1 | /******************************************************************************* |
2 | * Copyright 2020-2022 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_PRELU_HPP |
18 | #define CPU_REF_PRELU_HPP |
19 | |
20 | #include <assert.h> |
21 | |
22 | #include "common/c_types_map.hpp" |
23 | #include "common/dnnl_thread.hpp" |
24 | #include "common/memory_tracking.hpp" |
25 | #include "common/primitive.hpp" |
26 | #include "common/type_helpers.hpp" |
27 | #include "common/utils.hpp" |
28 | |
29 | #include "cpu/platform.hpp" |
30 | |
31 | #include "common/broadcast_strategy.hpp" |
32 | #include "cpu/cpu_prelu_pd.hpp" |
33 | |
34 | namespace dnnl { |
35 | namespace impl { |
36 | namespace cpu { |
37 | |
38 | namespace prelu { |
39 | void set_reduction_buffers( |
40 | const dim_t work_amount, dim_t &group_size, dim_t &buf_size); |
41 | dim_t get_scalar_scratchpad_offset(const std::size_t ithr, |
42 | const std::size_t nthr, const dim_t work_amount); |
43 | } // namespace prelu |
44 | |
45 | using byte = unsigned char; |
46 | |
47 | struct ref_prelu_fwd_t : public primitive_t { |
48 | struct pd_t : public cpu_prelu_fwd_pd_t { |
49 | using cpu_prelu_fwd_pd_t::cpu_prelu_fwd_pd_t; |
50 | |
51 | DECLARE_COMMON_PD_T("ref:any" , ref_prelu_fwd_t); |
52 | |
53 | status_t init(engine_t *engine) { |
54 | using namespace data_type; |
55 | bool ok = is_fwd() && src_md(0)->data_type == dst_md(0)->data_type |
56 | && platform::has_data_type_support(src_md(0)->data_type) |
57 | && platform::has_data_type_support(weights_md(0)->data_type) |
58 | && attr()->has_default_values() && set_default_formats() |
59 | && memory_desc_wrapper(src_md()) |
60 | == memory_desc_wrapper(dst_md()); |
61 | if (!ok) return status::unimplemented; |
62 | |
63 | return status::success; |
64 | } |
65 | }; |
66 | |
67 | ref_prelu_fwd_t(const pd_t *apd) : primitive_t(apd) {} |
68 | |
69 | status_t execute(const exec_ctx_t &ctx) const override { |
70 | return execute_forward(ctx); |
71 | } |
72 | |
73 | private: |
74 | status_t execute_forward(const exec_ctx_t &ctx) const; |
75 | const pd_t *pd() const { return (const pd_t *)primitive_t::pd().get(); } |
76 | }; |
77 | |
78 | struct ref_prelu_bwd_t : public primitive_t { |
79 | struct pd_t : public cpu_prelu_bwd_pd_t { |
80 | using cpu_prelu_bwd_pd_t::cpu_prelu_bwd_pd_t; |
81 | |
82 | DECLARE_COMMON_PD_T("ref:any" , ref_prelu_bwd_t); |
83 | |
84 | status_t init(engine_t *engine) { |
85 | using namespace data_type; |
86 | bool ok = !is_fwd() |
87 | && diff_src_md(0)->data_type == src_md(0)->data_type |
88 | && diff_weights_md(0)->data_type == weights_md(0)->data_type |
89 | && diff_dst_md(0)->data_type == diff_src_md(0)->data_type |
90 | && platform::has_data_type_support(src_md(0)->data_type) |
91 | && platform::has_data_type_support(weights_md(0)->data_type) |
92 | && attr()->has_default_values() && set_default_formats() |
93 | && memory_desc_wrapper(diff_dst_md()) |
94 | == memory_desc_wrapper(diff_src_md()); |
95 | if (!ok) return status::unimplemented; |
96 | |
97 | init_scratchpad(); |
98 | |
99 | return status::success; |
100 | } |
101 | |
102 | int nthr_; // To not exceed the limit in execute used for set up. |
103 | |
104 | private: |
105 | void init_scratchpad() { |
106 | auto scratchpad = this->scratchpad_registry().registrar(); |
107 | dim_t scratchpad_size; |
108 | const memory_desc_wrapper src_d(src_md()); |
109 | const memory_desc_wrapper weights_d(weights_md()); |
110 | auto broadcast_strategy |
111 | = get_rhs_arg_broadcasting_strategy(*weights_md(), src_d); |
112 | // Assign `nthr_` here since the amount needed maybe reduced. |
113 | nthr_ = dnnl_get_max_threads(); |
114 | // Scratchpad is needed to correctly reduce calculated diff_weights |
115 | // in cases where broadcast is used. |
116 | // |
117 | // example: if data tensor size is NxCxW and weight tensor is 1xCx1, |
118 | // diff_weight tensor would also be of size 1xCx1 and thus each value |
119 | // along C axis would equal: results summed up over N and W for given C. |
120 | // |
121 | // In current implementation reduction is 2 step: |
122 | // results are first copied to buffer and reduced, result is then |
123 | // stored in group buffer. Values in group buffer are then reduced |
124 | // to obtain final value. |
125 | if (broadcast_strategy == broadcasting_strategy_t::no_broadcast) { |
126 | return; |
127 | } else if (broadcast_strategy == broadcasting_strategy_t::scalar) { |
128 | int work_amount = static_cast<int>(src_d.nelems()); |
129 | nthr_ = nstl::min(nthr_, work_amount); |
130 | scratchpad_size = prelu::get_scalar_scratchpad_offset( |
131 | nthr_, nthr_, src_d.nelems()); |
132 | } else { |
133 | dim_t group_size, buf_size; |
134 | nthr_ = nstl::min(nthr_, static_cast<int>(weights_d.nelems())); |
135 | dim_t work_amount = src_d.nelems() / weights_d.nelems(); |
136 | prelu::set_reduction_buffers(work_amount, group_size, buf_size); |
137 | scratchpad_size = nthr_ * (group_size + buf_size); |
138 | } |
139 | scratchpad.book(memory_tracking::names::key_prelu_reduction, |
140 | scratchpad_size, types::data_type_size(dnnl_f32)); |
141 | } |
142 | }; |
143 | |
144 | ref_prelu_bwd_t(const pd_t *apd) : primitive_t(apd) {} |
145 | |
146 | status_t execute(const exec_ctx_t &ctx) const override { |
147 | return execute_backward(ctx); |
148 | } |
149 | |
150 | private: |
151 | status_t execute_backward(const exec_ctx_t &ctx) const; |
152 | const pd_t *pd() const { return (const pd_t *)primitive_t::pd().get(); } |
153 | |
154 | float ker(const byte *src, const byte *weights, const byte *diff_dst, |
155 | byte *diff_src, dim_t data_off, dim_t weight_off) const; |
156 | void calculate_scalar(const byte *src, const byte *weights, |
157 | byte *diff_weights, const byte *diff_dst, byte *diff_src, |
158 | float *scratchpad_buf) const; |
159 | void calculate_no_broadcast(const byte *src, const byte *weights, |
160 | byte *diff_weights, const byte *diff_dst, byte *diff_src, |
161 | float *scratchpad_buf) const; |
162 | void calculate_shared_axes(const byte *src, const byte *weights, |
163 | byte *diff_weights, const byte *diff_dst, byte *diff_src, |
164 | float *scratchpad_buf) const; |
165 | }; |
166 | |
167 | } // namespace cpu |
168 | } // namespace impl |
169 | } // namespace dnnl |
170 | |
171 | #endif |
172 | |
173 | // vim: et ts=4 sw=4 cindent cino+=l0,\:4,N-s |
174 | |