-// MDCT, overlap/add, save data for next overlapping FPMATH
-
- for(j=0;j<vc->audio_channels;++j) {
- uint_fast8_t step=vc->audio_channels;
- uint_fast16_t k;
- float *saved=vc->saved+j*vc->blocksize[1]/2;
- float *ret=vc->ret;
- const float *lwin=vc->win[1];
- const float *swin=vc->win[0];
- float *buf=vc->buf;
- float *buf_tmp=vc->buf_tmp;
-
- ch_floor_ptr=vc->channel_floors+j*blocksize/2;
-
- saved_start=vc->saved_start;
-
- vc->mdct[0].fft.imdct_calc(&vc->mdct[vc->modes[mode_number].blockflag], buf, ch_floor_ptr, buf_tmp);
-
- //FIXME process channels together, to allow faster simd vector_fmul_add_add?
- if (vc->modes[mode_number].blockflag) {
- // -- overlap/add
- if (previous_window) {
- vc->dsp.vector_fmul_add_add(ret+j, buf, lwin, saved, vc->add_bias, vc->blocksize[1]/2, step);
- retlen=vc->blocksize[1]/2;
- } else {
- int len = (vc->blocksize[1]-vc->blocksize[0])/4;
- buf += len;
- vc->dsp.vector_fmul_add_add(ret+j, buf, swin, saved, vc->add_bias, vc->blocksize[0]/2, step);
- k = vc->blocksize[0]/2*step + j;
- buf += vc->blocksize[0]/2;
- if(vc->exp_bias){
- for(i=0; i<len; i++, k+=step)
- ((uint32_t*)ret)[k] = ((uint32_t*)buf)[i] + vc->exp_bias; // ret[k]=buf[i]*(1<<bias)
- } else {
- for(i=0; i<len; i++, k+=step)
- ret[k] = buf[i] + fadd_bias;
- }
- buf=vc->buf;
- retlen=vc->blocksize[0]/2+len;
- }
- // -- save
- if (next_window) {
- buf += vc->blocksize[1]/2;
- vc->dsp.vector_fmul_reverse(saved, buf, lwin, vc->blocksize[1]/2);
- saved_start=0;
- } else {
- saved_start=(vc->blocksize[1]-vc->blocksize[0])/4;
- buf += vc->blocksize[1]/2;
- for(i=0; i<saved_start; i++)
- ((uint32_t*)saved)[i] = ((uint32_t*)buf)[i] + vc->exp_bias;
- vc->dsp.vector_fmul_reverse(saved+saved_start, buf+saved_start, swin, vc->blocksize[0]/2);
- }
+// Overlap/add, save data for next overlapping FPMATH
+
+ retlen = (blocksize + vc->blocksize[previous_window])/4;
+ for(j=0;j<vc->audio_channels;j++) {
+ uint_fast16_t bs0=vc->blocksize[0];
+ uint_fast16_t bs1=vc->blocksize[1];
+ float *residue=vc->channel_residues+res_chan[j]*blocksize/2;
+ float *saved=vc->saved+j*bs1/4;
+ float *ret=vc->channel_floors+j*retlen;
+ float *buf=residue;
+ const float *win=vc->win[blockflag&previous_window];
+
+ if(blockflag == previous_window) {
+ vc->dsp.vector_fmul_window(ret, saved, buf, win, fadd_bias, blocksize/4);
+ } else if(blockflag > previous_window) {
+ vc->dsp.vector_fmul_window(ret, saved, buf, win, fadd_bias, bs0/4);
+ copy_normalize(ret+bs0/2, buf+bs0/4, (bs1-bs0)/4, vc->exp_bias, fadd_bias);