luxfi/gpu

mirror of https://github.com/luxfi/gpu.git synced 2026-10-30 16:23:49 +00:00

Files

T

zeekay b9eb573818 refactor: consolidate GPU types and add session support

- Remove backend.go (duplicate Backend, Device types)
- Keep types in types.go with CGO-compatible fields
- Add Float64 to Dtype enum in ops.go
- Add newSession for CGO and non-CGO builds
- Fix gpu_cpu.go build tag (!cgo instead of cgo)
- Add cgoSessionHandle and cpuSessionHandle
- Remove deprecated metal/, mlx files

2026-01-09 14:50:48 -08:00

3.4 KiB

Raw Permalink Blame History

Lux GPU - Unified GPU Acceleration

Overview

Single package for GPU-accelerated operations with switchable backends:

Metal - Apple Silicon (macOS/iOS)
CUDA - NVIDIA GPUs (Linux/Windows)
Dawn - WebGPU via Dawn/wgpu (cross-platform)
CPU - SIMD-optimized fallback (always available)

Architecture

lux/gpu/
├── gpu.go           # Main API: Session, Backend switching
├── backend.go       # Backend enum and detection
├── device.go        # DeviceInfo
├── tensor.go        # Tensor type and operations
├── errors.go        # Error types
├── ops.go           # Unified compute operations
│
├── crypto/          # Cryptographic accelerations
│   ├── bls.go       # BLS12-381 operations
│   ├── poseidon.go  # Poseidon hash
│   ├── msm.go       # Multi-scalar multiplication
│   └── kzg.go       # KZG commitments
│
├── fhe/             # Fully Homomorphic Encryption
│   ├── tfhe.go      # TFHE gates
│   ├── ckks.go      # CKKS approximate arithmetic
│   └── ntt.go       # Number Theoretic Transform
│
├── ml/              # Machine Learning
│   ├── matmul.go    # Matrix multiplication
│   ├── attention.go # Attention mechanisms
│   └── quantize.go  # Quantization ops
│
└── internal/
    └── capi/        # CGO bindings to liblux_accel

Usage

import "github.com/luxfi/lux/gpu"

// Auto-detect best backend
sess := gpu.DefaultSession()

// Or explicitly select backend
sess, _ := gpu.NewSession(gpu.WithBackend(gpu.Metal))
sess, _ := gpu.NewSession(gpu.WithBackend(gpu.CUDA))
sess, _ := gpu.NewSession(gpu.WithBackend(gpu.Dawn))

// Runtime switching
sess.SetBackend(gpu.Metal)

// Check available backends
for _, b := range gpu.AvailableBackends() {
    fmt.Printf("%s: %v\n", b.Name, b.Available)
}

// Operations
tensor := sess.Zeros([]int{1024, 1024}, gpu.Float32)
result := sess.MatMul(a, b)

// Crypto operations
sess.Crypto().BLSVerifyBatch(sigs, msgs, pks)
sess.Crypto().PoseidonHash(inputs)

// FHE operations
sess.FHE().NTT(poly)
sess.FHE().TFHEBootstrap(ct)

Backend Selection Priority

Environment: LUX_GPU_BACKEND=metal|cuda|dawn|cpu
Explicit: gpu.WithBackend(gpu.Metal)
Auto-detect (in order):
- Metal (if macOS + Apple Silicon)
- CUDA (if NVIDIA GPU + driver)
- Dawn (if GPU available)
- CPU (always)

C++ Library (liblux_accel)

Single unified C library with backend plugins:

luxcpp/lux-accel/
├── include/lux/
│   ├── accel.h         # Main C API
│   ├── backend.h       # Backend types
│   ├── tensor.h        # Tensor operations
│   ├── crypto.h        # Crypto ops
│   └── fhe.h           # FHE ops
│
├── src/
│   ├── accel.cpp       # Core implementation
│   ├── backend.cpp     # Backend management
│   ├── metal/          # Metal backend
│   ├── cuda/           # CUDA backend
│   └── dawn/           # Dawn/WebGPU backend
│
└── kernels/            # GPU kernel sources
    ├── metal/*.metal
    ├── cuda/*.cu
    └── wgsl/*.wgsl

Build Configuration

# Build with all backends (auto-detect)
cmake -B build

# Specific backends
cmake -B build -DLUX_BACKEND_METAL=ON
cmake -B build -DLUX_BACKEND_CUDA=ON
cmake -B build -DLUX_BACKEND_DAWN=ON

# Go package
CGO_ENABLED=1 go build ./gpu/...

3.4 KiB Raw Permalink Blame History