TouhouVideoFeatureExtractor/worker.cc at master · PatchyVideo/TouhouVideoFeatureExtractor · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143

#include "worker.h"
#include "wrappers/CUDADeviceMemory.h"
#include "wrappers/CUDAStream.h"
#include "TransNet/transnet.cuh"
#include "CLIP/clip.cuh"

namespace worker_details {

std::mutex create_context_mutex;

void worker_thread(CUcontext cuda_context, Worker* self) {
	CUDAThreadContext cuda_thread_ctx(cuda_context);
	CUDAStream stream;

	std::unique_lock<std::mutex> ctx_creation_guard(create_context_mutex);
	nvinfer1::IExecutionContext *transnet_ctx(self->transnet_engine->createExecutionContext());
	nvinfer1::IExecutionContext* clip_ctx(self->clip_engine->createExecutionContext());
	ctx_creation_guard.unlock();

	assert(transnet_ctx->setOptimizationProfile(0));
	assert(clip_ctx->setOptimizationProfile(0));

	auto transnet_dims(nvinfer1::Dims{ .nbDims = 5 });
	transnet_dims.d[0] = 1;
	transnet_dims.d[1] = 950;
	transnet_dims.d[2] = 3;
	transnet_dims.d[3] = 27;
	transnet_dims.d[4] = 48;
	transnet_ctx->setBindingDimensions(0, transnet_dims);
	clip_ctx->setBindingDimensions(0, nvinfer1::Dims4(8, 3, CLIP_HEIGHT, CLIP_WIDTH));

	CUDADeviceMemoryUnique<f32> scratch_f32_1;
	CUDADeviceMemoryUnique<f32> scratch_f32_2;
	CUDADeviceMemoryUnique<i32> scratch_i32_1;
	CUDADeviceMemoryUnique<u8> scratch_u8_1;

	CUDADeviceMemoryUnique<f32> transnet_scores(950);
	CUDAHostMemoryUnique<f32> transnet_scores_cpu(950);

	CUDADeviceMemoryUnique<f32> clip_features;
	CUDAHostMemoryUnique<f32> clip_features_cpu;

	std::vector<u32> key_frames;
	key_frames.reserve(900);
	std::vector<i32> feature_key_frames;
	feature_key_frames.reserve(920);

	constexpr i32 max_key_frame_separation = 300;
	constexpr i32 min_key_frame_separation = 3;

	while (self->running) {
		auto task_opt(self->GetJob());
		if (!task_opt.has_value())
			continue; // self->running == false
		key_frames.clear();
		feature_key_frames.clear();
		FrameBatch fb(task_opt->frames);
		usize custom_data(task_opt->custom_data);
		run_transnet(
			transnet_ctx,
			transnet_scores.at_offset(0, 0),
			scratch_f32_1,
			scratch_f32_2,
			fb.frames_gpu,
			CLIP_WIDTH,
			CLIP_HEIGHT,
			fb.number_of_frames,
			stream
		);
		transnet_scores.download_partial(transnet_scores_cpu, fb.number_of_frames + 50, stream);
		cudaStreamSynchronize(stream);
		key_frames.push_back(0);
		i32 last_frame(0);
		for (usize i(0); i < fb.number_of_frames; ++i) {
			f32 cur_score(transnet_scores_cpu[i + 25]);
			if (cur_score > 0.1 && i > 0 && i < fb.number_of_frames - 1 && i - last_frame > min_key_frame_separation) {
				key_frames.push_back(i);
				last_frame = i;
			}
		}
		key_frames.push_back(fb.number_of_frames - 1);
		for (usize i(1); i < key_frames.size(); ++i) {
			i32 left(key_frames[i - 1]), right(key_frames[i]);
			if (right - left > max_key_frame_separation) {
				i32 n_divide(((right - left - 1) / max_key_frame_separation) + 1);
				i32 frame_advance((right - left) / n_divide);
				while (left + frame_advance / 2 <= right) {
					feature_key_frames.push_back(left + frame_advance / 2);
					left += frame_advance;
				}
			}
			else {
				feature_key_frames.push_back((left + right) / 2);
			}
		}
		usize num_features(feature_key_frames.size());
		run_clip(
			clip_ctx,
			clip_features,
			fb.frames_gpu,
			fb.number_of_frames,
			feature_key_frames.data(),
			num_features,
			CLIP_WIDTH,
			CLIP_HEIGHT,
			scratch_f32_1,
			scratch_u8_1,
			scratch_i32_1,
			stream
		);
		clip_features.download_partial(clip_features_cpu, num_features * CLIP_FEATURE_SIZE, stream);
		cudaStreamSynchronize(stream);
		fb.ConsumeBatch();
		for (auto& idx : feature_key_frames) {
			idx += fb.frame_offset;
		}
		self->SubmitResponse(
			WorkResponse(
				self,
				fb.end_of_video,
				fb.number_of_frames,
				num_features,
				reinterpret_cast<u8 const* const>(clip_features_cpu.at_offset(0, 0)),
				sizeof(f32) * CLIP_FEATURE_SIZE,
				feature_key_frames.data(),
				sizeof(*feature_key_frames.data()),
				custom_data
			)
		);
	}

	delete transnet_ctx;
	delete clip_ctx;
}

};

void WorkResponse::ConsumeResponse() {
	std::lock_guard<std::mutex> guard(*associated_worker->free_workers_mutex);
	associated_worker->free_workers->push_back(associated_worker->worker_id);
	associated_worker->doing_work = false;
}