reprospect.tools.binaries.elf module

This module provides tools to parse CUDA ELF files.

References:

class reprospect.tools.binaries.elf.CuInfo(note_version: int, virtual_sm: int, toolkit_version: int)View on GitHub 

Bases: object

Specialized decoder for .note.nv.cuinfo section.

Here is a typical section:

.section .note.nv.cuinfo
Entry: 1
    Owner                  Data size        Description
    NVIDIA Corp            8                NVIDIA CUDA Information
        Note Version: 2
        CUDA Virtual SM: sm_100
        CUDA Tool Kit Version: 13.0

References:

https://github.com/tek-life/cuda-gdb/blob/080b643333b48c1c3aba2ba01f9fc7408c7bf6df/gdb/cuda/cuda-sass-json.h##L64-L65

__init__(note_version: int, virtual_sm: int, toolkit_version: int) → None

classmethod decode(*, note: NoteSection) → Generator[CuInfo, None, None]View on GitHub : Iterate the note entries and decode them.

note_version: int

toolkit_version: int

virtual_sm: int

class reprospect.tools.binaries.elf.ELF(*, file: Path)View on GitHub 

Bases: object

Helper for reading ELF files and retrieve CUDA-specific information.

EF_CUDA_SM_OFFSET_POST_BLACKWELL: Final[int] = 8: Offset for compute capability field post BLACKWELL.

EF_CUDA_SM_POST_BLACKWELL: Final[int] = 65280: Mask for compute capability field post BLACKWELL.

EF_CUDA_SM_PRE_BLACKWELL: Final[int] = 255

Mask for compute capability field pre BLACKWELL.

References:

https://github.com/llvm/llvm-project/blob/12322b22c68a588caeee8702946695de0a8ba788/llvm/include/llvm/BinaryFormat/ELF.h#L932-L988

ELFABIVERSION_CUDA: Final[tuple[int, int]] = (7, 8)

References:

https://github.com/llvm/llvm-project/blob/46e9d6325a825b516826d0c56b6231abfaac16ab/llvm/include/llvm/BinaryFormat/ELF.h#L391-L392

ELFOSABI_CUDA: Final[tuple[int, ...]] = (41, 51, 65)

References:

https://github.com/llvm/llvm-project/blob/46e9d6325a825b516826d0c56b6231abfaac16ab/llvm/include/llvm/BinaryFormat/ELF.h#L364-L365

__enter__() → SelfView on GitHub 

__exit__(*args, **kwargs) → NoneView on GitHub 

__init__(*, file: Path) → NoneView on GitHub 

property arch: NVIDIAArchView on GitHub : Get compute capability encoded in header as NVIDIA architecture.

classmethod compute_capability(value) → ComputeCapabilityView on GitHub : Return compute capability encoded in e_flags.

elf: ELFFile | None

file: Path: Path to the ELF file.

property header: ContainerView on GitHub 

property is_cuda: boolView on GitHub : Return True if file is a valid CUDA binary file.

classmethod is_cuda_impl(*, header: Container) → boolView on GitHub 

nvinfo(mangled: str) → NvInfoView on GitHub : Extract and parse the .nv.info.<mangled> section.

class reprospect.tools.binaries.elf.NvInfo(attributes: tuple[NvInfoEntry, ...])View on GitHub 

Bases: object

Specialized decoder for a .nv.info.<mangled> section.

Here is a typical section as parsed by cuobjdump:

.nv.info._ZN6Kokkos4Impl33cuda_parallel_launch_local_memoryINS0_11ParallelForINS0_11ThenWrapperIN10reprospect8examples6kokkos5graph7FunctorINS_4ViewIiJNS_12CudaUVMSpaceEEEEEEEENS_11RangePolicyIJNS_4CudaENS0_16IsGraphKernelTagENS_12LaunchBoundsILj1ELj0EEEEEESF_EELj1ELj0EEEvT_
    <0x1>
    Attribute: EIATTR_CUDA_API_VERSION
    Format: EIFMT_SVAL
    Value: 0x82
    <0x2>
    Attribute: EIATTR_SW2861232_WAR
    Format: EIFMT_NVAL
    <0x3>
    Attribute: EIATTR_PARAM_CBANK
    Format: EIFMT_SVAL
    Value: 0x9 0x380160
    <0x4>
    Attribute: EIATTR_CBANK_PARAM_SIZE
    Format: EIFMT_HVAL
    Value: 0x38
    <0x5>
    Attribute: EIATTR_KPARAM_INFO
    Format: EIFMT_SVAL
    Value: Index : 0x0 Ordinal : 0x0 Offset  : 0x0 Size    : 0x38
        Pointee's logAlignment : 0x0 Space : 0x0 cbank : 0x1f Parameter Space : CBANK
    <0x6>
    Attribute: EIATTR_MAXREG_COUNT
    Format: EIFMT_HVAL
    Value: 0xff
    <0x7>
    Attribute: EIATTR_MERCURY_ISA_VERSION
    Format: EIFMT_HVAL
    Value: 0.0
    <0x8>
    Attribute: EIATTR_EXIT_INSTR_OFFSETS
    Format: EIFMT_SVAL
    Value: 0x60 0x170
    <0x9>
    Attribute: EIATTR_MAX_THREADS
    Format: EIFMT_SVAL
    Value: 0x1 0x1 0x1
    <0x10>
    Attribute: EIATTR_CRS_STACK_SIZE
    Format: EIFMT_SVAL
    Value: 0x0

References:

__init__(attributes: tuple[NvInfoEntry, ...]) → None

attributes: tuple[NvInfoEntry, ...]

classmethod decode(*, section: Section) → NvInfoView on GitHub 

iter(eiattr: NvInfoEIATTR) → Generator[NvInfoEntry, None, None]View on GitHub : Get attribute(s) of type eiattr.

classmethod parse(*, data: bytes) → NvInfoView on GitHub : Parse a single .nv.info.<mangled> section according to the per-format rules.

class reprospect.tools.binaries.elf.NvInfoEIATTR(*values)View on GitHub 

Bases: IntEnum

Attribute type that can be found in a .nv.info.<mangled> section.

References:

[HHH+19]

CBANK_PARAM_SIZE = 25

CRS_STACK_SIZE = 30

CUDA_API_VERSION = 55

EXIT_INSTR_OFFSETS = 28

KPARAM_INFO = 23

MAXREG_COUNT = 27

MAX_THREADS = 5

MERCURY_ISA_VERSION = 95

PARAM_CBANK = 10

SPARSE_MMA_MASK = 80

SW2861232_WAR = 53

SW_WAR = 54

VRC_CTA_INIT_COUNT = 74

WARP_WIDE_INSTR_OFFSETS = 49

class reprospect.tools.binaries.elf.NvInfoEIFMT(*values)View on GitHub 

Bases: IntEnum

Attribute format in a .nv.info.<mangled> section.

References:

[HHH+19]

BVAL = 2

HVAL = 3

NVAL = 1

SVAL = 4

class reprospect.tools.binaries.elf.NvInfoEntry(eifmt: NvInfoEIFMT, eiattr: NvInfoEIATTR, value: bytes | int | tuple[int, ...] | None = None)View on GitHub 

Bases: object

A single entry in a .nv.info.<mangled> section.

__init__(eifmt: NvInfoEIFMT, eiattr: NvInfoEIATTR, value: bytes | int | tuple[int, ...] | None = None) → None

eiattr: NvInfoEIATTR

eifmt: NvInfoEIFMT

value: bytes | int | tuple[int, ...] | None

class reprospect.tools.binaries.elf.TkInfo(note_version: int, object_filename: str, tool_name: str, tool_version: str, tool_branch: str, tool_options: str)View on GitHub 

Bases: object

Specialized decoder for .note.nv.tkinfo section.

Here is a typical section:

.section .note.nv.tkinfo
Entry: 1
    Owner                  Data size        Description
    NVIDIA Corp            140              NVIDIA CUDA Toolkit Information
        Note Version: 2
        Tool Name: ptxas
        Tool Version: Cuda compilation tools, release 13.0, V13.0.48
        Tool Branch: Build cuda_13.0.r13.0/compiler.36260728_0
        Tool Command Line Arguments: -arch sm_100 -m 64

References:

__init__(note_version: int, object_filename: str, tool_name: str, tool_version: str, tool_branch: str, tool_options: str) → None

classmethod decode(*, note: NoteSection) → Generator[TkInfo, None, None]View on GitHub : Iterate over the note entries and decode them.

static extract(arr: bytes, offset: int) → strView on GitHub : The strings section starts at offset 24.

note_version: int

object_filename: str

tool_branch: str

tool_name: str

tool_options: str

tool_version: str