xyf2002/CoSense

CoSense is a compiler that uses sensor information from datasheet to help with optimization. It is developed based on LLVM.

CoSense

Cosense is a part of the Noisy&Newton Project. Please check it for more details.

The main idea of CoSense is to use sensor specifications to help compilation optimization, including value range propagation, function overload & elimination, type compression, condition simplification, and constant substitution.

---

Getting started

The correct way to clone this repository to get the submodules is:

	git clone --recursive git@github.com:systems-nuts/CoSense.git

To update all submodules:

	git pull --recurse-submodules
	git submodule update --remote --recursive

If you forgot to clone with --recursive and end up with empty submodule directories, you can remedy this with

	git submodule update --init

Building the Noisy compiler and debug tools depends on the libflex, Wirth-tools, and DTrace-scripts repositories. These repositories are already included as submodules:

	Libflex:		git@github.com:phillipstanleymarbell/libflex.git
	Wirth tools:		git@github.com:phillipstanleymarbell/Wirth-tools.git
	DTrace-scripts:		git@github.com:phillipstanleymarbell/DTrace-scripts.git

For linear algebra in Newton, we use the Eigen library. This is also already linked to the repository as a submodule:

	Eigen:			git@github.com:eigenteam/eigen-git-mirror.git	

The build also depends on the C protobuf compiler, sloccount, and on Graphviz.

sudo apt install libprotobuf-c-dev protobuf-c-compiler sloccount graphviz-devel

Furthermore, LLVM is a build and runtime dependency on this project. Currently, passes related to LLVM are tested with LLVM 13 versions.

Make sure llvm-config is installed for one of the above versions. In case it is named differently, e.g., llvm-config-x you will need to create a symbolic link:

cd /location/of/llvm-config-x
ln -s llvm-config-x llvm-config

Once you have the above repositories,

1. Create a file config.local in the root of the Noisy tree and edit it to contain

	LIBFLEXPATH     = full-path-to-libflex-repository-clone
	CONFIGPATH      = full-path-to-libflex-repository-clone
	OSTYPE		= linux
	MACHTYPE	= x86_64

For example,

	LIBFLEXPATH=/home/me/Noisy-lang-compiler/submodules/libflex
	CONFIGPATH=/home/me/Noisy-lang-compiler/submodules/libflex
	OSTYPE		= linux
	MACHTYPE	= x86_64

2. Copy config.local to the libflex directory

	$ cp config.local submodules/libflex

3. In src/common/Makefile and src/newton/Makefile, change COMPILERVARIANT as necessary (default is gcc).

4. Build Libflex by going to the directory you cloned for Libflex and running make. The Makefile assumes the environment variables OSTYPE and MACHTYPE are set.

	$ cd submodules/libflex
	$ make

5. From the root of this top-level repository, build the Noisy and Newton compilers by running make. The makefile assumes the environment variables OSTYPE and MACHTYPE are set.

	$ make -j32

---

Experiments

Micro Benchmarks

Our micro-benchmark test cases are listed in applications/newton/llvm-ir.

1. Following the LLVM Benchmarking tips. We summarize it in a bash file.

sudo bash ./env.sh

In this bash file, we'll set the necessary benchmarking configs and then run the whole micro-benchmarks to check the compilation and execution. To reset the configs, run

sudo bash ./reset.sh

2. We have an auto_test.cpp to run and log the micro-benchmarks. To compile the test file and run:

make auto_test_compile
./auto_test 2> err.log

Users can get the average performance in average_speedup.log and the detailed performance numbers in perf.log. For the compilation log, please check compile.log, and please check err.log if any error happens.

3. Optional. We provide Python scripts to plot the figures as shown in our paper.
   • plot_sensor_ranges.py: Plot the heatmaps.
   • bar_plot.py: Plot the average speedup bar figure.