void CPushPinDesktop::CopyScreenToDataBlock(HDC hScrDC, BYTE *pData, BITMAPINFO *pHeader, IMediaSample *pSample)
{
HDC hMemDC; // screen DC and memory DC
HBITMAP hOldBitmap; // handles to device-dependent bitmaps
int nX, nY; // coordinates of rectangle to grab
int iFinalStretchHeight = getNegotiatedFinalHeight();
int iFinalStretchWidth = getNegotiatedFinalWidth();
ASSERT(!IsRectEmpty(&m_rScreen)); // that would be unexpected
// create a DC for the screen and create
// a memory DC compatible to screen DC
hMemDC = CreateCompatibleDC(hScrDC); // 0.02ms Anything else to reuse, this one's pretty fast...?
// determine points of where to grab from it, though I think we control these with m_rScreen
nX = m_rScreen.left;
nY = m_rScreen.top;
// sanity checks--except we don't want it apparently, to allow upstream to dynamically change the size? Can it do that?
ASSERT(m_rScreen.bottom - m_rScreen.top == iFinalStretchHeight);
ASSERT(m_rScreen.right - m_rScreen.left == iFinalStretchWidth);
// select new bitmap into memory DC
hOldBitmap = (HBITMAP) SelectObject(hMemDC, hRawBitmap);
doJustBitBltOrScaling(hMemDC, m_iCaptureConfigWidth, m_iCaptureConfigHeight, iFinalStretchWidth, iFinalStretchHeight, hScrDC, nX, nY);
AddMouse(hMemDC, &m_rScreen, hScrDC, m_iHwndToTrack);
// select old bitmap back into memory DC and get handle to
// bitmap of the capture...whatever this even means...
HBITMAP hRawBitmap2 = (HBITMAP) SelectObject(hMemDC, hOldBitmap);
BITMAPINFO tweakableHeader;
memcpy(&tweakableHeader, pHeader, sizeof(BITMAPINFO));
if(m_bConvertToI420) {
tweakableHeader.bmiHeader.biBitCount = 32;
tweakableHeader.bmiHeader.biCompression = BI_RGB;
tweakableHeader.bmiHeader.biHeight = -tweakableHeader.bmiHeader.biHeight; // prevent upside down conversion from i420...
tweakableHeader.bmiHeader.biSizeImage = GetBitmapSize(&tweakableHeader.bmiHeader);
}
if(m_bConvertToI420) {
// copy it to a temporary buffer first
doDIBits(hScrDC, hRawBitmap2, iFinalStretchHeight, pOldData, &tweakableHeader);
// memcpy(/* dest */ pOldData, pData, pSample->GetSize()); // 12.8ms for 1920x1080 desktop
// TODO smarter conversion/memcpy's here [?] we could combine scaling with rgb32_to_i420 for instance...
// or maybe we should integrate with libswscale here so they can request whatever they want LOL. (might be a higher quality i420 conversion...)
// now convert it to i420 into the "real" buffer
rgb32_to_i420(iFinalStretchWidth, iFinalStretchHeight, (const char *) pOldData, (char *) pData);// took 36.8ms for 1920x1080 desktop
} else {
doDIBits(hScrDC, hRawBitmap2, iFinalStretchHeight, pData, &tweakableHeader);
}
// clean up
DeleteDC(hMemDC);
}
//
// DecideBufferSize
//
// This will always be called after the format has been sucessfully
// negotiated (this is negotiatebuffersize). So we have a look at m_mt to see what size image we agreed.
// Then we can ask for buffers of the correct size to contain them.
//
HRESULT CPushPinDesktop::DecideBufferSize(IMemAllocator *pAlloc,
ALLOCATOR_PROPERTIES *pProperties)
{
CheckPointer(pAlloc,E_POINTER);
CheckPointer(pProperties,E_POINTER);
CAutoLock cAutoLock(m_pFilter->pStateLock());
HRESULT hr = NOERROR;
VIDEOINFO *pvi = (VIDEOINFO *) m_mt.Format();
BITMAPINFOHEADER header = pvi->bmiHeader;
ASSERT_RETURN(header.biPlanes == 1); // sanity check
// ASSERT_RAISE(header.biCompression == 0); // meaning "none" sanity check, unless we are allowing for BI_BITFIELDS [?] so leave commented out for now
// now try to avoid this crash [XP, VLC 1.1.11]: vlc -vvv dshow:// :dshow-vdev="screen-capture-recorder" :dshow-adev --sout "#transcode{venc=theora,vcodec=theo,vb=512,scale=0.7,acodec=vorb,ab=128,channels=2,samplerate=44100,audio-sync}:standard{access=file,mux=ogg,dst=test.ogv}" with 10x10 or 1000x1000
// LODO check if biClrUsed is passed in right for 16 bit [I'd guess it is...]
// pProperties->cbBuffer = pvi->bmiHeader.biSizeImage; // too small. Apparently *way* too small.
int bytesPerLine;
// there may be a windows method that would do this for us...GetBitmapSize(&header); but might be too small for VLC? LODO try it :)
// some pasted code...
int bytesPerPixel = (header.biBitCount/8);
if(m_bConvertToI420) {
bytesPerPixel = 32/8; // we convert from a 32 bit to i420, so need more space in this case
}
bytesPerLine = header.biWidth * bytesPerPixel;
/* round up to a dword boundary for stride */
if (bytesPerLine & 0x0003)
{
bytesPerLine |= 0x0003;
++bytesPerLine;
}
ASSERT_RETURN(header.biHeight > 0); // sanity check
ASSERT_RETURN(header.biWidth > 0); // sanity check
// NB that we are adding in space for a final "pixel array" (http://en.wikipedia.org/wiki/BMP_file_format#DIB_Header_.28Bitmap_Information_Header.29) even though we typically don't need it, this seems to fix the segfaults
// maybe somehow down the line some VLC thing thinks it might be there...weirder than weird.. LODO debug it LOL.
int bitmapSize = 14 + header.biSize + (long)(bytesPerLine)*(header.biHeight) + bytesPerLine*header.biHeight;
pProperties->cbBuffer = bitmapSize;
//pProperties->cbBuffer = max(pProperties->cbBuffer, m_mt.GetSampleSize()); // didn't help anything
if(m_bConvertToI420) {
pProperties->cbBuffer = header.biHeight * header.biWidth*3/2; // necessary to prevent an "out of memory" error for FMLE. Yikes. Oh wow yikes.
}
pProperties->cBuffers = 1; // 2 here doesn't seem to help the crashes...
// Ask the allocator to reserve us some sample memory. NOTE: the function
// can succeed (return NOERROR) but still not have allocated the
// memory that we requested, so we must check we got whatever we wanted.
ALLOCATOR_PROPERTIES Actual;
hr = pAlloc->SetProperties(pProperties,&Actual);
if(FAILED(hr))
{
return hr;
}
// Is this allocator unsuitable?
if(Actual.cbBuffer < pProperties->cbBuffer)
{
return E_FAIL;
}
// now some "once per run" setups
// LODO reset aer with each run...somehow...somehow...Stop method or something...
OSVERSIONINFOEX version;
ZeroMemory(&version, sizeof(OSVERSIONINFOEX));
version.dwOSVersionInfoSize = sizeof(OSVERSIONINFOEX);
GetVersionEx((LPOSVERSIONINFO)&version);
if(version.dwMajorVersion >= 6) { // meaning vista +
if(read_config_setting(TEXT("disable_aero_for_vista_plus_if_1"), 0, true) == 1) {
printf("turning aero off/disabling aero");
turnAeroOn(false);
}
else {
printf("leaving aero on");
turnAeroOn(true);
}
}
if(pOldData) {
free(pOldData);
pOldData = NULL;
}
pOldData = (BYTE *) malloc(max(pProperties->cbBuffer*pProperties->cBuffers, bitmapSize)); // we convert from a 32 bit to i420, so need more space, hence max
memset(pOldData, 0, pProperties->cbBuffer*pProperties->cBuffers); // reset it just in case :P
// create a bitmap compatible with the screen DC
if(hRawBitmap)
DeleteObject (hRawBitmap); // delete the old one in case it exists...
hRawBitmap = CreateCompatibleBitmap(hScrDc, getNegotiatedFinalWidth(), getNegotiatedFinalHeight());
previousFrameEndTime = 0; // reset
m_iFrameNumber = 0;
return NOERROR;
} // DecideBufferSize
void CPushPinDesktop::CopyScreenToDataBlock(HDC hScrDC, BYTE *pData, BITMAPINFO *pHeader, IMediaSample *pSample)
{
HDC hMemDC; // screen DC and memory DC
HBITMAP hOldBitmap; // handles to device-dependent bitmaps
int nX, nY; // coordinates of rectangle to grab
int iFinalStretchHeight = getNegotiatedFinalHeight();
int iFinalStretchWidth = getNegotiatedFinalWidth();
ASSERT_RAISE(!IsRectEmpty(&m_rScreen)); // that would be unexpected
// create a DC for the screen and create
// a memory DC compatible to screen DC
hMemDC = CreateCompatibleDC(hScrDC); // 0.02ms Anything else to reuse, this one's pretty fast...?
// determine points of where to grab from it, though I think we control these with m_rScreen
nX = m_rScreen.left;
nY = m_rScreen.top;
// sanity checks--except we don't want it apparently, to allow upstream to dynamically change the size? Can it do that?
ASSERT_RAISE(m_rScreen.bottom - m_rScreen.top == iFinalStretchHeight);
ASSERT_RAISE(m_rScreen.right - m_rScreen.left == iFinalStretchWidth);
// select new bitmap into memory DC
hOldBitmap = (HBITMAP) SelectObject(hMemDC, hRawBitmap);
doJustBitBltOrScaling(hMemDC, m_iCaptureConfigWidth, m_iCaptureConfigHeight, iFinalStretchWidth, iFinalStretchHeight, hScrDC, nX, nY);
if(m_bCaptureMouse)
AddMouse(hMemDC, &m_rScreen, hScrDC, m_iHwndToTrack);
// select old bitmap back into memory DC and get handle to
// bitmap of the capture...whatever this even means...
HBITMAP hRawBitmap2 = (HBITMAP) SelectObject(hMemDC, hOldBitmap);
BITMAPINFO tweakableHeader;
memcpy(&tweakableHeader, pHeader, sizeof(BITMAPINFO));
if(m_bConvertToI420) {
tweakableHeader.bmiHeader.biBitCount = 32;
tweakableHeader.bmiHeader.biCompression = BI_RGB;
tweakableHeader.bmiHeader.biHeight = -tweakableHeader.bmiHeader.biHeight; // prevent upside down conversion from i420...
tweakableHeader.bmiHeader.biSizeImage = GetBitmapSize(&tweakableHeader.bmiHeader);
}
if(m_bConvertToI420) {
// copy it to a temporary buffer first
doDIBits(hScrDC, hRawBitmap2, iFinalStretchHeight, pOldData, &tweakableHeader);
// memcpy(/* dest */ pOldData, pData, pSample->GetSize()); // 12.8ms for 1920x1080 desktop
// TODO smarter conversion/memcpy's here [?] we could combine scaling with rgb32_to_i420 for instance...
// or maybe we should integrate with libswscale here so they can request whatever they want LOL. (might be a higher quality i420 conversion...)
// now convert it to i420 into the "real" buffer
rgb32_to_i420(iFinalStretchWidth, iFinalStretchHeight, (const char *) pOldData, (char *) pData);// took 36.8ms for 1920x1080 desktop
} else {
doDIBits(hScrDC, hRawBitmap2, iFinalStretchHeight, pData, &tweakableHeader);
// if we're on vlc work around for odd pixel widths and 24 bit...<sigh>, like a width of 134 breaks vlc with 24bit. wow. see also GetMediaType comments
wchar_t buffer[MAX_PATH + 1]; // on the stack
GetModuleFileName(NULL, buffer, MAX_PATH);
if(wcsstr(buffer, L"vlc.exe") > 0) {
int bitCount = tweakableHeader.bmiHeader.biBitCount;
int stride = (iFinalStretchWidth * (bitCount / 8)) % 4; // see if lines have some padding at the end...
//int stride2 = (tweakableHeader.bmiHeader.biWidth * (tweakableHeader.bmiHeader.biBitCount / 8) + 3) & ~3; // ??
if(stride > 0) {
stride = 4 - stride; // they round up to 4 word boundary
// don't need to copy the first line :P
int lineSizeBytes = iFinalStretchWidth*(bitCount/8);
int lineSizeTotal = lineSizeBytes + stride;
for(int line = 1; line < iFinalStretchHeight; line++) {
//*dst, *src, size
// memmove required since these overlap...
memmove(&pData[line*lineSizeBytes], &pData[line*lineSizeTotal], lineSizeBytes);
}
}
}
}
// clean up
DeleteDC(hMemDC);
}
HRESULT CPushPinDesktop::FillBuffer(IMediaSample *pSample)
{
LocalOutput("video frame requested");
__int64 startThisRound = StartCounter();
BYTE *pData;
CheckPointer(pSample, E_POINTER);
if(m_bReReadRegistry) {
reReadCurrentStartXY(1);
}
if(!ever_started) {
// allow it to startup until Run is called...so StreamTime can work see http://stackoverflow.com/questions/2469855/how-to-get-imediacontrol-run-to-start-a-file-playing-with-no-delay/2470548#2470548
// since StreamTime anticipates that the graph's start time has already been set
FILTER_STATE myState;
CSourceStream::m_pFilter->GetState(INFINITE, &myState);
while(myState != State_Running) {
// TODO accomodate for pausing better, we're single run only currently [does VLC do pausing even?]
Sleep(1);
LocalOutput("sleeping till graph running for audio...");
m_pParent->GetState(INFINITE, &myState);
}
ever_started = true;
}
// Access the sample's data buffer
pSample->GetPointer(&pData);
// Make sure that we're still using video format
ASSERT_RETURN(m_mt.formattype == FORMAT_VideoInfo);
VIDEOINFOHEADER *pVih = (VIDEOINFOHEADER*) m_mt.pbFormat;
boolean gotNew = false; // dedupe stuff
while(!gotNew) {
CopyScreenToDataBlock(hScrDc, pData, (BITMAPINFO *) &(pVih->bmiHeader), pSample);
if(m_bDeDupe) {
if(memcmp(pData, pOldData, pSample->GetSize())==0) { // took desktop: 10ms for 640x1152, still 100 fps uh guess...
Sleep(m_millisToSleepBeforePollForChanges);
} else {
gotNew = true;
memcpy( /* dest */ pOldData, pData, pSample->GetSize()); // took 4ms for 640x1152, but it's worth it LOL.
// LODO memcmp and memcpy in the same loop LOL.
}
} else {
// it's always new for everyone else (the typical case)
gotNew = true;
}
}
// capture some debug stats (how long it took) before we add in our own arbitrary delay to enforce fps...
long double millisThisRoundTook = GetCounterSinceStartMillis(startThisRound);
fastestRoundMillis = min(millisThisRoundTook, fastestRoundMillis); // keep stats :)
sumMillisTook += millisThisRoundTook;
CRefTime now;
CRefTime endFrame;
now = 0;
CSourceStream::m_pFilter->StreamTime(now);
if((now > 0) && (now < previousFrameEndTime)) { // now > 0 to accomodate for if there is no reference graph clock at all...also at boot strap time to ignore it XXXX can negatives even ever happen anymore though?
while(now < previousFrameEndTime) { // guarantees monotonicity too :P
LocalOutput("sleeping because %llu < %llu", now, previousFrameEndTime);
Sleep(1);
CSourceStream::m_pFilter->StreamTime(now);
}
// avoid a tidge of creep since we sleep until [typically] just past the previous end.
endFrame = previousFrameEndTime + m_rtFrameLength;
previousFrameEndTime = endFrame;
} else {
// if there's no reference clock, it will "always" think it missed a frame
if(show_performance) {
if(now == 0)
LocalOutput("probable none reference clock, streaming fastly");
else
LocalOutput("it missed a frame--can't keep up %d %llu %llu", countMissed++, now, previousFrameEndTime); // we don't miss time typically I don't think, unless de-dupe is turned on, or aero, or slow computer, buffering problems downstream, etc.
}
// have to add a bit here, or it will always be "it missed a frame" for the next round...forever!
endFrame = now + m_rtFrameLength;
// most of this stuff I just made up because it "sounded right"
//LocalOutput("checking to see if I can catch up again now: %llu previous end: %llu subtr: %llu %i", now, previousFrameEndTime, previousFrameEndTime - m_rtFrameLength, previousFrameEndTime - m_rtFrameLength);
if(now > (previousFrameEndTime - (long long) m_rtFrameLength)) { // do I even need a long long cast?
// let it pretend and try to catch up, it's not quite a frame behind
previousFrameEndTime = previousFrameEndTime + m_rtFrameLength;
} else {
endFrame = now + m_rtFrameLength/2; // ?? seems to not hurt, at least...I guess
previousFrameEndTime = endFrame;
}
}
// accomodate for 0 to avoid startup negatives, which would kill our math on the next loop...
previousFrameEndTime = max(0, previousFrameEndTime);
pSample->SetTime((REFERENCE_TIME *) &now, (REFERENCE_TIME *) &endFrame);
//pSample->SetMediaTime((REFERENCE_TIME *)&now, (REFERENCE_TIME *) &endFrame);
LocalOutput("timestamping video packet as %lld -> %lld", now, endFrame);
m_iFrameNumber++;
// Set TRUE on every sample for uncompressed frames http://msdn.microsoft.com/en-us/library/windows/desktop/dd407021%28v=vs.85%29.aspx
pSample->SetSyncPoint(TRUE);
// only set discontinuous for the first...I think...
pSample->SetDiscontinuity(m_iFrameNumber <= 1);
#ifdef _DEBUG
// the swprintf costs like 0.04ms (25000 fps LOL)
double m_fFpsSinceBeginningOfTime = ((double) m_iFrameNumber)/(GetTickCount() - globalStart)*1000;
swprintf(out, L"done video frame! total frames: %d this one %dx%d -> (%dx%d) took: %.02Lfms, %.02f ave fps (%.02f is the theoretical max fps based on this round, ave. possible fps %.02f, fastest round fps %.02f, negotiated fps %.06f), frame missed %d",
m_iFrameNumber, m_iCaptureConfigHeight, m_iCaptureConfigWidth, getNegotiatedFinalWidth(), getNegotiatedFinalHeight(), millisThisRoundTook, m_fFpsSinceBeginningOfTime, 1.0*1000/millisThisRoundTook,
/* average */ 1.0*1000*m_iFrameNumber/sumMillisTook, 1.0*1000/fastestRoundMillis, GetFps(), countMissed);
LocalOutput(out);
set_config_string_setting(L"frame_stats", out);
#endif
return S_OK;
}
HRESULT CPushPinDesktop::FillBuffer(IMediaSample *pSample)
{
__int64 startThisRound = StartCounter();
BYTE *pData;
CheckPointer(pSample, E_POINTER);
if(m_bReReadRegistry) {
reReadCurrentPosition(1);
}
// Access the sample's data buffer
pSample->GetPointer(&pData);
// Make sure that we're still using video format
ASSERT(m_mt.formattype == FORMAT_VideoInfo);
VIDEOINFOHEADER *pVih = (VIDEOINFOHEADER*) m_mt.pbFormat;
// for some reason the timings are messed up initially, as there's no start time at all for the first frame (?) we don't start in State_Running ?
// race condition?
// so don't do some calculations unless we're in State_Running
FILTER_STATE myState;
CSourceStream::m_pFilter->GetState(INFINITE, &myState);
bool fullyStarted = myState == State_Running;
boolean gotNew = false;
while(!gotNew) {
CopyScreenToDataBlock(hScrDc, pData, (BITMAPINFO *) &(pVih->bmiHeader), pSample);
if(m_bDeDupe) {
if(memcmp(pData, pOldData, pSample->GetSize())==0) { // took desktop: 10ms for 640x1152, still 100 fps uh guess...
Sleep(m_millisToSleepBeforePollForChanges);
} else {
gotNew = true;
memcpy( /* dest */ pOldData, pData, pSample->GetSize()); // took 4ms for 640x1152, but it's worth it LOL.
// LODO memcmp and memcpy in the same loop LOL.
}
} else {
// it's always new for everyone else!
gotNew = true;
}
}
// capture how long it took before we add in our own arbitrary delay to enforce fps...
long double millisThisRoundTook = GetCounterSinceStartMillis(startThisRound);
fastestRoundMillis = min(millisThisRoundTook, fastestRoundMillis); // keep stats :)
sumMillisTook += millisThisRoundTook;
CRefTime now;
CRefTime endFrame;
CSourceStream::m_pFilter->StreamTime(now);
// wait until we "should" send this frame out...
if((now > 0) && (now < previousFrameEndTime)) { // now > 0 to accomodate for if there is no reference graph clock at all...also boot strap time ignore it :P
while(now < previousFrameEndTime) { // guarantees monotonicity too :P
Sleep(1);
CSourceStream::m_pFilter->StreamTime(now);
}
// avoid a tidge of creep since we sleep until [typically] just past the previous end.
endFrame = previousFrameEndTime + m_rtFrameLength;
previousFrameEndTime = endFrame;
} else {
if(show_performance)
LocalOutput("it missed a frame--can't keep up %d", countMissed++); // we don't miss time typically I don't think, unless de-dupe is turned on, or aero, or slow computer, buffering problems downstream, etc.
// have to add a bit here, or it will always be "it missed some time" for the next round...forever!
endFrame = now + m_rtFrameLength;
// most of this stuff I just made up because it "sounded right"
//LocalOutput("checking to see if I can catch up again now: %llu previous end: %llu subtr: %llu %i", now, previousFrameEndTime, previousFrameEndTime - m_rtFrameLength, previousFrameEndTime - m_rtFrameLength);
if(now > (previousFrameEndTime - (long long) m_rtFrameLength)) { // do I need a long long cast?
// let it pretend and try to catch up, it's not quite a frame behind
previousFrameEndTime = previousFrameEndTime + m_rtFrameLength;
} else {
endFrame = now + m_rtFrameLength/2; // ?? seems to work...I guess...
previousFrameEndTime = endFrame;
}
}
previousFrameEndTime = max(0, previousFrameEndTime);// avoid startup negatives, which would kill our math on the next loop...
// LocalOutput("marking frame with timestamps: %llu %llu", now, endFrame);
pSample->SetTime((REFERENCE_TIME *) &now, (REFERENCE_TIME *) &endFrame);
//pSample->SetMediaTime((REFERENCE_TIME *)&now, (REFERENCE_TIME *) &endFrame); //useless seemingly
if(fullyStarted) {
m_iFrameNumber++;
}
// Set TRUE on every sample for uncompressed frames http://msdn.microsoft.com/en-us/library/windows/desktop/dd407021%28v=vs.85%29.aspx
pSample->SetSyncPoint(TRUE);
// only set discontinuous for the first...I think...
pSample->SetDiscontinuity(m_iFrameNumber <= 1);
// the swprintf costs like 0.04ms (25000 fps LOL)
m_fFpsSinceBeginningOfTime = ((double) m_iFrameNumber)/(GetTickCount() - globalStart)*1000;
swprintf(out, L"done frame! total frames: %d this one %dx%d -> (%dx%d) took: %.02Lfms, %.02f ave fps (%.02f is the theoretical max fps based on this round, ave. possible fps %.02f, fastest round fps %.02f, negotiated fps %.06f), frame missed %d",
m_iFrameNumber, m_iCaptureConfigHeight, m_iCaptureConfigWidth, getNegotiatedFinalWidth(), getNegotiatedFinalHeight(), millisThisRoundTook, m_fFpsSinceBeginningOfTime, 1.0*1000/millisThisRoundTook,
/* average */ 1.0*1000*m_iFrameNumber/sumMillisTook, 1.0*1000/fastestRoundMillis, GetFps(), countMissed);
//#ifdef _DEBUG // probably not worth it but we do hit this a lot...hmm...
LocalOutput(out);
set_config_string_setting(L"frame_stats", out);
//#endif
return S_OK;
}