The SPARC/Solaris platform is just slow. I went through a lot of the workunits that were done on the machines and their processing time is way up there compared to other platforms. In the stats for Sparcs, not one is above 1000 in fpiops and Iops isn't much better, 2510 is tops with the top 10 percent in the 1500-1000 range.
Thanks, this looks useful. However I just checked again - -mcpu=ultrasparc doesn't gain anything with the current code. I expect some speedup once I get the autovect feature of GCC 4 to recognize our loops as vectorizable and use vis, but this may take some time.
Thanks, this looks useful. However I just checked again - -mcpu=ultrasparc doesn't gain anything with the current code. I expect some speedup once I get the autovect feature of GCC 4 to recognize our loops as vectorizable and use vis, but this may take some time.
BM
Another thing to look at, just released by Sun is their own SPARC optimizing backend for GCC. It uses the GCC ABI, so it should work with other GCC-built apps (i.e. boinc). http://cooltools.sunsource.net/gcc/index.html
Another thing to look at, just released by Sun is their own SPARC optimizing backend for GCC. It uses the GCC ABI, so it should work with other GCC-built apps (i.e. boinc). http://cooltools.sunsource.net/gcc/index.html
Thanks, this looks incredibly useful. However, this is only available for Solaris 9 and up.
I'm currently trying to compile the vector-code that's e.g. in the current Linux Beta Test App into a Solaris App that uses the VIS extension. However I have to use Suns linker and assembler, which forces me to compile _everything_ (all code and libraries used) in 64 Bit in order to properly link it together. Has anyone built a full 64-bit-enabled gcc (4.0.1 or newer) on Solaris 7?
The SPARC/Solaris platform is
)
The SPARC/Solaris platform is just slow. I went through a lot of the workunits that were done on the machines and their processing time is way up there compared to other platforms. In the stats for Sparcs, not one is above 1000 in fpiops and Iops isn't much better, 2510 is tops with the top 10 percent in the 1500-1000 range.
SPARC optimizations with GCC.
)
SPARC optimizations with GCC. I think this may help future development concerning Sparcs if it wasn't done already.
Article from OSNews.
SPARC Optimizations
RE: SPARC optimizations
)
Thanks, this looks useful. However I just checked again - -mcpu=ultrasparc doesn't gain anything with the current code. I expect some speedup once I get the autovect feature of GCC 4 to recognize our loops as vectorizable and use vis, but this may take some time.
BM
BM
RE: RE: SPARC
)
Another thing to look at, just released by Sun is their own SPARC optimizing backend for GCC. It uses the GCC ABI, so it should work with other GCC-built apps (i.e. boinc).
http://cooltools.sunsource.net/gcc/index.html
RE: Another thing to look
)
Thanks, this looks incredibly useful. However, this is only available for Solaris 9 and up.
I'm currently trying to compile the vector-code that's e.g. in the current Linux Beta Test App into a Solaris App that uses the VIS extension. However I have to use Suns linker and assembler, which forces me to compile _everything_ (all code and libraries used) in 64 Bit in order to properly link it together. Has anyone built a full 64-bit-enabled gcc (4.0.1 or newer) on Solaris 7?
BM
BM
A new App for Solaris SPARC
)
A new App for Solaris SPARC is available. See this thread.
BM
BM
RE: I expect some speedup
)
Hm. I just found out that VIS is integer only. Doesn't help for our purposes at all. Bad luck.
BM
BM