Merge pull request #11 from Xilinx/dev

finn-examples v0.0.2b

.github/workflows/python-publish.yml

@@ -0,0 +1,31 @@
+# This workflow will upload a Python Package using Twine when a release is created
+# For more information see: https://help.github.com/en/actions/language-and-framework-guides/using-python-with-github-actions#publishing-to-package-registries
+
+name: Upload Python Package
+
+on:
+  release:
+    types: [created]
+
+jobs:
+  deploy:
+
+    runs-on: ubuntu-latest
+
+    steps:
+    - uses: actions/checkout@v2
+    - name: Set up Python
+      uses: actions/setup-python@v2
+      with:
+        python-version: '3.x'
+    - name: Install dependencies
+      run: |
+        python -m pip install --upgrade pip
+        pip install setuptools wheel twine
+    - name: Build and publish
+      env:
+        TWINE_USERNAME: __token__
+        TWINE_PASSWORD: ${{ secrets.PYPI_API_TOKEN }}
+      run: |
+        python setup.py sdist
+        twine upload dist/*

AUTHORS.rst

@@ -0,0 +1,10 @@
+============
+Contributors
+============
+
+* Yaman Umuroglu (@maltanar) (maintainer)
+* Jakoba Petri-Koenig (@auphelia)
+* Lucian Petrica (@quetric)
+* Tobias Alonso (@Tobi-Alonso)
+* Hendrik Borras (@HenniOVP)
+* Felix Paul Jentzsch (@felixpj)

README.md

@@ -14,10 +14,13 @@ pre-built bitfiles, PYNQ Python drivers and Jupyter notebooks to get started,
 and you can rebuild them from source.
 Both PYNQ on Zynq and Alveo are supported.
 
+Need help with a problem in this repo, or got a question? Feel free to ask for help in the [FINN Gitter channel](https://gitter.im/xilinx-finn/community).
+
 ## Quickstart
 
 
 *For Alveo we recommend setting up everything inside a virtualenv as described [here](https://pynq.readthedocs.io/en/v2.6.1/getting_started/alveo_getting_started.html?highlight=alveo#install-conda).*
+*For PYNQ boards, all commands below must be prefixed with `sudo` or by first going into `sudo su`.*
 
 First, ensure that your `pip` and `setuptools` installations are up-to-date
 on your PYNQ board or Alveo server:
@@ -62,11 +65,11 @@ dummy_out = accel.execute(dummy_in)
 |----------------------------------------------------------------|-------------------------|------------------------------------------------------------|------------------|
 | <img src="docs/img/cifar-10.png" width="150"/><br/>CIFAR-10     | CNV (VGG-11-like)       | several variants:<br>1/2-bit weights/activations           | all              |
 | <img src="docs/img/mnist.jpg" width="150"/><br/><br>MNIST       | 3-layer fully-connected | several variants:<br>1/2-bit weights/activations           | all              |
-| <img src="docs/img/imagenet.jpg" width="150"/><br/><br>ImageNet | MobileNet-v1            | 4-bit weights and activations<br>8-bit first layer weights | Alveo U250       |
+| <img src="docs/img/imagenet.jpg" width="150"/><br/><br>ImageNet | MobileNet-v1            | 4-bit weights and activations<br>8-bit first layer weights | Alveo U250<br>ZCU104       |
 
 ## Supported Boards
 
-*Note that the larger NNs are only available on Alveo boards.*
+*Note that the larger NNs are only available on Alveo or selected Zynq boards.*
 
 `finn-examples` provides pre-built FPGA bitfiles for the following boards:
 

build/README.md

@@ -27,8 +27,7 @@ Please see the READMEs under the respective subfolders here for instructions on
 
 All examples in this repo use the same Python PYNQ driver, located under
 `finn_examples/driver.py` in the repo. This driver can support any FINN-generated
-accelerator that doesn't use external weights, the only thing that needs to be
-specified is the configuration for the input and output tensors in the `io_shape_dict`. Have a look at `finn_examples/models.py` to see how this is done for the example models in this repo:
+accelerator, the only thing that needs to be specified is the configuration for the input and output tensors in the `io_shape_dict`. Have a look at `finn_examples/models.py` to see how this is done for the example models in this repo:
 
 ```python
 _cifar10_cnv_io_shape_dict = {

build/get-finn.sh

@@ -30,7 +30,7 @@
 # URL for git repo to be cloned
 REPO_URL=https://github.com/Xilinx/finn
 # commit hash for repo
-REPO_COMMIT=4fee6ffd8e13f91314ec9086e9ce9b2ea9de15c7
+REPO_COMMIT=e5da788bdc74fc9c234bb0176521ad51e830c22e
 # directory (under the same folder as this script) to clone to
 REPO_DIR=finn
 

build/mobilenet-v1/README.md

@@ -17,7 +17,8 @@ It requires about 2 MB of weight storage and 1.1 GMACs per inference, yielding
 Due to the depthwise separable convolutions in MobileNet-v1,
 we use a specialized build script that replaces a few of the standard steps
 in FINN with custom ones.
-**MobileNet-v1 is currently only supported on Alveo U250.**
+**MobileNet-v1 is currently only supported on Alveo U250 and ZCU104.**
+We also provide a folding configuration for the **ZCU102**, but there is no pre-built Pynq image available for this board.
 
 0. Ensure you have performed the *Setup* steps in the top-level README for setting up the FINN requirements and environment variables.
 

build/mobilenet-v1/build.py

@@ -28,53 +28,147 @@
 
 import finn.builder.build_dataflow as build
 import finn.builder.build_dataflow_config as build_cfg
+from finn.util.basic import alveo_default_platform
+import os
+import shutil
 
 # custom steps for mobilenetv1
 from custom_steps import (
     step_mobilenet_streamline,
     step_mobilenet_convert_to_hls_layers,
+    step_mobilenet_convert_to_hls_layers_separate_th,
     step_mobilenet_lower_convs,
+    step_mobilenet_slr_floorplan,
 )
 
 model_name = "mobilenetv1-w4a4"
-board = "U250"
-vitis_platform = "xilinx_u250_xdma_201830_2"
-synth_clk_period_ns = 3.0
 
-mobilenet_build_steps = [
-    step_mobilenet_streamline,
-    step_mobilenet_lower_convs,
-    step_mobilenet_convert_to_hls_layers,
-    "step_create_dataflow_partition",
-    "step_apply_folding_config",
-    "step_generate_estimate_reports",
-    "step_hls_ipgen",
-    "step_set_fifo_depths",
-    "step_create_stitched_ip",
-    "step_make_pynq_driver",
-    "step_synthesize_bitfile",
-    "step_deployment_package",
-]
-
-
-cfg = build_cfg.DataflowBuildConfig(
-    steps=mobilenet_build_steps,
-    output_dir="output_%s_%s" % (model_name, board),
-    folding_config_file="folding_config/%s_folding_config.json" % board,
-    synth_clk_period_ns=synth_clk_period_ns,
-    board=board,
-    shell_flow_type=build_cfg.ShellFlowType.VITIS_ALVEO,
-    # folding config comes with FIFO depths already
-    auto_fifo_depths=False,
-    vitis_platform=vitis_platform,
-    # enable extra performance optimizations (physopt)
-    vitis_opt_strategy=build_cfg.VitisOptStrategyCfg.PERFORMANCE_BEST,
-    generate_outputs=[
-        build_cfg.DataflowOutputType.PYNQ_DRIVER,
-        build_cfg.DataflowOutputType.ESTIMATE_REPORTS,
-        build_cfg.DataflowOutputType.BITFILE,
-        build_cfg.DataflowOutputType.DEPLOYMENT_PACKAGE,
-    ],
-)
-model_file = "models/%s_pre_post_tidy.onnx" % model_name
-build.build_dataflow_cfg(model_file, cfg)
+# which platforms to build the networks for
+zynq_platforms = ["ZCU102", "ZCU104"]
+#alveo_platforms = ["U50", "U200", "U250", "U280"]
+alveo_platforms = ["U250"]
+platforms_to_build = zynq_platforms + alveo_platforms
+
+
+# determine which shell flow to use for a given platform
+def platform_to_shell(platform):
+    if platform in zynq_platforms:
+        return build_cfg.ShellFlowType.VIVADO_ZYNQ
+    elif platform in alveo_platforms:
+        return build_cfg.ShellFlowType.VITIS_ALVEO
+    else:
+        raise Exception("Unknown platform, can't determine ShellFlowType")
+
+
+# select target clock frequency
+def select_clk_period(platform):
+    if platform in zynq_platforms:
+        return 5.4
+    elif platform in alveo_platforms:
+        return 3.0
+
+
+# select build steps (ZCU104/102 folding config is based on separate thresholding nodes)
+def select_build_steps(platform):
+    if platform in zynq_platforms:
+        return [
+            step_mobilenet_streamline,
+            step_mobilenet_lower_convs,
+            step_mobilenet_convert_to_hls_layers_separate_th,
+            "step_create_dataflow_partition",
+            "step_apply_folding_config",
+            "step_generate_estimate_reports",
+            "step_hls_codegen",
+            "step_hls_ipgen",
+            "step_set_fifo_depths",
+            "step_create_stitched_ip",
+            "step_synthesize_bitfile",
+            "step_make_pynq_driver",
+            "step_deployment_package",
+        ]
+    elif platform in alveo_platforms:
+        return [
+            step_mobilenet_streamline,
+            step_mobilenet_lower_convs,
+            step_mobilenet_convert_to_hls_layers,
+            "step_create_dataflow_partition",
+            "step_apply_folding_config",
+            "step_generate_estimate_reports",
+            "step_hls_codegen",
+            "step_hls_ipgen",
+            "step_set_fifo_depths",
+            step_mobilenet_slr_floorplan,
+            "step_synthesize_bitfile",
+            "step_make_pynq_driver",
+            "step_deployment_package",
+        ]
+
+
+# create a release dir, used for finn-examples release packaging
+os.makedirs("release", exist_ok=True)
+
+
+for platform_name in platforms_to_build:
+    shell_flow_type = platform_to_shell(platform_name)
+    if shell_flow_type == build_cfg.ShellFlowType.VITIS_ALVEO:
+        vitis_platform = alveo_default_platform[platform_name]
+        # for Alveo, use the Vitis platform name as the release name
+        # e.g. xilinx_u250_xdma_201830_2
+        release_platform_name = vitis_platform
+    else:
+        vitis_platform = None
+        # for Zynq, use the board name as the release name
+        # e.g. ZCU104
+        release_platform_name = platform_name
+    platform_dir = "release/%s" % release_platform_name
+    os.makedirs(platform_dir, exist_ok=True)
+
+    cfg = build_cfg.DataflowBuildConfig(
+        steps=select_build_steps(platform_name),
+        output_dir="output_%s_%s" % (model_name, release_platform_name),
+        folding_config_file="folding_config/%s_folding_config.json" % platform_name,
+        synth_clk_period_ns=select_clk_period(platform_name),
+        board=platform_name,
+        shell_flow_type=shell_flow_type,
+        vitis_platform=vitis_platform,
+        # folding config comes with FIFO depths already
+        auto_fifo_depths=False,
+        # enable extra performance optimizations (physopt)
+        vitis_opt_strategy=build_cfg.VitisOptStrategyCfg.PERFORMANCE_BEST,
+        generate_outputs=[
+            build_cfg.DataflowOutputType.PYNQ_DRIVER,
+            build_cfg.DataflowOutputType.ESTIMATE_REPORTS,
+            build_cfg.DataflowOutputType.BITFILE,
+            build_cfg.DataflowOutputType.DEPLOYMENT_PACKAGE,
+        ],
+    )
+    model_file = "models/%s_pre_post_tidy.onnx" % model_name
+    build.build_dataflow_cfg(model_file, cfg)
+
+    # copy bitfiles and runtime weights into release dir if found
+    bitfile_gen_dir = cfg.output_dir + "/bitfile"
+    files_to_check_and_copy = [
+        "finn-accel.bit",
+        "finn-accel.hwh",
+        "finn-accel.xclbin",
+    ]
+    for f in files_to_check_and_copy:
+        src_file = bitfile_gen_dir + "/" + f
+        dst_file = platform_dir + "/" + f.replace("finn-accel", model_name)
+        if os.path.isfile(src_file):
+            shutil.copy(src_file, dst_file)
+
+    weight_gen_dir = cfg.output_dir + "/driver/runtime_weights"
+    weight_dst_dir = platform_dir + "/%s_runtime_weights" % model_name
+    if os.path.isdir(weight_gen_dir):
+        weight_files = os.listdir(weight_gen_dir)
+        if weight_files:
+            shutil.copytree(weight_gen_dir, weight_dst_dir)
+
+    # create zipfile for all examples for this platform
+    shutil.make_archive(
+        "release/" + release_platform_name,
+        "zip",
+        root_dir="release",
+        base_dir=release_platform_name,
+    )

build/mobilenet-v1/custom_steps.py

@@ -26,7 +26,10 @@
 # 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.core.modelwrapper import ModelWrapper
-from finn.builder.build_dataflow_config import DataflowBuildConfig
+from finn.builder.build_dataflow_config import (
+    DataflowBuildConfig,
+    ShellFlowType,
+)
 from finn.transformation.streamline import Streamline
 from finn.transformation.double_to_single_float import DoubleToSingleFloat
 import finn.transformation.streamline.absorb as absorb
@@ -39,6 +42,7 @@
 from finn.transformation.general import (
     GiveReadableTensorNames,
     GiveUniqueNodeNames,
+    ApplyConfig,
 )
 import finn.transformation.fpgadataflow.convert_to_hls_layers as to_hls
 from finn.transformation.infer_shapes import InferShapes
@@ -94,3 +98,36 @@ def step_mobilenet_convert_to_hls_layers(model: ModelWrapper, cfg: DataflowBuild
     model = model.transform(GiveUniqueNodeNames())
     model = model.transform(GiveReadableTensorNames())
     return model
+
+
+def step_mobilenet_slr_floorplan(model: ModelWrapper, cfg: DataflowBuildConfig):
+    if cfg.shell_flow_type == ShellFlowType.VITIS_ALVEO:
+        try:
+            from finn.analysis.partitioning import partition
+            # apply partitioning of the model, restricting the first and last layers to SLR0
+            default_slr = 0
+            abs_anchors = [(0,[default_slr]),(-1,[default_slr])]
+            floorplan = partition(model, cfg.synth_clk_period_ns, cfg.board, abs_anchors=abs_anchors, multivariant=False)[0]
+            # apply floorplan to model
+            model = model.transform(ApplyConfig(floorplan))
+            print("SLR floorplanning applied")
+        except:
+            print("No SLR floorplanning applied")
+    return model
+
+
+def step_mobilenet_convert_to_hls_layers_separate_th(
+    model: ModelWrapper, cfg: DataflowBuildConfig
+):
+    mem_mode = cfg.default_mem_mode.value
+    model = model.transform(to_hls.InferPool_Batch())
+    model = model.transform(to_hls.InferConvInpGen())
+    model = model.transform(to_hls.InferThresholdingLayer())
+    model = model.transform(to_hls.InferVVAU())
+    model = model.transform(to_hls.InferQuantizedStreamingFCLayer(mem_mode))
+    model = model.transform(to_hls.InferChannelwiseLinearLayer())
+    model = model.transform(to_hls.InferLabelSelectLayer())
+    model = model.transform(InferShapes())
+    model = model.transform(GiveUniqueNodeNames())
+    model = model.transform(GiveReadableTensorNames())
+    return model