mirror of
https://github.com/luxfi/dex.git
synced 2026-09-24 00:26:31 +00:00
zeekay
7d8c88cf8a
AMM unification (one Go entry point, per-platform impls):
- amm_gpu.go (cgo) — BatchEvalConstantProduct dispatcher with
LUX_GPU_DISABLE kill switch + fallback metrics
- amm_gpu_metal.go (cgo && darwin) — Metal driver via lux-dex-amm-metal
- amm_gpu_cuda.go (cgo && linux) — CUDA driver via lux-dex-amm-cuda
- amm_gpu_stub.go (cgo && other) — errAMMGPUUnsupported sentinel
- amm_nogpu.go (!cgo) — pure CPU path
CLOB CUDA wire (the LP-108 missing piece — Backend enum can come back):
- orderbook_cuda_types.go — Go mirrors of DEXOrder (64B), DEXTrade (80B),
DEXPrice; assertDEXABI() in init() panics
on any size/offset drift vs the C header
- orderbook_match_cpu.go — MatchOrderCPU canonical Go oracle
- orderbook_cuda.go (cgo && linux) — gpuMatchOrder via
lux-dex-clob-cuda → libdex_clob_cuda.a →
luxcpp/cuda/kernels/gpu/dex_swap.cu
- orderbook_cuda_stub.go (other) — errCLOBGPUUnsupported sentinel
- orderbook_gpu.go — MatchOrderGPU dispatcher with strict fallback
policy + backend.RecordFallback observability
- orderbook_cuda_parity_test.go — 8 randomized parity subtests using
DEXTrade.EqualFields / DEXOrder.EqualFields
(NOT struct ==) so padding bytes can't
cause false negatives
Padding normalization: _ blank pads renamed to _pad (addressable);
ClearPadding() methods on DEXOrder/DEXTrade; CPU oracle + CUDA host
driver both zero pad bytes for deterministic structural comparison.
Branding: all `Lux*`/`LUX_*` types in the C ABI mirror dropped (DEXOrder,
DEXTrade, DEXPrice, AMMReservePair, DEX_SIDE_BID, etc.).
Temporary replace github.com/luxfi/crypto => ../../lux/crypto pending
publication of crypto with backend.GPUDisabled / RecordFallback /
FallbackReason. Remove the replace once crypto is tagged.
60 lines
1.5 KiB
Go
60 lines
1.5 KiB
Go
// Copyright (C) 2025-2026, Lux Industries Inc. All rights reserved.
|
|
// See the file LICENSE for licensing terms.
|
|
|
|
//go:build cgo && linux
|
|
|
|
package lx
|
|
|
|
// CUDA backend for batch xy=k AMM evaluation. Linked via
|
|
// lux-dex-amm-cuda pkg-config bundle
|
|
// -> libamm_xyk_cuda.a (luxcpp/dex/gpu/cuda/amm_xyk_driver.cu
|
|
// + luxcpp/cuda/kernels/dex/amm_xyk.cu)
|
|
|
|
/*
|
|
#cgo pkg-config: lux-dex-amm-cuda
|
|
|
|
#include <stdint.h>
|
|
#include "amm_xyk_driver.h"
|
|
*/
|
|
import "C"
|
|
|
|
import (
|
|
"fmt"
|
|
"runtime"
|
|
"unsafe"
|
|
)
|
|
|
|
// gpuBatchEval is the CUDA implementation called by amm_gpu.go::BatchEvalConstantProduct.
|
|
// When no NVIDIA device is present at runtime, the driver returns -2 and we
|
|
// surface errAMMGPUUnsupported so the dispatcher falls back to CPU.
|
|
func gpuBatchEval(reserves []ReservePair, amounts []uint64) ([]uint64, error) {
|
|
n := len(reserves)
|
|
out := make([]uint64, n)
|
|
|
|
var pinner runtime.Pinner
|
|
defer pinner.Unpin()
|
|
pinner.Pin(&reserves[0])
|
|
pinner.Pin(&amounts[0])
|
|
pinner.Pin(&out[0])
|
|
|
|
rc := C.amm_xyk_batch_cuda_host(
|
|
(*C.AMMReservePair)(unsafe.Pointer(&reserves[0])),
|
|
(*C.uint64_t)(unsafe.Pointer(&amounts[0])),
|
|
(*C.uint64_t)(unsafe.Pointer(&out[0])),
|
|
C.uint32_t(n),
|
|
)
|
|
runtime.KeepAlive(reserves)
|
|
runtime.KeepAlive(amounts)
|
|
runtime.KeepAlive(out)
|
|
|
|
switch rc {
|
|
case 0:
|
|
return out, nil
|
|
case -2:
|
|
// No CUDA device / driver — fall back to CPU.
|
|
return nil, errAMMGPUUnsupported
|
|
default:
|
|
return nil, fmt.Errorf("lx: amm_xyk_batch_cuda_host rc=%d", int(rc))
|
|
}
|
|
}
|