Skip to main content

bmrt_test Model test tools

The bmrt_test tool is designed to test the correctness and actual runtime performance of a bmodel using the bmruntime interface.

Usage:
  --version          : Show version.
  --context_dir      : The directory containing compiled context files.
  --append_dir       : The directory of an additional compiled context.
  --bmodel           : The path to a `bmodel`; tests the `bmodel` without comparison.
  --bmodel_list      : File containing paths to one or more compiled contexts.
  --devid            : The target device number.
  --compare          : If set to 0, disables result comparison; otherwise, enables it.
  --accuracy_f       : Floating-point comparison accuracy.
  --accuracy_i       : Integer comparison accuracy.
  --loopnum          : Number of times to launch the network (default: 1).
  --calculate_times  : Number of inference loops (default: 1).
  --thread_num       : Number of threads for network launch.
  --prealloc         : Allocates memory before loading the `bmodel` (hex format).
  --export_neuron    : Exports neuron memory to a file.
  --export_output    : Exports output data to a file.
  --profile          : Enables profiling.
  --subnet_time      : Prints execution time for each subnet (if applicable).
  --net_idx          : Selects the network by index.
  --stage_idx        : Selects the stage by index.
  --debug_output     : Dumps output and reference data for debugging.
  --shapes           : Specifies input tensor shapes.
  --output_shapes    : Specifies output tensor shapes.
  --cascade_device   : Specifies devices for running a cascade model (e.g., `1,2`).
  --core_list        : Specifies core IDs used for inference.
                         Example:
                         - `0,1` means using cores 0 and 1 together for a multi-core compiled `bmodel`.
                         - `0:1` means using cores 0 and 1 to infer a single-core compiled `bmodel` in parallel.

Features:​

  1. Runs inference on a bmodel with random data to verify its integrity and executability.
  2. Runs inference with fixed input data and compares output against reference data for accuracy validation.
  3. Measures actual execution time of the bmodel.
  4. Generates performance profiling data using the bmprofile mechanism.

Usage​

Run a bmodel​

bmrt_test --bmodel xxx.bmodel

Run a Multi-Stage bmodel​

bmrt_test --bmodel xxx.bmodel --stage_idx 0 --shapes "[1,3,224,224]"

Run a bmodel from a Directory​

The bmodel_dir must contain compilation.bmodel.

bmrt_test --context_dir bmodel_dir --compare=0

Run a bmodel from a Directory with Data Comparison​

The bmodel_dir must contain compilation.bmodel, input_ref_data.dat, and output_ref_data.dat for comparison.

bmrt_test --context_dir bmodel_dir

Generate Profile Data​

BMRUNTIME_ENABLE_PROFILE=1 bmrt_test --bmodel xxx.bmodel

Save Inference Data​

Saves input and output inference data to input_ref_data.dat.bmrt and output_ref_data.dat.bmrt.

BMRT_SAVE_IO_TENSORS=1 bmrt_test --bmodel xxx.bmodel