README

This project is about doing general-purpose computing on GPUs at machine code
level. I am using those graphics processors from AMD for which documentation is
available, specifically these are the R6xx, R7xx, Evergreen and AMD HD 6900
series right now. For setting up the processing pipeline of these chips I am
using nothing but the Linux DRI and the documentation of these chips.

I do not believe that programming GPUs at machine code level is an interesting
option. This project is about gaining a deeper understanding of these chips and
exploring the possibilities of the instruction set may be a key element.

We have seen emerge in the past few years many programming models and languages
for GPGPU computing. Some of these, like CUDA and OpenCL, rely on components
and compilers supplied by hardware vendors. Others rely on translators from
higher-level languages to CUDA or OpenCL.

I feel that, for those of us who are interested in a higher-level programming
model, the translation approach has serious drawbacks. The CUDA and OpenCL
programming models have limitations which may not be inherent to the chip.
On the other hand, the way that the CUDA or OpenCL programming model is mapped
to the instruction set may not be the most efficient for a very different
programming model, such as that of a declarative programming language.


                                        Miguel Lopes Santos Ramos