Commit history for google/XNNPACK (bccfe733 -> ace56b61): - b3e7d5a1 mohammadmseet-hue: Fix stack buffer overflows in NCHW reduce rewrite and ynnpack shim functions - 1dfa3da5 mohammadmseet-hue: Address review: return error instead of clamping, use YNN_LOG_ERROR - 0cd97f2f Ken Unger: add rvv support for f16-vcmul - 1f8d093a Ken Unger: add rvv support for f16-vcmul - d4474df4 velonica0: rvv-f16-activation - e36eb021 Ken Unger: add fp16 rvv kernels for vsin,vcos,vexp - d3121ea9 velonica0: [RVV] add rvv f32 kernels for velu, vgelu, vapproxgelu - ae231328 velonica0: Alphabetize RVV elementwise entries in cmake/bzl lists - 0b6f61af velonica0: fix cmake bug - 803079cd Gregory Comer: Add AVX512 f32<->bf16 vcvt kernels - 80303ee9 Gregory Comer: Add native AVX512_BF16 f32->bf16 vcvt kernel - b0a078ab Volodymyr Kysenko: Add benchmarks for int4x2/int2x4 to int8_t conversions. - aca142ca Dillon Sharlet: Decide whether to constant fold heuristically - 172b7f4a Marie White: Add arm_neonbf16 binary kernels - b7ebadee MarkLee131: Reject qpint8 in xnn_define_dynamically_quantized_tensor_value - 524c06d2 MarkLee131: Detect size_t overflow in get_tensor_size and reject the tensor - 0d406ba6 Volodymyr Kysenko: Add 2-bit and 4-bit interleave kernels. - 3b773ce1 Quentin Khan: Don't produce an op when `Cast` is casting to the same type as the input type. - 16acef83 Volodymyr Kysenko: Fix typo in the comment. - c8ec3d63 Misha Gutman: Added dynamic_b support for qdu8-f32-qc8w operator. - 1c7554a8 Dillon Sharlet: Remove unnecessary assert - bf188a3d Nicolas Pitre: Add Zephyr RTOS (Generic) platform support - 55d4036f Frank Barchard: Increase tolerance for SUBGRAPH_FP16.fully_connected_qd8_f16_qc8w test to account for numerical deviation. - 8e06513 XNNPACK Team: Merge pull request #10060 from npitre:zephyr-support-pr - 6a9a9c50 Dillon Sharlet: Disable AMX kernels if msan is enabled - a9cb6cb9 XNNPACK Team: Merge pull request #10023 from GregoryComer:bf16-f32-vcvt-avx512 - cde7d935 Marie White: Improve tile sizes for arm_neonbf16 kernels. Tuned with AI agents. - fadd2dbe XNNPACK Team: Merge pull request #9986 from velonica0:rvv-f16-elementwise - 4ca5fb8d Dillon: Merge branch 'master' into f16-unary-trig-rvv - b2985573 Quentin Khan: Add wrappers for storage type of 2/4 bit int and 16 bit floats. - 9c70eb91 Quentin Khan: Add reverse data type to native type mapping. - 2c52c9fe Quentin Khan: Add a conversion function to be able to specialize buffer copy from a sequence. - c817561c Quentin Khan: Move declaration of `NativeStorage` and clarify comment of `StorageImpl`. - 7b375800 MarkLee131: Clarify qpint8 rejection wording - 1437d94 MarkLee131: Use xnn_safe_mul/xnn_safe_add in get_tensor_size - f074438f MarkLee131: Split xnn_safe_mul/xnn_safe_add into separate statements - 64081049 Dillon Sharlet: Resubmit #10069 - 56496fd6 Dillon Sharlet: Add int32 sum kernels - bbc68d90 XNNPACK Team: Merge pull request #9963 from velonica0:rvv-elementwise - 51759bd4 XNNPACK Team: Merge pull request #10102 from MarkLee131:fix/integer-overflow-tensor-size - 562e5274 Dillon Sharlet: Refactor `make_schedule` to allow building just the loop splits, and not a whole `scheduling_info` - 8e4e9d5b Dillon Sharlet: Change reduce to make the identity buffer in slinky, instead of in the subgraph - 64d21ff8 Ken Unger: handle unconfigured f16-vcmul kernel - 834051a2 XNNPACK Team: Merge pull request #10101 from MarkLee131:fix/qpint8-null-deref - b3a5d44f Ken Unger: merge master - 8b3bda45 Ken Unger: update-microkernels - e2da1edb Frank Barchard: Add f16_wasmrelaxedsimd SIMD headers - 5aa5d64e Quentin Khan: Add a parallel lib to `utils:matchers` for internal targets that are only compiled with OSS. - b2f46c0c Quentin Khan: Add a matcher to to check whether two graph are isomorphic. - f81e3eda Volodymyr Kysenko: Support channelwise zero points in YNNPACK quantized dot products. - 4a318ee8 Frank Barchard: Add portable SIMD template for f16-vsqrt - 4780ab70 Frank Barchard: Run generator to create rvv kernels - 3659dcf2 Ken Unger: merge master - f589c63c Jonathan Clohessy: Update CMakeLists.txt to match SME defaults from bazel - 6833e630 Dillon: Merge branch 'master' into f16-unary-trig-rvv - 26c61a7a XNNPACK Team: Merge pull request #9989 from ken-unger:f16-unary-trig-rvv - b0328fc2 Frank Barchard: Fix WAsm typo in XNNPACK by renaming to Wasm - 04b67752 Dillon Sharlet: Refactor tolerance calculations - 8c2df4d5 Dillon Sharlet: Parallelize reductions in YNNPACK - e9de2685 Dillon Sharlet: Add reference kernels for fp64 elementwise ops - a9390e5a Dillon Sharlet: Fix hexagon build - a3da013b Dillon Sharlet: Add benchmark coverage of reference fp64 elementwise ops - 8a3902dd Dillon Sharlet: Add optimized kernels for fp64 elementwise ops - c8c86398 Dillon Sharlet: Add fp64 fma rules to elementwise compiler - 807d9f9c Alexander Shaposhnikov: Introduce XNN_NO_SANITIZE_FUNCTION macro. - 8e406b86 Dillon Sharlet: Loosen tolerances for dequantize_dot test - bb6c6a48 Misha Gutman: Added convert from qint8 to qcint8. - 689c5c60 Misha Gutman: Removed convert qint8 to qcint8 tests from ynnpack test set. - a3664b21 Dillon Sharlet: Avoid capturing kernel in reduce ops - 0fc9e7e7 Volodymyr Kysenko: Disable subgraph_matcher_test when use_ynnpack is enabled. - b5bc455b Dillon Sharlet: Enable adding and removing dimensions via static_transpose - 6dfbf304 Frank Barchard: Optimize xnn_round_f32 for Hexagon HVX. - 52d94589 XNNPACK Team: Merge pull request #9851 from mohammadmseet-hue:fix/nchw-reduce-overflow-and-shim-bounds - 6bd50499 Misha Gutman: Fixed the crash due to unaligned read. - 94ce3bb6 Volodymyr Kysenko: Refactor extent handling in YNNPACK subgraph. - cceae52c Dillon Sharlet: Always constant fold pack_b ops - e0729a7c Dillon Sharlet: Add assert to catch infinite loop case - 50b01640 Frank Barchard: Fix Hexagon HVX build failure 'sf type used as qf32' on Clang 19 - b3daaef9 Dillon Sharlet: Enable sum(squared(x)) => sum_squared(x) for fp64 - 8f17e0c0 Dillon Sharlet: Relax tolerances of dequantize_dot more - a493bbeb Dillon Sharlet: Add missing benchmark - 95103d5b Frank Barchard: Enable f16 vsqrt wasmrelaxedsimd kernel and scalar fallbacks - d830cd16 Volodymyr Kysenko: Rewrite reduce(static_transpose(x)) into reduce(x) - 7b1bde34 Dillon Sharlet: Remove ternary multiply for purely float types - a571a74b Dillon Sharlet: Add tolerance for quantized int8 operations that may round differently - 58698bd6 Dillon Sharlet: Add `exp2_round` simd helper - ece55c6e Dillon Sharlet: Add rewrite for `sum(a*b)` => `dot(a, b)` where appropriate - fc7f8975 Frederic Rechtenstein: Fix alignment-related crash on AVX512 - 58a233a4 XNNPACK Team: Merge pull request #10167 from JonathanC-ARM:jonclo01/sync_bazel_cmake_defaults - 778408a8 Dillon Sharlet: Add exp_fp64 kernels - 16c63a38 Volodymyr Kysenko: Add benchmarks for fully connected with QC4W and QC2W weights. - 62f1d600 Misha Gutman: Added rewrite `bmm(a:f32, dequant(b:qint8):f32) -> f32` into - f6cf463c Volodymyr Kysenko: Disable BatchMatrixMultiplyDequantBmmRewrite test under ynnpack. - 5660b4b0 Dillon Sharlet: Implement `static_expand_dims` using `static_transpose` - 11e206b8 XNNPACK Team: Implement `static_expand_dims` using `static_transpose` - 8e4e78fd Quentin Khan: Don't use `graph::Tensor` in the XNNPack lowering interface. - b12ed13b Quentin Khan: Fix memory outdated planning optimization invalidated by reshapes. - c3d8c276 Misha Gutman: Disabled bmm rewrite by default as gemma4 fails precision. - fb152529 Volodymyr Kysenko: Rename QD8F32QC8W benchmark to QD8F32QC8WFullyConnected for consistency. - d8f5abe9 Dillon Sharlet: Rename svcnt => svcnts - 445e613a Dillon Sharlet: Fix spurious debug messages about sum(a*b) -> dot(a, b) rewrites - 48e1d0f0 Dillon Sharlet: Add test coverage of static and dynamic shapes - be45bb35 Dillon Sharlet: Add more test coverage for reduce operators - d48bc34c Dillon Sharlet: Add support for rewriting `sum(a*b, init_c)` => `dot(a, b, init_c)` - 4908d191 Marie White: Fix get_dot_kernel type bug - 84aa6a95 Dillon Sharlet: Move gemm, conv shapes hardcoded in benchmarks to text files - c5c413de Richard Townsend: [gn] Update DEPS - d877e1a1 Dillon Sharlet: Fix warning "unexpected tokens following preprocessor directive - expected a newline" - dbf04022 Volodymyr Kysenko: Fix handling of sub-byte types in packer. - 5039d217 Dillon Sharlet: Fix unsimplified slice extents - 6c8ac561 Frank Barchard: F16-VTANH for avx512, wasm and scalar - 3245ce20 Frank Barchard: Enable f16 vsin and vcos wasmrelaxedsimd kernel and scalar fallbacks - 091b9be6 XNNPACK Team: Enable f16 vsin and vcos wasmrelaxedsimd kernel and scalar fallbacks - 99e4485d Dillon Sharlet: Add `horizontal_sum` for floating point types - f919d369 Quentin Khan: Don't call optimize in fp16 rewrite tests. - c723a993 Quentin Khan: Prepare static_reduce test for upcoming fp16 to fp32 rewrite. - 0a27dcf1 Frank Barchard: Enable f16 vsin and vcos wasmrelaxedsimd kernel and scalar fallbacks - f43db489 Dillon Sharlet: Fix loss of precision for fp64 constants - 6e50ae9f Dillon Sharlet: Fix reshape -> slice pattern - 74daa88a Dillon Sharlet: Use internal define_static_expand_dims in define_dot - 28ef957f Dillon Sharlet: Disable sum(a*b) => dot(a, b) rewrite if there are no broadcast dimensions on either side - 016914cb Richard Townsend: [gn] Add pthreadpool for the Chromium config - 25d15607 Volodymyr Kysenko: Fix store in the tail of transpose kernels for sub-byte types. - 5d007c4c Volodymyr Kysenko: Make reference int2/int4 convert work with unaligned n. - 713c3b72 Dillon Sharlet: Require reshape strides to be the shape we need too - 2e6e343b Dillon Sharlet: Rewrite reduce kernels to optimize for numerical behavior - 73c5abb5 Marie White: Fix bug in `get_max_concurrency`. - 1dbb15fc Marie White: Fix fully-connected DynamicB tests to work with QP8. - cf96f77e Marie White: Fix fully-connected DynamicB tests to work with QP8. - 7829cd69 Quentin Khan: Move row sum rewrite to after other optimization rewrites. - 2d16035f Dillon Sharlet: Fix bugs with reduce fusion - 0b66c9f1 Dillon Sharlet: Fix slice bugs - 768003bd Marie White: Fix rank pollution in channelwise quantized scales for YNNPACK. - 26f5c9e1 Marie White: Fix logical extent calculation during constant folding for sub-byte types. - bb971d4d Dillon Sharlet: Refactor the implementation of `remove_static_broadcast_from_elementwise` - 860a6421 XNNPACK Team: Fix rank pollution in channelwise quantized scales for YNNPACK. - f3513194 Frank Barchard: Add rules for updating copyright for new files and removing trailing spaces on blank lines - 5dba5dad Dillon Sharlet: Improve static_slice test coverage - f569d17b Ken Unger: merge master - 12f71cd4 XNNPACK Team: Merge pull request #9971 from ken-unger:f16-vcmul-rvv - 2466b8c2 Dillon Sharlet: Update deps to get bug fixes - cc278f5c Dillon Sharlet: Add support for strides to static_slice - 53007d69 Dillon Sharlet: Add YNN_FLAG_NO_EXCESS_PRECISION - fe166973 Dillon Sharlet: Disable static_slice test until slinky bug is fixed - 4fad5b3 Dillon Sharlet: Disable static_slice test until slinky bug is fixed - d72fa85c Dillon Sharlet: Improve log_fp32 kernels - 95ee916a Dillon Sharlet: Use a better unroll factor for log2_fp32_sse2 - 9ab80cd6 Volodymyr Kysenko: Allow adding function own loops even if some of its non-trivial loops has been already fused. - 11fb8859 Dillon Sharlet: Implement round to nearest even for float -> bf16 conversions - 49e266f7 Volodymyr Kysenko: Add optimized convert int2/int4 to int8 kernels. - ace56b61 Dillon Sharlet: Improve `exp` kernel accuracy and correctness Updating cpuinfo. Project cpuinfo not found in 'third_party/tensorflow/workspace2.bzl'. Skipping. Commit history for pytorch/cpuinfo (8a921006 -> bc3c01e2): - 877328f1 Nikita Shulga: Update CMake minimum requirements to 3.18 (#330) - 37b484a5 Byoungchan Lee: Fix Samsung Exynos 2000 series chipset detection (#309) - 0d5985d6 Alexander Grund: Ignore disabled hyperthreads (#291) - 10a95e87 Gregory Comer: Update googletest and googlebench deps (#337) - f01ce870 Anthony Shaw: Add missing ARM-Cortex uarchs (#335) - d3a86a81 davidbienvenu: Fix compile warnings in arm/windows (#323) - f7b233b2 Anthony Shaw: Implement missing ISA feature flags for ARM64/Windows (#333) - 9a05d399 fbarchard: Fix windows arm64 build warnings (treated as error) (#332) - 735b2ac1 fbarchard: Rollback Ignore disabled hyperthreads (#340) - f1f25cbe Byoungchan Lee: Fix Unisoc T618 Chipset Detection (#307) - 952a5198 Jimothy Branson: Allow for longer CPU names (#320) - 1d1dbb4c Bin Bao: Fix missing AMD cache info when running on a VM (#338) - f858c30b Gregory Comer: Improve unisoc vendor detection, re-enable neon dot on unknown chipsets (#342) - 403d652d fbarchard: Fix overread asan error of brand_string (#344) - 161a9ec3 fbarchard: Fix ARM build -Wlogical-op-parentheses for unisoc detect (#347) - ff24ffee fbarchard: add Cortex A78C midr detect (#354) - e8b4def4 Vertexwahn: Bazel support: Sync with current status of BCR and apply fixes for Bazel 9 (#345) - 002c213d fbarchard: Add Oryon V3 uarch detect for pytorch/cpuinfo (#357) - 38b1af9c fbarchard: Fix typo cpofreq (#358) - e3f82e8f fbarchard: Blizzard duplicate return removed (#360) - b3b25967 fbarchard: Apple M4 uarch detect (#362) - 2846bafc fbarchard: Add Apple SME isa and vector length detects (#363) - 315f594 fbarchard: Add Apple M3, M5 and A19 uarch support (#366) - 0fea7f5f Alfredo Tupone: Upgrade C++ standard from CXX14 to CXX17 (#355) - c4b4f4bf fbarchard: Add ARM Cortex A520 little core uarch (#367) - 1ce292f fbarchard: Add ARM Lumex C1 uarch (#370) - e7b796cc Alfredo Tupone: cmake will require at least 3.10 compatibility (#319) - b8ecd2f4 Edward Chen: Add fallback for ARM Windows fp16 detection. (#348) - f9a03241 fbarchard: Detect AMD Family 26 extended models for cpuinfo_uarch_zen6 (#371) - 84818a41 Redwanul Haque Sourave: Fix buffer size for reading processor core id (#352) - 7364b490 Alexander Shaposhnikov: [cpuinfo] Add linux_ppc64le config (#373) - 7607ca50 Rohanjames1997: Report correct L2 cache size on ARM (Neoverse V1/V2) (#372) - bc3c01e2 fbarchard: Add Sapphire Rapids (Golden Cove) microarchitecture detection (#377) Commit history for dsharlet/slinky (1032be67 -> eb004cb3): - 63c773f3 Dillon: Simplify `make_buffer` with new broadcast dimensions to `transpose` (#802) - 66efc5ef Dillon: Fix `can_fuse` for broadcast dimensions (#803) - 6fcfed78 Dillon: Fix more instances of `fold_factor` that should have been changed to `stride` after #802 (#806) - 2af0a01 Dillon: Remove unnecessary branches for the rank of buffers when accessing dims (#807) - 70b443b7 Dillon: Add fast path to `for_each_element` for rank 0 buffers (#805) - dea32175 Dillon: Remove extent 1 dimensions in `optimize_dims` (#797) - 7e02995b Dillon: Fix out of bounds vector access when simplifying nested transpose ops (#808) - 9140d8ac Dillon: Change drop-loops to keep the loop but rewrite the extent (#809) - f3ab7b63 Dillon: Fix aliases that use buffer bounds before they are defined (#810) - 7bc45e1 Dillon: Add support for `slice_buffer`, `slice_dim`, and `transpose` in `alias_copies` (#811) - 0335d87e Dillon: Cast object instead of function pointer (#812) - 27f5d9d9 Dillon: Fix externally defined fold factors (#813) - c08ef409 Dillon: Fix copy aliasing for copies that remove dimensions (#814) - 284794e8 Dillon: Fix bugs uncovered by copying from a rank > 0 buffer to a scalar (#815) - 56f8638a Dillon: Fix crop simplification bug (#816) - 0fbea044 Dillon: Fix simplify of nested transposes (#817) - c01931be Dillon: Fix a straggler usage of `op->dims` => `dims` (#818) - eb004cb3 Dillon: Fix strided copies (#819) PiperOrigin-RevId: 916033850

reduce_sum (complex) on GPU segfaults #357

Description

Metadata

Metadata

Assignees

Labels

Type

Fields

Projects

Milestone

Relationships

Development

Issue actions