v53.1.0 Public Release

Feat

• Expose accumulation mode flag via Conv2dInfo
• Add Bf16 SME matmul and reshape kernels
• Enable BF16 I/O for CpuFullyConnected in the experimental Operator API

Fix

• Default ScaleKernelInfo to no padding
• Remove deprecated clang-tidy config option
• Detection of HW caps and num Cpus for FreeBSD(R) and OpenBSD(R)

Performance

• Enable signed requantized max pooling with differing qinfo
• Enable assembly kernel for QASYMM8 MAX pooling

Documentation (API, build guide, contribution guide, errata, etc.) available here:
https://artificial-intelligence.sites.arm.com/computelibrary/v53.1.0/index.xhtml

v53.0.0 Public Release

Breaking Changes

• We added tensor size checks to prevent very large tensors being fed to our functions. This may result in some previously working very large problem configurations not being supported anymore.

Feat

• Add SVE kernels for TopKV.
• Add experimental support for FreeBSD(R) OS
• Support lvalues in validation functions
• Add NETopKV function.

Fix

• Do not mutate shared _gemm_output_3d in CpuGemmConv2d::run()
• Remove epsilon from the comparisons in TopKV
• GPU only debug build errors in test fixtures
• Bump REUSE version to enable reading of unknown version files
• Add index validation to CpuMaxUnpoolingLayerKernel
• Relay --target in CCFLAGS to the assembler when building with Android(TM) NDK
• Use std::min in NECropKernel in addition to asserts-only macro to ensure valid index in production builds.
• Don't skip Im2Col if convolution has paddings
• Add index validation for batch_index in NECropKernel
• Add index checks to CLMaxUnpoolingLayerKernel

Documentation (API, build guide, contribution guide, errata, etc.) available here:

https://artificial-intelligence.sites.arm.com/computelibrary/v53.0.0/index.xhtml

v52.8.0 Public Minor Release

Feat

• Add SME GEMM and GEMV kernels.
• Add fine-grained SME feature flags.
• Relax supported size configuration checks for all operators
• Detect all available HW features except FEAT_FHM in Windows(R)

Fix

• Misuse of std::accumulate in TensorShape::collapse and total_size
• Invalid kernel heuristics and validation for small N in Fp16 MMUL kernel in CL backend:
• Prevent potential overflow when building with clang-cl
• Stop ignoring txt files in git history
• Potential -Wdeprecated-literal-operator error in half.hpp inclusion

Documentation (API, build guide, contribution guide, errata, etc.) available here:
https://artificial-intelligence.sites.arm.com/computelibrary/v52.8.0/index.xhtml

v52.7.0 Public Minor Release

Feat

• Add Single-ISA support and enable SME in Multi-ISA CMake build.
• Add SME1 FP32 for gemm kernels
• Provide libarm_compute.a in Linux(R) Cpu only multi-isa build via scons

Fix

• Exclude the FP16 MMUL Reshaped RHS kernel for MMUL extension version less than 1.1.0
• Remove private include from public header
• Detection of FEAT_SME2 in Android(TM) and Linux(R)

Refactor

• Flatten nested concat/combine/zip usage in validation/CPP and UNIT.
• Flatten nested concat usage in validation/CL and CPP.
• Flatten nested concat usage in validation/NEON.

Documentation (API, build guide, contribution guide, errata, etc.) available here:
https://artificial-intelligence.sites.arm.com/computelibrary/v52.7.0/index.xhtml

v52.6.0 Public Minor Release

Feat

• Enable F32 output in Quantized CpuGemmConv2d

Fix

• Invalidate certain Cpu operations if tensor sizes are large
• Missing output type validation in CpuGemmDirectConv2d
• Handle padding updates after configure() in CpuActivation

Refactor

• Flatten nested zip usage in validation/NEON.
• Flatten nested combine and zip usage in validation/CL directory.
• Flatten nested combine usage in validation/NEON directory.

Perf

• Do only one iteration of refinement for FP16 inv

Documentation (API, build guide, contribution guide, errata, etc.) available here:
https://artificial-intelligence.sites.arm.com/computelibrary/v52.6.0/index.xhtml

v52.5.0 Public Minor Release

Feat

• Add profiling tracepoints to CPU and GPU platforms
• Add Perfetto profiler as default backend
• Further modernization in CMake build
• Add CMakePresets.json

Fix

• Handle padding updates after configure() in CpuActivation
• Broken URLs in rendered non-released README.md
• Linker errors on macOS when building with CMake

Perf

• Add FP16 GEMM MMUL Reshaped Only Rhs Kernel

Documentation (API, build guide, contribution guide, errata, etc.) available here:
https://artificial-intelligence.sites.arm.com/computelibrary/v52.5.0/index.xhtml

v52.4.0 Public Minor Release

Notice

• The generation of pre-built binaries for macOS and Windows is currently under review and may be temporarily unavailable following this release.

Feat

• Updates to operator CpuGEMMLowp for static quantization, and associated tests.

Fix

• Potential null pointer access in CpuFullyConnected validate method.

Perf

• Remove switch statements in activation kernels.

Documentation (API, build guide, contribution guide, errata, etc.) available here:
https://artificial-intelligence.sites.arm.com/computelibrary/v52.4.0/index.xhtml

v52.3.0 Public Minor Release

Feat

• Support QSYMM8_PER_CHANNEL in NEQuantizationLayer.
• Add stateless wrapper for CpuFullyConnected.

Fix

• Support mixed-type quantized matmul when updating quantization after configure.
• Prevent overread when computing row sums in GEMM.
• Resolve out-of-bounds access in Dimensions::collapse().

Perf

• Remove switch in SVE activation.
• Remove switch in SVE2 activation.

Documentation (API, build guide, contribution guide, errata, etc.) available here:
https://artificial-intelligence.sites.arm.com/computelibrary/v52.3.0/index.xhtml

v52.2.0 Public Minor Release

Feat

• Enable non-transposed BF16 reorders.

Fix

• Reorder test failures on multi-isa builds.
• Over-eager read ahead of operands in a64_hgemm_8x24.

Documentation (API, build guide, contribution guide, errata, etc.) available here:
https://artificial-intelligence.sites.arm.com/computelibrary/v52.2.0/index.xhtml

v52.1.0 Public Minor Release

Feat

• Restrict GEMM stateless execution to fixed-format kernels only
• Add wrapper class to expose cpu::CpuPool2d functionality
• Enable non-transposed F32 reorders

Documentation (API, build guide, contribution guide, errata, etc.) available here:
https://artificial-intelligence.sites.arm.com/computelibrary/v52.1.0/index.xhtml