According to opencl 1.2 spec 6.1.5:
For arguments to a __kernel function declared to be a pointer to a
data type, the OpenCL compiler can assume that the pointee is always
appropriately aligned as required by the data type. The behavior of
an unaligned load or store is undefined, except for the
vloadn, vload_halfn, vstoren, and vstore_halfn functions defined in
section 6.12.7.
Original code read data of type T from address not aligned by multiple
of sizeof(T), so the result is incorrect. With this patch, the cases
./opencv_perf_imgproc
--gtest_filter=OCL_ImgSize_TmplSize_Method_MatType_MatchTemplate.MatchTemplate/*
could work well with beignet 0.9.3.
Signed-off-by: Chuanbo Weng <chuanbo.weng@intel.com>
The implicit narrowing in the initializer list throws a compiler error for some compilers with C++11 support turned on. The specific error message is: "error: narrowing conversion of 'PyInt_AsLong(((PyObject*)o))' from 'long int' to 'double' inside { }".
Tested on Clang 5.1.0 and Mac OS X 10.9.4.
- Accurate ASYC task handling for both GStreamer generations.
- Correct video duration request for videos.
- More accurate pipeline destruction.
- Correct color space conversion plugin for GStreamer 1.0.
Support for the custom sink is extended to non-WinRT not for compatibility as Windows Vista client is a minimum regardless, but because it offers more flexibility, could be faster and is able to be used as an optionally different code path during sink creation based on a future configuration parameter.
My discussion and proposal to finish this change:
Devices are so easily enumerated through WinRT Windows.Devices namespace that wrapping the calls in a library is quite a chore for little benefit though to get the various modes and formats could still be a worthwhile project. For now conditional compilation to remove videodevices and any offending non-video file related activity in videodevice. In my opinion, this is a different , far less fundamental and important change which can possibly be done as a future project and also much more easily implemented in C++/CX.
ImageGrabber has the IMFSampleGrabberSinkCallback replaced with a base class (SharedSampleGrabber) which also be is base class for ImageGrabberRT. This change is necessary as the custom sink does not require a thread to pump events which is done through MediaCapture already. IMFSampleGrabberSinkCallback is the common element between both models and that piece can be shared. Initializing the new ImageGrabberRT is as simple as passing an already initialized MediaCapture object and any video format/encoding parameters.
The concurrency event is necessary to wait for completion and is the way the underlying, IAsyncAction wrappers in the task library work as well. Native WIN32 event objects would be an option if HAVE_CONCURRENCY is not defined. I could even imagine doing it with sleep/thread yield and InterlockedCompareExchange yet I am not enthusiastic about that approach either. Since there is a specific compiler HAVE_ for concurrency, I do not like pulling it in though I think for WinRT it is safe to say we will always have it available though should probably conditionally compile with the Interlocked option as WIN32 events would require HAVE_WIN32.
It looks like C++/CX cannot be used for the IMediaExtension sink (which should not be a problem) as using COM objects requires WRL and though deriving from IMediaExtension can be done, there is little purpose without COM. Objects from C++/CX can be swapped to interact with objects from native C++ as Inspectable* can reinterpret_cast to the ref object IInspectable^ and vice-versa. A solution to the COM class with C++/CX would be great so we could have dual support. Also without #define for every WRL object in use, the code will get quite muddy given that the */^ would need to be ifdef'd everywhere.
Fixed bugs and completed the change. I believe the new classes need to be moved to a header file as the file has become to large and more classes need to be added for handling all the asynchronous problems (one wrapping IAsyncAction in a task and another for making a task out of IAsyncAction). Unfortunately, blocking on the UI thread is not an option in WinRT so a synchronous architecture is considered "illegal" by Microsoft's standards even if implementable (C++/CX ppltasks library throws errors if you try it). Worse, either by design or a bug in the MF MediaCapture class with Custom Sinks causes a crash if stop/start previewing without reinitializing (spPreferredPreviewMediaType is fatally nulled). After decompiling Windows.Media.dll, I worked around this in my own projects by using an activate-able custom sink ID which strangely assigns 1 to this pointer allowing it to be reinitialized in what can only be described as a hack by Microsoft. This would add additional overhead to the project to implement especially for static libraries as it requires IDL/DLL exporting followed by manifest declaration. Better to document that it is not supported.
Furthermore, an additional class for IMFAttributes should be implemented to make clean architecture for passing around attributes as opposed to directly calling non-COM interface calls on the objects and making use of SetProperties which would also be a set up for an object that uses the RuntimeClass activation ID.
The remaining changes are not difficult and will be complete soon along with debug tracing messages.
Update and rename cap_msmf.h to cap_msmf.hpp
Successful test - samples are grabbed
Library updated and cleaned up with comments, marshaling, exceptions and linker settings
Fixed trailing whitespace
VS 2013 support and cleanup consistency plus C++/CX new object fixed
Conflicts:
modules/highgui/src/cap_msmf.cpp
modules/highgui/src/cap_msmf.hpp
modules/highgui/src/ppltasks_winrt.h
Fix merge conflicts
VS 2013 Update 2 library bug fix integrated
a-wi's changed integrated
fixed dependency of stitching module on xfeatures2d module as optional
fixed the initModule_xfeatures2d function that was called in module stitching since it is in another namespace than the standard cv one.
