rifebench

Focused benchmarking tool for RIFE interpolation models running on VapourSynth R73+ with vs-mlrt.
Automates test execution, NVIDIA GPU telemetry collection, and Markdown report generation.

• Not universal — targets only RIFE models for simplicity and comparability.
• Windows + Linux support.
• Four components: Runner → Parser → GPU Sampler → Reporter.

---

Installation

Download latest release archive Latest release
Unpack .zip to python dir.

pip install rifebench

Required (auto-installed):

• pynvml — NVIDIA GPU telemetry

Optional:

pip install ffmpeg-python   # source file media-info probing

If use portable

1. Download:

   • rifebench
   • pynvml
   • ffmpeg-python
   • ffmpeg-python dependency future

2. Open portable python dir:

   • Unpack rifebench_vXXX.zip
   • Copy/move downloaded pynvml_XXX.whl and ffmpeg-python_XXX.whl to ./wheel

3. Install module:

python pip.pyz install --no-index --find-links=./wheel/pynvml
python pip.pyz install --no-index --find-links=./wheel/ffmpeg-python

4. Run(while in the Python directory):

# cmd
python -m rifebench.src --script ./bench.vpy --source D:/clip.mkv --models 425

# powershell
.\python -m rifebench.src --script ./bench.vpy --source D:/clip.mkv --models 425,4171

Tip

For portable use rifebench.src instead of rifebench

---

Quick Start

Minimal (vspipe mode)

python -m rifebench --script ./bench.vpy --source D:/clip.mkv --models 47,425,417 

FFmpeg pipe mode

python -m rifebench --script ./bench.vpy --source D:/clip.mkv --models 47,425 --mode ffmpeg

Full example

python -m rifebench \
  --script ./my_bench.vpy \
  --source D:/clip.mkv \
  --models 47,425,417 \
  --mode vspipe \
  --multi 2 \
  --rife_v=2 \
  --streams 2 \
  --trt_fp16 1 \
  --format RGBH \
  --trt_out_fmt=1 \
  --use_cuda_graph 1 \
  --sc_threshold=0.10 \
  --gpu_id 0 \
  --warmup_frames 1000 \
  --warmup_runs 1 \
  --tests 3 \
  --shuffle 1 \
  --cooldown_s 4 \
  --output_dir ./bench_results \
  --bench_name my_bench \
  --period_ms 100

---

CLI Reference

Parameter	Type	Default	Description
--source	path	required	Path to source video file
--models	string	required	Comma-separated RIFE model IDs, e.g. 47,425,417
--script	path	required	Path to VapourSynth script
--mode	string	vspipe	vspipe (render only) or ffmpeg (pipe through encoder)
--format	string	RGBH	Pixel format: RGBS (fp32) or RGBH (fp16)
--multi	int	2	Frame multiplier passed to RIFE
--rife_v	int	2	RIFE version .onnx
--streams	int	2	Inference streams (vs-mlrt)
--trt_fp16	0┆1	1	Enable TensorRT FP16
--trt_out_fmt	0┆1	1	output_format 0: fp32, 1: fp16
--use_cuda_graph	0┆1	1	Enable CUDA Graph capture
--gpu_id	int	0	CUDA device index (exits if not found)
--sc_threshold	FLOAT	0.10	Scene-cut threshold for RIFE
--warmup_frames	int	1000	Frames for warmup (0 = full clip)
--warmup_runs	int	1	Warmup iterations (results discarded)
--tests	int	1	Test iterations per model
--shuffle	0┆1	1	Randomize model order between repetitions
--cooldown_s	int	4	Pause between runs (seconds)
--output_dir	path	./bench_results	Output directory
--bench_name	string	bench_YY-MM-DD_HH-MM	Report file name (auto-suffixed _01, _02… on collision)
--period_ms	int	100	GPU sampling interval (ms)
--command	string	optional	Custom ffmpeg command (part after |). User is responsible for correctness
--ffmpeg_path	path	ffmpeg	Path to ffmpeg binary
--vspipe_path	path	vspipe	Path to vspipe binary
--s_frame	int	optional	First output frame (vspipe -s N)
--e_frame	int	optional	Last output frame (vspipe -e N)

sc_threshold

Scene-cut threshold for RIFE (lower=more sensitive, higher=less sensitive).
Typical: 0.08-0.10 anime, 0.10-0.14 live-action, "0.14-0.20 noisy sources

---

Modes

1. vspipe (default)

Renders frames via vspipe only, discarding output (--). All parameters are forwarded to the .vpy script:

vspipe -c y4m script.vpy -- \
  -a source=D:/clip.mkv \
  -a model=47 \
  -a gpu_id=0 \
  -a format=RGBH \
  -a multi=2 \
  -a rife_v=2 \
  -a streams=2 \
  -a trt_fp16=1 \
  -a trt_out_fmt=1 \
  -a use_cuda_graph=1 \
  -a sc_threshold=0.10

Parsed output: final stdout line → Output N frames in N seconds (N fps)

2. ffmpeg

Pipes vspipe y4m output into ffmpeg:

vspipe -c y4m script.vpy - -a ... | ffmpeg -benchmark -i - \
  -c:v hevc_nvenc -preset p5 -cq 18 \
  -profile:v main10 -pix_fmt p010le -f null NUL

• If --command is not provided, the default command above is used.
• If --command is provided, it replaces the part after |.
• The default uses hevc_nvenc / main10 — may not match all clip formats.

Parsed output:

• Progress lines → fps collected into array for running median
• After [out# → frames, fps, rtime extracted from final summary

---

.vpy Script Example

Open a sample benchmark file

rifebench passes parameters to your script via vspipe -a key=value. Your script must read them from globals(). Minimal example:

...
# All values arrive as strings via `vspipe -a key=value`.

source         = globals().get('source', None)
model          = int(globals().get("model", "47"))
gpu_id         = int(globals().get("gpu_id", "0"))
fmt            = globals().get("format", "RGBH")       # "RGBS" | "RGBH"
multi          = int(globals().get("multi", "2"))
streams        = int(globals().get("streams", "2"))
trt_fp16       = bool(int(globals().get("trt_fp16", "1")))
use_cuda_graph = bool(int(globals().get("use_cuda_graph", "1")))
...

Key point: globals() is a built-in dict populated by vspipe -a flags. All values are strings — cast them in your script.

---

Architecture

┌──────────┐     ┌──────────┐     ┌────────────-─┐     ┌──────────┐
│  Runner  │────▶│  Parser  │────▶│ GPU Sampler  │────▶│ Reporter │
└──────────┘     └──────────┘     └────────────-─┘     └──────────┘

Component	Module	Responsibility
Runner	runner.py	Builds commands, launches vspipe/ffmpeg subprocesses, manages warmup/test iterations, shuffle, cooldown
Parser	parser.py	Reads stdout/stderr line-by-line on the fly. Extracts frames, fps, time_s/rtime. Probes source & clip info
GPU Sampler	sampler.py	Background thread polling GPU via pynvml at period_ms intervals during test runs
Reporter	reporter.py	Writes incremental .md reports with specs, summary (medians), per-run details. Auto-suffixes filenames

---

Output Format

Reports are saved as {bench_name}.md in output_dir.

Specs Section

Collected automatically:

• OS — platform info (e.g. Windows 11 24H2)
• GPU model, Driver, CUDA version — via nvidia-smi / pynvml
• TDP (Power Limit), Total/Used/Free VRAM — via pynvml
• CPU — name, cores, clock
• Bench Mode — vspipe / ffmpeg
• VapourSynth version — parsed from vspipe -v (e.g. Core R73)
• Source — Width, Height, Frames, FPS, Format (via ffmpeg-python if available, else N/A)
• Clip — Width, Height, Frames, FPS, Format Name (via vspipe -i)

Summary Table

Aggregated per-model results from all successful iterations:

Model	FPS (avg┆median┆min)	Time s (avg┆mdn┆max)	GPU avg %	VRAM MB (mdn┆max)	PWR W (mdn┆max)	rxpci_max MB/s	txpci_max MB/s	Frames	Runs (tot┆succ)

All Runs Table

Per-run detail with GPU aggregates computed from samples within that run:

Model	Run	Frames	FPS	Time s	GPU util % (avg┆p99)	MEM util % (avg┆p99)	VRAM Mb (avg┆max)	PWR W (avg┆max)	rxpci Mb/s (avg┆max)	txpci Mb/s (avg┆max)	gpu_clock	mem_clock

• gpu_util_avg / gpu_util_p99 — mean and 99th percentile of samples within the run
• vram_max_mb — maximum across samples within the run
• gpu_clock / mem_clock — median of samples within the run

Errors Section

## Errors
**Error model=47 run=2** | vspipe exited with code 1: "Filter error: ..."

---

GPU Sampling

During each test run (not warmup), a background thread polls the GPU via pynvml:

Metric	Source	Unit
gpu_util	nvmlDeviceGetUtilizationRates	%
mem_util	nvmlDeviceGetUtilizationRates	%
power_w	nvmlDeviceGetPowerUsage / 1000	W
vram_mb	nvmlDeviceGetMemoryInfo.used / 1048576	MB
rxpci_mb_s	nvmlDeviceGetPcieThroughput (RX)	MB/s
txpci_mb_s	nvmlDeviceGetPcieThroughput (TX)	MB/s
gpu_clock	nvmlDeviceGetClockInfo (SM)	MHz
mem_clock	nvmlDeviceGetClockInfo (MEM)	MHz

Sampling interval is configurable via --period_ms (default: 100 ms).

---

Error Handling

Condition	Result
Process exits with non-zero code	Iteration = failed
Cannot extract frames / fps / time_s or rtime	Iteration = failed
Cannot collect GPU metrics	Reported as N/A, iteration still successful
Failed iterations	Excluded from Summary, listed in Errors section

---

Project Structure

rifebench/
│   src/
│   ├── __init__.py        # Package init, version
│   ├── __main__.py        # Entry point (python -m rifebench)
│   ├── cli.py             # Argument parser
│   ├── runner.py          # Process execution & orchestration
│   ├── parser.py          # Output parsing & media probing
│   ├── sampler.py         # GPU telemetry (pynvml)
│   └── reporter.py        # Markdown report generation
├── pyproject.toml     # Package metadata
├── README.md          # This file
└── rifebench.html     # Offline documentation (EN/RU)

MIT License