ANSCENTER/ANSCORE
2
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Fix double stop in ANSVideoPlayer

Browse Source
This commit is contained in:
Tuan Nghia Nguyen
2026-04-22 10:10:16 +10:00
parent 97d814936d
commit 57cc8e0a56
14 changed files with 492 additions and 70 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
engines/TensorRTAPI/include/engine/EngineBuildLoadNetwork.inl
View File

@@ -1167,7 +1167,12 @@ trt_cache_create_context:
// -- Pinned output buffers (CUDA graph prerequisite) -----------------------
// Invalidate any graphs captured by a previous loadNetwork() call on this instance.
for (auto& [bs, ge] : m_graphExecs) { if (ge) cudaGraphExecDestroy(ge); }
for (auto& [bs, ge] : m_graphExecs) {
if (ge) {
cudaGraphExecDestroy(ge);
m_trtGraphDestroys.fetch_add(1, std::memory_order_relaxed);
}
}
m_graphExecs.clear();
// Free any previously allocated pinned buffers.
for (T* p : m_pinnedOutputBuffers) { if (p) cudaFreeHost(p); }

36
engines/TensorRTAPI/include/engine/EngineRunInference.inl
View File

@@ -731,7 +731,10 @@ bool Engine<T>::runInference(const std::vector<std::vector<cv::cuda::GpuMat>>& i
if (!m_graphExecs.empty()) {
size_t destroyed = m_graphExecs.size();
for (auto& [bs, ge] : m_graphExecs) {
if (ge) cudaGraphExecDestroy(ge);
if (ge) {
cudaGraphExecDestroy(ge);
m_trtGraphDestroys.fetch_add(1, std::memory_order_relaxed);
}
}
m_graphExecs.clear();
ANS_DBG("TRT_Engine", "INVALIDATED %zu cached CUDA graphs after shape change (batch=%d)",
@@ -901,8 +904,10 @@ bool Engine<T>::runInference(const std::vector<std::vector<cv::cuda::GpuMat>>& i
if (captureOk) {
cudaGraphExec_t exec = nullptr;
if (cudaGraphInstantiate(&exec, graph, nullptr, nullptr, 0) == cudaSuccess)
if (cudaGraphInstantiate(&exec, graph, nullptr, nullptr, 0) == cudaSuccess) {
graphExec = exec;
m_trtGraphCreates.fetch_add(1, std::memory_order_relaxed);
}
cudaGraphDestroy(graph);
ANS_DBG("TRT_Engine", "CUDA graph CAPTURED OK for batch=%d exec=%p",
batchSize, (void*)graphExec);
@@ -1053,5 +1058,32 @@ bool Engine<T>::runInference(const std::vector<std::vector<cv::cuda::GpuMat>>& i
s_globalActiveInf.fetch_sub(1);
}
// Leak diagnostic — one [TRT_Leak] line per engine instance per 60 s.
// Reports CUDA graph create/destroy balance and current cache size.
// If (creates - destroys) climbs monotonically, graph execs are being
// leaked on every shape change; each leaked exec is tens of MB.
// Lock-free window claim via compare_exchange — concurrent inference
// threads race to log but only one wins per 60-s window.
{
using clk = std::chrono::steady_clock;
const long long tick = clk::now().time_since_epoch().count();
long long expected = m_trtLeakNextLogTick.load(std::memory_order_relaxed);
if (tick >= expected) {
const long long deadline = tick +
std::chrono::duration_cast<clk::duration>(
std::chrono::seconds(60)).count();
if (m_trtLeakNextLogTick.compare_exchange_strong(
expected, deadline, std::memory_order_relaxed)) {
const int64_t cr = m_trtGraphCreates.load(std::memory_order_relaxed);
const int64_t ds = m_trtGraphDestroys.load(std::memory_order_relaxed);
ANS_DBG("TRT_Leak",
"engine=%p creates=%lld destroys=%lld net=%lld cached=%zu",
(void*)this,
(long long)cr, (long long)ds, (long long)(cr - ds),
m_graphExecs.size());
}
}
}
return true;
}