zdoom/fluidsynth
0
0
Fork 0
mirror of https://github.com/ZDoom/fluidsynth.git synced 2024-12-12 13:42:38 +00:00
Code Issues Projects Releases Packages Wiki Activity

Merge branch '2.1.x' into master

Browse source
This commit is contained in:
derselbst 2021-03-04 20:19:09 +01:00
commit 73182866e0
1 changed files with 32 additions and 17 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

49
src/rvoice/fluid_rvoice_mixer.c
View file

@ -451,27 +451,42 @@ fluid_rvoice_buffers_mix(fluid_rvoice_buffers_t *buffers,
* But it seems like having two separate loops is easier for compilers to understand, and therefore * But it seems like having two separate loops is easier for compilers to understand, and therefore
* auto-vectorizing the loops. * auto-vectorizing the loops.
*/ */
#pragma omp simd aligned(dsp_buf,buf:FLUID_DEFAULT_ALIGNMENT) if(sample_count < FLUID_BUFSIZE)
for(dsp_i = 0; dsp_i < FLUID_BUFSIZE; dsp_i++)
{ {
// We cannot simply increment current_amp by amp_incr during every iteration, as this would create a dependency and prevent vectorization. // scalar loop variant, the voice will have finished afterwards
buf[start_block * FLUID_BUFSIZE + dsp_i] += (current_amp + amp_incr * dsp_i) * dsp_buf[start_block * FLUID_BUFSIZE + dsp_i]; for(dsp_i = 0; dsp_i < sample_count; 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[start_block * FLUID_BUFSIZE + dsp_i] += current_amp * dsp_buf[start_block * FLUID_BUFSIZE + dsp_i];
// buf and dsp_buf at the FLUID_BUFSIZE*sizeof(fluid_real_t) byte boundary and never somewhere current_amp += amp_incr;
// 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];
} }
} }
else
{
// here goes the vectorizable loop
#pragma omp simd aligned(dsp_buf,buf:FLUID_DEFAULT_ALIGNMENT)
for(dsp_i = 0; dsp_i < FLUID_BUFSIZE; dsp_i++)
{
// 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];
}
// we have reached the target_amp
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; buffers->bufs[i].current_amp = target_amp;
} }
} }