>> how can loongson 3 be (roughly) compared to x86 CPUs in performance? > >It's slower. A hell lot slower. > >3A systems are running at around 1GHz. The x86 code translation stuff >was benchmark-only and, to the best of my knowledge, has never been made >public (with full source code and acceptable licence terms). > >I don't see any point in this kind of benchmark. Either you are able to >recompile the x86 code on the Loongson system, and you need to compare >the speed of native Loongson code versus native X86 code. Or you can >only run the X86 code through the Loongson JIT-like code, and this is >the least of your problems because you have no idea what the original >code does and what the JIT does.
AFAIK the JIT is Qemu's; the extra instructions just help the translation from x86 -> "tiny code generator" bytecode (similar to LLVM) -> Loongson. I doubt there's much magic to it other than minimizing host CPU instructions but... I'm talking out of my ass. On the other hand you're right to question those benchmarks, after all nothing beats 1:1 x86 -> x86 translation (nop) so those were no doubt very theoretical. Even if the 70% of native speed were true it'd just mean that non-KVM Qemu is 30% slower on Loongson than on x86. So "not as lame as you'd think" seems a more accurate qualifier than "fast". -- p
