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

/*******************************************************************************
* 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/rnn/grid.hpp"

namespace dnnl {
namespace impl {
namespace gpu {
namespace intel {
namespace rnn {

#define PART_ONE 1
#define PART_TWO 2

using namespace utils;

template <prop_kind_t aprop>
cell_execution_sig((simple_common_t<aprop>::cell_execution_gru)) {
    const conf_t &conf = this->pd()->conf;
    const ocl_conf_t &ocl_conf = this->pd()->ocl_conf;
    const offsets_t &offsets = this->pd()->off;

    auto cell_layer = !conf.copy_src_layer && lay == 0
            ? user_data.src_layer(dir, iter)
            : workspace.states(lay - 1, dir, iter);
    auto gemm_cell_layer_fwd = !conf.copy_src_layer && lay == 0
            ? gemm_layer_fwd_src
            : gemm_layer_fwd;
    auto gemm_diff_wei_cell_layer = !conf.copy_src_layer && lay == 0
            ? gemm_diff_wei_layer_src
            : gemm_diff_wei_layer;
    auto cell_iter = workspace.states(lay, dir, iter - 1);
    auto cell_iter2 = workspace.states(lay, dir, iter);
    auto &scratch_cell = scratch.cell() ? *scratch.cell()
                                        : memory_storage_t::empty_storage();

    dim_t cell_wei_iter_offset2 = 2 * offsets.weights_iter[3]
            * types::data_type_size(conf.wei_iter_type);
    dim_t cell_scratch_offset2 = 2 * conf.dhc * conf.scratch_gates_elsz;

    auto scratch_gates = scratch.gates(iter);

    auto wei_layer = user_data.wei_layer(lay, dir);
    auto wei_iter = user_data.wei_iter(lay, dir);

    if (aprop == prop_kind::forward) {
        // 1. gemm Wx[0-2],x
        if (!conf.merge_gemm_layer)
            CHECK(gemm_primitive(engine, ctx, wei_layer, cell_layer,
                    scratch_gates, gemm_cell_layer_fwd));

        // 2. gemm Wh[0-1],h
        CHECK(gemm_primitive(engine, ctx, wei_iter, cell_iter, scratch_gates,
                gemm_iter_fwd));

        // 3. activation zt and rt + elemwise multiplication rt,ht-1
        CHECK((this->*elemwise_gru)(ctx, dir, lay, iter, conf.dhc, conf.mb, 1,
                user_data, workspace, scratch_gates, {}, scratch_cell, {}, {},
                {}, 0, {}, tm_scales, diff_bias, PART_ONE));

        // 4. gemm Wh[2],h~t
        CHECK(gemm_primitive(engine, ctx, {wei_iter, cell_wei_iter_offset2},
                cell_iter2, {scratch_gates, cell_scratch_offset2},
                gemm_iter_fwd_2));

        // 5. activation h~t + calculate ht
        CHECK((this->*elemwise_gru)(ctx, dir, lay, iter, conf.dhc, conf.mb, 1,
                user_data, workspace, scratch_gates, {}, scratch_cell, {}, {},
                {}, 0, {}, tm_scales, diff_bias, PART_TWO));

    } else {
        auto diff_states_iter = scratch.diff_states(lay, dir, 0, iter + 1);
        auto diff_states_layer
                = !conf.copy_diff_dst_layer && lay + 1 == conf.n_layer
                ? user_data.diff_dst_layer(dir, iter)
                : scratch.diff_states(lay + 1, dir, conf.n_states, iter);
        auto diff_states_layer_ld
                = !conf.copy_diff_dst_layer && lay + 1 == conf.n_layer
                ? offsets.diff_dst_layer[1]
                : conf.scratch_diff_states_ld;

        auto diff_states = scratch.diff_states(lay, dir, 0, iter);
        auto diff_states1 = !conf.copy_diff_src_layer && lay == 0
                ? user_data.diff_src_layer(dir, iter)
                : scratch.diff_states(lay, dir, conf.n_states, iter);

        auto diff_gates = scratch.diff_gates(iter);
        dim_t cell_scratch_diff_off2
                = 2 * conf.dhc * conf.scratch_diff_gates_elsz;
        sub_buffer_t scratch_diff_ht = *scratch.diff_ht();

        // 1. calculate dG2, dG1, and part of dht-1
        CHECK((this->*elemwise_gru)(ctx, dir, lay, iter, conf.dhc, conf.mb,
                ocl_conf.elemwise_bwd_batch_block, user_data, workspace,
                scratch_gates, diff_gates, scratch_cell, diff_states,
                diff_states_iter, diff_states_layer, diff_states_layer_ld,
                scratch_diff_ht, tm_scales, diff_bias, PART_ONE));

        // 2. calculate intermediate d(hG1)
        // d(hG1) = dG2 * W2h^t
        CHECK(gemm_primitive(engine, ctx, {wei_iter, cell_wei_iter_offset2},
                {diff_gates, cell_scratch_diff_off2}, scratch_diff_ht,
                gemm_iter_bwd_2));

        // 3. calculate dG1^ and part of dht-1
        // hg1 needs to be bf16 as it is used as gemm output
        CHECK((this->*elemwise_gru)(ctx, dir, lay, iter, conf.dhc, conf.mb,
                ocl_conf.elemwise_bwd_batch_block, user_data, workspace,
                scratch_gates, diff_gates, scratch_cell, diff_states,
                diff_states_iter, diff_states_layer, diff_states_layer_ld,
                scratch_diff_ht, tm_scales, diff_bias, PART_TWO));

        auto diff_wei_iter = user_data.diff_wei_iter(lay, dir);
        dim_t cell_diff_wei_iter_off2
                = 2 * offsets.diff_weights_iter[3] * sizeof(float);
        // 4. calculate diff weights
        // dWh1 += dG1 * h, dWh2 += dG2 * h, dWh3 += dG3 * (G1(*)h)
        CHECK(gemm_primitive(engine, ctx, diff_gates, cell_iter, diff_wei_iter,
                gemm_diff_wei_iter));

        CHECK(gemm_primitive(engine, ctx, {diff_gates, cell_scratch_diff_off2},
                scratch_cell, {diff_wei_iter, cell_diff_wei_iter_off2},
                gemm_diff_wei_iter_2));

        // 5. calculate diff states
        // dht-1 += dG1 * W1h + dG0 * W0h
        CHECK(gemm_primitive(
                engine, ctx, wei_iter, diff_gates, diff_states, gemm_iter_bwd));

        if (!conf.merge_gemm_layer) {
            // dWx += [dG0 dG1 dG2] * [x]
            CHECK(gemm_primitive(engine, ctx, diff_gates, cell_layer,
                    user_data.diff_wei_layer(lay, dir),
                    gemm_diff_wei_cell_layer));

            auto gemm_layer_cell_bwd = !conf.copy_diff_src_layer && lay == 0
                    ? gemm_layer_bwd_src
                    : gemm_layer_bwd;
            // dx = dG2 * W2x + dG1 * W1x + dG0 * W0x
            CHECK(gemm_primitive(engine, ctx, wei_layer, diff_gates,
                    diff_states1, gemm_layer_cell_bwd));
        }
    }
    return status::success;
}
template cell_execution_sig(simple_fwd_t::cell_execution_gru);
template cell_execution_sig(simple_bwd_t::cell_execution_gru);

} // namespace rnn
} // namespace intel
} // namespace gpu
} // namespace impl
} // namespace dnnl