Files
renderdoc/renderdoc/driver/shaders/spirv/spirv_reflect.cpp
T

2015 lines
68 KiB
C++

/******************************************************************************
* The MIT License (MIT)
*
* Copyright (c) 2019-2022 Baldur Karlsson
*
* Permission is hereby granted, free of charge, to any person obtaining a copy
* of this software and associated documentation files (the "Software"), to deal
* in the Software without restriction, including without limitation the rights
* to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
* copies of the Software, and to permit persons to whom the Software is
* furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice shall be included in
* all copies or substantial portions of the Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
* AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
* THE SOFTWARE.
******************************************************************************/
#include "spirv_reflect.h"
#include <limits.h>
#include <algorithm>
#include "common/formatting.h"
#include "replay/replay_driver.h"
#include "spirv_editor.h"
#include "spirv_op_helpers.h"
void FillSpecConstantVariables(ResourceId shader, const SPIRVPatchData &patchData,
const rdcarray<ShaderConstant> &invars,
rdcarray<ShaderVariable> &outvars,
const rdcarray<SpecConstant> &specInfo)
{
StandardFillCBufferVariables(shader, invars, outvars, bytebuf());
RDCASSERTEQUAL(invars.size(), outvars.size());
for(size_t v = 0; v < invars.size() && v < outvars.size(); v++)
outvars[v].value.u64v[0] = invars[v].defaultValue;
// find any actual values specified
for(size_t i = 0; i < specInfo.size(); i++)
{
for(size_t v = 0; v < invars.size() && v < outvars.size(); v++)
{
int32_t idx = patchData.specIDs.indexOf(specInfo[i].specID);
if(idx == -1)
continue;
if(idx * sizeof(uint64_t) == invars[v].byteOffset)
{
outvars[v].value.u64v[0] = specInfo[i].value;
}
}
}
}
void AddXFBAnnotations(const ShaderReflection &refl, const SPIRVPatchData &patchData,
const char *entryName, rdcarray<uint32_t> &modSpirv, uint32_t &xfbStride)
{
rdcspv::Editor editor(modSpirv);
editor.Prepare();
rdcarray<SigParameter> outsig = refl.outputSignature;
rdcarray<SPIRVInterfaceAccess> outpatch = patchData.outputs;
rdcspv::Id entryid;
for(const rdcspv::EntryPoint &entry : editor.GetEntries())
{
if(entry.name == entryName && MakeShaderStage(entry.executionModel) == refl.stage)
{
entryid = entry.id;
break;
}
}
bool hasXFB = false;
for(rdcspv::Iter it = editor.Begin(rdcspv::Section::ExecutionMode);
it < editor.End(rdcspv::Section::ExecutionMode); ++it)
{
rdcspv::OpExecutionMode execMode(it);
if(execMode.entryPoint == entryid && execMode.mode == rdcspv::ExecutionMode::Xfb)
{
hasXFB = true;
break;
}
}
if(hasXFB)
{
for(rdcspv::Iter it = editor.Begin(rdcspv::Section::Annotations);
it < editor.End(rdcspv::Section::Annotations); ++it)
{
// remove any existing xfb decorations
if(it.opcode() == rdcspv::Op::Decorate)
{
rdcspv::OpDecorate decorate(it);
if(decorate.decoration == rdcspv::Decoration::XfbBuffer ||
decorate.decoration == rdcspv::Decoration::XfbStride)
{
editor.Remove(it);
}
}
// offset is trickier, need to see if it'll match one we want later
if((it.opcode() == rdcspv::Op::Decorate &&
rdcspv::OpDecorate(it).decoration == rdcspv::Decoration::Offset) ||
(it.opcode() == rdcspv::Op::MemberDecorate &&
rdcspv::OpMemberDecorate(it).decoration == rdcspv::Decoration::Offset))
{
for(size_t i = 0; i < outsig.size(); i++)
{
if(outpatch[i].structID && it.opcode() == rdcspv::Op::MemberDecorate)
{
rdcspv::OpMemberDecorate decoded(it);
if(decoded.structureType == outpatch[i].structID &&
decoded.member == outpatch[i].structMemberIndex)
{
editor.Remove(it);
}
}
else if(!outpatch[i].structID && it.opcode() == rdcspv::Op::Decorate)
{
rdcspv::OpDecorate decoded(it);
if(decoded.target == outpatch[i].ID)
{
editor.Remove(it);
}
}
}
}
}
}
else
{
editor.AddExecutionMode(rdcspv::OpExecutionMode(entryid, rdcspv::ExecutionMode::Xfb));
}
editor.AddCapability(rdcspv::Capability::TransformFeedback);
// find the position output and move it to the front
for(size_t i = 0; i < outsig.size(); i++)
{
if(outsig[i].systemValue == ShaderBuiltin::Position)
{
outsig.insert(0, outsig[i]);
outsig.erase(i + 1);
outpatch.insert(0, outpatch[i]);
outpatch.erase(i + 1);
break;
}
}
for(size_t i = 0; i < outsig.size(); i++)
{
if(outpatch[i].isArraySubsequentElement)
{
// do not patch anything as we only patch the base array, but reserve space in the stride
}
else if(outpatch[i].structID && !outpatch[i].accessChain.empty())
{
editor.AddDecoration(
rdcspv::OpMemberDecorate(outpatch[i].structID, outpatch[i].structMemberIndex,
rdcspv::DecorationParam<rdcspv::Decoration::Offset>(xfbStride)));
}
else if(outpatch[i].ID)
{
editor.AddDecoration(rdcspv::OpDecorate(
outpatch[i].ID, rdcspv::DecorationParam<rdcspv::Decoration::Offset>(xfbStride)));
}
// components always get promoted to at least 32-bit
uint32_t compByteSize = RDCMAX(4U, VarTypeByteSize(outsig[i].varType));
xfbStride += outsig[i].compCount * compByteSize;
}
std::set<rdcspv::Id> vars;
for(size_t i = 0; i < outpatch.size(); i++)
{
if(outpatch[i].ID && !outpatch[i].isArraySubsequentElement &&
vars.find(outpatch[i].ID) == vars.end())
{
editor.AddDecoration(rdcspv::OpDecorate(
outpatch[i].ID, rdcspv::DecorationParam<rdcspv::Decoration::XfbBuffer>(0)));
editor.AddDecoration(rdcspv::OpDecorate(
outpatch[i].ID, rdcspv::DecorationParam<rdcspv::Decoration::XfbStride>(xfbStride)));
vars.insert(outpatch[i].ID);
}
}
}
const int32_t INVALID_BIND = -INT_MAX;
template <typename T>
struct bindpair
{
Bindpoint map;
T bindres;
bindpair() = default;
bindpair(const Bindpoint &m, const T &res) : map(m), bindres(res) {}
bool operator<(const bindpair &o) const
{
if(map.bindset != o.map.bindset)
return map.bindset < o.map.bindset;
// sort invalid/not set binds to the end
if(map.bind == INVALID_BIND && o.map.bind == INVALID_BIND) // equal
return false;
if(map.bind == INVALID_BIND) // invalid bind not less than anything
return false;
if(o.map.bind == INVALID_BIND) // anything is less than invalid bind
return true;
return map.bind < o.map.bind;
}
};
typedef bindpair<ConstantBlock> cblockpair;
typedef bindpair<ShaderResource> shaderrespair;
static uint32_t GetDescSet(uint32_t set)
{
return set == ~0U ? 0 : set;
}
static int32_t GetBinding(uint32_t binding)
{
return binding == ~0U ? INVALID_BIND : (uint32_t)binding;
}
static bool IsStrippableBuiltin(rdcspv::BuiltIn builtin)
{
return builtin == rdcspv::BuiltIn::PointSize || builtin == rdcspv::BuiltIn::ClipDistance ||
builtin == rdcspv::BuiltIn::CullDistance;
}
static uint32_t CalculateMinimumByteSize(const rdcarray<ShaderConstant> &variables)
{
if(variables.empty())
{
RDCERR("Unexpectedly empty array of shader constants!");
return 0;
}
const ShaderConstant &last = variables.back();
// find its offset
uint32_t byteOffset = last.byteOffset;
// arrays are easy
if(last.type.arrayByteStride > 0)
return byteOffset + last.type.arrayByteStride * last.type.elements;
if(last.type.members.empty())
{
// this is the last basic member
// now calculate its size and return offset + size
RDCASSERT(last.type.elements <= 1);
uint32_t basicTypeSize = 4;
if(last.type.baseType == VarType::Double)
basicTypeSize = 8;
uint32_t rows = last.type.rows;
uint32_t cols = last.type.columns;
// vectors are also easy
if(rows == 1)
return byteOffset + cols * basicTypeSize;
if(cols == 1)
return byteOffset + rows * basicTypeSize;
// for matrices we need to pad 3-column or 3-row up to 4
if(cols == 3 && last.type.RowMajor())
{
return byteOffset + rows * 4 * basicTypeSize;
}
else if(rows == 3 && last.type.ColMajor())
{
return byteOffset + cols * 4 * basicTypeSize;
}
else
{
// otherwise, it's a simple size
return byteOffset + rows * cols * basicTypeSize;
}
}
else
{
// if this is a struct type, recurse
return byteOffset + CalculateMinimumByteSize(last.type.members);
}
}
// Some generators output command-line arguments as OpModuleProcessed
static bool HasCommandLineInModuleProcessed(rdcspv::Generator gen)
{
return (gen == rdcspv::Generator::GlslangReferenceFrontEnd ||
gen == rdcspv::Generator::ShadercoverGlslang);
}
struct StructSizes
{
uint32_t scalarAlign = 1;
uint32_t baseAlign = 1;
uint32_t extendedAlign = 1;
uint32_t scalarSize = 0;
uint32_t baseSize = 0;
uint32_t extendedSize = 0;
};
StructSizes CalculateStructProps(uint32_t emptyStructSize, const ShaderConstant &c)
{
StructSizes ret;
if(c.type.baseType != VarType::Struct)
{
// A scalar of size N has a scalar alignment of N.
// A vector or matrix type has a scalar alignment equal to that of its component type.
// An array type has a scalar alignment equal to that of its element type.
ret.scalarAlign = VarTypeByteSize(c.type.baseType);
// A scalar has a base alignment equal to its scalar alignment.
ret.baseAlign = ret.scalarAlign;
// A row-major matrix of C columns has a base alignment equal to the base alignment of a vector
// of C matrix components.
uint8_t vecSize = c.type.columns;
uint8_t matSize = c.type.rows;
// A column-major matrix has a base alignment equal to the base alignment of the matrix column
// type.
if(c.type.rows > 1 && c.type.ColMajor())
{
vecSize = c.type.rows;
matSize = c.type.columns;
}
// A two-component vector has a base alignment equal to twice its scalar alignment.
if(vecSize == 2)
ret.baseAlign *= 2;
// A three- or four-component vector has a base alignment equal to four times its scalar
// alignment.
else if(vecSize == 3 || vecSize == 4)
ret.baseAlign *= 4;
// An array has a base alignment equal to the base alignment of its element type.
// N/A
// A scalar, vector or matrix type has an extended alignment equal to its base alignment.
ret.extendedAlign = ret.baseAlign;
// An array or structure type has an extended alignment equal to the largest extended alignment
// of any of its members, rounded up to a multiple of 16.
if(c.type.elements > 1)
ret.extendedAlign = AlignUp16(ret.extendedAlign);
if(matSize > 1)
ret.extendedAlign = ret.baseAlign = c.type.matrixByteStride;
ret.scalarSize = ret.scalarAlign * RDCMAX(c.type.rows, (uint8_t)1) *
RDCMAX(c.type.columns, (uint8_t)1) * RDCMAX(c.type.elements, 1U);
ret.baseSize = ret.baseAlign * matSize * RDCMAX(c.type.elements, 1U);
ret.extendedSize = ret.extendedAlign * matSize * RDCMAX(c.type.elements, 1U);
}
else
{
for(size_t i = 0; i < c.type.members.size(); i++)
{
const ShaderConstant &m = c.type.members[i];
StructSizes member = CalculateStructProps(emptyStructSize, m);
// A structure has a scalar alignment equal to the largest scalar alignment of any of its
// members.
ret.scalarAlign = RDCMAX(ret.scalarAlign, member.scalarAlign);
// A structure has a base alignment equal to the largest base alignment of any of its members.
ret.baseAlign = RDCMAX(ret.baseAlign, member.baseAlign);
// An array or structure type has an extended alignment equal to the largest extended
// alignment of any of its members, rounded up to a multiple of 16.
ret.extendedAlign = RDCMAX(ret.baseAlign, member.extendedAlign);
if(i + 1 == c.type.members.size())
{
ret.scalarSize = AlignUp(m.byteOffset + member.scalarSize, ret.scalarAlign);
ret.baseSize = AlignUp(m.byteOffset + member.baseSize, ret.baseAlign);
ret.extendedSize = AlignUp16(m.byteOffset + member.extendedSize);
}
}
ret.extendedAlign = AlignUp16(ret.extendedAlign);
// A structure has a base alignment equal to the largest base alignment of any of its members.
// An empty structure has a base alignment equal to the size of the smallest scalar type
// permitted by the capabilities declared in the SPIR-V module. (e.g., for a 1 byte aligned
// empty struct in the StorageBuffer storage class, StorageBuffer8BitAccess or
// UniformAndStorageBuffer8BitAccess must be declared in the SPIR-V module.)
if(c.type.members.empty())
{
ret.scalarSize = 0;
ret.scalarAlign = emptyStructSize;
ret.baseSize = emptyStructSize;
ret.baseAlign = emptyStructSize;
ret.extendedSize = AlignUp16(emptyStructSize);
ret.extendedAlign = AlignUp16(emptyStructSize);
}
}
return ret;
}
namespace rdcspv
{
Reflector::Reflector()
{
}
void Reflector::Parse(const rdcarray<uint32_t> &spirvWords)
{
Processor::Parse(spirvWords);
}
void Reflector::PreParse(uint32_t maxId)
{
Processor::PreParse(maxId);
strings.resize(idTypes.size());
}
void Reflector::RegisterOp(Iter it)
{
Processor::RegisterOp(it);
OpDecoder opdata(it);
if(opdata.op == Op::String)
{
OpString string(it);
strings[string.result] = string.string;
}
else if(opdata.op == Op::Name)
{
OpName name(it);
// technically you could name a string - in that case we ignore the name
if(strings[name.target].empty())
strings[name.target] = name.name;
}
else if(opdata.op == Op::MemberName)
{
OpMemberName memberName(it);
memberNames.push_back({memberName.type, memberName.member, memberName.name});
}
else if(opdata.op == Op::Variable)
{
OpVariable var(it);
// variables are always pointers
Id varType = dataTypes[var.resultType].InnerType();
// if we don't have a name for this variable but it's a pointer to a struct that is named then
// give the variable a name based on the type. This is a common pattern in GLSL for global
// blocks, and since the variable is how we access commonly we should give it a recognisable
// name.
if(strings[var.result].empty() && dataTypes[varType].type == DataType::StructType &&
!strings[varType].empty())
{
strings[var.result] = strings[varType] + "_var";
}
}
else if(opdata.op == Op::ModuleProcessed)
{
OpModuleProcessed processed(it);
if(HasCommandLineInModuleProcessed(m_Generator))
{
cmdline += " --" + processed.process;
}
}
else if(opdata.op == Op::Source)
{
OpSource source(it);
// glslang based tools output fake OpModuleProcessed comments at the start of pre-1.3
// shaders source before OpModuleProcessed existed (in SPIR-V 1.1)
if(m_MajorVersion == 1 && m_MinorVersion < 1 && HasCommandLineInModuleProcessed(m_Generator))
{
rdcstr &src = source.source;
const char compileFlagPrefix[] = "// OpModuleProcessed ";
const char endMarker[] = "#line 1\n";
if(src.find(compileFlagPrefix) == 0)
{
// process compile flags
int32_t nextLine = src.indexOf('\n');
while(nextLine > 0)
{
bool finished = false;
if(src.find(compileFlagPrefix) == 0)
{
size_t offs = sizeof(compileFlagPrefix) - 1;
cmdline += " --" + src.substr(offs, nextLine - offs);
}
else if(src.find(endMarker) == 0)
{
finished = true;
}
else
{
RDCERR("Unexpected preamble line with OpModuleProcessed: %s",
src.substr(0, nextLine).c_str());
break;
}
// erase this line
src.erase(0, nextLine + 1);
nextLine = src.indexOf('\n');
if(finished)
break;
}
}
}
sources.push_back({source.sourceLanguage, strings[source.file], source.source});
}
else if(opdata.op == Op::SourceContinued)
{
OpSourceContinued continued(it);
sources.back().contents += continued.continuedSource;
}
else if(opdata.op == Op::Label)
{
curBlock = opdata.result;
}
else if(opdata.op == Op::LoopMerge)
{
loopBlocks.insert(curBlock);
}
else if(opdata.op == Op::ExtInst)
{
OpShaderDbg dbg(it);
// we don't care about much debug info for just reflection. Only pay attention to source files,
// and potential names of global variables that might be missing.
if(dbg.set == knownExtSet[ExtSet_ShaderDbg])
{
if(dbg.inst == ShaderDbg::Source)
{
debugSources[dbg.result] = sources.size();
sources.push_back({
SourceLanguage::Unknown, strings[dbg.arg<Id>(0)],
dbg.params.size() > 1 ? strings[dbg.arg<Id>(1)] : rdcstr(),
});
}
else if(dbg.inst == ShaderDbg::SourceContinued)
{
sources.back().contents += strings[dbg.arg<Id>(0)];
}
else if(dbg.inst == ShaderDbg::CompilationUnit)
{
sources[debugSources[dbg.arg<Id>(2)]].lang =
(SourceLanguage)EvaluateConstant(dbg.arg<Id>(3), {}).value.u32v[0];
}
else if(dbg.inst == ShaderDbg::GlobalVariable)
{
// copy the name string to the variable string only if it's empty. If it has a name already,
// we prefer that. If the variable is DebugInfoNone then we don't care about it's name.
if(strings[dbg.arg<Id>(7)].empty())
strings[dbg.arg<Id>(7)] = strings[dbg.arg<Id>(0)];
}
}
}
}
void Reflector::UnregisterOp(Iter it)
{
RDCFATAL("Reflector should not be used for editing! UnregisterOp() call invalid");
}
void Reflector::PostParse()
{
Processor::PostParse();
// assign default names for types that we can
for(auto it = dataTypes.begin(); it != dataTypes.end(); ++it)
{
Id id = it->first;
DataType &type = it->second;
type.name = strings[id];
if(type.name.empty())
{
if(type.type == DataType::UnknownType)
{
// ignore
}
else if(type.scalar().type == Op::TypeVoid)
{
type.name = "void";
}
else if(type.scalar().type == Op::TypeBool)
{
type.name = "bool";
}
else if(type.type == DataType::StructType)
{
type.name = StringFormat::Fmt("struct%u", type.id.value());
}
else if(type.type == DataType::ArrayType)
{
// prefer the name
rdcstr lengthName;
if(type.length != Id())
{
lengthName = strings[type.length];
// if not, use the constant value
if(lengthName.empty())
lengthName = StringiseConstant(type.length);
// if not, it might be a spec constant, use the fallback
if(lengthName.empty())
lengthName = StringFormat::Fmt("_%u", type.length.value());
}
rdcstr basename = dataTypes[type.InnerType()].name;
// arrays are inside-out, so we need to insert our new array length before the first array
// length
int arrayCharIdx = basename.indexOf('[');
if(arrayCharIdx > 0)
{
type.name = StringFormat::Fmt("%s[%s]%s", basename.substr(0, arrayCharIdx).c_str(),
lengthName.c_str(), basename.substr(arrayCharIdx).c_str());
}
else
{
type.name = StringFormat::Fmt("%s[%s]", dataTypes[type.InnerType()].name.c_str(),
lengthName.c_str());
}
}
else if(type.type < DataType::StructType)
{
type.name = ToStr(type.scalar().Type());
if(type.type == DataType::VectorType)
{
type.name += StringFormat::Fmt("%u", type.vector().count);
}
else if(type.type == DataType::MatrixType)
{
type.name += StringFormat::Fmt("%ux%u", type.vector().count, type.matrix().count);
}
}
else if(type.type == DataType::ImageType)
{
const Image &img = imageTypes[type.id];
rdcstr name;
switch(img.dim)
{
case Dim::_1D: name = "1D"; break;
case Dim::_2D: name = "2D"; break;
case Dim::_3D: name = "3D"; break;
case Dim::Cube: name = "Cube"; break;
case Dim::Rect: name = "Rect"; break;
case Dim::SubpassData: name = "Subpass"; break;
case Dim::Buffer: name = "Buffer"; break;
case Dim::Invalid:
case Dim::Max: name = "Invalid"; break;
}
name = ToStr(img.retType.Type()) + name;
if(img.sampled == 2 && img.dim != Dim::SubpassData)
name = "Storage" + name;
if(img.ms)
name += "MS";
if(img.arrayed)
name += "Array";
type.name = StringFormat::Fmt("Image<%s>", name.c_str());
}
else if(type.type == DataType::SamplerType)
{
type.name = StringFormat::Fmt("sampler", type.id.value());
}
else if(type.type == DataType::SampledImageType)
{
type.name = StringFormat::Fmt("Sampled%s",
dataTypes[sampledImageTypes[type.id].baseId].name.c_str());
}
}
}
// do default names for pointer types in a second pass, because they can point forward at structs
// with higher IDs
for(auto it = dataTypes.begin(); it != dataTypes.end(); ++it)
{
if(it->second.type == DataType::PointerType && it->second.name.empty())
it->second.name = StringFormat::Fmt("%s*", dataTypes[it->second.InnerType()].name.c_str());
}
for(const MemberName &mem : memberNames)
dataTypes[mem.id].children[mem.member].name = mem.name;
memberNames.clear();
}
rdcarray<ShaderEntryPoint> Reflector::EntryPoints() const
{
rdcarray<ShaderEntryPoint> ret;
ret.reserve(entries.size());
for(const EntryPoint &e : entries)
ret.push_back({e.name, MakeShaderStage(e.executionModel)});
return ret;
}
void Reflector::MakeReflection(const GraphicsAPI sourceAPI, const ShaderStage stage,
const rdcstr &entryPoint, const rdcarray<SpecConstant> &specInfo,
ShaderReflection &reflection, ShaderBindpointMapping &mapping,
SPIRVPatchData &patchData) const
{
// set global properties
reflection.entryPoint = entryPoint;
reflection.stage = stage;
reflection.encoding = ShaderEncoding::SPIRV;
reflection.rawBytes.assign((byte *)m_SPIRV.data(), m_SPIRV.size() * sizeof(uint32_t));
CheckDebuggable(reflection.debugInfo.debuggable, reflection.debugInfo.debugStatus);
const EntryPoint *entry = NULL;
for(const EntryPoint &e : entries)
{
if(entryPoint == e.name && MakeShaderStage(e.executionModel) == stage)
{
entry = &e;
break;
}
}
bool multiEntryModule = entries.size() > 1;
if(!entry)
{
RDCERR("Entry point %s for stage %s not found in module", entryPoint.c_str(),
ToStr(stage).c_str());
return;
}
// pick up execution mode size
if(stage == ShaderStage::Compute)
{
const EntryPoint &e = *entry;
if(e.executionModes.localSizeId.x != Id())
{
reflection.dispatchThreadsDimension[0] =
EvaluateConstant(e.executionModes.localSizeId.x, specInfo).value.u32v[0];
reflection.dispatchThreadsDimension[1] =
EvaluateConstant(e.executionModes.localSizeId.y, specInfo).value.u32v[0];
reflection.dispatchThreadsDimension[2] =
EvaluateConstant(e.executionModes.localSizeId.z, specInfo).value.u32v[0];
}
else if(e.executionModes.localSize.x > 0)
{
reflection.dispatchThreadsDimension[0] = e.executionModes.localSize.x;
reflection.dispatchThreadsDimension[1] = e.executionModes.localSize.y;
reflection.dispatchThreadsDimension[2] = e.executionModes.localSize.z;
}
// vulkan spec says "If an object is decorated with the WorkgroupSize decoration, this must take
// precedence over any execution mode set for LocalSize."
for(auto it : constants)
{
const Constant &c = it.second;
if(decorations[c.id].builtIn == BuiltIn::WorkgroupSize)
{
RDCASSERT(c.children.size() == 3);
for(size_t i = 0; i < c.children.size() && i < 3; i++)
reflection.dispatchThreadsDimension[i] =
EvaluateConstant(c.children[i], specInfo).value.u32v[0];
}
}
}
else
{
reflection.dispatchThreadsDimension[0] = reflection.dispatchThreadsDimension[1] =
reflection.dispatchThreadsDimension[2] = 0;
}
if(!cmdline.empty())
reflection.debugInfo.compileFlags.flags = {{"@cmdline", cmdline}};
for(size_t i = 0; i < sources.size(); i++)
{
switch(sources[i].lang)
{
case SourceLanguage::ESSL:
case SourceLanguage::GLSL: reflection.debugInfo.encoding = ShaderEncoding::GLSL; break;
case SourceLanguage::HLSL: reflection.debugInfo.encoding = ShaderEncoding::HLSL; break;
case SourceLanguage::OpenCL_C:
case SourceLanguage::OpenCL_CPP:
case SourceLanguage::CPP_for_OpenCL:
case SourceLanguage::Unknown:
case SourceLanguage::Invalid:
case SourceLanguage::SYCL:
case SourceLanguage::Max: break;
}
if(!sources[i].name.empty())
reflection.debugInfo.files.push_back({sources[i].name, sources[i].contents});
}
PreprocessLineDirectives(reflection.debugInfo.files);
// we do a mini-preprocess of the files from the debug info to handle #line directives.
// This means that any lines that our source file declares to be in another filename via a #line
// get put in the right place for what the debug information hopefully matches.
// We also concatenate duplicate lines and display them all, to handle edge cases where #lines
// declare duplicates.
if(knownExtSet[ExtSet_ShaderDbg] != Id() && !reflection.debugInfo.files.empty())
{
reflection.debugInfo.compileFlags.flags.push_back({"preferSourceDebug", "1"});
reflection.debugInfo.sourceDebugInformation = true;
}
std::set<Id> usedIds;
std::map<Id, std::set<uint32_t>> usedStructChildren;
// build the static call tree from the entry point, and build a list of all IDs referenced
{
std::set<Id> processed;
rdcarray<Id> pending;
pending.push_back(entry->id);
while(!pending.empty())
{
Id func = pending.back();
pending.pop_back();
processed.insert(func);
ConstIter it(m_SPIRV, idOffsets[func]);
while(it.opcode() != Op::FunctionEnd)
{
OpDecoder::ForEachID(it, [&usedIds](Id id, bool result) { usedIds.insert(id); });
if(it.opcode() == Op::AccessChain || it.opcode() == Op::InBoundsAccessChain)
{
OpAccessChain access(it);
// save top-level children referenced in structs
if(dataTypes[dataTypes[idTypes[access.base]].InnerType()].type == DataType::StructType)
usedStructChildren[access.base].insert(
EvaluateConstant(access.indexes[0], specInfo).value.u32v[0]);
}
if(it.opcode() == Op::FunctionCall)
{
OpFunctionCall call(it);
if(processed.find(call.function) == processed.end())
pending.push_back(call.function);
}
it++;
}
}
}
if(m_MajorVersion > 1 || m_MinorVersion >= 4)
{
// from SPIR-V 1.4 onwards we can trust the entry point interface list to give us all used
// global variables. We still use the above heuristic so we can remove unused members of
// gl_PerVertex in structs.
usedIds.clear();
usedIds.insert(entry->usedIds.begin(), entry->usedIds.end());
}
// arrays of elements, which can be appended to in any order and then sorted
rdcarray<SigParameter> inputs;
rdcarray<SigParameter> outputs;
rdcarray<cblockpair> cblocks;
rdcarray<shaderrespair> samplers, roresources, rwresources;
// for pointer types, mapping of inner type ID to index in list (assigned sequentially)
SparseIdMap<uint16_t> pointerTypes;
// $Globals gathering - for GL global values
ConstantBlock globalsblock;
// specialisation constant gathering
ConstantBlock specblock;
for(const Variable &global : globals)
{
if(global.storage == StorageClass::Input || global.storage == StorageClass::Output)
{
// variable type must be a pointer of the same storage class
RDCASSERT(dataTypes[global.type].type == DataType::PointerType);
const DataType &baseType = dataTypes[dataTypes[global.type].InnerType()];
const bool isInput = (global.storage == StorageClass::Input);
rdcarray<SigParameter> &sigarray = (isInput ? inputs : outputs);
// try to use the instance/variable name
rdcstr name = strings[global.id];
// otherwise fall back to naming after the builtin or location
if(name.empty())
{
if(decorations[global.id].flags & Decorations::HasBuiltIn)
name = StringFormat::Fmt("_%s", ToStr(decorations[global.id].builtIn).c_str());
else if(decorations[global.id].flags & Decorations::HasLocation)
name = StringFormat::Fmt("_%s%u", isInput ? "input" : "output",
decorations[global.id].location);
else
name = StringFormat::Fmt("_sig%u", global.id.value());
}
const bool used = usedIds.find(global.id) != usedIds.end();
// if there are multiple entry points in this module only include signature parameters that
// are explicitly used.
if(multiEntryModule && !used)
continue;
// we want to skip any members of the builtin interface block that are completely unused and
// just came along for the ride (usually with gl_Position, but maybe declared and still
// unused). This is meaningless in SPIR-V and just generates useless noise, but some compilers
// from GLSL can generate the whole gl_PerVertex as a literal translation from the implicit
// GLSL declaration.
//
// Some compilers generate global variables instead of members of a global struct. If this is
// a directly decorated builtin variable which is never used, skip it
if(IsStrippableBuiltin(decorations[global.id].builtIn) && !used)
continue;
// if this is a struct variable then either all members must be builtins, or none of them, as
// per the SPIR-V Decoration rules:
//
// "When applied to a structure-type member, all members of that structure type must also be
// decorated with BuiltIn. (No allowed mixing of built-in variables and non-built-in variables
// within a single structure.)"
//
// Some old compilers might generate gl_PerVertex with unused variables having no decoration,
// so to handle this case we treat a struct with any builtin members as if all are builtin -
// which is still legal.
if(baseType.type == DataType::StructType)
{
// look to see if this struct contains a builtin member
bool hasBuiltins = false;
for(size_t i = 0; i < baseType.children.size(); i++)
{
hasBuiltins = (baseType.children[i].decorations.builtIn != BuiltIn::Invalid);
if(hasBuiltins)
break;
}
// if this is the builtin struct, explode the struct and call AddSignatureParameter for each
// member here, so we can skip unused children if we want
if(hasBuiltins)
{
const std::set<uint32_t> &usedchildren = usedStructChildren[global.id];
for(uint32_t i = 0; i < (uint32_t)baseType.children.size(); i++)
{
// skip this member if it's in a builtin struct but has no builtin decoration
if(baseType.children[i].decorations.builtIn == BuiltIn::Invalid)
continue;
// skip this member if it's unused and of a type that is commonly included 'by accident'
if(IsStrippableBuiltin(baseType.children[i].decorations.builtIn) &&
usedchildren.find(i) == usedchildren.end())
continue;
rdcstr childname = name;
if(!baseType.children[i].name.empty())
childname += "." + baseType.children[i].name;
else
childname += StringFormat::Fmt("._child%zu", i);
SPIRVInterfaceAccess patch;
patch.accessChain = {i};
uint32_t dummy = 0;
AddSignatureParameter(isInput, stage, global.id, baseType.id, dummy, patch, childname,
dataTypes[baseType.children[i].type],
baseType.children[i].decorations, sigarray, patchData, specInfo);
}
// move on now, we've processed this global struct
continue;
}
}
uint32_t dummy = 0;
AddSignatureParameter(isInput, stage, global.id, Id(), dummy, {}, name, baseType,
decorations[global.id], sigarray, patchData, specInfo);
}
else if(global.storage == StorageClass::Uniform ||
global.storage == StorageClass::UniformConstant ||
global.storage == StorageClass::AtomicCounter ||
global.storage == StorageClass::StorageBuffer ||
global.storage == StorageClass::PushConstant)
{
// variable type must be a pointer of the same storage class
RDCASSERT(dataTypes[global.type].type == DataType::PointerType);
RDCASSERT(dataTypes[global.type].pointerType.storage == global.storage);
const DataType *varType = &dataTypes[dataTypes[global.type].InnerType()];
// if the outer type is an array, get the length and peel it off.
bool isArray = false;
uint32_t arraySize = 1;
if(varType->type == DataType::ArrayType)
{
isArray = true;
// runtime arrays have no length
if(varType->length != Id())
arraySize = EvaluateConstant(varType->length, specInfo).value.u32v[0];
else
arraySize = ~0U;
varType = &dataTypes[varType->InnerType()];
}
// new SSBOs are in the storage buffer class, previously they were in uniform with BufferBlock
// decoration
const bool ssbo = (global.storage == StorageClass::StorageBuffer) ||
(decorations[varType->id].flags & Decorations::BufferBlock);
const bool pushConst = (global.storage == StorageClass::PushConstant);
const bool atomicCounter = (global.storage == StorageClass::AtomicCounter);
rdcspv::StorageClass effectiveStorage = global.storage;
if(ssbo)
effectiveStorage = StorageClass::StorageBuffer;
Bindpoint bindmap;
// set something crazy so this doesn't overlap with a real buffer binding
if(pushConst)
bindmap.bindset = PushConstantBindSet;
else
bindmap.bindset = GetDescSet(decorations[global.id].set);
bindmap.bind = GetBinding(decorations[global.id].binding);
// On GL if we have a location and no binding, put that in as the bind. It will be overwritten
// dynamically with the actual value read from glGetUniform. This should only happen for
// bare uniforms and not for texture/buffer type uniforms which should have a binding
if(sourceAPI == GraphicsAPI::OpenGL &&
(decorations[global.id].flags & (Decorations::HasLocation | Decorations::HasBinding)) ==
Decorations::HasLocation)
{
bindmap.bind = -int32_t(decorations[global.id].location);
}
bindmap.arraySize = isArray ? arraySize : 1;
bindmap.used = usedIds.find(global.id) != usedIds.end();
if(multiEntryModule && !bindmap.used)
{
// ignore this variable that's not in the entry point's used interface
}
else if(atomicCounter)
{
// GL style atomic counter variable
RDCASSERT(sourceAPI == GraphicsAPI::OpenGL);
ShaderResource res;
res.isReadOnly = false;
res.isTexture = false;
res.name = strings[global.id];
if(res.name.empty())
res.name = varType->name;
if(res.name.empty())
res.name = StringFormat::Fmt("atomic%u", global.id.value());
res.resType = TextureType::Buffer;
res.variableType.columns = 1;
res.variableType.rows = 1;
res.variableType.baseType = VarType::UInt;
res.variableType.name = varType->name;
bindmap.bindset = 0;
bindmap.bind = GetBinding(decorations[global.id].binding);
rwresources.push_back(shaderrespair(bindmap, res));
}
else if(varType->IsOpaqueType())
{
// on Vulkan should never have elements that have no binding declared but are used. On GL we
// should have gotten a location
// above, which will be rewritten later when looking up the pipeline state since it's
// mutable from action to action in theory.
RDCASSERT(!bindmap.used || bindmap.bind != INVALID_BIND);
// opaque type - buffers, images, etc
ShaderResource res;
res.name = strings[global.id];
if(res.name.empty())
res.name = StringFormat::Fmt("res%u", global.id.value());
if(varType->type == DataType::SamplerType)
{
res.resType = TextureType::Unknown;
res.isTexture = false;
res.isReadOnly = true;
samplers.push_back(shaderrespair(bindmap, res));
}
else
{
Id imageTypeId = varType->id;
if(varType->type == DataType::SampledImageType)
imageTypeId = sampledImageTypes[varType->id].baseId;
const Image &imageType = imageTypes[imageTypeId];
if(imageType.ms)
res.resType =
imageType.arrayed ? TextureType::Texture2DMSArray : TextureType::Texture2DMS;
else if(imageType.dim == rdcspv::Dim::_1D)
res.resType = imageType.arrayed ? TextureType::Texture1DArray : TextureType::Texture1D;
else if(imageType.dim == rdcspv::Dim::_2D)
res.resType = imageType.arrayed ? TextureType::Texture2DArray : TextureType::Texture2D;
else if(imageType.dim == rdcspv::Dim::Cube)
res.resType =
imageType.arrayed ? TextureType::TextureCubeArray : TextureType::TextureCube;
else if(imageType.dim == rdcspv::Dim::_3D)
res.resType = TextureType::Texture3D;
else if(imageType.dim == rdcspv::Dim::Rect)
res.resType = TextureType::TextureRect;
else if(imageType.dim == rdcspv::Dim::Buffer)
res.resType = TextureType::Buffer;
else if(imageType.dim == rdcspv::Dim::SubpassData)
res.resType = TextureType::Texture2D;
res.isTexture = res.resType != TextureType::Buffer;
res.isReadOnly = imageType.sampled != 2 || imageType.dim == rdcspv::Dim::SubpassData;
res.variableType.baseType = imageType.retType.Type();
if(res.isReadOnly)
roresources.push_back(shaderrespair(bindmap, res));
else
rwresources.push_back(shaderrespair(bindmap, res));
}
}
else
{
if(varType->type != DataType::StructType)
{
// global loose variable - add to $Globals block
RDCASSERT(varType->type == DataType::ScalarType || varType->type == DataType::VectorType ||
varType->type == DataType::MatrixType || varType->type == DataType::ArrayType);
RDCASSERT(sourceAPI == GraphicsAPI::OpenGL);
ShaderConstant constant;
MakeConstantBlockVariable(constant, pointerTypes, effectiveStorage, *varType,
strings[global.id], decorations[global.id], specInfo);
if(isArray)
constant.type.elements = arraySize;
else
constant.type.elements = 0;
constant.byteOffset = decorations[global.id].location;
globalsblock.variables.push_back(constant);
}
else
{
// on Vulkan should never have elements that have no binding declared but are used, unless
// it's push constants (which is handled elsewhere). On GL we should have gotten a
// location above, which will be rewritten later when looking up the pipeline state since
// it's mutable from action to action in theory.
RDCASSERT(!bindmap.used || pushConst || bindmap.bind != INVALID_BIND);
if(ssbo)
{
ShaderResource res;
res.isReadOnly = false;
res.isTexture = false;
res.name = strings[global.id];
if(res.name.empty())
res.name = StringFormat::Fmt("ssbo%u", global.id.value());
res.resType = TextureType::Buffer;
res.variableType.columns = 0;
res.variableType.rows = 0;
res.variableType.baseType = VarType::Float;
res.variableType.name = varType->name;
MakeConstantBlockVariables(effectiveStorage, *varType, 0, 0, res.variableType.members,
pointerTypes, specInfo);
rwresources.push_back(shaderrespair(bindmap, res));
}
else
{
ConstantBlock cblock;
cblock.name = strings[global.id];
if(cblock.name.empty())
cblock.name = StringFormat::Fmt("uniforms%u", global.id.value());
cblock.bufferBacked = !pushConst;
MakeConstantBlockVariables(effectiveStorage, *varType, 0, 0, cblock.variables,
pointerTypes, specInfo);
if(!varType->children.empty())
cblock.byteSize = CalculateMinimumByteSize(cblock.variables);
else
cblock.byteSize = 0;
cblocks.push_back(cblockpair(bindmap, cblock));
}
}
}
}
else if(global.storage == StorageClass::Private ||
global.storage == StorageClass::CrossWorkgroup ||
global.storage == StorageClass::Workgroup)
{
// silently allow
}
else
{
RDCWARN("Unexpected storage class for global: %s", ToStr(global.storage).c_str());
}
}
for(auto it : constants)
{
const Constant &c = it.second;
if(decorations[c.id].flags & Decorations::HasSpecId)
{
rdcstr name = strings[c.id];
if(name.empty())
name = StringFormat::Fmt("specID%u", decorations[c.id].specID);
ShaderConstant spec;
MakeConstantBlockVariable(spec, pointerTypes, rdcspv::StorageClass::PushConstant,
dataTypes[c.type], name, decorations[c.id], specInfo);
spec.byteOffset = uint32_t(specblock.variables.size() * sizeof(uint64_t));
spec.defaultValue = c.value.value.u64v[0];
specblock.variables.push_back(spec);
patchData.specIDs.push_back(decorations[c.id].specID);
}
}
if(!specblock.variables.empty())
{
specblock.name = "Specialization Constants";
specblock.bufferBacked = false;
specblock.compileConstants = true;
specblock.byteSize = 0;
Bindpoint bindmap;
// set something crazy so this doesn't overlap with a real buffer binding
// also identify this as specialization constant data
bindmap.bindset = SpecializationConstantBindSet;
bindmap.bind = INVALID_BIND;
bindmap.arraySize = 1;
bindmap.used = true;
cblocks.push_back(cblockpair(bindmap, specblock));
}
if(!globalsblock.variables.empty())
{
globalsblock.name = "$Globals";
globalsblock.bufferBacked = false;
globalsblock.byteSize = (uint32_t)globalsblock.variables.size();
globalsblock.bindPoint = (int)cblocks.size();
Bindpoint bindmap;
bindmap.bindset = 0;
bindmap.bind = INVALID_BIND;
bindmap.arraySize = 1;
bindmap.used = true;
cblocks.push_back(cblockpair(bindmap, globalsblock));
}
// look for execution modes that affect the reflection and apply them
{
const EntryPoint &e = *entry;
if(e.executionModes.depthMode == ExecutionModes::DepthGreater)
{
for(SigParameter &sig : outputs)
{
if(sig.systemValue == ShaderBuiltin::DepthOutput)
sig.systemValue = ShaderBuiltin::DepthOutputGreaterEqual;
}
}
else if(e.executionModes.depthMode == ExecutionModes::DepthLess)
{
for(SigParameter &sig : outputs)
{
if(sig.systemValue == ShaderBuiltin::DepthOutput)
sig.systemValue = ShaderBuiltin::DepthOutputLessEqual;
}
}
patchData.outTopo = e.executionModes.outTopo;
}
for(auto it = extSets.begin(); it != extSets.end(); it++)
if(it->second == "NonSemantic.DebugPrintf")
patchData.usesPrintf = true;
// sort system value semantics to the start of the list
struct sig_param_sort
{
sig_param_sort(const rdcarray<SigParameter> &arr) : sigArray(arr) {}
const rdcarray<SigParameter> &sigArray;
bool operator()(const size_t idxA, const size_t idxB)
{
const SigParameter &a = sigArray[idxA];
const SigParameter &b = sigArray[idxB];
if(a.systemValue == b.systemValue)
{
if(a.regIndex != b.regIndex)
return a.regIndex < b.regIndex;
if(a.regChannelMask != b.regChannelMask)
return a.regChannelMask < b.regChannelMask;
return a.varName < b.varName;
}
if(a.systemValue == ShaderBuiltin::Undefined)
return false;
if(b.systemValue == ShaderBuiltin::Undefined)
return true;
return a.systemValue < b.systemValue;
}
};
rdcarray<size_t> indices;
{
indices.resize(inputs.size());
for(size_t i = 0; i < inputs.size(); i++)
indices[i] = i;
std::sort(indices.begin(), indices.end(), sig_param_sort(inputs));
reflection.inputSignature.reserve(inputs.size());
for(size_t i = 0; i < inputs.size(); i++)
reflection.inputSignature.push_back(inputs[indices[i]]);
rdcarray<SPIRVInterfaceAccess> inPatch = patchData.inputs;
for(size_t i = 0; i < inputs.size(); i++)
patchData.inputs[i] = inPatch[indices[i]];
}
{
indices.resize(outputs.size());
for(size_t i = 0; i < outputs.size(); i++)
indices[i] = i;
std::sort(indices.begin(), indices.end(), sig_param_sort(outputs));
reflection.outputSignature.reserve(outputs.size());
for(size_t i = 0; i < outputs.size(); i++)
reflection.outputSignature.push_back(outputs[indices[i]]);
rdcarray<SPIRVInterfaceAccess> outPatch = patchData.outputs;
for(size_t i = 0; i < outputs.size(); i++)
patchData.outputs[i] = outPatch[indices[i]];
}
size_t numInputs = 16;
for(size_t i = 0; i < reflection.inputSignature.size(); i++)
if(reflection.inputSignature[i].systemValue == ShaderBuiltin::Undefined)
numInputs = RDCMAX(numInputs, (size_t)reflection.inputSignature[i].regIndex + 1);
mapping.inputAttributes.resize(numInputs);
for(size_t i = 0; i < numInputs; i++)
mapping.inputAttributes[i] = -1;
for(size_t i = 0; i < reflection.inputSignature.size(); i++)
if(reflection.inputSignature[i].systemValue == ShaderBuiltin::Undefined)
mapping.inputAttributes[reflection.inputSignature[i].regIndex] = (int32_t)i;
for(cblockpair &cb : cblocks)
{
// sort the variables within each block because we want them in offset order but they don't have
// to be declared in offset order in the SPIR-V.
std::sort(cb.bindres.variables.begin(), cb.bindres.variables.end());
}
std::sort(cblocks.begin(), cblocks.end());
std::sort(samplers.begin(), samplers.end());
std::sort(roresources.begin(), roresources.end());
std::sort(rwresources.begin(), rwresources.end());
mapping.constantBlocks.resize(cblocks.size());
reflection.constantBlocks.resize(cblocks.size());
mapping.samplers.resize(samplers.size());
reflection.samplers.resize(samplers.size());
mapping.readOnlyResources.resize(roresources.size());
reflection.readOnlyResources.resize(roresources.size());
mapping.readWriteResources.resize(rwresources.size());
reflection.readWriteResources.resize(rwresources.size());
for(size_t i = 0; i < cblocks.size(); i++)
{
mapping.constantBlocks[i] = cblocks[i].map;
// fix up any bind points marked with INVALID_BIND. They were sorted to the end
// but from here on we want to just be able to index with the bind point
// without any special casing.
if(mapping.constantBlocks[i].bind == INVALID_BIND)
mapping.constantBlocks[i].bind = 0;
reflection.constantBlocks[i] = cblocks[i].bindres;
reflection.constantBlocks[i].bindPoint = (int32_t)i;
}
for(size_t i = 0; i < samplers.size(); i++)
{
mapping.samplers[i] = samplers[i].map;
// fix up any bind points marked with INVALID_BIND. They were sorted to the end
// but from here on we want to just be able to index with the bind point
// without any special casing.
if(mapping.samplers[i].bind == INVALID_BIND)
mapping.samplers[i].bind = 0;
reflection.samplers[i].name = samplers[i].bindres.name;
reflection.samplers[i].bindPoint = (int32_t)i;
}
for(size_t i = 0; i < roresources.size(); i++)
{
mapping.readOnlyResources[i] = roresources[i].map;
// fix up any bind points marked with INVALID_BIND. They were sorted to the end
// but from here on we want to just be able to index with the bind point
// without any special casing.
if(mapping.readOnlyResources[i].bind == INVALID_BIND)
mapping.readOnlyResources[i].bind = 0;
reflection.readOnlyResources[i] = roresources[i].bindres;
reflection.readOnlyResources[i].bindPoint = (int32_t)i;
}
for(size_t i = 0; i < rwresources.size(); i++)
{
mapping.readWriteResources[i] = rwresources[i].map;
// fix up any bind points marked with INVALID_BIND. They were sorted to the end
// but from here on we want to just be able to index with the bind point
// without any special casing.
if(mapping.readWriteResources[i].bind == INVALID_BIND)
mapping.readWriteResources[i].bind = 0;
reflection.readWriteResources[i] = rwresources[i].bindres;
reflection.readWriteResources[i].bindPoint = (int32_t)i;
}
// populate the pointer types
reflection.pointerTypes.reserve(pointerTypes.size());
for(auto it = pointerTypes.begin(); it != pointerTypes.end(); ++it)
{
ShaderConstant dummy;
MakeConstantBlockVariable(dummy, pointerTypes, dataTypes[it->first].pointerType.storage,
dataTypes[it->first], rdcstr(), Decorations(), specInfo);
if(it->second >= reflection.pointerTypes.size())
reflection.pointerTypes.resize(it->second + 1);
reflection.pointerTypes[it->second] = dummy.type;
}
}
void Reflector::MakeConstantBlockVariables(rdcspv::StorageClass storage, const DataType &structType,
uint32_t arraySize, uint32_t arrayByteStride,
rdcarray<ShaderConstant> &cblock,
SparseIdMap<uint16_t> &pointerTypes,
const rdcarray<SpecConstant> &specInfo) const
{
// we get here for multi-dimensional arrays
if(structType.type == DataType::ArrayType)
{
uint32_t relativeOffset = 0;
if(arraySize == ~0U)
arraySize = 1;
cblock.resize(arraySize);
for(uint32_t i = 0; i < arraySize; i++)
{
MakeConstantBlockVariable(cblock[i], pointerTypes, storage, structType,
StringFormat::Fmt("[%u]", i), decorations[structType.id], specInfo);
cblock[i].byteOffset = relativeOffset;
relativeOffset += arrayByteStride;
}
return;
}
if(structType.children.empty())
return;
cblock.resize(structType.children.size());
for(size_t i = 0; i < structType.children.size(); i++)
{
MakeConstantBlockVariable(cblock[i], pointerTypes, storage,
dataTypes[structType.children[i].type], structType.children[i].name,
structType.children[i].decorations, specInfo);
}
uint32_t emptyStructSize = 4;
if(storage == rdcspv::StorageClass::StorageBuffer)
{
if(capabilities.find(rdcspv::Capability::StorageBuffer8BitAccess) != capabilities.end() ||
capabilities.find(rdcspv::Capability::UniformAndStorageBuffer8BitAccess) != capabilities.end())
emptyStructSize = 1;
else if(capabilities.find(rdcspv::Capability::StorageBuffer16BitAccess) != capabilities.end() ||
capabilities.find(rdcspv::Capability::UniformAndStorageBuffer16BitAccess) !=
capabilities.end())
emptyStructSize = 2;
}
else if(storage == rdcspv::StorageClass::Uniform)
{
if(capabilities.find(rdcspv::Capability::UniformAndStorageBuffer8BitAccess) != capabilities.end())
emptyStructSize = 1;
else if(capabilities.find(rdcspv::Capability::UniformAndStorageBuffer16BitAccess) !=
capabilities.end())
emptyStructSize = 2;
}
else if(storage == rdcspv::StorageClass::PushConstant)
{
if(capabilities.find(rdcspv::Capability::StoragePushConstant8) != capabilities.end())
emptyStructSize = 1;
else if(capabilities.find(rdcspv::Capability::StoragePushConstant16) != capabilities.end())
emptyStructSize = 2;
}
for(size_t i = 0; i < cblock.size(); i++)
{
// for structs that aren't in arrays, we need to define their byte size (stride). Without
// knowing the packing rules this shader is complying with, this is not fully possible.
//
// what we do is choose the most conservative size - so that this struct's size alone doesn't
// invalidate compliance with a particular ruleset (e.g. std140).
//
// we calculate the scalar, base, and extended sizes of the struct sizes of the struct. The
// largest one that fits between this struct and the next member is the one we use. If there is
// no next member, we always use the base size as it's impossible to tell how much trailing
// padding the shader expected.
//
// If we guess wrongly small, members after this struct will need an [[offset]] decoration,
// If we guess wrongly large the struct itself will need a [[size]] decoration
// Since we're choosing the largest valid size, it will always be just a [[size]] which might be
// unnecessary (if e.g. somewhere else the shader demonstrates scalar packing so the padded size
// is larger than the scalar calculated size) but that's only present in one place.
if(cblock[i].type.baseType == VarType::Struct && cblock[i].type.arrayByteStride == 0)
{
// this should not be an array - if it is SPIR-V requires an array byte stride, and this
// calculation below is also invalid.
RDCASSERTEQUAL(cblock[i].type.elements, 1);
StructSizes sizes = CalculateStructProps(emptyStructSize, cblock[i]);
uint32_t availSize = ~0U;
if(i + 1 < cblock.size())
availSize = cblock[i + 1].byteOffset - cblock[i].byteOffset;
else if(arrayByteStride != 0)
availSize = arrayByteStride - cblock[i].byteOffset;
// expect at least the scalar size to be available otherwise this struct seems to overlap
RDCASSERT(sizes.scalarSize <= availSize, sizes.scalarSize, availSize);
if(sizes.extendedSize <= availSize)
cblock[i].type.arrayByteStride = sizes.extendedSize;
else if(sizes.baseSize <= availSize)
cblock[i].type.arrayByteStride = sizes.baseSize;
else
cblock[i].type.arrayByteStride = sizes.scalarSize;
}
}
}
void Reflector::ApplyMatrixByteStride(const DataType &type, uint8_t matrixByteStride,
rdcarray<ShaderConstant> &members) const
{
const DataType &inner = dataTypes[type.InnerType()];
for(ShaderConstant &m : members)
{
if(m.type.matrixByteStride == 0)
m.type.matrixByteStride = matrixByteStride;
if(inner.type == DataType::ArrayType)
ApplyMatrixByteStride(inner, matrixByteStride, m.type.members);
}
}
void Reflector::MakeConstantBlockVariable(ShaderConstant &outConst,
SparseIdMap<uint16_t> &pointerTypes,
rdcspv::StorageClass storage, const DataType &type,
const rdcstr &name, const Decorations &varDecorations,
const rdcarray<SpecConstant> &specInfo) const
{
outConst.name = name;
outConst.defaultValue = 0;
if(varDecorations.offset != ~0U)
outConst.byteOffset = varDecorations.offset;
const DataType *curType = &type;
// if the type is an array, set array size and strides then unpeel the array
if(curType->type == DataType::ArrayType)
{
outConst.type.elements =
curType->length != Id() ? EvaluateConstant(curType->length, specInfo).value.u32v[0] : ~0U;
if(varDecorations.arrayStride != ~0U)
{
RDCASSERTMSG("Stride is too large for uint16_t", varDecorations.arrayStride <= 0xffff);
outConst.type.arrayByteStride = RDCMIN(varDecorations.arrayStride, 0xffffu) & 0xffff;
}
else if(decorations[curType->id].arrayStride != ~0U)
{
RDCASSERTMSG("Stride is too large for uint16_t",
decorations[curType->id].arrayStride <= 0xffff);
outConst.type.arrayByteStride = RDCMIN(decorations[curType->id].arrayStride, 0xffffu) & 0xffff;
}
if(varDecorations.matrixStride != ~0U)
outConst.type.matrixByteStride = varDecorations.matrixStride & 0xff;
else if(decorations[curType->id].matrixStride != ~0U)
outConst.type.matrixByteStride = decorations[curType->id].matrixStride & 0xff;
curType = &dataTypes[curType->InnerType()];
}
if(curType->type == DataType::VectorType || curType->type == DataType::MatrixType)
{
outConst.type.baseType = curType->scalar().Type();
if(curType->type == DataType::VectorType || (varDecorations.flags & Decorations::RowMajor))
outConst.type.flags |= ShaderVariableFlags::RowMajorMatrix;
if(varDecorations.matrixStride != ~0U)
outConst.type.matrixByteStride = varDecorations.matrixStride & 0xff;
if(curType->type == DataType::MatrixType)
{
outConst.type.rows = (uint8_t)curType->vector().count;
outConst.type.columns = (uint8_t)curType->matrix().count;
}
else
{
outConst.type.columns = (uint8_t)curType->vector().count;
}
outConst.type.name = curType->name;
}
else if(curType->type == DataType::ScalarType)
{
outConst.type.baseType = curType->scalar().Type();
outConst.type.flags |= ShaderVariableFlags::RowMajorMatrix;
outConst.type.name = curType->name;
}
else
{
if(curType->type == DataType::PointerType)
{
outConst.type.baseType = VarType::ULong;
outConst.type.rows = 1;
outConst.type.columns = 1;
outConst.type.name = curType->name;
// try to insert the inner type ID into the map. If it succeeds, it gets the next available
// pointer type index (size of the map), if not then we just get the previously added index
auto it =
pointerTypes.insert(std::make_pair(curType->InnerType(), (uint16_t)pointerTypes.size()));
outConst.type.pointerTypeID = it.first->second;
return;
}
RDCASSERT(curType->type == DataType::StructType || curType->type == DataType::ArrayType);
outConst.type.baseType = VarType::Struct;
outConst.type.rows = 0;
outConst.type.columns = 0;
outConst.type.name = curType->name;
MakeConstantBlockVariables(storage, *curType, outConst.type.elements,
outConst.type.arrayByteStride, outConst.type.members, pointerTypes,
specInfo);
if(curType->type == DataType::ArrayType)
{
// matrix byte stride is only applied on the root variable, so as we recurse down the type
// tree for multi-dimensional arrays of matrices we need to propagate down the stride
if(outConst.type.matrixByteStride != 0)
ApplyMatrixByteStride(*curType, outConst.type.matrixByteStride, outConst.type.members);
outConst.type.name = type.name;
// if the inner type is an array, it will be expanded in our members list. So don't also
// redundantly keep the element count
outConst.type.arrayByteStride *= outConst.type.elements;
outConst.type.elements = 1;
}
}
}
void Reflector::AddSignatureParameter(const bool isInput, const ShaderStage stage,
const Id globalID, const Id parentStructID, uint32_t &regIndex,
const SPIRVInterfaceAccess &parentPatch, const rdcstr &varName,
const DataType &type, const Decorations &varDecorations,
rdcarray<SigParameter> &sigarray, SPIRVPatchData &patchData,
const rdcarray<SpecConstant> &specInfo) const
{
SigParameter sig;
sig.needSemanticIndex = false;
SPIRVInterfaceAccess patch;
patch.accessChain = parentPatch.accessChain;
patch.ID = globalID;
patch.structID = parentStructID;
patch.isArraySubsequentElement = parentPatch.isArraySubsequentElement;
if(parentStructID)
patch.structMemberIndex = patch.accessChain.back();
const bool rowmajor = (varDecorations.flags & Decorations::RowMajor) != 0;
sig.regIndex = regIndex;
if(varDecorations.location != ~0U)
sig.regIndex = regIndex = varDecorations.location;
if(varDecorations.builtIn != BuiltIn::Invalid)
sig.systemValue = MakeShaderBuiltin(stage, varDecorations.builtIn);
// fragment shader outputs are implicitly colour outputs. All other builtin outputs do not have a
// register index
if(stage == ShaderStage::Fragment && !isInput && sig.systemValue == ShaderBuiltin::Undefined)
sig.systemValue = ShaderBuiltin::ColorOutput;
else if(sig.systemValue != ShaderBuiltin::Undefined)
sig.regIndex = 0;
const DataType *varType = &type;
bool isArray = false;
uint32_t arraySize = 1;
if(varType->type == DataType::ArrayType)
{
arraySize = EvaluateConstant(varType->length, specInfo).value.u32v[0];
isArray = true;
varType = &dataTypes[varType->InnerType()];
// if this is the first array level, we sometimes ignore it.
if(patch.accessChain.empty())
{
// for geometry/tessellation evaluation shaders, ignore the root level of array-ness for
// inputs
if((stage == ShaderStage::Geometry || stage == ShaderStage::Tess_Eval) && isInput)
arraySize = 1;
// for tessellation control shaders, ignore the root level of array-ness for both inputs and
// outputs
if(stage == ShaderStage::Tess_Control)
arraySize = 1;
// if this is a root array in the geometry shader, don't reflect it as an array either
if(stage == ShaderStage::Geometry && isInput)
isArray = false;
}
// arrays will need an extra access chain index
patch.accessChain.push_back(0U);
}
// if the current type is a struct, recurse for each member
if(varType->type == DataType::StructType)
{
for(uint32_t a = 0; a < arraySize; a++)
{
// push the member-index access chain value
patch.accessChain.push_back(0U);
for(size_t c = 0; c < varType->children.size(); c++)
{
rdcstr childName = varName;
if(isArray)
childName += StringFormat::Fmt("[%u]", a);
if(!varType->children[c].name.empty())
childName += "." + varType->children[c].name;
else
childName += StringFormat::Fmt("._child%zu", c);
AddSignatureParameter(isInput, stage, globalID, varType->id, regIndex, patch, childName,
dataTypes[varType->children[c].type],
varType->children[c].decorations, sigarray, patchData, specInfo);
// increment the member-index access chain value
patch.accessChain.back()++;
}
// pop the member-index access chain value
patch.accessChain.pop_back();
// increment the array-index access chain value
if(isArray)
{
patch.accessChain.back()++;
patch.isArraySubsequentElement = true;
}
}
return;
}
// similarly for arrays (this happens for multi-dimensional arrays
if(varType->type == DataType::ArrayType)
{
for(uint32_t a = 0; a < arraySize; a++)
{
AddSignatureParameter(isInput, stage, globalID, Id(), regIndex, patch,
varName + StringFormat::Fmt("[%u]", a), *varType, {}, sigarray,
patchData, specInfo);
// increment the array-index access chain value
patch.accessChain.back()++;
patch.isArraySubsequentElement = true;
}
return;
}
sig.varType = varType->scalar().Type();
sig.compCount = RDCMAX(1U, varType->vector().count);
sig.stream = 0;
sig.regChannelMask = sig.channelUsedMask = (1 << sig.compCount) - 1;
for(const DecorationAndParamData &d : varDecorations.others)
if(d.value == Decoration::Component)
sig.regChannelMask <<= d.component;
sig.channelUsedMask = sig.regChannelMask;
uint32_t regStep = 1;
if(sig.varType == VarType::Double || sig.varType == VarType::ULong || sig.varType == VarType::SLong)
regStep = 2;
for(uint32_t a = 0; a < arraySize; a++)
{
rdcstr n = varName;
if(isArray)
n += StringFormat::Fmt("[%u]", a);
sig.varName = n;
if(varType->matrix().count <= 1)
{
sigarray.push_back(sig);
regIndex += regStep;
if(isInput)
patchData.inputs.push_back(patch);
else
patchData.outputs.push_back(patch);
}
else
{
// use an extra access chain to get each vector out of the matrix.
patch.accessChain.push_back(0);
for(uint32_t m = 0; m < varType->matrix().count; m++)
{
SigParameter s = sig;
s.varName = StringFormat::Fmt("%s:%s%u", n.c_str(), rowmajor ? "row" : "col", m);
s.regIndex += m * regStep;
sigarray.push_back(s);
if(isInput)
patchData.inputs.push_back(patch);
else
patchData.outputs.push_back(patch);
regIndex += regStep;
// increment the matrix column access chain
patch.accessChain.back()++;
patch.isArraySubsequentElement = true;
}
// pop the matrix column access chain
patch.accessChain.pop_back();
}
sig.regIndex += RDCMAX(1U, varType->matrix().count) * regStep;
// increment the array index access chain (if it exists)
if(isArray)
{
patch.accessChain.back()++;
patch.isArraySubsequentElement = true;
}
}
}
}; // namespace rdcspv
#if ENABLED(ENABLE_UNIT_TESTS)
#include "catch/catch.hpp"
#include "data/glsl_shaders.h"
#include "glslang_compile.h"
TEST_CASE("Validate SPIR-V reflection", "[spirv][reflection]")
{
ShaderType type = ShaderType::Vulkan;
auto compiler = [&type](ShaderStage stage, const rdcstr &source, const rdcstr &entryPoint,
ShaderReflection &refl, ShaderBindpointMapping &mapping) {
rdcspv::Init();
RenderDoc::Inst().RegisterShutdownFunction(&rdcspv::Shutdown);
rdcarray<uint32_t> spirv;
rdcspv::CompilationSettings settings(type == ShaderType::Vulkan
? rdcspv::InputLanguage::VulkanGLSL
: rdcspv::InputLanguage::OpenGLGLSL,
rdcspv::ShaderStage(stage));
settings.debugInfo = true;
rdcstr errors = rdcspv::Compile(settings, {source}, spirv);
INFO("SPIR-V compile output: " << errors);
REQUIRE(!spirv.empty());
rdcspv::Reflector spv;
spv.Parse(spirv);
SPIRVPatchData patchData;
spv.MakeReflection(type == ShaderType::Vulkan ? GraphicsAPI::Vulkan : GraphicsAPI::OpenGL,
stage, entryPoint, {}, refl, mapping, patchData);
};
// test both Vulkan and GL SPIR-V reflection
SECTION("Vulkan GLSL reflection")
{
type = ShaderType::Vulkan;
TestGLSLReflection(type, compiler);
};
SECTION("OpenGL GLSL reflection")
{
type = ShaderType::GLSPIRV;
TestGLSLReflection(type, compiler);
};
}
#endif