I'm leaving this up for posterity, but users looking for a better Halide build experience should just use the new CMake build.

This is a simple tool for quickly creating Halide projects and generators with sane Makefiles. Current features include:

• Create new-style benchmark tools and runners by default.
• Manage multiple configurations per generator
• Precompiled headers for faster incremental builds

Planned features include:

• Merge all configurations into a single static lib with a suitable runtime
• Integrate with the autoscheduler (need upstream support -- weights are not included in the distributions yet)
• Out-of-tree builds

This project is hours old. Therefore expect:

• No support for non-GNU make or cmake
• No macOS support
• No support for compilers other than GCC (and maybe clang)
• Lots of bugs

Feel free to contribute, though I'm rapidly bringing this to alpha.

Here's the quickest way to see how hl works:

$ hl create project my_project
$ cd my_project
$ ls
Makefile  my_project.gen.cpp
$ make
...
$ ./run_my_project input=random:0:auto --output_extents=[64,64] --benchmarks=all
Benchmark for my_project produces best case of 4.49293e-07 sec/iter (over 10 samples, 904060 iterations, accuracy 5%).
Best output throughput is 8.7e+03 mpix/sec.
$ ls
Makefile  kernels/  my_project.gen.cpp  my_project.generator  run_my_project
$ ls kernels
RunGenMain.o  my_project.h     my_project.registration.cpp  stdafx.h
my_project.a  my_project.html  my_project.stmt              stdafx.h.gch

The default pipeline that's generated is a plain image copy. The Makefile automatically generates a precompiled header that includes just Halide.h. It includes the header via -include, which is supported by GCC and Clang.

TODO add a way of specifying additional headers to include in the PCH

This tool assumes you wish to use Halide in ahead-of-time mode. It will create and manage .gen.cpp files, each of which contain must contain a single generator class. Generators can have multiple configurations, each of which has a unique identifying name. The C functions are named <generator> for the default configuration, and <generator>__<config-name> for other configurations. These names are also used for the library files.

Running make will (by default) generate a run_<configuration> executable for each configuration specified in the Makefile. These are based on RunGenMain.cpp, a new tool in upstream Halide . Thus the "latest" release available from Github is incompatible with this tool. Intermediate files and libraries are collected in a directory called kernels/. You can also make just the libraries without the runners by calling make generate_<configuration>.

To consume the libraries produced by the Makefiles, just run Make recursively.

Just clone the repository to some location (maybe ~/bin) and add it to your path. For example, you could add the following line to your ~/.profile or ~/.bashrc:

export PATH="$PATH:$HOME/bin/halide-project-tool"

This project can use a system-installed Python 3 since it has no additional Python dependencies.

• Python 3.4+
• A recent Halide distribution or build (distrib target) with the environment variable HALIDE_DISTRIB_PATH set to the root. You can also install Halide to /opt/halide if you don't want to set an environment variable.