46 Pull requests merged by 3 people

• Add qd8-f32-qc4w AVX512 gemm microkernels

  #7935 merged Mar 5, 2025

• Remove tests for deprecated APIs

  #7948 merged Mar 5, 2025

• Explicitly subtract the stack pointer instead of subtracting from it.

  #7943 merged Mar 5, 2025

• Add c8 variants of qd8-f32-qc8w gemm microkernels

  #7825 merged Mar 5, 2025

• Add testing of reshaped shapes

  #7945 merged Mar 5, 2025

• Refactor static-reduce test

  #7944 merged Mar 5, 2025

• Copybara import of the project:

  #7942 merged Mar 4, 2025

• Update KleidiAI to use GitHub

  #7941 merged Mar 4, 2025

• Add sharding to slow test

  #7937 merged Mar 4, 2025

• Always inline functions that return partially initialized SIMD values

  #7936 merged Mar 3, 2025

• microkernel-utils: Increase tile size log level to debug

  #7934 merged Mar 3, 2025

• Clean up "unipass" remenants from dwconv

  #7931 merged Mar 3, 2025

• math_cvt_sat_u32_f64 for wasm fall back on C

  #7926 merged Mar 1, 2025

• Remove sse2 maxpool kernel

  #7910 merged Feb 28, 2025

• Add -fno-function-sections -Os to microkernel benchmark compile options to fix slow builds

  #7924 merged Feb 28, 2025

• Combine avgpool and pavgpool into one microkernel type

  #7920 merged Feb 28, 2025

• Make argmaxpool always multipass

  #7919 merged Feb 28, 2025

• Use a faster, albeit weaker, random number generator for tests on riscv, which are usually run in an emulator.

  #7905 merged Feb 28, 2025

• Reduce duplication in operator tester.

  #7915 merged Feb 28, 2025

• Make output_increment a stride for avgpool microkernels

  #7917 merged Feb 28, 2025

• Dedupe avgpool templates

  #7916 merged Feb 28, 2025

• Remove ubsan workaround, the UB was fixed

  #7908 merged Feb 28, 2025

• Rewrite maxpool kernels

  #7903 merged Feb 28, 2025

• [RVV] add support for qs8-vadd/vaddc and qu8-vadd/vaddc

  #7904 merged Feb 28, 2025

• Copybara import of the project:

  #7914 merged Feb 28, 2025

• Clean up comments. NFC.

  #7913 merged Feb 27, 2025

• Add min, max, signcomplement to s16 simd wrappers

  #7911 merged Feb 27, 2025

• Add simd wrappers for u8, s8

  #7912 merged Feb 27, 2025

• Fix more warnings on windows

  #7909 merged Feb 27, 2025

• wasmsimd preamble fix for to support timeOrigin

  #7901 merged Feb 27, 2025

• Increase shard count for //test:static_slice_test since it times out sometimes on wasm.

  #7906 merged Feb 27, 2025

• Rewrite avgpool microkernels

  #7892 merged Feb 27, 2025

• Only call KleidiAI functions if building with KleidiAI

  #7907 merged Feb 27, 2025

• Fix min/sub mixup in s32 simd wrappers, and add test coverage

  #7902 merged Feb 27, 2025

• Replace xnn_operator_type_to_string to xnn_operator_type_to_string_v2 where appropriate.

  #7898 merged Feb 27, 2025

• Switch on sme2 support by default.

  #7802 merged Feb 27, 2025

• union -> struct for {f16,f32}_minmax_params

  #7900 merged Feb 27, 2025

• Add .section .note.GNU-stack,"",@progbits to the generated x86 assembly.

  #7896 merged Feb 27, 2025

• Avoid UB in unaligned_int32_t

  #7895 merged Feb 26, 2025

• Shard slow static-slice-test on riscv64-rvv.

  #7899 merged Feb 26, 2025

• Add missing tags to additional SIMD wrapper tests.

  #7897 merged Feb 26, 2025

• Expose the operation type for unary and binary ops in the profiling data.

  #7894 merged Feb 26, 2025

• Integreate KleidiAI int4 SME into XNNPack.

  #7582 merged Feb 26, 2025

• Internal change only.

  #7893 merged Feb 26, 2025

• Remove qu8 average pooling

  #7884 merged Feb 26, 2025

• Refactor the gemm_compiler's BaseArchitecture class to store the m and n of the microkernels. This avoids passing the M and N paramters around.

  #7881 merged Feb 26, 2025

9 Pull requests opened by 5 people

• Modify f32-gemm for HVX

  #7918 opened Feb 28, 2025

• Use `xnn_gemm_best_tile_size` and flip the order of `mr`/`nr` in the iGEMM-based ops, analogously to what we did for the GEMM ops.

  #7921 opened Feb 28, 2025

• Update #7642 and #7632

  #7925 opened Feb 28, 2025

• F32-GEMM for HVX use Vsf for clamping instead of Vw

  #7928 opened Mar 1, 2025

• Added s8-rminmax microkernels.

  #7929 opened Mar 1, 2025

• RVV support qd8-f32-qc8w-gemm/igemm and qd8-f32-qc4w-gemm

  #7932 opened Mar 3, 2025

• Add support for bf16 * bf16 -> f32 batch_matrix_multiply

  #7933 opened Mar 3, 2025

• Added u8-rminmax microkernels.

  #7938 opened Mar 3, 2025

• Implement a tile dispatch in FC layer

  #7940 opened Mar 4, 2025

2 Issues opened by 2 people

• tags please

  #7930 opened Mar 1, 2025

• F32-GEMM on Hexagon uses 'Vw' for clamping

  #7927 opened Mar 1, 2025

6 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.

• Support RVV f32-dwconv-unipass

  #7402 commented on Mar 5, 2025 • 5 new comments

• Replace conv-hwc yaml with table header

  #7445 commented on Mar 5, 2025 • 3 new comments

• Add f32 rsum RVV implementation microkernels, tests and config changes for LMUL 2, 4 and 8

  #6533 commented on Feb 28, 2025 • 2 new comments

• Enable AVXVNNI microkernels for mobile platforms

  #7828 commented on Mar 5, 2025 • 2 new comments

• HVX x32 transposec

  #7395 commented on Mar 1, 2025 • 1 new comment

• Add QNX to recognized systems

  #7798 commented on Feb 28, 2025 • 0 new comments