mirror of
https://github.com/luxfi/precompile.git
synced 2026-09-24 00:26:21 +00:00
zeekayandHanzo Dev
2de521b835
The comparison circuit is the deepest one here, a bit-by-bit borrow chain, and under PN10QP27's 27-bit modulus its noise crossed the decryption threshold often enough to return the WRONG BOOLEAN. A soak of the comparison suite failed 4 of 8 runs, a different case each run -- lt, le, gt and ge all appeared, eq and ne never did. Under PN11QP54 the same soak failed 0 of 5 and the full suite passes. A comparison that is usually right is not a comparison, so the slower parameter set is the correct trade: Add 14.92s to 24.58s, Mul 77.74s to 239.99s on the same M1 Max the previous numbers were taken on. The gas constants are re-derived from those measurements rather than left describing a ring the code no longer uses. The literal was written out in four places, and the GPU NTT binding derived its own copy while its comment claimed to "track any future parameter change instead of hard-coding". It did not: it named the literal directly, so the CPU evaluator and the GPU context could bind different rings and silently disagree. There is one Params now, and the GPU test asserts N and Q derived from it rather than restating a pair that would keep passing after the ring moved. CI had never run any of this. The heavy tests skip under -short and `make test` passes -short, so the comparison suite -- the one that was failing half its runs -- was not executed. `make test-long` now runs in CI with a timeout that fits what these operations actually cost. Co-authored-by: Hanzo Dev <[email protected]>
75 lines
2.7 KiB
Makefile
75 lines
2.7 KiB
Makefile
# Lux precompile build wiring.
|
|
#
|
|
# accel v1.1.3+ auto-discovers the lux-gpu substrate (libluxgpu_hqc.a
|
|
# + lux/gpu/hqc.h headers). The default build does NOT call pkg-config;
|
|
# instead, `ops/code/code_cpu.go` hardcodes a list of `#cgo CFLAGS: -I`
|
|
# and `#cgo LDFLAGS: -L` directives covering every standard prefix, and
|
|
# the compiler/linker silently skip any -I/-L that doesn't exist on the
|
|
# host. The default `go build ./...` invocation works without any env
|
|
# var as long as ONE of the following has the install:
|
|
#
|
|
# 1. /usr/local (cmake default prefix); OR
|
|
# 2. /opt/homebrew (Homebrew on Apple Silicon); OR
|
|
# 3. /opt/homebrew/opt/lux-gpu (keg-only Homebrew); OR
|
|
# 4. /usr/local/opt/lux-gpu (Homebrew on Intel Mac); OR
|
|
# 5. /opt/lux (canonical Lux prefix); OR
|
|
# 6. The Go workspace has luxfi/mlx as a sibling repo with a
|
|
# ready-to-link `mlx/build/libluxgpu_hqc.a`.
|
|
#
|
|
# To target a non-listed install, set LUX_GPU_PREFIX:
|
|
#
|
|
# LUX_GPU_PREFIX=/opt/custom make ci
|
|
#
|
|
# That gets fed into CGO_CFLAGS / CGO_LDFLAGS automatically. Power
|
|
# users can still set CGO_CFLAGS / CGO_LDFLAGS directly.
|
|
#
|
|
# pkg-config is only consulted when the build tag `lux_gpu_pkgconfig`
|
|
# is set explicitly (`go build -tags=lux_gpu_pkgconfig ./...`). In
|
|
# that mode the build FAILS if `pkg-config lux-gpu` does not resolve;
|
|
# there is no hardcoded-prefix fallback. The runtime introspector
|
|
# `accel.GPUPaths()` also queries pkg-config, but only as a diagnostic.
|
|
|
|
# Optional override; empty means "use accel's auto-discovery".
|
|
LUX_GPU_PREFIX ?=
|
|
|
|
ifneq ($(LUX_GPU_PREFIX),)
|
|
export CGO_CFLAGS += -I$(LUX_GPU_PREFIX)/include
|
|
export CGO_LDFLAGS += -L$(LUX_GPU_PREFIX)/lib
|
|
endif
|
|
|
|
export GOWORK := off
|
|
|
|
.PHONY: build test test-long vet bench clean ci print-env
|
|
|
|
build:
|
|
go build ./...
|
|
|
|
test:
|
|
go test -count=1 -short -timeout 300s ./...
|
|
|
|
# The FHE tests are the reason this target exists and the reason it is slow. A
|
|
# single homomorphic multiply is ~240s of wall clock at the network's parameter
|
|
# set, so the full suite runs ~28 minutes and 600s could never have covered it.
|
|
# They skip under -short, so `make test` does not run them at all -- which is how
|
|
# a comparison that decrypted the wrong boolean in half of all runs stayed
|
|
# invisible while CI stayed green.
|
|
test-long:
|
|
go test -count=1 -timeout 3600s ./...
|
|
|
|
vet:
|
|
go vet ./...
|
|
|
|
bench:
|
|
go test -bench=. -benchmem -run=^$$ -timeout 600s ./...
|
|
|
|
ci: vet build test
|
|
|
|
# Diagnostic: print the resolved cgo env so a debugging dev can see
|
|
# what the build will use. With LUX_GPU_PREFIX empty, accel falls
|
|
# back to its built-in fallback chain (see accel.GPUPaths()).
|
|
print-env:
|
|
@echo "LUX_GPU_PREFIX = $(LUX_GPU_PREFIX)"
|
|
@echo "CGO_CFLAGS = $(CGO_CFLAGS)"
|
|
@echo "CGO_LDFLAGS = $(CGO_LDFLAGS)"
|
|
@echo "GOWORK = $(GOWORK)"
|