arsenm created this revision.
arsenm added reviewers: yaxunl, gregrodgers.
Herald added subscribers: kerbowa, tpr, nhaehnle, wdng, jvesely.
Apparently HIPToolChain does not subclass from AMDGPUToolChain, so
this was not applying the new denormal attributes. I'm not sure why
this doesn't subclass. Just copy the implementation for now.
https://reviews.llvm.org/D76862
Files:
clang/lib/Driver/ToolChains/AMDGPU.cpp
clang/lib/Driver/ToolChains/AMDGPU.h
clang/lib/Driver/ToolChains/HIP.cpp
clang/lib/Driver/ToolChains/HIP.h
clang/test/Driver/cuda-flush-denormals-to-zero.cu
Index: clang/test/Driver/cuda-flush-denormals-to-zero.cu
===================================================================
--- clang/test/Driver/cuda-flush-denormals-to-zero.cu
+++ clang/test/Driver/cuda-flush-denormals-to-zero.cu
@@ -7,6 +7,16 @@
// RUN: %clang -no-canonical-prefixes -### -target x86_64-linux-gnu -c -march=haswell --cuda-gpu-arch=sm_70 -fcuda-flush-denormals-to-zero -nocudainc -nocudalib %s 2>&1 | FileCheck -check-prefix=FTZ %s
// RUN: %clang -no-canonical-prefixes -### -target x86_64-linux-gnu -c -march=haswell --cuda-gpu-arch=sm_70 -fno-cuda-flush-denormals-to-zero -nocudainc -nocudalib %s 2>&1 | FileCheck -check-prefix=NOFTZ %s
+// Test explicit argument.
+// RUN: %clang -no-canonical-prefixes -### -target x86_64-linux-gnu -c -march=haswell --cuda-gpu-arch=gfx803 -fcuda-flush-denormals-to-zero -nocudainc -nogpulib %s 2>&1 | FileCheck -check-prefix=FTZ %s
+// RUN: %clang -no-canonical-prefixes -### -target x86_64-linux-gnu -c -march=haswell --cuda-gpu-arch=gfx803 -fno-cuda-flush-denormals-to-zero -nocudainc -nogpulib %s 2>&1 | FileCheck -check-prefix=NOFTZ %s
+// RUN: %clang -x hip -no-canonical-prefixes -### -target x86_64-linux-gnu -c -march=haswell --cuda-gpu-arch=gfx900 -fcuda-flush-denormals-to-zero -nocudainc -nogpulib %s 2>&1 | FileCheck -check-prefix=FTZ %s
+// RUN: %clang -x hip -no-canonical-prefixes -### -target x86_64-linux-gnu -c -march=haswell --cuda-gpu-arch=gfx900 -fno-cuda-flush-denormals-to-zero -nocudainc -nogpulib %s 2>&1 | FileCheck -check-prefix=NOFTZ %s
+
+// Test the default changing with no argument based on the subtarget.
+// RUN: %clang -x hip -no-canonical-prefixes -### -target x86_64-linux-gnu -c -march=haswell --cuda-gpu-arch=gfx803 -nocudainc -nogpulib %s 2>&1 | FileCheck -check-prefix=FTZ %s
+// RUN: %clang -x hip -no-canonical-prefixes -### -target x86_64-linux-gnu -c -march=haswell --cuda-gpu-arch=gfx900 -nocudainc -nogpulib %s 2>&1 | FileCheck -check-prefix=NOFTZ %s
+
// CPUFTZ-NOT: -fdenormal-fp-math
// FTZ-NOT: -fdenormal-fp-math-f32=
Index: clang/lib/Driver/ToolChains/HIP.h
===================================================================
--- clang/lib/Driver/ToolChains/HIP.h
+++ clang/lib/Driver/ToolChains/HIP.h
@@ -115,6 +115,11 @@
unsigned GetDefaultDwarfVersion() const override { return 4; }
+ llvm::DenormalMode getDefaultDenormalModeForType(
+ const llvm::opt::ArgList &DriverArgs,
+ Action::OffloadKind DeviceOffloadKind,
+ const llvm::fltSemantics *FPType = nullptr) const override;
+
const ToolChain &HostTC;
protected:
Index: clang/lib/Driver/ToolChains/HIP.cpp
===================================================================
--- clang/lib/Driver/ToolChains/HIP.cpp
+++ clang/lib/Driver/ToolChains/HIP.cpp
@@ -7,6 +7,7 @@
//===----------------------------------------------------------------------===//
#include "HIP.h"
+#include "AMDGPU.h"
#include "CommonArgs.h"
#include "InputInfo.h"
#include "clang/Basic/Cuda.h"
@@ -16,6 +17,7 @@
#include "clang/Driver/Options.h"
#include "llvm/Support/FileSystem.h"
#include "llvm/Support/Path.h"
+#include "llvm/Support/TargetParser.h"
using namespace clang::driver;
using namespace clang::driver::toolchains;
@@ -272,6 +274,34 @@
getProgramPaths().push_back(getDriver().Dir);
}
+// FIXME: Duplicated in AMDGPUToolChain
+llvm::DenormalMode HIPToolChain::getDefaultDenormalModeForType(
+ const llvm::opt::ArgList &DriverArgs, Action::OffloadKind DeviceOffloadKind,
+ const llvm::fltSemantics *FPType) const {
+ // Denormals should always be enabled for f16 and f64.
+ if (!FPType || FPType != &llvm::APFloat::IEEEsingle())
+ return llvm::DenormalMode::getIEEE();
+
+ if (DeviceOffloadKind == Action::OFK_Cuda) {
+ if (FPType && FPType == &llvm::APFloat::IEEEsingle() &&
+ DriverArgs.hasFlag(options::OPT_fcuda_flush_denormals_to_zero,
+ options::OPT_fno_cuda_flush_denormals_to_zero,
+ false))
+ return llvm::DenormalMode::getPreserveSign();
+ }
+
+ const StringRef GpuArch = DriverArgs.getLastArgValue(options::OPT_mcpu_EQ);
+ auto Kind = llvm::AMDGPU::parseArchAMDGCN(GpuArch);
+
+ // TODO: There are way too many flags that change this. Do we need to check
+ // them all?
+ bool DAZ = DriverArgs.hasArg(options::OPT_cl_denorms_are_zero) ||
+ !AMDGPUToolChain::getDefaultDenormsAreZeroForTarget(Kind);
+ // Outputs are flushed to zero, preserving sign
+ return DAZ ? llvm::DenormalMode::getPreserveSign() :
+ llvm::DenormalMode::getIEEE();
+}
+
void HIPToolChain::addClangTargetOptions(
const llvm::opt::ArgList &DriverArgs,
llvm::opt::ArgStringList &CC1Args,
Index: clang/lib/Driver/ToolChains/AMDGPU.h
===================================================================
--- clang/lib/Driver/ToolChains/AMDGPU.h
+++ clang/lib/Driver/ToolChains/AMDGPU.h
@@ -13,6 +13,8 @@
#include "clang/Driver/Options.h"
#include "clang/Driver/Tool.h"
#include "clang/Driver/ToolChain.h"
+#include "llvm/Support/TargetParser.h"
+
#include <map>
namespace clang {
@@ -67,6 +69,10 @@
llvm::opt::ArgStringList &CC1Args,
Action::OffloadKind DeviceOffloadKind) const override;
+ /// Return whether denormals should be flushed, and treated as 0 by default
+ /// for the subtarget.
+ static bool getDefaultDenormsAreZeroForTarget(llvm::AMDGPU::GPUKind GPUKind);
+
llvm::DenormalMode getDefaultDenormalModeForType(
const llvm::opt::ArgList &DriverArgs,
Action::OffloadKind DeviceOffloadKind,
Index: clang/lib/Driver/ToolChains/AMDGPU.cpp
===================================================================
--- clang/lib/Driver/ToolChains/AMDGPU.cpp
+++ clang/lib/Driver/ToolChains/AMDGPU.cpp
@@ -103,6 +103,19 @@
return DAL;
}
+bool AMDGPUToolChain::getDefaultDenormsAreZeroForTarget(
+ llvm::AMDGPU::GPUKind Kind) {
+ const unsigned ArchAttr = llvm::AMDGPU::getArchAttrAMDGCN(Kind);
+
+ // Default to enabling f32 denormals by default on subtargets where fma is
+ // fast with denormals
+ const bool DefaultDenormsAreZeroForTarget =
+ (ArchAttr & llvm::AMDGPU::FEATURE_FAST_FMA_F32) &&
+ (ArchAttr & llvm::AMDGPU::FEATURE_FAST_DENORMAL_F32);
+
+ return DefaultDenormsAreZeroForTarget;
+}
+
llvm::DenormalMode AMDGPUToolChain::getDefaultDenormalModeForType(
const llvm::opt::ArgList &DriverArgs, Action::OffloadKind DeviceOffloadKind,
const llvm::fltSemantics *FPType) const {
@@ -121,18 +134,10 @@
const StringRef GpuArch = DriverArgs.getLastArgValue(options::OPT_mcpu_EQ);
auto Kind = llvm::AMDGPU::parseArchAMDGCN(GpuArch);
- // Default to enabling f32 denormals by default on subtargets where fma is
- // fast with denormals
-
- const unsigned ArchAttr = llvm::AMDGPU::getArchAttrAMDGCN(Kind);
- const bool DefaultDenormsAreZeroForTarget =
- (ArchAttr & llvm::AMDGPU::FEATURE_FAST_FMA_F32) &&
- (ArchAttr & llvm::AMDGPU::FEATURE_FAST_DENORMAL_F32);
-
// TODO: There are way too many flags that change this. Do we need to check
// them all?
bool DAZ = DriverArgs.hasArg(options::OPT_cl_denorms_are_zero) ||
- !DefaultDenormsAreZeroForTarget;
+ !getDefaultDenormsAreZeroForTarget(Kind);
// Outputs are flushed to zero, preserving sign
return DAZ ? llvm::DenormalMode::getPreserveSign() :
llvm::DenormalMode::getIEEE();
_______________________________________________
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
