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

/*******************************************************************************
* Copyright 2019 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.
*******************************************************************************/

#include "gpu/intel/pool/ref.hpp"

namespace dnnl {
namespace impl {
namespace gpu {
namespace intel {
namespace pool {

static status_t init_conf_common(conf_t &conf, offsets_t &off, const pd_t *pd,
        impl::engine_t *engine, const bool is_bwd) {
    using namespace dnnl::impl::format_tag;

    const memory_desc_wrapper src_mdw(pd->invariant_src_md());
    const memory_desc_wrapper dst_mdw(pd->invariant_dst_md());

    set_default_conf(conf, *pd->desc(), *pd->invariant_src_md(),
            *pd->invariant_dst_md(), *pd->attr());
    conf.require_stateless_addressing = pd->has_large_buffers();

    set_offsets(src_mdw, off.src_off);
    set_offsets(dst_mdw, off.dst_off);

    auto *intel_engine = utils::downcast<intel::engine_t *>(engine);
    conf.dispatch = intel_engine->create_dispatch(
            conf.is_backward ? src_mdw.md_ : dst_mdw.md_);
    conf.dispatch.define_dim("MB", 0, conf.mb_padded);
    conf.dispatch.define_dim("OC", 1, conf.c_padded);
    int ndims = conf.ndims;
    if (is_bwd) {
        conf.dispatch.define_dim("ID", nstl::max(1, ndims - 3), conf.id);
        conf.dispatch.define_dim("IH", nstl::max(1, ndims - 2), conf.ih);
        conf.dispatch.define_dim("IW", nstl::max(1, ndims - 1), conf.iw);
    } else {
        conf.dispatch.define_dim("OD", nstl::max(1, ndims - 3), conf.od);
        conf.dispatch.define_dim("OH", nstl::max(1, ndims - 2), conf.oh);
        conf.dispatch.define_dim("OW", nstl::max(1, ndims - 1), conf.ow);
    }
    conf.dispatch.generate();

    conf.attr_info = attr_info_t::create(pd->attr());

    return status::success;
}

static status_t init_kernel_ctx_common(compute::kernel_ctx_t &kernel_ctx,
        const conf_t &conf, const offsets_t &off, const post_ops_t &post_ops,
        const memory_desc_t *dst_md) {
    using namespace dnnl::impl::alg_kind;
    kernel_ctx.set_data_type(conf.src_dt);
    kernel_ctx.require_stateless_addressing(conf.require_stateless_addressing);

    kernel_ctx.define_int("SUB_GROUP_SIZE", 1);
    kernel_ctx.define_int("NDIMS", conf.ndims);
    kernel_ctx.define_int("OC_WO_PADDING", conf.c);
    kernel_ctx.define_int("ID", conf.id);
    kernel_ctx.define_int("IH", conf.ih);
    kernel_ctx.define_int("IW", conf.iw);
    kernel_ctx.define_int("OD", conf.od);
    kernel_ctx.define_int("OH", conf.oh);
    kernel_ctx.define_int("OW", conf.ow);
    kernel_ctx.define_int("KD", conf.kd);
    kernel_ctx.define_int("KH", conf.kh);
    kernel_ctx.define_int("KW", conf.kw);
    kernel_ctx.define_int("DD", conf.dd);
    kernel_ctx.define_int("DH", conf.dh);
    kernel_ctx.define_int("DW", conf.dw);
    kernel_ctx.define_int("SD", conf.stride_d);
    kernel_ctx.define_int("SH", conf.stride_h);
    kernel_ctx.define_int("SW", conf.stride_w);
    kernel_ctx.define_int("PD", conf.f_pad);
    kernel_ctx.define_int("PH", conf.t_pad);
    kernel_ctx.define_int("PW", conf.l_pad);
    kernel_ctx.define_int("IS_TRAINING", conf.is_training);
    kernel_ctx.define_int("IS_BWD", conf.is_backward);
    kernel_ctx.define_int("IS_FWD", !conf.is_backward);

    kernel_ctx.define_int("ALG_MAX", (conf.alg == pooling_max));
    kernel_ctx.define_int(
            "ALG_AVG_NP", (conf.alg == pooling_avg_exclude_padding));
    kernel_ctx.define_int(
            "ALG_AVG_P", (conf.alg == pooling_avg_include_padding));

    CHECK(def_attr_info(kernel_ctx, conf.attr_info, post_ops, *dst_md));

    def_offsets(off.src_off, kernel_ctx, "SRC", conf.ndims);
    def_offsets(off.dst_off, kernel_ctx, "DST", conf.ndims);

    def_memory_desc_info(kernel_ctx, conf.src_md_info, "SRC");
    def_memory_desc_info(kernel_ctx, conf.dst_md_info, "DST");

    def_dispatch(kernel_ctx, conf.dispatch);

    return status::success;
}

status_t ref_fwd_t::pd_t::init_conf(impl::engine_t *engine) {
    return init_conf_common(conf, off, this, engine, false);
}

status_t ref_fwd_t::pd_t::init_kernel_ctx(
        compute::kernel_ctx_t &kernel_ctx) const {
    return init_kernel_ctx_common(
            kernel_ctx, conf, off, attr()->post_ops_, invariant_dst_md());
}

status_t ref_fwd_t::execute_forward(const exec_ctx_t &ctx) const {
    auto &src = CTX_IN_STORAGE(DNNL_ARG_SRC);
    auto &dst = CTX_OUT_STORAGE(DNNL_ARG_DST);
    auto &ws = CTX_OUT_STORAGE(DNNL_ARG_WORKSPACE);

    compute::kernel_arg_list_t arg_list;
    arg_list.set(0, src);
    arg_list.set(1, ws);
    arg_list.set(2, dst);
    append_post_ops_to_arg_list(
            ctx, arg_list, 3, pd()->attr()->post_ops_, *pd()->dst_md());

    auto nd_range = pd()->conf.dispatch.nd_range();

    return parallel_for(ctx, nd_range, kernel_, arg_list);
}

status_t ref_bwd_t::pd_t::init_conf(impl::engine_t *engine) {
    return init_conf_common(conf, off, this, engine, true);
}

status_t ref_bwd_t::pd_t::init_kernel_ctx(
        compute::kernel_ctx_t &kernel_ctx) const {
    return init_kernel_ctx_common(
            kernel_ctx, conf, off, attr()->post_ops_, invariant_dst_md());
}

status_t ref_bwd_t::execute_backward(const exec_ctx_t &ctx) const {
    auto &diff_src = CTX_OUT_STORAGE(DNNL_ARG_DIFF_SRC);
    auto &diff_dst = CTX_IN_STORAGE(DNNL_ARG_DIFF_DST);
    auto &ws = CTX_IN_STORAGE(DNNL_ARG_WORKSPACE);

    compute::kernel_arg_list_t arg_list;
    arg_list.set(0, diff_src);
    arg_list.set(1, ws);
    arg_list.set(2, diff_dst);

    auto nd_range = pd()->conf.dispatch.nd_range();

    return parallel_for(ctx, nd_range, kernel_, arg_list);
}

} // namespace pool
} // namespace intel
} // namespace gpu
} // namespace impl
} // namespace dnnl