Compute Library
 21.02
NEGenerateProposalsLayerKernel.cpp
Go to the documentation of this file.
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  */
25 
28 #include "arm_compute/core/Utils.h"
31 #include "src/core/CPP/Validate.h"
34 
35 #include <arm_neon.h>
36 
37 namespace arm_compute
38 {
39 namespace
40 {
41 Status validate_arguments(const ITensorInfo *anchors, const ITensorInfo *all_anchors, const ComputeAnchorsInfo &info)
42 {
43  ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(anchors, all_anchors);
45  ARM_COMPUTE_RETURN_ERROR_ON(anchors->dimension(0) != info.values_per_roi());
47  ARM_COMPUTE_RETURN_ERROR_ON(anchors->num_dimensions() > 2);
48  if(all_anchors->total_size() > 0)
49  {
50  const size_t feature_height = info.feat_height();
51  const size_t feature_width = info.feat_width();
52  const size_t num_anchors = anchors->dimension(1);
54  ARM_COMPUTE_RETURN_ERROR_ON(all_anchors->num_dimensions() > 2);
55  ARM_COMPUTE_RETURN_ERROR_ON(all_anchors->dimension(0) != info.values_per_roi());
56  ARM_COMPUTE_RETURN_ERROR_ON(all_anchors->dimension(1) != feature_height * feature_width * num_anchors);
57 
58  if(is_data_type_quantized(anchors->data_type()))
59  {
61  }
62  }
63  return Status{};
64 }
65 
66 } // namespace
67 
69  : _anchors(nullptr), _all_anchors(nullptr), _anchors_info(0.f, 0.f, 0.f)
70 {
71 }
72 
73 void NEComputeAllAnchorsKernel::configure(const ITensor *anchors, ITensor *all_anchors, const ComputeAnchorsInfo &info)
74 {
75  ARM_COMPUTE_ERROR_ON_NULLPTR(anchors, all_anchors);
76  ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(anchors->info(), all_anchors->info(), info));
77 
78  // Metadata
79  const size_t num_anchors = anchors->info()->dimension(1);
80  const DataType data_type = anchors->info()->data_type();
81  const float width = info.feat_width();
82  const float height = info.feat_height();
83 
84  // Initialize the output if empty
85  const TensorShape output_shape(info.values_per_roi(), width * height * num_anchors);
86  auto_init_if_empty(*all_anchors->info(), TensorInfo(output_shape, 1, data_type, anchors->info()->quantization_info()));
87 
88  // Set instance variables
89  _anchors = anchors;
90  _all_anchors = all_anchors;
91  _anchors_info = info;
92 
93  Window win = calculate_max_window(*all_anchors->info(), Steps(info.values_per_roi()));
94 
95  INEKernel::configure(win);
96 }
97 
99 {
100  ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(anchors, all_anchors, info));
101  return Status{};
102 }
103 
104 template <>
105 void NEComputeAllAnchorsKernel::internal_run<int16_t>(const Window &window)
106 {
107  Iterator all_anchors_it(_all_anchors, window);
108  Iterator anchors_it(_all_anchors, window);
109 
110  const size_t num_anchors = _anchors->info()->dimension(1);
111  const float stride = 1.f / _anchors_info.spatial_scale();
112  const size_t feat_width = _anchors_info.feat_width();
113 
115 
116  execute_window_loop(window, [&](const Coordinates & id)
117  {
118  const size_t anchor_offset = id.y() % num_anchors;
119 
120  const auto out_anchor_ptr = reinterpret_cast<int16_t *>(all_anchors_it.ptr());
121  const auto anchor_ptr = reinterpret_cast<int16_t *>(_anchors->ptr_to_element(Coordinates(0, anchor_offset)));
122 
123  const size_t shift_idy = id.y() / num_anchors;
124  const float shiftx = (shift_idy % feat_width) * stride;
125  const float shifty = (shift_idy / feat_width) * stride;
126 
127  const float new_anchor_x1 = dequantize_qsymm16(*anchor_ptr, qinfo.scale) + shiftx;
128  const float new_anchor_y1 = dequantize_qsymm16(*(1 + anchor_ptr), qinfo.scale) + shifty;
129  const float new_anchor_x2 = dequantize_qsymm16(*(2 + anchor_ptr), qinfo.scale) + shiftx;
130  const float new_anchor_y2 = dequantize_qsymm16(*(3 + anchor_ptr), qinfo.scale) + shifty;
131 
132  *out_anchor_ptr = quantize_qsymm16(new_anchor_x1, qinfo.scale);
133  *(out_anchor_ptr + 1) = quantize_qsymm16(new_anchor_y1, qinfo.scale);
134  *(out_anchor_ptr + 2) = quantize_qsymm16(new_anchor_x2, qinfo.scale);
135  *(out_anchor_ptr + 3) = quantize_qsymm16(new_anchor_y2, qinfo.scale);
136  },
137  all_anchors_it);
138 }
139 
140 template <typename T>
141 void NEComputeAllAnchorsKernel::internal_run(const Window &window)
142 {
143  Iterator all_anchors_it(_all_anchors, window);
144  Iterator anchors_it(_all_anchors, window);
145 
146  const size_t num_anchors = _anchors->info()->dimension(1);
147  const T stride = 1.f / _anchors_info.spatial_scale();
148  const size_t feat_width = _anchors_info.feat_width();
149 
150  execute_window_loop(window, [&](const Coordinates & id)
151  {
152  const size_t anchor_offset = id.y() % num_anchors;
153 
154  const auto out_anchor_ptr = reinterpret_cast<T *>(all_anchors_it.ptr());
155  const auto anchor_ptr = reinterpret_cast<T *>(_anchors->ptr_to_element(Coordinates(0, anchor_offset)));
156 
157  const size_t shift_idy = id.y() / num_anchors;
158  const T shiftx = (shift_idy % feat_width) * stride;
159  const T shifty = (shift_idy / feat_width) * stride;
160 
161  *out_anchor_ptr = *anchor_ptr + shiftx;
162  *(out_anchor_ptr + 1) = *(1 + anchor_ptr) + shifty;
163  *(out_anchor_ptr + 2) = *(2 + anchor_ptr) + shiftx;
164  *(out_anchor_ptr + 3) = *(3 + anchor_ptr) + shifty;
165  },
166  all_anchors_it);
167 }
168 
169 void NEComputeAllAnchorsKernel::run(const Window &window, const ThreadInfo &info)
170 {
171  ARM_COMPUTE_UNUSED(info);
174 
175  switch(_anchors->info()->data_type())
176  {
177  case DataType::QSYMM16:
178  {
179  internal_run<int16_t>(window);
180  break;
181  }
182 #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
183  case DataType::F16:
184  {
185  internal_run<float16_t>(window);
186  break;
187  }
188 #endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
189  case DataType::F32:
190  {
191  internal_run<float>(window);
192  break;
193  }
194  default:
195  {
196  ARM_COMPUTE_ERROR("Data type not supported");
197  }
198  }
199 }
200 } // namespace arm_compute
bool is_data_type_quantized(DataType dt)
Check if a given data type is of quantized type.
Definition: Utils.h:1168
Window calculate_max_window(const ValidRegion &valid_region, const Steps &steps, bool skip_border, BorderSize border_size)
const Window & window() const
The maximum window the kernel can be executed on.
Definition: IKernel.cpp:28
uint8_t * ptr_to_element(const Coordinates &id) const
Return a pointer to the element at the passed coordinates.
Definition: ITensor.h:63
int16_t quantize_qsymm16(float value, const UniformQuantizationInfo &qinfo, RoundingPolicy rounding_policy=RoundingPolicy::TO_NEAREST_UP)
Quantize a value given a 16-bit symmetric quantization scheme.
Shape of a tensor.
Definition: TensorShape.h:39
quantized, symmetric fixed-point 16-bit number
#define ARM_COMPUTE_RETURN_ERROR_ON_CPU_F16_UNSUPPORTED(tensor)
Definition: Validate.h:108
virtual size_t dimension(size_t index) const =0
Return the size of the requested dimension.
#define ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(...)
Definition: Validate.h:610
size_t values_per_roi() const
Definition: Types.h:1470
#define ARM_COMPUTE_ERROR(msg)
Print the given message then throw an std::runtime_error.
Definition: Error.h:352
#define ARM_COMPUTE_RETURN_ON_ERROR(status)
Checks if a status contains an error and returns it.
Definition: Error.h:204
virtual DataType data_type() const =0
Data type used for each element of the tensor.
1 channel, 1 F32 per channel
Store the tensor&#39;s metadata.
Definition: ITensorInfo.h:40
#define ARM_COMPUTE_ERROR_THROW_ON(status)
Definition: Error.h:455
Quantization info when assuming per layer quantization.
static Status validate(const ITensorInfo *anchors, const ITensorInfo *all_anchors, const ComputeAnchorsInfo &info)
Static function to check if given info will lead to a valid configuration of NEComputeAllAnchorsKerne...
Status class.
Definition: Error.h:52
void configure(const ITensor *anchors, ITensor *all_anchors, const ComputeAnchorsInfo &info)
Set the input and output tensors.
#define ARM_COMPUTE_RETURN_ERROR_ON(cond)
If the condition is true, an error is returned.
Definition: Error.h:296
Interface for Neon tensor.
Definition: ITensor.h:36
float spatial_scale() const
Definition: Types.h:1464
Copyright (c) 2017-2021 Arm Limited.
1 channel, 1 F16 per channel
#define ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(...)
Definition: Validate.h:163
const DataType data_type
Definition: Im2Col.cpp:150
#define ARM_COMPUTE_UNUSED(...)
To avoid unused variables warnings.
Definition: Error.h:152
float dequantize_qsymm16(int16_t value, const UniformQuantizationInfo &qinfo)
Dequantize a value given a 16-bit symmetric quantization scheme.
Class to describe a number of elements in each dimension.
Definition: Steps.h:40
Coordinates of an item.
Definition: Coordinates.h:37
UniformQuantizationInfo uniform() const
Return per layer quantization info.
bool auto_init_if_empty(ITensorInfo &info, const TensorShape &shape, int num_channels, DataType data_type, QuantizationInfo quantization_info=QuantizationInfo())
Auto initialize the tensor info (shape, number of channels and data type) if the current assignment i...
virtual ITensorInfo * info() const =0
Interface to be implemented by the child class to return the tensor&#39;s metadata.
constexpr uint8_t * ptr() const
Return a pointer to the current pixel.
Definition: Helpers.inl:139
virtual QuantizationInfo quantization_info() const =0
Get the quantization settings (scale and offset) of the tensor.
#define ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(k)
Definition: Validate.h:941
void run(const Window &window, const ThreadInfo &info) override
Execute the kernel on the passed window.
ScaleKernelInfo info(interpolation_policy, default_border_mode, PixelValue(), sampling_policy, false)
Information about executing thread and CPU.
Definition: CPPTypes.h:235
const QuantizationInfo qinfo
Definition: Im2Col.cpp:155
#define ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(...)
Definition: Validate.h:545
ComputeAnchors information class.
Definition: Types.h:1433
constexpr const Dimension & y() const
Alias to access the second dimension of the window.
Definition: Window.h:154
Status validate_arguments(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, const GEMMLowpOutputStageInfo *output_stage)
#define ARM_COMPUTE_ERROR_ON_NULLPTR(...)
Definition: Validate.h:161
Store the tensor&#39;s metadata.
Definition: TensorInfo.h:45
void execute_window_loop(const Window &w, L &&lambda_function, Ts &&... iterators)
Iterate through the passed window, automatically adjusting the iterators and calling the lambda_funct...
Definition: Helpers.inl:77
#define ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_NOT_IN(t,...)
Definition: Validate.h:694
Iterator updated by execute_window_loop for each window element.
Definition: Helpers.h:46
DataType
Available data types.
Definition: Types.h:77
Describe a multidimensional execution window.
Definition: Window.h:39
#define ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(f, s)
Definition: Validate.h:205