[Ringssss]

Summary

• Expose Hygon driver module loading, function lookup, launch, unload, and function attributes through the installed InfiniRT API.
• Add the device attributes used by native JIT launchers and generate infini/rt/hygon/driver_.h.

Motivation

InfiniOps can launch compiled Triton kernels from C++, but its Hygon implementation lacked the public driver interface. This supplies the interface needed by the Clover DCU integration without changing existing runtime operations.

Type of Change

• feat - new backend capability / public API

Platforms Affected

• Hygon (WITH_HYGON)
• Build system / CMake / generated headers
• Public headers / installed consumer API

Smoke Build and Test Result

Validated on Hygon BW / gfx936 with DTK 26.04, GCC 11.4, and CMake/Ninja:

cmake -S . -B build -G Ninja -DCMAKE_BUILD_TYPE=Release \
  -DCMAKE_INSTALL_PREFIX="$INFINI_ROOT" -DWITH_CPU=ON -DWITH_HYGON=ON \
  -DAUTO_DETECT_DEVICES=OFF -DINFINI_RT_BUILD_TESTING=ON
cmake --build build -j 12
cmake --install build
export LD_LIBRARY_PATH="/opt/dtk/cuda/cuda/lib64:$LD_LIBRARY_PATH"
ctest --test-dir build --output-on-failure

100% tests passed, 0 tests failed out of 13
Total Test time (real) = 2.18 sec

An installed-header consumer and real InfiniOps Hygon Triton Add, convolution, GDN and MoE launches also passed. The initial consumer run lacked the DTK CUDA compatibility library path; adding that runtime path resolved it.

Test Results on Supported Platforms

Platform	Affected	Build / Smoke Result	Full Result / Notes
CPU	Shared header generation	Passed	Included in the 13-test run
Hygon	Yes	Passed	Runtime, graph, installed consumer, and actual JIT launches
NVIDIA / Iluvatar / MetaX / Moore / Cambricon / Ascend	No backend implementation changes	N/A	Hardware unavailable; their header mappings are unchanged

Benchmark / Performance Impact

N/A: this exposes driver functionality; no standalone speedup is claimed.

Notes for Reviewers

The implementation uses DTK's CUDA driver compatibility API, consistent with the existing Hygon runtime. Tested compiler targets use 64-thread waves. InfiniOps must query the actual HIP compiler target rather than interpreting CUDA compute capability as a gfx architecture.