# HG changeset patch # User michael # Date 1273612255 0 # Node ID 6f9b4c452130e9e7471465354338f2d8036c227b # Parent 8d2e66ba12cf79c6c4f0fcfb2a1489f32995586e Optimize decoding high freqs. this is 10-20cpu cycles faster on duron (whole is about 50-60 cpu cylses) I wonder why gcc isnt doing this on its own ... diff -r 8d2e66ba12cf -r 6f9b4c452130 mpegaudiodec.c --- a/mpegaudiodec.c Tue May 11 20:33:48 2010 +0000 +++ b/mpegaudiodec.c Tue May 11 21:10:55 2010 +0000 @@ -1535,14 +1535,22 @@ g->sb_hybrid[s_index+3]= 0; while(code){ static const int idxtab[16]={3,3,2,2,1,1,1,1,0,0,0,0,0,0,0,0}; - INTFLOAT v; + int v; int pos= s_index+idxtab[code]; code ^= 8>>idxtab[code]; - v = RENAME(exp_table)[ exponents[pos] ]; -// v = RENAME(exp_table)[ (exponents[pos]&3) ] >> FFMIN(0 - (exponents[pos]>>2), 31); - if(get_bits1(&s->gb)) //FIXME try to flip the sign bit in int32_t, same above +/* Following is a optimized code for + INTFLOAT v = RENAME(exp_table)[ exponents[pos] ]; + if(get_bits1(&s->gb)) v = -v; g->sb_hybrid[pos] = v; +*/ +#if CONFIG_FLOAT + v = AV_RN32A(RENAME(exp_table)+exponents[pos]) ^ (get_bits1(&s->gb)<<31); + AV_WN32A(g->sb_hybrid+pos, v); +#else + v= -get_bits1(&s->gb); + g->sb_hybrid[pos] = (RENAME(exp_table)[ exponents[pos] ] ^ v) - v; +#endif } s_index+=4; }