mirror of
https://github.com/pytorch/pytorch.git
synced 2025-10-20 21:14:14 +08:00
576ed1e400
This PR is to upgrade submodule oneDNN to v3.7. ## Improvements - Improved performance of convolution and matmul primitives on Intel Xeon processors with Intel AMX instruction set support (formerly Sapphire Rapids and Granite Rapids). - Improved performance of int8 and fp32 forward convolution primitive on processors with Intel AVX2 instruction set support. - Improved performance of fp8 matmul primitives with bf16 and fp16 bias data type on Intel Xeon processors with Intel AMX instruction set support (formerly Sapphire Rapids and Granite Rapids). - Introduced initial optimizations for Intel GPUs based on Xe3 architecture. - Added bfloat16 support for SDPA, implemented fp16 and bf16 gemm kernel in SDPA. - Fixed f16 matmul accuracy, the issue of SDPA cannot dispatched to ukernel, bf16/fp16/fp32 conv performance, INT8 Kernel trigger page fault, deconvolution precision issue on complex128 and fp64 and gemm correctness issue in float16 issues. - Improved bf16 matmul performance with fp32 destination with Arm Compute Library (ACL). - Improved bf16 to fp32 reorder performance. - Improved bf16 reorder performance. - Improved bf16 convolution with ACL. Fixes https://github.com/pytorch/pytorch/issues/136348. ## Validation results on CPU 1. NLP models accuracy/inference/training   2. Torchbench cpu userbenchmark inference & training  3. Inductor quantization  4. Dynamo benchmarks         ## Validation results on XPU Accuracy is same as baseline. Performance is shown below.  ## Validation results on ARM   Pull Request resolved: https://github.com/pytorch/pytorch/pull/147498 Approved by: https://github.com/fadara01, https://github.com/mingfeima, https://github.com/atalman
This folder contains vendored copies of third-party libraries that we use.