DMIC: 16bits FIFOs record at double rate

[jajanusz]

DMIC works correctly only for 32bit (fifo_bits_a), for 16b it records at double rate - for example 10s squeezed into 5s with 2x tempo.

[lgirdwood]

@singalsu looks like DMA width is wrong for 16bit mode. probably reading back 32bits.

[singalsu]

I will do checks once I'm able to get this working again in my UP2 (code in dai.c line 671->) . I have still some issue with topology and kernel.

[singalsu]

I can repeat the bug now. With DMIC the 16 bit stereo data frame is packed to 32 bit word as L,R. And mono data is packed as L,L with two successive frames in a word. Is SSP similar? If SSP packs 16 bit data differently it could explain this.

[singalsu]

@lgirdwood I can hack this to work by doing this into dw-dma.c

			default:
				/* config the src width for 16 bit samples */
				//lli_desc->ctrl_lo |= DW_CTLL_SRC_WIDTH(1);
				lli_desc->ctrl_lo |= DW_CTLL_SRC_WIDTH(2);
				break;

But how can I make dai_config() in dai.c achieve the same cleanly? For DMIC the width can be always set to 32 bits.

[singalsu]

I tried to add this into dai_config() for SOF_DAI_INTEL_DMIC case

		dd->config.burst_elems = 8;
		dd->config.src_width = 4;
		dd->config.dest_width = 4;

But according to trace I get 8, 2, 2 so the setting here does not work. What puzzles me is that this code part gets executed 13 times for DMIC. It sounds like the count for SSPs in nocodec topology. So maybe a bug in function that calls dai_config.

[ranj063]

@singalsu I think this is an error. I've sent you a patch that addresses this issue by setting the config only for the comp who's dai_index matches the dai_index in the config. Please give it a try.

[singalsu]

Thanks, the patch reduces the number of calls to 3. Though it's still more than needed. I wonder if the list in ipc_comp_dai_config() in ipc.c contains duplicates.

The no impact to dw-dma.c from the dd->config settings attempt in dai_config() to achieve always 32 bit DMA mode for DMIC also remains. Like I said earlier ugly hacking in dw-dma.c fixes 16 bit capture for DMIC but naturally it would break 16 bit SSP modes.

Do you have @RanderWang @keyonjie suggestions what I could try next?

[singalsu]

I noticed that if I set "dd->config.burst_elems = 4;" for DMIC in dai_config() the change is seen in dw-dma.c. Therefore it is very likely that some later code overwrites dd->config.src_width to 2. Maybe some SSP specific thing kicks in?

Maybe dai_capture_params()?

[singalsu]

I can make DMIC 16bit capture work by commenting out these lines from dai_capture_params()

	/* set up DMA configuration */
	config->direction = DMA_DIR_DEV_TO_MEM;
	//config->src_width = comp_sample_bytes(dev);
	//config->dest_width = comp_sample_bytes(dev);
	config->cyclic = 1;
	config->src_dev = dd->dai->plat_data.fifo[1].handshake;

@lgirdwood My question is that can this SSP specific thing be done instead in dai_config for type SOF_DAI_INTEL_SSP?

[jajanusz]

@singalsu
I think we shouldn't set config fields in many places, it will get messy, better to set config once in *_params functions and then use them in dai_config as we do now. We have to find better way to figure out what widths to set before passing config to dai_config

[singalsu]

Then there should be DAI type check in _*params and apply differently src_width and dest_width.