Trace data can be captured from Memory Tiles using Xilinx Runtime (XRT) and Xilinx Software Debugger (XSDB) the same way it can be done for AI Engine Tiles, AI Engine Memory, and AI Engine Interface. These options are specific to the AI Engine-ML.
The channels metric set focuses on the start and end of buffer descriptors while the other metric sets add stalls to trace.
| Event Type / Metric Set | input_channels | input_channels_stalls | output_channels | output_channels_stalls |
|---|---|---|---|---|
| S2MM channel buffer descriptors transfers | Yes 2x | Yes | - | - |
| S2MM channel lock stalls | - | Yes | - | - |
| S2MM channel stream starvation | - | Yes | - | - |
| S2MM channel memory backpressure | - | Yes | - | - |
| MM2S channel buffer descriptors transfers | - | - | Yes 2x | Yes |
| MM2S channel lock stalls | - | - | - | Yes |
| MM2S channel stream backpressure | - | - | - | Yes |
| MM2S channel memory starvation | - | - | - | Yes |
| Option | Description |
|---|---|
-tile-based-memory-tile-metrics
|
Sets which AI Engine-ML Memory Tile event to store for specified tiles. |
-graph-based-memory-tile-metrics
|
Sets which AI Engine-ML Memory Tile event to store for specified buffers/graph. |
Here are some examples of event tracing for Memory Tiles using a XSDB flow:
# Sets event trace to "input_channels" for the buffer named buffer2 of graph1 (channels 1 and 2) and "input_channels_stalls" for buffer4 in graph3 for channles 3 and 4.
-graph-based-memory-tile-metrics "graph1:buffer2:input_channels:1:2; graph3:buffer4:input_channels_stalls:3:4"
# Sets event trace to "output_channels" for columns 1 to 4 only for channel 1 and 2, and "output_channels_stalls" for column 5 and channels 3 and 4.
-tile-based-memory-tile-metrics "1:4:output_channels:1:2; 5:output_channels_stalls:3:4"| Option | Description |
|---|---|
tile_based_memory_tile_metrics =
<column|all>:<off|channels|input_channels_stalls|output_channels_stalls>[:<channel
1>][:<channel 2>]
|
Configures the AI Engine Memory Tile event trace metric to be applied for a single column or all columns. |
tile_based_memory_tile_metrics =<mincolumn>
<maxcolumn>:<off|input_channels|input_channels_stalls|output_channels|output_channels_stalls>[:<channel
1>][:<channel 2>]
|
Configures the AI Engine Memory Tile event trace metric to be applied for all columns in a range. |
graph_based_memory_tile_metrics = <graph
name|all>:<buffer
name|all>:<off|input_channels|input_channels_stalls|output_channels|output_channels_stalls>[:<channel
1>][:<channel 2>]
|
Configures the AI Engine Memory Tile event trace metric to be applied for one or all buffers in a specific or all graphs. |
Following are some xrt.ini file
examples:
# Sets event trace to "input_channels" for the buffer named buffer2 of graph1 (channels 1 and 2) and "input_channels_stalls" for buffer4 in graph3 for channles 3 and 4.
graph-based-memory-tile-metrics=graph1:buffer2:input_channels:1:2; graph3:buffer4:input_channels_stalls:3:4
# Sets event trace to "output_channels" for columns 1 to 4 only for channel 1 and 2, and "output_channels_stalls" for column 5 and channels 3 and 4.
tile-based-memory-tile-metrics=1:4:output_channels:1:2; 5:output_channels_stalls:3:4