[jcwal1516]

CPU entropy decoding is most of a decode on both routes. On a 2048² RGB q8 image it is about 50 of 55 ms on the Metal route, and for batches of 256×256 tiles it is the preparation stage. Profiling showed that the HP coefficient loop, about 60% of entropy time, called each per-symbol VLC helper out of line. Each call reloaded and stored the bit reader's fields, returned its Result through memory, and re-checked its table indices. Those helpers were 21% of main-thread samples on that image.

• Inlined symbol decoding. decode_index, decode_abs_level, decode_first_index, and decode_run_cached are forced inline into the LP and HP block loops. The bit-serial run decoder they fall back to near a packet end is now cold and out of line.
• By-value refill and errors. The out-of-line cache refill takes the packet slice, position, and length and returns the new cache and bit count, and the end-of-packet error constructor takes values. Calling either no longer passes the reader by address. read_bits is forced inline. This mostly helps flexbit-heavy 16-bit images: on its own, the inlined read_bits gained nothing there, and the by-value helpers gave most of the gain.
• No behavior change. Decoding logic, error variants, and error positions are unchanged. Each #[inline(always)] carries #[expect(clippy::inline_always)] with its measured reason.

Performance. "Before" is main at 8a7ccbe, which includes #12–#18, and each figure is the median of 5 interleaved rounds. Single images, jxr-load-bench, ms. Prepare is the Metal stage (parse, plan, and entropy into staging); CPU is the full CPU-route decode. The M4 and M5 Max columns show speedups:

Image	M4 Pro prepare	M4 Pro CPU	M4 prepare	M4 CPU	M5 Max prepare	M5 Max CPU
gray 8-bit q10	27.3 → 24.7	25.2 → 23.1	1.07x	1.07x	1.08x	1.07x
gray 16-bit q0	38.6 → 34.2	36.1 → 32.0	1.12x	1.11x	1.10x	1.10x
gray 16-bit q20	36.6 → 32.7	34.1 → 30.4	1.10x	1.09x	1.09x	1.08x
4096² gray 16-bit	150.3 → 134.1	138.4 → 123.0	1.11x	1.10x	1.09x	1.09x
RGB q0	82.3 → 73.9	87.0 → 78.3	1.09x	1.09x	1.09x	1.08x
RGB q8	55.2 → 50.1	59.7 → 54.2	1.08x	1.07x	1.08x	1.07x
RGB q30	37.1 → 33.8	41.2 → 37.9	1.08x	1.06x	1.07x	1.06x
RGB q20 overlap 0	47.6 → 43.0	51.4 → 46.8	1.09x	1.08x	1.08x	1.07x
RGB q20 overlap 2	42.4 → 38.6	46.8 → 42.9	1.08x	1.07x	1.08x	1.07x
RGB q20 frequency	47.1 → 44.0	51.4 → 48.1	1.07x	1.06x	1.06x	1.06x
RGB q20 4:2:0	27.3 → 24.8	39.6 → 37.3	1.07x	1.04x	1.06x	1.04x
RGBA q10	74.5 → 67.9	80.5 → 74.0	1.08x	1.07x	1.08x	1.07x
4096² RGB, 4×4 tiles	30.3 → 27.6	45.6 → 43.3	1.07x	1.04x	1.06x	1.04x

Images are 2048² unless noted.

256×256 Boat tiles, jxr-pathology-bench, ms. Prepare is the Metal host route's preparation, and the other columns are route totals:

Batch	Machine	Prepare	CPU host	Metal host	Metal dense	Metal batch API
1	M4 Pro	0.40 → 0.31	0.67 → 0.61	0.66 → 0.58	0.67 → 0.59	0.69 → 0.60
1	M4	0.36 → 0.30	0.61 → 0.56	0.65 → 0.59	0.65 → 0.59	0.66 → 0.62
1	M5 Max	0.34 → 0.29	0.71 → 0.66	0.58 → 0.52	0.58 → 0.52	0.61 → 0.55
8	M4 Pro	0.77 → 0.64	1.45 → 1.26	1.36 → 1.18	1.37 → 1.19	1.85 → 1.77
8	M4	0.96 → 0.81	1.61 → 1.53	2.21 → 1.81	1.97 → 1.80	2.29 → 2.18
8	M5 Max	0.64 → 0.59	1.41 → 1.32	1.09 → 1.01	1.02 → 1.00	1.19 → 1.14
32	M4 Pro	2.44 → 2.02	4.08 → 3.80	4.39 → 3.83	4.53 → 4.09	4.31 → 3.78
32	M4	3.19 → 2.70	5.38 → 4.96	5.83 → 5.29	5.91 → 5.41	5.80 → 5.35
32	M5 Max	1.62 → 1.49	3.22 → 3.07	2.98 → 2.63	2.97 → 2.87	3.58 → 3.18
128	M4 Pro	8.69 → 7.19	15.31 → 12.82	14.76 → 12.59	14.53 → 13.90	12.38 → 10.81
128	M4	11.87 → 10.12	19.42 → 17.67	20.51 → 18.66	19.74 → 17.98	15.61 → 13.55
128	M5 Max	5.10 → 4.55	9.01 → 8.89	9.28 → 8.45	10.36 → 10.00	8.10 → 7.01

Batch preparation is 1.06–1.29x faster on every machine. The M5 Max CPU-host route at 64 tiles (not shown) measured 5.54 → 5.64 ms; its 32- and 128-tile rows are faster.

MpsGraphBatchDecoder::prepare then run_prepared_group on the same tiles with the rgb8_nhwc_reference graph, M4 Pro, ms:

Batch	Prepare	Prepare + run
1	0.43 → 0.38	0.83 → 0.73
8	0.84 → 0.74	1.53 → 1.42
32	2.49 → 2.07	4.29 → 3.79
128	8.98 → 7.43	14.60 → 12.66

Three variants measured slower or neutral and are not included: decoding each block from a local copy of the reader so its fields stay in registers (neutral), inlining flexbit decoding into the YUV HP loop (0.67–0.86x), and an inline fast path for refills away from the packet end (0.87–0.94x).

Validation on an M4 Pro, on the earlier main (e52d95c) with this change: cargo fmt --all -- --check; cargo clippy --workspace --all-targets --all-features -- -D warnings, and the same with --target x86_64-unknown-linux-gnu; cargo test --workspace --all-features (272 passed); cargo test -p jxr --no-default-features (24 passed); RUSTDOCFLAGS="-D warnings" cargo doc --workspace --all-features --no-deps; fuzz-package formatting and Clippy; cargo test -p jxr-mpsgraph --test metal -- --ignored --test-threads=1 (7 passed); T.834/T.835 conformance Metal 517/517 and CPU 517/517; jxr-diff --backend metal on the 13 large generated images, all identical. On the current main tree (with #12–#18) with this change, the workspace tests (278) and mpsgraph Metal tests (8) also passed, and T.834 Metal and CPU were 517/517 and the 13 large-image diffs identical on the M4 Pro, an M4, and an M5 Max. The owner-only metal-hardware workflow has not been run on this revision.