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
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236

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

/*
 * Cell execution GRU
 */

#include "common/dnnl_thread.hpp"
#include "common/math_utils.hpp"

#include "cpu/rnn/ref_rnn.hpp"

namespace dnnl {
namespace impl {
namespace cpu {

using namespace dnnl::impl::utils;
using namespace dnnl::impl::math;
using namespace rnn_utils;

#define AOC array_offset_calculator
template <data_type_t src_type, data_type_t weights_type, data_type_t acc_type>
rnn_cell_execution_sig(
        (ref_rnn_fwd_t<src_type, weights_type, acc_type>::cell_execution_gru)) {
    const ws_gates_aoc_t<gates_t> ws_gates(rnn, ws_gates_);
    const scratch_gates_aoc_t<scratch_t> scratch_gates(rnn, scratch_gates_);
    const auto weights_scales = this->pd_->attr()->rnn_weights_qparams_.scales_;

    const auto src_layer_ld = rnn.src_layer_ld(cell_position);
    const auto src_iter_ld = rnn.src_iter_ld(cell_position);
    const auto dst_iter_part2_ld = rnn.dst_iter_part2_ld(cell_position);

    // 1. gemm Wx[0-2],x
    if (rnn.need_gemm_layer(cell_position)) {
        if (rnn.use_matmul) {
            CHECK(this->execute_matmul(ctx,
                    this->get_matmul_layer(cell_position), w_layer_[0],
                    src_layer_, scratch_gates_));
        } else {
            CHECK((this->*gemm_layer_func)('N', 'N', rnn.n_gates * rnn.dhc,
                    rnn.mb, rnn.slc, 1.0, w_layer_[0], rnn.weights_layer_ld,
                    src_layer_, src_layer_ld, 0.0f, scratch_gates_,
                    rnn.scratch_gates_ld));
        }
    }

    // 2. gemm Wh[0-1],h
    if (rnn.use_matmul) {
        CHECK(this->execute_matmul(ctx, this->get_matmul_iter(cell_position),
                w_iter_[0], src_iter_, scratch_gates_));
    } else {
        CHECK((this->*gemm_iter_func)('N', 'N', (rnn.n_gates - 1) * rnn.dhc,
                rnn.mb, rnn.sic, 1.0, w_iter_[0], rnn.weights_iter_ld,
                src_iter_, src_iter_ld, 1.0f, scratch_gates_,
                rnn.scratch_gates_ld));
    }
    // 3. activation zt and rt + elemwise multiplication rt,ht-1
    this->rnn_postgemm_->execute(rnn, cell_position, ws_gates_, scratch_gates_,
            augru_attention_, dst_layer_, nullptr, src_iter_, nullptr,
            diff_src_layer_, diff_augru_attention_, diff_src_iter_, nullptr,
            diff_dst_layer_, diff_dst_iter_, nullptr, nullptr, bias_[0],
            nullptr, nullptr, dst_iter_, weights_scales, rnn.dhc);

    // 4. gemm Wh[2],h~t
    if (rnn.use_matmul) {
        CHECK(this->execute_matmul(ctx, this->get_matmul_part2(cell_position),
                w_iter_[1], dst_layer_, &(scratch_gates(0, 2, 0))));
    } else {
        CHECK((this->*gemm_iter_func)('N', 'N', rnn.dhc, rnn.mb, rnn.sic, 1.0,
                w_iter_[1], rnn.weights_iter_ld, dst_layer_, dst_iter_part2_ld,
                1.0, &(scratch_gates(0, 2, 0)), rnn.scratch_gates_ld));
    }
    // 5. activation h~t + calculate ht
    this->rnn_postgemm_->execute_part2(rnn, cell_position, ws_gates_,
            scratch_gates_, augru_attention_, dst_layer_, dst_iter_c_,
            src_iter_, src_iter_c_, diff_src_layer_, diff_augru_attention_,
            diff_src_iter_, nullptr, diff_dst_layer_, diff_dst_iter_, nullptr,
            nullptr, bias_[0], nullptr, nullptr, dst_iter_, weights_scales,
            rnn.dhc);

    return dnnl_success;
}

template rnn_cell_execution_sig(ref_rnn_fwd_f32_t::cell_execution_gru);
template rnn_cell_execution_sig(ref_rnn_fwd_bf16_t::cell_execution_gru);
template rnn_cell_execution_sig(ref_rnn_fwd_f16_t::cell_execution_gru);
template rnn_cell_execution_sig(ref_rnn_fwd_u8s8_t::cell_execution_gru);
template rnn_cell_execution_sig(ref_rnn_fwd_s8s8_t::cell_execution_gru);

template <typename T1, typename T2, typename T3, typename T4, typename T5,
        typename weights_data_t, typename src_data_t, typename acc_data_t,
        typename scratch_data_t>
dnnl_status_t gru_bwd_cell_exec_template(T1 gemm_layer_f, T2 gemm_iter_f,
        T3 gemm_weights_layer_f, T4 gemm_weights_iter_f, T5 rnn_postgemm_,
        const rnn_utils::rnn_conf_t &rnn, cell_position_t cell_position,
        src_data_t *ws_gates_, scratch_data_t *scratch_gates_,
        src_data_t *dst_layer_, const src_data_t *src_iter_,
        const src_data_t *src_layer_, const src_data_t *augru_attention_,
        weights_data_t **w_layer_, weights_data_t **w_iter_,
        acc_data_t *diff_w_layer_, acc_data_t *diff_w_iter_,
        acc_data_t *diff_src_layer_, acc_data_t *diff_augru_attention_,
        acc_data_t *diff_src_iter_, acc_data_t *diff_dst_iter_,
        acc_data_t *diff_dst_layer_, acc_data_t *diff_bias_,
        scratch_data_t *scratch_cell_, src_data_t *dst_iter_) {
    const ws_gates_aoc_t<src_data_t> ws_gates(rnn, ws_gates_);
    const scratch_gates_aoc_t<scratch_data_t> scratch_gates(
            rnn, scratch_gates_);

    const auto src_layer_ld = rnn.src_layer_ld(cell_position);
    const auto dst_iter_ld = rnn.dst_iter_ld(cell_position);
    const auto dst_layer_ld = rnn.dst_layer_ld(cell_position);
    const auto src_iter_ld = rnn.src_iter_ld(cell_position);
    const ws_states_layer_aoc_t<src_data_t> dst_layer(rnn, dst_layer_,
            (cell_position & last_layer) ? dst_layer_ld : dst_iter_ld);
    const ws_states_iter_aoc_t<const src_data_t> src_iter(
            rnn, src_iter_, src_iter_ld);
    const ws_diff_w_iter_aoc_t diff_w_iter(rnn, diff_w_iter_);

    // use state memory for intermediate computations
    // TODO: use cell ws for that
    float *dhG1_ = diff_src_layer_;
    const AOC<acc_data_t, 2> dhG1(
            dhG1_, rnn.ws_states_layer_nld, rnn.ws_states_layer_ld);
    // hg1 needs to be bf16 as it is used as gemm output
    // hence it cannot alias to dhG1, and should use scratch_cell
    const AOC<scratch_data_t, 2> hG1(
            scratch_cell_, rnn.ws_states_layer_nld, rnn.ws_states_layer_ld);

    // 1. calculate dG2, dG1, and part of dht-1
    rnn_postgemm_->execute(rnn, cell_position, ws_gates_, scratch_gates_,
            augru_attention_, dst_layer_, nullptr, src_iter_, nullptr,
            diff_src_layer_, diff_augru_attention_, diff_src_iter_, nullptr,
            diff_dst_layer_, diff_dst_iter_, nullptr, nullptr, nullptr, nullptr,
            scratch_cell_, dst_iter_, nullptr, 0);

    // 2. calculate intermediate d(hG1)
    // d(hG1) = dG2 * W2h^t
    CHECK(gemm_iter_f(rnn.sic, rnn.mb, rnn.dhc, w_iter_[1],
            &(scratch_gates(0, 2, 0)), 0.0f, dhG1_));

    // 3. calculate dG1^ and part of dht-1
    rnn_postgemm_->execute_part2(rnn, cell_position, ws_gates_, scratch_gates_,
            augru_attention_, dst_layer_, nullptr, src_iter_, nullptr,
            diff_src_layer_, diff_augru_attention_, diff_src_iter_, nullptr,
            diff_dst_layer_, diff_dst_iter_, nullptr, nullptr, nullptr, nullptr,
            scratch_cell_, dst_iter_, nullptr, 0);

    // 4. calculate diff weights
    // dWh1 += dG1 * h, dWh2 += dG2 * h, dWh3 += dG3 * (G1(*)h)
    CHECK(gemm_weights_iter_f((rnn.n_gates - 1) * rnn.dhc, rnn.sic, rnn.mb,
            scratch_gates_, src_iter_, src_iter_ld, diff_w_iter_));
    CHECK(gemm_weights_iter_f(rnn.dhc, rnn.sic, rnn.mb,
            &(scratch_gates(0, 2, 0)), scratch_cell_, rnn.ws_states_layer_ld,
            &(diff_w_iter(0, 2, 0))));

    // 5. calculate diff states
    // dht-1 += dG1 * W1h + dG0 * W0h
    CHECK(gemm_iter_f(rnn.sic, rnn.mb, (rnn.n_gates - 1) * rnn.dhc, w_iter_[0],
            scratch_gates_, 1.0f, diff_src_iter_));

    // dWx += [dG0 dG1 dG2] * [x]
    if (rnn.need_gemm_layer(cell_position))
        CHECK(gemm_weights_layer_f(
                scratch_gates_, src_layer_, src_layer_ld, diff_w_layer_));

    // dx = dG2 * W2x + dG1 * W1x + dG0 * W0x
    if (!rnn.merge_gemm_layer)
        CHECK(gemm_layer_f(w_layer_[0], scratch_gates_, diff_src_layer_));

    // 6. calculate diff bias
    gates_reduction(rnn, cell_position, scratch_gates_, diff_bias_);

    return dnnl_success;
}

template <data_type_t src_type, data_type_t weights_type, data_type_t acc_type>
rnn_cell_execution_sig(
        (ref_rnn_bwd_t<src_type, weights_type, acc_type>::cell_execution_gru)) {
    auto gemm_iter_f
            = [&](int m, int n, int k, const weights_t *A, const gemm_data_t *B,
                      float beta, gemm_acc_t *C) {
        return (this->*gemm_iter_func)('N', 'N', m, n, k, 1.0f, A,
                rnn.weights_iter_ld, B, rnn.scratch_gates_ld, beta, C,
                rnn.ws_diff_states_iter_ld);
    };
    auto gemm_layer_f
            = [&](const weights_t *A, const gemm_data_t *B, gemm_acc_t *C) {
        return (this->*gemm_layer_func)('N', 'N', rnn.slc, rnn.mb,
                rnn.n_gates * rnn.dhc, 1.0, A, rnn.weights_layer_ld, B,
                rnn.scratch_gates_ld, 0.0, C, rnn.ws_diff_states_layer_ld);
    };
    auto gemm_weights_layer_f = [&](const gemm_data_t *A, const weights_t *B,
                                        int ldb, gemm_acc_t *C) {
        const float beta = rnn.diff_weights_beta(cell_position);
        return gemm('N', 'T', rnn.n_gates * rnn.dhc, rnn.slc, rnn.mb, 1.0, A,
                rnn.scratch_gates_ld, B, ldb, beta, C,
                rnn.diff_weights_layer_ld);
    };
    auto gemm_weights_iter_f
            = [&](int m, int n, int k, const weights_t *A, const gemm_data_t *B,
                      int ldb, gemm_acc_t *C) {
        const float beta = rnn.diff_weights_beta(cell_position);
        return gemm('N', 'T', m, n, k, 1.0f, A, rnn.ws_gates_ld, B, ldb, beta,
                C, rnn.diff_weights_iter_ld);
    };

    return gru_bwd_cell_exec_template(gemm_layer_f, gemm_iter_f,
            gemm_weights_layer_f, gemm_weights_iter_f, this->rnn_postgemm_, rnn,
            cell_position, ws_gates_, scratch_gates_, dst_layer_, src_iter_,
            src_layer_, augru_attention_, w_layer_, w_iter_, diff_w_layer_,
            diff_w_iter_, diff_src_layer_, diff_augru_attention_,
            diff_src_iter_, diff_dst_iter_, diff_dst_layer_, diff_bias_,
            scratch_cell_, dst_iter_);
}

template rnn_cell_execution_sig(ref_rnn_bwd_f32_t::cell_execution_gru);
template rnn_cell_execution_sig(ref_rnn_bwd_bf16_t::cell_execution_gru);
template rnn_cell_execution_sig(ref_rnn_bwd_f16_t::cell_execution_gru);

#undef AOC
} // namespace cpu
} // namespace impl
} // namespace dnnl