Fix AMD ALPR_OCR and custom model

This commit is contained in:
2026-05-01 07:53:18 +10:00
parent e1f37ec767
commit 91efbb9a8f
25 changed files with 2841 additions and 429 deletions
+29 -12
View File
@@ -1,6 +1,7 @@
#include "ANSSRT.h"
#include "ANSMatRegistry.h"
#include "ANSGpuFrameOps.h"
#include "av_frame_ptr.h" // ansmedia::AvFramePtr — RAII for AVFrame*
#include "ANSCVVendorGate.h" // anscv_vendor_gate::IsNvidiaGpuAvailable()
#include "ANSLicense.h" // ANS_DBG macro
#include <memory>
@@ -826,17 +827,25 @@ extern "C" __declspec(dllexport) int GetSRTCVImage(ANSCENTER::ANSSRTClient** Han
// Thread-safe Mat pointer swap (anscv_mat_replace has its own internal lock)
anscv_mat_replace(image, std::move(img));
// NV12 GPU fast path (optional — disabled by default for stability)
// NV12 GPU fast path (optional — disabled by default for stability).
// AvFramePtr auto-frees clones if we don't reach the .release() handoff.
if ((*Handle)->IsNV12FastPath()) {
int gpuIdx = (*Handle)->GetHWDecodingGpuIndex();
AVFrame* cudaHW = (*Handle)->GetCudaHWFrame();
ansmedia::AvFramePtr cudaHW{ (*Handle)->GetCudaHWFrame() };
if (cudaHW) {
AVFrame* cpuNV12 = (*Handle)->GetNV12Frame();
gpu_frame_attach_cuda(*image, cudaHW, gpuIdx, timeStamp, cpuNV12);
ansmedia::AvFramePtr cpuNV12{ (*Handle)->GetNV12Frame() };
ANS_DBG("MEDIA_AvPtr",
"ANSSRT V1 attach_cuda cudaHW=%p cpuNV12=%p gpuIdx=%d",
(void*)cudaHW.get(), (void*)cpuNV12.get(), gpuIdx);
gpu_frame_attach_cuda(*image, cudaHW.release(), gpuIdx, timeStamp,
cpuNV12.release());
} else {
AVFrame* nv12 = (*Handle)->GetNV12Frame();
ansmedia::AvFramePtr nv12{ (*Handle)->GetNV12Frame() };
if (nv12) {
gpu_frame_attach(*image, nv12, gpuIdx, timeStamp);
ANS_DBG("MEDIA_AvPtr",
"ANSSRT V1 attach nv12=%p gpuIdx=%d",
(void*)nv12.get(), gpuIdx);
gpu_frame_attach(*image, nv12.release(), gpuIdx, timeStamp);
}
}
}
@@ -1077,16 +1086,24 @@ extern "C" __declspec(dllexport) int GetSRTCVImage_V2(uint64_t handleVal, int& w
anscv_mat_replace(image, std::move(img));
// Attach NV12 frame for GPU fast-path inference (side-table registry)
// Attach NV12 frame for GPU fast-path inference (side-table registry).
// AvFramePtr auto-frees clones if we don't reach the .release() handoff.
int gpuIdx = h->GetHWDecodingGpuIndex();
AVFrame* cudaHW = h->GetCudaHWFrame();
ansmedia::AvFramePtr cudaHW{ h->GetCudaHWFrame() };
if (cudaHW) {
AVFrame* cpuNV12 = h->GetNV12Frame();
gpu_frame_attach_cuda(*image, cudaHW, gpuIdx, timeStamp, cpuNV12);
ansmedia::AvFramePtr cpuNV12{ h->GetNV12Frame() };
ANS_DBG("MEDIA_AvPtr",
"ANSSRT V2 attach_cuda cudaHW=%p cpuNV12=%p gpuIdx=%d",
(void*)cudaHW.get(), (void*)cpuNV12.get(), gpuIdx);
gpu_frame_attach_cuda(*image, cudaHW.release(), gpuIdx, timeStamp,
cpuNV12.release());
} else {
AVFrame* nv12 = h->GetNV12Frame();
ansmedia::AvFramePtr nv12{ h->GetNV12Frame() };
if (nv12) {
gpu_frame_attach(*image, nv12, gpuIdx, timeStamp);
ANS_DBG("MEDIA_AvPtr",
"ANSSRT V2 attach nv12=%p gpuIdx=%d",
(void*)nv12.get(), gpuIdx);
gpu_frame_attach(*image, nv12.release(), gpuIdx, timeStamp);
}
}