Blame - src/core/CL/cl_kernels/normalization_layer.cl - ml/ComputeLibrary

Anthony Barbier

6ff3b19

2017-09-04 18:44:23 +0100

[diff] [blame]

/*

*

* 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

* SOFTWARE.

*/

#include "helpers.h"

/** Apply cross map normalization.

27

*

28

* @note Datatype should be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=short

29

*

30

* @param[in] input_ptr Pointer to the first source tensor. Supported data types: F16, F32

31

* @param[in] input_stride_x Stride of the first source tensor in X dimension (in bytes)

32

* @param[in] input_step_x input_stride_x * number of elements along X processed per workitem(in bytes)

33

* @param[in] input_stride_y Stride of the first source tensor in Y dimension (in bytes)

34

* @param[in] input_step_y input_stride_y * number of elements along Y processed per workitem(in bytes)

35

* @param[in] input_stride_z Stride of the first source tensor in Z dimension (in bytes)

36

* @param[in] input_step_z input_stride_z * number of elements along Z processed per workitem(in bytes)

37

* @param[in] input_offset_first_element_in_bytes The offset of the first element in the first source tensor

38

* @param[in] squared_input_ptr Pointer to the second source tensor. Supported data types: F16, F32

39

* @param[in] squared_input_stride_x Stride of the second source tensor in X dimension (in bytes)

40

* @param[in] squared_input_step_x input_stride_x * number of elements along X processed per workitem(in bytes)

41

* @param[in] squared_input_stride_y Stride of the second source tensor in Y dimension (in bytes)

42

* @param[in] squared_input_step_y input_stride_y * number of elements along Y processed per workitem(in bytes)

43

* @param[in] squared_input_stride_z Stride of the second source tensor in Z dimension (in bytes)

44

* @param[in] squared_input_step_z input_stride_z * number of elements along Z processed per workitem(in bytes)

45

* @param[in] squared_input_offset_first_element_in_bytes The offset of the second element in the second source tensor

46

* @param[out] output_ptr Pointer to the destination tensor. Supported data types: F16, F32

47

* @param[in] output_stride_x Stride of the destination tensor in X dimension (in bytes)

48

* @param[in] output_step_x output_stride_x * number of elements along X processed per workitem(in bytes)

49

* @param[in] output_stride_y Stride of the destination tensor in Y dimension (in bytes)

50

* @param[in] output_step_y output_stride_y * number of elements along Y processed per workitem(in bytes)

51

* @param[in] output_stride_z Stride of the destination tensor in Z dimension (in bytes)

52

* @param[in] output_step_z output_stride_z * number of elements along Z processed per workitem(in bytes)

53

* @param[in] output_offset_first_element_in_bytes The offset of the first element in the destination tensor

54

* @param[in] coeff Alpha parameter / norm_size

55

* @param[in] beta Beta parameter in the normalization equation

56

* @param[in] kappa Kappa parameter in the normalization equation

57

* @param[in] radius Number of elements on the right or left side to normalize across

58

*/

59

__kernel void normalization_layer_cross_map(TENSOR3D_DECLARATION(input),

60

TENSOR3D_DECLARATION(squared_input),

61

TENSOR3D_DECLARATION(output),

float coeff,

float beta,

float kappa,

uint radius)

{

Tensor3D in = CONVERT_TO_TENSOR3D_STRUCT(input);

68

Tensor3D squared_in = CONVERT_TO_TENSOR3D_STRUCT(squared_input);

69

Tensor3D out = CONVERT_TO_TENSOR3D_STRUCT(output);

DATA_TYPE acc = 0;

const int num_of_slices = get_global_size(2);

74

const int current_slice = get_global_id(2);

75

76

const int left_slice = max(current_slice - (int)radius, (int)0);

77

const int right_slice = min(current_slice + (int)radius, (int)(num_of_slices - 1));

78

79

for(int i = left_slice; i <= right_slice; i++)

80

{

81

acc += *(__global DATA_TYPE *)tensor3D_offset(&squared_in, 0, 0, i - current_slice);

82

}

83

84

const float normalized = pow(kappa + coeff * (float)acc, beta);

85

86

const float normalized_pixel = (float) * ((__global DATA_TYPE *)in.ptr) / normalized;

87

88

*(__global DATA_TYPE *)out.ptr = CONVERT(normalized_pixel, DATA_TYPE);

89

}

90

91

/** Apply in map normalization.

92

*

93

* @note Datatype should be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=short

94

*

95

* @param[in] input_ptr Pointer to the first source tensor. Supported data types: F16, F32

96

* @param[in] input_stride_x Stride of the first source tensor in X dimension (in bytes)

97

* @param[in] input_step_x input_stride_x * number of elements along X processed per workitem(in bytes)

98

* @param[in] input_stride_y Stride of the first source tensor in Y dimension (in bytes)

99

* @param[in] input_step_y input_stride_y * number of elements along Y processed per workitem(in bytes)

100

* @param[in] input_stride_z Stride of the first source tensor in Z dimension (in bytes)

101

* @param[in] input_step_z input_stride_z * number of elements along Z processed per workitem(in bytes)

102

* @param[in] input_offset_first_element_in_bytes The offset of the first element in the first source tensor

103

* @param[in] squared_input_ptr Pointer to the second source tensor. Supported data types: F16, F32

104

* @param[in] squared_input_stride_x Stride of the second source tensor in X dimension (in bytes)

105

* @param[in] squared_input_step_x input_stride_x * number of elements along X processed per workitem(in bytes)

106

* @param[in] squared_input_stride_y Stride of the second source tensor in Y dimension (in bytes)

107

* @param[in] squared_input_step_y input_stride_y * number of elements along Y processed per workitem(in bytes)

108

* @param[in] squared_input_stride_z Stride of the second source tensor in Z dimension (in bytes)

109

* @param[in] squared_input_step_z input_stride_z * number of elements along Z processed per workitem(in bytes)

110

* @param[in] squared_input_offset_first_element_in_bytes The offset of the second element in the second source tensor

111

* @param[out] output_ptr Pointer to the destination tensor. Supported data types: F16, F32

112

* @param[in] output_stride_x Stride of the destination tensor in X dimension (in bytes)

113

* @param[in] output_step_x output_stride_x * number of elements along X processed per workitem(in bytes)

114

* @param[in] output_stride_y Stride of the first destination tensor in Y dimension (in bytes)

115

* @param[in] output_step_y output_stride_y * number of elements along Y processed per workitem(in bytes)

116

* @param[in] output_stride_z Stride of the first source tensor in Z dimension (in bytes)

117

* @param[in] output_step_z output_stride_z * number of elements along Z processed per workitem(in bytes)

118

* @param[in] output_offset_first_element_in_bytes The offset of the first element in the destination tensor

119

* @param[in] coeff Alpha parameter / norm_size

120

* @param[in] beta Beta parameter in the normalization equation

121

* @param[in] kappa Kappa parameter in the normalization equation

122

* @param[in] radius Number of elements on the right or left side to normalize across

123

*/

124

__kernel void normalization_layer_in_map_1D(TENSOR3D_DECLARATION(input),

125

TENSOR3D_DECLARATION(squared_input),

126

TENSOR3D_DECLARATION(output),

float coeff,

float beta,

float kappa,

uint radius)

{

Tensor3D in = CONVERT_TO_TENSOR3D_STRUCT(input);

133

Tensor3D squared_in = CONVERT_TO_TENSOR3D_STRUCT(squared_input);

134

Tensor3D out = CONVERT_TO_TENSOR3D_STRUCT(output);

135

136

VEC_DATA_TYPE(DATA_TYPE, 4)

137

acc_vec = 0;

138

139

const int current_pos = get_global_id(0) << 2;

140

141

const int left_pos = max(current_pos - (int)radius, -3);

142

const int right_pos = min(current_pos + (int)radius, (int)((get_global_size(0) << 2) + 3 - 1));

143

144

for(int i = left_pos; i <= right_pos; i += 1)

145

{

146

acc_vec += vload4(0, (__global DATA_TYPE *)tensor3D_offset(&squared_in, i - current_pos, 0, 0));

147

}

148

149

const float4 normalized = pow((float4)kappa + coeff * (float4)acc_vec, beta);

150

151

const float4 normalized_pixel = CONVERT(vload4(0, (__global DATA_TYPE *)in.ptr), float4) / normalized;

152

153

vstore4(CONVERT(normalized_pixel, VEC_DATA_TYPE(DATA_TYPE, 4)), 0, (__global DATA_TYPE *)out.ptr);

154

}