pytest.mark.parametrize("exec_mode", ["cppsim", "rtlsim"])
+# whether to use 1D or 2D square testcases
+@pytest.mark.parametrize("is_1d", [False, True])
@pytest.mark.fpgadataflow
@pytest.mark.vivado
@pytest.mark.slow
-def test_fpgadataflow_upsampler(dt, IFMDim, scale, NumChannels, exec_mode):
+def test_fpgadataflow_upsampler(dt, IFMDim, scale, NumChannels, exec_mode, is_1d):
+ tmpdir = make_build_dir("upsample_export_")
atol = 1e-3
+ if is_1d:
+ input_shape = (1, NumChannels, IFMDim, 1)
+ upscale_factor = (scale, 1)
+ else:
+ input_shape = (1, NumChannels, IFMDim, IFMDim)
+ upscale_factor = (scale, scale)
# Create the test model and inputs for it
- torch_model = PyTorchTestModel(upscale_factor=scale)
- input_shape = (1, NumChannels, IFMDim, IFMDim)
+ torch_model = PyTorchTestModel(upscale_factor=upscale_factor)
test_in = torch.arange(0, np.prod(np.asarray(input_shape)))
# Limit the input to values valid for the given datatype
test_in %= dt.max() - dt.min() + 1
@@ -200,3 +210,4 @@ def test_fpgadataflow_upsampler(dt, IFMDim, scale, NumChannels, exec_mode):
assert output_matches, "Cppsim output doesn't match ONNX/PyTorch."
elif exec_mode == "rtlsim":
assert output_matches, "Rtlsim output doesn't match ONNX/PyTorch."
+ shutil.rmtree(tmpdir, ignore_errors=True)
diff --git a/tests/transformation/streamline/test_absorb_opposite_transposes.py b/tests/transformation/streamline/test_absorb_opposite_transposes.py
index 51ea5edfc420bf935de3e196df1b150934782a91..6d8d2b9f0cd4ad28c3ea0922d69b9b963a0deb08 100644
--- a/tests/transformation/streamline/test_absorb_opposite_transposes.py
+++ b/tests/transformation/streamline/test_absorb_opposite_transposes.py
@@ -29,8 +29,7 @@
import pytest
import numpy as np
-import onnx.helper as oh
-from onnx import TensorProto
+import onnx.parser as oprs
from qonnx.core.modelwrapper import ModelWrapper
from qonnx.transformation.infer_shapes import InferShapes
@@ -41,39 +40,42 @@ from finn.transformation.streamline.absorb import AbsorbConsecutiveTransposes
@pytest.mark.streamline
def test_absorb_opposite_transposes():
np.random.seed(0)
- input_shape = [1, 3, 4, 2]
- top_in = oh.make_tensor_value_info("top_in", TensorProto.FLOAT, input_shape)
- top_out = oh.make_tensor_value_info("top_out", TensorProto.FLOAT, input_shape)
- value_info = [oh.make_tensor_value_info("add_param_0", TensorProto.FLOAT, [1])]
- value_info += [oh.make_tensor_value_info("add_param_1", TensorProto.FLOAT, [1])]
- value_info += [oh.make_tensor_value_info("mul_param_0", TensorProto.FLOAT, [1])]
- modelproto = oh.make_model(
- oh.make_graph(
- name="test",
- inputs=[top_in],
- outputs=[top_out],
- value_info=value_info,
- nodes=[
- oh.make_node("Add", ["top_in", "add_param_0"], ["t0"]),
- oh.make_node("Transpose", ["t0"], ["t1"], perm=[0, 2, 3, 1]),
- oh.make_node("Transpose", ["t1"], ["t2"], perm=[0, 3, 1, 2]),
- oh.make_node("Add", ["t2", "add_param_1"], ["t3"]),
- oh.make_node("Transpose", ["t3"], ["t4"], perm=[0, 2, 3, 1]),
- oh.make_node("Transpose", ["t4"], ["t5"], perm=[0, 3, 1, 2]),
- oh.make_node("Add", ["t5", "t2"], ["t6"]),
- oh.make_node("Mul", ["t6", "mul_param_0"], ["top_out"]),
- ],
- )
- )
- model = ModelWrapper(modelproto)
+ shp = [1, 3, 4, 2]
+ shp_str = str(shp)
+ input = f"""
+ <
+ ir_version: 7,
+ opset_import: ["" : 9]
+ >
+ agraph (float{shp_str} in0) => (float{shp_str} out0)
+ <
+ float[1] add0_param = {{1.0}},
+ float[1] add1_param = {{3.0}},
+ float[1] mul0_param = {{2.0}}
+ >
+ {{
+ add0_out = Add(in0, add0_param)
+ t0_out = Transpose<perm=[0,2,3,1]>(add0_out)
+ t1_out = Transpose<perm=[0,3,1,2]>(t0_out)
+ add1_out = Add(t1_out, add1_param)
+ t2_out = Transpose<perm=[0,2,3,1]>(add1_out)
+ t3_out = Transpose<perm=[0,3,1,2]>(t2_out)
+ add2_out = Add(t1_out, t3_out)
+ t4_out = Transpose<perm=[0,2,3,1]>(add2_out)
+ t5_out = Transpose<perm=[0,3,1,2]>(t4_out)
+ t6_out = Transpose<perm=[0,3,1,2]>(t4_out)
+ m0_out = Mul(t5_out, mul0_param)
+ m1_out = Mul(t6_out, mul0_param)
+ out0 = Mul(m0_out, m1_out)
+ }}
+ """
+ model = oprs.parse_model(input)
+ model = ModelWrapper(model)
model = model.transform(InferShapes())
- model.set_initializer("add_param_0", np.asarray([1], dtype=np.float32))
- model.set_initializer("add_param_1", np.asarray([3], dtype=np.float32))
- model.set_initializer("mul_param_0", np.asarray([2], dtype=np.float32))
new_model = model.transform(AbsorbConsecutiveTransposes())
new_model = new_model.transform(InferShapes())
- inp_dict = {"top_in": np.random.rand(*input_shape).astype(np.float32)}
+ inp_dict = {"top_in": np.random.rand(*shp).astype(np.float32)}
assert ox.compare_execution(model, model, inp_dict)
- assert len(new_model.graph.node) == 4
+ assert len(new_model.graph.node) == 6
for n in new_model.graph.node:
assert new_model.graph.node[0].op_type != "Transpose"
diff --git a/tests/transformation/streamline/test_move_past_fork.py b/tests/transformation/streamline/test_move_past_fork.py
index 5064fa3fca869a245c87cf0c1680d1357e5de60b..7e77d7f9b3502429f08c40558e330b6261d0dbad 100644
--- a/tests/transformation/streamline/test_move_past_fork.py
+++ b/tests/transformation/streamline/test_move_past_fork.py
@@ -28,80 +28,113 @@
import pytest
import numpy as np
-from onnx import TensorProto, helper
+import onnx.parser as oprs
from qonnx.core.modelwrapper import ModelWrapper
+from qonnx.transformation.general import GiveUniqueNodeNames
from qonnx.transformation.infer_shapes import InferShapes
+from qonnx.util.basic import get_by_name
import finn.core.onnx_exec as oxe
-from finn.transformation.streamline.reorder import MoveLinearPastFork
+from finn.transformation.streamline.reorder import (
+ MoveLinearPastFork,
+ MoveTransposePastFork,
+)
+
+
+@pytest.mark.streamline
+def test_move_past_fork_transpose():
+ shp = [1, 3, 32, 32]
+ shp_str = str(shp)
+ input = f"""
+ <
+ ir_version: 7,
+ opset_import: ["" : 9]
+ >
+ agraph (float{shp_str} in0) => (float{shp_str} out0)
+ {{
+ t0_out = Transpose<perm=[0,2,3,1]>(in0)
+ t1_out = Transpose<perm=[0,3,1,2]>(t0_out)
+ t2_out = Transpose<perm=[0,3,1,2]>(t0_out)
+ out0 = Add(t1_out, t2_out)
+ }}
+ """
+ model = oprs.parse_model(input)
+ model = ModelWrapper(model)
+ model = model.transform(InferShapes())
+ new_model = model.transform(MoveTransposePastFork())
+ new_model = new_model.transform(GiveUniqueNodeNames())
+ nodes = new_model.graph.node
+ assert oxe.compare_execution(
+ model, new_model, {"in0": np.random.rand(*shp).astype(np.float32)}
+ )
+ assert len(nodes) == 5
+ assert not new_model.is_fork_node(get_by_name(nodes, "Transpose_0"))
@pytest.mark.streamline
@pytest.mark.parametrize("ch", [64, 1])
# ifmdim
@pytest.mark.parametrize("ifmdim", [-1, 7])
-def test_move_past_fork(ch, ifmdim):
- # generate test vectors of correct shape
+def test_move_past_fork_linear(ch, ifmdim):
if ifmdim == -1:
- input_shape = (1, ch)
+ shp = [1, ch]
else:
- input_shape = (1, ch, ifmdim, ifmdim)
+ shp = [1, ch, ifmdim, ifmdim]
+ shp_str = str(shp)
+ input = f"""
+ <
+ ir_version: 7,
+ opset_import: ["" : 9]
+ >
+ agraph (float{shp_str} in0) => (float{shp_str} out0)
+ <
+ float{shp_str} add0_param,
+ float{shp_str} mul_shared_param,
+ float{shp_str} add2_param,
+ float{shp_str} mul2_param,
+ float{shp_str} add3_param,
+ float{shp_str} add4_param,
+ float{shp_str} mul3_param,
+ float{shp_str} add6_param
+ >
+ {{
- top_in = helper.make_tensor_value_info("top_in", TensorProto.FLOAT, input_shape)
- top_out = helper.make_tensor_value_info("top_out", TensorProto.FLOAT, input_shape)
-
- num_of_params = 8
- value_info = []
- for i in range(num_of_params):
- value_info += [
- helper.make_tensor_value_info("p" + str(i), TensorProto.FLOAT, input_shape)
- ]
-
- add_1_to_move = helper.make_node("Add", ["top_in", "p0"], ["fork1"])
- mul_1_to_move = helper.make_node("Mul", ["t5", "p4"], ["fork2"])
- add_2_to_move = helper.make_node("Add", ["fork2", "p5"], ["t6"])
- mul_1_not_to_move = helper.make_node("Mul", ["t8", "p7"], ["fork3"])
- modelproto = helper.make_model(
- helper.make_graph(
- name="test",
- inputs=[top_in],
- outputs=[top_out],
- value_info=value_info,
- nodes=[
- # fork1
- add_1_to_move,
- helper.make_node("Mul", ["fork1", "p1"], ["t2"]),
- helper.make_node("Mul", ["fork1", "p2"], ["t3"]),
- helper.make_node("Add", ["t2", "t3"], ["t4"]),
- helper.make_node("Add", ["t4", "p3"], ["t5"]),
- # fork2
- mul_1_to_move,
- add_2_to_move,
- helper.make_node("Add", ["fork2", "p6"], ["t7"]),
- helper.make_node("Add", ["t6", "t7"], ["t8"]),
- # empty branches: do nothing
- mul_1_not_to_move,
- helper.make_node("Add", ["fork3", "fork3"], ["top_out"]),
- ],
- )
- )
- model = ModelWrapper(modelproto)
+ add0_out = Add(in0, add0_param)
+ mul0_out = Mul(add0_out, mul_shared_param)
+ mul1_out = Mul(add0_out, mul_shared_param)
+ add1_out = Add(mul0_out, mul1_out)
+ add2_out = Add(add1_out, add2_param)
+ mul2_out = Mul(add2_out, mul2_param)
+ add3_out = Add(mul2_out, add3_param)
+ add4_out = Add(mul2_out, add4_param)
+ add5_out = Add(add3_out, add4_out)
+ mul3_out = Mul(add5_out, mul3_param)
+ out0 = Add(mul3_out, add6_param)
+ }}
+ """
+ model = oprs.parse_model(input)
+ model = ModelWrapper(model)
model = model.transform(InferShapes())
np.random.seed(0)
- for i in range(num_of_params):
- model.set_initializer(
- "p" + str(i), np.random.rand(*input_shape).astype(np.float32)
- )
-
+ for tensor_name in model.get_all_tensor_names():
+ if tensor_name.endswith("_param"):
+ pshape = model.get_tensor_shape(tensor_name)
+ model.set_initializer(
+ tensor_name, np.random.rand(*pshape).astype(np.float32)
+ )
+ model = model.transform(GiveUniqueNodeNames())
# Transform
new_model = model.transform(MoveLinearPastFork())
- inp_dict = {"top_in": np.random.rand(*input_shape).astype(np.float32)}
-
+ new_model = new_model.transform(GiveUniqueNodeNames())
+ inp_dict = {"top_in": np.random.rand(*shp).astype(np.float32)}
# Test
assert oxe.compare_execution(model, new_model, inp_dict)
- assert not new_model.is_fork_node(add_1_to_move)
- assert not new_model.is_fork_node(mul_1_to_move)
- assert not new_model.is_fork_node(add_2_to_move)
- assert new_model.is_fork_node(mul_1_not_to_move)
+ nodes = new_model.graph.node
+ assert len(new_model.get_nodes_by_op_type("Add")) == 9
+ assert len(new_model.get_nodes_by_op_type("Mul")) == 5
+ assert not new_model.is_fork_node(get_by_name(nodes, "Add_0"))
+ assert new_model.is_join_node(get_by_name(nodes, "Add_2"))
+ assert not new_model.is_fork_node(get_by_name(nodes, "Mul_2"))
+ assert not new_model.is_join_node(get_by_name(nodes, "Add_5"))
assert len(new_model.graph.node) == 14
diff --git a/tests/transformation/test_qonnx_to_finn.py b/tests/transformation/test_qonnx_to_finn.py
index 43055f6704732866569ac4770202f1b4ff6bfb22..7e438b4b8ba9d9befca79100bb9727735afa27d3 100644
--- a/tests/transformation/test_qonnx_to_finn.py
+++ b/tests/transformation/test_qonnx_to_finn.py
@@ -94,6 +94,9 @@ def analysis_testing_for_no_quant_nodes(model):
@pytest.mark.parametrize("wbits", [1, 2])
@pytest.mark.parametrize("model_name", ["TFC", "SFC", "LFC", "CNV", "mobilenet"])
def test_QONNX_to_FINN(model_name, wbits, abits):
+ if model_name == "mobilenet":
+ pytest.xfail("MobileNet test is temporarily excluded from QONNX testing.")
+
if wbits > abits:
pytest.skip("No wbits > abits cases at the moment")
if model_name == "LFC" and wbits == 2 and abits == 2: