GPU Experiments - Metal Compute

Progressive GPU programming exercises using Metal on Apple Silicon.

Quick Start

# See all available commands
make help

# Build and run any problem using its prefix
make go-00          # Mandelbrot
make go-01          # Parallel scan

Structure

Each problem follows the pattern:

1. Sequential baseline - CPU reference implementation
2. v1-naive - Direct GPU translation, identifies bottlenecks
3. v2-optimized - Address memory/compute inefficiencies
4. v3-advanced - Architecture-specific optimizations

The framework provides (in lib/):

• MetalContext: Zero-boilerplate Metal setup, device management
• Timer: GPU-aware performance measurement with bandwidth/FLOPS tracking
• Visualizer: Debug arrays as heatmaps, correctness verification, access pattern analysis

Problem Progression

0. Mandelbrot - Embarrassingly parallel warm-up, visual debugging
1. Parallel Scan - Foundation for everything else
2. Bitonic Sort - Fixed comparison network, visualizable parallelism
3. Matrix Transpose - Bank conflicts, memory coalescing
4. Reduction - Warp divergence, atomic operations
5. Histogram - Atomic contention, privatization
6. Sparse Matrix - Irregular workloads, load balancing
7. Convolution - Constant memory, texture cache
8. Einstein Summation - Tensor contractions, index arithmetic

Implementation Notes

The framework uses direct Objective-C++ Metal API instead of metal-cpp for zero dependencies. Key abstractions:

• MetalContext: Device setup, shader loading, pipeline creation
• Timer: GPU-aware performance measurement with bandwidth tracking
• Visualizer: Array heatmaps, correctness verification, access pattern analysis
• ScopedBuffer: RAII buffer management with automatic cleanup

Key Concepts to Track

• Occupancy: Active threads vs hardware maximum (1024 threads/threadgroup on M2)
• Memory Bandwidth: Achieved vs theoretical (450 GB/s on M2 Max)
• Bank Conflicts: Threadgroup memory contention patterns
• Divergence: SIMD efficiency within simdgroups (32 threads on Apple Silicon)
• Coalescing: Sequential vs strided memory access

Profiling

# Quick benchmark
make benchmark

# Detailed profiling
xcrun xctrace record --template 'Metal System Trace' --launch ./build/problems/01-parallel-scan/benchmark

Philosophy

каждая оптимизация должна быть видимой - visualize access patterns, measure everything, understand why performance changes. The goal isn't just making things fast, but understanding exactly why they're fast.