Mercurial > libavcodec.hg
comparison vorbis_dec.c @ 7962:c0745c5b1bb4 libavcodec
Interleaved forward/backward channels processing in order to increase chances
of stepping on already cached data for the cores with extremely small data
cache.
Patch by Siarhei Siamashka: name dot surname at gmail
author | benoit |
---|---|
date | Wed, 01 Oct 2008 07:28:34 +0000 |
parents | a34931a8fea9 |
children | fb9a51e5d116 |
comparison
equal
deleted
inserted
replaced
7961:ef0616549c16 | 7962:c0745c5b1bb4 |
---|---|
1424 | 1424 |
1425 uint_fast8_t previous_window=vc->previous_window; | 1425 uint_fast8_t previous_window=vc->previous_window; |
1426 uint_fast8_t mode_number; | 1426 uint_fast8_t mode_number; |
1427 uint_fast8_t blockflag; | 1427 uint_fast8_t blockflag; |
1428 uint_fast16_t blocksize; | 1428 uint_fast16_t blocksize; |
1429 int_fast32_t i,j,dir; | 1429 int_fast32_t i,j; |
1430 uint_fast8_t no_residue[vc->audio_channels]; | 1430 uint_fast8_t no_residue[vc->audio_channels]; |
1431 uint_fast8_t do_not_decode[vc->audio_channels]; | 1431 uint_fast8_t do_not_decode[vc->audio_channels]; |
1432 vorbis_mapping *mapping; | 1432 vorbis_mapping *mapping; |
1433 float *ch_res_ptr=vc->channel_residues; | 1433 float *ch_res_ptr=vc->channel_residues; |
1434 float *ch_floor_ptr=vc->channel_floors; | 1434 float *ch_floor_ptr=vc->channel_floors; |
1516 mag=vc->channel_residues+res_chan[mapping->magnitude[i]]*blocksize/2; | 1516 mag=vc->channel_residues+res_chan[mapping->magnitude[i]]*blocksize/2; |
1517 ang=vc->channel_residues+res_chan[mapping->angle[i]]*blocksize/2; | 1517 ang=vc->channel_residues+res_chan[mapping->angle[i]]*blocksize/2; |
1518 vc->dsp.vorbis_inverse_coupling(mag, ang, blocksize/2); | 1518 vc->dsp.vorbis_inverse_coupling(mag, ang, blocksize/2); |
1519 } | 1519 } |
1520 | 1520 |
1521 // Dotproduct | 1521 // Dotproduct, MDCT |
1522 | 1522 |
1523 for(j=0, ch_floor_ptr=vc->channel_floors;j<vc->audio_channels;++j,ch_floor_ptr+=blocksize/2) { | 1523 for(j=vc->audio_channels-1;j>=0;j--) { |
1524 ch_floor_ptr=vc->channel_floors+j*blocksize/2; | |
1524 ch_res_ptr=vc->channel_residues+res_chan[j]*blocksize/2; | 1525 ch_res_ptr=vc->channel_residues+res_chan[j]*blocksize/2; |
1525 vc->dsp.vector_fmul(ch_floor_ptr, ch_res_ptr, blocksize/2); | 1526 vc->dsp.vector_fmul(ch_floor_ptr, ch_res_ptr, blocksize/2); |
1526 } | 1527 ff_imdct_half(&vc->mdct[blockflag], ch_res_ptr, ch_floor_ptr); |
1527 | 1528 } |
1528 // MDCT, overlap/add, save data for next overlapping FPMATH | 1529 |
1530 // Overlap/add, save data for next overlapping FPMATH | |
1529 | 1531 |
1530 retlen = (blocksize + vc->blocksize[previous_window])/4; | 1532 retlen = (blocksize + vc->blocksize[previous_window])/4; |
1531 dir = retlen <= blocksize/2; // pick an order so that ret[] can reuse floors[] without stepping on any data we need | 1533 for(j=0;j<vc->audio_channels;j++) { |
1532 for(j=dir?0:vc->audio_channels-1; (unsigned)j<vc->audio_channels; j+=dir*2-1) { | |
1533 uint_fast16_t bs0=vc->blocksize[0]; | 1534 uint_fast16_t bs0=vc->blocksize[0]; |
1534 uint_fast16_t bs1=vc->blocksize[1]; | 1535 uint_fast16_t bs1=vc->blocksize[1]; |
1535 float *residue=vc->channel_residues+res_chan[j]*blocksize/2; | 1536 float *residue=vc->channel_residues+res_chan[j]*blocksize/2; |
1536 float *floor=vc->channel_floors+j*blocksize/2; | |
1537 float *saved=vc->saved+j*bs1/4; | 1537 float *saved=vc->saved+j*bs1/4; |
1538 float *ret=vc->channel_floors+j*retlen; | 1538 float *ret=vc->channel_floors+j*retlen; |
1539 float *buf=residue; | 1539 float *buf=residue; |
1540 const float *win=vc->win[blockflag&previous_window]; | 1540 const float *win=vc->win[blockflag&previous_window]; |
1541 | |
1542 ff_imdct_half(&vc->mdct[blockflag], buf, floor); | |
1543 | 1541 |
1544 if(blockflag == previous_window) { | 1542 if(blockflag == previous_window) { |
1545 vc->dsp.vector_fmul_window(ret, saved, buf, win, fadd_bias, blocksize/4); | 1543 vc->dsp.vector_fmul_window(ret, saved, buf, win, fadd_bias, blocksize/4); |
1546 } else if(blockflag > previous_window) { | 1544 } else if(blockflag > previous_window) { |
1547 vc->dsp.vector_fmul_window(ret, saved, buf, win, fadd_bias, bs0/4); | 1545 vc->dsp.vector_fmul_window(ret, saved, buf, win, fadd_bias, bs0/4); |