henry/SimCore
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

ADD: added other eigen lib

Browse Source
This commit is contained in:
Henry Winkel
2022-12-21 16:19:04 +01:00
parent a570766dc6
commit 9e56c7f2c0
832 changed files with 36586 additions and 20006 deletions
Download Patch File Download Diff File Expand all files Collapse all files

267
libs/eigen/test/packetmath.cpp
View File

@@ -24,10 +24,30 @@ inline T REF_MUL(const T& a, const T& b) {
return a * b;
}
template <typename T>
inline T REF_MADD(const T& a, const T& b, const T& c) {
return a * b + c;
}
template <typename T>
inline T REF_MSUB(const T& a, const T& b, const T& c) {
return a * b - c;
}
template <typename T>
inline T REF_NMADD(const T& a, const T& b, const T& c) {
return (-a * b) + c;
}
template <typename T>
inline T REF_NMSUB(const T& a, const T& b, const T& c) {
return (-a * b) - c;
}
template <typename T>
inline T REF_DIV(const T& a, const T& b) {
return a / b;
}
template <typename T>
inline T REF_RECIPROCAL(const T& a) {
return T(1) / a;
}
template <typename T>
inline T REF_ABS_DIFF(const T& a, const T& b) {
return a > b ? a - b : b - a;
}
@@ -45,13 +65,23 @@ template <>
inline bool REF_MUL(const bool& a, const bool& b) {
return a && b;
}
template <>
inline bool REF_MADD(const bool& a, const bool& b, const bool& c) {
return (a && b) || c;
}
template <typename T>
inline T REF_FREXP(const T& x, T& exp) {
int iexp;
int iexp = 0;
EIGEN_USING_STD(frexp)
const T out = static_cast<T>(frexp(x, &iexp));
exp = static_cast<T>(iexp);
// The exponent value is unspecified if the input is inf or NaN, but MSVC
// seems to set it to 1. We need to set it back to zero for consistency.
if (!(numext::isfinite)(x)) {
exp = T(0);
}
return out;
}
@@ -235,7 +265,7 @@ struct packetmath_pcast_ops_runner {
// Only some types support cast from std::complex<>.
template <typename Scalar, typename Packet>
struct packetmath_pcast_ops_runner<Scalar, Packet, typename internal::enable_if<NumTraits<Scalar>::IsComplex>::type> {
struct packetmath_pcast_ops_runner<Scalar, Packet, std::enable_if_t<NumTraits<Scalar>::IsComplex>> {
static void run() {
test_cast_runner<Packet, std::complex<float> >::run();
test_cast_runner<Packet, std::complex<double> >::run();
@@ -353,10 +383,10 @@ template <typename Scalar, typename Packet>
void packetmath_minus_zero_add() {
const int PacketSize = internal::unpacket_traits<Packet>::size;
const int size = 2 * PacketSize;
EIGEN_ALIGN_MAX Scalar data1[size];
EIGEN_ALIGN_MAX Scalar data2[size];
EIGEN_ALIGN_MAX Scalar ref[size];
EIGEN_ALIGN_MAX Scalar data1[size] = {};
EIGEN_ALIGN_MAX Scalar data2[size] = {};
EIGEN_ALIGN_MAX Scalar ref[size] = {};
for (int i = 0; i < PacketSize; ++i) {
data1[i] = Scalar(-0.0);
data1[i + PacketSize] = Scalar(-0.0);
@@ -374,11 +404,11 @@ struct eigen_optimization_barrier_test {
};
template<typename Packet>
struct eigen_optimization_barrier_test<Packet, typename internal::enable_if<
struct eigen_optimization_barrier_test<Packet, std::enable_if_t<
!NumTraits<Packet>::IsComplex &&
!internal::is_same<Packet, Eigen::half>::value &&
!internal::is_same<Packet, Eigen::bfloat16>::value
>::type> {
>> {
static void run() {
typedef typename internal::unpacket_traits<Packet>::type Scalar;
Scalar s = internal::random<Scalar>();
@@ -428,6 +458,36 @@ void packetmath() {
VERIFY(test::areApprox(data1, data2 + offset, PacketSize) && "internal::pstoreu");
}
for (int M = 0; M < PacketSize; ++M) {
for (int N = 0; N <= PacketSize; ++N) {
for (int j = 0; j < size; ++j) {
data1[j] = internal::random<Scalar>() / RealScalar(PacketSize);
data2[j] = internal::random<Scalar>() / RealScalar(PacketSize);
refvalue = (std::max)(refvalue, numext::abs(data1[j]));
}
if (M == 0) {
internal::pstore_partial(data2, internal::pload_partial<Packet>(data1, N), N);
VERIFY(test::areApprox(data1, data2, N) && "aligned loadN/storeN");
for (int offset = 0; offset < PacketSize; ++offset) {
internal::pstore_partial(data2, internal::ploadu_partial<Packet>(data1 + offset, N), N);
VERIFY(test::areApprox(data1 + offset, data2, N) && "internal::ploadu_partial");
}
for (int offset = 0; offset < PacketSize; ++offset) {
internal::pstoreu_partial(data2 + offset, internal::pload_partial<Packet>(data1, N), N);
VERIFY(test::areApprox(data1, data2 + offset, N) && "internal::pstoreu_partial");
}
}
if (N + M > PacketSize) continue; // Don't read or write past end of Packet
internal::pstore_partial(data2, internal::pload_partial<Packet>(data1, N, M), N, M);
VERIFY(test::areApprox(data1, data2, N) && "aligned offset loadN/storeN");
}
}
if (internal::unpacket_traits<Packet>::masked_load_available) {
test::packet_helper<internal::unpacket_traits<Packet>::masked_load_available, Packet> h;
unsigned long long max_umask = (0x1ull << PacketSize);
@@ -464,8 +524,11 @@ void packetmath() {
CHECK_CWISE2_IF(PacketTraits::HasMul, REF_MUL, internal::pmul);
CHECK_CWISE2_IF(PacketTraits::HasDiv, REF_DIV, internal::pdiv);
if (PacketTraits::HasNegate) CHECK_CWISE1(internal::negate, internal::pnegate);
CHECK_CWISE1_IF(PacketTraits::HasNegate, internal::negate, internal::pnegate);
CHECK_CWISE1_IF(PacketTraits::HasReciprocal, REF_RECIPROCAL, internal::preciprocal);
CHECK_CWISE1(numext::conj, internal::pconj);
CHECK_CWISE1_IF(PacketTraits::HasSign, numext::sign, internal::psign);
for (int offset = 0; offset < 3; ++offset) {
for (int i = 0; i < PacketSize; ++i) ref[i] = data1[offset];
@@ -616,6 +679,23 @@ void packetmath() {
}
CHECK_CWISE1_IF(PacketTraits::HasSqrt, numext::sqrt, internal::psqrt);
CHECK_CWISE1_IF(PacketTraits::HasRsqrt, numext::rsqrt, internal::prsqrt);
CHECK_CWISE3_IF(true, REF_MADD, internal::pmadd);
if (!std::is_same<Scalar, bool>::value && NumTraits<Scalar>::IsSigned) {
CHECK_CWISE3_IF(true, REF_NMSUB, internal::pnmsub);
}
// For pmsub, pnmadd, the values can cancel each other to become near zero,
// which can lead to very flaky tests. Here we ensure the signs are such that
// they do not cancel.
for (int i = 0; i < PacketSize; ++i) {
data1[i] = numext::abs(internal::random<Scalar>());
data1[i + PacketSize] = numext::abs(internal::random<Scalar>());
data1[i + 2 * PacketSize] = -numext::abs(internal::random<Scalar>());
}
if (!std::is_same<Scalar, bool>::value && NumTraits<Scalar>::IsSigned) {
CHECK_CWISE3_IF(true, REF_MSUB, internal::pmsub);
CHECK_CWISE3_IF(true, REF_NMADD, internal::pnmadd);
}
}
// Notice that this definition works for complex types as well.
@@ -625,15 +705,104 @@ Scalar log2(Scalar x) {
return Scalar(EIGEN_LOG2E) * std::log(x);
}
// Create a functor out of a function so it can be passed (with overloads)
// to another function as an input argument.
#define CREATE_FUNCTOR(Name, Func) \
struct Name { \
template<typename T> \
T operator()(const T& val) const { \
return Func(val); \
} \
}
CREATE_FUNCTOR(psqrt_functor, internal::psqrt);
CREATE_FUNCTOR(prsqrt_functor, internal::prsqrt);
// TODO(rmlarsen): Run this test for more functions.
template <bool Cond, typename Scalar, typename Packet, typename RefFunctorT, typename FunctorT>
void packetmath_test_IEEE_corner_cases(const RefFunctorT& ref_fun,
const FunctorT& fun) {
const int PacketSize = internal::unpacket_traits<Packet>::size;
const Scalar norm_min = (std::numeric_limits<Scalar>::min)();
const Scalar norm_max = (std::numeric_limits<Scalar>::max)();
constexpr int size = PacketSize * 2;
EIGEN_ALIGN_MAX Scalar data1[size];
EIGEN_ALIGN_MAX Scalar data2[size];
EIGEN_ALIGN_MAX Scalar ref[size];
for (int i = 0; i < size; ++i) {
data1[i] = data2[i] = ref[i] = Scalar(0);
}
// Test for subnormals.
if (Cond && std::numeric_limits<Scalar>::has_denorm == std::denorm_present) {
for (int scale = 1; scale < 5; ++scale) {
// When EIGEN_FAST_MATH is 1 we relax the conditions slightly, and allow the function
// to return the same value for subnormals as the reference would return for zero with
// the same sign as the input.
#if EIGEN_FAST_MATH
data1[0] = Scalar(scale) * std::numeric_limits<Scalar>::denorm_min();
data1[1] = -data1[0];
test::packet_helper<Cond, Packet> h;
h.store(data2, fun(h.load(data1)));
for (int i=0; i < PacketSize; ++i) {
const Scalar ref_zero = ref_fun(data1[i] < 0 ? -Scalar(0) : Scalar(0));
const Scalar ref_val = ref_fun(data1[i]);
VERIFY(((std::isnan)(data2[i]) && (std::isnan)(ref_val)) || data2[i] == ref_zero ||
verifyIsApprox(data2[i], ref_val));
}
#else
CHECK_CWISE1_IF(Cond, ref_fun, fun);
#endif
}
}
// Test for smallest normalized floats.
data1[0] = norm_min;
data1[1] = -data1[0];
CHECK_CWISE1_IF(Cond, ref_fun, fun);
// Test for largest floats.
data1[0] = norm_max;
data1[1] = -data1[0];
CHECK_CWISE1_IF(Cond, ref_fun, fun);
// Test for zeros.
data1[0] = Scalar(0.0);
data1[1] = -data1[0];
CHECK_CWISE1_IF(Cond, ref_fun, fun);
// Test for infinities.
data1[0] = NumTraits<Scalar>::infinity();
data1[1] = -data1[0];
CHECK_CWISE1_IF(Cond, ref_fun, fun);
// Test for quiet NaNs.
data1[0] = std::numeric_limits<Scalar>::quiet_NaN();
data1[1] = -std::numeric_limits<Scalar>::quiet_NaN();
CHECK_CWISE1_IF(Cond, ref_fun, fun);
}
template <typename Scalar, typename Packet>
void packetmath_real() {
typedef internal::packet_traits<Scalar> PacketTraits;
const int PacketSize = internal::unpacket_traits<Packet>::size;
const int size = PacketSize * 4;
EIGEN_ALIGN_MAX Scalar data1[PacketSize * 4];
EIGEN_ALIGN_MAX Scalar data2[PacketSize * 4];
EIGEN_ALIGN_MAX Scalar ref[PacketSize * 4];
EIGEN_ALIGN_MAX Scalar data1[PacketSize * 4] = {};
EIGEN_ALIGN_MAX Scalar data2[PacketSize * 4] = {};
EIGEN_ALIGN_MAX Scalar ref[PacketSize * 4] = {};
// Negate with -0.
if (PacketTraits::HasNegate) {
test::packet_helper<PacketTraits::HasNegate,Packet> h;
data1[0] = Scalar{-0};
h.store(data2, internal::pnegate(h.load(data1)));
typedef typename internal::make_unsigned<typename internal::make_integer<Scalar>::type>::type Bits;
Bits bits = numext::bit_cast<Bits>(data2[0]);
VERIFY_IS_EQUAL(bits, static_cast<Bits>(Bits(1)<<(sizeof(Scalar)*CHAR_BIT - 1)));
}
for (int i = 0; i < size; ++i) {
data1[i] = Scalar(internal::random<double>(0, 1) * std::pow(10., internal::random<double>(-6, 6)));
@@ -658,6 +827,7 @@ void packetmath_real() {
CHECK_CWISE1_EXACT_IF(PacketTraits::HasCeil, numext::ceil, internal::pceil);
CHECK_CWISE1_EXACT_IF(PacketTraits::HasFloor, numext::floor, internal::pfloor);
CHECK_CWISE1_EXACT_IF(PacketTraits::HasRint, numext::rint, internal::print);
CHECK_CWISE1_IF(PacketTraits::HasSign, numext::sign, internal::psign);
packetmath_boolean_mask_ops_real<Scalar,Packet>();
@@ -713,6 +883,7 @@ void packetmath_real() {
for (int i = 0; i < size; ++i) {
data1[i] = Scalar(internal::random<double>(-87, 88));
data2[i] = Scalar(internal::random<double>(-87, 88));
data1[0] = -NumTraits<Scalar>::infinity();
}
CHECK_CWISE1_IF(PacketTraits::HasExp, std::exp, internal::pexp);
@@ -847,7 +1018,7 @@ void packetmath_real() {
}
}
#if EIGEN_HAS_C99_MATH && (EIGEN_COMP_CXXVER >= 11)
#if EIGEN_HAS_C99_MATH
data1[0] = NumTraits<Scalar>::infinity();
data1[1] = Scalar(-1);
CHECK_CWISE1_IF(PacketTraits::HasLog1p, std::log1p, internal::plog1p);
@@ -889,8 +1060,10 @@ void packetmath_real() {
data1[0] = std::numeric_limits<Scalar>::denorm_min();
data1[1] = -std::numeric_limits<Scalar>::denorm_min();
h.store(data2, internal::plog(h.load(data1)));
// TODO(rmlarsen): Reenable.
// VERIFY_IS_EQUAL(std::log(std::numeric_limits<Scalar>::denorm_min()), data2[0]);
// TODO(rmlarsen): Re-enable for bfloat16.
if (!internal::is_same<Scalar, bfloat16>::value) {
VERIFY_IS_APPROX(std::log(std::numeric_limits<Scalar>::denorm_min()), data2[0]);
}
VERIFY((numext::isnan)(data2[1]));
}
#endif
@@ -911,18 +1084,10 @@ void packetmath_real() {
VERIFY((numext::isnan)(data2[0]));
VERIFY((numext::isnan)(data2[1]));
}
if (PacketTraits::HasSqrt) {
test::packet_helper<PacketTraits::HasSqrt, Packet> h;
data1[0] = Scalar(-1.0f);
if (std::numeric_limits<Scalar>::has_denorm == std::denorm_present) {
data1[1] = -std::numeric_limits<Scalar>::denorm_min();
} else {
data1[1] = -NumTraits<Scalar>::epsilon();
}
h.store(data2, internal::psqrt(h.load(data1)));
VERIFY((numext::isnan)(data2[0]));
VERIFY((numext::isnan)(data2[1]));
}
packetmath_test_IEEE_corner_cases<PacketTraits::HasSqrt, Scalar, Packet>(numext::sqrt<Scalar>, psqrt_functor());
packetmath_test_IEEE_corner_cases<PacketTraits::HasRsqrt, Scalar, Packet>(numext::rsqrt<Scalar>, prsqrt_functor());
// TODO(rmlarsen): Re-enable for half and bfloat16.
if (PacketTraits::HasCos
&& !internal::is_same<Scalar, half>::value
@@ -972,6 +1137,23 @@ void packetmath_real() {
VERIFY_IS_EQUAL(data2[0], Scalar(1));
}
}
if (PacketTraits::HasReciprocal && PacketSize >= 2) {
test::packet_helper<PacketTraits::HasReciprocal, Packet> h;
const Scalar inf = NumTraits<Scalar>::infinity();
const Scalar zero = Scalar(0);
data1[0] = zero;
data1[1] = -zero;
h.store(data2, internal::preciprocal(h.load(data1)));
VERIFY_IS_EQUAL(data2[0], inf);
VERIFY_IS_EQUAL(data2[1], -inf);
data1[0] = inf;
data1[1] = -inf;
h.store(data2, internal::preciprocal(h.load(data1)));
VERIFY_IS_EQUAL(data2[0], zero);
VERIFY_IS_EQUAL(data2[1], -zero);
}
}
#define CAST_CHECK_CWISE1_IF(COND, REFOP, POP, SCALAR, REFTYPE) if(COND) { \
@@ -1170,6 +1352,7 @@ void packetmath_complex() {
data1[i] = Scalar(internal::random<RealScalar>(), internal::random<RealScalar>());
}
CHECK_CWISE1_N(numext::sqrt, internal::psqrt, size);
CHECK_CWISE1_IF(PacketTraits::HasSign, numext::sign, internal::psign);
// Test misc. corner cases.
const RealScalar zero = RealScalar(0);
@@ -1243,6 +1426,36 @@ void packetmath_scatter_gather() {
for (int i = 0; i < PacketSize; ++i) {
VERIFY(test::isApproxAbs(data1[i], buffer[i * 7], refvalue) && "pgather");
}
for (Index N = 0; N <= PacketSize; ++N) {
for (Index i = 0; i < N; ++i) {
data1[i] = internal::random<Scalar>() / RealScalar(PacketSize);
}
for (Index i = 0; i < N * 20; ++i) {
buffer[i] = Scalar(0);
}
packet = internal::pload_partial<Packet>(data1, N);
internal::pscatter_partial<Scalar, Packet>(buffer, packet, stride, N);
for (Index i = 0; i < N * 20; ++i) {
if ((i % stride) == 0 && i < stride * N) {
VERIFY(test::isApproxAbs(buffer[i], data1[i / stride], refvalue) && "pscatter_partial");
} else {
VERIFY(test::isApproxAbs(buffer[i], Scalar(0), refvalue) && "pscatter_partial");
}
}
for (Index i = 0; i < N * 7; ++i) {
buffer[i] = internal::random<Scalar>() / RealScalar(PacketSize);
}
packet = internal::pgather_partial<Scalar, Packet>(buffer, 7, N);
internal::pstore_partial(data1, packet, N);
for (Index i = 0; i < N; ++i) {
VERIFY(test::isApproxAbs(data1[i], buffer[i * 7], refvalue) && "pgather_partial");
}
}
}
namespace Eigen {