1 /*
2 * Copyright (c) 2017-2020 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/CLDepthConcatenateLayerKernel.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 depth_offset,const ITensorInfo * output)41 Status validate_arguments(const ITensorInfo *input, unsigned int depth_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_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::F16, DataType::F32);
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(2) + depth_offset > output->dimension(2));
51 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(3, input, output);
52
53 return Status{};
54 }
55 } // namespace
56
CLDepthConcatenateLayerKernel()57 CLDepthConcatenateLayerKernel::CLDepthConcatenateLayerKernel()
58 : _depth_offset(0)
59 {
60 }
61
configure(const CLCompileContext & compile_context,ITensorInfo * input,unsigned int depth_offset,ITensorInfo * output)62 void CLDepthConcatenateLayerKernel::configure(const CLCompileContext &compile_context, ITensorInfo *input, unsigned int depth_offset, ITensorInfo *output)
63 {
64 ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
65 ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input, depth_offset, output));
66
67 auto padding_info = get_padding_info({ input, output });
68
69 _depth_offset = depth_offset;
70
71 const unsigned int num_elems_processed_per_iteration = adjust_vec_size(16 / input->element_size(), input->dimension(0));
72
73 // Add build options
74 CLBuildOptions build_opts;
75 build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input->data_type()));
76 build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(num_elems_processed_per_iteration));
77 build_opts.add_option("-DVEC_SIZE_LEFTOVER=" + support::cpp11::to_string(input->dimension(0) % num_elems_processed_per_iteration));
78 if(is_data_type_quantized_asymmetric(input->data_type()) && input->quantization_info() != output->quantization_info())
79 {
80 const UniformQuantizationInfo iq_info = input->quantization_info().uniform();
81 const UniformQuantizationInfo oq_info = output->quantization_info().uniform();
82
83 build_opts.add_option("-DOFFSET_IN1=" + float_to_string_with_full_precision(iq_info.offset));
84 build_opts.add_option("-DOFFSET_OUT=" + float_to_string_with_full_precision(oq_info.offset));
85 build_opts.add_option("-DSCALE_IN1=" + float_to_string_with_full_precision(iq_info.scale));
86 build_opts.add_option("-DSCALE_OUT=" + float_to_string_with_full_precision(oq_info.scale));
87 }
88
89 // Create kernel
90 _kernel = create_kernel(compile_context, "concatenate", build_opts.options());
91
92 // Configure kernel window
93 auto win = calculate_max_window(*output, Steps(num_elems_processed_per_iteration));
94 win.set(Window::DimZ, Window::Dimension(0, input->tensor_shape().z(), 1));
95 ICLKernel::configure_internal(win);
96
97 // Set output valid region
98 output->set_valid_region(ValidRegion(Coordinates(), output->tensor_shape()));
99
100 ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info));
101 }
102
validate(const arm_compute::ITensorInfo * input,unsigned int depth_offset,const arm_compute::ITensorInfo * output)103 Status CLDepthConcatenateLayerKernel::validate(const arm_compute::ITensorInfo *input,
104 unsigned int depth_offset,
105 const arm_compute::ITensorInfo *output)
106 {
107 ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, depth_offset, output));
108 return Status{};
109 }
110
run_op(ITensorPack & tensors,const Window & window,cl::CommandQueue & queue)111 void CLDepthConcatenateLayerKernel::run_op(ITensorPack &tensors, const Window &window, cl::CommandQueue &queue)
112 {
113 ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
114 ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window);
115
116 const auto src = utils::cast::polymorphic_downcast<const ICLTensor *>(tensors.get_const_tensor(TensorType::ACL_SRC));
117 auto dst = utils::cast::polymorphic_downcast<ICLTensor *>(tensors.get_tensor(TensorType::ACL_DST));
118
119 Window slice = window.first_slice_window_3D();
120
121 const int offset_to_first_elements_in_bytes = _depth_offset * dst->info()->strides_in_bytes()[2];
122
123 unsigned int idx = 2 * num_arguments_per_3D_tensor(); // Skip the input and output parameters
124 _kernel.setArg<cl_int>(idx, offset_to_first_elements_in_bytes);
125
126 do
127 {
128 unsigned int idx = 0;
129 add_3D_tensor_argument(idx, src, slice);
130 add_3D_tensor_argument(idx, dst, slice);
131 enqueue(queue, *this, slice, lws_hint());
132 }
133 while(window.slide_window_slice_3D(slice));
134 }
135 } // namespace arm_compute
136