Compute Library
 21.02
qasymm8_signed.cpp
Go to the documentation of this file.
1 /*
2  * Copyright (c) 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  */
26 #include "arm_compute/core/Types.h"
30 
31 namespace arm_compute
32 {
33 namespace cpu
34 {
35 void sub_qasymm8_signed_neon(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window)
36 {
37  ARM_COMPUTE_UNUSED(policy);
38 
39  // Create input windows
40  Window input1_win = window.broadcast_if_dimension_le_one(src0->info()->tensor_shape());
41  Window input2_win = window.broadcast_if_dimension_le_one(src1->info()->tensor_shape());
42 
43  // Clear X Dimension on execution window as we handle manually
44  Window win = window;
45  win.set(Window::DimX, Window::Dimension(0, 1, 1));
46 
47  const int window_step_x = 16;
48  const auto window_start_x = static_cast<int>(window.x().start());
49  const auto window_end_x = static_cast<int>(window.x().end());
50  const bool is_broadcast_across_x = src0->info()->tensor_shape().x() != src1->info()->tensor_shape().x();
51 
52  const UniformQuantizationInfo iq1_info = src0->info()->quantization_info().uniform();
53  const UniformQuantizationInfo iq2_info = src1->info()->quantization_info().uniform();
54  const UniformQuantizationInfo oq_info = dst->info()->quantization_info().uniform();
55 
56  const float32x4_t invvscaleo = vdupq_n_f32(1.f / oq_info.scale);
57  const float32x4_t voffseto = vdupq_n_f32(oq_info.offset);
58 
59  if(is_broadcast_across_x)
60  {
61  const bool is_broadcast_input_2 = input2_win.x().step() == 0;
62  Window broadcast_win = is_broadcast_input_2 ? input2_win : input1_win;
63  Window non_broadcast_win = !is_broadcast_input_2 ? input2_win : input1_win;
64  const ITensor *broadcast_tensor = is_broadcast_input_2 ? src1 : src0;
65  const ITensor *non_broadcast_tensor = !is_broadcast_input_2 ? src1 : src0;
66  const UniformQuantizationInfo broadcast_qinfo = broadcast_tensor->info()->quantization_info().uniform();
67  const UniformQuantizationInfo non_broadcast_qinfo = non_broadcast_tensor->info()->quantization_info().uniform();
68  const float32x4_t vscale1 = is_broadcast_input_2 ? vdupq_n_f32(iq1_info.scale) : vdupq_n_f32(iq2_info.scale);
69  const float32x4_t vscale2 = is_broadcast_input_2 ? vdupq_n_f32(iq2_info.scale) : vdupq_n_f32(iq1_info.scale);
70  const int32x4_t voffset1 = is_broadcast_input_2 ? vdupq_n_s32(iq1_info.offset) : vdupq_n_s32(iq2_info.offset);
71  const int32x4_t voffset2 = is_broadcast_input_2 ? vdupq_n_s32(iq2_info.offset) : vdupq_n_s32(iq1_info.offset);
72 
73  // Clear X Dimension on execution window as we handle manually
74  non_broadcast_win.set(Window::DimX, Window::Dimension(0, 1, 1));
75 
76  Iterator broadcast_input(broadcast_tensor, broadcast_win);
77  Iterator non_broadcast_input(non_broadcast_tensor, non_broadcast_win);
78  Iterator output(dst, win);
79 
80  execute_window_loop(win, [&](const Coordinates &)
81  {
82  const auto non_broadcast_input_ptr = reinterpret_cast<const int8_t *>(non_broadcast_input.ptr());
83  const auto output_ptr = reinterpret_cast<int8_t *>(output.ptr());
84 
85  const auto broadcast_value = *reinterpret_cast<const int8_t *>(broadcast_input.ptr());
86  const auto broadcast_value_vec = wrapper::vdup_n(static_cast<int8_t>(broadcast_value), wrapper::traits::vector_128_tag{});
87 
88  const float32x4x4_t bf =
89  {
90  {
91  vmulq_f32(vcvtq_f32_s32(vsubq_s32(wrapper::vreinterpret(wrapper::vmovl(wrapper::vgetlow(wrapper::vmovl(wrapper::vgetlow(broadcast_value_vec))))), voffset2)), vscale2),
92  vmulq_f32(vcvtq_f32_s32(vsubq_s32(wrapper::vreinterpret(wrapper::vmovl(wrapper::vgethigh(wrapper::vmovl(wrapper::vgetlow(broadcast_value_vec))))), voffset2)), vscale2),
93  vmulq_f32(vcvtq_f32_s32(vsubq_s32(wrapper::vreinterpret(wrapper::vmovl(wrapper::vgetlow(wrapper::vmovl(wrapper::vgethigh(broadcast_value_vec))))), voffset2)), vscale2),
94  vmulq_f32(vcvtq_f32_s32(vsubq_s32(wrapper::vreinterpret(wrapper::vmovl(wrapper::vgethigh(wrapper::vmovl(wrapper::vgethigh(broadcast_value_vec))))), voffset2)), vscale2),
95  }
96  };
97 
98  // Compute S elements per iteration
99  int x = window_start_x;
100  for(; x <= (window_end_x - window_step_x); x += window_step_x)
101  {
102  const auto a = wrapper::vloadq(non_broadcast_input_ptr + x);
103 
104  const float32x4x4_t af =
105  {
106  {
107  vmulq_f32(vcvtq_f32_s32(vsubq_s32(wrapper::vreinterpret(wrapper::vmovl(wrapper::vgetlow(wrapper::vmovl(wrapper::vgetlow(a))))), voffset1)), vscale1),
108  vmulq_f32(vcvtq_f32_s32(vsubq_s32(wrapper::vreinterpret(wrapper::vmovl(wrapper::vgethigh(wrapper::vmovl(wrapper::vgetlow(a))))), voffset1)), vscale1),
109  vmulq_f32(vcvtq_f32_s32(vsubq_s32(wrapper::vreinterpret(wrapper::vmovl(wrapper::vgetlow(wrapper::vmovl(wrapper::vgethigh(a))))), voffset1)), vscale1),
110  vmulq_f32(vcvtq_f32_s32(vsubq_s32(wrapper::vreinterpret(wrapper::vmovl(wrapper::vgethigh(wrapper::vmovl(wrapper::vgethigh(a))))), voffset1)), vscale1),
111  }
112  };
113 
114  const int32x4x4_t rf =
115  {
116  {
117 #ifdef __aarch64_
118  vcvtnq_s32_f32(vmlaq_f32(voffseto, !is_broadcast_input_2 ? vsubq_f32(bf.val[0], af.val[0]) : vsubq_f32(af.val[0], bf.val[0]), invvscaleo)),
119  vcvtnq_s32_f32(vmlaq_f32(voffseto, !is_broadcast_input_2 ? vsubq_f32(bf.val[1], af.val[1]) : vsubq_f32(af.val[1], bf.val[1]), invvscaleo)),
120  vcvtnq_s32_f32(vmlaq_f32(voffseto, !is_broadcast_input_2 ? vsubq_f32(bf.val[2], af.val[2]) : vsubq_f32(af.val[2], bf.val[2]), invvscaleo)),
121  vcvtnq_s32_f32(vmlaq_f32(voffseto, !is_broadcast_input_2 ? vsubq_f32(bf.val[3], af.val[3]) : vsubq_f32(af.val[3], bf.val[3]), invvscaleo)),
122 #else //__aarch64__
123  vcvtq_s32_f32(vmlaq_f32(voffseto, !is_broadcast_input_2 ? vsubq_f32(bf.val[0], af.val[0]) : vsubq_f32(af.val[0], bf.val[0]), invvscaleo)),
124  vcvtq_s32_f32(vmlaq_f32(voffseto, !is_broadcast_input_2 ? vsubq_f32(bf.val[1], af.val[1]) : vsubq_f32(af.val[1], bf.val[1]), invvscaleo)),
125  vcvtq_s32_f32(vmlaq_f32(voffseto, !is_broadcast_input_2 ? vsubq_f32(bf.val[2], af.val[2]) : vsubq_f32(af.val[2], bf.val[2]), invvscaleo)),
126  vcvtq_s32_f32(vmlaq_f32(voffseto, !is_broadcast_input_2 ? vsubq_f32(bf.val[3], af.val[3]) : vsubq_f32(af.val[3], bf.val[3]), invvscaleo)),
127 #endif //__aarch64__
128  }
129  };
130 
131  const auto pa = vqmovn_s16(vcombine_s16(vqmovn_s32(rf.val[0]), vqmovn_s32(rf.val[1])));
132  const auto pb = vqmovn_s16(vcombine_s16(vqmovn_s32(rf.val[2]), vqmovn_s32(rf.val[3])));
133  wrapper::vstore(output_ptr + x, wrapper::vcombine(pa, pb));
134  }
135 
136  // Compute left-over elements
137  for(; x < window_end_x; ++x)
138  {
139  const float afs = static_cast<int32_t>(*(non_broadcast_input_ptr + x) - non_broadcast_qinfo.offset) * non_broadcast_qinfo.scale;
140  const float bfs = static_cast<int32_t>(broadcast_value - broadcast_qinfo.offset) * broadcast_qinfo.scale;
141  *(output_ptr + x) = quantize_qasymm8_signed(is_broadcast_input_2 ? afs - bfs : bfs - afs, dst->info()->quantization_info());
142  }
143  },
144  broadcast_input, non_broadcast_input, output);
145  }
146  else
147  {
148  const float32x4_t vscale1 = vdupq_n_f32(iq1_info.scale);
149  const float32x4_t vscale2 = vdupq_n_f32(iq2_info.scale);
150  const int32x4_t voffset1 = vdupq_n_s32(iq1_info.offset);
151  const int32x4_t voffset2 = vdupq_n_s32(iq2_info.offset);
152 
153  // Clear X Dimension on execution window as we handle manually
154  input1_win.set(Window::DimX, Window::Dimension(0, 1, 1));
155  input2_win.set(Window::DimX, Window::Dimension(0, 1, 1));
156 
157  Iterator input1(src0, input1_win);
158  Iterator input2(src1, input2_win);
159  Iterator output(dst, win);
160 
161  execute_window_loop(win, [&](const Coordinates &)
162  {
163  const auto input1_ptr = reinterpret_cast<const int8_t *>(input1.ptr());
164  const auto input2_ptr = reinterpret_cast<const int8_t *>(input2.ptr());
165  const auto output_ptr = reinterpret_cast<int8_t *>(output.ptr());
166 
167  // Compute S elements per iteration
168  int x = window_start_x;
169  for(; x <= (window_end_x - window_step_x); x += window_step_x)
170  {
171  const auto a = wrapper::vloadq(input1_ptr + x);
172  const auto b = wrapper::vloadq(input2_ptr + x);
173 
174  const float32x4x4_t af =
175  {
176  {
177  vmulq_f32(vcvtq_f32_s32(vsubq_s32(wrapper::vreinterpret(wrapper::vmovl(wrapper::vgetlow(wrapper::vmovl(wrapper::vgetlow(a))))), voffset1)), vscale1),
178  vmulq_f32(vcvtq_f32_s32(vsubq_s32(wrapper::vreinterpret(wrapper::vmovl(wrapper::vgethigh(wrapper::vmovl(wrapper::vgetlow(a))))), voffset1)), vscale1),
179  vmulq_f32(vcvtq_f32_s32(vsubq_s32(wrapper::vreinterpret(wrapper::vmovl(wrapper::vgetlow(wrapper::vmovl(wrapper::vgethigh(a))))), voffset1)), vscale1),
180  vmulq_f32(vcvtq_f32_s32(vsubq_s32(wrapper::vreinterpret(wrapper::vmovl(wrapper::vgethigh(wrapper::vmovl(wrapper::vgethigh(a))))), voffset1)), vscale1),
181  }
182  };
183 
184  const float32x4x4_t bf =
185  {
186  {
187  vmulq_f32(vcvtq_f32_s32(vsubq_s32(wrapper::vreinterpret(wrapper::vmovl(wrapper::vgetlow(wrapper::vmovl(wrapper::vgetlow(b))))), voffset2)), vscale2),
188  vmulq_f32(vcvtq_f32_s32(vsubq_s32(wrapper::vreinterpret(wrapper::vmovl(wrapper::vgethigh(wrapper::vmovl(wrapper::vgetlow(b))))), voffset2)), vscale2),
189  vmulq_f32(vcvtq_f32_s32(vsubq_s32(wrapper::vreinterpret(wrapper::vmovl(wrapper::vgetlow(wrapper::vmovl(wrapper::vgethigh(b))))), voffset2)), vscale2),
190  vmulq_f32(vcvtq_f32_s32(vsubq_s32(wrapper::vreinterpret(wrapper::vmovl(wrapper::vgethigh(wrapper::vmovl(wrapper::vgethigh(b))))), voffset2)), vscale2),
191  }
192  };
193 
194  const int32x4x4_t rf =
195  {
196  {
197 #ifdef __aarch64__
198  vcvtnq_s32_f32(vmlaq_f32(voffseto, vsubq_f32(af.val[0], bf.val[0]), invvscaleo)),
199  vcvtnq_s32_f32(vmlaq_f32(voffseto, vsubq_f32(af.val[1], bf.val[1]), invvscaleo)),
200  vcvtnq_s32_f32(vmlaq_f32(voffseto, vsubq_f32(af.val[2], bf.val[2]), invvscaleo)),
201  vcvtnq_s32_f32(vmlaq_f32(voffseto, vsubq_f32(af.val[3], bf.val[3]), invvscaleo)),
202 #else //__aarch64__
203  vcvtq_s32_f32(vmlaq_f32(voffseto, vsubq_f32(af.val[0], bf.val[0]), invvscaleo)),
204  vcvtq_s32_f32(vmlaq_f32(voffseto, vsubq_f32(af.val[1], bf.val[1]), invvscaleo)),
205  vcvtq_s32_f32(vmlaq_f32(voffseto, vsubq_f32(af.val[2], bf.val[2]), invvscaleo)),
206  vcvtq_s32_f32(vmlaq_f32(voffseto, vsubq_f32(af.val[3], bf.val[3]), invvscaleo)),
207 #endif //__aarch64__
208  }
209  };
210 
211  const auto pa = vqmovn_s16(vcombine_s16(vqmovn_s32(rf.val[0]), vqmovn_s32(rf.val[1])));
212  const auto pb = vqmovn_s16(vcombine_s16(vqmovn_s32(rf.val[2]), vqmovn_s32(rf.val[3])));
213  wrapper::vstore(output_ptr + x, wrapper::vcombine(pa, pb));
214  }
215 
216  // Compute left-over elements
217  for(; x < window_end_x; ++x)
218  {
219  const float afs = static_cast<int32_t>((*(input1_ptr + x)) - iq1_info.offset) * iq1_info.scale;
220  const float bfs = static_cast<int32_t>((*(input2_ptr + x)) - iq2_info.offset) * iq2_info.scale;
221 
222  *(output_ptr + x) = quantize_qasymm8_signed((afs - bfs), dst->info()->quantization_info());
223  }
224  },
225  input1, input2, output);
226  }
227 }
228 } // namespace cpu
229 } // namespace arm_compute
SimpleTensor< float > b
Definition: DFT.cpp:157
constexpr int step() const
Return the step of the dimension.
Definition: Window.h:104
uint8x16_t vloadq(const uint8_t *ptr)
Definition: load.h:58
Quantization info when assuming per layer quantization.
Describe one of the image&#39;s dimensions with a start, end and step.
Definition: Window.h:77
Interface for Neon tensor.
Definition: ITensor.h:36
Copyright (c) 2017-2021 Arm Limited.
T x() const
Alias to access the size of the first dimension.
Definition: Dimensions.h:87
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
virtual const TensorShape & tensor_shape() const =0
Size for each dimension of the tensor.
int8_t quantize_qasymm8_signed(float value, const INFO_TYPE &qinfo, RoundingPolicy rounding_policy=RoundingPolicy::TO_NEAREST_UP)
Quantize a value given a signed 8-bit asymmetric quantization scheme.
int16x4_t vreinterpret(const uint16x4_t &a)
Definition: reinterpret.h:44
Coordinates of an item.
Definition: Coordinates.h:37
UniformQuantizationInfo uniform() const
Return per layer quantization info.
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
uint8x8_t vgetlow(const uint8x16_t val)
Definition: getlow.h:39
void set(size_t dimension, const Dimension &dim)
Set the values of a given dimension.
Definition: Window.inl:49
uint8x16_t vcombine(const uint8x8_t &a, const uint8x8_t &b)
Definition: combine.h:39
Window broadcast_if_dimension_le_one(const TensorShape &shape) const
Don&#39;t advance in the dimension where shape is less equal to 1.
Definition: Window.inl:120
virtual QuantizationInfo quantization_info() const =0
Get the quantization settings (scale and offset) of the tensor.
uint8x8_t vgethigh(const uint8x16_t val)
Definition: gethigh.h:39
void sub_qasymm8_signed_neon(const ITensor *src0, const ITensor *src1, ITensor *dst, const ConvertPolicy &policy, const Window &window)
void vstore(uint8_t *ptr, uint8x8_t val)
Definition: store.h:39
uint8x8_t vdup_n(uint8_t value, traits::vector_64_tag)
Definition: dup_n.h:41
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
std::vector< NodeID > bfs(Graph &g)
Breadth first search traversal.
constexpr int end() const
Return the end of the dimension.
Definition: Window.h:99
Iterator updated by execute_window_loop for each window element.
Definition: Helpers.h:46
uint16x8_t vmovl(const uint8x8_t &a)
Definition: movl.h:39
constexpr int start() const
Return the start of the dimension.
Definition: Window.h:94
Describe a multidimensional execution window.
Definition: Window.h:39
ConvertPolicy
Policy to handle overflow.
Definition: Types.h:385
constexpr const Dimension & x() const
Alias to access the first dimension of the window.
Definition: Window.h:145