PPEM Stereo Matching - x86 & C6678 DSP Parallelization

Parallel implementation of a 9-stage stereo matching algorithm using OpenMP on x86 and TI C6678 DSP platforms.

Course: PPEM (Parallel Programming for Embedded Multicore) Institution: INSA Rennes Team Size: 3 members Duration: 3 weeks - 10 hours Target Platforms: x86 (Windows/Linux), TI C6678 8-core DSP

Project Overview

This project parallelizes a depth-from-stereo algorithm that processes stereo camera images to generate depth maps. The algorithm executes 9 sequential stages per frame, with parallelization opportunities in the computationally intensive stages.

Algorithm Pipeline

YUV Read → YUV2RGB → RGB2Gray → Census → Cost Construction (60 disparities)
→ Aggregate Costs → Disparity Select → Median Filter → MD5 Validation → Display

Objectives

1. ✅ Parallelize on x86 using OpenMP (≥2x speedup)
2. ✅ Port to C6678 DSP and parallelize (≥2x speedup)
3. ✅ Cross-platform validation (identical MD5 hashes)
4. ✅ Professional report (≤5 pages, English)
5. ✅ Live demonstration on both platforms

Quick Start

Prerequisites

For x86 development:

• Windows: Visual Studio 2022, CMake ≥3.15
• Linux: GCC/Clang with OpenMP, CMake, SDL2 development libraries

For C6678 development:

• Code Composer Studio (CCS) v12.x
• TI C6000 compiler tools
• C6678 hardware access (EVM or lab setup)

Build Instructions

x64 (Windows)

# Generate Visual Studio solution
cmake -G "Visual Studio 17 2022" -B bin

# Build
cmake --build bin --config Release

# Run
cd bin/Release
./stereo.exe

x86 (Linux)

# Generate Makefiles
cmake -B bin

# Build
cmake --build bin

# Run
cd bin
./stereo

C6678 DSP

See docs/issues/epic_04_c6678_sequential/ for detailed CCS project setup instructions.

Project Structure

PPEM-project/
├── src/                    # Algorithm implementation (9 stages)
│   ├── main.c             # Main loop, orchestrates pipeline
│   ├── costConstruction.c # Hotspot #1 for parallelization
│   ├── aggregateCost.c    # Hotspot #2 for parallelization
│   └── ...                # Other algorithm stages
├── include/               # Header files
│   ├── params.h          # Image dimensions, disparity range
│   └── ...               # Function declarations
├── dat/                   # Test stereo image datasets
├── lib/                   # SDL2 libraries (x86 only)
├── bin/                   # Build outputs, generated by CMake
├── CMakeLists.txt        # x86 build configuration
└── README.md             # This file

Development Workflow

1. Pick an Issue

Browse issues in docs/issues/ organized by epic:

• Epic 1: Foundation (setup, validation)
• Epic 2: x86 Baseline (profiling, documentation)
• Epic 3: x86 Parallel (OpenMP implementation)
• Epic 4: C6678 Sequential (DSP porting)
• Epic 5: C6678 Parallel (DSP parallelization)
• Epic 6: Validation & Report (cross-platform, demo, report)

2. Create a Branch

git checkout main
git pull origin main
git checkout -b epic-XX/issue-YY-description

3. Implement & Test

• Follow subtasks in the issue file (docs/issues/epic_XX_name/issue_YY.md)
• Test locally: build, run, validate MD5
• Commit frequently with clear messages

4. Submit Pull Request

• Use PR template (.github/PULL_REQUEST_TEMPLATE.md)
• Request code review from ≥1 team member
• Address feedback, then merge

5. Close Issue & Update Milestone

Key Technologies

• OpenMP: Parallel programming API (thread-level parallelism)
• CMake: Cross-platform build system generator
• SDL2: Simple DirectMedia Layer (visualization on x86)
• TI C6678: 8-core DSP (1.25 GHz C66x cores, 4 MB L2 cache)
• Code Composer Studio: TI's IDE for embedded development

Performance Targets

Platform	Sequential Time	Parallel Time	Speedup	Target
x86 (8 cores)	Baseline	TBD	≥2.0x	Required
C6678 (8 cores)	Baseline	TBD	≥2.0x	Required

Critical validation: MD5 hashes must match across:

• Sequential vs. parallel (same platform)
• x86 vs. C6678 (cross-platform)

Milestones

• M1 - Foundation Complete (Week 1, Dec 1)

  • Development environments working
  • Sequential code validated with baseline MD5

• M2 - x86 Parallelization Complete (Week 2-3, Dec 12)

  • x86 parallelized with ≥2x speedup
  • MD5 validation passing

• M3 - C6678 Sequential Port Complete (Week 2-3, Dec 12)

  • C6678 compiles and runs sequentially
  • MD5 matches x86 baseline

• M4 - C6678 Parallelization Complete (Week 3-4, Dec 19)

  • C6678 parallelized with ≥2x speedup
  • MD5 validation passing

• M5 - Project Delivery (Week 4-5, Dec 27)

  • Report complete (≤5 pages)
  • Demo rehearsed and successful
  • Code packaged for submission

See docs/project_milestones.md for detailed timeline.

Testing & Validation

MD5 Checksum Validation

The algorithm computes an MD5 hash of the output depth map for correctness validation.

Baseline MD5: Stored in bin/md5.txt after first sequential run (Issue #4)

Validation rules:

• Parallel code must produce identical MD5 as sequential
• C6678 must produce identical MD5 as x86
• Use schedule(static) in OpenMP to ensure determinism