From c865298e985e890498be6f09cffed72f469f3631 Mon Sep 17 00:00:00 2001 From: Allard Hendriksen Date: Tue, 17 Dec 2024 14:41:09 +0100 Subject: [PATCH] fence: Add .sync_restrict, .proxy.async.sync_restrict --- docs/libcudacxx/ptx/instructions/fence.rst | 10 ++++++++++ libcudacxx/include/cuda/__ptx/instructions/fence.h | 2 ++ .../libcudacxx/cuda/ptx/ptx.fence.compile.pass.cpp | 2 ++ 3 files changed, 14 insertions(+) diff --git a/docs/libcudacxx/ptx/instructions/fence.rst b/docs/libcudacxx/ptx/instructions/fence.rst index 82de170f63b..4d9126be62f 100644 --- a/docs/libcudacxx/ptx/instructions/fence.rst +++ b/docs/libcudacxx/ptx/instructions/fence.rst @@ -13,6 +13,11 @@ fence .. include:: generated/fence.rst +fence.sync_restrict +------------------- + +.. include:: generated/fence_sync_restrict.rst + fence.mbarrier_init ------------------- @@ -29,6 +34,11 @@ fence.proxy.async .. include:: generated/fence_proxy_async.rst +fence.proxy.async.sync_restrict +------------------------------- + +.. include:: generated/fence_proxy_async_generic_sync_restrict.rst + fence.proxy.tensormap --------------------- diff --git a/libcudacxx/include/cuda/__ptx/instructions/fence.h b/libcudacxx/include/cuda/__ptx/instructions/fence.h index a8dccf979c2..3c123840797 100644 --- a/libcudacxx/include/cuda/__ptx/instructions/fence.h +++ b/libcudacxx/include/cuda/__ptx/instructions/fence.h @@ -36,7 +36,9 @@ _LIBCUDACXX_BEGIN_NAMESPACE_CUDA_PTX #include #include #include +#include #include +#include _LIBCUDACXX_END_NAMESPACE_CUDA_PTX diff --git a/libcudacxx/test/libcudacxx/cuda/ptx/ptx.fence.compile.pass.cpp b/libcudacxx/test/libcudacxx/cuda/ptx/ptx.fence.compile.pass.cpp index c439720b8f8..aa2c9ec6152 100644 --- a/libcudacxx/test/libcudacxx/cuda/ptx/ptx.fence.compile.pass.cpp +++ b/libcudacxx/test/libcudacxx/cuda/ptx/ptx.fence.compile.pass.cpp @@ -20,7 +20,9 @@ #include "generated/fence_mbarrier_init.h" #include "generated/fence_proxy_alias.h" #include "generated/fence_proxy_async.h" +#include "generated/fence_proxy_async_generic_sync_restrict.h" #include "generated/fence_proxy_tensormap_generic.h" +#include "generated/fence_sync_restrict.h" int main(int, char**) {