666fe19293
Similarly, make tf.reduce_min, and tf.reduce_max return NaN if any element among thos reduced are NaN. PiperOrigin-RevId: 338194828 Change-Id: I337729b083468500694bfd7b5846633415be1710
99 lines
3.8 KiB
Diff
99 lines
3.8 KiB
Diff
diff -ru a/Eigen/src/Geometry/arch/Geometry_SSE.h b/Eigen/src/Geometry/arch/Geometry_SSE.h
|
|
--- a/Eigen/src/Geometry/arch/Geometry_SSE.h
|
|
+++ b/Eigen/src/Geometry/arch/Geometry_SSE.h
|
|
@@ -33,13 +33,14 @@
|
|
Packet4f b = be.template packet<BAlignment,Packet4f>(0);
|
|
Packet4f s1 = pmul(vec4f_swizzle1(a,1,2,0,2),vec4f_swizzle1(b,2,0,1,2));
|
|
Packet4f s2 = pmul(vec4f_swizzle1(a,3,3,3,1),vec4f_swizzle1(b,0,1,2,1));
|
|
- pstoret<float,Packet4f,ResAlignment>(
|
|
- &res.x(),
|
|
- padd(psub(pmul(a,vec4f_swizzle1(b,3,3,3,3)),
|
|
- pmul(vec4f_swizzle1(a,2,0,1,0),
|
|
- vec4f_swizzle1(b,1,2,0,0))),
|
|
- pxor(mask,padd(s1,s2))));
|
|
-
|
|
+ pstoret<float, Packet4f, ResAlignment>(
|
|
+ &res.x(),
|
|
+ padd<Packet4f>(
|
|
+ psub<Packet4f>(pmul<Packet4f>(a, vec4f_swizzle1(b, 3, 3, 3, 3)),
|
|
+ pmul<Packet4f>(vec4f_swizzle1(a, 2, 0, 1, 0),
|
|
+ vec4f_swizzle1(b, 1, 2, 0, 0))),
|
|
+ pxor<Packet4f>(mask, padd(s1, s2))));
|
|
+
|
|
return res;
|
|
}
|
|
};
|
|
diff -ru a/Eigen/src/Core/GenericPacketMath.h b/Eigen/src/Core/GenericPacketMath.h
|
|
--- a/Eigen/src/Core/GenericPacketMath.h
|
|
+++ b/Eigen/src/Core/GenericPacketMath.h
|
|
@@ -255,49 +255,43 @@
|
|
return std::complex<RealScalar>(b, b);
|
|
}
|
|
|
|
-template <typename Packet, typename Op>
|
|
-EIGEN_DEVICE_FUNC inline Packet bitwise_helper(const Packet& a, const Packet& b, Op op) {
|
|
+/** \internal \returns the bitwise and of \a a and \a b */
|
|
+template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
|
|
+pand(const Packet& a, const Packet& b) {
|
|
const unsigned char* a_ptr = reinterpret_cast<const unsigned char*>(&a);
|
|
const unsigned char* b_ptr = reinterpret_cast<const unsigned char*>(&b);
|
|
Packet c;
|
|
unsigned char* c_ptr = reinterpret_cast<unsigned char*>(&c);
|
|
for (size_t i = 0; i < sizeof(Packet); ++i) {
|
|
- *c_ptr++ = op(*a_ptr++, *b_ptr++);
|
|
+ *c_ptr++ = *a_ptr++ & *b_ptr++;
|
|
}
|
|
return c;
|
|
}
|
|
|
|
-/** \internal \returns the bitwise and of \a a and \a b */
|
|
-template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
|
|
-pand(const Packet& a, const Packet& b) {
|
|
-#if defined(EIGEN_HIP_DEVICE_COMPILE)
|
|
- return bitwise_helper(a ,b, std::bit_and<unsigned char>());
|
|
-#else
|
|
- EIGEN_USING_STD(bit_and);
|
|
- return bitwise_helper(a ,b, bit_and<unsigned char>());
|
|
-#endif
|
|
-}
|
|
-
|
|
/** \internal \returns the bitwise or of \a a and \a b */
|
|
template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
|
|
por(const Packet& a, const Packet& b) {
|
|
-#if defined(EIGEN_HIP_DEVICE_COMPILE)
|
|
- return bitwise_helper(a ,b, std::bit_or<unsigned char>());
|
|
-#else
|
|
- EIGEN_USING_STD(bit_or);
|
|
- return bitwise_helper(a ,b, bit_or<unsigned char>());
|
|
-#endif
|
|
+ const unsigned char* a_ptr = reinterpret_cast<const unsigned char*>(&a);
|
|
+ const unsigned char* b_ptr = reinterpret_cast<const unsigned char*>(&b);
|
|
+ Packet c;
|
|
+ unsigned char* c_ptr = reinterpret_cast<unsigned char*>(&c);
|
|
+ for (size_t i = 0; i < sizeof(Packet); ++i) {
|
|
+ *c_ptr++ = *a_ptr++ | *b_ptr++;
|
|
+ }
|
|
+ return c;
|
|
}
|
|
|
|
/** \internal \returns the bitwise xor of \a a and \a b */
|
|
template<typename Packet> EIGEN_DEVICE_FUNC inline Packet
|
|
pxor(const Packet& a, const Packet& b) {
|
|
-#if defined(EIGEN_HIP_DEVICE_COMPILE)
|
|
- return bitwise_helper(a ,b, std::bit_xor<unsigned char>());
|
|
-#else
|
|
- EIGEN_USING_STD(bit_xor);
|
|
- return bitwise_helper(a ,b, bit_xor<unsigned char>());
|
|
-#endif
|
|
+ const unsigned char* a_ptr = reinterpret_cast<const unsigned char*>(&a);
|
|
+ const unsigned char* b_ptr = reinterpret_cast<const unsigned char*>(&b);
|
|
+ Packet c;
|
|
+ unsigned char* c_ptr = reinterpret_cast<unsigned char*>(&c);
|
|
+ for (size_t i = 0; i < sizeof(Packet); ++i) {
|
|
+ *c_ptr++ = *a_ptr++ ^ *b_ptr++;
|
|
+ }
|
|
+ return c;
|
|
}
|
|
|
|
/** \internal \returns the bitwise and of \a a and not \a b */
|