PULP-SDK
This is the latest version of the PULP SDK, which is under active development. The previous (now legacy) version, which is no longer supported, is on the v1 branch.
Citing
If you intend to use or reference GVSoC for an academic publication, please consider citing it:
@INPROCEEDINGS{9643828,
author={Bruschi, Nazareno and Haugou, Germain and Tagliavini, Giuseppe and Conti, Francesco and Benini, Luca and Rossi, Davide},
booktitle={2021 IEEE 39th International Conference on Computer Design (ICCD)},
title={GVSoC: A Highly Configurable, Fast and Accurate Full-Platform Simulator for RISC-V based IoT Processors},
year={2021},
volume={},
number={},
pages={409-416},
doi={10.1109/ICCD53106.2021.00071}}
PULP-SDK completely supports Dory as QNNs deployment tool. You can find an example of deployment [here]. If you intend to use or reference Dory for an academic publication, please consider citing it:
@ARTICLE{9381618,
author={Burrello, Alessio and Garofalo, Angelo and Bruschi, Nazareno and Tagliavini, Giuseppe and Rossi, Davide and Conti, Francesco},
journal={IEEE Transactions on Computers},
title={DORY: Automatic End-to-End Deployment of Real-World DNNs on Low-Cost IoT MCUs},
year={2021},
volume={70},
number={8},
pages={1253-1268},
doi={10.1109/TC.2021.3066883}}
Getting started
These instructions were developed using a fresh Ubuntu 18.04 Bionic Beaver 64-Bit.
The following packages needed to be installed:
sudo apt-get install -y build-essential git libftdi-dev libftdi1 doxygen python3-pip libsdl2-dev curl cmake libusb-1.0-0-dev scons gtkwave libsndfile1-dev rsync autoconf automake texinfo libtool pkg-config libsdl2-ttf-dev
The SDK also requires the argcomplete and pyelftools Python package. You can install them for the local user with:
pip install --user argcomplete pyelftools
Omit --user to install at system level instead, which will probably require admin rights.
This version requires PULP toolchain to compile the application exploiting pulp features. PULP toolchain is available at: https://github.com/pulp-platform/pulp-riscv-gnu-toolchain
You can choose also its precompiled version, exploring: https://github.com/pulp-platform/pulp-riscv-gnu-toolchain/releases/tag/v1.0.16
Please, refer to the corresponding README for the installation.
Once PULP toolchain is correctly installed, define the path in which there is toolchain bin folder:
export PULP_RISCV_GCC_TOOLCHAIN=<INSTALL_DIR>
Source the file corresponding to the desired configuration:
cd pulp-sdk
source configs/pulp-open.sh
At least gcc 4.9.1 is needed. If the default one is not correct, CC and CXX can be set to point to a correct one. To check if gcc has the right version:
gcc --version
Please, refer to official guide to update gcc if is needed.
GVSoC
GVSoC is the PULP chips simulator that is natively included in the SDK and is described and evaluated fully in Bruschi et al. [arXiv:2201.08166v1].
Compile GVSoC with this command:
make build
Test execution
Some examples are availaible at https://github.com/GreenWaves-Technologies/pmsis_tests
Then, go to a test, for example pmsis_tests/quick/cluster/fork/, and execute:
make clean all run
This will by default execute it on GVSoC (platform=gvsoc), and you can configure the RTL platform with this command:
make clean all run platform=rtl
Notice that the environment variable VSIM_PATH should be set to the directory where the RTL platform has been built. This is typically done by sourcing the setup/vsim.sh file from the main folder of the RTL platform.
To run pre-generated real-world networks, such as MobileNetV1:
cd applications/MobileNetV1
make clean all run platform=<PLATFORM> CORE=<NUM_CORES>
Nemo + Dory + Pulp-NN
Our vertical flow allows to deploy optimized QNNs on low-power and low-resources MCUs, starting from a Pytorch model.
Nemo
[Nemo] is a framework for Deep Neural Networks layer-wise quantization. He starts from a common Pytorch project and produces an equivalent quantized model, which well suits the usually integer MCUs. Its output are a .onnx as quantized model and several .txt as set of input and weigths of the network, also including the golden activations to checks the output of every network's layer. Please refer to its README for more details and [here] you can find a Colab project and a very detailed tutorial on how to get started with Nemo.
Dory
[Dory] is an automatic tool to generate and directly deploy MLP/CNNs on PULP family boards, exploiting [Pulp-NN] as optimized back-end.
Dory has a complete and autonomous testsuite, named [Dory-Example], which is periodically updated, and please refer to its README for more details. To generate the code and run one of these examples:
cd dory/dory_examples/
python3 network_generate --network_dir <e.g., ./examples/MobileNetV1/>
cd application
make clean all run platform=<PLATFORM> CORE=<NUM_CORES>
where you should choose CORE=8 if you want to test the network on pulp cluster with all of the eight cores active (by default only 1 is set).
To set up and execute a custom application, firstly, copy your file network.onnx and files out_layer{i}.txt in a single folder (e.g., pulp-sdk/application/MyCustomNetwork/) and then:
cd dory/dory_examples/
python3 network_generate --network_dir <pulp-sdk/application/MyCustomNetwork/>
cd application
make clean all run platform=<PLATFORM> CORE=<NUM_CORES>
You can use L1 and L2 memory constraints to specify the amount of memory used inside the application. Please refer to Dory and Dory-example READMEs for more details.
but I have the following error :
![pulp-run: Error while building component [/sys/board/chip/soc/cluster/l1_ico/interleaver]](https://avatars.githubusercontent.com/u/15718670?v=4)
Thank you!
I would expect that the CS stays low after the first function call, as described in :
Indeed, the signal generated is the following (yellow: CLK, blue: CS):
53 Dec 04, 2022
215 Dec 07, 2022
10 Oct 10, 2022
4.3k Jan 08, 2023
4 Feb 15, 2022
243 Dec 26, 2022
21 Nov 24, 2022
31 Oct 19, 2019
53 Dec 30, 2022
132 Sep 23, 2022
423 Dec 07, 2022
1.9k Dec 15, 2022
68 Dec 13, 2022
37 Oct 26, 2022
2 Dec 15, 2020
34 Jan 01, 2023
26 Nov 23, 2022
9 Dec 08, 2022
47 Jan 01, 2023
3.8k Dec 29, 2022