• 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/CLWidthConcatenateLayerKernel.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 width_offset,const ITensorInfo * output)41 Status validate_arguments(const ITensorInfo *input, unsigned int width_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 
47     ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
48     ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(0) + width_offset > output->dimension(0));
49 
50     for(size_t i = 1; i < Coordinates::num_max_dimensions; ++i)
51     {
52         ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(i) != output->dimension(i));
53     }
54     ARM_COMPUTE_RETURN_ERROR_ON(input->num_dimensions() > 4);
55 
56     return Status{};
57 }
58 } // namespace
59 
CLWidthConcatenateLayerKernel()60 CLWidthConcatenateLayerKernel::CLWidthConcatenateLayerKernel()
61 {
62 }
63 
validate(const ITensorInfo * input,unsigned int width_offset,const ITensorInfo * output)64 Status CLWidthConcatenateLayerKernel::validate(const ITensorInfo *input, unsigned int width_offset, const ITensorInfo *output)
65 {
66     ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, width_offset, output));
67     return Status{};
68 }
69 
configure(const CLCompileContext & compile_context,ITensorInfo * input,unsigned int width_offset,ITensorInfo * output)70 void CLWidthConcatenateLayerKernel::configure(const CLCompileContext &compile_context, ITensorInfo *input, unsigned int width_offset, ITensorInfo *output)
71 {
72     ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
73     ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input, width_offset, output));
74 
75     auto padding_info = get_padding_info({ input, output });
76 
77     const unsigned int num_elems_processed_per_iteration = adjust_vec_size(16, input->dimension(0));
78 
79     // Add build options
80     CLBuildOptions build_opts;
81     build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input->data_type()));
82     build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(num_elems_processed_per_iteration));
83     build_opts.add_option("-DVEC_SIZE_LEFTOVER=" + support::cpp11::to_string(input->dimension(0) % num_elems_processed_per_iteration));
84     build_opts.add_option("-DWIDTH_OFFSET=" + support::cpp11::to_string(width_offset));
85     build_opts.add_option("-DDEPTH=" + support::cpp11::to_string(input->dimension(2)));
86 
87     if(is_data_type_quantized_asymmetric(input->data_type()) && input->quantization_info() != output->quantization_info())
88     {
89         const UniformQuantizationInfo iqinfo = input->quantization_info().uniform();
90         const UniformQuantizationInfo oqinfo = output->quantization_info().uniform();
91 
92         build_opts.add_option("-DOFFSET_IN1=" + float_to_string_with_full_precision(iqinfo.offset));
93         build_opts.add_option("-DOFFSET_OUT=" + float_to_string_with_full_precision(oqinfo.offset));
94         build_opts.add_option("-DSCALE_IN1=" + float_to_string_with_full_precision(iqinfo.scale));
95         build_opts.add_option("-DSCALE_OUT=" + float_to_string_with_full_precision(oqinfo.scale));
96     }
97 
98     // Create kernel
99     _kernel = create_kernel(compile_context, "concatenate_width", build_opts.options());
100     // Configure kernel window
101     Window win = calculate_max_window(*input, Steps(num_elems_processed_per_iteration));
102     ICLKernel::configure_internal(win.collapse(win, Window::DimZ));
103 
104     // Set output valid region
105     output->set_valid_region(ValidRegion(Coordinates(), output->tensor_shape()));
106 
107     ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info));
108 }
109 
run_op(ITensorPack & tensors,const Window & window,cl::CommandQueue & queue)110 void CLWidthConcatenateLayerKernel::run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue)
111 {
112     ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
113     ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);
114 
115     const auto src = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC));
116     auto       dst = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_DST));
117 
118     unsigned int idx = 0;
119     add_4D_tensor_argument(idx, src, window);
120     add_4D_tensor_argument(idx, dst, window);
121     enqueue(queue, *this, window, lws_hint());
122 }
123 } // namespace arm_compute
124