GitHub-Proxy/eigen
1
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2025-06-04 18:54:00 +08:00
Code Issues Packages Projects Releases Wiki Activity

Fix boolean scatter and random generation for tensors.

Browse Source
This commit is contained in:
Antonio Sánchez 2025-02-25 21:37:09 +00:00
parent 9935396b15
commit d79bac0d3c
3 changed files with 27 additions and 6 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

22
Eigen/src/Core/arch/SSE/PacketMath.h
View File

@ -1694,10 +1694,24 @@ EIGEN_STRONG_INLINE void pscatter<uint32_t, Packet4ui>(uint32_t* to, const Packe
} }
template <> template <>
EIGEN_STRONG_INLINE void pscatter<bool, Packet16b>(bool* to, const Packet16b& from, Index stride) { EIGEN_STRONG_INLINE void pscatter<bool, Packet16b>(bool* to, const Packet16b& from, Index stride) {
to[4 * stride * 0] = _mm_cvtsi128_si32(from); EIGEN_ALIGN16 bool tmp[16];
to[4 * stride * 1] = _mm_cvtsi128_si32(_mm_shuffle_epi32(from, 1)); pstore(tmp, from);
to[4 * stride * 2] = _mm_cvtsi128_si32(_mm_shuffle_epi32(from, 2)); to[stride * 0] = tmp[0];
to[4 * stride * 3] = _mm_cvtsi128_si32(_mm_shuffle_epi32(from, 3)); to[stride * 1] = tmp[1];
to[stride * 2] = tmp[2];
to[stride * 3] = tmp[3];
to[stride * 4] = tmp[4];
to[stride * 5] = tmp[5];
to[stride * 6] = tmp[6];
to[stride * 7] = tmp[7];
to[stride * 8] = tmp[8];
to[stride * 9] = tmp[9];
to[stride * 10] = tmp[10];
to[stride * 11] = tmp[11];
to[stride * 12] = tmp[12];
to[stride * 13] = tmp[13];
to[stride * 14] = tmp[14];
to[stride * 15] = tmp[15];
} }
// some compilers might be tempted to perform multiple moves instead of using a vector path. // some compilers might be tempted to perform multiple moves instead of using a vector path.

5
test/packetmath.cpp
View File

@ -1635,7 +1635,7 @@ void packetmath_scatter_gather() {
EIGEN_ALIGN_MAX Scalar data1[PacketSize]; EIGEN_ALIGN_MAX Scalar data1[PacketSize];
RealScalar refvalue = RealScalar(0); RealScalar refvalue = RealScalar(0);
for (int i = 0; i < PacketSize; ++i) { for (int i = 0; i < PacketSize; ++i) {
data1[i] = internal::random<Scalar>() / RealScalar(PacketSize); data1[i] = internal::random<Scalar>();
} }
int stride = internal::random<int>(1, 20); int stride = internal::random<int>(1, 20);
@ -1655,7 +1655,7 @@ void packetmath_scatter_gather() {
} }
for (int i = 0; i < PacketSize * 7; ++i) { for (int i = 0; i < PacketSize * 7; ++i) {
buffer[i] = internal::random<Scalar>() / RealScalar(PacketSize); buffer[i] = internal::random<Scalar>();
} }
packet = internal::pgather<Scalar, Packet>(buffer, 7); packet = internal::pgather<Scalar, Packet>(buffer, 7);
internal::pstore(data1, packet); internal::pstore(data1, packet);
@ -1745,6 +1745,7 @@ EIGEN_DECLARE_TEST(packetmath) {
CALL_SUBTEST_12(test::runner<std::complex<double>>::run()); CALL_SUBTEST_12(test::runner<std::complex<double>>::run());
CALL_SUBTEST_13(test::runner<half>::run()); CALL_SUBTEST_13(test::runner<half>::run());
CALL_SUBTEST_14((packetmath<bool, internal::packet_traits<bool>::type>())); CALL_SUBTEST_14((packetmath<bool, internal::packet_traits<bool>::type>()));
CALL_SUBTEST_14((packetmath_scatter_gather<bool, internal::packet_traits<bool>::type>()));
CALL_SUBTEST_15(test::runner<bfloat16>::run()); CALL_SUBTEST_15(test::runner<bfloat16>::run());
g_first_pass = false; g_first_pass = false;
} }

6
unsupported/Eigen/CXX11/src/Tensor/TensorRandom.h
View File

@ -49,6 +49,12 @@ EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T RandomToTypeUniform(uint64_t* state, uin
return static_cast<T>(rnd); return static_cast<T>(rnd);
} }
template <>
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool RandomToTypeUniform<bool>(uint64_t* state, uint64_t stream) {
unsigned rnd = PCG_XSH_RS_generator(state, stream);
return (rnd & 0x1) != 0;
}
template <> template <>
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Eigen::half RandomToTypeUniform<Eigen::half>(uint64_t* state, uint64_t stream) { EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Eigen::half RandomToTypeUniform<Eigen::half>(uint64_t* state, uint64_t stream) {
// Generate 10 random bits for the mantissa, merge with exponent. // Generate 10 random bits for the mantissa, merge with exponent.