1 /*******************************************************************************
2 * Copyright 2016-2018 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 *******************************************************************************/
17 #ifndef CPU_REF_SOFTMAX_HPP
18 #define CPU_REF_SOFTMAX_HPP
22 #include "c_types_map.hpp"
23 #include "memory_tracking.hpp"
24 #include "type_helpers.hpp"
27 #include "cpu_softmax_pd.hpp"
33 template <impl::data_type_t data_type>
34 struct ref_softmax_fwd_t: public cpu_primitive_t {
35 struct pd_t: public cpu_softmax_fwd_pd_t {
36 pd_t(engine_t *engine, const softmax_desc_t *adesc,
37 const primitive_attr_t *attr,
38 const softmax_fwd_pd_t *hint_fwd_pd)
39 : cpu_softmax_fwd_pd_t(engine, adesc, attr, hint_fwd_pd) {}
41 DECLARE_COMMON_PD_T("ref:any", ref_softmax_fwd_t);
43 virtual status_t init() override {
44 using namespace prop_kind;
45 assert(engine()->kind() == engine_kind::cpu);
47 && utils::one_of(desc()->prop_kind, forward_inference,
49 && data_pd_.desc()->data_type == data_type
50 && attr()->has_default_values();
51 if (!ok) return status::unimplemented;
55 return status::success;
59 void init_scratchpad() {
60 const int inner_size = utils::array_product(
61 desc()->data_desc.dims + desc()->softmax_axis + 1,
62 desc()->data_desc.ndims - desc()->softmax_axis - 1);
65 auto scratchpad = scratchpad_registry().registrar();
66 scratchpad.book(memory_tracking::names::key_softmax_reduction,
67 sizeof(data_t) * 2 * inner_size);
72 ref_softmax_fwd_t(const pd_t *apd, const input_vector &inputs,
73 const output_vector &outputs)
74 : cpu_primitive_t(apd, inputs, outputs)
76 auto ndims = pd()->desc()->data_desc.ndims;
77 auto dims = pd()->desc()->data_desc.dims;
78 auto axis = pd()->desc()->softmax_axis;
80 outer_size_ = utils::array_product(dims, axis);
81 channels_ = dims[axis];
82 inner_size_ = utils::array_product(dims + axis + 1, ndims - axis - 1);
84 const memory_desc_wrapper data_d(pd()->src_pd());
85 use_dense_ = inner_size_ == 1 && data_d.is_dense()
86 && data_d.blocking_desc().block_dims[axis] == 1
87 && data_d.blocking_desc().strides[0][axis] == 1;
89 ~ref_softmax_fwd_t() {}
91 typedef typename prec_traits<data_type>::type data_t;
93 virtual void execute(event_t *e) const {
94 if (use_dense_) execute_forward_dense();
95 else execute_forward_generic();
96 e->set_state(event_t::ready);
100 void execute_forward_dense() const;
101 void execute_forward_generic() const;
103 void _max(int n, const data_t *x, data_t *max_data) const;
104 void _sub(int n, data_t alpha, const data_t *x, data_t *y) const;
105 void _exp(int n, const data_t *a, data_t *r) const;
106 void _exp_parallel(int n, const data_t *a, data_t *r) const;
107 void _sum(int n, const data_t *x, data_t *sum_data) const;
108 void _scal(int n, data_t alpha, data_t *x) const;
109 void _scal_parallel(int n, data_t alpha, data_t *x) const;
111 const pd_t *pd() const { return (const pd_t *)primitive_t::pd(); }
114 int outer_size_, channels_, inner_size_;
117 template <impl::data_type_t data_type>
118 struct ref_softmax_bwd_t: public cpu_primitive_t {
119 struct pd_t: public cpu_softmax_bwd_pd_t {
120 pd_t(engine_t *engine, const softmax_desc_t *adesc,
121 const primitive_attr_t *attr,
122 const softmax_fwd_pd_t *hint_fwd_pd)
123 : cpu_softmax_bwd_pd_t(engine, adesc, attr, hint_fwd_pd) {}
125 DECLARE_COMMON_PD_T("ref:any", ref_softmax_bwd_t);
127 virtual status_t init() override {
128 using namespace prop_kind;
129 assert(engine()->kind() == engine_kind::cpu);
131 && utils::one_of(desc()->prop_kind, backward_data)
132 && diff_src_pd_.desc()->data_type == data_type
133 && diff_dst_pd_.desc()->data_type == data_type
134 && attr()->has_default_values();
135 if (!ok) return status::unimplemented;
137 return status::success;
141 ref_softmax_bwd_t(const pd_t *apd, const input_vector &inputs,
142 const output_vector &outputs)
143 : cpu_primitive_t(apd, inputs, outputs) {
144 auto dims = pd()->desc()->diff_desc.dims;
145 auto axis = pd()->desc()->softmax_axis;
146 auto ndims = pd()->desc()->diff_desc.ndims;
148 outer_size_ = utils::array_product(dims, axis);
149 channels_ = dims[axis];
150 inner_size_ = utils::array_product(dims + axis + 1, ndims - axis - 1);
152 // Diff desc as well as data desc whould be checked
153 const memory_desc_wrapper data_d(pd()->dst_pd());
154 const memory_desc_wrapper diff_d(pd()->diff_dst_pd());
159 && diff_d.blocking_desc().block_dims[axis] == 1
160 && diff_d.blocking_desc().strides[0][axis] == 1;
162 ~ref_softmax_bwd_t() {}
164 typedef typename prec_traits<data_type>::type data_t;
166 virtual void execute(event_t *e) const {
167 if (use_dense_) execute_backward_dense();
168 else execute_backward_generic();
169 e->set_state(event_t::ready);
173 void execute_backward_dense() const;
174 void execute_backward_generic() const;
175 const pd_t *pd() const { return (const pd_t *)primitive_t::pd(); }
178 int outer_size_, channels_, inner_size_;
188 // vim: et ts=4 sw=4 cindent cino^=l0,\:0,N-s