On Fri, 17 Nov 2023 13:05:54 +0100, Cezary Rojewski wrote:
Patchset aims to address format selection restrictions present currently in the HDAudio library. Formats which we are concerned about are 20 and 24 valid bits per sample within 32 bit depth container. One may identify them as S20_LE and S24_LE except that those, according to comments found in include/uapi/sound/asound.h, are for LSB-aligned scenarios. HDAudio streams expect MSB-aligned data, no matter if we are speaking of HOST (SDxFMT) or LINK (PPLCxFMT) side - chapter 4.5.1 of the public HDAudio specification. In short, S20_LE and S24_LE are invalid options.
Right now, given the implementation of snd_hdac_query_supported_pcm() within sound/hda/hdac_device.c, even if a codec responds with: "I support all possible formats specified within HDAudio specification", there will be no option to open a 20/32 or 24/32 stream. The kernel will force the stream to be opened using the highest available bit depth.
After discussing subject initially with Jaroslav and Takashi, suggestion was made to utilize 'subformat' option to address the problem. The eye-opening discussion begun much earlier though, in 2019 [1].
Paired with PRs for alsa-utils [2] and alsa-lib [3].
Flow of changes:
The very first patch adds MSBITS subformat options to allow for granular 20/32, 24/32 and 32/32 format selection. The next two make sure subformat is actually honored during runtime. Most of that code is based on format-related API.
Follow up is upgrade to the hda stream-format interface - several functions are added to make the granular format selection simple in the HDAudio world. Core of the implementation is based on the existing snd_hdac_calc_stream_format(). The next ten patches are straightforward switch from one interface to another with cleanup of now-unsed function as a finishing touch.
Last but not least - the avs-driver, on which the problem analyzed and debugged, is updated to no longer acknowledge S24_LE as a valid format option.
Results with skylake-driver and snd_hda_intel show status quo on our RVPs. PR filed on SOF github shows promising results too [4].
Changes in v5:
- reworded 'bps' to 'bits' in all occurrences.
- fixed an issue of MSBITS_MAX not being reported for S32_LE as reported by Jaroslav
- snd_pcm_subformat_width() has been inlined into its only user: snd_pcm_hw_params_bits() as suggested by Jaroslav.
- updated commit messege for patch 01 as it was out of date given the recent updates.
Changes in v4:
- fixed compilation issues in sof-driver, patch 12/16, reported by ikp
- fixed sparse warnings in patch 01/16, reported by ikp
- updated commit message for patch 03: "ASoC pcm: Honor subformat when configuring runtime", as snd_pcm_hw_copy() is gone since revision v3.
Changes in v3:
- merged the first two patches as suggested by Jaroslav
- re-authored patch 01 to Jaroslav, added my Co-developed-by.
- added Jaroslav' Co-developed-by to patch 02.
- 'subformats' field now S32_LE-specific. Given the fact that it is the only format currently requiring subformat-intervention, functionality is narrowed to reduce amount of memory allocations and cleanup. Suggested by Jaroslav.
- note to the above: the hdaudio part converted 1:1 as requested, patch 02/16
- note #2: alsa part converted to S32_LE-specific yet without addition of the chicken bit. Instead, struct snd_pcm_hardware is updated with u32 subformat mask to do the job.
- ALSA-core additions in form of snd_pcm_subformat_width() and snd_pcm_hw_params_bps() relocated from 01/16 to the user, patch 05.
Changes in v2:
- patch 01/17, introduced struct snd_pcm_subformat which task is to represent subformat-mask on per format basis. Expectation is that manipulated arrays of subformats always end with a sentinel entry
- patch 01/17, added snd_pcm_hw_copy() as the copying snd_pcm_hardware becomes non-trivial
- patch 02/17, added hw_rule that produces final subformat mask based on provided formats as suggested by Jaroslav
- patch 04/17, soc_pcm_hw_update_subformat() refactored as the subformat intersection becomes non-trivial
- relevant functions releasing resources occupied by hda_pcm and snd_pcm_runtime updated to also kfree() subformats
- except for 16/17, no changes to patches past 04/17, retaining acks for these
Changes in v1:
- fixed UBSAN due to missing snd_pcm_subformat_names[] entries for new subformats
- as HDMI stream capabilities are assigned on PCM open, patch 16/17 has been updated to ignore such codecs for now. A separate patchset will take care of this case
- params_bps() reworded to snd_pcm_hw_params_bps()
- fixed compilation issues in sof-driver, patch 13/17
Cezary Rojewski (15): ALSA: hda: Honor subformat when querying PCMs ASoC: pcm: Honor subformat when configuring runtime ALSA: hda: Upgrade stream-format infrastructure ALSA: hda: Switch to new stream-format interface ALSA: hda/hdmi: Switch to new stream-format interface ALSA: hda/ca0132: Switch to new stream-format interface ASoC: codecs: hda: Switch to new stream-format interface ASoC: codecs: hdac_hda: Switch to new stream-format interface ASoC: codecs: hdac_hdmi: Switch to new stream-format interface ASoC: Intel Skylake: Switch to new stream-format interface ASoC: SOF: Intel: Switch to new stream-format interface ASoC: Intel: avs: Switch to new stream-format interface ALSA: hda: Drop snd_hdac_calc_stream_format() ASoC: Intel: avs: Kill S24_LE format ASoC: Intel: avs: Unhardcode HDAudio BE DAI drivers description
Jaroslav Kysela (1): ALSA: pcm: Introduce MSBITS subformat interface
Now the patch 1 became the one from Jaroslav, it's a good move. Jaroslav, could you take a look at the rest patches? I'm going to merge this in this week unless any objection comes up.
thanks,
Takashi