The clean up of input parsing and encoder initialization code had caused the
real source input dimensions to be used for array initialization, instead of
the dimensions modified to be multiples of the smallest coding unit.
We will also maintain the Makefile for basic building. This is for more
complicated stuff, like building the project on command line with
Visual Studio.
Because selecting references "far away" costs bits, this should help with the issue that multiple reference frames actually increases coded bits with no quality gain.
- Working towards issue #11.
- Removed intra_get_block_mode as unused.
- Removed unused parameters from functions. Many of them were remnants from
earlier data structures and earlier features of HEVC that have been removed.
- Lots of implicit conversions from larger types to smaller ones. I tried to
avoid turning all of them to explicit ones this time and opted for changing
the original data type instead. Had to do it in few cases though to stop the
changes from propagating too widely.
The intent was to remove the flag that enabled intra in inter frames, as that
is not really that usefull anymore, but it ended up disabling intra instead.
In NxN mode, chroma predictions were pushed to buffer when chroma should not have been used at all. (Because it is processed only on first of the four NxN luma blocks)
- Working towards issue #11.
- Moved all const arrays from .h to the .c. These are not used anywhere else.
- Moved entropy_bits array and its helper macro to rdo.c. They are not used
anywhere else.
- Implicit conversions to explicit ones.
- Working towards issue #11.
- Change lambda cost multiplier for intra NxN to from 256 to 4.
- Add 0.5 to the lambda multipliers so it's rounded instead of truncated.
- Change stderr back to text mode. Can't see what harm it should allow us to print
correct newlines for different platforms.
- Remove copy-pasted function calls from usage printing.
No need to be so clever for something this simple.
I moved the cost initialization outside the recursing function because it
relied on the clever recursion to work. It should eventually be moved to its
own function that also initializes all the other fields just to be safe. I
didn't do that yet because I want to do it per-LCU and these functions are
still working on per-frame basis.
The search_buildReferenceBorder was an ugly hack and a place for bugs to hide
that should never have existed. Now it doesn't.
The change reduces PSNR a little, but also reduces the bitrate, when the
expected result was to have no change in either. I'm guessing there was still
some bug in the search_buildReferenceBorder, but the bug could also be in
intra_build_reference_border. Will have to do more testing to be sure, but
having one place to look at will be better than having two.
Enables the output of spec-compliant byte streams, as the specification
notes that an additional zero_byte has to be added under certain
circuimstances.
Intra search was using faulty border data and selected modes were a bit random. Around -0.5% (LP) and -2% (AI) change in BD-rate was seen in limited testing conditions.
Name mangling is causing problems on different platforms (issues #2 and #3) and some of it was solved modifying the Makefile. Separate 32bit and 64bit assembly functions were also causing problems and since we were going to move to x264asm abstraction layer anyway, we decided to do it now before spending time on the old implementation.
Conflicts:
src/encoder.c
- Chroma RDOQ changes conflicted because I had moved the chroma
quantization/dequantization to it's own function.
- Merged to master because I want my code to show up in github. =)
All the old stuff still works, even though NxN doesn't work, so there
is no reason not to merge anyway.
- Re-enable intra search based on reconstructed image.
- This didn't have as much of an effect as I thought it would.
- Re-enable SAO and deblocking.
- Disable NxN searching. (4x4 luma coding is still broken)
Still doesn't work. I have no idea what the problem is. Probably somehow related to the coefficient coding, since the bitstream seems to work, the prediction is correct and the error is not very severe.
- Change scan order selection to be more verbose and based on the correct mode for 4x4. Didn't affect the problem with 4x4 luma in any way although it should have.
- Re-enable residual coding as everything seems to work now besides 4x4 luma.
New implementation uses precalculated tables to look up number of reference
pixels available in coded CUs. Otherwise it works just like the previous
version.
- NxN mostly works. Prediction appears to be almost correct but there is a
slight error in the exact values, probably related to filtering.
- Set part-size for Inter.
- Change to Intra Only mode for testing.
- Many small changes here and there. Should have been separate commits probably, but too late.
- Disable SAO and deblocking to be able to see problems with reconstruction better.
- Adjust predictor list to take modes from PUs in addition to 2Nx2N CUs.
- Change intra_get_dir_luma_predictor to take PU index instead of CU index.
- Comment prediction encoding now that I've had to look it up.
- Clean up code and comment.
- Change terminology to match H.265 specification where possible.
- Move transform splitting for depth==0 out of the coding part. It's not
possible to do it here anyway because intra reconstruction is different
if the transform is split.
- Add checking for transform hierarchy depth when coding split flag.
- Fixes bug with cu_data.tr_depth being set. The CU struct was being reused
for inter coded CUs, which did not initialize the tr_depth.
- Add room to cu_data.coeff_top_yuv arrays for the 4x4 PUs data. Will probably have to do the same to other coeff flags. The flags could also probably be combined as they are a bit redundant.
- Doesn't work yet so it's disabled.
- Change encode_transform_coeff to accept PU (Prediction Unit) coordinates
instead of CU coordinates because CUs are 8x8.
Should work exactly the same, but with the prediction mode selection done
separately from the binarization it's easier to see that the implementation
is correct.