1 /*
2 * Copyright (c) 2019-2020 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/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/ICLTensor.h"
29 #include "arm_compute/core/Helpers.h"
30 #include "arm_compute/core/Utils.h"
31 #include "src/core/CL/CLValidate.h"
32 #include "src/core/helpers/WindowHelpers.h"
33 #include "support/Cast.h"
34
35 #include "support/StringSupport.h"
36
37 namespace arm_compute
38 {
39 namespace
40 {
validate_arguments(const ITensorInfo * input,unsigned int batch_offset,const ITensorInfo * output)41 Status validate_arguments(const ITensorInfo *input, unsigned int batch_offset, const ITensorInfo *output)
42 {
43 ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output);
44 ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input);
45 ARM_COMPUTE_RETURN_ERROR_ON(input->data_type() == DataType::UNKNOWN);
46 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
47
48 ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(Window::DimX) != output->dimension(Window::DimX));
49 ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(Window::DimY) != output->dimension(Window::DimY));
50 ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(Window::DimZ) != output->dimension(Window::DimZ));
51 ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(3) + batch_offset > output->dimension(3));
52 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(4, input, output);
53
54 return Status{};
55 }
56 } // namespace
57
CLBatchConcatenateLayerKernel()58 CLBatchConcatenateLayerKernel::CLBatchConcatenateLayerKernel()
59 : _batch_offset(0)
60 {
61 }
62
configure(const CLCompileContext & compile_context,ITensorInfo * input,unsigned int batch_offset,ITensorInfo * output)63 void CLBatchConcatenateLayerKernel::configure(const CLCompileContext &compile_context, ITensorInfo *input, unsigned int batch_offset, ITensorInfo *output)
64 {
65 ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
66 ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input, batch_offset, output));
67
68 auto padding_info = get_padding_info({ input, output });
69
70 _batch_offset = batch_offset;
71
72 const unsigned int num_elems_processed_per_iteration = adjust_vec_size(16 / input->element_size(), input->dimension(0));
73
74 // Add build options
75 CLBuildOptions build_opts;
76 build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input->data_type()));
77 build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(num_elems_processed_per_iteration));
78 build_opts.add_option("-DVEC_SIZE_LEFTOVER=" + support::cpp11::to_string(input->dimension(0) % num_elems_processed_per_iteration));
79 if(is_data_type_quantized_asymmetric(input->data_type()) && input->quantization_info() != output->quantization_info())
80 {
81 const UniformQuantizationInfo iq_info = input->quantization_info().uniform();
82 const UniformQuantizationInfo oq_info = output->quantization_info().uniform();
83
84 build_opts.add_option("-DOFFSET_IN1=" + float_to_string_with_full_precision(iq_info.offset));
85 build_opts.add_option("-DOFFSET_OUT=" + float_to_string_with_full_precision(oq_info.offset));
86 build_opts.add_option("-DSCALE_IN1=" + float_to_string_with_full_precision(iq_info.scale));
87 build_opts.add_option("-DSCALE_OUT=" + float_to_string_with_full_precision(oq_info.scale));
88 }
89
90 // Create kernel
91 _kernel = create_kernel(compile_context, "concatenate", build_opts.options());
92
93 // Configure kernel window
94 auto win = calculate_max_window(*output, Steps(num_elems_processed_per_iteration));
95 win.set(3, Window::Dimension(0, input->tensor_shape()[3], 1));
96 ICLKernel::configure_internal(win);
97
98 // Set output valid region
99 output->set_valid_region(ValidRegion(Coordinates(), output->tensor_shape()));
100
101 // Set config_id for enabling LWS tuning
102 _config_id = "concatenate_";
103 _config_id += support::cpp11::to_string(3);
104 _config_id += "_";
105 _config_id += support::cpp11::to_string(batch_offset);
106 _config_id += "_";
107 _config_id += support::cpp11::to_string(input->dimension(0));
108 _config_id += "_";
109 _config_id += support::cpp11::to_string(input->dimension(1));
110 _config_id += "_";
111 _config_id += support::cpp11::to_string(input->dimension(2));
112 _config_id += "_";
113 _config_id += support::cpp11::to_string(input->dimension(3));
114
115 ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info));
116 }
117
validate(const arm_compute::ITensorInfo * input,unsigned int batch_offset,const arm_compute::ITensorInfo * output)118 Status CLBatchConcatenateLayerKernel::validate(const arm_compute::ITensorInfo *input,
119 unsigned int batch_offset,
120 const arm_compute::ITensorInfo *output)
121 {
122 ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, batch_offset, output));
123 return Status{};
124 }
125
run_op(ITensorPack & tensors,const Window & window,cl::CommandQueue & queue)126 void CLBatchConcatenateLayerKernel::run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue)
127 {
128 ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
129 ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);
130
131 const auto src = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC));
132 auto dst = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_DST));
133
134 Window slice = window.first_slice_window_3D();
135
136 const int offset_to_first_elements_in_bytes = _batch_offset * dst->info()->strides_in_bytes()[3];
137
138 unsigned int idx = 2 * num_arguments_per_3D_tensor(); // Skip the input and output parameters
139 _kernel.setArg<cl_int>(idx, offset_to_first_elements_in_bytes);
140
141 do
142 {
143 unsigned int idx = 0;
144 add_3D_tensor_argument(idx, src, slice);
145 add_3D_tensor_argument(idx, dst, slice);
146 enqueue(queue, *this, slice, lws_hint());
147 }
148 while(window.slide_window_slice_3D(slice));
149 }
150 } // namespace arm_compute
151