gpu::xetla::subgroup::tile_desc_t< tile_size_x_, tile_size_y_, block_size_x_, block_size_y_, reg_layout_ > Struct Template Reference

Is to illustrate the tile information about a sub matrix. More...

#include <api.hpp>

Static Public Attributes
static constexpr uint32_t	tile_size_x = tile_size_x_
static constexpr uint32_t	tile_size_y = tile_size_y_
static constexpr uint32_t	block_size_x = block_size_x_
static constexpr uint32_t	block_size_y = block_size_y_
static constexpr uint32_t	remained_size_y = tile_size_y % block_size_y
static constexpr reg_layout	register_layout = reg_layout_
static constexpr bool	reg_transpose = reg_layout_ == reg_layout::transpose_tiled
static constexpr uint32_t	num_block_x = tile_size_x / block_size_x
static constexpr uint32_t	num_block_y = tile_size_y / block_size_y
static constexpr uint32_t	num_block = num_block_x * (num_block_y + (remained_size_y > 0 ? 1 : 0))
static constexpr uint32_t	block_elems = block_size_x * block_size_y
static constexpr uint32_t	tile_elems = tile_size_x * tile_size_y

Detailed Description

template<uint32_t tile_size_x_, uint32_t tile_size_y_, uint32_t block_size_x_, uint32_t block_size_y_, reg_layout reg_layout_ = reg_layout::tiled>
struct gpu::xetla::subgroup::tile_desc_t< tile_size_x_, tile_size_y_, block_size_x_, block_size_y_, reg_layout_ >

Is to illustrate the tile information about a sub matrix.

Template Parameters

tile_size_x_	Is the horizon tile size.
tile_size_y_	Is the vertical tile size.
block_size_x_	Is the horizon block size.
block_size_y_	Is the vertical block size.
reg_layout_	Is the register layout i.e. tiled, vnni_tiled and so on.
gpu_arch_	Is the hardware architecture tag.

Member Data Documentation

block_elems

template<uint32_t tile_size_x_, uint32_t tile_size_y_, uint32_t block_size_x_, uint32_t block_size_y_, reg_layout reg_layout_ = reg_layout::tiled>

constexpr uint32_t gpu::xetla::subgroup::tile_desc_t< tile_size_x_, tile_size_y_, block_size_x_, block_size_y_, reg_layout_ >::block_elems = block_size_x * block_size_y

staticconstexpr

block_size_x

template<uint32_t tile_size_x_, uint32_t tile_size_y_, uint32_t block_size_x_, uint32_t block_size_y_, reg_layout reg_layout_ = reg_layout::tiled>

constexpr uint32_t gpu::xetla::subgroup::tile_desc_t< tile_size_x_, tile_size_y_, block_size_x_, block_size_y_, reg_layout_ >::block_size_x = block_size_x_

staticconstexpr

block_size_y

template<uint32_t tile_size_x_, uint32_t tile_size_y_, uint32_t block_size_x_, uint32_t block_size_y_, reg_layout reg_layout_ = reg_layout::tiled>

constexpr uint32_t gpu::xetla::subgroup::tile_desc_t< tile_size_x_, tile_size_y_, block_size_x_, block_size_y_, reg_layout_ >::block_size_y = block_size_y_

staticconstexpr

num_block

template<uint32_t tile_size_x_, uint32_t tile_size_y_, uint32_t block_size_x_, uint32_t block_size_y_, reg_layout reg_layout_ = reg_layout::tiled>

constexpr uint32_t gpu::xetla::subgroup::tile_desc_t< tile_size_x_, tile_size_y_, block_size_x_, block_size_y_, reg_layout_ >::num_block = num_block_x * (num_block_y + (remained_size_y > 0 ? 1 : 0))

staticconstexpr

num_block_x

template<uint32_t tile_size_x_, uint32_t tile_size_y_, uint32_t block_size_x_, uint32_t block_size_y_, reg_layout reg_layout_ = reg_layout::tiled>

constexpr uint32_t gpu::xetla::subgroup::tile_desc_t< tile_size_x_, tile_size_y_, block_size_x_, block_size_y_, reg_layout_ >::num_block_x = tile_size_x / block_size_x

staticconstexpr

num_block_y

template<uint32_t tile_size_x_, uint32_t tile_size_y_, uint32_t block_size_x_, uint32_t block_size_y_, reg_layout reg_layout_ = reg_layout::tiled>

constexpr uint32_t gpu::xetla::subgroup::tile_desc_t< tile_size_x_, tile_size_y_, block_size_x_, block_size_y_, reg_layout_ >::num_block_y = tile_size_y / block_size_y

staticconstexpr

reg_transpose

template<uint32_t tile_size_x_, uint32_t tile_size_y_, uint32_t block_size_x_, uint32_t block_size_y_, reg_layout reg_layout_ = reg_layout::tiled>

constexpr bool gpu::xetla::subgroup::tile_desc_t< tile_size_x_, tile_size_y_, block_size_x_, block_size_y_, reg_layout_ >::reg_transpose = reg_layout_ == reg_layout::transpose_tiled

staticconstexpr

register_layout

template<uint32_t tile_size_x_, uint32_t tile_size_y_, uint32_t block_size_x_, uint32_t block_size_y_, reg_layout reg_layout_ = reg_layout::tiled>

constexpr reg_layout gpu::xetla::subgroup::tile_desc_t< tile_size_x_, tile_size_y_, block_size_x_, block_size_y_, reg_layout_ >::register_layout = reg_layout_

staticconstexpr

remained_size_y

template<uint32_t tile_size_x_, uint32_t tile_size_y_, uint32_t block_size_x_, uint32_t block_size_y_, reg_layout reg_layout_ = reg_layout::tiled>

constexpr uint32_t gpu::xetla::subgroup::tile_desc_t< tile_size_x_, tile_size_y_, block_size_x_, block_size_y_, reg_layout_ >::remained_size_y = tile_size_y % block_size_y

staticconstexpr

tile_elems

template<uint32_t tile_size_x_, uint32_t tile_size_y_, uint32_t block_size_x_, uint32_t block_size_y_, reg_layout reg_layout_ = reg_layout::tiled>

constexpr uint32_t gpu::xetla::subgroup::tile_desc_t< tile_size_x_, tile_size_y_, block_size_x_, block_size_y_, reg_layout_ >::tile_elems = tile_size_x * tile_size_y

staticconstexpr

tile_size_x

template<uint32_t tile_size_x_, uint32_t tile_size_y_, uint32_t block_size_x_, uint32_t block_size_y_, reg_layout reg_layout_ = reg_layout::tiled>

constexpr uint32_t gpu::xetla::subgroup::tile_desc_t< tile_size_x_, tile_size_y_, block_size_x_, block_size_y_, reg_layout_ >::tile_size_x = tile_size_x_

staticconstexpr

tile_size_y

template<uint32_t tile_size_x_, uint32_t tile_size_y_, uint32_t block_size_x_, uint32_t block_size_y_, reg_layout reg_layout_ = reg_layout::tiled>

constexpr uint32_t gpu::xetla::subgroup::tile_desc_t< tile_size_x_, tile_size_y_, block_size_x_, block_size_y_, reg_layout_ >::tile_size_y = tile_size_y_

staticconstexpr