v3.9

Performance Optimizations

Intel Architecture Processors

• Introduced initial support for 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.
• Introduced initial support for 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 initialization time for convolution primitive when a large number of threads is used by introducing a new thread partition estimation and adjusting several blocking parameters.
• Improved performance of fp8 convolution primitive with scales and bf16 output
• Improved performance of matmul primitive with post-ops on processors with Intel AMX support
• Improved performance of RNN primitive for LBR_GRU and VANILLA_LSTM cell types on processors with Intel AVX2 instruction set support
• Improved performance of the following subgraphs with Graph API:
  • Scaled Dot Product Attention (SDPA) with implicit causal mask.
  • Grouped Query Attention (GQA) flavor specific for GEMMA models.

Intel Graphics Products

• Improved performance on Intel GPUs based on Xe3 architecture.
• Improved matmul performance for Intel Arc Graphics for Intel Core Ultra processors (Series 2) (formerly Lunar Lake).
• Improved RNN primitive performance with LBR_GRU cell type.
• Improved int8 convolution performance with plain weights and trivial filter.
• Improved convolution performance with NCHW activations with 1x1 filter and unit strides.
• Improved fp32 softmax performance.
• Improved performance of reorder when used with USM host memory.
• Improved performance of the following subgraphs with Graph API:
  • fp32 SDPA with implicit causal mask.
  • fp16 SDPA on Intel GPUs without Intel XMX cores.

AArch64-based Processors

• Improved int8 convolution performance.
• Improved bf16 depthwise convolution performance.
• Improved f16 matmul performance with Arm Compute Library (ACL).

Functionality

Functional API

• Introduced Root Mean Square Normalization (RMSNorm) mode for layer normalization primitive. This functionality is optimized for Intel CPUs and Intel GPUs.
• Sparse memory objects and sparse matmul are promoted to production status.

Graph API

• Introduced support for tanh approximation in GELU operation.
• Extended Graph API Softmax operation to support optional stats output.
• Introduced fusion support for SDPA training forward and backward propagation.
• Introduced fusion support for SDPA with bottom-right implicit causal mask.
• Introduced make_scalar_tensor() API for engine-agnostic scalar tensor creation.

Microkernel API

• Introduced support for fp8 data type.

Intel Architecture Processors

• Introduced support for select algorithm in binary post-op.
• Introduced source, destination, and weight scales support in fp8 convolution and deconvolution primitives.

Intel Graphics Products

• Introduced support for select algorithm in binary primitive.

Generic GPU Vendor

• Introduced support for RNN Vanilla backward propagation.

Usability

• Enabled build with -Wundef compiler flag.
• [Experimental] Introduced support for kernel compilation with SYCL kernel compiler extension.

Validation

• Improved benchdnn performance by optimizing input data filling and testing results comparison steps.
• Improved benchdnn graph driver performance mode via adding CPU memory pool for allocator.

Known Limitations

• The group normalization with normalization_flags::use_scale specified produces incorrect results for backward propagation kind in oneDNN v3.9 and earlier.
• Binary primitive with certain shapes and Graph API SDPA with bottom right causal mask may hang with SYCL debug runtime on Windows.
• fp8 matmul primitive may sporadically produce incorrect results on Intel Arc B-series graphics.
• int8 inner product primitive with tensors exceeding 4 Gb in size may produce incorrect results on Intel Data Сenter GPU Max series.
• bf16 pooling with tensors exceeding 4 Gb in size may produce incorrect results on Intel Data Сenter GPU Max series.
• bf16/fp16 matmul with large inner dimension has a performance regression on Intel Data Сenter GPU Max Series.
• bf16/fp16 convolution with NCHW activations has a performance regression on Intel Data Сenter GPU Max Series.
• Softmax with non-trivial strides and blocked format may produce incorrect results.
• bf16 layer normalization backpropagation may produce incorrect results on Intel Data Сenter GPU Max Series.

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.

Thanks to our Contributors

This release contains contributions from the project core team as well as Aditya Tewari @aditew01, Alexander Simonov @asimonov1, @Anallear, Anna Sztukowska @asztukow, Avanish Tiwari @Tiwari-Avanish, Dmitriy Ovchinnikov @inteldimitrius, Kasture Deeksha, Krishna Sai @krishnasai-mcw, Manaal @manaalmj, Marek Michalowski @michalowski-arm, Orel Yehuda @yehudaorel, Ruqiu Cao @rcao8, Tsao Zhong @CaoZhongZ, Viktoriia Gvozdeva @vgvozdeva, Yair Obodovsky @yair-obodovsky, Ye Tao @taoye9, Yuanyuan Chen @cyyever, @gausah-arm, @karmeh01, @pmanczak, and @zhangfeiv0. We would also like to thank everyone who asked questions and reported issues.