Extend nvrtc workaround for missing SM_100 target macros

NVIDIA · Jan 30, 2025 · 6f3a3a9 · 6f3a3a9
1 parent 554ce9b
commit 6f3a3a9
Show file tree

Hide file tree

Showing 3 changed files with 25 additions and 1 deletion.
diff --git a/libcudacxx/test/libcudacxx/cuda/ptx/nvrtc_workaround.h b/libcudacxx/test/libcudacxx/cuda/ptx/nvrtc_workaround.h
@@ -13,6 +13,25 @@
 // required for the PTX tests, we define them here outside the header guards.
 // TODO(bgruber): limit this workaround to NVRTC versions older than the first one shipping those macros
 #ifdef __CUDACC_RTC__
+
+// missing SM_100
+#  define _NV_TARGET_VAL_SM_100             1000
+#  define _NV_TARGET___NV_PROVIDES_SM_100   (_NV_TARGET_PROVIDES(_NV_TARGET_VAL_SM_100))
+#  define _NV_TARGET___NV_IS_EXACTLY_SM_100 (_NV_TARGET_IS_EXACTLY(_NV_TARGET_VAL_SM_100))
+#  define NV_PROVIDES_SM_100                __NV_PROVIDES_SM_100
+#  define NV_IS_EXACTLY_SM_100              __NV_IS_EXACTLY_SM_100
+#  if (_NV_TARGET___NV_IS_EXACTLY_SM_100)
+#    define _NV_TARGET_BOOL___NV_IS_EXACTLY_SM_100 1
+#  else
+#    define _NV_TARGET_BOOL___NV_IS_EXACTLY_SM_100 0
+#  endif
+#  if (_NV_TARGET___NV_PROVIDES_SM_100)
+#    define _NV_TARGET_BOOL___NV_PROVIDES_SM_100 1
+#  else
+#    define _NV_TARGET_BOOL___NV_PROVIDES_SM_100 0
+#  endif
+
+// missing SM_100a
 #  ifndef NV_HAS_FEATURE_SM_100a
 #    define NV_HAS_FEATURE_SM_100a __NV_HAS_FEATURE_SM_100a
 #    if (defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 1000) && defined(__CUDA_ARCH_FEAT_SM100_ALL))
@@ -22,7 +41,7 @@
 #    endif
 #  endif // NV_HAS_FEATURE_SM_100a
 
-// Re-enable sm_101a support in nvcc.
+// missing SM_101a
 #  ifndef NV_HAS_FEATURE_SM_101a
 #    define NV_HAS_FEATURE_SM_101a __NV_HAS_FEATURE_SM_101a
 #    if (defined(__CUDA_ARCH__) && (__CUDA_ARCH__ >= 1010) && defined(__CUDA_ARCH_FEAT_SM101_ALL))
@@ -31,4 +50,5 @@
 #      define _NV_TARGET_BOOL___NV_HAS_FEATURE_SM_101a 0
 #    endif
 #  endif // NV_HAS_FEATURE_SM_101a
+
 #endif // __CUDACC_RTC__
diff --git a/libcudacxx/test/libcudacxx/cuda/ptx/ptx.clusterlaunchcontrol.compile.pass.cpp b/libcudacxx/test/libcudacxx/cuda/ptx/ptx.clusterlaunchcontrol.compile.pass.cpp
@@ -14,6 +14,8 @@
 #include <cuda/ptx>
 #include <cuda/std/utility>
 
+#include "nvrtc_workaround.h"
+// above header needs to be included before the generated test header
 #include "generated/clusterlaunchcontrol.h"
 
 int main(int, char**)

diff --git a/libcudacxx/test/libcudacxx/cuda/ptx/ptx.st.bulk.compile.pass.cpp b/libcudacxx/test/libcudacxx/cuda/ptx/ptx.st.bulk.compile.pass.cpp
@@ -14,6 +14,8 @@
 #include <cuda/ptx>
 #include <cuda/std/utility>
 
+#include "nvrtc_workaround.h"
+// above header needs to be included before the generated test header
 #include "generated/st_bulk.h"
 
 int main(int, char**)