# HG changeset patch
# User nickols_k
# Date 991997764 0
# Node ID 9b0bb5c805b21753f57cdda3c0b815ccc2964cdc
# Parent  970fbd433564c4cc6d349f2eb5780ec38f8ed670
Last minute improvements before release. Added banchmark

diff -r 970fbd433564 -r 9b0bb5c805b2 mp3lib/decode_k7.s
--- a/mp3lib/decode_k7.s	Fri Jun 08 09:41:22 2001 +0000
+++ b/mp3lib/decode_k7.s	Fri Jun 08 10:56:04 2001 +0000
@@ -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