#include <kernel_func.hpp>
|
| static void | run (sycl::nd_item< 3 > &item, input_T *layer_ptr, input_T *h0_ptr, input_T *W_ir_ptr, input_T *W_hr_ptr, input_T *W_iz_ptr, input_T *W_hz_ptr, input_T *W_in_ptr, input_T *W_hn_ptr, input_T *layer_out_ptr, input_T *hidden_out_ptr, input_T *ping_pong_buffer, input_T *ping_pong_cell, int batch_size, int input_size, int hidden_size, int sequence_length, int layer_size) |
| |
◆ run()
template<typename input_T , typename Act_T , uint32_t wg_tile_m_t, uint32_t wg_tile_n_t, uint32_t sg_tile_m_t, uint32_t sg_tile_n_t, uint32_t sg_tile_k_t>
| static void kernel_xcoder_gru_fusion< input_T, Act_T, wg_tile_m_t, wg_tile_n_t, sg_tile_m_t, sg_tile_n_t, sg_tile_k_t >::run |
( |
sycl::nd_item< 3 > & |
item, |
|
|
input_T * |
layer_ptr, |
|
|
input_T * |
h0_ptr, |
|
|
input_T * |
W_ir_ptr, |
|
|
input_T * |
W_hr_ptr, |
|
|
input_T * |
W_iz_ptr, |
|
|
input_T * |
W_hz_ptr, |
|
|
input_T * |
W_in_ptr, |
|
|
input_T * |
W_hn_ptr, |
|
|
input_T * |
layer_out_ptr, |
|
|
input_T * |
hidden_out_ptr, |
|
|
input_T * |
ping_pong_buffer, |
|
|
input_T * |
ping_pong_cell, |
|
|
int |
batch_size, |
|
|
int |
input_size, |
|
|
int |
hidden_size, |
|
|
int |
sequence_length, |
|
|
int |
layer_size |
|
) |
| |
|
inlinestatic |
- Parameters
-
| item | Is the sycl::nd_item |
| layer_ptr | input from previous layer i.e X_t |
| h0_ptr | hx_ptr input i.e. h_{0} shape = layer_size x batch_size x hidden_size weights |
| W_ir_ptr | weights with input of reset gate, (input_weight_size, hidden_weight_size, ...) |
| W_hr_ptr | weights with hidden input of reset gate, shape = layer_size x hidden_weight_size |
| W_iz_ptr | weights with input of input gate, (input_weight_size, hidden_weight_size, ...) |
| W_hz_ptr | weights with hidden input of input gate, shape = layer_size x hidden_weight_size |
| W_in_ptr | weights with input of new gate, (input_weight_size, hidden_weight_size, ...) |
| W_hn_ptr | weights with hidden input of new gate, shape = layer_size x hidden_weight_size output |
| layer_out_ptr | the last cell per layer output, shape = layer_size x batch_size x hidden_size |
| hidden_out_ptr | the last layer output for per gru cell, shape = sequence_length x batch_size x hidden_size |
1.9.7