2 * Copyright (c) 2017 ARM Limited.
4 * SPDX-License-Identifier: MIT
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:
13 * The above copyright notice and this permission notice shall be included in all
14 * copies or substantial portions of the Software.
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
26 #if defined(FIXED_POINT_POSITION)
28 #include "fixed_point.h"
29 #define MUL_OP(x, y) MUL_SAT_OP_EXPAND((x), (y), DATA_TYPE, VEC_SIZE, FIXED_POINT_POSITION)
30 #define ADD_OP(x, y) ADD_SAT_OP_EXPAND((x), (y), DATA_TYPE, VEC_SIZE)
31 #define DIV_OP(x, y) DIV_SAT_OP_VEC_EXPAND((x), (y), DATA_TYPE, VEC_SIZE, FIXED_POINT_POSITION)
32 #define EXP_OP(x) EXP_OP_EXPAND((x), DATA_TYPE, VEC_SIZE, FIXED_POINT_POSITION)
33 #define LOG_OP(x) LOG_OP_EXPAND((x), DATA_TYPE, VEC_SIZE, FIXED_POINT_POSITION)
34 #define POW_OP(x, y) EXP_OP(MUL_OP(LOG_OP((x)), (y)))
35 #define SQCVT_SAT(a) SQCVT_SAT_OP_EXPAND((a), DATA_TYPE, FIXED_POINT_POSITION)
37 #define LOAD_OP(offset, ptr) vload16(offset, ptr)
38 #define STORE_OP(data, offset, ptr) vstore16(data, offset, ptr)
40 #else // FIXED_POINT_POSITION
42 #define MUL_OP(x, y) ((x) * (y))
43 #define ADD_OP(x, y) ((x) + (y))
44 #define DIV_OP(x, y) ((x) / (y))
45 #define POW_OP(x, y) pow((x), (y))
46 #define SQCVT_SAT(a) (a)
48 #define LOAD_OP(offset, ptr) vload4(offset, ptr)
49 #define STORE_OP(data, offset, ptr) vstore4(data, offset, ptr)
51 #endif // FIXED_POINT_POSITION
53 /** Apply cross-map normalization.
55 * @note Datatype should be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=short
56 * @note Vector size should be given as a preprocessor argument using -DVEC_SIZE=size, e.g. -DVEC_SIZE=16
57 * @note The radius should be given as a preprocessor argument using -DRADIUS=size. e.g. -DRADIUS=5
58 * @note The number of slices should be given as a preprocessor argument using -DNUM_SLICES=size. e.g. -DNUM_SLICES=192
59 * @note In case of fixed-point operation -DFIXED_POINT_POSITION=fixed_point_position must be provided: e.g. -DFIXED_POINT_POSITION=3
60 * @note Scaling coefficient (= alpha/norm_size), beta and kappa need to be passed at compile time using -DCOEFF, -DALPHA and -DKAPPA
62 * @param[in] input_ptr Pointer to the first source tensor. Supported data types: QS8/QS16/F16/F32
63 * @param[in] input_stride_x Stride of the first source tensor in X dimension (in bytes)
64 * @param[in] input_step_x input_stride_x * number of elements along X processed per workitem(in bytes)
65 * @param[in] input_stride_y Stride of the first source tensor in Y dimension (in bytes)
66 * @param[in] input_step_y input_stride_y * number of elements along Y processed per workitem(in bytes)
67 * @param[in] input_stride_z Stride of the first source tensor in Z dimension (in bytes)
68 * @param[in] input_step_z input_stride_z * number of elements along Z processed per workitem(in bytes)
69 * @param[in] input_offset_first_element_in_bytes The offset of the first element in the first source tensor
70 * @param[out] output_ptr Pointer to the destination tensor. Supported data types: same as @p input_ptr
71 * @param[in] output_stride_x Stride of the destination tensor in X dimension (in bytes)
72 * @param[in] output_step_x output_stride_x * number of elements along X processed per workitem(in bytes)
73 * @param[in] output_stride_y Stride of the destination tensor in Y dimension (in bytes)
74 * @param[in] output_step_y output_stride_y * number of elements along Y processed per workitem(in bytes)
75 * @param[in] output_stride_z Stride of the destination tensor in Z dimension (in bytes)
76 * @param[in] output_step_z output_stride_z * number of elements along Z processed per workitem(in bytes)
77 * @param[in] output_offset_first_element_in_bytes The offset of the first element in the destination tensor
79 __kernel void normalization_layer_cross_map(TENSOR3D_DECLARATION(input),
80 TENSOR3D_DECLARATION(output))
82 Tensor3D in = CONVERT_TO_TENSOR3D_STRUCT(input);
83 Tensor3D out = CONVERT_TO_TENSOR3D_STRUCT(output);
85 VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
86 acc = (VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE))0;
87 const VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
88 coeff_v = (VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE))SQCVT_SAT(COEFF);
89 const VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
90 beta_v = (VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE))SQCVT_SAT(BETA);
91 const VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
92 kappa_v = (VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE))SQCVT_SAT(KAPPA);
94 const int current_slice = get_global_id(2);
95 const int left_slice = max(-(int)RADIUS, -current_slice);
96 const int right_slice = min((int)RADIUS, (int)NUM_SLICES - 1 - current_slice);
98 for(int i = left_slice; i <= right_slice; i++)
100 VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
101 values = LOAD_OP(0, (__global DATA_TYPE *)tensor3D_offset(&in, 0, 0, i));
102 acc = ADD_OP(acc, MUL_OP(values, values));
105 acc = ADD_OP(MUL_OP(acc, coeff_v), kappa_v);
106 const VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
107 normalized = POW_OP(acc, beta_v);
108 const VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
109 normalized_pixel = DIV_OP(LOAD_OP(0, (__global DATA_TYPE *)in.ptr), normalized);
111 STORE_OP(normalized_pixel, 0, (__global DATA_TYPE *)out.ptr);
114 /** Apply in-map normalization.
116 * @note Datatype should be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=short
117 * @note Vector size should be given as a preprocessor argument using -DVEC_SIZE=size, e.g. -DVEC_SIZE=16
118 * @note The radius should be given as a preprocessor argument using -DRADIUS=size. e.g. -DRADIUS=5
119 * @note In case of fixed-point operation -DFIXED_POINT_POSITION=fixed_point_position must be provided: e.g. -DFIXED_POINT_POSITION=3
120 * @note Scaling coefficient (= alpha/norm_size), beta and kappa need to be passed at compile time using -DCOEFF, -DALPHA and -DKAPPA
122 * @param[in] input_ptr Pointer to the first source tensor. Supported data types: QS8/F16/F32
123 * @param[in] input_stride_x Stride of the first source tensor in X dimension (in bytes)
124 * @param[in] input_step_x input_stride_x * number of elements along X processed per workitem(in bytes)
125 * @param[in] input_stride_y Stride of the first source tensor in Y dimension (in bytes)
126 * @param[in] input_step_y input_stride_y * number of elements along Y processed per workitem(in bytes)
127 * @param[in] input_stride_z Stride of the first source tensor in Z dimension (in bytes)
128 * @param[in] input_step_z input_stride_z * number of elements along Z processed per workitem(in bytes)
129 * @param[in] input_offset_first_element_in_bytes The offset of the first element in the first source tensor
130 * @param[out] output_ptr Pointer to the destination tensor. Supported data types: same as @p input_ptr
131 * @param[in] output_stride_x Stride of the destination tensor in X dimension (in bytes)
132 * @param[in] output_step_x output_stride_x * number of elements along X processed per workitem(in bytes)
133 * @param[in] output_stride_y Stride of the first destination tensor in Y dimension (in bytes)
134 * @param[in] output_step_y output_stride_y * number of elements along Y processed per workitem(in bytes)
135 * @param[in] output_stride_z Stride of the first source tensor in Z dimension (in bytes)
136 * @param[in] output_step_z output_stride_z * number of elements along Z processed per workitem(in bytes)
137 * @param[in] output_offset_first_element_in_bytes The offset of the first element in the destination tensor
139 __kernel void normalization_layer_in_map(TENSOR3D_DECLARATION(input),
140 TENSOR3D_DECLARATION(output))
142 Tensor3D in = CONVERT_TO_TENSOR3D_STRUCT(input);
143 Tensor3D out = CONVERT_TO_TENSOR3D_STRUCT(output);
145 VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
146 acc = (VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE))0;
147 const VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
148 coeff_v = (VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE))SQCVT_SAT(COEFF);
149 const VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
150 beta_v = (VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE))SQCVT_SAT(BETA);
151 const VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
152 kappa_v = (VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE))SQCVT_SAT(KAPPA);
154 const int current_col = get_global_id(0) << 2;
155 const int left_pos = max(-(int)RADIUS, -3 - current_col);
156 const int right_pos = min((int)RADIUS, (int)((get_global_size(0) << 2) + 3 - 1 - current_col));
158 #if defined(IN_MAP_2D)
159 const int current_row = get_global_id(1);
160 const int first_row = max(-(int)RADIUS, -current_row);
161 const int last_row = min((int)RADIUS, (int)get_global_size(1) - 1 - current_row);
162 #endif /* defined(IN_MAP_2D) */
164 #if defined(IN_MAP_2D)
165 for(int j = first_row; j <= last_row; ++j)
167 #endif /* defined(IN_MAP_2D) */
168 for(int i = left_pos; i <= right_pos; ++i)
170 #if defined(IN_MAP_2D)
171 VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
172 values = LOAD_OP(0, (__global DATA_TYPE *)tensor3D_offset(&in, i, j, 0));
173 #else /* defined(IN_MAP_2D) */
174 VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
175 values = LOAD_OP(0, (__global DATA_TYPE *)tensor3D_offset(&in, i, 0, 0));
176 #endif /* defined(IN_MAP_2D) */
177 acc = ADD_OP(acc, MUL_OP(values, values));
179 #if defined(IN_MAP_2D)
181 #endif /* defined(IN_MAP_2D) */
183 acc = ADD_OP(MUL_OP(acc, coeff_v), kappa_v);
184 const VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
185 normalized = POW_OP(acc, beta_v);
186 const VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
187 normalized_pixel = DIV_OP(LOAD_OP(0, (__global DATA_TYPE *)in.ptr), normalized);
189 STORE_OP(normalized_pixel, 0, (__global DATA_TYPE *)out.ptr);