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/NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel.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/NESymm.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::QSYMM16);
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::QSYMM16));
74
75 // Configure kernel window
76 Window win = calculate_max_window(*input, Steps());
77
78 // NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel 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 class Coordinates;
88
89 template <bool is_bounded_relu>
run(const Window & window)90 void NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel::run(const Window &window)
91 {
92 const int16x8_t min_s16 = vdupq_n_s16(static_cast<int16_t>(_min));
93 const int16x8_t max_s16 = vdupq_n_s16(static_cast<int16_t>(_max));
94
95 ARM_COMPUTE_UNUSED(min_s16);
96 ARM_COMPUTE_UNUSED(max_s16);
97
98 const int window_step_x = 8;
99 const auto window_start_x = static_cast<int>(window.x().start());
100 const auto window_end_x = static_cast<int>(window.x().end());
101
102 Window win_collapsed = window.collapse_if_possible(window, Window::DimZ);
103 win_collapsed.set(Window::DimX, Window::Dimension(0, 1, 1));
104
105 Iterator in(_input, win_collapsed);
106 Iterator out(_output, win_collapsed);
107 if(_bias != nullptr)
108 {
109 Window win_biases;
110 win_biases.set(Window::DimX, Window::Dimension(0, 1, 1));
111 win_biases.set(Window::DimY, Window::Dimension(0, 1, 1));
112
113 Iterator bias(_bias, win_biases);
114 execute_window_loop(win_collapsed, [&](const Coordinates &)
115 {
116 // Compute 16 elements per iteration
117 int x = window_start_x;
118 for(; x <= (window_end_x - window_step_x); x += window_step_x)
119 {
120 int32x4x2_t in_s32 =
121 {
122 {
123 vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 0),
124 vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 4)
125 }
126 };
127
128 const int32x4x2_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 }
134 };
135
136 // Add the bias to GEMM's result
137 in_s32.val[0] = vaddq_s32(in_s32.val[0], bias_s32.val[0]);
138 in_s32.val[1] = vaddq_s32(in_s32.val[1], bias_s32.val[1]);
139
140 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));
141 }
142
143 // Compute left-over elements
144 for(; x < window_end_x; ++x)
145 {
146 const int32_t bias_value = *(reinterpret_cast<const int32_t *>(bias.ptr()) + x);
147 int32_t in_value = *(reinterpret_cast<const int32_t *>(in.ptr()) + x);
148
149 // Add bias
150 in_value += bias_value;
151 // Finalize and store the result
152 *(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),
153 static_cast<int16_t>(_max));
154 }
155 },
156 in, out, bias);
157 }
158 else
159 {
160 execute_window_loop(win_collapsed, [&](const Coordinates &)
161 {
162 // Compute 16 elements per iteration
163 int x = window_start_x;
164 for(; x <= (window_end_x - window_step_x); x += window_step_x)
165 {
166 int32x4x2_t in_s32 =
167 {
168 {
169 vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 0),
170 vld1q_s32(reinterpret_cast<const int32_t *>(in.ptr()) + x + 4)
171 }
172 };
173
174 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));
175 }
176
177 // Compute left-over elements
178 for(; x < window_end_x; ++x)
179 {
180 const int32_t in_value = *(reinterpret_cast<const int32_t *>(in.ptr()) + x);
181 ARM_COMPUTE_UNUSED(in_value);
182 // Finalize and store the result
183 *(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),
184 static_cast<int16_t>(_max));
185 }
186 },
187 in, out);
188 }
189 }
190
NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel()191 NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel::NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel()
192 : _func(nullptr), _input(nullptr), _bias(nullptr), _output(nullptr), _result_fixedpoint_multiplier(0), _result_shift(0), _min(0), _max(0)
193 {
194 }
195
configure(const ITensor * input,const ITensor * bias,ITensor * output,int result_fixedpoint_multiplier,int result_shift,int min,int max)196 void NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel::configure(const ITensor *input, const ITensor *bias, ITensor *output, int result_fixedpoint_multiplier, int result_shift,
197 int min, int max)
198 {
199 // Perform validate step
200 ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
201 ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), (bias != nullptr) ? bias->info() : nullptr, output->info(), min, max));
202
203 _input = input;
204 _bias = bias;
205 _output = output;
206 _result_fixedpoint_multiplier = result_fixedpoint_multiplier;
207 _result_shift = result_shift;
208 _min = min;
209 _max = max;
210
211 // Configure kernel window
212 auto win_config = validate_and_configure_window(input->info(), output->info());
213 ARM_COMPUTE_ERROR_THROW_ON(win_config.first);
214 INEKernel::configure(win_config.second);
215
216 // Check if we need to clamp the result using min and max
217 const bool is_bounded_relu = !(min <= -32768 && max >= 32767);
218 _func = is_bounded_relu ? &NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel::run<true> : &NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel::run<false>;
219 }
220
validate(const ITensorInfo * input,const ITensorInfo * bias,const ITensorInfo * output,int min,int max)221 Status NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel::validate(const ITensorInfo *input, const ITensorInfo *bias, const ITensorInfo *output, int min, int max)
222 {
223 ARM_COMPUTE_ERROR_ON_NULLPTR(input, output);
224 ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, bias, output, min, max));
225 ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), output->clone().get()).first);
226
227 return Status{};
228 }
229
run(const Window & window,const ThreadInfo & info)230 void NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointKernel::run(const Window &window, const ThreadInfo &info)
231 {
232 ARM_COMPUTE_UNUSED(info);
233 ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this);
234 ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(INEKernel::window(), window);
235
236 (this->*_func)(window);
237 }
238 } // namespace arm_compute
239