xref: /aosp_15_r20/external/ComputeLibrary/src/runtime/CL/functions/CLUnstack.cpp (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 "arm_compute/runtime/CL/functions/CLUnstack.h"
25 
26 #include "arm_compute/core/CL/ICLTensor.h"
27 #include "arm_compute/core/Error.h"
28 #include "arm_compute/core/TensorInfo.h"
29 #include "arm_compute/core/Types.h"
30 #include "arm_compute/core/utils/misc/ShapeCalculator.h"
31 
32 #include "src/common/utils/Log.h"
33 
34 namespace arm_compute
35 {
36 namespace
37 {
wrap_axis(int axis,const ITensorInfo * const tensor)38 inline unsigned int wrap_axis(int axis, const ITensorInfo *const tensor)
39 {
40     return wrap_around(axis, static_cast<int>(tensor->num_dimensions()));
41 }
42 
setup_slice_coordinates_and_mask(Coordinates & slice_start,int32_t & slice_end_mask,const unsigned int input_num_dimensions)43 inline void setup_slice_coordinates_and_mask(Coordinates &slice_start, int32_t &slice_end_mask, const unsigned int input_num_dimensions)
44 {
45     // Setups up coordinates to slice the input tensor: start coordinates to all 0s and the unstacking axis of both Start/End to slice just one 2d tensor at a time.
46     Coordinates slice_end;
47     slice_start.set_num_dimensions(input_num_dimensions);
48     slice_end.set_num_dimensions(input_num_dimensions);
49     for(size_t k = 0; k < input_num_dimensions; ++k)
50     {
51         slice_start.set(k, 0);
52         slice_end.set(k, -1);
53     }
54     slice_end_mask = arm_compute::helpers::tensor_transform::construct_slice_end_mask(slice_end);
55 }
56 } // namespace
57 
CLUnstack()58 CLUnstack::CLUnstack() // NOLINT
59     : _num_slices(0),
60       _strided_slice_vector()
61 {
62 }
63 
configure(const ICLTensor * input,const std::vector<ICLTensor * > & output_vector,int axis)64 void CLUnstack::configure(const ICLTensor *input, const std::vector<ICLTensor *> &output_vector, int axis)
65 {
66     configure(CLKernelLibrary::get().get_compile_context(), input, output_vector, axis);
67 }
68 
configure(const CLCompileContext & compile_context,const ICLTensor * input,const std::vector<ICLTensor * > & output_vector,int axis)69 void CLUnstack::configure(const CLCompileContext &compile_context, const ICLTensor *input, const std::vector<ICLTensor *> &output_vector, int axis)
70 {
71     ARM_COMPUTE_LOG_PARAMS(input, output_vector, axis);
72     std::vector<ITensorInfo *> outputs_vector_info(output_vector.size());
73     std::transform(output_vector.begin(), output_vector.end(), outputs_vector_info.begin(), [](ICLTensor * t)
74     {
75         ARM_COMPUTE_ERROR_ON_NULLPTR(t);
76         return t->info();
77     });
78 
79     ARM_COMPUTE_ERROR_ON_NULLPTR(input);
80     ARM_COMPUTE_ERROR_THROW_ON(CLUnstack::validate(input->info(), outputs_vector_info, axis));
81 
82     // Wrap around negative values
83     const unsigned int axis_u = wrap_axis(axis, input->info());
84     _num_slices               = std::min(outputs_vector_info.size(), input->info()->dimension(axis_u));
85     _strided_slice_vector.resize(_num_slices);
86 
87     Coordinates slice_start;
88     int32_t     slice_end_mask;
89     setup_slice_coordinates_and_mask(slice_start, slice_end_mask, input->info()->tensor_shape().num_dimensions());
90     for(unsigned int slice = 0; slice < _num_slices; ++slice)
91     {
92         // Adjusts start and end coordinates to take a 2D slice at a time
93         slice_start.set(axis_u, slice);
94         _strided_slice_vector[slice].configure(compile_context, input, output_vector[slice], slice_start, Coordinates(), BiStrides(), 0, slice_end_mask, (1 << axis_u));
95     }
96 }
97 
validate(const ITensorInfo * input,const std::vector<ITensorInfo * > & output_vector,int axis)98 Status CLUnstack::validate(const ITensorInfo *input, const std::vector<ITensorInfo *> &output_vector, int axis)
99 {
100     ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input);
101     ARM_COMPUTE_RETURN_ERROR_ON(output_vector.empty());
102     ARM_COMPUTE_RETURN_ERROR_ON(axis < (-static_cast<int>(input->tensor_shape().num_dimensions())));
103     ARM_COMPUTE_RETURN_ERROR_ON(axis >= static_cast<int>(input->tensor_shape().num_dimensions()));
104     const unsigned int num_slices = std::min(output_vector.size(), input->dimension(wrap_axis(axis, input)));
105     ARM_COMPUTE_RETURN_ERROR_ON(num_slices > input->dimension(wrap_axis(axis, input)));
106     ARM_COMPUTE_RETURN_ERROR_ON(num_slices > output_vector.size());
107     Coordinates slice_start;
108     int32_t     slice_end_mask;
109     for(size_t k = 0; k < num_slices; ++k)
110     {
111         slice_start.set(wrap_axis(axis, input), k);
112         setup_slice_coordinates_and_mask(slice_start, slice_end_mask, input->tensor_shape().num_dimensions());
113         ARM_COMPUTE_RETURN_ON_ERROR(CLStridedSlice::validate(input, output_vector[k], slice_start, Coordinates(), BiStrides(), 0, slice_end_mask, (1 << wrap_axis(axis, input))));
114     }
115     return Status{};
116 }
117 
run()118 void CLUnstack::run()
119 {
120     for(unsigned i = 0; i < _num_slices; ++i)
121     {
122         _strided_slice_vector[i].run();
123     }
124 }
125 
126 } // namespace arm_compute
127