cuda-swift - GitLite

KO | EN

gitlite — search

#javascript #hacktoberfest #llm #go #android #machine-learning #rust #nodejs #awesome #kubernetes #pytorch #security #ios #mcp

cuda-swift

(★ 82)

Parallel Computing Library for Linux and macOS & NVIDIA CUDA Wrapper

파일 탐색기

최종 버전 다운로드 (.zip)

- - CUDA.xml
- CUDA.iml
- misc.xml
- modules.xml
- vcs.xml
- xcode.xml
- - contents.xcworkspacedata
- CuBLAS_Info.plist
- CuBLASTests_Info.plist
- CUDADriver_Info.plist
- CUDADriverTests_Info.plist
- CUDARuntime_Info.plist
- CUDARuntimeTests_Info.plist
- NVRTC_Info.plist
- NVRTCTests_Info.plist
- project.pbxproj
- Warp_Info.plist
- WarpTests_Info.plist
- - module.modulemap
- - Info.plist
  - module.modulemap
- - module.modulemap
- - module.modulemap
- - module.modulemap
- - module.modulemap
- README.md
- cublas_example.swift
- kernel_example.swift
- - BLAS.swift
  - Error.swift
  - Level1.swift
  - Level2.swift
  - Level3.swift
  - Types.swift
- - CHandleCarrier.swift
  - Context.swift
  - Device.swift
  - Driver.swift
  - Error.swift
  - Function.swift
  - Memory.swift
  - Module.swift
  - PTX.swift
  - Stream.swift
- - Device.swift
  - Error.swift
  - Function.swift
  - Memory.swift
  - Stream.swift
- - Compiler.swift
  - Error.swift
  - Option.swift
- - Array.swift
  - ArrayBuffer.swift
  - CollectionOperations.swift
  - DeviceTypes.swift
  - KernelDataTypes.swift
  - KernelManager.swift
  - KernelSources.swift
  - Value.swift
  - ValueBuffer.swift
- - CuBLASTests.swift
- - CUDADriverTests.swift
- - CUDARuntimeTests.swift
- - NVRTCTests.swift
- - WarpTests.swift
- LinuxMain.swift
.gitignore
LICENSE
Makefile
Package.pins
Package.swift
README.md

# 설치 가이드

1. 코드 내려받기

git clone https://github.com/rxwei/cuda-swift

깃허브에서 프로젝트 코드 전체를 내 컴퓨터로 내려받습니다.

cd cuda-swift

방금 내려받은 프로젝트 폴더 안으로 이동합니다.

2. Make

보통 추천

사전 준비물

Git GitHub에서 프로젝트 코드를 내려받으려면 필요합니다.
Make Linux/macOS는 보통 기본 설치되어 있습니다. Windows는 별도 설치(예: MSYS2, WSL)가 필요합니다.

make

생성된 빌드 설정을 바탕으로 실제 컴파일을 진행해 실행 파일을 만듭니다.

✅ 에러 없이 끝나면 성공입니다. 생성된 실행 파일을 직접 실행해보세요.

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# cuda-swift

This project provides a native Swift interface to CUDA with the following
modules:

- [x] CUDA Driver API `import CUDADriver`
- [x] CUDA Runtime API `import CUDARuntime`
- [x] NVRTC - CUDA Runtime Compiler `import NVRTC`
- [x] cuBLAS - CUDA Basic Linear Algebra Subprograms `import CuBLAS`
- [x] Warp - GPU Acceleration Library `import Warp` ([Thrust](https://github.com/thrust/thrust) counterpart)

Any machine with CUDA 7.0+ and a CUDA-capable GPU is supported. Xcode Playground
is supported as well. Please refer to [Usage](#Usage)
and [Components](#Components).

## Quick look

### Value types

CUDA Driver, Runtime, cuBLAS, and NVRTC (real-time compiler) are wrapped in
native Swift types. Warp provides higher level value types, `DeviceArray` and
`DeviceValue`, with copy-on-write semantics.

```swift
import Warp

/// Initialize two arrays on device
var x: DeviceArray<Float> = [1.0, 2.0, 3.0, 4.0, 5.0]
let y: DeviceArray<Float> = [1.0, 2.0, 3.0, 4.0, 5.0]

/// Scalar map operations
x.incrementElements(by: 2) // x => [2.0, 3.0, 4.0, 5.0, 6.0] on device
x.multiplyElements(by: 2) // x => [2.0, 4.0, 6.0, 8.0, 10.0] on device

/// Addition
x.formElementwise(.addition, with: y) // x => [3.0, 6.0, 9.0, 12.0, 15.0] on device

/// Dot product
x • y // => 165.0

/// Sum
x.sum() // => 15

/// Absolute sum
x.sumOfAbsoluteValues() // => 15

/// Transform by 1-place math functions
x.transform(by: .sin)
x.transform(by: .tanh)
x.transform(by: .ceil)

/// Elementwise operation
x.formElementwise(.addition, with: y)
x.formElementwise(.subtraction, with: y)
x.formElementwise(.multiplication, with: y)
x.formElementwise(.division, with: y)

/// Fill with the same value
var z = y
z.fill(with: 10.0)

/// Composite assignment
x.assign(from: .subtraction, left: y, multipliedBy: 100.0, right: z)
```

### Real-time compilation

#### Compile source string to PTX
```swift
import NVRTC
import CUDADriver
import Warp

let source: String =
  + "extern \"C\" __global__ void saxpy(float a, float *x, float *y, float *out, int n) {"
  + "    size_t tid = blockIdx.x * blockDim.x + threadIdx.x;"
  + "    if (tid < n) out[tid] = a * x[tid] + y[tid];"
  + "}";
let ptx = try Compiler.compile(source)
```

#### JIT-compile and load PTX using Driver API within a device context
```swift
try Device.main.withContext { context in
    let module = try Module(ptx: ptx)
    let function = module.function(named: "saxpy")!
    
    let x: DeviceArray<Float> = [1, 2, 3, 4, 5, 6, 7, 8]
    let y: DeviceArray<Float> = [2, 3, 4, 5, 6, 7, 8, 9]
    var result = DeviceArray<Float>(capacity: 8)

try function<<<(1, 8)>>>[.float(1.0), .constPointer(to: x), .constPointer(to: y), .pointer(to: &result), .int(8)]
    /// result => [3, 5, 7, 9, 11, 13, 15, 17] on device
}
```

## Package Information

Add a dependency:

```swift
.Package(url: "https://github.com/rxwei/cuda-swift", majorVersion: 1)
```

You may use the `Makefile` in this repository for you own project. No extra path
configuration is needed.

Otherwise, specify the path to your CUDA headers and library at `swift build`.

#### macOS
```
swift build -Xcc -I/usr/local/cuda/include -Xlinker -L/usr/local/cuda/lib
```

#### Linux
```
swift build -Xcc -I/usr/local/cuda/include -Xlinker -L/usr/local/cuda/lib64
```

## Components

### Core

- [x] CUDADriver - CUDA Driver API
    - [x] `Context`
    - [x] `Device`
    - [x] `Function`
    - [x] `PTX`
    - [x] `Module`
    - [x] `Stream`
    - [x] `Unsafe(Mutable)DevicePointer<T>`
    - [x] `DriverError` (all error codes from CUDA C API)
- [x] CUDARuntime - CUDA Runtime API
    - [x] `Unsafe(Mutable)DevicePointer<T>`
    - [x] `Device`
    - [x] `Stream`
    - [x] `RuntimeError` (all error codes from CUDA C API)
- [x] NVRTC - CUDA Runtime Compiler
    - [x] `Compiler`
- [x] CuBLAS - GPU Basic Linear Algebra Subprograms (in-progress)
    - [x] Level 1 BLAS operations
    - [x] Level 2 BLAS operations (GEMV)
    - [x] Level 3 BLAS operations (GEMM)
- [x] Warp - GPU Acceleration Library ([Thrust](https://github.com/thrust/thrust) counterpart)
    - [x] `DeviceArray<T>` (generic array in device memory)
    - [x] `DeviceValue<T>` (generic value in device memory)
    - [x] Acclerated vector operations
    - [x] Type-safe kernel argument helpers

### Optional

- [x] Swift Playground
  - CUDADriver works in the playground. But other modules cause the "couldn't lookup
    symbols" problem for which we don't have a solution until Xcode is fixed.
  - To use the playground, open the Xcode workspace file, and add a library for
    every modulemap under `Frameworks`.

## Dependencies

- [CCUDA (CUDA C System Module)](https://github.com/rxwei/CCUDA)

## License

MIT License

CUDA is a registered trademark of NVIDIA Corporation.