[Feature Request]: MLIR CodeGen 支持libdevice相关符号__nv_**

[howin98]

Background and motivation

MLIR CodeGen中 _nv**符号在使用SerializeToCubinPass中会出现相关符号的未定义错误，通过链接libdevice进行解决

API Proposal

def new_api(value: Tensor) -> Tensor:
  pass

API Usage

No response

Alternatives

No response

Risks

No response

[howin98]

byteir的实现为 https://github.com/bytedance/byteir/tree/main/compiler/lib/Target/PTX
megengine的实现为 https://github.com/MegEngine/MegEngine/blob/551cad4955eb55f96a1974808c00c0dcfa3c07e1/src/jit/impl/mlir/compiler.cpp#L59
bladedisc的实现为 https://github.com/alibaba/BladeDISC/blob/ea2fff1091f3785edbd9dd5f79d210538e9d090e/tao_compiler/mlir/disc/transforms/disc_gpu_source_to_lib.cc#L37

[howin98]

当前路径尝试1 ：将bc编译成ptx，然后参与culinkaddfile

cp /usr/local/cuda/nvvm/libdevice/libdevice.10.bc .
llc -mcpu=sm_60   -mattr=+ptx60 libdevice.10.bc -o libdevice.10.ptx

  RETURN_ON_CUDA_ERROR(cuLinkAddFile(linkState, CUjitInputType::CU_JIT_INPUT_PTX,
                                     "libdevice.10.ptx",
                                     0,       /* number of jit options */
                                     nullptr, /* jit options */
                                     nullptr  /* jit option values */
                                     ));

结果1：成功链接，但是pass后端出现新报错

[howin98]

https://github.com/llvm/llvm-project/blob/main/mlir/lib/Dialect/GPU/Transforms/SerializeToBlob.cpp#L64