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

/*******************************************************************************
* 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/jit/ir/gemm_schedule.hpp"

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

layout_t bmnk_mapper_t::map_to_bmnk(abc_kind_t abc_kind,
        const std::vector<bmnk_kind_t> &bmnk_kinds, const view_t &view) const {
    auto layout = view.create_pseudo_vlayout();
    return map_to_bmnk(abc_kind, bmnk_kinds, layout);
}

layout_t bmnk_mapper_t::map_to_bmnk(abc_kind_t abc_kind,
        const std::vector<bmnk_kind_t> &bmnk_kinds,
        const layout_t &layout) const {
    std::vector<layout_block_t> blocks;
    for (auto &b : layout.blocks()) {
        auto b_bmnk_kind = bmnk_kind(abc_kind, b.idx);
        bool found = false;
        for (int i = 0; i < int(bmnk_kinds.size()); i++) {
            if (bmnk_kinds[i] == b_bmnk_kind) {
                blocks.emplace_back(i, b.size, b.stride);
                found = true;
                break;
            }
        }
        if (!found) gpu_error_not_expected() << "MNK dimension not found.";
    }
    return layout_t(layout.type(), blocks, 0, int(bmnk_kinds.size()));
}

layout_t bmnk_mapper_t::map_from_bmnk(abc_kind_t abc_kind,
        const std::vector<bmnk_kind_t> &bmnk_kinds, const layout_t &bmnk_layout,
        const layout_t &abc_layout) const {
    bmnk_block_mapper_t m(*this);
    m.push_blocks(abc_kind, abc_layout.blocks());
    return m.map_from_bmnk(abc_kind, bmnk_kinds, bmnk_layout);
}

void bmnk_block_mapper_t::push_block(
        abc_kind_t abc_kind, const layout_block_t &b) {
    auto bmnk_kind = bmnk_mapper_.bmnk_kind(abc_kind, b.idx);
    switch (bmnk_kind) {
        case bmnk_kind_t::b: b_blocks_.emplace_back(abc_kind, b); break;
        case bmnk_kind_t::m: m_blocks_.emplace_back(abc_kind, b); break;
        case bmnk_kind_t::n: n_blocks_.emplace_back(abc_kind, b); break;
        case bmnk_kind_t::k: k_blocks_.emplace_back(abc_kind, b); break;
        default: gpu_error_not_expected() << "Unknown MNK kind.";
    }
}

layout_t bmnk_block_mapper_t::map_from_bmnk(abc_kind_t abc_kind,
        const std::vector<bmnk_kind_t> &bmnk_kinds,
        const layout_t &bmnk_layout) const {
    gpu_assert(bmnk_layout.ndims() <= 3);
    gpu_assert(bmnk_layout.offset().is(0));
    std::vector<layout_block_t> blocks;
    std::vector<std::vector<layout_block_t>> tmp_blocks(
            static_cast<int>(bmnk_kind_t::k) + 1);
    tmp_blocks[static_cast<int>(bmnk_kind_t::b)]
            = create_prb_blocks(abc_kind, b_blocks_);
    tmp_blocks[static_cast<int>(bmnk_kind_t::m)]
            = create_prb_blocks(abc_kind, m_blocks_);
    tmp_blocks[static_cast<int>(bmnk_kind_t::n)]
            = create_prb_blocks(abc_kind, n_blocks_);
    tmp_blocks[static_cast<int>(bmnk_kind_t::k)]
            = create_prb_blocks(abc_kind, k_blocks_);
    for (auto &b : bmnk_layout.blocks()) {
        auto &bmnk_blocks = tmp_blocks[static_cast<int>(bmnk_kinds[b.idx])];
        bool ok = pop_block(bmnk_blocks, blocks, b);
        gpu_assert(ok) << "Can't map from bmnk layout to problem layout.";
        MAYBE_UNUSED(ok);
    }
    for (auto bmnk_kind : bmnk_kinds) {
        auto &bmnk_blocks = tmp_blocks[static_cast<int>(bmnk_kind)];
        pop_size_1_blocks(bmnk_blocks);
        gpu_assert(bmnk_blocks.empty());
    }

    // Fix strides to make them dense.
    dim_t dense_stride = 1;
    for (auto &b : blocks) {
        b.stride = stride_t(dense_stride);
        dense_stride *= b.size;
    }

    return layout_t(
            bmnk_layout.type(), blocks, 0, bmnk_mapper_.ndims(abc_kind));
}

bool bmnk_block_mapper_t::pop_block(std::vector<layout_block_t> &bmnk_blocks,
        std::vector<layout_block_t> &prb_blocks,
        const layout_block_t &bmnk_block) const {
    if (bmnk_block.size == 1) return true;

    pop_size_1_blocks(bmnk_blocks);
    if (bmnk_blocks.empty()) return false;

    auto &next_block = bmnk_blocks.front();
    dim_t common_size = math::gcd(next_block.size, bmnk_block.size);
    if (common_size == bmnk_block.size) {
        prb_blocks.emplace_back(next_block.idx, common_size, next_block.stride);
        next_block.size /= common_size;
        next_block.stride *= common_size;
        return true;
    } else if (common_size == next_block.size) {
        prb_blocks.emplace_back(next_block.idx, common_size, next_block.stride);
        bmnk_blocks.erase(bmnk_blocks.begin());
        auto tmp_block = bmnk_block;
        tmp_block.size /= common_size;
        return pop_block(bmnk_blocks, prb_blocks, tmp_block);
    }
    return false;
}

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