Resolve __noinline__ conflicts between gcc 12/13 libstdc++ and CUDA/H…

…IP code

.github/workflows/hip.yml

@@ -21,7 +21,7 @@ jobs:
         torch-version: [2.2.1] # [1.12.0, 1.13.0, 2.0.0, 2.1.0, 2.1.1, 2.1.2, 2.2.0, 2.2.1]
         cuda-version: [11.8.0] # [12.3.1, 12.1.1, 11.8.0, 11.7.1, 11.6.2, 11.5.2,11.4.4, 11.3.1, 11.2.2, 11.1.1, 11.0.3, cpu]
         rocm-version: [6.0.2] # [5.4.2, 5.6.1, 5.7.3, 6.0.1, 6.0.2]
-        llvm-version: [16] # [12, 13, 14, 15, 16]
+        llvm-version: [15] # [12, 13, 14, 15, 16, 17]
         cmake-build-type: [Release] # [Debug, ClangTidy]
     env:
       CCACHE_DIR: ${{ github.workspace }}/ccache
@@ -155,7 +155,19 @@ jobs:
           source .github/workflows/cuda/${{ runner.os }}-env.sh ${CUDA_VER_SHORT}
           export PATH=$PATH:/opt/rocm/bin
           clang --version
-          hipify-clang ./vendor/gsplat/backward.cu ./vendor/gsplat/bindings.cu ./vendor/gsplat/forward.cu --cuda-path=${CUDA_HOME} -I /usr/local/${CUDA_HOME}/include -I ./vendor/gsplat -I /mnt/deps/libtorch/include --print-stats
+          find /opt/rocm -name __clang_cuda_runtime_wrapper.h
+          sudo ln -s /opt/rocm/llvm/lib/clang/17.0.0 /opt/rocm/llvm/lib/clang/17
+          hipify-clang -v ./vendor/gsplat/backward.cu ./vendor/gsplat/bindings.cu ./vendor/gsplat/forward.cu \
+            --cuda-path=${CUDA_HOME} \
+            -I /opt/rocm/include \
+            -I ${CUDA_HOME}/include \
+            -I ${CUDA_HOME}/targets/x86_64-linux/include \
+            -I ./vendor/gsplat \
+            -I /mnt/deps/libtorch/include \
+            -I /mnt/deps/libtorch/include/torch/csrc/api/include/torch \
+            --print-stats -- -include ./vendor/gsplat/noinline.h
+#          -I /usr/lib/llvm-15/lib/clang/15.0.7/include \
+
 
       - name: Clean Compiler Cache
         run: |

vendor/gsplat/noinline.h

@@ -0,0 +1,5 @@
+// Workaround addressing __noinline__ conflicts between gcc 12/13 libstdc++ and CUDA/HIP code
+// https://github.com/llvm/llvm-project/issues/57544
+#if defined(__clang__) && defined(__CUDA__)
+#undef __noinline__
+#endif