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/NEON/kernels/NEDepthConcatenateLayerKernel.h"
25
26 #include "arm_compute/core/Error.h"
27 #include "arm_compute/core/Helpers.h"
28 #include "arm_compute/core/IAccessWindow.h"
29 #include "arm_compute/core/ITensor.h"
30 #include "arm_compute/core/TensorInfo.h"
31 #include "arm_compute/core/Utils.h"
32 #include "arm_compute/core/Validate.h"
33 #include "arm_compute/core/Window.h"
34 #include "src/core/NEON/NEAsymm.h"
35 #include "src/core/NEON/NEFixedPoint.h"
36 #include "src/core/NEON/wrapper/wrapper.h"
37 #include "src/core/helpers/AutoConfiguration.h"
38 #include "src/core/helpers/WindowHelpers.h"
39
40 #include <cstdint>
41
42 namespace arm_compute
43 {
44 namespace
45 {
46 template <typename T>
depth_concat(const ITensor * in,ITensor * out,unsigned int depth_offset,const Window & window)47 void depth_concat(const ITensor *in, ITensor *out, unsigned int depth_offset, const Window &window)
48 {
49 // Offset input
50 uint8_t *input_ptr = in->buffer() + in->info()->offset_first_element_in_bytes();
51
52 // Offset output
53 uint8_t *output_ptr = out->buffer() + out->info()->offset_first_element_in_bytes() + depth_offset * out->info()->strides_in_bytes()[2];
54
55 const auto window_start_x = static_cast<int>(window.x().start());
56 const auto window_end_x = static_cast<int>(window.x().end());
57 const int window_step_x = 16 / out->info()->element_size();
58
59 Window win{ window };
60 win.set(Window::DimX, Window::Dimension(0, 1, 1));
61 win.set(Window::DimZ, Window::Dimension(0, in->info()->tensor_shape().z(), 1));
62
63 Iterator input(in, win);
64 Iterator output(out, win);
65
66 const DataType dt = in->info()->data_type();
67 const UniformQuantizationInfo input_qinfo = in->info()->quantization_info().uniform();
68 const UniformQuantizationInfo output_qinfo = out->info()->quantization_info().uniform();
69 if(dt == DataType::QASYMM8 && input_qinfo != output_qinfo)
70 {
71 execute_window_loop(win, [&](const Coordinates &)
72 {
73 const auto in_ptr = reinterpret_cast<const uint8_t *>(input_ptr + input.offset());
74 const auto out_ptr = reinterpret_cast<uint8_t *>(output_ptr + output.offset());
75 int x = window_start_x;
76 for(; x <= (window_end_x - window_step_x); x += window_step_x)
77 {
78 wrapper::vstore(out_ptr + x, vquantize(vdequantize(wrapper::vloadq(in_ptr + x), input_qinfo), output_qinfo));
79 }
80
81 // Compute left-over elements
82 for(; x < window_end_x; ++x)
83 {
84 *(out_ptr + x) = quantize_qasymm8(dequantize_qasymm8(*(in_ptr + x), input_qinfo), output_qinfo);
85 }
86 },
87 input, output);
88 }
89 else if(dt == DataType::QASYMM8_SIGNED && input_qinfo != output_qinfo)
90 {
91 execute_window_loop(win, [&](const Coordinates &)
92 {
93 const auto in_ptr = reinterpret_cast<const int8_t *>(input_ptr + input.offset());
94 const auto out_ptr = reinterpret_cast<int8_t *>(output_ptr + output.offset());
95 int x = window_start_x;
96 for(; x <= (window_end_x - window_step_x); x += window_step_x)
97 {
98 wrapper::vstore(out_ptr + x, vquantize_signed(vdequantize(wrapper::vloadq(in_ptr + x), input_qinfo), output_qinfo));
99 }
100
101 // Compute left-over elements
102 for(; x < window_end_x; ++x)
103 {
104 *(out_ptr + x) = quantize_qasymm8_signed(dequantize_qasymm8_signed(*(in_ptr + x), input_qinfo), output_qinfo);
105 }
106 },
107 input, output);
108 }
109 else
110 {
111 execute_window_loop(win, [&](const Coordinates &)
112 {
113 const auto in_ptr = reinterpret_cast<const T *>(input_ptr + input.offset());
114 const auto out_ptr = reinterpret_cast<T *>(output_ptr + output.offset());
115 int x = window_start_x;
116 for(; x <= (window_end_x - window_step_x); x += window_step_x)
117 {
118 wrapper::vstore(out_ptr + x, wrapper::vloadq(in_ptr + x));
119 }
120 // Compute left-over elements
121 for(; x < window_end_x; ++x)
122 {
123 *(out_ptr + x) = *(in_ptr + x);
124 }
125 },
126 input, output);
127 }
128 }
129
validate_arguments(const ITensorInfo * input,unsigned int depth_offset,const ITensorInfo * output)130 Status validate_arguments(const ITensorInfo *input, unsigned int depth_offset, const ITensorInfo *output)
131 {
132 ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output);
133 //Note: ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(input) is not needed here as this kernel doesn't use NEON FP16 instructions.
134 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::F16, DataType::F32);
135 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output);
136
137 ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(Window::DimX) != output->dimension(Window::DimX));
138 ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(Window::DimY) != output->dimension(Window::DimY));
139 ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(2) + depth_offset > output->dimension(2));
140 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(3, input, output);
141
142 return Status{};
143 }
144 } // namespace
145
NEDepthConcatenateLayerKernel()146 NEDepthConcatenateLayerKernel::NEDepthConcatenateLayerKernel()
147 : _func(nullptr), _depth_offset(0)
148 {
149 }
150
configure(const ITensorInfo * input,unsigned int depth_offset,ITensorInfo * output)151 void NEDepthConcatenateLayerKernel::configure(const ITensorInfo *input, unsigned int depth_offset, ITensorInfo *output)
152 {
153 ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
154 ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input, depth_offset, output));
155
156 _func = nullptr;
157 _depth_offset = depth_offset;
158
159 switch(input->data_type())
160 {
161 case DataType::QASYMM8:
162 _func = &depth_concat<uint8_t>;
163 break;
164 case DataType::QASYMM8_SIGNED:
165 _func = &depth_concat<int8_t>;
166 break;
167 case DataType::F16:
168 _func = &depth_concat<uint16_t>;
169 break;
170 case DataType::F32:
171 _func = &depth_concat<uint32_t>;
172 break;
173 default:
174 ARM_COMPUTE_ERROR("Unsupported data type.");
175 }
176
177 // Configure kernel window
178 Window win = calculate_max_window(*output, Steps());
179 Coordinates coord;
180 coord.set_num_dimensions(output->num_dimensions());
181
182 output->set_valid_region(ValidRegion(coord, output->tensor_shape()));
183 INEKernel::configure(win);
184 }
185
validate(const arm_compute::ITensorInfo * input,unsigned int depth_offset,const arm_compute::ITensorInfo * output)186 Status NEDepthConcatenateLayerKernel::validate(const arm_compute::ITensorInfo *input,
187 unsigned int depth_offset,
188 const arm_compute::ITensorInfo *output)
189 {
190 ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, depth_offset, output));
191 return Status{};
192 }
193
run_op(ITensorPack & tensors,const Window & window,const ThreadInfo & info)194 void NEDepthConcatenateLayerKernel::run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info)
195 {
196 ARM_COMPUTE_UNUSED(info);
197 ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
198 ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window);
199 ARM_COMPUTE_ERROR_ON(_func == nullptr);
200
201 (*_func)(tensors.get_const_tensor(TensorType::ACL_SRC),
202 tensors.get_tensor(TensorType::ACL_DST),
203 _depth_offset,
204 window);
205 }
206 } // namespace arm_compute
207