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
|
/*
* Copyright (c) 2016-2018 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"
/** Calculate square sum of a vector
*
* @param[in] input Pointer to the first pixel.
*
* @return square sum of vector.
*/
inline DATA_TYPE square_sum(__global const DATA_TYPE *input)
{
VEC_DATA_TYPE(DATA_TYPE, 16)
in = vload16(0, input);
in *= in;
in.s01234567 += in.s89ABCDEF;
in.s0123 += in.s4567;
in.s01 += in.s23;
return (in.s0 + in.s1);
}
/** Calculate sum of a vector
*
* @param[in] input Pointer to the first pixel.
*
* @return sum of vector.
*/
inline DATA_TYPE sum(__global const DATA_TYPE *input)
{
VEC_DATA_TYPE(DATA_TYPE, 16)
in = vload16(0, input);
in.s01234567 += in.s89ABCDEF;
in.s0123 += in.s4567;
in.s01 += in.s23;
return (in.s0 + in.s1);
}
/** This kernel performs reduction given an operation.
*
* @note The data type must be passed at compile time using -DDATA_TYPE: e.g. -DDATA_TYPE=float
* @note The data size must be passed at compile time using -DDATA_SIZE e.g. -DDATA_SIZE=32
* @note The operation we want to perform must be passed at compile time using -DOPERATION e.g. -DOPERATION=square_sum
*
* @param[in] src_ptr Pointer to the source tensor. Supported data types: F32
* @param[in] src_stride_x Stride of the 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 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_offset_first_element_in_bytes The offset of the first element in the source tensor
* @param[in] partial_sum_ptr The local buffer to hold sumed values. Supported data types: same as @p src_ptt
* @param[in] partial_sum_stride_x Stride of the output tensor in X dimension (in bytes)
* @param[in] partial_sum_step_x partial_sum_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] partial_sum_stride_y Stride of the output tensor in Y dimension (in bytes)
* @param[in] partial_sum_step_y partial_sum_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] partial_sum_offset_first_element_in_bytes The offset of the first element in the source tensor
* @param[in] local_sums Local buffer for storing the partial sum
*/
__kernel void reduction_operation(
IMAGE_DECLARATION(src),
IMAGE_DECLARATION(partial_sum),
__local DATA_TYPE *local_sums)
{
Image src = CONVERT_TO_IMAGE_STRUCT(src);
Image partial_sum = CONVERT_TO_IMAGE_STRUCT(partial_sum);
unsigned int lsize = get_local_size(0);
unsigned int lid = get_local_id(0);
for(unsigned int y = 0; y < get_local_size(1); ++y)
{
local_sums[lid] = OPERATION((__global DATA_TYPE *)offset(&src, 0, y));
barrier(CLK_LOCAL_MEM_FENCE);
// Perform parallel reduction
for(unsigned int i = lsize >> 1; i > 0; i >>= 1)
{
if(lid < i)
{
local_sums[lid] += local_sums[lid + i];
}
barrier(CLK_LOCAL_MEM_FENCE);
}
if(lid == 0)
{
((__global DATA_TYPE *)offset(&partial_sum, get_group_id(0), y))[0] = local_sums[0];
}
}
}
|
