1 /*******************************************************************************
2 * Copyright 2017-2018 Intel Corporation
4 * Licensed under the Apache License, Version 2.0 (the "License");
5 * you may not use this file except in compliance with the License.
6 * You may obtain a copy of the License at
8 * http://www.apache.org/licenses/LICENSE-2.0
10 * Unless required by applicable law or agreed to in writing, software
11 * distributed under the License is distributed on an "AS IS" BASIS,
12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 * See the License for the specific language governing permissions and
14 * limitations under the License.
15 *******************************************************************************/
17 #include "mkldnn_thread.hpp"
19 #include "simple_concat.hpp"
25 using namespace memory_tracking::names;
27 template <data_type_t data_type>
28 void simple_concat_t<data_type>::execute() const {
29 auto scratchpad = this->scratchpad();
30 auto iptrs = scratchpad.template get<const data_t *>(key_concat_iptrs);
31 auto optrs = scratchpad.template get<data_t *>(key_concat_optrs);
32 auto nelems_to_copy = scratchpad.template get<size_t>(key_concat_nelems);
33 auto is = scratchpad.template get<strides_t>(key_concat_istrides);
35 const int num_arrs = pd()->n_inputs();
36 const ptrdiff_t *perm = pd()->perm_, *iperm = pd()->iperm_;
37 const int concat_dim = pd()->concat_dim();
38 auto o_base_ptr = reinterpret_cast<data_t *>(this->memory());
40 for (int a = 0; a < num_arrs; ++a) {
41 const memory_desc_wrapper i_d(pd()->src_pd(a));
42 const memory_desc_wrapper o_d(pd()->src_image_pd(a));
44 iptrs[a] = reinterpret_cast<const data_t *>(
45 this->input_memory(a)) + i_d.blk_off(0);
46 optrs[a] = o_base_ptr + o_d.blk_off(0);
47 nelems_to_copy[a] = pd()->nelems_to_concat(i_d);
48 for (int i = 0; i < TENSOR_MAX_DIMS; i++) {
49 if (i < perm[concat_dim])
50 is[a][i] = size_t(i_d.blocking_desc().strides[0][iperm[i]]);
56 const memory_desc_wrapper o_d(pd()->src_image_pd());
57 auto &blk = o_d.blocking_desc();
60 for (int i = 0; i < perm[concat_dim]; i++)
61 os[i] = o_d.blocking_desc().strides[0][iperm[i]];
64 for (size_t i = 0; i < sizeof(phys_dims)/sizeof(phys_dims[0]); i++)
65 phys_dims[i] = (i < (size_t)perm[concat_dim])
66 ? o_d.dims()[iperm[i]] / blk.block_dims[iperm[i]] : 1;
68 if (perm[concat_dim] == 0) {
69 for (int a = 0; a < num_arrs; ++a) {
70 const data_t *i = &iptrs[a][0];
71 data_t *o = &optrs[a][0];
72 parallel_nd((ptrdiff_t)nelems_to_copy[a],
73 [&](ptrdiff_t e) { o[e] = i[e]; });
76 parallel_nd(phys_dims[0], phys_dims[1], phys_dims[2], phys_dims[3],
77 phys_dims[4], num_arrs,
78 [&](int n0, int n1, int n2, int n3, int n4, int a) {
79 // XXX: this code may access uninitialized values in is[*][0-4] --
80 // that's why we have to set them to zero although this is
82 size_t in_off = is[a][0] * n0 + is[a][1] * n1 + is[a][2] * n2
83 + is[a][3] * n3 + is[a][4] * n4;
84 size_t out_off = os[0] * n0 + os[1] * n1 + os[2] * n2
85 + os[3] * n3 + os[4] * n4;
86 const data_t *i = &iptrs[a][in_off];
87 data_t *o = &optrs[a][out_off];
88 #if defined(__GNUC__) && !defined(__INTEL_COMPILER)
89 // The code below performs data copying: o[e] = i[e]
90 // and uses a workaround to make GNU compilers optimize it
91 uint8_t *ptro = reinterpret_cast<uint8_t *>(o);
92 const uint8_t *ptri = reinterpret_cast<const uint8_t *>(i);
93 const size_t main_part =
94 nelems_to_copy[a] * sizeof(data_t) / sizeof(uint32_t);
95 const size_t tail_part =
96 nelems_to_copy[a] * sizeof(data_t) % sizeof(uint32_t);
99 for (size_t e = 0; e < main_part; ++e) {
100 *(reinterpret_cast<uint32_t *>(ptro))
101 = *(reinterpret_cast<const uint32_t *>(ptri));
102 ptro += sizeof(uint32_t);
103 ptri += sizeof(uint32_t);
105 for (size_t e = 0; e < tail_part; ++e) {
112 for (size_t e = 0; e < nelems_to_copy[a]; ++e) o[e] = i[e];
118 template struct simple_concat_t<data_type::f32>;
119 template struct simple_concat_t<data_type::u8>;
120 template struct simple_concat_t<data_type::s8>;
121 template struct simple_concat_t<data_type::s32>;