xref: /aosp_15_r20/external/ComputeLibrary/src/core/CL/cl_kernels/common/elementwise_operation.cl (revision c217d954acce2dbc11938adb493fc0abd69584f3)
1/*
2 * Copyright (c) 2018-2021 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#if defined(OP) && defined(VEC_SIZE_IN1) && defined(VEC_SIZE_IN2) && defined(VEC_SIZE_OUT) && defined(DATA_TYPE)
27
28/** List of all the operations supported by this kernel.
29 * @note ADD and SUB operations, when executed on integers, support saturation */
30#ifdef SATURATE
31#define ADD(x, y) add_sat((x), (y))
32#define SUB(x, y) sub_sat((x), (y))
33#else /* SATURATE */
34#define ADD(x, y) (x) + (y)
35#define SUB(x, y) (x) - (y)
36#endif /* SATURATE */
37
38#define MAX(x, y) max(x, y)
39#define MIN(x, y) min(x, y)
40#define SQUARED_DIFF(x, y) (x - y) * (x - y)
41#define POWER(x, y) pow(x, y)
42
43#if VEC_SIZE_OUT == 1
44#define PRELU(x, y) (x > 0 ? x : x * y)
45#else // VEC_SIZE_OUT == 1
46#define PRELU(x, y) (select(y * x, x, CONVERT((x > (DATA_TYPE)0), SELECT_VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE_OUT))))
47#endif // VEC_SIZE_OUT == 1
48
49#if defined(S32)
50#define DIV(x, y) CONVERT(floor(CONVERT(x, VEC_DATA_TYPE(float, VEC_SIZE_OUT)) / CONVERT(y, VEC_DATA_TYPE(float, VEC_SIZE_OUT))), VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE_OUT));
51#else /* S32 */
52#define DIV(x, y) (x / y)
53#endif /* S32 */
54
55#define AND(x, y) (CONVERT((x && y), VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE_OUT)) & ((VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE_OUT))1))
56#define OR(x, y) (CONVERT((x || y), VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE_OUT)) & ((VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE_OUT))1))
57
58#define OP_FUN_NAME_STR(op) elementwise_operation_##op
59#define OP_FUN_NAME(op) OP_FUN_NAME_STR(op)
60
61#if defined(ACTIVATION_TYPE)
62#include "activation_float_helpers.h"
63#endif // defined(ACTIVATION_TYPE)
64
65/** This function executes an element-wise operation among two tensors.
66 *
67 * @note Vector sizes of inputs and output have to be passed at compile time using -DVEC_SIZE_IN1, -DVEC_SIZE_IN2, -DVEC_SIZE_OUT.
68 * @note Leftover vector size has to be passed at compile time using -DVEC_SIZE_LEFTOVER. e.g. -DVEC_SIZE_OUT=3. It is defined as the remainder between the input's first dimension and VEC_SIZE_OUT
69 * @note The input and output data_types need to be passed at compile time using -DDATA_TYPE: e.g. -DDATA_TYPE=uchar
70 * @note To perform saturating operation -DSATURATE has to be passed to the compiler otherwise wrapping policy will be used.
71 * @note The element-wise operation to be executed has to be passed at compile time using -DOP (e.g., -DOP=ADD)
72 *
73 * @param[in]  in1_ptr                           Pointer to the source tensor. Supported data types: U8/S16/F16/F32
74 * @param[in]  in1_stride_x                      Stride of the source tensor in X dimension (in bytes)
75 * @param[in]  in1_step_x                        in1_stride_x * number of elements along X processed per workitem(in bytes)
76 * @param[in]  in1_stride_y                      Stride of the source tensor in Y dimension (in bytes)
77 * @param[in]  in1_step_y                        in1_stride_y * number of elements along Y processed per workitem(in bytes)
78 * @param[in]  in1_stride_z                      Stride of the source tensor in Z dimension (in bytes)
79 * @param[in]  in1_step_z                        in1_stride_z * number of elements along Z processed per workitem(in bytes)
80 * @param[in]  in1_offset_first_element_in_bytes The offset of the first element in the source tensor
81 * @param[in]  in2_ptr                           Pointer to the source tensor. Supported data types: U8/S16/F16/F32
82 * @param[in]  in2_stride_x                      Stride of the source tensor in X dimension (in bytes)
83 * @param[in]  in2_step_x                        in2_stride_x * number of elements along X processed per workitem(in bytes)
84 * @param[in]  in2_stride_y                      Stride of the source tensor in Y dimension (in bytes)
85 * @param[in]  in2_step_y                        in2_stride_y * number of elements along Y processed per workitem(in bytes)
86 * @param[in]  in2_stride_z                      Stride of the source tensor in Z dimension (in bytes)
87 * @param[in]  in2_step_z                        in2_stride_z * number of elements along Z processed per workitem(in bytes)
88 * @param[in]  in2_offset_first_element_in_bytes The offset of the first element in the source tensor
89 * @param[out] out_ptr                           Pointer to the destination tensor. Supported data types: U8 (only if both inputs are U8), S16/F16/F32
90 * @param[in]  out_stride_x                      Stride of the destination tensor in X dimension (in bytes)
91 * @param[in]  out_step_x                        out_stride_x * number of elements along X processed per workitem(in bytes)
92 * @param[in]  out_stride_y                      Stride of the destination tensor in Y dimension (in bytes)
93 * @param[in]  out_step_y                        out_stride_y * number of elements along Y processed per workitem(in bytes)
94 * @param[in]  out_stride_z                      Stride of the source tensor in Z dimension (in bytes)
95 * @param[in]  out_step_z                        out_stride_z * number of elements along Z processed per workitem(in bytes)
96 * @param[in]  out_offset_first_element_in_bytes The offset of the first element in the destination tensor
97 */
98__kernel void OP_FUN_NAME(OP)(
99    TENSOR3D_DECLARATION(in1),
100    TENSOR3D_DECLARATION(in2)
101#if !defined(IN_PLACE)
102    ,
103    TENSOR3D_DECLARATION(out)
104#endif // !defined(IN_PLACE)
105)
106{
107#if VEC_SIZE_IN1 == 1
108    uint in1_x_offs = 0;
109#else  // VEC_SIZE_IN1 == 1
110    uint in1_x_offs = max((int)(get_global_id(0) * VEC_SIZE_IN1 - (VEC_SIZE_IN1 - VEC_SIZE_LEFTOVER) % VEC_SIZE_IN1), 0);
111#endif // VEC_SIZE_IN1 == 1
112#if VEC_SIZE_IN2 == 1
113    uint in2_x_offs = 0;
114#else  // VEC_SIZE_IN2 == 1
115    uint in2_x_offs = max((int)(get_global_id(0) * VEC_SIZE_IN2 - (VEC_SIZE_IN2 - VEC_SIZE_LEFTOVER) % VEC_SIZE_IN2), 0);
116#endif // VEC_SIZE_IN2 == 1
117#if !defined(IN_PLACE)
118    uint out_x_offs = max((int)(get_global_id(0) * VEC_SIZE_OUT - (VEC_SIZE_OUT - VEC_SIZE_LEFTOVER) % VEC_SIZE_OUT), 0);
119#endif // !defined(IN_PLACE)
120
121    // Get pixels pointer
122    __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;
123    __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;
124    __global        uchar *
125#if !defined(IN_PLACE)
126    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;
127#else // !defined(IN_PLACE)
128#if defined(SRC1_IN_PLACE)
129    out_addr    = in1_addr;
130#else  //defined(SRC1_IN_PLACE)
131    out_addr = in2_addr;
132#endif //defined(SRC1_IN_PLACE)
133#endif // !defined(IN_PLACE)
134
135    // Load values
136    VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE_OUT)
137    in_a = CONVERT((VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE_OUT))(VLOAD(VEC_SIZE_IN1)(0, (__global DATA_TYPE *)in1_addr)), VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE_OUT));
138    VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE_OUT)
139    in_b = CONVERT((VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE_OUT))(VLOAD(VEC_SIZE_IN2)(0, (__global DATA_TYPE *)in2_addr)), VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE_OUT));
140
141    // Calculate and store result
142    VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE_OUT)
143    res0 = OP(in_a, in_b);
144#if defined(ACTIVATION_TYPE)
145    res0 = ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, VEC_SIZE_OUT, res0, A_VAL, B_VAL);
146#endif // defined(ACTIVATION_TYPE)
147
148    STORE_VECTOR_SELECT(res, DATA_TYPE, out_addr, VEC_SIZE_OUT, VEC_SIZE_LEFTOVER, VEC_SIZE_LEFTOVER != 0 && get_global_id(0) == 0)
149}
150#endif /* defined(OP) && defined(VEC_SIZE_IN1) && defined(VEC_SIZE_IN2) && defined(VEC_SIZE_OUT) && defined(DATA_TYPE) */
151