30 Pull requests merged by 6 people

• Add f32-vgelu microkernels.

  #6677 merged Jul 4, 2024

• QB4W Neondot F16/F32 GEMM Kernels

  #6664 merged Jul 4, 2024

• Mark generate_build_identifier_py incompatible with wasm platforms.

  #6678 merged Jul 3, 2024

• QB4W AVX GEMM Kernels

  #6621 merged Jul 3, 2024

• Update test and benchmark generation for blockwise kr>2

  #6557 merged Jul 3, 2024

• QB4W SSE2/SSE41 GEMM Kernels

  #6576 merged Jul 3, 2024

• Remove unused variable in binary-elementwise-config

  #6667 merged Jul 2, 2024

• F32 copysign subgraph API

  #6586 merged Jul 2, 2024

• F32 copysign op

  #6585 merged Jul 2, 2024

• Add SIMD copysign microkernels

  #6581 merged Jul 2, 2024

• Clean up #includes in the src/ subdirectory.

  #6661 merged Jul 2, 2024

• Sort order of #include statements in test/*.cc and test/*.h files.

  #6659 merged Jul 2, 2024

• NEON qs8-rsum use 16 byte one's and zero's instead of 15

  #6662 merged Jul 2, 2024

• Change avx256 build for Visual Studio to /arch:AVX512

  #6660 merged Jul 1, 2024

• Fix simd/f32-wasmsimd.h static initializer types.

  #6658 merged Jul 1, 2024

• qs8-rsum ssse3 use load not loadu for params

  #6647 merged Jul 1, 2024

• Detect avx10 and enable avx256skx

  #6648 merged Jul 1, 2024

• qs8-rsum avx512 vpmaddwd to sum 16 bit accumulators to 32 bit

  #6649 merged Jul 1, 2024

• AVX10 QS8/QD8 GEMM/IGEMM

  #6650 merged Jul 1, 2024

• Added standalone rsum HVX

  #6615 merged Jul 1, 2024

• Only copy cvt params if they are set

  #6655 merged Jul 1, 2024

• Fix a few #include <xnnpack/...> to #include "src/xnnpack/..." stragglers.

  #6652 merged Jul 1, 2024

• Fix FullyConnectedTestQD8F32QC4W which was accidentally setting the kernel to zeros.

  #6651 merged Jul 1, 2024

• Add f32 vrelu RVV implementation microkernels, tests and config changes

  #6597 merged Jul 1, 2024

• Fix generate-enum.py to use #include "..." instead of #include <...>.

  #6643 merged Jul 1, 2024

• AVX10 build files support and qs8-rsum

  #6645 merged Jun 29, 2024

• Add scripts for running Hexagon executable on the simulator and on an adb-connected device.

  #6644 merged Jun 28, 2024

• Comment out KleidiAI repository until they have official Bazel support.

  #6642 merged Jun 28, 2024

• Fix a few #include <xnnpack/...> to #include "src/xnnpack/..." stragglers.

  #6640 merged Jun 28, 2024

• Add vectorized f32-vlog microkernels.

  #6614 merged Jun 28, 2024

15 Pull requests opened by 5 people

• HVX `F32-raddstoreexpminusmax` microkernels for Softmax

  #6646 opened Jun 30, 2024

• Add `xnn_cmpeq_f32` to the portable SIMD wrappers.

  #6654 opened Jul 1, 2024

• Rename `xnn_shift(l|r)_f32` to `xnn_shift_(left|right)_f32`, add `xnn_shift_right_signed_f32`.

  #6656 opened Jul 1, 2024

• Fix `f32-vlog-rational-3-3` microkernels to return `NaN` on negative inputs and `-Inf` on `0.0f`.

  #6657 opened Jul 1, 2024

• Add f32 vlrelu RVV implementation microkernels, tests and config changes.

  #6665 opened Jul 2, 2024

• Remove duplicate test generator

  #6669 opened Jul 2, 2024

• Fix duplicate test and benchmark generation

  #6670 opened Jul 2, 2024

• X16-PACKW for avx512 GEMM's that need NR=32 and NR=64

  #6671 opened Jul 2, 2024

• Add AVX10 qd8/qs8 GEMM microkernels and enable 5x8 for avx10 vs 3x8 for avx2

  #6672 opened Jul 2, 2024

• QB4W Fully Connected Operator

  #6673 opened Jul 3, 2024

• QB4W Fully Connected Subgraph

  #6674 opened Jul 3, 2024

• HVX `f32-qs8-vcvt` microkernels

  #6675 opened Jul 3, 2024

• Add `f32-vgelu` kernels using Newton-Raphson iteration instead of division.

  #6680 opened Jul 3, 2024

• QB4W Neoni8mm F16/F32 GEMM Kernels

  #6681 opened Jul 4, 2024

• Use a single `max_abx_x` instead of `min_x` and `max_x`.

  #6682 opened Jul 4, 2024

2 Issues opened by 2 people

• `f32-vgelu` for HVX build fails

  #6683 opened Jul 4, 2024

• Low XNNPACK speedup for ARM Cortex A73

  #6653 opened Jul 1, 2024

4 Unresolved conversations

Sometimes conversations happen on old items that aren’t yet closed. Here is a list of all the Issues and Pull Requests with unresolved conversations.

• QB4W AVX2 GEMM Kernels

  #6618 commented on Jul 4, 2024 • 1 new comment

• QB4W Development

  #6502 commented on Jul 4, 2024 • 0 new comments

• Hook up the new KleidiAI GEMM microkernels to the `fully-connected` operator.

  #6575 commented on Jul 3, 2024 • 0 new comments

• AVX512SKX QB4 Kernels [F16]

  #6636 commented on Jul 1, 2024 • 0 new comments