Eyevinn / mp4ff / 37003905289

83%

push

github

perf(bits,avc): read EBSP without allocating per byte

EBSPReader.Read fetched every byte with binary.Read, which allocates since the byte escapes through an interface: 13 allocations to parse an AVC slice header and 28 for an SPS. Read now uses ReadByte when the reader is an io.ByteReader, as the bytes.Reader and bytes.Buffer of all AVC and HEVC parameter set and slice header parsers are, and io.ReadFull into a local array otherwise.

NewEBSPReaderFromSlice reads directly from a slice and goes back in it for MoreRbspData without seeking. ParseSliceHeader uses it, and the new avc.SliceHeaderSize returns only the header size without allocating a SliceHeader, which is all that cbcs subsample encryption needs.

BenchmarkParseNALUnits (Apple M4 Pro, n=10):
- ParseSliceHeader: 341 ns -> 263 ns, 13 -> 1 allocations
- ParseSPSNALUnit: 568 ns -> 359 ns, 28 -> 3 allocations
- SliceHeaderSize: 244 ns, 0 allocations

TestEBSPReaderKinds runs the same reads, Exp-Golomb codes, MoreRbspData and trailing bits, across emulation prevention bytes and the end of data, through a slice, an io.ReadSeeker and a plain io.Reader, and requires identical values, positions and errors. TestEBSPReaderAllocations fails on master with 101 allocations for 100 bytes. TestSliceHeaderSize checks that SliceHeaderSize returns ParseSliceHeader's size and error for the test streams and for truncated, weighted-prediction and unknown-PPS slices.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

44 of 48 new or added lines in 2 files covered. (91.67%)

23058 of 27927 relevant lines covered (82.57%)

0.83 hits per line