From 49afe84db7be02a175c490b1221dd215bbfddf35 Mon Sep 17 00:00:00 2001 From: nickols_k Date: Fri, 8 Jun 2001 10:56:04 +0000 Subject: [PATCH] Last minute improvements before release. Added banchmark git-svn-id: svn://svn.mplayerhq.hu/mplayer/trunk@1055 b3059339-0415-0410-9bf9-f77b7e298cf2 --- mp3lib/decode_k7.s | 9 +++++++-- 1 file 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 /// / synth_1to1_3dnow works the same way as the c version of