1 /*
2 * Copyright (c) 2017-2020 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/NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel.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);
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 inizialitation if not yet initialized
73 auto_init_if_empty(*output, input->clone()->set_data_type(DataType::QASYMM8));
74
75 // Configure kernel window
76 Window win = calculate_max_window(*input, Steps());
77
78 // NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel 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 namespace arm_compute
88 {
89 class Coordinates;
90 } // namespace arm_compute
91
92 template <bool is_bounded_relu>
run(const Window & window)93 void NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::run(const Window &window)
94 {
95 const int32x4_t result_offset_after_shift_s32 = vdupq_n_s32(_result_offset_after_shift);
96 const uint8x16_t min_u8 = vdupq_n_u8(static_cast<uint8_t>(_min));
97 const uint8x16_t max_u8 = vdupq_n_u8(static_cast<uint8_t>(_max));
98
99 ARM_COMPUTE_UNUSED(min_u8);
100 ARM_COMPUTE_UNUSED(max_u8);
101
102 const int window_step_x = 16;
103 const auto window_start_x = static_cast<int>(window.x().start());
104 const auto window_end_x = static_cast<int>(window.x().end());
105
106 Window win_collapsed = window.collapse_if_possible(window, Window::DimZ);
107 win_collapsed.set(Window::DimX, Window::Dimension(0, 1, 1));
108
109 Iterator in(_input, win_collapsed);
110 Iterator out(_output, win_collapsed);
111 if(_bias != nullptr)
112 {
113 Window win_biases;
114 win_biases.set(Window::DimX, Window::Dimension(0, 1, 1));
115 win_biases.set(Window::DimY, Window::Dimension(0, 1, 1));
116
117 Iterator bias(_bias, win_biases);
118 execute_window_loop(win_collapsed, [&](const Coordinates &)
119 {
120 // Compute 16 elements per iteration
121 int x = window_start_x;
122 for(; x <= (window_end_x - window_step_x); x += window_step_x)
123 {
124 int32x4x4_t in_s32 =
125 {
126 {
127 vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 0),
128 vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 4),
129 vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 8),
130 vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 12)
131 }
132 };
133
134 const int32x4x4_t bias_s32 =
135 {
136 {
137 vld1q_s32(reinterpret_cast<const int32_t *>(bias.ptr()) + x + 0),
138 vld1q_s32(reinterpret_cast<const int32_t *>(bias.ptr()) + x + 4),
139 vld1q_s32(reinterpret_cast<const int32_t *>(bias.ptr()) + x + 8),
140 vld1q_s32(reinterpret_cast<const int32_t *>(bias.ptr()) + x + 12)
141 }
142 };
143
144 // Add the bias to GEMM's result
145 in_s32.val[0] = vaddq_s32(in_s32.val[0], bias_s32.val[0]);
146 in_s32.val[1] = vaddq_s32(in_s32.val[1], bias_s32.val[1]);
147 in_s32.val[2] = vaddq_s32(in_s32.val[2], bias_s32.val[2]);
148 in_s32.val[3] = vaddq_s32(in_s32.val[3], bias_s32.val[3]);
149
150 vst1q_u8(out.ptr() + x, finalize_quantization(in_s32, _result_fixedpoint_multiplier, _result_shift, result_offset_after_shift_s32, min_u8, max_u8, is_bounded_relu));
151 }
152
153 // Compute left-over elements
154 for(; x < window_end_x; ++x)
155 {
156 const int32_t bias_value = *(reinterpret_cast<const int32_t *>(bias.ptr()) + x);
157 int32_t in_value = *(reinterpret_cast<const int32_t *>(in.ptr()) + x);
158
159 // Add bias
160 in_value += bias_value;
161 // Finalize and store the result
162 *(out.ptr() + x) = finalize_quantization(in_value, _result_fixedpoint_multiplier, _result_shift, _result_offset_after_shift, static_cast<uint8_t>(_min), static_cast<uint8_t>(_max), is_bounded_relu);
163 }
164 },
165 in, out, bias);
166 }
167 else
168 {
169 execute_window_loop(win_collapsed, [&](const Coordinates &)
170 {
171 // Compute 16 elements per iteration
172 int x = window_start_x;
173 for(; x <= (window_end_x - window_step_x); x += window_step_x)
174 {
175 int32x4x4_t in_s32 =
176 {
177 {
178 vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 0),
179 vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 4),
180 vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 8),
181 vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 12)
182 }
183 };
184
185 vst1q_u8(out.ptr() + x, finalize_quantization(in_s32, _result_fixedpoint_multiplier, _result_shift, result_offset_after_shift_s32, min_u8, max_u8, is_bounded_relu));
186 }
187
188 // Compute left-over elements
189 for(; x < window_end_x; ++x)
190 {
191 const int32_t in_value = *(reinterpret_cast<const int32_t *>(in.ptr()) + x);
192
193 // Finalize and store the result
194 *(out.ptr() + x) = finalize_quantization(in_value, _result_fixedpoint_multiplier, _result_shift, _result_offset_after_shift, static_cast<uint8_t>(_min), static_cast<uint8_t>(_max), is_bounded_relu);
195 }
196 },
197 in, out);
198 }
199 }
200
NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel()201 NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel()
202 : _func(nullptr), _input(nullptr), _bias(nullptr), _output(nullptr), _result_fixedpoint_multiplier(0), _result_shift(0), _result_offset_after_shift(0), _min(0), _max(0)
203 {
204 }
205
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)206 void NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::configure(const ITensor *input, const ITensor *bias, ITensor *output, int result_fixedpoint_multiplier, int result_shift,
207 int result_offset_after_shift, int min, int max)
208 {
209 // Perform validate step
210 ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
211 ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), (bias != nullptr) ? bias->info() : nullptr, output->info(), min, max));
212
213 _input = input;
214 _bias = bias;
215 _output = output;
216 _result_fixedpoint_multiplier = result_fixedpoint_multiplier;
217 _result_shift = result_shift;
218 _result_offset_after_shift = result_offset_after_shift;
219 _min = min;
220 _max = max;
221
222 // Configure kernel window
223 auto win_config = validate_and_configure_window(input->info(), output->info());
224 ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
225 INEKernel::configure(win_config.second);
226
227 // Check if we need to clamp the result using min and max
228 const bool is_bounded_relu = !(min <= 0 && max >= 255);
229 _func = is_bounded_relu ? &NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::run<true> : &NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::run<false>;
230 }
231
validate(const ITensorInfo * input,const ITensorInfo * bias,const ITensorInfo * output,int min,int max)232 Status NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, int min, int max)
233 {
234 ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
235 ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, bias, output, min, max));
236 ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), output->clone().get()).first);
237
238 return Status{};
239 }
240
run(const Window & window,const ThreadInfo & info)241 void NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointKernel::run(const Window &window, const ThreadInfo &info)
242 {
243 ARM_COMPUTE_UNUSED(info);
244 ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
245 ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window);
246
247 (this->*_func)(window);
248 }
249 } // namespace arm_compute
250