[mlir][nvvm] NVPTXSerializer does not support FATBIN target in compileToBinaryNVPTX flow

[MikaOvO]

In lib/Target/LLVM/NVVM/Target.cpp, NVPTXSerializer compile PTX to binary with two different flows controlled by MLIR_ENABLE_NVPTXCOMPILER.

If building mlir with -DMLIR_ENABLE_NVPTXCOMPILER=ON, the flow does not check if the target is gpu::CompilationTarget::Fatbin, and compile PTX to cubin directly, which is not consistent with another flow.

Is this expected?

[joker-eph]

I don't think it is, feel free to send a patch!

[MikaOvO]

The compileToBinary flow converts PTX to Cubin using ptxas, and converts Cubin to Fatbin using fatbinary, all through files.

The compileToBinaryNVPTX flow converts PTX to Cubin using libnvptxcompiler_static.a, which is done in memory. So is it a bad idea to generate Fatbin through fatbinary (writing the Cubin result and PTX code to file first)? Is it acceptable to generate Fatbin only through libnvfatbin_static.a?

Thanks!

[joker-eph]

Seems reasonable to me.

[MikaOvO]

PR here: #118044