Re: [mpg123-devel] fixed point decoders

[Thomas O. <tho...@or...>]

Am Mon, 01 Jun 2009 22:26:03 +0900
schrieb Taihei Monma <tm...@ma...>: 

> The behavior when using -msse without --mfpmath=sse is: basic maths  
> are done on x87 and fp-to-int truncations are done on SSE. And on  
> Athlon, in most cases the x87 code is faster than the SSE code. So  
> yeah, your findings are almost expected.

Yeah... that nearly sounds reasonable;-) It's justified, then, not to use mfpmath=sse wherever it's possible.
Some people just use it per default because they can.
But anyhow, good that we have hand-tuned SSE code so that these oddities of compiler flags are minor. Perhaps there may be a time where you can trust a compiler to produce optimal code *cough*.

> Hmm... Generally 1.2ghz core2 should run faster than 1.4ghz athlon. Is  
> your core2 running in down-clocked mode, isn't it? I had a experience  
> that a Pentium M didn't get a full speed when running a very short  
> benchmark.

It is clocking up for the test over 17 seconds. And even if it didn't, its low-clock mode is 800MHz.
Even for that I would expect more than half the performance of an old K7 core at less than doubled clock. Tempted to pull out the Alpha EV67 again, to ask it about it's opinion on the performance of old designs;-)
Somehow I would not be too surprised if the 667MHz EV67 punts the Core2 at 800MHz.
But, well, this Core2 U7600 seems to be different enough from your average desktop Core2.


Alrighty then,

Thomas.