[CUDA] provide wrapper functions for new NVCC builtins.

For sm_80 NVCC introduced a handful of builtins with the names that deviate from
the historic __nvvm_/__nv naming convention. Clang/LLVM does provide equivalent
builtins, but using different names. This patch maps NVCC-style builtins to
their clang counterparts.

Differential Revision: https://reviews.llvm.org/D150894

diff --git a/clang/lib/Headers/__clang_cuda_intrinsics.h b/clang/lib/Headers/__clang_cuda_intrinsics.h
index b87413e..d085bf6 100644 (file)
--- a/clang/lib/Headers/__clang_cuda_intrinsics.h
+++ b/clang/lib/Headers/__clang_cuda_intrinsics.h
@@ -512,6 +512,78 @@ __device__ inline void *__nv_cvta_local_to_generic_impl(size_t __ptr) {
 __device__ inline cuuint32_t __nvvm_get_smem_pointer(void *__ptr) {
   return __nv_cvta_generic_to_shared_impl(__ptr);
 }
+
+#if !defined(__CUDA_ARCH__) || __CUDA_ARCH__ >= 800
+__device__ inline unsigned __reduce_add_sync_unsigned_impl(unsigned __mask,
+                                                           unsigned __value) {
+  return __nvvm_redux_sync_add(__mask, __value);
+}
+__device__ inline int __reduce_add_sync_signed_impl(unsigned __mask,
+                                                    int __value) {
+  return __nvvm_redux_sync_add(__mask, __value);
+}
+__device__ inline unsigned __reduce_min_sync_unsigned_impl(unsigned __mask,
+                                                           unsigned __value) {
+  return __nvvm_redux_sync_umin(__mask, __value);
+}
+__device__ inline unsigned __reduce_max_sync_unsigned_impl(unsigned __mask,
+                                                           unsigned __value) {
+  return __nvvm_redux_sync_umax(__mask, __value);
+}
+__device__ inline int __reduce_min_sync_signed_impl(unsigned __mask,
+                                                    int __value) {
+  return __nvvm_redux_sync_min(__mask, __value);
+}
+__device__ inline int __reduce_max_sync_signed_impl(unsigned __mask,
+                                                    int __value) {
+  return __nvvm_redux_sync_max(__mask, __value);
+}
+__device__ inline unsigned __reduce_or_sync_unsigned_impl(unsigned __mask,
+                                                          unsigned __value) {
+  return __nvvm_redux_sync_or(__mask, __value);
+}
+__device__ inline unsigned __reduce_and_sync_unsigned_impl(unsigned __mask,
+                                                           unsigned __value) {
+  return __nvvm_redux_sync_and(__mask, __value);
+}
+__device__ inline unsigned __reduce_xor_sync_unsigned_impl(unsigned __mask,
+                                                           unsigned __value) {
+  return __nvvm_redux_sync_xor(__mask, __value);
+}
+
+__device__ inline void
+__nv_memcpy_async_shared_global_4_impl(void *__dst, const void *__src,
+                                       unsigned __src_size) {
+  __nvvm_cp_async_ca_shared_global_4(
+      (void __attribute__((address_space(3))) *)__dst,
+      (const void __attribute__((address_space(1))) *)__src, __src_size);
+}
+__device__ inline void
+__nv_memcpy_async_shared_global_8_impl(void *__dst, const void *__src,
+                                       unsigned __src_size) {
+  __nvvm_cp_async_ca_shared_global_8(
+      (void __attribute__((address_space(3))) *)__dst,
+      (const void __attribute__((address_space(1))) *)__src, __src_size);
+}
+__device__ inline void
+__nv_memcpy_async_shared_global_16_impl(void *__dst, const void *__src,
+                                        unsigned __src_size) {
+  __nvvm_cp_async_ca_shared_global_16(
+      (void __attribute__((address_space(3))) *)__dst,
+      (const void __attribute__((address_space(1))) *)__src, __src_size);
+}
+
+__device__ inline void *
+__nv_associate_access_property_impl(const void *__ptr,
+                                    unsigned long long __prop) {
+  // TODO: it appears to provide compiler with some sort of a hint. We do not
+  // know what exactly it is supposed to do. However, CUDA headers suggest that
+  // just passing through __ptr should not affect correctness. They do so on
+  // pre-sm80 GPUs where this builtin is not available.
+  return (void*)__ptr;
+}
+#endif // !defined(__CUDA_ARCH__) || __CUDA_ARCH__ >= 800
+
 } // extern "C"
 #endif // CUDA_VERSION >= 11000