Commit history for google/XNNPACK (76de1380 -> 37e59cdb):
- 70889f15 Vaisakh K V: Added SME1 support for qp8-f32-qc4w gemm
- ab8714dd Vaisakh K V: Merge remote-tracking branch 'google/master' into sme1/qp8-f32-qc4w-gemm
- 8649fdd1 aizu-m: validate reduction axes before nchw remap in static reduce
- 383840f8 Frank Barchard: Update XNNPACK elementwise benchmarks to use consistent N elements.
- cced44b5 Dillon Sharlet: Add fp16 and bf16 implementations of `exp`, `expm1`, `log`, `log1p`, `erf`, `tanh`
- 6062dc08 waris ):  fix-concat-oob-write
- d4312603 Frank Barchard: Fix undeclared identifier XNN_SIMD_NUM_RCP_ITER_F16 in wasm relaxed simd fp16
- 2fef1fcf Alexander Shaposhnikov: Change cache key schema for bazel builds.
- a1b0202e Frank Barchard: Add 2 bit SSE GEMM microkernels
- be7bb974 Dillon Sharlet: Don't rewrite sum(a * b) => dot(a, b) if the dot would be a vector-vector multiply
- 85a6530d Misha Gutman: Added f32_qc8w operator level support for batch matrix multiply.
- 932ea64e Quentin Khan: Make hardware configuration and initialization guards (re)setable.
- 207dab62 Misha Gutman: Enabled f32_qc8w bmm on the subgraph level.
- 4bc0e6ab Quentin Khan: Remove useless test main function.
- 799de5c9 XNNPACK Team: Merge pull request #10373 from sin99xx:master
- 859e1a51 XNNPACK Team: Merge pull request #10369 from aizu-m:reduce-axis-bounds
- 6f2d84ec Quentin Khan: Fix isomorphic matcher.
- 51af7f59 Quentin Khan: Make external tensor order follow topological traversal of the graph.
- 5fa627d5 Frank Barchard: f16 vlog switch from rational-3-3 to rational-1-3
- eb93f24d Quentin Khan: Correctly zero-initialize `xnn_unary_params`.
- 98930bab Dillon Sharlet: Test reduce kernels with pi summation in both ascending and descending order
- cf112cb8 Dillon Sharlet: Minor loop fusion improvements
- 68373b6a Dillon Sharlet: Disable F32QC8W tests when using YNNPACK
- ce3b7c1b Frank Barchard: Fix XNNPACK compilation failure on Windows ARM64.
- 1f4e631e Misha Gutman: Rewrote bmm(f32, dequant(qint8)) -> f32 to bmm(f32, qint8 -> qcint8) -> f32.
- 43c4f7b3 Quentin Khan: Add fine-grain detection of unsupported fp16 ops when falling back to fp32.
- 3ed73f57 Mohammadreza Heydary: Rewrote bmm(f32, dequant(qint8)) -> f32 to bmm(f32, qint8 -> qcint8) -> f32.
- 12f3b6e9 Alexander Shaposhnikov: On Hexagon std::int32_t is defined as long, which is distinct from int. Because slinky::thread_pool uses int in its base class signature, using int32_t in the derived class causes a type mismatch.
- dc0d8821 Alexander Shaposhnikov: Fix wrong types in hexagon_hvx.h
- ce684eb2 Misha Gutman: Rewrote bmm(f32, dequant(qint8)) -> f32 to bmm(f32, qint8 -> qcint8) -> f32.
- 3e5f1c7e Byungchul Kim: Internal changes only
- 5be8df86 Alexander Shaposhnikov: Adjust define_transpose_a declaration.
- e504f4e0 Dillon Sharlet: Add rewrite for add(square(x), y) => multiply_add(x, x, y)
- 32df615f Dillon Sharlet: Move `sum_kn` kernels from `neondot` to `neon`
- f4271c2a Dillon Sharlet: Remove concatenating constructor from `vec` types
- f6486e3e Dillon Sharlet: Explicitly number enumerations in ynnpack.h for ABI stability
- bd205c6b Dillon Sharlet: Refactor precision for unary ops
- 00f16e11 Dillon Sharlet: Clean up target suffix for elementwise ops
- 95a93bc1 Dillon Sharlet: Add `isnan`, `isinf`, `isfinite` for YNNPACK float types
- d1576216 Dillon Sharlet: Refactor precision for unary ops
- f5d61330 XNNPACK Team: Merge pull request #10230 from qualcomm:sme1/qp8-f32-qc4w-gemm
- 8d8fa1fe aizu-m: reject overflowing output dimensions in constant pad reshape
- a9a5a662 aizu-m: Guard indirection buffer size against overflow in igemm reshape
- a9c19098 aizu-m: Validate expand-dims axes in resize_expand_dims_output_tensor
- 337236cb aizu-m: bound rope token count to weights tensor size in reshape
- 463c3e99 aizu-m: validate input rank in softmax and qp8 convert reshape
- a505fe2a Dillon Sharlet: Only define `isinf`, `isnan`, `isfinite` if not on MSVC
- 0e469aba Byungchul Kim: Internal changes only
- 4bd2e73b XNNPACK Team: Merge pull request #10357 from aizu-m:constant-pad-dim-overflow
- 494ddd06 Misha Gutman: Fixed underallocating memory for qint8->qcint8 conversion which only could have been seen in combination with weight_cache.
- 13c32f5e Frank Barchard: Add AVX512SKX F16-F32ACC GEMM/IGEMM microkernels to XNNPACK
- 9d225af7 XNNPACK Team: Merge pull request #10389 from aizu-m:igemm-indirection-overflow
- 843528da XNNPACK Team: Merge pull request #10424 from aizu-m:softmax-convert-input-rank
- a80e3c43 XNNPACK Team: Merge pull request #10404 from aizu-m:expand-dims-axis-validation
- 85115d25 XNNPACK Team: Merge pull request #10412 from aizu-m:rope-tokens-weights-bound
- c6f7e32e Dillon Sharlet: Relax tolerance of batch_matrix_multiply test
- 71b2a887 Dillon Sharlet: Add tracing to tsl::profiler traces in tensorflow
- f42612c6 Frank Barchard: Fix SSE2 gemm being used when it should not be AVX/AVX512
- 463600be Dillon Sharlet: Add avx512 fp16 convert kernels
- 3912146d XNNPACK Team: Add tracing to tsl::profiler traces in tensorflow
- 508ca2d6 Dillon Sharlet: Fix new tests in YNNPACK
- 1f742eb0 Volodymyr Kysenko: Remove packing output buffer scheduling.
- 1e14f367 Dillon Sharlet: Add tracing to tsl::profiler traces in tensorflow
- 87730c12 Dillon Sharlet: Fix bf16 gemm config when there are no kernels
- 11852443 Dillon Sharlet: Add fp8 types and reference convert kernels to YNNPACK
- f0e859a4 Byungchul Kim: XNNPACK runner allows external tensor to grow even when it was added without copy
- e71d1ae4 Byungchul Kim: Fix memory leak
- 199b0e5c Byungchul Kim: Internal changes only
- f774d42b Dillon Sharlet: Add `ynn_define_dynamic_quantization`
- ff27209a Frank Barchard: f16-f32acc-approxgelu for WebAssembly and native
- c6fc8008 foodlook: softmax: reject input/output datatype mismatch
- afa8f9c Gerardo Carranza: Add BatchMatmul Generator to ATS.
- cc3f6ba9 Dillon Sharlet: Add `ynn_define_gather` and remove `ynn_define_lut`
- 15e4a6e6 Dillon Sharlet: Fix attention benchmark to be more realistic and test longer sequences
- 7f2b83e6 Gerardo Carranza: Add FullyConnected Generator to ATS.
- 8084a7f9 Frank Barchard: Update ISA guards for F16-VLOG, F16-VSIN, and F16-VCOS in XNNPACK.
- 0a6f23d2 Gerardo Carranza: Add support for asymmetric quantization and weights format in FullyConnected.
- f0b3a299 Richard Townsend: [gn] Experimental CI for x64 Windows
- 7c92728e aizu-m: size both argmax pooling outputs before returning reallocation
- 4d7d0a38 Dillon Sharlet: Improve precision of compute_qd8_params
- 9c202cae Dillon Sharlet: Add `YNN_NODE_FLAG_UNIQUE_DIMS`
- 79ae5230 Dillon Sharlet: Don't return negative dimension indices from `axis_to_slinky_dim`
- 37099ac2 XNNPACK Team: Merge pull request #10471 from aizu-m:argmax-pooling-index-realloc
- c6e5f040 XNNPACK Team: Merge pull request #10467 from foodlook:fix-softmax-datatype-mismatch
- a51b41a4 Gregory Comer: Add bf16-qd8 convert operator support
- 210d8cee Gregory Comer: Add bf16-qu8 convert operator support
- b92d2986 Dillon Sharlet: Add bf16 and fp16 dequantize_dot kernels
- 78beab26 Frank Barchard: Fix build error in unary when AVX is disabled
- 7a54dd8b Frank Barchard: Support FP16 in Convolution 2D, Depthwise Convolution 2D, and Fully Connected subgraphs.
- 7d47dd87 Dillon Sharlet: Fix FP16-to-FP32 fallback rewrite for convert nodes in XNNPACK.
- c69adf79 aizu-m: free temporary f32 buffers in pf32_f16 conv create error path
- f24b7d89 aizu-m: validate index tensor shape in unpooling reshape
- 25b9c41d XNNPACK Team: Merge pull request #10022 from GregoryComer:bf16-qd8-operator
- 71708b85 Dillon Sharlet: SIMD wrapper implementation improvements
- 7f16ce87 aizu-m: validate input_id before values lookup in global pooling define
- a62374fa Volodymyr Kysenko: Use backward propagated extents to decide when loops should be fused.
- cba17d66 Dillon Sharlet: Change sme2 docker image to a version of qemu that supports fp8
- 2278f2c9 Dillon Sharlet: Attempt to fix docker QEMU build
- 9b5fee0b Frank Barchard: Improve precision of f16-f32acc vapproxgelu microkernels using NR division.
- 19655824 Dillon Sharlet: Add patch to allow enable FPMR register access in our qemu build
- e99ee83 Dillon Sharlet: Apparently ubuntu-latest-16core doesn't exist
- 5f4cf8b5 Frank Barchard: 4-bit weight packw microkernel Scalar, SSE2 and AVX2
- ad262b84 Dillon Sharlet: Add neonfp8 cast support
- 9c2b0e11 Frank Barchard: vtanh f16-f32acc kernels across multiple architectures
- e8b818d9 aizu-m: reject overflowing axis in fuse_dims and split_dim define
- 54143bc4 XNNPACK Team: Add Slice, ResizeBilinear, and DepthwiseConv2D ops for WebGpu TensorAPI backend
- b936b7ac Dillon Sharlet: Fix GitHub Actions cache configuration for XNNPACK.
- f5a3ba01 Richard Townsend: [gn] add standalone modules
- d4ea4e0b Dillon Sharlet: Use input shape instead of output shape to determine the number of channels
- 82601d27 Dillon Sharlet: Add SIMD types using GNU vector extension types
- 0f219961 Dillon Sharlet: Implement dot tests by widening sub-32-bit floats to float
- 4b7c368f Dillon Sharlet: Add ARM fp8 dot kernels
- 702d9390 Volodymyr Kysenko: Allow fusing loops with different required steps by using their least common multiple.
- 964559ae XNNPACK Team: Merge pull request #10489 from aizu-m:unpooling-index-shape-check
- ea34996a XNNPACK Team: Merge pull request #10495 from aizu-m:global-pooling-input-id-bounds
- 48b9ad77 XNNPACK Team: Merge pull request #10487 from aizu-m:pf32-f16-conv-bias-leak
- 987df4b9 Dillon Sharlet: Cleanup/simplification pass of dot subgraph and tests
- d8fe82d3 Frank Barchard: GIO weight packing ukernels for KR=1 x8/x16/x32
- b1e3c597 Richard Townsend: [gn] DEPS update for June, 2026
- a5fe7ac0 Dillon Sharlet: Catch out of bounds indices in gather
- 1eb6eb37 Dillon Sharlet: Add CMake build for YNNPACK
- e25ed4b2 Volodymyr Kysenko: Update slinky dependency to a newer commit.
- a50d9b38 Volodymyr Kysenko: Add XNN_FLAG_NO_BROADCAST to batch matrix multiply in attention benchmark.
- e53f11a4 XNNPACK Team: Merge pull request #10520 from aizu-m:fuse-split-axis-overflow
- 4bfa3bb8 Volodymyr Kysenko: Change the order of the dot loops in schedule_info.
- d0aa6a4 Dillon Sharlet: Fix required dot tiling
- 697543f6 Dillon Sharlet: Fix split-fuse in YNNPACK
- 44b9b41f aizu-m: goto error on convolution_op alloc failure in deconv create
- 3f731006 aizu-m: free partial operator on alloc failure in create paths
- d6e4de13 Minh Vu: Fix benchmark minmax params order
- fb4e3d2d aizu-m: reject input rank mismatch in transpose reshape
- a9c1b094 aizu-m: Guard transpose rank-mismatch test with XNNPACK_USE_YNNPACK
- d49f39ee XNNPACK Team: Merge pull request #10543 from fallintoplace:fix/f16-benchmark-minmax-params
- a27ef624 XNNPACK Team: Merge pull request #10542 from aizu-m:free-partial-op-on-alloc-failure
- b0e7c33f XNNPACK Team: Merge pull request #10545 from aizu-m:transpose-rank-mismatch
- 737a5564 XNNPACK Team: Merge pull request #10541 from aizu-m:deconv-create-oom-leak
- 608aa8b9 Dillon Sharlet: LUT kernel improvements
- f9e04bf2 Dillon Sharlet: Improve `ynn_define_gather` to support gathering multiple dimensions
- 23718f1f Dillon Sharlet: Fix static_transpose when running with --define xnnpack_use_ynnpack=true
- a2b924c2 Frank Barchard: Update generated scalar ukernel for qs8-qc4w-packw x8c8 gemm goi
- 3ce6c680 Frank Barchard: Enable SSSE3 GEMM microkernels for qs8_qc2w and qdu8_f32_qc2w in XNNPACK.
- 7359ebdc Frank Barchard: Add f16-vexp FP32 accumulation kernels for performance on AVX512F/F16C/NEON/Wasm
- 617815ff Alexander Shaposhnikov: Add missing f32x32 operators and reduce kernels on Hexagon.
- 108e8348 Ping Yu: Update Gemma3 example with slicing and layernorm adjustments.
- 944466f5 Dillon Sharlet: Remove some excessive axis validation in split/fuse/broadcast ops
- 08712133 Dillon Sharlet: Change `ynn_define_unary_polynomial` arguments from float to double
- ce360caa Byungchul Kim: Allow weights share among xnnpack runners.
- 7d99d465 Frank Barchard: Add AVX512SKX microkernels for BF16/F16/F32 to QS8/QU8 vcvt.
- 74e718e8 Dillon Sharlet: Don't let arithmetic folding change the type of an operation
- 671aee6b Dillon Sharlet: Add some safety checks
- 0882ef3d Volodymyr Kysenko: Add memory usage tracking to XNNPACK subgraph benchmarks.
- 3a991096 Dillon Sharlet: Fix bug reinterpreting `xnn_status` as `bool`
- bf52497e Dillon Sharlet: Add `YNN_VALUE_FLAG_NO_EXCESS_PRECISION`
- 9a7ac40a Volodymyr Kysenko: Extract source region inference into a helper function.
- 68c99f5c Dillon Sharlet: Add composite op library in YNNPACK
- 75d2c7d4 Frank Barchard: Fix ASan buffer overflow in AVX packw microkernel.
- 890ecdd4 Frank Barchard: Add NEON and AVX2 GEMM packing micro-kernels for QS8 and X8.
- 96ee1373 Dillon Sharlet: Move more ops to composites library from XNNPACK compatibility layer
- 653cbf90 Dillon Sharlet: Reduce max rank of tests for reduce
- 37e59cdb Dillon Sharlet: Add `define_dot_quantization` to the composites library

PiperOrigin-RevId: 929008850