Kernel vector trial: correctness passes, card budgets still fail

The physical trial passes representative vector context preservation and 192 scalar-reference pixel comparisons. Six matched live-card runs show small, mixed CPU-cost changes and fail the unchanged frame CPU and tracking cadence budgets in both modes. Keep the trial optional; the ordinary image remains the baseline. Normal system, shell and Wi-Fi recovery passed after the final trial.

• Trial boot and vector context evidence
• Physical pixel comparisons and optional package
• Paired card measurements, failures and decision

Original build checkpoints

The checkpoints below preserve what was known at each build stage. Their unverified board/rendering limits describe those earlier checkpoints; the linked physical reports above record the later observations.

The normal kernel reports v in the ISA string but omits V from the standard hwprobe response; see the read-only board probe. The running /proc/config.gz contains neither TOOLCHAIN_HAS_V nor RISCV_ISA_V. The normal kernel’s regenerated config reproduces that absence.

The pinned Xuantie arch/riscv/Kconfig tests -march=rv64iv / rv32iv to detect vector-capable compilers. GCC 15.3 rejects those combinations because its V implementation requires M. This makes the hidden TOOLCHAIN_HAS_V false, which makes vector context support unavailable even though the processor’s ISA description contains V. The actual compiler probes are retained in compiler-probes.json: both old checks fail, and adding M makes both pass. This establishes the build/configuration defect; booting the correction is a separate step.

kernel-rvv-trial patches the two checks to rv64imv and rv32imv, and explicitly requests RISCV_ISA_V and RISCV_ISA_V_DEFAULT_ENABLE. It intercepts the board kernel’s buildLinux inputs so its internally-owned src/config are actually changed. The normal kernel, DTB, image and package set are unchanged.

nix build .#kernel-rvv-trial.configfile --max-jobs 1 --cores 4 --no-link --print-out-paths
nix build .#kernel-rvv-trial --max-jobs 1 --cores 8 --no-link --print-out-paths

The configuration build passes. config.diff is the complete difference from the regenerated normal config: TOOLCHAIN_HAS_V, RISCV_ISA_V, RISCV_ISA_V_DEFAULT_ENABLE, DYNAMIC_SIGFRAME, and ERRATA_THEAD_VECTOR become y. The latter two follow kernel dependencies/defaults; the trial does not select a specific vendor instruction set unconditionally. The resulting source is /nix/store/l6jdpbzp53y5n24ky6602f41gzs6f6ik-linux-xuantie-k230-rvv-src and config /nix/store/yp5qydyrl86li5bmx25l6vaykn6yfj64-linux-config-riscv64-unknown-linux-gnu-6.6.36-xuantie.

The full kernel cross-build passes; kernel-build.json records the artifact, SHA256, finish time and final built configuration. The first build client was lost before producing a valid output; the resumed build completed successfully. No boot, vector execution, signal or context preservation on the physical board, framebuffer correctness, speedup or image integration is verified. The board remains on the previous normal image. Do not force RVV in userspace while the running kernel omits its support.

Before integrating: build a matching system closure, including its external Wi-Fi module, rather than swapping the kernel beneath old modules; use a recoverable board trial; require hwprobe V, vector computation across scheduling/signals, normal shell/touch/Wi-Fi recovery and pixel comparisons; then compare the same live card workload with the corrected Pixman’s RVV enabled and disabled. Preserve all existing interaction and frame-cost budgets. A kernel build alone does not close card task 4.2 or any physical acceptance gate.

The context probe now has passing full Linux guest checks at two vector lengths and a deliberate corruption control. The same compiled probe safely skips on the unchanged physical board. These validate the diagnostic and fallback; they do not verify the trial kernel on hardware.

default-identity.json verifies the trial branch’s ordinary system derivation still produces the exact /run/current-system store path observed on the physical board. The vector additions are separate package outputs; the ordinary image has not acquired either the trial kernel or the probe. This identity check is not verification of the optional kernel’s full build or boot.

Matching system configuration

nixosConfigurations.k230-rvv-trial extends the ordinary board configuration with the opt-in kernel and context probe. External modules and the initrd derive from that same selected kernel. trial-system-evaluation.json records the exact kernel, Wi-Fi module and system derivations: the selected kernel matches the already-built standalone trial, the Wi-Fi module changes, the renderer remains Pixman, and the ordinary system derivation stays unchanged.

The kernel helper is imported directly to preserve the returned kernel’s override API. Wrapping it in callPackage intercepted NixOS’s kernel override arguments and caused an evaluation error for features.

nix build .#nixosConfigurations.k230-rvv-trial.config.system.build.toplevel \
  --max-jobs 1 --cores 8 --no-link --print-out-paths

The matching system cross-build passes. trial-system-build.json records its store path, boot-artifact links, rebuilt Wi-Fi module and context-probe closure members, plus the installed service’s Pixman selection. The board trial still requires matching boot artifacts, a known recovery route, and the checks above; no board boot is claimed.

sdImage-rvv-trial packages that trial configuration with its matching kernel and initrd using the existing image layout and stage 1. The shared image helper preserves the ordinary sdImage derivation exactly; before/after evaluation is recorded in trial-image-evaluation.json.

nix build .#sdImage-rvv-trial --max-jobs 1 --cores 8 --no-link --print-out-paths

The trial image build passes. trial-image-build.json records the image hash and a host inspection of its boot partition: Image matches the trial kernel, the wrapped initrd matches the trial system and has valid U-Boot CRCs, and bootargs selects that same system. No card has been flashed or rebooted with it. Preserve the known ordinary image as the recovery artifact before a trial.

Trial on the expanded root

The storage-capacity change is now archived after two successful physical boots. The normal system is now the growth-enabled system recorded in docs/evidence/storage-capacity/board-repeat.json. The earlier ordinary-system identity comparisons above describe their original checkpoint.

The next trial artifacts rebuild the complete RVV system and image on master 648eab07, retaining the same trial kernel and its matching modules/initrd while including the verified root-growth service. Actual host boot-partition inspection passes again. This does not establish physical vector execution or improve any card performance acceptance result.