diff mbox

[alsa-lib,3/4] pcm: linear, route: handle linear formats with 20-bit sample on 4 bytes

Message ID 06db78be-2d91-c996-5a5b-d408ff4a84fa@maciej.szmigiero.name (mailing list archive)
State New, archived
Headers show

Commit Message

Maciej S. Szmigiero Nov. 29, 2017, 10:17 p.m. UTC
The previous patch has added 20-bit PCM formats SND_PCM_FORMAT_{S,U}20 to
alsa-lib.
We need to extend the linear format conversion code with handling of these
sample formats so they can also be utilized by applications that only
recognize the more typical ones like SND_PCM_FORMAT_S16.

Since the conversion arrays are indexed by a format bit width divided by 8
the easiest way to handle these formats is to treat them like they were
40-bit wide (the next free integer multiple of 8 bits).
This doesn't create a collision risk with a future format since there can't
really be a 40-bit sample format that occupies 4 bytes.

Make sure we use the getput conversion method for these formats since a
direct conversion from / to them is not supported.

While we are at it let's also add asserts on a format bit width value
in snd_pcm_linear_get_index() and snd_pcm_linear_put_index() received from
snd_pcm_format_width() so we won't try to access memory outside a
conversion array if for some reason this function had returned a value that
is not in the expected range.

Signed-off-by: Maciej S. Szmigiero <mail@maciej.szmigiero.name>
---
 src/pcm/pcm_linear.c | 16 +++++++++++++---
 src/pcm/pcm_route.c  |  6 ++++--
 src/pcm/plugin_ops.h | 50 ++++++++++++++++++++++++++++++++++++++++++++++----
 3 files changed, 63 insertions(+), 9 deletions(-)

Comments

Takashi Sakamoto Dec. 5, 2017, 5:50 a.m. UTC | #1
Hi,

On Nov 30 2017 07:17, Maciej S. Szmigiero wrote:
> The previous patch has added 20-bit PCM formats SND_PCM_FORMAT_{S,U}20 to
> alsa-lib.
> We need to extend the linear format conversion code with handling of these
> sample formats so they can also be utilized by applications that only
> recognize the more typical ones like SND_PCM_FORMAT_S16.
> 
> Since the conversion arrays are indexed by a format bit width divided by 8
> the easiest way to handle these formats is to treat them like they were
> 40-bit wide (the next free integer multiple of 8 bits).
> This doesn't create a collision risk with a future format since there can't
> really be a 40-bit sample format that occupies 4 bytes.
> 
> Make sure we use the getput conversion method for these formats since a
> direct conversion from / to them is not supported.
> 
> While we are at it let's also add asserts on a format bit width value
> in snd_pcm_linear_get_index() and snd_pcm_linear_put_index() received from
> snd_pcm_format_width() so we won't try to access memory outside a
> conversion array if for some reason this function had returned a value that
> is not in the expected range.
> 
> Signed-off-by: Maciej S. Szmigiero <mail@maciej.szmigiero.name>
> ---
>   src/pcm/pcm_linear.c | 16 +++++++++++++---
>   src/pcm/pcm_route.c  |  6 ++++--
>   src/pcm/plugin_ops.h | 50 ++++++++++++++++++++++++++++++++++++++++++++++----
>   3 files changed, 63 insertions(+), 9 deletions(-)
> 
> diff --git a/src/pcm/pcm_linear.c b/src/pcm/pcm_linear.c
> index 0d61e927323f..e4973a308004 100644
> --- a/src/pcm/pcm_linear.c
> +++ b/src/pcm/pcm_linear.c
> @@ -90,6 +90,7 @@ int snd_pcm_linear_get_index(snd_pcm_format_t src_format, snd_pcm_format_t dst_f
>   		endian = 0;
>   	pwidth = snd_pcm_format_physical_width(src_format);
>   	width = snd_pcm_format_width(src_format);
> +	assert(width >= 8 && width <= 32);
> 
> ...
> > @@ -121,6 +125,7 @@ int snd_pcm_linear_put_index(snd_pcm_format_t 
src_format, snd_pcm_format_t dst_f
>   		endian = 0;
>   	pwidth = snd_pcm_format_physical_width(dst_format);
>   	width = snd_pcm_format_width(dst_format);
> +	assert(width >= 8 && width <= 32);
>   	if (pwidth == 24) {
>   		switch (width) {
>   		case 24:

I can understand your intention to insert these assert(), while it's 
better not to include them in this series of patch.

A call of 'snd_pcm_format_physical_width()' and 'snd_pcm_format_width()'
returns 4 for 'SNDRV_PCM_FORMAT_IMA_ADPCM'. Furthermore, both of them
returns -EINVAL for 'SNDRV_PCM_FORMAT_MPEG', 'SNDRV_PCM_FORMAT_GSM' and
'SNDRV_PCM_FORMAT_SPECIAL'. They can bring wrong calculation to the
resampling function, in the worse case result in any fault.

However, this is another issue, at least out of your aim for this
patchset. It's a kind of bug for wider influence. I think to have
another opportunity to discuss about the way to handle such special
formats in the resampling layer. Would you please post this patchset
again without these assertions?

(Additionally, I suspect that current implementation of the resampling
layer itself is not enough proper for non-PCM samples such as float, DSD
and so on.)

And next time, please follow a below instruction.
1.use '--subject-prefix' option with 'alsa-lib][PATCH v3' for
   git-format-patch so that subject of your patches includes better
   information for reviewers.
   (If it's for kernel stuffs, just use '--v')
2. use '--cover-letter' option for git-format-patch so that you can
   write changelog of your successive work.
3.use '--thread' option for git-send-email (or add 'sendemail.thread'
   entry into your local repository) so that posted patches spin in one
   message thread.


Thanks

Takashi Sakamoto
Maciej S. Szmigiero Dec. 5, 2017, 1:22 p.m. UTC | #2
Hi Takashi,

On 05.12.2017 06:50, Takashi Sakamoto wrote:
> Hi,
> 
> On Nov 30 2017 07:17, Maciej S. Szmigiero wrote:
(..)
>> diff --git a/src/pcm/pcm_linear.c b/src/pcm/pcm_linear.c
>> index 0d61e927323f..e4973a308004 100644
>> --- a/src/pcm/pcm_linear.c
>> +++ b/src/pcm/pcm_linear.c
>> @@ -90,6 +90,7 @@ int snd_pcm_linear_get_index(snd_pcm_format_t src_format, snd_pcm_format_t dst_f
>>           endian = 0;
>>       pwidth = snd_pcm_format_physical_width(src_format);
>>       width = snd_pcm_format_width(src_format);
>> +    assert(width >= 8 && width <= 32);
>>
>> ...
>> > @@ -121,6 +125,7 @@ int snd_pcm_linear_put_index(snd_pcm_format_t 
> src_format, snd_pcm_format_t dst_f
>>           endian = 0;
>>       pwidth = snd_pcm_format_physical_width(dst_format);
>>       width = snd_pcm_format_width(dst_format);
>> +    assert(width >= 8 && width <= 32);
>>       if (pwidth == 24) {
>>           switch (width) {
>>           case 24:
> 
> I can understand your intention to insert these assert(), while it's better not to include them in this series of patch.
> 
> A call of 'snd_pcm_format_physical_width()' and 'snd_pcm_format_width()'
> returns 4 for 'SNDRV_PCM_FORMAT_IMA_ADPCM'. Furthermore, both of them
> returns -EINVAL for 'SNDRV_PCM_FORMAT_MPEG', 'SNDRV_PCM_FORMAT_GSM' and
> 'SNDRV_PCM_FORMAT_SPECIAL'. They can bring wrong calculation to the
> resampling function, in the worse case result in any fault.

snd_pcm_linear_get_index() and snd_pcm_linear_put_index() should
be called for linear formats only (like their names suggest).

If they were called for 'SNDRV_PCM_FORMAT_IMA_ADPCM' this will result
in a negative conversion arrays index.
The same goes for formats that return '-EINVAL' as their width.

This will mean that an application will try to access these arrays
outside of their bounds, which may produce a hard to find bugs.
I think it is really better to fail with an assertion failure in such
cases.

> However, this is another issue, at least out of your aim for this
> patchset. It's a kind of bug for wider influence. I think to have
> another opportunity to discuss about the way to handle such special
> formats in the resampling layer. Would you please post this patchset
> again without these assertions?

If these assertions are really controversial I will remove it, but please
think again about it in the light of the paragraph above.
Asserts can also be split out to a separate patch.

> (Additionally, I suspect that current implementation of the resampling
> layer itself is not enough proper for non-PCM samples such as float, DSD
> and so on.)
> 
> And next time, please follow a below instruction.
> 1.use '--subject-prefix' option with 'alsa-lib][PATCH v3' for
>   git-format-patch so that subject of your patches includes better
>   information for reviewers.
>   (If it's for kernel stuffs, just use '--v')

Hmm, but this series already had this prefix ('[alsa-lib][PATCH]')?
Had no version number, since it was the first iteration.

> 2. use '--cover-letter' option for git-format-patch so that you can
>   write changelog of your successive work.

Will do.

> 3.use '--thread' option for git-send-email (or add 'sendemail.thread'
>   entry into your local repository) so that posted patches spin in one
>   message thread.
> 

Will do.

Also, will add commas at the ends of last entries in enums, as you had
suggested in your other message.

> Thanks
> 
> Takashi Sakamoto

Thanks,
Maciej
Takashi Sakamoto Dec. 5, 2017, 2:37 p.m. UTC | #3
Hi,

On Dec 5 2017 22:22, Maciej S. Szmigiero wrote:
>>>> @@ -121,6 +125,7 @@ int snd_pcm_linear_put_index(snd_pcm_format_t
>> src_format, snd_pcm_format_t dst_f
>>>            endian = 0;
>>>        pwidth = snd_pcm_format_physical_width(dst_format);
>>>        width = snd_pcm_format_width(dst_format);
>>> +    assert(width >= 8 && width <= 32);
>>>        if (pwidth == 24) {
>>>            switch (width) {
>>>            case 24:
>>
>> I can understand your intention to insert these assert(), while it's better not to include them in this series of patch.
>>
>> A call of 'snd_pcm_format_physical_width()' and 'snd_pcm_format_width()'
>> returns 4 for 'SNDRV_PCM_FORMAT_IMA_ADPCM'. Furthermore, both of them
>> returns -EINVAL for 'SNDRV_PCM_FORMAT_MPEG', 'SNDRV_PCM_FORMAT_GSM' and
>> 'SNDRV_PCM_FORMAT_SPECIAL'. They can bring wrong calculation to the
>> resampling function, in the worse case result in any fault.
> 
> snd_pcm_linear_get_index() and snd_pcm_linear_put_index() should
> be called for linear formats only (like their names suggest).
> 
> If they were called for 'SNDRV_PCM_FORMAT_IMA_ADPCM' this will result
> in a negative conversion arrays index.
> The same goes for formats that return '-EINVAL' as their width.
> 
> This will mean that an application will try to access these arrays
> outside of their bounds, which may produce a hard to find bugs.
> I think it is really better to fail with an assertion failure in such
> cases.
> 
>> However, this is another issue, at least out of your aim for this
>> patchset. It's a kind of bug for wider influence. I think to have
>> another opportunity to discuss about the way to handle such special
>> formats in the resampling layer. Would you please post this patchset
>> again without these assertions?
> 
> If these assertions are really controversial I will remove it, but please
> think again about it in the light of the paragraph above.
> Asserts can also be split out to a separate patch.

Your have correct understanding of this issue. However, it's better for
us to achieve your aim (addition of the new pcm formats) at first. Then
we start discussion about better solution to handle the non-Linear
formats.

In my opinion, addition of constraint of PCM hardware parameters to
these PCM plugins is preferable to insertion of the assertions, because 
abort inner library is bad idea itself. Proper errors should be returned
to applications in protocol via alsa-lib PCM APIs. It's more friendly to
usual developers and users.

>> (Additionally, I suspect that current implementation of the resampling
>> layer itself is not enough proper for non-PCM samples such as float, DSD
>> and so on.)
>>
>> And next time, please follow a below instruction.
>> 1.use '--subject-prefix' option with 'alsa-lib][PATCH v3' for
>>    git-format-patch so that subject of your patches includes better
>>    information for reviewers.
>>    (If it's for kernel stuffs, just use '--v')
> 
> Hmm, but this series already had this prefix ('[alsa-lib][PATCH]')?
> Had no version number, since it was the first iteration.

This patchset[1] should have v2, because I did comment to your first
version[2]. So the next version should have v3.

>> 2. use '--cover-letter' option for git-format-patch so that you can
>>    write changelog of your successive work.
> 
> Will do.
> 
>> 3.use '--thread' option for git-send-email (or add 'sendemail.thread'
>>    entry into your local repository) so that posted patches spin in one
>>    message thread.
>>
> 
> Will do.
> 
> Also, will add commas at the ends of last entries in enums, as you had
> suggested in your other message.

OK. Thanks.

[1] 
http://mailman.alsa-project.org/pipermail/alsa-devel/2017-November/128335.html
http://mailman.alsa-project.org/pipermail/alsa-devel/2017-November/128336.html
http://mailman.alsa-project.org/pipermail/alsa-devel/2017-November/128337.html
http://mailman.alsa-project.org/pipermail/alsa-devel/2017-November/128338.html
[2] 
http://mailman.alsa-project.org/pipermail/alsa-devel/2017-November/128278.html

Takashi Sakamoto
Takashi Iwai Dec. 5, 2017, 2:57 p.m. UTC | #4
On Tue, 05 Dec 2017 15:37:09 +0100,
Takashi Sakamoto wrote:
> 
> Hi,
> 
> On Dec 5 2017 22:22, Maciej S. Szmigiero wrote:
> >>>> @@ -121,6 +125,7 @@ int snd_pcm_linear_put_index(snd_pcm_format_t
> >> src_format, snd_pcm_format_t dst_f
> >>>            endian = 0;
> >>>        pwidth = snd_pcm_format_physical_width(dst_format);
> >>>        width = snd_pcm_format_width(dst_format);
> >>> +    assert(width >= 8 && width <= 32);
> >>>        if (pwidth == 24) {
> >>>            switch (width) {
> >>>            case 24:
> >>
> >> I can understand your intention to insert these assert(), while it's better not to include them in this series of patch.
> >>
> >> A call of 'snd_pcm_format_physical_width()' and 'snd_pcm_format_width()'
> >> returns 4 for 'SNDRV_PCM_FORMAT_IMA_ADPCM'. Furthermore, both of them
> >> returns -EINVAL for 'SNDRV_PCM_FORMAT_MPEG', 'SNDRV_PCM_FORMAT_GSM' and
> >> 'SNDRV_PCM_FORMAT_SPECIAL'. They can bring wrong calculation to the
> >> resampling function, in the worse case result in any fault.
> >
> > snd_pcm_linear_get_index() and snd_pcm_linear_put_index() should
> > be called for linear formats only (like their names suggest).
> >
> > If they were called for 'SNDRV_PCM_FORMAT_IMA_ADPCM' this will result
> > in a negative conversion arrays index.
> > The same goes for formats that return '-EINVAL' as their width.
> >
> > This will mean that an application will try to access these arrays
> > outside of their bounds, which may produce a hard to find bugs.
> > I think it is really better to fail with an assertion failure in such
> > cases.
> >
> >> However, this is another issue, at least out of your aim for this
> >> patchset. It's a kind of bug for wider influence. I think to have
> >> another opportunity to discuss about the way to handle such special
> >> formats in the resampling layer. Would you please post this patchset
> >> again without these assertions?
> >
> > If these assertions are really controversial I will remove it, but please
> > think again about it in the light of the paragraph above.
> > Asserts can also be split out to a separate patch.
> 
> Your have correct understanding of this issue. However, it's better for
> us to achieve your aim (addition of the new pcm formats) at first. Then
> we start discussion about better solution to handle the non-Linear
> formats.
> 
> In my opinion, addition of constraint of PCM hardware parameters to
> these PCM plugins is preferable to insertion of the assertions,
> because abort inner library is bad idea itself. Proper errors should
> be returned
> to applications in protocol via alsa-lib PCM APIs. It's more friendly to
> usual developers and users.

Well, it's basically moot to discuss about it, as
snd_pcm_lib_linear_get_index() is an internal function, i.e. it's not
supposed to be called from applications.

In that sense, assert() wouldn't be too bad.   However, the primary
question is -- would it really catch anything in practice?  What drove
you to put that assert() at all?

If a check of the valid formats, then calling snd_pcm_format_linear()
is the safest.


thanks,

Takashi
Maciej S. Szmigiero Dec. 5, 2017, 4:14 p.m. UTC | #5
On 05.12.2017 15:57, Takashi Iwai wrote:
> On Tue, 05 Dec 2017 15:37:09 +0100,
> Takashi Sakamoto wrote:
>>
>> Hi,
>>
(..)
>> In my opinion, addition of constraint of PCM hardware parameters to
>> these PCM plugins is preferable to insertion of the assertions,
>> because abort inner library is bad idea itself. Proper errors should
>> be returned
>> to applications in protocol via alsa-lib PCM APIs. It's more friendly to
>> usual developers and users.
> 
> Well, it's basically moot to discuss about it, as
> snd_pcm_lib_linear_get_index() is an internal function, i.e. it's not
> supposed to be called from applications.
> 
> In that sense, assert() wouldn't be too bad.   However, the primary
> question is -- would it really catch anything in practice?  What drove
> you to put that assert() at all?
> 
> If a check of the valid formats, then calling snd_pcm_format_linear()
> is the safest.

The reason I added these asserts is that snd_pcm_linear_{get,put}_index()
use result from snd_pcm_format_width() directly to calculate an array
index.
If, for any reason, this function returns an unexpected width a subtle
bug will emerge.

Three reasons that I could think of why an unexpected width would be
returned is a bug in a plugin that uses these functions (currently 9
plugins use them) that fail to properly restrict source / destination
format space to linear formats only, a bug in general format refining
code or somebody adding a format to linear formats mask without
adapting conversion arrays, too.

I know that these may sound like rare occurrences but since this
check is close to zero-cost (it only happens on plugin stack
initialization time and not, for example, at every sample) I think it
is worth the extra time (and one can always compile the library with
NDEBUG to skip it).

> thanks,
> 
> Takashi
> 

Thanks,
Maciej
Takashi Iwai Dec. 5, 2017, 4:28 p.m. UTC | #6
On Tue, 05 Dec 2017 17:14:18 +0100,
Maciej S. Szmigiero wrote:
> 
> On 05.12.2017 15:57, Takashi Iwai wrote:
> > On Tue, 05 Dec 2017 15:37:09 +0100,
> > Takashi Sakamoto wrote:
> >>
> >> Hi,
> >>
> (..)
> >> In my opinion, addition of constraint of PCM hardware parameters to
> >> these PCM plugins is preferable to insertion of the assertions,
> >> because abort inner library is bad idea itself. Proper errors should
> >> be returned
> >> to applications in protocol via alsa-lib PCM APIs. It's more friendly to
> >> usual developers and users.
> > 
> > Well, it's basically moot to discuss about it, as
> > snd_pcm_lib_linear_get_index() is an internal function, i.e. it's not
> > supposed to be called from applications.
> > 
> > In that sense, assert() wouldn't be too bad.   However, the primary
> > question is -- would it really catch anything in practice?  What drove
> > you to put that assert() at all?
> > 
> > If a check of the valid formats, then calling snd_pcm_format_linear()
> > is the safest.
> 
> The reason I added these asserts is that snd_pcm_linear_{get,put}_index()
> use result from snd_pcm_format_width() directly to calculate an array
> index.
> If, for any reason, this function returns an unexpected width a subtle
> bug will emerge.
> 
> Three reasons that I could think of why an unexpected width would be
> returned is a bug in a plugin that uses these functions (currently 9
> plugins use them) that fail to properly restrict source / destination
> format space to linear formats only, a bug in general format refining
> code or somebody adding a format to linear formats mask without
> adapting conversion arrays, too.
> 
> I know that these may sound like rare occurrences but since this
> check is close to zero-cost (it only happens on plugin stack
> initialization time and not, for example, at every sample) I think it
> is worth the extra time (and one can always compile the library with
> NDEBUG to skip it).

So it has nothing to do with the 20bit PCM support itself.
Please put in another patch with more contexts.


thanks,

Takashi
diff mbox

Patch

diff --git a/src/pcm/pcm_linear.c b/src/pcm/pcm_linear.c
index 0d61e927323f..e4973a308004 100644
--- a/src/pcm/pcm_linear.c
+++ b/src/pcm/pcm_linear.c
@@ -90,6 +90,7 @@  int snd_pcm_linear_get_index(snd_pcm_format_t src_format, snd_pcm_format_t dst_f
 		endian = 0;
 	pwidth = snd_pcm_format_physical_width(src_format);
 	width = snd_pcm_format_width(src_format);
+	assert(width >= 8 && width <= 32);
 	if (pwidth == 24) {
 		switch (width) {
 		case 24:
@@ -100,8 +101,11 @@  int snd_pcm_linear_get_index(snd_pcm_format_t src_format, snd_pcm_format_t dst_f
 		default:
 			width = 2; break;
 		}
-		return width * 4 + endian * 2 + sign + 16;
+		return width * 4 + endian * 2 + sign + 20;
 	} else {
+		if (width == 20)
+			width = 40;
+
 		width = width / 8 - 1;
 		return width * 4 + endian * 2 + sign;
 	}
@@ -121,6 +125,7 @@  int snd_pcm_linear_put_index(snd_pcm_format_t src_format, snd_pcm_format_t dst_f
 		endian = 0;
 	pwidth = snd_pcm_format_physical_width(dst_format);
 	width = snd_pcm_format_width(dst_format);
+	assert(width >= 8 && width <= 32);
 	if (pwidth == 24) {
 		switch (width) {
 		case 24:
@@ -131,8 +136,11 @@  int snd_pcm_linear_put_index(snd_pcm_format_t src_format, snd_pcm_format_t dst_f
 		default:
 			width = 2; break;
 		}
-		return width * 4 + endian * 2 + sign + 16;
+		return width * 4 + endian * 2 + sign + 20;
 	} else {
+		if (width == 20)
+			width = 40;
+
 		width = width / 8 - 1;
 		return width * 4 + endian * 2 + sign;
 	}
@@ -303,7 +311,9 @@  static int snd_pcm_linear_hw_params(snd_pcm_t *pcm, snd_pcm_hw_params_t *params)
 	if (err < 0)
 		return err;
 	linear->use_getput = (snd_pcm_format_physical_width(format) == 24 ||
-			      snd_pcm_format_physical_width(linear->sformat) == 24);
+			      snd_pcm_format_physical_width(linear->sformat) == 24 ||
+			      snd_pcm_format_width(format) == 20 ||
+			      snd_pcm_format_width(linear->sformat) == 20);
 	if (linear->use_getput) {
 		if (pcm->stream == SND_PCM_STREAM_PLAYBACK) {
 			linear->get_idx = snd_pcm_linear_get_index(format, SND_PCM_FORMAT_S32);
diff --git a/src/pcm/pcm_route.c b/src/pcm/pcm_route.c
index 1f485a8e79f9..bbcc6118b593 100644
--- a/src/pcm/pcm_route.c
+++ b/src/pcm/pcm_route.c
@@ -565,10 +565,12 @@  static int snd_pcm_route_hw_params(snd_pcm_t *pcm, snd_pcm_hw_params_t * params)
 	}
 	if (err < 0)
 		return err;
-	/* 3 bytes formats? */
+	/* 3 bytes or 20-bit formats? */
 	route->params.use_getput =
 		(snd_pcm_format_physical_width(src_format) + 7) / 8 == 3 ||
-		(snd_pcm_format_physical_width(dst_format) + 7) / 8 == 3;
+		(snd_pcm_format_physical_width(dst_format) + 7) / 8 == 3 ||
+		snd_pcm_format_width(src_format) == 20 ||
+		snd_pcm_format_width(dst_format) == 20;
 	route->params.get_idx = snd_pcm_linear_get_index(src_format, SND_PCM_FORMAT_S32);
 	route->params.put_idx = snd_pcm_linear_put_index(SND_PCM_FORMAT_S32, dst_format);
 	route->params.conv_idx = snd_pcm_linear_convert_index(src_format, dst_format);
diff --git a/src/pcm/plugin_ops.h b/src/pcm/plugin_ops.h
index 5a6016adb13a..a784d9c2b320 100644
--- a/src/pcm/plugin_ops.h
+++ b/src/pcm/plugin_ops.h
@@ -21,6 +21,12 @@ 
 
 #ifndef SX_INLINES
 #define SX_INLINES
+static inline uint32_t sx20(uint32_t x)
+{
+	if(x&0x00080000)
+		return x|0xFFF00000;
+	return x&0x000FFFFF;
+}
 static inline uint32_t sx24(uint32_t x)
 {
 	if(x&0x00800000)
@@ -341,7 +347,7 @@  conv_1234_123C: as_u32(dst) = as_u32c(src) ^ 0x80; goto CONV_END;
 
 #ifdef GET16_LABELS
 /* src_wid src_endswap sign_toggle */
-static void *const get16_labels[4 * 2 * 2 + 4 * 3] = {
+static void *const get16_labels[5 * 2 * 2 + 4 * 3] = {
 	&&get16_1_10,	 /*  8h -> 16h */
 	&&get16_1_90,	 /*  8h ^> 16h */
 	&&get16_1_10,	 /*  8s -> 16h */
@@ -350,6 +356,7 @@  static void *const get16_labels[4 * 2 * 2 + 4 * 3] = {
 	&&get16_12_92,	 /* 16h ^> 16h */
 	&&get16_12_21,	 /* 16s -> 16h */
 	&&get16_12_A1,	 /* 16s ^> 16h */
+	/* 4 byte formats */
 	&&get16_0123_12, /* 24h -> 16h */
 	&&get16_0123_92, /* 24h ^> 16h */
 	&&get16_1230_32, /* 24s -> 16h */
@@ -358,6 +365,10 @@  static void *const get16_labels[4 * 2 * 2 + 4 * 3] = {
 	&&get16_1234_92, /* 32h ^> 16h */
 	&&get16_1234_43, /* 32s -> 16h */
 	&&get16_1234_C3, /* 32s ^> 16h */
+	&&get16_0123_12_20, /* 20h -> 16h */
+	&&get16_0123_92_20, /* 20h ^> 16h */
+	&&get16_1230_32_20, /* 20s -> 16h */
+	&&get16_1230_B2_20, /* 20s ^> 16h */
 	/* 3bytes format */
 	&&get16_123_12,	 /* 24h -> 16h */
 	&&get16_123_92,	 /* 24h ^> 16h */
@@ -390,6 +401,10 @@  get16_1234_12: sample = as_u32c(src) >> 16; goto GET16_END;
 get16_1234_92: sample = (as_u32c(src) >> 16) ^ 0x8000; goto GET16_END;
 get16_1234_43: sample = bswap_16(as_u32c(src)); goto GET16_END;
 get16_1234_C3: sample = bswap_16(as_u32c(src) ^ 0x80); goto GET16_END;
+get16_0123_12_20: sample = as_u32c(src) >> 4; goto GET16_END;
+get16_0123_92_20: sample = (as_u32c(src) >> 4) ^ 0x8000; goto GET16_END;
+get16_1230_32_20: sample = bswap_32(as_u32c(src)) >> 4; goto GET16_END;
+get16_1230_B2_20: sample = (bswap_32(as_u32c(src)) >> 4) ^ 0x8000; goto GET16_END;
 get16_123_12: sample = _get_triple(src) >> 8; goto GET16_END;
 get16_123_92: sample = (_get_triple(src) >> 8) ^ 0x8000; goto GET16_END;
 get16_123_32: sample = _get_triple_s(src) >> 8; goto GET16_END;
@@ -407,7 +422,7 @@  get16_123_B2_18: sample = (_get_triple_s(src) >> 2) ^ 0x8000; goto GET16_END;
 
 #ifdef PUT16_LABELS
 /* dst_wid dst_endswap sign_toggle */
-static void *const put16_labels[4 * 2 * 2 + 4 * 3] = {
+static void *const put16_labels[5 * 2 * 2 + 4 * 3] = {
 	&&put16_12_1,		 /* 16h ->  8h */
 	&&put16_12_9,		 /* 16h ^>  8h */
 	&&put16_12_1,		 /* 16h ->  8s */
@@ -416,6 +431,7 @@  static void *const put16_labels[4 * 2 * 2 + 4 * 3] = {
 	&&put16_12_92,		 /* 16h ^> 16h */
 	&&put16_12_21,		 /* 16h -> 16s */
 	&&put16_12_29,		 /* 16h ^> 16s */
+	/* 4 byte formats */
 	&&put16_12_0120,	 /* 16h -> 24h */
 	&&put16_12_0920,	 /* 16h ^> 24h */
 	&&put16_12_0210,	 /* 16h -> 24s */
@@ -424,6 +440,10 @@  static void *const put16_labels[4 * 2 * 2 + 4 * 3] = {
 	&&put16_12_9200,	 /* 16h ^> 32h */
 	&&put16_12_0021,	 /* 16h -> 32s */
 	&&put16_12_0029,	 /* 16h ^> 32s */
+	&&put16_12_0120_20,	 /* 16h -> 20h */
+	&&put16_12_0920_20,	 /* 16h ^> 20h */
+	&&put16_12_0210_20,	 /* 16h -> 20s */
+	&&put16_12_0290_20,	 /* 16h ^> 20s */
 	/* 3bytes format */
 	&&put16_12_120,		 /* 16h -> 24h */
 	&&put16_12_920,		 /* 16h ^> 24h */
@@ -456,6 +476,10 @@  put16_12_1200: as_u32(dst) = (uint32_t)sample << 16; goto PUT16_END;
 put16_12_9200: as_u32(dst) = (uint32_t)(sample ^ 0x8000) << 16; goto PUT16_END;
 put16_12_0021: as_u32(dst) = (uint32_t)bswap_16(sample); goto PUT16_END;
 put16_12_0029: as_u32(dst) = (uint32_t)bswap_16(sample) ^ 0x80; goto PUT16_END;
+put16_12_0120_20: as_u32(dst) = sx20((uint32_t)sample << 4); goto PUT16_END;
+put16_12_0920_20: as_u32(dst) = sx20((uint32_t)(sample ^ 0x8000) << 4); goto PUT16_END;
+put16_12_0210_20: as_u32(dst) = bswap_32(sx20((uint32_t)sample << 4)); goto PUT16_END;
+put16_12_0290_20: as_u32(dst) = bswap_32(sx20((uint32_t)(sample ^ 0x8000) << 4)); goto PUT16_END;
 put16_12_120: _put_triple(dst, (uint32_t)sample << 8); goto PUT16_END;
 put16_12_920: _put_triple(dst, (uint32_t)(sample ^ 0x8000) << 8); goto PUT16_END;
 put16_12_021: _put_triple_s(dst, (uint32_t)sample << 8); goto PUT16_END;
@@ -478,7 +502,7 @@  put16_12_029_18: _put_triple_s(dst, (uint32_t)(sample ^ 0x8000) << 2); goto PUT1
 
 #ifdef GET32_LABELS
 /* src_wid src_endswap sign_toggle */
-static void *const get32_labels[4 * 2 * 2 + 4 * 3] = {
+static void *const get32_labels[5 * 2 * 2 + 4 * 3] = {
 	&&get32_1_1000,	 	 /*  8h -> 32h */
 	&&get32_1_9000,	 	 /*  8h ^> 32h */
 	&&get32_1_1000,		 /*  8s -> 32h */
@@ -487,6 +511,7 @@  static void *const get32_labels[4 * 2 * 2 + 4 * 3] = {
 	&&get32_12_9200,	 /* 16h ^> 32h */
 	&&get32_12_2100,	 /* 16s -> 32h */
 	&&get32_12_A100,	 /* 16s ^> 32h */
+	/* 4 byte formats */
 	&&get32_0123_1230,	 /* 24h -> 32h */
 	&&get32_0123_9230,	 /* 24h ^> 32h */
 	&&get32_1230_3210,	 /* 24s -> 32h */
@@ -495,6 +520,10 @@  static void *const get32_labels[4 * 2 * 2 + 4 * 3] = {
 	&&get32_1234_9234,	 /* 32h ^> 32h */
 	&&get32_1234_4321,	 /* 32s -> 32h */
 	&&get32_1234_C321,	 /* 32s ^> 32h */
+	&&get32_0123_1230_20,	 /* 20h -> 32h */
+	&&get32_0123_9230_20,	 /* 20h ^> 32h */
+	&&get32_1230_3210_20,	 /* 20s -> 32h */
+	&&get32_1230_B210_20,	 /* 20s ^> 32h */
 	/* 3bytes format */
 	&&get32_123_1230,	 /* 24h -> 32h */
 	&&get32_123_9230,	 /* 24h ^> 32h */
@@ -531,6 +560,10 @@  get32_1234_1234: sample = as_u32c(src); goto GET32_END;
 get32_1234_9234: sample = as_u32c(src) ^ 0x80000000; goto GET32_END;
 get32_1234_4321: sample = bswap_32(as_u32c(src)); goto GET32_END;
 get32_1234_C321: sample = bswap_32(as_u32c(src) ^ 0x80); goto GET32_END;
+get32_0123_1230_20: sample = as_u32c(src) << 12; goto GET32_END;
+get32_0123_9230_20: sample = (as_u32c(src) << 12) ^ 0x80000000; goto GET32_END;
+get32_1230_3210_20: sample = bswap_32(as_u32c(src)) << 12; goto GET32_END;
+get32_1230_B210_20: sample = (bswap_32(as_u32c(src)) << 12) ^ 0x80000000; goto GET32_END;
 get32_123_1230: sample = _get_triple(src) << 8; goto GET32_END;
 get32_123_9230: sample = (_get_triple(src) << 8) ^ 0x80000000; goto GET32_END;
 get32_123_3210: sample = _get_triple_s(src) << 8; goto GET32_END;
@@ -553,7 +586,7 @@  __conv24_get: goto *put;
 
 #ifdef PUT32_LABELS
 /* dst_wid dst_endswap sign_toggle */
-static void *const put32_labels[4 * 2 * 2 + 4 * 3] = {
+static void *const put32_labels[5 * 2 * 2 + 4 * 3] = {
 	&&put32_1234_1,	 	 /* 32h ->  8h */
 	&&put32_1234_9,	 	 /* 32h ^>  8h */
 	&&put32_1234_1,	 	 /* 32h ->  8s */
@@ -562,6 +595,7 @@  static void *const put32_labels[4 * 2 * 2 + 4 * 3] = {
 	&&put32_1234_92,	 /* 32h ^> 16h */
 	&&put32_1234_21,	 /* 32h -> 16s */
 	&&put32_1234_29,	 /* 32h ^> 16s */
+	/* 4 byte formats */
 	&&put32_1234_0123,	 /* 32h -> 24h */
 	&&put32_1234_0923,	 /* 32h ^> 24h */
 	&&put32_1234_3210,	 /* 32h -> 24s */
@@ -570,6 +604,10 @@  static void *const put32_labels[4 * 2 * 2 + 4 * 3] = {
 	&&put32_1234_9234,	 /* 32h ^> 32h */
 	&&put32_1234_4321,	 /* 32h -> 32s */
 	&&put32_1234_4329,	 /* 32h ^> 32s */
+	&&put32_1234_0123_20,	 /* 32h -> 20h */
+	&&put32_1234_0923_20,	 /* 32h ^> 20h */
+	&&put32_1234_3210_20,	 /* 32h -> 20s */
+	&&put32_1234_3290_20,	 /* 32h ^> 20s */
 	/* 3bytes format */
 	&&put32_1234_123,	 /* 32h -> 24h */
 	&&put32_1234_923,	 /* 32h ^> 24h */
@@ -607,6 +645,10 @@  put32_1234_1234: as_u32(dst) = sample; goto PUT32_END;
 put32_1234_9234: as_u32(dst) = sample ^ 0x80000000; goto PUT32_END;
 put32_1234_4321: as_u32(dst) = bswap_32(sample); goto PUT32_END;
 put32_1234_4329: as_u32(dst) = bswap_32(sample) ^ 0x80; goto PUT32_END;
+put32_1234_0123_20: as_u32(dst) = sx20(sample >> 12); goto PUT32_END;
+put32_1234_0923_20: as_u32(dst) = sx20((sample ^ 0x80000000) >> 12); goto PUT32_END;
+put32_1234_3210_20: as_u32(dst) = bswap_32(sx20(sample >> 12)); goto PUT32_END;
+put32_1234_3290_20: as_u32(dst) = bswap_32(sx20((sample ^ 0x80000000) >> 12)); goto PUT32_END;
 put32_1234_123: _put_triple(dst, sample >> 8); goto PUT32_END;
 put32_1234_923: _put_triple(dst, (sample ^ 0x80000000) >> 8); goto PUT32_END;
 put32_1234_321: _put_triple_s(dst, sample >> 8); goto PUT32_END;