ParseSliceHeader silently drops the sole SPS long-term reference and misreads headers

perfloop-oss/mp4ff · Uncatalogued correctness

https://perfloop.ai/t/oss/case_6fbmm873n3

Verdict

VERIFIED · settled 2026-11-14

What happened: The check fails before this change and passes with it.

Hypothesis

A slice that reuses the only long-term picture in its sequence parameter set (SPS) can be returned with that reference replaced by zero values and no error. The resulting in-use picture count can shift later header fields and the byte boundary used for CBCS encryption. Occurrence in encoder-produced streams and real-stream prevalence remain unknown.

In hevc/slice.go:234-255, ParseSliceHeader initializes lt to zero and copies sps.LongTermRefPicSets[LtIdxSps] only inside the sps.NumLongTermRefPics > 1 condition at lines 237-243. For an inherited entry when the SPS has exactly one long-term picture, neither PocLsbLt nor UsedByCurrPicLtFlag is copied. That flag contributes to NumPicTotalCurr at lines 248-250. The count gates reference-list-modification parsing at lines 289-300, before QpDelta is read at line 334 and Size is calculated at lines 424-427. Readers of LongTermRefPicSets, QpDelta or Size are affected for such inputs. The package's slice-header parsing is documented at README.md:137; mp4/crypto.go:313-320 calls this parser and uses Size as the CBCS clear-header range.

The submitter describes the smallest failing input as an SPS with long_term_ref_pics_present_flag=1 and exactly one long-term entry, lt_ref_pic_poc_lsb_sps[0]=100 and used_by_curr_pic_lt_sps_flag[0]=1. Its hex is 42010101600000030090000003000003005da02081059657aa481ae55a6484. The reported P-slice inputs use num_long_term_sps=1: - With PPS 4401c0718012 and slice 0201d02ca208c0a55a80, the submitter reports a nil error and LongTermRefPicSets[0]={PocLsbLt:0 UsedByCurrPicLtFlag:false}, instead of {PocLsbLt:100 UsedByCurrPicLtFlag:true}. - With PPS 4401c0718032, which enables lists_modification_present_flag, and slice 0201d02ca10f80a55a80, the submitter reports a nil error, QpDelta=-3 and Size=6 bytes, instead of QpDelta=-7 and Size=7 bytes.

A focused unchanged-code probe during anchoring reproduced the zeroed entry in both examples and the reported QpDelta=-3, Size=6, nil-error result in the second example. This does not independently establish the expected values or exposure in real streams.

The submitter cites H.265 7.4.7.1: 'If i is less than num_long_term_sps, PocLsbLt[ i ] is set equal to lt_ref_pic_poc_lsb_sps[ lt_idx_sps[ i ] ]'. They state that lt_idx_sps is not coded when the SPS has one entry and is then inferred as 0. They also cite FFmpeg n7.1 libavcodec/hevc/hevcdec.c:314-320, https://github.com/FFmpeg/FFmpeg/blob/n7.1/libavcodec/hevc/hevcdec.c#L314-L320: 'uint8_t lt_idx_sps = 0; if (sps->num_long_term_ref_pics_sps > 1) lt_idx_sps = get_bits(...)'. These are the reported independent semantic references to check.

The submitter reports that using index 0 for one entry makes all 680 tried header variants parse as expected; that sweep has not been rerun here. They also report that none of the 147 JCT-VC HEVC_v1 conformance streams has exactly one SPS long-term entry. The supplied input is built from the syntax, not from an encoder. The sweep and conformance-coverage reports remain attributed, unconfirmed observations, not evidence of stream prevalence.

A provisional first check is a focused Go regression that parses the supplied SPS/PPS/slice triples and compares inherited long-term fields, QpDelta, Size and errors with an independent syntax-derived oracle, including whether list-modification bits must be consumed. The stated expected results confirm the report if the syntax check agrees; contrary independent results reject those expectations. Then check the reported variant sweep and the downstream CBCS boundary, retaining multiple-entry and invalid-index controls. This proposed test layout, placement and oracle implementation are provisional proof methods, not additional frozen requirements.

Current upstream checks returned no open pull requests at https://github.com/Eyevinn/mp4ff/pulls, and the current https://github.com/Eyevinn/mp4ff/blob/master/hevc/slice.go still has the guarded copy. The inspected recent bounds change, https://github.com/Eyevinn/mp4ff/commit/aa5c752772a2a65be0a0d2a73e1857b0c622ea08, adds count and loop limits rather than fixing single-entry inheritance; this outcome must retain its checks. History coverage is bounded: the history API returned 403, and a 40-commit git deepening supplied only recent history. CONTRIBUTING.md names issues and pull requests as the contribution venue; no separate review venue was identified.

Change to test: Restore SPS-derived long-term picture values and downstream header alignment when the SPS has exactly one long-term entry, while retaining index bounds validation. In ParseSliceHeader's long-term loop, default the index to 0, read lt_idx_sps only when sps.NumLongTermRefPics > 1, and perform the bounds check and sps.LongTermRefPicSets[index] copy outside that condition for every entry with i < num_long_term_sps.

Where it lives

perfloop-oss/mp4ff

Evidence

The check fails before this change and passes with it: `For the supplied sole-long-term-reference P-slices and syntax-built I/P/B slices with 0, 1, 2, 3 or 32 SPS long-term pictures, ParseSliceHeader returns the syntax-derived long-term reference fields, list entries, QP and byte-aligned header size; supplied CBCS samples keep the seven-byte slice header clear with external and in-band parameter sets, and invalid inherited indexes are rejected.`

The check fails before this change and passes with it: `For syntax-built I/P/B slices with zero, one, two or three SPS short-term reference sets, ParseSliceHeader returns the inline or inferred or signaled set and syntax-derived list entries, QP -5 and aligned size; CBCS ranges exclude the complete header with external or in-band parameter sets, three-set invalid indexes are rejected, repeated calls preserve parameter sets, and combined sole-short/long inheritance returns both reference kinds.`

ParseSliceHeader SPS EmptyI QP0 · 21 sample pairs

metric baseline candidate paired median change confidence range required result
ns/op 188.5 194.6 +7.3 −8.7 to +32.3 loss ≤ stated 0.5 × |baseline median| PASSED
B/op 400 400 0 0 to 0 loss ≤ 0.05 × |baseline median| PASSED
allocs/op 2 2 0 0 to 0 loss ≤ 0.05 × |baseline median| PASSED

ParseSliceHeader SPS EmptyI QP-5 · 21 sample pairs

metric baseline candidate paired median change confidence range required result
ns/op 209.7 219.2 +7.5 −4.2 to +27.3 loss ≤ stated 0.5 × |baseline median| PASSED
B/op 400 400 0 0 to 0 loss ≤ 0.05 × |baseline median| PASSED
allocs/op 2 2 0 0 to 0 loss ≤ 0.05 × |baseline median| PASSED

ParseSliceHeader SPS EmptyP QP0 · 21 sample pairs

metric baseline candidate paired median change confidence range required result
ns/op 193.9 197.7 +9 −12.1 to +23.1 loss ≤ stated 0.5 × |baseline median| PASSED
B/op 400 400 0 0 to 0 loss ≤ 0.05 × |baseline median| PASSED
allocs/op 2 2 0 0 to 0 loss ≤ 0.05 × |baseline median| PASSED

ParseSliceHeader SPS EmptyP QP-5 · 21 sample pairs

metric baseline candidate paired median change confidence range required result
ns/op 222.1 226.3 +4.2 −7.7 to +12.6 loss ≤ stated 0.1 × |baseline median| PASSED
B/op 400 400 0 0 to 0 loss ≤ 0.05 × |baseline median| PASSED
allocs/op 2 2 0 0 to 0 loss ≤ 0.05 × |baseline median| PASSED

ParseSliceHeader SPS EmptyB QP0 · 21 sample pairs

metric baseline candidate paired median change confidence range required result
ns/op 202.4 204.5 +3.5 −18.5 to +29.9 loss ≤ stated 0.65 × |baseline median| PASSED
B/op 400 400 0 0 to 0 loss ≤ 0.05 × |baseline median| PASSED
allocs/op 2 2 0 0 to 0 loss ≤ 0.05 × |baseline median| PASSED

ParseSliceHeader SPS EmptyB QP-5 · 21 sample pairs

metric baseline candidate paired median change confidence range required result
ns/op 215.6 222.9 +4.8 −7.6 to +34.6 loss ≤ stated 0.65 × |baseline median| PASSED
B/op 400 400 0 0 to 0 loss ≤ 0.05 × |baseline median| PASSED
allocs/op 2 2 0 0 to 0 loss ≤ 0.05 × |baseline median| PASSED

ParseSliceHeader SPS TwoP QP0 · 21 sample pairs

metric baseline candidate paired median change confidence range required result
ns/op 198.7 209.9 +15.4 −3 to +50.5 loss ≤ stated 0.5 × |baseline median| PASSED
B/op 400 400 0 0 to 0 loss ≤ 0.05 × |baseline median| PASSED
allocs/op 2 2 0 0 to 0 loss ≤ 0.05 × |baseline median| PASSED

ParseSliceHeader SPS TwoP QP-5 · 21 sample pairs

metric baseline candidate paired median change confidence range required result
ns/op 210.8 229.2 +16.2 −5.2 to +34.2 loss ≤ stated 0.5 × |baseline median| PASSED
B/op 400 400 0 0 to 0 loss ≤ 0.05 × |baseline median| PASSED
allocs/op 2 2 0 0 to 0 loss ≤ 0.05 × |baseline median| PASSED

ParseSliceHeader SPS TwoPLists QP0 · 21 sample pairs

metric baseline candidate paired median change confidence range required result
ns/op 216.5 263.3 +21.1% (+45.7) +34.5 to +65.4 loss ≤ stated 0.6 × |baseline median| PASSED
B/op 416 465 +11.8% (+49) +49 to +49 loss ≤ stated 0.25 × |baseline median| PASSED
allocs/op 3 4 +33.3% (+1) +1 to +1 loss ≤ stated 0.5 × |baseline median| PASSED

ParseSliceHeader SPS TwoPLists QP-5 · 21 sample pairs

metric baseline candidate paired median change confidence range required result
ns/op 201.6 282 +39.9% (+80.4) +66.5 to +103.6 loss ≤ stated 0.8 × |baseline median| PASSED
B/op 400 465 +16.3% (+65) +65 to +65 loss ≤ stated 0.25 × |baseline median| PASSED
allocs/op 2 4 +100% (+2) +2 to +2 loss ≤ stated 1.25 × |baseline median| PASSED

ParseSliceHeader SPS 32P QP0 · 21 sample pairs

metric baseline candidate paired median change confidence range required result
ns/op 196.3 206.1 +11.3 −0.5 to +27.8 loss ≤ stated 0.5 × |baseline median| PASSED
B/op 400 400 0 0 to 0 loss ≤ 0.05 × |baseline median| PASSED
allocs/op 2 2 0 0 to 0 loss ≤ 0.05 × |baseline median| PASSED

ParseSliceHeader SPS 32P QP-5 · 21 sample pairs

metric baseline candidate paired median change confidence range required result
ns/op 222.7 234.6 +11.1 −7.5 to +51.2 loss ≤ stated 0.5 × |baseline median| PASSED
B/op 400 400 0 0 to 0 loss ≤ 0.05 × |baseline median| PASSED
allocs/op 2 2 0 0 to 0 loss ≤ 0.05 × |baseline median| PASSED

ParseSliceHeader sole long-term SPS with lists · 21 sample pairs

metric baseline candidate paired median change confidence range required result
ns/op 289.4 356.1 +22.7% (+65.7) +34 to +86.6 loss ≤ stated 0.5 × |baseline median| PASSED
B/op 416 480 +15.4% (+64) +64 to +64 loss ≤ stated 0.25 × |baseline median| PASSED
allocs/op 3 4 +33.3% (+1) +1 to +1 loss ≤ stated 0.5 × |baseline median| PASSED

Timeline