Configuration

CIMFlow uses JSON configuration files to define tool paths, batch runs, and hardware parameters.

Configuration Directory

After installation, configuration files are stored in config/:

config/
├── tool_paths.json     # Tool paths and global settings
└── batch.json          # Batch run configurations (user-created)

tool_paths.json

Defines paths to CIMFlow tools and hardware configuration directories.

batch.json

User-created file for defining multiple pipeline runs with different parameters.

Tool Paths Configuration

The tool_paths.json file defines paths to CIMFlow tools and resources.

config/tool_paths.json

{
  "config_dir": "../modules/CIMFlow-Simulator/config/example",
  "profiler_cfg": "../modules/CIMFlow-Simulator/config/profiler_config.json",
  "compiler_bin": "cim-compiler",
  "simulator_bin": "cim-simulator"
}

Field	Description
`config_dir`	Directory containing hardware configuration files
`profiler_cfg`	Path to profiler configuration (optional)
`compiler_bin`	Compiler executable name or path
`simulator_bin`	Simulator executable name or path

Relative paths in tool_paths.json are resolved relative to the file's directory.

This file is generated during installation. Modify it to use custom tool paths or configuration directories.

Batch Configuration

Batch files define multiple pipeline runs with different parameters.

config/batch.json

{
  "paths_cfg": "tool_paths.json",
  "output_dir": "output",
  "keep_ir": false,
  "run_name": "batch",
  "runs": [
    {
      "model_path": "path/to/model.onnx",
      "t": 8,
      "k": 16,
      "b": 16,
      "c": 64,
      "batch_size": 8,
      "strategy": "dp",
      "visualize": false,
      "log_level": null,
      "config_path": null
    }
  ]
}

Global Settings

Field	Description
`paths_cfg`	Path to tool_paths.json (optional, defaults to `config/tool_paths.json`)
`output_dir`	Shared output directory for all runs
`keep_ir`	Keep intermediate IR files for debugging
`run_name`	Name for batch output folder

Per-Run Settings

Field	Default	Description
`model_path`	(required)	Path to ONNX model
`t`	8	Macro group size
`k`	16	Macro group number
`b`	16	NoC bandwidth (flit size)
`c`	64	Number of cores
`batch_size`	8	Inference batch size
`strategy`	dp	Partition strategy
`visualize`	false	Visualize partition result
`log_level`	null	Logging level (null uses default)
`config_path`	null	Custom hardware config (null auto-resolves)
`profile`	""	Profiling: `""` (disabled), `"from-config"`, or path to config

Model paths can be absolute or relative to the batch file location.

Hardware Parameters

CIMFlow's hardware model is controlled by four primary parameters:

T-t

Macro group size — number of macros per group

K-k

Macro group number — number of macro groups per core

B-b

Bandwidth — NoC flit size in bytes

C-c

Core count — number of parallel execution cores

These parameters affect:

Compilation: Operator partitioning and memory allocation
Simulation: Timing accuracy and resource utilization

For theoretical background and architectural details, see Hardware Abstractions.

Auto-Resolution

When --config is not specified, CIMFlow automatically resolves the hardware configuration based on T and B values. Configuration files are named config-mg{T}-b{B}.json in the config directory.

# Uses config-mg8-b16.json automatically
cimflow run pipeline -m model.onnx -o output -t 8 -b 16

Hardware Configuration Files

Hardware configuration files define the complete accelerator architecture.

Profiling Configuration

Enable profiling for detailed performance breakdowns.

# Use profiler config from tool_paths.json
cimflow run pipeline -m model.onnx -o output --profile=from-config

# Use custom profiler config
cimflow run pipeline -m model.onnx -o output --profile=path/to/profiler.json

Profiler Configuration File

profiler_config.json

{
  "profiling": false,
  "report_to_json": true,
  "json_flat": true,
  "json_file": "../report/profiling.json",
  "hardware_profiler_config": {
    "profiling": false,
    "record_timing_segments": false,
    "each_core_profiling": false,
    "report_level_cnt_": 2
  },
  "inst_profiler_config": {
    "single_inst_profiling": false,
    "inst_type_profiling": false,
    "inst_group_profiling": false,
    "inst_groups": []
  }
}

Field	Description
`profiling`	Enable profiling globally
`report_to_json`	Output results in JSON format
`json_file`	Path for JSON output file
`hardware_profiler_config`	Hardware-level profiling options
`inst_profiler_config`	Instruction-level profiling options

Output Directory Structure

CIMFlow organizes output based on the --keep-ir flag:

Default Output (--no-keep-ir)

output/
└── 20250101_120000_example_run/       # Timestamped folder with run_name
    ├── simulation_report_*.txt        # Per-model simulation reports
    ├── profiling_*.json               # Profiling output (if enabled)
    └── run_info.json                  # Batch run metadata

Debug Output (--keep-ir)

output/
└── 20250101_120000_resnet18/          # Timestamped folder with model name
    ├── instructions_*.json            # CG-level instructions
    ├── isa_instructions_*.json        # OP-level ISA output
    ├── simulation_report.txt          # Simulation report
    ├── profiling_*.json               # Profiling output (if enabled)
    └── run_info.json                  # Run metadata

For log level options and debugging flags, see CLI Reference.

Configuration

tool_paths.json

batch.json

Chip Configuration

CIM Unit Configuration

Memory Configuration

Power Configuration

Simulation Configuration

On this page