Skip to content

README.md

Latest commit

 

History

History
74 lines (48 loc) · 3.75 KB

README.md

File metadata and controls

74 lines (48 loc) · 3.75 KB

Image Processing

Image convolution with configurable filters on CPU (sequential / parallel) and GPU via OpenCL. Includes a MailboxProcessor-based streaming pipeline for batch processing.

Backends

Implementation Function Parallelism
CPU Sequential applyFilter Single-threaded pixel loop
CPU Parallel (per-pixel) applyFilterCpuParallel Array.Parallel.iter over flat pixel array
CPU Parallel (per-row) applyFilterCpuParallelRows Array.Parallel.iter over rows, sequential within each row
GPU (any OpenCL device) applyFiltersGPU OpenCL kernel, configurable local work size

Streaming mode: a MailboxProcessor-based pipeline that loads images from a directory, distributes them across multiple filter workers (each potentially on a different platform), and saves results — all concurrently.

CLI: Argu-based argument parser supports --input, --output, --platform (6 backends), --work-group-size, and --workers for streaming.

Benchmarks

The benchmarks/ImageProcessing.Benchmarks/ project uses BenchmarkDotNet to measure filter processing times across all available backends — CPU sequential, CPU parallel (per-pixel and per-row), and GPU (POCL, Nvidia, Intel GPU) — for square images from 100×100 up to 8000×8000 pixels.

Benchmark classes

Class Device param Extra params
CpuFilterBench CPUSequential, CPUParallel, CPUParallelRows —
GpuFilterBench POCL, Nvidia, IntelGPU LWS: 8, 16, 32, 64, 128, 256

Common parameter across all classes:

  • Size — image side in pixels: 100, 200, 500, 1000, 2000, 4000, 8000

Design

  • Image generation: each benchmark generates a random square image (deterministic seed Random 42) in [GlobalSetup] — excluded from measurement
  • Filter: all benchmarks use gaussianBlurKernel (5×5 normalized) — a module-level constant, not recreated per invocation
  • Measurement: the [Benchmark] method applies exactly one filter pass and discards the result; only the processing time is captured
  • GPU setup: ClContext and GPU applier are created once in [GlobalSetup]; the device name is printed at startup
  • No redundant combinations: LWS is only parameterized for GPU benchmarks — CPU benchmarks have zero useless LWS configurations

How to run

# Full run (interactive menu selects CPU or GPU benchmarks):
dotnet run -c Release --project benchmarks/ImageProcessing.Benchmarks

# Quick smoke test (CPU only, ShortRun):
dotnet run -c Release --project benchmarks/ImageProcessing.Benchmarks -- --job short --filter '*CpuFilterBench*'

# GPU benchmarks:
dotnet run -c Release --project benchmarks/ImageProcessing.Benchmarks -- --filter '*GpuFilterBench*'

# CPU benchmarks:
dotnet run -c Release --project benchmarks/ImageProcessing.Benchmarks -- --filter '*CPUParallel*'

BenchmarkDotNet passes remaining CLI arguments (like --filter, --job, --stopOnFirstError) through to its own parser. Results are exported as CSV, Markdown, and HTML to BenchmarkDotNet.Artifacts/results/.

Analysis script

The Python script benchmarks/analyze_benchmarks.py reads the CSV results and generates two comparison plots:

  • GPU work-group size comparison (left): compares Intel UHD Graphics 620, NVIDIA GeForce MX150, and POCL (CPU OpenCL) across all work-group sizes (8–256) at the largest image (8000×8000)
  • CPU vs best GPU configurations (right): compares all three CPU variants (sequential, pixel-parallel, rows-parallel) against the best-performing LWS per GPU device across all image sizes
python benchmarks/analyze_benchmarks.py

Example of plot: Benchmark comparison