On Fri, Dec 1, 2017 at 9:35 AM, Andreas Müller <schnitzelt...@gmail.com> wrote: > On Fri, Dec 1, 2017 at 3:49 PM, Khem Raj <raj.k...@gmail.com> wrote: >> >> On Fri, Dec 1, 2017 at 3:08 AM, Andreas Müller <schnitzelt...@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 <schnitzelt...@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.m...@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?= <schnitzelt...@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 <schnitzelt...@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
interesting. does it mean that vectorization sucks in gcc form arm ? and works ok on other architectures -- _______________________________________________ Openembedded-devel mailing list Openembedded-devel@lists.openembedded.org http://lists.openembedded.org/mailman/listinfo/openembedded-devel
