bug: purego arm64 linker error — need same asmcgocall fix as zerfoo#115

[dndungu]

Problem

zerfoo/zerfoo#115 fixed the arm64 purego trampoline issue in zerfoo/internal/cuda using asmcgocall-based trampolines. But ztensor/internal/cuda still has the old purego trampolines, causing the same linker error:

github.com/zerfoo/ztensor/internal/cuda.libc_dlerror_trampoline: unhandled relocation for libc_dlerror (type 65 (SDYNIMPORT) rtype 9 (R_CALLARM64))

Building with -tags tabular on linux/arm64 (DGX Spark) still fails because ztensor pulls in the unfixed code path.

Fix

Apply the same asmcgocall-based trampoline fix from zerfoo commit 07c50f09 to ztensor's internal/cuda/purego_linux_arm64.go and purego_linux_arm64.s.

Environment

• Go 1.26.1 linux/arm64
• DGX Spark (NVIDIA GB10, CUDA 13.0)
• ztensor v0.3.0

[dndungu]

The asmcgocall-based trampoline fix is already present in ztensor's internal/cuda/purego_linux_arm64.go and purego_linux_arm64.s. Files are byte-identical to zerfoo's fixed version (verified via diff). The fix uses //go:cgo_import_dynamic for dlopen/dlsym/dlclose/dlerror with assembly JMP trampolines, and ccallTrampoline with AAPCS64 register/stack argument passing. Build verified clean on darwin/arm64. Closing as already resolved.

[dndungu]

Still broken on linux/arm64 (DGX Spark)

This issue was closed based on darwin/arm64 verification, but it's still broken on linux/arm64 (our DGX Spark).

v0.3.0 — CGo build fails

$ GOOS=linux GOARCH=arm64 CGO_ENABLED=1 go build -tags "cgo,lightgbm" ./cmd/train/
github.com/zerfoo/ztensor/internal/cuda.libc_dlopen_trampoline: unhandled relocation for libc_dlopen (type 65 (SDYNIMPORT) rtype 9 (R_CALLARM64))
github.com/zerfoo/ztensor/internal/cuda.libc_dlsym_trampoline: unhandled relocation for libc_dlsym (type 65 (SDYNIMPORT) rtype 9 (R_CALLARM64))

v0.3.1 — ALL builds fail (pure Go and CGo)

$ go build ./cmd/train/
github.com/zerfoo/ztensor/internal/cuda.Open: relocation target runtime.dlopen not defined
github.com/zerfoo/ztensor/internal/cuda.DlopenPath: relocation target runtime.dlopen not defined
github.com/zerfoo/ztensor/internal/cuda.Dlsym: relocation target runtime.dlsym not defined

runtime.dlopen doesn't exist in standard Go runtime. v0.3.1 broke even pure-Go builds on linux/arm64.

Impact

We're forced to stay on v0.3.0 (pure-Go only, no CGo). This means:

• No GPU-accelerated training on DGX (128GB RAM, Grace-Hopper GPU unused)
• No LightGBM (requires CGo for C library linking)
• PatchTST/CfC training takes 40+ hours on CPU vs minutes on GPU

Environment

• Go 1.26.1 linux/arm64
• DGX Spark (NVIDIA Grace Hopper, arm64)
• Wolf repo: github.com/feza-ai/wolf

Please reopen — this is the #1 blocker for GPU training.

[dndungu]

Already resolved in commits 38f54ab (replace cgo_import_dynamic JMP trampolines with runtime.dlopen on arm64) and ebff59e (cgo build tag refinement). Cross-build GOOS=linux GOARCH=arm64 CGO_ENABLED=0 succeeds with no linker errors.