blob: 9c2b68acc0e0c8b6c1661cb84832114d92f000f8 [file] [log] [blame]
Vidhya Sudhan Loganathan338595b2019-06-28 14:09:53 +01001/*
Michele Di Giorgiod9eaf612020-07-08 11:12:57 +01002 * Copyright (c) 2019-2020 Arm Limited.
Vidhya Sudhan Loganathan338595b2019-06-28 14:09:53 +01003 *
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/core/CL/kernels/CLBatchConcatenateLayerKernel.h"
25
26#include "arm_compute/core/CL/CLHelpers.h"
27#include "arm_compute/core/CL/CLKernelLibrary.h"
28#include "arm_compute/core/CL/CLValidate.h"
29#include "arm_compute/core/CL/ICLTensor.h"
Vidhya Sudhan Loganathan338595b2019-06-28 14:09:53 +010030#include "arm_compute/core/Helpers.h"
Vidhya Sudhan Loganathan338595b2019-06-28 14:09:53 +010031#include "arm_compute/core/TensorInfo.h"
32#include "arm_compute/core/Utils.h"
33#include "arm_compute/core/Window.h"
34
Matthew Bentham758b5ba2020-03-05 23:37:48 +000035#include "support/StringSupport.h"
Vidhya Sudhan Loganathan338595b2019-06-28 14:09:53 +010036
Michele Di Giorgiof6f78762020-07-06 11:27:21 +010037namespace arm_compute
38{
Vidhya Sudhan Loganathan338595b2019-06-28 14:09:53 +010039namespace
40{
41std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, unsigned int batch_offset, ITensorInfo *output)
42{
43 ARM_COMPUTE_UNUSED(batch_offset);
44
45 const unsigned int num_elems_processed_per_iteration = 16 / input->element_size();
46
47 // The window needs to be based on output, except for the batch size
48 Window win = calculate_max_window(*output, Steps(num_elems_processed_per_iteration));
49 // The total batch size is the concatenation of the batch size of the inputs
50 win.set(3, Window::Dimension(0, input->tensor_shape()[3], 1));
51
52 AccessWindowHorizontal input_access(input, 0, num_elems_processed_per_iteration);
53 AccessWindowHorizontal output_access(output, 0, num_elems_processed_per_iteration);
54 bool window_changed = update_window_and_padding(win, input_access, output_access);
55 output_access.set_valid_region(win, ValidRegion(Coordinates(), output->tensor_shape()));
56
57 Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{};
58 return std::make_pair(err, win);
59}
60Status validate_arguments(const ITensorInfo *input, unsigned int batch_offset, const ITensorInfo *output)
61{
62 ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output);
63 ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input);
Manuel Bottini8481d832019-12-10 15:28:40 +000064 ARM_COMPUTE_RETURN_ERROR_ON(input->data_type() == DataType::UNKNOWN);
Vidhya Sudhan Loganathan338595b2019-06-28 14:09:53 +010065 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
66
67 ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(Window::DimX) != output->dimension(Window::DimX));
68 ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(Window::DimY) != output->dimension(Window::DimY));
69 ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(Window::DimZ) != output->dimension(Window::DimZ));
70 ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(3) + batch_offset > output->dimension(3));
71 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(4, input, output);
72
73 return Status{};
74}
75} // namespace
76
77CLBatchConcatenateLayerKernel::CLBatchConcatenateLayerKernel()
78 : _input(nullptr), _output(nullptr), _batch_offset(0)
79{
80}
81
82void CLBatchConcatenateLayerKernel::configure(const ICLTensor *input, unsigned int batch_offset, ICLTensor *output)
83{
Manuel Bottini4c6bd512020-04-08 10:15:51 +010084 configure(CLKernelLibrary::get().get_compile_context(), input, batch_offset, output);
85}
86
Manuel Bottini256c0b92020-04-21 13:29:30 +010087void CLBatchConcatenateLayerKernel::configure(const CLCompileContext &compile_context, const ICLTensor *input, unsigned int batch_offset, ICLTensor *output)
Manuel Bottini4c6bd512020-04-08 10:15:51 +010088{
Vidhya Sudhan Loganathan338595b2019-06-28 14:09:53 +010089 ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
90 ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), batch_offset, output->info()));
91
92 _input = input;
93 _output = output;
94 _batch_offset = batch_offset;
95
96 const unsigned int num_elems_processed_per_iteration = 16 / input->info()->element_size();
97
98 // Add build options
99 CLBuildOptions build_opts;
Michele Di Giorgiof6f78762020-07-06 11:27:21 +0100100 build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type()));
Vidhya Sudhan Loganathan338595b2019-06-28 14:09:53 +0100101 build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(num_elems_processed_per_iteration));
102 if(is_data_type_quantized_asymmetric(input->info()->data_type()) && input->info()->quantization_info() != output->info()->quantization_info())
103 {
104 const UniformQuantizationInfo iq_info = input->info()->quantization_info().uniform();
105 const UniformQuantizationInfo oq_info = output->info()->quantization_info().uniform();
106
107 build_opts.add_option("-DOFFSET_IN1=" + float_to_string_with_full_precision(iq_info.offset));
108 build_opts.add_option("-DOFFSET_OUT=" + float_to_string_with_full_precision(oq_info.offset));
109 build_opts.add_option("-DSCALE_IN1=" + float_to_string_with_full_precision(iq_info.scale));
110 build_opts.add_option("-DSCALE_OUT=" + float_to_string_with_full_precision(oq_info.scale));
111 }
112
113 // Create kernel
Manuel Bottini4c6bd512020-04-08 10:15:51 +0100114 _kernel = create_kernel(compile_context, "concatenate", build_opts.options());
Vidhya Sudhan Loganathan338595b2019-06-28 14:09:53 +0100115
116 // Configure kernel window
117 auto win_config = validate_and_configure_window(input->info(), batch_offset, output->info());
118 ARM_COMPUTE_ERROR_THROW_ON(std::get<0>(win_config));
119
120 ICLKernel::configure_internal(std::get<1>(win_config));
Isabella Gottardif59b16f2019-07-25 12:03:39 +0100121
122 // Set output valid region
123 output->info()->set_valid_region(ValidRegion(Coordinates(), output->info()->tensor_shape()));
124
Vidhya Sudhan Loganathan338595b2019-06-28 14:09:53 +0100125 // Set config_id for enabling LWS tuning
126 _config_id = "concatenate_";
127 _config_id += support::cpp11::to_string(3);
128 _config_id += "_";
129 _config_id += support::cpp11::to_string(batch_offset);
130 _config_id += "_";
131 _config_id += support::cpp11::to_string(input->info()->dimension(0));
132 _config_id += "_";
133 _config_id += support::cpp11::to_string(input->info()->dimension(1));
134 _config_id += "_";
135 _config_id += support::cpp11::to_string(input->info()->dimension(2));
136 _config_id += "_";
137 _config_id += support::cpp11::to_string(input->info()->dimension(3));
138}
139
140Status CLBatchConcatenateLayerKernel::validate(const arm_compute::ITensorInfo *input,
141 unsigned int batch_offset,
142 const arm_compute::ITensorInfo *output)
143{
144 ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, batch_offset, output));
145 ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), batch_offset, output->clone().get()).first);
146 return Status{};
147}
148
149void CLBatchConcatenateLayerKernel::run(const Window &window, cl::CommandQueue &queue)
150{
151 ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
152 ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);
153
154 Window slice = window.first_slice_window_3D();
155
156 const int offset_to_first_elements_in_bytes = _batch_offset * _output->info()->strides_in_bytes()[3];
157
158 unsigned int idx = 2 * num_arguments_per_3D_tensor(); // Skip the input and output parameters
159 _kernel.setArg<cl_int>(idx, offset_to_first_elements_in_bytes);
160
161 do
162 {
163 unsigned int idx = 0;
164 add_3D_tensor_argument(idx, _input, slice);
165 add_3D_tensor_argument(idx, _output, slice);
166 enqueue(queue, *this, slice, lws_hint());
167 }
168 while(window.slide_window_slice_3D(slice));
169}
Michele Di Giorgiof6f78762020-07-06 11:27:21 +0100170} // namespace arm_compute