Implement 4 AMD-specific extensions.
- Support GL_AMD_shader_ballot (SPV_AMD_shader_ballot).
- Support GL_AMD_shader_trinary_minmax (SPV_AMD_shader_trinary_minmax).
- Support GL_AMD_shader_explicit_vertex_parameter
(SPV_AMD_shader_explicit_vertex_parameter).
- Support GL_AMD_gcn_shader (SPV_AMD_gcn_shader).
diff --git a/SPIRV/GlslangToSpv.cpp b/SPIRV/GlslangToSpv.cpp
index dd11304..0fc5730 100755
--- a/SPIRV/GlslangToSpv.cpp
+++ b/SPIRV/GlslangToSpv.cpp
@@ -43,6 +43,9 @@
#include "SpvBuilder.h"
namespace spv {
#include "GLSL.std.450.h"
+#ifdef AMD_EXTENSIONS
+ #include "GLSL.ext.AMD.h"
+#endif
}
// Glslang includes
@@ -147,9 +150,9 @@
spv::Id createConversion(glslang::TOperator op, spv::Decoration precision, spv::Decoration noContraction, spv::Id destTypeId, spv::Id operand, glslang::TBasicType typeProxy);
spv::Id makeSmearedConstant(spv::Id constant, int vectorSize);
spv::Id createAtomicOperation(glslang::TOperator op, spv::Decoration precision, spv::Id typeId, std::vector<spv::Id>& operands, glslang::TBasicType typeProxy);
- spv::Id createInvocationsOperation(glslang::TOperator, spv::Id typeId, spv::Id operand);
+ spv::Id createInvocationsOperation(glslang::TOperator, spv::Id typeId, spv::Id operand, glslang::TBasicType typeProxy);
spv::Id createMiscOperation(glslang::TOperator op, spv::Decoration precision, spv::Id typeId, std::vector<spv::Id>& operands, glslang::TBasicType typeProxy);
- spv::Id createNoArgOperation(glslang::TOperator op);
+ spv::Id createNoArgOperation(glslang::TOperator op, spv::Decoration precision, spv::Id typeId);
spv::Id getSymbolId(const glslang::TIntermSymbol* node);
void addDecoration(spv::Id id, spv::Decoration dec);
void addDecoration(spv::Id id, spv::Decoration dec, unsigned value);
@@ -160,6 +163,7 @@
bool isTrivialLeaf(const glslang::TIntermTyped* node);
bool isTrivial(const glslang::TIntermTyped* node);
spv::Id createShortCircuit(glslang::TOperator, glslang::TIntermTyped& left, glslang::TIntermTyped& right);
+ spv::Id getExtBuiltins(const char* name);
spv::Function* shaderEntry;
spv::Instruction* entryPoint;
@@ -175,6 +179,7 @@
std::set<spv::Id> iOSet; // all input/output variables from either static use or declaration of interface
const glslang::TIntermediate* glslangIntermediate;
spv::Id stdBuiltins;
+ std::unordered_map<const char*, spv::Id> extBuiltinMap;
std::unordered_map<int, spv::Id> symbolValues;
std::unordered_set<int> constReadOnlyParameters; // set of formal function parameters that have glslang qualifier constReadOnly, so we know they are not local function "const" that are write-once
@@ -371,6 +376,10 @@
return spv::DecorationNoPerspective;
else if (qualifier.flat)
return spv::DecorationFlat;
+#ifdef AMD_EXTENSIONS
+ else if (qualifier.explicitInterp)
+ return spv::DecorationExplicitInterpAMD;
+#endif
else
return spv::DecorationMax;
}
@@ -508,6 +517,15 @@
// TODO: Add SPIR-V builtin ID.
logger->missingFunctionality("shader ballot");
return spv::BuiltInMax;
+#ifdef AMD_EXTENSIONS
+ case glslang::EbvBaryCoordNoPersp: return spv::BuiltInBaryCoordNoPerspAMD;
+ case glslang::EbvBaryCoordNoPerspCentroid: return spv::BuiltInBaryCoordNoPerspCentroidAMD;
+ case glslang::EbvBaryCoordNoPerspSample: return spv::BuiltInBaryCoordNoPerspSampleAMD;
+ case glslang::EbvBaryCoordSmooth: return spv::BuiltInBaryCoordSmoothAMD;
+ case glslang::EbvBaryCoordSmoothCentroid: return spv::BuiltInBaryCoordSmoothCentroidAMD;
+ case glslang::EbvBaryCoordSmoothSample: return spv::BuiltInBaryCoordSmoothSampleAMD;
+ case glslang::EbvBaryCoordPullModel: return spv::BuiltInBaryCoordPullModelAMD;
+#endif
default: return spv::BuiltInMax;
}
}
@@ -628,6 +646,10 @@
child.invariant = true;
if (parent.nopersp)
child.nopersp = true;
+#ifdef AMD_EXTENSIONS
+ if (parent.explicitInterp)
+ child.explicitInterp = true;
+#endif
if (parent.flat)
child.flat = true;
if (parent.centroid)
@@ -1484,6 +1506,9 @@
break;
case glslang::EOpInterpolateAtSample:
case glslang::EOpInterpolateAtOffset:
+#ifdef AMD_EXTENSIONS
+ case glslang::EOpInterpolateAtVertex:
+#endif
if (arg == 0)
lvalue = true;
break;
@@ -1524,7 +1549,7 @@
// Pass through to generic operations.
switch (glslangOperands.size()) {
case 0:
- result = createNoArgOperation(node->getOp());
+ result = createNoArgOperation(node->getOp(), precision, convertGlslangToSpvType(node->getType()));
break;
case 1:
result = createUnaryOperation(
@@ -3169,6 +3194,7 @@
spv::Id TGlslangToSpvTraverser::createUnaryOperation(glslang::TOperator op, spv::Decoration precision, spv::Decoration noContraction, spv::Id typeId, spv::Id operand, glslang::TBasicType typeProxy)
{
spv::Op unaryOp = spv::OpNop;
+ int extBuiltins = -1;
int libCall = -1;
bool isUnsigned = typeProxy == glslang::EbtUint || typeProxy == glslang::EbtUint64;
bool isFloat = typeProxy == glslang::EbtFloat || typeProxy == glslang::EbtDouble;
@@ -3448,7 +3474,32 @@
case glslang::EOpAnyInvocation:
case glslang::EOpAllInvocations:
case glslang::EOpAllInvocationsEqual:
- return createInvocationsOperation(op, typeId, operand);
+#ifdef AMD_EXTENSIONS
+ case glslang::EOpMinInvocations:
+ case glslang::EOpMaxInvocations:
+ case glslang::EOpAddInvocations:
+ case glslang::EOpMinInvocationsNonUniform:
+ case glslang::EOpMaxInvocationsNonUniform:
+ case glslang::EOpAddInvocationsNonUniform:
+#endif
+ return createInvocationsOperation(op, typeId, operand, typeProxy);
+
+#ifdef AMD_EXTENSIONS
+ case glslang::EOpMbcnt:
+ extBuiltins = getExtBuiltins(spv::E_SPV_AMD_shader_ballot);
+ libCall = spv::MbcntAMD;
+ break;
+
+ case glslang::EOpCubeFaceIndex:
+ extBuiltins = getExtBuiltins(spv::E_SPV_AMD_gcn_shader);
+ libCall = spv::CubeFaceIndexAMD;
+ break;
+
+ case glslang::EOpCubeFaceCoord:
+ extBuiltins = getExtBuiltins(spv::E_SPV_AMD_gcn_shader);
+ libCall = spv::CubeFaceCoordAMD;
+ break;
+#endif
default:
return 0;
@@ -3458,7 +3509,7 @@
if (libCall >= 0) {
std::vector<spv::Id> args;
args.push_back(operand);
- id = builder.createBuiltinCall(typeId, stdBuiltins, libCall, args);
+ id = builder.createBuiltinCall(typeId, extBuiltins >= 0 ? extBuiltins : stdBuiltins, libCall, args);
} else {
id = builder.createUnaryOp(unaryOp, typeId, operand);
}
@@ -3765,12 +3816,20 @@
}
// Create group invocation operations.
-spv::Id TGlslangToSpvTraverser::createInvocationsOperation(glslang::TOperator op, spv::Id typeId, spv::Id operand)
+spv::Id TGlslangToSpvTraverser::createInvocationsOperation(glslang::TOperator op, spv::Id typeId, spv::Id operand, glslang::TBasicType typeProxy)
{
+ bool isUnsigned = typeProxy == glslang::EbtUint || typeProxy == glslang::EbtUint64;
+ bool isFloat = typeProxy == glslang::EbtFloat || typeProxy == glslang::EbtDouble;
+
builder.addCapability(spv::CapabilityGroups);
std::vector<spv::Id> operands;
operands.push_back(builder.makeUintConstant(spv::ScopeSubgroup));
+#ifdef AMD_EXTENSIONS
+ if (op == glslang::EOpMinInvocations || op == glslang::EOpMaxInvocations || op == glslang::EOpAddInvocations ||
+ op == glslang::EOpMinInvocationsNonUniform || op == glslang::EOpMaxInvocationsNonUniform || op == glslang::EOpAddInvocationsNonUniform)
+ operands.push_back(spv::GroupOperationReduce);
+#endif
operands.push_back(operand);
switch (op) {
@@ -3786,6 +3845,74 @@
return builder.createBinOp(spv::OpLogicalOr, typeId, groupAll,
builder.createUnaryOp(spv::OpLogicalNot, typeId, groupAny));
}
+#ifdef AMD_EXTENSIONS
+ case glslang::EOpMinInvocations:
+ case glslang::EOpMaxInvocations:
+ case glslang::EOpAddInvocations:
+ {
+ spv::Op spvOp = spv::OpNop;
+ if (op == glslang::EOpMinInvocations) {
+ if (isFloat)
+ spvOp = spv::OpGroupFMin;
+ else {
+ if (isUnsigned)
+ spvOp = spv::OpGroupUMin;
+ else
+ spvOp = spv::OpGroupSMin;
+ }
+ } else if (op == glslang::EOpMaxInvocations) {
+ if (isFloat)
+ spvOp = spv::OpGroupFMax;
+ else {
+ if (isUnsigned)
+ spvOp = spv::OpGroupUMax;
+ else
+ spvOp = spv::OpGroupSMax;
+ }
+ } else {
+ if (isFloat)
+ spvOp = spv::OpGroupFAdd;
+ else
+ spvOp = spv::OpGroupIAdd;
+ }
+
+ return builder.createOp(spvOp, typeId, operands);
+ }
+ case glslang::EOpMinInvocationsNonUniform:
+ case glslang::EOpMaxInvocationsNonUniform:
+ case glslang::EOpAddInvocationsNonUniform:
+ {
+ spv::Op spvOp = spv::OpNop;
+ if (op == glslang::EOpMinInvocationsNonUniform) {
+ if (isFloat)
+ spvOp = spv::OpGroupFMinNonUniformAMD;
+ else {
+ if (isUnsigned)
+ spvOp = spv::OpGroupUMinNonUniformAMD;
+ else
+ spvOp = spv::OpGroupSMinNonUniformAMD;
+ }
+ }
+ else if (op == glslang::EOpMaxInvocationsNonUniform) {
+ if (isFloat)
+ spvOp = spv::OpGroupFMaxNonUniformAMD;
+ else {
+ if (isUnsigned)
+ spvOp = spv::OpGroupUMaxNonUniformAMD;
+ else
+ spvOp = spv::OpGroupSMaxNonUniformAMD;
+ }
+ }
+ else {
+ if (isFloat)
+ spvOp = spv::OpGroupFAddNonUniformAMD;
+ else
+ spvOp = spv::OpGroupIAddNonUniformAMD;
+ }
+
+ return builder.createOp(spvOp, typeId, operands);
+ }
+#endif
default:
logger->missingFunctionality("invocation operation");
return spv::NoResult;
@@ -3798,6 +3925,7 @@
bool isFloat = typeProxy == glslang::EbtFloat || typeProxy == glslang::EbtDouble;
spv::Op opCode = spv::OpNop;
+ int extBuiltins = -1;
int libCall = -1;
size_t consumedOperands = operands.size();
spv::Id typeId0 = 0;
@@ -3941,6 +4069,60 @@
libCall = spv::GLSLstd450Bad;
break;
+#ifdef AMD_EXTENSIONS
+ case glslang::EOpSwizzleInvocations:
+ extBuiltins = getExtBuiltins(spv::E_SPV_AMD_shader_ballot);
+ libCall = spv::SwizzleInvocationsAMD;
+ break;
+ case glslang::EOpSwizzleInvocationsMasked:
+ extBuiltins = getExtBuiltins(spv::E_SPV_AMD_shader_ballot);
+ libCall = spv::SwizzleInvocationsMaskedAMD;
+ break;
+ case glslang::EOpWriteInvocation:
+ extBuiltins = getExtBuiltins(spv::E_SPV_AMD_shader_ballot);
+ libCall = spv::WriteInvocationAMD;
+ break;
+
+ case glslang::EOpMin3:
+ extBuiltins = getExtBuiltins(spv::E_SPV_AMD_shader_trinary_minmax);
+ if (isFloat)
+ libCall = spv::FMin3AMD;
+ else {
+ if (isUnsigned)
+ libCall = spv::UMin3AMD;
+ else
+ libCall = spv::SMin3AMD;
+ }
+ break;
+ case glslang::EOpMax3:
+ extBuiltins = getExtBuiltins(spv::E_SPV_AMD_shader_trinary_minmax);
+ if (isFloat)
+ libCall = spv::FMax3AMD;
+ else {
+ if (isUnsigned)
+ libCall = spv::UMax3AMD;
+ else
+ libCall = spv::SMax3AMD;
+ }
+ break;
+ case glslang::EOpMid3:
+ extBuiltins = getExtBuiltins(spv::E_SPV_AMD_shader_trinary_minmax);
+ if (isFloat)
+ libCall = spv::FMid3AMD;
+ else {
+ if (isUnsigned)
+ libCall = spv::UMid3AMD;
+ else
+ libCall = spv::SMid3AMD;
+ }
+ break;
+
+ case glslang::EOpInterpolateAtVertex:
+ extBuiltins = getExtBuiltins(spv::E_SPV_AMD_shader_explicit_vertex_parameter);
+ libCall = spv::InterpolateAtVertexAMD;
+ break;
+#endif
+
default:
return 0;
}
@@ -3951,7 +4133,7 @@
// Construct the call arguments, without modifying the original operands vector.
// We might need the remaining arguments, e.g. in the EOpFrexp case.
std::vector<spv::Id> callArguments(operands.begin(), operands.begin() + consumedOperands);
- id = builder.createBuiltinCall(typeId, stdBuiltins, libCall, callArguments);
+ id = builder.createBuiltinCall(typeId, extBuiltins >= 0 ? extBuiltins : stdBuiltins, libCall, callArguments);
} else {
switch (consumedOperands) {
case 0:
@@ -3997,8 +4179,8 @@
return builder.setPrecision(id, precision);
}
-// Intrinsics with no arguments, no return value, and no precision.
-spv::Id TGlslangToSpvTraverser::createNoArgOperation(glslang::TOperator op)
+// Intrinsics with no arguments (or no return value, and no precision).
+spv::Id TGlslangToSpvTraverser::createNoArgOperation(glslang::TOperator op, spv::Decoration precision, spv::Id typeId)
{
// TODO: get the barrier operands correct
@@ -4045,6 +4227,14 @@
// Control barrier with non-"None" semantic is also a memory barrier.
builder.createControlBarrier(spv::ScopeWorkgroup, spv::ScopeWorkgroup, spv::MemorySemanticsWorkgroupMemoryMask);
return 0;
+#ifdef AMD_EXTENSIONS
+ case glslang::EOpTime:
+ {
+ std::vector<spv::Id> args; // Dummy arguments
+ spv::Id id = builder.createBuiltinCall(typeId, getExtBuiltins(spv::E_SPV_AMD_gcn_shader), spv::TimeAMD, args);
+ return builder.setPrecision(id, precision);
+ }
+#endif
default:
logger->missingFunctionality("unknown operation with no arguments");
return 0;
@@ -4437,6 +4627,20 @@
return builder.createOp(spv::OpPhi, boolTypeId, phiOperands);
}
+// Return type Id of the imported set of extended instructions corresponds to the name.
+// Import this set if it has not been imported yet.
+spv::Id TGlslangToSpvTraverser::getExtBuiltins(const char* name)
+{
+ if (extBuiltinMap.find(name) != extBuiltinMap.end())
+ return extBuiltinMap[name];
+ else {
+ builder.addExtensions(name);
+ spv::Id extBuiltins = builder.import(name);
+ extBuiltinMap[name] = extBuiltins;
+ return extBuiltins;
+ }
+}
+
}; // end anonymous namespace
namespace glslang {