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 */
24 #include "src/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel.h"
25
26 #include "arm_compute/core/Error.h"
27 #include "arm_compute/core/Helpers.h"
28 #include "arm_compute/core/ITensor.h"
29 #include "arm_compute/core/TensorInfo.h"
30 #include "arm_compute/core/Types.h"
31 #include "arm_compute/core/Utils.h"
32 #include "arm_compute/core/Validate.h"
33 #include "arm_compute/core/Window.h"
34 #include "arm_compute/core/utils/misc/ShapeCalculator.h"
35 #include "src/core/AccessWindowStatic.h"
36 #include "src/core/NEON/NEAsymm.h"
37 #include "src/core/helpers/AutoConfiguration.h"
38 #include "src/core/helpers/WindowHelpers.h"
39
40 #include <arm_neon.h>
41 #include <cstddef>
42 #include <cstdint>
43
44 namespace arm_compute
45 {
46 namespace
47 {
validate_arguments(const ITensorInfo * input,const ITensorInfo * bias,const ITensorInfo * output,int min,int max)48 Status validate_arguments(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, int min, int max)
49 {
50 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::S32);
51 ARM_COMPUTE_RETURN_ERROR_ON(min > max);
52
53 // Check biases if exist
54 if(bias != nullptr)
55 {
56 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, bias);
57 ARM_COMPUTE_RETURN_ERROR_ON(bias->num_dimensions() > 1);
58 ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(0) != bias->dimension(0));
59 }
60
61 if(output->total_size() != 0)
62 {
63 ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::QASYMM8_SIGNED);
64 ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(output, input);
65 }
66
67 return Status{};
68 }
69
validate_and_configure_window(ITensorInfo * input,ITensorInfo * output)70 std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *output)
71 {
72 // Output auto initialization if not yet initialized
73 auto_init_if_empty(*output, input->clone()->set_data_type(DataType::QASYMM8_SIGNED));
74
75 // Configure kernel window
76 Window win = calculate_max_window(*input, Steps());
77
78 // NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel doesn't need padding so update_window_and_padding() can be skipped
79 Coordinates coord;
80 coord.set_num_dimensions(output->num_dimensions());
81 output->set_valid_region(ValidRegion(coord, output->tensor_shape()));
82
83 return std::make_pair(Status{}, win);
84 }
85 } // namespace
86
87 template <bool is_bounded_relu>
run(const Window & window)88 void NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel::run(const Window &window)
89 {
90 const int32x4_t result_offset_after_shift_s32 = vdupq_n_s32(_result_offset_after_shift);
91 const int8x16_t min_s8 = vdupq_n_s8(static_cast<int8_t>(_min));
92 const int8x16_t max_s8 = vdupq_n_s8(static_cast<int8_t>(_max));
93
94 ARM_COMPUTE_UNUSED(min_s8, max_s8);
95
96 const int window_step_x = 16;
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 int32x4x4_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 vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 8),
124 vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 12)
125 }
126 };
127
128 const int32x4x4_t bias_s32 =
129 {
130 {
131 vld1q_s32(reinterpret_cast<const int32_t *>(bias.ptr()) + x + 0),
132 vld1q_s32(reinterpret_cast<const int32_t *>(bias.ptr()) + x + 4),
133 vld1q_s32(reinterpret_cast<const int32_t *>(bias.ptr()) + x + 8),
134 vld1q_s32(reinterpret_cast<const int32_t *>(bias.ptr()) + x + 12)
135 }
136 };
137
138 // Add the bias to GEMM's result
139 in_s32.val[0] = vaddq_s32(in_s32.val[0], bias_s32.val[0]);
140 in_s32.val[1] = vaddq_s32(in_s32.val[1], bias_s32.val[1]);
141 in_s32.val[2] = vaddq_s32(in_s32.val[2], bias_s32.val[2]);
142 in_s32.val[3] = vaddq_s32(in_s32.val[3], bias_s32.val[3]);
143
144 vst1q_s8(reinterpret_cast<int8_t *>(out.ptr() + x),
145 finalize_quantization(in_s32, _result_fixedpoint_multiplier, _result_shift, result_offset_after_shift_s32, min_s8, max_s8, is_bounded_relu));
146 }
147
148 // Compute left-over elements
149 for(; x < window_end_x; ++x)
150 {
151 const int32_t bias_value = *(reinterpret_cast<const int32_t *>(bias.ptr()) + x);
152 int32_t in_value = *(reinterpret_cast<const int32_t *>(in.ptr()) + x);
153
154 // Add bias
155 in_value += bias_value;
156 // Finalize and store the result
157 *reinterpret_cast<int8_t *>(out.ptr() + x) = finalize_quantization(in_value, _result_fixedpoint_multiplier, _result_shift, _result_offset_after_shift,
158 static_cast<int8_t>(_min), static_cast<int8_t>(_max), is_bounded_relu);
159 }
160 },
161 in, out, bias);
162 }
163 else
164 {
165 execute_window_loop(win_collapsed, [&](const Coordinates &)
166 {
167 // Compute 16 elements per iteration
168 int x = window_start_x;
169 for(; x <= (window_end_x - window_step_x); x += window_step_x)
170 {
171 int32x4x4_t in_s32 =
172 {
173 {
174 vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 0),
175 vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 4),
176 vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 8),
177 vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 12)
178 }
179 };
180
181 vst1q_s8(reinterpret_cast<int8_t *>(out.ptr() + x),
182 finalize_quantization(in_s32, _result_fixedpoint_multiplier, _result_shift, result_offset_after_shift_s32, min_s8, max_s8, is_bounded_relu));
183 }
184
185 // Compute left-over elements
186 for(; x < window_end_x; ++x)
187 {
188 const int32_t in_value = *(reinterpret_cast<const int32_t *>(in.ptr()) + x);
189
190 // Finalize and store the result
191 *reinterpret_cast<int8_t *>(out.ptr() + x) = finalize_quantization(in_value, _result_fixedpoint_multiplier, _result_shift, _result_offset_after_shift,
192 static_cast<int8_t>(_min), static_cast<int8_t>(_max), is_bounded_relu);
193 }
194 },
195 in, out);
196 }
197 }
198
NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel()199 NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel::NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel()
200 : _func(nullptr), _input(nullptr), _bias(nullptr), _output(nullptr), _result_fixedpoint_multiplier(0), _result_shift(0), _result_offset_after_shift(0), _min(0), _max(0)
201 {
202 }
203
configure(const ITensor * input,const ITensor * bias,ITensor * output,int result_fixedpoint_multiplier,int result_shift,int result_offset_after_shift,int min,int max)204 void NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel::configure(const ITensor *input, const ITensor *bias, ITensor *output, int result_fixedpoint_multiplier, int result_shift,
205 int result_offset_after_shift, int min, int max)
206 {
207 // Perform validate step
208 ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
209 ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), (bias != nullptr) ? bias->info() : nullptr, output->info(), min, max));
210
211 _input = input;
212 _bias = bias;
213 _output = output;
214 _result_fixedpoint_multiplier = result_fixedpoint_multiplier;
215 _result_shift = result_shift;
216 _result_offset_after_shift = result_offset_after_shift;
217 _min = min;
218 _max = max;
219
220 // Configure kernel window
221 auto win_config = validate_and_configure_window(input->info(), output->info());
222 ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
223 INEKernel::configure(win_config.second);
224
225 // Check if we need to clamp the result using min and max
226 const bool is_bounded_relu = !(min <= -128 && max >= 127);
227 _func = is_bounded_relu ? &NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel::run<true> : &NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel::run<false>;
228 }
229
validate(const ITensorInfo * input,const ITensorInfo * bias,const ITensorInfo * output,int min,int max)230 Status NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel::validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, int min, int max)
231 {
232 ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
233 ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, bias, output, min, max));
234 ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), output->clone().get()).first);
235
236 return Status{};
237 }
238
run(const Window & window,const ThreadInfo & info)239 void NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointKernel::run(const Window &window, const ThreadInfo &info)
240 {
241 ARM_COMPUTE_UNUSED(info);
242 ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
243 ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window);
244
245 (this->*_func)(window);
246 }
247 } // namespace arm_compute
248