Skip to content

Commit

Permalink
Remove the check for major>6 (>=8)
Browse files Browse the repository at this point in the history 
Remove the conditon of major > 6 or major >= 8.
Considering `supportDpp()` is always true, remove the check for
`supportDpp`.
  • Loading branch information
@xuechen417
xuechen417 committed Nov 2, 2023
1 parent 51a4c5c commit feb8bd5
Show file tree
Hide file tree
Showing 7 changed files with 274 additions and 523 deletions.
6 changes: 0 additions & 6 deletions lgc/builder/BuilderImpl.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -245,12 +245,6 @@ Value *BuilderImpl::CreateIntegerDotProduct(Value *vector1, Value *vector2, Valu
return computedResult;
}

// =====================================================================================================================
// Get whether the context we are building in supports DPP operations.
bool BuilderImpl::supportDpp() const {
return getPipelineState()->getTargetInfo().getGfxIpVersion().major >= 8;
}

// =====================================================================================================================
// Get whether the context we are building in supports DPP ROW_XMASK operations.
bool BuilderImpl::supportDppRowXmask() const {
Expand Down
7 changes: 2 additions & 5 deletions lgc/builder/DescBuilder.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -389,11 +389,8 @@ Value *BuilderImpl::getDescPtr(ResourceNodeType concreteType, const ResourceNode
// @param isNonUniform : Whether value is marked as non-uniform
Value *BuilderImpl::scalarizeIfUniform(Value *value, bool isNonUniform) {
assert(value->getType()->isIntegerTy(32));
if (!isNonUniform && !isa<Constant>(value)) {
// NOTE: GFX6 encounters GPU hang with this optimization enabled. So we should skip it.
if (getPipelineState()->getTargetInfo().getGfxIpVersion().major > 6)
value = CreateIntrinsic(getInt32Ty(), Intrinsic::amdgcn_readfirstlane, value);
}
if (!isNonUniform && !isa<Constant>(value))
value = CreateIntrinsic(getInt32Ty(), Intrinsic::amdgcn_readfirstlane, value);
return value;
}

Expand Down
102 changes: 35 additions & 67 deletions lgc/builder/MiscBuilder.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -173,73 +173,41 @@ Instruction *BuilderImpl::CreateReadClock(bool realtime, const Twine &instName)
Value *BuilderImpl::CreateDerivative(Value *value, bool isDirectionY, bool isFine, const Twine &instName) {
unsigned tableIdx = isDirectionY * 2 + isFine;
Value *result = nullptr;
if (supportDpp()) {
// DPP (GFX8+) version.
// For quad pixels, quad_perm:[pix0,pix1,pix2,pix3] = [0,1,2,3]
// Table of first dpp_ctrl, in order coarseX, fineX, coarseY, fineY
static const unsigned FirstDppCtrl[4] = {
0x55, // CoarseX: [0,1,2,3] -> [1,1,1,1]
0xF5, // FineX: [0,1,2,3]->[1,1,3,3]
0xAA, // CoarseY: [0,1,2,3] -> [2,2,2,2]
0xEE, // FineY: [0,1,2,3]->[2,3,2,3]
};
// Table of second dpp_ctrl, in order coarseX, fineX, coarseY, fineY
static const unsigned SecondDppCtrl[4] = {
0x00, // CoarseX: [0,1,2,3]->[0,0,0,0]
0xA0, // FineX: [0,1,2,3]->[0,0,2,2]
0x00, // CoarseY: [0,1,2,3]->[0,0,0,0]
0x44, // FineY: [0,1,2,3]->[0,1,0,1]
};
unsigned perm1 = FirstDppCtrl[tableIdx];
unsigned perm2 = SecondDppCtrl[tableIdx];
result = scalarize(value, [this, perm1, perm2](Value *value) {
Type *valTy = value->getType();
value = CreateBitCast(value, getIntNTy(valTy->getPrimitiveSizeInBits()));
value = CreateZExtOrTrunc(value, getInt32Ty());
Value *firstVal = CreateIntrinsic(Intrinsic::amdgcn_mov_dpp, getInt32Ty(),
{value, getInt32(perm1), getInt32(15), getInt32(15), getTrue()});
firstVal = CreateZExtOrTrunc(firstVal, getIntNTy(valTy->getPrimitiveSizeInBits()));
firstVal = CreateBitCast(firstVal, valTy);
Value *secondVal = CreateIntrinsic(Intrinsic::amdgcn_mov_dpp, getInt32Ty(),
{value, getInt32(perm2), getInt32(15), getInt32(15), getTrue()});
secondVal = CreateZExtOrTrunc(secondVal, getIntNTy(valTy->getPrimitiveSizeInBits()));
secondVal = CreateBitCast(secondVal, valTy);
Value *result = CreateFSub(firstVal, secondVal);
return CreateUnaryIntrinsic(Intrinsic::amdgcn_wqm, result);
});
} else {
// ds_swizzle (pre-GFX8) version

// Table of first swizzle control, in order coarseX, fineX, coarseY, fineY
static const unsigned FirstSwizzleCtrl[4] = {
0x8055, // CoarseX: Broadcast channel 1 to whole quad
0x80F5, // FineX: Swizzle channels in quad (1 -> 0, 1 -> 1, 3 -> 2, 3 -> 3)
0x80AA, // CoarseY: Broadcast channel 2 to whole quad
0x80EE, // FineY: Swizzle channels in quad (2 -> 0, 3 -> 1, 2 -> 2, 3 -> 3)
};
// Table of second swizzle control, in order coarseX, fineX, coarseY, fineY
static const unsigned SecondSwizzleCtrl[4] = {
0x8000, // CoarseX: Broadcast channel 0 to whole quad
0x80A0, // FineX: Swizzle channels in quad (0 -> 0, 0 -> 1, 2 -> 2, 2 -> 3)
0x8000, // CoarseY: Broadcast channel 0 to whole quad
0x8044, // FineY: Swizzle channels in quad (0 -> 0, 1 -> 1, 0 -> 2, 1 -> 3)
};
unsigned perm1 = FirstSwizzleCtrl[tableIdx];
unsigned perm2 = SecondSwizzleCtrl[tableIdx];
result = scalarize(value, [this, perm1, perm2](Value *value) {
Type *valTy = value->getType();
value = CreateBitCast(value, getIntNTy(valTy->getPrimitiveSizeInBits()));
value = CreateZExtOrTrunc(value, getInt32Ty());
Value *firstVal = CreateIntrinsic(Intrinsic::amdgcn_ds_swizzle, {}, {value, getInt32(perm1)});
firstVal = CreateZExtOrTrunc(firstVal, getIntNTy(valTy->getPrimitiveSizeInBits()));
firstVal = CreateBitCast(firstVal, valTy);
Value *secondVal = CreateIntrinsic(Intrinsic::amdgcn_ds_swizzle, {}, {value, getInt32(perm2)});
secondVal = CreateZExtOrTrunc(secondVal, getIntNTy(valTy->getPrimitiveSizeInBits()));
secondVal = CreateBitCast(secondVal, valTy);
Value *result = CreateFSub(firstVal, secondVal);
return CreateUnaryIntrinsic(Intrinsic::amdgcn_wqm, result);
});
}

// DPP (GFX9+) version.
// For quad pixels, quad_perm:[pix0,pix1,pix2,pix3] = [0,1,2,3]
// Table of first dpp_ctrl, in order coarseX, fineX, coarseY, fineY
static const unsigned FirstDppCtrl[4] = {
0x55, // CoarseX: [0,1,2,3] -> [1,1,1,1]
0xF5, // FineX: [0,1,2,3]->[1,1,3,3]
0xAA, // CoarseY: [0,1,2,3] -> [2,2,2,2]
0xEE, // FineY: [0,1,2,3]->[2,3,2,3]
};
// Table of second dpp_ctrl, in order coarseX, fineX, coarseY, fineY
static const unsigned SecondDppCtrl[4] = {
0x00, // CoarseX: [0,1,2,3]->[0,0,0,0]
0xA0, // FineX: [0,1,2,3]->[0,0,2,2]
0x00, // CoarseY: [0,1,2,3]->[0,0,0,0]
0x44, // FineY: [0,1,2,3]->[0,1,0,1]
};
unsigned perm1 = FirstDppCtrl[tableIdx];
unsigned perm2 = SecondDppCtrl[tableIdx];
result = scalarize(value, [this, perm1, perm2](Value *value) {
Type *valTy = value->getType();
value = CreateBitCast(value, getIntNTy(valTy->getPrimitiveSizeInBits()));
value = CreateZExtOrTrunc(value, getInt32Ty());
Value *firstVal = CreateIntrinsic(Intrinsic::amdgcn_mov_dpp, getInt32Ty(),
{value, getInt32(perm1), getInt32(15), getInt32(15), getTrue()});
firstVal = CreateZExtOrTrunc(firstVal, getIntNTy(valTy->getPrimitiveSizeInBits()));
firstVal = CreateBitCast(firstVal, valTy);
Value *secondVal = CreateIntrinsic(Intrinsic::amdgcn_mov_dpp, getInt32Ty(),
{value, getInt32(perm2), getInt32(15), getInt32(15), getTrue()});
secondVal = CreateZExtOrTrunc(secondVal, getIntNTy(valTy->getPrimitiveSizeInBits()));
secondVal = CreateBitCast(secondVal, valTy);
Value *result = CreateFSub(firstVal, secondVal);
return CreateUnaryIntrinsic(Intrinsic::amdgcn_wqm, result);
});

result->setName(instName);
return result;
}
Loading

0 comments on commit feb8bd5

Please sign in to comment.