Parallelism, communication overhead tradeoffs

[lukehatpadl]

I'm working on mixer that will use an XE232 to ingest 4x TDM16 streams, mixing into 2 independent stereo outputs. (The first two tiles are reserved for the control plane and Ethernet.)

The current architecture is:

• tile[2] 2xTDM16 I/O (1-bit ports), 2xI2S output (4-bit port) (<-> shared memory)
• tile[2] 4x submixer tasks (shared memory -> channel)
• tile[2] summing/output mixer task feeding I/O task (channel -> shared memory)
• tile[3] 2xTDM16 I/O (-> shared memory)
• tile[3] 4x submixer tasks (shared memory -> channel)

I have prototyped this on the xCORE-200 explorerKIT with fewer channels and it appears to be working. Also, I can build (but not yet test) for XE232.

Other permutations would have better parallelism and less communication overhead (for example, having a DAC per tile and optionally mult'ing the TDM streams to both tiles, avoiding any cross-tile signal plane communication), but with slightly less flexibility and greater CPU overhead.

I'm wondering if anyone here has built something similar has any general advice on these tradeoffs.

[Ross]

Sounds like a mixing beast that should be able to have a huge amount of routing flexibility, excellent task for an xcore! Ive not built anything so large - topping out at 18 x 8 (but with flexibility to route many more channels in/out)