yaxunl created this revision.
yaxunl added reviewers: tra, b-sumner, ashi1, scchan, t-tye.
Herald added a subscriber: mgorny.
This patch introduces a new option -offloading-target-id for HIP.
Offloading target id is a generalization of CUDA/HIP GPU arch.
It is a device name plus optional feature strings delimited by
plus sign, e.g. gfx906+xnack+sram-ecc. GPU arch is the degenerated
case of offloading target id where there is no feature string. For
each device name, there is a limited number of predefined feature
strings which are allowed to show up in offloading target it. When
feature strings show up in offloading target id, they must follow
predefined order. Therefore offloading target id is a unique id
to convey device name and enabled target feature.
For each offloading target id, a device compilation will be performed
by the driver. If the device compilation results in a device object,
the offloading target id is used in the fat binary to uniquely identify
the device object. This is to allow runtime to load the device
object suitable for the device configuration.
This patches changes HIP action builder to handle -offloading-target-id
option. It generalizes GPUArchList in CUDA/HIP action builder so that
it can handle both CUDA GPU arch and HIP offloading target id. It changes
HIP toolchain to handle offloading target id as bound arch.
This patch is NFC for CUDA toolchain.
https://reviews.llvm.org/D60620
Files:
include/clang/Basic/DiagnosticDriverKinds.td
include/clang/Basic/HIP.h
include/clang/Driver/Options.td
lib/Basic/CMakeLists.txt
lib/Basic/HIP.cpp
lib/Driver/Driver.cpp
lib/Driver/ToolChains/HIP.cpp
test/Driver/hip-invalid-offloading-target-id.hip
test/Driver/hip-offloading-target-id.hip
Index: test/Driver/hip-offloading-target-id.hip
===================================================================
--- /dev/null
+++ test/Driver/hip-offloading-target-id.hip
@@ -0,0 +1,55 @@
+// REQUIRES: clang-driver
+// REQUIRES: x86-registered-target
+// REQUIRES: amdgpu-registered-target
+
+// RUN: %clang -### -target x86_64-linux-gnu \
+// RUN: -x hip --offloading-target-id=gfx906 \
+// RUN: --offloading-target-id=gfx906+xnack \
+// RUN: --offloading-target-id=gfx906+xnack+sram-ecc \
+// RUN: --hip-device-lib-path=%S/Inputs/hip_dev_lib \
+// RUN: %s 2>&1 | FileCheck %s
+
+// CHECK: [[CLANG:".*clang.*"]] "-cc1" "-triple" "amdgcn-amd-amdhsa"
+// CHECK-SAME: "-aux-triple" "x86_64-unknown-linux-gnu"
+// CHECK-SAME: "-emit-llvm-bc"
+// CHECK-SAME: {{.*}} "-target-cpu" "gfx906" "-mno-xnack" "-mno-sram-ecc"
+
+// CHECK: [[OPT:".*opt"]] {{".*-gfx906-linked.*bc"}} "-mtriple=amdgcn-amd-amdhsa"
+// CHECK-SAME: "-mcpu=gfx906"
+// CHECK-SAME: "-o" [[OPT_906_BC:".*-gfx906-optimized.*bc"]]
+
+// CHECK: [[LLC: ".*llc"]] [[OPT_906_BC]]
+// CHECK-SAME: "-mtriple=amdgcn-amd-amdhsa" "-filetype=obj"
+// CHECK-SAME: {{.*}} "-mcpu=gfx906"
+// CHECK-SAME: "-o" {{".*-gfx906-.*o"}}
+
+// CHECK: [[CLANG]] "-cc1" "-triple" "amdgcn-amd-amdhsa"
+// CHECK-SAME: "-aux-triple" "x86_64-unknown-linux-gnu"
+// CHECK-SAME: "-emit-llvm-bc"
+// CHECK-SAME: {{.*}} "-target-cpu" "gfx906" "-mxnack" "-mno-sram-ecc"
+
+// CHECK: [[OPT]] {{".*-gfx906\+xnack.*bc"}} "-mtriple=amdgcn-amd-amdhsa"
+// CHECK-SAME: "-mcpu=gfx906"
+// CHECK-SAME: "-o" [[OPT_906XN_BC:".*-gfx906\+xnack.*bc"]]
+
+// CHECK: [[LLC]] [[OPT_906XN_BC]]
+// CHECK-SAME: "-mtriple=amdgcn-amd-amdhsa" "-filetype=obj"
+// CHECK-SAME: {{.*}} "-mcpu=gfx906"
+// CHECK-SAME: "-o" {{".*-gfx906\+xnack.*o"}}
+
+// CHECK: [[CLANG]] "-cc1" "-triple" "amdgcn-amd-amdhsa"
+// CHECK-SAME: "-aux-triple" "x86_64-unknown-linux-gnu"
+// CHECK-SAME: "-emit-llvm-bc"
+// CHECK-SAME: {{.*}} "-target-cpu" "gfx906" "-mxnack" "-msram-ecc"
+
+// CHECK: [[OPT]] {{".*-gfx906\+xnack\+sram-ecc.*bc"}} "-mtriple=amdgcn-amd-amdhsa"
+// CHECK-SAME: "-mcpu=gfx906"
+// CHECK-SAME: "-o" [[OPT_906XE_BC:".*-gfx906\+xnack\+sram-ecc.*bc"]]
+
+// CHECK: [[LLC]] [[OPT_906XE_BC]]
+// CHECK-SAME: "-mtriple=amdgcn-amd-amdhsa" "-filetype=obj"
+// CHECK-SAME: {{.*}} "-mcpu=gfx906"
+// CHECK-SAME: "-o" {{".*-gfx906\+xnack\+sram-ecc.*o"}}
+
+// CHECK: {{".*clang-offload-bundler"}}
+// CHECK-SAME: "-targets=host-x86_64-unknown-linux,hip-amdgcn-amd-amdhsa-gfx906,hip-amdgcn-amd-amdhsa-gfx906+xnack,hip-amdgcn-amd-amdhsa-gfx906+xnack+sram-ecc"
Index: test/Driver/hip-invalid-offloading-target-id.hip
===================================================================
--- /dev/null
+++ test/Driver/hip-invalid-offloading-target-id.hip
@@ -0,0 +1,48 @@
+// REQUIRES: clang-driver
+// REQUIRES: x86-registered-target
+// REQUIRES: amdgpu-registered-target
+
+// RUN: %clang -### -target x86_64-linux-gnu \
+// RUN: -x hip --offloading-target-id=gfx906 \
+// RUN: --offloading-target-id=gfx906xnack \
+// RUN: --hip-device-lib-path=%S/Inputs/hip_dev_lib \
+// RUN: %s 2>&1 | FileCheck -check-prefix=NOPLUS %s
+
+// NOPLUS: error: Invalid HIP offloading target id: gfx906xnack
+
+// RUN: %clang -### -target x86_64-linux-gnu \
+// RUN: -x hip --offloading-target-id=gfx906 \
+// RUN: --offloading-target-id=gfx906+sram-ecc+xnack \
+// RUN: --hip-device-lib-path=%S/Inputs/hip_dev_lib \
+// RUN: %s 2>&1 | FileCheck -check-prefix=ORDER %s
+
+// ORDER: error: Invalid HIP offloading target id: gfx906+sram-ecc+xnack
+
+// RUN: %clang -### -target x86_64-linux-gnu \
+// RUN: -x hip --offloading-target-id=gfx906 \
+// RUN: --offloading-target-id=gfx906+unknown \
+// RUN: --offloading-target-id=gfx906+sram-ecc+unknown \
+// RUN: --offloading-target-id=gfx900+xnack \
+// RUN: --hip-device-lib-path=%S/Inputs/hip_dev_lib \
+// RUN: %s 2>&1 | FileCheck -check-prefix=UNK %s
+
+// UNK: error: Invalid HIP offloading target id: gfx906+unknown
+
+// RUN: %clang -### -target x86_64-linux-gnu \
+// RUN: -x hip --offloading-target-id=gfx906 \
+// RUN: --offloading-target-id=gfx906+sram-ecc+unknown \
+// RUN: --offloading-target-id=gfx900+xnack \
+// RUN: --hip-device-lib-path=%S/Inputs/hip_dev_lib \
+// RUN: %s 2>&1 | FileCheck -check-prefix=MIXED %s
+
+// MIXED: error: Invalid HIP offloading target id: gfx906+sram-ecc+unknown
+
+// RUN: %clang -### -target x86_64-linux-gnu \
+// RUN: -x hip --offloading-target-id=gfx906 \
+// RUN: --offloading-target-id=gfx900+xnack \
+// RUN: --hip-device-lib-path=%S/Inputs/hip_dev_lib \
+// RUN: %s 2>&1 | FileCheck -check-prefix=UNSUP %s
+
+// UNSUP: error: Invalid HIP offloading target id: gfx900+xnack
+
+
Index: lib/Driver/ToolChains/HIP.cpp
===================================================================
--- lib/Driver/ToolChains/HIP.cpp
+++ lib/Driver/ToolChains/HIP.cpp
@@ -10,10 +10,12 @@
#include "CommonArgs.h"
#include "InputInfo.h"
#include "clang/Basic/Cuda.h"
+#include "clang/Basic/HIP.h"
#include "clang/Driver/Compilation.h"
#include "clang/Driver/Driver.h"
#include "clang/Driver/DriverDiagnostic.h"
#include "clang/Driver/Options.h"
+#include "llvm/ADT/StringSet.h"
#include "llvm/Support/FileSystem.h"
#include "llvm/Support/Path.h"
@@ -77,6 +79,7 @@
Compilation &C, const JobAction &JA, const InputInfoList &Inputs,
const llvm::opt::ArgList &Args, llvm::StringRef SubArchName,
llvm::StringRef OutputFilePrefix, const char *InputFileName) const {
+ SubArchName = SubArchName.split('+').first;
// Construct opt command.
ArgStringList OptArgs;
// The input to opt is the output from llvm-link.
@@ -125,6 +128,9 @@
Compilation &C, const JobAction &JA, const InputInfoList &Inputs,
const llvm::opt::ArgList &Args, llvm::StringRef SubArchName,
llvm::StringRef OutputFilePrefix, const char *InputFileName) const {
+
+ SubArchName = SubArchName.split('+').first;
+
// Construct llc command.
// FIXME: -disable-promote-alloca-to-lds is a workaround for issues in
// AMDGPUPromoteAlloca pass which cause invalid memory access in PyTorch.
@@ -263,6 +269,9 @@
HostTC.addClangTargetOptions(DriverArgs, CC1Args, DeviceOffloadingKind);
StringRef GpuArch = DriverArgs.getLastArgValue(options::OPT_march_EQ);
+ auto Splits = GpuArch.split('+');
+ StringRef FeatureStr = Splits.second;
+ GpuArch = Splits.first;
assert(!GpuArch.empty() && "Must have an explicit GPU arch.");
(void) GpuArch;
assert(DeviceOffloadingKind == Action::OFK_HIP &&
@@ -270,6 +279,26 @@
CC1Args.push_back("-target-cpu");
CC1Args.push_back(DriverArgs.MakeArgStringRef(GpuArch));
+
+ llvm::StringSet<> FeatureSet;
+ while (!FeatureStr.empty()) {
+ auto Splits = FeatureStr.split('+');
+ FeatureSet.insert(Splits.first);
+ FeatureStr = Splits.second;
+ }
+
+ // Iterate through all possible target id features for the given GPU,
+ // If it is contained in the target id, pass -mfeature to clang -cc1,
+ // otherwise pass -mno-feature to clang -cc1.
+ for (auto Feature : HIP::getAllPossibleTargetIdFeatures(GpuArch)) {
+ std::string Opt = "-m";
+ if (FeatureSet.count(Feature))
+ Opt = Opt + Feature.str();
+ else
+ Opt = Opt + "no-" + Feature.str();
+ CC1Args.push_back(DriverArgs.MakeArgStringRef(Opt));
+ }
+
CC1Args.push_back("-fcuda-is-device");
if (DriverArgs.hasFlag(options::OPT_fcuda_flush_denormals_to_zero,
Index: lib/Driver/Driver.cpp
===================================================================
--- lib/Driver/Driver.cpp
+++ lib/Driver/Driver.cpp
@@ -43,6 +43,7 @@
#include "ToolChains/TCE.h"
#include "ToolChains/WebAssembly.h"
#include "ToolChains/XCore.h"
+#include "clang/Basic/HIP.h"
#include "clang/Basic/Version.h"
#include "clang/Config/config.h"
#include "clang/Driver/Action.h"
@@ -2282,8 +2283,20 @@
bool CompileHostOnly = false;
bool CompileDeviceOnly = false;
+ /// Id to identify each device compilation. For CUDA it is simply the
+ /// GPU arch string. For HIP it is either the GPU arch string or GPU
+ /// arch string plus feature strings delimited by a plus sign, e.g.
+ /// gfx906+xnack.
+ struct OffloadingTargetId {
+ /// Target id string which is persistent throughout the compilation.
+ const char *Id;
+ OffloadingTargetId(CudaArch Arch) { Id = CudaArchToString(Arch); }
+ OffloadingTargetId(const char *Id) : Id(Id) {}
+ operator const char *() { return Id; }
+ operator StringRef() { return StringRef(Id); }
+ };
/// List of GPU architectures to use in this compilation.
- SmallVector<CudaArch, 4> GpuArchList;
+ SmallVector<OffloadingTargetId, 4> GpuArchList;
/// The CUDA actions for the current input.
ActionList CudaDeviceActions;
@@ -2362,7 +2375,7 @@
for (auto Arch : GpuArchList) {
CudaDeviceActions.push_back(UA);
- UA->registerDependentActionInfo(ToolChains[0], CudaArchToString(Arch),
+ UA->registerDependentActionInfo(ToolChains[0], Arch,
AssociatedOffloadKind);
}
return ABRT_Success;
@@ -2373,10 +2386,9 @@
void appendTopLevelActions(ActionList &AL) override {
// Utility to append actions to the top level list.
- auto AddTopLevel = [&](Action *A, CudaArch BoundArch) {
+ auto AddTopLevel = [&](Action *A, OffloadingTargetId TargetId) {
OffloadAction::DeviceDependences Dep;
- Dep.add(*A, *ToolChains.front(), CudaArchToString(BoundArch),
- AssociatedOffloadKind);
+ Dep.add(*A, *ToolChains.front(), TargetId, AssociatedOffloadKind);
AL.push_back(C.MakeAction<OffloadAction>(Dep, A->getType()));
};
@@ -2482,7 +2494,7 @@
// Default to sm_20 which is the lowest common denominator for
// supported GPUs. sm_20 code should work correctly, if
// suboptimally, on all newer GPUs.
- if (GpuArchList.empty())
+ if (GpuArchList.empty() && AssociatedOffloadKind != Action::OFK_HIP)
GpuArchList.push_back(CudaArch::SM_20);
return Error;
@@ -2556,8 +2568,7 @@
for (auto &A : {AssembleAction, BackendAction}) {
OffloadAction::DeviceDependences DDep;
- DDep.add(*A, *ToolChains.front(), CudaArchToString(GpuArchList[I]),
- Action::OFK_Cuda);
+ DDep.add(*A, *ToolChains.front(), GpuArchList[I], Action::OFK_Cuda);
DeviceActions.push_back(
C.MakeAction<OffloadAction>(DDep, A->getType()));
}
@@ -2606,6 +2617,7 @@
class HIPActionBuilder final : public CudaActionBuilderBase {
/// The linker inputs obtained for each device arch.
SmallVector<ActionList, 8> DeviceLinkerInputs;
+ const char *const DefaultTargetId = "gfx900";
public:
HIPActionBuilder(Compilation &C, DerivedArgList &Args,
@@ -2654,8 +2666,8 @@
// device arch of the next action being propagated to the above link
// action.
OffloadAction::DeviceDependences DDep;
- DDep.add(*CudaDeviceActions[I], *ToolChains.front(),
- CudaArchToString(GpuArchList[I]), AssociatedOffloadKind);
+ DDep.add(*CudaDeviceActions[I], *ToolChains.front(), GpuArchList[I],
+ AssociatedOffloadKind);
CudaDeviceActions[I] = C.MakeAction<OffloadAction>(
DDep, CudaDeviceActions[I]->getType());
}
@@ -2710,11 +2722,73 @@
for (auto &LI : DeviceLinkerInputs) {
auto *DeviceLinkAction =
C.MakeAction<LinkJobAction>(LI, types::TY_Image);
- DA.add(*DeviceLinkAction, *ToolChains[0],
- CudaArchToString(GpuArchList[I]), AssociatedOffloadKind);
+ DA.add(*DeviceLinkAction, *ToolChains[0], GpuArchList[I],
+ AssociatedOffloadKind);
++I;
}
}
+ bool initialize() override {
+ if (CudaActionBuilderBase::initialize())
+ return true;
+
+ auto IsValidTargetId = [](StringRef IdStr) {
+ auto Split = IdStr.split('+');
+ const StringRef ArchStr = Split.first;
+ CudaArch Arch = StringToCudaArch(ArchStr);
+ if (Arch == CudaArch::UNKNOWN)
+ return false;
+ StringRef Features = Split.second;
+ if (Features.empty())
+ return true;
+ auto &AllFeatures = HIP::getAllPossibleTargetIdFeatures(ArchStr);
+ unsigned CurIndex = 0;
+ while (!Features.empty()) {
+ auto Splits = Features.split('+');
+ for (; CurIndex < AllFeatures.size(); ++CurIndex) {
+ if (Splits.first == AllFeatures[CurIndex])
+ break;
+ }
+ if (CurIndex == AllFeatures.size())
+ return false;
+ Features = Splits.second;
+ }
+ return true;
+ };
+ // Collect all offloading_target_id parameters, removing duplicates.
+ std::set<StringRef> TargetIds;
+ for (Arg *A : Args) {
+ if (!(A->getOption().matches(options::OPT_offloading_target_id_EQ) ||
+ A->getOption().matches(options::OPT_no_offloading_target_id_EQ)))
+ continue;
+ A->claim();
+
+ const StringRef IdStr = A->getValue();
+ if (A->getOption().matches(options::OPT_offloading_target_id_EQ) &&
+ IdStr == "all") {
+ TargetIds.clear();
+ continue;
+ }
+ if (!IsValidTargetId(IdStr)) {
+ C.getDriver().Diag(clang::diag::err_drv_hip_bad_target_id) << IdStr;
+ return true;
+ } else if (A->getOption().matches(options::OPT_offloading_target_id_EQ))
+ TargetIds.insert(IdStr);
+ else if (A->getOption().matches(
+ options::OPT_no_offloading_target_id_EQ))
+ TargetIds.erase(IdStr);
+ else
+ llvm_unreachable("Unexpected option.");
+ }
+
+ // Collect list of target ids remaining in the set.
+ for (auto Id : TargetIds)
+ GpuArchList.push_back(Id.data());
+
+ if (GpuArchList.empty())
+ GpuArchList.push_back(DefaultTargetId);
+
+ return false;
+ }
};
/// OpenMP action builder. The host bitcode is passed to the device frontend
Index: lib/Basic/HIP.cpp
===================================================================
--- /dev/null
+++ lib/Basic/HIP.cpp
@@ -0,0 +1,21 @@
+//===--- HIP.cpp - Utilities for compiling HIP code ----------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+
+#include "clang/Basic/HIP.h"
+
+namespace clang {
+
+namespace HIP {
+
+const llvm::SmallVector<llvm::StringRef, 4> &
+getAllPossibleTargetIdFeatures(llvm::StringRef Device) {
+ static OffloadingTargetIdFeatures F;
+ return F.Features[Device];
+}
+} // namespace HIP
+} // namespace clang
Index: lib/Basic/CMakeLists.txt
===================================================================
--- lib/Basic/CMakeLists.txt
+++ lib/Basic/CMakeLists.txt
@@ -48,6 +48,7 @@
FileManager.cpp
FileSystemStatCache.cpp
FixedPoint.cpp
+ HIP.cpp
IdentifierTable.cpp
LangOptions.cpp
Module.cpp
Index: include/clang/Driver/Options.td
===================================================================
--- include/clang/Driver/Options.td
+++ include/clang/Driver/Options.td
@@ -587,6 +587,11 @@
def fcuda_short_ptr : Flag<["-"], "fcuda-short-ptr">, Flags<[CC1Option]>,
HelpText<"Use 32-bit pointers for accessing const/local/shared address spaces.">;
def fno_cuda_short_ptr : Flag<["-"], "fno-cuda-short-ptr">;
+def offloading_target_id_EQ : Joined<["--"], "offloading-target-id=">, Flags<[DriverOption]>,
+ HelpText<"HIP offloading target id (e.g. gfx906+xnack). May be specified more than once.">;
+def no_offloading_target_id_EQ : Joined<["--"], "no-offloading-target-id=">, Flags<[DriverOption]>,
+ HelpText<"Remove HIP offloading target id (e.g. gfx906+xnack) from the list of target ids to compile for."
+ "'all' resets the list to its default value.">;
def hip_device_lib_path_EQ : Joined<["--"], "hip-device-lib-path=">, Group<Link_Group>,
HelpText<"HIP device library path">;
def hip_device_lib_EQ : Joined<["--"], "hip-device-lib=">, Group<Link_Group>,
Index: include/clang/Basic/HIP.h
===================================================================
--- /dev/null
+++ include/clang/Basic/HIP.h
@@ -0,0 +1,38 @@
+//===--- HIP.h - Utilities for compiling HIP code --------------*- C++ -*-===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef LLVM_CLANG_BASIC_HIP_H
+#define LLVM_CLANG_BASIC_HIP_H
+
+#include "llvm/ADT/SmallVector.h"
+#include "llvm/ADT/StringMap.h"
+
+namespace clang {
+
+namespace HIP {
+/// Defines all feature strings that can be used in offloading target id
+/// for each device. Offloading target id is a device name with optional
+/// target feature strings delimited by a plus sign, e.g.
+/// gfx906+xnack+sram-ecc. Target feature itself cannot contain plus sign.
+/// Each device have limited number of predefined target features which
+/// have to follow predefined order when showing up in a offloading target
+/// id.
+struct OffloadingTargetIdFeatures {
+ llvm::StringMap<llvm::SmallVector<llvm::StringRef, 4>> Features;
+ OffloadingTargetIdFeatures() { Features["gfx906"] = {"xnack", "sram-ecc"}; }
+};
+
+/// Get all feature strings that can be used in code object target id for
+/// \m Device.
+const llvm::SmallVector<llvm::StringRef, 4> &
+getAllPossibleTargetIdFeatures(llvm::StringRef Device);
+
+} // namespace HIP
+} // namespace clang
+
+#endif
Index: include/clang/Basic/DiagnosticDriverKinds.td
===================================================================
--- include/clang/Basic/DiagnosticDriverKinds.td
+++ include/clang/Basic/DiagnosticDriverKinds.td
@@ -41,6 +41,7 @@
"--no-cuda-version-check.">;
def err_drv_cuda_host_arch : Error<"unsupported architecture '%0' for host compilation.">;
def err_drv_mix_cuda_hip : Error<"Mixed Cuda and HIP compilation is not supported.">;
+def err_drv_hip_bad_target_id : Error<"Invalid HIP offloading target id: %0">;
def err_drv_invalid_thread_model_for_target : Error<
"invalid thread model '%0' in '%1' for this target">;
def err_drv_invalid_linker_name : Error<
_______________________________________________
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
