2 * Copyright (c) 2020 Samsung Electronics Co., Ltd. All Rights Reserved
3 * Copyright 2019 The TensorFlow Authors. All Rights Reserved.
5 * Licensed under the Apache License, Version 2.0 (the "License");
6 * you may not use this file except in compliance with the License.
7 * You may obtain a copy of the License at
9 * http://www.apache.org/licenses/LICENSE-2.0
11 * Unless required by applicable law or agreed to in writing, software
12 * distributed under the License is distributed on an "AS IS" BASIS,
13 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
14 * See the License for the specific language governing permissions and
15 * limitations under the License.
18 #include "kernels/Conv2D.h"
20 #include "kernels/Utils.h"
22 #include <tensorflow/lite/kernels/internal/optimized/legacy_optimized_ops.h>
27 namespace luci_interpreter
32 Conv2D::Conv2D(const Tensor *input, const Tensor *filter, const Tensor *bias, Tensor *output,
33 const Conv2DParams ¶ms)
34 : KernelWithParams<Conv2DParams>({input, filter, bias}, {output}, params)
38 void Conv2D::configure()
40 // TensorFlow Lite (as of v2.2.0) supports the following combinations of types:
41 // | input filter bias output |
42 // ----+---------------------------+
43 // (1) | float float float float |
44 // (2) | float int8 float float | hybrid
45 // (3) | uint8 uint8 int32 uint8 | quantized
46 // (4) | int8 int8 int32 int8 | quantized per channel
48 // We only support (1) and (3) for now, and additionally the following:
49 // | input filter bias output |
50 // ----+---------------------------+
51 // (5) | int16 int16 int64 int16 |
53 if (input()->element_type() == DataType::FLOAT32 && filter()->element_type() == DataType::FLOAT32)
55 LUCI_INTERPRETER_CHECK(bias() == nullptr || bias()->element_type() == DataType::FLOAT32);
57 else if (input()->element_type() == DataType::U8 && filter()->element_type() == DataType::U8)
59 LUCI_INTERPRETER_CHECK(bias() == nullptr || bias()->element_type() == DataType::S32);
61 else if (input()->element_type() == DataType::S16 && filter()->element_type() == DataType::S16)
63 LUCI_INTERPRETER_CHECK(bias() == nullptr || bias()->element_type() == DataType::S64);
67 throw std::runtime_error("Unsupported type.");
69 LUCI_INTERPRETER_CHECK(output()->element_type() == input()->element_type());
71 const Shape &input_shape = input()->shape();
72 const Shape &filter_shape = filter()->shape();
73 LUCI_INTERPRETER_CHECK(input_shape.num_dims() == 4 && filter_shape.num_dims() == 4);
75 const int32_t batches = input_shape.dim(0);
76 const int32_t input_height = input_shape.dim(1);
77 const int32_t input_width = input_shape.dim(2);
78 const int32_t output_depth = filter_shape.dim(0);
79 const int32_t filter_height = filter_shape.dim(1);
80 const int32_t filter_width = filter_shape.dim(2);
81 LUCI_INTERPRETER_CHECK(filter_shape.dim(3) == input_shape.dim(3));
83 LUCI_INTERPRETER_CHECK(bias() == nullptr || (bias()->shape().num_dims() == 1 &&
84 bias()->shape().dim(0) == output_depth));
86 const int32_t output_height =
87 computeOutputSize(_params.padding, input_height, filter_height, _params.stride_height,
88 _params.dilation_height_factor);
89 const int32_t output_width =
90 computeOutputSize(_params.padding, input_width, filter_width, _params.stride_width,
91 _params.dilation_width_factor);
93 _padding_height = computePadding(_params.stride_height, _params.dilation_height_factor,
94 input_height, filter_height, output_height);
95 _padding_width = computePadding(_params.stride_width, _params.dilation_width_factor, input_width,
96 filter_width, output_width);
98 output()->resize({batches, output_height, output_width, output_depth});
100 // Allocate tensor for Im2Col, if needed.
101 // The checks here should be aligned with the actual implementation.
102 const bool need_dilated_im2col =
103 _params.dilation_height_factor != 1 || _params.dilation_width_factor != 1;
104 const bool need_non_dilated_im2col = _params.stride_height != 1 || _params.stride_width != 1 ||
105 filter_height != 1 || filter_width != 1;
106 const bool need_im2col =
107 input()->element_type() != DataType::S16 && (need_dilated_im2col || need_non_dilated_im2col);
110 const int input_depth = input_shape.dim(3);
111 Shape im2col_shape{batches, output_height, output_width,
112 input_depth * filter_height * filter_width};
116 std::make_unique<Tensor>(input()->element_type(), im2col_shape, AffineQuantization{}, "");
118 catch (std::bad_alloc &ba)
120 // Failed memory allocation
126 void Conv2D::execute() const
128 switch (input()->element_type())
130 case DataType::FLOAT32:
131 if (filter()->element_type() == DataType::FLOAT32)
136 throw std::runtime_error("Unsupported type.");
144 throw std::runtime_error("Unsupported type.");
147 _im2col->deallocate();
150 void Conv2D::evalFloat() const
152 float activation_min{};
153 float activation_max{};
154 calculateActivationRange(_params.activation, &activation_min, &activation_max);
156 tflite::ConvParams params{};
157 params.padding_values.height = _padding_height;
158 params.padding_values.width = _padding_width;
159 params.stride_height = _params.stride_height;
160 params.stride_width = _params.stride_width;
161 params.dilation_height_factor = _params.dilation_height_factor;
162 params.dilation_width_factor = _params.dilation_width_factor;
163 params.float_activation_min = activation_min;
164 params.float_activation_max = activation_max;
167 tflite::optimized_ops::Conv(params, getTensorShape(input()), getTensorData<float>(input()),
168 getTensorShape(filter()), getTensorData<float>(filter()),
169 getTensorShape(bias()), getTensorData<float>(bias()),
170 getTensorShape(output()), getTensorData<float>(output()),
171 getTensorShape(_im2col.get()), getTensorData<float>(_im2col.get()));
173 tflite::reference_ops::Conv(
174 params, getTensorShape(input()), getTensorData<float>(input()), getTensorShape(filter()),
175 getTensorData<float>(filter()), getTensorShape(bias()), getTensorData<float>(bias()),
176 getTensorShape(output()), getTensorData<float>(output()), tflite::RuntimeShape(), nullptr);
179 void Conv2D::evalQuantized() const
181 const auto input_scale = static_cast<double>(input()->scale());
182 const auto filter_scale = static_cast<double>(filter()->scale());
183 const auto output_scale = static_cast<double>(output()->scale());
185 const double real_multiplier = input_scale * filter_scale / output_scale;
186 int32_t output_multiplier{};
188 quantizeMultiplier(real_multiplier, &output_multiplier, &output_shift);
190 int32_t activation_min{};
191 int32_t activation_max{};
192 calculateActivationRangeQuantized(_params.activation, output(), &activation_min, &activation_max);
194 tflite::ConvParams params{};
195 params.padding_values.height = _padding_height;
196 params.padding_values.width = _padding_width;
197 params.stride_height = _params.stride_height;
198 params.stride_width = _params.stride_width;
199 params.dilation_height_factor = _params.dilation_height_factor;
200 params.dilation_width_factor = _params.dilation_width_factor;
201 // The kernel expects input and filter zero points to be negated.
202 params.input_offset = -input()->zero_point(); // Note the '-'.
203 params.weights_offset = -filter()->zero_point(); // Note the '-'.
204 params.output_offset = output()->zero_point();
205 params.output_multiplier = output_multiplier;
206 params.output_shift = output_shift;
207 params.quantized_activation_min = activation_min;
208 params.quantized_activation_max = activation_max;
210 // TODO This should only be done once (although it takes only a few microseconds).
211 // Also, the user should be able to adjust the number of threads.
212 auto gemmlowp_context = std::make_unique<gemmlowp::GemmContext>();
213 gemmlowp_context->set_max_num_threads(static_cast<int>(std::thread::hardware_concurrency()));
215 tflite::optimized_ops::Conv(
216 params, getTensorShape(input()), getTensorData<uint8_t>(input()), getTensorShape(filter()),
217 getTensorData<uint8_t>(filter()), getTensorShape(bias()), getTensorData<int32_t>(bias()),
218 getTensorShape(output()), getTensorData<uint8_t>(output()), getTensorShape(_im2col.get()),
219 getTensorData<uint8_t>(_im2col.get()), gemmlowp_context.get());
222 void Conv2D::evalQuantizedS16() const
224 const auto *input_data = getTensorData<int16_t>(input());
225 const auto *filter_data = getTensorData<int16_t>(filter());
226 const auto *bias_data = getTensorData<int64_t>(bias());
227 auto *output_data = getTensorData<int16_t>(output());
229 const Shape &input_shape = input()->shape();
230 const Shape &filter_shape = filter()->shape();
231 const Shape &output_shape = output()->shape();
233 const int32_t batches = input_shape.dim(0);
234 const int32_t input_height = input_shape.dim(1);
235 const int32_t input_width = input_shape.dim(2);
236 const int32_t input_depth = input_shape.dim(3);
237 const int32_t output_depth = filter_shape.dim(0);
238 const int32_t filter_height = filter_shape.dim(1);
239 const int32_t filter_width = filter_shape.dim(2);
240 const int32_t output_height = output_shape.dim(1);
241 const int32_t output_width = output_shape.dim(2);
243 const int32_t stride_height = _params.stride_height;
244 const int32_t stride_width = _params.stride_width;
245 const int32_t dilation_height_factor = _params.dilation_height_factor;
246 const int32_t dilation_width_factor = _params.dilation_width_factor;
248 int32_t activation_min{};
249 int32_t activation_max{};
250 calculateActivationRangeQuantized(_params.activation, output(), &activation_min, &activation_max);
252 const std::vector<double> effective_output_scale =
253 getQuantizedConvolutionMultiplers(input()->scale(), filter()->scales(), output()->scale());
255 const std::vector<ChannelQuantMultipliers> multipliers_raw =
256 quantizeMultipliers(effective_output_scale);
257 BroadcastableWrapper<ChannelQuantMultipliers> multipliers(multipliers_raw);
259 for (int32_t batch = 0; batch < batches; ++batch)
261 for (int32_t out_y = 0; out_y < output_height; ++out_y)
263 for (int32_t out_x = 0; out_x < output_width; ++out_x)
265 for (int32_t out_c = 0; out_c < output_depth; ++out_c)
267 const int32_t in_y_origin = out_y * stride_height - _padding_height;
268 const int32_t in_x_origin = out_x * stride_width - _padding_width;
270 for (int32_t filter_y = 0; filter_y < filter_height; ++filter_y)
272 for (int32_t filter_x = 0; filter_x < filter_width; ++filter_x)
274 const int32_t in_y = in_y_origin + dilation_height_factor * filter_y;
275 const int32_t in_x = in_x_origin + dilation_width_factor * filter_x;
276 if ((in_y >= 0 && in_y < input_height) && (in_x >= 0 && in_x < input_width))
278 for (int32_t in_c = 0; in_c < input_depth; ++in_c)
280 const int16_t input_val =
281 input_data[calcOffset(input_shape, batch, in_y, in_x, in_c)];
282 const int16_t filter_val =
283 filter_data[calcOffset(filter_shape, out_c, filter_y, filter_x, in_c)];
284 acc += static_cast<int64_t>(input_val) * static_cast<int64_t>(filter_val);
291 acc += bias_data[out_c];
294 int32_t scaled_acc = tflite::MultiplyByQuantizedMultiplier(
295 acc, multipliers[out_c].multiplier, multipliers[out_c].shift);
297 scaled_acc = std::max(scaled_acc, activation_min);
298 scaled_acc = std::min(scaled_acc, activation_max);
300 output_data[calcOffset(output_shape, batch, out_y, out_x, out_c)] = scaled_acc;
307 } // namespace kernels
308 } // namespace luci_interpreter