From 947f79f97a5fa6547d99bff282606026632e010b Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Andreas=20M=C3=BCller?= <schnitzeltony@gmail.com>
Date: Sat, 13 Oct 2018 23:01:11 +0200
Subject: [PATCH] Use ARM-NEON accelaration for float-multithreaded setups

Profiling shows a considerable performance win. See estimated number of voices:
They increase from ~471 to ~513 which makes a ~9% win:

*******************************************************************************
WITHOUT ARM NEON:
*******************************************************************************

morona@raspberrypi3:~$ fluidsynth -o synth.cpu-cores=4 -o synth.chorus.active=0 -o synth.reverb.active=0 /usr/share/sf2/fluidr3gm.sf2
> prof_set_print 1
> prof_set_notes 10
> prof_start 3 10000
Generating 10 notes, generated voices:20
Number of measures(n_prof):3, duration of one mesure(dur):10000ms

Profiling time(mm:ss): Total=0:30  Remainder=0:30, press <ENTER> to cancel
 ------------------------------------------------------------------------------
 Duration(microsecond) and cpu loads(%) (sr: 48000 Hz, sp: 20.83 microsecond)
 ------------------------------------------------------------------------------
 Code under profiling       |Voices|       Duration (microsecond)   |  Load(%)
                            |   nbr|       min|       avg|       max|
 ---------------------------|------|--------------------------------|----------
 synth_write_* ------------>|    20|    112.00|    118.96|    206.00|   4.461
 synth_one_block ---------->|    20|    109.00|    116.44|    196.00|   4.367
 synth_one_block:clear ---->|    20|      1.00|      1.67|     18.00|   0.063
 synth_one_block:one voice->|     1|     11.00|     12.36|     58.00|   0.463
 synth_one_block:all voices>|    20|    107.00|    113.47|    187.00|   4.255
 synth_one_block:reverb --->| no profiling available
 synth_one_block:chorus --->| no profiling available
 voice:note --------------->| no profiling available
 voice:release ------------>| no profiling available
 ------------------------------------------------------------------------------
 Cpu loads(%) (sr: 48000 Hz, sp: 20.83 microsecond) and maximum voices
 ------------------------------------------------------------------------------
 nVoices| total(%)|voices(%)| reverb(%)|chorus(%)| voice(%)|estimated maxVoices
 -------|---------|---------|----------|---------|---------|-------------------
      20|    4.461|    4.461|     0.000|    0.000|    0.213|              470

Profiling time(mm:ss): Total=0:30  Remainder=0:20, press <ENTER> to cancel
 ------------------------------------------------------------------------------
 Duration(microsecond) and cpu loads(%) (sr: 48000 Hz, sp: 20.83 microsecond)
 ------------------------------------------------------------------------------
 Code under profiling       |Voices|       Duration (microsecond)   |  Load(%)
                            |   nbr|       min|       avg|       max|
 ---------------------------|------|--------------------------------|----------
 synth_write_* ------------>|    20|    112.00|    118.42|    216.00|   4.441
 synth_one_block ---------->|    20|    109.00|    115.91|    205.00|   4.347
 synth_one_block:clear ---->|    20|      1.00|      1.65|     18.00|   0.062
 synth_one_block:one voice->|     1|     11.00|     12.30|     58.00|   0.461
 synth_one_block:all voices>|    20|    107.00|    112.98|    197.00|   4.237
 synth_one_block:reverb --->| no profiling available
 synth_one_block:chorus --->| no profiling available
 voice:note --------------->| no profiling available
 voice:release ------------>| no profiling available
 ------------------------------------------------------------------------------
 Cpu loads(%) (sr: 48000 Hz, sp: 20.83 microsecond) and maximum voices
 ------------------------------------------------------------------------------
 nVoices| total(%)|voices(%)| reverb(%)|chorus(%)| voice(%)|estimated maxVoices
 -------|---------|---------|----------|---------|---------|-------------------
      20|    4.441|    4.441|     0.000|    0.000|    0.212|              472

Profiling time(mm:ss): Total=0:30  Remainder=0:10, press <ENTER> to cancel
 ------------------------------------------------------------------------------
 Duration(microsecond) and cpu loads(%) (sr: 48000 Hz, sp: 20.83 microsecond)
 ------------------------------------------------------------------------------
 Code under profiling       |Voices|       Duration (microsecond)   |  Load(%)
                            |   nbr|       min|       avg|       max|
 ---------------------------|------|--------------------------------|----------
 synth_write_* ------------>|    20|    112.00|    118.64|    244.00|   4.449
 synth_one_block ---------->|    20|    109.00|    116.12|    234.00|   4.355
 synth_one_block:clear ---->|    20|      1.00|      1.67|     37.00|   0.062
 synth_one_block:one voice->|     1|     11.00|     12.31|     63.00|   0.462
 synth_one_block:all voices>|    20|    107.00|    113.18|    214.00|   4.244
 synth_one_block:reverb --->| no profiling available
 synth_one_block:chorus --->| no profiling available
 voice:note --------------->| no profiling available
 voice:release ------------>| no profiling available
 ------------------------------------------------------------------------------
 Cpu loads(%) (sr: 48000 Hz, sp: 20.83 microsecond) and maximum voices
 ------------------------------------------------------------------------------
 nVoices| total(%)|voices(%)| reverb(%)|chorus(%)| voice(%)|estimated maxVoices
 -------|---------|---------|----------|---------|---------|-------------------
      20|    4.449|    4.449|     0.000|    0.000|    0.212|              471
Stopping 20 voices...voices stopped.
> quit
cheers!
JackTemporaryException : now quits...
Jack main caught signal 2
fluid_profiling_print
fluidsynth: Estimated times: min/avg/max (micro seconds)
fluidsynth: synth_write_* ------------>: 112.000/118.636/244.000
fluidsynth: synth_one_block ---------->: 109.000/116.124/234.000
fluidsynth: synth_one_block:clear ---->: 1.000/1.665/37.000
fluidsynth: synth_one_block:one voice->: 11.000/12.309/63.000
fluidsynth: synth_one_block:all voices>: 107.000/113.180/214.000

*******************************************************************************
WITH ARM NEON:
*******************************************************************************

morona@raspberrypi3:~$ fluidsynth -o synth.cpu-cores=4 -o synth.chorus.active=0 -o synth.reverb.active=0 /usr/share/sf2/fluidr3gm.sf2
> prof_set_print 1
> prof_set_notes 10
> prof_start 3 10000
Generating 10 notes, generated voices:20
Number of measures(n_prof):3, duration of one mesure(dur):10000ms

Profiling time(mm:ss): Total=0:30  Remainder=0:30, press <ENTER> to cancel
 ------------------------------------------------------------------------------
 Duration(microsecond) and cpu loads(%) (sr: 48000 Hz, sp: 20.83 microsecond)
 ------------------------------------------------------------------------------
 Code under profiling       |Voices|       Duration (microsecond)   |  Load(%)
                            |   nbr|       min|       avg|       max|
 ---------------------------|------|--------------------------------|----------
 synth_write_* ------------>|    20|    102.00|    109.21|    213.00|   4.095
 synth_one_block ---------->|    20|     99.00|    106.68|    201.00|   4.001
 synth_one_block:clear ---->|    20|      1.00|      1.64|     18.00|   0.062
 synth_one_block:one voice->|     1|     11.00|     12.30|     54.00|   0.461
 synth_one_block:all voices>|    20|     97.00|    103.71|    188.00|   3.889
 synth_one_block:reverb --->| no profiling available
 synth_one_block:chorus --->| no profiling available
 voice:note --------------->| no profiling available
 voice:release ------------>| no profiling available
 ------------------------------------------------------------------------------
 Cpu loads(%) (sr: 48000 Hz, sp: 20.83 microsecond) and maximum voices
 ------------------------------------------------------------------------------
 nVoices| total(%)|voices(%)| reverb(%)|chorus(%)| voice(%)|estimated maxVoices
 -------|---------|---------|----------|---------|---------|-------------------
      20|    4.095|    4.095|     0.000|    0.000|    0.194|              514

Profiling time(mm:ss): Total=0:30  Remainder=0:20, press <ENTER> to cancel
 ------------------------------------------------------------------------------
 Duration(microsecond) and cpu loads(%) (sr: 48000 Hz, sp: 20.83 microsecond)
 ------------------------------------------------------------------------------
 Code under profiling       |Voices|       Duration (microsecond)   |  Load(%)
                            |   nbr|       min|       avg|       max|
 ---------------------------|------|--------------------------------|----------
 synth_write_* ------------>|    20|    102.00|    109.46|    278.00|   4.105
 synth_one_block ---------->|    20|     99.00|    106.91|    265.00|   4.009
 synth_one_block:clear ---->|    20|      1.00|      1.67|     22.00|   0.062
 synth_one_block:one voice->|     1|     11.00|     12.30|     54.00|   0.461
 synth_one_block:all voices>|    20|     97.00|    103.94|    251.00|   3.898
 synth_one_block:reverb --->| no profiling available
 synth_one_block:chorus --->| no profiling available
 voice:note --------------->| no profiling available
 voice:release ------------>| no profiling available
 ------------------------------------------------------------------------------
 Cpu loads(%) (sr: 48000 Hz, sp: 20.83 microsecond) and maximum voices
 ------------------------------------------------------------------------------
 nVoices| total(%)|voices(%)| reverb(%)|chorus(%)| voice(%)|estimated maxVoices
 -------|---------|---------|----------|---------|---------|-------------------
      20|    4.105|    4.105|     0.000|    0.000|    0.195|              513

Profiling time(mm:ss): Total=0:30  Remainder=0:10, press <ENTER> to cancel
 ------------------------------------------------------------------------------
 Duration(microsecond) and cpu loads(%) (sr: 48000 Hz, sp: 20.83 microsecond)
 ------------------------------------------------------------------------------
 Code under profiling       |Voices|       Duration (microsecond)   |  Load(%)
                            |   nbr|       min|       avg|       max|
 ---------------------------|------|--------------------------------|----------
 synth_write_* ------------>|    20|    102.00|    109.22|    278.00|   4.096
 synth_one_block ---------->|    20|     99.00|    106.65|    265.00|   3.999
 synth_one_block:clear ---->|    20|      1.00|      1.67|     22.00|   0.062
 synth_one_block:one voice->|     1|     11.00|     12.31|     57.00|   0.462
 synth_one_block:all voices>|    20|     97.00|    103.68|    251.00|   3.888
 synth_one_block:reverb --->| no profiling available
 synth_one_block:chorus --->| no profiling available
 voice:note --------------->| no profiling available
 voice:release ------------>| no profiling available
 ------------------------------------------------------------------------------
 Cpu loads(%) (sr: 48000 Hz, sp: 20.83 microsecond) and maximum voices
 ------------------------------------------------------------------------------
 nVoices| total(%)|voices(%)| reverb(%)|chorus(%)| voice(%)|estimated maxVoices
 -------|---------|---------|----------|---------|---------|-------------------
      20|    4.096|    4.096|     0.000|    0.000|    0.194|              514
Stopping 20 voices...voices stopped.
> quit
cheers!
JackTemporaryException : now quits...
Jack main caught signal 2
fluid_profiling_print
fluidsynth: Estimated times: min/avg/max (micro seconds)
fluidsynth: synth_write_* ------------>: 102.000/109.216/278.000
fluidsynth: synth_one_block ---------->: 99.000/106.649/265.000
fluidsynth: synth_one_block:clear ---->: 1.000/1.666/22.000
fluidsynth: synth_one_block:one voice->: 11.000/12.307/57.000
fluidsynth: synth_one_block:all voices>: 97.000/103.681/251.000

Upstream-Status: Inappropriate [embedded-specific]

---
 src/rvoice/fluid_rvoice_mixer.c | 55 ++++++++++++++++++++++++++++++++-
 1 file changed, 54 insertions(+), 1 deletion(-)

diff --git a/src/rvoice/fluid_rvoice_mixer.c b/src/rvoice/fluid_rvoice_mixer.c
index af0ef75d..07a357c7 100644
--- a/src/rvoice/fluid_rvoice_mixer.c
+++ b/src/rvoice/fluid_rvoice_mixer.c
@@ -27,6 +27,9 @@
 #include "fluid_ladspa.h"
 #include "fluid_synth.h"
 
+#if defined(__ARM_NEON__)
+#include "arm_neon.h"
+#endif
 
 // If less than x voices, the thread overhead is larger than the gain,
 // so don't activate the thread(s).
@@ -1053,9 +1056,15 @@ fluid_mixer_buffers_mix(fluid_mixer_buffers_t *dst, fluid_mixer_buffers_t *src,
     int i, j;
     int scount = current_blockcount * FLUID_BUFSIZE;
     int minbuf;
+#if defined(__ARM_NEON__) && defined(WITH_FLOAT)
+    fluid_real_t *FLUID_RESTRICT base_src_left;
+    fluid_real_t *FLUID_RESTRICT base_src_right;
+    fluid_real_t *FLUID_RESTRICT base_dst_left;
+    fluid_real_t *FLUID_RESTRICT base_dst_right;
+#else
     fluid_real_t *FLUID_RESTRICT base_src;
     fluid_real_t *FLUID_RESTRICT base_dst;
-
+#endif
     minbuf = dst->buf_count;
 
     if(minbuf > src->buf_count)
@@ -1063,6 +1072,27 @@ fluid_mixer_buffers_mix(fluid_mixer_buffers_t *dst, fluid_mixer_buffers_t *src,
         minbuf = src->buf_count;
     }
 
+#if defined(__ARM_NEON__) && defined(WITH_FLOAT)
+    base_src_left = fluid_align_ptr(src->left_buf, FLUID_DEFAULT_ALIGNMENT);
+    base_dst_left = fluid_align_ptr(dst->left_buf, FLUID_DEFAULT_ALIGNMENT);
+    base_src_right = fluid_align_ptr(src->right_buf, FLUID_DEFAULT_ALIGNMENT);
+    base_dst_right = fluid_align_ptr(dst->right_buf, FLUID_DEFAULT_ALIGNMENT);
+
+    for(i = 0; i < minbuf; i++)
+    {
+        for(j = 0; j < scount; j+=4)
+        {
+            int dsp_i = i * FLUID_MIXER_MAX_BUFFERS_DEFAULT * FLUID_BUFSIZE + j;
+
+            float32x4_t vleft = vld1q_f32(&base_dst_left[dsp_i]);
+            float32x4_t vright = vld1q_f32(&base_dst_right[dsp_i]);
+            vleft = vaddq_f32(vleft, vld1q_f32(&base_src_left[dsp_i]));
+            vright = vaddq_f32(vright, vld1q_f32(&base_src_right[dsp_i]));
+            vst1q_f32(&base_dst_left[dsp_i], vleft);
+            vst1q_f32(&base_dst_right[dsp_i], vright);
+        }
+    }
+#else
     base_src = fluid_align_ptr(src->left_buf, FLUID_DEFAULT_ALIGNMENT);
     base_dst = fluid_align_ptr(dst->left_buf, FLUID_DEFAULT_ALIGNMENT);
 
@@ -1090,6 +1120,7 @@ fluid_mixer_buffers_mix(fluid_mixer_buffers_t *dst, fluid_mixer_buffers_t *src,
             base_dst[dsp_i] += base_src[dsp_i];
         }
     }
+#endif
 
     minbuf = dst->fx_buf_count;
 
@@ -1098,6 +1129,27 @@ fluid_mixer_buffers_mix(fluid_mixer_buffers_t *dst, fluid_mixer_buffers_t *src,
         minbuf = src->fx_buf_count;
     }
 
+#if defined(__ARM_NEON__) && defined(WITH_FLOAT)
+    base_src_left = fluid_align_ptr(src->fx_left_buf, FLUID_DEFAULT_ALIGNMENT);
+    base_dst_left = fluid_align_ptr(dst->fx_left_buf, FLUID_DEFAULT_ALIGNMENT);
+    base_src_right = fluid_align_ptr(src->fx_right_buf, FLUID_DEFAULT_ALIGNMENT);
+    base_dst_right = fluid_align_ptr(dst->fx_right_buf, FLUID_DEFAULT_ALIGNMENT);
+
+    for(i = 0; i < minbuf; i++)
+    {
+        for(j = 0; j < scount; j+=4)
+        {
+            int dsp_i = i * FLUID_MIXER_MAX_BUFFERS_DEFAULT * FLUID_BUFSIZE + j;
+
+            float32x4_t vleft = vld1q_f32(&base_dst_left[dsp_i]);
+            float32x4_t vright = vld1q_f32(&base_dst_right[dsp_i]);
+            vleft = vaddq_f32(vleft, vld1q_f32(&base_src_left[dsp_i]));
+            vright = vaddq_f32(vright, vld1q_f32(&base_src_right[dsp_i]));
+            vst1q_f32(&base_dst_left[dsp_i], vleft);
+            vst1q_f32(&base_dst_right[dsp_i], vright);
+        }
+    }
+#else
     base_src = fluid_align_ptr(src->fx_left_buf, FLUID_DEFAULT_ALIGNMENT);
     base_dst = fluid_align_ptr(dst->fx_left_buf, FLUID_DEFAULT_ALIGNMENT);
 
@@ -1125,6 +1177,7 @@ fluid_mixer_buffers_mix(fluid_mixer_buffers_t *dst, fluid_mixer_buffers_t *src,
             base_dst[dsp_i] += base_src[dsp_i];
         }
     }
+#endif
 }
 
 
-- 
2.20.1