marfrit/daedalus-decoder
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

phase1: IDCT 8x8 dispatch (High profile transform_8x8_size_flag)

Browse Source 
Adds the High-profile 8x8 luma transform path alongside the existing
4x4 dispatch.  flush_frame now partitions macroblocks by each MB's
transform_8x8 flag and issues a separate luma dispatch per partition:

  - mb.transform_8x8 == 0 (Baseline/Main) → coeffs[0..256) interpreted
    as 16 4x4 blocks, fed to daedalus_recipe_dispatch_h264_idct4
    (existing behaviour, unchanged).
  - mb.transform_8x8 == 1 (High)          → coeffs[0..256) interpreted
    as 4 8x8 blocks (64 int16 each, column-major), fed to the new
    daedalus_recipe_dispatch_h264_idct8 call.

Both luma partitions can be non-empty in the same frame (FFmpeg sets
the flag per-MB).  Each non-empty partition costs one
vkQueueSubmit + vkQueueWaitIdle; empty partitions are skipped
(common case: Baseline streams skip the 8x8 dispatch entirely).

Chroma is unchanged — 4:2:0 chroma always uses the 4x4 transform.

API surface:
  - New uint8_t `transform_8x8` field in `struct daedalus_decoder_mb_input`
    (after deblock_*).  Backwards-compatible at the source level
    because the field defaults to 0 with C99 designated initialisers
    or {0} struct zeroing, both of which select the existing 4x4
    path.  ABI is pre-0.1 (per the header doc) so structural change
    is fine.
  - Mirrored in `struct daedalus_decoder_mb_desc` (internal layout).

Test changes:

  - test_idct_bitexact now exercises a mixed-mode frame: every odd
    raster MB uses 8x8, every even uses 4x4 (so flush_frame's
    partitioning is also under test, not just the underlying shaders).
  - 8x8 C reference (h264_idct8_butterfly + ref_idct8_add)
    transcribed from daedalus-fourier tests/h264_idct8_ref.c per
    H.264 §8.5.13.2.  Block layout column-major; +32 >> 6 rounding;
    add-to-predicted; clip255.
  - Reference luma compute branches per MB on the same mb_8x8[]
    array used to set the input flag.

Verified on hertz (Pi 5 / V3D 7.1 / daedalus-fourier 0.1.0):

  $ ./build/test_idct_bitexact
  test_idct_bitexact: 320x240 (300 MBs), seed=0xfeedface5a5a5a5a
  MB mix: 150 4x4 MBs, 150 8x8 MBs
  Y bytes total:  76800
  Y bytes diff:   0 (0.0000%)
  Cb bytes total: 19200  diff: 0 (0.0000%)
  Cr bytes total: 19200  diff: 0 (0.0000%)
  BIT-EXACT PASS (Y + Cb + Cr)

  $ ctest --test-dir build
  100% tests passed, 0 tests failed out of 2

Bit-exact PASS first try for the 8x8 path — 150 8x8 MBs × 4 blocks =
600 8x8 IDCTs against the spec C reference, identical output.
Validates both the daedalus-fourier IDCT 8x8 shader (already gated
by its own cycle-7 bit-exact test, now also gated end-to-end through
our flush_frame), and our 8x8 layout assumptions (column-major coeffs,
raster sb_y*2+sb_x block order, top-left = mb*16 + sb*8).

What's NOT covered yet (deferred):

  - Z-scan permutation for FFmpeg compatibility (libavcodec intercept
    patch's concern; both 4x4 and 8x8 z-scans differ).
  - Chroma DC / luma Intra16x16 DC Hadamard pre-pass.
  - Mixed intra/inter MB handling — currently all MBs treated as
    residual-only (predicted=0).

Closes the "IDCT 8x8 (High profile)" item from PR #3's deferred list.
This commit is contained in:
Claude (noether)
2026-05-24 22:41:05 +02:00
parent 5fa495964d
commit adaabb1f63
4 changed files with 203 additions and 73 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/daedalus_decoder.c
+82 -49
View File
@@ -119,6 +119,7 @@ int daedalus_decoder_append_mb(daedalus_decoder *dec,
d->deblock_disable = mb->deblock_disable;
d->deblock_alpha_c0 = mb->deblock_alpha_c0;
d->deblock_beta = mb->deblock_beta;
d->transform_8x8 = mb->transform_8x8;
memcpy(&dec->coeffs[(size_t) expected * 384],
mb->coeffs,
@@ -179,74 +180,104 @@ int daedalus_decoder_flush_frame(daedalus_decoder *dec,
int rc = 0;
/* ---- Build frame-scaled luma-4x4 dispatch ---- */
/* ---- Build frame-scaled luma dispatches (4x4 + 8x8) ---- */
const size_t n_luma_blocks_per_mb = 16;
const size_t n_luma_blocks = (size_t) dec->n_mbs * n_luma_blocks_per_mb;
/* Scratch Y plane — coded-size byte buffer. Zero-initialised so
* the IDCT-ADD-clip operation reduces to clip255(IDCT) per block
* (predicted=0 because no intra/MC has run yet). */
/* Two partitions of the per-MB luma section based on each MB's
* transform_8x8 flag:
*
* transform_8x8 == 0 → 16 4x4 blocks contribute to the 4x4
* dispatch (16 coeffs each).
* transform_8x8 == 1 → 4 8x8 blocks contribute to the 8x8
* dispatch (64 coeffs each).
*
* Both partitions can be non-empty in the same frame (FFmpeg sets
* transform_8x8_size_flag per MB), so we allocate worst-case for
* each and track actual counts.
*/
const size_t y_stride_int = (size_t) dec->width;
const size_t y_size = y_stride_int * (size_t) dec->height;
uint8_t *scratch_y = calloc(1, y_size);
int16_t *flat_coeffs = malloc(n_luma_blocks * 16 * sizeof(int16_t));
daedalus_h264_block_meta *meta = malloc(
n_luma_blocks * sizeof(daedalus_h264_block_meta));
if (!scratch_y || !flat_coeffs || !meta) {
const size_t worst_4x4 = (size_t) dec->n_mbs * 16;
const size_t worst_8x8 = (size_t) dec->n_mbs * 4;
int16_t *coeffs4 = malloc(worst_4x4 * 16 * sizeof(int16_t));
int16_t *coeffs8 = malloc(worst_8x8 * 64 * sizeof(int16_t));
daedalus_h264_block_meta *meta4 = malloc(worst_4x4 * sizeof(*meta4));
daedalus_h264_block_meta *meta8 = malloc(worst_8x8 * sizeof(*meta8));
if (!scratch_y || !coeffs4 || !coeffs8 || !meta4 || !meta8) {
rc = -1;
goto cleanup;
}
/* Raster-order layout: walk each MB, then each of its 16 luma 4×4
* sub-blocks in raster order (sb_y=0..3 outer, sb_x=0..3 inner).
/* Walk MBs in raster order, append each MB's luma blocks to the
* partition selected by its transform_8x8 flag.
*
* NB: H.264's actual per-MB 4×4 coefficient scan order is the
* z-scan from spec §6.4.3 / fig 6-10. We're using a flat raster
* here because Phase 1 stage 1 only validates the dispatch
* round-trip; bit-exact against an FFmpeg reference requires the
* z-scan permutation and is a follow-on test. The per-MB
* coeffs[] field's first 256 entries are interpreted as 16
* consecutive 4×4 blocks in the same raster order on the input
* side, so this is self-consistent for the validation. */
size_t bi = 0;
* NB: per-MB 4x4 / 8x8 coefficient ORDER inside the H.264 bitstream
* follows the z-scan from spec §6.4.3 / fig 6-10. We're using
* flat raster on the input side too (sb_y outer, sb_x inner) for
* Phase 1 self-consistency; the z-scan permutation is the
* libavcodec-intercept patch's responsibility.
*/
size_t bi4 = 0, bi8 = 0;
for (int mb_y = 0; mb_y < dec->mb_height; mb_y++) {
for (int mb_x = 0; mb_x < dec->mb_width; mb_x++) {
int mb_idx = mb_y * dec->mb_width + mb_x;
const struct daedalus_decoder_mb_desc *d = &dec->mb_descs[mb_idx];
const int16_t *mb_coeffs = &dec->coeffs[(size_t) mb_idx * 384];
for (int sb_y = 0; sb_y < 4; sb_y++) {
for (int sb_x = 0; sb_x < 4; sb_x++) {
/* Block top-left pixel in the coded Y plane. */
size_t px_y = (size_t) mb_y * 16 + (size_t) sb_y * 4;
size_t px_x = (size_t) mb_x * 16 + (size_t) sb_x * 4;
meta[bi].dst_off = (uint32_t) (px_y * y_stride_int + px_x);
/* Copy 16 coeffs for this block from the per-MB
* coeffs[] (luma offset = block_idx * 16). */
int block_in_mb = sb_y * 4 + sb_x;
memcpy(&flat_coeffs[bi * 16],
&mb_coeffs[block_in_mb * 16],
16 * sizeof(int16_t));
bi++;
if (d->transform_8x8) {
/* 4 luma 8x8 blocks, raster sb_y*2+sb_x. */
for (int sb_y = 0; sb_y < 2; sb_y++) {
for (int sb_x = 0; sb_x < 2; sb_x++) {
size_t px_y = (size_t) mb_y * 16 + (size_t) sb_y * 8;
size_t px_x = (size_t) mb_x * 16 + (size_t) sb_x * 8;
meta8[bi8].dst_off = (uint32_t)
(px_y * y_stride_int + px_x);
int block_in_mb = sb_y * 2 + sb_x;
memcpy(&coeffs8[bi8 * 64],
&mb_coeffs[block_in_mb * 64],
64 * sizeof(int16_t));
bi8++;
}
}
} else {
/* 16 luma 4x4 blocks, raster sb_y*4+sb_x. */
for (int sb_y = 0; sb_y < 4; sb_y++) {
for (int sb_x = 0; sb_x < 4; sb_x++) {
size_t px_y = (size_t) mb_y * 16 + (size_t) sb_y * 4;
size_t px_x = (size_t) mb_x * 16 + (size_t) sb_x * 4;
meta4[bi4].dst_off = (uint32_t)
(px_y * y_stride_int + px_x);
int block_in_mb = sb_y * 4 + sb_x;
memcpy(&coeffs4[bi4 * 16],
&mb_coeffs[block_in_mb * 16],
16 * sizeof(int16_t));
bi4++;
}
}
}
}
}
/* assert bi == n_luma_blocks; the loop math guarantees it */
/* assert bi4 + bi8*4 == n_mbs*16; loop math guarantees it */
/* ---- One Vulkan submit + wait for the whole frame's luma IDCT.
/* ---- One Vulkan submit + wait per non-empty luma partition.
* AUTO substrate picks QPU per the post-decree recipe table; falls
* back to CPU NEON if the daedalus-fourier ctx wasn't QPU-capable. */
int dr = daedalus_recipe_dispatch_h264_idct4(dec->dctx,
scratch_y, y_stride_int,
flat_coeffs,
n_luma_blocks,
meta);
if (dr != 0) {
rc = -3; /* GPU dispatch failure */
goto cleanup;
* back to CPU NEON if the daedalus-fourier ctx wasn't QPU-capable.
* Skipping the dispatch when the partition is empty avoids the
* shader-pool warm-up cost on the common case (a typical Baseline
* stream is all-4x4 → 8x8 dispatch is no-op). */
if (bi4 > 0) {
int dr = daedalus_recipe_dispatch_h264_idct4(dec->dctx,
scratch_y, y_stride_int,
coeffs4, bi4, meta4);
if (dr != 0) { rc = -3; goto cleanup; }
}
if (bi8 > 0) {
int dr = daedalus_recipe_dispatch_h264_idct8(dec->dctx,
scratch_y, y_stride_int,
coeffs8, bi8, meta8);
if (dr != 0) { rc = -3; goto cleanup; }
}
/* ---- Copy Y out to caller's plane at the requested stride. ---- */
@@ -362,8 +393,10 @@ int daedalus_decoder_flush_frame(daedalus_decoder *dec,
}
cleanup:
free(meta);
free(flat_coeffs);
free(meta8);
free(meta4);
free(coeffs8);
free(coeffs4);
free(scratch_y);
dec->mbs_appended = 0;
return rc;
src/internal.h
+2 -1
View File
@@ -41,7 +41,8 @@ struct daedalus_decoder_mb_desc {
uint8_t deblock_disable;
int8_t deblock_alpha_c0;
int8_t deblock_beta;
uint8_t _pad1;
uint8_t transform_8x8; /* 0 = 4 luma blocks of 4x4 (16 total),
* 1 = 4 luma blocks of 8x8. */
};
struct daedalus_decoder {