On Sun, Mar 29, 2015 at 9:51 PM, Connor Abbott <cwabbo...@gmail.com> wrote: > On Sun, Mar 29, 2015 at 5:54 PM, Thomas Helland > <thomashellan...@gmail.com> wrote: >> Does anyone have suggestions for good literature on compilers? >> Since GPU's and CPU's are a bit different there are probably >> books that are better suited than others for GPUs? >> I have what is probably Norway's biggest library on the subject to rent >> books from, so I guess I should be able to find most suggestions there. >> >> Regards >> Thomas >> >> >> _______________________________________________ >> mesa-dev mailing list >> mesa-dev@lists.freedesktop.org >> http://lists.freedesktop.org/mailman/listinfo/mesa-dev >> > > Hi, > > Unfortunately there seems to be a bit of a dearth of books when it > comes to compiler optimizations and backend things, especially when it > comes to SSA -- which is the most interesting part! I would skip the > dragon book unless you're interested in the front-end things (parsing, > symbol tables, etc.); I haven't read a lot of it myself, but > apparently the latest edition only has a passing mention of SSA and > others don't mention it at all. GCC has a list of books: > https://gcc.gnu.org/wiki/ListOfCompilerBooks and apparently some of > them are better in that regard. Personally, I learned about a lot of > this stuff through papers. I have a list of them here: > > http://cwabbottdev.blogspot.com/2013/06/compiler-theory-links.html > > although it may be a little out of date, and some of the things might > have been more useful back when I was working on lima. Suggestions for > adding to the list are very welcome.
Ok, since that page was a little lacking I added a few more links. Also, as to GPU-specific things... well, there really isn't much of anything I'm aware of that's out there. There's one paper called "Divergance Analysis and Optimizations" that's specific to SIMD machines that run multiple threads execution at once like GPU's, and we'd like to implement that in the future to give us more precise information about how we can move derivatives and textures that take an implicit derivative, since they can't be moved out of uniform control flow, but other than that there isn't anything I'm aware of. Most of the other things carry over to both CPU's and GPU's. One other GPU-specific problem I'm aware of is how to go out of SSA on classic vector-based architectures like i965 vec4 VS without introducing extra copies due to writemasked operations, but afaik there isn't a publically-available description of someone's solution -- the paper has yet to be written :). _______________________________________________ mesa-dev mailing list mesa-dev@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/mesa-dev
