Skip to content

GettingStarted.md

Latest commit

 

History

History
80 lines (55 loc) · 1.85 KB

GettingStarted.md

File metadata and controls

80 lines (55 loc) · 1.85 KB
Raw

Getting Started with Metal Compute Kernels

Installation

Add the following to your Package.swift:

.package(url: "https://github.com/PawScale/kernels.git", from: "1.0.0")

Then add it as a dependency to your target:

.target(
    name: "YourTarget",
    dependencies: ["Kernels"]
)

Quick Start

1. Write a Metal Kernel

Create a .metal file with your compute kernel:

kernel void vector_add(
    device const float* a [[buffer(0)]],
    device const float* b [[buffer(1)]],
    device float* c [[buffer(2)]],
    uint id [[thread_position_in_grid]]
) {
    c[id] = a[id] + b[id];
}

2. Load and Execute in Swift

import Kernels

let kernel = MetalKernel(name: "vector_add")
let result = kernel.execute(inputs: [vectorA, vectorB])

Architecture

The framework provides:

  • MetalKernel: Main class for loading and executing Metal compute kernels
  • GPUBuffer: Wrapper for GPU memory management
  • ComputeContext: GPU device and command queue management

Performance Tips

  1. Batch operations to reduce kernel launch overhead
  2. Use appropriate grid and threadgroup sizes
  3. Minimize data transfer between CPU and GPU
  4. Profile your kernels using Xcode's Metal debugger

Examples

See the Examples directory for complete working examples.

Troubleshooting

Kernel not found

Ensure your Metal library is properly bundled and the kernel name matches exactly.

GPU memory errors

Check that buffer sizes match your kernel's expectations and that you're not exceeding GPU memory limits.

Performance degradation

Use Xcode's Metal profiler to identify bottlenecks and optimize accordingly.

Further Reading