summaryrefslogtreecommitdiffstats
path: root/mp3lib
diff options
context:
space:
mode:
Diffstat (limited to 'mp3lib')
-rw-r--r--mp3lib/decode_k7.s9
1 files changed, 7 insertions, 2 deletions
diff --git a/mp3lib/decode_k7.s b/mp3lib/decode_k7.s
index edea66d567..93803d8625 100644
--- a/mp3lib/decode_k7.s
+++ b/mp3lib/decode_k7.s
@@ -7,15 +7,20 @@
/// - Added new opcode PFNACC
/// - decreased number of opcodes (as it was suggested by k7 manual)
/// (using memory reference as operand of instructions)
-/// - added PREFETCHW opcode. It has different semantic than k6-2
+/// - added PREFETCHW opcode. It has different semantic on k7 than on k6-2
/// and saves 15-25 cpu clocks for athlon.
/// - partial unrolling loops for removing slower MOVW insns.
-/// (Note probably same operation should be done for decode_3dnow)
+/// (Note: probably same operation should be done for decode_3dnow.s)
/// - change function name for support 3DNowEx! automatic detect
///
/// note: because K7 processors are an aggresive out-of-order three-way
/// superscalar ones instruction order is not significand for them.
///
+/// Benchmark: measured by mplayer on Duron-700:
+/// 3dNow! optimized code - 1.5% of cpu usage
+/// k7 optimized code - 1.1% of cpu usage
+/// Note: K6-2 users have an chance with partial loops unrolling
+///
/// Modified by Nick Kurshev <nickols_k@mail.ru>
///
/ synth_1to1_3dnow works the same way as the c version of