Skip to content

Commit

Permalink
[SYCL][Fusion] Kernel Fusion support for CUDA backend (intel#8747)
Browse files Browse the repository at this point in the history 
Extend kernel fusion for the CUDA backend. 

In contrast to the existing SPIR-V based backends, the default binary
format for the CUDA backend (PTX or CUBIN) is not suitable as input for
the kernel fusion JIT compiler.

This PR therefore extends the driver to **additionally** embed LLVM IR
in the fat binary if the user specifies the `-fsycl-embed-ir` during
compilation, by taking the output of the `sycl-post-link` step for the
CUDA backend.

The JIT compiler has been extended to handle LLVM IR as input format and
PTX assembly as output format (including translation via the NVPTX
backend). Target-specific parts of the fusion process have been
refactored to `TargetFusionInformation`.

The connecting logic to the JIT compiler in the SYCL RT has been
extended to produce valid PI device binaries for the CUDA backend/PI.

Heterogeneous ND ranges are not yet supported for the CUDA backend.

---------

Signed-off-by: Lukas Sommer <lukas.sommer@codeplay.com>
  • Loading branch information
@sommerlukas
sommerlukas authored May 8, 2023
1 parent 6953c46 commit a93e59d
Show file tree
Hide file tree
Showing 83 changed files with 1,805 additions and 712 deletions.
7 changes: 6 additions & 1 deletion clang/include/clang/Driver/Action.h
View file
Original file line number Diff line number Diff line change
Expand Up @@ -660,9 +660,14 @@ class OffloadUnbundlingJobAction final : public JobAction {
class OffloadWrapperJobAction : public JobAction {
void anchor() override;

bool EmbedIR;

public:
OffloadWrapperJobAction(ActionList &Inputs, types::ID Type);
OffloadWrapperJobAction(Action *Input, types::ID OutputType);
OffloadWrapperJobAction(Action *Input, types::ID OutputType,
bool EmbedIR = false);

bool isEmbeddedIR() const { return EmbedIR; }

static bool classof(const Action *A) {
return A->getKind() == OffloadWrapperJobClass;
Expand Down
2 changes: 2 additions & 0 deletions clang/include/clang/Driver/Options.td
View file
Original file line number Diff line number Diff line change
Expand Up @@ -2973,6 +2973,8 @@ def fintelfpga : Flag<["-"], "fintelfpga">, Group<f_Group>,
HelpText<"Perform ahead-of-time compilation for FPGA">;
def fsycl_device_only : Flag<["-"], "fsycl-device-only">, Flags<[CoreOption]>,
HelpText<"Compile SYCL kernels for device">;
def fsycl_embed_ir : Flag<["-"], "fsycl-embed-ir">, Flags<[CoreOption]>,
HelpText<"Embed LLVM IR for runtime kernel fusion">;
defm sycl_esimd_force_stateless_mem : BoolFOption<"sycl-esimd-force-stateless-mem",
LangOpts<"SYCLESIMDForceStatelessMem">, DefaultFalse,
PosFlag<SetTrue, [], "Enforce using stateless memory accesses. "
Expand Down
8 changes: 4 additions & 4 deletions clang/lib/Driver/Action.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -478,11 +478,11 @@ void OffloadWrapperJobAction::anchor() {}

OffloadWrapperJobAction::OffloadWrapperJobAction(ActionList &Inputs,
types::ID Type)
: JobAction(OffloadWrapperJobClass, Inputs, Type) {}
: JobAction(OffloadWrapperJobClass, Inputs, Type), EmbedIR(false) {}

OffloadWrapperJobAction::OffloadWrapperJobAction(Action *Input,
types::ID Type)
: JobAction(OffloadWrapperJobClass, Input, Type) {}
OffloadWrapperJobAction::OffloadWrapperJobAction(Action *Input, types::ID Type,
bool IsEmbeddedIR)
: JobAction(OffloadWrapperJobClass, Input, Type), EmbedIR(IsEmbeddedIR) {}

void OffloadPackagerJobAction::anchor() {}

Expand Down
116 changes: 64 additions & 52 deletions clang/lib/Driver/Driver.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -5516,6 +5516,8 @@ class OffloadingActionBuilder final {
// s - device code split requested
// r - relocatable device code is requested
// f - link object output type is TY_Tempfilelist (fat archive)
// e - Embedded IR for fusion (-fsycl-embed-ir) was requested
// and target is NVPTX.
// * - "all other cases"
// - no condition means output/input is "always" present
// First symbol indicates output/input type
Expand All @@ -5535,58 +5537,58 @@ class OffloadingActionBuilder final {
// | |
// | |
// .---------------------------------------.
// | PostLink |
// .---------------------------------------.
// [+*] [+]
// | |
// | |
// |--------- |
// | | |
// | | |
// | [+!rf] |
// | .-------------. |
// | | llvm-foreach| |
// | .-------------. |
// | | |
// [+*] [+!rf] |
// .-----------------. |
// | FileTableTform | |
// | (extract "Code")| |
// .-----------------. |
// [-] |-----------
// --------------------| |
// | | |
// | |----------------- |
// | | | |
// | | [-!rf] |
// | | .--------------. |
// | | |FileTableTform| |
// | | | (merge) | |
// | | .--------------. |
// | | [-] |-------
// | | | | |
// | | | ------| |
// | | --------| | |
// [.] [-*] [-!rf] [+!rf] |
// .---------------. .-------------------. .--------------. |
// | finalizeNVPTX | | SPIRVTranslator | |FileTableTform| |
// | finalizeAMDGCN | | | | (merge) | |
// .---------------. .-------------------. . -------------. |
// [.] [-as] [-!a] | |
// | | | | |
// | [-s] | | |
// | .----------------. | | |
// | | BackendCompile | | | |
// | .----------------. | ------| |
// | [-s] | | |
// | | | | |
// | [-a] [-!a] [-!rf] |
// | .--------------------. |
// -----------[-n]| FileTableTform |[+*]--------------|
// | (replace "Code") |
// .--------------------.
// |
// [+*]
// | PostLink |[+e]----------------
// .---------------------------------------. |
// [+*] [+] |
// | | |
// | | |
// |--------- | |
// | | | |
// | | | |
// | [+!rf] | |
// | .-------------. | |
// | | llvm-foreach| | |
// | .-------------. | |
// | | | |
// [+*] [+!rf] | |
// .-----------------. | |
// | FileTableTform | | |
// | (extract "Code")| | |
// .-----------------. | |
// [-] |----------- |
// --------------------| | |
// | | | |
// | |----------------- | |
// | | | | |
// | | [-!rf] | |
// | | .--------------. | |
// | | |FileTableTform| | |
// | | | (merge) | | |
// | | .--------------. | |
// | | [-] |------- |
// | | | | | |
// | | | ------| | |
// | | --------| | | |
// [.] [-*] [-!rf] [+!rf] | |
// .---------------. .-------------------. .--------------. | |
// | finalizeNVPTX | | SPIRVTranslator | |FileTableTform| | |
// | finalizeAMDGCN | | | | (merge) | | |
// .---------------. .-------------------. . -------------. | |
// [.] [-as] [-!a] | | |
// | | | | | |
// | [-s] | | | |
// | .----------------. | | | |
// | | BackendCompile | | | | |
// | .----------------. | ------| | |
// | [-s] | | | |
// | | | | | |
// | [-a] [-!a] [-!rf] | |
// | .--------------------. | |
// -----------[-n]| FileTableTform |[+*]--------------| |
// | (replace "Code") | |
// .--------------------. |
// | -------------------------
// [+*] | [+e]
// .--------------------------------------.
// | OffloadWrapper |
// .--------------------------------------.
Expand Down Expand Up @@ -5693,6 +5695,16 @@ class OffloadingActionBuilder final {
return TypedPostLinkAction;
};
Action *PostLinkAction = createPostLinkAction();
if (isNVPTX && Args.hasArg(options::OPT_fsycl_embed_ir)) {
// When compiling for Nvidia/CUDA devices and the user requested the
// IR to be embedded in the application (via option), run the output
// of sycl-post-link (filetable referencing LLVM Bitcode + symbols)
// through the offload wrapper and link the resulting object to the
// application.
auto *WrapBitcodeAction = C.MakeAction<OffloadWrapperJobAction>(
PostLinkAction, types::TY_Object, true);
DA.add(*WrapBitcodeAction, *TC, BoundArch, Action::OFK_SYCL);
}
bool NoRDCFatStaticArchive =
!IsRDC &&
FullDeviceLinkAction->getType() == types::TY_Tempfilelist;
Expand Down
10 changes: 9 additions & 1 deletion clang/lib/Driver/ToolChains/Clang.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -9272,6 +9272,14 @@ void OffloadWrapper::ConstructJob(Compilation &C, const JobAction &JA,
createArgString("-link-opts=");
}

bool IsEmbeddedIR = cast<OffloadWrapperJobAction>(JA).isEmbeddedIR();
if (IsEmbeddedIR) {
// When the offload-wrapper is called to embed LLVM IR, add a prefix to
// the target triple to distinguish the LLVM IR from the actual device
// binary for that target.
TargetTripleOpt = ("llvm_" + TargetTripleOpt).str();
}

WrapperArgs.push_back(
C.getArgs().MakeArgString(Twine("-target=") + TargetTripleOpt));

Expand All @@ -9293,7 +9301,7 @@ void OffloadWrapper::ConstructJob(Compilation &C, const JobAction &JA,
assert(I.isFilename() && "Invalid input.");

if (I.getType() == types::TY_Tempfiletable ||
I.getType() == types::TY_Tempfilelist)
I.getType() == types::TY_Tempfilelist || IsEmbeddedIR)
// wrapper actual input files are passed via the batch job file table:
WrapperArgs.push_back(C.getArgs().MakeArgString("-batch"));
WrapperArgs.push_back(C.getArgs().MakeArgString(I.getFilename()));
Expand Down
2 changes: 1 addition & 1 deletion sycl-fusion/common/include/Kernel.h
View file
Original file line number Diff line number Diff line change
Expand Up @@ -34,7 +34,7 @@ enum class ParameterKind : uint32_t {
};

/// Different binary formats supported as input to the JIT compiler.
enum class BinaryFormat : uint32_t { INVALID, LLVM, SPIRV };
enum class BinaryFormat : uint32_t { INVALID, LLVM, SPIRV, PTX };

/// Information about a device intermediate representation module (e.g., SPIR-V,
/// LLVM IR) from DPC++.
Expand Down
1 change: 1 addition & 0 deletions sycl-fusion/common/lib/KernelIO.h
View file
Original file line number Diff line number Diff line change
Expand Up @@ -47,6 +47,7 @@ template <> struct ScalarEnumerationTraits<jit_compiler::BinaryFormat> {
static void enumeration(IO &IO, jit_compiler::BinaryFormat &BF) {
IO.enumCase(BF, "LLVM", jit_compiler::BinaryFormat::LLVM);
IO.enumCase(BF, "SPIRV", jit_compiler::BinaryFormat::SPIRV);
IO.enumCase(BF, "PTX", jit_compiler::BinaryFormat::PTX);
IO.enumCase(BF, "INVALID", jit_compiler::BinaryFormat::INVALID);
}
};
Expand Down
12 changes: 11 additions & 1 deletion sycl-fusion/jit-compiler/CMakeLists.txt
View file
Original file line number Diff line number Diff line change
Expand Up @@ -2,13 +2,15 @@
add_llvm_library(sycl-fusion
lib/KernelFusion.cpp
lib/JITContext.cpp
lib/translation/KernelTranslation.cpp
lib/translation/SPIRVLLVMTranslation.cpp
lib/fusion/FusionPipeline.cpp
lib/fusion/FusionHelper.cpp
lib/fusion/ModuleHelper.cpp
lib/helper/ConfigHelper.cpp

LINK_COMPONENTS
LINK_COMPONENTS
BitReader
Core
Support
Analysis
Expand All @@ -18,6 +20,10 @@ add_llvm_library(sycl-fusion
Linker
ScalarOpts
InstCombine
Target
TargetParser
MC
${LLVM_TARGETS_TO_BUILD}
)

target_include_directories(sycl-fusion
Expand All @@ -40,6 +46,10 @@ target_link_libraries(sycl-fusion
${CMAKE_THREAD_LIBS_INIT}
)

if("NVPTX" IN_LIST LLVM_TARGETS_TO_BUILD)
target_compile_definitions(sycl-fusion PRIVATE FUSION_JIT_SUPPORT_PTX)
endif()

if (BUILD_SHARED_LIBS)
if(NOT MSVC AND NOT APPLE)
# Manage symbol visibility through the linker to make sure no LLVM symbols
Expand Down
17 changes: 12 additions & 5 deletions sycl-fusion/jit-compiler/include/JITContext.h
View file
Original file line number Diff line number Diff line change
Expand Up @@ -36,17 +36,21 @@ using CacheKeyT =
std::optional<std::vector<NDRange>>>;

///
/// Wrapper around a SPIR-V binary.
class SPIRVBinary {
/// Wrapper around a kernel binary.
class KernelBinary {
public:
explicit SPIRVBinary(std::string Binary);
explicit KernelBinary(std::string &&Binary, BinaryFormat Format);

jit_compiler::BinaryAddress address() const;

size_t size() const;

BinaryFormat format() const;

private:
std::string Blob;

BinaryFormat Format;
};

///
Expand All @@ -61,7 +65,10 @@ class JITContext {

llvm::LLVMContext *getLLVMContext();

SPIRVBinary &emplaceSPIRVBinary(std::string Binary);
template <typename... Ts> KernelBinary &emplaceKernelBinary(Ts &&...Args) {
WriteLockT WriteLock{BinariesMutex};
return Binaries.emplace_back(std::forward<Ts>(Args)...);
}

std::optional<SYCLKernelInfo> getCacheEntry(CacheKeyT &Identifier) const;

Expand All @@ -79,7 +86,7 @@ class JITContext {

MutexT BinariesMutex;

std::vector<SPIRVBinary> Binaries;
std::vector<KernelBinary> Binaries;

mutable MutexT CacheMutex;

Expand Down
7 changes: 6 additions & 1 deletion sycl-fusion/jit-compiler/include/Options.h
View file
Original file line number Diff line number Diff line change
Expand Up @@ -9,12 +9,14 @@
#ifndef SYCL_FUSION_JIT_COMPILER_OPTIONS_H
#define SYCL_FUSION_JIT_COMPILER_OPTIONS_H

#include "Kernel.h"

#include <memory>
#include <unordered_map>

namespace jit_compiler {

enum OptionID { VerboseOutput, EnableCaching };
enum OptionID { VerboseOutput, EnableCaching, TargetFormat };

class OptionPtrBase {};

Expand Down Expand Up @@ -78,6 +80,9 @@ struct JITEnableVerbose : public OptionBase<OptionID::VerboseOutput, bool> {};

struct JITEnableCaching : public OptionBase<OptionID::EnableCaching, bool> {};

struct JITTargetFormat
: public OptionBase<OptionID::TargetFormat, BinaryFormat> {};

} // namespace option
} // namespace jit_compiler

Expand Down
17 changes: 6 additions & 11 deletions sycl-fusion/jit-compiler/lib/JITContext.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -11,29 +11,24 @@

using namespace jit_compiler;

SPIRVBinary::SPIRVBinary(std::string Binary) : Blob{std::move(Binary)} {}
KernelBinary::KernelBinary(std::string &&Binary, BinaryFormat Fmt)
: Blob{std::move(Binary)}, Format{Fmt} {}

jit_compiler::BinaryAddress SPIRVBinary::address() const {
jit_compiler::BinaryAddress KernelBinary::address() const {
// FIXME: Verify it's a good idea to perform this reinterpret_cast here.
return reinterpret_cast<jit_compiler::BinaryAddress>(Blob.c_str());
}

size_t SPIRVBinary::size() const { return Blob.size(); }
size_t KernelBinary::size() const { return Blob.size(); }

BinaryFormat KernelBinary::format() const { return Format; }

JITContext::JITContext() : LLVMCtx{new llvm::LLVMContext}, Binaries{} {}

JITContext::~JITContext() = default;

llvm::LLVMContext *JITContext::getLLVMContext() { return LLVMCtx.get(); }

SPIRVBinary &JITContext::emplaceSPIRVBinary(std::string Binary) {
WriteLockT WriteLock{BinariesMutex};
// NOTE: With C++17, which returns a reference from emplace_back, the
// following code would be even simpler.
Binaries.emplace_back(std::move(Binary));
return Binaries.back();
}

std::optional<SYCLKernelInfo>
JITContext::getCacheEntry(CacheKeyT &Identifier) const {
ReadLockT ReadLock{CacheMutex};
Expand Down
Loading

0 comments on commit a93e59d

Please sign in to comment.