Checksum error when using biprediction, TMVP and --gop=8

[aryla]

Steps to reproduce:

$ kvazaar -i test_seqs/hevc-F/BasketballDrillText_832x480_50_500.yuv --preset ultrafast --bipred --no-deblock --no-wpp --threads 0 --gop 8 -n8 -o x.hevc
Using preset ultrafast: --pu-depth-intra=2-3 --pu-depth-inter=2-3 --rd=0 --me=hexbs --ref=1 --deblock=0:0 --signhide=0 --subme=0 --sao=0 --rdoq=0 --rdoq-skip=1 --transform-skip=0 --full-intra-search=0 --mv-rdo=0 --smp=0 --amp=0 --cu-split-termination=zero --me-early-termination=sensitive --gop=lp-g4d3t1 
Compiled: INTEL, flags: MMX SSE SSE2
Detected: INTEL, flags: MMX SSE SSE2 SSE3 SSSE3 SSE41 SSE42 AVX
Available: avx(11) sse2(2) sse41(1) 
In use: avx(11) 
--owf=auto value set to 0.
Input: test_seqs/hevc-F/BasketballDrillText_832x480_50_500.yuv, output: x.hevc
  Video size: 832x480 (input=832x480)
POC    0 QP 22 (I-frame)     489216 bits PSNR: 41.4679 43.6355 44.2168
POC    4 QP 24 (B-frame)     140704 bits PSNR: 39.6001 42.3790 42.9969 [L0 0 ] [L1 ]
POC    2 QP 25 (B-frame)      87432 bits PSNR: 39.2088 42.4912 43.2453 [L0 0 ] [L1 4 ]
POC    1 QP 26 (B-frame)       2464 bits PSNR: 41.4656 43.6372 44.2162 [L0 0 ] [L1 2 4 ]
POC    3 QP 26 (B-frame)      51736 bits PSNR: 39.1799 42.3489 43.0546 [L0 2 0 ] [L1 4 ]
POC    6 QP 25 (B-frame)      92600 bits PSNR: 39.2871 42.1256 42.7286 [L0 4 0 ] [L1 ]
POC    5 QP 26 (B-frame)      72072 bits PSNR: 38.7578 41.9125 42.6012 [L0 4 0 ] [L1 6 ]
POC    7 QP 26 (B-frame)      72552 bits PSNR: 38.8464 41.8501 42.4821 [L0 6 4 0 ] [L1 ]
 Processed 8 frames,    1008776 bits AVG PSNR: 39.7267 42.5475 43.1927
 Total CPU time: 1.832 s.
 Encoding time: 1.829 s.
 Encoding wall time: 1.829 s.
 Encoding CPU usage: 100.01%
 FPS: 4.37
$ TAppDecoderStatic -b x.hevc

HM software: Decoder Version [16.14] (including RExt)[Linux][GCC 6.3.1][64 bit] 
POC    0 TId: 0 ( I-SLICE, QP 22 ) [DT  0.070] [L0 ] [L1 ] [Checksum:0310c9d0,00c1b849,00c0e688,(OK)] 
POC    4 TId: 0 ( B-SLICE, QP 24 ) [DT  0.026] [L0 0 ] [L1 0 ] [Checksum:0312e7d3,00c10fa3,00c31741,(OK)] 
POC    2 TId: 0 ( B-SLICE, QP 25 ) [DT  0.018] [L0 0 ] [L1 4 ] [Checksum:030ea932,00c1a5d2,00c12340,(OK)] 
POC    1 TId: 0 ( B-SLICE, QP 26 ) [DT  0.010] [L0 0 ] [L1 2 4 ] [Checksum:0310c9fb,00c1ba09,00c0e4ea,(OK)] 
POC    3 TId: 0 ( B-SLICE, QP 26 ) [DT  0.014] [L0 2 0 ] [L1 4 ] [Checksum:030eac45,00c19207,00c21182,(***ERROR***)] [rxChecksum:030eab71,00c1921e,00c21166] 
POC    6 TId: 0 ( B-SLICE, QP 25 ) [DT  0.017] [L0 4 0 ] [L1 4 ] [Checksum:03148de4,00bf86b8,00c44e90,(OK)] 
POC    5 TId: 0 ( B-SLICE, QP 26 ) [DT  0.016] [L0 4 0 ] [L1 6 ] [Checksum:03171826,00c008d3,00c40421,(***ERROR***)] [rxChecksum:0317191e,00c0090e,00c40422] 
POC    7 TId: 0 ( B-SLICE, QP 26 ) [DT  0.015] [L0 6 4 0 ] [L1 6 ] [Checksum:030fb377,00bea647,00c4b22e,(OK)] 


***ERROR*** A decoding mismatch occured: signalled md5sum does not match

 Total Time:        0.248 sec.

The first differing pixel is the luma pixel at (336, 168) in frame with POC 3.

[aryla]

Seems to have first appeared in 4270d45. Before that TMVP was disabled whenever GOP was used.