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 v
year.
index and with
an optional -dev
suffix to indicate work in progress. For example, the
following versions are ordered from oldest to newest:
v2016.0
v2016.1-dev
v2016.1
v2016.2-dev
v2016.2
Use 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.
- Support for SPIR-V 1.0, through 1.5
- Based on SPIR-V syntax described by JSON grammar files in the SPIRV-Headers repository.
- Usually, support for a new version of SPIR-V is ready within days after publication.
- Support for extended instruction sets:
- GLSL std450 version 1.0 Rev 3
- OpenCL version 1.0 Rev 2
- Assembler only does basic syntax checking. No cross validation of
IDs or types is performed, except to check literal arguments to
OpConstant
,OpSpecConstant
, andOpSwitch
.
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:
- To restructure, simplify, or normalize the code for further processing.
- To eliminate undesirable code.
- To improve code quality in some metric such as size or performance. Note: These transforms are not guaranteed to actually improve any given metric. Users should always measure results for their own situation.
As of this writing, there are 67 transforms including examples such as:
- Simplification
- Strip debug info
- Strip reflection info
- Specialization Constants
- Set spec constant default value
- Freeze spec constant to default value
- Fold
OpSpecConstantOp
andOpSpecConstantComposite
- Unify constants
- Eliminate dead constant
- Code Reduction
- Inline all function calls exhaustively
- Convert local access chains to inserts/extracts
- Eliminate local load/store in single block
- Eliminate local load/store with single store
- Eliminate local load/store with multiple stores
- Eliminate local extract from insert
- Eliminate dead instructions (aggressive)
- Eliminate dead branches
- Merge single successor / single predecessor block pairs
- Eliminate common uniform loads
- Remove duplicates: Capabilities, extended instruction imports, types, and decorations.
- Normalization
- Compact IDs
- CFG cleanup
- Flatten decorations
- Merge returns
- Convert AMD-specific instructions to KHR instructions
- Code improvement
- Conditional constant propagation
- If-conversion
- Loop fission
- Loop fusion
- Loop-invariant code motion
- Loop unroll
- Other
- Graphics robust access
- Upgrade memory model to VulkanKHR
Additionally, certain sets of transformations have been packaged into higher-level recipes. These include:
- Optimization for size (
spirv-opt -Os
) - Optimization for performance (
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:
- Combine multiple SPIR-V binary modules together.
- Combine into a library (exports are retained) or an executable (no symbols are exported).
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.
Note: The diff tool is still under development.
The diff tool takes two SPIR-V files, either in binary or text format and produces a diff-style comparison between the two. The instructions between the src and dst modules are matched as best as the tool can, and output is produced (in src id-space) that shows which instructions are removed in src, added in dst or modified between them. The order of instructions are not retained.
Matching instructions between two SPIR-V modules is not trivial, and thus a number of heuristics are applied in this tool. In particular, without debug information, match functions is nontrivial as they can be reordered. As such, this tool is primarily useful to produce the diff of two SPIR-V modules derived from the same source, for example before and after a modification to the shader, before and after a transformation, or SPIR-V produced from different tools.
- Utility filters
- Build target
spirv-tools-vimsyntax
generates filespvasm.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:
- Be provided under the Apache 2.0.
- You'll be prompted with a one-time "click-through" Khronos Open Source Contributor License Agreement (CLA) dialog as part of submitting your pull request or other contribution to GitHub.
- Include tests to cover updated functionality.
- C++ code should follow the Google C++ Style Guide.
- Code should be formatted with
clang-format
. kokoro/check-format/build.sh shows how to download it. Note that we currently useclang-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.
- If SPIRV-Tools is configured as part of a larger project that already uses Effcee, then that project should include Effcee before SPIRV-Tools.
- Otherwise, SPIRV-Tools expects Effcee sources to appear in
external/effcee
and RE2 sources to appear inexternal/re2
.
example
: demo code of using SPIRV-Tools APIsexternal/googletest
: Intended location for the googletest sources, not providedexternal/effcee
: Location of Effcee sources, if theeffcee
library is not already configured by an enclosing project.external/re2
: Location of RE2 sources, if there2
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 executables
The 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:
- If SPIR-V Tools is configured as part of an enclosing project, then the
enclosing project should configure
googletest
before configuring SPIR-V Tools. - If SPIR-V Tools is configured as a standalone project, then download the
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:
- CMake: if using CMake for generating compilation targets, you need to install CMake Version 2.8.12 or later.
- Python 3: for utility scripts and running the test suite.
- Bazel (optional): if building the source with Bazel, you need to install Bazel Version 5.0.0 on your machine. Other versions may also work, but are not verified.
- Emscripten SDK (optional): if building the WebAssembly module.
SPIRV-Tools is regularly tested with the following compilers:
On Linux
- GCC version 9.3
- Clang version 10.0
On MacOS
- AppleClang 11.0
On Windows
- Visual Studio 2015
- Visual Studio 2017
Other compilers or later versions may work, but they are not tested.
The following CMake options are supported:
SPIRV_BUILD_FUZZER={ON|OFF}
, defaultOFF
- Build the spirv-fuzz tool.SPIRV_COLOR_TERMINAL={ON|OFF}
, defaultON
- Enables color console output.SPIRV_SKIP_TESTS={ON|OFF}
, defaultOFF
- Build only the library and the command line tools. This will prevent the tests from being built.SPIRV_SKIP_EXECUTABLES={ON|OFF}
, defaultOFF
- 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 ofclang
this option enables the use of the sanitizers documented here. This should only be used with a debug build.SPIRV_WARN_EVERYTHING={ON|OFF}
, defaultOFF
- On UNIX platforms enable more strict warnings. The code might not compile with this option enabled. For Clang, enables-Weverything
. For GCC, enables-Wpedantic
. SeeCMakeLists.txt
for details.SPIRV_WERROR={ON|OFF}
, defaultON
- 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 ofspvBinaryParse
.spvValidate
implements the validator functionality. IncompletespvValidateBinary
implements the validator functionality. Incomplete
The C++ interface is comprised of three classes, SpirvTools
, Optimizer
and
Linker
, all in the spvtools
namespace.
SpirvTools
providesAssemble
,Disassemble
, andValidate
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/as
Use 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/dis
Use 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/link
The 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/opt
Warning: 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/val
The 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/reduce
Run 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/fuzz
Run 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/cfg
Warning: This functionality is under development, and is incomplete.
The diff tool produces a diff-style comparison between two SPIR-V modules.
spirv-diff
- the standalone diff tool<spirv-dir>
/tools/diff`
-
spirv-lesspipe.sh
- Automatically disassembles.spv
binary files for theless
program, on compatible systems. For example, set theLESSOPEN
environment variable as follows, assuming bothspirv-lesspipe.sh
andspirv-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
- The
spirv-lesspipe.sh
script will pass through any extra arguments tospirv-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'
- The
-
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 theInstruction
operand ofOpExtInst
is used without an appropriateOpExtInstImport
. -
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.
- The disassembler could emit helpful annotations in comments. For example:
- Use variable name information from debug instructions to annotate key operations on variables.
- Show control flow information by annotating
OpLabel
instructions with that basic block's predecessors.
- Error messages could be improved.
This is a work in progress.
- The linker could accept math transformations such as allowing MADs, or other math flags passed at linking-time in OpenCL.
- Linkage attributes can not be applied through a group.
- Check decorations of linked functions attributes.
- Remove dead instructions, such as OpName targeting imported symbols.
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.