2 * Copyright (c) 2017-2018 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(VEC_SIZE) && defined(DATA_TYPE)
28 #if defined(FIXED_POINT_POSITION)
29 #include "fixed_point.h"
31 #define ADD_OP(a, b) ADD_SAT_OP_EXPAND((a), (b), DATA_TYPE, VEC_SIZE)
32 #define SUB_OP(a, b) SUB_SAT_OP_EXPAND((a), (b), DATA_TYPE, VEC_SIZE)
33 #define MUL_OP(a, b) MUL_SAT_OP_EXPAND((a), (b), DATA_TYPE, VEC_SIZE, FIXED_POINT_POSITION)
34 #define INVSQRT_OP(a) INVSQRT_OP_EXPAND((a), DATA_TYPE, VEC_SIZE, FIXED_POINT_POSITION)
35 #define SQCVT_SAT(a) SQCVT_SAT_OP_EXPAND((a), DATA_TYPE, FIXED_POINT_POSITION)
37 #else /* FIXED_POINT_POSITION */
39 #define ADD_OP(a, b) ((a) + (b))
40 #define SUB_OP(a, b) ((a) - (b))
41 #define MUL_OP(a, b) ((a) * (b))
42 #define INVSQRT_OP(a) rsqrt((a))
43 #define SQCVT_SAT(a) (a)
45 #endif /* FIXED_POINT_POSITION */
48 #define ACTIVATION_FUNC(x) CLAMP(x, (DATA_TYPE)B_VAL, (DATA_TYPE)A_VAL)
50 #define ACTIVATION_FUNC(x) CLAMP(x, (DATA_TYPE)0, (DATA_TYPE)A_VAL)
52 #define ACTIVATION_FUNC(x) max(x, (DATA_TYPE)0)
54 #define ACTIVATION_FUNC(x) (x)
55 #endif /* FUSED_ACT */
57 /** Apply batch normalization.
59 * @param[in] input_ptr Pointer to the first source tensor. Supported data types: QS8/QS16/F16/F32
60 * @param[in] input_stride_x Stride of the first source tensor in X dimension (in bytes)
61 * @param[in] input_step_x input_stride_x * number of elements along X processed per workitem(in bytes)
62 * @param[in] input_stride_y Stride of the first source tensor in Y dimension (in bytes)
63 * @param[in] input_step_y input_stride_y * number of elements along Y processed per workitem(in bytes)
64 * @param[in] input_stride_z Stride of the first source tensor in Z dimension (in bytes)
65 * @param[in] input_step_z input_stride_z * number of elements along Z processed per workitem(in bytes)
66 * @param[in] input_offset_first_element_in_bytes The offset of the first element in the first source tensor
67 * @param[out] output_ptr Pointer to the destination tensor. Supported data types: same as @p input_ptr
68 * @param[in] output_stride_x Stride of the destination tensor in X dimension (in bytes)
69 * @param[in] output_step_x output_stride_x * number of elements along X processed per workitem(in bytes)
70 * @param[in] output_stride_y Stride of the destination tensor in Y dimension (in bytes)
71 * @param[in] output_step_y output_stride_y * number of elements along Y processed per workitem(in bytes)
72 * @param[in] output_stride_z Stride of the destination tensor in Z dimension (in bytes)
73 * @param[in] output_step_z output_stride_z * number of elements along Z processed per workitem(in bytes)
74 * @param[in] output_offset_first_element_in_bytes The offset of the first element in the destination tensor
75 * @param[in] mean_ptr Pointer to the mean source tensor. Supported data types: same as @p input_ptr
76 * @param[in] mean_stride_x Stride of the mean source tensor in X dimension (in bytes)
77 * @param[in] mean_step_x mean_stride_x * number of elements along X processed per workitem(in bytes)
78 * @param[in] mean_offset_first_element_in_bytes The offset of the first element in the mean source tensor
79 * @param[in] var_ptr Pointer to the var tensor. Supported data types: same as @p input_ptr
80 * @param[in] var_stride_x Stride of the var tensor in X dimension (in bytes)
81 * @param[in] var_step_x var_stride_x * number of elements along X processed per workitem(in bytes)
82 * @param[in] var_offset_first_element_in_bytes The offset of the first element in the var source tensor
83 * @param[in] beta_ptr Pointer to the beta source tensor. Supported data types: same as @p input_ptr
84 * @param[in] beta_stride_x Stride of the beta source tensor in X dimension (in bytes)
85 * @param[in] beta_step_x beta_stride_x * number of elements along X processed per workitem(in bytes)
86 * @param[in] beta_offset_first_element_in_bytes The offset of the first element in the beta source tensor
87 * @param[in] gamma_ptr Pointer to the gamma source tensor. Supported data types: same as @p input_ptr
88 * @param[in] gamma_stride_x Stride of the gamma source tensor in X dimension (in bytes)
89 * @param[in] gamma_step_x gamma_stride_x * number of elements along X processed per workitem(in bytes)
90 * @param[in] gamma_offset_first_element_in_bytes The offset of the first element in the gamma source tensor
91 * @param[in] epsilon Epsilon parameter in the batch normalization equation
93 __kernel void batchnormalization_layer(TENSOR3D_DECLARATION(input),
95 TENSOR3D_DECLARATION(output),
96 #endif /* not IN_PLACE */
97 VECTOR_DECLARATION(mean),
98 VECTOR_DECLARATION(var),
99 VECTOR_DECLARATION(beta),
100 VECTOR_DECLARATION(gamma),
103 Tensor3D in = CONVERT_TO_TENSOR3D_STRUCT(input);
107 Tensor3D out = CONVERT_TO_TENSOR3D_STRUCT(output);
108 #endif /* IN_PLACE */
109 Vector mean = CONVERT_TO_VECTOR_STRUCT(mean);
110 Vector var = CONVERT_TO_VECTOR_STRUCT(var);
111 Vector beta = CONVERT_TO_VECTOR_STRUCT(beta);
112 Vector gamma = CONVERT_TO_VECTOR_STRUCT(gamma);
114 VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
116 VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
118 VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
120 VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
122 VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
124 VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
127 const int current_slice = get_global_id(2);
129 data = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)in.ptr);
130 denominator = *((__global DATA_TYPE *)(var.ptr + current_slice * var.stride_x));
131 denominator = INVSQRT_OP(ADD_OP(denominator, ((VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE))SQCVT_SAT(epsilon))));
133 // Calculate x bar and store results
134 numerator = *((__global DATA_TYPE *)(mean.ptr + current_slice * mean.stride_x));
135 numerator = SUB_OP(data, numerator);
136 x_bar = MUL_OP(numerator, denominator);
138 gamma_vec = *((__global DATA_TYPE *)(gamma.ptr + current_slice * gamma.stride_x));
139 beta_vec = *((__global DATA_TYPE *)(beta.ptr + current_slice * beta.stride_x));
141 VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
142 res = ADD_OP(MUL_OP(gamma_vec, x_bar), beta_vec);
144 res = ACTIVATION_FUNC(res);
147 (res, 0, (__global DATA_TYPE *)out.ptr);
150 #endif /* defined(VEC_SIZE) && defined(DATA_TYPE) */