H264-Encoder | Implementation of a subset of CBP of h.264 encoder | Machine Learning library

Background: I'm encoding NV12 buffers into h264 video stream wrapped in MPEG4 container using SinkWriter. Everything works fine but one problem though, as the SinkWriter abstracts the low-level encoder configurations I'm unable to control the properties like GOP size, B-picture count, CODECAPI_AVEncCommonRateControlMode, etc.

The problem is due to the fact that the SinkWriter instantiates the encoder transform only after the SetInputMediaType call, and we will be able to get the CodecAPI instance only after that point. So, we have no way to control the encoder and configure necessary props before all this happens, it also never honours further changes to the encoder through CodecAPI instance.

Experiments: I tried the PropertyStore(MF_SINK_WRITER_ENCODER_CONFIG) approach but nothing seemed to be changing (It could be a platform/encoder specific behaviour), I could also see a lot of people complaining about unpredictable behaviours with these APIs. Then, I came across this MSDN thread (Almost 7 years old post) in which the user described how he dealt with this problem by locally registering a custom class factory on Windows7 machine.

Problem: Having the MSDN thread as a reference, I tried implementing IClassFactory and registering it through MFTRegisterLocal but the CreateInstance function is never getting called for me (Windows 10 machine). I'm only getting the QueryInterface method called for IID_IClassFactory and IID_IMFAttributes interfaces instead. And, the SinkWriter seems to be proceeding with fetching the MFT on its own.

I understand I might be doing something wrong, and I'm not an expert in COM. Is there any other way I could achieve this?

Custom class factory implementation:

I'm capturing the desktop using DesktopDuplication API and converting the samples from RGBA to NV12 in GPU and feeding the same to MediaFoundation hardware H264 MFT. This works fine with Nvidia graphics, and also with software encoders but fails when only intel graphics hardware MFT is available. The code works fine on the same intel graphics machine if I fallback to Software MFT. I also have ensured that the encoding is actually done in hardware on Nvidia graphics machines.

On Intel graphics, MFT returns MEError ("Unspecified error"), which happens just only after the first sample is fed, and subsequent calls to ProcessInput (when event generator triggers METransformNeedInput) returns "The callee is currently not accepting further input". It's rare that MFT consumes a few more samples before returning these errors. This behavior is confusing, I'm feeding a sample only when the event generator triggers METransformNeedInput asynchronously through IMFAsyncCallback, and also checking properly whether METransformHaveOutput is triggered as soon as a sample is fed. This really baffles me when the same asynchronous logic works fine with Nvidia hardware MFT & Microsoft software encoders.

There is also a similar unresolved question in the intel forum itself. My code is similar to the one mentioned in the intel thread, except for the fact that I'm also setting d3d device manager to the encoder like below.

And, there are three other stack overflow threads reporting a similar issue with no solution given (MFTransform encoder->ProcessInput returns E_FAIL & How to create IMFSample from D11 texture for Intel MFT encoder & Asynchronous MFT is not sending MFTransformHaveOutput Event(Intel Hardware MJPEG Decoder MFT)). I have tried every possible option with no improvement on this.

Color converter code is taken from intel media sdk samples. I have also uploaded my complete code here.

Method to set d3d manager: