All pull requests

Pull requests

• Add learned cost models for dot kernels

  #11568·copybara-service[bot] opened Oct 4, 2026Bot·

  ·

• gate unpooling indirection reuse on channels and pixel strides

  #11567·aizu-m opened Oct 4, 2026Contributor·

  ·

• Validate every blockwise scale the qb4w kernels read

  #11566·anagnorisis2peripeteia opened Oct 4, 2026Contributor·

  ·

• Require static-slice num_dims to match the input rank

  #11565·anagnorisis2peripeteia opened Oct 4, 2026Contributor·

  ·

• Avoid trap on non-FP32 filters in NCHW sparsity scan

  #11564·anagnorisis2peripeteia opened Oct 4, 2026Contributor·

  ·

• [WIP] Add per-node control of BMM dequantization fusion and benchmark harness

  #11563·copybara-service[bot] opened Oct 4, 2026Bot·

  ·

• Re-evaluate the in-place skip on every setup of a copy operator

  #11562·anagnorisis2peripeteia opened Oct 4, 2026Contributor·

  ·

• Release the weights cache descriptors when create fails

  #11561·anagnorisis2peripeteia opened Oct 4, 2026Contributor·

  ·

• Re-arm the OOM status before the allocations that follow a microkernel path

  #11560·anagnorisis2peripeteia opened Oct 4, 2026Contributor·

  ·

• Report the real status from create_batch_matrix_multiply_nc_helper

  #11558·anagnorisis2peripeteia opened Oct 3, 2026Contributor·

  ·

• Make xnn_datatype_size_bits/bytes well-defined for xnn_datatype_invalid

  #11557·anagnorisis2peripeteia opened Oct 3, 2026Contributor·

  ·

• Do not fail a weights-cache-locked section in create_fully_connected_nc

  #11556·anagnorisis2peripeteia opened Oct 3, 2026Contributor·

  ·

• Enforce the fingerprint vector bound at run time, not just with an assert

  #11555·anagnorisis2peripeteia opened Oct 3, 2026Contributor·

  ·

• Fix a data race in the fingerprint cache: return a copy, not a pointer

  #11554·anagnorisis2peripeteia opened Oct 3, 2026Contributor·

  ·

• Avoid left-shifting negative values in the binary shift_left operator

  #11553·anagnorisis2peripeteia opened Oct 3, 2026Contributor·

  ·

• Avoid left-shifting negative values in saturating_rounding_shift_left_s32

  #11552·anagnorisis2peripeteia opened Oct 3, 2026Contributor·

  ·

• Fix misaligned 4-byte stores in the qb4 reference packers

  #11551·anagnorisis2peripeteia opened Oct 3, 2026Contributor·

  ·

• Report an allocation failure from create_fingerprint_context

  #11550·anagnorisis2peripeteia opened Oct 3, 2026Contributor·

  ·

• Take ownership of rewritten static values before node creation can fail

  #11549·anagnorisis2peripeteia opened Oct 3, 2026Contributor·

  ·

• Check the NULL from xnn_subgraph_new_internal_value in the subgraph rewrites

  #11548·anagnorisis2peripeteia opened Oct 3, 2026Contributor·

  ·

• Check the status of xnn_allocate_extra_params at every call site

  #11547·anagnorisis2peripeteia opened Oct 3, 2026Contributor·

  ·

• Return out-of-memory when the runtime workspace allocation fails

  #11546·anagnorisis2peripeteia opened Oct 3, 2026Contributor·

  ·
  1

• Fix ambiguous std::isnan call on xnn_float16 in vbinary tester

  #11545·anagnorisis2peripeteia opened Oct 3, 2026Contributor·

  ·

• Fix f32-vrsqrt NaN on zero inputs, including x86 SSE2/AVX2

  #11544·anagnorisis2peripeteia opened Oct 3, 2026Contributor·

  ·

• Fail operator create when packed weights miss a finalized cache

  #11543·anagnorisis2peripeteia opened Oct 3, 2026Contributor·

  ·
  1