Compute Library
 22.11
CpuGemmLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel.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/NESymm.h"
38 
39 #include <arm_neon.h>
40 
41 namespace arm_compute
42 {
43 namespace cpu
44 {
45 namespace kernels
46 {
47 namespace
48 {
49 Status validate_arguments(const ITensorInfo *src, const ITensorInfo *bias, const ITensorInfo *dst, int min, int max)
50 {
53  ARM_COMPUTE_RETURN_ERROR_ON(min > max);
54 
55  // Check biases if exist
56  if(bias != nullptr)
57  {
59  ARM_COMPUTE_RETURN_ERROR_ON(bias->num_dimensions() > 1);
60  ARM_COMPUTE_RETURN_ERROR_ON(src->dimension(0) != bias->dimension(0));
61  }
62 
63  if(dst->total_size() != 0)
64  {
67  }
68 
69  return Status{};
70 }
71 } // namespace
72 
73 template <bool is_bounded_relu>
74 void CpuGemmLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel::run_internal(const ITensor *src, const ITensor *bias, ITensor *dst, const Window &window)
75 {
76  const int16x8_t min_s16 = vdupq_n_s16(static_cast<int16_t>(_min));
77  const int16x8_t max_s16 = vdupq_n_s16(static_cast<int16_t>(_max));
78 
79  ARM_COMPUTE_UNUSED(min_s16);
80  ARM_COMPUTE_UNUSED(max_s16);
81 
82  const int window_step_x = 8;
83  const auto window_start_x = static_cast<int>(window.x().start());
84  const auto window_end_x = static_cast<int>(window.x().end());
85 
86  Window win_collapsed = window.collapse_if_possible(window, Window::DimZ);
87  win_collapsed.set(Window::DimX, Window::Dimension(0, 1, 1));
88 
89  Iterator in(src, win_collapsed);
90  Iterator out(dst, win_collapsed);
91  if(bias != nullptr)
92  {
93  Window win_biases;
94  win_biases.set(Window::DimX, Window::Dimension(0, 1, 1));
95  win_biases.set(Window::DimY, Window::Dimension(0, 1, 1));
96 
97  Iterator bias_i(bias, win_biases);
98  execute_window_loop(win_collapsed, [&](const Coordinates &)
99  {
100  // Compute 16 elements per iteration
101  int x = window_start_x;
102  for(; x <= (window_end_x - window_step_x); x += window_step_x)
103  {
104  int32x4x2_t in_s32 =
105  {
106  {
107  vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 0),
108  vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 4)
109  }
110  };
111 
112  const int32x4x2_t bias_s32 =
113  {
114  {
115  vld1q_s32(reinterpret_cast<const int32_t *>(bias_i.ptr()) + x + 0),
116  vld1q_s32(reinterpret_cast<const int32_t *>(bias_i.ptr()) + x + 4)
117  }
118  };
119 
120  // Add the bias to GEMM's result
121  in_s32.val[0] = vaddq_s32(in_s32.val[0], bias_s32.val[0]);
122  in_s32.val[1] = vaddq_s32(in_s32.val[1], bias_s32.val[1]);
123 
124  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));
125  }
126 
127  // Compute left-over elements
128  for(; x < window_end_x; ++x)
129  {
130  const int32_t bias_value = *(reinterpret_cast<const int32_t *>(bias_i.ptr()) + x);
131  int32_t in_value = *(reinterpret_cast<const int32_t *>(in.ptr()) + x);
132 
133  // Add bias
134  in_value += bias_value;
135  // Finalize and store the result
136  *(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),
137  static_cast<int16_t>(_max));
138  }
139  },
140  in, out, bias_i);
141  }
142  else
143  {
144  execute_window_loop(win_collapsed, [&](const Coordinates &)
145  {
146  // Compute 16 elements per iteration
147  int x = window_start_x;
148  for(; x <= (window_end_x - window_step_x); x += window_step_x)
149  {
150  int32x4x2_t in_s32 =
151  {
152  {
153  vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 0),
154  vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 4)
155  }
156  };
157 
158  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));
159  }
160 
161  // Compute left-over elements
162  for(; x < window_end_x; ++x)
163  {
164  const int32_t in_value = *(reinterpret_cast<const int32_t *>(in.ptr()) + x);
165  ARM_COMPUTE_UNUSED(in_value);
166  // Finalize and store the result
167  *(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),
168  static_cast<int16_t>(_max));
169  }
170  },
171  in, out);
172  }
173 }
174 
175 void CpuGemmLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel::configure(ITensorInfo *src, ITensorInfo *bias, ITensorInfo *dst, int result_fixedpoint_multiplier, int result_shift,
176  int min, int max)
177 {
178  // Perform validate step
179  ARM_COMPUTE_UNUSED(bias, dst);
181  ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(src, bias, dst, min, max));
182 
183  _result_fixedpoint_multiplier = result_fixedpoint_multiplier;
184  _result_shift = result_shift;
185  _min = min;
186  _max = max;
187 
188  // Output auto inizialitation if not yet initialized
189  auto_init_if_empty(*src, src->clone()->set_data_type(DataType::QSYMM16));
190  // Configure kernel window
191  Window win_config = calculate_max_window(*src, Steps());
192  ICpuKernel::configure(win_config);
193 
194  // Check if we need to clamp the result using min and max
195  const bool is_bounded_relu = !(min <= -32768 && max >= 32767);
196  _func = is_bounded_relu ? &CpuGemmLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel::run_internal<true> :
197  &CpuGemmLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel::run_internal<false>;
198 }
199 
201 {
202  ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
203  ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, bias, output, min, max));
204  return Status{};
205 }
206 
208 {
209  ARM_COMPUTE_UNUSED(info);
212  ARM_COMPUTE_ERROR_ON_MSG(tensors.empty(), "No inputs provided");
213 
214  auto src = tensors.get_const_tensor(TensorType::ACL_SRC);
215  auto bias = tensors.get_const_tensor(TensorType::ACL_BIAS);
216  auto dst = tensors.get_tensor(TensorType::ACL_DST);
217 
218  (this->*_func)(src, bias, dst, window);
219 }
220 
222 {
223  return "CpuGemmLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel";
224 }
225 } // namespace kernels
226 } // namespace cpu
227 } // namespace arm_compute
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
quantized, symmetric fixed-point 16-bit number
void configure(ITensorInfo *src, ITensorInfo *bias, ITensorInfo *dst, int result_fixedpoint_multiplier, int result_shift, int min=0, int max=0)
Initialise the kernel&#39;s input and output.
void run_op(ITensorPack &tensors, const Window &window, const ThreadInfo &info) override
Execute the kernel on the passed window.
bool empty() const
Checks if pack is empty.
Definition: ITensorPack.cpp:80
#define ARM_COMPUTE_RETURN_ON_ERROR(status)
Checks if a status contains an error and returns it.
Definition: Error.h:204
Store the tensor&#39;s metadata.
Definition: ITensorInfo.h:40
#define ARM_COMPUTE_ERROR_THROW_ON(status)
Definition: Error.h:455
Status class.
Definition: Error.h:52
Status validate_arguments(const ITensorInfo *src, const ITensorInfo *weights, const ITensorInfo *dst, const PadStrideInfo &conv_info)
#define ARM_COMPUTE_RETURN_ERROR_ON(cond)
If the condition is true, an error is returned.
Definition: Error.h:296
SimpleTensor< float > src
Definition: DFT.cpp:155
Copyright (c) 2017-2022 Arm Limited.
1 channel, 1 S32 per channel
const ITensor * get_const_tensor(int id) const
Get constant tensor of a given id.
Definition: ITensorPack.cpp:54
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
Class to describe a number of elements in each dimension.
Definition: Steps.h:40
#define ARM_COMPUTE_ERROR_ON_MSG(cond, msg)
Definition: Error.h:456
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.
#define ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(k)
Definition: Validate.h:915
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)
ITensor * get_tensor(int id)
Get tensor of a given id from the pac.
Definition: ITensorPack.cpp:64
Information about executing thread and CPU.
Definition: CPPTypes.h:179
#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
Tensor packing service.
Definition: ITensorPack.h:39
static Status validate(const ITensorInfo *src, const ITensorInfo *bias, const ITensorInfo *dst, int min=0, int max=0)
Static function to check if given info will lead to a valid configuration.
#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
Describe a multidimensional execution window.
Definition: Window.h:39
#define ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(f, s)
Definition: Validate.h:201
const int32_t * bias