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

Improve ANSCV

Browse Source
This commit is contained in:
Tuan Nghia Nguyen
2026-04-21 09:26:02 +10:00
parent 9f0a10a4c8
commit 7e772f76bc
15 changed files with 749 additions and 421 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
modules/ANSOCR/ANSRTOCR/RTOCRRecognizer.cpp
View File

@@ -28,8 +28,11 @@ bool RTOCRRecognizer::Initialize(const std::string& onnxPath, const std::string&
ANSCENTER::Options options;
options.deviceIndex = gpuId;
options.precision = ANSCENTER::Precision::FP16;
options.maxBatchSize = 1;
options.optBatchSize = 1;
// maxBatch=4 matches FaceRecognizer / ALPR configuration — allows the
// recognizer to process up to 4 detected text lines in one call,
// amortizing per-invocation overhead while keeping TRT workspace small.
options.maxBatchSize = 4;
options.optBatchSize = 4;
// Fixed height, dynamic width for recognition
options.minInputHeight = imgH_;

40
modules/ANSOCR/dllmain.cpp
View File

@@ -185,11 +185,22 @@ extern "C" ANSOCR_API int CreateANSOCRHandleEx(ANSCENTER::ANSOCRBase** Handle,
ANSCENTER::ANSLibsLoader::Initialize();
ANSCENTER::EngineType engineType = ANSCENTER::ANSLicenseHelper::CheckHardwareInformation();
{
// Describe the backend the engine-selector below will actually choose
// for this (hardware, engineMode) combination. Previous versions of
// this log claimed "TensorRT OCR enabled" based on hardware alone,
// which was misleading because engineMode=0 (auto) unconditionally
// picked ONNX — users saw the log and assumed TRT was running.
const bool isNvidia = (engineType == ANSCENTER::EngineType::NVIDIA_GPU);
const bool willUseTRT =
isNvidia && (engineMode == 0 /* auto → TRT on NVIDIA */ ||
engineMode == 1 /* GPU → TRT on NVIDIA */);
const char* vendorTag =
engineType == ANSCENTER::EngineType::NVIDIA_GPU ? "NVIDIA_GPU (TensorRT OCR enabled)" :
engineType == ANSCENTER::EngineType::AMD_GPU ? "AMD_GPU (ONNX Runtime / DirectML, TensorRT OCR DISABLED)" :
engineType == ANSCENTER::EngineType::OPENVINO_GPU ? "OPENVINO_GPU (ONNX Runtime / OpenVINO, TensorRT OCR DISABLED)" :
"CPU (ONNX Runtime, TensorRT OCR DISABLED)";
engineType == ANSCENTER::EngineType::NVIDIA_GPU
? (willUseTRT ? "NVIDIA_GPU (TensorRT OCR active)"
: "NVIDIA_GPU (TensorRT available, but engineMode forces ONNX)")
: engineType == ANSCENTER::EngineType::AMD_GPU ? "AMD_GPU (ONNX Runtime / DirectML, TensorRT OCR unavailable)"
: engineType == ANSCENTER::EngineType::OPENVINO_GPU ? "OPENVINO_GPU (ONNX Runtime / OpenVINO, TensorRT OCR unavailable)"
: "CPU (ONNX Runtime, TensorRT OCR unavailable)";
char buf[192];
snprintf(buf, sizeof(buf),
"[ANSOCR] CreateANSOCRHandleEx: detected engineType=%d [%s], engineMode=%d\n",
@@ -230,10 +241,23 @@ extern "C" ANSOCR_API int CreateANSOCRHandleEx(ANSCENTER::ANSOCRBase** Handle,
// select, including DirectML for AMD).
const bool isNvidia = (engineType == ANSCENTER::EngineType::NVIDIA_GPU);
switch (engineMode) {
case 0:// Auto-detect, always use ONNX for better compatibility, especially on AMD GPUs and high-res images
(*Handle) = new ANSCENTER::ANSONNXOCR();
case 0: // Auto-detect — prefer TensorRT on NVIDIA, ONNX elsewhere.
// Previous policy was "always ONNX" for cross-platform safety,
// but on NVIDIA that defeated the point: each ANSONNXOCR handle
// allocates its own cls/dec/rec OrtSessions (no dedupe), which
// wasted ~300–600 MB VRAM per extra instance and ran ~2× slower
// than ANSRTOCR's shared-engine path via EnginePoolManager.
if (isNvidia) {
limitSideLen = 960;
(*Handle) = new ANSCENTER::ANSRTOCR();
} else {
// AMD / Intel / CPU — ANSRTOCR hard-requires CUDA and would
// crash. ANSONNXOCR auto-picks the correct ORT EP
// (DirectML on AMD, OpenVINO on Intel, CPU otherwise).
(*Handle) = new ANSCENTER::ANSONNXOCR();
}
break;
case 1:// GPU — use TensorRT engine ONLY on NVIDIA hardware.
case 1: // GPU — use TensorRT engine ONLY on NVIDIA hardware.
if (isNvidia) {
limitSideLen = 960;
(*Handle) = new ANSCENTER::ANSRTOCR();
@@ -244,7 +268,7 @@ extern "C" ANSOCR_API int CreateANSOCRHandleEx(ANSCENTER::ANSOCRBase** Handle,
(*Handle) = new ANSCENTER::ANSONNXOCR();
}
break;
case 2:// CPU
case 2: // CPU
(*Handle) = new ANSCENTER::ANSONNXOCR();
break;
default: