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
|
/*
* Copyright (c) 2018-2021 Arm Limited.
*
* SPDX-License-Identifier: MIT
*
* Permission is hereby granted, free of charge, to any person obtaining a copy
* of this software and associated documentation files (the "Software"), to
* deal in the Software without restriction, including without limitation the
* rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
* sell copies of the Software, and to permit persons to whom the Software is
* furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice shall be included in all
* copies or substantial portions of the Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
* AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
* SOFTWARE.
*/
#include "helpers.h"
#include "tile_helpers.h"
#if defined(DATA_TYPE) && defined(VEC_SIZE) && defined(NUM_GROUPS) && defined(K) && defined(SRC_DIM_Z)
// Check valid VEC_SIZES
#if VEC_SIZE != 1 && VEC_SIZE != 2 && VEC_SIZE != 3 && VEC_SIZE != 4 && VEC_SIZE != 8 && VEC_SIZE != 16
#error "Only vector sizes 1, 2, 3, 4, 8 and 16 are supported"
#endif // VEC_SIZE != 1 && VEC_SIZE != 2 && VEC_SIZE != 3 && VEC_SIZE != 4 && VEC_SIZE != 8 && VEC_SIZE != 16
#define DIV_MOD_UINT(x, y, div_res, mod_res) \
({ \
div_res = (uint)((x) * (float)(1.0f / (float)(y))); \
uint r = div_res * (y); \
mod_res = (x)-r; \
})
#if defined(VEC_SIZE) && defined(VEC_SIZE_LEFTOVER) && defined(SRC_DIM_X)
/** Performs channel shuffle when the data layout is NHWC. See https://arxiv.org/pdf/1707.01083.pdf for details.
*
* @note The vector size must be given as a preprocessor argument using -DVEC_SIZE=num. e.g. -DVEC_SIZE=4
* @note The third dimension of the tensor must be given as a preprocessor argument using -DSRC_DIM_Z=num. e.g. -DSRC_DIM_Z=64
* @note The first dimension of the tensor must be given as a preprocessor argument using -DSRC_DIM_X=num. e.g. -DSRC_DIM_X=64
* @note The number of groups must be given as a preprocessor argument using -DNUM_GROUPS=num_groups. e.g. -DNUM_GROUPS=2
* @note The number of channels in each group must be given as a preprocessor argument using -DK=num. e.g. -DK=1
* K is equal to num_channels / num_groups.
* @note The leftover size in the X dimension shoud be given as preprocessor argument using -DVEC_SIZE_LEFTOVER is; x_dimension % VEC_SIZE. e.g. -DVEC_SIZE_LEFTOVER=1
*
* @param[in] src_ptr Pointer to the source matrix. Supported data types: All
* @param[in] src_stride_x Stride of the first source tensor in X dimension (in bytes)
* @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] src_stride_y Stride of the first source tensor in Y dimension (in bytes)
* @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] src_stride_z Stride of the first source tensor in Z dimension (in bytes)
* @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes)
* @param[in] src_stride_w Stride of the first source tensor in Z dimension (in bytes)
* @param[in] src_step_w src_stride_z * number of elements along Z processed per workitem(in bytes)
* @param[in] src_offset_first_element_in_bytes The offset of the first element in the first source tensor
* @param[out] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr
* @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes)
* @param[in] dst_step_x output_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes)
* @param[in] dst_step_y output_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes)
* @param[in] dst_step_z output_stride_z * number of elements along Z processed per workitem(in bytes)
* @param[in] dst_stride_w Stride of the destination tensor in Z dimension (in bytes)
* @param[in] dst_step_w output_stride_z * number of elements along Z processed per workitem(in bytes)
* @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor
*/
__kernel void channel_shuffle_nhwc(TENSOR4D_DECLARATION(src),
TENSOR4D_DECLARATION(dst))
{
// Offset computation
const uint curr_out_channel = GET_SPATIAL_IDX(0, VEC_SIZE, VEC_SIZE_LEFTOVER); // output feature map
uint z = 0;
uint batch_id = 0;
// Compute curr_channel and batch_id
DIV_MOD_UINT(get_global_id(2), (uint)SRC_DIM_Z, batch_id, z);
VEC_DATA_TYPE(uint, VEC_SIZE)
curr_out_channels = (VEC_DATA_TYPE(uint, VEC_SIZE))(curr_out_channel) + VEC_OFFS(uint, VEC_SIZE);
VEC_DATA_TYPE(uint, VEC_SIZE)
in_channels = (curr_out_channels * (VEC_DATA_TYPE(uint, VEC_SIZE))(K)) % (VEC_DATA_TYPE(uint, VEC_SIZE))(SRC_DIM_X) + (curr_out_channels / (VEC_DATA_TYPE(uint, VEC_SIZE))(NUM_GROUPS));
// Load the values
const __global DATA_TYPE *input_ptr = (const __global DATA_TYPE *)(src_ptr + src_offset_first_element_in_bytes + get_global_id(1) * src_stride_y + z * src_stride_z + batch_id * src_stride_w);
#if VEC_SIZE == 1
DATA_TYPE out0 = *((const __global * DATA_TYPE)(input_ptr) + in_channels);
#elif VEC_SIZE == 2
VEC_DATA_TYPE(DATA_TYPE, 2)
out0 =
{
*(input_ptr + in_channels.s0),
*(input_ptr + in_channels.s1)
};
#elif VEC_SIZE == 3
VEC_DATA_TYPE(DATA_TYPE, 3)
out0 =
{
*(input_ptr + in_channels.s0),
*(input_ptr + in_channels.s1),
*(input_ptr + in_channels.s2)
};
#elif VEC_SIZE == 4
VEC_DATA_TYPE(DATA_TYPE, 4)
out0 =
{
*(input_ptr + in_channels.s0),
*(input_ptr + in_channels.s1),
*(input_ptr + in_channels.s2),
*(input_ptr + in_channels.s3)
};
#elif VEC_SIZE == 8
VEC_DATA_TYPE(DATA_TYPE, 8)
out0 =
{
*(input_ptr + in_channels.s0),
*(input_ptr + in_channels.s1),
*(input_ptr + in_channels.s2),
*(input_ptr + in_channels.s3),
*(input_ptr + in_channels.s4),
*(input_ptr + in_channels.s5),
*(input_ptr + in_channels.s6),
*(input_ptr + in_channels.s7)
};
#elif VEC_SIZE == 16
VEC_DATA_TYPE(DATA_TYPE, 16)
out0 =
{
*(input_ptr + in_channels.s0),
*(input_ptr + in_channels.s1),
*(input_ptr + in_channels.s2),
*(input_ptr + in_channels.s3),
*(input_ptr + in_channels.s4),
*(input_ptr + in_channels.s5),
*(input_ptr + in_channels.s6),
*(input_ptr + in_channels.s7),
*(input_ptr + in_channels.s8),
*(input_ptr + in_channels.s9),
*(input_ptr + in_channels.sa),
*(input_ptr + in_channels.sb),
*(input_ptr + in_channels.sc),
*(input_ptr + in_channels.sd),
*(input_ptr + in_channels.se),
*(input_ptr + in_channels.sf)
};
#endif // VEC_SIZE == 1
__global uchar *output_ptr = dst_ptr + curr_out_channel * sizeof(DATA_TYPE) + dst_offset_first_element_in_bytes + get_global_id(1) * dst_stride_y + z * dst_stride_z + batch_id * dst_stride_w;
STORE_VECTOR_SELECT(out, DATA_TYPE, output_ptr, VEC_SIZE, VEC_SIZE_LEFTOVER, VEC_SIZE_LEFTOVER != 0 && get_global_id(0) == 0);
}
#endif // defined(VEC_SIZE) && defined(VEC_SIZE_LEFTOVER) && defined(SRC_DIM_X)
#endif // defined(DATA_TYPE) && defined(VEC_SIZE) && defined(NUM_GROUPS) && defined(K) && defined(SRC_DIM_Z)
|
