tdls::TileOperations< T, tile_size, unroll_inner > Struct Template Reference

tile_size x tile_size register-tile micro-kernels shared by the solver families. More...

#include <tile_operations.hpp>

Inheritance diagram for tdls::TileOperations< T, tile_size, unroll_inner >:

Static Public Member Functions
template<int k_extent>
TDLS_HOST_DEVICE static TDLS_FORCEINLINE constexpr void	swap_rows (T *TDLS_RESTRICT t, int k, int r) noexcept
	Row swap k <-> r inside the KExKE active part of the tile, compile-time indexed on both sides.
template<int diag_extent, int col_extent>
TDLS_HOST_DEVICE static TDLS_FORCEINLINE constexpr void	trsm_left_unit (const T *TDLS_RESTRICT lu, T *TDLS_RESTRICT B) noexcept
	B := L^-1 B, with L the unit lower part of the factored diagonal tile. L is KDxKD, B is KDxC.
template<int diag_extent, int row_extent>
TDLS_HOST_DEVICE static TDLS_FORCEINLINE constexpr void	trsm_right (const T *TDLS_RESTRICT lu, T *TDLS_RESTRICT B) noexcept
	B := B U^-1, with U the upper part of the factored diagonal tile. U is KDxKD, B is RxKD.
template<int row_extent, int col_extent, int k_extent>
TDLS_HOST_DEVICE static TDLS_FORCEINLINE constexpr void	gemm_sub (T *TDLS_RESTRICT Ct, const T *TDLS_RESTRICT At, const T *TDLS_RESTRICT Bt) noexcept
	Ct -= At*Bt with per-element dot-product accumulation. At is RxK, Bt is KxC, Ct is RxC.

Detailed Description

template<typename T, int tile_size, bool unroll_inner>
struct tdls::TileOperations< T, tile_size, unroll_inner >

tile_size x tile_size register-tile micro-kernels shared by the solver families.

Template Parameters

T	scalar type
tile_size	tile size (int, row stride of the register tiles)
unroll_inner	unroll knob, forwarded from the solver configuration

Member Function Documentation

gemm_sub()

template<typename T , int tile_size, bool unroll_inner>

template<int row_extent, int col_extent, int k_extent>

TDLS_HOST_DEVICE static TDLS_FORCEINLINE constexpr void tdls::TileOperations< T, tile_size, unroll_inner >::gemm_sub ( T *TDLS_RESTRICT Ct, const T *TDLS_RESTRICT At, const T *TDLS_RESTRICT Bt )

inlinestaticconstexprnoexcept

Ct -= At*Bt with per-element dot-product accumulation. At is RxK, Bt is KxC, Ct is RxC.

Template Parameters

row_extent	row extent of Ct and At
col_extent	column extent of Ct and Bt
k_extent	inner extent (columns of At, rows of Bt)

Parameters

[in,out]	Ct	accumulator tile
[in]	At	left factor tile
[in]	Bt	right factor tile

swap_rows()

template<typename T , int tile_size, bool unroll_inner>

template<int k_extent>

TDLS_HOST_DEVICE static TDLS_FORCEINLINE constexpr void tdls::TileOperations< T, tile_size, unroll_inner >::swap_rows ( T *TDLS_RESTRICT t, int k, int r )

inlinestaticconstexprnoexcept

Row swap k <-> r inside the KExKE active part of the tile, compile-time indexed on both sides.

The equality test against every unrolled row index keeps the tile addressing static; a dynamic t[r*tile_size+j] would spill the tile. k_extent bounds both loops, so the phantom slots of trailing tiles are never touched, as in every other micro-kernel.

Template Parameters

k_extent

active extent of the tile

Parameters

[in,out]	t	register tile
[in]	k	destination row
[in]	r	source row to swap in

trsm_left_unit()

template<typename T , int tile_size, bool unroll_inner>

template<int diag_extent, int col_extent>

TDLS_HOST_DEVICE static TDLS_FORCEINLINE constexpr void tdls::TileOperations< T, tile_size, unroll_inner >::trsm_left_unit ( const T *TDLS_RESTRICT lu, T *TDLS_RESTRICT B )

inlinestaticconstexprnoexcept

B := L^-1 B, with L the unit lower part of the factored diagonal tile. L is KDxKD, B is KDxC.

Template Parameters

diag_extent	extent of the factored diagonal tile
col_extent	column extent of B

Parameters

[in]	lu	factored diagonal tile (L\U)
[in,out]	B	updated register tile

trsm_right()

template<typename T , int tile_size, bool unroll_inner>

template<int diag_extent, int row_extent>

TDLS_HOST_DEVICE static TDLS_FORCEINLINE constexpr void tdls::TileOperations< T, tile_size, unroll_inner >::trsm_right ( const T *TDLS_RESTRICT lu, T *TDLS_RESTRICT B )

inlinestaticconstexprnoexcept

B := B U^-1, with U the upper part of the factored diagonal tile. U is KDxKD, B is RxKD.

The diagonal of lu must already hold the pivot RECIPROCALS: no divisions here.

Template Parameters

diag_extent	extent of the factored diagonal tile
row_extent	row extent of B

Parameters

[in]	lu	factored diagonal tile (L\U, reciprocal diagonal)
[in,out]	B	updated register tile

---

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

• include/tdls/solvers/tile_operations.hpp