diff options
| author |  Kan Chen <chenkan0734@163.com> | 2020-05-28 20:26:09 +0800 |
|---|---|---|
| committer |  Kan Chen <chenkan5@huawei.com> | 2020-05-29 00:33:45 +0000 |
| commit | 8d1302f566f70c29d0ba864972ab50460a96cf2d (patch) | |
| tree | 674936e743a34b74e596cf7df27c9467acd9884a /Eigen/src/Core/arch/NEON/PacketMath.h | |
| parent | 8719b9c5bc1a97e62d675c02495ed72dda6fae73 (diff) | |
Add support for PacketBlock<Packet8s,4> and PacketBlock<Packet16uc,4> ptranspose on NEON
Diffstat (limited to 'Eigen/src/Core/arch/NEON/PacketMath.h')
| -rw-r--r-- | Eigen/src/Core/arch/NEON/PacketMath.h | 29 |
1 files changed, 29 insertions, 0 deletions
diff --git a/Eigen/src/Core/arch/NEON/PacketMath.h b/Eigen/src/Core/arch/NEON/PacketMath.h index e11af1dca..065c8100f 100644 --- a/Eigen/src/Core/arch/NEON/PacketMath.h +++ b/Eigen/src/Core/arch/NEON/PacketMath.h @@ -2869,6 +2869,35 @@ EIGEN_DEVICE_FUNC inline void ptranspose(PacketBlock<Packet4s, 4>& kernel) kernel.packet[2] = vreinterpret_s16_u32(zip32_2.val[0]); kernel.packet[3] = vreinterpret_s16_u32(zip32_2.val[1]); } + +EIGEN_DEVICE_FUNC inline void ptranspose(PacketBlock<Packet8s, 4>& kernel) +{ + const int16x8x2_t zip16_1 = vzipq_s16(kernel.packet[0], kernel.packet[1]); + const int16x8x2_t zip16_2 = vzipq_s16(kernel.packet[2], kernel.packet[3]); + + const uint32x4x2_t zip32_1 = vzipq_u32(vreinterpretq_u32_s16(zip16_1.val[0]), vreinterpretq_u32_s16(zip16_2.val[0])); + const uint32x4x2_t zip32_2 = vzipq_u32(vreinterpretq_u32_s16(zip16_1.val[1]), vreinterpretq_u32_s16(zip16_2.val[1])); + + kernel.packet[0] = vreinterpretq_s16_u32(zip32_1.val[0]); + kernel.packet[1] = vreinterpretq_s16_u32(zip32_1.val[1]); + kernel.packet[2] = vreinterpretq_s16_u32(zip32_2.val[0]); + kernel.packet[3] = vreinterpretq_s16_u32(zip32_2.val[1]); +} + +EIGEN_DEVICE_FUNC inline void ptranspose(PacketBlock<Packet16uc, 4>& kernel) +{ + const uint8x16x2_t zip8_1 = vzipq_u8(kernel.packet[0], kernel.packet[1]); + const uint8x16x2_t zip8_2 = vzipq_u8(kernel.packet[2], kernel.packet[3]); + + const uint16x8x2_t zip16_1 = vzipq_u16(vreinterpretq_u16_u8(zip8_1.val[0]), vreinterpretq_u16_u8(zip8_2.val[0])); + const uint16x8x2_t zip16_2 = vzipq_u16(vreinterpretq_u16_u8(zip8_1.val[1]), vreinterpretq_u16_u8(zip8_2.val[1])); + + kernel.packet[0] = vreinterpretq_u8_u16(zip16_1.val[0]); + kernel.packet[1] = vreinterpretq_u8_u16(zip16_1.val[1]); + kernel.packet[2] = vreinterpretq_u8_u16(zip16_2.val[0]); + kernel.packet[3] = vreinterpretq_u8_u16(zip16_2.val[1]); +} + EIGEN_DEVICE_FUNC inline void ptranspose(PacketBlock<Packet8s, 8>& kernel) { const int16x8x2_t zip16_1 = vzipq_s16(kernel.packet[0], kernel.packet[1]); |