diff -U 3 -H -d -r -N -- mpg123-0.59r-15/decode_i486.c mpg123-0.59r/decode_i486.c --- mpg123-0.59r-15/decode_i486.c 1998-11-19 11:42:36.000000000 +0100 +++ mpg123-0.59r/decode_i486.c 2004-09-20 20:57:44.000000000 +0200 @@ -246,3 +246,108 @@ return clip; } +int synth_2to1_486(real *bandPtr,int channel,unsigned char *out, int nb_blocks) +{ + static int buffs[2][2][17*FIR_BUFFER_SIZE]; + static int bo[2] = { FIR_SIZE-1, FIR_SIZE-1 }; + short *samples = (short *) out; + int *b0,(*buf)[17*FIR_BUFFER_SIZE]; + int clip = 0; + int block,b,bo_start; + + /* samples address */ + samples+=channel; + + bo_start=bo[channel]; + buf = buffs[channel]; + + b=bo_start; + for(block=0;block= FIR_BUFFER_SIZE) { + int *p,*q; + int c,i,j; + + /* we shift the buffers */ + for(c=0;c<2;c++) { + p=&buf[c][0]+1; + q=p+(FIR_BUFFER_SIZE-FIR_SIZE); + for(i=0;i<17;i++) { + for(j=0;j= FIR_BUFFER_SIZE) b=FIR_SIZE; + if(b & 1) { + b0 = buf[0] + b - (FIR_SIZE-1); + } else { + b0 = buf[1] + b - (FIR_SIZE-1); + } + + FIR16_1(0,-7,53,-114,509,-1288,1643,-9372,18759,9372,1643,1288,509,114,53,7,0); + b0+=FIR_BUFFER_SIZE;\ + FIR16_2(1,-6,50,-86,520,-1106,1910,-8447,18714,10294,1322,1469,488,145,55,8,0, + 15,0,-8,55,-145,488,-1469,1322,-10294,18714,8447,1910,1106,520,86,50,6); + b0+=FIR_BUFFER_SIZE;\ + FIR16_2(2,-4,47,-61,521,-926,2123,-7528,18578,11205,944,1647,455,177,56,10,0, + 14,0,-10,56,-177,455,-1647,944,-11205,18578,7528,2123,926,521,61,47,4); + b0+=FIR_BUFFER_SIZE;\ + FIR16_2(3,-4,44,-38,514,-751,2284,-6620,18353,12097,509,1817,411,212,57,12,0, + 13,0,-12,57,-212,411,-1817,509,-12097,18353,6620,2284,751,514,38,44,4); + b0+=FIR_BUFFER_SIZE;\ + + samples+=32; + } + samples-=32*nb_blocks; + + /* filter bank: part 2 */ + + b=bo_start; + for(block=0;block= FIR_BUFFER_SIZE) b=FIR_SIZE; + if(b & 1) { + b0 = buf[0] + b - (FIR_SIZE-1) + 8*FIR_BUFFER_SIZE; + } else { + b0 = buf[1] + b - (FIR_SIZE-1) + 8*FIR_BUFFER_SIZE; + } + + FIR16_2(4,-3,40,-18,500,-582,2398,-5732,18042,12963,17,1977,353,247,56,14,0, + 12,0,-14,56,-247,353,-1977,17,-12963,18042,5732,2398,582,500,18,40,3); + b0+=FIR_BUFFER_SIZE;\ + FIR16_2(5,-2,36,0,479,-423,2465,-4869,17647,13794,-530,2122,282,284,53,17,0, + 11,0,-17,53,-284,282,-2122,-530,-13794,17647,4869,2465,423,479,0,36,2); + b0+=FIR_BUFFER_SIZE;\ + FIR16_2(6,-2,33,14,454,-273,2491,-4038,17173,14583,-1133,2249,198,320,50,19,0, + 10,0,-19,50,-320,198,-2249,-1133,-14583,17173,4038,2491,273,454,-14,33,2); + b0+=FIR_BUFFER_SIZE;\ + FIR16_2(7,-1,29,26,424,-136,2479,-3245,16623,15322,-1788,2354,100,357,44,22,-1, + 9,-1,-22,44,-357,100,-2354,-1788,-15322,16623,3245,2479,136,424,-26,29,1); + b0+=FIR_BUFFER_SIZE;\ + FIR16_1(8,-1,0,36,0,-11,0,-2493,0,16004,0,2431,0,391,0,26,0); + + samples+=32; + } + + return clip; +} diff -U 3 -H -d -r -N -- mpg123-0.59r-15/layer3.c mpg123-0.59r/layer3.c --- mpg123-0.59r-15/layer3.c 2004-09-20 22:21:18.000000000 +0200 +++ mpg123-0.59r/layer3.c 2004-09-15 00:58:45.000000000 +0200 @@ -1818,7 +1818,7 @@ } #ifdef I486_OPT - if (fr->synth != synth_1to1 || single >= 0) { + if ((fr->synth != synth_1to1 && fr->synth != synth_2to1) || single >= 0) { #endif for(ss=0;ss= 0) { @@ -1845,16 +1845,27 @@ #ifdef I486_OPT } else { /* Only stereo, 16 bits benefit from the 486 optimization. */ + /* However, a i486dx4@75Mhz still needs *optimized* 2to1 downsampling */ + int block_size; + + block_size=2*32; + if (fr->synth == synth_1to1) block_size=block_size*2; + ss=0; while (ss < SSLIMIT) { int n; - n=(audiobufsize - pcm_point) / (2*2*32); + n=(audiobufsize - pcm_point) / block_size; if (n > (SSLIMIT-ss)) n=SSLIMIT-ss; - - synth_1to1_486(hybridOut[0][ss],0,pcm_sample+pcm_point,n); - synth_1to1_486(hybridOut[1][ss],1,pcm_sample+pcm_point,n); + + if (fr->synth == synth_1to1) { + synth_1to1_486(hybridOut[0][ss],0,pcm_sample+pcm_point,n); + synth_1to1_486(hybridOut[1][ss],1,pcm_sample+pcm_point,n); + } else { + synth_2to1_486(hybridOut[0][ss],0,pcm_sample+pcm_point,n); + synth_2to1_486(hybridOut[1][ss],1,pcm_sample+pcm_point,n); + } ss+=n; - pcm_point+=(2*2*32)*n; + pcm_point+=block_size*n; if(pcm_point >= audiobufsize) audio_flush(outmode,ai); diff -U 3 -H -d -r -N -- mpg123-0.59r-15/mpg123.h mpg123-0.59r/mpg123.h --- mpg123-0.59r-15/mpg123.h 2004-09-20 22:21:18.000000000 +0200 +++ mpg123-0.59r/mpg123.h 2004-09-22 20:58:34.000000000 +0200 @@ -384,6 +384,6 @@ #define FIR_BUFFER_SIZE 128 extern void dct64_486(int *a,int *b,real *c); extern int synth_1to1_486(real *bandPtr,int channel,unsigned char *out,int nb_blocks); - +extern int synth_2to1_486(real *bandPtr,int channel,unsigned char *out,int nb_blocks); diff -U 3 -H -d -r -N -- mpg123-0.59r-15/README mpg123-0.59r/README --- mpg123-0.59r-15/README 2004-09-20 22:21:18.000000000 +0200 +++ mpg123-0.59r/README 2004-09-20 22:51:16.000000000 +0200 @@ -184,6 +184,7 @@ Copyrights (c) 1997 of several enhancements by Oliver Fromme. See the packages 'maplay 1.2(+)' and 'mpegaudio' for their copy-policies. Copyrights (c) 1997 of the OS2 specific parts by Niclas Lindstrom +Copyrights (c) 1998 of the i486 optimization code by Fabrice Bellard. known bugs: ----------- diff -U 3 -H -d -r -N -- mpg123-0.59r-15/README.i486 mpg123-0.59r/README.i486 --- mpg123-0.59r-15/README.i486 1970-01-01 01:00:00.000000000 +0100 +++ mpg123-0.59r/README.i486 2004-09-20 22:48:31.000000000 +0200 @@ -0,0 +1,42 @@ + +i486 optimization : +------------------ + +You can compile the code by 'make linux-i486' to enable the 486 +specific optimizations. DO NOT USE IT ON A PENTIUM, IT WILL BE +SLOWER THAN THE NORMAL CODE. + +The i486 optimizations are *only* enabled when playing in layer 3 at +44.1 kHz or 22.1 kHz in 16 bits. The other modes fallback to the +default pentium optimized routines. + +In other words, if you use a 8 bit sound card, the i486 optimizations +won't be used ! + +Try to increase the buffer size with the option '-b' if your CPU is +not fast enough. + +'Benchmarks': (+/- 2%) +---------------------- +MPEG 1.0 layer III, 112 kbit/s, 44100 Hz joint-stereo + +Processor 22kHz/stereo 44kHz/stereo + +486DX2/66 OK - (almost!) +486DX4/100 OK OK + +Please send me some feedback to tell if it works on your i486 ! + +Fabrice Bellard. +-- +bellard@email.enst.fr - http://www-stud.enst.fr/~bellard + +----------------------------------------------------- +merged back i486 optimized 2to1 downsampling +MPEG 1.0 Layer III, 128 kbit/s, 44100 Hz joint-stereo + +486DX4/75 OK - (90 %) + +160 kbit/s and up work as well. +Christoph Jungegger (disdos%traum404,de) 20.09.2004 +-----------------------------------------------------