mirror of
https://github.com/luxfi/accel.git
synced 2026-10-30 17:03:45 +00:00
zeekay
7c7258fbd0
Reads PLUGIN_PATH (canonical); honors deprecated LUX_PLUGIN_PATH for one release with deprecation log. Doc comments mention GPU_BACKEND as the canonical override. crypto_secp256k1_cgo.go: rename C symbols to match brand-neutral C-ABI in luxcpp/crypto (lux_secp256k1_* -> secp256k1_*, LUX_SECP256K1_OK -> SECP256K1_OK).
99 lines
3.2 KiB
Go
99 lines
3.2 KiB
Go
// Package accel provides GPU-accelerated operations for blockchain and ML workloads.
|
|
//
|
|
// The package supports multiple GPU backends (Metal, WebGPU, CUDA) via runtime
|
|
// plugin discovery. When built without CGO or when no backends are available,
|
|
// operations return ErrNoBackends.
|
|
//
|
|
// # Architecture
|
|
//
|
|
// accel wraps the lux-accel C++ library which provides:
|
|
// - ML operations: matmul, attention, convolution, normalization
|
|
// - Crypto operations: batch signature verification, hashing, Merkle trees
|
|
// - ZK operations: NTT, MSM, polynomial arithmetic
|
|
// - Lattice crypto: Kyber, Dilithium post-quantum operations
|
|
// - FHE operations: BFV/CKKS homomorphic encryption
|
|
// - DEX operations: AMM swaps, TWAP, order matching
|
|
//
|
|
// # Backend Selection
|
|
//
|
|
// Backends are automatically detected and selected in this priority order:
|
|
// - CUDA (NVIDIA GPUs)
|
|
// - Metal (Apple Silicon)
|
|
// - WebGPU (cross-platform fallback)
|
|
//
|
|
// You can override with environment variable GPU_BACKEND or via API:
|
|
//
|
|
// session, _ := accel.NewSessionWithBackend(accel.BackendMetal)
|
|
//
|
|
// The deprecated names LUX_BACKEND, LUX_ACCEL_BACKEND, and CRYPTO_BACKEND
|
|
// (when set to a backend name) are read for one transition release with a
|
|
// deprecation log message.
|
|
//
|
|
// # Runtime Backend Selection
|
|
//
|
|
// For intelligent backend selection based on required operations:
|
|
//
|
|
// // Select best backend for ZK operations
|
|
// backend, _ := accel.SelectBackend(accel.OpNTT, accel.OpMSM)
|
|
// session, _ := accel.NewSessionWithBackend(backend)
|
|
//
|
|
// // Query capabilities
|
|
// caps, _ := accel.Capabilities(accel.BackendWebGPU)
|
|
// if caps.Supports(accel.OpMSM) {
|
|
// // Use MSM on WebGPU
|
|
// }
|
|
//
|
|
// // Compare backends for an operation
|
|
// comparison, _ := accel.CompareBackends(accel.OpNTT, 10)
|
|
// fmt.Printf("Fastest backend for NTT: %s\n", comparison.Fastest)
|
|
//
|
|
// // Print all capabilities
|
|
// accel.PrintCapabilities()
|
|
//
|
|
// # Pure Go Mode
|
|
//
|
|
// When built with CGO_ENABLED=0, the package compiles in pure Go mode.
|
|
// All operations return ErrNoBackends but the package remains importable,
|
|
// allowing graceful fallback to CPU implementations.
|
|
//
|
|
// # Basic Usage
|
|
//
|
|
// // Initialize library
|
|
// if err := accel.Init(); err != nil {
|
|
// log.Printf("GPU accel not available: %v", err)
|
|
// }
|
|
// defer accel.Shutdown()
|
|
//
|
|
// // Check availability
|
|
// if !accel.Available() {
|
|
// // Use CPU fallback
|
|
// return
|
|
// }
|
|
//
|
|
// // Create session
|
|
// session, err := accel.NewSession()
|
|
// if err != nil {
|
|
// log.Fatal(err)
|
|
// }
|
|
// defer session.Close()
|
|
//
|
|
// // Create tensors
|
|
// a, _ := accel.NewTensor[float32](session, []int{1024, 1024})
|
|
// b, _ := accel.NewTensor[float32](session, []int{1024, 1024})
|
|
// c, _ := accel.NewTensor[float32](session, []int{1024, 1024})
|
|
//
|
|
// // Perform GPU operation
|
|
// if err := session.ML().MatMul(a.Untyped(), b.Untyped(), c.Untyped()); err != nil {
|
|
// log.Fatal(err)
|
|
// }
|
|
//
|
|
// # Integration with Lux Node
|
|
//
|
|
// The accel package integrates with lux-node for:
|
|
// - Batch signature verification in consensus
|
|
// - Merkle tree computation for state sync
|
|
// - Post-quantum cryptography for future-proofing
|
|
//
|
|
// See the node/consensus and precompile packages for integration examples.
|
|
package accel
|