All pull requests

Pull requests

• Add bitwise SIMD and scalar fallbacks for f16 vabs (& 0x7FFF) across scalar, NEON, AVX2, AVX-512 SKX, WASM SIMD, RVV, and HVX so xnn_init_f16_abs_config() always succeeds without requiring hardware FP16 instructions.

  #11492·copybara-service[bot] opened Oct 2, 2026Bot·

  ·

• Added SME1 GEMM/GEMV support for qp8-f32-qb4w

  #11491·debsaha1998arch opened Oct 2, 2026·

  ·

• F32_QC8W weight packing for NEON and AVX2

  #11490·copybara-service[bot] opened Oct 2, 2026Bot·

  ·

• Guard unchecked overflow in convolution strides and bilinear indirection offsets

  #11485·anagnorisis2peripeteia opened Oct 1, 2026Contributor·

  ·

• Harden operator creation and reshape against null dereference and overflow

  #11484·anagnorisis2peripeteia opened Oct 1, 2026Contributor·

  ·

• Harden subgraph operator definition and reshape against overflow and crashes

  #11483·anagnorisis2peripeteia opened Oct 1, 2026Contributor·

  ·

• Separate required blocking alignment from preferred loop step in the YNNPACK subgraph scheduler.

  #11482·copybara-service[bot] opened Oct 1, 2026Bot·

  ·

• Harden runtime, memory planner, and subgraph against overflow and leaks

  #11481·anagnorisis2peripeteia opened Oct 1, 2026Contributor·

  ·

• Harden KleidiAI packing against integer overflow and pointer overread

  #11480·anagnorisis2peripeteia opened Oct 1, 2026Contributor·

  ·

• guard missing producer in packed-lhs and row-sum rewrites

  #11478·aizu-m opened Oct 1, 2026Contributor·

  ·

• [RISC-V] Harden vector configuration, guard vlenb, and add fallbacks

  #11476·anagnorisis2peripeteia opened Oct 1, 2026Contributor·

  ·

• configs: harden GEMM mr/nr bounds and CPU uarch logging

  #11475·anagnorisis2peripeteia opened Oct 1, 2026Contributor·

  ·

• reference: harden integer division, modulus, pow, and bitshift operations

  #11474·anagnorisis2peripeteia opened Oct 1, 2026Contributor·

  ·

• operators: harden reshape validation for pooling and unpooling

  #11473·anagnorisis2peripeteia opened Oct 1, 2026Contributor·

  ·

• subgraph: harden operator definition validation in deconv, rope, pool, and split

  #11472·anagnorisis2peripeteia opened Oct 1, 2026Contributor·

  ·

• Change softmax beta support test.

  #11471·copybara-service[bot] opened Oct 1, 2026Bot·

  ·

• Allow NnpackRunner::SetNumThreads to fail.

  #11470·copybara-service[bot] opened Oct 1, 2026Bot·

  ·

• subgraph: guard SIZE_MAX from get_size in reduce/concatenate/copy reshape

  #11468·anagnorisis2peripeteia opened Oct 1, 2026Contributor·

  ·

• Fix split_dims and convert reshape validation and arithmetic overflows

  #11467·anagnorisis2peripeteia opened Oct 1, 2026Contributor·

  ·

• Fix RISC-V vector config gating and qu8 dwconv initialization typo

  #11466·anagnorisis2peripeteia opened Oct 1, 2026Contributor·

  ·

• [subgraph] Fix heap double-free in cvt_to_fp32 fallback on early abort

  #11465·anagnorisis2peripeteia opened Oct 1, 2026Contributor·

  ·

• Guard RISC-V vector configurations with runtime arch flags check

  #11464·anagnorisis2peripeteia opened Oct 1, 2026Contributor·

  ·

• Fix integer overflow and undefined behavior in reference kernels

  #11463·anagnorisis2peripeteia opened Oct 1, 2026Contributor·

  ·

• [subgraph, operators] Harden reduce, transpose, depth-to-space, and space-to-depth

  #11462·anagnorisis2peripeteia opened Oct 1, 2026Contributor·

  ·

• [operators, subgraph, packing] Harden spatial operator kernel bounds, zero-batch handling, and qu8 packing stride

  #11461·anagnorisis2peripeteia opened Oct 1, 2026Contributor·

  ·