9 Mar
2008
9 Mar
'08
11:37 p.m.
On Sunday 09 March 2008 19:59, Charles Henry wrote:
Pd for 64-bit processors could potentially redefine t_sample as a double, with no loss in performance (with nearly twice as much memory usage).
Mmmmmno...
It just so happens that x87s always compute a double, so it makes little difference there. What you're forgetting is that if you're using SIMD (for example SSE2) using doubles means you can only fit two samples into a vector register instead of four. Half the throughput. On top of that, if you check instruction tables, packed double instructions have a slightly higher latency than packed singles. For the core2 at least.
Not everything (even 64bit) uses an x87.
robert.