1 | /******************************************************************************* |
2 | * Copyright 2016-2022 Intel Corporation |
3 | * |
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 |
7 | * |
8 | * http://www.apache.org/licenses/LICENSE-2.0 |
9 | * |
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 | *******************************************************************************/ |
16 | |
17 | #ifndef CPU_REF_BATCH_NORMALIZATION_HPP |
18 | #define CPU_REF_BATCH_NORMALIZATION_HPP |
19 | |
20 | #include <assert.h> |
21 | |
22 | #include "common/c_types_map.hpp" |
23 | #include "common/primitive.hpp" |
24 | #include "common/type_helpers.hpp" |
25 | #include "common/utils.hpp" |
26 | |
27 | #include "cpu/platform.hpp" |
28 | |
29 | #include "cpu/cpu_batch_normalization_pd.hpp" |
30 | |
31 | namespace dnnl { |
32 | namespace impl { |
33 | namespace cpu { |
34 | |
35 | template <data_type_t d_type> |
36 | struct ref_batch_normalization_fwd_t : public primitive_t { |
37 | struct pd_t : public cpu_batch_normalization_fwd_pd_t { |
38 | pd_t(const batch_normalization_desc_t *adesc, |
39 | const primitive_attr_t *attr, |
40 | const batch_normalization_fwd_pd_t *hint_fwd_pd) |
41 | : cpu_batch_normalization_fwd_pd_t(adesc, attr, hint_fwd_pd) {} |
42 | |
43 | DECLARE_COMMON_PD_T("ref:any" , ref_batch_normalization_fwd_t); |
44 | |
45 | status_t init(engine_t *engine) { |
46 | using namespace data_type; |
47 | bool ok = is_fwd() |
48 | && utils::everyone_is( |
49 | d_type, src_md()->data_type, dst_md()->data_type) |
50 | && platform::has_data_type_support(d_type) |
51 | && IMPLICATION(is_training(), |
52 | platform::has_training_support(d_type)) |
53 | && check_scale_shift_data_type() |
54 | && (attr()->has_default_values() |
55 | || with_relu_post_op(is_training())) |
56 | && set_default_formats_common() |
57 | && memory_desc_wrapper(src_md()) |
58 | == memory_desc_wrapper(dst_md()); |
59 | if (!ok) return status::unimplemented; |
60 | |
61 | // BN+Add+Relu fusion is not currently implemented |
62 | if (fuse_norm_add_relu()) return status::unimplemented; |
63 | |
64 | if (src_md()->data_type == s8 && !stats_is_src()) |
65 | return status::unimplemented; |
66 | |
67 | if (is_training() && fuse_norm_relu()) init_default_ws(8); |
68 | |
69 | return status::success; |
70 | } |
71 | }; |
72 | |
73 | ref_batch_normalization_fwd_t(const pd_t *apd) : primitive_t(apd) {} |
74 | |
75 | typedef typename prec_traits<d_type>::type data_t; |
76 | |
77 | status_t execute(const exec_ctx_t &ctx) const override { |
78 | return execute_forward(ctx); |
79 | } |
80 | |
81 | private: |
82 | status_t execute_forward(const exec_ctx_t &ctx) const; |
83 | const pd_t *pd() const { return (const pd_t *)primitive_t::pd().get(); } |
84 | }; |
85 | |
86 | template <data_type_t d_type> |
87 | struct ref_batch_normalization_bwd_t : public primitive_t { |
88 | struct pd_t : public cpu_batch_normalization_bwd_pd_t { |
89 | pd_t(const batch_normalization_desc_t *adesc, |
90 | const primitive_attr_t *attr, |
91 | const batch_normalization_fwd_pd_t *hint_fwd_pd) |
92 | : cpu_batch_normalization_bwd_pd_t(adesc, attr, hint_fwd_pd) {} |
93 | |
94 | DECLARE_COMMON_PD_T("ref:any" , ref_batch_normalization_bwd_t); |
95 | |
96 | status_t init(engine_t *engine) { |
97 | using namespace data_type; |
98 | |
99 | bool ok = !is_fwd() |
100 | && utils::everyone_is(d_type, src_md()->data_type, |
101 | diff_dst_md()->data_type, diff_src_md()->data_type) |
102 | && platform::has_data_type_support(d_type) |
103 | && platform::has_training_support(d_type) |
104 | && check_scale_shift_data_type() |
105 | && attr()->has_default_values() |
106 | && set_default_formats_common() |
107 | && memory_desc_wrapper(diff_src_md()) |
108 | == memory_desc_wrapper(diff_dst_md()); |
109 | if (!ok) return status::unimplemented; |
110 | |
111 | // BN+Add+Relu fusion is not currently implemented |
112 | if (fuse_norm_add_relu()) return status::unimplemented; |
113 | |
114 | if (fuse_norm_relu()) { |
115 | init_default_ws(8); |
116 | if (!compare_ws(hint_fwd_pd_)) return status::unimplemented; |
117 | } |
118 | |
119 | return status::success; |
120 | } |
121 | }; |
122 | |
123 | ref_batch_normalization_bwd_t(const pd_t *apd) : primitive_t(apd) {} |
124 | typedef typename prec_traits<d_type>::type data_t; |
125 | |
126 | status_t execute(const exec_ctx_t &ctx) const override { |
127 | return execute_backward(ctx); |
128 | } |
129 | |
130 | private: |
131 | status_t execute_backward(const exec_ctx_t &ctx) const; |
132 | const pd_t *pd() const { return (const pd_t *)primitive_t::pd().get(); } |
133 | }; |
134 | |
135 | } // namespace cpu |
136 | } // namespace impl |
137 | } // namespace dnnl |
138 | |
139 | #endif |
140 | |
141 | // vim: et ts=4 sw=4 cindent cino+=l0,\:4,N-s |
142 | |