Work Queues

On CAVS platforms, the wall clock is used as a time source for multiple work queues (one work queue instance per active core).

Since enough comparators are not available, all instances register to a shared interrupt where one comparator is used to wake up all. The primary core re-programs the wall clock to the next wake event. Its work queue operates in primary mode. Work queues running on other cores are attached to the shared time source on CAVS SMP platforms; these are configured to the secondary mode. On other SMP platforms where multiple independent time sources are available, all queues can be configured in primary mode.

Synchronous SysTick on All Cores

The shared time source aligns work scheduling on all cores as all synchronously wake up on the same periodic event via a system tick timer, or SysTick.

SysTick periods are configurable. While no resolution is guaranteed, the default value of 1ms is acceptable for works that are typically scheduled on this system, specifically low latency works that are enabled and can be run on multiple cores in sync. For ultra low latency configurations, the SysTick period can be configured to a value of < 1ms.

An HD/A DMA running in circular buffer mode (@ dai) is already registered in the work queue with a specified timeout period of 1ms. Other DMAs can be switched from their individual interrupt sources (buffer completion) to work queues, thus making pipelines scheduling fully systick aligned.

In the case of more complex topologies, pipelines that start/terminate with a component other then dai can be also driven by work queues.

Note

Work queue primary/secondary mode vs. independent mode configurable by CONFIG @ compile time. The work queue min tick (1ms/0.33ms/1us) is configurable @ run-time so that the current mode is still fully supported.

scale max 800 height skinparam rectangle { backgroundColor<<dai>> #6fccdd backgroundColor<<dma>> #f6ed80 backgroundColor<<stream>> #d6d6de borderColor<<stream>> #d6d6de borderColor<<ppl>> #a1a1ca backgroundColor<<event>> #f05772 stereotypeFontColor<<event>> #ffffff fontColor<<event>> #ffffff backgroundColor<<cpu>> #f0f0f0 } together { ' ssp dais rectangle "ssp-dai #0" as ssp_0 <<dai>> rectangle "ssp-dai #1" as ssp_1 <<dai>> ' hda dais rectangle "hda-dai #0" as hda_0 <<dai>> } rectangle "core #0" <<cpu>> { together { rectangle "hda-dma\nhost output #0" as hda_dma_ho_0 <<dma>> rectangle "hda-dma\nhost input #0" as hda_dma_hi_0 <<dma>> } together { rectangle "dw-dma #0" as dw_dma_0 <<dma>> rectangle "dw-dma #1" as dw_dma_1 <<dma>> } rectangle "playback #0" <<stream>> { rectangle "ppl #0" as ppl_0 <<ppl>> { rectangle "host comp" as host_0 rectangle "vol" as vol_0 rectangle "mix-in" as mix_in_0 host_0 --> vol_0 vol_0 --> mix_in_0 } rectangle "ppl #1" as ppl_1 <<ppl>> { rectangle "mix-out" as mix_out_1 rectangle "dai comp" as dai_1 mix_out_1 --> dai_1 } mix_in_0 --> mix_out_1 } hda_dma_ho_0 --> host_0 dai_1 --> dw_dma_0 dw_dma_0 --> ssp_0 rectangle "capture #0" <<stream>> { rectangle "ppl #3" as ppl_3 <<ppl>> { rectangle "host comp" as host_3 rectangle "dai comp" as dai_3 host_3 <-- dai_3 } } hda_dma_hi_0 <-- host_3 dai_3 <-- dw_dma_1 dw_dma_1 <-- ssp_1 ' now let's show who's driving rectangle "work queue #0\n@core_0" as wq_0 <<event>> wq_0 .[#green]> mix_in_0 wq_0 .[#green]> dw_dma_0 : new wq_0 .[#green]> dw_dma_1 : new } rectangle "core #1" <<cpu>> { rectangle "hda-dma\nhost output #1" as hda_dma_ho_1 <<dma>> rectangle "hda-dma\nlink input #0" as hda_dma_li_0 <<dma>> rectangle "playback #1" <<stream>> { rectangle "ppl #2" as ppl_2 <<ppl>> { rectangle "host comp" as host_2 rectangle "dai comp" as dai_2 host_2 --> dai_2 } } hda_dma_ho_1 --> host_2 dai_2 --> hda_dma_li_0 hda_dma_li_0 --> hda_0 rectangle "work queue #1\n@core_1" as wq_1 <<event>> wq_1 .[#green]> hda_dma_li_0 }

Figure 10 Work queue dependecies for CAVS SMP

participant ipc participant "core 0" as core_0 participant "core 1" as core_1 participant wallclk ipc -> core_0 : new pipeline 0 (@core 0) activate core_0 core_0 -> wallclk : timer_register() core_0 -> core_0 : work_schedule(ppl 0 dai, +1ms) core_0 -> wallclk : work_set_timer(period = +1ms) ipc <-- core_0 deactivate core_0 wallclk -> core_0 : cb() @ next ms activate core_0 core_0 -> core_0 : ppl 0 dai cb() activate core_0 core_0 -> core_0 : ppl 0 copy core_0 -> core_0 : work_schedule(ppl 0 dai, +1ms) deactivate core_0 core_0 -> wallclk : work_set_timer() deactivate core_0 ipc -> core_0 : new pipeline 1 (@core 1) core_0 -> core_1 activate core_1 core_1 -> wallclk : timer_register() core_1 -> core_1 : work_schedule(ppl 1 dai) core_0 <-- core_1 deactivate core_1 wallclk -> core_0 : cb() @ next ms activate core_0 activate core_1 core_0 -> core_0 : ppl 0 dai cb() activate core_0 core_0 -> core_0 : ppl 0 copy core_0 -> core_0 : work_schedule(ppl 0 dai, +1ms) deactivate core_0 core_1 -> core_1 : ppl 1 dai cb() activate core_1 core_1 -> core_1 : ppl 1 copy core_1 -> core_1 : work_schedule(ppl 1 dai, +1ms) deactivate core_1 deactivate core_1 core_0 -> wallclk : work_set_timer() deactivate core_0

Figure 11 Work queue flow for CAVS SMP