mewin/glslang
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

WIP: HLSL: add f16tof32 and f32tof16 decompositions.

Browse Source
This commit is contained in:
steve-lunarg
2017-02-27 13:09:32 -07:00
parent 42e33c9bec
commit 86b510efd1
7 changed files with 1079 additions and 177 deletions
Download Patch File Download Diff File Expand all files Collapse all files

101
hlsl/hlslParseHelper.cpp
View File

@@ -3564,10 +3564,107 @@ void HlslParseContext::decomposeIntrinsic(const TSourceLoc& loc, TIntermTyped*&
}
case EOpF16tof32:
{
// input uvecN with low 16 bits of each component holding a float16. convert to float32.
TIntermTyped* argValue = node->getAsUnaryNode()->getOperand();
TIntermTyped* zero = intermediate.addConstantUnion(0.0, EbtFloat, loc, true);
const int vecSize = argValue->getType().getVectorSize();
TOperator constructOp = EOpNull;
switch (vecSize) {
case 1: constructOp = EOpNull; break; // direct use, no construct needed
case 2: constructOp = EOpConstructVec2; break;
case 3: constructOp = EOpConstructVec3; break;
case 4: constructOp = EOpConstructVec4; break;
default: assert(0); break;
}
// For scalar case, we don't need to construct another type.
TIntermAggregate* result = (vecSize > 1) ? new TIntermAggregate(constructOp) : nullptr;
if (result) {
result->setType(TType(EbtFloat, EvqTemporary, vecSize));
result->setLoc(loc);
}
for (int idx = 0; idx < vecSize; ++idx) {
TIntermTyped* idxConst = intermediate.addConstantUnion(idx, loc, true);
TIntermTyped* component = argValue->getType().isVector() ?
intermediate.addIndex(EOpIndexDirect, argValue, idxConst, loc) : argValue;
if (component != argValue)
component->setType(TType(argValue->getBasicType(), EvqTemporary));
TIntermTyped* unpackOp = new TIntermUnary(EOpUnpackHalf2x16);
unpackOp->setType(TType(EbtFloat, EvqTemporary, 2));
unpackOp->getAsUnaryNode()->setOperand(component);
unpackOp->setLoc(loc);
TIntermTyped* lowOrder = intermediate.addIndex(EOpIndexDirect, unpackOp, zero, loc);
if (result != nullptr) {
result->getSequence().push_back(lowOrder);
node = result;
} else {
node = lowOrder;
}
}
break;
}
case EOpF32tof16:
{
// Temporary until decomposition is available.
error(loc, "unimplemented intrinsic: handle natively", "f32tof16", "");
// input floatN converted to 16 bit float in low order bits of each component of uintN
TIntermTyped* argValue = node->getAsUnaryNode()->getOperand();
TIntermTyped* zero = intermediate.addConstantUnion(0.0, EbtFloat, loc, true);
const int vecSize = argValue->getType().getVectorSize();
TOperator constructOp = EOpNull;
switch (vecSize) {
case 1: constructOp = EOpNull; break; // direct use, no construct needed
case 2: constructOp = EOpConstructUVec2; break;
case 3: constructOp = EOpConstructUVec3; break;
case 4: constructOp = EOpConstructUVec4; break;
default: assert(0); break;
}
// For scalar case, we don't need to construct another type.
TIntermAggregate* result = (vecSize > 1) ? new TIntermAggregate(constructOp) : nullptr;
if (result) {
result->setType(TType(EbtUint, EvqTemporary, vecSize));
result->setLoc(loc);
}
for (int idx = 0; idx < vecSize; ++idx) {
TIntermTyped* idxConst = intermediate.addConstantUnion(idx, loc, true);
TIntermTyped* component = argValue->getType().isVector() ?
intermediate.addIndex(EOpIndexDirect, argValue, idxConst, loc) : argValue;
if (component != argValue)
component->setType(TType(argValue->getBasicType(), EvqTemporary));
TIntermAggregate* vec2ComponentAndZero = new TIntermAggregate(EOpConstructVec2);
vec2ComponentAndZero->getSequence().push_back(component);
vec2ComponentAndZero->getSequence().push_back(zero);
vec2ComponentAndZero->setType(TType(EbtFloat, EvqTemporary, 2));
vec2ComponentAndZero->setLoc(loc);
TIntermTyped* packOp = new TIntermUnary(EOpPackHalf2x16);
packOp->getAsUnaryNode()->setOperand(vec2ComponentAndZero);
packOp->setLoc(loc);
packOp->setType(TType(EbtUint, EvqTemporary));
if (result != nullptr) {
result->getSequence().push_back(packOp);
node = result;
} else {
node = packOp;
}
}
break;
}