diff options
Diffstat (limited to 'llvm/lib/Target/AMDGPU/MCTargetDesc')
5 files changed, 135 insertions, 5 deletions
diff --git a/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUCodeObjectMetadata.h b/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUCodeObjectMetadata.h index d25f131529a..da893163406 100644 --- a/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUCodeObjectMetadata.h +++ b/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUCodeObjectMetadata.h @@ -274,6 +274,81 @@ struct Metadata final { } // end namespace Arg +//===----------------------------------------------------------------------===// +// Kernel Code Properties Metadata. +//===----------------------------------------------------------------------===// +namespace CodeProps { + +namespace Key { +/// \brief Key for Kernel::CodeProps::mKernargSegmentSize. +constexpr char KernargSegmentSize[] = "KernargSegmentSize"; +/// \brief Key for Kernel::CodeProps::mWorkgroupGroupSegmentSize. +constexpr char WorkgroupGroupSegmentSize[] = "WorkgroupGroupSegmentSize"; +/// \brief Key for Kernel::CodeProps::mWorkitemPrivateSegmentSize. +constexpr char WorkitemPrivateSegmentSize[] = "WorkitemPrivateSegmentSize"; +/// \brief Key for Kernel::CodeProps::mWavefrontNumSGPRs. +constexpr char WavefrontNumSGPRs[] = "WavefrontNumSGPRs"; +/// \brief Key for Kernel::CodeProps::mWorkitemNumVGPRs. +constexpr char WorkitemNumVGPRs[] = "WorkitemNumVGPRs"; +/// \brief Key for Kernel::CodeProps::mKernargSegmentAlign. +constexpr char KernargSegmentAlign[] = "KernargSegmentAlign"; +/// \brief Key for Kernel::CodeProps::mGroupSegmentAlign. +constexpr char GroupSegmentAlign[] = "GroupSegmentAlign"; +/// \brief Key for Kernel::CodeProps::mPrivateSegmentAlign. +constexpr char PrivateSegmentAlign[] = "PrivateSegmentAlign"; +/// \brief Key for Kernel::CodeProps::mWavefrontSize. +constexpr char WavefrontSize[] = "WavefrontSize"; +} // end namespace Key + +/// \brief In-memory representation of kernel code properties metadata. +struct Metadata final { + /// \brief Size in bytes of the kernarg segment memory. Kernarg segment memory + /// holds the values of the arguments to the kernel. Optional. + uint64_t mKernargSegmentSize = 0; + /// \brief Size in bytes of the group segment memory required by a workgroup. + /// This value does not include any dynamically allocated group segment memory + /// that may be added when the kernel is dispatched. Optional. + uint32_t mWorkgroupGroupSegmentSize = 0; + /// \brief Size in bytes of the private segment memory required by a workitem. + /// Private segment memory includes arg, spill and private segments. Optional. + uint32_t mWorkitemPrivateSegmentSize = 0; + /// \brief Total number of SGPRs used by a wavefront. Optional. + uint16_t mWavefrontNumSGPRs = 0; + /// \brief Total number of VGPRs used by a workitem. Optional. + uint16_t mWorkitemNumVGPRs = 0; + /// \brief Maximum byte alignment of variables used by the kernel in the + /// kernarg memory segment. Expressed as a power of two. Optional. + uint8_t mKernargSegmentAlign = 0; + /// \brief Maximum byte alignment of variables used by the kernel in the + /// group memory segment. Expressed as a power of two. Optional. + uint8_t mGroupSegmentAlign = 0; + /// \brief Maximum byte alignment of variables used by the kernel in the + /// private memory segment. Expressed as a power of two. Optional. + uint8_t mPrivateSegmentAlign = 0; + /// \brief Wavefront size. Expressed as a power of two. Optional. + uint8_t mWavefrontSize = 0; + + /// \brief Default constructor. + Metadata() = default; + + /// \returns True if kernel code properties metadata is empty, false + /// otherwise. + bool empty() const { + return !notEmpty(); + } + + /// \returns True if kernel code properties metadata is not empty, false + /// otherwise. + bool notEmpty() const { + return mKernargSegmentSize || mWorkgroupGroupSegmentSize || + mWorkitemPrivateSegmentSize || mWavefrontNumSGPRs || + mWorkitemNumVGPRs || mKernargSegmentAlign || mGroupSegmentAlign || + mPrivateSegmentAlign || mWavefrontSize; + } +}; + +} // end namespace CodeProps + namespace Key { /// \brief Key for Kernel::Metadata::mName. constexpr char Name[] = "Name"; @@ -285,6 +360,8 @@ constexpr char LanguageVersion[] = "LanguageVersion"; constexpr char Attrs[] = "Attrs"; /// \brief Key for Kernel::Metadata::mArgs. constexpr char Args[] = "Args"; +/// \brief Key for Kernel::Metadata::mCodeProps. +constexpr char CodeProps[] = "CodeProps"; } // end namespace Key /// \brief In-memory representation of kernel metadata. @@ -299,6 +376,8 @@ struct Metadata final { Attrs::Metadata mAttrs = Attrs::Metadata(); /// \brief Arguments metadata. Optional. std::vector<Arg::Metadata> mArgs = std::vector<Arg::Metadata>(); + /// \brief Code properties metadata. Optional. + CodeProps::Metadata mCodeProps = CodeProps::Metadata(); /// \brief Default constructor. Metadata() = default; diff --git a/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUCodeObjectMetadataStreamer.cpp b/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUCodeObjectMetadataStreamer.cpp index 304cf5252ae..4a247677211 100644 --- a/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUCodeObjectMetadataStreamer.cpp +++ b/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUCodeObjectMetadataStreamer.cpp @@ -154,6 +154,30 @@ struct MappingTraits<Kernel::Arg::Metadata> { }; template <> +struct MappingTraits<Kernel::CodeProps::Metadata> { + static void mapping(IO &YIO, Kernel::CodeProps::Metadata &MD) { + YIO.mapOptional(Kernel::CodeProps::Key::KernargSegmentSize, + MD.mKernargSegmentSize, uint64_t(0)); + YIO.mapOptional(Kernel::CodeProps::Key::WorkgroupGroupSegmentSize, + MD.mWorkgroupGroupSegmentSize, uint32_t(0)); + YIO.mapOptional(Kernel::CodeProps::Key::WorkitemPrivateSegmentSize, + MD.mWorkitemPrivateSegmentSize, uint32_t(0)); + YIO.mapOptional(Kernel::CodeProps::Key::WavefrontNumSGPRs, + MD.mWavefrontNumSGPRs, uint16_t(0)); + YIO.mapOptional(Kernel::CodeProps::Key::WorkitemNumVGPRs, + MD.mWorkitemNumVGPRs, uint16_t(0)); + YIO.mapOptional(Kernel::CodeProps::Key::KernargSegmentAlign, + MD.mKernargSegmentAlign, uint8_t(0)); + YIO.mapOptional(Kernel::CodeProps::Key::GroupSegmentAlign, + MD.mGroupSegmentAlign, uint8_t(0)); + YIO.mapOptional(Kernel::CodeProps::Key::PrivateSegmentAlign, + MD.mPrivateSegmentAlign, uint8_t(0)); + YIO.mapOptional(Kernel::CodeProps::Key::WavefrontSize, + MD.mWavefrontSize, uint8_t(0)); + } +}; + +template <> struct MappingTraits<Kernel::Metadata> { static void mapping(IO &YIO, Kernel::Metadata &MD) { YIO.mapRequired(Kernel::Key::Name, MD.mName); @@ -164,6 +188,8 @@ struct MappingTraits<Kernel::Metadata> { YIO.mapOptional(Kernel::Key::Attrs, MD.mAttrs); if (!MD.mArgs.empty() || !YIO.outputting()) YIO.mapOptional(Kernel::Key::Args, MD.mArgs); + if (!MD.mCodeProps.empty() || !YIO.outputting()) + YIO.mapOptional(Kernel::Key::CodeProps, MD.mCodeProps); } }; @@ -531,13 +557,31 @@ void MetadataStreamer::emitKernelArg(const DataLayout &DL, Type *Ty, Arg.mTypeName = TypeName; } +void MetadataStreamer::emitKernelCodeProps( + const amd_kernel_code_t &KernelCode) { + auto &CodeProps = CodeObjectMetadata.mKernels.back().mCodeProps; + + CodeProps.mKernargSegmentSize = KernelCode.kernarg_segment_byte_size; + CodeProps.mWorkgroupGroupSegmentSize = + KernelCode.workgroup_group_segment_byte_size; + CodeProps.mWorkitemPrivateSegmentSize = + KernelCode.workitem_private_segment_byte_size; + CodeProps.mWavefrontNumSGPRs = KernelCode.wavefront_sgpr_count; + CodeProps.mWorkitemNumVGPRs = KernelCode.workitem_vgpr_count; + CodeProps.mKernargSegmentAlign = KernelCode.kernarg_segment_alignment; + CodeProps.mGroupSegmentAlign = KernelCode.group_segment_alignment; + CodeProps.mPrivateSegmentAlign = KernelCode.private_segment_alignment; + CodeProps.mWavefrontSize = KernelCode.wavefront_size; +} + void MetadataStreamer::begin(const FeatureBitset &Features, const Module &Mod) { emitVersion(); emitIsa(Features); emitPrintf(Mod); } -void MetadataStreamer::emitKernel(const Function &Func) { +void MetadataStreamer::emitKernel(const Function &Func, + const amd_kernel_code_t &KernelCode) { if (Func.getCallingConv() != CallingConv::AMDGPU_KERNEL) return; @@ -548,6 +592,7 @@ void MetadataStreamer::emitKernel(const Function &Func) { emitKernelLanguage(Func); emitKernelAttrs(Func); emitKernelArgs(Func); + emitKernelCodeProps(KernelCode); } ErrorOr<std::string> MetadataStreamer::toYamlString() { diff --git a/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUCodeObjectMetadataStreamer.h b/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUCodeObjectMetadataStreamer.h index 3c1e51e9271..5d51d6f4655 100644 --- a/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUCodeObjectMetadataStreamer.h +++ b/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUCodeObjectMetadataStreamer.h @@ -17,6 +17,7 @@ #define LLVM_LIB_TARGET_AMDGPU_MCTARGETDESC_AMDGPUCODEOBJECTMETADATASTREAMER_H #include "AMDGPUCodeObjectMetadata.h" +#include "AMDKernelCodeT.h" #include "llvm/ADT/StringRef.h" #include "llvm/Support/ErrorOr.h" @@ -72,6 +73,9 @@ private: StringRef TypeQual = "", StringRef BaseTypeName = "", StringRef AccQual = "", StringRef Name = "", StringRef TypeName = ""); + + void emitKernelCodeProps(const amd_kernel_code_t &KernelCode); + public: MetadataStreamer() = default; ~MetadataStreamer() = default; @@ -80,7 +84,7 @@ public: void end() {} - void emitKernel(const Function &Func); + void emitKernel(const Function &Func, const amd_kernel_code_t &KernelCode); ErrorOr<std::string> toYamlString(); diff --git a/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUTargetStreamer.cpp b/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUTargetStreamer.cpp index 963171f4462..100f1ba84b9 100644 --- a/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUTargetStreamer.cpp +++ b/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUTargetStreamer.cpp @@ -47,8 +47,9 @@ void AMDGPUTargetStreamer::EmitStartOfCodeObjectMetadata( CodeObjectMetadataStreamer.begin(Features, Mod); } -void AMDGPUTargetStreamer::EmitKernelCodeObjectMetadata(const Function &Func) { - CodeObjectMetadataStreamer.emitKernel(Func); +void AMDGPUTargetStreamer::EmitKernelCodeObjectMetadata( + const Function &Func, const amd_kernel_code_t &KernelCode) { + CodeObjectMetadataStreamer.emitKernel(Func, KernelCode); } void AMDGPUTargetStreamer::EmitEndOfCodeObjectMetadata( diff --git a/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUTargetStreamer.h b/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUTargetStreamer.h index 0307d3c5507..0d0fcbfb230 100644 --- a/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUTargetStreamer.h +++ b/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUTargetStreamer.h @@ -52,7 +52,8 @@ public: virtual void EmitStartOfCodeObjectMetadata(const FeatureBitset &Features, const Module &Mod); - virtual void EmitKernelCodeObjectMetadata(const Function &Func); + virtual void EmitKernelCodeObjectMetadata( + const Function &Func, const amd_kernel_code_t &KernelCode); virtual void EmitEndOfCodeObjectMetadata(const FeatureBitset &Features); |
