v0.4.3

🎉 Features

• Add support for multi threaded interpreter (#512) @lgeiger
• Add support for iterators in Interpreter.predict (#511) @lgeiger
• Add Python wrapper for LCE interpreter (#507) @lgeiger

👷‍♂️ Internal Improvements

• Remove redundant code paths: reference BGemm, BGemm functor. (#510) @AdamHillier
• Make Interpreter::get_shapes and get_types more generic (#509) @lgeiger
• Add padding to temporary arrays to ensure we don't read beyond bounds. (#505) @AdamHillier
• Simplify MAKE_ZERO macro (#503) @lgeiger

⬆️ Dependencies

• Bump DoozyX/clang-format-lint-action from v0.9 to v0.10 (#504) @dependabot

v0.4.2

🐛 Bug Fixes

• Fix bug introduced by #497. Make sure OPT kernels run. (#501) @AdamHillier

v0.4.1

🎉 Features

• Add bitpacked output support to the Aarch64 optimised int16 kernel. (#496) @AdamHillier
• Add Int8 output support to the optimised binary convolution kernels. (#492) @AdamHillier

🚀 Performance

• Add bitpacked output support to the Aarch64 optimised int16 kernel. (#496) @AdamHillier
• Update int8 reference bitpack code (#487) @Tombana
• Add an optimised Aarch64 assembly implementation of Int8 bitpacking. (#494) @AdamHillier
• Add Int8 output support to the optimised binary convolution kernels. (#492) @AdamHillier
• Fuse part of the back-transformation into the multiplier and bias. (#490) @AdamHillier

🐛 Bug Fixes

• Fix bug in 32-bit accumulator kernel introduced by 346f72e. (#491) @AdamHillier

📖 Documentation

• Fix typo in README.md (#485) @leonoverweel

👷‍♂️ Internal Improvements

• Simplify path selection and remove unnecessary kernel registrations. (#497) @AdamHillier
• Add inline to round and saturate (#499) @Tombana
• Remove Preprocessed from OutputTransformDetails (#495) @Tombana
• Use Bazel test_filter argument instead of action_env for tests (#488) @AdamHillier
• Clean-up bconv2d by simplifying Init/Prepare/OneTimeSetup. (#486) @AdamHillier
• Simplify BConv2DParams by removing flags. (#484) @AdamHillier

⬆️ Dependencies

• Bump DoozyX/clang-format-lint-action from v0.8 to v0.9 (#489) @dependabot

v0.4.0

⚠️ Breaking Changes ⚠️

• Split BSign into Quantize/Dequantize and separate from BConv (#457) @lgeiger, @AdamHillier @Tombana
• Change custom option enums to match flatbuffer schema (#464) @lgeiger
• Set padding and fused activations as enums (#386) @lgeiger
• Rename BConv attributes to follow TFLite more closely (#373) @lgeiger

🎉 Features

• Support full int8 QAT models in converter (#449) @lgeiger
• Add ARM32 kernel implementation (#432) @honglh
• Allow to set fake default ranges to enable latency test of int8 models (#357) @lgeiger
• Publish prebuilt benchmark binaries with each new release (#398) @lgeiger
• Publish dev docker images for releases (#397) @lgeiger
• Use thresholds for bitpacked output (#387) @Tombana
• Support dilated convolutions in converter (#385) @lgeiger

🚀 Performance

• Improve performance with optimised Ruy matrix packing. (#462) @AdamHillier
• Improve Aarch64 performance by removing NEON pipeline stalls. (#394) @AdamHillier

🐛 Bug Fixes

• Simplify make build (#480) @Tombana
• Fix int8 outputtransform (#422) @Tombana
• Fix trailing quantization op with experimental_bitpacking fusion (#402) @lgeiger
• Fix BMaxPool IR definition (#404) @lgeiger
• Fix sequential model in end2end test: add missing quantisers. (#390) @AdamHillier

📖 Documentation

• Update URLs in readme to adapt to the build docs changes (#481) @lgeiger
• Update benchmark table in Readme (#470) @lgeiger
• Fix broken links (#413) @koenhelwegen

👷‍♂️ Internal Improvements

• Remove zero_point argument from im2col (#478) @lgeiger
• Remove lce_benchmark_all binary (#477) @lgeiger
• Add profiler scopes for bitpacking functions and quantization/maxpool ops. (#471) @AdamHillier
• Remove Windows builds from release workflow (#476) @lgeiger
• Remove input/output dimensions from BConvParams (#475) @lgeiger
• Directly pass packed filer shape to bconv kernel (#474) @lgeiger
• Remove output shape argument from bitpack_tensor (#473) @lgeiger
• Simplify shape handling in kernel prepare method (#472) @lgeiger
• Fix unittest linkopts for macos (#465) @lgeiger
• Rename packbits to bitpack for consistency. (#463) @AdamHillier
• Automatically upload benchmark binaries to new releases (#456) @lgeiger
• Standardise on 32-bit bitpacking. Closes #446. (#461) @AdamHillier
• add extra braces around padding_buffer (#455) @andrewstanfordjason
• Prebuild AArch32 benchmark binary for new releases (#453) @lgeiger
• Add optimised canonical bitpacking for Aarch64. Closes #435. (#443) @AdamHillier
• Add Android AAR build to release workflow (#451) @lgeiger
• Add make_unsigned to unpack_matrix type (#442) @Tombana
• Use a 32-column C++ kernel layout when writing bitpacked output. (#441) @AdamHillier
• Test against Python 3.8 by default on CI (#438) @lgeiger
• Remove unused packbits_arm32 (#437) @lgeiger
• Test converter against TF 2.3 (#436) @lgeiger
• Update the unittest error threshold (#430) @Tombana
• Refactor OutputTransform with template types (#426) @Tombana
• Replace context->ReportError calls with TF_LITE_KERNEL_LOG (#403) @lgeiger
• Improve error message when passing wrong converter arg (#414) @lgeiger
• Use custom larq MLIR dialect for our ops (#384) @lgeiger
• Add misc updates for Micro (#401) @Tombana
• Upgrade TensorFlow to eaacee173897b77cdb6afd22d5e78154177a10f3 (#363) @lgeiger
• Slightly optimize docker image size (#399) @lgeiger
• Add new performance label to release notes (#395) @lgeiger
• Move weight bitpacking into its own converter pass. (#393) @AdamHillier
• Add custom dev docker file (#392) @lgeiger
• Remove RUY_ASM_FLAG_HAS_BIAS (#389) @lgeiger
• Remove filter_format attribute from IR and bconv kernel (#382) @lgeiger
• Add pattern to ensure post_activation_multipliers are always positiv (#375) @lgeiger
• Upgrade github/actions (#381) @lgeiger
• Add testing script for faster model conversions debugging (#380) @lgeiger

⬆️ Dependencies

• Update lint dependencies (#479) @lgeiger
• Bump actions/setup-python from v2.1.1 to v2.1.2 (#460) @dependabot
• Update actions/upload-artifact requirement to v2.1.4 (#419, #425, #440, #448, #454, #459) @dependabot
• Bump GoogleCloudPlatform/github-actions from 0.1.2 to 0.1.3 (#444) @dependabot
• Upgrade tensorflow to 2.3 stable (#439) @lgeiger
• Update lce_register.cc (#434) @lgeiger
• Bump DoozyX/clang-format-lint-action from v0.6 to v0.8 (#427, #433) @dependabot
• Bump actions/setup-python from v2 to v2.1.1 (#429) @dependabot
• Upgrade TensorFlow to 2.3.0rc0 (#415) @lgeiger
• Bump actions/download-artifact from v1 to v2 (#418) @dependabot
• Bump toolmantim/release-drafter from v5.8.0 to v5.11.0 (#417) @dependabot
• Create Dependabot config file (#416) @dependabot-preview
• Upgrade TensorFlow to tensorflow@9a70ab8 (#405) @lgeiger

v0.3.1

🐛 Bug Fixes

• Fix weight bitpacking which could lead to non-deterministic behaviour (#377) @lgeiger

v0.3.0

🎉 Features

• Add experimental support for bitpacked output tensors to the converter (#352) @AdamHillier
• Build wheels for Windows (#355) @lgeiger
• Add support for prebuilt Python 3.8 wheels (#347) @lgeiger
• Upgrade to TF 2.2 (#258) @lgeiger

📖 Documentation

• Update benchmark results (#359) @lgeiger
• Reformat summaries in python.converter.py (#343) @leonoverweel
• Update README with QuickNet Correction (#341) @jamescook106

👷‍♂️ Internal Improvements

• Add 8x4 Arm64 kernel to improve performance (#350) @AdamHillier
• Build wheels using AVX support (#353) @lgeiger
• Use pre-installed bazelisk on CI (#354) @lgeiger
• Upgrade ARM gcc compiler to 9.2.1 (#331) @lgeiger
• Add TF 2.2 to converter test matrix (#348) @lgeiger
• Setup possibility to build nightly wheels on CI (#346) @lgeiger
• ⬆️ lint dependencies (#342) @lgeiger

v0.2.1

🎉 Features

• Add 8-bit quantization support to kernels (#327) @Tombana
• Support reading and writing bitpacked activations in C++ kernels. (#305) @AdamHillier
• Support weight scaling in converter (#326) @lgeiger
• Add ImageNet evaluation tool from TFLite (#313) @AdamHillier
• Add reference implementation of binary convolution. (#314) @arashb

📖 Documentation

• Add new benchmark numbers (#339) @Tombana
• Fix broken links (#335) @lgeiger
• Fix typo (#334) @lgeiger
• Correct top-1 accuracies according to #315 (#319) @lgeiger
• Correct model accuracies (#315) @lgeiger
• Fix docs link in contributing guide (#299) @lgeiger

👷‍♂️ Internal Improvements

• Fix GTEST_FILTER on CI (#338) @Tombana
• Add bitpack order argument to packbits_tensor (#337) @Tombana
• Remove bitpacking dynamic memory allocations (#336) @Tombana
• AArch64 BMLA: Use pairwise addition to accumulate popcount results (#332) @lgeiger
• Simplify ARM compiler build files (#330) @lgeiger
• Consistently use bazel test and --test_output=all on CI (#329) @lgeiger
• Run AAR test outside the custom op container (#322) @lgeiger
• Switch CI testing from mac to linux (#321) @lgeiger
• fix dependency for ref. bconv (#320) @arashb
• set bazel as default builder for android AAR (#318) @arashb
• fix the header guard for packbit utils (#312) @arashb
• Add MLIR status handler to log converter errors (#310) @lgeiger
• Consolidate four vector register loads into one instruction. (#308) @AdamHillier
• add header guard for bconv impl (#304) @arashb
• Fix typos and standardise std::namespace usage. (#303) @AdamHillier
• Support a dynamic bitpacking bitwidth. (#302) @AdamHillier
• Don't use cache if authentication fails (#301) @lgeiger
• Consolidate MLIR and End2End tests (#298) @lgeiger

v0.2.0

⚠️ Breaking Changes ⚠️

• Rename op to LceBconv2d (a3f3838) @Tombana

🎉 Features

• Fused activation (#267) @Tombana
• Add support for one-padding (#252) @Tombana

🐛 Bug Fixes

• Fix typo that caused 64-bit op on ARM32 (#295) @Tombana

📖 Documentation

• update the benchmark results (#294) @arashb

👷‍♂️ Internal Improvements

• Simplify ci builds (#297) @Tombana
• Add preprocessor statement to fix warning (#296) @Tombana
• Enable Bazel GCS remote cache (#287) @lgeiger
• Use fastbuild instead of dbg for tests. (#288) @AdamHillier
• Add +1 padding to the end2end test (#283) @lgeiger
• (ci-skip) Simplify pip package publishing (#282) @lgeiger
• Switch converter tests to be Python only (#280) @lgeiger
• Add end2end test (#269) @Tombana
• Merge some bconv test params (as the # of params is limited). (#277) @AdamHillier
• Replace mul + add with shl + sub (#268) @lgeiger
• remove armv7 android ABI from LCE Lite AAR (#274) @arashb
• Cleanup LCE AAR build (#273) @lgeiger
• Remove un-used variable, fix typo. (#272) @AdamHillier
• Run ARM tests with debug builds. Closes #270. (#271) @AdamHillier
• Reduce normal register loads (#263) @AdamHillier
• Split fused multiply into backtransform and post multiply (#260) @Tombana
• Simplify padding constraint (#264) @lgeiger
• Run CI on PRs (including from forks). (#265) @AdamHillier
• remove legacy reference impl. of LCE bconv op (#261) @arashb
• Add core tests back to CI (#259) @arashb
• Simplify is binary check (#253) @lgeiger
• Cleanup MLIR tests (#255) @lgeiger
• Support running lit tests on macOS without needing to install GNU find (#254) @lgeiger
• Update BConv TableGen definition to match TFLite Op (#256) @lgeiger

v0.1.2

🎉 Features

• build LCE compatible TensorFlow Lite android AAR target (#238) @arashb

🐛 Bug Fixes

• Use vectors instead of tflite temp buffers (#247) @arashb
• Bitpack weights only on the first run (#246) @Tombana

📖 Documentation

• update the pixel phone benchmark results (#249) @arashb
• Move docs to larq/docs (#239) @lgeiger
• Improve converter API docs (#237) @lgeiger
• fix links in second table (#236) @koenhelwegen
• QuickNet .h5 links (#235) @jamescook106

👷‍♂️ Internal Improvements

• Add tests for MLIR passes (#248) @lgeiger
• remove legacy TF ops (#243) @arashb
• remove LCE tflite legacy python API (#241) @arashb

v0.1.1

🐛 Bug Fixes

• Build wheels for macOS 10.13 (#233) @lgeiger
• Fix is-binary check (#230) @lgeiger

📖 Documentation

• simplify the android quickstart guide (#231) @arashb

👷‍♂️ Internal Improvements

• Update actions/checkout@v2 (#232) @lgeiger