[Mlir-commits] [mlir] [mlir][Target] Support Fatbin target for static nvptxcompiler (PR #118044)

[Zichen Lu]

================
@@ -538,6 +555,30 @@ NVPTXSerializer::compileToBinaryNVPTX(const std::string &ptxCode) {
   });
 #undef DEBUG_TYPE
   RETURN_ON_NVPTXCOMPILER_ERROR(nvPTXCompilerDestroy(&compiler));
+
+  if (targetOptions.getCompilationTarget() == gpu::CompilationTarget::Fatbin) {
+    const char *cubinOpts[1] = {"-64"};
+    if (useFatbin32) {
+      cubinOpts[0] = {"-32"};
+    }
----------------
MikaOvO wrote:

A better way is to pass options, but to be consistent with another flow, we only consider the `-32` option for now, so I write it this way for now.

> Using const char * in C++ often hints at some design flaw.

I agree with you, but I referred to the [official documentation](https://docs.nvidia.com/cuda/nvfatbin/index.html) here.

https://github.com/llvm/llvm-project/pull/118044