gpu::xetla::subgroup::gelu_fwd_w_op_t< dtype_out, arch_tag, enable > Struct Template Reference

Is the element-wise gelu training forward op functor. More...

#include <tile_op_functor.hpp>

Detailed Description

template<typename dtype_out, gpu_arch arch_tag, class enable = void>
struct gpu::xetla::subgroup::gelu_fwd_w_op_t< dtype_out, arch_tag, enable >

Is the element-wise gelu training forward op functor.

Get the gelu input from matAcc, update the the gelu output in place, and dump the intermediate buffer_w to memory for backward purpose. Used in epilogue::tile_op or chained_tile_op.

Template Parameters

dtype_out	Is the data type of the intermediate buffer_w.
arch_tag	Is the hardware architecture tag.