By directly creating the OpImage instructions we were not propagating
the appropriate decorations.
Since this had a lot of cases I centralized the OpImage creation
a bit too.
The direct pointer argument to stores has to use
the NonUniform decoration but we were not using
qualifiers at all to decorate the NonUniform pointer.
(Test fixes by Greg Fischer <greg@lunarg.com>)
We need separate concepts for
- total set of extensions ever enabled, for the back end
- current state of extensions, for parsing
- the set of features currently enabled for building the AST
This change strips a few features similar to GLSLANG_WEB but doesn't
remove every detail like the latter. It also hardcodes profile/version
to core/450.
In particular, TBuiltIns::initialize is specialized to remove most of
what is not supported or won't be supported by ANGLE. The result of
this function is parsed with TParseContext::parseShaderStrings which is
a performance bottleneck.
This change shaves about 300KB off of ANGLE's binary size and reduces
the cost of SetupBuiltinSymbolTable to nearly a sixth.
Signed-off-by: Shahbaz Youssefi <ShabbyX@gmail.com>
iOS 8 does not support `thread_local`, which is still in use.
Another approach will have to be found.
This change is a revert of the following changes:
a3845240 - "Simplify PoolAlloc with use of thread_local."
abf92c80 - "Deprecate InitializeDll functions"
33585c87 - "Limit visibility of symbols for internal libraries"
Issue: #2346
Also remove `SPIRV/doc.cpp` from the `SPVRemapper` target as this
is part of `SPIRV`, causing ODR violations. Instead have
`SPVRemapper` link against `SPIRV`.
Fixes ODR violations.
This PR significantly reworks the way glslang is versioned.
Instead of committing changes to the `GLSLANG_MINOR_VERSION` define in
`glslang/Public/ShaderLang.h`, and using `make-revision` to generate
`GLSLANG_PATCH_LEVEL` in `glslang/Include/revision.h`, all version
information is now derived from the new `CHANGES.md` file.
`CHANGES.md` acts as the single source of truth for glslang version
information, along with a convenient place to put all release notes for
each notable change made.
`CHANGES.md` is parsed using the new `build_info.py` python script.
This script can read basic template files to produce new source files,
which it does to read the new `build_info.h.tmpl` to generate (at build
time) a glslang private header at
`<build-dir>/include/glslang/build_info.h`.
I've written generators for each of the CMake, Bazel, gn, and
`Android.mk` build scripts.
The new version code conforms to the Semantic Versioning 2.0 spec.
This new version is also used by the CMake rules to produce versioned
shared objects, including a major-versioned SONAME.
New APIs:
---------
* `glslang::GetVersion()` returns a `Version` struct with the version
major, minor, patch and flavor.
Breaking API changes:
---------------------
* The public defines `GLSLANG_MINOR_VERSION` and `GLSLANG_PATCH_LEVEL`
have been entirely removed.
* `glslang/Public/ShaderLang.h` and `glslang/Include/revision.h` have
been deleted.
* Instead, `<build-dir>/include/glslang/build_info.h` is created in
the build directory, and `<build-dir>/include` is a CMake `PUBLIC`
(dependee-inherited) include directory for the glslang targets.
* `<build-dir>/include/glslang/build_info.h` contains the following
new #defines:
`GLSLANG_VERSION_MAJOR`, `GLSLANG_VERSION_MINOR`,
`GLSLANG_VERSION_PATCH`, `GLSLANG_VERSION_FLAVOR`,
`GLSLANG_VERSION_GREATER_THAN(major, minor, patch)`,
`GLSLANG_VERSION_GREATER_OR_EQUAL_TO(major, minor, patch)`,
`GLSLANG_VERSION_LESS_THAN(major, minor, patch)`,
`GLSLANG_VERSION_LESS_OR_EQUAL_TO(major, minor, patch)`
* The CMake install output directory contains a copy of
`build_info.h` at: `include/glslang/build_info.h`
* Python3 is now always required to build glslang (likely always
required for transitive dependency builds).
This change strips a few features similar to GLSLANG_WEB but doesn't
remove every detail like the latter. It also hardcodes profile/version
to core/450.
In particular, TBuiltIns::initialize is specialized to remove most of
what is not supported or won't be supported by ANGLE. The result of
this function is parsed with TParseContext::parseShaderStrings which is
a performance bottleneck.
This change shaves about 300KB off of ANGLE's binary size and reduces
the cost of SetupBuiltinSymbolTable to nearly a sixth.
Signed-off-by: Shahbaz Youssefi <ShabbyX@gmail.com>
Fix#2298.
The AST has two precisions, an operation precision and a result precision.
Actual use of GLSL with mediump samplers wants the result precision, so
pick that up instead of the operation precision.
This simplifies and enforces use of precision in many more places,
to help avoid accidental loss of RelaxedPrecision through intermediate
operations. Known fixes are:
- ?:
- function return values with mis-matched precision
- precision of function return values when a copy was needed to fix types
When arguments are copied to make space for a writable formal parameter,
and the formal parameter is relaxed precision, make the copy also
relaxed precision.
The normal IEEE not equal operation tests whether operands are unordered
or not equal (so comparison with a NaN returns true). This corresponds
to the SPIR-V OpFUnordNotEqual, so change to using that.
Analogous to gcc, -g0 would strip all debug info. This is done
regardless of whether optimizations are enabled.
Signed-off-by: Shahbaz Youssefi <ShabbyX@gmail.com>
If a uniform has an initializer it will now be given as the optional
initializer operand to the OpVariable instruction.
Fixes: https://github.com/KhronosGroup/glslang/issues/1259
Signed-off-by: Neil Roberts <nroberts@igalia.com> (the code)
Signed-off-by: Alejandro Piñeiro <apinheiro@igalia.com> (the tests)
Signed-off-by: Arcady Goldmints-Orlov <agoldmints@igalia.com>
Co-authored-by: Neil Roberts <nroberts@igalia.com>
* HLSL: Fix handling of uniforms in entry point parameters
* HLSL: fix handling of "uniform in"
* Tests: Update baseResults of hlsl.function.frag.out for #2254
* HLSL: fix uniforms in function parameters for opaque types
This CL updates the build scripts to only install to glslang/SPIRV
instead of also installing to the SPIRV/ folder. The deprecation notice
is also removed.
Note, this may cause downstream build issues if include directories have
not been updated
Fixes#1964#2216
Swizzled out operands were added in bbbd9a2a. This was sufficient
for most tests, but we ran into problems with umulExtended and
imulExtended, which have two.
This CL converts the tracking values to vectors so multiple operands
can be supported.
Test: KHR-GLES31.core.shader_bitfield_operation.*
Test: ctest
For GLSL function imageAtomicStore, it will be translated to
OpImageTexelPointer + OpAtomicStore. The result type of
OpImageTexelPointer must be the same as the sampled type of OpTypeImage.
On translation, the result type is mistakenly fetched from operand list
operands[2] while operands[2] corresponds to sampleNum whose type is
always uint. This leads to an error if the image type is iimageXXX that is
int image.
Root cause:
GlslangToSpv use symbol structure's ptr as a map key, but multiple shader object can build a new AST.
In the AST the the same symbol has different ptr point to their structure, so indext map faild.
solution:
Add a new map glslangTypeToIdMap to map ptr to symbol id, and use symbol id to index memberRemapper.
v2: Move addCapability and addExtension extension calls from
TGlslangToSpvTraverser::createBinaryOperation to
TGlslangToSpvTraverser::visitAggregate. Suggested by JohnK.
There is a flag to disable the SPVRemapper during the GLSLang build.
That flag is check in some, but not all spots so if you try to build
with SPVRemapper disabled you get CMake errors and compile errors.
This CL fixs up the build so building with -DENABLE_SPVREMAPPER=0 will
complete correclty.
The SPIR-V post-processing to discover capabilities and
extensions does not apply to WebGPU compilation. So don't include
that code.
This reclaims some of the code space added by #1943
More aggressively prune unreachable code as follows.
When no control flow edges reach a merge block or continue target:
- delete their contents so that:
- a merge block becomes OpLabel, then OpUnreachable
- a continue target becomes OpLabel, then an OpBranch back to the
loop header
- any basic block which is dominated by such a merge block or continue
target is removed as well.
- decorations targeting the removed instructions are removed.
Enables the SPIR-V builder post-processing step the GLSLANG_WEB case.
If the semantics that require the KHR form over the EXT form are seen
(OpBitcast between a vector and a pointer) promote the requested extension
from the EXT to the KHR.
Currently the SPIRV/ folder will get installed into the include
directory. This folder is part of GLSLang, so it makes more sense under
glslang/SPIRV.
Currently, GLSLang will install a SPIRV/ folder while spirv-headers will
install a spirv/ folder. This is confusing and will cause issues on a
case sensitive filesystem if both are installed at the same time.
Saved about 21K, size down to 380K of MSVC x86 code.
Fixed one bug that needs to be looked at on the master branch:
The test for needing a Vulkan binding has a bug in it, "!layoutAttachment"
which does not mean "no layoutAttachment", because that is non-zero.
This is why some test and test results changed.
About 50 fewer #ifdefs.
About 14K smaller.
Note, the base size is ill-defined due to optimizer settings (size vs. performance),
compression, and target architecture. Some recent %'s are accidentally reported as
3X the real savings. Early %'s were accurate. What matters though is that each
step got worthwhile gains, and what the final size ends up being.
Focus was on the front end (not SPIR-V), minus the grammar.
Reduces #ifdef count by around 320 and makes the web build 270K smaller,
which is about 90% the target size.
The grammar and scanner will be another step, as will the SPIR-V backend.
This makes heavy use of methods #ifdef'd to return false as a global way
of turning off code, relying on C++ DCE to do the rest.
Save about 100K.
N.B.: This is done by eliminating a function call, at a high level,
not by #ifdef'ing a bunch of code.
Also, removed no longer needed *_EXTENSION #ifdef in the code not
needed by GLSLANG_WEB.
On reading built-in variables SubgroupEqMask, SubgroupGeMask, SubgroupGtMask,
SubgroupLeMask, and SubgroupLtMask, the AST expects 64-bit ints, while SPIR-V
is defined as vectors of 32-bit ints.
The declaration type has to be translated in the opposite direction.
Including spirv and AST tests
Also increase size of TBuiltInVariable bitfields since we've now exceeded 127
and add a static_assert to make this easier to find next time it happens!
Closes#1735
GlslangToSpv.cpp
- minor formatting cleanup
BaseTypes.h
- minor formatting cleanup
- add subgroup builtins to GetBuiltInVariableString
(was resulting in "unknown built-in variable" messages in test output)
Initialize.cpp
- better naming and re-use of strings for subgroup builtin variable declarations
- define subgroup builtin variables in ray-tracing shaders
intermOut.cpp
- add handling of the EOpSubgroupParition* variables
(was resulting in "ERROR: Bad aggregation op" messages in test output)
Update test results.
One variable was only used in an 'assert' call. MSVC flagged this
as unused in Release. Suppress the warning and also add a static
cast to void so the variable becomes referenced.
The spvtools::Optimizer::Run method glslang is using constructs a default set
of spvtools::OptimizerOptions. This default set of options instructs the
validator to run. That is not quite correct since glslang will invoke the
validator _explicitly_ after the optimization pass.
Change-Id: I30f458304c6e7f81e89fc4ebd25eabbbd8348063
This is a better place for it logically, since it is not specific to
glsl->spirv translation. And in a future change I want to use it outside
of glslangtospv.
This is an alternate fix for the issue described in commit be63facd, whose
solution didn't work if there were non-trivial operations involved in computing
a constant initializer which caused the 'constant unfolding' code to kick in
(addConstantReferenceConversion). Instead, this change does the 'unfolding'
later in createSpvConstantFromConstUnionArray. If a reference-type constant has
survived that long, then folding is already done, this must be a 'real' (inside
a function) use of the constant, and it should be safe to unfold and apply the
bitcast.
Last year we changed 'volatile' to also act as 'coherent', but when I
resolved the memory model changes against that change I missed handling
volatile in a couple places that we check for coherent. There was also
a place in post-processing that acted as if the volatile memory access
flag has a literal number associated with it, when it doesn't.
Without this commit, if the XfbStride was explicitly set, the
decoration was added twice on the shader.
v2 (changes after Jonh Kessenich first review)
* Simplified by just removing the firs assignment
* Removed assert
including SPV generation using SPV_EXT_fragment_invocation_density.
This is an alias of the functionality in SPV_NV_shading_rate, and thus in some
cases we can only have one set of the tokens present (switch statements), so
we have picked the EXT version. This required updating the expected test
results for SPV_NV_shading_rate.
Also updated the known-good for spirv-headers so that the validator in
spirv-tools knows about the new extension.
Consider the following code:
layout(constant_id=0) const int Y = 1;
layout(constant_id=1) const int Z = 2;
layout(constant_id=3) const int X = Y + Z;
Previously, it would produce SPIR-V decorations like this:
Decorate 21(Y) SpecId 1
Decorate 22 SpecId 3
Decorate 33(Z) SpecId 0
This seems inaccurate, since the spec constant `X` that is dependent on
the two others did not get a name in the SPIR-V decorations. This behavior
may potentially negatively affect shader introspection capabilities.
This change alters the behavior to always add a name, which results in the code
above producing the following decorations:
Decorate 21(Y) SpecId 1
Decorate 22(X) SpecId 3
Decorate 33(Z) SpecId 0
UniformAndStorageBuffer8BitAccess capability.
When using the 8-bit storage extension it basically always used the
`UniformAndStorageBuffer8BitAccess` capability, even in cases where it
wasn't required. For instance if we are targeting Vulkan 1.1 (SPIR-V 1.3
or higher), and we are only using 8-bit types in an SSBO, we only need
the `StorageBuffer8BitAccess` capability.
I fixed this by enabling storage buffer use in Vulkan 1.1 / SPIR-V 1.3
or higher, and then changing the logic to match.
I also added some tests that will output different capabilities when run
on Vulkan 1.0 and 1.1, thus they are added twice to the test list (one
for each version).
Fixes#1539
- Emit relevant capability/extension for use of perprimitiveNV in fragment shader
- Remove redundant checks for mesh shader qualifiers in glslang.y
- Add profile version check for use of extension GL_NV_mesh_shader
- Add a new gtest for use of perprimitiveNV in fragment shader
This CL adds the necessary configuration to build glslang inside a
Chromium checkout. Two build warnings were fixed in the process to
make things compile.
the following warning gets emitted:
```
In file included from ./Vulkan/glslang/SPIRV/hex_float.h:39,
from ./Vulkan/glslang/SPIRV/SpvBuilder.cpp:49:
./Vulkan/glslang/SPIRV/bitutils.h: In instantiation of ‘Dest spvutils::BitwiseCast(Src) [with Dest = spvutils::Float16; Src = short unsigned int]’:
./Vulkan/glslang/SPIRV/hex_float.h:138:47: required from ‘T spvutils::FloatProxy<T>::getAsFloat() const [with T = spvutils::Float16]’
./Vulkan/glslang/SPIRV/hex_float.h:821:52: required from here
./Vulkan/glslang/SPIRV/bitutils.h:29:14: warning: ‘void* memcpy(void*, const void*, size_t)’ writing to an object of non-trivially copyable type ‘class spvutils::Float16’; use copy-assignment or copy-initialization instead [-Wclass-memaccess]
std::memcpy(&dest, &source, sizeof(dest));
~~~~~~~~~~~^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
In file included from ./Vulkan/glslang/SPIRV/SpvBuilder.cpp:49:
./Vulkan/glslang/SPIRV/hex_float.h:43:7: note: ‘class spvutils::Float16’ declared here
class Float16 {
^~~~~~~
In file included from ./Vulkan/glslang/SPIRV/hex_float.h:39,
from ./Vulkan/glslang/SPIRV/SpvBuilder.cpp:49:
./Vulkan/glslang/SPIRV/bitutils.h: In instantiation of ‘Dest spvutils::BitwiseCast(Src) [with Dest = spvutils::FloatProxy<spvutils::Float16>; Src = short unsigned int]’:
./Vulkan/glslang/SPIRV/hex_float.h:431:28: required from ‘void spvutils::HexFloat<T, Traits>::setFromSignUnbiasedExponentAndNormalizedSignificand(bool, spvutils::HexFloat<T, Traits>::int_type, spvutils::HexFloat<T, Traits>::uint_type, bool) [with T = spvutils::FloatProxy<spvutils::Float16>; Traits = spvutils::HexFloatTraits<spvutils::FloatProxy<spvutils::Float16> >; spvutils::HexFloat<T, Traits>::int_type = short int; spvutils::HexFloat<T, Traits>::uint_type = short unsigned int]’
./Vulkan/glslang/SPIRV/hex_float.h:633:5: required from ‘void spvutils::HexFloat<T, Traits>::castTo(other_T&, spvutils::round_direction) [with other_T = spvutils::HexFloat<spvutils::FloatProxy<spvutils::Float16>, spvutils::HexFloatTraits<spvutils::FloatProxy<spvutils::Float16> > >; T = spvutils::FloatProxy<float>; Traits = spvutils::HexFloatTraits<spvutils::FloatProxy<float> >]’
./Vulkan/glslang/SPIRV/hex_float.h:817:39: required from here
./Vulkan/glslang/SPIRV/bitutils.h:29:14: warning: ‘void* memcpy(void*, const void*, size_t)’ copying an object of non-trivial type ‘class spvutils::FloatProxy<spvutils::Float16>’ from an array of ‘short unsigned int’ [-Wclass-memaccess]
std::memcpy(&dest, &source, sizeof(dest));
~~~~~~~~~~~^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
In file included from ./Vulkan/glslang/SPIRV/SpvBuilder.cpp:49:
./Vulkan/glslang/SPIRV/hex_float.h:115:7: note: ‘class spvutils::FloatProxy<spvutils::Float16>’ declared here
class FloatProxy {
^~~~~~~~~~
```
When capabilities are needed for specific SPIR-V instructions, it is
fragile to do so based on GLSL/AST usage; it should be based on actual
instructions they got translated to.
When constructing a matrix from another matrix with smaller dimensions,
there's no need to extract the scalars out of columns and rebuild the
resulting matrix from scalars - instead, we can just construct shorter
vectors with OpShuffle and combine them to the final result.
This keeps the common casts such as mat3(mat4) in vector registers,
which may improve performance for some GPUs, and cleans up output of
translation tools like SPIRV-Cross.
Fixes#1412.
These introduce limited support for 8/16-bit types such that they can only be accessed in buffer memory and converted to/from 32-bit types.
Contributed from Khronos-internal work.
