zdoom/fluidsynth
0
0
Fork 0
mirror of https://github.com/ZDoom/fluidsynth.git synced 2024-11-10 06:51:54 +00:00
Code Issues Projects Releases Packages Wiki Activity

Simplify control flow to allow auto-vectorization by compiler

Browse source
This commit is contained in:
derselbst 2021-02-06 16:23:03 +01:00 committed by Tom M
parent 25b0503ba7
commit aebc4837dd
1 changed files with 26 additions and 19 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

45
src/rvoice/fluid_rvoice_mixer.c
View file

@ -405,29 +405,36 @@ fluid_rvoice_buffers_mix(fluid_rvoice_buffers_t *buffers,
FLUID_ASSERT((uintptr_t)buf % FLUID_DEFAULT_ALIGNMENT == 0);
/* mixdown sample_count samples in the current buffer buf
Note, that this loop could be unrolled by FLUID_BUFSIZE elements */
/* Mixdown sample_count samples in the current buffer buf
*
* For the first FLUID_BUFSIZE samples, we linearly interpolate the buffers amplitude to
* avoid clicks/pops when rapidly changing the channels panning (issue 768).
*
* We could have squashed this into one single loop by using an if clause within the loop body.
* But it seems like having two separate loops is easier for compilers to understand, and therefore
* auto-vectorizing the loops.
*/
#pragma omp simd aligned(dsp_buf,buf:FLUID_DEFAULT_ALIGNMENT)
for(dsp_i = 0; dsp_i < sample_count; dsp_i++)
for(dsp_i = 0; dsp_i < FLUID_BUFSIZE; dsp_i++)
{
fluid_real_t samp;
if(dsp_i < FLUID_BUFSIZE)
{
samp = (current_amp + amp_incr * dsp_i) * dsp_buf[start_block * FLUID_BUFSIZE + dsp_i];
}
else
{
samp = target_amp * dsp_buf[start_block * FLUID_BUFSIZE + dsp_i];
}
// Index by blocks (not by samples) to let the compiler know that we always start accessing
// buf and dsp_buf at the FLUID_BUFSIZE*sizeof(fluid_real_t) byte boundary and never somewhere
// in between.
// A good compiler should understand: Aha, so I don't need to add a peel loop when vectorizing
// this loop. Great.
buf[start_block * FLUID_BUFSIZE + dsp_i] += samp;
// We cannot simply increment current_amp by amp_incr during every iteration, as this would create a dependency and prevent vectorization.
buf[start_block * FLUID_BUFSIZE + dsp_i] += (current_amp + amp_incr * dsp_i) * dsp_buf[start_block * FLUID_BUFSIZE + dsp_i];
}
if(target_amp > 0)
{
/* Note, that this loop could be unrolled by FLUID_BUFSIZE elements */
#pragma omp simd aligned(dsp_buf,buf:FLUID_DEFAULT_ALIGNMENT)
for(dsp_i = FLUID_BUFSIZE; dsp_i < sample_count; dsp_i++)
{
// Index by blocks (not by samples) to let the compiler know that we always start accessing
// buf and dsp_buf at the FLUID_BUFSIZE*sizeof(fluid_real_t) byte boundary and never somewhere
// in between.
// A good compiler should understand: Aha, so I don't need to add a peel loop when vectorizing
// this loop. Great.
buf[start_block * FLUID_BUFSIZE + dsp_i] += target_amp * dsp_buf[start_block * FLUID_BUFSIZE + dsp_i];
}
}
buffers->bufs[i].current_amp = target_amp;
}
}