Patchwork pulseaudio: rescale input being passed to float method of speex

login
register
mail settings
Submitter Fahad Arslan
Date Dec. 24, 2013, 2:37 p.m.
Message ID <1387895862-15110-1-git-send-email-fahad_arslan@mentor.com>
Download mbox | patch
Permalink /patch/63709/
State New
Headers show

Comments

Fahad Arslan - Dec. 24, 2013, 2:37 p.m.
From: Fahad Arslan <Fahad_Arslan@mentor.com>

Pulseaudio uses Speex to do resampling. Default Pulseaudio resampler
is speex-float-1. However, Speex recipe in poky configures speex in
fixed point. This scenario creates a situation in which audio streams
that need to be resampled are not playedback since input to speex is
zeroed out when flaot input in range of +/-1 is converted to int.
So we are rescaling the input before invoking speex flaot method.

Upstream-Status: Submitted [pulseaudio-discuss@lists.freedesktop.org]

Signed-off-by: Fahad Arslan <Fahad_Arslan@mentor.com>
---
 .../pulseaudio/rescale_input_to_speex_float.patch  |   24 ++++++++++++++++++++
 .../pulseaudio/pulseaudio_4.0.bb                   |    3 ++-
 2 files changed, 26 insertions(+), 1 deletion(-)
 create mode 100644 meta/recipes-multimedia/pulseaudio/pulseaudio/rescale_input_to_speex_float.patch
Saul Wold - Dec. 24, 2013, 4:54 p.m.
On 12/24/2013 06:37 AM, Fahad Arslan wrote:
> From: Fahad Arslan <Fahad_Arslan@mentor.com>
>
> Pulseaudio uses Speex to do resampling. Default Pulseaudio resampler
> is speex-float-1. However, Speex recipe in poky configures speex in
> fixed point. This scenario creates a situation in which audio streams
> that need to be resampled are not playedback since input to speex is
> zeroed out when flaot input in range of +/-1 is converted to int.
> So we are rescaling the input before invoking speex flaot method.
>
> Upstream-Status: Submitted [pulseaudio-discuss@lists.freedesktop.org]
>
> Signed-off-by: Fahad Arslan <Fahad_Arslan@mentor.com>
> ---
>   .../pulseaudio/rescale_input_to_speex_float.patch  |   24 ++++++++++++++++++++
>   .../pulseaudio/pulseaudio_4.0.bb                   |    3 ++-
>   2 files changed, 26 insertions(+), 1 deletion(-)
>   create mode 100644 meta/recipes-multimedia/pulseaudio/pulseaudio/rescale_input_to_speex_float.patch
>
> diff --git a/meta/recipes-multimedia/pulseaudio/pulseaudio/rescale_input_to_speex_float.patch b/meta/recipes-multimedia/pulseaudio/pulseaudio/rescale_input_to_speex_float.patch
> new file mode 100644
> index 0000000..fe5e6a7
> --- /dev/null
> +++ b/meta/recipes-multimedia/pulseaudio/pulseaudio/rescale_input_to_speex_float.patch

Fahad,

The patch itself is missing the Upstream-Status: and Signed-off-by: 
Tags, I know you included them in the commit message above, but they 
need to be in this patch file also.

Thanks
	Sau!

> @@ -0,0 +1,24 @@
> +diff -Naur a/src/pulsecore/resampler.c b/src/pulsecore/resampler.c
> +--- a/src/pulsecore/resampler.c	2013-12-24 18:25:08.164787401 +0500
> ++++ b/src/pulsecore/resampler.c	2013-12-24 18:24:43.056787026 +0500
> +@@ -1347,7 +1347,7 @@
> +
> + static void speex_resample_float(pa_resampler *r, const pa_memchunk *input, unsigned in_n_frames, pa_memchunk *output, unsigned *out_n_frames) {
> +     float *in, *out;
> +-    uint32_t inf = in_n_frames, outf = *out_n_frames;
> ++    uint32_t inf = in_n_frames, outf = *out_n_frames, i;
> +
> +     pa_assert(r);
> +     pa_assert(input);
> +@@ -1357,6 +1357,11 @@
> +     in = pa_memblock_acquire_chunk(input);
> +     out = pa_memblock_acquire_chunk(output);
> +
> ++    /* Speex float API scale range is +/-32768 instead of +/-1.
> ++       So rescale input before passing it to Speex. */
> ++    for (i = 0; i < inf; i++)
> ++        in[i] = 32768.*in[i];
> ++
> +     pa_assert_se(speex_resampler_process_interleaved_float(r->speex.state, in, &inf, out, &outf) == 0);
> +
> +     pa_memblock_release(input->memblock);
> diff --git a/meta/recipes-multimedia/pulseaudio/pulseaudio_4.0.bb b/meta/recipes-multimedia/pulseaudio/pulseaudio_4.0.bb
> index b419c54..5afa20b 100644
> --- a/meta/recipes-multimedia/pulseaudio/pulseaudio_4.0.bb
> +++ b/meta/recipes-multimedia/pulseaudio/pulseaudio_4.0.bb
> @@ -1,7 +1,8 @@
>   require pulseaudio.inc
>
>   SRC_URI = "http://freedesktop.org/software/pulseaudio/releases/pulseaudio-${PV}.tar.xz \
> -           file://volatiles.04_pulse"
> +           file://volatiles.04_pulse \
> +           file://rescale_input_to_speex_float.patch"
>
>   SRC_URI[md5sum] = "591f211db2790a7e4d222f2dc6858db3"
>   SRC_URI[sha256sum] = "35ceb36bb1822fe54f0b5e4863b4f486769fdfb8ff2111f01fd8778928f9cdae"
>
Koen Kooi - Dec. 26, 2013, 11:57 a.m.
Op 24 dec. 2013, om 15:37 heeft Fahad Arslan <fahad_arslan@mentor.com> het volgende geschreven:

> From: Fahad Arslan <Fahad_Arslan@mentor.com>
> 
> Pulseaudio uses Speex to do resampling. Default Pulseaudio resampler
> is speex-float-1. However, Speex recipe in poky configures speex in

s/poky/oe-core/

> fixed point. This scenario creates a situation in which audio streams
> that need to be resampled are not playedback since input to speex is
> zeroed out when flaot input in range of +/-1 is converted to int.
> So we are rescaling the input before invoking speex flaot method.

s/flaot/float/g

> 
> Upstream-Status: Submitted [pulseaudio-discuss@lists.freedesktop.org]
> 
> Signed-off-by: Fahad Arslan <Fahad_Arslan@mentor.com>
> ---
> .../pulseaudio/rescale_input_to_speex_float.patch  |   24 ++++++++++++++++++++
> .../pulseaudio/pulseaudio_4.0.bb                   |    3 ++-
> 2 files changed, 26 insertions(+), 1 deletion(-)
> create mode 100644 meta/recipes-multimedia/pulseaudio/pulseaudio/rescale_input_to_speex_float.patch
> 
> diff --git a/meta/recipes-multimedia/pulseaudio/pulseaudio/rescale_input_to_speex_float.patch b/meta/recipes-multimedia/pulseaudio/pulseaudio/rescale_input_to_speex_float.patch
> new file mode 100644
> index 0000000..fe5e6a7
> --- /dev/null
> +++ b/meta/recipes-multimedia/pulseaudio/pulseaudio/rescale_input_to_speex_float.patch
> @@ -0,0 +1,24 @@
> +diff -Naur a/src/pulsecore/resampler.c b/src/pulsecore/resampler.c
> +--- a/src/pulsecore/resampler.c	2013-12-24 18:25:08.164787401 +0500
> ++++ b/src/pulsecore/resampler.c	2013-12-24 18:24:43.056787026 +0500
> +@@ -1347,7 +1347,7 @@
> + 
> + static void speex_resample_float(pa_resampler *r, const pa_memchunk *input, unsigned in_n_frames, pa_memchunk *output, unsigned *out_n_frames) {
> +     float *in, *out;
> +-    uint32_t inf = in_n_frames, outf = *out_n_frames;
> ++    uint32_t inf = in_n_frames, outf = *out_n_frames, i;
> + 
> +     pa_assert(r);
> +     pa_assert(input);
> +@@ -1357,6 +1357,11 @@
> +     in = pa_memblock_acquire_chunk(input);
> +     out = pa_memblock_acquire_chunk(output);
> + 
> ++    /* Speex float API scale range is +/-32768 instead of +/-1.
> ++       So rescale input before passing it to Speex. */
> ++    for (i = 0; i < inf; i++)
> ++        in[i] = 32768.*in[i];
> ++
> +     pa_assert_se(speex_resampler_process_interleaved_float(r->speex.state, in, &inf, out, &outf) == 0);
> + 
> +     pa_memblock_release(input->memblock);
> diff --git a/meta/recipes-multimedia/pulseaudio/pulseaudio_4.0.bb b/meta/recipes-multimedia/pulseaudio/pulseaudio_4.0.bb
> index b419c54..5afa20b 100644
> --- a/meta/recipes-multimedia/pulseaudio/pulseaudio_4.0.bb
> +++ b/meta/recipes-multimedia/pulseaudio/pulseaudio_4.0.bb
> @@ -1,7 +1,8 @@
> require pulseaudio.inc
> 
> SRC_URI = "http://freedesktop.org/software/pulseaudio/releases/pulseaudio-${PV}.tar.xz \
> -           file://volatiles.04_pulse"
> +           file://volatiles.04_pulse \
> +           file://rescale_input_to_speex_float.patch"
> 
> SRC_URI[md5sum] = "591f211db2790a7e4d222f2dc6858db3"
> SRC_URI[sha256sum] = "35ceb36bb1822fe54f0b5e4863b4f486769fdfb8ff2111f01fd8778928f9cdae"
> -- 
> 1.7.9.5
> 
> _______________________________________________
> Openembedded-core mailing list
> Openembedded-core@lists.openembedded.org
> http://lists.openembedded.org/mailman/listinfo/openembedded-core
>
Fahad Arslan - Jan. 1, 2014, 4:01 p.m.
Pulseaudio upstream says that this patch breaks speex-float if 
it isn't compiled with FIXED_POINT defined, which is valid concern. 
So what is the best option now to avoid the no sound issue:
a. replace --enable-fixed-point and --disable-float-api with --enable-float-api in Speex's recipe
b. replace default resampler (speex-float-1) used by Pulseaudio to speex-fixed-1

In my opinion option (a) above is better. Thoughts/suggestions? 

Thanks,
Fahad
Saul Wold - Jan. 2, 2014, 6:47 p.m.
On 01/01/2014 08:01 AM, Arslan, Fahad wrote:
> Pulseaudio upstream says that this patch breaks speex-float if
> it isn't compiled with FIXED_POINT defined, which is valid concern.
> So what is the best option now to avoid the no sound issue:
> a. replace --enable-fixed-point and --disable-float-api with --enable-float-api in Speex's recipe

There is a commit message from 2008 by Marcin J. that seems to disable 
this, I am not sure why that choice was made at that time.

What affect does it have on the overall size of speex

> b. replace default resampler (speex-float-1) used by Pulseaudio to speex-fixed-1
>
> In my opinion option (a) above is better. Thoughts/suggestions?
>
> Thanks,
> Fahad
>
> ________________________________________
> From: Saul Wold [sgw@linux.intel.com]
> Sent: Tuesday, December 24, 2013 9:54 PM
> To: Arslan, Fahad; openembedded-core@lists.openembedded.org
> Subject: Re: [OE-core] [PATCH] pulseaudio: rescale input being passed to float method of speex
>
> On 12/24/2013 06:37 AM, Fahad Arslan wrote:
>> From: Fahad Arslan <Fahad_Arslan@mentor.com>
>>
>> Pulseaudio uses Speex to do resampling. Default Pulseaudio resampler
>> is speex-float-1. However, Speex recipe in poky configures speex in
>> fixed point. This scenario creates a situation in which audio streams
>> that need to be resampled are not playedback since input to speex is
>> zeroed out when flaot input in range of +/-1 is converted to int.
>> So we are rescaling the input before invoking speex flaot method.
>>
>> Upstream-Status: Submitted [pulseaudio-discuss@lists.freedesktop.org]
>>
>> Signed-off-by: Fahad Arslan <Fahad_Arslan@mentor.com>
>> ---
>>    .../pulseaudio/rescale_input_to_speex_float.patch  |   24 ++++++++++++++++++++
>>    .../pulseaudio/pulseaudio_4.0.bb                   |    3 ++-
>>    2 files changed, 26 insertions(+), 1 deletion(-)
>>    create mode 100644 meta/recipes-multimedia/pulseaudio/pulseaudio/rescale_input_to_speex_float.patch
>>
>> diff --git a/meta/recipes-multimedia/pulseaudio/pulseaudio/rescale_input_to_speex_float.patch b/meta/recipes-multimedia/pulseaudio/pulseaudio/rescale_input_to_speex_float.patch
>> new file mode 100644
>> index 0000000..fe5e6a7
>> --- /dev/null
>> +++ b/meta/recipes-multimedia/pulseaudio/pulseaudio/rescale_input_to_speex_float.patch
>
> Fahad,
>
> The patch itself is missing the Upstream-Status: and Signed-off-by:
> Tags, I know you included them in the commit message above, but they
> need to be in this patch file also.
>
> Thanks
>          Sau!
>
>> @@ -0,0 +1,24 @@
>> +diff -Naur a/src/pulsecore/resampler.c b/src/pulsecore/resampler.c
>> +--- a/src/pulsecore/resampler.c      2013-12-24 18:25:08.164787401 +0500
>> ++++ b/src/pulsecore/resampler.c      2013-12-24 18:24:43.056787026 +0500
>> +@@ -1347,7 +1347,7 @@
>> +
>> + static void speex_resample_float(pa_resampler *r, const pa_memchunk *input, unsigned in_n_frames, pa_memchunk *output, unsigned *out_n_frames) {
>> +     float *in, *out;
>> +-    uint32_t inf = in_n_frames, outf = *out_n_frames;
>> ++    uint32_t inf = in_n_frames, outf = *out_n_frames, i;
>> +
>> +     pa_assert(r);
>> +     pa_assert(input);
>> +@@ -1357,6 +1357,11 @@
>> +     in = pa_memblock_acquire_chunk(input);
>> +     out = pa_memblock_acquire_chunk(output);
>> +
>> ++    /* Speex float API scale range is +/-32768 instead of +/-1.
>> ++       So rescale input before passing it to Speex. */
>> ++    for (i = 0; i < inf; i++)
>> ++        in[i] = 32768.*in[i];
>> ++
>> +     pa_assert_se(speex_resampler_process_interleaved_float(r->speex.state, in, &inf, out, &outf) == 0);
>> +
>> +     pa_memblock_release(input->memblock);
>> diff --git a/meta/recipes-multimedia/pulseaudio/pulseaudio_4.0.bb b/meta/recipes-multimedia/pulseaudio/pulseaudio_4.0.bb
>> index b419c54..5afa20b 100644
>> --- a/meta/recipes-multimedia/pulseaudio/pulseaudio_4.0.bb
>> +++ b/meta/recipes-multimedia/pulseaudio/pulseaudio_4.0.bb
>> @@ -1,7 +1,8 @@
>>    require pulseaudio.inc
>>
>>    SRC_URI = "http://freedesktop.org/software/pulseaudio/releases/pulseaudio-${PV}.tar.xz \
>> -           file://volatiles.04_pulse"
>> +           file://volatiles.04_pulse \
>> +           file://rescale_input_to_speex_float.patch"
>>
>>    SRC_URI[md5sum] = "591f211db2790a7e4d222f2dc6858db3"
>>    SRC_URI[sha256sum] = "35ceb36bb1822fe54f0b5e4863b4f486769fdfb8ff2111f01fd8778928f9cdae"
>>
>
>
Fahad Arslan - Jan. 7, 2014, 9:59 a.m.
> What affect does it have on the overall size of speex

If we enable floating point support, there is decrease in size of libs. 
libspeexdsp.so is ~480 KB in fixed point configuration
libspeexdsp.so is ~220 KB in floating point configuration

Further details are shown below:


Case-1 (current configuration)
======

EXTRA_OECONF = " --enable-fixed-point --with-ogg-libraries=${STAGING_LIBDIR} \
                 --disable-float-api --disable-vbr \
                 --with-ogg-includes=${STAGING_INCDIR} --disable-oggtest"

build$ 
build$ ls -l src/.libs/speex*
-rwxr-xr-x 1 farslan farslan 75112 Jan  7 14:09 src/.libs/speexdec
-rwxr-xr-x 1 farslan farslan 76989 Jan  7 14:09 src/.libs/speexenc
build$ file src/.libs/speex*
src/.libs/speexdec: ELF 32-bit LSB executable, ARM, version 1 (SYSV), dynamically linked (uses shared libs), for GNU/Linux 2.6.16, not stripped
src/.libs/speexenc: ELF 32-bit LSB executable, ARM, version 1 (SYSV), dynamically linked (uses shared libs), for GNU/Linux 2.6.16, not stripped
build$ 
build$ 
build$ ls -l libspeex/.libs/libspeex*.so.1.5.0 
-rwxr-xr-x 1 farslan farslan 484940 Jan  7 14:09 libspeex/.libs/libspeexdsp.so.1.5.0
-rwxr-xr-x 1 farslan farslan 370309 Jan  7 14:09 libspeex/.libs/libspeex.so.1.5.0
build$ 
build$ file libspeex/.libs/libspeex*.so.1.5.0 
libspeex/.libs/libspeexdsp.so.1.5.0: ELF 32-bit LSB shared object, ARM, version 1 (SYSV), dynamically linked, not stripped
libspeex/.libs/libspeex.so.1.5.0:    ELF 32-bit LSB shared object, ARM, version 1 (SYSV), dynamically linked, not stripped
build$ 


Case-2 (suggested configuration)
======

EXTRA_OECONF = " --with-ogg-libraries=${STAGING_LIBDIR} \
                 --enable-float-api --disable-vbr \
                 --with-ogg-includes=${STAGING_INCDIR} --disable-oggtest"

build$ 
build$ ls -l src/.libs/speex*
-rwxr-xr-x 1 farslan farslan 75112 Jan  7 14:40 src/.libs/speexdec
-rwxr-xr-x 1 farslan farslan 76989 Jan  7 14:40 src/.libs/speexenc
build$ file src/.libs/speex*
src/.libs/speexdec: ELF 32-bit LSB executable, ARM, version 1 (SYSV), dynamically linked (uses shared libs), for GNU/Linux 2.6.16, not stripped
src/.libs/speexenc: ELF 32-bit LSB executable, ARM, version 1 (SYSV), dynamically linked (uses shared libs), for GNU/Linux 2.6.16, not stripped
build$
build$
build$ ls -l libspeex/.libs/libspeex*.so.1.5.0 
-rwxr-xr-x 1 farslan farslan 222582 Jan  7 14:40 libspeex/.libs/libspeexdsp.so.1.5.0
-rwxr-xr-x 1 farslan farslan 280029 Jan  7 14:39 libspeex/.libs/libspeex.so.1.5.0
build$ 
build$ file libspeex/.libs/libspeex*.so.1.5.0 
libspeex/.libs/libspeexdsp.so.1.5.0: ELF 32-bit LSB shared object, ARM, version 1 (SYSV), dynamically linked, not stripped
libspeex/.libs/libspeex.so.1.5.0:    ELF 32-bit LSB shared object, ARM, version 1 (SYSV), dynamically linked, not stripped
build$



Thanks,
Fahad
Koen Kooi - Jan. 7, 2014, 10:03 a.m.
Op 7 jan. 2014, om 10:59 heeft Arslan, Fahad <Fahad_Arslan@mentor.com> het volgende geschreven:

>> What affect does it have on the overall size of speex
> 
> If we enable floating point support, there is decrease in size of libs. 
> libspeexdsp.so is ~480 KB in fixed point configuration
> libspeexdsp.so is ~220 KB in floating point configuration
> 
> Further details are shown below:
> 
> 
> Case-1 (current configuration)
> ======
> 
> EXTRA_OECONF = " --enable-fixed-point --with-ogg-libraries=${STAGING_LIBDIR} \
>                 --disable-float-api --disable-vbr \
>                 --with-ogg-includes=${STAGING_INCDIR} --disable-oggtest"
> 
> build$ 
> build$ ls -l src/.libs/speex*
> -rwxr-xr-x 1 farslan farslan 75112 Jan  7 14:09 src/.libs/speexdec
> -rwxr-xr-x 1 farslan farslan 76989 Jan  7 14:09 src/.libs/speexenc
> build$ file src/.libs/speex*
> src/.libs/speexdec: ELF 32-bit LSB executable, ARM, version 1 (SYSV), dynamically linked (uses shared libs), for GNU/Linux 2.6.16, not stripped
> src/.libs/speexenc: ELF 32-bit LSB executable, ARM, version 1 (SYSV), dynamically linked (uses shared libs), for GNU/Linux 2.6.16, not stripped
> build$ 
> build$ 
> build$ ls -l libspeex/.libs/libspeex*.so.1.5.0 
> -rwxr-xr-x 1 farslan farslan 484940 Jan  7 14:09 libspeex/.libs/libspeexdsp.so.1.5.0
> -rwxr-xr-x 1 farslan farslan 370309 Jan  7 14:09 libspeex/.libs/libspeex.so.1.5.0
> build$ 
> build$ file libspeex/.libs/libspeex*.so.1.5.0 
> libspeex/.libs/libspeexdsp.so.1.5.0: ELF 32-bit LSB shared object, ARM, version 1 (SYSV), dynamically linked, not stripped
> libspeex/.libs/libspeex.so.1.5.0:    ELF 32-bit LSB shared object, ARM, version 1 (SYSV), dynamically linked, not stripped
> build$ 
> 
> 
> Case-2 (suggested configuration)
> ======
> 
> EXTRA_OECONF = " --with-ogg-libraries=${STAGING_LIBDIR} \
>                 --enable-float-api --disable-vbr \
>                 --with-ogg-includes=${STAGING_INCDIR} --disable-oggtest"

Or case 3, which I implemented 3 years ago in https://github.com/openembedded/openembedded/commit/e06553979d23531397af3dd71870abb80718c681 :

def get_speex_fpu_setting(bb, d):
        if bb.data.getVar('TARGET_FPU', d, 1) in [ 'soft' ]:
                return "--enable-fixed-point --disable-float-api --disable-vbr"
        return ""


EXTRA_OECONF += "${@get_speex_fpu_setting(bb, d)}

That OE-classic recipe also has support for arm asm which speed things up a lot.

regards,

Koen
Koen Kooi - Jan. 7, 2014, 10:22 a.m.
Marcins linaro mail address has expired, adding the other one

Op 7 jan. 2014, om 11:03 heeft Koen Kooi <koen@dominion.thruhere.net> het volgende geschreven:

> 
> Op 7 jan. 2014, om 10:59 heeft Arslan, Fahad <Fahad_Arslan@mentor.com> het volgende geschreven:
> 
>>> What affect does it have on the overall size of speex
>> 
>> If we enable floating point support, there is decrease in size of libs. 
>> libspeexdsp.so is ~480 KB in fixed point configuration
>> libspeexdsp.so is ~220 KB in floating point configuration
>> 
>> Further details are shown below:
>> 
>> 
>> Case-1 (current configuration)
>> ======
>> 
>> EXTRA_OECONF = " --enable-fixed-point --with-ogg-libraries=${STAGING_LIBDIR} \
>>                --disable-float-api --disable-vbr \
>>                --with-ogg-includes=${STAGING_INCDIR} --disable-oggtest"
>> 
>> build$ 
>> build$ ls -l src/.libs/speex*
>> -rwxr-xr-x 1 farslan farslan 75112 Jan  7 14:09 src/.libs/speexdec
>> -rwxr-xr-x 1 farslan farslan 76989 Jan  7 14:09 src/.libs/speexenc
>> build$ file src/.libs/speex*
>> src/.libs/speexdec: ELF 32-bit LSB executable, ARM, version 1 (SYSV), dynamically linked (uses shared libs), for GNU/Linux 2.6.16, not stripped
>> src/.libs/speexenc: ELF 32-bit LSB executable, ARM, version 1 (SYSV), dynamically linked (uses shared libs), for GNU/Linux 2.6.16, not stripped
>> build$ 
>> build$ 
>> build$ ls -l libspeex/.libs/libspeex*.so.1.5.0 
>> -rwxr-xr-x 1 farslan farslan 484940 Jan  7 14:09 libspeex/.libs/libspeexdsp.so.1.5.0
>> -rwxr-xr-x 1 farslan farslan 370309 Jan  7 14:09 libspeex/.libs/libspeex.so.1.5.0
>> build$ 
>> build$ file libspeex/.libs/libspeex*.so.1.5.0 
>> libspeex/.libs/libspeexdsp.so.1.5.0: ELF 32-bit LSB shared object, ARM, version 1 (SYSV), dynamically linked, not stripped
>> libspeex/.libs/libspeex.so.1.5.0:    ELF 32-bit LSB shared object, ARM, version 1 (SYSV), dynamically linked, not stripped
>> build$ 
>> 
>> 
>> Case-2 (suggested configuration)
>> ======
>> 
>> EXTRA_OECONF = " --with-ogg-libraries=${STAGING_LIBDIR} \
>>                --enable-float-api --disable-vbr \
>>                --with-ogg-includes=${STAGING_INCDIR} --disable-oggtest"
> 
> Or case 3, which I implemented 3 years ago in https://github.com/openembedded/openembedded/commit/e06553979d23531397af3dd71870abb80718c681 :
> 
> def get_speex_fpu_setting(bb, d):
>        if bb.data.getVar('TARGET_FPU', d, 1) in [ 'soft' ]:
>                return "--enable-fixed-point --disable-float-api --disable-vbr"
>        return ""
> 
> 
> EXTRA_OECONF += "${@get_speex_fpu_setting(bb, d)}
> 
> That OE-classic recipe also has support for arm asm which speed things up a lot.
> 
> regards,
> 
> Koen
Fahad Arslan - Jan. 10, 2014, 7:48 a.m.
>def get_speex_fpu_setting(bb, d):
>        if bb.data.getVar('TARGET_FPU', d, 1) in [ 'soft' ]:
>                return "--enable-fixed-point --disable-float-api --disable-vbr"
>        return ""

If we are using Pulseaudio with default settings (speex-float-1) and 
the if condition in above method is evaluated true, we will end up facing 
the original issue: streams that need to be resampled are not playedback 
since input to speex is zeroed out when float input passed by Pulseaudio 
in range of +/-1 is converted to int.

So isn't Case 2 appropriate, thoughts?

Regards,
Fahad
Koen Kooi - Jan. 10, 2014, 3:33 p.m.
Op 10 jan. 2014, om 08:48 heeft Arslan, Fahad <Fahad_Arslan@mentor.com> het volgende geschreven:

>> def get_speex_fpu_setting(bb, d):
>>       if bb.data.getVar('TARGET_FPU', d, 1) in [ 'soft' ]:
>>               return "--enable-fixed-point --disable-float-api --disable-vbr"
>>       return ""
> 
> If we are using Pulseaudio with default settings (speex-float-1) and 
> the if condition in above method is evaluated true, we will end up facing 
> the original issue: streams that need to be resampled are not playedback 
> since input to speex is zeroed out when float input passed by Pulseaudio 
> in range of +/-1 is converted to int.
> 
> So isn't Case 2 appropriate, thoughts?

It's also 10-50x slower on ARM, so always enabling float is a bad option. Just patch it to work, no matter what upstream says :)


> 
> Regards,
> Fahad
> 
> ________________________________________
> From: Koen Kooi [koen@dominion.thruhere.net]
> Sent: Tuesday, January 07, 2014 3:22 PM
> To: Arslan, Fahad
> Cc: Saul Wold; openembedded-core@lists.openembedded.org; marcin@juszkiewicz.pl
> Subject: Re: [OE-core] [PATCH] pulseaudio: rescale input being passed to float method of speex
> 
> Marcins linaro mail address has expired, adding the other one
> 
> Op 7 jan. 2014, om 11:03 heeft Koen Kooi <koen@dominion.thruhere.net> het volgende geschreven:
> 
>> 
>> Op 7 jan. 2014, om 10:59 heeft Arslan, Fahad <Fahad_Arslan@mentor.com> het volgende geschreven:
>> 
>>>> What affect does it have on the overall size of speex
>>> 
>>> If we enable floating point support, there is decrease in size of libs.
>>> libspeexdsp.so is ~480 KB in fixed point configuration
>>> libspeexdsp.so is ~220 KB in floating point configuration
>>> 
>>> Further details are shown below:
>>> 
>>> 
>>> Case-1 (current configuration)
>>> ======
>>> 
>>> EXTRA_OECONF = " --enable-fixed-point --with-ogg-libraries=${STAGING_LIBDIR} \
>>>               --disable-float-api --disable-vbr \
>>>               --with-ogg-includes=${STAGING_INCDIR} --disable-oggtest"
>>> 
>>> build$
>>> build$ ls -l src/.libs/speex*
>>> -rwxr-xr-x 1 farslan farslan 75112 Jan  7 14:09 src/.libs/speexdec
>>> -rwxr-xr-x 1 farslan farslan 76989 Jan  7 14:09 src/.libs/speexenc
>>> build$ file src/.libs/speex*
>>> src/.libs/speexdec: ELF 32-bit LSB executable, ARM, version 1 (SYSV), dynamically linked (uses shared libs), for GNU/Linux 2.6.16, not stripped
>>> src/.libs/speexenc: ELF 32-bit LSB executable, ARM, version 1 (SYSV), dynamically linked (uses shared libs), for GNU/Linux 2.6.16, not stripped
>>> build$
>>> build$
>>> build$ ls -l libspeex/.libs/libspeex*.so.1.5.0
>>> -rwxr-xr-x 1 farslan farslan 484940 Jan  7 14:09 libspeex/.libs/libspeexdsp.so.1.5.0
>>> -rwxr-xr-x 1 farslan farslan 370309 Jan  7 14:09 libspeex/.libs/libspeex.so.1.5.0
>>> build$
>>> build$ file libspeex/.libs/libspeex*.so.1.5.0
>>> libspeex/.libs/libspeexdsp.so.1.5.0: ELF 32-bit LSB shared object, ARM, version 1 (SYSV), dynamically linked, not stripped
>>> libspeex/.libs/libspeex.so.1.5.0:    ELF 32-bit LSB shared object, ARM, version 1 (SYSV), dynamically linked, not stripped
>>> build$
>>> 
>>> 
>>> Case-2 (suggested configuration)
>>> ======
>>> 
>>> EXTRA_OECONF = " --with-ogg-libraries=${STAGING_LIBDIR} \
>>>               --enable-float-api --disable-vbr \
>>>               --with-ogg-includes=${STAGING_INCDIR} --disable-oggtest"
>> 
>> Or case 3, which I implemented 3 years ago in https://github.com/openembedded/openembedded/commit/e06553979d23531397af3dd71870abb80718c681 :
>> 
>> def get_speex_fpu_setting(bb, d):
>>       if bb.data.getVar('TARGET_FPU', d, 1) in [ 'soft' ]:
>>               return "--enable-fixed-point --disable-float-api --disable-vbr"
>>       return ""
>> 
>> 
>> EXTRA_OECONF += "${@get_speex_fpu_setting(bb, d)}
>> 
>> That OE-classic recipe also has support for arm asm which speed things up a lot.
>> 
>> regards,
>> 
>> Koen
> 
>
Fahad Arslan - Jan. 16, 2014, 12:08 p.m.
>It's also 10-50x slower on ARM, so always enabling float is a bad option. Just patch it to work, no matter what upstream says :)

Seems I misunderstood the statement earlier. So to confirm, do you mean I should submit patch here for original bug in Speex source code?
Koen Kooi - Jan. 16, 2014, 7:16 p.m.
Op 16 jan. 2014, om 13:08 heeft Arslan, Fahad <Fahad_Arslan@mentor.com> het volgende geschreven:

>> It's also 10-50x slower on ARM, so always enabling float is a bad option. Just patch it to work, no matter what upstream says :)
> 
> Seems I misunderstood the statement earlier. So to confirm, do you mean I should submit patch here for original bug in Speex source code?

Whatever allows us to use fixed-point mode on fpuless systems

> 
> ________________________________________
> From: Koen Kooi [koen@dominion.thruhere.net]
> Sent: Friday, January 10, 2014 8:33 PM
> To: Arslan, Fahad
> Cc: Saul Wold; openembedded-core@lists.openembedded.org; Marcin Juszkiewicz
> Subject: Re: [OE-core] [PATCH] pulseaudio: rescale input being passed to float method of speex
> 
> Op 10 jan. 2014, om 08:48 heeft Arslan, Fahad <Fahad_Arslan@mentor.com> het volgende geschreven:
> 
>>> def get_speex_fpu_setting(bb, d):
>>>      if bb.data.getVar('TARGET_FPU', d, 1) in [ 'soft' ]:
>>>              return "--enable-fixed-point --disable-float-api --disable-vbr"
>>>      return ""
>> 
>> If we are using Pulseaudio with default settings (speex-float-1) and
>> the if condition in above method is evaluated true, we will end up facing
>> the original issue: streams that need to be resampled are not playedback
>> since input to speex is zeroed out when float input passed by Pulseaudio
>> in range of +/-1 is converted to int.
>> 
>> So isn't Case 2 appropriate, thoughts?
> 
> It's also 10-50x slower on ARM, so always enabling float is a bad option. Just patch it to work, no matter what upstream says :)
> 
> 
>> 
>> Regards,
>> Fahad
>> 
>> ________________________________________
>> From: Koen Kooi [koen@dominion.thruhere.net]
>> Sent: Tuesday, January 07, 2014 3:22 PM
>> To: Arslan, Fahad
>> Cc: Saul Wold; openembedded-core@lists.openembedded.org; marcin@juszkiewicz.pl
>> Subject: Re: [OE-core] [PATCH] pulseaudio: rescale input being passed to float method of speex
>> 
>> Marcins linaro mail address has expired, adding the other one
>> 
>> Op 7 jan. 2014, om 11:03 heeft Koen Kooi <koen@dominion.thruhere.net> het volgende geschreven:
>> 
>>> 
>>> Op 7 jan. 2014, om 10:59 heeft Arslan, Fahad <Fahad_Arslan@mentor.com> het volgende geschreven:
>>> 
>>>>> What affect does it have on the overall size of speex
>>>> 
>>>> If we enable floating point support, there is decrease in size of libs.
>>>> libspeexdsp.so is ~480 KB in fixed point configuration
>>>> libspeexdsp.so is ~220 KB in floating point configuration
>>>> 
>>>> Further details are shown below:
>>>> 
>>>> 
>>>> Case-1 (current configuration)
>>>> ======
>>>> 
>>>> EXTRA_OECONF = " --enable-fixed-point --with-ogg-libraries=${STAGING_LIBDIR} \
>>>>              --disable-float-api --disable-vbr \
>>>>              --with-ogg-includes=${STAGING_INCDIR} --disable-oggtest"
>>>> 
>>>> build$
>>>> build$ ls -l src/.libs/speex*
>>>> -rwxr-xr-x 1 farslan farslan 75112 Jan  7 14:09 src/.libs/speexdec
>>>> -rwxr-xr-x 1 farslan farslan 76989 Jan  7 14:09 src/.libs/speexenc
>>>> build$ file src/.libs/speex*
>>>> src/.libs/speexdec: ELF 32-bit LSB executable, ARM, version 1 (SYSV), dynamically linked (uses shared libs), for GNU/Linux 2.6.16, not stripped
>>>> src/.libs/speexenc: ELF 32-bit LSB executable, ARM, version 1 (SYSV), dynamically linked (uses shared libs), for GNU/Linux 2.6.16, not stripped
>>>> build$
>>>> build$
>>>> build$ ls -l libspeex/.libs/libspeex*.so.1.5.0
>>>> -rwxr-xr-x 1 farslan farslan 484940 Jan  7 14:09 libspeex/.libs/libspeexdsp.so.1.5.0
>>>> -rwxr-xr-x 1 farslan farslan 370309 Jan  7 14:09 libspeex/.libs/libspeex.so.1.5.0
>>>> build$
>>>> build$ file libspeex/.libs/libspeex*.so.1.5.0
>>>> libspeex/.libs/libspeexdsp.so.1.5.0: ELF 32-bit LSB shared object, ARM, version 1 (SYSV), dynamically linked, not stripped
>>>> libspeex/.libs/libspeex.so.1.5.0:    ELF 32-bit LSB shared object, ARM, version 1 (SYSV), dynamically linked, not stripped
>>>> build$
>>>> 
>>>> 
>>>> Case-2 (suggested configuration)
>>>> ======
>>>> 
>>>> EXTRA_OECONF = " --with-ogg-libraries=${STAGING_LIBDIR} \
>>>>              --enable-float-api --disable-vbr \
>>>>              --with-ogg-includes=${STAGING_INCDIR} --disable-oggtest"
>>> 
>>> Or case 3, which I implemented 3 years ago in https://github.com/openembedded/openembedded/commit/e06553979d23531397af3dd71870abb80718c681 :
>>> 
>>> def get_speex_fpu_setting(bb, d):
>>>      if bb.data.getVar('TARGET_FPU', d, 1) in [ 'soft' ]:
>>>              return "--enable-fixed-point --disable-float-api --disable-vbr"
>>>      return ""
>>> 
>>> 
>>> EXTRA_OECONF += "${@get_speex_fpu_setting(bb, d)}
>>> 
>>> That OE-classic recipe also has support for arm asm which speed things up a lot.
>>> 
>>> regards,
>>> 
>>> Koen
>> 
>> 
> 
>

Patch

diff --git a/meta/recipes-multimedia/pulseaudio/pulseaudio/rescale_input_to_speex_float.patch b/meta/recipes-multimedia/pulseaudio/pulseaudio/rescale_input_to_speex_float.patch
new file mode 100644
index 0000000..fe5e6a7
--- /dev/null
+++ b/meta/recipes-multimedia/pulseaudio/pulseaudio/rescale_input_to_speex_float.patch
@@ -0,0 +1,24 @@ 
+diff -Naur a/src/pulsecore/resampler.c b/src/pulsecore/resampler.c
+--- a/src/pulsecore/resampler.c	2013-12-24 18:25:08.164787401 +0500
++++ b/src/pulsecore/resampler.c	2013-12-24 18:24:43.056787026 +0500
+@@ -1347,7 +1347,7 @@
+ 
+ static void speex_resample_float(pa_resampler *r, const pa_memchunk *input, unsigned in_n_frames, pa_memchunk *output, unsigned *out_n_frames) {
+     float *in, *out;
+-    uint32_t inf = in_n_frames, outf = *out_n_frames;
++    uint32_t inf = in_n_frames, outf = *out_n_frames, i;
+ 
+     pa_assert(r);
+     pa_assert(input);
+@@ -1357,6 +1357,11 @@
+     in = pa_memblock_acquire_chunk(input);
+     out = pa_memblock_acquire_chunk(output);
+ 
++    /* Speex float API scale range is +/-32768 instead of +/-1.
++       So rescale input before passing it to Speex. */
++    for (i = 0; i < inf; i++)
++        in[i] = 32768.*in[i];
++
+     pa_assert_se(speex_resampler_process_interleaved_float(r->speex.state, in, &inf, out, &outf) == 0);
+ 
+     pa_memblock_release(input->memblock);
diff --git a/meta/recipes-multimedia/pulseaudio/pulseaudio_4.0.bb b/meta/recipes-multimedia/pulseaudio/pulseaudio_4.0.bb
index b419c54..5afa20b 100644
--- a/meta/recipes-multimedia/pulseaudio/pulseaudio_4.0.bb
+++ b/meta/recipes-multimedia/pulseaudio/pulseaudio_4.0.bb
@@ -1,7 +1,8 @@ 
 require pulseaudio.inc
 
 SRC_URI = "http://freedesktop.org/software/pulseaudio/releases/pulseaudio-${PV}.tar.xz \
-           file://volatiles.04_pulse"
+           file://volatiles.04_pulse \
+           file://rescale_input_to_speex_float.patch"
 
 SRC_URI[md5sum] = "591f211db2790a7e4d222f2dc6858db3"
 SRC_URI[sha256sum] = "35ceb36bb1822fe54f0b5e4863b4f486769fdfb8ff2111f01fd8778928f9cdae"