-
Building glslang as a DLL or shared library is now possible and supported.
-
The
GenericCodeGen
,MachineIndependent
,OSDependent
, andSPIRV
libraries have been integrated into the mainglslang
library. The old separate libraries have replaced with empty stubs for a temporary compatibility period, and they will be removed entirely in the future. -
A new CMake
ENABLE_SPIRV
option has been added to control whether glslang is built with SPIR-V support. Its default value isON
. -
OGLCompiler
andHLSL
stub libraries have been fully removed from the build.
There are several components:
An OpenGL GLSL and OpenGL|ES GLSL (ESSL) front-end for reference validation and translation of GLSL/ESSL into an internal abstract syntax tree (AST).
Status: Virtually complete, with results carrying similar weight as the specifications.
An HLSL front-end for translation of an approximation of HLSL to glslang's AST form.
Status: Partially complete. Semantics are not reference quality and input is not validated. This is in contrast to the DXC project, which receives a much larger investment and attempts to have definitive/reference-level semantics.
See issue 362 and issue 701 for current status.
Translates glslang's AST to the Khronos-specified SPIR-V intermediate language.
Status: Virtually complete.
An API for getting reflection information from the AST, reflection types/variables/etc. from the HLL source (not the SPIR-V).
Status: There is a large amount of functionality present, but no specification/goal to measure completeness against. It is accurate for the input HLL and AST, but only approximate for what would later be emitted for SPIR-V.
glslang
is command-line tool for accessing the functionality above.
Status: Complete.
Tasks waiting to be done are documented as GitHub issues.
Also see the Khronos landing page for glslang as a reference front end:
https://www.khronos.org/opengles/sdk/tools/Reference-Compiler/
The above page, while not kept up to date, includes additional information regarding glslang as a reference validator.
To use the standalone binary form, execute glslang
, and it will print
a usage statement. Basic operation is to give it a file containing a shader,
and it will print out warnings/errors and optionally an AST.
The applied stage-specific rules are based on the file extension:
.vert
for a vertex shader.tesc
for a tessellation control shader.tese
for a tessellation evaluation shader.geom
for a geometry shader.frag
for a fragment shader.comp
for a compute shader
For ray tracing pipeline shaders:
.rgen
for a ray generation shader.rint
for a ray intersection shader.rahit
for a ray any-hit shader.rchit
for a ray closest-hit shader.rmiss
for a ray miss shader.rcall
for a callable shader
There is also a non-shader extension:
.conf
for a configuration file of limits, see usage statement for example
Instead of building manually, you can also download the binaries for your platform directly from the main-tot release on GitHub. Those binaries are automatically uploaded by the buildbots after successful testing and they always reflect the current top of the tree of the main branch.
- A C++17 compiler. (For MSVS: use 2019 or later.)
- CMake: for generating compilation targets.
- make: Linux, ninja is an alternative, if configured.
- Python 3.x: for executing SPIRV-Tools scripts. (Optional if not using SPIRV-Tools and the 'External' subdirectory does not exist.)
- bison: optional, but needed when changing the grammar (glslang.y).
- googletest: optional, but should use if making any changes to glslang.
The following steps assume a Bash shell. On Windows, that could be the Git Bash shell or some other shell of your choosing.
cd <parent of where you want glslang to be>
git clone https://github.com/KhronosGroup/glslang.git
./update_glslang_sources.py
Assume the source directory is $SOURCE_DIR
and the build directory is $BUILD_DIR
.
CMake will create the $BUILD_DIR
for the user if it doesn't exist.
First change your working directory:
cd $SOURCE_DIR
For building on Linux:
cmake -B $BUILD_DIR -DCMAKE_BUILD_TYPE=Release -DCMAKE_INSTALL_PREFIX="$(pwd)/install"
# "Release" (for CMAKE_BUILD_TYPE) could also be "Debug" or "RelWithDebInfo"
For building on Android:
cmake -B $BUILD_DIR -G "Unix Makefiles" -DCMAKE_INSTALL_PREFIX="$(pwd)/install" -DANDROID_ABI=arm64-v8a -DCMAKE_BUILD_TYPE=Release -DANDROID_STL=c++_static -DANDROID_PLATFORM=android-24 -DCMAKE_SYSTEM_NAME=Android -DANDROID_TOOLCHAIN=clang -DANDROID_ARM_MODE=arm -DCMAKE_MAKE_PROGRAM=$ANDROID_NDK_HOME/prebuilt/linux-x86_64/bin/make -DCMAKE_TOOLCHAIN_FILE=$ANDROID_NDK_HOME/build/cmake/android.toolchain.cmake
# If on Windows will be -DCMAKE_MAKE_PROGRAM=%ANDROID_NDK_HOME%\prebuilt\windows-x86_64\bin\make.exe
# -G is needed for building on Windows
# -DANDROID_ABI can also be armeabi-v7a for 32 bit
For building on Windows:
cmake -B $BUILD_DIR -DCMAKE_INSTALL_PREFIX="$(pwd)/install"
# The CMAKE_INSTALL_PREFIX part is for testing (explained later).
Also, consider using git config --global core.fileMode false
(or with --local
) on Windows
to prevent the addition of execution permission on files.
# for Linux:
make -j4 install
# for Windows:
cmake --build . --config Release --target install
# "Release" (for --config) could also be "Debug", "MinSizeRel", or "RelWithDebInfo"
If using MSVC, after running CMake to configure, use the
Configuration Manager to check the INSTALL
project.
glslang can also be built with the GN build system.
Download depot_tools.zip,
extract to a directory, and add this directory to your PATH
.
This only needs to be done once after updating glslang
.
With the current directory set to your glslang
checkout, type:
./update_glslang_sources.py
gclient sync --gclientfile=standalone.gclient
gn gen out/Default
With the current directory set to your glslang
checkout, type:
cd out/Default
ninja
The grammar in glslang/MachineIndependent/glslang.y
has to be recompiled with
bison if it changes, the output files are committed to the repo to avoid every
developer needing to have bison configured to compile the project when grammar
changes are quite infrequent. For windows you can get binaries from
GnuWin32.
The command to rebuild is:
bison --defines=MachineIndependent/glslang_tab.cpp.h
-t MachineIndependent/glslang.y
-o MachineIndependent/glslang_tab.cpp
The above command is also available in the bash script in updateGrammar
,
when executed from the glslang subdirectory of the glslang repository.
Use the steps in Build Steps, with the following notes/exceptions:
emsdk
needs to be present in your executable search path, PATH for Bash-like environments:- Wrap cmake call:
emcmake cmake
- Set
-DENABLE_OPT=OFF
. - Set
-DENABLE_HLSL=OFF
if HLSL is not needed. - For a standalone JS/WASM library, turn on
-DENABLE_GLSLANG_JS=ON
. - To get a fully minimized build, make sure to use
brotli
to compress the .js and .wasm files - Note that by default, Emscripten allocates a very small stack size, which may cause stack overflows when compiling large shaders. Use the STACK_SIZE compiler setting to increase the stack size.
Example:
emcmake cmake -DCMAKE_BUILD_TYPE=Release -DENABLE_GLSLANG_JS=ON \
-DENABLE_HLSL=OFF -DENABLE_OPT=OFF ..
You can download and install glslang using the vcpkg dependency manager:
git clone https://github.com/Microsoft/vcpkg.git
cd vcpkg
./bootstrap-vcpkg.sh
./vcpkg integrate install
./vcpkg install glslang
The glslang port in vcpkg is kept up to date by Microsoft team members and community contributors. If the version is out of date, please create an issue or pull request on the vcpkg repository.
Right now, there are two test harnesses existing in glslang: one is Google
Test, one is the runtests
script. The former
runs unit tests and single-shader single-threaded integration tests, while
the latter runs multiple-shader linking tests and multi-threaded tests.
Tests may erroneously fail or pass if using ALLOW_EXTERNAL_SPIRV_TOOLS
with
any commit other than the one specified in known_good.json
.
The runtests
script requires compiled binaries to be
installed into $BUILD_DIR/install
. Please make sure you have supplied the
correct configuration to CMake (using -DCMAKE_INSTALL_PREFIX
) when building;
otherwise, you may want to modify the path in the runtests
script.
Running Google Test-backed tests:
cd $BUILD_DIR
# for Linux:
ctest
# for Windows:
ctest -C {Debug|Release|RelWithDebInfo|MinSizeRel}
# or, run the test binary directly
# (which gives more fine-grained control like filtering):
<dir-to-glslangtests-in-build-dir>/glslangtests
Running runtests
script-backed tests:
cd $SOURCE_DIR/Test && ./runtests
If some tests fail with validation errors, there may be a mismatch between the
version of spirv-val
on the system and the version of glslang. In this
case, it is necessary to run update_glslang_sources.py
. See "Check-Out
External Projects" above for more details.
Test results should always be included with a pull request that modifies functionality.
If you are writing unit tests, please use the Google Test framework and
place the tests under the gtests/
directory.
Integration tests are placed in the Test/
directory. It contains test input
and a subdirectory baseResults/
that contains the expected results of the
tests. Both the tests and baseResults/
are under source-code control.
Google Test runs those integration tests by reading the test input, compiling
them, and then compare against the expected results in baseResults/
. The
integration tests to run via Google Test is registered in various
gtests/*.FromFile.cpp
source files. glslangtests
provides a command-line
option --update-mode
, which, if supplied, will overwrite the golden files
under the baseResults/
directory with real output from that invocation.
For more information, please check gtests/
directory's
README.
For the runtests
script, it will generate current results in the
localResults/
directory and diff
them against the baseResults/
.
When you want to update the tracked test results, they need to be
copied from localResults/
to baseResults/
. This can be done by
the bump
shell script.
You can add your own private list of tests, not tracked publicly, by using
localtestlist
to list non-tracked tests. This is automatically read
by runtests
and included in the diff
and bump
process.
Another piece of software can programmatically translate shaders to an AST using one of two different interfaces:
- A new C++ class-oriented interface, or
- The original C functional interface
The main()
in StandAlone/StandAlone.cpp
shows examples using both styles.
This interface is in roughly the last 1/3 of ShaderLang.h
. It is in the
glslang namespace and contains the following, here with suggested calls
for generating SPIR-V:
const char* GetEsslVersionString();
const char* GetGlslVersionString();
bool InitializeProcess();
void FinalizeProcess();
class TShader
setStrings(...);
setEnvInput(EShSourceHlsl or EShSourceGlsl, stage, EShClientVulkan or EShClientOpenGL, 100);
setEnvClient(EShClientVulkan or EShClientOpenGL, EShTargetVulkan_1_0 or EShTargetVulkan_1_1 or EShTargetOpenGL_450);
setEnvTarget(EShTargetSpv, EShTargetSpv_1_0 or EShTargetSpv_1_3);
bool parse(...);
const char* getInfoLog();
class TProgram
void addShader(...);
bool link(...);
const char* getInfoLog();
Reflection queries
For just validating (not generating code), substitute these calls:
setEnvInput(EShSourceHlsl or EShSourceGlsl, stage, EShClientNone, 0);
setEnvClient(EShClientNone, 0);
setEnvTarget(EShTargetNone, 0);
See ShaderLang.h
and the usage of it in StandAlone/StandAlone.cpp
for more
details. There is a block comment giving more detail above the calls for
setEnvInput, setEnvClient, and setEnvTarget
.
This interface is in roughly the first 2/3 of ShaderLang.h
, and referred to
as the Sh*()
interface, as all the entry points start Sh
.
The Sh*()
interface takes a "compiler" call-back object, which it calls after
building call back that is passed the AST and can then execute a back end on it.
The following is a simplified resulting run-time call stack:
ShCompile(shader, compiler) -> compiler(AST) -> <back end>
In practice, ShCompile()
takes shader strings, default version, and
warning/error and other options for controlling compilation.
This interface is located glslang_c_interface.h
and exposes functionality similar to the C++ interface. The following snippet is a complete example showing how to compile GLSL into SPIR-V 1.5 for Vulkan 1.2.
#include <glslang/Include/glslang_c_interface.h>
// Required for use of glslang_default_resource
#include <glslang/Public/resource_limits_c.h>
typedef struct SpirVBinary {
uint32_t *words; // SPIR-V words
int size; // number of words in SPIR-V binary
} SpirVBinary;
SpirVBinary compileShaderToSPIRV_Vulkan(glslang_stage_t stage, const char* shaderSource, const char* fileName) {
const glslang_input_t input = {
.language = GLSLANG_SOURCE_GLSL,
.stage = stage,
.client = GLSLANG_CLIENT_VULKAN,
.client_version = GLSLANG_TARGET_VULKAN_1_2,
.target_language = GLSLANG_TARGET_SPV,
.target_language_version = GLSLANG_TARGET_SPV_1_5,
.code = shaderSource,
.default_version = 100,
.default_profile = GLSLANG_NO_PROFILE,
.force_default_version_and_profile = false,
.forward_compatible = false,
.messages = GLSLANG_MSG_DEFAULT_BIT,
.resource = glslang_default_resource(),
};
glslang_shader_t* shader = glslang_shader_create(&input);
SpirVBinary bin = {
.words = NULL,
.size = 0,
};
if (!glslang_shader_preprocess(shader, &input)) {
printf("GLSL preprocessing failed %s\n", fileName);
printf("%s\n", glslang_shader_get_info_log(shader));
printf("%s\n", glslang_shader_get_info_debug_log(shader));
printf("%s\n", input.code);
glslang_shader_delete(shader);
return bin;
}
if (!glslang_shader_parse(shader, &input)) {
printf("GLSL parsing failed %s\n", fileName);
printf("%s\n", glslang_shader_get_info_log(shader));
printf("%s\n", glslang_shader_get_info_debug_log(shader));
printf("%s\n", glslang_shader_get_preprocessed_code(shader));
glslang_shader_delete(shader);
return bin;
}
glslang_program_t* program = glslang_program_create();
glslang_program_add_shader(program, shader);
if (!glslang_program_link(program, GLSLANG_MSG_SPV_RULES_BIT | GLSLANG_MSG_VULKAN_RULES_BIT)) {
printf("GLSL linking failed %s\n", fileName);
printf("%s\n", glslang_program_get_info_log(program));
printf("%s\n", glslang_program_get_info_debug_log(program));
glslang_program_delete(program);
glslang_shader_delete(shader);
return bin;
}
glslang_program_SPIRV_generate(program, stage);
bin.size = glslang_program_SPIRV_get_size(program);
bin.words = malloc(bin.size * sizeof(uint32_t));
glslang_program_SPIRV_get(program, bin.words);
const char* spirv_messages = glslang_program_SPIRV_get_messages(program);
if (spirv_messages)
printf("(%s) %s\b", fileName, spirv_messages);
glslang_program_delete(program);
glslang_shader_delete(shader);
return bin;
}
-
Initial lexical analysis is done by the preprocessor in
MachineIndependent/Preprocessor
, and then refined by a GLSL scanner inMachineIndependent/Scan.cpp
. There is currently no use of flex. -
Code is parsed using bison on
MachineIndependent/glslang.y
with the aid of a symbol table and an AST. The symbol table is not passed on to the back-end; the intermediate representation stands on its own. The tree is built by the grammar productions, many of which are offloaded intoParseHelper.cpp
, and byIntermediate.cpp
. -
The intermediate representation is very high-level, and represented as an in-memory tree. This serves to lose no information from the original program, and to have efficient transfer of the result from parsing to the back-end. In the AST, constants are propagated and folded, and a very small amount of dead code is eliminated.
To aid linking and reflection, the last top-level branch in the AST lists all global symbols.
-
The primary algorithm of the back-end compiler is to traverse the tree (high-level intermediate representation), and create an internal object code representation. There is an example of how to do this in
MachineIndependent/intermOut.cpp
. -
Reduction of the tree to a linear byte-code style low-level intermediate representation is likely a good way to generate fully optimized code.
-
There is currently some dead old-style linker-type code still lying around.
-
Memory pool: parsing uses types derived from C++
std
types, using a custom allocator that puts them in a memory pool. This makes allocation of individual container/contents just few cycles and deallocation free. This pool is popped after the AST is made and processed.The use is simple: if you are going to call
new
, there are three cases:-
the object comes from the pool (its base class has the macro
POOL_ALLOCATOR_NEW_DELETE
in it) and you do not have to calldelete
-
it is a
TString
, in which case callNewPoolTString()
, which gets it from the pool, and there is no correspondingdelete
-
the object does not come from the pool, and you have to do normal C++ memory management of what you
new
-
-
Features can be protected by version/extension/stage/profile: See the comment in
glslang/MachineIndependent/Versions.cpp
.