Clang Offload Bundler

Introduction

For heterogeneous single source programming languages, use one or more --offload-arch=<target-id> Clang options to specify the target IDs of the code to generate for the offload code regions.

The tool chain may perform multiple compilations of a translation unit to produce separate code objects for the host and potentially multiple offloaded devices. The clang-offload-bundler tool may be used as part of the tool chain to combine these multiple code objects into a single bundled code object.

The tool chain may use a bundled code object as an intermediate step so that each tool chain step consumes and produces a single file as in traditional non-heterogeneous tool chains. The bundled code object contains the code objects for the host and all the offload devices.

A bundled code object may also be used to bundle just the offloaded code objects, and embedded as data into the host code object. The host compilation includes an init function that will use the runtime corresponding to the offload kind (see Bundled Code Object Offload Kind) to load the offload code objects appropriate to the devices present when the host program is executed.

clang-offload-bundler is located in clang/tools/clang-offload-bundler.

$ clang-offload-bundler -help
OVERVIEW: A tool to bundle several input files of the specified type <type>
referring to the same source file but different targets into a single
one. The resulting file can also be unbundled into different files by
this tool if -unbundle is provided.

USAGE: clang-offload-bundler [options]

OPTIONS:

Generic Options:

  --help                  - Display available options (--help-hidden for more)
  --help-list             - Display list of available options (--help-list-hidden for more)
  --version               - Display the version of this program

clang-offload-bundler options:

  --###                   - Print any external commands that are to be executed instead of actually executing them - for testing purposes.
  --allow-missing-bundles - Create empty files if bundles are missing when unbundling.
  --bundle-align=<uint>   - Alignment of bundle for binary files
  --check-input-archive   - Check if input heterogeneous archive is valid in terms of TargetID rules.
  --inputs=<string>       - [<input file>,...]
  --list                  - List bundle IDs in the bundled file.
  --outputs=<string>      - [<output file>,...]
  --targets=<string>      - [<offload kind>-<target triple>,...]
  --type=<string>         - Type of the files to be bundled/unbundled.
                            Current supported types are:
                              i   - cpp-output
                              ii  - c++-cpp-output
                              cui - cuda/hip-output
                              d   - dependency
                              ll  - llvm
                              bc  - llvm-bc
                              s   - assembler
                              o   - object
                              a   - archive of bundled files
                              gch - precompiled-header
                              ast - clang AST file
  --unbundle              - Unbundle bundled file into several output files.

Usage

This tool can be used as follows for bundling:

clang-offload-bundler -targets=triple1,triple2 -type=ii -inputs=a.triple1.ii,a.triple2.ii -outputs=a.ii

or, it can be used as follows for unbundling:

clang-offload-bundler -targets=triple1,triple2 -type=ii -outputs=a.triple1.ii,a.triple2.ii -inputs=a.ii -unbundle

Supported File Formats

Multiple text and binary file formats are supported for bundling/unbundling. See Supported File Formats for a list of currently supported input formats. Use the File Type column to determine the value to pass to the --type option based on the type of input files while bundling/unbundling.

Supported File Formats

File Format

File Type

Text/Binary

CPP output

i

Text

C++ CPP output

ii

Text

CUDA/HIP output

cui

Text

Dependency

d

Text

LLVM

ll

Text

LLVM Bitcode

bc

Binary

Assembler

s

Text

Object

o

Binary

Archive of bundled files

a

Binary

Precompiled header

gch

Binary

Clang AST file

ast

Binary

Bundled Text File Layout

The text file formats are concatenated with comments that have a magic string and bundle entry ID in between. The BNF syntax to represent a code object bundle file is:

<file>    ::== <bundle> | <bundle> <file>
<bundle>  ::== <comment> <start> <bundle_id> <eol> <bundle> <eol>
               <comment> end <bundle_id> <eol>
<start>   ::== OFFLOAD_BUNDLER_MAGIC_STR__START__
<end>     ::== OFFLOAD_BUNDLER_MAGIC_STR__END__
comment

The symbol used for starting single-line comment in the file type of constituting bundles. E.g. it is “;” for ll File Type and “#” for “s” File Type.

bundle_id

The Bundle Entry ID for the enclosing bundle.

eol

The end of line character.

bundle

The code object stored in one of the supported text file formats.

OFFLOAD_BUNDLER_MAGIC_STR__

Magic string that marks the existence of offloading data i.e. “__CLANG_OFFLOAD_BUNDLE__”.

Bundled Binary File Layout

The layout of a bundled code object is defined by the following table:

Bundled Code Object Layout

Field

Type

Size in Bytes

Description

Magic String

string

24

__CLANG_OFFLOAD_BUNDLE__

Number Of Bundle Entries

integer

8

Number of bundle entries.

1st Bundle Entry Code Object Offset

integer

8

Byte offset from beginning of bundled code object to 1st code object.

1st Bundle Entry Code Object Size

integer

8

Byte size of 1st code object.

1st Bundle Entry ID Length

integer

8

Character length of bundle entry ID of 1st code object.

1st Bundle Entry ID

string

1st Bundle Entry ID Length

Bundle entry ID of 1st code object. This is not NUL terminated. See Bundle Entry ID.

...

Nth Bundle Entry Code Object Offset

integer

8

Nth Bundle Entry Code Object Size

integer

8

Nth Bundle Entry ID Length

integer

8

Nth Bundle Entry ID

string

1st Bundle Entry ID Length

1st Bundle Entry Code Object

bytes

1st Bundle Entry Code Object Size

...

Nth Bundle Entry Code Object

bytes

Nth Bundle Entry Code Object Size

Bundle Entry ID

Each entry in a bundled code object (see Bundled Text File Layout and Bundled Binary File Layout) has a bundle entry ID that indicates the kind of the entry’s code object and the runtime that manages it.

Bundle entry ID syntax is defined by the following BNF syntax:

<bundle-entry-id> ::== <offload-kind> "-" <target-triple> [ "-" <target-id> ]

Where:

offload-kind

The runtime responsible for managing the bundled entry code object. See Bundled Code Object Offload Kind.

Bundled Code Object Offload Kind

Offload Kind

Description

host

Host code object. clang-offload-bundler always includes this entry as the first bundled code object entry. For an embedded bundled code object this entry is not used by the runtime and so is generally an empty code object.

hip

Offload code object for the HIP language. Used for all HIP language offload code objects when the clang-offload-bundler is used to bundle code objects as intermediate steps of the tool chain. Also used for AMD GPU code objects before ABI version V4 when the clang-offload-bundler is used to create a fat binary to be loaded by the HIP runtime. The fat binary can be loaded directly from a file, or be embedded in the host code object as a data section with the name .hip_fatbin.

hipv4

Offload code object for the HIP language. Used for AMD GPU code objects with at least ABI version V4 and above when the clang-offload-bundler is used to create a fat binary to be loaded by the HIP runtime. The fat binary can be loaded directly from a file, or be embedded in the host code object as a data section with the name .hip_fatbin.

openmp

Offload code object for the OpenMP language extension.

Note: The distinction between the hip and hipv4 offload kinds is historically based. Originally, these designations might have indicated different versions of the code object ABI. However, as the system has evolved, the ABI version is now embedded directly within the code object itself, making these historical distinctions irrelevant during the unbundling process. Consequently, hip and hipv4 are treated as compatible in current implementations, facilitating interchangeable handling of code objects without differentiation based on offload kind.

target-triple

The target triple of the code object. See Target Triple.

The bundler accepts target triples with or without the optional environment field:

<arch><sub>-<vendor>-<sys>, or <arch><sub>-<vendor>-<sys>-<env>

However, in order to standardize outputs for tools that consume bitcode bundles, bundles written by the bundler internally use only the 4-field target triple:

<arch><sub>-<vendor>-<sys>-<env>

target-id

The canonical target ID of the code object. Present only if the target supports a target ID. See Target ID.

Bundled Code Object Composition

  • Each entry of a bundled code object must have a different bundle entry ID.

  • There can be multiple entries for the same processor provided they differ in target feature settings.

  • If there is an entry with a target feature specified as Any, then all entries must specify that target feature as Any for the same processor.

There may be additional target specific restrictions.

Compatibility Rules for Bundle Entry ID

A code object, specified using its Bundle Entry ID, can be loaded and executed on a target processor, if:

  • Their offload kinds are the same or comptible.

  • Their target triples are compatible.

  • Their Target IDs are compatible as defined in Compatibility Rules for Target ID.

Target ID

A target ID is used to indicate the processor and optionally its configuration, expressed by a set of target features, that affect ISA generation. It is target specific if a target ID is supported, or if the target triple alone is sufficient to specify the ISA generation.

It is used with the -mcpu=<target-id> and --offload-arch=<target-id> Clang compilation options to specify the kind of code to generate.

It is also used as part of the bundle entry ID to identify the code object. See Bundle Entry ID.

Target ID syntax is defined by the following BNF syntax:

<target-id> ::== <processor> ( ":" <target-feature> ( "+" | "-" ) )*

Where:

processor

Is a the target specific processor or any alternative processor name.

target-feature

Is a target feature name that is supported by the processor. Each target feature must appear at most once in a target ID and can have one of three values:

Any

Specified by omitting the target feature from the target ID. A code object compiled with a target ID specifying the default value of a target feature can be loaded and executed on a processor configured with the target feature on or off.

On

Specified by +, indicating the target feature is enabled. A code object compiled with a target ID specifying a target feature on can only be loaded on a processor configured with the target feature on.

Off

specified by -, indicating the target feature is disabled. A code object compiled with a target ID specifying a target feature off can only be loaded on a processor configured with the target feature off.

Compatibility Rules for Target ID

A code object compiled for a Target ID is considered compatible for a target, if:

  • Their processor is same.

  • Their feature set is compatible as defined above.

There are two forms of target ID:

Non-Canonical Form

The non-canonical form is used as the input to user commands to allow the user greater convenience. It allows both the primary and alternative processor name to be used and the target features may be specified in any order.

Canonical Form

The canonical form is used for all generated output to allow greater convenience for tools that consume the information. It is also used for internal passing of information between tools. Only the primary and not alternative processor name is used and the target features are specified in alphabetic order. Command line tools convert non-canonical form to canonical form.

Target Specific information

Target specific information is available for the following:

AMD GPU

AMD GPU supports target ID and target features. See User Guide for AMDGPU Backend which defines the processors and target features supported.

Most other targets do not support target IDs.

Archive Unbundling

Unbundling of a heterogeneous device archive (HDA) is done to create device specific archives. HDA is in a format compatible with GNU ar utility and contains a collection of bundled device binaries where each bundle file will contain device binaries for a host and one or more targets. The output device-specific archive is in a format compatible with GNU ar utility and contains a collection of device binaries for a specific target.

Heterogeneous Device Archive, HDA = {F1.X, F2.X, ..., FN.Y}
where, Fi = Bundle{Host-DeviceBinary, T1-DeviceBinary, T2-DeviceBinary, ...,
                   Tm-DeviceBinary},
       Ti = {Target i, qualified using Bundle Entry ID},
       X/Y = \*.bc for AMDGPU and \*.cubin for NVPTX

Device Specific Archive, DSA(Tk) = {F1-Tk-DeviceBinary.X, F2-Tk-DeviceBinary.X, ...
                                    FN-Tk-DeviceBinary.Y}
where, Fi-Tj-DeviceBinary.X represents device binary of i-th bundled device
binary file for target Tj.

The clang-offload-bundler extracts compatible device binaries for a given target from the bundled device binaries in a heterogeneous device archive and creates a target-specific device archive without bundling.

The clang-offload-bundler determines whether a device binary is compatible with a target by comparing bundle IDs. Two bundle IDs are considered compatible if:

  • Their offload kinds are the same

  • Their target triples are the same

  • Their Target IDs are the same

Creating a Heterogeneous Device Archive

  1. Compile source file(s) to generate object file(s)

clang -O2 -fopenmp -fopenmp-targets=amdgcn-amd-amdhsa,amdgcn-amd-amdhsa,\
   nvptx64-nvidia-cuda, nvptx64-nvidia-cuda \
  -Xopenmp-target=amdgcn-amd-amdhsa -march=gfx906:sramecc-:xnack+ \
  -Xopenmp-target=amdgcn-amd-amdhsa -march=gfx906:sramecc+:xnack+ \
  -Xopenmp-target=nvptx64-nvidia-cuda -march=sm_70 \
  -Xopenmp-target=nvptx64-nvidia-cuda -march=sm_80 \
  -c func_1.c -o func_1.o

clang -O2 -fopenmp -fopenmp-targets=amdgcn-amd-amdhsa,amdgcn-amd-amdhsa,
  nvptx64-nvidia-cuda, nvptx64-nvidia-cuda \
  -Xopenmp-target=amdgcn-amd-amdhsa -march=gfx906:sramecc-:xnack+ \
  -Xopenmp-target=amdgcn-amd-amdhsa -march=gfx906:sramecc+:xnack+ \
  -Xopenmp-target=nvptx64-nvidia-cuda -march=sm_70 \
  -Xopenmp-target=nvptx64-nvidia-cuda -march=sm_80 \
  -c func_2.c -o func_2.o
  1. Create a heterogeneous device archive by combining all the object file(s)

llvm-ar cr libFatArchive.a func_1.o func_2.o

Extracting a Device Specific Archive

UnbundleArchive takes a heterogeneous device archive file (“.a”) as input containing bundled device binary files, and a list of offload targets (not host), and extracts the device binaries into a new archive file for each offload target. Each resulting archive file contains all device binaries compatible with that particular offload target. Compatibility between a device binary in HDA and a target is based on the compatibility between their bundle entry IDs as defined in Compatibility Rules for Bundle Entry ID.

Following cases may arise during compatibility testing:

  • A binary is compatible with one or more targets: Insert the binary into the device-specific archive of each compatible target.

  • A binary is not compatible with any target: Skip the binary.

  • One or more binaries are compatible with a target: Insert all binaries into the device-specific archive of the target. The insertion need not be ordered.

  • No binary is compatible with a target: If allow-missing-bundles option is present then create an empty archive for the target. Otherwise, produce an error without creating an archive.

The created archive file does not contain an index of the symbols and device binary files are named as <<Parent Bundle Name>-<DeviceBinary’s TargetID>>, with ‘:’ replaced with ‘_’.

Usage

clang-offload-bundler --unbundle --inputs=libFatArchive.a -type=a \
 -targets=openmp-amdgcn-amdhsa-gfx906:sramecc+:xnack+, \
          openmp-amdgcn-amdhsa-gfx908:sramecc-:xnack+  \
 -outputs=devicelib-gfx906.a,deviceLib-gfx908.a

Additional Options while Archive Unbundling

-allow-missing-bundles

Create an empty archive file if no compatible device binary is found.

-check-input-archive

Check if input heterogeneous device archive follows rules for composition as defined in Bundled Code Object Composition before creating device-specific archive(s).

-debug-only=CodeObjectCompatibility

Verbose printing of matched/unmatched comparisons between bundle entry id of a device binary from HDA and bundle entry ID of a given target processor (see Compatibility Rules for Bundle Entry ID).

Compression and Decompression

clang-offload-bundler provides features to compress and decompress the full bundle, leveraging inherent redundancies within the bundle entries. Use the -compress command-line option to enable this compression capability.

The compressed offload bundle begins with a header followed by the compressed binary data:

  • Magic Number (4 bytes):

    This is a unique identifier to distinguish compressed offload bundles. The value is the string ‘CCOB’ (Compressed Clang Offload Bundle).

  • Version Number (16-bit unsigned int):

    This denotes the version of the compressed offload bundle format. The current version is 2.

  • Compression Method (16-bit unsigned int):

    This field indicates the compression method used. The value corresponds to either zlib or zstd, represented as a 16-bit unsigned integer cast from the LLVM compression enumeration.

  • Total File Size (32-bit unsigned int):

    This is the total size (in bytes) of the file, including the header. Available in version 2 and above.

  • Uncompressed Binary Size (32-bit unsigned int):

    This is the size (in bytes) of the binary data before it was compressed.

  • Hash (64-bit unsigned int):

    This is a 64-bit truncated MD5 hash of the uncompressed binary data. It serves for verification and caching purposes.

  • Compressed Data:

    The actual compressed binary data follows the header. Its size can be inferred from the total size of the file minus the header size.