lib/mpg123
1
0
Fork 0
mirror of http://mpg123.de/trunk/.git synced 2025-08-07 21:02:55 +03:00
Code Activity

libsyn123: add SYN123_NO_CASES switch for mono and stereo

Browse Source 
Using scripts/syn123-channelscaling.sh, I see about 15% benefit
for mono/stereo operation when treating it as special cases in
the resampler. Fun fact: Even the higher channel counts profit
a tiny bit (1% to 2%, not sure how significant) from the
separation.

Giving the compiler information about what can be likely values
for those inner loop rounds still makes sense.



git-svn-id: svn://scm.orgis.org/mpg123/trunk@4494 35dc7657-300d-0410-a2e5-dc2837fedb53
This commit is contained in:
thor
2019-07-22 09:21:32 +00:00
parent 5bd474dc8b
commit 15a3788e92
2 changed files with 110 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files

32
scripts/syn123-channelscaling.sh Normal file
View File

@@ -0,0 +1,32 @@
#!/bin/sh
# This gives a view at the scaling of resampling runtime with channel count.
# The tested resampling mode involves 2X upsampling, lowpass, and interpolation.
set -e
export LANG=C
export LC_NUMERIC=C
chan_min=1
chan_max=10
chan_fit=3
out123=src/out123
generate="--wave-freq 300 --inputrate 44100 --timelimit 4410000 -q"
wd=$(mktemp -d channelscaling.XXXX)
echo "workdir: $wd"
for n in $(seq $chan_min $chan_max)
do
printf "generate with %d channels\n" "$n" >&2
/usr/bin/time -f "$n\t%e" $out123 $generate -c $n --rate 44100 -t 2>&1
done > $wd/generate.txd
for n in $(seq $chan_min $chan_max)
do
printf "resample with %d channels\n" "$n" >&2
/usr/bin/time -f "$n\t%e" $out123 $generate -c $n --rate 44101 -t 2>&1
done > $wd/resample.txd
txdcalc '[3]=[2]-[1,2]' $wd/generate.txd < $wd/resample.txd > $wd/resampling-overhead.txd
gpfit --plot -g=1 -r='[3:]' $wd/resampling-overhead.txd
echo "Check results in $wd/."

84
src/libsyn123/resample.c
View File

@@ -739,6 +739,7 @@ static size_t decimate(struct decimator_state *rd, struct resample_data *rrd, fl
} }
float *out = in; float *out = in;
size_t outs = 0; size_t outs = 0;
#ifndef SYN123_NO_CASES
switch(rrd->channels) switch(rrd->channels)
{ {
case 1: for(size_t i=0; i<ins; ++i) case 1: for(size_t i=0; i<ins; ++i)
@@ -806,7 +807,9 @@ static size_t decimate(struct decimator_state *rd, struct resample_data *rrd, fl
rd->sflags |= decimate_store; rd->sflags |= decimate_store;
} }
break; break;
default: for(size_t i=0; i<ins; ++i) default:
#endif
for(size_t i=0; i<ins; ++i)
{ {
int ni[LPF_4_ORDER]; int ni[LPF_4_ORDER];
for(int j=0; j<LPF_4_ORDER; ++j) for(int j=0; j<LPF_4_ORDER; ++j)
@@ -840,7 +843,9 @@ static size_t decimate(struct decimator_state *rd, struct resample_data *rrd, fl
} else } else
rd->sflags |= decimate_store; rd->sflags |= decimate_store;
} }
#ifndef SYN123_NO_CASES
} }
#endif
return outs; return outs;
} }
@@ -955,6 +960,7 @@ static float df2_initval(unsigned int order, float *filter_a, float insample)
// The parameter determines the number of repeated applications of the same // The parameter determines the number of repeated applications of the same
// low pass. // low pass.
#ifndef SYN123_NO_CASES
#define LOWPASS_DF2_FUNCSX(times) \ #define LOWPASS_DF2_FUNCSX(times) \
\ \
static void lowpass##times##_df2_preemp_2x(struct resample_data *rd, float *in, size_t ins, float *out) \ static void lowpass##times##_df2_preemp_2x(struct resample_data *rd, float *in, size_t ins, float *out) \
@@ -1046,10 +1052,54 @@ static void lowpass##times##_df2_preemp(struct resample_data *rd, float *in, siz
in += rd->channels; \ in += rd->channels; \
out += rd->channels; \ out += rd->channels; \
} \ } \
break; \
} \ } \
LPF_DF2_END \ LPF_DF2_END \
} }
#else
#define LOWPASS_DF2_FUNCSX(times) \
\
static void lowpass##times##_df2_preemp_2x(struct resample_data *rd, float *in, size_t ins, float *out) \
{ \
if(!ins) \
return; \
LPF_DF2_BEGIN(times,in,rd->channels,) \
PREEMP_DF2_BEGIN(in,rd->channels,); \
for(size_t i=0; i<ins; ++i) \
{ \
PREEMP_DF2_SAMPLE(in, rd->frame, rd->channels) \
/* Zero-stuffing! Insert zero after making up for energy loss. */ \
for(unsigned int c=0; c<rd->channels; ++c) \
rd->frame[c] *= 2; \
LPF_DF2_SAMPLE(times, rd->frame, out, rd->channels) \
out += rd->channels; \
for(unsigned int c=0; c<rd->channels; ++c) \
rd->frame[c] = 0; \
LPF_DF2_SAMPLE(times, rd->frame, out, rd->channels) \
out += rd->channels; \
in += rd->channels; \
} \
LPF_DF2_END \
} \
\
static void lowpass##times##_df2_preemp(struct resample_data *rd, float *in, size_t ins) \
{ \
if(!ins) \
return; \
float *out = in; \
LPF_DF2_BEGIN(times, in, rd->channels,) \
PREEMP_DF2_BEGIN(in, rd->channels,); \
for(size_t i=0; i<ins; ++i) \
{ \
PREEMP_DF2_SAMPLE(in, rd->frame, rd->channels) \
LPF_DF2_SAMPLE(times, rd->frame, out, rd->channels) \
in += rd->channels; \
out += rd->channels; \
} \
LPF_DF2_END \
}
#endif
// Need that indirection so that times is expanded properly for concatenation. // Need that indirection so that times is expanded properly for concatenation.
#define LOWPASS_DF2_FUNCS(times) \ #define LOWPASS_DF2_FUNCS(times) \
LOWPASS_DF2_FUNCSX(times) LOWPASS_DF2_FUNCSX(times)
@@ -1130,6 +1180,7 @@ static size_t resample_opt4p4o(struct resample_data *rd, float*in
if(!ins) if(!ins)
return outs; return outs;
OPT4P4O_BEGIN(in) OPT4P4O_BEGIN(in)
#ifndef SYN123_NO_CASES
switch(rd->channels) switch(rd->channels)
{ {
case 1: for(size_t i=0; i<ins; ++i) case 1: for(size_t i=0; i<ins; ++i)
@@ -1144,12 +1195,16 @@ static size_t resample_opt4p4o(struct resample_data *rd, float*in
in += 2; in += 2;
} }
break; break;
default: for(size_t i=0; i<ins; ++i) default:
#endif
for(size_t i=0; i<ins; ++i)
{ {
OPT4P4O_INTERPOL(in, out, outs, rd->channels) OPT4P4O_INTERPOL(in, out, outs, rd->channels)
in += rd->channels; in += rd->channels;
} }
#ifndef SYN123_NO_CASES
} }
#endif
OPT4P4O_END OPT4P4O_END
return outs; return outs;
} }
@@ -1158,6 +1213,7 @@ static size_t resample_opt4p4o_2batch(struct resample_data *rd, float*in, float
{ {
size_t outs = 0; size_t outs = 0;
OPT4P4O_BEGIN(in) OPT4P4O_BEGIN(in)
#ifndef SYN123_NO_CASES
switch(rd->channels) switch(rd->channels)
{ {
case 1: for(size_t i=0; i<2*BATCH; ++i) case 1: for(size_t i=0; i<2*BATCH; ++i)
@@ -1172,12 +1228,16 @@ static size_t resample_opt4p4o_2batch(struct resample_data *rd, float*in, float
in += 2; in += 2;
} }
break; break;
default: for(size_t i=0; i<2*BATCH; ++i) default:
#endif
for(size_t i=0; i<2*BATCH; ++i)
{ {
OPT4P4O_INTERPOL(in, out, outs, rd->channels) OPT4P4O_INTERPOL(in, out, outs, rd->channels)
in += rd->channels; in += rd->channels;
} }
#ifndef SYN123_NO_CASES
} }
#endif
OPT4P4O_END OPT4P4O_END
return outs; return outs;
} }
@@ -1266,6 +1326,7 @@ static size_t resample_opt6p5o(struct resample_data *rd, float*in
if(!ins) if(!ins)
return outs; return outs;
OPT6P5O_BEGIN(in) OPT6P5O_BEGIN(in)
#ifndef SYN123_NO_CASES
switch(rd->channels) switch(rd->channels)
{ {
case 1: for(size_t i=0; i<ins; ++i) case 1: for(size_t i=0; i<ins; ++i)
@@ -1280,12 +1341,16 @@ static size_t resample_opt6p5o(struct resample_data *rd, float*in
in += 2; in += 2;
} }
break; break;
default: for(size_t i=0; i<ins; ++i) default:
#endif
for(size_t i=0; i<ins; ++i)
{ {
OPT6P5O_INTERPOL(in, out, outs, rd->channels) OPT6P5O_INTERPOL(in, out, outs, rd->channels)
in += rd->channels; in += rd->channels;
} }
#ifndef SYN123_NO_CASES
} }
#endif
OPT6P5O_END OPT6P5O_END
return outs; return outs;
} }
@@ -1296,6 +1361,7 @@ static size_t resample_opt6p5o_2batch(struct resample_data *rd, float*in
{ {
size_t outs = 0; size_t outs = 0;
OPT6P5O_BEGIN(in) OPT6P5O_BEGIN(in)
#ifndef SYN123_NO_CASES
switch(rd->channels) switch(rd->channels)
{ {
case 1: for(size_t i=0; i<2*BATCH; ++i) case 1: for(size_t i=0; i<2*BATCH; ++i)
@@ -1310,12 +1376,16 @@ static size_t resample_opt6p5o_2batch(struct resample_data *rd, float*in
in += 2; in += 2;
} }
break; break;
default: for(size_t i=0; i<2*BATCH; ++i) default:
#endif
for(size_t i=0; i<2*BATCH; ++i)
{ {
OPT6P5O_INTERPOL(in, out, outs, rd->channels) OPT6P5O_INTERPOL(in, out, outs, rd->channels)
in += rd->channels; in += rd->channels;
} }
#ifndef SYN123_NO_CASES
} }
#endif
OPT6P5O_END OPT6P5O_END
return outs; return outs;
} }
@@ -1888,7 +1958,9 @@ syn123_setup_resample( syn123_handle *sh, long inrate, long outrate
rd->decim_hist = NULL; rd->decim_hist = NULL;
rd->channels = channels; rd->channels = channels;
rd->frame = NULL; rd->frame = NULL;
#ifndef SYN123_NO_CASES
if(channels > 2) if(channels > 2)
#endif
{ {
rd->frame = malloc(sizeof(float)*channels); rd->frame = malloc(sizeof(float)*channels);
if(!rd->frame) if(!rd->frame)