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
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273

/*******************************************************************************
* Copyright 2016 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 <assert.h>

#include "common/c_types_map.hpp"
#include "common/compiler_workarounds.hpp"
#include "common/dnnl_thread.hpp"
#include "common/math_utils.hpp"
#include "common/type_helpers.hpp"

#include "cpu/ref_eltwise.hpp"
#include "cpu/ref_io_helper.hpp"
#include "cpu/simple_q10n.hpp"

namespace dnnl {
namespace impl {
namespace cpu {

#define DATA_OFF(f, n, c, d, h, w) \
    (ndims == 1) \
            ? (f).off(n) \
            : ((ndims == 2) ? (f).off(n, c) \
                            : ((ndims == 3) ? (f).off(n, c, w) \
                                            : ((ndims == 4) ? (f).off(n, c, h, \
                                                                      w) \
                                                            : (f).off(n, c, d, \
                                                                      h, w))))

status_t ref_eltwise_fwd_t::execute_forward_generic(
        const exec_ctx_t &ctx) const {
    /* fast return */
    if (pd()->has_zero_dim_memory()) return status::success;

    status_t status = status::success;
    auto src = CTX_IN_MEM(const void *, DNNL_ARG_SRC);
    auto dst = CTX_OUT_CLEAN_MEM(void *, DNNL_ARG_DST, status);
    CHECK(status);

    const auto dropout_p
            = CTX_IN_MEM(const float *, DNNL_ARG_ATTR_DROPOUT_PROBABILITY);
    const auto dropout_seed
            = CTX_IN_MEM(const void *, DNNL_ARG_ATTR_DROPOUT_SEED);
    const auto dropout_offset
            = CTX_IN_MEM(const int64_t *, DNNL_ARG_ATTR_DROPOUT_OFFSET);
    auto dropout_mask = CTX_OUT_CLEAN_MEM(
            unsigned char *, DNNL_ARG_ATTR_DROPOUT_MASK, status);
    CHECK(status);

    const memory_desc_wrapper src_d(pd()->src_md());
    const memory_desc_wrapper dst_d(pd()->dst_md());

    const dim_t MB = pd()->MB();
    const dim_t C = pd()->C();
    const dim_t D = pd()->D();
    const dim_t H = pd()->H();
    const dim_t W = pd()->W();
    const auto alg_kind = pd()->desc()->alg_kind;
    const float alpha = pd()->desc()->alpha;
    const float beta = pd()->desc()->beta;
    const int ndims = pd()->ndims();
    const bool non_default_attrs = !pd()->attr()->has_default_values();
    const bool with_dropout = !pd()->attr()->dropout_.has_default_values();

    parallel_nd(MB, C, D, H, W,
            [= COMPAT_THIS_CAPTURE](
                    dim_t n, dim_t c, dim_t d, dim_t h, dim_t w) {
        auto data_p_off = DATA_OFF(src_d, n, c, d, h, w);
        const float s
                = io::load_float_value(src_d.data_type(), src, data_p_off);
        float res = compute_eltwise_scalar_fwd(alg_kind, s, alpha, beta);
        dim_t data_l_off = (((n * C + c) * D + d) * H + h) * W + w;

        int64_t dropout_seed_val = with_dropout
                ? io::load_int64_value(
                          pd()->attr()->dropout_.seed_dt_, dropout_seed, 0)
                : 0;
        float dropout_p_val = with_dropout ? dropout_p[0] : 0.0f;
        int64_t dropout_offset_val
                = with_dropout && pd()->attr()->dropout_.use_offset_
                ? dropout_offset[0]
                : 0;

        if (non_default_attrs) {
            if (with_dropout) {
                res = ref_dropout(res, dropout_mask, data_p_off, dropout_p_val,
                        dropout_seed_val, dropout_offset_val);
            }
            ref_post_ops_t::args_t args;
            args.ctx = &ctx;
            args.l_offset = data_l_off;
            args.dst_md = pd()->dst_md();
            ref_post_ops->execute(res, args);
        }

        io::store_float_value(dst_d.data_type(), res, dst, data_p_off);
    });
    return status::success;
}

status_t ref_eltwise_fwd_t::execute_forward_dense(const exec_ctx_t &ctx) const {
    status_t status = status::success;
    auto src = CTX_IN_MEM(const void *, DNNL_ARG_SRC);
    auto dst = CTX_OUT_CLEAN_MEM(void *, DNNL_ARG_DST, status);
    CHECK(status);

    const memory_desc_wrapper src_d(pd()->src_md());
    const memory_desc_wrapper dst_d(pd()->dst_md());

    const auto nelems = src_d.nelems(true);
    const auto alg_kind = pd()->desc()->alg_kind;
    const float alpha = pd()->desc()->alpha;
    const float beta = pd()->desc()->beta;

    src = static_cast<const char *>(src)
            + src_d.data_type_size() * src_d.offset0();
    dst = static_cast<char *>(dst) + dst_d.data_type_size() * dst_d.offset0();

    // a fast path for relu as the most popular activation
    if (alg_kind == alg_kind::eltwise_relu && alpha == 0) {
        parallel_nd(nelems, [=](dim_t e) {
            const float s = io::load_float_value(src_d.data_type(), src, e);
            float res = math::relu_fwd(s, alpha);
            io::store_float_value(dst_d.data_type(), res, dst, e);
        });
        return status::success;
    }

    parallel_nd(nelems, [=](dim_t e) {
        const float s = io::load_float_value(src_d.data_type(), src, e);
        float res = compute_eltwise_scalar_fwd(alg_kind, s, alpha, beta);
        io::store_float_value(dst_d.data_type(), res, dst, e);
    });
    return status::success;
}

status_t ref_eltwise_bwd_t::execute_backward_generic(
        const exec_ctx_t &ctx) const {
    /* fast return */
    if (pd()->has_zero_dim_memory()) return status::success;

    status_t status = status::success;
    auto src = CTX_IN_MEM(
            const void *, pd()->use_dst() ? DNNL_ARG_DST : DNNL_ARG_SRC);
    auto diff_dst = CTX_IN_MEM(const void *, DNNL_ARG_DIFF_DST);
    auto diff_src = CTX_OUT_CLEAN_MEM(void *, DNNL_ARG_DIFF_SRC, status);
    CHECK(status);

    const memory_desc_wrapper data_d(pd()->data_md());
    const memory_desc_wrapper diff_data_d(pd()->diff_src_md());

    const dim_t MB = pd()->MB();
    const dim_t C = pd()->C();
    const dim_t D = pd()->D();
    const dim_t H = pd()->H();
    const dim_t W = pd()->W();
    const auto alg_kind = pd()->desc()->alg_kind;
    const float alpha = pd()->desc()->alpha;
    const float beta = pd()->desc()->beta;
    const int ndims = pd()->ndims();

    parallel_nd(
            MB, C, D, H, W, [=](dim_t n, dim_t c, dim_t d, dim_t h, dim_t w) {
        auto data_off = DATA_OFF(data_d, n, c, d, h, w);
        auto diff_data_off = DATA_OFF(diff_data_d, n, c, d, h, w);
        const float s = io::load_float_value(data_d.data_type(), src, data_off);
        const float dd = io::load_float_value(
                diff_data_d.data_type(), diff_dst, diff_data_off);
        float res = compute_eltwise_scalar_bwd(alg_kind, dd, s, alpha, beta);
        io::store_float_value(
                diff_data_d.data_type(), res, diff_src, diff_data_off);
    });
    return status::success;
}

status_t ref_eltwise_bwd_t::execute_backward_dense(
        const exec_ctx_t &ctx) const {
    status_t status = status::success;
    const void *src = pd()->use_dst() ? CTX_IN_MEM(const void *, DNNL_ARG_DST)
                                      : CTX_IN_MEM(const void *, DNNL_ARG_SRC);
    const void *diff_dst = CTX_IN_MEM(const void *, DNNL_ARG_DIFF_DST);
    void *diff_src = CTX_OUT_CLEAN_MEM(void *, DNNL_ARG_DIFF_SRC, status);
    CHECK(status);

    const memory_desc_wrapper data_d(pd()->data_md());
    const memory_desc_wrapper diff_data_d(pd()->diff_src_md());

    const auto nelems = data_d.nelems(true);
    const auto alg_kind = pd()->desc()->alg_kind;
    const float alpha = pd()->desc()->alpha;
    const float beta = pd()->desc()->beta;

    src = static_cast<const char *>(src)
            + data_d.data_type_size() * data_d.offset0();
    diff_dst = static_cast<const char *>(diff_dst)
            + diff_data_d.data_type_size() * diff_data_d.offset0();
    diff_src = static_cast<char *>(diff_src)
            + diff_data_d.data_type_size() * diff_data_d.offset0();

    if (data_d.data_type() == data_type::f32) {
        parallel(0, [=](const int ithr, const int nthr) {
            dim_t start = 0, end = 0;
            balance211(nelems, nthr, ithr, start, end);
            if (start == end) return;

            for (dim_t i = start; i < end; i++) {
                const float s
                        = io::load_float_value(data_d.data_type(), src, i);
                const float dd = io::load_float_value(
                        diff_data_d.data_type(), diff_dst, i);
                float res = compute_eltwise_scalar_bwd(
                        alg_kind, dd, s, alpha, beta);
                io::store_float_value(
                        diff_data_d.data_type(), res, diff_src, i);
            }
        });
    } else if (utils::one_of(data_d.data_type(), data_type::bf16,
                       data_type::f16, data_type::f8_e5m2,
                       data_type::f8_e4m3)) {
        using namespace memory_tracking::names;
        const auto &scratchpad = ctx.get_scratchpad_grantor();
        auto *src_f32 = scratchpad.template get<float>(key_eltwise_src);
        auto *diff_dst_f32
                = scratchpad.template get<float>(key_eltwise_diff_dst);

        parallel(0, [=](const int ithr, const int nthr) {
            dim_t start = 0, end = 0;
            balance211(nelems, nthr, ithr, start, end);
            if (start == end) return;

            types::cvt_to_float(data_d.data_type(), src_f32 + start,
                    static_cast<const char *>(src)
                            + data_d.data_type_size() * start,
                    end - start);
            types::cvt_to_float(diff_data_d.data_type(), diff_dst_f32 + start,
                    static_cast<const char *>(diff_dst)
                            + diff_data_d.data_type_size() * start,
                    end - start);

            for (dim_t i = start; i < end; i++) {
                diff_dst_f32[i] = compute_eltwise_scalar_bwd(
                        alg_kind, diff_dst_f32[i], src_f32[i], alpha, beta);
            }

            types::cvt_from_float(data_d.data_type(),
                    static_cast<char *>(diff_src)
                            + diff_data_d.data_type_size() * start,
                    diff_dst_f32 + start, end - start);
        });
    } else {
        assert(!"unsupported data type");
    }
    return status::success;
}

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

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