Eigen::internal::unrolls::transB< Scalar > Class Template Reference

#include <Eigen/src/Core/arch/AVX512/TrsmKernel.h>

Detailed Description

template<typename Scalar>
class Eigen::internal::unrolls::transB< Scalar >

Unrolls for copyBToRowMajor

Idea: 1) Load a block of right-hand sides to registers (using loadB). 2) Convert the block from column-major to row-major (transposeLxL) 3) Store the blocks from register either to a temp array (toTemp == true), or back to B (toTemp == false).

We use at most EIGEN_AVX_MAX_NUM_ACC avx registers to store the blocks of B. The remaining registers are used as temps for transposing.

Blocks will be of size Lx{U1,U2,U3}. packetIndexOffset is used to index between these subblocks For fp32, PacketSize = 2*EIGEN_AVX_MAX_NUM_ROW, so we reinterpret packets as packets half the size (zmm -> ymm).

Static Public Member Functions
template<int64_t endN, int64_t counter, int64_t packetIndexOffset, bool remM, int64_t remN_>
static EIGEN_ALWAYS_INLINE std::enable_if_t<(counter > 0)>	aux_loadB (Scalar *B_arr, int64_t LDB, PacketBlock< vecHalf, EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS > &ymm, int64_t remM_=0)
template<int64_t endN, int64_t counter, bool toTemp, bool remM, int64_t remN_>
static EIGEN_ALWAYS_INLINE std::enable_if_t<(counter > 0)>	aux_loadBBlock (Scalar *B_arr, int64_t LDB, Scalar *B_temp, int64_t LDB_, PacketBlock< vecHalf, EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS > &ymm, int64_t remM_=0)
template<int64_t endN, int64_t counter, int64_t packetIndexOffset, bool remK, bool remM>
static EIGEN_ALWAYS_INLINE std::enable_if_t<(counter > 0)>	aux_storeB (Scalar *B_arr, int64_t LDB, PacketBlock< vecHalf, EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS > &ymm, int64_t rem_=0)
template<int64_t endN, int64_t counter, bool toTemp, bool remM, int64_t remK_>
static EIGEN_ALWAYS_INLINE std::enable_if_t<(counter > 0)>	aux_storeBBlock (Scalar *B_arr, int64_t LDB, Scalar *B_temp, int64_t LDB_, PacketBlock< vecHalf, EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS > &ymm, int64_t remM_=0)
template<int64_t unrollN, bool toTemp, bool remM>
static EIGEN_ALWAYS_INLINE void	transB_kernel (Scalar *B_arr, int64_t LDB, Scalar *B_temp, int64_t LDB_, PacketBlock< vecHalf, EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS > &ymm, int64_t remM_=0)

Member Function Documentation

aux_loadB()

template<typename Scalar>

template<int64_t endN, int64_t counter, int64_t packetIndexOffset, bool remM, int64_t remN_>

static EIGEN_ALWAYS_INLINE std::enable_if_t<(counter > 0)> Eigen::internal::unrolls::transB< Scalar >::aux_loadB ( Scalar * B_arr, int64_t LDB, PacketBlock< vecHalf, EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS > & ymm, int64_t remM_ = 0 )

inlinestatic

aux_loadB

1-D unroll for(startN = 0; startN < endN; startN++)

aux_loadBBlock()

template<typename Scalar>

template<int64_t endN, int64_t counter, bool toTemp, bool remM, int64_t remN_>

static EIGEN_ALWAYS_INLINE std::enable_if_t<(counter > 0)> Eigen::internal::unrolls::transB< Scalar >::aux_loadBBlock ( Scalar * B_arr, int64_t LDB, Scalar * B_temp, int64_t LDB_, PacketBlock< vecHalf, EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS > & ymm, int64_t remM_ = 0 )

inlinestatic

aux_loadBBlock

1-D unroll for(startN = 0; startN < endN; startN += EIGEN_AVX_MAX_NUM_ROW)

aux_storeB()

template<typename Scalar>

template<int64_t endN, int64_t counter, int64_t packetIndexOffset, bool remK, bool remM>

static EIGEN_ALWAYS_INLINE std::enable_if_t<(counter > 0)> Eigen::internal::unrolls::transB< Scalar >::aux_storeB ( Scalar * B_arr, int64_t LDB, PacketBlock< vecHalf, EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS > & ymm, int64_t rem_ = 0 )

inlinestatic

aux_storeB

1-D unroll for(startN = 0; startN < endN; startN++)

aux_storeBBlock()

template<typename Scalar>

template<int64_t endN, int64_t counter, bool toTemp, bool remM, int64_t remK_>

static EIGEN_ALWAYS_INLINE std::enable_if_t<(counter > 0)> Eigen::internal::unrolls::transB< Scalar >::aux_storeBBlock ( Scalar * B_arr, int64_t LDB, Scalar * B_temp, int64_t LDB_, PacketBlock< vecHalf, EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS > & ymm, int64_t remM_ = 0 )

inlinestatic

aux_storeBBlock

1-D unroll for(startN = 0; startN < endN; startN += EIGEN_AVX_MAX_NUM_ROW)

transB_kernel()

template<typename Scalar>

template<int64_t unrollN, bool toTemp, bool remM>

static EIGEN_ALWAYS_INLINE void Eigen::internal::unrolls::transB< Scalar >::transB_kernel ( Scalar * B_arr, int64_t LDB, Scalar * B_temp, int64_t LDB_, PacketBlock< vecHalf, EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS > & ymm, int64_t remM_ = 0 )

inlinestatic

Unrolls needed for each case:

• AVX512 fp32 48 32 16 8 4 2 1
• AVX512 fp64 24 16 8 4 2 1

For fp32 L and U1 are 1:2 so for U3/U2 cases the loads/stores need to be split up.

---

The documentation for this class was generated from the following file:

• /builds/libeigen/eigen/Eigen/src/Core/arch/AVX512/TrsmKernel.h