[oe] [PATCH 04/10] fluidsynth: performance improvements

[Andreas Müller]

On Fri, Dec 1, 2017 at 3:49 PM, Khem Raj <raj.khem at gmail.com> wrote:

> On Fri, Dec 1, 2017 at 3:08 AM, Andreas Müller <schnitzeltony at gmail.com>
> wrote:
> > * Use floats instead of double for sound calculations. This improves
> >   performance notable and was default for version 1.1.6 using autotools.
> > * Fix buffer overrun when using floats
> > * Make use of ARM NEON for multithreading enabled
> >
> > Performance and sound correctnes was tested with qtractor and a reworked
> > version of fluidsynth-dssi [1-2]. Tests were performed for single- and
> > multithreading enabled.
> >
> > [1] https://github.com/schnitzeltony/fluidsynth-dssi/commit/
> bad09c6f5c5508c5f5330aa5188510f975e50c50
> > [2] https://github.com/schnitzeltony/meta-qt5-extra/
> blob/master/recipes-misc/recipes-multimedia/fluidsynth/
> fluidsynth-dssi_1.0.0.bb
> >
> > Signed-off-by: Andreas Müller <schnitzeltony at gmail.com>
> > ---
> >  ...uffer-overrun-in-fluid_synth_nwrite_float.patch | 32 +++++++++
> >  ...N-accelaration-for-float-multithreaded-se.patch | 76
> ++++++++++++++++++++++
> >  .../fluidsynth/fluidsynth_1.1.8.bb                 |  8 ++-
> >  3 files changed, 114 insertions(+), 2 deletions(-)
> >  create mode 100644 meta-multimedia/recipes-multimedia/fluidsynth/files/
> 0001-avoid-buffer-overrun-in-fluid_synth_nwrite_float.patch
> >  create mode 100644 meta-multimedia/recipes-multimedia/fluidsynth/files/
> 0002-Use-ARM-NEON-accelaration-for-float-multithreaded-se.patch
> >
> > diff --git a/meta-multimedia/recipes-multimedia/fluidsynth/files/
> 0001-avoid-buffer-overrun-in-fluid_synth_nwrite_float.patch
> b/meta-multimedia/recipes-multimedia/fluidsynth/files/
> 0001-avoid-buffer-overrun-in-fluid_synth_nwrite_float.patch
> > new file mode 100644
> > index 0000000..dda76cf
> > --- /dev/null
> > +++ b/meta-multimedia/recipes-multimedia/fluidsynth/files/
> 0001-avoid-buffer-overrun-in-fluid_synth_nwrite_float.patch
> > @@ -0,0 +1,32 @@
> > +From a13cb63103aa56b5e8bad816c7d13d6e01c0cd9f Mon Sep 17 00:00:00 2001
> > +From: derselbst <tom.mbrt at googlemail.com>
> > +Date: Sun, 26 Nov 2017 22:12:12 +0100
> > +Subject: [PATCH 1/2] avoid buffer overrun in fluid_synth_nwrite_float()
> > +
> > +Upstream-Status: Backport [1]
> > +
> > +[1] https://github.com/FluidSynth/fluidsynth/commit/
> a13cb63103aa56b5e8bad816c7d13d6e01c0cd9f
> > +---
> > + src/synth/fluid_synth.c | 4 ++--
> > + 1 file changed, 2 insertions(+), 2 deletions(-)
> > +
> > +diff --git a/src/synth/fluid_synth.c b/src/synth/fluid_synth.c
> > +index 266d759..14f6b21 100644
> > +--- a/src/synth/fluid_synth.c
> > ++++ b/src/synth/fluid_synth.c
> > +@@ -2752,10 +2752,10 @@ fluid_synth_nwrite_float(fluid_synth_t* synth,
> int len,
> > +     {
> > + #ifdef WITH_FLOAT
> > +       if(fx_left != NULL)
> > +-        FLUID_MEMCPY(fx_left[i + count], fx_left_in[i], bytes);
> > ++        FLUID_MEMCPY(fx_left[i] + count, fx_left_in[i], bytes);
> > +
> > +       if(fx_right != NULL)
> > +-        FLUID_MEMCPY(fx_right[i + count], fx_right_in[i], bytes);
> > ++        FLUID_MEMCPY(fx_right[i] + count, fx_right_in[i], bytes);
> > + #else //WITH_FLOAT
> > +       int j;
> > +       if(fx_left != NULL) {
> > +--
> > +2.9.5
> > +
> > diff --git a/meta-multimedia/recipes-multimedia/fluidsynth/files/
> 0002-Use-ARM-NEON-accelaration-for-float-multithreaded-se.patch
> b/meta-multimedia/recipes-multimedia/fluidsynth/files/0002-Use-ARM-NEON-
> accelaration-for-float-multithreaded-se.patch
> > new file mode 100644
> > index 0000000..0e1846e
> > --- /dev/null
> > +++ b/meta-multimedia/recipes-multimedia/fluidsynth/files/
> 0002-Use-ARM-NEON-accelaration-for-float-multithreaded-se.patch
> > @@ -0,0 +1,76 @@
> > +From 2de7e128fbdf528716b500cf27ed9a4358c931c9 Mon Sep 17 00:00:00 2001
> > +From: =?UTF-8?q?Andreas=20M=C3=BCller?= <schnitzeltony at gmail.com>
> > +Date: Fri, 24 Nov 2017 00:05:35 +0100
> > +Subject: [PATCH 2/2] Use ARM-NEON accelaration for float-multithreaded
> setups
> > +MIME-Version: 1.0
> > +Content-Type: text/plain; charset=UTF-8
> > +Content-Transfer-Encoding: 8bit
> > +
> > +Upstream-Status: Pending
> > +
> > +Signed-off-by: Andreas Müller <schnitzeltony at gmail.com>
> > +---
> > + src/rvoice/fluid_rvoice_mixer.c | 26 ++++++++++++++++++++++++++
> > + 1 file changed, 26 insertions(+)
> > +
> > +diff --git a/src/rvoice/fluid_rvoice_mixer.c b/src/rvoice/fluid_rvoice_
> mixer.c
> > +index 9616518..dbf8057 100644
> > +--- a/src/rvoice/fluid_rvoice_mixer.c
> > ++++ b/src/rvoice/fluid_rvoice_mixer.c
> > +@@ -27,6 +27,10 @@
> > + #include "fluid_ladspa.h"
> > + #include "fluid_synth.h"
> > +
> > ++#if defined(__ARM_NEON__)
> > ++#include "arm_neon.h"
> > ++#endif
> > ++
> > +
> > + #define ENABLE_MIXER_THREADS 1
> > +
> > +@@ -794,20 +798,42 @@ fluid_mixer_buffers_mix(fluid_mixer_buffers_t*
> dest, fluid_mixer_buffers_t* src)
> > +   if (minbuf > src->buf_count)
> > +     minbuf = src->buf_count;
> > +   for (i=0; i < minbuf; i++) {
> > ++#if defined(__ARM_NEON__) && defined(WITH_FLOAT)
> > ++    for (j=0; j < scount; j+=4) {
> > ++        float32x4_t vleft = vld1q_f32(&dest->left_buf[i][j]);
> > ++        float32x4_t vright = vld1q_f32(&dest->right_buf[i][j]);
> > ++        vleft = vaddq_f32(vleft, vld1q_f32(&src->left_buf[i][j]));
> > ++        vright = vaddq_f32(vright, vld1q_f32(&src->right_buf[i][j]));
> > ++        vst1q_f32(&dest->left_buf[i][j], vleft);
> > ++        vst1q_f32(&dest->right_buf[i][j], vright);
> > ++    }
>
> > I wonder if gcc could be initiated to see if it can vectorize the code,
> here
> > have you tried setting cmdline options or may be pragma hints
>
>
I've sent patches like these to some projects. If I remember correctly
first one was jack2 [1]. For that I wrote some performance tests [2] and
tested vectorizing either but with disappointing results. This is ~ a year
ago and I think intrinsics are still the option performing better.

[1]
https://github.com/jackaudio/jack2/commit/77bb8be12e0d856fbc004cf57185ab36a2df04c2
[2]
https://github.com/jackaudio/jack2/commit/c32b823860bc8e887774ab10bc3e9bd76e85e3f3

Andreas