- c7c8a48 Merge pull request #11545 from anagnorisis2peripeteia:fix-f16-isnan-overload-ambiguity by XNNPACK Team · 5 hours ago upstream/master
- 96eaa53 Scope AVX-512 target attributes in SIMD unit tests to SIMD headers and test bodies. by Frank Barchard · 2 days ago
- 82c6c83 Use AVX256VNNI x64c4 packw microkernel for qdu8-f32-qc8w AVX512VNNI GEMM config. by Frank Barchard · 2 days ago
- c71f5af Add f32-qc4w-gemm AArch64 NEON `fmagic` (`TBL2 + FSUB`) microkernels (`1x8`, `4x8`, `6x8`, `8x8`) and enable `1x8`, `4x8`, and `6x8` in `gemm-config.c`. by Frank Barchard · 2 days ago
- 2143d99 Name anonymous XNNPACK weights cache mappings on Linux and Android. by Jared Duke · 2 days ago
- 6b90b2f Add ynn_enable_fp64 build option. by Dillon Sharlet · 2 days ago
- 0ecadc4 Update batch_size and channels before early return in unary elementwise reshape. by Dillon Sharlet · 2 days ago
- 3743883 Merge pull request #11553 from anagnorisis2peripeteia:fix-reference-shiftleft-negative by XNNPACK Team · 2 days ago
- b8b6b48 Map missing AMD, Intel, ARM, and Apple microarchitectures in `cpuinfo_to_xnn_uarch`. by Frank Barchard · 2 days ago
- 00b8773 Avoid left-shifting negative values in the binary shift_left operator by Cameron Beeley · 3 days ago
- e599a31 Generate SSE2 x16c8 qs8-qc4w packw microkernels and enable for AVX512SKX GEMM configs. by Frank Barchard · 3 days ago
- 4f70383 Skip HardwareConfigResetTest when FP16 config macros are disabled at build time. by Frank Barchard · 3 days ago
- 207799d Make `ynn_enable_perfetto` opt-in in `ynnpack/subgraph` by Ce Zheng CA · 3 days ago
- 8abb390 Support XNN_FLAG_TRANSIENT_INDIRECTION_BUFFER for deconvolution-2d by Jared Duke · 3 days ago
- 6afb369 Disable XNNPACK specific test under YNNPACK by Dillon Sharlet · 3 days ago
- 568365c Add `transpose_a` option to `dot_bench` by Dillon Sharlet · 3 days ago
- e6e692f Add 8-way 8-bit interleave kernels for ARM by Dillon Sharlet · 3 days ago
- 3444ef2 Merge pull request #11558 from anagnorisis2peripeteia:fix-bmm-create-helper-oom-status by XNNPACK Team · 3 days ago
- f0afaf5 Merge pull request #11686 from aizu-m:unpooling-zero-output-dimension by XNNPACK Team · 3 days ago
- a40d373 Merge pull request #11562 from anagnorisis2peripeteia:copy-skip-bug by XNNPACK Team · 3 days ago
- 41ffd85 Merge pull request #11565 from anagnorisis2peripeteia:fix-static-slice-rank-mismatch by XNNPACK Team · 3 days ago
- ed0db3e Use cpuinfo_has_arm_neon_f8dot4() instead of deprecated cpuinfo_has_arm_f8dot() in YNNPACK. by Frank Barchard · 3 days ago
- be94192 Restore `ynn_value::physical_extent(i)` to delegate to `extent(i)` (returning `1` for broadcast/extent-1 dimensions, partially rolling back cl/995207384), use `extents` directly for `elementwise_bounds` in `dot.cc` and `runtime.cc`, and set `func_input.bounds[d] = slinky::point(dims[d]) - pre_padding` in `define_stencil_copy` so static extent-1 stencil dimensions are not treated as broadcast dimensions. by Gerardo Carranza · 3 days ago
- f6514b1 [YNNPACK] Use `_mm256_dpbusd_avx_epi32` in `x86_avxvnni` dot kernel generators by Dmitri Latushko · 3 days ago
- 70873bd Constant-fold non-aliasing `static_transpose` nodes. by Misha Gutman · 4 days ago
- a7d8c2b Roll cpuinfo to 66ee79c038d70dad9f08705b2c9b3e58f6d8f512 in XNNPACK. by Frank Barchard · 4 days ago
- 6d97a1d skip unpooling reshape when output height or width is zero by Aizal Khan · 4 days ago
- 953ee73 Compile `x86_avx512_fp16_fp16_fp32` dot kernels with `x86_avx512` instead of `x86_avx512fp16` in CMake. by Changming Sun · 4 days ago
- 33dbaa1 Update XNNPACK dependencies by Dillon Sharlet · 4 days ago
- c35adb3 [Tensor API] Preserve hierarchical layer/composite scopes (graph::Scope & ScopeGuard) on Operation DAGs. by Ping Yu · 4 days ago
- d0f55f8 Add fp64 support to `get_tensor_shape` by Dillon Sharlet · 4 days ago
- c800c9d Add casts to avoid errors on some compilers by Dillon Sharlet · 4 days ago
- 4953674 Check `#if XNN_ENABLE_SSE41` and `xnn_arch_x86_sse4_1` before selecting SSE4.1 GEMM microkernels; fall back to scalar for `f32_qc4w`/`f32_qc8w`. by Frank Barchard · 4 days ago
- 0b8bc17 Document ML Drift Tensor API backend and add OSS status test macros by Changming Sun CA · 4 days ago
- 377ec06 Fix `XNN_ENABLE_SSE41` and `XNN_ENABLE_AVX512VBMI` macro typos in `unary-elementwise-config.c` and `x8-lut-config.c`. by Frank Barchard · 5 days ago
- bd5a566 Optimize AArch64 NEON `bf8 -> f16` (`f8_e5m2x16 -> f16x16`) conversion using `ZIP1`/`ZIP2` with zero in YNNPACK. by Frank Barchard · 5 days ago
- 5d12504 Add dot kernel cost model for Oryon V3. by Volodymyr Kysenko · 5 days ago
- 512faee Correctly propagate NaN in min/max operations by Dillon Sharlet · 5 days ago
- c985d80 Fix qc4w packw microkernels dropping the last weight when KC is odd. by Frank Barchard · 5 days ago
- 10dd55b Return NULL from xnn_init_qd8_f16_qc4w_gemm_config when no microkernel is available. by Frank Barchard · 5 days ago
- f699167 Merge pull request #11478 from aizu-m:subgraph-producer-guard by XNNPACK Team · 5 days ago
- 1ca8ba6 Added agents instructions to the open source repo. by Misha Gutman · 6 days ago
- ea06ed3 guard missing producer in packed-lhs and row-sum rewrites by Aizal Khan · 6 days ago
- 071c399 Set arch on the qd8_f16_qc8w AVX512AMX gemm and igemm configs. by Frank Barchard · 6 days ago
- 3a1d6ef Automated Code Change by XNNPACK Team · 6 days ago
- a0c37a0 Merge pull request #11270 from anagnorisis2peripeteia:max-pooling-stride-hardening by XNNPACK Team · 6 days ago
- 0a66ebb Add f32-qc8w GEMM SVE2 kernels: LD1SB Z.S + SCVTF weight conversion via the SVE/NEON bridge. by Frank Barchard · 6 days ago
- 6c451a1 Fix int4 test case of FullyConnectedQuantizedOp weights data. by Quentin Khan · 6 days ago
- e6cc6c2 Document how `TensorInit` fields are used to initialize a `TensorHandle`. by Quentin Khan · 6 days ago
- 46eae5c Don't copy the buffer and consumers of the input of Quantize and Dequantize. by Quentin Khan · 6 days ago
- 0b12c6a Allocate independent raw_buffer descriptors for external values in ynn_runtime. by Dillon Sharlet · 6 days ago
- 520e7d3 Guard against out-of-bounds dimension accesses on output.extents in remove_broadcast. by Dillon Sharlet · 6 days ago
- 49f0759 Merge pull request #11243 from jasperjonkhans:fix/ynnpack-broadcast-partial-reduction by XNNPACK Team · 6 days ago
- fdd49a8 [YNNPACK:SIMD] Improve float32 log accuracy with relative-error minimax [2/2] rational coefficients by Sean Talts · 6 days ago
- b778121 ci: re-trigger CI checks by Cameron Beeley · 7 days ago
- 24d7f10 max-pooling: harden stride and dimension calculations against integer overflow by Cameron Beeley · 7 days ago
- 663dae0 Fix YNNPACK broadcast partial reductions by jasper jonkhans · 8 days ago
- 4a5fe37 Add more block shapes of int8 x int2 dot kernels on x86 by Dillon Sharlet · 9 days ago
- 44dda8a Use same strategy for loads in AVX512 and AVX2 for int8 k1 kernels by Dillon Sharlet · 9 days ago
- c3c545f Disable arm64_neoni8mm int8_int2_int32 dot kernels in YNNPACK in CMake. by Volodymyr Kysenko · 9 days ago
- ddec1f3 Remove `num_threads` argument to `NnpackRunner::CreateRuntime()` by Quentin Khan · 9 days ago
- 231b418 Add AArch64 NEON fmagic (`TBL + FSUB`) and 8x8 `f32-qc8w` GEMM microkernels, and update `gemm-config.c` across ARM64 cores. by Zichuan Wei · 9 days ago
- e9e6662 Validate that begin_id and stride_id match output type and add check for stride tensor extent in ynn_define_iota. by Dillon Sharlet · 9 days ago
- e9a56a9 Add mean and max relative error to unary operator test by Dillon Sharlet · 9 days ago
- 1e2ff02 Update XNNPACk dependencies by Dillon Sharlet · 9 days ago
- 15d78ef Add Comparison Generator to ATS. by Gerardo Carranza · 9 days ago
- 993db3a Merge pull request #11624 from ipranjalks:avgpool-include by XNNPACK Team · 9 days ago
- c356837 Include avgpool.h instead of maxpool.h in RVV avgpool template by Pranjal Srivastava · 9 days ago
- 5d067b8 [Tensor API] Preserve inferred reshape dimensions in TFLite exports. by Changming Sun · 10 days ago chromium/8096 chromium/8097 chromium/8098 chromium/8099 chromium/8100 chromium/8101 chromium/8102 chromium/8103 chromium/8104 chromium/8105 chromium/8106 chromium/8107 chromium/8108 chromium/8109 chromium/8110
- b8d2d9a Merge pull request #11592 from shubhamg13:ohos-cmake-system-name by XNNPACK Team · 10 days ago
- 3755947 Prevent rewrite_reduce_static_transpose from running when reduce has an init operand. by Dillon Sharlet · 10 days ago
- a64e248 Guard against n == 0 in Hexagon HVX partial_load and partial_store. by Dillon Sharlet · 10 days ago
- 4b8dbe0 Explicitly set tensor dimension extents to 0 when logical extent is 0. by Dillon Sharlet · 10 days ago
- 7a4a3ad Validate output tensor rank and extents in outputs_are_compatible during common subexpression elimination. by Dillon Sharlet · 10 days ago
- ebd43ff Optimize AArch64 NEON `qs8`/`qu8 -> f32` (`TBL2 + FSUB`) and `bf16 -> f32` (`ZIP1`/`ZIP2` with zero) conversions in XNNPACK and YNNPACK. by Frank Barchard · 10 days ago
- 0182e24 Validate output_rank and axes pointer in ynn_define_gather. by Dillon Sharlet · 10 days ago
- 8cdb2c5 Validate that all input tensors have matching data types in ynn_define_concatenate and ynn_define_stack. by Dillon Sharlet · 10 days ago
- 0422578 Default padding_id to a scalar zero value in ynn_define_static_pad when not specified. by Dillon Sharlet · 10 days ago
- 587c101 Fix shape validation check in ynn_define_broadcast_like by capturing input_extent before updating output_extent. by Dillon Sharlet · 10 days ago
- 6eccdc5 Fix build and inference on x86 without AVX2; add FMA3/F16C F16 GEMM/IGEMM, F16C F16 DWCONV and AVX (VEX128) madd QD8/F32/QC4W GEMM microkernels, and AVX256SKX packw microkernels. by Frank Barchard · 10 days ago
- d77d9d3 Use YNN_NODE_FLAG_KEEP_DIMS with permutation {-1, -2} to transpose the last two dimensions in batch_matrix_multiply and fully_connected. by Dillon Sharlet · 10 days ago
- e3549f4 Do not retype or reshape external output values in move_broadcast_to_output fusion rewrite. by Dillon Sharlet · 10 days ago
- 16de61d Validate that the axes array in static_transpose contains no duplicate entries. by Dillon Sharlet · 10 days ago
- 349b5bc Validate stencil dimensions, strides, and dilations against zero and integer overflow in ynn_define_stencil_copy. by Dillon Sharlet · 10 days ago
- 52903dd Validate that the reduction operation has exactly one reduction dimension and check the return status of ynn_define_dot in rewrite_sum_to_dot. by Dillon Sharlet · 10 days ago
- 74299b7 Validate that the padding tensor data type matches the input tensor data type in ynn_define_static_pad. by Dillon Sharlet · 10 days ago
- e9853ad Validate num_axes and axes pointer in ynn_define_broadcast, and use negative indexing for new_axes in define_xnn_stencil. by Dillon Sharlet · 10 days ago
- e28d46b Ensure external values produced by operations reallocate data buffer to the inferred rank and clamp dimension iteration to buffer rank. by Dillon Sharlet · 10 days ago
- 64609c3 Validate that pre-bound scale and zero_point output tensors have correct datatypes in ynn_define_dynamic_quantization. by Dillon Sharlet · 10 days ago
- b4272ae Add AArch64 NEON fmagic (`TBL + FSUB`) and 8x8 `f32-qc8w` GEMM microkernels, and update `gemm-config.c` across ARM64 cores. by Frank Barchard · 11 days ago
- 6269103 Add OHOS to the recognized CMAKE_SYSTEM_NAME values by Shubham Gupta · 11 days ago
- d297c79 [YNNPACK] Preserve undefined extents in ynn_value::physical_extent matching ynn_value::physical_extents() by Volodymyr Kysenko · 11 days ago
- d7c395d Report the real status from create_batch_matrix_multiply_nc_helper by Cameron Beeley · 11 days ago
- 58760a2 Fix ambiguous std::isnan call on xnn_float16 in vbinary tester by Cameron Beeley · 11 days ago
- 8387207 Optimize B-tile loads in ARM NEON and x86 AVX2 int8 dot generators by Esun Kim · 11 days ago
- cfafb43 Require static-slice num_dims to match the input rank by Cameron Beeley · 12 days ago
- 81a1a7a Re-evaluate the in-place skip on every setup of a copy operator by Cameron Beeley · 12 days ago
- 087cc73 Add learned cost models for dot kernels by Dillon Sharlet · 12 days ago
- 63bb37f [YNNPACK] Mark provably single-iteration loops serial in compute_workers by Volodymyr Kysenko · 12 days ago
- 798ab80 Merge pull request #10934 from prasanna8585:fix/fc-qc2w-requant-scale-overflow by XNNPACK Team · 12 days ago
