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

/*******************************************************************************
* Copyright 2023 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 "cpu/x64/brgemm/brgemm_containers.hpp"
#include "cpu/x64/brgemm/jit_brdgmm_kernel.hpp"

namespace dnnl {
namespace impl {
namespace cpu {
namespace x64 {

using namespace dnnl::impl::utils;

namespace brgemm_containers {

std::set<std::shared_ptr<brgemm_kernel_t>,
        decltype(brgemm_kernel_container_t::brgemm_kernel_cmp) *> &
brgemm_kernel_container_t::get_set() {
#ifdef BRGEMM_KERNEL_GLOBAL_STORAGE
    static std::set<std::shared_ptr<brgemm_kernel_t>,
            decltype(brgemm_kernel_container_t::brgemm_kernel_cmp) *>
            set_ = std::set<std::shared_ptr<brgemm_kernel_t>,
                    decltype(brgemm_kernel_container_t::brgemm_kernel_cmp) *>(
                    brgemm_kernel_container_t::brgemm_kernel_cmp);
#endif
    return set_;
}

bool brgemm_desc_container_t::insert(int idx, brgemm_desc_t &brg,
        const std::vector<char> &bd_mask,
        const std::vector<brgemm_batch_element_t> &static_offsets) {
    bd_mask_list_.push_back(bd_mask);
    brg.brgattr.bd_mask = bd_mask_list_.back().data();

    static_offsets_list_.push_back(static_offsets);
    brg.brgattr.static_offsets = static_offsets_list_.back().data();

    const auto ret = map_.insert({brg, idx});
    const int ref_size = refs_.size();
    if (idx > ref_size - 1) { refs_.resize(idx + 1); }
    refs_[idx] = &(ret.first->first);
    // if there was no insertion then clean bd_mask and static_offsets
    if (!ret.second) {
        bd_mask_list_.pop_back();
        static_offsets_list_.pop_back();
    }
    return ret.second;
}

int brgemm_desc_container_t::insert(brgemm_desc_t &brg,
        const std::vector<char> &bd_mask,
        const std::vector<brgemm_batch_element_t> &static_offsets) {
    bd_mask_list_.push_back(bd_mask);
    brg.brgattr.bd_mask = bd_mask_list_.back().data();

    static_offsets_list_.push_back(static_offsets);
    brg.brgattr.static_offsets = static_offsets_list_.back().data();
    const int ref_size = refs_.size();
    const auto ret = map_.insert({brg, -1});
    if (!ret.second) {
        // if there was no insertion then clean bd_mask and static_offsets
        bd_mask_list_.pop_back();
        static_offsets_list_.pop_back();
        return ret.first->second;
    }

    int idx = map_.size() - 1;
    if (idx > ref_size - 1) {
        if (ref_size == 0)
            refs_.resize(1);
        else
            refs_.resize(2 * ref_size);
    }
    refs_[idx] = &(ret.first->first);
    ret.first->second = idx;
    return idx;
}

bool brgemm_kernel_container_t::brgemm_kernel_cmp(
        const std::shared_ptr<brgemm_kernel_t> &lhs,
        const std::shared_ptr<brgemm_kernel_t> &rhs) {
    const auto lsz = lhs->get_jit_generator()->getSize();
    const auto rsz = rhs->get_jit_generator()->getSize();
    if (lsz != rsz) return (lsz < rsz);
    const auto lcode = lhs->get_jit_generator()->CodeGenerator::getCode();
    const auto rcode = rhs->get_jit_generator()->CodeGenerator::getCode();
    return (std::memcmp(lcode, rcode, lsz) < 0);
}

status_t brgemm_kernel_container_t::insert(int idx, const brgemm_desc_t *brg) {
    // Use two level hashing of brgemm kernels:
    // 1. Try to find entry in local brgemm_map_ using brgemm descriptor as a
    // key (we can check if brgemm descriptor is unique inside brgemm primitive)
    // 2. Only if we do not find entry in local brgemm_map_  then try to find
    // entry in kernel storage using kernel code as key
    const auto brgemm_it = brgemm_map_.find(brg);
    if (brgemm_it == brgemm_map_.end()) {
        brgemm_kernel_t *brg_kernel = nullptr;
        CHECK(brgemm_kernel_create(&brg_kernel, *brg));
        std::shared_ptr<brgemm_kernel_t> sptr(brg_kernel);
        lock_write();
        const auto kernel_ret = get_set().insert(std::move(sptr));
        refs_[idx] = kernel_ret.first->get();
        unlock_write();
        const auto brgemm_ret = brgemm_map_.insert({brg, refs_[idx]});
        if (!brgemm_ret.second) return status::runtime_error;
    } else {
        refs_[idx] = brgemm_it->second;
    }
    return status::success;
}

bool brgemm_palette_container_t::insert(int idx, const brgemm_desc_t *brg) {
    S_t kernel_palette;
    auto status = brgemm_init_tiles(*brg, kernel_palette.data());
    if (status != status::success) return false;

    const auto ret = set_.insert(kernel_palette);
    refs_[idx] = &(*ret.first);
    return ret.second;
}

} // namespace brgemm_containers
} // namespace x64
} // namespace cpu
} // namespace impl
} // namespace dnnl

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