GitHub-Proxy/eigen
1
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2025-04-22 09:39:34 +08:00
Code Issues Packages Projects Releases Wiki Activity

Revert "Fix rint for SSE/NEON."

Browse Source 
This reverts commit e72dfeb8b9fa5662831b5d0bb9d132521f9173dd
This commit is contained in:
Antonio Sánchez 2021-03-03 18:51:51 +00:00
parent e72dfeb8b9
commit 9a663973b4
4 changed files with 29 additions and 83 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

34
Eigen/src/Core/arch/NEON/PacketMath.h
View File

@ -3207,34 +3207,20 @@ template<> EIGEN_STRONG_INLINE Packet4f pceil<Packet4f>(const Packet4f& a)
template<> EIGEN_STRONG_INLINE Packet4f print(const Packet4f& a) { template<> EIGEN_STRONG_INLINE Packet4f print(const Packet4f& a) {
// Adds and subtracts signum(a) * 2^23 to force rounding. // Adds and subtracts signum(a) * 2^23 to force rounding.
const Packet4f limit = pset1<Packet4f>(static_cast<float>(1<<23)); const Packet4f offset =
const Packet4f abs_a = pabs(a); pselect(pcmp_lt(a, pzero(a)),
// Inline asm to prevent the compiler from optimizing away the pset1<Packet4f>(-static_cast<float>(1<<23)),
// addition and subtraction. pset1<Packet4f>(+static_cast<float>(1<<23)));
// Packet4f r = psub(padd(abs_a, limit), limit); return psub(padd(a, offset), offset);
Packet4f r = abs_a;
__asm__ ("vadd.f32 %[r], %[r], %[limit]\n\t"
"vsub.f32 %[r], %[r], %[limit]" : [r] "+x" (r) : [limit] "x" (limit));
// If greater than limit, simply return a. Otherwise, account for sign.
r = pselect(pcmp_lt(abs_a, limit),
pselect(pcmp_lt(a, pzero(a)), pnegate(r), r), a);
return r;
} }
template<> EIGEN_STRONG_INLINE Packet2f print(const Packet2f& a) { template<> EIGEN_STRONG_INLINE Packet2f print(const Packet2f& a) {
// Adds and subtracts signum(a) * 2^23 to force rounding. // Adds and subtracts signum(a) * 2^23 to force rounding.
const Packet2f limit = pset1<Packet2f>(static_cast<float>(1<<23)); const Packet2f offset =
const Packet2f abs_a = pabs(a); pselect(pcmp_lt(a, pzero(a)),
// Inline asm to prevent the compiler from optimizing away the pset1<Packet2f>(-static_cast<float>(1<<23)),
// addition and subtraction. pset1<Packet2f>(+static_cast<float>(1<<23)));
// Packet4f r = psub(padd(abs_a, limit), limit); return psub(padd(a, offset), offset);
Packet2f r = abs_a;
__asm__ ("vadd.f32 %[r], %[r], %[limit]\n\t"
"vsub.f32 %[r], %[r], %[limit]" : [r] "+x" (r) : [limit] "x" (limit));
// If greater than limit, simply return a. Otherwise, account for sign.
r = pselect(pcmp_lt(abs_a, limit),
pselect(pcmp_lt(a, pzero(a)), pnegate(r), r), a);
return r;
} }
template<> EIGEN_STRONG_INLINE Packet4f pfloor<Packet4f>(const Packet4f& a) template<> EIGEN_STRONG_INLINE Packet4f pfloor<Packet4f>(const Packet4f& a)

35
Eigen/src/Core/arch/SSE/PacketMath.h
View File

@ -646,35 +646,20 @@ template<> EIGEN_STRONG_INLINE Packet2d pfloor<Packet2d>(const Packet2d& a) { re
#else #else
template<> EIGEN_STRONG_INLINE Packet4f print(const Packet4f& a) { template<> EIGEN_STRONG_INLINE Packet4f print(const Packet4f& a) {
// Adds and subtracts signum(a) * 2^23 to force rounding. // Adds and subtracts signum(a) * 2^23 to force rounding.
const Packet4f limit = pset1<Packet4f>(static_cast<float>(1<<23)); const Packet4f offset =
const Packet4f abs_a = pabs(a); pselect(pcmp_lt(a, pzero(a)),
// Inline asm to prevent the compiler from optimizing away the pset1<Packet4f>(-static_cast<float>(1<<23)),
// addition and subtraction. pset1<Packet4f>(+static_cast<float>(1<<23)));
// Packet4f r = psub(padd(abs_a, limit), limit); return psub(padd(a, offset), offset);
Packet4f r = abs_a;
__asm__ ("addps %[limit], %[r]\n\t"
"subps %[limit], %[r]" : [r] "+x" (r) : [limit] "x" (limit));
// If greater than limit, simply return a. Otherwise, account for sign.
r = pselect(pcmp_lt(abs_a, limit),
pselect(pcmp_lt(a, pzero(a)), pnegate(r), r), a);
return r;
} }
template<> EIGEN_STRONG_INLINE Packet2d print(const Packet2d& a) { template<> EIGEN_STRONG_INLINE Packet2d print(const Packet2d& a) {
// Adds and subtracts signum(a) * 2^52 to force rounding. // Adds and subtracts signum(a) * 2^52 to force rounding.
const Packet2d limit = pset1<Packet2d>(static_cast<double>(1ull<<52)); const Packet2d offset =
const Packet2d abs_a = pabs(a); pselect(pcmp_lt(a, pzero(a)),
// Inline asm to prevent the compiler from optimizing away the pset1<Packet2d>(-static_cast<double>(1ull<<52)),
// addition and subtraction. pset1<Packet2d>(+static_cast<double>(1ull<<52)));
// Packet2d r = psub(padd(abs_a, limit), limit); return psub(padd(a, offset), offset);
Packet2d r = abs_a;
asm("addpd %[limit], %[r] \n\t"
"subpd %[limit], %[r]" : [r] "+x" (r) : [limit] "x" (limit));
// If greater than limit, simply return a. Otherwise, account for sign.
r = pselect(pcmp_lt(abs_a, limit),
pselect(pcmp_lt(a, pzero(a)), pnegate(r), r), a);
return r;
} }
template<> EIGEN_STRONG_INLINE Packet4f pfloor<Packet4f>(const Packet4f& a) template<> EIGEN_STRONG_INLINE Packet4f pfloor<Packet4f>(const Packet4f& a)

18
test/packetmath.cpp
View File

@ -543,10 +543,10 @@ void packetmath_real() {
CHECK_CWISE1_IF(PacketTraits::HasCos, std::cos, internal::pcos); CHECK_CWISE1_IF(PacketTraits::HasCos, std::cos, internal::pcos);
CHECK_CWISE1_IF(PacketTraits::HasTan, std::tan, internal::ptan); CHECK_CWISE1_IF(PacketTraits::HasTan, std::tan, internal::ptan);
CHECK_CWISE1_EXACT_IF(PacketTraits::HasRound, numext::round, internal::pround); CHECK_CWISE1_IF(PacketTraits::HasRound, numext::round, internal::pround);
CHECK_CWISE1_EXACT_IF(PacketTraits::HasCeil, numext::ceil, internal::pceil); CHECK_CWISE1_IF(PacketTraits::HasCeil, numext::ceil, internal::pceil);
CHECK_CWISE1_EXACT_IF(PacketTraits::HasFloor, numext::floor, internal::pfloor); CHECK_CWISE1_IF(PacketTraits::HasFloor, numext::floor, internal::pfloor);
CHECK_CWISE1_EXACT_IF(PacketTraits::HasRint, numext::rint, internal::print); CHECK_CWISE1_IF(PacketTraits::HasRint, numext::rint, internal::print);
// Rounding edge cases. // Rounding edge cases.
if (PacketTraits::HasRound || PacketTraits::HasCeil || PacketTraits::HasFloor || PacketTraits::HasRint) { if (PacketTraits::HasRound || PacketTraits::HasCeil || PacketTraits::HasFloor || PacketTraits::HasRint) {
@ -583,10 +583,10 @@ void packetmath_real() {
for (size_t k=0; k<values.size(); ++k) { for (size_t k=0; k<values.size(); ++k) {
data1[0] = values[k]; data1[0] = values[k];
CHECK_CWISE1_EXACT_IF(PacketTraits::HasRound, numext::round, internal::pround); CHECK_CWISE1_IF(PacketTraits::HasRound, numext::round, internal::pround);
CHECK_CWISE1_EXACT_IF(PacketTraits::HasCeil, numext::ceil, internal::pceil); CHECK_CWISE1_IF(PacketTraits::HasCeil, numext::ceil, internal::pceil);
CHECK_CWISE1_EXACT_IF(PacketTraits::HasFloor, numext::floor, internal::pfloor); CHECK_CWISE1_IF(PacketTraits::HasFloor, numext::floor, internal::pfloor);
CHECK_CWISE1_EXACT_IF(PacketTraits::HasRint, numext::rint, internal::print); CHECK_CWISE1_IF(PacketTraits::HasRint, numext::rint, internal::print);
} }
} }
@ -644,7 +644,7 @@ void packetmath_real() {
if (PacketTraits::HasExp) { if (PacketTraits::HasExp) {
data1[0] = Scalar(-1); data1[0] = Scalar(-1);
// underflow to zero // underflow to zero
data1[PacketSize] = Scalar(std::numeric_limits<Scalar>::min_exponent-55); data1[PacketSize] = Scalar(std::numeric_limits<Scalar>::min_exponent-10);
CHECK_CWISE2_IF(PacketTraits::HasExp, REF_LDEXP, internal::pldexp); CHECK_CWISE2_IF(PacketTraits::HasExp, REF_LDEXP, internal::pldexp);
// overflow to inf // overflow to inf
data1[PacketSize] = Scalar(std::numeric_limits<Scalar>::max_exponent+10); data1[PacketSize] = Scalar(std::numeric_limits<Scalar>::max_exponent+10);

25
test/packetmath_test_shared.h
View File

@ -108,23 +108,6 @@ template<typename Scalar> bool areApprox(const Scalar* a, const Scalar* b, int s
return true; return true;
} }
template<typename Scalar> bool areEqual(const Scalar* a, const Scalar* b, int size)
{
for (int i=0; i<size; ++i)
{
if (a[i] != b[i])
{
if((numext::isnan)(a[i]) && (numext::isnan)(b[i]))
{
continue;
}
std::cout << "ref: [" << Map<const Matrix<Scalar,1,Dynamic> >(a,size) << "]" << " != vec: [" << Map<const Matrix<Scalar,1,Dynamic> >(b,size) << "]\n";
return false;
}
}
return true;
}
#define CHECK_CWISE1(REFOP, POP) { \ #define CHECK_CWISE1(REFOP, POP) { \
for (int i=0; i<PacketSize; ++i) \ for (int i=0; i<PacketSize; ++i) \
ref[i] = REFOP(data1[i]); \ ref[i] = REFOP(data1[i]); \
@ -195,14 +178,6 @@ struct packet_helper<false,Packet>
VERIFY(test::areApprox(ref, data2, PacketSize) && #POP); \ VERIFY(test::areApprox(ref, data2, PacketSize) && #POP); \
} }
#define CHECK_CWISE1_EXACT_IF(COND, REFOP, POP) if(COND) { \
test::packet_helper<COND,Packet> h; \
for (int i=0; i<PacketSize; ++i) \
ref[i] = Scalar(REFOP(data1[i])); \
h.store(data2, POP(h.load(data1))); \
VERIFY(test::areEqual(ref, data2, PacketSize) && #POP); \
}
#define CHECK_CWISE2_IF(COND, REFOP, POP) if(COND) { \ #define CHECK_CWISE2_IF(COND, REFOP, POP) if(COND) { \
test::packet_helper<COND,Packet> h; \ test::packet_helper<COND,Packet> h; \
for (int i=0; i<PacketSize; ++i) \ for (int i=0; i<PacketSize; ++i) \