| Message ID | 20180426163007.5632-2-ckeepax@opensource.cirrus.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
On Thu, Apr 26, 2018 at 05:30:05PM +0100, Charles Keepax wrote: > It is proposed that the model adopted for compressed component > support currently should be that multiple components are supported > on a compressed DAI but that they must provide a unified set of > capabilities. So for example having multiple components involved in > the decode is fine but the core will not presently attempt to make > smart decisions like offloading MP3 to this component and AAC another. Well this is supposed to be entirely a userspace call, we just present devices with capabilities and the usespace decides... Btw capabilities are supposed to be dynamic. Looking at the code again now, I realized that we are treating compress like PCM. It makes little sense to me to have multiple components for a compressed device, does that model on your systems..? > To implement this it is suggested that callbacks configuring the state > of the components (trigger, set_params, ack and set_metadata) should be > called on all components and required to succeed on all components > before being considered to have succeeded. However for callbacks that > return information from the driver (copy, get_metadata, pointer, > get_codec_caps, get_caps and get_params) it is expected that either > there is only a single provider on the link or that all components > will return identical results. > > Essentially this matches the current implementation of the code and > only small clean ups are undertaken here. > For callbacks configuring the state of the components simplify the > code a little and make intention clearer by aborting as soon as an > error is encountered. The operation has already failed and there is > nothing to be gained from processing the additional callbacks. > > For callbacks returning information from the driver only look for the > first callback provided, currently the code will call every callback > only returning the information provided by the last. Again this makes > the currently supported feature set a little more clear. Btw code changes look fine but I think we need broader cleanup here...
On Thu, May 03, 2018 at 09:57:14PM +0530, Vinod Koul wrote: > On Thu, Apr 26, 2018 at 05:30:05PM +0100, Charles Keepax wrote: > > It is proposed that the model adopted for compressed component > > support currently should be that multiple components are supported > > on a compressed DAI but that they must provide a unified set of > > capabilities. So for example having multiple components involved in > > the decode is fine but the core will not presently attempt to make > > smart decisions like offloading MP3 to this component and AAC another. > > Well this is supposed to be entirely a userspace call, we just present > devices with capabilities and the usespace decides... Btw capabilities are > supposed to be dynamic. My intention was to not suggest otherwise. The only point I was really making is that if there are multiple components on the link the core won't make attempt to amalgamate output from those components, but we will inform all components of activity on the DAI. > Looking at the code again now, I realized that we are treating compress like > PCM. It makes little sense to me to have multiple components for a > compressed device, does that model on your systems..? So that was very much my initial reaction as well [1], and certainly for our devices it only really makes sense to have a single component handle the compressed ops. Hence why I initially started with basically just returning the first component with compressed ops. That said however, thinking about it more I do think there are pretty reasonable systems that have multiple components on a compressed DAI. For example I could imagine a system where you have a DSP on both the AP and CODEC ends of the DAI link and they split the decode doing some work on each. In that case one would want calls like open and set_params to go to both components. As you say separate decode engines probably belong on separate DAIs and that kinda is what led me to the current series. It should implement enough to enable basic multi-component use-cases but makes it more clear that we are not supporting multiplexing multiple offloads onto a single DAI at the moment. Thanks, Charles [1] https://patchwork.kernel.org/patch/10182603/
On Fri, May 04, 2018 at 12:59:22PM +0100, Charles Keepax wrote: > On Thu, May 03, 2018 at 09:57:14PM +0530, Vinod Koul wrote: > > On Thu, Apr 26, 2018 at 05:30:05PM +0100, Charles Keepax wrote: > > > It is proposed that the model adopted for compressed component > > > support currently should be that multiple components are supported > > > on a compressed DAI but that they must provide a unified set of > > > capabilities. So for example having multiple components involved in > > > the decode is fine but the core will not presently attempt to make > > > smart decisions like offloading MP3 to this component and AAC another. > > > > Well this is supposed to be entirely a userspace call, we just present > > devices with capabilities and the usespace decides... Btw capabilities are > > supposed to be dynamic. > > My intention was to not suggest otherwise. The only point I > was really making is that if there are multiple components on > the link the core won't make attempt to amalgamate output from > those components, but we will inform all components of activity > on the DAI. > > > Looking at the code again now, I realized that we are treating compress like > > PCM. It makes little sense to me to have multiple components for a > > compressed device, does that model on your systems..? > > So that was very much my initial reaction as well [1], and > certainly for our devices it only really makes sense to have a > single component handle the compressed ops. Hence why I initially > started with basically just returning the first component with > compressed ops. > > That said however, thinking about it more I do think there are > pretty reasonable systems that have multiple components on a > compressed DAI. For example I could imagine a system where you > have a DSP on both the AP and CODEC ends of the DAI link and > they split the decode doing some work on each. In that case > one would want calls like open and set_params to go to both > components. > > As you say separate decode engines probably belong on separate > DAIs and that kinda is what led me to the current series. It > should implement enough to enable basic multi-component use-cases > but makes it more clear that we are not supporting multiplexing > multiple offloads onto a single DAI at the moment. > > Thanks, > Charles > > [1] https://patchwork.kernel.org/patch/10182603/ Just adding Vinod's kernel.org address. Thanks, Charles
On 04-05-18, 12:59, Charles Keepax wrote: > On Thu, May 03, 2018 at 09:57:14PM +0530, Vinod Koul wrote: > > On Thu, Apr 26, 2018 at 05:30:05PM +0100, Charles Keepax wrote: > > > It is proposed that the model adopted for compressed component > > > support currently should be that multiple components are supported > > > on a compressed DAI but that they must provide a unified set of > > > capabilities. So for example having multiple components involved in > > > the decode is fine but the core will not presently attempt to make > > > smart decisions like offloading MP3 to this component and AAC another. > > > > Well this is supposed to be entirely a userspace call, we just present > > devices with capabilities and the usespace decides... Btw capabilities are > > supposed to be dynamic. > > My intention was to not suggest otherwise. The only point I > was really making is that if there are multiple components on > the link the core won't make attempt to amalgamate output from > those components, but we will inform all components of activity > on the DAI. > > > Looking at the code again now, I realized that we are treating compress like > > PCM. It makes little sense to me to have multiple components for a > > compressed device, does that model on your systems..? > > So that was very much my initial reaction as well [1], and > certainly for our devices it only really makes sense to have a > single component handle the compressed ops. Hence why I initially > started with basically just returning the first component with > compressed ops. > > That said however, thinking about it more I do think there are > pretty reasonable systems that have multiple components on a > compressed DAI. For example I could imagine a system where you > have a DSP on both the AP and CODEC ends of the DAI link and > they split the decode doing some work on each. In that case > one would want calls like open and set_params to go to both > components. Am not sure if we can split the decode into multiple DSPs like this :) Yes one can do processing and one can decode if both have the capability but I don't forsee that being split, so not sure if we need it!!! > As you say separate decode engines probably belong on separate > DAIs and that kinda is what led me to the current series. It > should implement enough to enable basic multi-component use-cases > but makes it more clear that we are not supporting multiplexing > multiple offloads onto a single DAI at the moment. > > Thanks, > Charles > > [1] https://patchwork.kernel.org/patch/10182603/
On Tue, May 08, 2018 at 02:03:19PM +0530, Vinod Koul wrote: > On 04-05-18, 12:59, Charles Keepax wrote: > > On Thu, May 03, 2018 at 09:57:14PM +0530, Vinod Koul wrote: > > > On Thu, Apr 26, 2018 at 05:30:05PM +0100, Charles Keepax wrote: > > That said however, thinking about it more I do think there are > > pretty reasonable systems that have multiple components on a > > compressed DAI. For example I could imagine a system where you > > have a DSP on both the AP and CODEC ends of the DAI link and > > they split the decode doing some work on each. In that case > > one would want calls like open and set_params to go to both > > components. > > Am not sure if we can split the decode into multiple DSPs like > this :) Yes one can do processing and one can decode if both have > the capability but I don't forsee that being split, so not sure > if we need it!!! > I think you definitely could split the decode across multiple DSPs like that, we have had processing split across multiple cores for various things on the CODECs. That said "Need it" is very strong, I would say such a system is plausible but not something I am aware anyone is actually building right now. I guess my thinking was that the system is basically already supporting multiple components so it seemed like a step backwards to rip it out given there are plausible systems were it might be useful. And this patch is really just tidying up the already submitted code a little rather than changing the functionality in any substancial way. That said if you feel strongly about it, I certainly don't need the support in the foreseeable future. I am happy to go back to something similar to my earlier patch that just locates the first device with compressed ops and calls the ops on that? Although it might still be worth merging this one as an intermediate tidy up in that case. Thanks, Charles
On 08-05-18, 11:52, Charles Keepax wrote: > On Tue, May 08, 2018 at 02:03:19PM +0530, Vinod Koul wrote: > > On 04-05-18, 12:59, Charles Keepax wrote: > > > On Thu, May 03, 2018 at 09:57:14PM +0530, Vinod Koul wrote: > > > > On Thu, Apr 26, 2018 at 05:30:05PM +0100, Charles Keepax wrote: > > > That said however, thinking about it more I do think there are > > > pretty reasonable systems that have multiple components on a > > > compressed DAI. For example I could imagine a system where you > > > have a DSP on both the AP and CODEC ends of the DAI link and > > > they split the decode doing some work on each. In that case > > > one would want calls like open and set_params to go to both > > > components. > > > > Am not sure if we can split the decode into multiple DSPs like > > this :) Yes one can do processing and one can decode if both have > > the capability but I don't forsee that being split, so not sure > > if we need it!!! > > > > I think you definitely could split the decode across multiple > DSPs like that, we have had processing split across multiple > cores for various things on the CODECs. That said "Need it" > is very strong, I would say such a system is plausible but > not something I am aware anyone is actually building right now. well processing is different, we are dealing with PCM data. Decode is one shot, you get PCM output and then can do whatever you want with it. If decode is split what is the output format ;-) > I guess my thinking was that the system is basically already > supporting multiple components so it seemed like a step > backwards to rip it out given there are plausible systems were > it might be useful. And this patch is really just tidying up > the already submitted code a little rather than changing the > functionality in any substancial way. > > That said if you feel strongly about it, I certainly don't > need the support in the foreseeable future. I am happy to > go back to something similar to my earlier patch that just > locates the first device with compressed ops and calls the > ops on that? Although it might still be worth merging this > one as an intermediate tidy up in that case. Going back would be better, I leave it upto Mark on how he wants to do this, either way am fine if end goal is met :) > > Thanks, > Charles
On Tue, May 08, 2018 at 05:34:53PM +0530, Vinod Koul wrote: > On 08-05-18, 11:52, Charles Keepax wrote: > > On Tue, May 08, 2018 at 02:03:19PM +0530, Vinod Koul wrote: > > > On 04-05-18, 12:59, Charles Keepax wrote: > > > > On Thu, May 03, 2018 at 09:57:14PM +0530, Vinod Koul wrote: > > > > > On Thu, Apr 26, 2018 at 05:30:05PM +0100, Charles Keepax wrote: > > > > That said however, thinking about it more I do think there are > > > > pretty reasonable systems that have multiple components on a > > > > compressed DAI. For example I could imagine a system where you > > > > have a DSP on both the AP and CODEC ends of the DAI link and > > > > they split the decode doing some work on each. In that case > > > > one would want calls like open and set_params to go to both > > > > components. > > > > > > Am not sure if we can split the decode into multiple DSPs like > > > this :) Yes one can do processing and one can decode if both have > > > the capability but I don't forsee that being split, so not sure > > > if we need it!!! > > > > > > > I think you definitely could split the decode across multiple > > DSPs like that, we have had processing split across multiple > > cores for various things on the CODECs. That said "Need it" > > is very strong, I would say such a system is plausible but > > not something I am aware anyone is actually building right now. > > well processing is different, we are dealing with PCM data. > Decode is one shot, you get PCM output and then can do whatever > you want with it. > > If decode is split what is the output format ;-) > The end output format is still PCM, but what is passed from the DSP on the AP to the DSP on the CODEC could be anything. And decode is hardly one shot, most decodes are made up of many steps (frequency domain transforms, lookup tables, etc.). It seems like we are slightly talking at cross purposes here, I wonder is this because you are thinking mostly of the DPCM compressed case where it looks like it is mostly hard-coded to use a PCM backend? Whereas I am thinking more of how our systems is used where it is a direct compressed DAI so there isn't really a backend? > > I guess my thinking was that the system is basically already > > supporting multiple components so it seemed like a step > > backwards to rip it out given there are plausible systems were > > it might be useful. And this patch is really just tidying up > > the already submitted code a little rather than changing the > > functionality in any substancial way. > > > > That said if you feel strongly about it, I certainly don't > > need the support in the foreseeable future. I am happy to > > go back to something similar to my earlier patch that just > > locates the first device with compressed ops and calls the > > ops on that? Although it might still be worth merging this > > one as an intermediate tidy up in that case. > > Going back would be better, I leave it upto Mark on how he wants > to do this, either way am fine if end goal is met :) > Alright will wait and see what Mark thinks, but will dig out my original patch on the assumption that is probably the way we are going. Thanks, Charles
diff --git a/sound/soc/soc-compress.c b/sound/soc/soc-compress.c index 62875c6a93a14..b411143f21ccc 100644 --- a/sound/soc/soc-compress.c +++ b/sound/soc/soc-compress.c @@ -363,7 +363,7 @@ static int soc_compr_trigger(struct snd_compr_stream *cstream, int cmd) struct snd_soc_rtdcom_list *rtdcom; struct snd_soc_dai *codec_dai = rtd->codec_dai; struct snd_soc_dai *cpu_dai = rtd->cpu_dai; - int ret = 0, __ret; + int ret = 0; mutex_lock_nested(&rtd->pcm_mutex, rtd->pcm_subclass); @@ -374,12 +374,10 @@ static int soc_compr_trigger(struct snd_compr_stream *cstream, int cmd) !component->driver->compr_ops->trigger) continue; - __ret = component->driver->compr_ops->trigger(cstream, cmd); - if (__ret < 0) - ret = __ret; + ret = component->driver->compr_ops->trigger(cstream, cmd); + if (ret < 0) + goto out; } - if (ret < 0) - goto out; if (cpu_dai->driver->cops && cpu_dai->driver->cops->trigger) cpu_dai->driver->cops->trigger(cstream, cmd, cpu_dai); @@ -404,7 +402,7 @@ static int soc_compr_trigger_fe(struct snd_compr_stream *cstream, int cmd) struct snd_soc_component *component; struct snd_soc_rtdcom_list *rtdcom; struct snd_soc_dai *cpu_dai = fe->cpu_dai; - int ret = 0, __ret, stream; + int ret, stream; if (cmd == SND_COMPR_TRIGGER_PARTIAL_DRAIN || cmd == SND_COMPR_TRIGGER_DRAIN) { @@ -416,9 +414,9 @@ static int soc_compr_trigger_fe(struct snd_compr_stream *cstream, int cmd) !component->driver->compr_ops->trigger) continue; - __ret = component->driver->compr_ops->trigger(cstream, cmd); - if (__ret < 0) - ret = __ret; + ret = component->driver->compr_ops->trigger(cstream, cmd); + if (ret < 0) + return ret; } return ret; } @@ -444,12 +442,10 @@ static int soc_compr_trigger_fe(struct snd_compr_stream *cstream, int cmd) !component->driver->compr_ops->trigger) continue; - __ret = component->driver->compr_ops->trigger(cstream, cmd); - if (__ret < 0) - ret = __ret; + ret = component->driver->compr_ops->trigger(cstream, cmd); + if (ret < 0) + goto out; } - if (ret < 0) - goto out; fe->dpcm[stream].runtime_update = SND_SOC_DPCM_UPDATE_FE; @@ -483,7 +479,7 @@ static int soc_compr_set_params(struct snd_compr_stream *cstream, struct snd_soc_component *component; struct snd_soc_rtdcom_list *rtdcom; struct snd_soc_dai *cpu_dai = rtd->cpu_dai; - int ret = 0, __ret; + int ret; mutex_lock_nested(&rtd->pcm_mutex, rtd->pcm_subclass); @@ -507,12 +503,10 @@ static int soc_compr_set_params(struct snd_compr_stream *cstream, !component->driver->compr_ops->set_params) continue; - __ret = component->driver->compr_ops->set_params(cstream, params); - if (__ret < 0) - ret = __ret; + ret = component->driver->compr_ops->set_params(cstream, params); + if (ret < 0) + goto err; } - if (ret < 0) - goto err; if (rtd->dai_link->compr_ops && rtd->dai_link->compr_ops->set_params) { ret = rtd->dai_link->compr_ops->set_params(cstream); @@ -533,7 +527,7 @@ static int soc_compr_set_params(struct snd_compr_stream *cstream, cancel_delayed_work_sync(&rtd->delayed_work); - return ret; + return 0; err: mutex_unlock(&rtd->pcm_mutex); @@ -549,7 +543,7 @@ static int soc_compr_set_params_fe(struct snd_compr_stream *cstream, struct snd_soc_component *component; struct snd_soc_rtdcom_list *rtdcom; struct snd_soc_dai *cpu_dai = fe->cpu_dai; - int ret = 0, __ret, stream; + int ret, stream; if (cstream->direction == SND_COMPRESS_PLAYBACK) stream = SNDRV_PCM_STREAM_PLAYBACK; @@ -571,12 +565,10 @@ static int soc_compr_set_params_fe(struct snd_compr_stream *cstream, !component->driver->compr_ops->set_params) continue; - __ret = component->driver->compr_ops->set_params(cstream, params); - if (__ret < 0) - ret = __ret; + ret = component->driver->compr_ops->set_params(cstream, params); + if (ret < 0) + goto out; } - if (ret < 0) - goto out; if (fe->dai_link->compr_ops && fe->dai_link->compr_ops->set_params) { ret = fe->dai_link->compr_ops->set_params(cstream); @@ -618,7 +610,7 @@ static int soc_compr_get_params(struct snd_compr_stream *cstream, struct snd_soc_component *component; struct snd_soc_rtdcom_list *rtdcom; struct snd_soc_dai *cpu_dai = rtd->cpu_dai; - int ret = 0, __ret; + int ret = 0; mutex_lock_nested(&rtd->pcm_mutex, rtd->pcm_subclass); @@ -635,9 +627,8 @@ static int soc_compr_get_params(struct snd_compr_stream *cstream, !component->driver->compr_ops->get_params) continue; - __ret = component->driver->compr_ops->get_params(cstream, params); - if (__ret < 0) - ret = __ret; + ret = component->driver->compr_ops->get_params(cstream, params); + break; } err: @@ -651,7 +642,7 @@ static int soc_compr_get_caps(struct snd_compr_stream *cstream, struct snd_soc_pcm_runtime *rtd = cstream->private_data; struct snd_soc_component *component; struct snd_soc_rtdcom_list *rtdcom; - int ret = 0, __ret; + int ret = 0; mutex_lock_nested(&rtd->pcm_mutex, rtd->pcm_subclass); @@ -662,9 +653,8 @@ static int soc_compr_get_caps(struct snd_compr_stream *cstream, !component->driver->compr_ops->get_caps) continue; - __ret = component->driver->compr_ops->get_caps(cstream, caps); - if (__ret < 0) - ret = __ret; + ret = component->driver->compr_ops->get_caps(cstream, caps); + break; } mutex_unlock(&rtd->pcm_mutex); @@ -677,7 +667,7 @@ static int soc_compr_get_codec_caps(struct snd_compr_stream *cstream, struct snd_soc_pcm_runtime *rtd = cstream->private_data; struct snd_soc_component *component; struct snd_soc_rtdcom_list *rtdcom; - int ret = 0, __ret; + int ret = 0; mutex_lock_nested(&rtd->pcm_mutex, rtd->pcm_subclass); @@ -688,9 +678,9 @@ static int soc_compr_get_codec_caps(struct snd_compr_stream *cstream, !component->driver->compr_ops->get_codec_caps) continue; - __ret = component->driver->compr_ops->get_codec_caps(cstream, codec); - if (__ret < 0) - ret = __ret; + ret = component->driver->compr_ops->get_codec_caps(cstream, + codec); + break; } mutex_unlock(&rtd->pcm_mutex); @@ -703,7 +693,7 @@ static int soc_compr_ack(struct snd_compr_stream *cstream, size_t bytes) struct snd_soc_component *component; struct snd_soc_rtdcom_list *rtdcom; struct snd_soc_dai *cpu_dai = rtd->cpu_dai; - int ret = 0, __ret; + int ret = 0; mutex_lock_nested(&rtd->pcm_mutex, rtd->pcm_subclass); @@ -720,9 +710,9 @@ static int soc_compr_ack(struct snd_compr_stream *cstream, size_t bytes) !component->driver->compr_ops->ack) continue; - __ret = component->driver->compr_ops->ack(cstream, bytes); - if (__ret < 0) - ret = __ret; + ret = component->driver->compr_ops->ack(cstream, bytes); + if (ret < 0) + goto err; } err: @@ -736,7 +726,7 @@ static int soc_compr_pointer(struct snd_compr_stream *cstream, struct snd_soc_pcm_runtime *rtd = cstream->private_data; struct snd_soc_component *component; struct snd_soc_rtdcom_list *rtdcom; - int ret = 0, __ret; + int ret = 0; struct snd_soc_dai *cpu_dai = rtd->cpu_dai; mutex_lock_nested(&rtd->pcm_mutex, rtd->pcm_subclass); @@ -751,9 +741,8 @@ static int soc_compr_pointer(struct snd_compr_stream *cstream, !component->driver->compr_ops->pointer) continue; - __ret = component->driver->compr_ops->pointer(cstream, tstamp); - if (__ret < 0) - ret = __ret; + ret = component->driver->compr_ops->pointer(cstream, tstamp); + break; } mutex_unlock(&rtd->pcm_mutex); @@ -792,7 +781,7 @@ static int soc_compr_set_metadata(struct snd_compr_stream *cstream, struct snd_soc_component *component; struct snd_soc_rtdcom_list *rtdcom; struct snd_soc_dai *cpu_dai = rtd->cpu_dai; - int ret = 0, __ret; + int ret; if (cpu_dai->driver->cops && cpu_dai->driver->cops->set_metadata) { ret = cpu_dai->driver->cops->set_metadata(cstream, metadata, cpu_dai); @@ -807,12 +796,13 @@ static int soc_compr_set_metadata(struct snd_compr_stream *cstream, !component->driver->compr_ops->set_metadata) continue; - __ret = component->driver->compr_ops->set_metadata(cstream, metadata); - if (__ret < 0) - ret = __ret; + ret = component->driver->compr_ops->set_metadata(cstream, + metadata); + if (ret < 0) + return ret; } - return ret; + return 0; } static int soc_compr_get_metadata(struct snd_compr_stream *cstream, @@ -822,7 +812,7 @@ static int soc_compr_get_metadata(struct snd_compr_stream *cstream, struct snd_soc_component *component; struct snd_soc_rtdcom_list *rtdcom; struct snd_soc_dai *cpu_dai = rtd->cpu_dai; - int ret = 0, __ret; + int ret; if (cpu_dai->driver->cops && cpu_dai->driver->cops->get_metadata) { ret = cpu_dai->driver->cops->get_metadata(cstream, metadata, cpu_dai); @@ -837,12 +827,11 @@ static int soc_compr_get_metadata(struct snd_compr_stream *cstream, !component->driver->compr_ops->get_metadata) continue; - __ret = component->driver->compr_ops->get_metadata(cstream, metadata); - if (__ret < 0) - ret = __ret; + return component->driver->compr_ops->get_metadata(cstream, + metadata); } - return ret; + return 0; } /* ASoC Compress operations */
It is proposed that the model adopted for compressed component support currently should be that multiple components are supported on a compressed DAI but that they must provide a unified set of capabilities. So for example having multiple components involved in the decode is fine but the core will not presently attempt to make smart decisions like offloading MP3 to this component and AAC another. To implement this it is suggested that callbacks configuring the state of the components (trigger, set_params, ack and set_metadata) should be called on all components and required to succeed on all components before being considered to have succeeded. However for callbacks that return information from the driver (copy, get_metadata, pointer, get_codec_caps, get_caps and get_params) it is expected that either there is only a single provider on the link or that all components will return identical results. Essentially this matches the current implementation of the code and only small clean ups are undertaken here. For callbacks configuring the state of the components simplify the code a little and make intention clearer by aborting as soon as an error is encountered. The operation has already failed and there is nothing to be gained from processing the additional callbacks. For callbacks returning information from the driver only look for the first callback provided, currently the code will call every callback only returning the information provided by the last. Again this makes the currently supported feature set a little more clear. Signed-off-by: Charles Keepax <ckeepax@opensource.cirrus.com> --- sound/soc/soc-compress.c | 105 +++++++++++++++++++++-------------------------- 1 file changed, 47 insertions(+), 58 deletions(-)