Skip to content

[SYCL][ESIMD] Add support for addc and subb operations #8758

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 9 commits into from
Mar 30, 2023
Merged

[SYCL][ESIMD] Add support for addc and subb operations #8758

Show file tree
Hide file tree
Changes from 3 commits
Commits
Show all changes
9 commits
Select commit Hold shift + click to select a range
8bebcfc
Initial version
fineg74 Mar 23, 2023
c5d83bb
Fix build crash
fineg74 Mar 23, 2023
40debc1
Add support for addc and subb
fineg74 Mar 23, 2023
0b5a6bd
Remove unnecessary changes
fineg74 Mar 23, 2023
59398ea
Address PR comments
fineg74 Mar 24, 2023
fe01d7c
Merge remote-tracking branch 'origin/sycl' into addc
fineg74 Mar 24, 2023
9051181
Address PR comments
fineg74 Mar 25, 2023
89bb384
Merge remote-tracking branch 'origin/sycl' into addc
fineg74 Mar 29, 2023
bc9b818
Move e2e tests from test repository
fineg74 Mar 29, 2023
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
68 changes: 54 additions & 14 deletions llvm/lib/SYCLLowerIR/ESIMD/LowerESIMD.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -663,12 +663,21 @@ class ESIMDIntrinDescTable {
{"test.src.tmpl.arg", {t(0), t1(1), t8(2), t16(3), t32(4), c8(17)}}},
{"slm_init", {"slm.init", {a(0)}}},
{"bf_cvt", {"bf.cvt", {a(0)}}},
{"tf32_cvt", {"tf32.cvt", {a(0)}}}};
{"tf32_cvt", {"tf32.cvt", {a(0)}}},
{"addc", {"addc", {l(0)}}},
{"subb", {"subb", {l(0)}}}};
}

const IntrinTable &getTable() { return Table; }
};

static bool isStructureReturningFunction(StringRef FunctionName) {
return llvm::StringSwitch<bool>(FunctionName)
.Case("addc", true)
.Case("subb", true)
.Default(false);
}

// The C++11 "magic static" idiom to lazily initialize the ESIMD intrinsic table
static const IntrinTable &getIntrinTable() {
static ESIMDIntrinDescTable TheTable;
Expand Down Expand Up @@ -1409,6 +1418,8 @@ static void translateESIMDIntrinsicCall(CallInst &CI) {
SmallVector<Value *, 16> GenXArgs;
createESIMDIntrinsicArgs(Desc, GenXArgs, CI, FE);
Function *NewFDecl = nullptr;
bool IsStructureReturningFunction =
isStructureReturningFunction(Desc.GenXSpelling);
if (Desc.GenXSpelling.rfind("test.src.", 0) == 0) {
// Special case for testing purposes
NewFDecl = createTestESIMDDeclaration(Desc, GenXArgs, CI);
Expand All @@ -1417,12 +1428,21 @@ static void translateESIMDIntrinsicCall(CallInst &CI) {
GenXIntrinsic::getGenXIntrinsicPrefix() + Desc.GenXSpelling + Suffix);

SmallVector<Type *, 16> GenXOverloadedTypes;
if (GenXIntrinsic::isOverloadedRet(ID))
GenXOverloadedTypes.push_back(CI.getType());
if (GenXIntrinsic::isOverloadedRet(ID)) {
if (IsStructureReturningFunction) {
// TODO implement more generic handling of returned structure
// current code assumes that returned code has 2 members of the
// same type as arguments.
GenXOverloadedTypes.push_back(GenXArgs[1]->getType());
GenXOverloadedTypes.push_back(GenXArgs[1]->getType());
} else {
GenXOverloadedTypes.push_back(CI.getType());
}
}
for (unsigned i = 0; i < GenXArgs.size(); ++i)
if (GenXIntrinsic::isOverloadedArg(ID, i))
if (GenXIntrinsic::isOverloadedArg(ID, i)) {
GenXOverloadedTypes.push_back(GenXArgs[i]->getType());

}
NewFDecl = GenXIntrinsic::getGenXDeclaration(CI.getModule(), ID,
GenXOverloadedTypes);
}
Expand All @@ -1432,15 +1452,35 @@ static void translateESIMDIntrinsicCall(CallInst &CI) {
NewFDecl->getFnAttribute(llvm::Attribute::ReadNone).isValid();
if (FixReadNone)
NewFDecl->removeFnAttr(llvm::Attribute::ReadNone);
CallInst *NewCI = IntrinsicInst::Create(
NewFDecl, GenXArgs,
NewFDecl->getReturnType()->isVoidTy() ? "" : CI.getName() + ".esimd",
&CI);
if (FixReadNone)
NewCI->setMemoryEffects(MemoryEffects::none());
NewCI->setDebugLoc(CI.getDebugLoc());

Instruction *NewInst = addCastInstIfNeeded(&CI, NewCI);
Instruction *NewInst = nullptr;
if (IsStructureReturningFunction) {
AddrSpaceCastInst *a = static_cast<AddrSpaceCastInst *>(GenXArgs[0]);

GenXArgs.erase(GenXArgs.begin());
CallInst *NewCI = IntrinsicInst::Create(
NewFDecl, GenXArgs,
NewFDecl->getReturnType()->isVoidTy() ? "" : CI.getName() + ".esimd",
&CI);
if (FixReadNone)
NewCI->setMemoryEffects(MemoryEffects::none());
NewCI->setDebugLoc(CI.getDebugLoc());

IRBuilder<> Builder(&CI);

NewInst = Builder.CreateStore(
NewCI, Builder.CreateBitCast(a->getPointerOperand(),
NewCI->getType()->getPointerTo()));
} else {
CallInst *NewCI = IntrinsicInst::Create(
NewFDecl, GenXArgs,
NewFDecl->getReturnType()->isVoidTy() ? "" : CI.getName() + ".esimd",
&CI);
if (FixReadNone)
NewCI->setMemoryEffects(MemoryEffects::none());
NewCI->setDebugLoc(CI.getDebugLoc());

NewInst = addCastInstIfNeeded(&CI, NewCI);
}
CI.replaceAllUsesWith(NewInst);
CI.eraseFromParent();
}
Expand Down
41 changes: 41 additions & 0 deletions sycl/include/sycl/ext/intel/experimental/esimd/detail/math_intrin.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -714,6 +714,47 @@ __esimd_dpasw_nosrc0(__ESIMD_DNS::vector_type_t<T1, N1> src1,
}
#endif // !__SYCL_DEVICE_ONLY__

template <typename T, int N>
__ESIMD_INTRIN std::pair<__ESIMD_DNS::vector_type_t<T, N>,
__ESIMD_DNS::vector_type_t<T, N>>
__esimd_addc(__ESIMD_DNS::vector_type_t<T, N> src0,
__ESIMD_DNS::vector_type_t<T, N> src1)
#ifdef __SYCL_DEVICE_ONLY__
;
#else // !__SYCL_DEVICE_ONLY__
{
__ESIMD_NS::simd<uint64_t, N> Result64 = __ESIMD_NS::simd<T, N>(src0);
Result64 += __ESIMD_NS::simd<T, N>(src1);
auto Result32 = Result64.template bit_cast_view<T>();
__ESIMD_NS::simd<uint32_t, N> CarryV = Result32.template select<N, 2>(1);
__ESIMD_NS::simd<uint32_t, N> ResV = Result32.template select<N, 2>(0);
std::pair<__ESIMD_DNS::vector_type_t<T, N>, __ESIMD_DNS::vector_type_t<T, N>>
ReturnValue = std::make_pair(CarryV.data(), ResV.data());
return ReturnValue;
}
#endif // !__SYCL_DEVICE_ONLY__

template <typename T, int N>
__ESIMD_INTRIN std::pair<__ESIMD_DNS::vector_type_t<T, N>,
__ESIMD_DNS::vector_type_t<T, N>>
__esimd_subb(__ESIMD_DNS::vector_type_t<T, N> src0,
__ESIMD_DNS::vector_type_t<T, N> src1)
#ifdef __SYCL_DEVICE_ONLY__
;
#else // !__SYCL_DEVICE_ONLY__
{
__ESIMD_NS::simd<uint64_t, N> Result64 = __ESIMD_NS::simd<T, N>(src0);
Result64 -= __ESIMD_NS::simd<T, N>(src1);
auto Result32 = Result64.template bit_cast_view<T>();
__ESIMD_NS::simd<uint32_t, N> BorrowV =
__ESIMD_NS::simd<T, N>(src0) < __ESIMD_NS::simd<T, N>(src1);
__ESIMD_NS::simd<uint32_t, N> ResV = Result32.template select<N, 2>(0);
std::pair<__ESIMD_DNS::vector_type_t<T, N>, __ESIMD_DNS::vector_type_t<T, N>>
ReturnValue = std::make_pair(BorrowV.data(), ResV.data());
return ReturnValue;
}
#endif // !__SYCL_DEVICE_ONLY__

#undef __ESIMD_raw_vec_t
#undef __ESIMD_cpp_vec_t

Expand Down
74 changes: 74 additions & 0 deletions sycl/include/sycl/ext/intel/experimental/esimd/math.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -496,6 +496,80 @@ imul(T &rmd, T0 src0, T1 src1) {
return Res[0];
}

template <int N>
__ESIMD_API __ESIMD_NS::simd<uint32_t, N>
addc(__ESIMD_NS::simd<uint32_t, N> &carry, __ESIMD_NS::simd<uint32_t, N> src0,
__ESIMD_NS::simd<uint32_t, N> src1) {
std::pair<__ESIMD_DNS::vector_type_t<uint32_t, N>,
__ESIMD_DNS::vector_type_t<uint32_t, N>>
Result = __esimd_addc<uint32_t, N>(src0.data(), src1.data());

carry = Result.first;
return Result.second;
}

template <int N>
__ESIMD_API __ESIMD_NS::simd<uint32_t, N>
addc(__ESIMD_NS::simd<uint32_t, N> &carry, __ESIMD_NS::simd<uint32_t, N> src0,
uint32_t src1) {
__ESIMD_NS::simd<uint32_t, N> Src1V = src1;
return addc(carry, src0, Src1V);
}

template <int N>
__ESIMD_API __ESIMD_NS::simd<uint32_t, N>
addc(__ESIMD_NS::simd<uint32_t, N> &carry, uint32_t src0,
__ESIMD_NS::simd<uint32_t, N> src1) {
__ESIMD_NS::simd<uint32_t, N> Src0V = src0;
return addc(carry, Src0V, src1);
}

__ESIMD_API uint32_t addc(uint32_t &carry, uint32_t src0, uint32_t src1) {
__ESIMD_NS::simd<uint32_t, 1> CarryV = carry;
__ESIMD_NS::simd<uint32_t, 1> Src0V = src0;
__ESIMD_NS::simd<uint32_t, 1> Src1V = src1;
__ESIMD_NS::simd<uint32_t, 1> Res = addc(CarryV, Src0V, Src1V);
carry = CarryV[0];
return Res[0];
}

template <int N>
__ESIMD_API __ESIMD_NS::simd<uint32_t, N>
subb(__ESIMD_NS::simd<uint32_t, N> &borrow, __ESIMD_NS::simd<uint32_t, N> src0,
__ESIMD_NS::simd<uint32_t, N> src1) {
std::pair<__ESIMD_DNS::vector_type_t<uint32_t, N>,
__ESIMD_DNS::vector_type_t<uint32_t, N>>
Result = __esimd_subb<uint32_t, N>(src0.data(), src1.data());

borrow = Result.first;
return Result.second;
}

template <int N>
__ESIMD_API __ESIMD_NS::simd<uint32_t, N>
subb(__ESIMD_NS::simd<uint32_t, N> &borrow, __ESIMD_NS::simd<uint32_t, N> src0,
uint32_t src1) {
__ESIMD_NS::simd<uint32_t, N> Src1V = src1;
return subb(borrow, src0, Src1V);
}

template <int N>
__ESIMD_API __ESIMD_NS::simd<uint32_t, N>
subb(__ESIMD_NS::simd<uint32_t, N> &borrow, uint32_t src0,
__ESIMD_NS::simd<uint32_t, N> src1) {
__ESIMD_NS::simd<uint32_t, N> Src0V = src0;
return subb(borrow, Src0V, src1);
}

__ESIMD_API uint32_t subb(uint32_t &borrow, uint32_t src0, uint32_t src1) {
__ESIMD_NS::simd<uint32_t, 1> BorrowV = borrow;
__ESIMD_NS::simd<uint32_t, 1> Src0V = src0;
__ESIMD_NS::simd<uint32_t, 1> Src1V = src1;
__ESIMD_NS::simd<uint32_t, 1> Res = subb(BorrowV, Src0V, Src1V);
borrow = BorrowV[0];
return Res[0];
}

/// Integral quotient (vector version)
/// @tparam T element type of the input and return vectors.
/// @tparam SZ size of the input and returned vectors.
Expand Down