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

Use software decoder by default

Browse Source
This commit is contained in:
Tuan Nghia Nguyen
2026-04-04 20:19:54 +11:00
parent 3a21026790
commit e134ebdf15
24 changed files with 693 additions and 215 deletions
Download Patch File Download Diff File Expand all files Collapse all files

16
modules/ANSODEngine/ANSODEngine.cpp
View File

@@ -3,6 +3,7 @@
#include <cmath>
#include <json.hpp>
#include "ANSODEngine.h"
#include "ANSLicense.h" // ANS_DBG macro
#include "ANSYOLOOD.h"
#include "ANSTENSORRTOD.h"
#include "ANSTENSORRTCL.h"
@@ -879,6 +880,9 @@ namespace ANSCENTER
std::vector<Object> allResults;
allResults.clear();
try {
ANS_DBG("ODEngine", "SAHI START: %dx%d tile=%dx%d overlap=%.1f cam=%s",
input.cols, input.rows, tiledWidth, tiledHeight, overLap, camera_id.c_str());
auto _sahiStart = std::chrono::steady_clock::now();
cv::Mat image = input.clone();
if (image.empty() || !image.data || !image.u) {
return allResults;
@@ -920,6 +924,16 @@ namespace ANSCENTER
//4. Apply Non-Maximum Suppression (NMS) to merge overlapping results
float iouThreshold = 0.1;
std::vector<Object> finalResults = ANSUtilityHelper::ApplyNMS(allResults, iouThreshold);
{
double _sahiMs = std::chrono::duration<double, std::milli>(
std::chrono::steady_clock::now() - _sahiStart).count();
ANS_DBG("ODEngine", "SAHI DONE: %.1fms patches=%zu results=%zu cam=%s",
_sahiMs, patches.size() + 1, finalResults.size(), camera_id.c_str());
if (_sahiMs > 2000.0) {
ANS_DBG("ODEngine", "SAHI SLOW: %.1fms — %zu patches held _mutex entire time!",
_sahiMs, patches.size() + 1);
}
}
image.release();
return finalResults;
}
@@ -2103,6 +2117,8 @@ namespace ANSCENTER
// No coarse _mutex — sub-components (engines, trackers) have their own locks.
// LabVIEW semaphore controls concurrency at the caller level.
try {
ANS_DBG("ODEngine", "RunInferenceWithOption: cam=%s %dx%d mode=%s",
camera_id.c_str(), input.cols, input.rows, activeROIMode.c_str());
int mode = 0;
double confidenceThreshold = 0.35;
std::vector<int> trackingObjectIds;

19
modules/ANSODEngine/ANSRTYOLO.cpp
View File

@@ -1,5 +1,6 @@
#include "ANSRTYOLO.h"
#include "Utility.h"
#include "ANSLicense.h" // ANS_DBG macro for DebugView
#include <future>
#include <numeric>
#include <cmath>
@@ -903,7 +904,6 @@ namespace ANSCENTER {
return {};
}
// Check if model is classification (output ndims <= 2)
const auto& outputDims = m_trtEngine->getOutputDims();
const bool isClassification = !outputDims.empty() && outputDims[0].nbDims <= 2;
@@ -914,11 +914,8 @@ namespace ANSCENTER {
cv::cuda::GpuMat resized;
if (imgRGB.rows != inputH || imgRGB.cols != inputW) {
if (isClassification) {
// Classification: direct resize (no letterbox padding)
// Must use explicit stream to avoid conflict with CUDA Graph capture on null stream
cv::cuda::resize(imgRGB, resized, cv::Size(inputW, inputH), 0, 0, cv::INTER_LINEAR, stream);
} else {
// Detection/Seg/Pose/OBB: letterbox resize + right-bottom pad
resized = Engine<float>::resizeKeepAspectRatioPadRightBottom(imgRGB, inputH, inputW);
}
}
@@ -1831,8 +1828,7 @@ namespace ANSCENTER {
}
// --- 2. Preprocess under lock ---
// Try NV12 fast path first (12MB upload vs 24MB BGR for 4K)
// Falls back to standard GPU preprocessing if no NV12 data available.
ANS_DBG("YOLO", "Preprocess START %dx%d", inputImage.cols, inputImage.rows);
ImageMetadata meta;
std::vector<std::vector<cv::cuda::GpuMat>> input;
bool usedNV12 = false;
@@ -1874,11 +1870,22 @@ namespace ANSCENTER {
}
// --- 3. TRT Inference (mutex released for concurrent GPU slots) ---
ANS_DBG("YOLO", "TRT inference START nv12=%d inputSize=%dx%d",
(int)usedNV12,
input.empty() ? 0 : (input[0].empty() ? 0 : input[0][0].cols),
input.empty() ? 0 : (input[0].empty() ? 0 : input[0][0].rows));
auto _trtStart = std::chrono::steady_clock::now();
std::vector<std::vector<std::vector<float>>> featureVectors;
if (!m_trtEngine->runInference(input, featureVectors)) {
ANS_DBG("YOLO", "ERROR: TRT runInference FAILED");
_logger.LogError("ANSRTYOLO::DetectObjects", "Error running inference", __FILE__, __LINE__);
return {};
}
auto _trtEnd = std::chrono::steady_clock::now();
double _trtMs = std::chrono::duration<double, std::milli>(_trtEnd - _trtStart).count();
if (_trtMs > 500.0) {
ANS_DBG("YOLO", "SLOW TRT inference: %.1fms", _trtMs);
}
double msInference = dbg ? elapsed() : 0;
// --- 4. Transform output ---

1
modules/ANSODEngine/ANSRTYOLO.h
View File

@@ -81,6 +81,7 @@ namespace ANSCENTER {
std::vector<std::vector<cv::cuda::GpuMat>> PreprocessBatch(
const std::vector<cv::Mat>& inputImages, BatchMetadata& outMetadata);
// ── Detection pipeline ───────────────────────────────────────────
std::vector<Object> DetectObjects(const cv::Mat& inputImage,
const std::string& camera_id);

1
modules/ANSODEngine/NV12PreprocessHelper.cpp
View File

@@ -1,6 +1,7 @@
#include "NV12PreprocessHelper.h"
#include "ANSGpuFrameRegistry.h"
#include "ANSEngineCommon.h"
#include "ANSLicense.h" // ANS_DBG macro
#include <opencv2/cudaimgproc.hpp>
#include <opencv2/cudawarping.hpp>
#include <opencv2/core/cuda_stream_accessor.hpp>

16
modules/ANSODEngine/dllmain.cpp
View File

@@ -6,6 +6,7 @@
#include "engine/TRTEngineCache.h" // clearAll() on DLL_PROCESS_DETACH
#include "engine/EnginePoolManager.h" // clearAll() on DLL_PROCESS_DETACH
#include <climits> // INT_MIN
#include "ANSLicense.h" // ANS_DBG macro for DebugView
// Process-wide flag: when true, all engines force single-GPU path (no pool, no idle timers).
// Defined here, declared extern in EngineBuildLoadNetwork.inl.
@@ -1696,6 +1697,8 @@ static int RunInferenceComplete_LV_Impl(
auto* engine = guard.get();
try {
auto _t0 = std::chrono::steady_clock::now();
// Save/restore thread-local to support nested calls (custom model DLLs
// calling back into ANSODEngine via ANSLIB.dll).
GpuFrameData* savedFrame = tl_currentGpuFrame();
@@ -1708,6 +1711,10 @@ static int RunInferenceComplete_LV_Impl(
int originalWidth = localImage.cols;
int originalHeight = localImage.rows;
ANS_DBG("LV_Inference", "START cam=%s %dx%d gpuFrame=%p nv12=%s",
cameraId ? cameraId : "?", originalWidth, originalHeight,
(void*)gpuFrame, gpuFrame ? "YES" : "NO");
if (originalWidth == 0 || originalHeight == 0) {
tl_currentGpuFrame() = savedFrame;
return -2;
@@ -1717,8 +1724,17 @@ static int RunInferenceComplete_LV_Impl(
// Safe: *cvImage holds a refcount, keeping gpuFrame alive during inference.
// Only use OWN gpuFrame — never inherit outer caller's frame (dimension mismatch on crops).
tl_currentGpuFrame() = gpuFrame;
auto _t1 = std::chrono::steady_clock::now();
std::vector<ANSCENTER::Object> outputs = engine->RunInferenceWithOption(localImage, cameraId, activeROIMode);
auto _t2 = std::chrono::steady_clock::now();
tl_currentGpuFrame() = savedFrame;
double prepMs = std::chrono::duration<double, std::milli>(_t1 - _t0).count();
double infMs = std::chrono::duration<double, std::milli>(_t2 - _t1).count();
if (infMs > 500.0) {
ANS_DBG("LV_Inference", "SLOW cam=%s prep=%.1fms inf=%.1fms results=%zu",
cameraId ? cameraId : "?", prepMs, infMs, outputs.size());
}
bool getJpeg = (getJpegString == 1);
std::string stImage;
// NOTE: odMutex was removed here. All variables in this scope are local

5
modules/ANSODEngine/engine.h
View File

@@ -402,6 +402,9 @@ private:
cudaStream_t m_memoryStream; // ADD THIS - separate stream for memory operations
std::vector<cv::cuda::GpuMat> m_preprocessedInputs; // Keep inputs alive
// Note: blobFromGpuMats and resizeKeepAspectRatioPadRightBottom are static,
// so cached buffers use thread_local inside the functions themselves.
// Thermal management (ADD THESE)
//int m_consecutiveInferences;
@@ -431,7 +434,7 @@ private:
Logger m_logger;
bool m_verbose{ true }; // false for non-probe pool slots
bool m_disableGraphs{ false }; // true for pool slots — concurrent graph captures corrupt CUDA context
bool m_disableGraphs{ true }; // DISABLED by default — concurrent graph launches + uploads cause GPU deadlock on WDDM
// -- Multi-GPU pool data ---------------------------------------------------