gtbercea created this revision.
gtbercea added reviewers: ABataev, hfinkel, caomhin.
Herald added subscribers: cfe-commits, jdoerfert, guansong, mgorny.
Herald added a project: clang.
gtbercea added a reviewer: tra.
gtbercea added parent revisions: D60906: [OpenMP][libomptarget][WIP] Add math
functions support in OpenMP offloading, D60905: [OpenMP][LLVM][WIP] Add math
functions support to OpenMP.
gtbercea edited the summary of this revision.
This patch adds an OpenMP specific math functions header to the lib/Headers
folder and ensures it is passed to Clang.
Note:
This is an example of how support for math functions could be implemented.
Before expanding this to include other math functions please let me know if you
have any comments, concerns or proposed changes.
Repository:
rC Clang
https://reviews.llvm.org/D60907
Files:
include/clang/Driver/ToolChain.h
lib/Driver/ToolChains/Clang.cpp
lib/Driver/ToolChains/Cuda.cpp
lib/Driver/ToolChains/Cuda.h
lib/Headers/CMakeLists.txt
lib/Headers/__clang_openmp_math.h
Index: lib/Headers/__clang_openmp_math.h
===================================================================
--- /dev/null
+++ lib/Headers/__clang_openmp_math.h
@@ -0,0 +1,33 @@
+
+#ifndef __CLANG_OMP_CMATH_H__
+#define __CLANG_OMP_CMATH_H__
+
+#ifdef __NVPTX__
+#pragma omp declare target
+
+// Declarations of function in libomptarget
+#if defined(__cplusplus)
+extern "C" {
+#endif
+
+double __kmpc_pow(double, double);
+double __kmpc_sin(double);
+
+#if defined(__cplusplus)
+}
+#endif
+
+// Define existing function to call kmpc functions.
+__attribute__((always_inline, used)) static double pow(double a, double b) {
+ return __kmpc_pow(a, b);
+}
+
+__attribute__((always_inline, used)) static double sin(double a) {
+ return __kmpc_sin(a);
+}
+
+#pragma omp end declare target
+#endif
+
+#endif
+
Index: lib/Headers/CMakeLists.txt
===================================================================
--- lib/Headers/CMakeLists.txt
+++ lib/Headers/CMakeLists.txt
@@ -31,6 +31,7 @@
avxintrin.h
bmi2intrin.h
bmiintrin.h
+ __clang_openmp_math.h
__clang_cuda_builtin_vars.h
__clang_cuda_cmath.h
__clang_cuda_complex_builtins.h
Index: lib/Driver/ToolChains/Cuda.h
===================================================================
--- lib/Driver/ToolChains/Cuda.h
+++ lib/Driver/ToolChains/Cuda.h
@@ -48,6 +48,9 @@
void AddCudaIncludeArgs(const llvm::opt::ArgList &DriverArgs,
llvm::opt::ArgStringList &CC1Args) const;
+ void AddMathDeviceFunctions(const llvm::opt::ArgList &DriverArgs,
+ llvm::opt::ArgStringList &CC1Args) const;
+
/// Emit an error if Version does not support the given Arch.
///
/// If either Version or Arch is unknown, does not emit an error. Emits at
@@ -165,6 +168,9 @@
void AddCudaIncludeArgs(const llvm::opt::ArgList &DriverArgs,
llvm::opt::ArgStringList &CC1Args) const override;
+ void AddMathDeviceFunctions(const llvm::opt::ArgList &DriverArgs,
+ llvm::opt::ArgStringList &CC1Args) const override;
+
void addClangWarningOptions(llvm::opt::ArgStringList &CC1Args) const override;
CXXStdlibType GetCXXStdlibType(const llvm::opt::ArgList &Args) const override;
void
Index: lib/Driver/ToolChains/Cuda.cpp
===================================================================
--- lib/Driver/ToolChains/Cuda.cpp
+++ lib/Driver/ToolChains/Cuda.cpp
@@ -255,6 +255,16 @@
CC1Args.push_back("__clang_cuda_runtime_wrapper.h");
}
+void CudaInstallationDetector::AddMathDeviceFunctions(
+ const ArgList &DriverArgs, ArgStringList &CC1Args) const {
+ CC1Args.push_back("-internal-isystem");
+ CC1Args.push_back(DriverArgs.MakeArgString(getIncludePath()));
+ CC1Args.push_back("-include");
+ CC1Args.push_back("__clang_openmp_math.h");
+ CC1Args.push_back("-I");
+ CC1Args.push_back(DriverArgs.MakeArgString(getIncludePath()));
+}
+
void CudaInstallationDetector::CheckCudaVersionSupportsArch(
CudaArch Arch) const {
if (Arch == CudaArch::UNKNOWN || Version == CudaVersion::UNKNOWN ||
@@ -898,6 +908,11 @@
CudaInstallation.AddCudaIncludeArgs(DriverArgs, CC1Args);
}
+void CudaToolChain::AddMathDeviceFunctions(
+ const ArgList &DriverArgs, ArgStringList &CC1Args) const {
+ CudaInstallation.AddMathDeviceFunctions(DriverArgs, CC1Args);
+}
+
llvm::opt::DerivedArgList *
CudaToolChain::TranslateArgs(const llvm::opt::DerivedArgList &Args,
StringRef BoundArch,
Index: lib/Driver/ToolChains/Clang.cpp
===================================================================
--- lib/Driver/ToolChains/Clang.cpp
+++ lib/Driver/ToolChains/Clang.cpp
@@ -1150,6 +1150,14 @@
if (JA.isOffloading(Action::OFK_Cuda))
getToolChain().AddCudaIncludeArgs(Args, CmdArgs);
+ // If we are offloading to a target via OpenMP and this target happens
+ // to be an NVIDIA GPU then we need to include the CUDA runtime wrapper
+ // to ensure the correct math functions are called in the offloaded
+ // code.
+ if (JA.isDeviceOffloading(Action::OFK_OpenMP) &&
+ getToolChain().getTriple().isNVPTX())
+ getToolChain().AddMathDeviceFunctions(Args, CmdArgs);
+
// Add -i* options, and automatically translate to
// -include-pch/-include-pth for transparent PCH support. It's
// wonky, but we include looking for .gch so we can support seamless
Index: include/clang/Driver/ToolChain.h
===================================================================
--- include/clang/Driver/ToolChain.h
+++ include/clang/Driver/ToolChain.h
@@ -572,6 +572,10 @@
virtual void AddCudaIncludeArgs(const llvm::opt::ArgList &DriverArgs,
llvm::opt::ArgStringList &CC1Args) const;
+ /// Add arguments to use system-specific CUDA includes.
+ virtual void AddMathDeviceFunctions(const llvm::opt::ArgList &DriverArgs,
+ llvm::opt::ArgStringList &CC1Args) const {};
+
/// Add arguments to use MCU GCC toolchain includes.
virtual void AddIAMCUIncludeArgs(const llvm::opt::ArgList &DriverArgs,
llvm::opt::ArgStringList &CC1Args) const;
_______________________________________________
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
