Run-Time Parameter Support Summary

Run-Time Parameter Support Summary - 2021.2 English

Versal ACAP AI Engine Programming Environment User Guide (UG1076)

Document ID

UG1076

Release Date

2021-12-17

Version

2021.2 English

This section summarizes the AI Engine run-time parameter (RTP) support status.

Table 1. AI Engine RTP Default and Support Status
AI Engine RTP (from/to PS)	Input		Output
AI Engine RTP (from/to PS)	Synchronous	Asynchronous	Synchronous	Asynchronous
Scalar	Default	Supported	Supported	Default
Array	Default	Supported	Supported	Default

Code snippets for RTP connections from or to the PS:

connect<parameter>(fromPS, first.in[0]); //Synchronous RTP, default for input
connect<parameter>(fromPS, sync(first.in[0])); //Synchronous RTP
connect<parameter>(fromPS, async(first.in[0])); //Asynchronous RTP
connect<parameter>(second.inout[0], toPS); //Asynchronous RTP, default for output
connect<parameter>(async(second.inout[0]), toPS); //Asynchronous RTP
connect<parameter>(sync(second.inout[0]), toPS); //Synchronous RTP

Table 2. AI Engine RTP to AI Engine RTP Connection Support Status
AI Engine RTP to AI Engine RTP		To
AI Engine RTP to AI Engine RTP		Synchronous	Asynchronous	Not Specified
From	Synchronous	Synchronous	Not Supported	Synchronous
	Asynchronous	Not Supported	Asynchronous	Asynchronous
	Not Specified	Synchronous	Asynchronous	Synchronous

Code snippets for RTP connections between AI Engines:

connect<parameter>(first.inout[0], second.in[0]); //Not specified for output and input. Synchronous RTP from first.inout to second.in
connect<parameter>(sync(first.inout[0]), second.in[0]); //Specify “sync” for output. Synchronous RTP from first.inout to second.in
connect<parameter>(first.inout[0], sync(second.in[0])); //Specify “sync” for input. Synchronous RTP from first.inout to second.in
connect<parameter>(sync(first.inout[0]), sync(second.in[0])); //Specify “sync” for both. Synchronous RTP from first.inout to second.in
connect<parameter>(async(first.inout[0]), async(second.in[0])); //Specify “async” for both. Asynchronous RTP from first.inout to second.in
connect<parameter>(first.inout[0], async(second.in[0])); //Specify “async” for input. Asynchronous RTP from first.inout to second.in
connect<parameter>(async(first.inout[0]), second.in[0]); //Specify “async” for output. Asynchronous RTP from first.inout to second.in
connect<parameter>(async(first.inout[0]), sync(second.in[0])); //Not supported
connect<parameter>(sync(first.inout[0]), async(second.in[0])); //Not supported

Note: When using AI Engine RTP to AI Engine RTP connections, the source and destination kernels use shared memory for data communication. DMA is not involved in data communication. So, AI Engine kernels with RTP connections must be placed in the same AI Engine tile or adjacent tiles.

Table 3. AI Engine RTP Broadcast to Multiple AI Engine RTP Destinations Support Status
AI Engine RTP to AI Engine RTP		To
AI Engine RTP to AI Engine RTP		D1: Sync D2:Sync	D1:Async D2:Async	D1:Sync D2:Async	D1:Sync D2:Not Specified	D1:Async D2:Not Specified
From	S1:Sync S2:Sync	Synchronous	Not Supported	Not Supported	S1-D1:Sync S2-D2:Sync	Not Supported
	S1:Async S2:Async	Not Supported	Asynchronous	Not Supported	Not Supported	S1-D1:Async S2-D2:Async
	S1:Sync S2:Async	Not Supported	Not Supported	S1-D1:Sync S2-D2:Async	S1-D1:Sync S2-D2:Async	Not Supported
	S1:Sync S2:Not Specified	Not Supported	Not Supported	S1-D1:Sync S2-D2:Async	Synchronous	Not Supported
	S1:Async S2:Not Specified	Not Supported	Asynchronous	Not Supported	Not Supported	S1-D1:Async S2-D2:Sync
`S1` and `S2` in this table denote the RTP source for the first and second connections. `D1` and `D2` denote the RTP ports of multiple destination kernels. `Sync` and `Async` denote the synchronous and asynchronous modes specified in the graph. `S1-D1:Sync` denotes the connection from `S1` to `D1` is synchronous mode.

Similar to the AI Engine RTP to AI Engine RTP connection, the source and destination kernels for the RTP broadcast must be placed in the same AI Engine tile or adjacent tiles. Code snippets are similar too. Only the source RTP port is used in multiple connections. The RTP buffers are shared for the source and destination kernels.

Note:

There is a limitation in this version that only synchronous RTP is allowed between AI Engine kernel to AI Engine kernel when they are placed in the same AI Engine tile.
Array RTP broadcast is not supported in the current version.
Single buffer is not supported for RTP broadcast.