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

Fix packetmath tests on M* macs.

Browse Source 
libeigen/eigen!2120
This commit is contained in:
Antonio Sánchez
2026-02-08 18:07:24 +00:00
committed by Rasmus Munk Larsen
parent 752911927f
commit 4d05fcf8da
4 changed files with 52 additions and 15 deletions
Download Patch File Download Diff File Expand all files Collapse all files

38
Eigen/src/Core/arch/NEON/PacketMath.h
View File

@@ -2496,38 +2496,60 @@ template <>
EIGEN_STRONG_INLINE Packet4f ploadquad<Packet4f>(const float* from) {
return vld1q_dup_f32(from);
}
// WORKAROUND: Apple Clang 17.0.0 (and Homebrew Clang 21.1.8) at -O0 optimization
// generate incorrect code for vld1_dup_[su]8, ignoring the pointer offset.
// We use vdup_n_s8(*from) to force a safe scalar load before broadcast.
EIGEN_ALWAYS_INLINE int8x8_t eigen_vld1_dup_s8(const int8_t* ptr) {
#if EIGEN_COMP_CLANGAPPLE && EIGEN_ARCH_ARM64
return vdup_n_s8(*ptr);
#else
return vld1_dup_s8(ptr);
#endif
}
EIGEN_ALWAYS_INLINE uint8x8_t eigen_vld1_dup_u8(const uint8_t* ptr) {
#if EIGEN_COMP_CLANGAPPLE && EIGEN_ARCH_ARM64
return vdup_n_u8(*ptr);
#else
return vld1_dup_u8(ptr);
#endif
}
template <>
EIGEN_STRONG_INLINE Packet4c ploadquad<Packet4c>(const int8_t* from) {
return vget_lane_s32(vreinterpret_s32_s8(vld1_dup_s8(from)), 0);
return vget_lane_s32(vreinterpret_s32_s8(eigen_vld1_dup_s8(from)), 0);
}
template <>
EIGEN_STRONG_INLINE Packet8c ploadquad<Packet8c>(const int8_t* from) {
return vreinterpret_s8_u32(
vzip_u32(vreinterpret_u32_s8(vld1_dup_s8(from)), vreinterpret_u32_s8(vld1_dup_s8(from + 1))).val[0]);
vzip_u32(vreinterpret_u32_s8(eigen_vld1_dup_s8(from)), vreinterpret_u32_s8(eigen_vld1_dup_s8(from + 1))).val[0]);
}
template <>
EIGEN_STRONG_INLINE Packet16c ploadquad<Packet16c>(const int8_t* from) {
const int8x8_t a = vreinterpret_s8_u32(
vzip_u32(vreinterpret_u32_s8(vld1_dup_s8(from)), vreinterpret_u32_s8(vld1_dup_s8(from + 1))).val[0]);
vzip_u32(vreinterpret_u32_s8(eigen_vld1_dup_s8(from)), vreinterpret_u32_s8(eigen_vld1_dup_s8(from + 1))).val[0]);
const int8x8_t b = vreinterpret_s8_u32(
vzip_u32(vreinterpret_u32_s8(vld1_dup_s8(from + 2)), vreinterpret_u32_s8(vld1_dup_s8(from + 3))).val[0]);
vzip_u32(vreinterpret_u32_s8(eigen_vld1_dup_s8(from + 2)), vreinterpret_u32_s8(eigen_vld1_dup_s8(from + 3)))
.val[0]);
return vcombine_s8(a, b);
}
template <>
EIGEN_STRONG_INLINE Packet4uc ploadquad<Packet4uc>(const uint8_t* from) {
return vget_lane_u32(vreinterpret_u32_u8(vld1_dup_u8(from)), 0);
return vget_lane_u32(vreinterpret_u32_u8(eigen_vld1_dup_u8(from)), 0);
}
template <>
EIGEN_STRONG_INLINE Packet8uc ploadquad<Packet8uc>(const uint8_t* from) {
return vreinterpret_u8_u32(
vzip_u32(vreinterpret_u32_u8(vld1_dup_u8(from)), vreinterpret_u32_u8(vld1_dup_u8(from + 1))).val[0]);
vzip_u32(vreinterpret_u32_u8(eigen_vld1_dup_u8(from)), vreinterpret_u32_u8(eigen_vld1_dup_u8(from + 1))).val[0]);
}
template <>
EIGEN_STRONG_INLINE Packet16uc ploadquad<Packet16uc>(const uint8_t* from) {
const uint8x8_t a = vreinterpret_u8_u32(
vzip_u32(vreinterpret_u32_u8(vld1_dup_u8(from)), vreinterpret_u32_u8(vld1_dup_u8(from + 1))).val[0]);
vzip_u32(vreinterpret_u32_u8(eigen_vld1_dup_u8(from)), vreinterpret_u32_u8(eigen_vld1_dup_u8(from + 1))).val[0]);
const uint8x8_t b = vreinterpret_u8_u32(
vzip_u32(vreinterpret_u32_u8(vld1_dup_u8(from + 2)), vreinterpret_u32_u8(vld1_dup_u8(from + 3))).val[0]);
vzip_u32(vreinterpret_u32_u8(eigen_vld1_dup_u8(from + 2)), vreinterpret_u32_u8(eigen_vld1_dup_u8(from + 3)))
.val[0]);
return vcombine_u8(a, b);
}
template <>