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

/*******************************************************************************
* 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 <memory>

#include "graph/interface/partition.hpp"
#include "graph/interface/partition_hashing.hpp"

#include "common/dnnl_thread.hpp"

namespace dnnl {
namespace impl {
namespace graph {
namespace partition_hashing {

key_t::key_t(const impl::engine_t *engine,
        const std::vector<std::shared_ptr<op_t>> &ops,
        const std::vector<const logical_tensor_t *> &ins,
        const std::vector<const logical_tensor_t *> &outs,
        const impl::graph::fpmath_t &fpmath)
    : ops_(get_raw_ptrs(ops))
    , nthread_(dnnl_get_max_threads())
    // Here we use engine as a member of partition_hashing key_t, because for
    // CPU engine and nativa runtime, the engine_id is nullptr for all engine
    // instances. The compiled partition would be hit under different engine
    // instances. For example, first compile with engine 1 -> execute with
    // engine1 -> second compile with engine 2 (cache hit) -> execute with
    // engine 2(fail). So we need to use engine as a member of key_t to avoid
    // execution crash.
    , engine_(engine)
    , fpmath_(fpmath)
    , thread_id_(std::this_thread::get_id()) {
    ins_.reserve(ins.size());
    outs_.reserve(outs.size());
    for (auto &in : ins) {
        ins_.emplace_back(*in);
    }
    for (auto &out : outs) {
        outs_.emplace_back(*out);
    }
}

key_t::key_t(const partition_t *partition, const impl::engine_t *engine,
        const std::vector<const logical_tensor_t *> &ins,
        const std::vector<const logical_tensor_t *> &outs)
    : key_t(engine, partition->get_ops(), ins, outs,
              partition->get_fpmath_mode()) {}

bool key_t::operator==(const key_t &rhs) const {
    if (this == &rhs) return true;

    const size_t lhs_num_ops = ops_.size();
    const size_t rhs_num_ops = rhs.ops_.size();
    const size_t lhs_num_ins = ins_.size();
    const size_t rhs_num_ins = rhs.ins_.size();
    const size_t lhs_num_outs = outs_.size();
    const size_t rhs_num_outs = rhs.outs_.size();

    bool ret = true && lhs_num_ops == rhs_num_ops && lhs_num_ins == rhs_num_ins
            && lhs_num_outs == rhs_num_outs && nthread_ == rhs.nthread_
            && engine_ == rhs.engine_ && fpmath_ == rhs.fpmath_;
    if (!ret) return false;

    for (size_t i = 0; i < lhs_num_ops; ++i) {
        const op_t *op = ops_[i];
        if (std::find_if(rhs.ops_.begin(), rhs.ops_.end(),
                    [op](const op_t *rhs_op) { return *op == *rhs_op; })
                == rhs.ops_.end())
            return false;
    }

    for (size_t i = 0; i < lhs_num_ins; ++i) {
        const logical_tensor_wrapper_t lhs_lt {ins_[i]};
        if (std::find_if(rhs.ins_.begin(), rhs.ins_.end(),
                    [&lhs_lt](const logical_tensor_t &rhs_lt) {
            return logical_tensor_wrapper_t(rhs_lt) == lhs_lt;
        }) == rhs.ins_.end())
            return false;
    }

    for (size_t i = 0; i < lhs_num_outs; ++i) {
        const logical_tensor_wrapper_t lhs_lt {outs_[i]};
        if (std::find_if(rhs.outs_.begin(), rhs.outs_.end(),
                    [&lhs_lt](const logical_tensor_t &rhs_lt) {
            return logical_tensor_wrapper_t(rhs_lt) == lhs_lt;
        }) == rhs.outs_.end())
            return false;
    }

    return true;
}

size_t get_logical_tensors_hash(
        const std::vector<std::shared_ptr<value_t>> &values) {
    size_t seed = 0;
    for (auto &value : values) {
        seed = hash_combine(seed,
                logical_tensor_wrapper_t(value->get_logical_tensor()).hash());
    }
    return seed;
}

using attribute_value_t = dnnl::impl::graph::utils::attribute_value_t;

size_t get_attributes_hash(
        const std::unordered_map<op_attr_t, attribute_value_t> &attributes) {
    size_t seed = 0;
    for (auto &attr : attributes) {
        seed = hash_combine(seed, static_cast<size_t>(attr.first));
        auto kind = attr.second.get_kind();
        try {
            switch (kind) {
                case attribute_kind::f:
                    seed = hash_combine(seed, attr.second.get<float>());
                    break;
                case attribute_kind::i:
                    seed = hash_combine(seed, attr.second.get<int64_t>());
                    break;
                case attribute_kind::s:
                    seed = hash_combine(seed, attr.second.get<std::string>());
                    break;
                case attribute_kind::b:
                    seed = hash_combine(seed, attr.second.get<bool>());
                    break;
                case attribute_kind::fs:
                    seed = get_array_hash(seed,
                            attr.second.get<std::vector<float>>().data(),
                            attr.second.get<std::vector<float>>().size());
                    break;
                case attribute_kind::is:
                    seed = get_array_hash(seed,
                            attr.second.get<std::vector<int64_t>>().data(),
                            attr.second.get<std::vector<int64_t>>().size());
                    break;
                default: break;
            }
        } catch (...) { assert(!"should not reach here"); }
    }
    return seed;
}

size_t get_op_hash(const op_t &op) {
    size_t seed = 0;
    seed = hash_combine(seed, op.get_id());
    seed = hash_combine(seed, static_cast<size_t>(op.get_kind()));
    seed = hash_combine(seed, get_logical_tensors_hash(op.get_input_values()));
    seed = hash_combine(seed, get_logical_tensors_hash(op.get_output_values()));
    seed = hash_combine(seed, get_attributes_hash(op.get_attributes()));
    return seed;
}

} // namespace partition_hashing
} // namespace graph
} // namespace impl
} // namespace dnnl