Compute Library
 22.08
batchnormalization_layer.cl
Go to the documentation of this file.
1 /*
2  * Copyright (c) 2017-2021 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 #define ADD_OP(a, b) ((a) + (b))
27 #define SUB_OP(a, b) ((a) - (b))
28 #define MUL_OP(a, b) ((a) * (b))
29 #define INVSQRT_OP(a) rsqrt((a))
30 #define SQCVT_SAT(a) (a)
31 
32 #if defined(VEC_SIZE) && defined(DATA_TYPE) && defined(ACTIVATION_TYPE)
34 
35 /** Apply batch normalization.
36  *
37  * @note It is possible to select the activation function to apply using -DACTIVATION_TYPE e.g. -DACTIVATION_TYPE=relu
38  * @note A, B variables required by some activation functions are set using -DA_VAL= and -DB_VAL= respectively
39  *
40  * @param[in] input_ptr Pointer to the first source tensor. Supported data types: F16/F32
41  * @param[in] input_stride_x Stride of the first source tensor in X dimension (in bytes)
42  * @param[in] input_step_x input_stride_x * number of elements along X processed per workitem(in bytes)
43  * @param[in] input_stride_y Stride of the first source tensor in Y dimension (in bytes)
44  * @param[in] input_step_y input_stride_y * number of elements along Y processed per workitem(in bytes)
45  * @param[in] input_stride_z Stride of the first source tensor in Z dimension (in bytes)
46  * @param[in] input_step_z input_stride_z * number of elements along Z processed per workitem(in bytes)
47  * @param[in] input_offset_first_element_in_bytes The offset of the first element in the first source tensor
48  * @param[out] output_ptr Pointer to the destination tensor. Supported data types: same as @p input_ptr
49  * @param[in] output_stride_x Stride of the destination tensor in X dimension (in bytes)
50  * @param[in] output_step_x output_stride_x * number of elements along X processed per workitem(in bytes)
51  * @param[in] output_stride_y Stride of the destination tensor in Y dimension (in bytes)
52  * @param[in] output_step_y output_stride_y * number of elements along Y processed per workitem(in bytes)
53  * @param[in] output_stride_z Stride of the destination tensor in Z dimension (in bytes)
54  * @param[in] output_step_z output_stride_z * number of elements along Z processed per workitem(in bytes)
55  * @param[in] output_offset_first_element_in_bytes The offset of the first element in the destination tensor
56  * @param[in] mean_ptr Pointer to the mean source tensor. Supported data types: same as @p input_ptr
57  * @param[in] mean_stride_x Stride of the mean source tensor in X dimension (in bytes)
58  * @param[in] mean_step_x mean_stride_x * number of elements along X processed per workitem(in bytes)
59  * @param[in] mean_offset_first_element_in_bytes The offset of the first element in the mean source tensor
60  * @param[in] var_ptr Pointer to the var tensor. Supported data types: same as @p input_ptr
61  * @param[in] var_stride_x Stride of the var tensor in X dimension (in bytes)
62  * @param[in] var_step_x var_stride_x * number of elements along X processed per workitem(in bytes)
63  * @param[in] var_offset_first_element_in_bytes The offset of the first element in the var source tensor
64  * @param[in] beta_ptr Pointer to the beta source tensor. Supported data types: same as @p input_ptr
65  * @param[in] beta_stride_x Stride of the beta source tensor in X dimension (in bytes)
66  * @param[in] beta_step_x beta_stride_x * number of elements along X processed per workitem(in bytes)
67  * @param[in] beta_offset_first_element_in_bytes The offset of the first element in the beta source tensor
68  * @param[in] gamma_ptr Pointer to the gamma source tensor. Supported data types: same as @p input_ptr
69  * @param[in] gamma_stride_x Stride of the gamma source tensor in X dimension (in bytes)
70  * @param[in] gamma_step_x gamma_stride_x * number of elements along X processed per workitem(in bytes)
71  * @param[in] gamma_offset_first_element_in_bytes The offset of the first element in the gamma source tensor
72  * @param[in] epsilon Epsilon parameter in the batch normalization equation
73  */
74 __kernel void batchnormalization_layer_nchw(TENSOR3D_DECLARATION(input),
75 #ifndef IN_PLACE
76  TENSOR3D_DECLARATION(output),
77 #endif /* not IN_PLACE */
78  VECTOR_DECLARATION(mean),
79  VECTOR_DECLARATION(var),
80 #ifndef USE_DEFAULT_BETA
81  VECTOR_DECLARATION(beta),
82 #endif /* USE_DEFAULT_BETA */
83 #ifndef USE_DEFAULT_GAMMA
84  VECTOR_DECLARATION(gamma),
85 #endif /* USE_DEFAULT_GAMMA */
86  float epsilon)
87 {
89 #ifdef IN_PLACE
90  Tensor3D out = in;
91 #else /* IN_PLACE */
93 #endif /* IN_PLACE */
94  Vector mean = CONVERT_TO_VECTOR_STRUCT(mean);
96 #ifndef USE_DEFAULT_BETA
97  Vector beta = CONVERT_TO_VECTOR_STRUCT(beta);
98 #endif /* USE_DEFAULT_BETA */
99 #ifndef USE_DEFAULT_GAMMA
100  Vector gamma = CONVERT_TO_VECTOR_STRUCT(gamma);
101 #endif /* USE_DEFAULT_GAMMA */
102 
103  VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
104  data = 0;
105  VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
106  denominator = 0;
107  VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
108  numerator = 0;
109  VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
110  x_bar = 0;
111  VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
112  res = 0;
113 
114  const int current_slice = get_global_id(2);
115 
116  data = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)in.ptr);
117  denominator = *((__global DATA_TYPE *)(var.ptr + current_slice * var.stride_x));
118  denominator = INVSQRT_OP(ADD_OP(denominator, ((VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE))SQCVT_SAT(epsilon))));
119 
120  // Calculate x bar and store results
121  numerator = *((__global DATA_TYPE *)(mean.ptr + current_slice * mean.stride_x));
122  numerator = SUB_OP(data, numerator);
123  x_bar = MUL_OP(numerator, denominator);
124 
125 #ifndef USE_DEFAULT_GAMMA
126  VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
127  gamma_vec = *((__global DATA_TYPE *)(gamma.ptr + current_slice * gamma.stride_x));
128 
129  res = MUL_OP(gamma_vec, x_bar);
130 #else /* USE_DEFAULT_GAMMA */
131  // gamma is equal to 1, no need to perform multiplications
132  res = x_bar;
133 #endif /* USE_DEFAULT_GAMMA */
134 
135 #ifndef USE_DEFAULT_BETA
136  VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
137  beta_vec = *((__global DATA_TYPE *)(beta.ptr + current_slice * beta.stride_x));
138  // beta is not zero, hence we need to perform the addition
139  res = ADD_OP(res, beta_vec);
140 #endif /* USE_DEFAULT_BETA */
141 
142  res = ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, res, A_VAL, B_VAL);
143 
145  (res, 0, (__global DATA_TYPE *)out.ptr);
146 }
147 #endif /* defined(VEC_SIZE) && defined(DATA_TYPE) && defined(DATA_TYPE)*/
Structure to hold Vector information.
Definition: helpers.h:889
#define VEC_SIZE
#define MUL_OP(a, b)
#define SQCVT_SAT(a)
Structure to hold 3D tensor information.
Definition: helpers.h:906
#define ADD_OP(a, b)
#define CONVERT_TO_VECTOR_STRUCT(name)
Definition: helpers.h:849
#define VECTOR_DECLARATION(name)
Definition: helpers.h:799
int stride_x
Stride of the image in X dimension (in bytes)
Definition: helpers.h:893
__global uchar * ptr
Pointer to the starting postion of the buffer.
Definition: helpers.h:891
#define CONVERT_TO_TENSOR3D_STRUCT(name)
Definition: helpers.h:870
#define SUB_OP(a, b)
#define VSTORE(size)
Definition: helpers.h:458
__global uchar * ptr
Pointer to the starting postion of the buffer.
Definition: helpers.h:908
#define VLOAD(size)
Definition: helpers.h:204
#define TENSOR3D_DECLARATION(name)
Definition: helpers.h:813
#define ACTIVATION(op, DATA_TYPE, VEC_SIZE, x, A_VAL, B_VAL)
#define INVSQRT_OP(a)
#define VEC_DATA_TYPE(type, size)
Definition: helpers.h:728