marfrit
a18ba53d6b
Phase 3 deep strace: only meaningful SPS diff is bytes 10-11.
libva bytes 10-11 = 00 00 (sps_max_num_reorder_pics=0, latency=0)
kdirect bytes 10-11 = 02 04 (reorder=2, latency=4)
Hardcoded at h265.c:110-111 with comment "/* not exposed */". VAAPI's
VAPictureParameterBufferHEVC doesn't forward these; kdirect parses
SPS NAL directly. sps_max_num_reorder_pics = 0 tells rkvdec "no
reordering" -> B-frame decode blocked -> all-zero output (Bug 5 fits).
Secondary diffs (Phase 4b candidates if α-13 doesn't close):
- SLICE_PARAMS num_entry_point_offsets = 0 (hardcoded at h265.c:356
with "iter2 doesn't do tiles" comment); kdirect submits 22.
- PPS UNIFORM_SPACING flag bit 20 (don't-care for non-tiled).
Phase 4 α-13: ~2 LOC fix. Set sps_max_num_reorder_pics =
sps_max_dec_pic_buffering_minus1 (safe upper bound per H.265 §A.4.2).
Leave sps_max_latency_increase_plus1 = 0 (spec "unconstrained").
Phase 5b review required before Phase 6b implementation per
"reviews never skippable".
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
105 lines
4.6 KiB
Markdown
105 lines
4.6 KiB
Markdown
# Iteration 11 — Phase 3 (HEVC libva vs kdirect deep strace)
|
|
|
|
Captured 2026-05-14 on fresnel (bare hostname, back on LAN). Backend SHA `a17e3c39…` (iter9 close). Fixture `bbb_720p10s_hevc.mp4` 3-frame sweep.
|
|
|
|
## Anchors
|
|
|
|
| Path | Hash |
|
|
|---|---|
|
|
| libva HEVC | `06b2c5a0c01e515d…` (all-zero, Bug 5 reproduced) |
|
|
| kdirect HEVC | `9340b832e08c23c4d4b2f732cecffb1d4104f2fc3257e561fac217c2d70d84c2` (correct decoded BBB) |
|
|
|
|
## HEVC control payload comparison — frame 1
|
|
|
|
Per-request `S_EXT_CTRLS` (ctrl_class=0xf010000):
|
|
|
|
| Control | libva size | kdirect size | Diff? |
|
|
|---|---|---|---|
|
|
| HEVC_SPS (0xa40a90) | 40 | 40 | **DIFF at bytes 10-11** |
|
|
| HEVC_PPS (0xa40a91) | 64 | 64 | **DIFF at bytes 56-57** |
|
|
| HEVC_SLICE_PARAMS (0xa40a92) | 280 | 280 | **DIFF at bytes 0-3 + 8-11 + 22** |
|
|
| HEVC_SCALING_MATRIX (0xa40a93) | 1000 | 1000 | identical (all-zero) |
|
|
| HEVC_DECODE_PARAMS (0xa40a94) | 328 | 328 | DIFF at tail (poc / flags) |
|
|
|
|
### Diff 1 — SPS bytes 10-11
|
|
|
|
```
|
|
libva bytes 0..15: 00 00 00 05 d0 02 00 00 04 04 [00 00] 01 01 00 03
|
|
kdirect bytes 0..15: 00 00 00 05 d0 02 00 00 04 04 [02 04] 01 01 00 03
|
|
```
|
|
|
|
Field mapping per `struct v4l2_ctrl_hevc_sps`:
|
|
- byte 10: `sps_max_num_reorder_pics` (u8)
|
|
- libva: 0
|
|
- kdirect: 2
|
|
- byte 11: `sps_max_latency_increase_plus1` (u8)
|
|
- libva: 0
|
|
- kdirect: 4
|
|
|
|
**Highly suspect**. `sps_max_num_reorder_pics = 0` tells the decoder "no reordering needed in this stream" — but BBB HEVC has B-frames in IBBP order. The kernel may refuse to set up reorder buffers, producing zero output for B-frames.
|
|
|
|
### Diff 2 — PPS bytes 56-63 (flags u64)
|
|
|
|
```
|
|
libva PPS flags: 0x0000_0000_0000_7040
|
|
kdirect PPS flags: 0x0000_0000_0010_7040
|
|
```
|
|
|
|
Diff = bit 20 = `V4L2_HEVC_PPS_FLAG_UNIFORM_SPACING` (set in kdirect, missing in libva). UNIFORM_SPACING is meaningful only when `num_tile_columns_minus1 > 0 || num_tile_rows_minus1 > 0`. For non-tiled streams it's don't-care. BBB HEVC may or may not use tiles — check `num_entry_point_offsets` in slice (Diff 3).
|
|
|
|
### Diff 3 — SLICE_PARAMS bytes 0-31
|
|
|
|
```
|
|
libva head32: d0d30b00 2d000000 00000000 14010200 00000000 0000ff00 07000000 00000000
|
|
kdirect head32: 80d20b00 2d000000 16000000 14010200 00000000 00000000 07000000 00000000
|
|
^^^^^^^^ ^^^^^^^^ ^^
|
|
bit_size num_entry_pt byte 22
|
|
```
|
|
|
|
- bytes 0-3 `bit_size` u32 LE: libva=0x000bd3d0 (774096), kdirect=0x000bd280 (773760). Diff = 336 bits ≈ 42 bytes. Small but present.
|
|
- bytes 8-11 `num_entry_point_offsets` u32 LE: **libva=0, kdirect=22**. BBB HEVC slice IS tiled / wavefront-parallel — kdirect submits 22 entry points; libva hardcodes 0.
|
|
- byte 22: libva=0xff, kdirect=0x00. Per struct layout this is `slice_qp_delta` (s8): libva=-1, kdirect=0.
|
|
|
|
### libva source code — where the zeros come from
|
|
|
|
`src/h265.c::h265_fill_sps` lines 110-111:
|
|
|
|
```c
|
|
sps->sps_max_num_reorder_pics = 0; /* not exposed */
|
|
sps->sps_max_latency_increase_plus1 = 0; /* not exposed */
|
|
```
|
|
|
|
VAAPI's `VAPictureParameterBufferHEVC` does not include these fields. Libva fills 0 by default.
|
|
|
|
`src/h265.c::h265_fill_slice_params` line 356:
|
|
|
|
```c
|
|
slice_params->num_entry_point_offsets = 0; /* iter2 doesn't do tiles */
|
|
```
|
|
|
|
Hardcoded 0 with a comment from iter2 acknowledging the scope-cut.
|
|
|
|
## Conclusion
|
|
|
|
Bug 5 is caused by **at least one** of:
|
|
- SPS `sps_max_num_reorder_pics` hardcoded 0 (vs kdirect's 2 from SPS-NAL parse). Strongest suspect — directly blocks B-frame reordering.
|
|
- SLICE_PARAMS `num_entry_point_offsets` hardcoded 0 (vs kdirect's 22). Affects tile/WPP parallel decode.
|
|
- PPS UNIFORM_SPACING flag missing. Don't-care for non-tiled but BBB does use tiles per Diff 3.
|
|
|
|
The classes are analogous to iter8's H.264 `constraint_set_flags` situation: VAAPI doesn't forward the field, libva fills a placeholder, the kernel cares.
|
|
|
|
## Phase 4 plan (preview)
|
|
|
|
Single iter11 commit, 5-10 LOC change in h265.c:
|
|
- Set `sps->sps_max_num_reorder_pics` from `picture->sps_max_dec_pic_buffering_minus1` (safe upper bound).
|
|
- Leave `sps->sps_max_latency_increase_plus1 = 0` (spec default: "no constraint").
|
|
|
|
Secondary attempts if first doesn't fix:
|
|
- Set PPS UNIFORM_SPACING flag when tiles are used (`num_tile_columns_minus1 > 0 || num_tile_rows_minus1 > 0`).
|
|
- Parse slice header for `num_entry_point_offsets` (analogous to h264_slice_header.c). Costlier; defer to Phase 4b if needed.
|
|
|
|
Phase 5 review will scrutinize:
|
|
- Whether `sps_max_num_reorder_pics = sps_max_dec_pic_buffering_minus1` is safe vs producing too many reorder pictures.
|
|
- Whether reaching for the bitstream's true value (SPS NAL parse) is needed for correctness.
|
|
- That the change doesn't regress VP9 / VP8 / H.264 / MPEG-2.
|