void CPushPinDesktop::doJustBitBltOrScaling(HDC hMemDC, int nWidth, int nHeight, int iFinalWidth, int iFinalHeight, HDC hScrDC, int nX, int nY) {
__int64 start = StartCounter();
boolean notNeedStretching = (iFinalWidth == nWidth) && (iFinalHeight == nHeight);
if(m_iHwndToTrack != NULL)
ASSERT_RAISE(notNeedStretching); // we don't support HWND plus scaling...hmm... LODO move assertion LODO implement this (low prio since they probably are just needing that window, not with scaling too [?])
int captureType = SRCCOPY;
if(m_bUseCaptureBlt)
captureType = captureType | CAPTUREBLT; // CAPTUREBLT here [last param] is for layered (transparent) windows in non-aero I guess (which happens to include the mouse, but we do that elsewhere)
if (notNeedStretching) {
if(m_iHwndToTrack != NULL) {
// make sure we only capture 'not too much' i.e. not past the border of this HWND, for the case of Aero being turned off, it shows other windows that we don't want
// a bit confusing...
RECT p;
if (m_bHwndTrackDecoration)
GetWindowRectIncludingAero(m_iHwndToTrack, &p); // 0.05 ms
else
GetClientRect(m_iHwndToTrack, &p); // 0.005 ms
//GetWindowRect(m_iHwndToTrack, &p); // 0.005 ms
nWidth = min(p.right-p.left, nWidth);
nHeight = min(p.bottom-p.top, nHeight);
}
// Bit block transfer from screen our compatible memory DC. Apparently this is faster than stretching.
BitBlt(hMemDC, 0, 0, nWidth, nHeight, hScrDC, nX, nY, captureType);
// 9.3 ms 1920x1080 -> 1920x1080 (100 fps) (11 ms? 14? random?)
}
else {
if (m_iStretchMode == 0)
{
// low quality scaling -- looks terrible
SetStretchBltMode (hMemDC, COLORONCOLOR); // the SetStretchBltMode call itself takes 0.003ms
// COLORONCOLOR took 92ms for 1920x1080 -> 1000x1000, 69ms/80ms for 1920x1080 -> 500x500 aero
// 20 ms 1920x1080 -> 500x500 without aero
// LODO can we get better results with good speed? it is sooo ugly!
}
else
{
SetStretchBltMode (hMemDC, HALFTONE);
// high quality stretching
// HALFTONE took 160ms for 1920x1080 -> 1000x1000, 107ms/120ms for 1920x1080 -> 1000x1000
// 50 ms 1920x1080 -> 500x500 without aero
SetBrushOrgEx(hMemDC, 0, 0, 0); // MSDN says I should call this after using HALFTONE
}
StretchBlt(hMemDC, 0, 0, iFinalWidth, iFinalHeight, hScrDC, nX, nY, nWidth, nHeight, captureType);
}
if(show_performance)
LocalOutput("%s took %.02f ms", notNeedStretching ? "bitblt" : "stretchblt", GetCounterSinceStartMillis(start));
}
void CPushPinDesktop::reReadCurrentStartXY(int isReRead) {
__int64 start = StartCounter();
// assume 0 means not set...negative ignore :)
// TODO no overflows, that's a bad value too... they cause a crash, I think! [position youtube too far bottom right, track it...]
int old_x = m_rScreen.left;
int old_y = m_rScreen.top;
int config_start_x = read_config_setting(TEXT("start_x"), m_rScreen.left, true);
m_rScreen.left = config_start_x;
// is there a better way to do this registry stuff?
int config_start_y = read_config_setting(TEXT("start_y"), m_rScreen.top, true);
m_rScreen.top = config_start_y;
if(old_x != m_rScreen.left || old_y != m_rScreen.top) {
if(isReRead) {
m_rScreen.right = m_rScreen.left + m_iCaptureConfigWidth;
m_rScreen.bottom = m_rScreen.top + m_iCaptureConfigHeight;
}
}
if(show_performance) {
wchar_t out[1000];
swprintf(out, 1000, L"new screen pos from reg: %d %d\n", config_start_x, config_start_y);
LocalOutput("[re]readCurrentPosition (including swprintf call) took %.02fms", GetCounterSinceStartMillis(start)); // takes 0.42ms (2000 fps)
LocalOutput(out);
}
}
void CPushPinDesktop::doDIBits(HDC hScrDC, HBITMAP hRawBitmap, int nHeightScanLines, BYTE *pData, BITMAPINFO *pHeader) {
__int64 start = StartCounter();
// Copy the bitmap data into the provided BYTE buffer, in the right format I guess.
GetDIBits(hScrDC, hRawBitmap, 0, nHeightScanLines, pData, pHeader, DIB_RGB_COLORS); // just copies raw bits to pData, I guess, from an HBITMAP handle. "like" GetObject, but also does conversions [?]
if(show_performance)
LocalOutput("doDiBits took %.02fms", GetCounterSinceStartMillis(start)); // took 1.1/3.8ms total, so this brings us down to 80fps compared to max 251...but for larger things might make more difference...
}
// This is where we insert the DIB bits into the video stream.
// FillBuffer is called once for every sample in the stream.
HRESULT CPushPinDesktop::FillBuffer(IMediaSample *pSample)
{
__int64 startThisRound = StartCounter();
BYTE *pData;
long cbData;
CheckPointer(pSample, E_POINTER);
// Access the sample's data buffer
pSample->GetPointer(&pData);
cbData = pSample->GetSize();
// Make sure that we're still using video format
ASSERT(m_mt.formattype == FORMAT_VideoInfo);
VIDEOINFOHEADER *pVih = (VIDEOINFOHEADER*)m_mt.pbFormat;
// Copy the DIB bits over into our filter's output buffer.
// Since sample size may be larger than the image size, bound the copy size.
int nSize = min(pVih->bmiHeader.biSizeImage, (DWORD) cbData); // cbData is the size of pData
HDIB hDib = CopyScreenToBitmap(hScrDc, &m_rScreen, pData, (BITMAPINFO *) &(pVih->bmiHeader));
if (hDib)
DeleteObject(hDib);
// for some reason the timings are messed up initially, as there's no start time.
// race condition?
// so don't count them unless they seem valid...
FILTER_STATE myState;
CSourceStream::m_pFilter->GetState(INFINITE, &myState);
bool fullyStarted = myState == State_Running;
CRefTime now;
CSourceStream::m_pFilter->StreamTime(now);
// calculate how long it took before we add in our own arbitrary delay to enforce fps...
long double millisThisRoundTook = GetCounterSinceStartMillis(startThisRound);
// wait until we "should" send this frame out...TODO...more precise et al...
if(m_iFrameNumber > 0 && (now > 0)) { // now > 0 to accomodate for if there is no clock at all...
while(now < previousFrameEndTime) { // guarantees monotonicity too :P
Sleep(1);
CSourceStream::m_pFilter->StreamTime(now);
}
}
REFERENCE_TIME endFrame = now + m_rtFrameLength;
pSample->SetTime((REFERENCE_TIME *) &now, &endFrame);
if(fullyStarted)
m_iFrameNumber++;
// Set TRUE on every sample for uncompressed frames
pSample->SetSyncPoint(TRUE);
// only set discontinuous for the first...I think...
pSample->SetDiscontinuity(m_iFrameNumber == 1);
double fpsSinceBeginningOfTime = ((double) m_iFrameNumber)/(GetTickCount() - globalStart)*1000;
LocalOutput("end total frames %d %.020Lfms, total since beginning of time %f fps (theoretical max fps %f)", m_iFrameNumber, millisThisRoundTook,
fpsSinceBeginningOfTime, 1.0*1000/millisThisRoundTook);
previousFrameEndTime = endFrame;
return S_OK;
}
// the default child constructor...
CPushPinDesktop::CPushPinDesktop(HRESULT *phr, CPushSourceDesktop *pFilter)
: CSourceStream(NAME("Push Source CPushPinDesktop child/pin"), phr, pFilter, L"Capture"),
m_bReReadRegistry(0),
m_bDeDupe(0),
m_iFrameNumber(0),
pOldData(NULL),
m_bConvertToI420(false),
m_pParent(pFilter),
m_bFormatAlreadySet(false),
hRawBitmap(NULL),
m_bUseCaptureBlt(false),
previousFrameEndTime(0)
{
// Get the device context of the main display, just to get some metrics for it...
globalStart = GetTickCount();
m_iHwndToTrack = (HWND) read_config_setting(TEXT("hwnd_to_track"), NULL, false);
if(m_iHwndToTrack) {
LocalOutput("using specified hwnd no decoration: %d", m_iHwndToTrack);
hScrDc = GetDC(m_iHwndToTrack); // using GetDC here seemingly allows you to capture "just a window" without decoration
m_bHwndTrackDecoration = false;
} else {
m_iHwndToTrack = (HWND) read_config_setting(TEXT("hwnd_to_track_with_window_decoration"), NULL, false);
if(m_iHwndToTrack) {
LocalOutput("using specified hwnd with decoration: %d", m_iHwndToTrack);
hScrDc = GetWindowDC(m_iHwndToTrack);
m_bHwndTrackDecoration = true;
} else {
int useForeGroundWindow = read_config_setting(TEXT("capture_foreground_window_if_1"), 0, true);
if(useForeGroundWindow) {
LocalOutput("using foreground window %d", GetForegroundWindow());
hScrDc = GetDC(GetForegroundWindow());
} else {
// the default, just capture desktop
// hScrDc = CreateDC(TEXT("DISPLAY"), NULL, NULL, NULL); // possibly better than GetDC(0), supposed to be multi monitor?
// LocalOutput("using the dangerous CreateDC DISPLAY\n");
// danger, CreateDC DC is only good as long as this particular thread is still alive...hmm...is it better for directdraw
hScrDc = GetDC(NULL);
}
}
}
//m_iScreenBitDepth = GetTrueScreenDepth(hScrDc);
ASSERT_RAISE(hScrDc != 0); // 0 implies failure... [if using hwnd, can mean the window is gone!]
// Get the dimensions of the main desktop window as the default
m_rScreen.left = m_rScreen.top = 0;
m_rScreen.right = GetDeviceCaps(hScrDc, HORZRES); // NB this *fails* for dual monitor support currently... but we just get the wrong width by default, at least with aero windows 7 both can capture both monitors
m_rScreen.bottom = GetDeviceCaps(hScrDc, VERTRES);
// now read some custom settings...
WarmupCounter();
if(!m_iHwndToTrack) {
reReadCurrentStartXY(0);
} else {
LocalOutput("ignoring startx, starty since hwnd was specified");
}
int config_width = read_config_setting(TEXT("capture_width"), 0, false);
ASSERT_RAISE(config_width >= 0); // negatives not allowed...
int config_height = read_config_setting(TEXT("capture_height"), 0, false);
ASSERT_RAISE(config_height >= 0); // negatives not allowed, if it's set :)
if(config_width > 0) {
int desired = m_rScreen.left + config_width;
//int max_possible = m_rScreen.right; // disabled check until I get dual monitor working. or should I allow off screen captures anyway?
//if(desired < max_possible)
m_rScreen.right = desired;
//else
// m_rScreen.right = max_possible;
} else {
// leave full screen
}
m_iCaptureConfigWidth = m_rScreen.right - m_rScreen.left;
ASSERT_RAISE(m_iCaptureConfigWidth > 0);
if(config_height > 0) {
int desired = m_rScreen.top + config_height;
//int max_possible = m_rScreen.bottom; // disabled, see above.
//if(desired < max_possible)
m_rScreen.bottom = desired;
//else
// m_rScreen.bottom = max_possible;
} else {
// leave full screen
}
m_iCaptureConfigHeight = m_rScreen.bottom - m_rScreen.top;
ASSERT_RAISE(m_iCaptureConfigHeight > 0);
m_iStretchToThisConfigWidth = read_config_setting(TEXT("stretch_to_width"), 0, false);
m_iStretchToThisConfigHeight = read_config_setting(TEXT("stretch_to_height"), 0, false);
m_iStretchMode = read_config_setting(TEXT("stretch_mode_high_quality_if_1"), 0, true); // guess it's either stretch mode 0 or 1
ASSERT_RAISE(m_iStretchToThisConfigWidth >= 0 && m_iStretchToThisConfigHeight >= 0 && m_iStretchMode >= 0); // sanity checks
m_bUseCaptureBlt = read_config_setting(TEXT("capture_transparent_windows_including_mouse_in_non_aero_if_1_causes_annoying_mouse_flicker"), 0, true) == 1;
m_bCaptureMouse = read_config_setting(TEXT("capture_mouse_default_1"), 1, true) == 1;
// default 30 fps...hmm...
int config_max_fps = read_config_setting(TEXT("default_max_fps"), 30, false); // TODO allow floats [?] when ever requested
ASSERT_RAISE(config_max_fps > 0);
// m_rtFrameLength is also re-negotiated later...
m_rtFrameLength = UNITS / config_max_fps;
if(is_config_set_to_1(TEXT("track_new_x_y_coords_each_frame_if_1"))) {
m_bReReadRegistry = 1; // takes 0.416880ms, but I thought it took more when I made it off by default :P
}
if(is_config_set_to_1(TEXT("dedup_if_1"))) {
m_bDeDupe = 1; // takes 10 or 20ms...but useful to me! :)
}
m_millisToSleepBeforePollForChanges = read_config_setting(TEXT("millis_to_sleep_between_poll_for_dedupe_changes"), 10, true);
wchar_t out[10000];
swprintf(out, 10000, L"default/from reg read config as: %dx%d -> %dx%d (%d top %d bottom %d l %d r) %dfps, dedupe? %d, millis between dedupe polling %d, m_bReReadRegistry? %d hwnd:%d \n",
m_iCaptureConfigHeight, m_iCaptureConfigWidth, getCaptureDesiredFinalHeight(), getCaptureDesiredFinalWidth(), m_rScreen.top, m_rScreen.bottom, m_rScreen.left, m_rScreen.right, config_max_fps, m_bDeDupe, m_millisToSleepBeforePollForChanges, m_bReReadRegistry, m_iHwndToTrack);
// warmup the debugging message system
__int64 measureDebugOutputSpeed = StartCounter();
LocalOutput(out);
LocalOutput("writing a large-ish debug itself took: %.02Lf ms", GetCounterSinceStartMillis(measureDebugOutputSpeed));
set_config_string_setting(L"last_init_config_was", out);
}
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;
}
// the default child constructor...
CPushPinDesktop::CPushPinDesktop(HRESULT *phr, CPushSourceDesktop *pFilter)
: CSourceStream(NAME("Push Source CPushPinDesktop child/pin"), phr, pFilter, L"Capture"),
m_FramesWritten(0),
m_bReReadRegistry(0),
m_bDeDupe(0),
m_iFrameNumber(0),
pOldData(NULL),
m_bConvertToI420(false),
//m_nCurrentBitDepth(32), // negotiated later...
m_pParent(pFilter),
m_bFormatAlreadySet(false),
hRawBitmap(NULL)
{
// Get the device context of the main display, just to get some metrics for it...
globalStart = GetTickCount();
m_iHwndToTrack = (HWND) read_config_setting(TEXT("hwnd_to_track"), NULL);
hScrDc = GetDC(m_iHwndToTrack);
m_iScreenBitDepth = GetTrueScreenDepth(hScrDc);
ASSERT(hScrDc != 0);
GdiSetBatchLimit(1); // disable any GDI...just in case this helps anybody...
// Get the dimensions of the main desktop window as the default
m_rScreen.left = m_rScreen.top = 0;
m_rScreen.right = GetDeviceCaps(hScrDc, HORZRES); // NB this *fails* for dual monitor support currently... but we just get the wrong width by default, at least with aero windows 7 both can capture both monitors
m_rScreen.bottom = GetDeviceCaps(hScrDc, VERTRES);
// now read some custom settings...
WarmupCounter();
reReadCurrentPosition(0);
int config_width = read_config_setting(TEXT("capture_width"), 0);
ASSERT(config_width >= 0); // negatives not allowed...
int config_height = read_config_setting(TEXT("capture_height"), 0);
ASSERT(config_height >= 0); // negatives not allowed, if it's set :)
if(config_width > 0) {
int desired = m_rScreen.left + config_width;
//int max_possible = m_rScreen.right; // disabled check until I get dual monitor working. or should I allow off screen captures anyway?
//if(desired < max_possible)
m_rScreen.right = desired;
//else
// m_rScreen.right = max_possible;
} else {
// leave full screen
}
m_iCaptureConfigWidth = m_rScreen.right - m_rScreen.left;
ASSERT(m_iCaptureConfigWidth > 0);
if(config_height > 0) {
int desired = m_rScreen.top + config_height;
//int max_possible = m_rScreen.bottom; // disabled, see above.
//if(desired < max_possible)
m_rScreen.bottom = desired;
//else
// m_rScreen.bottom = max_possible;
} else {
// leave full screen
}
m_iCaptureConfigHeight = m_rScreen.bottom - m_rScreen.top;
ASSERT(m_iCaptureConfigHeight > 0);
m_iStretchToThisConfigWidth = read_config_setting(TEXT("stretch_to_width"), 0);
m_iStretchToThisConfigHeight = read_config_setting(TEXT("stretch_to_height"), 0);
m_iStretchMode = read_config_setting(TEXT("stretch_mode_high_quality_if_1"), 0);
ASSERT(m_iStretchToThisConfigWidth >= 0 && m_iStretchToThisConfigHeight >= 0 && m_iStretchMode >= 0); // sanity checks
// default 30 fps...hmm...
int config_max_fps = read_config_setting(TEXT("default_max_fps"), 30); // TODO allow floats [?] when ever requested
ASSERT(config_max_fps >= 0);
// m_rtFrameLength is also re-negotiated later...
m_rtFrameLength = UNITS / config_max_fps;
if(is_config_set_to_1(TEXT("track_new_x_y_coords_each_frame_if_1"))) {
m_bReReadRegistry = 1; // takes 0.416880ms, but I thought it took more when I made it off by default :P
}
if(is_config_set_to_1(TEXT("dedup_if_1"))) {
m_bDeDupe = 1; // takes 10 or 20ms...but useful to me! :)
}
m_millisToSleepBeforePollForChanges = read_config_setting(TEXT("millis_to_sleep_between_poll_for_dedupe_changes"), 10);
wchar_t out[1000];
swprintf(out, 1000, L"default/from reg read config as: %dx%d -> %dx%d (%dtop %db %dl %dr) %dfps, dedupe? %d, millis between dedupe polling %d, m_bReReadRegistry? %d \n",
m_iCaptureConfigHeight, m_iCaptureConfigWidth, getCaptureDesiredFinalHeight(), getCaptureDesiredFinalWidth(), m_rScreen.top, m_rScreen.bottom, m_rScreen.left, m_rScreen.right, config_max_fps, m_bDeDupe, m_millisToSleepBeforePollForChanges, m_bReReadRegistry);
LocalOutput(out); // warmup for the below debug :)
__int64 measureDebugOutputSpeed = StartCounter();
LocalOutput(out);
LocalOutput("writing a large-ish debug itself took: %.0Lf ms", GetCounterSinceStartMillis(measureDebugOutputSpeed));
// does this work with flash?
set_config_string_setting(L"last_init_config_was", out);
}
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;
}
