composable_kernel/script/RetinaNet/README_CONVERTER.md

PyTorch to CK Profiler Conversion Tools

This directory contains two Python scripts to convert PyTorch convolution operations to CK Profiler format and execute them.

Overview

1. convert_pytorch_to_ck.py - Converts PyTorch JSON to CK Profiler configuration JSON
2. run_ck_profiler.py - Executes CK profilers for each configuration

Workflow

# Step 1: Convert PyTorch JSON to CK Profiler JSON
python convert_pytorch_to_ck.py build/test-data/conv_repros_ir.json ck_profiler_configs.json

# Step 2: Execute profilers
python run_ck_profiler.py ck_profiler_configs.json --profiler-path ./build/bin

---

Script 1: convert_pytorch_to_ck.py

Description

Converts PyTorch convolution operations from conv_repros_ir.json to CK Profiler configuration format.

Supported Operations

• aten::miopen_convolution → Forward convolution
• aten::convolution_backward → Backward data and/or backward weight

Configuration

The script uses these hardcoded settings:

• Data type: FP16 (data_type=1)
• Layout: NGCHW_GKCYX_NGKHW (layout=3)
• Index type: 32-bit (index_type=0, for forward only)
• Verification: Disabled (verify=0)
• Initialization: Integer values (init=1)
• Logging: Disabled (log=0)
• Timing: Enabled (time=1)
• Split-K: "all" (for backward operations)

Usage

# Basic usage (uses defaults)
python convert_pytorch_to_ck.py

# Specify input and output files
python convert_pytorch_to_ck.py input.json output.json

# Help
python convert_pytorch_to_ck.py --help

Default Files

• Input: build/test-data/conv_repros_ir.json
• Output: ck_profiler_configs.json

Output Format

The output JSON contains an array of configurations:

[
  {
    "operation_type": "profile_grouped_conv_fwd",
    "profiler_args": {
      "data_type": 1,
      "layout": 3,
      "index_type": 0,
      "verify": 0,
      "init": 1,
      "log": 0,
      "time": 1,
      "num_dim_spatial": 2,
      "G": 32,
      "N": 32,
      "K": 4,
      "C": 4,
      "filter_spatial": [3, 3],
      "input_spatial": [200, 200],
      "strides": [1, 1],
      "dilations": [1, 1],
      "left_pads": [1, 1],
      "right_pads": [1, 1]
    },
    "metadata": {
      "priority_rank": 1,
      "pytorch_op": "aten::miopen_convolution",
      "description": "Forward conv: G=32, N=32, K=4, C=4, [3, 3] kernel, [200, 200] input"
    }
  }
]

Example Output

Converting PyTorch convolutions to CK Profiler format...

✓ [1/78] Converted: aten::miopen_convolution (rank 1) → fwd
✓ [2/78] Converted: aten::convolution_backward (rank 2) → bwd_data
✓ [2/78] Converted: aten::convolution_backward (rank 2) → bwd_weight
...

============================================================
Conversion Summary
============================================================
Total PyTorch operations:     78
  ✓ Forward convolutions:       30
  ✓ Backward data convolutions: 35
  ✓ Backward weight convolutions: 35
  ⚠ Skipped operations:         0

Total CK profiler configs:    100
Output file:                  ck_profiler_configs.json
============================================================

Conversion completed successfully!

---

Script 2: run_ck_profiler.py

Description

Executes CK Profiler binaries for each configuration in the JSON file, with support for dry-run, filtering, and result collection.

Usage

# Run all configurations
python run_ck_profiler.py ck_profiler_configs.json

# Run first 10 only (for testing)
python run_ck_profiler.py ck_profiler_configs.json --max-ops 10

# Dry run (show commands without executing)
python run_ck_profiler.py ck_profiler_configs.json --dry-run

# Specify profiler path
python run_ck_profiler.py ck_profiler_configs.json --profiler-path ./build/bin

# Quiet mode (less verbose output)
python run_ck_profiler.py ck_profiler_configs.json --quiet

# Don't save results
python run_ck_profiler.py ck_profiler_configs.json --no-save

# Custom results output file
python run_ck_profiler.py ck_profiler_configs.json --output my_results.json

Command Line Options

Option	Description	Default
config_file	CK Profiler configuration JSON file	(required)
--profiler-path	Path to CK profiler binaries	./build/bin
--max-ops	Maximum number of operations to execute	All
--dry-run	Show commands without executing	False
--quiet	Reduce output verbosity	False
--no-save	Do not save results to JSON	False
--output	Output file for results	profiler_results.json

Example Output

Executing CK Profiler for 100 configurations...

======================================================================
[1/100] Forward conv: G=32, N=32, K=4, C=4, [3, 3] kernel, [200, 200] input
Priority Rank: 1
PyTorch Op:    aten::miopen_convolution
======================================================================
Command: ./build/bin/profile_grouped_conv_fwd 1 3 0 0 1 0 1 2 32 32 4 4 3 3 200 200 1 1 1 1 1 1 1 1

✓ SUCCESS (completed in 2.34s)

Output:
...profiler output...

...

======================================================================
Execution Summary
======================================================================
Total configurations: 100
  ✓ Successful:  95
  ✗ Failed:      5

Success rate: 95.0%
======================================================================

Results saved to: profiler_results.json

Results File Format

The profiler_results.json file contains:

{
  "timestamp": "2026-01-19T05:48:00",
  "stats": {
    "total": 100,
    "success": 95,
    "failed": 5,
    "skipped": 0
  },
  "results": [
    {
      "operation": "profile_grouped_conv_fwd",
      "description": "Forward conv: G=32, N=32, K=4, C=4, [3, 3] kernel, [200, 200] input",
      "priority_rank": 1,
      "success": true,
      "returncode": 0,
      "elapsed_time": 2.34,
      "command": "./build/bin/profile_grouped_conv_fwd ...",
      "error": null,
      "stdout": "...truncated..."
    }
  ]
}

---

Requirements

• Python 3.6+
• CK profiler binaries built in ./build/bin (or specify with --profiler-path)
• Input JSON file with PyTorch convolution operations

Notes

Layout Mapping

PyTorch uses NCHW layout, which maps to CK's NGCHW_GKCYX_NGKHW (layout=3):

• Input: [N, G, C, Hi, Wi]
• Weight: [G, K, C, Y, X]
• Output: [N, G, K, Ho, Wo]

Per-Group Channels

For grouped convolutions:

• C_per_group = C_total / groups
• K_per_group = K_total / groups

The converter automatically computes these values.

Backward Operations

The aten::convolution_backward operation can compute:

• Backward data (gradient w.r.t. input) when output_mask[0]=true
• Backward weight (gradient w.r.t. weight) when output_mask[1]=true
• Both if both flags are true

Split-K Parameter

For backward operations, split_k is set to "all" which instructs the profiler to test all split-K values: -1, 1, 2, 4, 8, 16, 32, 64, 128.

---

Troubleshooting

"Profiler executable not found"

Ensure CK profilers are built:

mkdir -p build && cd build
cmake .. -DCMAKE_BUILD_TYPE=Release
make profile_grouped_conv_fwd profile_grouped_conv_bwd_data profile_grouped_conv_bwd_weight

"Input file not found"

Check that the PyTorch JSON file exists:

ls -l build/test-data/conv_repros_ir.json

Conversion warnings

Yellow warnings indicate operations that couldn't be converted (e.g., non-convolution operations). These are expected and don't indicate errors.

---