Delphi FFMPEG CUDA hardware accelerated video decoding generating shadows - delphi

Question

I'm trying to use cuda hardware accelerated video decoding, loading from a .mp4 file, according to the code bellow, but getting the image with shadows
Specification:
FFMPEG 4.3
Delphi 10.3
NVIDIA GeForce RTX 3050 Laptop GPU 4Gb DDR6
Intel Core i5 11400H 16 Gb RAM
unit UVideoDecoding;
interface
uses
Vcl.Graphics, Winapi.Windows,
System.SysUtils, System.Classes,
libavcodec, libavdevice,
libavfilter, libswresample, libavformat, libavutil,
libswscale;
type
TDecoder = class
protected
fFormatCtx: PAVFormatContext;
fCodec: PAVCodec;
fCodecParams, fLocalCodecParams: PAVCodecParameters;
fCodecCtx: libavcodec.PAVCodecContext;
fPacket: PAVPacket;
fFrameIn, fFrameOut: PAVFrame;
fScaler: PSwsContext;
fVideoStreamIndex: Integer;
fHWPixelFormat: AVPixelFormat;
fOutputWidth: Integer;
fOutputHeight: Integer;
function DecodeVideoPacket: Integer;
public
function Load(pFileName: string): Integer;
end;
implementation
function TDecoder.Load(pFileName: string): Integer;
var
vCount: Integer;
vLocalCodec: PAVCodec;
vFileName: AnsiString;
vType: AVHWDeviceType;
vDecPackResponse: Integer;
begin
vType := av_hwdevice_find_type_by_name('cuda');
{vType = AV_HWDEVICE_TYPE_CUDA}
fFormatCtx := avformat_alloc_context;
vFileName := AnsiString(pFileName);
avformat_open_input(fFormatCtx, PAnsiChar(vFileName), nil, nil);
avformat_find_stream_info(fFormatCtx, nil);
for vCount := 0 to fFormatCtx.nb_streams - 1 do
begin
fLocalCodecParams := fFormatCtx.streams[vCount].codecpar;
case fLocalCodecParams.codec_id of
AVCodecID.AV_CODEC_ID_H264: vLocalCodec := avcodec_find_decoder_by_name('h264_cuvid');
AVCodecID.AV_CODEC_ID_HEVC: vLocalCodec := avcodec_find_decoder_by_name('hevc_cuvid');
AVCodecID.AV_CODEC_ID_MJPEG: vLocalCodec := avcodec_find_decoder_by_name('mjpeg_cuvid');
AVCodecID.AV_CODEC_ID_MPEG4: vLocalCodec := avcodec_find_decoder_by_name('mpeg4_cuvid');
else
vLocalCodec := avcodec_find_decoder(fLocalCodecParams.codec_id);
end;
if fLocalCodecParams.codec_type = AVMEDIA_TYPE_VIDEO then
begin
fVideoStreamIndex := vCount;
fCodec := vLocalCodec;
fCodecParams := fLocalCodecParams;
end;
end;
{fCodec.name = h264_cuvid}
fCodecCtx := avcodec_alloc_context3(fCodec);
var vHWDeviceCtx: PAVBufferRef;
try
if av_hwdevice_ctx_create(vHWDeviceCtx, vType, nil, nil, 0) >= 0 then
fCodecCtx.hw_device_ctx := av_buffer_ref(vHWDeviceCtx);
finally
av_buffer_unref(vHWDeviceCtx);
end;
avcodec_open2(fCodecCtx, fCodec, nil);
{
fCodecCtx.codec_id = AV_CODEC_ID_H264
fCodecCtx.pix_fmt = AV_PIX_FMT_MMAL
}
fFrameIn := av_frame_alloc;
fFrameOut := av_frame_alloc;
fFrameOut.format := Integer(AV_PIX_FMT_BGRA);
fFrameOut.width := 640;
fFrameOut.height := 480;
//On getting sws_context I've tried srcFormat = AV_PIX_FMT_MMAL but the result was nil
fScaler := sws_getContext(fCodecCtx.Width, fCodecCtx.Height, AV_PIX_FMT_NV12{fCodecCtx.pix_fmt},
fOutputWidth, fOutputHeight, AV_PIX_FMT_BGRA, SWS_BILINEAR, nil, nil, nil);
fPacket := av_packet_alloc;
while (av_read_frame(fFormatCtx, fPacket) >= 0) do
begin
if fPacket.stream_index = fVideoStreamIndex then
begin
vDecPackResponse := DecodeVideoPacket;
av_packet_unref(fPacket);
if vDecPackResponse < 0 then
Break;
end
end;
Exit(0);
end;
function TDecoder.DecodeVideoPacket: Integer;
var
vResponse: Integer;
vBmp: Vcl.Graphics.TBitmap;
vScaleResult, vSize: Integer;
vBuffer: PByte;
vSWFrame, vTMPFrame: pAVFrame;
begin
Result := 0;
vResponse := avcodec_send_packet(fCodecCtx, fPacket);
if vResponse < 0 then
Exit(vResponse);
while (vResponse >= 0) do
begin
vResponse := avcodec_receive_frame(fCodecCtx, fFrameIn);
if (vResponse = AVERROR_EAGAIN) or (vResponse = AVERROR_EOF) then
Break
else if vResponse < 0 then
Exit(vResponse);
if vResponse >= 0 then
begin
vSWFrame := av_frame_alloc;
if av_hwframe_transfer_data(vSWFrame, fFrameIn, 0) >= 0 then
vTMPFrame := vSWFrame
else
vTMPFrame := fFrameIn;
vSize := av_image_get_buffer_size(AVPixelFormat(vTMPFrame.format), vTMPFrame.width, vTMPFrame.height, 1);
vBuffer := av_malloc(vSize);
if Assigned(vBuffer) then
av_image_copy_to_buffer(vBuffer, vSize, #vTMPFrame.data, #vTMPFrame.linesize, AVPixelFormat(vTMPFrame.format), vTMPFrame.width, vTMPFrame.height, 1);
MoveMemory(vTMPFrame.data[0], vBuffer, vSize);
vTMPFrame.data[0] := vTMPFrame.data[0] + vTMPFrame.linesize[0] * (fCodecCtx.height - 1);
vTMPFrame.linesize[0] := vTMPFrame.linesize[0] * -1;
vTMPFrame.data[1] := vTMPFrame.data[1] + vTMPFrame.linesize[1] * (fCodecCtx.height div 2 - 1);
vTMPFrame.linesize[1] := vTMPFrame.linesize[1] * -1;
vTMPFrame.data[2] := vTMPFrame.data[2] + vTMPFrame.linesize[2] * (fCodecCtx.height div 2 - 1);
vTMPFrame.linesize[2] := vTMPFrame.linesize[2] * -1;
vScaleResult := sws_scale(fScaler, #vTMPFrame.Data, #vTMPFrame.Linesize, 0,
fCodecCtx.Height, #fFrameOut.data, #fFrameOut.Linesize);
if vScaleResult <= 0 then
begin
Break;
end;
vBmp := Vcl.Graphics.TBitmap.Create;
try
vBmp.Height := fOutputHeight;
vBmp.Width := fOutputWidth;
vBmp.PixelFormat := TPixelFormat.pf32bit;
MoveMemory(PByte(vBmp.ScanLine[vBmp.Height -1]),
fFrameOut.data[0], fOutputHeight * fFrameOut.linesize[0]);
//Renders BMP on main thread
// if Assigned(fOnBitmap) then
// fOnBitmap(vBmp);
finally
av_frame_unref(vSWFrame);
av_frame_free(vTMPFrame);
av_freep(#vBuffer);
FreeAndNil(vBmp);
Result := 1;
end;
end;
end;
end;
end.
I've tried using different pixelformats but none of them properly decoded the video.