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

/*******************************************************************************
* 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.
*******************************************************************************/

#include "gpu/intel/block_structure.hpp"

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

block_layout_t block_layout_t::normalized(bool remove_size_1_blocks) const {
    if (num_blocks == 0) return block_layout_t();
    block_layout_t res;

    std::vector<block_t> block_vec(num_blocks);
    memcpy(&block_vec[0], &blocks[0], num_blocks * sizeof(block_t));

    std::vector<block_t> new_blocks
            = normalize_blocks(block_vec, remove_size_1_blocks);
    for (const block_t &block : new_blocks) {
        res.append(block);
    }

    return res;
}

std::vector<block_t> normalize_blocks(
        const std::vector<block_t> &blocks, bool remove_size_1_blocks) {
    if (blocks.empty()) return {};
    std::vector<block_t> res;

    for (const block_t &block : blocks) {
        if (remove_size_1_blocks && block.block == 1) continue;

        if (!res.empty() && res.back().can_merge(block)) {
            res.back().block *= block.block;
        } else {
            res.emplace_back(block);
        }
    }

    return res;
}

block_layout_t::block_layout_t(
        const memory_desc_wrapper &mdw, bool inner_only, bool do_normalize) {
    if (mdw.format_kind() == format_kind::undef) return;

    const dim_idx_t ndims = into<uint32_t>(mdw.ndims());
    auto &blocking = mdw.blocking_desc();
    auto *padded_dims = mdw.padded_dims();

    dim_t stride = 1;
    std::vector<dim_t> full_blocks(ndims, 1);
    for (int i = blocking.inner_nblks - 1; i >= 0; i--) {
        dim_idx_t dim_idx = into<uint32_t>(blocking.inner_idxs[i]);
        dim_t block = blocking.inner_blks[i];
        append(block_t(dim_idx, block, stride));
        stride *= block;
        full_blocks[dim_idx] *= block;
    }

    if (!inner_only) {
        for (dim_idx_t i = 0; i < ndims; i++) {
            dim_t block = padded_dims[i] / full_blocks[i];
            append(block_t(i, block, blocking.strides[i]));
        }

        // Sort outer blocks by their stride.
        std::sort(begin() + blocking.inner_nblks, end(),
                [](const block_t &a, const block_t &b) {
            return a.stride < b.stride
                    || (a.stride == b.stride && a.block < b.block);
        });
    }

    if (do_normalize) *this = normalized();
}

block_layout_t get_inner_layout(const memory_desc_wrapper &md) {
    block_layout_t inner_layout(md, /* inner_only */ true);

    block_layout_t ret;
    // Explicitly initialize to size-1 blocks
    for (int d = 0; d < MAX_NDIMS; d++) {
        ret.append(block_t(d, 1, 0));
    }

    // Overwrite inner blocks with their actual values
    for (const auto &block : inner_layout) {
        ret[block.dim_idx] = block;
    }

    return ret;
}

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