Compute Library
 21.05
NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel.cpp
Go to the documentation of this file.
1 /*
2  * Copyright (c) 2019-2021 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 
26 #include "arm_compute/core/Error.h"
30 #include "arm_compute/core/Types.h"
31 #include "arm_compute/core/Utils.h"
35 #include "src/core/NEON/NEAsymm.h"
38 
39 #include <arm_neon.h>
40 #include <cstddef>
41 #include <cstdint>
42 
43 namespace arm_compute
44 {
45 namespace
46 {
47 Status validate_arguments(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, int min, int max)
48 {
50  ARM_COMPUTE_RETURN_ERROR_ON(min > max);
51 
52  // Check biases if exist
53  if(bias != nullptr)
54  {
56  ARM_COMPUTE_RETURN_ERROR_ON(bias->num_dimensions() > 1);
57  ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(0) != bias->dimension(0));
58  }
59 
60  if(output->total_size() != 0)
61  {
64  }
65 
66  return Status{};
67 }
68 
69 std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *output)
70 {
71  // Output auto initialization if not yet initialized
72  auto_init_if_empty(*output, input->clone()->set_data_type(DataType::QASYMM8_SIGNED));
73 
74  // Configure kernel window
75  Window win = calculate_max_window(*input, Steps());
76 
77  // NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel doesn't need padding so update_window_and_padding() can be skipped
78 
79  return std::make_pair(Status{}, win);
80 }
81 } // namespace
82 
83 template <bool is_bounded_relu>
85 {
86  const int32x4_t result_offset_after_shift_s32 = vdupq_n_s32(_result_offset_after_shift);
87  const int8x16_t min_s8 = vdupq_n_s8(static_cast<int8_t>(_min));
88  const int8x16_t max_s8 = vdupq_n_s8(static_cast<int8_t>(_max));
89 
90  ARM_COMPUTE_UNUSED(min_s8, max_s8);
91 
92  const int window_step_x = 16;
93  const auto window_start_x = static_cast<int>(window.x().start());
94  const auto window_end_x = static_cast<int>(window.x().end());
95 
96  Window win_collapsed = window.collapse_if_possible(window, Window::DimZ);
97  win_collapsed.set(Window::DimX, Window::Dimension(0, 1, 1));
98 
99  Iterator in(_input, win_collapsed);
100  Iterator out(_output, win_collapsed);
101  if(_bias != nullptr)
102  {
103  Window win_biases;
104  win_biases.set(Window::DimX, Window::Dimension(0, 1, 1));
105  win_biases.set(Window::DimY, Window::Dimension(0, 1, 1));
106 
107  Iterator bias(_bias, win_biases);
108  execute_window_loop(win_collapsed, [&](const Coordinates &)
109  {
110  // Compute 16 elements per iteration
111  int x = window_start_x;
112  for(; x <= (window_end_x - window_step_x); x += window_step_x)
113  {
114  int32x4x4_t in_s32 =
115  {
116  {
117  vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 0),
118  vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 4),
119  vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 8),
120  vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 12)
121  }
122  };
123 
124  const int32x4x4_t bias_s32 =
125  {
126  {
127  vld1q_s32(reinterpret_cast<const int32_t *>(bias.ptr()) + x + 0),
128  vld1q_s32(reinterpret_cast<const int32_t *>(bias.ptr()) + x + 4),
129  vld1q_s32(reinterpret_cast<const int32_t *>(bias.ptr()) + x + 8),
130  vld1q_s32(reinterpret_cast<const int32_t *>(bias.ptr()) + x + 12)
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  in_s32.val[2] = vaddq_s32(in_s32.val[2], bias_s32.val[2]);
138  in_s32.val[3] = vaddq_s32(in_s32.val[3], bias_s32.val[3]);
139 
140  vst1q_s8(reinterpret_cast<int8_t *>(out.ptr() + x),
141  finalize_quantization(in_s32, _result_fixedpoint_multiplier, _result_shift, result_offset_after_shift_s32, min_s8, max_s8, is_bounded_relu));
142  }
143 
144  // Compute left-over elements
145  for(; x < window_end_x; ++x)
146  {
147  const int32_t bias_value = *(reinterpret_cast<const int32_t *>(bias.ptr()) + x);
148  int32_t in_value = *(reinterpret_cast<const int32_t *>(in.ptr()) + x);
149 
150  // Add bias
151  in_value += bias_value;
152  // Finalize and store the result
153  *reinterpret_cast<int8_t *>(out.ptr() + x) = finalize_quantization(in_value, _result_fixedpoint_multiplier, _result_shift, _result_offset_after_shift,
154  static_cast<int8_t>(_min), static_cast<int8_t>(_max), is_bounded_relu);
155  }
156  },
157  in, out, bias);
158  }
159  else
160  {
161  execute_window_loop(win_collapsed, [&](const Coordinates &)
162  {
163  // Compute 16 elements per iteration
164  int x = window_start_x;
165  for(; x <= (window_end_x - window_step_x); x += window_step_x)
166  {
167  int32x4x4_t in_s32 =
168  {
169  {
170  vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 0),
171  vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 4),
172  vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 8),
173  vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 12)
174  }
175  };
176 
177  vst1q_s8(reinterpret_cast<int8_t *>(out.ptr() + x),
178  finalize_quantization(in_s32, _result_fixedpoint_multiplier, _result_shift, result_offset_after_shift_s32, min_s8, max_s8, is_bounded_relu));
179  }
180 
181  // Compute left-over elements
182  for(; x < window_end_x; ++x)
183  {
184  const int32_t in_value = *(reinterpret_cast<const int32_t *>(in.ptr()) + x);
185 
186  // Finalize and store the result
187  *reinterpret_cast<int8_t *>(out.ptr() + x) = finalize_quantization(in_value, _result_fixedpoint_multiplier, _result_shift, _result_offset_after_shift,
188  static_cast<int8_t>(_min), static_cast<int8_t>(_max), is_bounded_relu);
189  }
190  },
191  in, out);
192  }
193 }
194 
196  : _func(nullptr), _input(nullptr), _bias(nullptr), _output(nullptr), _result_fixedpoint_multiplier(0), _result_shift(0), _result_offset_after_shift(0), _min(0), _max(0)
197 {
198 }
199 
200 void NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel::configure(const ITensor *input, const ITensor *bias, ITensor *output, int result_fixedpoint_multiplier, int result_shift,
201  int result_offset_after_shift, int min, int max)
202 {
203  // Perform validate step
205  ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), (bias != nullptr) ? bias->info() : nullptr, output->info(), min, max));
206 
207  _input = input;
208  _bias = bias;
209  _output = output;
210  _result_fixedpoint_multiplier = result_fixedpoint_multiplier;
211  _result_shift = result_shift;
212  _result_offset_after_shift = result_offset_after_shift;
213  _min = min;
214  _max = max;
215 
216  // Configure kernel window
217  auto win_config = validate_and_configure_window(input->info(), output->info());
218  ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
219  INEKernel::configure(win_config.second);
220 
221  // Check if we need to clamp the result using min and max
222  const bool is_bounded_relu = !(min <= -128 && max >= 127);
223  _func = is_bounded_relu ? &NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel::run<true> : &NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel::run<false>;
224 }
225 
227 {
229  ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, bias, output, min, max));
230  ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), output->clone().get()).first);
231 
232  return Status{};
233 }
234 
236 {
240 
241  (this->*_func)(window);
242 }
243 } // namespace arm_compute
void run(const Window &window, const ThreadInfo &info) override
Execute the kernel on the passed window.
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
#define ARM_COMPUTE_RETURN_ON_ERROR(status)
Checks if a status contains an error and returns it.
Definition: Error.h:204
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 CPU tensor.
Definition: ITensor.h:36
Copyright (c) 2017-2021 Arm Limited.
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
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 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 configure(const ITensor *input, const ITensor *bias, ITensor *output, int result_fixedpoint_multiplier, int result_shift, int result_offset_after_shift, int min=0, int max=0)
Initialise the kernel's input and output.
void set(size_t dimension, const Dimension &dim)
Set the values of a given dimension.
Definition: Window.inl:49
#define ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(k)
Definition: Validate.h:915
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...
static constexpr size_t DimY
Alias for dimension 1 also known as Y dimension.
Definition: Window.h:45
ScaleKernelInfo info(interpolation_policy, default_border_mode, PixelValue(), sampling_policy, false)
Information about executing thread and CPU.
Definition: CPPTypes.h:252
#define ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(...)
Definition: Validate.h:439
static constexpr size_t DimZ
Alias for dimension 2 also known as Z dimension.
Definition: Window.h:47
#define ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(...)
Definition: Validate.h:541
#define ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(t, c,...)
Definition: Validate.h:788
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:157
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
quantized, asymmetric fixed-point 8-bit number signed
constexpr int end() const
Return the end of the dimension.
Definition: Window.h:99
constexpr int start() const
Return the start of the dimension.
Definition: Window.h:94
Describe a multidimensional execution window.
Definition: Window.h:39
wrapper::traits::neon_vector< T, 16 >::type finalize_quantization(int32x4x4_t &in_s32, int32x4_t result_shift_s32, typename wrapper::traits::neon_vector< T, 16 >::type min, typename wrapper::traits::neon_vector< T, 16 >::type max)
#define ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(f, s)
Definition: Validate.h:201
constexpr const Dimension & x() const
Alias to access the first dimension of the window.
Definition: Window.h:145