2 * Copyright (c) 2017-2018 ARM Limited.
4 * SPDX-License-Identifier: MIT
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:
13 * The above copyright notice and this permission notice shall be included in all
14 * copies or substantial portions of the Software.
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
26 #if defined(FIXED_POINT_POSITION)
27 #include "fixed_point.h"
28 #endif // FIXED_POINT_POSITION
30 #if defined(DATA_TYPE) && defined(WIDTH_OUTPUT) && defined(ELEMENT_SIZE) && defined(WIDTH_INPUT)
31 #if !defined(FIXED_POINT_POSITION)
34 #define COND_DATA_TYPE char
35 #elif ELEMENT_SIZE == 2
36 #define COND_DATA_TYPE short
37 #elif ELEMENT_SIZE == 4
38 #define COND_DATA_TYPE int
40 #error "Element size not support"
41 #endif // ELEMENT_SIZE
43 /** This kernel performs a reshaping of the output of the convolution layer
45 * @note The data type must be passed at compile time using -DDATA_TYPE: e.g. -DDATA_TYPE=float
46 * @note The width of the input tensor must be passed at compile time using -DWIDTH_INPUT: e.g. -DWIDTH_INPUT=320
47 * @note The width of the output tensor must be passed at compile time using -DWIDTH_OUTPUT: e.g. -DWIDTH_OUTPUT=600
48 * @note The element size must be passed at compile time using -DELEMENT_SIZE: e.g. -DELEMENT_SIZE=4
50 * @param[in] src_ptr Pointer to the source tensor. Supported data types: QASYMM8/F16/F32
51 * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes)
52 * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
53 * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes)
54 * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
55 * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes)
56 * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes)
57 * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor
58 * @param[out] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr
59 * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes)
60 * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes)
61 * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes)
62 * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes)
63 * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes)
64 * @param[in] dst_step_z dst_stride_z * number of elements along Z processed per workitem(in bytes)
65 * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor
66 * @param[in] dst_stride_w Stride of the destination tensor in W dimension (in bytes)
69 TENSOR3D_DECLARATION(src),
70 TENSOR3D_DECLARATION(dst),
73 Tensor3D src = CONVERT_TO_TENSOR3D_STRUCT(src);
75 VEC_DATA_TYPE(DATA_TYPE, 8)
76 data = vload8(0, (__global DATA_TYPE *)src.ptr);
78 uint x = get_global_id(0) * 8;
79 uint8 x_clamped = x + (uint8)(0, 1, 2, 3, 4, 5, 6, 7);
81 VEC_DATA_TYPE(COND_DATA_TYPE, 8)
82 cond0 = CONVERT((x_clamped < WIDTH_INPUT), VEC_DATA_TYPE(COND_DATA_TYPE, 8));
84 // Clamp x if out-of-bounds
85 x_clamped = select((uint8)x, x_clamped, convert_int8(cond0));
87 // If out-of-bound, overwrite with the first element
88 data = select((VEC_DATA_TYPE(DATA_TYPE, 8))data.s0, data, cond0);
90 __global uchar *output_ptr = dst_ptr + dst_offset_first_element_in_bytes;
92 // Compute output offset
93 int idx = (get_global_id(1) / WIDTH_OUTPUT) * dst_stride_y + (get_global_id(1) % WIDTH_OUTPUT) * dst_stride_x + get_global_id(2) * dst_stride_w;
96 *((__global DATA_TYPE *)(output_ptr + idx + x_clamped.s0 * dst_stride_z)) = data.s0;
97 *((__global DATA_TYPE *)(output_ptr + idx + x_clamped.s1 * dst_stride_z)) = data.s1;
98 *((__global DATA_TYPE *)(output_ptr + idx + x_clamped.s2 * dst_stride_z)) = data.s2;
99 *((__global DATA_TYPE *)(output_ptr + idx + x_clamped.s3 * dst_stride_z)) = data.s3;
100 *((__global DATA_TYPE *)(output_ptr + idx + x_clamped.s4 * dst_stride_z)) = data.s4;
101 *((__global DATA_TYPE *)(output_ptr + idx + x_clamped.s5 * dst_stride_z)) = data.s5;
102 *((__global DATA_TYPE *)(output_ptr + idx + x_clamped.s6 * dst_stride_z)) = data.s6;
103 *((__global DATA_TYPE *)(output_ptr + idx + x_clamped.s7 * dst_stride_z)) = data.s7;
105 #else // !defined(FIXED_POINT_POSITION)
106 /** This kernel performs a reshaping of the output of the convolution layer.
108 * @note The data type must be passed at compile time using -DDATA_TYPE: e.g. -DDATA_TYPE=qs8
109 * @note The width of the output tensor must be passed at compile time using -DWIDTH_OUTPUT: e.g. -DWIDTH_OUTPUT=320
111 * @param[in] src_ptr Pointer to the source tensor. Supported data types: QS8/QS16
112 * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes)
113 * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes)
114 * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes)
115 * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes)
116 * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes)
117 * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes)
118 * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor
119 * @param[out] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr
120 * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes)
121 * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes)
122 * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes)
123 * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes)
124 * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes)
125 * @param[in] dst_step_z dst_stride_z * number of elements along Z processed per workitem(in bytes)
126 * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor
127 * @param[in] dst_stride_w Stride of the destination tensor in W dimension (in bytes)
129 __kernel void col2im(
130 TENSOR3D_DECLARATION(src),
131 TENSOR3D_DECLARATION(dst),
134 Tensor3D src = CONVERT_TO_TENSOR3D_STRUCT(src);
135 Tensor3D dst = CONVERT_TO_TENSOR3D_STRUCT_NO_STEP(dst);
137 // Compute output offset
138 int idx = get_global_id(0) * dst.stride_z + (get_global_id(1) / WIDTH_OUTPUT) * dst_stride_y + (get_global_id(1) % WIDTH_OUTPUT) * dst_stride_x + get_global_id(2) * dst_stride_w;
141 *((__global DATA_TYPE *)(dst.ptr + idx)) = *((__global DATA_TYPE *)(src.ptr));
143 #endif // !defined(FIXED_POINT_POSITION)
144 #endif // defined(DATA_TYPE) && defined(WIDTH_OUTPUT) && defined(ELEMENT_SIZE) && defined(WIDTH_INPUT)