Compute Library
 20.08
NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel.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 
27 #include "arm_compute/core/Error.h"
32 #include "arm_compute/core/Types.h"
33 #include "arm_compute/core/Utils.h"
37 
38 #include <arm_neon.h>
39 #include <cstddef>
40 #include <cstdint>
41 
42 namespace arm_compute
43 {
44 namespace
45 {
46 Status validate_arguments(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, int min, int max)
47 {
49  ARM_COMPUTE_RETURN_ERROR_ON(min > max);
50 
51  // Check biases if exist
52  if(bias != nullptr)
53  {
55  ARM_COMPUTE_RETURN_ERROR_ON(bias->num_dimensions() > 1);
56  ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(0) != bias->dimension(0));
57  }
58 
59  if(output->total_size() != 0)
60  {
63  }
64 
65  return Status{};
66 }
67 
68 std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *output)
69 {
70  // Output auto inizialitation if not yet initialized
71  auto_init_if_empty(*output, input->clone()->set_data_type(DataType::QSYMM16));
72 
73  // Configure kernel window
74  Window win = calculate_max_window(*input, Steps());
75 
76  // NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel doesn't need padding so update_window_and_padding() can be skipped
77  Coordinates coord;
78  coord.set_num_dimensions(output->num_dimensions());
79  output->set_valid_region(ValidRegion(coord, output->tensor_shape()));
80 
81  return std::make_pair(Status{}, win);
82 }
83 } // namespace
84 
85 class Coordinates;
86 
87 template <bool is_bounded_relu>
89 {
90  const int16x8_t min_s16 = vdupq_n_s16(static_cast<int16_t>(_min));
91  const int16x8_t max_s16 = vdupq_n_s16(static_cast<int16_t>(_max));
92 
93  ARM_COMPUTE_UNUSED(min_s16);
94  ARM_COMPUTE_UNUSED(max_s16);
95 
96  const int window_step_x = 8;
97  const auto window_start_x = static_cast<int>(window.x().start());
98  const auto window_end_x = static_cast<int>(window.x().end());
99 
100  Window win_collapsed = window.collapse_if_possible(window, Window::DimZ);
101  win_collapsed.set(Window::DimX, Window::Dimension(0, 1, 1));
102 
103  Iterator in(_input, win_collapsed);
104  Iterator out(_output, win_collapsed);
105  if(_bias != nullptr)
106  {
107  Window win_biases;
108  win_biases.set(Window::DimX, Window::Dimension(0, 1, 1));
109  win_biases.set(Window::DimY, Window::Dimension(0, 1, 1));
110 
111  Iterator bias(_bias, win_biases);
112  execute_window_loop(win_collapsed, [&](const Coordinates &)
113  {
114  // Compute 16 elements per iteration
115  int x = window_start_x;
116  for(; x <= (window_end_x - window_step_x); x += window_step_x)
117  {
118  int32x4x2_t in_s32 =
119  {
120  {
121  vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 0),
122  vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 4)
123  }
124  };
125 
126  const int32x4x2_t bias_s32 =
127  {
128  {
129  vld1q_s32(reinterpret_cast<const int32_t *>(bias.ptr()) + x + 0),
130  vld1q_s32(reinterpret_cast<const int32_t *>(bias.ptr()) + x + 4)
131  }
132  };
133 
134  // Add the bias to GEMM's result
135  in_s32.val[0] = vaddq_s32(in_s32.val[0], bias_s32.val[0]);
136  in_s32.val[1] = vaddq_s32(in_s32.val[1], bias_s32.val[1]);
137 
138  vst1q_s16(reinterpret_cast<int16_t *>(out.ptr()) + x, finalize_quantization_int16<is_bounded_relu>(in_s32, _result_fixedpoint_multiplier, _result_shift, min_s16, max_s16));
139  }
140 
141  // Compute left-over elements
142  for(; x < window_end_x; ++x)
143  {
144  const int32_t bias_value = *(reinterpret_cast<const int32_t *>(bias.ptr()) + x);
145  int32_t in_value = *(reinterpret_cast<const int32_t *>(in.ptr()) + x);
146 
147  // Add bias
148  in_value += bias_value;
149  // Finalize and store the result
150  *(reinterpret_cast<int16_t *>(out.ptr()) + x) = finalize_quantization_int16<is_bounded_relu>(in_value, _result_fixedpoint_multiplier, _result_shift, static_cast<int16_t>(_min),
151  static_cast<int16_t>(_max));
152  }
153  },
154  in, out, bias);
155  }
156  else
157  {
158  execute_window_loop(win_collapsed, [&](const Coordinates &)
159  {
160  // Compute 16 elements per iteration
161  int x = window_start_x;
162  for(; x <= (window_end_x - window_step_x); x += window_step_x)
163  {
164  int32x4x2_t in_s32 =
165  {
166  {
167  vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 0),
168  vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 4)
169  }
170  };
171 
172  vst1q_s16(reinterpret_cast<int16_t *>(out.ptr()) + x, finalize_quantization_int16<is_bounded_relu>(in_s32, _result_fixedpoint_multiplier, _result_shift, min_s16, max_s16));
173  }
174 
175  // Compute left-over elements
176  for(; x < window_end_x; ++x)
177  {
178  const int32_t in_value = *(reinterpret_cast<const int32_t *>(in.ptr()) + x);
179  ARM_COMPUTE_UNUSED(in_value);
180  // Finalize and store the result
181  *(reinterpret_cast<int16_t *>(out.ptr()) + x) = finalize_quantization_int16<is_bounded_relu>(in_value, _result_fixedpoint_multiplier, _result_shift, static_cast<int16_t>(_min),
182  static_cast<int16_t>(_max));
183  }
184  },
185  in, out);
186  }
187 }
188 
190  : _func(nullptr), _input(nullptr), _bias(nullptr), _output(nullptr), _result_fixedpoint_multiplier(0), _result_shift(0), _min(0), _max(0)
191 {
192 }
193 
194 void NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel::configure(const ITensor *input, const ITensor *bias, ITensor *output, int result_fixedpoint_multiplier, int result_shift,
195  int min, int max)
196 {
197  // Perform validate step
199  ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), (bias != nullptr) ? bias->info() : nullptr, output->info(), min, max));
200 
201  _input = input;
202  _bias = bias;
203  _output = output;
204  _result_fixedpoint_multiplier = result_fixedpoint_multiplier;
205  _result_shift = result_shift;
206  _min = min;
207  _max = max;
208 
209  // Configure kernel window
210  auto win_config = validate_and_configure_window(input->info(), output->info());
211  ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
212  INEKernel::configure(win_config.second);
213 
214  // Check if we need to clamp the result using min and max
215  const bool is_bounded_relu = !(min <= -32768 && max >= 32767);
216  _func = is_bounded_relu ? &NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel::run<true> : &NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel::run<false>;
217 }
218 
220 {
223  ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), output->clone().get()).first);
224 
225  return Status{};
226 }
227 
229 {
233 
234  (this->*_func)(window);
235 }
236 } // namespace arm_compute
static Status validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, int min=0, int max=0)
Static function to check if given info will lead to a valid configuration of NEGEMMLowpQuantizeDownIn...
const Window & window() const
The maximum window the kernel can be executed on.
Definition: IKernel.cpp:28
quantized, symmetric fixed-point 16-bit number
#define ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(...)
Definition: Validate.h:545
void configure(const ITensor *input, const ITensor *bias, ITensor *output, int result_fixedpoint_multiplier, int result_shift, int min=0, int max=0)
Initialise the kernel's input and output.
#define ARM_COMPUTE_RETURN_ON_ERROR(status)
Checks if a status contains an error and returns it.
Definition: Error.h:204
#define ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(t, c,...)
Definition: Validate.h:792
Store the tensor's metadata.
Definition: ITensorInfo.h:40
#define ARM_COMPUTE_ERROR_THROW_ON(status)
Definition: Error.h:455
Status class.
Definition: Error.h:52
#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
Window calculate_max_window(const ValidRegion &valid_region, const Steps &steps=Steps(), bool skip_border=false, BorderSize border_size=BorderSize())
Calculate the maximum window for a given tensor shape and border setting.
Definition: Helpers.cpp:28
Copyright (c) 2017-2020 Arm Limited.
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...
Definition: Helpers.inl:207
ITensorInfo * info() const override
Interface to be implemented by the child class to return the tensor's metadata.
Definition: Tensor.cpp:33
1 channel, 1 S32 per channel
static constexpr size_t DimX
Alias for dimension 0 also known as X dimension.
Definition: Window.h:43
#define ARM_COMPUTE_UNUSED(...)
To avoid unused variables warnings.
Definition: Error.h:152
Window collapse_if_possible(const Window &full_window, size_t first, size_t last, bool *has_collapsed=nullptr) const
Collapse the dimensions between first and last if possible.
Definition: Window.inl:68
#define ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(...)
Definition: Validate.h:443
virtual std::unique_ptr< T > clone() const =0
Provide a clone of the current object of class T.
virtual ITensorInfo * info() const =0
Interface to be implemented by the child class to return the tensor's metadata.
void set(size_t dimension, const Dimension &dim)
Set the values of a given dimension.
Definition: Window.inl:49
static constexpr size_t DimY
Alias for dimension 1 also known as Y dimension.
Definition: Window.h:45
#define ARM_COMPUTE_ERROR_ON_NULLPTR(...)
Definition: Validate.h:161
Information about executing thread and CPU.
Definition: CPPTypes.h:235
static constexpr size_t DimZ
Alias for dimension 2 also known as Z dimension.
Definition: Window.h:47
Status validate_arguments(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, const GEMMLowpOutputStageInfo *output_stage)
void run(const Window &window, const ThreadInfo &info) override
Execute the kernel on the passed window.
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:128
constexpr int end() const
Return the end of the dimension.
Definition: Window.h:97
#define ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(f, s)
Definition: Validate.h:205
constexpr int start() const
Return the start of the dimension.
Definition: Window.h:92
Describe a multidimensional execution window.
Definition: Window.h:39
#define ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(k)
Definition: Validate.h:941
constexpr const Dimension & x() const
Alias to access the first dimension of the window.
Definition: Window.h:143