1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
|
/*
* Copyright (c) 2018-2021 Arm Limited.
*
* SPDX-License-Identifier: MIT
*
* Permission is hereby granted, free of charge, to any person obtaining a copy
* of this software and associated documentation files (the "Software"), to
* deal in the Software without restriction, including without limitation the
* rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
* sell copies of the Software, and to permit persons to whom the Software is
* furnished to do so, subject to the following conditions:
*
* The above copyright notice and this permission notice shall be included in all
* copies or substantial portions of the Software.
*
* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
* AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
* SOFTWARE.
*/
#include "helpers.h"
#define SUB(x, y) (x - y)
#define ADD(x, y) (x + y)
#define MAX(x, y) max((x), (y))
#define MIN(x, y) min((x), (y))
#define SQUARED_DIFF(x, y) (x - y) * (x - y)
#define PRELU(x, y) (select(y * x, x, CONVERT((x > (DATA_TYPE)0), SELECT_VEC_DATA_TYPE(float, VEC_SIZE_OUT))))
#define DIV(x, y) (x / y)
#define CONVERT_RTE(x, type) (convert_##type##_rte((x)))
#define CONVERT_DOWN(x, type) CONVERT_RTE(x, type)
#define OP_FUN_NAME_STR(op) elementwise_operation_##op##_quantized
#define OP_FUN_NAME(op) OP_FUN_NAME_STR(op)
#if defined(OP) && defined(VEC_SIZE_IN1) && defined(VEC_SIZE_IN2) && defined(VEC_SIZE_OUT) && defined(OFFSET_IN1) && defined(OFFSET_IN2) && defined(OFFSET_OUT) && defined(SCALE_IN1) && defined(SCALE_IN2) && defined(SCALE_OUT) && defined(DATA_TYPE)
#define VEC_FLOAT VEC_DATA_TYPE(float, VEC_SIZE_OUT)
#define VEC_INT VEC_DATA_TYPE(int, VEC_SIZE_OUT)
#define VEC_TYPE VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE_OUT)
/** This function executes an element-wise operation among two tensors.
*
* @note Vector sizes of inputs and output have to be passed at compile time using -DVEC_SIZE_IN1, -DVEC_SIZE_IN2, -DVEC_SIZE_OUT.
* @note Leftover vector size has to be passed at compile time using -DVEC_SIZE_LEFTOVER. e.g. -DVEC_SIZE=3. It is defined as the remainder between the input's first dimension and VEC_SIZE
* @note In case of broadcasting along the X dimension the proper preprocessor argument should be passed depending on the input (e.g. -DIS_IN1_X_BROADCASTING, -DIS_IN2_X_BROADCASTING)
* @note The quantization offset of the first operand must be passed at compile time using -DOFFSET_IN1, i.e. -DOFFSET_IN1=10
* @note The quantization offset of the second operand must be passed at compile time using -DOFFSET_IN2, i.e. -DOFFSET_IN2=10
* @note The quantization offset of the output must be passed at compile time using -DOFFSET_OUT, i.e. -DOFFSET_OUT=10
* @note The quantization scale of the first operand must be passed at compile time using -DSCALE_IN1, i.e. -DSCALE_IN1=10
* @note The quantization scale of the second operand must be passed at compile time using -DSCALE_IN2, i.e. -DSCALE_IN2=10
* @note The quantization scale of the output must be passed at compile time using -DSCALE_OUT, i.e. -DSCALE_OUT=10
* @note To perform saturating operation -DSATURATE has to be passed to the compiler otherwise wrapping policy will be used.
* @note The element-wise operation to be executed has to be passed at compile time using -DOP (e.g., -DOP=ADD)
* @note For QSYMM16 operations OFFSET_IN1, OFFSET_IN2 and OFFSET_OUT must be set to zero
* @note The data type must be passed at compile time using -DDATA_TYPE, i.e. -DDATA_TYPE=uchar
*
* @param[in] in1_ptr Pointer to the source tensor. Supported data types: QASYMM8/QSYMM16
* @param[in] in1_stride_x Stride of the source tensor in X dimension (in bytes)
* @param[in] in1_step_x in1_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] in1_stride_y Stride of the source tensor in Y dimension (in bytes)
* @param[in] in1_step_y in1_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] in1_stride_z Stride of the source tensor in Z dimension (in bytes)
* @param[in] in1_step_z in1_stride_z * number of elements along Z processed per workitem(in bytes)
* @param[in] in1_offset_first_element_in_bytes The offset of the first element in the source tensor
* @param[in] in2_ptr Pointer to the source tensor. Supported data types: same as @p in1_ptr
* @param[in] in2_stride_x Stride of the source tensor in X dimension (in bytes)
* @param[in] in2_step_x in2_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] in2_stride_y Stride of the source tensor in Y dimension (in bytes)
* @param[in] in2_step_y in2_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] in2_stride_z Stride of the source tensor in Z dimension (in bytes)
* @param[in] in2_step_z in2_stride_z * number of elements along Z processed per workitem(in bytes)
* @param[in] in2_offset_first_element_in_bytes The offset of the first element in the source tensor
* @param[out] out_ptr Pointer to the destination tensor. Supported data types: same as @p in1_ptr
* @param[in] out_stride_x Stride of the destination tensor in X dimension (in bytes)
* @param[in] out_step_x out_stride_x * number of elements along X processed per workitem(in bytes)
* @param[in] out_stride_y Stride of the destination tensor in Y dimension (in bytes)
* @param[in] out_step_y out_stride_y * number of elements along Y processed per workitem(in bytes)
* @param[in] out_stride_z Stride of the source tensor in Z dimension (in bytes)
* @param[in] out_step_z out_stride_z * number of elements along Z processed per workitem(in bytes)
* @param[in] out_offset_first_element_in_bytes The offset of the first element in the destination tensor
*/
__kernel void OP_FUN_NAME(OP)(
TENSOR3D_DECLARATION(in1),
TENSOR3D_DECLARATION(in2)
#if !defined(IN_PLACE)
,
TENSOR3D_DECLARATION(out)
#endif // !defined(IN_PLACE)
)
{
#if VEC_SIZE_IN1 == 1
uint in1_x_offs = 0;
#else // VEC_SIZE_IN1 == 1
uint in1_x_offs = max((int)(get_global_id(0) * VEC_SIZE_IN1 - (VEC_SIZE_IN1 - VEC_SIZE_LEFTOVER) % VEC_SIZE_IN1), 0);
#endif // VEC_SIZE_IN1 == 1
#if VEC_SIZE_IN2 == 1
uint in2_x_offs = 0;
#else // VEC_SIZE_IN2 == 1
uint in2_x_offs = max((int)(get_global_id(0) * VEC_SIZE_IN2 - (VEC_SIZE_IN2 - VEC_SIZE_LEFTOVER) % VEC_SIZE_IN2), 0);
#endif // VEC_SIZE_IN2 == 1
#if !defined(IN_PLACE)
uint out_x_offs = max((int)(get_global_id(0) * VEC_SIZE_OUT - (VEC_SIZE_OUT - VEC_SIZE_LEFTOVER) % VEC_SIZE_OUT), 0);
#endif // !defined(IN_PLACE)
// Get pixels pointer
__global uchar *in1_addr = in1_ptr + in1_offset_first_element_in_bytes + in1_x_offs * sizeof(DATA_TYPE) + get_global_id(1) * in1_step_y + get_global_id(2) * in1_step_z;
__global uchar *in2_addr = in2_ptr + in2_offset_first_element_in_bytes + in2_x_offs * sizeof(DATA_TYPE) + get_global_id(1) * in2_step_y + get_global_id(2) * in2_step_z;
__global uchar *
#if !defined(IN_PLACE)
out_addr = out_ptr + out_offset_first_element_in_bytes + out_x_offs * sizeof(DATA_TYPE) + get_global_id(1) * out_step_y + get_global_id(2) * out_step_z;
#else // !defined(IN_PLACE)
#if defined(SRC1_IN_PLACE)
out_addr = in1_addr;
#else //defined(SRC1_IN_PLACE)
out_addr = in2_addr;
#endif //defined(SRC1_IN_PLACE)
#endif // !defined(IN_PLACE)
VEC_INT in_a = CONVERT((VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE_OUT))(VLOAD(VEC_SIZE_IN1)(0, (__global DATA_TYPE *)in1_addr)), VEC_INT);
VEC_INT in_b = CONVERT((VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE_OUT))(VLOAD(VEC_SIZE_IN2)(0, (__global DATA_TYPE *)in2_addr)), VEC_INT);
in_a = SUB(in_a, (VEC_INT)((int)OFFSET_IN1));
in_b = SUB(in_b, (VEC_INT)((int)OFFSET_IN2));
const VEC_FLOAT in1f32 = CONVERT(in_a, VEC_FLOAT) * (VEC_FLOAT)((float)SCALE_IN1);
const VEC_FLOAT in2f32 = CONVERT(in_b, VEC_FLOAT) * (VEC_FLOAT)((float)SCALE_IN2);
const VEC_FLOAT qresf32 = OP(in1f32, in2f32) / ((VEC_FLOAT)(float)SCALE_OUT) + ((VEC_FLOAT)((float)OFFSET_OUT));
const VEC_TYPE res0 = CONVERT_SAT(CONVERT_DOWN(qresf32, VEC_INT), VEC_TYPE);
// Store result
STORE_VECTOR_SELECT(res, DATA_TYPE, out_addr, VEC_SIZE_OUT, VEC_SIZE_LEFTOVER, VEC_SIZE_LEFTOVER != 0 && get_global_id(0) == 0)
}
#endif /* defined(OP) && defined(VEC_SIZE_IN1) && defined(VEC_SIZE_IN2) && defined(VEC_SIZE_OUT) && defined(OFFSET_IN1) && defined(OFFSET_IN2) && defined(OFFSET_OUT) && defined(SCALE_IN1) && defined(SCALE_IN2) && defined(SCALE_OUT) && defined(DATA_TYPE) */
|
