[PATCH] D38883: [CMake][OpenMP] Customize default offloading arch
Jonas Hahnfeld via Phabricator via cfe-commits
cfe-commits at lists.llvm.org
Fri Oct 13 08:07:27 PDT 2017
Hahnfeld created this revision.
Herald added a subscriber: mgorny.
For the shuffle instructions in reductions we need at least sm_30
but the user may want to customize the default architecture.
Also remove some code that went in while troubleshooting broken
tests on external build bots.
https://reviews.llvm.org/D38883
Files:
CMakeLists.txt
include/clang/Config/config.h.cmake
lib/Driver/ToolChains/Cuda.cpp
lib/Driver/ToolChains/Cuda.h
Index: lib/Driver/ToolChains/Cuda.h
===================================================================
--- lib/Driver/ToolChains/Cuda.h
+++ lib/Driver/ToolChains/Cuda.h
@@ -76,17 +76,6 @@
std::string getLibDeviceFile(StringRef Gpu) const {
return LibDeviceMap.lookup(Gpu);
}
- /// \brief Get lowest available compute capability
- /// for which a libdevice library exists.
- std::string getLowestExistingArch() const {
- std::string LibDeviceFile;
- for (auto key : LibDeviceMap.keys()) {
- LibDeviceFile = LibDeviceMap.lookup(key);
- if (!LibDeviceFile.empty())
- return key;
- }
- return "sm_20";
- }
};
namespace tools {
Index: lib/Driver/ToolChains/Cuda.cpp
===================================================================
--- lib/Driver/ToolChains/Cuda.cpp
+++ lib/Driver/ToolChains/Cuda.cpp
@@ -167,19 +167,6 @@
}
}
- // This code prevents IsValid from being set when
- // no libdevice has been found.
- bool allEmpty = true;
- std::string LibDeviceFile;
- for (auto key : LibDeviceMap.keys()) {
- LibDeviceFile = LibDeviceMap.lookup(key);
- if (!LibDeviceFile.empty())
- allEmpty = false;
- }
-
- if (allEmpty)
- continue;
-
IsValid = true;
break;
}
@@ -565,12 +552,8 @@
StringRef Arch = DAL->getLastArgValue(options::OPT_march_EQ);
if (Arch.empty()) {
- // Default compute capability for CUDA toolchain is the
- // lowest compute capability supported by the installed
- // CUDA version.
- DAL->AddJoinedArg(nullptr,
- Opts.getOption(options::OPT_march_EQ),
- CudaInstallation.getLowestExistingArch());
+ DAL->AddJoinedArg(nullptr, Opts.getOption(options::OPT_march_EQ),
+ CLANG_OPENMP_NVPTX_DEFAULT_ARCH);
}
return DAL;
Index: include/clang/Config/config.h.cmake
===================================================================
--- include/clang/Config/config.h.cmake
+++ include/clang/Config/config.h.cmake
@@ -20,6 +20,9 @@
/* Default OpenMP runtime used by -fopenmp. */
#define CLANG_DEFAULT_OPENMP_RUNTIME "${CLANG_DEFAULT_OPENMP_RUNTIME}"
+/* Default architecture for OpenMP offloading to Nvidia GPUs. */
+#define CLANG_OPENMP_NVPTX_DEFAULT_ARCH "${CLANG_OPENMP_NVPTX_DEFAULT_ARCH}"
+
/* Multilib suffix for libdir. */
#define CLANG_LIBDIR_SUFFIX "${CLANG_LIBDIR_SUFFIX}"
Index: CMakeLists.txt
===================================================================
--- CMakeLists.txt
+++ CMakeLists.txt
@@ -235,6 +235,16 @@
set(CLANG_DEFAULT_OPENMP_RUNTIME "libomp" CACHE STRING
"Default OpenMP runtime used by -fopenmp.")
+# OpenMP offloading requires at least sm_30 because we use shuffle instructions
+# to generate efficient code for reductions.
+set(CLANG_OPENMP_NVPTX_DEFAULT_ARCH "sm_30" CACHE STRING
+ "Default architecture for OpenMP offloading to Nvidia GPUs.")
+if (NOT("${CLANG_OPENMP_NVPTX_DEFAULT_ARCH}" MATCHES "^sm_[0-9]+$"))
+ message(WARNING "Resetting default architecture for OpenMP offloading to Nvidia GPUs to sm_30")
+ set(CLANG_OPENMP_NVPTX_DEFAULT_ARCH "sm_30" CACHE STRING
+ "Default architecture for OpenMP offloading to Nvidia GPUs." FORCE)
+endif()
+
set(CLANG_VENDOR ${PACKAGE_VENDOR} CACHE STRING
"Vendor-specific text for showing with version information.")
-------------- next part --------------
A non-text attachment was scrubbed...
Name: D38883.118912.patch
Type: text/x-patch
Size: 3340 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/cfe-commits/attachments/20171013/e2da685b/attachment.bin>
More information about the cfe-commits
mailing list