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

Disable NV12 path for ANSCV by default. Currenly use cv::Mat** directly

Browse Source
This commit is contained in:
Tuan Nghia Nguyen
2026-04-04 10:09:47 +11:00
parent 445abefebe
commit 3a21026790
19 changed files with 575 additions and 232 deletions
Download Patch File Download Diff File Expand all files Collapse all files

28
modules/ANSODEngine/NV12PreprocessHelper.cpp
View File

@@ -269,6 +269,15 @@ namespace ANSCENTER {
return result;
}
// Ensure async D2D copy (NVDEC → pool buffer) has completed before
// reading yPlane/uvPlane. The copy was queued in gpu_frame_attach_cuda()
// on a non-blocking stream. By the time inference runs (~50-200ms later),
// the copy (~0.3ms) has long finished, so this sync returns immediately.
if (gpuData->d2dCopyStream) {
cudaStreamSynchronize(static_cast<cudaStream_t>(gpuData->d2dCopyStream));
gpuData->d2dCopyStream = nullptr; // Only sync once per frame
}
const bool isCudaDevice = gpuData->isCudaDevicePtr;
const bool gpuMatch = !isCudaDevice ||
gpuData->gpuIndex < 0 ||
@@ -367,7 +376,6 @@ namespace ANSCENTER {
cv::cuda::GpuMat gpuY, gpuUV;
if (useZeroCopy) {
// CUDA zero-copy: wrap pool buffer device pointers directly
gpuY = cv::cuda::GpuMat(frameH, frameW, CV_8UC1,
effYPlane, static_cast<size_t>(effYLinesize));
gpuUV = cv::cuda::GpuMat(frameH / 2, frameW, CV_8UC1,
@@ -641,6 +649,12 @@ namespace ANSCENTER {
return result;
}
// Ensure async D2D copy has completed before reading NV12 buffers
if (gpuData->d2dCopyStream) {
cudaStreamSynchronize(static_cast<cudaStream_t>(gpuData->d2dCopyStream));
gpuData->d2dCopyStream = nullptr;
}
const bool isCudaDevice = gpuData->isCudaDevicePtr;
const bool gpuMatch = !isCudaDevice ||
gpuData->gpuIndex < 0 ||
@@ -775,6 +789,12 @@ namespace ANSCENTER {
if (!gpuData->isCudaDevicePtr || !gpuData->yPlane || !gpuData->uvPlane)
return result; // NV12 not on GPU
// Ensure async D2D copy has completed before reading NV12 buffers
if (gpuData->d2dCopyStream) {
cudaStreamSynchronize(static_cast<cudaStream_t>(gpuData->d2dCopyStream));
gpuData->d2dCopyStream = nullptr;
}
const int frameW = gpuData->width;
const int frameH = gpuData->height;
@@ -890,6 +910,12 @@ namespace ANSCENTER {
return result;
}
// Ensure async D2D copy has completed before reading NV12 buffers
if (gpuData->d2dCopyStream) {
cudaStreamSynchronize(static_cast<cudaStream_t>(gpuData->d2dCopyStream));
gpuData->d2dCopyStream = nullptr;
}
const bool isCudaDevice = gpuData->isCudaDevicePtr;
const bool gpuMatch = !isCudaDevice ||
gpuData->gpuIndex < 0 ||