File openhtj2k.changes of Package openhtj2k
-------------------------------------------------------------------
Tue Apr 7 09:23:08 UTC 2026 - Michael Vetter <mvetter@suse.com>
- Update to 0.8.0:
New Features:
* Add JPH file format decoding with automatic colorspace detection. The decoder
now accepts .jph input files, parses the JP2 box structure (signature →
ftyp → jp2h → jp2c), extracts the embedded J2K codestream, and reads the
colr box EnumCS field. When EnumCS is 18 (YCbCr) and the output is PPM,
YCbCr→RGB conversion (BT.601) is applied automatically — no -ycbcr flag
needed. The explicit -ycbcr bt601|bt709 flag still overrides the
auto-detected standard.
* Move JPH box parsing (parse_jph_boxes, get_color_space) into the library
(jph.hpp / jph.cpp) and expose it via decoder.hpp; the app-layer
now calls the library API instead of duplicating box-parsing logic.
Performance:
* AVX-512 IDWT: add 9 new functions (idwt_avx512.cpp) covering horizontal
lifting (irrev 9/7 and rev 5/3) and vertical deinterleave / lifting steps;
dispatched when OPENHTJ2K_TRY_AVX2 && __AVX512F__ — falls back to AVX2
on non-AVX-512 x86-64 CPUs with no code change required
* NEON: vextq+carry optimization for horizontal IDWT eliminates per-row
boundary-extension copies; irrev 9/7 lossy: batch −8 %, streaming −34 %;
rev 5/3 lossless: −10 %
* AVX2: widen HT cleanup decode MagSgn path to full 256-bit in
ht_cleanup_decode, processing 8 coefficients per iteration instead of 4
* FDWT: optimize streaming cascade and adv_step_f to reduce per-row
overhead; add SIMD dispatch (AVX2/NEON) for the reversible 5/3 adv_step_f
vertical lifting step
Bug Fixes:
* Fix segfault on truncated JPH/J2K codestreams: guard box and marker
parsers against premature end-of-stream
* Fix streaming decoder vertical upsampling for subsampled components
(e.g. 4:2:0 / 4:2:2): chroma rows are now correctly replicated when the
output stride does not match the subsampled component height
* Fix scalar color transform fallback guard: condition now matches the
dispatch table order, preventing wrong-path selection on non-AVX2 / non-NEON
builds
WASM:
* Vectorize fused YCbCr→RGB color transforms and interleave step in
color_wasm.cpp using WASM SIMD 128-bit intrinsics; add
OPENHTJ2K_ENABLE_WASM_SIMD build flags for the new translation unit
* Add YCbCr→RGB conversion (BT.601 / BT.709) to the Node.js decoder CLI
(open_htj2k_dec.mjs)
* Auto-detect YCbCr colorspace from JPH inputs in index.html; update
dropzone UI with JPH support note and local-decode privacy text; promote
privacy notice to a standalone banner
-------------------------------------------------------------------
Tue Apr 7 09:22:02 UTC 2026 - Michael Vetter <mvetter@suse.com>
- Update to 0.7.1:
* Fix undefined behaviour in ht_magref_decode (scalar and NEON paths):
left-shifting a negative int32_t value is UB in C++; combined into a single
unsigned-arithmetic shift (static_cast<int32_t>((0xFFFFFFFEU | ...) <<
pLSB))
* Work around clang-18 code-generation bug: clang 18 miscompiles certain
inline functions in coding_units.cpp when building for AArch64 NEON,
producing wrong decoder output. Add -fno-inline to coding_units.cpp when
CMAKE_CXX_COMPILER_VERSION < 19 and the compiler is Clang. Clang 19+, GCC,
and MSVC are unaffected. All 445 conformance tests pass on both clang-18 and clang-20.
* Fix streaming encoder (-i file.pgx,...) rejecting subsampled (4:2:2, 4:2:0,
etc.) PGX inputs. PgxStreamReader now accepts component files whose
dimensions are integer sub-multiples of the luma component, computes the
correct XRsiz/YRsiz factors for the SIZ marker, seeks each chroma file to
the correct row (y / YRsiz), and skips redundant push_line_enc calls for
chroma on luma rows that carry no new chroma data.
-------------------------------------------------------------------
Tue Apr 7 09:21:03 UTC 2026 - Michael Vetter <mvetter@suse.com>
- Update to 0.7.0:
JPEG 2000 Part 2: DFS and ATK kernel support:
* Add encoder and decoder support for Part 2 Downsampling Factor Structures
(DFS) markers
* Add decoder support for arbitrary ATK (Arbitrary Transform Kernel) DWT
kernels: irreversible 9/7-based and reversible 5/3-based ATK
* Add scalar FDWT and AVX2 IDWT implementations for ATK kernels
* Add 5 conformance test cases for Part 2 DFS+ATK bitstreams (all pass)
* Fix AVX2 dequantize: use transformation == 1 (not truthy) to distinguish
lossless from ATK kernels (transformation >= 2 is lossy but truthy)
* Fix ATK finalize downshift: (transformation==1)?0:FRACBITS-bitdepth
* Fix bounds-safe color transform dispatch for ATK (transformation >= 2)
* Fix LL0 subband dimensions for DFS codestreams in streaming decode
* Fix ATK encode OOB color dispatch; add line-based DFS/ATK tests
Decoder:
* -reduce now respects DFS markers: the maximum reduce level is clamped to
the number of consecutive bidirectional DWT levels, preventing nonsensical
HONLY/VONLY reduced-resolution outputs
* Fix PPM output for subsampled (4:2:2) codestreams: the streaming path
correctly upsamples chroma with nearest-neighbour interpolation; the batch
write_ppm path uses a scalar fallback for mismatched component dimensions
* Add experimental -ycbcr bt601|bt709 flag: converts YCbCr to RGB during
PPM output using full-range ITU-R BT.601 or BT.709 coefficients
(fixed-point 2^14); handles 4:2:2 nearest-neighbour chroma upsampling;
applies to PPM output only
Encoder:
* Fix streaming (line-based) path for PGX input: a new PgxStreamReader
opens one file per component and reads rows on demand; previously PGX input
always triggered "Failed to open input file for streaming"
WASM:
* Fix 4:2:2 chroma subsampling in invoke_decoder_to_rgba() WASM wrapper
* Add YCbCr→RGB conversion buttons (BT.601 / BT.709) to the web demo at
https://htj2k-demo.pages.dev/
* Modernize button styling in the web demo
-------------------------------------------------------------------
Tue Apr 7 09:20:05 UTC 2026 - Michael Vetter <mvetter@suse.com>
- Update to 0.6.0:
WASM:
* Add invoke_decoder_stream() C export: streaming line-based decode via
callback, eliminating the 96 MB W×H×C int32 output buffer entirely
* Add invoke_decoder_to_rgba() C export: converts decoded samples to a
packed uint8/uint16 big-endian buffer inside WASM, eliminating the JS
per-sample pixel loop
* Add open_htj2k_dec.mjs: Node.js CLI decoder built on the WASM library;
supports J2C / J2K / JPH input, PGM / PPM / PGX output, --reduce,
--num_threads, and --iter options
* Fix WASM SIMD build: add -O3 -flto to SIMD object compilation
* Fix open_htj2k_dec.mjs: compute correct output dimensions when --reduce > 0
* Update index.html to use the new streaming and RGBA export APIs
Performance (native):
* Fused MCT + finalize: color inverse transform and int32 output writeback
combined into a single pass, eliminating an intermediate float buffer
* In-place horizontal IDWT: eliminates ext_buf memcpy per row
* NEON: port lossy dequantize fast path; fix color_neon build failure
* NEON: 2× unroll idwt_level_src_fn interleave loop
* IDWT cascade: eliminate redundant get_dl / is_lp calls in hot loop
* AlignedLargePool: extend macOS support; route file I/O buffers through pool
Stack-allocate Eline / rholine scratch arrays; route large buffers through
* AlignedLargePool to reduce heap fragmentation
* Use aligned AVX2 loads (_mm256_load_ps) where 32-byte alignment is
guaranteed, replacing unaligned variants in hot DWT paths
Portability:
* Lower minimum compiler requirement from C++17 to C++11
* Replace all [[nodiscard]] / [[maybe_unused]] raw attributes with
OPENHTJ2K_NODISCARD / OPENHTJ2K_MAYBE_UNUSED macros that expand
to the real attributes under C++17 and to nothing under C++14/11
* Replace if constexpr with plain if in HT block decoding (all four
variants: generic, AVX2, NEON, WASM)
* ThreadPool: split pre-C++17 enqueue() into separate C++14 and C++11
branches using std::decay_t and typename std::decay<T>::type
respectively
* Remove pre-existing malformed #define[[maybe_unsed]] line in utils.hpp
* All three standard levels (C++11, C++14, C++17) verified
-------------------------------------------------------------------
Tue Apr 7 09:19:26 UTC 2026 - Michael Vetter <mvetter@suse.com>
- Update to 0.5.1:
WASM SIMD:
* Add color transform vectorization: new color_wasm.cpp with all 8
color transform functions (forward/inverse RCT and ICT, integer and
float-domain variants) implemented using WASM SIMD 128-bit intrinsics,
processing 4 elements per iteration
-------------------------------------------------------------------
Tue Apr 7 09:18:14 UTC 2026 - Michael Vetter <mvetter@suse.com>
- Update to 0.5.0:
WASM SIMD:
* Complete separation of WASM-SIMD from NEON: dedicated source files
(fdwt_wasm.cpp, idwt_wasm.cpp, ht_block_encoding_wasm.cpp,
ht_block_decoding_wasm.cpp) guarded solely by
OPENHTJ2K_ENABLE_WASM_SIMD; no NEON flag is used in WASM builds
* Translate HT block encoding and decoding to native WASM-SIMD intrinsics
(v128_t); encoder state classes in ht_block_encoding_wasm.hpp
* Add dedicated WASM fwd_buf implementation in ht_block_decoding.hpp
subprojects/CMakeLists.txt: compile library twice — scalar build
linked to libopen_htj2k.js, SIMD build linked to libopen_htj2k_simd.js
Performance:
* IDWT: eliminate redundant memcpy in idwt_level_src_fn; remove hp_tmp
scratch buffer; reduce streaming ring buffer depth from 12 to 8
* IDWT: vectorize LP/HP interleave with AVX2 (unpacklo/hi + permute2f128)
and NEON; vectorize 5/3 reversible vertical lifting steps with AVX2
* FDWT/IDWT: AVX2 and NEON odd-width interleave/deinterleave fixes
* Decoder finalization: AVX2 and NEON paths for float→int32 conversion
in decode_line_based_stream; NEON ds==0 fast path; 16-element/iter loop
* NEON: 2× unroll reversible vertical lifting steps and FDWT deinterleave
* DWT cascade: specialize 5/3 path; fix O(n²) scan-window growth
* HT block decoding: skip sigma stores for single-pass codeblocks;
skip redundant memset for single-pass codeblocks
* Allocation: eliminate per-codeblock allocation storm in
j2k_precinct_subband; replace per-strip and per-resolution allocation;
hoist tree_path vector out of per-codeblock loop; replace per-block
all_segments std::vector with stack array in htj2k_decode()
* ThreadPool: batch-push tasks, reserve marker list, combine ring buffer
and predecoded pool; pre-fault ring/prefetch buffers; reduce
task-queue thundering herd
* LB encoder: overlap DWT computation with HT block encoding
* Eliminate bulk ring_buf/prefetch_buf zeroing; zero only empty regions
* Replace scratch array memset with per-row 8-byte guard zeros
Fixes:
* Fix O(n²) scan window growth in DWT cascade 5/3 specialization
* Add include for std::swap in fdwt.cpp and idwt.cpp
-------------------------------------------------------------------
Tue Mar 31 07:14:43 UTC 2026 - Michael Vetter <mvetter@suse.com>
- Update to 0.4.1:
* Introduce a temporary workaround for the line-based validation test in
Windows ARM64.
-------------------------------------------------------------------
Tue Mar 31 07:13:23 UTC 2026 - Michael Vetter <mvetter@suse.com>
- Update to 0.4.0:
New APIs:
* Add invoke_line_based() decoder API: row-by-row output via callback
using per-subband ring buffers; avoids full-image intermediate allocation
* Add invoke_line_based_stream() decoder API: multi-tile scatter path
assembling full-width output rows from per-tile strip buffers
* Add encode_line_based_stream() encoder API: push-row streaming encoder
driven by a user-supplied source callback
* Add lb_compare utility to validate line-based output against invoke()
Architecture:
* Change sprec_t from int32_t to float: DWT pipeline now operates
natively in float32, eliminating integer↔float round-trip conversions
* fdwt_2d_state / idwt_2d_state: new stateful streaming DWT classes
that process one row at a time for the line-based encode/decode paths
* Line-based mode is now the default for encoder and decoder apps;
the previous full-image (batch) path is selected with the -batch flag
Performance:
* AVX2 horizontal DWT: 2× loop unrolling on all lifting step loops
* NEON: full float32 horizontal and vertical DWT kernels
* ICT/RCT color transforms: AVX2 and NEON variants with runtime dispatch
* LB decoder: pipeline MCT+clamp of row Y with pool pull of row Y+1,
achieving 18–23 % throughput improvement at T≥4 threads
* Encoder: per-thread cblk_data_pool bump allocator replaces per-block
malloc/free in the HTCOT hot path; PSE scratch buffer hoisted out
of per-tile DWT calls
Build / Tests:
* Decoder and encoder apps reject unknown command-line flag names
* Introduce OPENHTJ2K_EXPORT macro, removing duplicated __declspec
guards from encoder.hpp
* Add lb_stream_validation.cmake: conformance tests for
invoke_line_based_stream() across all HT profile-0/1 and HF bitstreams
* Remove redundant lb_* cmake test files (duplicate of non-lb equivalents);
* CTest count reduced from 480 to 277
* WASM: dark UI redesign, SIMD badge, decode-time display, GitHub link
-------------------------------------------------------------------
Mon Jan 19 08:26:51 UTC 2026 - Michael Vetter <mvetter@suse.com>
- Update to 0.3.1:
* Add Windows ARM + MSVC support to CI workflow
-------------------------------------------------------------------
Mon Jul 28 09:17:39 UTC 2025 - Michael Vetter <mvetter@suse.com>
- Add openhtj2k-0.3.0-tiff.patch:
make pkg-config based requires for the devel package actually
match our libtiff package
-------------------------------------------------------------------
Mon Jul 28 05:18:12 UTC 2025 - Michael Vetter <mvetter@suse.com>
- Update to 0.3.0:
* SOVERSION is introduced for versioning of shared library
- Drop openhtj2k-0.2.9-soversion.patch
-------------------------------------------------------------------
Thu Jul 24 07:55:16 UTC 2025 - Michael Vetter <mvetter@suse.com>
- Add openhtj2k-0.2.9-soversion.patch:
Add so versioning
-------------------------------------------------------------------
Mon Jul 7 05:46:24 UTC 2025 - Michael Vetter <mvetter@suse.com>
- Update to 0.2.9:
* Fix bug for the wrong stride of dwt coefficients.
* Add HEAP32 into EXPORTED_RUNTIME_METHODS for was-build.
-------------------------------------------------------------------
Wed Nov 27 08:07:09 UTC 2024 - Michael Vetter <mvetter@suse.com>
- Update to 0.2.8:
* Fix incorrect packet parsing for RPCL, PCRL, CPRL
* Introduce stride access into DWT
* Change cmake configuration for MinGW environments
-------------------------------------------------------------------
Fri Jun 14 08:49:14 UTC 2024 - Michael Vetter <mvetter@suse.com>
- Update to 0.2.7:
* Refactor non-SIMD HT cleanup decoding
* Fix CMakeLists.txt for windows-latest runner (GitHub actions)
* Fix PGX parsing
-------------------------------------------------------------------
Thu Jun 13 06:43:50 UTC 2024 - Michael Vetter <mvetter@suse.com>
- Update to 0.2.6:
* Fix unnecessary assignment of pass_length in packet header parsing
* Remove CR (=0xd) from the delimiter in imgcmp
-------------------------------------------------------------------
Fri Jan 19 09:26:56 UTC 2024 - Michael Vetter <mvetter@suse.com>
- Update to 0.2.5:
* Fix memory leak in decoder with '-reduce' parameter greater
than actual DWT levels
* Fix buffer overrun with image width which is not multiple
of vector length in IDWT and block-decoding for ARM NEON
* Improve UI for WASM demo
* Enable WASM SIMD (using NEON)
* Fix wrong line break in encoder usage (#162)
-------------------------------------------------------------------
Fri Jan 19 09:26:37 UTC 2024 - Michael Vetter <mvetter@suse.com>
- Update to 0.2.4:
* small fix for wasm wrapper
-------------------------------------------------------------------
Fri Jan 19 09:26:28 UTC 2024 - Michael Vetter <mvetter@suse.com>
- Update to 0.2.3:
* Experimental support of emscripten
* Fix compilation error on aarch64 with gcc
* Small editorial changes
-------------------------------------------------------------------
Mon Nov 27 12:43:54 UTC 2023 - Michael Vetter <mvetter@suse.com>
- Update to 0.2.2:
* Fix compilation errors in aarch64 and gcc 9 or earlier
-------------------------------------------------------------------
Mon Nov 13 08:01:25 UTC 2023 - Michael Vetter <mvetter@suse.com>
- Update to 0.2.1:
* Add installation part to CMakeLists.txt (#154)
* Allow space between comma separated input file names
-------------------------------------------------------------------
Thu Nov 2 09:58:23 UTC 2023 - Michael Vetter <mvetter@suse.com>
- Initial package of OpenHTJ2K 0.2 for openSUSE
