The SPIR-V Tools project provides an API and commands for processing SPIR-V modules.
The project includes an assembler, binary module parser, disassembler, validator, and optimizer for SPIR-V. Except for the optimizer, all are based on a common static library. The library contains all of the implementation details, and is used in the standalone tools whilst also enabling integration into other code bases directly. The optimizer implementation resides in its own library, which depends on the core library.
The interfaces have stabilized: We don't anticipate making a breaking change for existing features.
SPIR-V is defined by the Khronos Group Inc. See the SPIR-V Registry for the SPIR-V specification, headers, and XML registry.
See CHANGES for a high level summary of recent changes, by version.
SPIRV-Tools project version numbers are of the form vyear.index and with an optional -dev suffix to indicate work in progress. For example, the following versions are ordered from oldest to newest:
v2016.0v2016.1-devv2016.1v2016.2-devv2016.2Use the --version option on each command line tool to see the software version. An API call reports the software version as a C-style string.
Some versions of SPIRV-Tools are tagged as stable releases (see tags on github). These versions undergo extra testing. Releases are not directly related to releases (or versions) of SPIRV-Headers. Releases of SPIRV-Tools are tested against the version of SPIRV-Headers listed in the DEPS file. The release generally uses the most recent compatible version of SPIRV-Headers available at the time of release. No version of SPIRV-Headers other than the one listed in the DEPS file is guaranteed to work with the SPIRV-Tools release.
OpConstant, OpSpecConstant, and OpSwitch.See docs/syntax.md for the assembly language syntax.
The validator checks validation rules described by the SPIR-V specification.
Khronos recommends that tools that create or transform SPIR-V modules use the validator to ensure their outputs are valid, and that tools that consume SPIR-V modules optionally use the validator to protect themselves from bad inputs. This is especially encouraged for debug and development scenarios.
The validator has one-sided error: it will only return an error when it has implemented a rule check and the module violates that rule.
The validator is incomplete. See the CHANGES file for reports on completed work, and the Validator sub-project for planned and in-progress work.
Note: The validator checks some Universal Limits, from section 2.17 of the SPIR-V spec. The validator will fail on a module that exceeds those minimum upper bound limits. It is future work to parameterize the validator to allow larger limits accepted by a more than minimally capable SPIR-V consumer.
The optimizer is a collection of code transforms, or “passes”. Transforms are written for a diverse set of reasons:
As of this writing, there are 67 transforms including examples such as:
OpSpecConstantOp and OpSpecConstantCompositeAdditionally, certain sets of transformations have been packaged into higher-level recipes. These include:
spirv-opt -Os)spirv-opt -O)For the latest list with detailed documentation, please refer to include/spirv-tools/optimizer.hpp.
For suggestions on using the code reduction options, please refer to this white paper.
Note: The linker is still under development.
Current features:
See the CHANGES file for reports on completed work, and the General sub-project for planned and in-progress work.
Note: The reducer is still under development.
The reducer simplifies and shrinks a SPIR-V module with respect to a user-supplied interestingness function. For example, given a large SPIR-V module that cause some SPIR-V compiler to fail with a given fatal error message, the reducer could be used to look for a smaller version of the module that causes the compiler to fail with the same fatal error message.
To suggest an additional capability for the reducer, file an issue with “Reducer:” as the start of its title.
Note: The fuzzer is still under development.
The fuzzer applies semantics-preserving transformations to a SPIR-V binary module, to produce an equivalent module. The original and transformed modules should produce essentially identical results when executed on identical inputs: their results should differ only due to floating-point round-off, if at all. Significant differences in results can pinpoint bugs in tools that process SPIR-V binaries, such as miscompilations. This metamorphic testing approach is similar to the method used by the GraphicsFuzz project for fuzzing of GLSL shaders.
To suggest an additional capability for the fuzzer, file an issue with “Fuzzer:” as the start of its title.
spirv-tools-vimsyntax generates file spvasm.vim. Copy that file into your $HOME/.vim/syntax directory to get SPIR-V assembly syntax highlighting in Vim. This build target is not built by default.The SPIR-V Tools project is maintained by members of the The Khronos Group Inc., and is hosted at https://github.com/KhronosGroup/SPIRV-Tools.
Consider joining the public_spirv_tools_dev@khronos.org mailing list, via https://www.khronos.org/spir/spirv-tools-mailing-list/. The mailing list is used to discuss development plans for the SPIRV-Tools as an open source project. Once discussion is resolved, specific work is tracked via issues and sometimes in one of the projects.
(To provide feedback on the SPIR-V specification, file an issue on the SPIRV-Headers GitHub repository.)
See docs/projects.md to see how we use the GitHub Project feature to organize planned and in-progress work.
Contributions via merge request are welcome. Changes should:
clang-format. kokoro/check-format/build.sh shows how to download it. Note that we currently use clang-format version 5.0.0 for SPIRV-Tools. Settings are defined by the included .clang-format file.We intend to maintain a linear history on the GitHub master branch.
Example of getting sources, assuming SPIRV-Tools is configured as a standalone project:
git clone https://github.com/KhronosGroup/SPIRV-Tools.git spirv-tools cd spirv-tools # Check out sources for dependencies, at versions known to work together, # as listed in the DEPS file. python3 utils/git-sync-deps
For some kinds of development, you may need the latest sources from the third-party projects:
git clone https://github.com/KhronosGroup/SPIRV-Headers.git spirv-tools/external/spirv-headers git clone https://github.com/google/googletest.git spirv-tools/external/googletest git clone https://github.com/google/effcee.git spirv-tools/external/effcee git clone https://github.com/google/re2.git spirv-tools/external/re2
Some tests depend on the Effcee library for stateful matching. Effcee itself depends on RE2.
external/effcee and RE2 sources to appear in external/re2.example: demo code of using SPIRV-Tools APIsexternal/googletest: Intended location for the googletest sources, not providedexternal/effcee: Location of Effcee sources, if the effcee library is not already configured by an enclosing project.external/re2: Location of RE2 sources, if the re2 library is not already configured by an enclosing project. (The Effcee project already requires RE2.)include/: API clients should add this directory to the include search pathexternal/spirv-headers: Intended location for SPIR-V headers, not providedinclude/spirv-tools/libspirv.h: C API public interfacesource/: API implementationtest/: Tests, using the googletest frameworktools/: Command line executablesThe project contains a number of tests, used to drive development and ensure correctness. The tests are written using the googletest framework. The googletest source is not provided with this project. There are two ways to enable tests:
googletest before configuring SPIR-V Tools.googletest source into the <spirv-dir>/external/googletest directory before configuring and building the project.Note: Prebuilt binaries are available from the downloads page.
First get the sources. Then build using CMake, Bazel, Android ndk-build, or the Emscripten SDK.
You can build the project using CMake:
cd <spirv-dir> mkdir build && cd build cmake [-G <platform-generator>] <spirv-dir>
Once the build files have been generated, build using the appropriate build command (e.g. ninja, make, msbuild, etc.; this depends on the platform generator used above), or use your IDE, or use CMake to run the appropriate build command for you:
cmake --build . [--config Debug] # runs `make` or `ninja` or `msbuild` etc.
The SPIR-V fuzzer, spirv-fuzz, can only be built via CMake, and is disabled by default. To build it, clone protobuf and use the SPIRV_BUILD_FUZZER CMake option, like so:
# In <spirv-dir> (the SPIRV-Tools repo root): git clone --depth=1 --branch v3.13.0.1 https://github.com/protocolbuffers/protobuf external/protobuf # In your build directory: cmake [-G <platform-generator>] <spirv-dir> -DSPIRV_BUILD_FUZZER=ON cmake --build . --config Debug
You can also add -DSPIRV_ENABLE_LONG_FUZZER_TESTS=ON to build additional fuzzer tests.
You can also use Bazel to build the project.
cd <spirv-dir> bazel build :all
The SPIRV-Tools core library can be built to a WebAssembly node.js module. The resulting SpirvTools WebAssembly module only exports methods to assemble and disassemble SPIR-V modules.
First, make sure you have the Emscripten SDK. Then:
cd <spirv-dir> ./source/wasm/build.sh
The resulting node package, with JavaScript and TypeScript bindings, is written to <spirv-dir>/out/web.
Note: This builds the package locally. It does not publish it to npm.
To test the result:
node ./test/wasm/test.js
For building and testing SPIRV-Tools, the following tools should be installed regardless of your OS:
SPIRV-Tools is regularly tested with the following compilers:
On Linux
On MacOS
On Windows
Other compilers or later versions may work, but they are not tested.
The following CMake options are supported:
SPIRV_BUILD_FUZZER={ON|OFF}, default OFF - Build the spirv-fuzz tool.SPIRV_COLOR_TERMINAL={ON|OFF}, default ON - Enables color console output.SPIRV_SKIP_TESTS={ON|OFF}, default OFF- Build only the library and the command line tools. This will prevent the tests from being built.SPIRV_SKIP_EXECUTABLES={ON|OFF}, default OFF- Build only the library, not the command line tools and tests.SPIRV_USE_SANITIZER=<sanitizer>, default is no sanitizing - On UNIX platforms with an appropriate version of clang this option enables the use of the sanitizers documented here. This should only be used with a debug build.SPIRV_WARN_EVERYTHING={ON|OFF}, default OFF - On UNIX platforms enable more strict warnings. The code might not compile with this option enabled. For Clang, enables -Weverything. For GCC, enables -Wpedantic. See CMakeLists.txt for details.SPIRV_WERROR={ON|OFF}, default ON - Forces a compilation error on any warnings encountered by enabling the compiler-specific compiler front-end option. No compiler front-end options are enabled when this option is OFF.Additionally, you can pass additional C preprocessor definitions to SPIRV-Tools via setting SPIRV_TOOLS_EXTRA_DEFINITIONS. For example, by setting it to /D_ITERATOR_DEBUG_LEVEL=0 on Windows, you can disable checked iterators and iterator debugging.
SPIR-V Tools supports building static libraries libSPIRV-Tools.a and libSPIRV-Tools-opt.a for Android:
cd <spirv-dir>
export ANDROID_NDK=/path/to/your/ndk
mkdir build && cd build
mkdir libs
mkdir app
$ANDROID_NDK/ndk-build -C ../android_test \
NDK_PROJECT_PATH=. \
NDK_LIBS_OUT=`pwd`/libs \
NDK_APP_OUT=`pwd`/app
Occasionally the entries in DEPS will need to be updated. This is done on demand when there is a request to do this, often due to downstream breakages. To update DEPS, run utils/roll_deps.sh and confirm that tests pass. The script requires Chromium's depot_tools.
The internals of the library use C++11 features, and are exposed via both a C and C++ API.
In order to use the library from an application, the include path should point to <spirv-dir>/include, which will enable the application to include the header <spirv-dir>/include/spirv-tools/libspirv.h{|pp} then linking against the static library in <spirv-build-dir>/source/libSPIRV-Tools.a or <spirv-build-dir>/source/SPIRV-Tools.lib. For optimization, the header file is <spirv-dir>/include/spirv-tools/optimizer.hpp, and the static library is <spirv-build-dir>/source/libSPIRV-Tools-opt.a or <spirv-build-dir>/source/SPIRV-Tools-opt.lib.
SPIRV-Tools CMake target: Creates the static library:<spirv-build-dir>/source/libSPIRV-Tools.a on Linux and OS X.<spirv-build-dir>/source/libSPIRV-Tools.lib on Windows.SPIRV-Tools-opt CMake target: Creates the static library:<spirv-build-dir>/source/libSPIRV-Tools-opt.a on Linux and OS X.<spirv-build-dir>/source/libSPIRV-Tools-opt.lib on Windows.The interfaces are still under development, and are expected to change.
There are five main entry points into the library in the C interface:
spvTextToBinary: An assembler, translating text to a binary SPIR-V module.spvBinaryToText: A disassembler, translating a binary SPIR-V module to text.spvBinaryParse: The entry point to a binary parser API. It issues callbacks for the header and each parsed instruction. The disassembler is implemented as a client of spvBinaryParse.spvValidate implements the validator functionality. IncompletespvValidateBinary implements the validator functionality. IncompleteThe C++ interface is comprised of three classes, SpirvTools, Optimizer and Linker, all in the spvtools namespace.
SpirvTools provides Assemble, Disassemble, and Validate methods.Optimizer provides methods for registering and running optimization passes.Linker provides methods for combining together multiple binaries.Command line tools, which wrap the above library functions, are provided to assemble or disassemble shader files. It's a convention to name SPIR-V assembly and binary files with suffix .spvasm and .spv, respectively.
The assembler reads the assembly language text, and emits the binary form.
The standalone assembler is the executable called spirv-as, and is located in <spirv-build-dir>/tools/spirv-as. The functionality of the assembler is implemented by the spvTextToBinary library function.
spirv-as - the standalone assembler<spirv-dir>/tools/asUse option -h to print help.
The disassembler reads the binary form, and emits assembly language text.
The standalone disassembler is the executable called spirv-dis, and is located in <spirv-build-dir>/tools/spirv-dis. The functionality of the disassembler is implemented by the spvBinaryToText library function.
spirv-dis - the standalone disassembler<spirv-dir>/tools/disUse option -h to print help.
The output includes syntax colouring when printing to the standard output stream, on Linux, Windows, and OS X.
The linker combines multiple SPIR-V binary modules together, resulting in a single binary module as output.
This is a work in progress. The linker does not support OpenCL program linking options related to math flags. (See section 5.6.5.2 in OpenCL 1.2)
spirv-link - the standalone linker<spirv-dir>/tools/linkThe optimizer processes a SPIR-V binary module, applying transformations in the specified order.
This is a work in progress, with initially only few available transformations.
spirv-opt - the standalone optimizer<spirv-dir>/tools/optWarning: This functionality is under development, and is incomplete.
The standalone validator is the executable called spirv-val, and is located in <spirv-build-dir>/tools/spirv-val. The functionality of the validator is implemented by the spvValidate library function.
The validator operates on the binary form.
spirv-val - the standalone validator<spirv-dir>/tools/valThe reducer shrinks a SPIR-V binary module, guided by a user-supplied interestingness test.
This is a work in progress, with initially only shrinks a module in a few ways.
spirv-reduce - the standalone reducer<spirv-dir>/tools/reduceRun spirv-reduce --help to see how to specify interestingness.
The fuzzer transforms a SPIR-V binary module into a semantically-equivalent SPIR-V binary module by applying transformations in a randomized fashion.
This is a work in progress, with initially only a few semantics-preserving transformations.
spirv-fuzz - the standalone fuzzer<spirv-dir>/tools/fuzzRun spirv-fuzz --help for a detailed list of options.
The control flow dumper prints the control flow graph for a SPIR-V module as a GraphViz graph.
This is experimental.
spirv-cfg - the control flow graph dumper<spirv-dir>/tools/cfgspirv-lesspipe.sh - Automatically disassembles .spv binary files for the less program, on compatible systems. For example, set the LESSOPEN environment variable as follows, assuming both spirv-lesspipe.sh and spirv-dis are on your executable search path:
export LESSOPEN='| spirv-lesspipe.sh "%s"'
Then you page through a disassembled module as follows:
less foo.spv
spirv-lesspipe.sh script will pass through any extra arguments to spirv-dis. So, for example, you can turn off colours and friendly ID naming as follows:export LESSOPEN='| spirv-lesspipe.sh "%s" --no-color --raw-id'
vim-spirv - A vim plugin which supports automatic disassembly of .spv files using the :edit command and assembly using the :write command. The plugin also provides additional features which include; syntax highlighting; highlighting of all ID's matching the ID under the cursor; and highlighting errors where the Instruction operand of OpExtInst is used without an appropriate OpExtInstImport.
50spirv-tools.el - Automatically disassembles ‘.spv’ binary files when loaded into the emacs text editor, and re-assembles them when saved, provided any modifications to the file are valid. This functionality must be explicitly requested by defining the symbol SPIRV_TOOLS_INSTALL_EMACS_HELPERS as follows:
cmake -DSPIRV_TOOLS_INSTALL_EMACS_HELPERS=true ...
In addition, this helper is only installed if the directory /etc/emacs/site-start.d exists, which is typically true if emacs is installed on the system.
Note that symbol IDs are not currently preserved through a load/edit/save operation. This may change if the ability is added to spirv-as.
Tests are only built when googletest is found.
Use ctest -j <num threads> to run all the tests. To run tests using all threads:
ctest -j$(nproc)
To run a single test target, use ctest [-j <N>] -R <test regex>. For example, you can run all opt tests with:
ctest -R 'spirv-tools-test_opt'
Use bazel test :all to run all tests. This will run tests in parallel by default.
To run a single test target, specify :my_test_target instead of :all. Test target names get printed when you run bazel test :all. For example, you can run opt_def_use_test with:
bazel test :opt_def_use_test
See the projects pages for more information.
OpLabel instructions with that basic block's predecessors.This is a work in progress.
Full license terms are in LICENSE
Copyright (c) 2015-2016 The Khronos Group Inc.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.