Skip to content

[CUDA] refactor fatbinary to use --image3 #20996

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Draft
bratpiorka wants to merge 1 commit into
base: sycl
Choose a base branch
from
+53 −21
Draft

[CUDA] refactor fatbinary to use --image3 #20996

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
4 changes: 2 additions & 2 deletions clang/test/Driver/clang-linker-wrapper.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -121,7 +121,7 @@
// CHK-CMDS-AOT-NV-NEXT: sycl-post-link{{.*}} SYCL_POST_LINK_OPTIONS -o [[SYCLPOSTLINKOUT:.*]].table [[SECONDLLVMLINKOUT]].bc
// CHK-CMDS-AOT-NV-NEXT: clang{{.*}} -o [[CLANGOUT:.*]] -dumpdir a.out.nvptx64.sm_50.img. --target=nvptx64-nvidia-cuda -march={{.*}}
// CHK-CMDS-AOT-NV-NEXT: ptxas{{.*}} --output-file [[PTXASOUT:.*]] [[CLANGOUT]]
// CHK-CMDS-AOT-NV-NEXT: fatbinary{{.*}} --create [[FATBINOUT:.*]] --image=profile={{.*}},file=[[CLANGOUT]] --image=profile={{.*}},file=[[PTXASOUT]]
// CHK-CMDS-AOT-NV-NEXT: fatbinary{{.*}} --create [[FATBINOUT:[^ ]+]]{{.*}}[[CLANGOUT]]{{.*}}[[PTXASOUT]]
// CHK-CMDS-AOT-NV-NEXT: offload-wrapper: output: [[WRAPPEROUT:.*]].bc, input: [[FATBINOUT]]
// CHK-CMDS-AOT-NV-NEXT: clang{{.*}} -c -o [[LLCOUT:.*]] [[WRAPPEROUT]]
// CHK-CMDS-AOT-NV-NEXT: "{{.*}}/ld" -- HOST_LINKER_FLAGS -dynamic-linker HOST_DYN_LIB -o a.out [[LLCOUT]] HOST_LIB_PATH HOST_STAT_LIB {{.*}}.o
Expand Down Expand Up @@ -166,7 +166,7 @@
// CHK-CMDS-AOT-NV-EMBED-IR-NEXT: clang{{.*}} -c -o [[LLCOUT1:.*]] [[WRAPPEROUT1]]
// CHK-CMDS-AOT-NV-EMBED-IR-NEXT: clang{{.*}} -o [[CLANGOUT:.*]] -dumpdir a.out.nvptx64.sm_50.img. --target=nvptx64-nvidia-cuda -march={{.*}}
// CHK-CMDS-AOT-NV-EMBED-IR-NEXT: ptxas{{.*}} --output-file [[PTXASOUT:.*]] [[CLANGOUT]]
// CHK-CMDS-AOT-NV-EMBED-IR-NEXT: fatbinary{{.*}} --create [[FATBINOUT:.*]] --image=profile={{.*}},file=[[CLANGOUT]] --image=profile={{.*}},file=[[PTXASOUT]]
// CHK-CMDS-AOT-NV-EMBED-IR-NEXT: fatbinary{{.*}}--create [[FATBINOUT:[^ ]+]]{{.*}}[[CLANGOUT]]{{.*}}[[PTXASOUT]]
// CHK-CMDS-AOT-NV-EMBED-IR-NEXT: offload-wrapper: output: [[WRAPPEROUT:.*]].bc, input: [[FATBINOUT]]
// CHK-CMDS-AOT-NV-EMBED-IR-NEXT: clang{{.*}} -c -o [[LLCOUT2:.*]] [[WRAPPEROUT]]
// CHK-CMDS-AOT-NV-EMBED-IR-NEXT: "{{.*}}/ld" -- HOST_LINKER_FLAGS -dynamic-linker HOST_DYN_LIB -o a.out [[LLCOUT1]] [[LLCOUT2]] HOST_LIB_PATH HOST_STAT_LIB {{.*}}.o
Expand Down
4 changes: 2 additions & 2 deletions clang/test/Driver/linker-wrapper.c
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -106,7 +106,7 @@ __attribute__((visibility("protected"), used)) int x;

// CUDA: clang{{.*}} -o [[IMG_SM70:.+]] -dumpdir a.out.nvptx64.sm_70.img. --target=nvptx64-nvidia-cuda -march=sm_70
// CUDA: clang{{.*}} -o [[IMG_SM52:.+]] -dumpdir a.out.nvptx64.sm_52.img. --target=nvptx64-nvidia-cuda -march=sm_52
// CUDA: fatbinary{{.*}}-64 --create {{.*}}.fatbin --image=profile=sm_70,file=[[IMG_SM70]] --image=profile=sm_52,file=[[IMG_SM52]]
// CUDA: fatbinary{{.*}}-64 --create {{.*}}.fatbin --image{{(3)?}}={{(profile=sm_70,file=|kind=elf,sm=70,file=)}}[[IMG_SM70]] --image{{(3)?}}={{(profile=sm_52,file=|kind=elf,sm=52,file=)}}[[IMG_SM52]]
// CUDA: usr/bin/ld{{.*}} {{.*}}.openmp.image.{{.*}}.o {{.*}}.cuda.image.{{.*}}.o

// RUN: llvm-offload-binary -o %t.out \
Expand Down Expand Up @@ -240,7 +240,7 @@ __attribute__((visibility("protected"), used)) int x;
// RUN: %t.o -o a.out 2>&1 | FileCheck %s --check-prefix=RELOCATABLE-LINK-CUDA

// RELOCATABLE-LINK-CUDA: clang{{.*}} -o {{.*}}.img -dumpdir a.out.nvptx64.sm_89.img. --target=nvptx64-nvidia-cuda
// RELOCATABLE-LINK-CUDA: fatbinary{{.*}} -64 --create {{.*}}.fatbin --image=profile=sm_89,file={{.*}}.img
// RELOCATABLE-LINK-CUDA: fatbinary{{.*}} -64 --create {{.*}}.fatbin --image{{(3)?}}={{(profile=sm_89,file=|kind=elf,sm=89,file=)}}{{.*}}.img
// RELOCATABLE-LINK-CUDA: /usr/bin/ld.lld{{.*}}-r
// RELOCATABLE-LINK-CUDA: llvm-objcopy{{.*}}a.out --remove-section .llvm.offloading

Expand Down
66 changes: 49 additions & 17 deletions clang/tools/clang-linker-wrapper/ClangLinkerWrapper.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -457,9 +457,8 @@ void printVersion(raw_ostream &OS) {
}

namespace nvptx {
Expected<StringRef>
fatbinary(ArrayRef<std::pair<StringRef, StringRef>> InputFiles,
const ArgList &Args) {
Expected<StringRef> fatbinary(ArrayRef<OffloadingImage> Images,
const ArgList &Args) {
llvm::TimeTraceScope TimeScope("NVPTX fatbinary");
// NVPTX uses the fatbinary program to bundle the linked images.
Expected<std::string> FatBinaryPath =
Expand All @@ -481,9 +480,26 @@ fatbinary(ArrayRef<std::pair<StringRef, StringRef>> InputFiles,
CmdArgs.push_back(Triple.isArch64Bit() ? "-64" : "-32");
CmdArgs.push_back("--create");
CmdArgs.push_back(*TempFileOrErr);
for (const auto &[File, Arch] : InputFiles)
CmdArgs.push_back(
Args.MakeArgString("--image=profile=" + Arch + ",file=" + File));
for (const OffloadingImage &Image : Images) {
StringRef File = Image.Image->getBufferIdentifier();
StringRef Arch = Image.StringData.lookup("arch");

// Determine the kind based on image type
const char *Kind = "elf";
if (Image.TheImageKind == ImageKind::IMG_PTX)
Kind = "ptx";

// Extract numeric SM value from arch
// Arch can be "sm_75", "compute_75", or just "75"
StringRef SMValue = Arch;
if (Arch.starts_with("sm_"))
SMValue = Arch.drop_front(3);
else if (Arch.starts_with("compute_"))
SMValue = Arch.drop_front(8);

CmdArgs.push_back(Args.MakeArgString("--image3=kind=" + Twine(Kind) +
",sm=" + SMValue + ",file=" + File));
}

if (Error Err = executeCommands(*FatBinaryPath, CmdArgs))
return std::move(Err);
Expand Down Expand Up @@ -1992,12 +2008,7 @@ bundleSYCL(ArrayRef<OffloadingImage> Images) {

Expected<SmallVector<std::unique_ptr<MemoryBuffer>>>
bundleCuda(ArrayRef<OffloadingImage> Images, const ArgList &Args) {
SmallVector<std::pair<StringRef, StringRef>, 4> InputFiles;
for (const OffloadingImage &Image : Images)
InputFiles.emplace_back(std::make_pair(Image.Image->getBufferIdentifier(),
Image.StringData.lookup("arch")));

auto FileOrErr = nvptx::fatbinary(InputFiles, Args);
auto FileOrErr = nvptx::fatbinary(Images, Args);
if (!FileOrErr)
return FileOrErr.takeError();

Expand Down Expand Up @@ -2279,7 +2290,7 @@ linkAndWrapDeviceFiles(ArrayRef<SmallVector<OffloadFile>> LinkerInputFiles,
}
for (size_t I = 0, E = SplitModules.size(); I != E; ++I) {
SmallVector<StringRef> Files = {SplitModules[I].ModuleFilePath};
SmallVector<std::pair<StringRef, StringRef>, 4> BundlerInputFiles;
SmallVector<OffloadingImage, 4> BundlerImages;
auto ClangOutputOrErr =
linkDevice(Files, LinkerArgs, true /* IsSYCLKind */,
CompileLinkOptionsOrErr->first);
Expand All @@ -2292,14 +2303,35 @@ linkAndWrapDeviceFiles(ArrayRef<SmallVector<OffloadFile>> LinkerInputFiles,
nvptx::ptxas(*ClangOutputOrErr, LinkerArgs, Arch);
if (!PtxasOutputOrErr)
return PtxasOutputOrErr.takeError();
BundlerInputFiles.emplace_back(*ClangOutputOrErr, VirtualArch);
BundlerInputFiles.emplace_back(*PtxasOutputOrErr, Arch);
auto BundledFileOrErr =
nvptx::fatbinary(BundlerInputFiles, LinkerArgs);

// Create OffloadingImage for PTX output
OffloadingImage PtxImage;
PtxImage.TheImageKind = ImageKind::IMG_PTX;
PtxImage.TheOffloadKind = OffloadKind::OFK_Cuda;
PtxImage.StringData["arch"] = VirtualArch;
auto PtxBuffer = MemoryBuffer::getFile(*ClangOutputOrErr);
if (!PtxBuffer)
return createFileError(*ClangOutputOrErr, PtxBuffer.getError());
PtxImage.Image = std::move(*PtxBuffer);
BundlerImages.push_back(std::move(PtxImage));

// Create OffloadingImage for Cubin output
OffloadingImage CubinImage;
CubinImage.TheImageKind = ImageKind::IMG_Cubin;
CubinImage.TheOffloadKind = OffloadKind::OFK_Cuda;
CubinImage.StringData["arch"] = Arch;
auto CubinBuffer = MemoryBuffer::getFile(*PtxasOutputOrErr);
if (!CubinBuffer)
return createFileError(*PtxasOutputOrErr, CubinBuffer.getError());
CubinImage.Image = std::move(*CubinBuffer);
BundlerImages.push_back(std::move(CubinImage));

auto BundledFileOrErr = nvptx::fatbinary(BundlerImages, LinkerArgs);
if (!BundledFileOrErr)
return BundledFileOrErr.takeError();
SplitModules[I].ModuleFilePath = *BundledFileOrErr;
} else if (Triple.isAMDGCN()) {
SmallVector<std::pair<StringRef, StringRef>, 4> BundlerInputFiles;
BundlerInputFiles.emplace_back(*ClangOutputOrErr, Arch);
auto BundledFileOrErr =
amdgcn::fatbinary(BundlerInputFiles, LinkerArgs);
Expand Down
Loading