summaryrefslogtreecommitdiff
path: root/thirdparty/oidn/mkl-dnn/src/cpu/gemm_convolution.hpp
blob: 302e46369a3fb940923d494d8f89bf397b3f0034 (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
/*******************************************************************************
* Copyright 2016-2018 Intel Corporation
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
*     http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*******************************************************************************/

#ifndef CPU_JIT_GEMM_CONVOLUTION_HPP
#define CPU_JIT_GEMM_CONVOLUTION_HPP

#include "c_types_map.hpp"
#include "memory_tracking.hpp"

#include "gemm_convolution_utils.hpp"
#include "gemm/gemm.hpp"
#include "ref_eltwise.hpp"

#include "cpu_convolution_pd.hpp"

namespace mkldnn {
namespace impl {
namespace cpu {

struct gemm_convolution_fwd_t: public cpu_primitive_t {
    struct pd_t: public cpu_convolution_fwd_pd_t {
        pd_t(engine_t *engine,
                const convolution_desc_t *adesc, const primitive_attr_t *attr,
                const typename pd_t::base_class *hint_fwd_pd)
            : cpu_convolution_fwd_pd_t(engine, adesc, attr, hint_fwd_pd)
            , jcp_() {}

        DECLARE_COMMON_PD_T(GEMM_IMPL_STR, gemm_convolution_fwd_t);

        status_t init() {
            bool ok = true
                && is_fwd()
                && set_default_alg_kind(alg_kind::convolution_direct)
                && expect_data_types(data_type::f32, data_type::f32,
                        data_type::f32, data_type::f32, data_type::f32)
                && !has_zero_dim_memory()
                && set_default_formats_common(dat_tag(), wei_tag(), dat_tag())
                && post_ops_ok()
                && memory_desc_matches_tag(*src_md(), dat_tag())
                && memory_desc_matches_tag(*dst_md(), dat_tag())
                && memory_desc_matches_tag(*weights_md(), wei_tag());
            if (!ok) return status::unimplemented;

            auto scratchpad = scratchpad_registry().registrar();
            return jit_gemm_convolution_utils::init_conf(jcp_, scratchpad,
                    *desc(), src_md(), weights_md(0), dst_md(),
                    mkldnn_get_max_threads());
        }

        jit_gemm_conv_conf_t jcp_;

    protected:
        format_tag_t dat_tag() const {
            using namespace format_tag;
            return utils::pick(ndims() - 3, ncw, nchw, ncdhw);
        }

        format_tag_t wei_tag() const {
            using namespace format_tag;
            return with_groups()
                ? utils::pick(ndims() - 3, goiw, goihw, goidhw)
                : utils::pick(ndims() - 3, oiw, oihw, oidhw);
        }

        bool post_ops_ok() const {
            auto const &po = attr()->post_ops_;
            auto is_eltwise = [&](int idx)
            { return po.entry_[idx].is_eltwise(); };
            auto is_sum = [&](int idx) { return po.entry_[idx].is_sum(); };

            switch (po.len_) {
            case 0: return true; // no post_ops
            case 1: return is_eltwise(0) || is_sum(0); // sum OR eltwise
            case 2: return is_sum(0) && is_eltwise(1); // sum -> eltwise
            default: return false;
            }
            return false;
        }
    };

    gemm_convolution_fwd_t(const pd_t *apd)
        : cpu_primitive_t(apd, true)
        , eltwise_(nullptr)
    {
        const auto &post_ops = pd()->attr()->post_ops_;
        const data_t one = 1.0, zero = 0.0;
        beta_ = post_ops.find(primitive_kind::sum) >= 0 ? one : zero;

        const int entry_idx = post_ops.find(primitive_kind::eltwise);
        if (entry_idx != -1) eltwise_ = new ref_eltwise_scalar_fwd_t(
                post_ops.entry_[entry_idx].eltwise);
    }

    ~gemm_convolution_fwd_t() { delete eltwise_; }

    typedef typename prec_traits<data_type::f32>::type data_t;

    virtual status_t execute(const exec_ctx_t &ctx) const override {
        execute_forward(ctx);
        return status::success;
    }

private:
    void execute_forward(const exec_ctx_t &ctx) const;
    const pd_t *pd() const { return (const pd_t *)primitive_t::pd(); }

    data_t beta_;

    ref_eltwise_scalar_fwd_t* eltwise_;
};

struct gemm_convolution_bwd_data_t: public cpu_primitive_t {
    struct pd_t: public cpu_convolution_bwd_data_pd_t {
        pd_t(engine_t *engine,
                const convolution_desc_t *adesc, const primitive_attr_t *attr,
                const convolution_fwd_pd_t *hint_fwd_pd)
            : cpu_convolution_bwd_data_pd_t(engine, adesc, attr, hint_fwd_pd)
            , jcp_() {}

        DECLARE_COMMON_PD_T(GEMM_IMPL_STR, gemm_convolution_bwd_data_t);

        status_t init() {
            bool ok = true
                && desc()->prop_kind == prop_kind::backward_data
                && set_default_alg_kind(alg_kind::convolution_direct)
                && expect_data_types(data_type::f32, data_type::f32,
                        data_type::undef, data_type::f32, data_type::f32)
                && !has_zero_dim_memory()
                && set_default_formats_common(dat_tag(), wei_tag(), dat_tag())
                && memory_desc_matches_tag(*diff_src_md(), dat_tag())
                && memory_desc_matches_tag(*diff_dst_md(), dat_tag())
                && memory_desc_matches_tag(*weights_md(), wei_tag());
            if (!ok) return status::unimplemented;

            auto scratchpad = scratchpad_registry().registrar();
            return jit_gemm_convolution_utils::init_conf(jcp_, scratchpad,
                    *desc(), diff_src_md(), weights_md(0), diff_dst_md(),
                    mkldnn_get_max_threads());
        }

        jit_gemm_conv_conf_t jcp_;

    protected:
        format_tag_t dat_tag() const {
            using namespace format_tag;
            return utils::pick(ndims() - 3, ncw, nchw, ncdhw);
        }

        format_tag_t wei_tag() const {
            using namespace format_tag;
            return with_groups()
                ? utils::pick(ndims() - 3, goiw, goihw, goidhw)
                : utils::pick(ndims() - 3, oiw, oihw, oidhw);
        }
    };

    gemm_convolution_bwd_data_t(const pd_t *apd)
        : cpu_primitive_t(apd, true) {}

    typedef typename prec_traits<data_type::f32>::type data_t;

    virtual status_t execute(const exec_ctx_t &ctx) const override {
        execute_backward_data(ctx);
        return status::success;
    }

private:
    void execute_backward_data(const exec_ctx_t &ctx) const;
    const pd_t *pd() const { return (const pd_t *)primitive_t::pd(); }
};

struct gemm_convolution_bwd_weights_t: public cpu_primitive_t {
    struct pd_t: public cpu_convolution_bwd_weights_pd_t {
        pd_t(engine_t *engine,
                const convolution_desc_t *adesc,
                const primitive_attr_t *attr,
                const convolution_fwd_pd_t *hint_fwd_pd)
            : cpu_convolution_bwd_weights_pd_t(engine, adesc, attr, hint_fwd_pd)
            , jcp_() {}

        DECLARE_COMMON_PD_T(GEMM_IMPL_STR, gemm_convolution_bwd_weights_t);

        status_t init() {
            bool ok = true
                && desc()->prop_kind == prop_kind::backward_weights
                && set_default_alg_kind(alg_kind::convolution_direct)
                && expect_data_types(data_type::f32, data_type::f32,
                        data_type::f32, data_type::f32, data_type::f32)
                && !has_zero_dim_memory()
                && set_default_formats_common(dat_tag(), wei_tag(), dat_tag())
                && memory_desc_matches_tag(*src_md(), dat_tag())
                && memory_desc_matches_tag(*diff_dst_md(), dat_tag())
                && memory_desc_matches_tag(*diff_weights_md(), wei_tag());
            if (!ok) return status::unimplemented;

            auto scratchpad = scratchpad_registry().registrar();
            return jit_gemm_convolution_utils::init_conf(jcp_, scratchpad,
                    *desc(), src_md(), diff_weights_md(0), diff_dst_md(),
                    mkldnn_get_max_threads());
        }

        jit_gemm_conv_conf_t jcp_;

    protected:
        format_tag_t dat_tag() const {
            using namespace format_tag;
            return utils::pick(ndims() - 3, ncw, nchw, ncdhw);
        }

        format_tag_t wei_tag() const {
            using namespace format_tag;
            return with_groups()
                ? utils::pick(ndims() - 3, goiw, goihw, goidhw)
                : utils::pick(ndims() - 3, oiw, oihw, oidhw);
        }
    };

    gemm_convolution_bwd_weights_t(const pd_t *apd)
        : cpu_primitive_t(apd, true) {}

    typedef typename prec_traits<data_type::f32>::type data_t;

    virtual status_t execute(const exec_ctx_t &ctx) const override {
        execute_backward_weights(ctx);
        return status::success;
    }

private:
    void execute_backward_weights(const exec_ctx_t &ctx) const;
    const pd_t *pd() const { return (const pd_t *)primitive_t::pd(); }
};

}
}
}

#endif