blob: d46a22600f8541c939f04ea89df79aefbbc5dffd [file] [log] [blame]
Michalis Spyrou04f089c2017-08-08 17:42:38 +01001/*
2 * Copyright (c) 2016, 2017 ARM Limited.
3 *
4 * SPDX-License-Identifier: MIT
5 *
6 * Permission is hereby granted, free of charge, to any person obtaining a copy
7 * of this software and associated documentation files (the "Software"), to
8 * deal in the Software without restriction, including without limitation the
9 * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
10 * sell copies of the Software, and to permit persons to whom the Software is
11 * furnished to do so, subject to the following conditions:
12 *
13 * The above copyright notice and this permission notice shall be included in all
14 * copies or substantial portions of the Software.
15 *
16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
19 * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
20 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
21 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
22 * SOFTWARE.
23 */
24#include "helpers.h"
25
26/** Calculate square sum of a vector
27 *
28 * @param[in] input Pointer to the first pixel.
29 *
30 * @return square sum of vector.
31 */
32inline DATA_TYPE square_sum(__global const DATA_TYPE *input)
33{
34 VEC_DATA_TYPE(DATA_TYPE, 16)
35 in = vload16(0, input);
36
37 in *= in;
38
39 in.s01234567 += in.s89ABCDEF;
40 in.s0123 += in.s4567;
41 in.s01 += in.s23;
42
43 return (in.s0 + in.s1);
44}
45
46/** Calculate sum of a vector
47 *
48 * @param[in] input Pointer to the first pixel.
49 *
50 * @return sum of vector.
51 */
52inline DATA_TYPE sum(__global const DATA_TYPE *input)
53{
54 VEC_DATA_TYPE(DATA_TYPE, 16)
55 in = vload16(0, input);
56
57 in.s01234567 += in.s89ABCDEF;
58 in.s0123 += in.s4567;
59 in.s01 += in.s23;
60
61 return (in.s0 + in.s1);
62}
63
64/** This kernel performs reduction given an operation.
65 *
66 * @note The data type must be passed at compile time using -DDATA_TYPE: e.g. -DDATA_TYPE=float
67 * @note The data size must be passed at compile time using -DDATA_SIZE e.g. -DDATA_SIZE=32
68 * @note The operation we want to perform must be passed at compile time using -DOPERATION e.g. -DOPERATION=square_sum
69 *
70 * @param[in] src_ptr Pointer to the source tensor. Supported data types: F32
71 * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes)
72 * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
73 * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor
74 * @param[in] partial_sum_ptr The local buffer to hold sumed values. Supported data types: same as @p src_ptt
75 * @param[in] partial_sum_stride_x Stride of the source tensor in X dimension (in bytes)
76 * @param[in] partial_sum_step_x partial_sum_stride_x * number of elements along X processed per workitem(in bytes)
77 * @param[in] partial_sum_offset_first_element_in_bytes The offset of the first element in the source tensor
78 * @param[in] local_sums Local buffer for storing the partioal sum
79 */
80__kernel void reduction_operation(
81 VECTOR_DECLARATION(src),
82 VECTOR_DECLARATION(partial_sum),
83 __local DATA_TYPE *local_sums)
84{
85 Vector src = CONVERT_TO_VECTOR_STRUCT(src);
86 Vector partial_sum = CONVERT_TO_VECTOR_STRUCT(partial_sum);
87
88 unsigned int lsize = get_local_size(0);
89 unsigned int lid = get_local_id(0);
90
91 local_sums[lid] = OPERATION((__global DATA_TYPE *)src.ptr);
92 barrier(CLK_LOCAL_MEM_FENCE);
93
94 // Perform parallel reduction
95 for(unsigned int i = lsize >> 1; i > 0; i >>= 1)
96 {
97 if(lid < i)
98 {
99 local_sums[lid] += local_sums[lid + i];
100 }
101 barrier(CLK_LOCAL_MEM_FENCE);
102 }
103
104 if(lid == 0)
105 {
106 ((__global DATA_TYPE *)partial_sum.ptr + get_group_id(0))[0] = local_sums[0];
107 }
108}