All pull requests

Pull requests

• [XNNPACK] Add F16 raddstoreexpminusmax SIMD microkernels

  #11796·copybara-service[bot] opened Oct 24, 2026Bot·

  ·

• Split the vld1q_s8 + vzip1q_u64/vzip2q_u64 4x16c8 NEON I8MM GEMM/IGEMM microkernels into a separate _zip variant (VARIANT=ZIP) and update the existing c8-neoni8mm microkernels to use vld2q_dup_u64 for lane 0 instead of zero-initializing a loop-carried uint64x2x2_t register pair for vld2q_lane_u64(..., 0).

  #11794·copybara-service[bot] opened Oct 23, 2026Bot·

  ·

• packq: harden packed-LHS size helpers against integer overflow; reject SIZE_MAX at call sites

  #11792·Subhan1794 opened Oct 23, 2026·

  ·

• subgraph: harden qs8->qc8 convert channelwise-quantization size arithmetic against integer overflow

  #11791·Subhan1794 opened Oct 23, 2026·

  ·

• deconvolution-nhwc: harden packed-weights size arithmetic against integer overflow

  #11790·Subhan1794 opened Oct 23, 2026·

  ·

• Improve performance on large square shapes.

  #11788·copybara-service[bot] opened Oct 23, 2026Bot·

  ·

• Copybara import of the project:

  #11787·copybara-service[bot] opened Oct 23, 2026Bot·

  ·

• reject input2 rank below 2 in xnn_define_batch_matrix_multiply

  #11785·aizu-m opened Oct 22, 2026Contributor·

  ·
  2

• Support std::optional<T> in ScopeGuard::Bind.

  #11784·copybara-service[bot] opened Oct 22, 2026Bot·

  ·

• Dequantize 8-bit integer constants on the GPU in ML Drift.

  #11782·copybara-service[bot] opened Oct 22, 2026Bot·

  ·

• Lower ML Drift quantizations to real float to int conversions.

  #11781·copybara-service[bot] opened Oct 22, 2026Bot·

  ·

• Use kr = 4 dot-product GEMMs for qs8/qc8 convolutions with <= 4 input channels

  #11780·gadgilrajeev opened Oct 22, 2026·

  ·

• Speed up 4x16c8 I8MM qs8/qd8 GEMMs with vld1q+vzip activation loads

  #11779·gadgilrajeev opened Oct 22, 2026·

  ·
  7

• RISC-V: block f32 dwconv unipass kernels by two output pixels

  #11778·liyifei-li opened Oct 22, 2026·

  ·
  2

• Support transposed inputs in pack_b and absorb transposes.

  #11777·copybara-service[bot] opened Oct 22, 2026Bot·

  ·

• Add ARM SME and SME2 ACLE x8-packq (f32qp8) LHS dynamic quantization and packing microkernels (xnn_x8_packq_f32qp8_ukernel__sme and xnn_x8_packq_f32qp8_ukernel__sme2) with svprfb software prefetching, along with SME/SME2 C microkernel build targets (-march=armv9.2-a+sme+nosve and -march=armv9.2-a+sme2+nosve) and weak SME ABI runtime fallback stubs (arm-sme-abi-stubs.c).

  #11774·copybara-service[bot] opened Oct 22, 2026Bot·

  ·

• Migrated Gemma 3 onto TransformerBlock

  #11772·copybara-service[bot] opened Oct 21, 2026Bot·

  ·

• Added the first iteration of parametric transformer block.

  #11771·copybara-service[bot] opened Oct 21, 2026Bot·

  ·

• Lower ML Drift dequantizations of 8-bit integers to real int to float conversions.

  #11770·copybara-service[bot] opened Oct 21, 2026Bot·

  ·

• Preserve broadcast strides when aliasing reshape buffers in YNNPACK.

  #11764·copybara-service[bot] opened Oct 21, 2026Bot·

  ·

• [HVX] Replace unnecessary vsplat + vector ops with immediate, scalar R, and unary instructions.

  #11762·copybara-service[bot] opened Oct 21, 2026Bot·

  ·

• subgraph: reject out-of-range split/concat axes in reshape

  #11728·rootkiller6788 opened Oct 19, 2026·

  ·

• Remove redundant software NaN checks in XNNPACK bfloat16 conversions.

  #11725·copybara-service[bot] opened Oct 19, 2026Bot·

  ·

• Add Qualcomm Hexagon HMX and HVX c2 f16-gemm and x16-packw microkernels

  #11722·copybara-service[bot] opened Oct 19, 2026Bot·

  ·

• Use square dot tiles when B is packed per split of n and A doesn't fit in cache.

  #11713·copybara-service[bot] opened Oct 16, 2026Bot·

  ·