1 /*******************************************************************************
2 * Copyright 2018 Intel Corporation
4 * Licensed under the Apache License, Version 2.0 (the "License");
5 * you may not use this file except in compliance with the License.
6 * You may obtain a copy of the License at
8 * http://www.apache.org/licenses/LICENSE-2.0
10 * Unless required by applicable law or agreed to in writing, software
11 * distributed under the License is distributed on an "AS IS" BASIS,
12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 * See the License for the specific language governing permissions and
14 * limitations under the License.
15 *******************************************************************************/
22 #include "c_types_map.hpp"
23 #include "memory_pd.hpp"
24 #include "primitive_desc.hpp"
29 // struct rnn_fwd_pd_t;
31 struct rnn_pd_t : public primitive_desc_t {
32 static constexpr auto base_pkind = primitive_kind::rnn;
34 rnn_pd_t(mkldnn::impl::engine_t *engine, const rnn_desc_t *adesc,
35 const primitive_attr_t *attr, const rnn_pd_t *hint_pd)
36 : primitive_desc_t(engine, attr, primitive_kind::rnn)
38 , hint_pd_(hint_pd) {}
39 virtual ~rnn_pd_t() {}
41 const rnn_desc_t *desc() const { return &desc_; }
42 virtual const op_desc_t *op_desc() const override {
43 return reinterpret_cast<const op_desc_t *>(this->desc());
45 virtual void init_info() override { init_info_rnn(this, this->info_); }
47 virtual status_t query(query_t what, int idx, void *result) const override {
49 case query::rnn_d: *(const rnn_desc_t **)result = desc(); break;
50 default: return primitive_desc_t::query(what, idx, result);
52 return status::success;
55 inline bool is_training() const {
56 return utils::one_of(desc_.prop_kind, prop_kind::forward_training,
60 inline size_t ws_states_size() {
61 int wic = nstl::max(SLC(), nstl::max(SIC(), DIC()));
62 return (size_t)(L() + 1) * D() * (T() + 1) * S() * MB() * wic;
65 inline size_t ws_diff_states_size() {
66 int wic = nstl::max(SLC(), nstl::max(SIC(), DIC()));
67 return (size_t)(L() + 1) * D() * (T() + 1) * (S() + 1) * MB() * wic;
70 inline size_t ws_gates_size() {
72 int n_direction = D();
78 return (size_t)n_layer * n_direction * n_iter * batch * n_gates
82 inline size_t ws_cell_comp_size() {
86 return (size_t)is_lbr() * n_gates * batch * s_size;
89 inline size_t ws_grid_comp_size() {
91 int n_direction = D();
95 return (size_t)is_lbr() * is_training() * n_layer * n_direction * n_iter
99 inline int ws_per_cell() {
102 return is_lbr() * is_training() * batch * s_size;
105 inline void set_offsets(size_t &ws_gates_offset, size_t &ws_states_offset,
106 size_t &ws_diff_states_offset, size_t &ws_grid_comp_offset,
107 size_t &ws_cell_comp_offset) {
108 const size_t page_size = 4096; // 2097152;
110 = 0; // assumes the workspace base pointer is page aligned
111 ws_states_offset = utils::rnd_up(ws_gates_size(), page_size);
112 ws_diff_states_offset
113 = utils::rnd_up(ws_states_offset + ws_states_size(), page_size);
114 ws_grid_comp_offset = utils::rnd_up(ws_diff_states_offset
115 + ws_diff_states_size(), page_size);
117 ws_cell_comp_offset = utils::rnd_up(ws_grid_comp_offset
118 + ws_grid_comp_size(), page_size);
121 inline size_t get_ws_size() {
122 size_t ws_gates_offset, ws_states_offset, ws_diff_states_offset,
123 ws_grid_comp_offset, ws_cell_comp_offset;
125 ws_gates_offset, ws_states_offset, ws_diff_states_offset,
126 ws_grid_comp_offset, ws_cell_comp_offset);
127 return ws_grid_comp_offset + ws_grid_comp_size();
130 inline size_t get_scratchpad_size() {
131 size_t ws_gates_offset, ws_states_offset, ws_diff_states_offset,
132 ws_grid_comp_offset, ws_cell_comp_offset;
134 ws_gates_offset, ws_states_offset, ws_diff_states_offset,
135 ws_grid_comp_offset, ws_cell_comp_offset);
136 if (desc_.prop_kind == prop_kind::forward_inference)
137 return ws_cell_comp_offset + ws_cell_comp_size();
139 return ws_cell_comp_size();
142 int T() const { return desc_.src_layer_desc.dims[0]; }
143 int MB() const { return desc_.src_layer_desc.dims[1]; }
145 int L() const { return desc_.weights_layer_desc.dims[0]; }
146 int D() const { return desc_.weights_layer_desc.dims[1]; }
148 int SIC() const { return desc_.weights_iter_desc.dims[2]; }
150 int SLC() const { return desc_.weights_layer_desc.dims[2]; }
151 int G() const { return desc_.weights_layer_desc.dims[3]; }
152 int DIC() const { return desc_.weights_layer_desc.dims[4]; }
154 int DLC() const { return desc_.dst_layer_desc.dims[2]; }
156 int S() const { return mkldnn_rnn_cell_get_states_count(&desc_.cell_desc); }
158 bool with_bias() const {
159 return !memory_desc_wrapper(desc_.bias_desc).is_zero();
162 bool with_src_iter() const {
163 return !(memory_desc_wrapper(desc_.src_iter_desc).is_zero());
166 bool with_dst_iter() const {
167 return !memory_desc_wrapper(desc_.dst_iter_desc).is_zero();
170 mkldnn::impl::alg_kind_t cell_kind() const {
171 return desc_.cell_desc.cell_kind;
173 mkldnn::impl::alg_kind_t activation_kind() const {
174 return desc_.cell_desc.activation_kind;
177 bool is_lbr() const {
178 return cell_kind() == mkldnn_gru_linear_before_reset;
181 mkldnn_rnn_direction_t direction() const { return desc_.direction; }
185 const rnn_pd_t *hint_pd_;
188 struct rnn_fwd_pd_t : public rnn_pd_t {
189 typedef rnn_fwd_pd_t base_class;
190 typedef rnn_fwd_pd_t hint_class;
192 using rnn_pd_t::rnn_pd_t;
193 virtual ~rnn_fwd_pd_t() {}
195 virtual const memory_pd_t *input_pd(int index = 0) const override {
197 case 0: return src_pd(0);
198 case 1: return src_pd(1);
199 case 2: return weights_pd(0);
200 case 3: return weights_pd(1);
201 case 4: return weights_pd(2);
202 default: return nullptr;
206 virtual const memory_pd_t *output_pd(int index = 0) const override {
208 case 0: return dst_pd(0);
209 case 1: return dst_pd(1);
210 case 2: return workspace_pd();
211 default: return nullptr;
215 virtual int n_inputs() const override {
216 return 3 + with_bias() + with_src_iter();
219 virtual int n_outputs() const override {
220 return 1 + with_dst_iter() + is_training();
223 int ws_idx() const { return 1 + with_dst_iter(); }
226 struct rnn_bwd_pd_t : public rnn_pd_t {
227 typedef rnn_bwd_pd_t base_class;
228 typedef rnn_bwd_pd_t hint_class;
230 using rnn_pd_t::rnn_pd_t;
231 virtual ~rnn_bwd_pd_t() {}
233 virtual const memory_pd_t *input_pd(int index = 0) const override {
235 case 0: return src_pd(0);
236 case 1: return src_pd(1);
237 case 2: return weights_pd(0);
238 case 3: return weights_pd(1);
239 case 4: return weights_pd(2);
240 case 5: return dst_pd(0);
241 case 6: return dst_pd(1);
242 case 7: return diff_dst_pd(0);
243 case 8: return diff_dst_pd(1);
244 case 9: return workspace_pd();
245 default: return nullptr;
249 virtual const memory_pd_t *output_pd(int index = 0) const override {
251 case 0: return diff_src_pd(0);
252 case 1: return diff_src_pd(1);
253 case 2: return diff_weights_pd(0);
254 case 3: return diff_weights_pd(1);
255 case 4: return diff_weights_pd(2);
256 default: return nullptr;
260 virtual int n_inputs() const override {
261 return 6 + with_src_iter() + with_bias() + 2 * with_dst_iter();
263 virtual int n_outputs() const override {
264 return 3 + with_src_iter() + with_bias();
268 return 5 + with_src_iter() + with_bias() + 2 * with_dst_iter();