8 .. _clang-offload-bundler:
13 For heterogeneous single source programming languages, use one or more
14 ``--offload-arch=<target-id>`` Clang options to specify the target IDs of the
15 code to generate for the offload code regions.
17 The tool chain may perform multiple compilations of a translation unit to
18 produce separate code objects for the host and potentially multiple offloaded
19 devices. The ``clang-offload-bundler`` tool may be used as part of the tool
20 chain to combine these multiple code objects into a single bundled code object.
22 The tool chain may use a bundled code object as an intermediate step so that
23 each tool chain step consumes and produces a single file as in traditional
24 non-heterogeneous tool chains. The bundled code object contains the code objects
25 for the host and all the offload devices.
27 A bundled code object may also be used to bundle just the offloaded code
28 objects, and embedded as data into the host code object. The host compilation
29 includes an ``init`` function that will use the runtime corresponding to the
30 offload kind (see :ref:`clang-offload-kind-table`) to load the offload code
31 objects appropriate to the devices present when the host program is executed.
33 :program:`clang-offload-bundler` is located in
34 `clang/tools/clang-offload-bundler`.
36 .. code-block:: console
38 $ clang-offload-bundler -help
39 OVERVIEW: A tool to bundle several input files of the specified type <type>
40 referring to the same source file but different targets into a single
41 one. The resulting file can also be unbundled into different files by
42 this tool if -unbundle is provided.
44 USAGE: clang-offload-bundler [options]
50 --help - Display available options (--help-hidden for more)
51 --help-list - Display list of available options (--help-list-hidden for more)
52 --version - Display the version of this program
54 clang-offload-bundler options:
56 --### - Print any external commands that are to be executed instead of actually executing them - for testing purposes.
57 --allow-missing-bundles - Create empty files if bundles are missing when unbundling.
58 --bundle-align=<uint> - Alignment of bundle for binary files
59 --check-input-archive - Check if input heterogeneous archive is valid in terms of TargetID rules.
60 --inputs=<string> - [<input file>,...]
61 --list - List bundle IDs in the bundled file.
62 --outputs=<string> - [<output file>,...]
63 --targets=<string> - [<offload kind>-<target triple>,...]
64 --type=<string> - Type of the files to be bundled/unbundled.
65 Current supported types are:
74 a - archive of bundled files
75 gch - precompiled-header
77 --unbundle - Unbundle bundled file into several output files.
82 This tool can be used as follows for bundling:
86 clang-offload-bundler -targets=triple1,triple2 -type=ii -inputs=a.triple1.ii,a.triple2.ii -outputs=a.ii
88 or, it can be used as follows for unbundling:
92 clang-offload-bundler -targets=triple1,triple2 -type=ii -outputs=a.triple1.ii,a.triple2.ii -inputs=a.ii -unbundle
95 Supported File Formats
96 ======================
98 Multiple text and binary file formats are supported for bundling/unbundling. See
99 :ref:`supported-file-formats-table` for a list of currently supported input
100 formats. Use the ``File Type`` column to determine the value to pass to the
101 ``--type`` option based on the type of input files while bundling/unbundling.
103 .. table:: Supported File Formats
104 :name: supported-file-formats-table
106 +--------------------------+----------------+-------------+
107 | File Format | File Type | Text/Binary |
108 +==========================+================+=============+
109 | CPP output | i | Text |
110 +--------------------------+----------------+-------------+
111 | C++ CPP output | ii | Text |
112 +--------------------------+----------------+-------------+
113 | CUDA/HIP output | cui | Text |
114 +--------------------------+----------------+-------------+
115 | Dependency | d | Text |
116 +--------------------------+----------------+-------------+
118 +--------------------------+----------------+-------------+
119 | LLVM Bitcode | bc | Binary |
120 +--------------------------+----------------+-------------+
121 | Assembler | s | Text |
122 +--------------------------+----------------+-------------+
123 | Object | o | Binary |
124 +--------------------------+----------------+-------------+
125 | Archive of bundled files | a | Binary |
126 +--------------------------+----------------+-------------+
127 | Precompiled header | gch | Binary |
128 +--------------------------+----------------+-------------+
129 | Clang AST file | ast | Binary |
130 +--------------------------+----------------+-------------+
132 .. _clang-bundled-code-object-layout-text:
134 Bundled Text File Layout
135 ========================
137 The text file formats are concatenated with comments that have a magic string
138 and bundle entry ID in between. The BNF syntax to represent a code object
143 <file> ::== <bundle> | <bundle> <file>
144 <bundle> ::== <comment> <start> <bundle_id> <eol> <bundle> <eol>
145 <comment> end <bundle_id> <eol>
146 <start> ::== OFFLOAD_BUNDLER_MAGIC_STR__START__
147 <end> ::== OFFLOAD_BUNDLER_MAGIC_STR__END__
150 The symbol used for starting single-line comment in the file type of
151 constituting bundles. E.g. it is ";" for ll ``File Type`` and "#" for "s"
155 The :ref:`clang-bundle-entry-id` for the enclosing bundle.
158 The end of line character.
161 The code object stored in one of the supported text file formats.
163 **OFFLOAD_BUNDLER_MAGIC_STR__**
164 Magic string that marks the existence of offloading data i.e.
165 "__CLANG_OFFLOAD_BUNDLE__".
167 .. _clang-bundled-code-object-layout:
169 Bundled Binary File Layout
170 ==========================
172 The layout of a bundled code object is defined by the following table:
174 .. table:: Bundled Code Object Layout
175 :name: bundled-code-object-layout-table
177 =================================== ======= ================ ===============================
178 Field Type Size in Bytes Description
179 =================================== ======= ================ ===============================
180 Magic String string 24 ``__CLANG_OFFLOAD_BUNDLE__``
181 Number Of Bundle Entries integer 8 Number of bundle entries.
182 1st Bundle Entry Code Object Offset integer 8 Byte offset from beginning of
183 bundled code object to 1st code
185 1st Bundle Entry Code Object Size integer 8 Byte size of 1st code object.
186 1st Bundle Entry ID Length integer 8 Character length of bundle
187 entry ID of 1st code object.
188 1st Bundle Entry ID string 1st Bundle Entry Bundle entry ID of 1st code
189 ID Length object. This is not NUL
191 :ref:`clang-bundle-entry-id`.
193 Nth Bundle Entry Code Object Offset integer 8
194 Nth Bundle Entry Code Object Size integer 8
195 Nth Bundle Entry ID Length integer 8
196 Nth Bundle Entry ID string 1st Bundle Entry
198 1st Bundle Entry Code Object bytes 1st Bundle Entry
201 Nth Bundle Entry Code Object bytes Nth Bundle Entry
203 =================================== ======= ================ ===============================
205 .. _clang-bundle-entry-id:
210 Each entry in a bundled code object (see :ref:`clang-bundled-code-object-layout-text`
211 and :ref:`clang-bundled-code-object-layout`) has a bundle entry ID that indicates
212 the kind of the entry's code object and the runtime that manages it.
214 Bundle entry ID syntax is defined by the following BNF syntax:
218 <bundle-entry-id> ::== <offload-kind> "-" <target-triple> [ "-" <target-id> ]
223 The runtime responsible for managing the bundled entry code object. See
224 :ref:`clang-offload-kind-table`.
226 .. table:: Bundled Code Object Offload Kind
227 :name: clang-offload-kind-table
229 ============= ==============================================================
230 Offload Kind Description
231 ============= ==============================================================
232 host Host code object. ``clang-offload-bundler`` always includes
233 this entry as the first bundled code object entry. For an
234 embedded bundled code object this entry is not used by the
235 runtime and so is generally an empty code object.
237 hip Offload code object for the HIP language. Used for all
238 HIP language offload code objects when the
239 ``clang-offload-bundler`` is used to bundle code objects as
240 intermediate steps of the tool chain. Also used for AMD GPU
241 code objects before ABI version V4 when the
242 ``clang-offload-bundler`` is used to create a *fat binary*
243 to be loaded by the HIP runtime. The fat binary can be
244 loaded directly from a file, or be embedded in the host code
245 object as a data section with the name ``.hip_fatbin``.
247 hipv4 Offload code object for the HIP language. Used for AMD GPU
248 code objects with at least ABI version V4 and above when the
249 ``clang-offload-bundler`` is used to create a *fat binary*
250 to be loaded by the HIP runtime. The fat binary can be
251 loaded directly from a file, or be embedded in the host code
252 object as a data section with the name ``.hip_fatbin``.
254 openmp Offload code object for the OpenMP language extension.
255 ============= ==============================================================
257 Note: The distinction between the `hip` and `hipv4` offload kinds is historically based.
258 Originally, these designations might have indicated different versions of the
259 code object ABI. However, as the system has evolved, the ABI version is now embedded
260 directly within the code object itself, making these historical distinctions irrelevant
261 during the unbundling process. Consequently, `hip` and `hipv4` are treated as compatible
262 in current implementations, facilitating interchangeable handling of code objects
263 without differentiation based on offload kind.
266 The target triple of the code object. See `Target Triple
267 <https://clang.llvm.org/docs/CrossCompilation.html#target-triple>`_.
269 The bundler accepts target triples with or without the optional environment
272 ``<arch><sub>-<vendor>-<sys>``, or
273 ``<arch><sub>-<vendor>-<sys>-<env>``
275 However, in order to standardize outputs for tools that consume bitcode
276 bundles, bundles written by the bundler internally use only the 4-field
279 ``<arch><sub>-<vendor>-<sys>-<env>``
282 The canonical target ID of the code object. Present only if the target
283 supports a target ID. See :ref:`clang-target-id`.
285 .. _code-object-composition:
287 Bundled Code Object Composition
288 -------------------------------
290 * Each entry of a bundled code object must have a different bundle entry ID.
291 * There can be multiple entries for the same processor provided they differ
292 in target feature settings.
293 * If there is an entry with a target feature specified as *Any*, then all
294 entries must specify that target feature as *Any* for the same processor.
296 There may be additional target specific restrictions.
298 .. _compatibility-bundle-entry-id:
300 Compatibility Rules for Bundle Entry ID
301 ---------------------------------------
303 A code object, specified using its Bundle Entry ID, can be loaded and
304 executed on a target processor, if:
306 * Their offload kinds are the same or comptible.
307 * Their target triples are compatible.
308 * Their Target IDs are compatible as defined in :ref:`compatibility-target-id`.
315 A target ID is used to indicate the processor and optionally its configuration,
316 expressed by a set of target features, that affect ISA generation. It is target
317 specific if a target ID is supported, or if the target triple alone is
318 sufficient to specify the ISA generation.
320 It is used with the ``-mcpu=<target-id>`` and ``--offload-arch=<target-id>``
321 Clang compilation options to specify the kind of code to generate.
323 It is also used as part of the bundle entry ID to identify the code object. See
324 :ref:`clang-bundle-entry-id`.
326 Target ID syntax is defined by the following BNF syntax:
330 <target-id> ::== <processor> ( ":" <target-feature> ( "+" | "-" ) )*
335 Is a the target specific processor or any alternative processor name.
338 Is a target feature name that is supported by the processor. Each target
339 feature must appear at most once in a target ID and can have one of three
343 Specified by omitting the target feature from the target ID.
344 A code object compiled with a target ID specifying the default
345 value of a target feature can be loaded and executed on a processor
346 configured with the target feature on or off.
349 Specified by ``+``, indicating the target feature is enabled. A code
350 object compiled with a target ID specifying a target feature on
351 can only be loaded on a processor configured with the target feature on.
354 specified by ``-``, indicating the target feature is disabled. A code
355 object compiled with a target ID specifying a target feature off
356 can only be loaded on a processor configured with the target feature off.
358 .. _compatibility-target-id:
360 Compatibility Rules for Target ID
361 ---------------------------------
363 A code object compiled for a Target ID is considered compatible for a
366 * Their processor is same.
367 * Their feature set is compatible as defined above.
369 There are two forms of target ID:
372 The non-canonical form is used as the input to user commands to allow the user
373 greater convenience. It allows both the primary and alternative processor name
374 to be used and the target features may be specified in any order.
377 The canonical form is used for all generated output to allow greater
378 convenience for tools that consume the information. It is also used for
379 internal passing of information between tools. Only the primary and not
380 alternative processor name is used and the target features are specified in
381 alphabetic order. Command line tools convert non-canonical form to canonical
384 Target Specific information
385 ===========================
387 Target specific information is available for the following:
390 AMD GPU supports target ID and target features. See `User Guide for AMDGPU Backend
391 <https://llvm.org/docs/AMDGPUUsage.html>`_ which defines the `processors
392 <https://llvm.org/docs/AMDGPUUsage.html#amdgpu-processors>`_ and `target
393 features <https://llvm.org/docs/AMDGPUUsage.html#amdgpu-target-features>`_
396 Most other targets do not support target IDs.
401 Unbundling of a heterogeneous device archive (HDA) is done to create device specific
402 archives. HDA is in a format compatible with GNU ``ar`` utility and contains a
403 collection of bundled device binaries where each bundle file will contain
404 device binaries for a host and one or more targets. The output device-specific
405 archive is in a format compatible with GNU ``ar`` utility and contains a
406 collection of device binaries for a specific target.
410 Heterogeneous Device Archive, HDA = {F1.X, F2.X, ..., FN.Y}
411 where, Fi = Bundle{Host-DeviceBinary, T1-DeviceBinary, T2-DeviceBinary, ...,
413 Ti = {Target i, qualified using Bundle Entry ID},
414 X/Y = \*.bc for AMDGPU and \*.cubin for NVPTX
416 Device Specific Archive, DSA(Tk) = {F1-Tk-DeviceBinary.X, F2-Tk-DeviceBinary.X, ...
417 FN-Tk-DeviceBinary.Y}
418 where, Fi-Tj-DeviceBinary.X represents device binary of i-th bundled device
419 binary file for target Tj.
421 The clang-offload-bundler extracts compatible device binaries for a given target
422 from the bundled device binaries in a heterogeneous device archive and creates
423 a target-specific device archive without bundling.
425 The clang-offload-bundler determines whether a device binary is compatible
426 with a target by comparing bundle IDs. Two bundle IDs are considered
429 * Their offload kinds are the same
430 * Their target triples are the same
431 * Their Target IDs are the same
433 Creating a Heterogeneous Device Archive
434 ---------------------------------------
436 1. Compile source file(s) to generate object file(s)
440 clang -O2 -fopenmp -fopenmp-targets=amdgcn-amd-amdhsa,amdgcn-amd-amdhsa,\
441 nvptx64-nvidia-cuda, nvptx64-nvidia-cuda \
442 -Xopenmp-target=amdgcn-amd-amdhsa -march=gfx906:sramecc-:xnack+ \
443 -Xopenmp-target=amdgcn-amd-amdhsa -march=gfx906:sramecc+:xnack+ \
444 -Xopenmp-target=nvptx64-nvidia-cuda -march=sm_70 \
445 -Xopenmp-target=nvptx64-nvidia-cuda -march=sm_80 \
446 -c func_1.c -o func_1.o
448 clang -O2 -fopenmp -fopenmp-targets=amdgcn-amd-amdhsa,amdgcn-amd-amdhsa,
449 nvptx64-nvidia-cuda, nvptx64-nvidia-cuda \
450 -Xopenmp-target=amdgcn-amd-amdhsa -march=gfx906:sramecc-:xnack+ \
451 -Xopenmp-target=amdgcn-amd-amdhsa -march=gfx906:sramecc+:xnack+ \
452 -Xopenmp-target=nvptx64-nvidia-cuda -march=sm_70 \
453 -Xopenmp-target=nvptx64-nvidia-cuda -march=sm_80 \
454 -c func_2.c -o func_2.o
456 2. Create a heterogeneous device archive by combining all the object file(s)
460 llvm-ar cr libFatArchive.a func_1.o func_2.o
462 Extracting a Device Specific Archive
463 ------------------------------------
465 UnbundleArchive takes a heterogeneous device archive file (".a") as input
466 containing bundled device binary files, and a list of offload targets (not
467 host), and extracts the device binaries into a new archive file for each
468 offload target. Each resulting archive file contains all device binaries
469 compatible with that particular offload target. Compatibility between a
470 device binary in HDA and a target is based on the compatibility between their
471 bundle entry IDs as defined in :ref:`compatibility-bundle-entry-id`.
473 Following cases may arise during compatibility testing:
475 * A binary is compatible with one or more targets: Insert the binary into the
476 device-specific archive of each compatible target.
477 * A binary is not compatible with any target: Skip the binary.
478 * One or more binaries are compatible with a target: Insert all binaries into
479 the device-specific archive of the target. The insertion need not be ordered.
480 * No binary is compatible with a target: If ``allow-missing-bundles`` option is
481 present then create an empty archive for the target. Otherwise, produce an
482 error without creating an archive.
484 The created archive file does not contain an index of the symbols and device
485 binary files are named as <<Parent Bundle Name>-<DeviceBinary's TargetID>>,
486 with ':' replaced with '_'.
493 clang-offload-bundler --unbundle --inputs=libFatArchive.a -type=a \
494 -targets=openmp-amdgcn-amdhsa-gfx906:sramecc+:xnack+, \
495 openmp-amdgcn-amdhsa-gfx908:sramecc-:xnack+ \
496 -outputs=devicelib-gfx906.a,deviceLib-gfx908.a
498 .. _additional-options-archive-unbundling:
500 Additional Options while Archive Unbundling
501 -------------------------------------------
503 **-allow-missing-bundles**
504 Create an empty archive file if no compatible device binary is found.
506 **-check-input-archive**
507 Check if input heterogeneous device archive follows rules for composition
508 as defined in :ref:`code-object-composition` before creating device-specific
511 **-debug-only=CodeObjectCompatibility**
512 Verbose printing of matched/unmatched comparisons between bundle entry id of
513 a device binary from HDA and bundle entry ID of a given target processor
514 (see :ref:`compatibility-bundle-entry-id`).
516 Compression and Decompression
517 =============================
519 ``clang-offload-bundler`` provides features to compress and decompress the full
520 bundle, leveraging inherent redundancies within the bundle entries. Use the
521 `-compress` command-line option to enable this compression capability.
523 The compressed offload bundle begins with a header followed by the compressed binary data:
525 - **Magic Number (4 bytes)**:
526 This is a unique identifier to distinguish compressed offload bundles. The value is the string 'CCOB' (Compressed Clang Offload Bundle).
528 - **Version Number (16-bit unsigned int)**:
529 This denotes the version of the compressed offload bundle format. The current version is `2`.
531 - **Compression Method (16-bit unsigned int)**:
532 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.
534 - **Total File Size (32-bit unsigned int)**:
535 This is the total size (in bytes) of the file, including the header. Available in version 2 and above.
537 - **Uncompressed Binary Size (32-bit unsigned int)**:
538 This is the size (in bytes) of the binary data before it was compressed.
540 - **Hash (64-bit unsigned int)**:
541 This is a 64-bit truncated MD5 hash of the uncompressed binary data. It serves for verification and caching purposes.
543 - **Compressed Data**:
544 The actual compressed binary data follows the header. Its size can be inferred from the total size of the file minus the header size.