> > Do we still care about people not having sse2 ? > > No, I believe (and someone will likely correct me) that its > assumed that people who are using SSE processors have > processors that are SSE2 capable. This pretty much includes > most SSE enabled processors this decade (with very few exceptions)
Intel introduced SSE2 in 2000 (Netburst Architecture, P4 and P4M and up) and AMD in 2001 (K7 Architecture, Athlon 64 and up). It is safe to assume all AMD/Intel x86-CPUs sold after 2003 have SSE2 support (Crusoe and VIA may not). For plain SSE the respective introduction years were 1999 (Intel P III and up) and 2001 (Athlon XP/MP and up). In my benchmarks, the SSE2 sigmoid code alone added about 10% performance to a NN based batch eval. Once you unconditionally use SSE2 anyway, it also makes sense to completely wave goodbye to x387 code by switching GCC SSE math mode. That avoids some float conversion instructions needed when you mix both FPU units, and probably saves another bit of time. => "gcc -mfpmath=sse -msse2 -msse" should be a no-brainer for the binary distribution. Ingo _______________________________________________ Bug-gnubg mailing list [email protected] http://lists.gnu.org/mailman/listinfo/bug-gnubg
