Here are detailed the steps necessary to use the tools included in the Docker image to execute and compile EDDL code.
The Docker image can be cloned into any device with the command:
docker pull solverml/eddl_riscv:artifacts
Once cloned, we can run the image as a container by executing the command:
docker run -t -d solverml/eddl_riscv:artifacts /bin/bash
To access the Docker container from a terminal run the following command:
docker exec -i -t container_id /bin/bash where 'container_id' needs to be replaced by the ID assign to the Docker container we have just created. To check this ID simply run the command docker ps. This will keep open a terminal allowing to execute commands as a root user inside the docker container.
Starting from the terminal to the container and then moving to the directory root/artifacts will show the files necessary to run the RISC-V emulation, as well as a shellscript file which will initiate the emulation.
To start the emulation run the following command: sh run_qemu.sh. Once the startup of the emulation is complete it will prompt us to introduce a user and password. Type 'root' in both to access the RISC-V system.
The terminal then will change its focus from the container to the RISC-V emulation, where its possible to run and compile code using the EDDL library since it's already been installed.
In the RISC-V emulation provided here is possible to compile C++ code using the EDDL library by writing a CMakeList.txt file along with the main code of a program and using the cmake command.
To illustrate this process and to show the capabilities of the EDDL various example programs are include in the emulation, each appliying the train and test of neural networks to different types of datasets. Inside the emulation, in the folder ~/eddl_examples/ the different examples can be found. The description of every example can be found in this GitHub repository in the section Code Examples.
For example, to compile the code from the MNIST train example we would move to the folder ~/eddl_example/mnist_eddl_train where the essential files would be the main code (main.cpp) and a CMakeList.txt file. In this folder we would create a new folder called build and move to it. Here we would execute the commands cmake .. and make and as a result we would obtain a new executable file which will run the compiled code. To compile and run custom code simply imitate the same structure of folders and commands describe here.
In the emulation all code examples and programs are already compiled.
To test custom code it may be necessary to transfer files in and out of the RISC-V emulation. Its possible to do this using ssh commands. The IP correponding the host machine running the Docker cotainer is visble to the RISC-V emulation as 127.17.0.1, so by executing a command similar to scp {host_user}@127.17.0.1 {path_in_host} {path_in_emulation} is possible to transfer a file from the host machine to the emulation. Inverting the paths is possible to transfer a file from the emulation to the host.
Since QEMU has been use to create the emulation its possible to type special commands that will affect the state of the emulation by typing Ctrl + A, C. For example:
- By default, when closing the emulation all changes made to it will be lost. To commit those changes before exiting the emulation use the special QEMU command
commit all. - To stop the emulation use the command
quit - The full list of QEMU commands can be found here https://en.wikibooks.org/wiki/QEMU/Monitor.