onednn-src 0.1.13

Source of oneAPI Deep Neural Network Library (oneDNN)
Documentation 
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

/*******************************************************************************
* Copyright 2020 Intel Corporation
* Copyright 2020 Codeplay Software Limited
*
* 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 GPU_AMD_MIOPEN_BATCH_NORMALIZATION_HPP
#define GPU_AMD_MIOPEN_BATCH_NORMALIZATION_HPP

#include <miopen/miopen.h>

#include "common/batch_normalization_pd.hpp"
#include "common/c_types_map.hpp"
#include "common/type_helpers.hpp"
#include "gpu/amd/engine.hpp"
#include "gpu/amd/miopen_batch_normalization_executor.hpp"
#include "gpu/amd/miopen_batch_normalization_impl.hpp"
#include "gpu/amd/stream.hpp"
#include "gpu/amd/sycl_hip_utils.hpp"
#include "gpu/gpu_primitive.hpp"

namespace dnnl {
namespace impl {
namespace gpu {
namespace amd {

struct miopen_batch_normalization_common_t {
    template <typename pd_t>
    static status_t execute(
            const exec_ctx_t &ctx, impl::engine_t *engine, const pd_t *pd) {
        if (memory_desc_wrapper(pd->src_md()).has_zero_dim()) {
            return status::success;
        }
        return pd->executor_->execute(ctx, engine, pd->bnorm_impl_);
    }

    template <typename pd_t>
    static void init_ws(const pd_t *pd, memory_desc_t &ws_md) {

        const auto wrap = memory_desc_wrapper(pd->src_md());
        const auto y_size
                = wrap.nelems() * types::data_type_size(data_type::f32);
        // Mean and variance are always f32.
        const size_t mean_invvar_size
                = 2 * pd->C() * types::data_type_size(data_type::f32);
        const dims_t ws_size
                = {(dim_t)(y_size * pd->fuse_norm_relu() + mean_invvar_size)};

        memory_desc_init_by_tag(
                ws_md, 1, ws_size, data_type::u8, format_tag::x);
    }
};

struct miopen_batch_normalization_fwd_t : public gpu::primitive_t {
    using gpu::primitive_t::primitive_t;
    struct pd_t : public batch_normalization_fwd_pd_t {
        using batch_normalization_fwd_pd_t::batch_normalization_fwd_pd_t;

        DECLARE_COMMON_PD_T("hip:miopen:any", miopen_batch_normalization_fwd_t);

        status_t init(impl::engine_t *) {
            using namespace data_type;
            using namespace types;

            const auto norm_flags_supported
                    = normalization_flags::use_global_stats
                    | normalization_flags::fuse_norm_relu
                    | normalization_flags::use_scale
                    | normalization_flags::use_shift;
            if ((~norm_flags_supported & desc()->flags) != 0)
                return status::unimplemented;

            const auto attr_skip_mask = primitive_attr_t::skip_mask_t::post_ops;

            bool ok = is_fwd() && utils::one_of(src_md()->data_type, f32, f16)
                    && src_md()->data_type == dst_md()->data_type
                    && check_scale_shift_data_type()
                    && attr()->has_default_values(attr_skip_mask)
                    && IMPLICATION(!attr()->has_default_values(),
                            attr()->post_ops_.len() == 1 && with_relu_post_op())
                    && IMPLICATION(utils::one_of(src_md()->data_type, f16),
                            !is_training() && stats_is_src())
                    && set_default_formats_common()
                    && memory_desc_wrapper(src_md())
                            == memory_desc_wrapper(dst_md())
                    && src_md()->format_desc.blocking.inner_nblks == 0
                    && check_format();

            if (!ok) return status::unimplemented;

            if (is_training()) {
                miopen_batch_normalization_common_t::init_ws(this, ws_md_);
            }

            if (use_global_stats()) {
                bnorm_impl_.reset(
                        new miopen_batch_normalization_fwd_stats_impl_t());
            } else {
                bnorm_impl_.reset(new miopen_batch_normalization_fwd_impl_t());
            }

            executor_.reset(new bnorm_exec_fwd_t());

            return bnorm_impl_->init(this);
        }

        bool check_format() const {
            // Only abx formats are supported
            return (memory_desc_wrapper(src_md()).matches_one_of_tag(
                            format_tag::a, format_tag::ab, format_tag::abc,
                            format_tag::abcd, format_tag::abcde)
                    && memory_desc_wrapper(dst_md()).matches_one_of_tag(
                            format_tag::a, format_tag::ab, format_tag::abc,
                            format_tag::abcd, format_tag::abcde));
        }

        std::shared_ptr<miopen_batch_normalization_impl_base_t> bnorm_impl_;
        std::shared_ptr<bnorm_exec_base_t> executor_;
    };

    status_t execute(const exec_ctx_t &ctx) const override;

private:
    const pd_t *pd() const { return (const pd_t *)primitive_t::pd().get(); }
};

struct miopen_batch_normalization_bwd_t : public gpu::primitive_t {
    using gpu::primitive_t::primitive_t;

    struct pd_t : public batch_normalization_bwd_pd_t {
        using batch_normalization_bwd_pd_t::batch_normalization_bwd_pd_t;

        DECLARE_COMMON_PD_T("hip:miopen:any", miopen_batch_normalization_bwd_t);

        status_t init(impl::engine_t *) {
            using namespace data_type;
            using namespace types;

            const auto norm_flags_supported
                    = normalization_flags::fuse_norm_relu
                    | normalization_flags::use_scale
                    | normalization_flags::use_shift;
            if ((~norm_flags_supported & desc()->flags) != 0)
                return status::unimplemented;

            bool ok = !is_fwd()
                    && (utils::everyone_is(f32, src_md()->data_type,
                            diff_src_md()->data_type, diff_dst_md()->data_type))
                    && check_scale_shift_data_type()
                    && attr()->has_default_values()
                    && set_default_formats_common()
                    && memory_desc_wrapper(diff_src_md())
                            == memory_desc_wrapper(diff_dst_md())
                    && src_md()->format_desc.blocking.inner_nblks == 0
                    && diff_src_md()->format_desc.blocking.inner_nblks == 0
                    && check_format();
            if (!ok) return status::unimplemented;

            miopen_batch_normalization_common_t::init_ws(this, ws_md_);
            if (!compare_ws(hint_fwd_pd_)) return status::unimplemented;

            if (fuse_norm_relu()) {
                bnorm_impl_.reset(
                        new miopen_batch_normalization_bwd_relu_impl_t());
            } else {
                bnorm_impl_.reset(new miopen_batch_normalization_bwd_impl_t());
            }

            executor_.reset(new bnorm_exec_bwd_t());

            return bnorm_impl_->init(this);
        }

        bool check_format() const {
            // Only abx formats are supported
            return (memory_desc_wrapper(diff_src_md())
                            .matches_one_of_tag(format_tag::a, format_tag::ab,
                                    format_tag::abc, format_tag::abcd,
                                    format_tag::abcde)
                    && memory_desc_wrapper(diff_dst_md())
                               .matches_one_of_tag(format_tag::a,
                                       format_tag::ab, format_tag::abc,
                                       format_tag::abcd, format_tag::abcde));
        }

        std::shared_ptr<miopen_batch_normalization_impl_base_t> bnorm_impl_;
        std::shared_ptr<bnorm_exec_base_t> executor_;
    };

    status_t execute(const exec_ctx_t &ctx) const override;

private:
    const pd_t *pd() const { return (const pd_t *)primitive_t::pd().get(); }
};

} // namespace amd
} // namespace gpu
} // namespace impl
} // namespace dnnl

#endif