Performance Optimizations

Intel Architecture Processors

• Improved performance on future Intel Xeon processors with Intel AVX 10.2 and Intel AMX instruction sets support.
  This functionality is not dispatched by default and requires opt-in with environment variable ONEDNN_MAX_CPU_ISA=AVX10_2_512_AMX_2.
• Improved performance on future Intel Core processors with Intel AVX 10.2 instruction set support. This functionality is not dispatched by default and requires opt-in with environment variable ONEDNN_MAX_CPU_ISA=AVX10_2_512.
• Improved performance of matmul primitive on processors with Intel AMX support.
• Improved performance of f32 matmul primitive for GEMV cases on on processors with Intel AVX2 instruction set support.
• Improved matmul performance with int4 and int8 compressed weights and per-channel zero-points.
• Improved f32 matmul performance with int4 and int8 compressed weights on processors with Intel AVX2 and Intel AVX512 instruction set support.
• Improved bf16 matmul performance with int4 and int8 compressed weights on processors with Intel AVX512, Intel DL Boost and bfloat16 instruction set support.
• Improved performance of int8 convolution primitive when using zero points.
• Improved performance of int8 matmul and inner product primitives with fp16 destination.
• Improved performance of f32 and bf16 convolution primitive with int8 destination.
• Improved performance of RNN primitive on processors with Intel AVX2 instruction set support when using OpenMP runtime.
• Improved performance of subgraphs containing sequence of multiple binary ops with Graph API.

Intel Graphics Products

• Improved GEMM performance for small batch size on Intel Core Ultra processors (Series 2) (formerly Lunar Lake).
• Improved matmul performance for Qwen2-7B shapes on Intel Arc graphics (formerly DG2) and Intel Arc Graphics for Intel Core Ultra processors (formerly Arrow Lake-H).
• Improved int8 matmul performance with int4 weights and per-tensor zero-points.
• Improved bf16 matmul performance with fp8 weights.
• Graph API optimizations:
  • Improved Scaled Dot Product Attention (SDPA) subgraph performance for inference when relaxed accumulation mode is enabled on Intel Core Ultra processors (formerly Meteor Lake).
  • Improved SDPA and GQA subgraphs performance when using host-side scalars.
  • Improved performance of GQA subgraph for 2nd token scenarios.
  • Improved performance of subgraphs containing sequence of multiple binary ops.
  • Improved performance of Grouped Query Attention (GQA) subgraphs for training forward and backward propagation.

AArch64-based Processors

• Improved performance of reorder primitive
• Improved performance of bf16 convolutions
• Improved performance of convolutions on 128-bit SVE platforms
• Improved performance of eltwise on Arm® Neoverse™ N1

Functionality

Functional API

• Introduced host-side scalar memory objects. This functionality allows passing host-side scalars instead of device memory objects when using oneDNN with OpenCL or SYCL runtimes. Host-side scalars are currently supported in matmul and convolution primitives on Intel GPUs.

• Introduced support for pre-computed reductions in matmul primitive. This functionality is intended to improve performance in case of int8 activations and int8 weights with zero-point.

Graph API

• Introduced host_scalar property for logical tensors. This functionality allows passing host-side scalars instead of device memory objects when using oneDNN with OpenCL or SYCL runtimes. Host-side scalars are currently supported to define attention scale, sequence length, and the negative infinity value in SDPA/GQA subgraphs.
• Introduced accumulation mode attribute support in Matmul op. This attribute allows relaxing fp32 accumulation requirements to achieve performance benefits on some platforms.

Intel Graphics Products

• Introduced support for fp4 weights in matmul primitive.
• Introduced support for grouped quantization with group size 16 in matmul with int8 compressed weights.
• Introduced support group size 16 int8 for decompressed weight with regular weights decompression.

Intel Architecture Processors

• Introduced fp4 weights support for fp32 matmul and convolution for future Intel Xeon processors with Intel AVX10.2 instruction set support.

Usability

• Extended diagnostics available in verbose mode for primitive descriptor creation issues.
• Extended dispatch diagnostics in verbose mode output for primitives implementations on Intel GPUs.

AArch64-based Processors

• Fixed crashes in backward-pass convolutions
• Fixed numerical errors in 4D matmul primitives
• Fixed numerical errors in low-precision convolutions
• Fixed numerical errors in reorders with compensation
• Fixed illegal-instruction crashes on Arm® Neoverse™ N1
• Fixed crashes in binary primitive in Debug builds
• Fixed segmentation fault in eltwise_log post-ops for large kernels

Deprecated Functionality

• BLAS-like API including dnnl::sgemm, dnnl::gemm_u8s8s32, and dnnl::gemm_s8s8s32 functions is deprecated and will be removed in future releases. If you are using this API consider switching to matmul primitive.

Breaking Changes

AArch64-based Processors

• Bumped the minimum required Arm® Compute Library 52.4.0

Thanks to our Contributors

This release contains contributions from the project core team as well as Andrei Hutu @Anndrey24,
Anna Sztukowska @asztukow, Arseniy Obolenskiy @aobolensk, Avanish Tiwari @Tiwari-Avanish, Daniel Kuts @apach301, Daniel Whittaker @danwhittaker-arm, Deeksha Kasture @kasturedeeksha, George Nash @georgen117, Henry Gardiner @henry-gar, Keanu Czirjak @keanucz, Krishna Sai @krishnasai-mcw, Marek Michalowski @michalowski-arm, Sheldon Robinson @sheldonrobinson, @Shreyas-fuj, Viktoriia Gvozdeva @vgvozdeva, Xiang1 Guo, Yejing Lai @Yejing-Lai, Yonghao Gu, Yusuf Butt @UseTheForce007, Zhibo Li @zhili03, @almayne, @co63oc, @focusunsink, @gassan-arm, @jstachowintel, @pmanczak, @puneetmatharu, @raistefintel, @vishwascm, @vyevtyus, @zhangfeiv0, @zhangjian29, and @xiazhuozhao.