index ec183286182124ee94991f84a5d4592ed608ff6d..3e7238d1659d8da4303fda5f1f001a46df2a9b21 100644 (file)
* THE POSSIBILITY OF SUCH DAMAGE.
*****************************************************************************/
#include <signal.h>
-#include <getopt.h>
#include <iostream>
#include <iomanip>
#include <fstream>
#include <queue>
#include <vector>
#include <cstdio>
+#include <chrono>
#include "executor.h"
#include "execution_object.h"
#include "configuration.h"
-#include "object_classes.h"
-
-#include "opencv2/core.hpp"
-#include "opencv2/imgproc.hpp"
-#include "opencv2/highgui.hpp"
-#include "opencv2/videoio.hpp"
-
-#define NUM_VIDEO_FRAMES 100
-#define DEFAULT_CONFIG "jseg21_tiscapes"
-#define DEFAULT_INPUT "../test/testvecs/input/000100_1024x512_bgr.y"
-
-bool __TI_show_debug_ = false;
-bool is_default_input = false;
-bool is_preprocessed_input = false;
-bool is_camera_input = false;
-int orig_width;
-int orig_height;
-object_class_table_t *object_class_table;
+#include "../common/object_classes.h"
+#include "../common/utils.h"
+#include "../common/video_utils.h"
+using namespace std;
using namespace tidl;
using namespace cv;
-bool RunConfiguration(const std::string& config_file, int num_devices,
- DeviceType device_type, std::string& input_file);
-bool RunAllConfigurations(int32_t num_devices, DeviceType device_type);
-
-bool ReadFrame(ExecutionObject& eo, int frame_idx,
- const Configuration& configuration, int num_frames,
- std::string& image_file, VideoCapture &cap);
-
-bool WriteFrameOutput(const ExecutionObject &eo,
- const Configuration& configuration);
+#define NUM_VIDEO_FRAMES 300
+#define DEFAULT_CONFIG "jseg21_tiscapes"
+#define DEFAULT_INPUT "../test/testvecs/input/000100_1024x512_bgr.y"
+#define DEFAULT_INPUT_FRAMES (9)
+#define DEFAULT_OBJECT_CLASSES_LIST_FILE "jseg21_objects.json"
-static void ProcessArgs(int argc, char *argv[],
- std::string& config,
- int& num_devices,
- DeviceType& device_type,
- std::string& input_file);
+std::unique_ptr<ObjectClasses> object_classes;
+uint32_t orig_width;
+uint32_t orig_height;
-static void DisplayHelp();
-static double ms_diff(struct timespec &t0, struct timespec &t1)
-{ return (t1.tv_sec - t0.tv_sec) * 1e3 + (t1.tv_nsec - t0.tv_nsec) / 1e6; }
+bool RunConfiguration(const cmdline_opts_t& opts);
+Executor* CreateExecutor(DeviceType dt, uint32_t num, const Configuration& c);
+bool ReadFrame(ExecutionObjectPipeline& eop,
+ uint32_t frame_idx, const Configuration& c,
+ const cmdline_opts_t& opts, VideoCapture &cap);
+bool WriteFrameOutput(const ExecutionObjectPipeline &eop,
+ const Configuration& c, const cmdline_opts_t& opts);
+void DisplayHelp();
int main(int argc, char *argv[])
signal(SIGTERM, exit);
// If there are no devices capable of offloading TIDL on the SoC, exit
- uint32_t num_dla = Executor::GetNumDevices(DeviceType::DLA);
- uint32_t num_dsp = Executor::GetNumDevices(DeviceType::DSP);
- if (num_dla == 0 && num_dsp == 0)
+ uint32_t num_eves = Executor::GetNumDevices(DeviceType::EVE);
+ uint32_t num_dsps = Executor::GetNumDevices(DeviceType::DSP);
+ if (num_eves == 0 && num_dsps == 0)
{
- std::cout << "TI DL not supported on this SoC." << std::endl;
+ cout << "TI DL not supported on this SoC." << endl;
return EXIT_SUCCESS;
}
// Process arguments
- std::string config = DEFAULT_CONFIG;
- std::string input_file = DEFAULT_INPUT;
- int num_devices = 1;
- DeviceType device_type = (num_dla > 0 ? DeviceType::DLA:DeviceType::DSP);
- ProcessArgs(argc, argv, config, num_devices, device_type, input_file);
+ cmdline_opts_t opts;
+ opts.config = DEFAULT_CONFIG;
+ opts.object_classes_list_file = DEFAULT_OBJECT_CLASSES_LIST_FILE;
+ if (num_eves != 0) { opts.num_eves = 1; opts.num_dsps = 0; }
+ else { opts.num_eves = 0; opts.num_dsps = 1; }
+ if (! ProcessArgs(argc, argv, opts))
+ {
+ DisplayHelp();
+ exit(EXIT_SUCCESS);
+ }
+ assert(opts.num_dsps != 0 || opts.num_eves != 0);
+ if (opts.num_frames == 0)
+ opts.num_frames = (opts.is_camera_input || opts.is_video_input) ?
+ NUM_VIDEO_FRAMES :
+ (opts.input_file.empty() ? DEFAULT_INPUT_FRAMES : 1);
+ if (opts.input_file.empty())
+ cout << "Input: " << DEFAULT_INPUT << endl;
+ else
+ cout << "Input: " << opts.input_file << endl;
- if ((object_class_table = GetObjectClassTable(config)) == nullptr)
+ // Get object classes list
+ object_classes = std::unique_ptr<ObjectClasses>(
+ new ObjectClasses(opts.object_classes_list_file));
+ if (object_classes->GetNumClasses() == 0)
{
- std::cout << "No object classes defined for this config." << std::endl;
+ cout << "No object classes defined for this config." << endl;
return EXIT_FAILURE;
}
- if (input_file == DEFAULT_INPUT) is_default_input = true;
- if (input_file == "camera") is_camera_input = true;
- if (input_file.length() > 2 &&
- input_file.compare(input_file.length() - 2, 2, ".y") == 0)
- is_preprocessed_input = true;
- std::cout << "Input: " << input_file << std::endl;
- std::string config_file = "../test/testvecs/config/infer/tidl_config_"
- + config + ".txt";
- bool status = RunConfiguration(config_file, num_devices, device_type,
- input_file);
-
+ // Run network
+ bool status = RunConfiguration(opts);
if (!status)
{
- std::cout << "segmentation FAILED" << std::endl;
+ cout << "segmentation FAILED" << endl;
return EXIT_FAILURE;
}
- std::cout << "segmentation PASSED" << std::endl;
+ cout << "segmentation PASSED" << endl;
return EXIT_SUCCESS;
}
-bool RunConfiguration(const std::string& config_file, int num_devices,
- DeviceType device_type, std::string& input_file)
+bool RunConfiguration(const cmdline_opts_t& opts)
{
- DeviceIds ids;
- for (int i = 0; i < num_devices; i++)
- ids.insert(static_cast<DeviceId>(i));
-
// Read the TI DL configuration file
- Configuration configuration;
- bool status = configuration.ReadFromFile(config_file);
+ Configuration c;
+ std::string config_file = "../test/testvecs/config/infer/tidl_config_"
+ + opts.config + ".txt";
+ bool status = c.ReadFromFile(config_file);
if (!status)
{
- std::cerr << "Error in configuration file: " << config_file
- << std::endl;
+ cerr << "Error in configuration file: " << config_file << endl;
return false;
}
+ c.enableApiTrace = opts.verbose;
- // setup input
- int num_frames = is_default_input ? 3 : 1;
+ // setup camera/video input/output
VideoCapture cap;
- std::string image_file;
- if (is_camera_input)
- {
- cap = VideoCapture(1); // cap = VideoCapture("test.mp4");
- if (! cap.isOpened())
- {
- std::cerr << "Cannot open camera input." << std::endl;
- return false;
- }
- num_frames = NUM_VIDEO_FRAMES;
- namedWindow("Segmentation", WINDOW_AUTOSIZE | CV_GUI_NORMAL);
- }
- else
- {
- image_file = input_file;
- }
+ if (! SetVideoInputOutput(cap, opts, "Segmentation")) return false;
try
{
- // Create a executor with the approriate core type, number of cores
+ // Create Executors with the approriate core type, number of cores
// and configuration specified
- Executor executor(device_type, ids, configuration);
-
- // Query Executor for set of ExecutionObjects created
- const ExecutionObjects& execution_objects =
- executor.GetExecutionObjects();
- int num_eos = execution_objects.size();
-
- // Allocate input and output buffers for each execution object
- std::vector<void *> buffers;
- for (auto &eo : execution_objects)
- {
- size_t in_size = eo->GetInputBufferSizeInBytes();
- size_t out_size = eo->GetOutputBufferSizeInBytes();
- ArgInfo in = { ArgInfo(malloc(in_size), in_size)};
- ArgInfo out = { ArgInfo(malloc(out_size), out_size)};
- eo->SetInputOutputBuffer(in, out);
-
- buffers.push_back(in.ptr());
- buffers.push_back(out.ptr());
- }
-
- #define MAX_NUM_EOS 4
- struct timespec t0[MAX_NUM_EOS], t1;
-
- // Process frames with available execution objects in a pipelined manner
+ Executor* e_eve = CreateExecutor(DeviceType::EVE, opts.num_eves, c);
+ Executor* e_dsp = CreateExecutor(DeviceType::DSP, opts.num_dsps, c);
+
+ // Get ExecutionObjects from Executors
+ vector<ExecutionObject*> eos;
+ for (uint32_t i = 0; i < opts.num_eves; i++) eos.push_back((*e_eve)[i]);
+ for (uint32_t i = 0; i < opts.num_dsps; i++) eos.push_back((*e_dsp)[i]);
+ uint32_t num_eos = eos.size();
+
+ // Use duplicate EOPs to do double buffering on frame input/output
+ // because each EOP has its own set of input/output buffers,
+ // so that host ReadFrame() can be overlapped with device processing
+ // Use one EO as an example, with different buffer_factor,
+ // we have different execution behavior:
+ // If buffer_factor is set to 1 -> single buffering
+ // we create one EOP: eop0 (eo0)
+ // pipeline execution of multiple frames over time is as follows:
+ // --------------------- time ------------------->
+ // eop0: [RF][eo0.....][WF]
+ // eop0: [RF][eo0.....][WF]
+ // eop0: [RF][eo0.....][WF]
+ // If buffer_factor is set to 2 -> double buffering
+ // we create two EOPs: eop0 (eo0), eop1(eo0)
+ // pipeline execution of multiple frames over time is as follows:
+ // --------------------- time ------------------->
+ // eop0: [RF][eo0.....][WF]
+ // eop1: [RF] [eo0.....][WF]
+ // eop0: [RF] [eo0.....][WF]
+ // eop1: [RF] [eo0.....][WF]
+ vector<ExecutionObjectPipeline *> eops;
+ uint32_t buffer_factor = 2; // set to 1 for single buffering
+ for (uint32_t j = 0; j < buffer_factor; j++)
+ for (uint32_t i = 0; i < num_eos; i++)
+ eops.push_back(new ExecutionObjectPipeline({eos[i]}));
+ uint32_t num_eops = eops.size();
+
+ // Allocate input and output buffers for each EOP
+ AllocateMemory(eops);
+
+ chrono::time_point<chrono::steady_clock> tloop0, tloop1;
+ tloop0 = chrono::steady_clock::now();
+
+ // Process frames with available eops in a pipelined manner
// additional num_eos iterations to flush the pipeline (epilogue)
- for (int frame_idx = 0;
- frame_idx < num_frames + num_eos; frame_idx++)
+ for (uint32_t frame_idx = 0;
+ frame_idx < opts.num_frames + num_eops; frame_idx++)
{
- ExecutionObject* eo = execution_objects[frame_idx % num_eos].get();
+ ExecutionObjectPipeline* eop = eops[frame_idx % num_eops];
- // Wait for previous frame on the same eo to finish processing
- if (eo->ProcessFrameWait())
+ // Wait for previous frame on the same eop to finish processing
+ if (eop->ProcessFrameWait())
{
- clock_gettime(CLOCK_MONOTONIC, &t1);
- double elapsed_host =
- ms_diff(t0[eo->GetFrameIndex() % num_eos], t1);
- double elapsed_device = eo->GetProcessTimeInMilliSeconds();
- double overhead = 100 - (elapsed_device/elapsed_host*100);
-
- std::cout << "frame[" << eo->GetFrameIndex() << "]: "
- << "Time on device: "
- << std::setw(6) << std::setprecision(4)
- << elapsed_device << "ms, "
- << "host: "
- << std::setw(6) << std::setprecision(4)
- << elapsed_host << "ms ";
- std::cout << "API overhead: "
- << std::setw(6) << std::setprecision(3)
- << overhead << " %" << std::endl;
-
- WriteFrameOutput(*eo, configuration);
+ WriteFrameOutput(*eop, c, opts);
}
- // Read a frame and start processing it with current eo
- if (ReadFrame(*eo, frame_idx, configuration, num_frames,
- image_file, cap))
- {
- clock_gettime(CLOCK_MONOTONIC, &t0[frame_idx % num_eos]);
- eo->ProcessFrameStartAsync();
- }
+ // Read a frame and start processing it with current eop
+ if (ReadFrame(*eop, frame_idx, c, opts, cap))
+ eop->ProcessFrameStartAsync();
}
- for (auto b : buffers)
- free(b);
+ tloop1 = chrono::steady_clock::now();
+ chrono::duration<float> elapsed = tloop1 - tloop0;
+ cout << "Loop total time (including read/write/opencv/print/etc): "
+ << setw(6) << setprecision(4)
+ << (elapsed.count() * 1000) << "ms" << endl;
+ FreeMemory(eops);
+ for (auto eop : eops) delete eop;
+ delete e_eve;
+ delete e_dsp;
}
catch (tidl::Exception &e)
{
- std::cerr << e.what() << std::endl;
+ cerr << e.what() << endl;
status = false;
}
return status;
}
+// Create an Executor with the specified type and number of EOs
+Executor* CreateExecutor(DeviceType dt, uint32_t num, const Configuration& c)
+{
+ if (num == 0) return nullptr;
+
+ DeviceIds ids;
+ for (uint32_t i = 0; i < num; i++)
+ ids.insert(static_cast<DeviceId>(i));
+
+ return new Executor(dt, ids, c);
+}
-bool ReadFrame(ExecutionObject &eo, int frame_idx,
- const Configuration& configuration, int num_frames,
- std::string& image_file, VideoCapture &cap)
+bool ReadFrame(ExecutionObjectPipeline &eop,
+ uint32_t frame_idx, const Configuration& c,
+ const cmdline_opts_t& opts, VideoCapture &cap)
{
- if (frame_idx >= num_frames)
+ if (frame_idx >= opts.num_frames)
return false;
- eo.SetFrameIndex(frame_idx);
+ eop.SetFrameIndex(frame_idx);
- char* frame_buffer = eo.GetInputBufferPtr();
+ char* frame_buffer = eop.GetInputBufferPtr();
assert (frame_buffer != nullptr);
- int channel_size = configuration.inWidth * configuration.inHeight;
+ int channel_size = c.inWidth * c.inHeight;
Mat image;
- if (! image_file.empty())
+ if (! opts.is_camera_input && ! opts.is_video_input)
{
- if (is_preprocessed_input)
+ if (opts.input_file.empty())
{
- std::ifstream ifs(image_file, std::ios::binary);
- ifs.seekg(frame_idx * channel_size * 3);
+ ifstream ifs(DEFAULT_INPUT, ios::binary);
+ ifs.seekg((frame_idx % DEFAULT_INPUT_FRAMES) * channel_size * 3);
ifs.read(frame_buffer, channel_size * 3);
bool ifs_status = ifs.good();
ifs.close();
- orig_width = configuration.inWidth;
- orig_height = configuration.inHeight;
+ orig_width = c.inWidth;
+ orig_height = c.inHeight;
return ifs_status; // already PreProc-ed
}
else
{
- image = cv::imread(image_file, CV_LOAD_IMAGE_COLOR);
+ image = cv::imread(opts.input_file, CV_LOAD_IMAGE_COLOR);
if (image.empty())
{
- std::cerr << "Unable to read from: " << image_file << std::endl;
+ cerr << "Unable to read from: " << opts.input_file << endl;
return false;
}
}
Mat s_image, bgr_frames[3];
orig_width = image.cols;
orig_height = image.rows;
- cv::resize(image, s_image,
- Size(configuration.inWidth, configuration.inHeight),
+ cv::resize(image, s_image, Size(c.inWidth, c.inHeight),
0, 0, cv::INTER_AREA);
cv::split(s_image, bgr_frames);
memcpy(frame_buffer, bgr_frames[0].ptr(), channel_size);
{
for (int i = 0; i < channel_size; i++)
{
- object_class_t *object_class = GetObjectClass(object_class_table,
- classes[i]);
- mb[i] = object_class->color.blue;
- mg[i] = object_class->color.green;
- mr[i] = object_class->color.red;
+ const ObjectClass& object_class = object_classes->At(classes[i]);
+ mb[i] = object_class.color.blue;
+ mg[i] = object_class.color.green;
+ mr[i] = object_class.color.red;
}
}
// Create frame overlayed with pixel-level segmentation
-bool WriteFrameOutput(const ExecutionObject &eo,
- const Configuration& configuration)
+bool WriteFrameOutput(const ExecutionObjectPipeline &eop,
+ const Configuration& c,
+ const cmdline_opts_t& opts)
{
- unsigned char *out = (unsigned char *) eo.GetOutputBufferPtr();
- int out_size = eo.GetOutputBufferSizeInBytes();
- int width = configuration.inWidth;
- int height = configuration.inHeight;
+ unsigned char *out = (unsigned char *) eop.GetOutputBufferPtr();
+ int width = c.inWidth;
+ int height = c.inHeight;
int channel_size = width * height;
Mat mask, frame, blend, r_blend, bgr[3];
cv::merge(bgr, 3, mask);
// Asseembly original frame
- unsigned char *in = (unsigned char *) eo.GetInputBufferPtr();
+ unsigned char *in = (unsigned char *) eop.GetInputBufferPtr();
bgr[0] = Mat(height, width, CV_8UC(1), in);
bgr[1] = Mat(height, width, CV_8UC(1), in + channel_size);
bgr[2] = Mat(height, width, CV_8UC(1), in + channel_size*2);
// Create overlayed frame
cv::addWeighted(frame, 0.7, mask, 0.3, 0.0, blend);
- cv::resize(blend, r_blend, Size(orig_width, orig_height));
- if (is_camera_input)
+ // Resize to output width/height, keep aspect ratio
+ uint32_t output_width = opts.output_width;
+ if (output_width == 0) output_width = orig_width;
+ uint32_t output_height = (output_width*1.0f) / orig_width * orig_height;
+ cv::resize(blend, r_blend, Size(output_width, output_height));
+
+ if (opts.is_camera_input || opts.is_video_input)
{
cv::imshow("Segmentation", r_blend);
waitKey(1);
}
else
{
- int frame_index = eo.GetFrameIndex();
+ int frame_index = eop.GetFrameIndex();
char outfile_name[64];
- if (is_preprocessed_input)
+ if (opts.input_file.empty())
{
snprintf(outfile_name, 64, "frame_%d.png", frame_index);
cv::imwrite(outfile_name, frame);
return true;
}
-
-void ProcessArgs(int argc, char *argv[], std::string& config,
- int& num_devices, DeviceType& device_type,
- std::string& input_file)
-{
- const struct option long_options[] =
- {
- {"config", required_argument, 0, 'c'},
- {"num_devices", required_argument, 0, 'n'},
- {"device_type", required_argument, 0, 't'},
- {"image_file", required_argument, 0, 'i'},
- {"help", no_argument, 0, 'h'},
- {"verbose", no_argument, 0, 'v'},
- {0, 0, 0, 0}
- };
-
- int option_index = 0;
-
- while (true)
- {
- int c = getopt_long(argc, argv, "c:n:t:i:hv", long_options, &option_index);
-
- if (c == -1)
- break;
-
- switch (c)
- {
- case 'c': config = optarg;
- break;
-
- case 'n': num_devices = atoi(optarg);
- assert (num_devices > 0 && num_devices <= 4);
- break;
-
- case 't': if (*optarg == 'e')
- device_type = DeviceType::DLA;
- else if (*optarg == 'd')
- device_type = DeviceType::DSP;
- else
- {
- std::cerr << "Invalid argument to -t, only e or d"
- " allowed" << std::endl;
- exit(EXIT_FAILURE);
- }
- break;
-
- case 'i': input_file = optarg;
- break;
-
- case 'v': __TI_show_debug_ = true;
- break;
-
- case 'h': DisplayHelp();
- exit(EXIT_SUCCESS);
- break;
-
- case '?': // Error in getopt_long
- exit(EXIT_FAILURE);
- break;
-
- default:
- std::cerr << "Unsupported option: " << c << std::endl;
- break;
- }
- }
-}
-
void DisplayHelp()
{
- std::cout << "Usage: segmentation\n"
- " Will run segmentation network to perform pixel-level"
- " classification.\n Use -c to run a different"
- " segmentation network. Default is jseg21_tiscapes.\n"
- "Optional arguments:\n"
- " -c <config> Valid configs: jseg21_tiscapes, jseg21\n"
- " -n <number of cores> Number of cores to use (1 - 4)\n"
- " -t <d|e> Type of core. d -> DSP, e -> DLA\n"
- " -i <image> Path to the image file\n"
- " Default are 3 frames in testvecs\n"
- " -i camera Use camera as input\n"
- " -v Verbose output during execution\n"
- " -h Help\n";
+ std::cout <<
+ "Usage: segmentation\n"
+ " Will run segmentation network to perform pixel-level"
+ " classification.\n Use -c to run a different"
+ " segmentation network. Default is jseg21_tiscapes.\n"
+ "Optional arguments:\n"
+ " -c <config> Valid configs: jseg21_tiscapes, jseg21\n"
+ " -d <number> Number of dsp cores to use\n"
+ " -e <number> Number of eve cores to use\n"
+ " -i <image> Path to the image file as input\n"
+ " Default are 9 frames in testvecs\n"
+ " -i camera<number> Use camera as input\n"
+ " video input port: /dev/video<number>\n"
+ " -i <name>.{mp4,mov,avi} Use video file as input\n"
+ " -l <objects_list> Path to the object classes list file\n"
+ " -f <number> Number of frames to process\n"
+ " -w <number> Output image/video width\n"
+ " -v Verbose output during execution\n"
+ " -h Help\n";
}