lovefawn / harlan

A language for GPU computing.

Harlan

Harlan is a declarative, domain specific language for programming GPUs. This project serves primarily as a testbed for implementation and optimization techniques. The language is intentionally small, in order to simplify the process of exploring new analyses and optimizations.

Getting Started

Harlan is known to build and run on the following operating systems.

• Mac OS X 10.6 (Snow Leopard)
• Mac OS X 10.7 (Lion)
• Mac OS X 10.8 (Mountain Lion)
• Various flavors of Linux

Others will probably work as well.

Harlan requires an OpenCL implementation as well as Petite Chez Scheme. Below are several OpenCL implementations that should work.

• Intel OpenCL SDK
• NVIDIA CUDA Toolkit
• AMD Accelerated Parallel Processing (APP) SDK

Petite Cheze Scheme can be downloaded from http://www.scheme.com/download.

Once all the prerequisites are installed, you can compile and run the test suite as follows.

make check

If the tests are successful, you will see the following at the end of all the output:

All tests succeeded.

The test programs are available in the test directory. End-to-end test programs have the .kfc extension. Other extensions represent code that is valid at various intermediate passes in the compiler.

Make puts test binaries in the test.bin directory, and also saves output from test programs here. Programs may be run directly from this directory for easier debugging.

Harlan programs can be compiled manually as follows.

./harlanc hello.kfc

For debugging purposes, the -v flag can be used.

./harlanc -v hello.kfc

This causes the compiler to write out the intermediate results from each compiler pass.

Assuming the Harlan compiler is successful, the compiler will produce an executable based on the program's filename. For example, the previous example will produce the file hello, which can be executed directly.

If Latex is installed, you can build the Harlan User's Guide as follows.

make docs