Fix OSS performance on FP8 AMD kernels OS (#4462)

Summary:
X-link: facebookresearch/FBGEMM#1521

 This adds some missing hipcc compiler flags, without these flags certain kernel instances suffer from major performance issues.

Pull Request resolved: #4462

Reviewed By: jwfromm

Differential Revision: D78010696

Pulled By: cthi

fbshipit-source-id: e5d35944f236d90d5c0aa9cb2587a7d6d45540b6

Author: cthi

fbgemm_gpu/experimental/gen_ai/CMakeLists.txt

@@ -149,7 +149,18 @@ gpu_cpp_library(
     ${experimental_gen_ai_cpp_source_files_hip}
   TORCH_LIBS
     # Used when building as part of PyTorch
-    ${FBGEMM_GENAI_TORCH_LIBS})
+    ${FBGEMM_GENAI_TORCH_LIBS}
+  HIPCC_FLAGS
+    # Below flags are required for strong CK performance
+    # on certain kernel instances
+    -mllvm
+    # Reduce register spillage on certain kernels
+    -amdgpu-coerce-illegal-types=1
+    -mllvm
+    -enable-post-misched=0
+    -mllvm
+    -greedy-reverse-local-assignment=1
+    -fhip-new-launch-api)