gpu::xetla::kernel::dispatch_policy_int4_dequantize_kslicing< group_swizzle_policy_, num_global_kslicing_, num_local_kslicing_ > Struct Template Reference

4bit kslicing GEMM implementation. More...

#include <dispatch_policy.hpp>

Public Types
using	group_swizzle_policy = group_swizzle_policy_

Static Public Attributes
static constexpr int	num_global_kslicing = num_global_kslicing_
static constexpr int	num_local_kslicing = num_local_kslicing_
static constexpr gpu_arch	arch_tag = group_swizzle_policy::arch_tag

Detailed Description

template<typename group_swizzle_policy_, int num_global_kslicing_ = 1, int num_local_kslicing_ = 1>
struct gpu::xetla::kernel::dispatch_policy_int4_dequantize_kslicing< group_swizzle_policy_, num_global_kslicing_, num_local_kslicing_ >

4bit kslicing GEMM implementation.

A special GEMM implementation to increase the hardware occupancy by splitting the GEMM task along k dimension. It includes inter-group reduction (by using global atomic) and intra-group reduction (by using local memory for data exchange).

Note

The difference compare with dispatch_policy_kslicing is we will add additional handling for 4bit.

Template Parameters

num_global_kslicing_	Is the k dim split ratio between groups.
num_local_kslicing_	Is the k dim split ratio within a group.
arch_tag_	Is the HW architecture.

Member Typedef Documentation

group_swizzle_policy

template<typename group_swizzle_policy_ , int num_global_kslicing_ = 1, int num_local_kslicing_ = 1>

using gpu::xetla::kernel::dispatch_policy_int4_dequantize_kslicing< group_swizzle_policy_, num_global_kslicing_, num_local_kslicing_ >::group_swizzle_policy = group_swizzle_policy_

Member Data Documentation

arch_tag

template<typename group_swizzle_policy_ , int num_global_kslicing_ = 1, int num_local_kslicing_ = 1>

constexpr gpu_arch gpu::xetla::kernel::dispatch_policy_int4_dequantize_kslicing< group_swizzle_policy_, num_global_kslicing_, num_local_kslicing_ >::arch_tag = group_swizzle_policy::arch_tag

staticconstexpr

num_global_kslicing

template<typename group_swizzle_policy_ , int num_global_kslicing_ = 1, int num_local_kslicing_ = 1>

constexpr int gpu::xetla::kernel::dispatch_policy_int4_dequantize_kslicing< group_swizzle_policy_, num_global_kslicing_, num_local_kslicing_ >::num_global_kslicing = num_global_kslicing_

staticconstexpr

num_local_kslicing

template<typename group_swizzle_policy_ , int num_global_kslicing_ = 1, int num_local_kslicing_ = 1>

constexpr int gpu::xetla::kernel::dispatch_policy_int4_dequantize_kslicing< group_swizzle_policy_, num_global_kslicing_, num_local_kslicing_ >::num_local_kslicing = num_local_kslicing_

staticconstexpr