• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  * Copyright (c) 2018-2020 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 "src/core/CL/kernels/CLReverseKernel.h"
25 
26 #include "arm_compute/core/CL/CLHelpers.h"
27 #include "arm_compute/core/CL/CLKernelLibrary.h"
28 #include "arm_compute/core/CL/ICLTensor.h"
29 #include "arm_compute/core/Helpers.h"
30 #include "arm_compute/core/TensorInfo.h"
31 #include "src/core/CL/CLValidate.h"
32 #include "src/core/helpers/AutoConfiguration.h"
33 #include "src/core/helpers/WindowHelpers.h"
34 #include "support/StringSupport.h"
35 
36 namespace arm_compute
37 {
38 namespace
39 {
validate_arguments(const ITensorInfo * input,const ITensorInfo * output,const ITensorInfo * axis)40 Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, const ITensorInfo *axis)
41 {
42     ARM_COMPUTE_ERROR_ON_NULLPTR(input, output, axis);
43     ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input);
44     ARM_COMPUTE_RETURN_ERROR_ON(input->data_type() == DataType::UNKNOWN);
45     ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(axis, 1, DataType::U32);
46     ARM_COMPUTE_RETURN_ERROR_ON_MSG(axis->num_dimensions() > 1, "Axis must be a 1D tensor");
47     ARM_COMPUTE_RETURN_ERROR_ON_MSG(axis->dimension(0) > 4, "Only up to 4 dimensions can be reversed");
48 
49     // Checks performed when output is configured
50     if(output->total_size() != 0)
51     {
52         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input, output);
53         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
54         ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(input, output);
55     }
56 
57     return Status{};
58 }
59 } // namespace
60 
CLReverseKernel()61 CLReverseKernel::CLReverseKernel()
62     : _input(nullptr), _output(nullptr), _axis(nullptr)
63 {
64 }
65 
configure(const ICLTensor * input,ICLTensor * output,const ICLTensor * axis)66 void CLReverseKernel::configure(const ICLTensor *input, ICLTensor *output, const ICLTensor *axis)
67 {
68     configure(CLKernelLibrary::get().get_compile_context(), input, output, axis);
69 }
70 
configure(const CLCompileContext & compile_context,const ICLTensor * input,ICLTensor * output,const ICLTensor * axis)71 void CLReverseKernel::configure(const CLCompileContext &compile_context, const ICLTensor *input, ICLTensor *output, const ICLTensor *axis)
72 {
73     ARM_COMPUTE_ERROR_ON_NULLPTR(input, output, axis);
74 
75     _input  = input;
76     _output = output;
77     _axis   = axis;
78 
79     // Output tensor auto initialization if not yet initialized
80     auto_init_if_empty(*output->info(), *input->info()->clone());
81 
82     ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), axis->info()));
83 
84     // Set kernel build options
85     CLBuildOptions build_opts;
86     build_opts.add_option("-DNUM_REVERSE_DIMS=" + support::cpp11::to_string(axis->info()->dimension(0)));
87     build_opts.add_option("-DDATA_TYPE=" + get_cl_unsigned_type_from_element_size(input->info()->element_size()));
88 
89     // Create kernel
90     _kernel = create_kernel(compile_context, "reverse", build_opts.options());
91 
92     // Set static kernel arguments
93     unsigned int idx = 2 * num_arguments_per_4D_tensor() + num_arguments_per_1D_tensor();
94     add_argument<cl_uint>(idx, input->info()->dimension(0));
95     add_argument<cl_uint>(idx, input->info()->dimension(1));
96     add_argument<cl_uint>(idx, input->info()->dimension(2));
97     add_argument<cl_uint>(idx, input->info()->dimension(3));
98 
99     // Configure kernel window
100     Window win = calculate_max_window(*output->info(), Steps());
101     ICLKernel::configure_internal(win);
102 
103     // Set config_id for enabling LWS tuning
104     _config_id += "reverse_";
105     _config_id += lower_string(string_from_data_type(input->info()->data_type()));
106     _config_id += "_";
107     _config_id += support::cpp11::to_string(input->info()->dimension(0));
108     _config_id += "_";
109     _config_id += support::cpp11::to_string(input->info()->dimension(1));
110     _config_id += "_";
111     _config_id += support::cpp11::to_string(input->info()->dimension(2));
112 }
113 
validate(const ITensorInfo * input,const ITensorInfo * output,const ITensorInfo * axis)114 Status CLReverseKernel::validate(const ITensorInfo *input, const ITensorInfo *output, const ITensorInfo *axis)
115 {
116     ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, axis));
117     return Status{};
118 }
119 
run(const Window & window,cl::CommandQueue & queue)120 void CLReverseKernel::run(const Window &window, cl::CommandQueue &queue)
121 {
122     ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
123     ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);
124 
125     Window collapsed  = window.collapse(ICLKernel::window(), Window::DimZ);
126     Window slice      = collapsed.first_slice_window_4D();
127     Window axis_slice = collapsed.first_slice_window_1D();
128 
129     do
130     {
131         unsigned int idx = 0;
132         add_4D_tensor_argument(idx, _input, slice);
133         add_1D_tensor_argument(idx, _axis, axis_slice);
134         add_4D_tensor_argument(idx, _output, slice);
135         enqueue(queue, *this, slice, lws_hint());
136     }
137     while(collapsed.slide_window_slice_4D(slice));
138 }
139 } // namespace arm_compute
140