All pull requests

Pull requests

• Add qp8_f32_qc2w_gemm benchmark for SME2 and add k_large unit tests

  #11937·copybara-service[bot] opened Nov 7, 2026Bot·

  ·

• require 4D input in convolution and depthwise convolution reshape

  #11936·aizu-m opened Nov 6, 2026Contributor·

  ·

• Fix x8_packq benchmark argument order and m_idx_start buffer overrun

  #11935·copybara-service[bot] opened Nov 6, 2026Bot·

  ·

• Add Hexagon HMX qd8-f32-qc4w/qc8w GEMM and qd8/qs8 IGEMM microkernels

  #11934·copybara-service[bot] opened Nov 6, 2026Bot·

  ·

• Add x86 ACE qp8_f32_qc8w GEMM microkernels and qp8 LHS packer

  #11933·copybara-service[bot] opened Nov 6, 2026Bot·

  ·

• Add SME/SME2 packed-LHS IGEMM for dynamically quantized convolutions

  #11931·copybara-service[bot] opened Nov 6, 2026Bot·

  ·

• Add AVX2, AVX-512 SKX, Wasm Relaxed SIMD, and HVX SIMD wrapper ops

  #11930·copybara-service[bot] opened Nov 6, 2026Bot·

  ·

• Fix YNNPACK average pooling for quantized and narrow-float types

  #11929·copybara-service[bot] opened Nov 6, 2026Bot·

  ·

• Add qs8/qu8 average pooling and AVX2/AVX512/Wasm/HVX max/avg pooling

  #11928·copybara-service[bot] opened Nov 6, 2026Bot·

  ·

• Add native bf16 binary elementwise ops (NEON, AVX512)

  #11926·JakeStevens opened Nov 5, 2026Contributor·

  ·

• Internal build change

  #11920·copybara-service[bot] opened Nov 5, 2026Bot·

  ·

• Increase XNN_MAX_UARCH_TYPES to 4 on ARM

  #11919·JakeStevens opened Nov 4, 2026Contributor·

  ·

• Cap block_n in define_pack_b to kernel.max_block_n for all data types rather than scaling it up to a 128 KiB cache budget for multi-byte types. Keeping packed B blocks narrow reduces B load strides and lets schedule_dot sub-tile across L1 and L2 caches without evicting packed A panels.

  #11918·copybara-service[bot] opened Nov 4, 2026Bot·

  ·

• Zero-pad and pack unaligned k1 panels of A once per m-block so each packed panel is reused across all n-blocks and reduced across (k3, k2) in a single kernel call, and select kernels using the padded k1 extent. When k1 has an unaligned tail and a row of A is under 384 bytes, pack the entire k1 extent in one pass instead of running a separate tail pass over C.

  #11917·copybara-service[bot] opened Nov 4, 2026Bot·

  ·

• Refactor s16/s32 SIMD tests into single .cc files and fix SIMD bugs

  #11916·copybara-service[bot] opened Nov 4, 2026Bot·

  ·

• docs: fix relative links to the XNNPACK compatibility layer in ynnpack/README.md

  #11910·lxcxjxhx opened Nov 3, 2026·

  ·
  1

• Query ynn_runtime_property_may_block in YnnFusionThunk::ExecuteMayBlock.

  #11905·copybara-service[bot] opened Nov 2, 2026Bot·

  ·

• Add ynn_runtime_property_may_block to ynn_query_runtime so callers with task-graph schedulers can determine whether invoking a synchronous runtime is expensive enough to offload alongside independent work, even when medium single-tile dots stay at single-thread concurrency. Estimate runtime duration from dot kernel cost models (seconds_per_mac) and operand/output streaming bandwidths (task_bytes_per_second, task_output_bytes_per_second), updating dynamic-shape estimates during ynn_reshape_runtime and comparing against min_task_seconds.

  #11904·copybara-service[bot] opened Nov 2, 2026Bot·

  ·

• Add ARM SME2 and SME ACLE GEMM microkernels for pf32-qc8w and pf32-qc4w along with SME x32 LHS packing microkernels, unit tests, and benchmarks.

  #11899·copybara-service[bot] opened Nov 2, 2026Bot·

  ·

• Add ynn_define_ragged_dot, a dot product where each row i of A selects its weight slice B[..., g, :, :] via g = row_group(batch..., i, 0).

  #11890·copybara-service[bot] opened Oct 30, 2026Bot·

  ·

• Implement pure C99 ARM SME and SME2 ACLE intrinsics fallbacks for the

  #11889·copybara-service[bot] opened Oct 30, 2026Bot·

  ·

• skip static reshape rewrite when element count changes

  #11885·aizu-m opened Oct 30, 2026Contributor·

  ·

• Disable clang floating-point contraction in xnn_fma_f16, xnn_fnma_f16, and xnn_fms_f16 in f16-scalar.h so scalar _Float16 helpers do not fuse separate multiply and add operations when FMA is enabled.

  #11883·copybara-service[bot] opened Oct 29, 2026Bot·

  ·

• Fuse fp16/bf16 zero static_constant_pad into [depthwise] convolution 2D

  #11846·JakeStevens opened Oct 29, 2026Contributor·

  ·
  1

• Keep runtime memory_planned unset when xnn_plan_memory fails

  #11845·professor-moody opened Oct 29, 2026·

  ·