- In InterPrediction::rightShiftMSB, use existing floorLog2 function

- In addAvg_SSE, avoid unnecessary use of 32-bit path

- In copyBufferSimd, reduce number of loops

- In paddingSimd, take advantage of padding extent being either 1 or 2

- In addBIOAvg4_SSE, reduce number of operations and avoid nasty
Xmm register -> memory -> integer register path

Overall, a decoder runtime reduction of about 3% is expected

Note that lines 450-457 should already support the loosened
restriction

Overall decoder run time is reduced by about 5%

SIMD code rewritten to run 2x faster

Clean up virtual boundary related checks:
- Replace modulo operation by bitwise AND
- Define local variable instead of having same computation appear
  dozens of times in code
- Use const type qualifiter where appropriate

Read 32 bits instead of 64 bits for blocks of width 2 such as to
not read past array boundaries

Adopt JVET-O0070 method 4-2.1a

Jonathan Taquet authored 5 years ago and Karsten Suehring committed 5 years ago

includes offline discussed replacement of tb(v) by ue(v) coding for alf_chroma_num_alt_filters_minus1

  - Use different SIMD operations as requested.

  - Multiplication of "d" and "hv" terms overlows signed 32-bit data type.

Extended the HAD calculation with SIMD to be performed in 32-bit space rather than 16-bit. This fixes #339

Fix ticket #267 on wrong SAD calculation in xGetSAD_NxN_SIMD(...) if rcDtParam.subShift is not equal to zero and input size is equal to 4x4.