xref: /aosp_15_r20/external/ComputeLibrary/src/core/CL/cl_kernels/common/tile.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#if defined(DATA_TYPE) && defined(SRC_WIDTH) && defined(SRC_HEIGHT) && defined(SRC_DEPTH) && defined(DST_DEPTH)
26/** Perform a floor operation on an input tensor.
27 *
28 * @attention Data type can be passed using the -DDATA_TYPE compile flag, e.g. -DDATA_TYPE=float
29 * @attention Vector size should be given as a preprocessor argument using -DVEC_SIZE=size. e.g. -DVEC_SIZE=16
30 * @note Can only take floating point data types.
31 *
32 * @param[in]  input_ptr                            Pointer to the source image. Supported data types: All
33 * @param[in]  input_stride_x                       Stride of the source image in X dimension (in bytes)
34 * @param[in]  input_step_x                         input_stride_x * number of elements along X processed per workitem(in bytes)
35 * @param[in]  input_stride_y                       Stride of the source image in Y dimension (in bytes)
36 * @param[in]  input_step_y                         input_stride_y * number of elements along Y processed per workitem(in bytes)
37 * @param[in]  input_stride_z                       Stride of the source tensor in Z dimension (in bytes)
38 * @param[in]  input_step_z                         input_stride_z * number of elements along Z processed per workitem(in bytes)
39 * @param[in]  input_offset_first_element_in_bytes  The offset of the first element in the source image
40 * @param[out] output_ptr                           Pointer to the destination image. Supported data types: same as @p input_ptr
41 * @param[in]  output_stride_x                      Stride of the destination image in X dimension (in bytes)
42 * @param[in]  output_step_x                        output_stride_x * number of elements along X processed per workitem(in bytes)
43 * @param[in]  output_stride_y                      Stride of the destination image in Y dimension (in bytes)
44 * @param[in]  output_step_y                        output_stride_y * number of elements along Y processed per workitem(in bytes)
45 * @param[in]  output_stride_z                      Stride of the source tensor in Z dimension (in bytes)
46 * @param[in]  output_step_z                        output_stride_z * number of elements along Z processed per workitem(in bytes)
47 * @param[in]  output_offset_first_element_in_bytes The offset of the first element in the destination image
48 */
49__kernel void tile(
50    TENSOR4D_DECLARATION(input),
51    TENSOR4D_DECLARATION(output))
52{
53    Tensor4D output = CONVERT_TO_TENSOR4D_STRUCT(output, DST_DEPTH);
54    Tensor4D input  = CONVERT_TO_TENSOR4D_STRUCT_NO_STEP(input, SRC_DEPTH);
55
56    // For all coordinates but x, each tile copies from the input
57    const int y     = get_global_id(1);
58    const int z     = get_global_id(2) % DST_DEPTH;
59    const int batch = get_global_id(2) / DST_DEPTH;
60
61#if defined(VEC_SIZE) && defined(OFFSET)
62    // If we are loading/storing multiple elements at time, we need to
63    // not exceed the input boundaries. The last threads need to backtrack
64    // of OFFSET elements. Those elements cumulates for previous tiles
65    const int id = (int)(get_global_id(0));
66    int       x  = id * VEC_SIZE;
67
68    // Shift x based on the previous offsets
69    const int tile_number = x / SRC_WIDTH;
70    x -= (tile_number) * OFFSET;
71    int x_input = x % SRC_WIDTH;
72
73    // Shift x based on being the last tile
74    const int last_tile = (int)(x_input + VEC_SIZE > SRC_WIDTH);
75    x -= last_tile * OFFSET;
76    x_input = x % SRC_WIDTH;
77    output.ptr -= (tile_number + last_tile) * OFFSET * output_stride_x;
78
79    // Update the input pointer
80    input.ptr = tensor4D_offset(&input, x_input, y % SRC_HEIGHT, z % SRC_DEPTH, batch % SRC_BATCHES);
81
82    // Copy the data
83    VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE)
84    data = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)input.ptr);
85
86    VSTORE(VEC_SIZE)
87    (data, 0, (__global DATA_TYPE *)output.ptr);
88#else  // !defined(VEC_SIZE) || !defined(OFFSET)
89    const int x = get_global_id(0);
90
91    // Update the input pointer
92    input.ptr = tensor4D_offset(&input, x % SRC_WIDTH, y % SRC_HEIGHT, z % SRC_DEPTH, batch % SRC_BATCHES);
93
94    *((__global DATA_TYPE *)(output.ptr)) = *((__global DATA_TYPE *)(input.ptr));
95#endif // defined(VEC_SIZE) && defined(OFFSET)
96}
97#endif // defined(DATA_TYPE) && defined(SRC_WIDTH) && defined(SRC_HEIGHT) && defined(SRC_DEPTH) && defined(DST_DEPTH)
98