On 23. 08. 23 18:29, Amadeusz Sławiński wrote:
On 8/23/2023 3:42 PM, Jaroslav Kysela wrote:
On 23. 08. 23 13:08, Takashi Iwai wrote:
On Wed, 23 Aug 2023 12:47:33 +0200, Jaroslav Kysela wrote:
On 23. 08. 23 12:20, Amadeusz Sławiński wrote:
On 8/23/2023 12:00 PM, Jaroslav Kysela wrote:
On 23. 08. 23 11:53, Takashi Iwai wrote: > On Wed, 23 Aug 2023 11:10:38 +0200, > Jaroslav Kysela wrote: >> >> On 23. 08. 23 10:11, Cezary Rojewski wrote: >>> On 2023-08-22 9:03 PM, Jaroslav Kysela wrote: >>>> On 22. 08. 23 17:38, Takashi Iwai wrote: >>>>> On Tue, 22 Aug 2023 17:29:47 +0200, >>>>> Jaroslav Kysela wrote: >>>>>> >>>>>> On 22. 08. 23 17:07, Takashi Iwai wrote: >>>>>>> On Tue, 22 Aug 2023 17:03:02 +0200, >>>>>>> Jaroslav Kysela wrote: >>>>>>>> >>>>>>>> On 11. 08. 23 18:48, Cezary Rojewski wrote: >>>>>>>> >>>>>>>>> +#define SNDRV_PCM_SUBFMTBIT_MSBITS_32 >>>>>>>>> _SNDRV_PCM_SUBFMTBIT(MSBITS_32) >>>>>>>> >>>>>>>> What was reason to add 32/32 format ? Subformat STD + >>>>>>>> msbits == 32 >>>>>>>> should already handle the maximal resolution. Until we do not >>>>>>>> have 64 >>>>>>>> bit formats, it seems like an useless extension. >>>>>>> >>>>>>> My understanding is to distinguish the cases "we do fully >>>>>>> support >>>>>>> 32bit" and "we don't care". But, the end effect is same for >>>>>>> both, >>>>>>> user-space would handle 32bit in both cases, so this difference >>>>>>> won't >>>>>>> help much, indeed. >>>>>> >>>>>> I don't think that we have a "do not care" situation. The >>>>>> applications >>>>>> currently expects to use the maximal msbits for STD >>>>>> subformat. The >>>>>> subformat should be used only to refine (downgrade) the >>>>>> resolution on >>>>>> the driver / hw side on demand. I would just add only >>>>>> necessary API >>>>>> extensions and save one bit for now. >>>>> >>>>> Well, the current behavior (with STD) is to choose whatever 32bit >>>>> format the driver supports, and the driver may set a different >>>>> value >>>>> of hw_params.msbits at hw_params. The explicit MSBITS_32 would >>>>> enforce the hw_params.msbits to be 32, otherwise hw_refine would >>>>> fail. So I see a potential difference. >>>> >>>> I see. But if our target is to create a complete query/set >>>> msbits API, >>>> we should cover all cases also for other formats. >>>> >>>> I vote to replace SUBFMTBIT_MSBITS_32 to SUBFMTBIT_MSBITS_MAX >>>> as the >>>> second bit (right after STD). The format hw parameter already >>>> defines >>>> the maximal width. We can add SUBFMTBIT_MSBITS_32 when it's really >>>> required. Note that MAX should be handled for all cases (not >>>> only for >>>> S32_LE or so). >>> >>> In my opinion STD already states "max". The word is not explicit >>> either >>> - max in the eyes of whom? The driver'? Then the driver may >>> reply: max >>> allowed e.g.: 24/32. And that translates to: fallback to STD. >> >> Max in the contents of the physical sample format (S32 = 32 bits, >> S24 >> = 24 bits, S8 = 8 bits etc). It would mean, if the driver >> supports S32 >> but only with 24-bit resolution, this bit should not be >> set/allowed. We can also use word full or something other. If we >> like >> to extend the API in this way (force the specific msbits with the >> error handling), all formats should be covered. For STD - see >> Takashi's reply. > > I think MAX can be problematic when the device supports multiple > formats, say, 16bit and 32bit. Then it's not clear which MAX points > to: is 16bit max or 32bit max.
I don't take this point. The subformat depends on the format, thus if one format support max, it should be set for queries.
Theoretically, this problem is in this API extension proposal too. Imagine that driver/hw support S24 and S32 formats and 20-bit msbits for one of them. How do you handle this? The subformat depends on format and should be refined when the format is known (single choice).
> I find the subformat extension OK, as this doesn't need much > change in > API. OTOH, if we want to be more consistent way, we may extend > hw_params for a new interval, e.g. SNDRV_PCM_HW_PARAM_MSBITS, and let > the driver choosing it. This will need more hw_params rules and > become more complex, but it allows drivers really exotic setups (like > 19bit PCM :) But my gut feeling is that the subformat extension > should suffice.
I'm not ok with 32 == 32. We should handle this case universally or discard.
Jaroslav
The reason for MSBITS32 is that, when only MSBITS_20 and MSBITS_24 are defined, when userspace (in this case aplay) asks for usable formats and subformat it gets something like:
ACCESS: MMAP_INTERLEAVED RW_INTERLEAVED FORMAT: S16_LE S32_LE SUBFORMAT: STD MSBITS_20 MSBITS_24 SAMPLE_BITS: [16 32] FRAME_BITS: [32 64] CHANNELS: 2 RATE: 48000
when MSBITS_32 is not defined it is not clear if HW supports 24 or 32 bits as maximum possible value of msbits. However when MSBITS_32 is defined it is clear - in the above case maximum possible bps then is 24, because MSBITS_32 is missing in output.
STD behaves as before and takes maximum possible value - in above case it is 24.
This example is nice for S32_LE but not S16_LE. With the max/full bit, we can already cover also S16_LE (so that the application can ask for the maximal msbits which fits to the physical format for S16_LE). It would be also a preparation for future, when we need to deal with more msbits combos (like 14bit or 15bit samples stored in the bigger physical words) etc.
So application can set those parameters for in your case:
S16_LE + STD (maximum bits handled by driver - 16 in this case) S16_LE + MSBITS_MAX (maximum physical bits for the format - 16) S32_LE + STD (maximum bits handled by driver - 24 in this case) S32_LE + MSBITS_MAX (maximum physical bits for the format - 32) S32_LE + MSBITS_24 S32_LE + MSBITS_20
Dtto for other format like S8, S24 etc. Another way is to define MSBITS_8, MSBITS_16 etc. But I'd prefer to save subformat bits. The MSBITS_MAX would cover almost all cases for now.
It becomes a bit tricky if we have a device that has 24bit bps on 32bit format and 16bit bps / 16bit format. Both formats and subformats are bitmaps in hw_params, and initially formats bitmap is 16|24 and subformats bitmap is MAX|24.
Now, suppose that app determines to use 16bit. Then we need to we need to update subformats bitmap to MAX by dropping 24. OTOH, if app chooses 32bit format, subformats will be 24 by dropping MAX, as we don't support 32 bps. And, it's not so trivial to achieve this commonly only with the single subformats bitmap of snd_pcm_hardware, as the meaning of MAX depends on the chosen format.
It's easy to implement using a code which will go through all set format bits and do bitwise OR for all corresponding subformat bits. We can use a callback which the driver may override.
Something like (for the above HDA example): - SUBFMTBIT_STD is required so it would be handled at the upper level):
snd_pcm_subformat_t (*get_subformat)(struct snd_pcm_substream *substream, snd_pcm_format_t format) { switch (format) { case SNDRV_PCM_FMTBIT_S16_LE: return SNDRV_PCM_SUBFMTBIT_MSBITS_MAX; case SNDRV_PCM_FMTBIT_S32_LE: return SNDRV_PCM_SUBFMTBIT_MSBITS_24 | SNDRV_PCM_SUBFMTBIT_MSBITS_20; default: return 0; } }
SUBFMTBIT_STD is required so it would be handled at the upper level.
Meanwhile, if the subformats bitmap is with explicit bit flags, i.e. 24|16, we can reduce the bitmap easily depending on the format.
I don't think that the current proposed code does any reduction. It will return MSBITS_24 when only S16_LE format is selected, too. The refining mechanism is ignored. We need to handle subformat bits differently than format bits because the dependency.
Jaroslav
Problem with MSBITS_MAX is that if kernel reports something like this:
FORMAT: S16_LE S32_LE SUBFORMAT: STD MSBITS_20 MSBITS_MAX
to userspace, is that userspace can't really tell if you should only apply it to S16_LE or to S32_LE, or both. On the other hand if at some point someone adds S64_LE format, something like:
Unfortunately, you've not got the point that the subformat contents depends on the selected format. So the subformat mask is for ALL formats selected in the configuration space. The only valid contents for one format is when application or kernel reduces the format to single one. And applications can do:
1) set format to S32_LE 2) call refine 3) get subformat bits for single S32_LE format from the refined cfg space
In this case, queries and specific msbits selection will work. It's the standard refine mechanism which works also for all other fields from the parameter configuration space etc. If you look to all other fields from the parameter configuration space, you cannot predict the exact parameters (buffer size, period size, channels) until you do more refining to set all parameters to exact values (single value).
In other words, the above example:
FORMAT: S16_LE S32_LE SUBFORMAT: STD MSBITS_20 MSBITS_MAX
.. means - at least one format supports maximal msbits for the given format.
Jaroslav