2 * Copyright (c) 2016, 2017 ARM Limited.
4 * SPDX-License-Identifier: MIT
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:
13 * The above copyright notice and this permission notice shall be included in all
14 * copies or substantial portions of the Software.
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
24 #include "arm_compute/runtime/NEON/NEFunctions.h"
26 #include "arm_compute/core/Types.h"
27 #include "arm_compute/runtime/Allocator.h"
28 #include "arm_compute/runtime/BlobLifetimeManager.h"
29 #include "arm_compute/runtime/MemoryManagerOnDemand.h"
30 #include "arm_compute/runtime/PoolManager.h"
31 #include "utils/Utils.h"
33 using namespace arm_compute;
34 using namespace utils;
36 void main_cnn(int argc, const char **argv)
38 ARM_COMPUTE_UNUSED(argc);
39 ARM_COMPUTE_UNUSED(argv);
41 // Create NEON allocator
44 // Create memory manager components
45 // We need 2 memory managers: 1 for handling the tensors within the functions (mm_layers) and 1 for handling the input and output tensors of the functions (mm_transitions))
46 auto lifetime_mgr0 = std::make_shared<BlobLifetimeManager>(); // Create lifetime manager
47 auto lifetime_mgr1 = std::make_shared<BlobLifetimeManager>(); // Create lifetime manager
48 auto pool_mgr0 = std::make_shared<PoolManager>(); // Create pool manager
49 auto pool_mgr1 = std::make_shared<PoolManager>(); // Create pool manager
50 auto mm_layers = std::make_shared<MemoryManagerOnDemand>(lifetime_mgr0, pool_mgr0); // Create the memory manager
51 auto mm_transitions = std::make_shared<MemoryManagerOnDemand>(lifetime_mgr1, pool_mgr1); // Create the memory manager
53 // The src tensor should contain the input image
56 // The weights and biases tensors should be initialized with the values inferred with the training
74 // Create layers and set memory manager where allowed to manage internal memory requirements
75 NEConvolutionLayer conv0(mm_layers);
76 NEConvolutionLayer conv1(mm_layers);
79 NEFullyConnectedLayer fc0(mm_layers);
80 NEActivationLayer act0;
81 NEActivationLayer act1;
82 NEActivationLayer act2;
83 NESoftmaxLayer softmax(mm_layers);
85 /* [Initialize tensors] */
87 // Initialize src tensor
88 constexpr unsigned int width_src_image = 32;
89 constexpr unsigned int height_src_image = 32;
90 constexpr unsigned int ifm_src_img = 1;
92 const TensorShape src_shape(width_src_image, height_src_image, ifm_src_img);
93 src.allocator()->init(TensorInfo(src_shape, 1, DataType::F32));
95 // Initialize tensors of conv0
96 constexpr unsigned int kernel_x_conv0 = 5;
97 constexpr unsigned int kernel_y_conv0 = 5;
98 constexpr unsigned int ofm_conv0 = 8;
100 const TensorShape weights_shape_conv0(kernel_x_conv0, kernel_y_conv0, src_shape.z(), ofm_conv0);
101 const TensorShape biases_shape_conv0(weights_shape_conv0[3]);
102 const TensorShape out_shape_conv0(src_shape.x(), src_shape.y(), weights_shape_conv0[3]);
104 weights0.allocator()->init(TensorInfo(weights_shape_conv0, 1, DataType::F32));
105 biases0.allocator()->init(TensorInfo(biases_shape_conv0, 1, DataType::F32));
106 out_conv0.allocator()->init(TensorInfo(out_shape_conv0, 1, DataType::F32));
108 // Initialize tensor of act0
109 out_act0.allocator()->init(TensorInfo(out_shape_conv0, 1, DataType::F32));
111 // Initialize tensor of pool0
112 TensorShape out_shape_pool0 = out_shape_conv0;
113 out_shape_pool0.set(0, out_shape_pool0.x() / 2);
114 out_shape_pool0.set(1, out_shape_pool0.y() / 2);
115 out_pool0.allocator()->init(TensorInfo(out_shape_pool0, 1, DataType::F32));
117 // Initialize tensors of conv1
118 constexpr unsigned int kernel_x_conv1 = 3;
119 constexpr unsigned int kernel_y_conv1 = 3;
120 constexpr unsigned int ofm_conv1 = 16;
122 const TensorShape weights_shape_conv1(kernel_x_conv1, kernel_y_conv1, out_shape_pool0.z(), ofm_conv1);
124 const TensorShape biases_shape_conv1(weights_shape_conv1[3]);
125 const TensorShape out_shape_conv1(out_shape_pool0.x(), out_shape_pool0.y(), weights_shape_conv1[3]);
127 weights1.allocator()->init(TensorInfo(weights_shape_conv1, 1, DataType::F32));
128 biases1.allocator()->init(TensorInfo(biases_shape_conv1, 1, DataType::F32));
129 out_conv1.allocator()->init(TensorInfo(out_shape_conv1, 1, DataType::F32));
131 // Initialize tensor of act1
132 out_act1.allocator()->init(TensorInfo(out_shape_conv1, 1, DataType::F32));
134 // Initialize tensor of pool1
135 TensorShape out_shape_pool1 = out_shape_conv1;
136 out_shape_pool1.set(0, out_shape_pool1.x() / 2);
137 out_shape_pool1.set(1, out_shape_pool1.y() / 2);
138 out_pool1.allocator()->init(TensorInfo(out_shape_pool1, 1, DataType::F32));
140 // Initialize tensor of fc0
141 constexpr unsigned int num_labels = 128;
143 const TensorShape weights_shape_fc0(out_shape_pool1.x() * out_shape_pool1.y() * out_shape_pool1.z(), num_labels);
144 const TensorShape biases_shape_fc0(num_labels);
145 const TensorShape out_shape_fc0(num_labels);
147 weights2.allocator()->init(TensorInfo(weights_shape_fc0, 1, DataType::F32));
148 biases2.allocator()->init(TensorInfo(biases_shape_fc0, 1, DataType::F32));
149 out_fc0.allocator()->init(TensorInfo(out_shape_fc0, 1, DataType::F32));
151 // Initialize tensor of act2
152 out_act2.allocator()->init(TensorInfo(out_shape_fc0, 1, DataType::F32));
154 // Initialize tensor of softmax
155 const TensorShape out_shape_softmax(out_shape_fc0.x());
156 out_softmax.allocator()->init(TensorInfo(out_shape_softmax, 1, DataType::F32));
158 /* -----------------------End: [Initialize tensors] */
160 /* [Configure functions] */
162 // in:32x32x1: 5x5 convolution, 8 output features maps (OFM)
163 conv0.configure(&src, &weights0, &biases0, &out_conv0, PadStrideInfo(1 /* stride_x */, 1 /* stride_y */, 2 /* pad_x */, 2 /* pad_y */));
165 // in:32x32x8, out:32x32x8, Activation function: relu
166 act0.configure(&out_conv0, &out_act0, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
168 // in:32x32x8, out:16x16x8 (2x2 pooling), Pool type function: Max
169 pool0.configure(&out_act0, &out_pool0, PoolingLayerInfo(PoolingType::MAX, 2, PadStrideInfo(2 /* stride_x */, 2 /* stride_y */)));
171 // in:16x16x8: 3x3 convolution, 16 output features maps (OFM)
172 conv1.configure(&out_pool0, &weights1, &biases1, &out_conv1, PadStrideInfo(1 /* stride_x */, 1 /* stride_y */, 1 /* pad_x */, 1 /* pad_y */));
174 // in:16x16x16, out:16x16x16, Activation function: relu
175 act1.configure(&out_conv1, &out_act1, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
177 // in:16x16x16, out:8x8x16 (2x2 pooling), Pool type function: Average
178 pool1.configure(&out_act1, &out_pool1, PoolingLayerInfo(PoolingType::AVG, 2, PadStrideInfo(2 /* stride_x */, 2 /* stride_y */)));
180 // in:8x8x16, out:128
181 fc0.configure(&out_pool1, &weights2, &biases2, &out_fc0);
183 // in:128, out:128, Activation function: relu
184 act2.configure(&out_fc0, &out_act2, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU));
187 softmax.configure(&out_act2, &out_softmax);
189 /* -----------------------End: [Configure functions] */
191 /*[ Add tensors to memory manager ]*/
193 // We need 2 memory groups for handling the input and output
194 // We call explicitly allocate after manage() in order to avoid overlapping lifetimes
195 MemoryGroup memory_group0(mm_transitions);
196 MemoryGroup memory_group1(mm_transitions);
198 memory_group0.manage(&out_conv0);
199 out_conv0.allocator()->allocate();
200 memory_group1.manage(&out_act0);
201 out_act0.allocator()->allocate();
202 memory_group0.manage(&out_pool0);
203 out_pool0.allocator()->allocate();
204 memory_group1.manage(&out_conv1);
205 out_conv1.allocator()->allocate();
206 memory_group0.manage(&out_act1);
207 out_act1.allocator()->allocate();
208 memory_group1.manage(&out_pool1);
209 out_pool1.allocator()->allocate();
210 memory_group0.manage(&out_fc0);
211 out_fc0.allocator()->allocate();
212 memory_group1.manage(&out_act2);
213 out_act2.allocator()->allocate();
214 memory_group0.manage(&out_softmax);
215 out_softmax.allocator()->allocate();
217 /* -----------------------End: [ Add tensors to memory manager ] */
219 /* [Allocate tensors] */
221 // Now that the padding requirements are known we can allocate all tensors
222 src.allocator()->allocate();
223 weights0.allocator()->allocate();
224 weights1.allocator()->allocate();
225 weights2.allocator()->allocate();
226 biases0.allocator()->allocate();
227 biases1.allocator()->allocate();
228 biases2.allocator()->allocate();
230 /* -----------------------End: [Allocate tensors] */
232 // Finalize layers memory manager
234 // Set allocator that the memory manager will use
235 mm_layers->set_allocator(&allocator);
237 // Number of pools that the manager will create. This specifies how many layers you want to run in parallel
238 mm_layers->set_num_pools(1);
240 // Finalize the manager. (Validity checks, memory allocations etc)
241 mm_layers->finalize();
243 // Finalize transitions memory manager
245 // Set allocator that the memory manager will use
246 mm_transitions->set_allocator(&allocator);
248 // Number of pools that the manager will create. This specifies how many models we can run in parallel.
249 // Setting to 2 as we need one for the input and one for the output at any given time
250 mm_transitions->set_num_pools(2);
252 // Finalize the manager. (Validity checks, memory allocations etc)
253 mm_transitions->finalize();
255 /* [Initialize weights and biases tensors] */
257 // Once the tensors have been allocated, the src, weights and biases tensors can be initialized
260 /* -----------------------[Initialize weights and biases tensors] */
262 /* [Execute the functions] */
264 // Acquire memory for the memory groups
265 memory_group0.acquire();
266 memory_group1.acquire();
279 memory_group0.release();
280 memory_group1.release();
282 /* -----------------------End: [Execute the functions] */
285 /** Main program for cnn test
287 * The example implements the following CNN architecture:
289 * Input -> conv0:5x5 -> act0:relu -> pool:2x2 -> conv1:3x3 -> act1:relu -> pool:2x2 -> fc0 -> act2:relu -> softmax
291 * @param[in] argc Number of arguments
292 * @param[in] argv Arguments
294 int main(int argc, const char **argv)
296 return utils::run_example(argc, argv, main_cnn);