cv::va_intel::convertFromVASurface always fails on AMD devices using mesa #21536
Description
System information (version)
- OpenCV => 4.5.5
- Operating System / Platform => Slackware-current
Detailed description
cv::va_intel::convertFromVASurface always fails on AMD devices using mesa
Steps to reproduce
I am trying to convert a VAAPI image from FFMPEG in an AVFrame to a UMat, so my code is like this:
bool MotionOpenCV::process_frame(const AVFrame *frame, bool video){
//check if the incoming frame is VAAPI
if (video && frame->format == AV_PIX_FMT_VAAPI && frame->hw_frames_ctx){
AVHWFramesContext *hw_frames = (AVHWFramesContext *)frame->hw_frames_ctx->data;
AVHWDeviceContext *device_ctx = hw_frames->device_ctx;
AVVAAPIDeviceContext *hwctx;
if (device_ctx->type != AV_HWDEVICE_TYPE_VAAPI){
LERROR("VAAPI frame does not have VAAPI device");
return false;
}
hwctx = static_cast<AVVAAPIDeviceContext*>(device_ctx->hwctx);
cv::UMat tmp;
const VASurfaceID surf = reinterpret_cast<uintptr_t const>(frame->data[3]);
try {
cv::va_intel::convertFromVASurface(hwctx->display, surf, cv::Size(frame->width, frame->height), tmp);
} catch (cv::Exception &e){
LWARN("Error converting image from VA-API To OpenCV: " + e.msg);
return false;
}
Digging into it, modules/core/src/va_intel.cpp calls vaDeriveImage which if using mesa, calls vlVaDeriveImage. but per the comment in mesa, that always fails on AMD devices unless your program name is "vlc", "h264encode", or "hevcencode"
/* This function is used by some programs to test for hardware decoding, but on
* AMD devices, the buffers default to interlaced, which causes this function to fail.
* Some programs expect this function to fail, while others, assume this means
* hardware acceleration is not available and give up without trying the fall-back
* vaCreateImage + vaPutImage
*/
https://gitlab.freedesktop.org/mesa/mesa/-/blob/main/src/gallium/frontends/va/image.c#L213
The solution then, is modules/core/src/va_intel.cpp should have the vaCreateImage + vaPutImage fallback that mesa requires.