Feat: AICPU launch via dispatcher bootstrap and per-task rtsLaunchCpuKernel

[puddingfjz]

Summary

AICPU kernel loading for CANN 9.0+ — no tar.gz, no sudo, no pre-deployment.

Bootstrap (one-time per (process, device, runtime fingerprint))

Host bundles dispatcher SO bytes + runtime AICPU kernel SO bytes into a single rtAicpuKernelLaunchExWithArgs (kernel_type = KERNEL_TYPE_AICPU_KFC) targeting CANN's preinstalled libaicpu_extend_kernels.so. The dispatcher runs once on the AICPU sched thread (HwHiAiUser) and writes the runtime SO bytes to:

/usr/lib64/aicpu_kernels/0/aicpu_kernels_device/simpler_inner_<fp>.so

…using sched-thread write permission. The dispatcher SO itself is never persisted to disk.

The runtime SO basename embeds an ELF Build-ID-derived 64-bit fingerprint (elf_build_id_64, with FNV-1a-over-full-buffer fallback when the SO was linked without -Wl,--build-id). Host and dispatcher compute the same fingerprint from the same bytes, so the preinstall basename is agreed without any other channel of communication. Writes go via atomic tmp+rename inside the dispatcher — no truncation window visible to concurrent aicpu_scheduler readers. A process-level mutex-protected fingerprint cache in LoadAicpuOp short-circuits redundant libaicpu_extend_kernels invocations across DeviceRunner instances.

Per-task launch

LoadAicpuOp::Init() JSON-registers the runtime SO via rtsBinaryLoadFromFile (cpuKernelMode=0, kernelSo points at the preinstall basename), then resolves simpler_aicpu_init and simpler_aicpu_exec to rtFuncHandles via rtsFuncGetByName. JSON is per-process (/tmp/simpler_inner_<fp>_<pid>.json) so concurrent multi-chip / multi-worker tests don't race on a shared file. opType is suffixed with the fingerprint so multiple LoadAicpuOp instances in the same process register non-colliding entries even though the underlying symbol names are identical across runtimes.

Per-task launches go through rtsLaunchCpuKernel on the cached rtFuncHandles — no per-call string marshalling, no global op registry lookups, no dispatcher hop.

Steady-state ownership

After bootstrap completes, dispatcher_so_binary_ and aicpu_so_binary_ are released on DeviceRunner. Steady-state state held per DeviceRunner is the cached rtFuncHandles + the aicore kernel binary (still needed by per-run rtRegisterAllKernel). AicpuSoInfo and the DeviceArgs::aicpu_so_bin / aicpu_so_len fields are gone — nothing in the dispatcher SO, our runtime AICPU SO, or libaicpu_extend_kernels reads them under this load path. DeviceArgs is kept as a 96-byte zero placeholder for layout stability with any device-side code that walks offsets within it.

Build layout

• libsimpler_aicpu_dispatcher.so is built per-arch (a2a3, a5) and staged once at build/lib/<arch>/dispatcher/. All runtimes on the same arch share that copy — the dispatcher carries no runtime-specific code.
• RuntimeBinaries.dispatcher_path surfaces the path to ChipWorker.init, which threads the bytes through the simpler_init ABI explicitly (no dladdr-based sibling discovery).

Cleanup

• Removes BUILD_WITH_NEW_CANN CMake option and all ifdef branches.
• Deletes the legacy AicpuLoader stub.

Fixes #356.

[gemini-code-assist]

Code Review

This pull request introduces an AicpuLoader abstraction to support both legacy and new CANN 7.0+ interfaces for launching AICPU kernels across the a2a3 and a5 platforms. The implementation includes build system updates, runtime JSON descriptor generation, and integration into the DeviceRunner. Feedback focuses on improving build portability by avoiding hardcoded architecture paths and enhancing the robustness of manual JSON construction. Additionally, the removal of a default parameter in the a2a3 platform's header is identified as a breaking change that violates cross-platform consistency. Suggestions were also made to reduce coupling in the kernel name mapping.