4bcc707984bae927bf2c521e45c2138a89061872
1 /******************************************************************************
2 * Copyright (c) 2018, Texas Instruments Incorporated - http://www.ti.com/
3 * All rights reserved.
4 *
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions are met:
7 * * Redistributions of source code must retain the above copyright
8 * notice, this list of conditions and the following disclaimer.
9 * * Redistributions in binary form must reproduce the above copyright
10 * notice, this list of conditions and the following disclaimer in the
11 * documentation and/or other materials provided with the distribution.
12 * * Neither the name of Texas Instruments Incorporated nor the
13 * names of its contributors may be used to endorse or promote products
14 * derived from this software without specific prior written permission.
15 *
16 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
17 * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
18 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
19 * ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
20 * LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
21 * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
22 * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
23 * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
24 * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
25 * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF
26 * THE POSSIBILITY OF SUCH DAMAGE.
27 *****************************************************************************/
28 #include <signal.h>
29 #include <iostream>
30 #include <iomanip>
31 #include <fstream>
32 #include <cassert>
33 #include <string>
34 #include <functional>
35 #include <algorithm>
36 #include <time.h>
37 #include <unistd.h>
39 #include <queue>
40 #include <vector>
41 #include <cstdio>
42 #include <chrono>
44 #include "executor.h"
45 #include "execution_object.h"
46 #include "execution_object_pipeline.h"
47 #include "configuration.h"
48 #include "../segmentation/object_classes.h"
49 #include "../common/utils.h"
50 #include "../common/video_utils.h"
52 using namespace std;
53 using namespace tidl;
54 using namespace cv;
57 #define NUM_VIDEO_FRAMES 100
58 #define DEFAULT_CONFIG "jdetnet"
59 #define DEFAULT_INPUT "../test/testvecs/input/preproc_0_768x320.y"
60 #define DEFAULT_INPUT_FRAMES (1)
62 object_class_table_t *object_class_table;
63 uint32_t orig_width;
64 uint32_t orig_height;
67 bool RunConfiguration(const cmdline_opts_t& opts);
68 Executor* CreateExecutor(DeviceType dt, uint32_t num, const Configuration& c,
69 int layers_group_id);
70 bool ReadFrame(ExecutionObjectPipeline& eop, uint32_t frame_idx,
71 const Configuration& c, const cmdline_opts_t& opts,
72 VideoCapture &cap);
73 bool WriteFrameOutput(const ExecutionObjectPipeline& eop,
74 const Configuration& c, const cmdline_opts_t& opts);
75 static void DisplayHelp();
78 int main(int argc, char *argv[])
79 {
80 // Catch ctrl-c to ensure a clean exit
81 signal(SIGABRT, exit);
82 signal(SIGTERM, exit);
84 // If there are no devices capable of offloading TIDL on the SoC, exit
85 uint32_t num_eves = Executor::GetNumDevices(DeviceType::EVE);
86 uint32_t num_dsps = Executor::GetNumDevices(DeviceType::DSP);
87 if (num_eves == 0 || num_dsps == 0)
88 {
89 cout << "ssd_multibox requires both EVE and DSP for execution." << endl;
90 return EXIT_SUCCESS;
91 }
93 // Process arguments
94 cmdline_opts_t opts;
95 opts.config = DEFAULT_CONFIG;
96 opts.num_eves = 1;
97 opts.num_dsps = 1;
98 if (! ProcessArgs(argc, argv, opts))
99 {
100 DisplayHelp();
101 exit(EXIT_SUCCESS);
102 }
103 assert(opts.num_dsps != 0 && opts.num_eves != 0);
104 if (opts.num_frames == 0)
105 opts.num_frames = (opts.is_camera_input || opts.is_video_input) ?
106 NUM_VIDEO_FRAMES :
107 (opts.input_file.empty() ? DEFAULT_INPUT_FRAMES : 1);
108 if (opts.input_file.empty())
109 cout << "Input: " << DEFAULT_INPUT << endl;
110 else
111 cout << "Input: " << opts.input_file << endl;
113 // Get object class table
114 if ((object_class_table = GetObjectClassTable(opts.config)) == nullptr)
115 {
116 cout << "No object classes defined for this config." << endl;
117 return EXIT_FAILURE;
118 }
120 // Run network
121 bool status = RunConfiguration(opts);
122 if (!status)
123 {
124 cout << "ssd_multibox FAILED" << endl;
125 return EXIT_FAILURE;
126 }
128 cout << "ssd_multibox PASSED" << endl;
129 return EXIT_SUCCESS;
130 }
132 bool RunConfiguration(const cmdline_opts_t& opts)
133 {
134 // Read the TI DL configuration file
135 Configuration c;
136 std::string config_file = "../test/testvecs/config/infer/tidl_config_"
137 + opts.config + ".txt";
138 bool status = c.ReadFromFile(config_file);
139 if (!status)
140 {
141 cerr << "Error in configuration file: " << config_file << endl;
142 return false;
143 }
144 c.enableApiTrace = opts.verbose;
146 // setup camera/video input
147 VideoCapture cap;
148 if (! SetVideoInputOutput(cap, opts, "SSD_Multibox")) return false;
150 try
151 {
152 // Create Executors with the approriate core type, number of cores
153 // and configuration specified
154 // EVE will run layersGroupId 1 in the network, while
155 // DSP will run layersGroupId 2 in the network
156 Executor* e_eve = CreateExecutor(DeviceType::EVE, opts.num_eves, c, 1);
157 Executor* e_dsp = CreateExecutor(DeviceType::DSP, opts.num_dsps, c, 2);
159 // Construct ExecutionObjectPipeline that utilizes multiple
160 // ExecutionObjects to process a single frame, each ExecutionObject
161 // processes one layerGroup of the network
162 //
163 // Pipeline depth can enable more optimized pipeline execution:
164 // Given one EVE and one DSP as an example, with different
165 // pipeline_depth, we have different execution behavior:
166 // If pipeline_depth is set to 1,
167 // we create one EOP: eop0 (eve0, dsp0)
168 // pipeline execution of multiple frames over time is as follows:
169 // --------------------- time ------------------->
170 // eop0: [eve0...][dsp0]
171 // eop0: [eve0...][dsp0]
172 // eop0: [eve0...][dsp0]
173 // eop0: [eve0...][dsp0]
174 // If pipeline_depth is set to 2,
175 // we create two EOPs: eop0 (eve0, dsp0), eop1(eve0, dsp0)
176 // pipeline execution of multiple frames over time is as follows:
177 // --------------------- time ------------------->
178 // eop0: [eve0...][dsp0]
179 // eop1: [eve0...][dsp0]
180 // eop0: [eve0...][dsp0]
181 // eop1: [eve0...][dsp0]
182 // Additional benefit of setting pipeline_depth to 2 is that
183 // it can also overlap host ReadFrame() with device processing:
184 // --------------------- time ------------------->
185 // eop0: [RF][eve0...][dsp0]
186 // eop1: [RF] [eve0...][dsp0]
187 // eop0: [RF][eve0...][dsp0]
188 // eop1: [RF][eve0...][dsp0]
189 vector<ExecutionObjectPipeline *> eops;
190 uint32_t pipeline_depth = 2; // 2 EOs in EOP -> depth 2
191 for (uint32_t j = 0; j < pipeline_depth; j++)
192 for (uint32_t i = 0; i < max(opts.num_eves, opts.num_dsps); i++)
193 eops.push_back(new ExecutionObjectPipeline(
194 {(*e_eve)[i%opts.num_eves], (*e_dsp)[i%opts.num_dsps]}));
195 uint32_t num_eops = eops.size();
197 // Allocate input/output memory for each EOP
198 AllocateMemory(eops);
200 chrono::time_point<chrono::steady_clock> tloop0, tloop1;
201 tloop0 = chrono::steady_clock::now();
203 // Process frames with available eops in a pipelined manner
204 // additional num_eops iterations to flush pipeline (epilogue)
205 for (uint32_t frame_idx = 0;
206 frame_idx < opts.num_frames + num_eops; frame_idx++)
207 {
208 ExecutionObjectPipeline* eop = eops[frame_idx % num_eops];
210 // Wait for previous frame on the same eop to finish processing
211 if (eop->ProcessFrameWait())
212 {
213 ReportTime(eop);
214 WriteFrameOutput(*eop, c, opts);
215 }
217 // Read a frame and start processing it with current eo
218 if (ReadFrame(*eop, frame_idx, c, opts, cap))
219 eop->ProcessFrameStartAsync();
220 }
222 tloop1 = chrono::steady_clock::now();
223 chrono::duration<float> elapsed = tloop1 - tloop0;
224 cout << "Loop total time (including read/write/opencv/print/etc): "
225 << setw(6) << setprecision(4)
226 << (elapsed.count() * 1000) << "ms" << endl;
228 FreeMemory(eops);
229 for (auto eop : eops) delete eop;
230 delete e_eve;
231 delete e_dsp;
232 }
233 catch (tidl::Exception &e)
234 {
235 cerr << e.what() << endl;
236 status = false;
237 }
239 return status;
240 }
242 // Create an Executor with the specified type and number of EOs
243 Executor* CreateExecutor(DeviceType dt, uint32_t num, const Configuration& c,
244 int layers_group_id)
245 {
246 if (num == 0) return nullptr;
248 DeviceIds ids;
249 for (uint32_t i = 0; i < num; i++)
250 ids.insert(static_cast<DeviceId>(i));
252 return new Executor(dt, ids, c, layers_group_id);
253 }
255 bool ReadFrame(ExecutionObjectPipeline& eop, uint32_t frame_idx,
256 const Configuration& c, const cmdline_opts_t& opts,
257 VideoCapture &cap)
258 {
259 if ((uint32_t)frame_idx >= opts.num_frames)
260 return false;
262 eop.SetFrameIndex(frame_idx);
264 char* frame_buffer = eop.GetInputBufferPtr();
265 assert (frame_buffer != nullptr);
266 int channel_size = c.inWidth * c.inHeight;
268 Mat image;
269 if (!opts.is_camera_input && !opts.is_video_input)
270 {
271 if (opts.input_file.empty())
272 {
273 ifstream ifs(DEFAULT_INPUT, ios::binary);
274 ifs.seekg((frame_idx % DEFAULT_INPUT_FRAMES) * channel_size * 3);
275 ifs.read(frame_buffer, channel_size * 3);
276 bool ifs_status = ifs.good();
277 ifs.close();
278 orig_width = c.inWidth;
279 orig_height = c.inHeight;
280 return ifs_status; // already PreProc-ed
281 }
282 else
283 {
284 image = cv::imread(opts.input_file, CV_LOAD_IMAGE_COLOR);
285 if (image.empty())
286 {
287 cerr << "Unable to read from: " << opts.input_file << endl;
288 return false;
289 }
290 }
291 }
292 else
293 {
294 // 640x480 camera input, process one in every 5 frames,
295 // can adjust number of skipped frames to match real time processing
296 if (! cap.grab()) return false;
297 if (! cap.grab()) return false;
298 if (! cap.grab()) return false;
299 if (! cap.grab()) return false;
300 if (! cap.grab()) return false;
301 if (! cap.retrieve(image)) return false;
302 }
304 // scale to network input size
305 Mat s_image, bgr_frames[3];
306 orig_width = image.cols;
307 orig_height = image.rows;
308 cv::resize(image, s_image, Size(c.inWidth, c.inHeight),
309 0, 0, cv::INTER_AREA);
310 cv::split(s_image, bgr_frames);
311 memcpy(frame_buffer, bgr_frames[0].ptr(), channel_size);
312 memcpy(frame_buffer+1*channel_size, bgr_frames[1].ptr(), channel_size);
313 memcpy(frame_buffer+2*channel_size, bgr_frames[2].ptr(), channel_size);
314 return true;
315 }
317 // Create frame with boxes drawn around classified objects
318 bool WriteFrameOutput(const ExecutionObjectPipeline& eop,
319 const Configuration& c, const cmdline_opts_t& opts)
320 {
321 // Asseembly original frame
322 int width = c.inWidth;
323 int height = c.inHeight;
324 int channel_size = width * height;
325 Mat frame, r_frame, bgr[3];
327 unsigned char *in = (unsigned char *) eop.GetInputBufferPtr();
328 bgr[0] = Mat(height, width, CV_8UC(1), in);
329 bgr[1] = Mat(height, width, CV_8UC(1), in + channel_size);
330 bgr[2] = Mat(height, width, CV_8UC(1), in + channel_size*2);
331 cv::merge(bgr, 3, frame);
333 int frame_index = eop.GetFrameIndex();
334 char outfile_name[64];
335 if (opts.input_file.empty())
336 {
337 snprintf(outfile_name, 64, "frame_%d.png", frame_index);
338 cv::imwrite(outfile_name, frame);
339 printf("Saving frame %d to: %s\n", frame_index, outfile_name);
340 }
342 // Draw boxes around classified objects
343 float *out = (float *) eop.GetOutputBufferPtr();
344 int num_floats = eop.GetOutputBufferSizeInBytes() / sizeof(float);
345 for (int i = 0; i < num_floats / 7; i++)
346 {
347 int index = (int) out[i * 7 + 0];
348 if (index < 0) break;
350 int label = (int) out[i * 7 + 1];
351 int xmin = (int) (out[i * 7 + 3] * width);
352 int ymin = (int) (out[i * 7 + 4] * height);
353 int xmax = (int) (out[i * 7 + 5] * width);
354 int ymax = (int) (out[i * 7 + 6] * height);
356 object_class_t *object_class = GetObjectClass(object_class_table,
357 label);
358 if (object_class == nullptr) continue;
360 #if 0
361 printf("(%d, %d) -> (%d, %d): %s, score=%f\n",
362 xmin, ymin, xmax, ymax, object_class->label, score);
363 #endif
365 cv::rectangle(frame, Point(xmin, ymin), Point(xmax, ymax),
366 Scalar(object_class->color.blue,
367 object_class->color.green,
368 object_class->color.red), 2);
369 }
371 // Resize to output width/height, keep aspect ratio
372 uint32_t output_width = opts.output_width;
373 if (output_width == 0) output_width = orig_width;
374 uint32_t output_height = (output_width*1.0f) / orig_width * orig_height;
375 cv::resize(frame, r_frame, Size(output_width, output_height));
377 if (opts.is_camera_input || opts.is_video_input)
378 {
379 cv::imshow("SSD_Multibox", r_frame);
380 waitKey(1);
381 }
382 else
383 {
384 snprintf(outfile_name, 64, "multibox_%d.png", frame_index);
385 cv::imwrite(outfile_name, r_frame);
386 printf("Saving frame %d with SSD multiboxes to: %s\n",
387 frame_index, outfile_name);
388 }
390 return true;
391 }
393 void DisplayHelp()
394 {
395 std::cout <<
396 "Usage: ssd_multibox\n"
397 " Will run partitioned ssd_multibox network to perform "
398 "multi-objects detection\n"
399 " and classification. First part of network "
400 "(layersGroupId 1) runs on EVE,\n"
401 " second part (layersGroupId 2) runs on DSP.\n"
402 " Use -c to run a different segmentation network. Default is jdetnet.\n"
403 "Optional arguments:\n"
404 " -c <config> Valid configs: jdetnet \n"
405 " -d <number> Number of dsp cores to use\n"
406 " -e <number> Number of eve cores to use\n"
407 " -i <image> Path to the image file as input\n"
408 " Default are 9 frames in testvecs\n"
409 " -i camera<number> Use camera as input\n"
410 " video input port: /dev/video<number>\n"
411 " -i <name>.{mp4,mov,avi} Use video file as input\n"
412 " -f <number> Number of frames to process\n"
413 " -w <number> Output image/video width\n"
414 " -v Verbose output during execution\n"
415 " -h Help\n";
416 }