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

/*******************************************************************************
* Copyright 2021 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_REF_CONVOLUTION_INT8_HPP
#define CPU_REF_CONVOLUTION_INT8_HPP

#include <assert.h>

#include "common/c_types_map.hpp"
#include "common/primitive.hpp"
#include "common/type_helpers.hpp"
#include "common/utils.hpp"

#include "cpu/cpu_convolution_pd.hpp"
#include "cpu/primitive_attr_postops.hpp"

namespace dnnl {
namespace impl {
namespace cpu {

struct ref_convolution_int8_fwd_t : public primitive_t {
    struct pd_t : public cpu_convolution_fwd_pd_t {
        using cpu_convolution_fwd_pd_t::cpu_convolution_fwd_pd_t;

        DECLARE_COMMON_PD_T("ref:any", ref_convolution_int8_fwd_t);

        status_t init(engine_t *engine) {
            using namespace data_type;
            using smask_t = primitive_attr_t::skip_mask_t;
            const auto src_type = src_md(0)->data_type;
            const auto wei_type = weights_md(0)->data_type;
            const auto bia_type = weights_md(1)->data_type;
            const auto dst_type = dst_md(0)->data_type;

            VDISPATCH_CONV(is_fwd(), VERBOSE_BAD_PROPKIND);
            VDISPATCH_CONV(set_default_alg_kind(alg_kind::convolution_direct),
                    VERBOSE_BAD_ALGORITHM);
            VDISPATCH_CONV(
                    utils::one_of(src_type, s8, u8), VERBOSE_UNSUPPORTED_DT);
            VDISPATCH_CONV(wei_type == s8, VERBOSE_UNSUPPORTED_DT);
            VDISPATCH_CONV(
                    IMPLICATION(with_bias(),
                            utils::one_of(bia_type, f32, bf16, s32, s8, u8)),
                    VERBOSE_UNSUPPORTED_BIAS_CFG);
            VDISPATCH_CONV(utils::one_of(dst_type, f32, bf16, s32, s8, u8),
                    VERBOSE_UNSUPPORTED_DT);
            VDISPATCH_CONV(set_default_formats(), VERBOSE_UNSUPPORTED_TAG);
            VDISPATCH_CONV(
                    attr()->has_default_values(smask_t::scales
                                    | smask_t::zero_points | smask_t::post_ops
                                    | smask_t::sum_dt,
                            dst_type),
                    VERBOSE_UNSUPPORTED_ATTR);
            VDISPATCH_CONV(attr()->post_ops_.check_sum_consistency(dst_type,
                                   /* is_int8 */ true),
                    VERBOSE_UNSUPPORTED_POSTOP);
            CHECK(attr_scales_ok());
            CHECK(attr_zero_points_ok(
                    {{DNNL_ARG_SRC, {0, 2}}, {DNNL_ARG_DST, {0, 2}}}));
            VDISPATCH_CONV(post_ops_ok(), VERBOSE_UNSUPPORTED_POSTOP);
            VDISPATCH_CONV(
                    attr_.set_default_formats(dst_md(0)) == status::success,
                    VERBOSE_UNSUPPORTED_POSTOP);
            return status::success;
        }

    protected:
        bool set_default_formats() {
            using namespace format_tag;
            auto dat_tag = utils::pick(ndims() - 3, nwc, nhwc, ndhwc);
            auto wei_tag = with_groups()
                    ? utils::pick(ndims() - 3, goiw, goihw, goidhw)
                    : utils::pick(ndims() - 3, oiw, oihw, oidhw);
            return set_default_formats_common(dat_tag, wei_tag, dat_tag);
        }

        bool post_ops_ok() const {
            return ref_post_ops_t::post_ops_ok(attr()->post_ops_);
        }
    };

    ref_convolution_int8_fwd_t(const pd_t *apd) : primitive_t(apd) {}

    status_t init(engine_t *engine) override {
        ref_post_ops
                = utils::make_unique<ref_post_ops_t>(pd()->attr()->post_ops_);
        if (!ref_post_ops) return status::out_of_memory;
        CHECK(ref_post_ops->init(pd()->dst_md()));
        return status::success;
    }

    status_t execute(const exec_ctx_t &ctx) const override {
        return execute_forward(ctx);
    }

private:
    status_t execute_forward(const exec_ctx_t &ctx) const;
    const pd_t *pd() const { return (const pd_t *)primitive_t::pd().get(); }
    std::unique_ptr<ref_post_ops_t> ref_post_ops;
};

struct ref_convolution_int8_bwd_data_t : public primitive_t {
    struct pd_t : public cpu_convolution_bwd_data_pd_t {
        using cpu_convolution_bwd_data_pd_t::cpu_convolution_bwd_data_pd_t;

        DECLARE_COMMON_PD_T("ref:any", ref_convolution_int8_bwd_data_t);

        status_t init(engine_t *engine) {
            using namespace data_type;
            const auto diff_src_type = diff_src_md(0)->data_type;
            const auto wei_type = weights_md(0)->data_type;
            const auto diff_dst_type = diff_dst_md(0)->data_type;

            VDISPATCH_CONV(desc()->prop_kind == prop_kind::backward_data,
                    VERBOSE_BAD_PROPKIND);
            VDISPATCH_CONV(set_default_alg_kind(alg_kind::convolution_direct),
                    VERBOSE_BAD_ALGORITHM);
            VDISPATCH_CONV(utils::one_of(diff_dst_type, s8, u8),
                    VERBOSE_UNSUPPORTED_DT);
            VDISPATCH_CONV(wei_type == s8, VERBOSE_UNSUPPORTED_DT);
            VDISPATCH_CONV(utils::one_of(diff_src_type, f32, bf16, s32, s8, u8),
                    VERBOSE_UNSUPPORTED_DT);
            VDISPATCH_CONV(set_default_formats(), VERBOSE_UNSUPPORTED_TAG);
            VDISPATCH_CONV(attr()->has_default_values(
                                   primitive_attr_t::skip_mask_t::scales),
                    VERBOSE_UNSUPPORTED_ATTR);
            CHECK(attr_scales_ok());

            return status::success;
        }

    protected:
        bool set_default_formats() {
            using namespace format_tag;
            auto dat_tag = utils::pick(ndims() - 3, nwc, nhwc, ndhwc);
            auto wei_tag = with_groups()
                    ? utils::pick(ndims() - 3, goiw, goihw, goidhw)
                    : utils::pick(ndims() - 3, oiw, oihw, oidhw);
            return set_default_formats_common(dat_tag, wei_tag, dat_tag);
        }
    };

    ref_convolution_int8_bwd_data_t(const pd_t *apd) : primitive_t(apd) {}

    status_t execute(const exec_ctx_t &ctx) const override {
        return execute_backward_data(ctx);
    }

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

} // namespace cpu
} // namespace impl
} // namespace dnnl

#endif

// vim: et ts=4 sw=4 cindent cino+=l0,\:4,N-s