Extra Options - 1.1 English

Vitis AI User Guide (UG1414)

Document ID
Release Date
1.1 English

In addition to the common options, VAI_C supports some extra options available only for edge DPU, which are specified by ‘--options’ or '-e' option of VAI_C compiler and described in the following table.

Table 1. VAI_C Extra Options for Edge DPU
Parameters Description
--help Show all available options of VAI_C for edge DPU.
--version Show VAI_C version information.
--save_kernel To save kernel description info into file.
--mode Compilation mode for DPU kernel: debug or normal. By default, network models are compiled into DPU kernels under normal mode. Note that the debug mode is default for edge VAI_C before version 5.0.

Debug: DPU nodes (or supper layers) of the network model run one by one under the scheduling of runtime N2Cube. With the help of DExplorer, the users can perform debugging or performance profiling for each node of DPU kernel under debug mode.

Normal: All layers/operators of the network model are packaged into one single DPU execution unit and there is no interruption involved during launching. Compared with debug mode, normal mode DPU kernel delivers better performance and should be used during production release phase.

--dump Dump different types of info to facilitate debugging and use commas as delimiter when multiple types are given:
  • Graph: Original graph and transformed graph in DOT format. The dump files’ names are ended with “.gv” suffix.
  • Weights: Weights and bias data for different layers. The dump files’ names are ended with “.weights” or “.bias” suffix.
  • ir: Immediate representation for different layer in VAI_C. The dump files’ names are ended with “.ir” suffix.
  • quant_info: Quantization information for different layers. The dump file’s name is “quant.info”.
  • dcf: DPU configuration parameters specified during DPU IP block design. The dump file’s name is “dpu.dcf.dump”. Note: The dpu.dcf.dump file is just used for dump purposes and should not be fed to VAI_C by “--dcf" option for compilation purpose.
  • log: Other compilation log generated by VAI_C.
  • fused_graph_info: VAI_C graph IR informaton to describe the relationship between model’s original layers/operators and DPU nodes (or super layer).
  • all: Dump all listed above.
Note: All dumped files except for graph, weights, bias, and fushed_graph_info type are decrypted by VAI_C. In case of network compilation errors, these dump files can be delivered to Xilinx AI support team for further analysis.
--split_io_mem Enable the using of split IO memory model for DPU kernel to be built. For this model, the DPU memory buffer for input/output tensors are separated from the memory buffer for intermediate feature maps. If not specified, the unique memory model is used for DPU kernel by default. About unique memory model and split IO memory model, refer to section Advanced Programming for Edge for more details.