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.
- Use much simpler method to update implicit array sizes.
The previous overly complicated method was error prone.
- Rationalize all use of unsized arrays.
- Combine decorations when generating SPIR-V, to simplify
adding extensions.
There a couple functional problems, which when reduced down also led to
some good simplifications and rationalization. So, this commit:
- corrects "mixed" functionality: int[A] f[B] -> f[B][A]
- correct multi-identifier decls: int[A] f[B], g[C] -> f and g are independently sized.
- increases symmetry between different places in the code that do this
- makes fewer ways to do the same thing; several methods are just gone now
- makes more clear when something is copied or shared
glslang/SPIRV/SpvBuilder.cpp:2533:27: error: comparison of integers of different signs: 'int' and 'size_type' (aka 'unsigned long') [-Werror,-Wsign-compare]
for (int c = 0; c < accessChain.swizzle.size(); ++c)
~ ^ ~~~~~~~~~~~~~~~~~~~~~~~~~~
1 error generated.
glslang/hlsl/hlslParseHelper.cpp:69:5: error: field 'cullDistanceInput' will be initialized after field 'clipDistanceOutput' [-Werror,-Wreorder]
cullDistanceInput(nullptr),
^
1 error generated.
glslang/glslang/MachineIndependent/attribute.cpp:85:16: error: comparison of integers of different signs: 'int' and 'size_type' (aka 'unsigned long') [-Werror,-Wsign-compare]
if (argNum >= args->getSequence().size())
~~~~~~ ^ ~~~~~~~~~~~~~~~~~~~~~~~~~~
1 error generated.
When a matrix is divided by a scalar it tries to take the reciprocal
of the scalar to convert the operation into a multiply. However it was
always doing this by making a 32-bit constant. If the scalar is a
double then this would end up making an FDiv instruction with
different types in the operands.
This patch adds a helper method called makeFpConstant which makes a
floating-point constant of the given type. The code to take the
reciprocal now uses it to make the same type as the result.
Fixes https://github.com/KhronosGroup/glslang/issues/1278
This factored computeTypeLocationSize() out of needing the TIntermediate contents,
and uses it to show how to know how many locations an object needs.
However, it still does not do cross stage, or mixed location/no-location
analysis.
HLSL allows image and texture types to be templatized on sub-vec4 types,
or even structures. This was mostly handled already during creation of
sampling operations. However, for operator[] which can generate image
loads, this wasn't happening.
It also isn't very easy to do at that point in time, because operator[]
does not know where the results it produces will end up. They may be
an lvalue or an rvalue, and there's a post-process to convert loads to
stores. They may end up in atomic ops.
To bypass that difficulty, GlslangToSpv now looks for this case and
adds the appropriate conversion. LIMITATION: this only works for
cases for which a simple conversion opcode suffices. That is to say,
it will not work if the type is templatized on a struct.
- make it sharable with GLSL
- correct the case insensitivity
- remove the map; queries are not needed, all entries need processing
- make it easier to build bottom up (will help GLSL parsing)
- support semantic checking and reporting
- allow front-end dependent semantics and attribute name mapping
TGlslangToSpvTraverser::getExtBuiltins is only used when AMD_EXTENSIONS
is defined, so only define it in that case to avoid an unused function
warning.
Also, only emit this XFB information where the SPIR-V spec says
it should be emitted: essentially, on objects.
This and the previous commit together fix#1185.
This continues to prevent writing output buffers (out from a function),
but fixes the problem where the copy-in/out was not getting done.
Making everything work will require knowing both in/out-ness and bufferness,
but these are currently mutually exclusive, because both are storage
qualifiers.
Also, remove assumption that if something is opaque that it
must be in the UniformConstant storage class.
This allows function declarations to know all parameters will
be in the Function storage class.
Also added known-good mechanism to fetch latest validated spirv-tools.
Also added -Od and -Os to disable optimizer and optimize for size.
Fetching spirv-tools is optional for both glsl and hlsl. Legalization
of hlsl is done by default if spirv-opt is present at cmake time.
Optimization for glsl is currently done through the option -Os.
Legalization testing is currently only done on four existing shaders.
A separate baseLegalResults directory holds those results. All previous
testing is done with the optimizer disabled.
Remapper errors are generally fatal: there has been some unexpected situation while
parsing the SPV binary, and there is no reasonable way to carry on. The
errorHandler() function is called in this case, which by default exits, but
it is possible to submit a handler which does not. In that case the remapper would
carry on in a bad state.
This change ensures a graceful termination of the remap() function.
While a try {} catch {} construct would be the ideal and safe way to do this,
that's off limits for certain environments, so this tries to do the same thing
with explicit code, to catch all the bailout paths.
Hull shaders have an implicitly arrayed output. This is handled by creating an arrayed form of the
provided output type, and writing to the element of it indexed by InvocationID.
The implicit indirection into that array was causing some troubles when copying to a split
structure. handleAssign was able to handle simple symbol lvalues, but not an lvalue composed
of an indirection into an array.
There was some ambiguity/contradiction in this behavior, and
Khronos decided glslang should always have these outside blocks,
rather than have stage/vendor/target variations.
There were several locations in TGlslangToSpvTraverser::handleUserFunctionCall testing for
whether a fn argument should be in the lvalue or rvalue array. They must get the same
result for indexing sanity, but had slightly different logic.
They're now forced into the same test.
