devtools/eigen
1
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2026-04-10 11:34:33 +08:00
Code Issues Packages Projects Releases Wiki Activity

Clean up packetmath tests and fix various bugs to make bfloat16 pass (almost) all packetmath tests with SSE, AVX, and AVX512.

Browse Source
This commit is contained in:
Rasmus Munk Larsen
2020-10-09 20:05:49 +00:00
parent 7a8d3d5b81
commit 4e4d3f32d1
8 changed files with 114 additions and 70 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
Eigen/src/Core/arch/AVX512/PacketMath.h
View File

@@ -1626,8 +1626,6 @@ template <> struct is_arithmetic<Packet16bf> { enum { value = true }; };
template <>
struct packet_traits<bfloat16> : default_packet_traits {
typedef Packet16bf type;
// There is no half-size packet for current Packet16bf.
// TODO: support as SSE path.
typedef Packet8bf half;
enum {
Vectorizable = 1,
@@ -1883,6 +1881,11 @@ EIGEN_STRONG_INLINE Packet16bf pmax<Packet16bf>(const Packet16bf& a,
return F32ToBf16(pmax<Packet16f>(Bf16ToF32(a), Bf16ToF32(b)));
}
template <>
EIGEN_STRONG_INLINE Packet16bf plset<Packet16bf>(const bfloat16& a) {
return F32ToBf16(plset<Packet16f>(static_cast<float>(a)));
}
template <>
EIGEN_STRONG_INLINE Packet8bf predux_half_dowto4<Packet16bf>(const Packet16bf& a) {
Packet8bf lane0 = _mm256_extractf128_si256(a, 0);