DXGI doesn't take into account the SDRWhiteLevel that has already been
applied to the monitor when it converts to HDR10 which results in clipping.
This change set implements a HLSL shader to reverse this while at the same
time converting to HDR10.
This is still not perfect but far better then doing nothing.
There is no performance penalty to using ARGB10 as it's still a 32-bit copy per
pixel and the nvidia driver internally handles any conversions required to make
this work for both SDR and HDR.
@quantum has observed nvfbc under rare circumstances fail to initialize,
this adds a retry to the init with a short delay to hopefully recover
from this situation.
This is an experimental & incomplete feature for those using
supersampling. Anything > 1200p will be downsampled by 50% before
copying out of the GPU to save on memory bandwidth.
Unfinished! Has issues with damage tracking and currently can not
be configured. Only dx11 has been tested at this point, everything
else will likely have problems/crash.
The nsleep() call lets d3d12 sleep for a more precise amount of
time while maintaining the current millisecond-scale sleep
interface in the configuration file.
To avoid client showing "Using : NVFBC (NVidia Frame Buffer Capt".
This happens because the string is truncated to 31 characters to fit
in the char capture[32]; member of KVMFRRecord_VMInfo.
For Windows 10, it so happens that the major.minor is 10.0. This is not
usually a given, e.g. on Windows 7 where it would read 6.1, on
Windows 8 it would read 6.2, and on Windows 8.1 it would read 6.3.
This is obviously undesirable, so we should just read the ProductName
from registry if possible. This results in something like:
OS Name: Windows 10 Pro for Workstations (Build: 19043)
This will cache up to 10 handles, in practice I have never seen DXGI
return anything but the same resource each time but we allow for more
anyway should MS change something in the future.
Should the cache get over filled it is disabled entirely and we revert
to the original behaviour.
The buffer input sizes to the `IDXGIOutputDuplication` methods are measured
in bytes. This dramatically increases the number of dirty/move rects that
can be handled.
This new feature while helps on some systems, others using freesync or
higher refresh rates where the capture can't keep up will limit to
fractions of the refresh rate. Better to disable and allow users to
opt-in.
This change reduces the host GPU and CPU load by a large margin
improving guest system performance along with removing latency spikes
when moving the mouse. This is default enabled but can be disabled with
the new option `dxgi:dwmFlush=no` as it limits the capture rate to the
refresh rate of the guests output which may not be desireable.
This change allows the host to provide information to the client about
how the VM is configured, information such as the UUID, CPU
configuration and capture method both for informational display in the
client as well as debugging in the client's logs.
The format of the records allows this to be extended later with new
record types without needing to bump the KVMFR version.
If the wait times out, we used to simply restart the loop, which causes
it to check this->stop and exit if set to true. However, nvfbc_stop
already calls lgSignalEvent, which would wake up the pointer thread to
perform the check, so there is no need to set a timeout on the wait.
The host is a 64-bit application, so requiring 64-bit Windows isn't an
issue. However, not using 32-bit installers has an advantage: we don't
need to require WoW64.
This commit adds damage tracking to the DXGI textures, and only copies the
damaged areas to the textures with ID3D11DeviceContext::CopySubresourceRegion.
The sleep logic in waitFrame makes it difficult for this to reduce the
latency, but removing it shows significant improvements (6-7 ms to ~3 ms)
when a tiny portion of the screen is damaged, while showing no difference on
full screen damage.
This implementation uses a line sweep algorithm to copy the precisely the
intersection of all accumulated damage rectangles, ensuring that every
pixel is copied exactly once, and no pixel is ever copied multiple times.
Furthermore, once a row has been swept, we update the framebuffer write
pointer immediately.
Certain drivers do not support pitches that are not multiples of 128 bytes,
and instead just does some kind of rounding internally. On DXGI, this is not
a problem because the API rounds pixel pitch, but NvFBC does not. This causes
certain resolutions to simply not work with dmabuf, most notably 3440x1440,
which is 1440p ultrawide.
Since we are copying pixels with the CPU anyways, we might as well round the
pitch up to 128 bytes (32 pixels).
This commit adds a new host configuration option, nvfbc:diffRes, which
specifies the dimensions of every block in the diff map. This defaults to
128, meaning the default 128x128 block size.
Since block sizes other than 128x128 is not guaranteed to be supported by
NvFBC, the function NvFBCGetDiffMapBlockSize was introduced to query the
support and output the actual block size used.
When our window is destroyed, our timers are also destroyed. This causes our
attempt at destruction to fail. Instead, set MessageHWND to NULL in the
WM_DESTROY handler and don't try destroying the timers if the window is gone.
DestroyWindow can only be invoked on the thread that created the window.
All other threads must use WM_CLOSE or another message to signal tell the
window to destroy itself.
MinGW seems to decide at random whether it wants to use memcpy from
mscvrt.dll or ntdll.dll. Currently, on Debian buster, ntdll.dll is chosen,
while on sid, mscvrt.dll is chosen.
This commit declares a new .def file for ntdll containing only the
functions we want to link from ntdll.dll, and generates ntdll.a from it
with dlltool. This way, MinGW will never be tempted to link functions
like memcpy from ntdll.dll.
This function is sometimes flaky and may fail for no apparent reason,
see https://stackoverflow.com/q/3945003. This has also been experienced
during the development of #610.
This commit adds logging so we may see if it ever fails for no reason
and work out some way to fix it.
We were using an auto-reset event to signal the mousehook exit. This was
fine when there was only one thread, but with the addition of the update
thread, only one thread is signaled, causing the wait to last forever.
The fix is switching to a manual reset event and call ResetEvent after
the threads have exited.
The type of the QuadPart member of the LARGE_INTEGER union is actually
LONGLONG, so we should cast to LONGLONG instead of int.
This avoids truncation should (ms * 10000.0f) exceed 2^31-1.
This function is available since Windows Vista and can therefore be used
directly without going through GetProcAddress. Unfortunately, MinGW does
not have d3dkmthk.h, but we can declare the prototype ourselves and link
against gdi32.dll.
There is no need to LoadLibrary and GetProcAddress to get pointers to
NtDelayExecution or NtSetTimerResolution. These functions don't have
prototypes in any SDK header, but they are exported in ntdll.dll and
we can simply declare the prototype and link ntdll.
There is also no chance that the functions do not exist: I checked an
old install of Windows NT 4.0 and both of these functions exist.
Also used NtSetTimerResolution instead of ZeSetTimerResolution for
consistency (they are the same).
Also changed system timer resolution log message units to μs with
one decimal digit for readability. This is the actual amount of
precision available to us.
According to MSDN documentation for CreateEnvironmentBlock, "[i]f the
environment block is passed to CreateProcessAsUser, you must also
specify the CREATE_UNICODE_ENVIRONMENT flag."
Also pass DETACHED_PROCESS because the host is a GUI application and
doesn't use the console.
Since with the service, we are already running as SYSTEM, we don't need
to use dupeSystemProcessToken to get the token for SYSTEM. This removes
the need for having SeDebugPrivilege, SeTcbPrivilege, and
SeAssignPrimaryTokenPrivilege, or otherwise doing sketchy things.
Furthermore, we now only open the token with the privileges we actually
need.
This allows the process to be terminated without resorting to
TerminateProcess. With some fixes, this allows the notification icon to be
removed when the service is restarted.
Furthermore, instead of sending WM_DESTROY to fool the window into believing
it's being destroyed, we actually call DestroyWindow now.
For adjacent changed regions, we actually use the bounding box for the
entire polygon. This may result in more area being damaged than strictly
necessary, but is nevertheless desirable since it reduces the number of
rectangles.
The windows hook WH_MOUSE_LL is called in such a way that any delay in
processing causes a system wide stall. This change spawns an extra
thread which waits on an event set by the hook which is then used to
call the callback with an artifical limit of 1000Hz.
Before we try and perhaps fail to init DXGI, we should print out what
the device is so that when there is an error report we can immediately
see if the user has the QXL device attached still.
While it's correct for DXGI to use a asyncronous waitFrame model, other
capture interfaces such as NvFBC it is not correct. This change allows
the capture interface to specify which is more correct for it and moves
the waitFrame/post into the main thread if async is not desired.
