xref: /aosp_15_r20/external/ComputeLibrary/src/core/CL/cl_kernels/nhwc/batchnormalization_layer.cl (revision c217d954acce2dbc11938adb493fc0abd69584f3)
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)
33#include "activation_float_helpers.h"
34
35/** Apply batch normalization on tensors with NHWC format.
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_nhwc(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{
88    uint x_offs = max((int)(get_global_id(0) * VEC_SIZE * sizeof(DATA_TYPE) - (VEC_SIZE - VEC_SIZE_LEFTOVER) % VEC_SIZE * sizeof(DATA_TYPE)), 0);
89
90    __global uchar *input_addr = input_ptr + input_offset_first_element_in_bytes + x_offs + get_global_id(1) * input_stride_y + get_global_id(2) * input_stride_z;
91#ifdef IN_PLACE
92    __global uchar *output_addr = input_ptr;
93#else  /* IN_PLACE */
94    __global uchar *output_addr = output_ptr + output_offset_first_element_in_bytes + x_offs + get_global_id(1) * output_stride_y + get_global_id(2) * output_stride_z;
95#endif /* IN_PLACE */
96    __global uchar *mean_addr = mean_ptr + mean_offset_first_element_in_bytes + x_offs;
97    __global uchar *var_addr  = var_ptr + var_offset_first_element_in_bytes + x_offs;
98#ifndef USE_DEFAULT_BETA
99    __global uchar *beta_addr = beta_ptr + beta_offset_first_element_in_bytes + x_offs;
100#endif /* USE_DEFAULT_BETA */
101#ifndef USE_DEFAULT_GAMMA
102    __global uchar *gamma_addr = gamma_ptr + gamma_offset_first_element_in_bytes + x_offs;
103#endif /* USE_DEFAULT_GAMMA */
104
105    VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
106    data = 0;
107    VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
108    denominator = 0;
109    VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
110    numerator = 0;
111    VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
112    x_bar = 0;
113    VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
114    res0 = 0;
115
116    data        = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)input_addr);
117    denominator = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)var_addr);
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 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)mean_addr);
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 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)gamma_addr);
128
129    res0 = MUL_OP(gamma_vec, x_bar);
130#else  /* USE_DEFAULT_GAMMA */
131    // gamma is equal to 1, no need to perform multiplications
132    res0 = x_bar;
133#endif /* USE_DEFAULT_GAMMA */
134
135#ifndef USE_DEFAULT_BETA
136    VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
137    beta_vec = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)beta_addr);
138    // beta is not zero, hence we need to perform the addition
139    res0 = ADD_OP(res0, beta_vec);
140#endif /* USE_DEFAULT_BETA */
141
142    res0 = ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE, res0, A_VAL, B_VAL);
143
144    STORE_VECTOR_SELECT(res, DATA_TYPE, output_addr, VEC_SIZE, VEC_SIZE_LEFTOVER, VEC_SIZE_LEFTOVER != 0 && get_global_id(0) == 0)
145}
146#endif /* defined(VEC_SIZE) && defined(DATA_TYPE) && defined(DATA_TYPE)*/