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
|
/*******************************************************************************
* Copyright 2018 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 "mkldnn_types.h"
#include "c_types_map.hpp"
#include "jit_uni_x8s8s32x_convolution.hpp"
#include "utils.hpp"
#include "mkldnn_thread.hpp"
#include "type_helpers.hpp"
namespace mkldnn {
namespace impl {
namespace cpu {
using namespace mkldnn::impl::status;
using namespace mkldnn::impl::memory_format;
using namespace mkldnn::impl::utils;
template <cpu_isa_t isa, bool with_relu, impl::data_type_t src_type, data_type_t dst_type>
void _jit_uni_x8s8s32x_convolution_fwd_t<isa, with_relu, src_type, dst_type>::execute_forward() {
auto src = reinterpret_cast<const src_data_t *>(this->input_memory(0));
auto weights = reinterpret_cast<const wei_data_t *>(this->input_memory(1));
auto bias = reinterpret_cast<const char *>(this->input_memory(2));
auto dst = reinterpret_cast<dst_data_t *>(this->memory());
const memory_desc_wrapper src_d(conf_.src_pd());
const memory_desc_wrapper dst_d(conf_.dst_pd());
const memory_desc_wrapper weights_d(conf_.weights_pd(0));
const memory_desc_wrapper bias_d(conf_.weights_pd(1));
const auto &jcp = kernel_->jcp;
size_t offset = (size_t)jcp.ngroups * rnd_up(jcp.oc, jcp.oc_block) * rnd_up(jcp.ic, jcp.ic_block) * jcp.kh * jcp.kw;
auto w = const_cast<wei_data_t *>(weights);
int32_t* compensation = (jcp.signed_input) ? reinterpret_cast<int32_t *>(&w[offset]) : 0;
const size_t bia_dt_size = conf_.with_bias() ? types::data_type_size(conf_.cdesc()->bias_desc.data_type) : 0;
float* scales = conf_.attr()->output_scales_.scales_;
int ocb_work = div_up(jcp.nb_oc, jcp.nb_oc_blocking);
const size_t work_amount = jcp.mb * jcp.ngroups * ocb_work * jcp.oh;
auto ker = [&](const int ithr, const int nthr) {
size_t start{0}, end{0};
balance211(work_amount, nthr, ithr, start, end);
size_t n{0}, g{0}, ocbb{0}, oh{0};
nd_iterator_init(start, n, jcp.mb, g, jcp.ngroups, ocbb, ocb_work,
oh, jcp.oh);
for (size_t iwork = start; iwork < end; ++iwork) {
int ocb = ocbb * jcp.nb_oc_blocking;
int ocb_num = jcp.nb_oc_blocking;
jit_conv_call_s par_conv = {};
const int ij = oh * jcp.stride_h;
const int i_t_overflow = nstl::min(jcp.kh, div_up(nstl::max(0, jcp.t_pad - ij), (jcp.dilate_h+1)));
const int i_b_overflow = nstl::min(jcp.kh, div_up(nstl::max(jcp.ih, ij + (jcp.kh-1) * (jcp.dilate_h+1) -
jcp.t_pad+1) - jcp.ih, (jcp.dilate_h + 1)));
const size_t _oc = g * jcp.nb_oc + ocb;
const size_t _ic = g * jcp.nb_ic;
const int ih = nstl::max(ij - jcp.t_pad + i_t_overflow * (jcp.dilate_h + 1), 0);
par_conv.src = &src[src_d.blk_off(n, _ic*jcp.ic_block, ih, 0)];
size_t dst_off = dst_d.blk_off(n, _oc*jcp.oc_block, oh, 0);
par_conv.dst = &dst[dst_off];
const int wh = (!jcp.signed_input) ? i_t_overflow : 0;
par_conv.filt = &weights[conf_.with_groups()
? weights_d.blk_off(g, ocb, 0, wh, 0)
: weights_d.blk_off(ocb, 0, wh, 0)];
if (bias)
par_conv.bias = &bias[bias_d.blk_off(_oc * jcp.oc_block*bia_dt_size)];
par_conv.oc_work =
nstl::min((ocb + ocb_num) * jcp.oc_block, jcp.oc) - ocb*jcp.oc_block;
par_conv.kw_padding = 0;
const int kh_padding = jcp.kh - i_t_overflow - i_b_overflow;
par_conv.kh_padding = nstl::max(0, kh_padding);
par_conv.scales = (jcp.signed_input) ? &local_scales_[jcp.is_oc_scale * _oc * jcp.oc_block]
: &scales[jcp.is_oc_scale * _oc * jcp.oc_block];
par_conv.compensation = (jcp.signed_input) ? compensation + _oc * jcp.oc_block : 0;
par_conv.t_overflow = i_t_overflow;
par_conv.b_overflow = i_b_overflow;
kernel_->jit_ker(&par_conv);
nd_iterator_step(n, jcp.mb, g, jcp.ngroups, ocbb, ocb_work, oh, jcp.oh);
}
};
parallel(0, ker);
}
template void _jit_uni_x8s8s32x_convolution_fwd_t<avx2, true, data_type::u8, data_type::u8>::execute_forward();
template void _jit_uni_x8s8s32x_convolution_fwd_t<avx2, true, data_type::u8, data_type::s8>::execute_forward();
template void _jit_uni_x8s8s32x_convolution_fwd_t<avx2, true, data_type::u8, data_type::s32>::execute_forward();
template void _jit_uni_x8s8s32x_convolution_fwd_t<avx2, true, data_type::u8, data_type::f32>::execute_forward();
template void _jit_uni_x8s8s32x_convolution_fwd_t<avx2, false, data_type::u8, data_type::u8>::execute_forward();
template void _jit_uni_x8s8s32x_convolution_fwd_t<avx2, false, data_type::u8, data_type::s8>::execute_forward();
template void _jit_uni_x8s8s32x_convolution_fwd_t<avx2, false, data_type::u8, data_type::s32>::execute_forward();
template void _jit_uni_x8s8s32x_convolution_fwd_t<avx2, false, data_type::u8, data_type::f32>::execute_forward();
template void _jit_uni_x8s8s32x_convolution_fwd_t<avx2, true, data_type::s8, data_type::u8>::execute_forward();
template void _jit_uni_x8s8s32x_convolution_fwd_t<avx2, true, data_type::s8, data_type::s8>::execute_forward();
template void _jit_uni_x8s8s32x_convolution_fwd_t<avx2, true, data_type::s8, data_type::s32>::execute_forward();
template void _jit_uni_x8s8s32x_convolution_fwd_t<avx2, true, data_type::s8, data_type::f32>::execute_forward();
template void _jit_uni_x8s8s32x_convolution_fwd_t<avx2, false, data_type::s8, data_type::u8>::execute_forward();
template void _jit_uni_x8s8s32x_convolution_fwd_t<avx2, false, data_type::s8, data_type::s8>::execute_forward();
template void _jit_uni_x8s8s32x_convolution_fwd_t<avx2, false, data_type::s8, data_type::s32>::execute_forward();
template void _jit_uni_x8s8s32x_convolution_fwd_t<avx2, false, data_type::s8, data_type::f32>::execute_forward();
template void _jit_uni_x8s8s32x_convolution_fwd_t<sse42, true, data_type::u8, data_type::u8>::execute_forward();
template void _jit_uni_x8s8s32x_convolution_fwd_t<sse42, true, data_type::u8, data_type::s8>::execute_forward();
template void _jit_uni_x8s8s32x_convolution_fwd_t<sse42, true, data_type::u8, data_type::s32>::execute_forward();
template void _jit_uni_x8s8s32x_convolution_fwd_t<sse42, true, data_type::u8, data_type::f32>::execute_forward();
template void _jit_uni_x8s8s32x_convolution_fwd_t<sse42, false, data_type::u8, data_type::u8>::execute_forward();
template void _jit_uni_x8s8s32x_convolution_fwd_t<sse42, false, data_type::u8, data_type::s8>::execute_forward();
template void _jit_uni_x8s8s32x_convolution_fwd_t<sse42, false, data_type::u8, data_type::s32>::execute_forward();
template void _jit_uni_x8s8s32x_convolution_fwd_t<sse42, false, data_type::u8, data_type::f32>::execute_forward();
template void _jit_uni_x8s8s32x_convolution_fwd_t<sse42, true, data_type::s8, data_type::u8>::execute_forward();
template void _jit_uni_x8s8s32x_convolution_fwd_t<sse42, true, data_type::s8, data_type::s8>::execute_forward();
template void _jit_uni_x8s8s32x_convolution_fwd_t<sse42, true, data_type::s8, data_type::s32>::execute_forward();
template void _jit_uni_x8s8s32x_convolution_fwd_t<sse42, true, data_type::s8, data_type::f32>::execute_forward();
template void _jit_uni_x8s8s32x_convolution_fwd_t<sse42, false, data_type::s8, data_type::u8>::execute_forward();
template void _jit_uni_x8s8s32x_convolution_fwd_t<sse42, false, data_type::s8, data_type::s8>::execute_forward();
template void _jit_uni_x8s8s32x_convolution_fwd_t<sse42, false, data_type::s8, data_type::s32>::execute_forward();
template void _jit_uni_x8s8s32x_convolution_fwd_t<sse42, false, data_type::s8, data_type::f32>::execute_forward();
}
}
}
|
