Hello,
On Mon, Mar 10 2025, Zhang lv via Gcc wrote: > Hi here, > > I'm a Master of Computer Science student at the University of Melbourne. > Previously, I worked on implementing a GCC optimization prefetching pass > (which involved loop unrolling) for an Alpha-like computer architecture. > I'm set to complete my research project in July and graduate soon. great, that sounds you are definitely qualified. Is this prefething project available anywhere? > > I'm very interested in applying for the 2025 GSoC project and discussing > potential ideas with the community. However, I'm still unfamiliar with the > best way to engage in these discussions and apply for the project, so this > is my first attempt at reaching out. I think emailing gcc@gcc.gnu.org and fort...@gcc.gnu.org is exactly the right thing to do. > > My primary interest is in projects related to auto-parallelization, > particularly the Fortran *DO CONCURRENT* project. Below, I outline my > initial understanding of the project and would appreciate any feedback from > the community to ensure I'm on the right track: > I am not a Fortran front-end developer but... > 1. The *front-end parser* processes the Fortran *DO CONCURRENT* syntax > and converts it into a language-independent IR—*GIMPLE*. This sounds correct. The important bit, AFAIU, is that the IR would utilize a gimple statement for parallel execution that is so far created for OpenMP input. > 2. The *middle-end* then applies tree optimizations, utilizing SSA > passes to optimize the code for auto-parallelization. Not necessarily, some OpenMP bits are processed in gimple form that is not yet in SSA. I also don't think any auto-parallelization is involved, the whole point of the task is to implement manual parallelization, I believe? > 3. This project will involve both *front-end* work (parser and > parameterized command-line flags) and *middle-end* optimizations > (optimization passes). Right, though the middle end bits might be mostly about proper "lowering" rather than optimization. > > Loop unrolling is influenced by multiple factors, such as loop nesting and > whether iteration counts are deterministic. A significant performance gain > comes from reducing array memory access delays, which techniques like > prefetching can help with. Since *DO CONCURRENT* assumes iteration > independence, we have more flexibility to unroll loops and implement > parallelization efficiently. Frankly, I don't quite understand the above. If anything, I'd expect iteration independence makes unrolling much harder. But I fail to see how unrolling is in any way relevant. Good luck! Martin > > One of the most exciting advantages of this project is that it enables > auto-parallelization for a broader range of code without requiring > developers to be an OpenMP expert. *DO CONCURRENT* in Fortran is much > easier to use, and previous auto-parallelization techniques have been quite > limited in their applicability, as only specific loop structures could > benefit from them. > > I look forward to engaging with the community and gaining insights on how > best to contribute to this project. > > Best regards, > > Chenlu Zhang
