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
|
/*
* Copyright (c) 2018-2020 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"
#if defined(DATA_TYPE) && defined(VEC_SIZE) && defined(OFFSET) && defined(SCALE)
#define TYPE VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
#define OFFSET_FLT ((float)OFFSET)
#define SCALE_FLT ((float)SCALE)
#if defined(NUM_CHANNELS)
/** Apply normalize_planar_yuv layer on tensors with NCHW data layout.
*
* @note Data type should be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=float
* @note Vector size should be given as a preprocessor argument using -DVEC_SIZE e.g. -DVEC_SIZE=8
* @note The depth of the input tensor should be given as a preprocessor argument using -DNUM_CHANNELS e.g. -DNUM_CHANNELS=8
* @note The quantization offset should be given as a preprocessor argument using -DOFFSET e.g. -DOFFSET=8
* @note The quantization scale should be given as a preprocessor argument using -DSCALE e.g. -DSCALE=8
*
* @param[in] src_ptr Pointer to the first source tensor. Supported data types: QASYMM8/QASYMM8_SIGNED
* @param[in] src_stride_x Stride of the first source tensor in X dimension (in bytes)
* @param[in] src_step_x input_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 input_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 input_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_offset_first_element_in_bytes The offset of the first element in the destination tensor
* @param[in] mean_ptr Pointer to the mean source tensor. Supported data types: same as @p src_ptr
* @param[in] mean_stride_x Stride of the mean source tensor in X dimension (in bytes)
* @param[in] mean_step_x mean_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] mean_offset_first_element_in_bytes The offset of the first element in the mean source tensor
* @param[in] std_ptr Pointer to the std tensor. Supported data types: same as @p src_ptr
* @param[in] std_stride_x Stride of the std tensor in X dimension (in bytes)
* @param[in] std_step_x std_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] std_offset_first_element_in_bytes The offset of the first element in the var source tensor
*/
__kernel void normalize_planar_yuv_layer_q8_nchw(TENSOR3D_DECLARATION(src),
TENSOR3D_DECLARATION(dst),
VECTOR_DECLARATION(mean),
VECTOR_DECLARATION(std))
{
Tensor3D src = CONVERT_TO_TENSOR3D_STRUCT(src);
Tensor3D dst = CONVERT_TO_TENSOR3D_STRUCT(dst);
Vector mean = CONVERT_TO_VECTOR_STRUCT(mean);
Vector std = CONVERT_TO_VECTOR_STRUCT(std);
const uint current_slice = get_global_id(2) % NUM_CHANNELS;
float16 curr_mean_flt = (float16)(*((__global DATA_TYPE *)(mean.ptr + current_slice * sizeof(DATA_TYPE))));
curr_mean_flt = round(curr_mean_flt - OFFSET_FLT) * SCALE_FLT;
float16 curr_std_flt = (float16)(*((__global DATA_TYPE *)(std.ptr + current_slice * sizeof(DATA_TYPE))));
curr_std_flt = round(curr_std_flt - OFFSET_FLT) * SCALE_FLT;
float16 data_flt = CONVERT(VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)src.ptr), float16);
data_flt = round(data_flt - OFFSET_FLT) * SCALE_FLT;
// Perform normalization
float16 res_flt = (data_flt - curr_mean_flt) / curr_std_flt;
const TYPE res_u8 = CONVERT_SAT(round(res_flt / SCALE_FLT) + OFFSET_FLT, TYPE);
VSTORE(VEC_SIZE)
(res_u8, 0, (__global DATA_TYPE *)dst.ptr);
}
#endif // defined(NUM_CHANNELS)
/** Apply normalize_planar_yuv layer on tensors with NHWC data layout.
*
* @note Data type should be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=float
* @note Vector size should be given as a preprocessor argument using -DVEC_SIZE e.g. -DVEC_SIZE=8
* @note The quantization offset should be given as a preprocessor argument using -DOFFSET e.g. -DOFFSET=8
* @note The quantization scale should be given as a preprocessor argument using -DSCALE e.g. -DSCALE=8
*
* @param[in] src_ptr Pointer to the first source tensor. Supported data types: QASYMM8/QASYMM8_SIGNED
* @param[in] src_stride_x Stride of the first source tensor in X dimension (in bytes)
* @param[in] src_step_x input_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 input_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 input_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_offset_first_element_in_bytes The offset of the first element in the destination tensor
* @param[in] mean_ptr Pointer to the mean source tensor. Supported data types: same as @p src_ptr
* @param[in] mean_stride_x Stride of the mean source tensor in X dimension (in bytes)
* @param[in] mean_step_x mean_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] mean_offset_first_element_in_bytes The offset of the first element in the mean source tensor
* @param[in] std_ptr Pointer to the std tensor. Supported data types: same as @p src_ptr
* @param[in] std_stride_x Stride of the std tensor in X dimension (in bytes)
* @param[in] std_step_x std_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] std_offset_first_element_in_bytes The offset of the first element in the var source tensor
*/
__kernel void normalize_planar_yuv_layer_q8_nhwc(TENSOR3D_DECLARATION(src),
TENSOR3D_DECLARATION(dst),
VECTOR_DECLARATION(mean),
VECTOR_DECLARATION(std))
{
Tensor3D src = CONVERT_TO_TENSOR3D_STRUCT(src);
Tensor3D dst = CONVERT_TO_TENSOR3D_STRUCT(dst);
Vector mean = CONVERT_TO_VECTOR_STRUCT(mean);
Vector std = CONVERT_TO_VECTOR_STRUCT(std);
const uint current_slice = get_global_id(0);
float16 curr_mean_flt = CONVERT(VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(mean.ptr + current_slice * VEC_SIZE * sizeof(DATA_TYPE))), float16);
curr_mean_flt = round(curr_mean_flt - OFFSET_FLT) * SCALE_FLT;
float16 curr_std_flt = CONVERT(VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(std.ptr + current_slice * VEC_SIZE * sizeof(DATA_TYPE))), float16);
curr_std_flt = round(curr_std_flt - OFFSET_FLT) * SCALE_FLT;
float16 data_flt = CONVERT(VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)src.ptr), float16);
data_flt = round(data_flt - OFFSET_FLT) * (SCALE_FLT);
// Perform normalization
float16 res_flt = (data_flt - curr_mean_flt) / curr_std_flt;
const TYPE res_u8 = CONVERT_SAT(round(res_flt / SCALE_FLT) + OFFSET_FLT, TYPE);
VSTORE(VEC_SIZE)
(res_u8, 0, (__global DATA_TYPE *)dst.ptr);
}
#endif // defined(DATA_TYPE) && defined(VEC_SIZE) && defined(OFFSET) && defined(SCALE)
|
