Files
crypto/hash/blake3/gpu/gpu.go
T
Zach Kelling 9f8cb81a58 fix: update GPU build tags for cgo/pure Go separation
- Change CGO files to use luxgpu tag (opt-in)
- Change pure Go files to use !luxgpu tag (default)
- Fix relative paths to luxcpp/crypto headers

This allows CGO=0 and CGO=1 builds to work without requiring
the full GPU libraries to be installed.
2025-12-26 22:14:43 -08:00

178 lines
4.7 KiB
Go

// Copyright (C) 2025, Lux Industries, Inc. All rights reserved.
// See the file LICENSE for licensing terms.
//go:build cgo
// Package gpu provides GPU-accelerated Blake3 hash functions via Metal.
// This package links to luxcpp/crypto for hardware acceleration.
// The C++ library handles automatic fallback to CPU when GPU is unavailable.
package gpu
/*
#cgo CFLAGS: -I${SRCDIR}/../../../../../luxcpp/crypto/include
#cgo LDFLAGS: -L${SRCDIR}/../../../../../luxcpp/crypto/build-local -lluxcrypto -framework Metal -framework Foundation
#include <stdint.h>
#include <stdlib.h>
// C API from luxcpp/crypto
extern void crypto_blake3(uint8_t* out, const uint8_t* in, size_t len);
extern int crypto_batch_hash(uint8_t** outs, const uint8_t* const* ins, const size_t* lens, uint32_t count, int hash_type);
extern int crypto_gpu_available(void);
extern const char* crypto_get_backend(void);
// Metal-specific Blake3 functions
extern int metal_blake3_hash256(const uint8_t* data, uint32_t len, uint8_t* out);
extern int metal_blake3_hash512(const uint8_t* data, uint32_t len, uint8_t* out);
extern int metal_blake3_xof(const uint8_t* data, uint32_t inLen, uint8_t* out, uint32_t outLen);
extern int metal_blake3_merkle_root(const uint8_t* leaves, uint32_t numLeaves, uint8_t* out);
extern int metal_blake3_derive_key(const uint8_t* context, uint32_t contextLen, const uint8_t* key, uint32_t keyLen, uint8_t* out);
*/
import "C"
import (
"unsafe"
)
const (
// Size256 is the standard Blake3 output size
Size256 = 32
// Size512 is extended Blake3 output size
Size512 = 64
// BatchThreshold is minimum batch size for GPU acceleration
BatchThreshold = 64
)
// Available returns true if GPU acceleration is available.
func Available() bool {
return C.crypto_gpu_available() != 0
}
// Backend returns the active backend name ("Metal", "CUDA", or "CPU").
func Backend() string {
return C.GoString(C.crypto_get_backend())
}
// Sum256 computes a 32-byte Blake3 hash using GPU if available.
func Sum256(data []byte) [Size256]byte {
var out [Size256]byte
if len(data) == 0 {
C.crypto_blake3((*C.uint8_t)(&out[0]), nil, 0)
return out
}
C.crypto_blake3(
(*C.uint8_t)(&out[0]),
(*C.uint8_t)(&data[0]),
C.size_t(len(data)),
)
return out
}
// Sum512 computes a 64-byte Blake3 hash using Metal GPU.
func Sum512(data []byte) [Size512]byte {
var out [Size512]byte
var dataPtr *C.uint8_t
if len(data) > 0 {
dataPtr = (*C.uint8_t)(&data[0])
}
C.metal_blake3_hash512(dataPtr, C.uint32_t(len(data)), (*C.uint8_t)(&out[0]))
return out
}
// SumXOF computes an extendable output hash of arbitrary length.
func SumXOF(data []byte, outLen int) []byte {
if outLen <= 0 {
return nil
}
out := make([]byte, outLen)
var dataPtr *C.uint8_t
if len(data) > 0 {
dataPtr = (*C.uint8_t)(&data[0])
}
C.metal_blake3_xof(
dataPtr,
C.uint32_t(len(data)),
(*C.uint8_t)(&out[0]),
C.uint32_t(outLen),
)
return out
}
// BatchSum256 computes Blake3 hashes for multiple inputs in parallel on GPU.
// More efficient than calling Sum256 repeatedly for batch operations.
func BatchSum256(inputs [][]byte) [][Size256]byte {
n := len(inputs)
if n == 0 {
return nil
}
// Prepare C arrays for batch processing
outs := make([]*C.uint8_t, n)
ins := make([]*C.uint8_t, n)
lens := make([]C.size_t, n)
results := make([][Size256]byte, n)
for i, data := range inputs {
outs[i] = (*C.uint8_t)(&results[i][0])
if len(data) > 0 {
ins[i] = (*C.uint8_t)(&data[0])
}
lens[i] = C.size_t(len(data))
}
// Call batch hash (hash_type 2 = BLAKE3)
C.crypto_batch_hash(
(**C.uint8_t)(unsafe.Pointer(&outs[0])),
(**C.uint8_t)(unsafe.Pointer(&ins[0])),
(*C.size_t)(&lens[0]),
C.uint32_t(n),
2, // BLAKE3
)
return results
}
// MerkleRoot computes the Merkle root of leaves using GPU-accelerated Blake3.
func MerkleRoot(leaves [][]byte) [Size256]byte {
var out [Size256]byte
n := len(leaves)
if n == 0 {
return out
}
// Flatten leaves to contiguous buffer
leafData := make([]byte, n*Size256)
for i, leaf := range leaves {
hash := Sum256(leaf)
copy(leafData[i*Size256:], hash[:])
}
C.metal_blake3_merkle_root(
(*C.uint8_t)(&leafData[0]),
C.uint32_t(n),
(*C.uint8_t)(&out[0]),
)
return out
}
// DeriveKey derives a key from context and key material using Blake3 KDF.
func DeriveKey(context string, keyMaterial []byte) [Size256]byte {
var out [Size256]byte
contextBytes := []byte(context)
var contextPtr, keyPtr *C.uint8_t
if len(contextBytes) > 0 {
contextPtr = (*C.uint8_t)(&contextBytes[0])
}
if len(keyMaterial) > 0 {
keyPtr = (*C.uint8_t)(&keyMaterial[0])
}
C.metal_blake3_derive_key(
contextPtr, C.uint32_t(len(contextBytes)),
keyPtr, C.uint32_t(len(keyMaterial)),
(*C.uint8_t)(&out[0]),
)
return out
}