Changelog

v1.0.3 (May 2026)

API unification + ARM Linux aarch64 platform — single ship gate covering Android arm64-v8a + Linux aarch64.

API change — type-only template parameter slots

61 operators previously exposed enum, bool, or int non-type template parameters as part of their public template signature. v1.0.3 demotes them all to runtime function arguments with sensible defaults.

cpp

// v1.0.1
acl::cvtcolor::RGB2Gray<uint8_t, ColorCvtGrayMode::GRAY_MAX>(src, dst, w, h);

// v1.0.3
acl::cvtcolor::RGB2Gray<uint8_t>(src, dst, w, h,
    /*rgbStride*/ 0, /*grayStride*/ 0,
    /*cR*/ 0.299f, /*cG*/ 0.587f, /*cB*/ 0.114f,
    ColorCvtGrayMode::GRAY_MAX);

v1.0.3 main batch — 47 ops: RGB2Gray family, normalize, histogram, histMatch, blockAverage, sobel3x3, scharr, boxFilter, filter2D, sepFilter2D, rotate (incl. NV/YV12/YUV444 NEON variants), resize (incl. NV NEON), warpAffine, warpPerspective, remap, yuvRemap, plus the full RGB↔YUV family (24 ops bundled into the new acl::YUVConvertParams struct)
v1.0.3-extension — 14 ops: mul, threshold, adaptiveThreshold, matchTemplate, bayer2RGB, bayer2RGBA, connectedComponent_8n_dfs, bilateralFilter NEON, canny NEON, geometric::rotate (blockW/blockH → runtime), neon::resize SHIFT_RIGHT
Behavior bit-identical — algorithms unchanged, kernel template specializations unchanged, identical results for matching parameter values
license.dat fully backward compatible — licenses issued for v1.0.1 / v1.0.2 work directly on v1.0.3
Header set unchanged: acl.h, api.h, err.h, typeDef.h

New: `acl::YUVConvertParams` struct

The 24 RGB↔YUV operators (rgb2YV12, rgb2NV21, rgb2YUV444, yv122RGB, nv212RGB, yuv4442RGB and their _fixed / RGBA siblings) previously required 7-9 non-type template parameters. They now accept a single struct argument with defaulted fields:

cpp

struct YUVConvertParams {
    YUVEncodeStandard yuv_std = STD_BT601;
    bool yuv444_fmt = true;
    bool nv21_fmt = false;
    bool rgb_fmt = true;
    int  bit_depth = 0;        // 0 = auto: sizeof(YUV_T)*8
    int  shift_right = 0;
    int  shift_left = 0;
    bool yuv_full_range = true;
    bool rgb_full_range = true;
};

// All defaults — common BT.601 8-bit full-range pipeline
acl::cvtcolor::rgb2YV12_fixed<uint8_t, uint8_t>(rgb, yv12, w, h);

// Override individual fields
acl::cvtcolor::rgb2YV12_fixed<uint8_t, uint8_t>(rgb, yv12, w, h,
    {.yuv_std = STD_BT709, .yuv_full_range = false});

New platform: ARM Linux aarch64

v1.0.3 ships a second platform alongside Android arm64-v8a — Linux aarch64 static libraries, validated end-to-end via QEMU + license isolation tests:

Platform	Toolchain	Validated on
Android arm64-v8a	NDK 26.3 (Clang 17)	DX-4 / DX-5 / SD8550 / SD8650
Linux aarch64	aarch64-linux-gnu-g++ 9.4	qemu-aarch64-static 4.2

Both platforms share the same acl:: public API — single source of truth
Same licenseId ledger and per-slot key binding across platforms
Both platforms pass the same 5-axis stock_qa gate: zip structure / positive init (rv=0) / negative init(nullptr) (rv=-1001) / pre-init operator call (rv=-1001) / 6-operator demo PASS

Release packaging — 3 paid SKUs × 2 platforms

commercial_release/ ships 6 paid tier zips (measured sizes; Linux aarch64 is ~30% smaller than Android — GCC produces more compact NEON code):

Tier	Android arm64-v8a zip	libacl.a (Android)	Linux aarch64 zip	libacl.a (Linux)
starter	1.51 MB	6.18 MB	0.82 MB	3.94 MB
pro	1.80 MB	7.13 MB	1.13 MB	4.89 MB
business	1.88 MB	7.52 MB	1.22 MB	5.29 MB

Trial ships through the GitHub release channel (see v1.0.3 GitHub release notes), not via the Taobao stock pipeline.

Tier rename effective v1.0.2: core → starter / advanced → pro / full → business. Trial naming unchanged.

Regression highlights

Android (4 device × 3 paid tier × 7 size): 84 paid runs, 0 mismatch vs v1.0.1 release baseline; 6,032 PNG byte-identical across cross-device pairs
Linux aarch64 (3 paid tier × 2 slot × 5 axis): stock_qa 6/6 PASS via qemu-aarch64-static
Trial smoke (4 Android devices + Linux qemu): 6/6 PASS each — init returns -1001 without license, returns 0 with valid license, demo output carries trial watermark
Cross-platform consistency: Android and Linux libacl.a built from one source tree, ABI distinguished only by toolchain

Migration

scripts/migrate_to_v1.0.3.py — automated rewrite of v1.0.1/v1.0.2-style template-arg call sites to v1.0.3 runtime-arg form. YUV family emits manual-review warnings (struct literals cannot be auto-derived)
docs/MIGRATION_v1.0.3.md — per-operator migration recipes and YUV struct field mapping

Internals

Generated api.h excludes _v102kernel internal templates (audit cleanup)
All tiers rebuilt with fresh per-slot keys; per-slot key binding mechanism unchanged

v1.0.2 (May 2026)

Footprint reduction + tier rename.

Tier rename

SKU names aligned with marketing tiers:

v1.0.1 name	v1.0.2 name
core	starter
advanced	pro
full	business
trial	trial (unchanged)

license.dat already issued under the old names continues to work — the licenseId ledger maps both old and new tier strings to the same per-slot key.

Footprint reduction

BorderType demoted from template parameter to runtime argument — one function instantiation now covers all five border modes (BORDER_REPLICATE / BORDER_REFLECT / BORDER_REFLECT_101 / BORDER_WRAP / BORDER_CONSTANT) instead of 5× template specializations
Kernel layer unchanged — wrapper folds the BORDER template dimension into a runtime switch; bit-identical for matching (op, T, DT, BT) combos
Commercial release builds add -fno-exceptions -fno-rtti
ld -r post-processing dedupes weak symbols across translation units
llvm-strip + .eh_frame removal drops unwind metadata not needed for static linkage

Validation

4 device × 4 tier × 7 size = 112 runs all bit-identical vs v1.0.1 release baseline
Cross-device 6 pairs × 7 size = 42 pairs bit-identical
IP audit: 0 public symbols lost vs v1.0.1
DX-4 customer-style link smoke — single-translation-unit #include <acl/api.h> link + run PASS

v1.0.1 (April 2026)

Stability release — 109-operator set hardened through multi-device regression + license delivery pivot to Taobao auto-fulfillment.

License Delivery (Session 68)

Removed: per-app binding (Android package name + certificate SHA-256 verification)
Added: licenseId (ACL-YYYY-NNNNNN) recorded against Taobao buyer ID + order number for lifetime traceability
New purchase flow: Taobao auto-fulfillment — order triggers delivery of pre-generated license.dat + tier-matching SDK zip via Taobao message channel; no sign-up, no signing-key collection, no email round-trip
Anti-redistribution: legal notice shipped with every SDK (README.md / README_CN.md in the delivery zip); each licenseId in the ledger identifies the source of any leaked copy
API surface: acl::init(const char* licensePath) — the public contract only requires a license path; JNIEnv* / jobject context remain as trailing parameters defaulted to nullptr for backward compatibility with v1.0.0 integrations and are no longer consulted by the license gate
Error code -1004 (ACL_ERR_LICENSE_MISMATCH) repurposed: License tier doesn't match library tier (e.g. core license linked against libacl_full.a), instead of package-name / certificate mismatch
Runtime gate: license magic word check is injected into every operator — if acl::init() never ran or failed, any operator returns -1001 immediately

Release Packaging

4 standard tier zips shipped from commercial/package/:
- acl-pack-trial-v1.0.0-android-arm64.zip (~725 KB)
- acl-pack-core-v1.0.0-android-arm64.zip (~1.2 MB)
- acl-pack-advanced-v1.0.0-android-arm64.zip (~1.4 MB)
- acl-pack-full-v1.0.0-android-arm64.zip (~1.5 MB)
Each zip layout: lib/arm64-v8a/libacl.a + include/acl/ + demo/ + README.md (EN) + README_CN.md (CN) + LICENSE

Regression Highlights

4 devices × 7 size tiers = 7,813 rows/device = 31,252 benchmark rows total, all devices bit-identical
Pass rate 7,421 / 7,813 = 95.0% per device (was 7,561 / 7,717 = 98.0% at S67; new NEON sub_tests added to regression matrix pull this down); 154 failures all in documented categories
+89 PASS vs S67 / 0 new FAIL / +96 rows across the 7 size tiers (15 new NEON sub_tests for bilateralFilter / boxFilter / filter2D / sepFilter2D / resize variants)
Commercial v2b regression (all 4 SoCs) — 30,753 rows/device × 4 devices = 123,012 bit-identical tests, 0 PASS flips vs release baseline and 0 flips cross-device (4 SoCs × 4 SKU tiers × 7 image sizes; ship gate)
Overall M-tier speedup: 1.56× over OpenCV 4.13.0 (3,217 timed rows aggregated across 4 devices)

Bug Fixes (S47 → S67)

S60: houghLines NMS (CPP + NEON); opticalFlowLK rewrite
S62: 3 crash fixes — gabor vld1_u8, YUV odd-dimension, extractPixelPerUxV ceil boundary
S64: Geometric YUV takes YUV_STD + range template params (shares cvtcolor infra); BT.601/709/2020 × full/limited coverage 24/24 PASS
S65: cvtcolor roundtripTolerance 3 → 160 + BT.601 decoder range alignment (4-device FAIL 352 → 156)
S66-67: houghCircles CPP + NEON rewrite (gradient-line vote + top-K NMS + NEON); coverage tuned 0.10 → 0.12 (ACL 4 circles vs OCV 5, was 12 over-detect)

Release Hardening (2026-04-27)

Trial zip repackaged — commercial/package/acl-pack-trial-v1.0.0-android-arm64.zip (239 KB). acl.h auto-defines ACL_BUILT_TIER_TRIAL when linking libacl_trial.a, so trial consumers need no manual macro setup. Now ships a pre-signed license.dat at the zip root so acl::init("license.dat") succeeds out-of-the-box (2-year validity from release date)
watermark.h shipped in trial zip (3 ship-path fixes) — trial guard now activates automatically
Dead-code cleanup: removed jni_verify, strcasecmp_portable, license_manager.py, stray test_license_tier.cpp, and stale build.sh echo text
Gate negative test: acl::init(nullptr) returns -1001 (verified on DX-5 with fresh binary)
Plaintext audit across 4 tier .a files: 0 ACE11CDA magic leak, 0 acl::internal leak; only mangled C++ symbols exposed (acceptable per obfuscation design)
push_test_all.sh SD8550 trap diagnosed: a past adb root push left /data/local/tmp/acl_test/bin/ owned by root with write blocked for uid=2000 (shell), silently freezing binaries at the old build. One-shot fix documented; script defensive-purge is a follow-up TODO

Release Chain End-to-End Validation (2026-04-28)

Zip-only smoke 4/4 tier on DX-4: each tier zip is unzipped, compiled against its own bundled libacl.a + acl.h via NDK clang++ (the same path a customer takes), pushed to device, and run. All 4 tiers: 5/5 PASS. Trial correctly returns -1006 for 1921×1080 (resolution cap) and applies watermark; core/advanced/full allow 2048×1536 and leave output clean
Trial demo scoped to 3 operators: new demo_trial.cpp ships only the trial-whitelisted operators (gaussianBlur / threshold / resize) and demonstrates the cap + watermark. Other tiers continue to ship the full demo_acl.cpp
Package tooling hardening: package_sdk.sh now selects tier-appropriate demo (demo_trial.cpp for trial, demo_acl.cpp for core/advanced/full)
Second-device commercial validation (DX-4 × 4 tiers × M size): 4/4 pairs non-timing bit-identical to release DX-4 M baseline (1130 rows, 1110 grep-PASS, 19 grep-FAIL each). Combined with SD8550's 28/28, commercial validated on 2 physical devices × 32 (tier, size) pairs = 32/32 bit-identical; the 4-device cross-device invariant (ISA determinism) carries the remaining 2 devices

Test / Infra Improvements

7-tier sizing: S 640×480, M 1920×1280, L 4096×3072, E1 639×479, E2 1277×717, E3 17×17, E4 1919×1279
Human review complete: 2,550+ rows across S, E1, and M tiers visually verified
Cross-device pairwise diff = 0 on all 6 device pairs

Performance (Session 67 new Top)

Top operators on DX-5 M tier (ACL > 0.5 ms):

Speedup	Operator
27.16×	`nlmeans_h10_p3_s5` NEON
14.96×	`gammaTransform_2.2_1ch` CPP
13.79×	`nlmeans_h5_p2_s3` NEON
13.53×	`bgr2Lab` NEON

v1.1.0 (April 2026) — superseded by v1.0.1

P1 operator expansion + YUV renaming (integrated into v1.0.1 final release).

New Operators (10)

Category	Operator	Tier	NEON
Feature	houghLines	Advanced	Yes
Feature	houghLinesP	Advanced	Yes
Feature	houghCircles	Advanced	Yes
Feature	opticalFlowLK	Advanced	Yes
Feature	bfMatch (L2)	Advanced	Yes
Feature	bfMatchBinary (Hamming)	Advanced	Yes
Feature	bfKnnMatch (L2)	Advanced	Yes
Feature	bfKnnMatchBinary (Hamming)	Advanced	Yes
Contour	minAreaRect	Advanced	No
Contour	fitEllipse	Advanced	No

Changes

Total operators: 99 → 109 (Core 53, Advanced 93 cumulative, Full 109)
YUV pipeline naming: The NV21/NV12 resize+rotate+YUV→RGB pipeline is exposed as YUV_utilities / YUV_utilities_crop / YUV_utilities_float / YUV_utilities_crop_float (Enterprise Only). An earlier experimental naming scheme (nv21ResizeRotate2RGB / yuvResizeRotate*) was reverted; only the YUV_utilities* headers ship.
New test files: test_feature_houghLines, test_feature_houghCircles, test_feature_opticalFlowLK, test_feature_descriptorMatch, test_contour_minAreaRect

v1.0.0 (April 2026)

Initial commercial release.

Highlights

89 operators across 8 categories: filter, color conversion, geometric, arithmetic, analysis, feature detection, transform, math
ARM NEON acceleration for all performance-critical operators
83% of NEON operators faster than OpenCV on ARM64 (140 faster / 7 tie / 21 slower)
Peak speedups: resize area 37.5x, NL-Means denoising 27.8x, Scharr 24.1x, LUT 6.9x (v1.0.0 single-device baseline; see v1.0.1 for multi-device S67 figures)
Zero dependencies — static library, no OpenCV/third-party runtime required
~200KB library size (vs OpenCV ~50MB)

Operators by Category

Category	Count	Key Operators
Filter	18	gaussianBlur, boxFilter, medianFilter, bilateralFilter, canny, sobel, morphology, guidedFilter, nlMeansDenoising
Color Conversion	11	RGB/YUV (float+fixed), Bayer, grayscale, gamma, channel swap
Geometric	7	resize (NN/bilinear/area), rotate, NV21 compound ops, pyramid
Arithmetic	14	threshold, adaptiveThreshold, LUT, normalize, bitwise, addWeighted, inRange
Analysis	16	histogram, integral, matchTemplate (6 methods), CLAHE, moments, minMaxLoc
Feature	8	Harris, ORB, SIFT, SURF, shiTomasiDetect, HOG
Transform	8	warpAffine, warpPerspective, remap, findHomography
Math	5	DFT/IDFT (1D/2D), mulSpectrums

Platform Support

Android ARM64 (arm64-v8a), NDK r26+
ARM Linux (aarch64) — same library, different binding

Tier System

Tier	Operators	Data Types
Trial	2 (gaussianBlur + resize)	uint8_t
Starter (Core)	48	uint8_t
Pro (Advanced)	71	uint8_t, uint16_t, int16_t
Business (Full)	89	uint8_t, uint16_t, int16_t, float

License System

RSA-2048 signed license files (PKCS#1 v1.5 + SHA-256)
Per-application binding (Android package name + certificate SHA-256) — replaced in v1.0.1 by Taobao licenseId traceability
Expiration date enforcement
Tier validation (license tier must match library tier)

Known Limitations

24 test cases with known deviations (canny edge thresholds, Bayer demosaic edge pixels, pyramid upscale boundary, NV21 color space rounding)
matchTemplate NEON: slower than OpenCV for 6 of 12 variants (TM_CCORR_NORMED, TM_CCOEFF_NORMED on large templates)
No iOS/macOS support in v1.0 (planned for v2.0)

Changelog ​

v1.0.3 (May 2026) ​

API change — type-only template parameter slots ​

New: acl::YUVConvertParams struct ​

New platform: ARM Linux aarch64 ​

Release packaging — 3 paid SKUs × 2 platforms ​

Regression highlights ​

Migration ​

Internals ​

v1.0.2 (May 2026) ​

Tier rename ​

Footprint reduction ​

Validation ​

v1.0.1 (April 2026) ​

License Delivery (Session 68) ​

Release Packaging ​

Regression Highlights ​

Bug Fixes (S47 → S67) ​

Release Hardening (2026-04-27) ​

Release Chain End-to-End Validation (2026-04-28) ​

Test / Infra Improvements ​

Performance (Session 67 new Top) ​

v1.1.0 (April 2026) — superseded by v1.0.1 ​

New Operators (10) ​

Changes ​

v1.0.0 (April 2026) ​

Highlights ​

Operators by Category ​

Platform Support ​

Tier System ​

License System ​

Known Limitations ​