xref: /aosp_15_r20/external/ComputeLibrary/src/core/CL/cl_kernels/common/col2im.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#if defined(DATA_TYPE) && defined(WIDTH_OUTPUT) && defined(ELEMENT_SIZE) && defined(WIDTH_INPUT) && defined(NUM_GROUPS)
27
28#if ELEMENT_SIZE == 1
29#define COND_DATA_TYPE char
30#elif ELEMENT_SIZE == 2
31#define COND_DATA_TYPE short
32#elif ELEMENT_SIZE == 4
33#define COND_DATA_TYPE int
34#else // ELEMENT_SIZE
35#error "Element size not support"
36#endif // ELEMENT_SIZE
37
38/** This kernel performs a reshaping of the output of the convolution layer
39 *
40 * @note The data type must be passed at compile time using -DDATA_TYPE: e.g. -DDATA_TYPE=float
41 * @note The width of the input tensor must be passed at compile time using -DWIDTH_INPUT: e.g. -DWIDTH_INPUT=320
42 * @note The width of the output tensor must be passed at compile time using -DWIDTH_OUTPUT: e.g. -DWIDTH_OUTPUT=600
43 * @note The element size must be passed at compile time using -DELEMENT_SIZE: e.g. -DELEMENT_SIZE=4
44 * @note The number of groups must be passed at compile time using  -DNUM_GROUPS: e.g. -DNUM_GROUPS=4
45 *
46 * @param[in]  src_ptr                           Pointer to the source tensor. Supported data types: QASYMM8/QASYMM8_SIGNED/F16/F32
47 * @param[in]  src_stride_x                      Stride of the source tensor in X dimension (in bytes)
48 * @param[in]  src_step_x                        src_stride_x * number of elements along X processed per workitem(in bytes)
49 * @param[in]  src_stride_y                      Stride of the source tensor in Y dimension (in bytes)
50 * @param[in]  src_step_y                        src_stride_y * number of elements along Y processed per workitem(in bytes)
51 * @param[in]  src_stride_z                      Stride of the source tensor in Z dimension (in bytes)
52 * @param[in]  src_step_z                        src_stride_z * number of elements along Z processed per workitem(in bytes)
53 * @param[in]  src_offset_first_element_in_bytes The offset of the first element in the source tensor
54 * @param[out] dst_ptr                           Pointer to the destination tensor. Supported data types: same as @p src_ptr
55 * @param[in]  dst_stride_x                      Stride of the destination tensor in X dimension (in bytes)
56 * @param[in]  dst_step_x                        dst_stride_x * number of elements along X processed per workitem(in bytes)
57 * @param[in]  dst_stride_y                      Stride of the destination tensor in Y dimension (in bytes)
58 * @param[in]  dst_step_y                        dst_stride_y * number of elements along Y processed per workitem(in bytes)
59 * @param[in]  dst_stride_z                      Stride of the destination tensor in Z dimension (in bytes)
60 * @param[in]  dst_step_z                        dst_stride_z * number of elements along Z processed per workitem(in bytes)
61 * @param[in]  dst_stride_w                      Stride of the destination tensor in W dimension (in bytes)
62 * @param[in]  dst_step_w                        dst_stride_w * number of elements along W processed per workitem(in bytes)
63 * @param[in]  dst_offset_first_element_in_bytes The offset of the first element in the destination tensor
64 */
65__kernel void col2im(
66    TENSOR3D_DECLARATION(src),
67    TENSOR4D_DECLARATION(dst))
68{
69    Tensor3D src = CONVERT_TO_TENSOR3D_STRUCT(src);
70    Tensor4D dst = CONVERT_TO_TENSOR4D_STRUCT_NO_STEP(dst, 0);
71
72    const uint xd = get_global_id(1) % WIDTH_OUTPUT; // x coordinate of the destination tensor
73    const uint yd = get_global_id(1) / WIDTH_OUTPUT; // y coordinate of the destination tensor
74
75    VEC_DATA_TYPE(DATA_TYPE, 8)
76    data = vload8(0, (__global DATA_TYPE *)src.ptr);
77
78    uint  x         = get_global_id(0) * 8;
79    uint8 x_clamped = x + (uint8)(0, 1, 2, 3, 4, 5, 6, 7);
80
81    VEC_DATA_TYPE(COND_DATA_TYPE, 8)
82    cond0 = CONVERT((x_clamped < WIDTH_INPUT), VEC_DATA_TYPE(COND_DATA_TYPE, 8));
83
84    // Clamp x if out-of-bounds
85    x_clamped = select((uint8)x, x_clamped, convert_int8(cond0));
86
87    // If out-of-bound, overwrite with the first element
88    data = select((VEC_DATA_TYPE(DATA_TYPE, 8))data.s0, data, cond0);
89
90#if NUM_GROUPS > 1
91    // Compute output offset (batches on 4th dimension)
92    int idx = yd * dst_stride_y + xd * dst_stride_x + (get_global_id(2) / NUM_GROUPS) * dst.stride_w;
93
94    const uint group = get_global_id(2) % NUM_GROUPS; // group ID
95    x_clamped += group * WIDTH_INPUT;
96#else  /* defined(NUM_GROUPS > 1 ) */
97    // Compute output offset (batches on 3rd dimension)
98    int idx = yd * dst.stride_y + xd * dst.stride_x + get_global_id(2) * dst.stride_w;
99#endif /* NUM_GROUPS > 1 */
100
101    // Store value
102    *((__global DATA_TYPE *)(dst.ptr + idx + x_clamped.s0 * dst.stride_z)) = data.s0;
103    *((__global DATA_TYPE *)(dst.ptr + idx + x_clamped.s1 * dst.stride_z)) = data.s1;
104    *((__global DATA_TYPE *)(dst.ptr + idx + x_clamped.s2 * dst.stride_z)) = data.s2;
105    *((__global DATA_TYPE *)(dst.ptr + idx + x_clamped.s3 * dst.stride_z)) = data.s3;
106    *((__global DATA_TYPE *)(dst.ptr + idx + x_clamped.s4 * dst.stride_z)) = data.s4;
107    *((__global DATA_TYPE *)(dst.ptr + idx + x_clamped.s5 * dst.stride_z)) = data.s5;
108    *((__global DATA_TYPE *)(dst.ptr + idx + x_clamped.s6 * dst.stride_z)) = data.s6;
109    *((__global DATA_TYPE *)(dst.ptr + idx + x_clamped.s7 * dst.stride_z)) = data.s7;
110}
111#endif // defined(DATA_TYPE) && defined(WIDTH_OUTPUT) && defined(ELEMENT_SIZE) && defined(WIDTH_INPUT) && defined(NUM_GROUPS)
112