v3.12

Performance Optimizations

Intel 64/AMD64 Processors

• Improved performance on future Intel Core Ultra processors with Intel AVX10.2 instruction set support (code name Nova Lake). These optimizations are now enabled by default on compatible processors.
• Improved performance on future Intel Xeon processors with Intel AVX10.2 and Intel AMX instruction set support (code name Diamond Rapids). These optimizations are now enabled by default on compatible processors.
• Improved performance of fp8 and int8 matmul with transposed source on processors with Intel AMX instruction set support.
• Improved performance of bf16 and f16 matmul with transposed source on processors with Intel AVX2 instruction set support.

Intel Graphics

• Introduced initial performance optimizations for future integrated GPUs based on Xe3p-LPG architecture.
• Introduced initial performance optimizations for future discrete GPUs based on Xe3p-XPC architecture. This is a preview functionality not recommended for production use.
• Improved f16 matmul performance on Intel Arc Graphics for Intel Core Ultra processor Series 3 (formerly Panther Lake).
• Improved performance of matmul with host-side scalar arguments.
• Improved matmul performance for cases with small M/N and large K.
• Improved SDPA forward and backpropagation subgraph performance with Graph API.

AArch64 Processors

• Improved f16 and f32 softmax performance across Arm Neoverse cores.
• Improved eltwise performance on Arm Neoverse N1 cores.
• Improved matmul and convolution performance on Arm Neoverse V2 cores.
• Improved performance of multiple primitives by quering processor cache sizes.

RISC-V Processors

• Improved f32 matmul, inner product, convolution, softmax and layer normalization primitives performance on processors with V extension support.
• Improved f16 softmax primitive performance on processors with Zvfh extension support.

Functionality

Functional API

• [experimental] Introduced grouped memory format and grouped matmul support to improve performance of AI models based on Mixture-of-Experts (MoE) architecture. This is an experimental feature that requires opt-in with ONEDNN_EXPERIMENTAL_GROUPED_MEMORY=ON build option. Optimized version of this functionality is implemented for Intel GPUs.
• [experimental] Extended grouped matmul with optional execution-time hint DNNL_ARG_HINT_MAX_GROUP_SIZE to communicate the maximum size of the group across the variable dimension for the execution call.

Graph API

• Introduced Dropout operation. Extended supported fusion patterns to enable fusion of Dropout with Matmul, Softmax, and elementwise operations.

Usability

Common

• Extended information about primitive execution available in VTune Profiler with the same level of details as reported by oneDNN verbose mode. This feature requires VTune Profiler 2025.7 or later.

Intel Graphics

• [experimental] Introduced support for Level Zero runtime on Intel GPUs. New functionality includes Level Zero interoperability API and build knob ONEDNN_GPU_RUNTIME=ZE.

AArch64 Processors

• Reduced memory usage of certain convolutions on Arm Neoverse V1/V2 cores.
• Fixed a bug causing high-memory usage and crashes in convolution with certain post-ops.

Validation

• Extended benchdnn with support for integer masks in quantization attributes.
• Improved consistency of benchdnn performance results when data compression is enabled by default on Intel Graphics.

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.
• f4_e3m0 data type is deprecated and will be removed in future releases.

Thanks to our Contributors

This release contains contributions from the project core team as well as Alexandre de Limas Santana @alexandrelimassantana, Andrei (Andrey) Khropov @andrey-khropov, Andrei Hutu @Anndrey24, Fadi Arafeh @fadara01, George Nash @georgen117, Kamil Wieloch @kwieloch-intel, Kasture Deeksha, MarkVeerasingam @MarkVeerasingam, Nikhil Gupta @nikhil-arm, @pmanczak, @vishwascm, and Xia Zhuozhao @xiazhuozhao.

v3.11.3

This is a patch release containing the following changes to v3.11.2:

• Fixed undefined behavior in matmul implementation on Intel64/AMD64 CPUs (bd117e4)
• Fixed performance regression in f32 reorder on Intel64/AMD64 CPUs (a4acece)
• Fixed a SEGFAULT in binary primitive with large sizes on Intel GPUs (157cba5, 1a8bc11)
• Fixed performance regression in f32 convolution with small number of input channels on processors with Intel AVX-512 instruction set support (e1f4a61)

v3.12-rc

Performance Optimizations

Intel 64/AMD64 Processors

• Improved performance on future Intel Core Ultra processors with Intel AVX10.2 instruction set support (code name Nova Lake). These optimizations are now enabled by default on compatible processors.
• Improved performance on future Intel Xeon processors with Intel AVX10.2 and Intel AMX instruction set support (code name Diamond Rapids). These optimizations are now enabled by default on compatible processors.
• Improved performance of fp8 and int8 matmul with transposed source on processors with Intel AMX instruction set support.
• Improved performance of bf16 and f16 matmul with transposed source on processors with Intel AVX2 instruction set support.

Intel Graphics

• Introduced initial performance optimizations for future integrated GPUs based on Xe3p-LPG architecture.
• Introduced initial performance optimizations for future discrete GPUs based on Xe3p-XPC architecture.
• Improved f16 matmul performance on Intel Arc Graphics for Intel Core Ultra processor Series 3 (formerly Panther Lake).
• Improved performance of matmul with host-side scalar arguments.
• Improved matmul performance for cases with small M/N and large K.
• Improved SDPA forward and backpropagation subgraph performance with Graph API.

AArch64 Processors

• Improved f16 and f32 softmax performance across Arm Neoverse cores.
• Improved eltwise performance on Arm Neoverse N1 cores.
• Improved matmul and convolution performance on Arm Neoverse V2 cores.

RISC-V Processors

• Improved f32 matmul, inner product, convolution, softmax and layer normalization primitives performance on processors with V extension support.
• Improved f16 softmax primitive performance on processors with Zvfh extension support.

Functionality

Functional API

• [experimental] Introduced grouped memory format and grouped matmul support to improve performance of AI models based on Mixture-of-Experts (MoE) architecture. This is an experimental feature that requires opt-in with ONEDNN_EXPERIMENTAL_GROUPED_MEMORY=ON build option. Optimized version of this functionality is implemented for Intel GPUs.
• [experimental] Extended grouped matmul with optional execution-time hint DNNL_ARG_HINT_MAX_GROUP_SIZE to communicate the maximum size of the group across the variable dimension for the execution call.

Graph API

• Introduced Dropout operation. Extended supported fusion patterns to enable fusion of Dropout with Matmul, Softmax, and elementwise operations.

Usability

Common

• Extended information about primitive execution available in VTune Profiler with the same level of details as reported by oneDNN verbose mode. This feature requires VTune Profiler 2025.7 or later.

Intel Graphics

• [experimental] Introduced support for Level Zero runtime on Intel GPUs. New functionality includes Level Zero interoperability API and build knob ONEDNN_GPU_RUNTIME=ZE.

AArch64 Processors

• Introduced support for the library to correctly query processor cache sizes.
• Reduced memory usage of certain convolutions on Arm Neoverse V1/V2 cores.
• Fixed a bug causing high-memory usage and crashes in convolution with certain post-ops.

Validation

• Extended benchdnn with support for integer masks in quantization attributes.
• Improved consistency of benchdnn performance results when data compression is enabled by default on Intel Graphics.

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.
• f4_e3m0 data type is deprecated and will be removed in future releases.

Thanks to our Contributors

This release contains contributions from the project core team as well as Alexandre de Limas Santana @alexandrelimassantana, Andrei (Andrey) Khropov @andrey-khropov, Andrei Hutu @Anndrey24, Fadi Arafeh @fadara01, George Nash @georgen117, Kamil Wieloch @kwieloch-intel, Kasture Deeksha, MarkVeerasingam @MarkVeerasingam, Nikhil Gupta @nikhil-arm, @pmanczak, @vishwascm, and Xia Zhuozhao @xiazhuozhao.

v3.11.2

This is a patch release containing the following changes to v3.11.1:

• Fixed an issue with unintentionally exposed internal symbols in Graph API (d369c5f, 8fc3ec3)
• Fixed an integer overflow in memory descriptor size computation for humungous tensors (265704b, 98e2011)
• Fixed a potential heap corruption in f32 GEMM kernels on x64 CPUs (6191181)
• Added support for bf16 and f16 matmul with transposed source on x64 CPUs with Intel AVX2 instruction set support (5af82d4, f0c6428, eef1cf0)
• Updated benchdnn to use non-compressible random data in performance benchmarking mode on Intel GPUs (8051b37)

v3.11.1

This is a patch release containing the following changes to v3.11:

• Fixed performance regression in bf16 matmul with int4 weights on Intel GPUs based on Xe2 architecture (d4d4d7a)
• Fixed performance regression in inner product primitive with transposed weights on x64 CPUs (c5d2d09)
• Updated benchdnn input files for matmul and convolution performance benchmarking (e80a1a8, 96d72a9, b9c9bce)
• Fixed an out of registers issue in SDPA fusion with Graph API on Intel GPUs (ba81382)
• Fixed integer overflow in softmax primitive implementation for Intel GPUs (4a711d7, b02cfa0, c557f33, ab64a9b)
• Fixed incorrect results in f64 convolution weight gradient on Intel GPUs based on Xe-LPG architecture (adcb323, 3d1a7e4)
• Removed in-place optimization for reorder in Graph API to avoid correctness issues (a6c3630)
• Improved performance of int8, f16, and bf16 convolution on processors with Intel AMX support (a418949)
• Fixed a correctness issue in f32 convolution with small number of input channels (3d1d9b4, ada85c5)
• Fixed a correctness issue in matmul with binary post-op and non-trivial strides on x64 CPUs (f49f470, 265df18, 5892570)
• Fixed benchdnn graph driver test to support non-trivial strides (0232763, 662cbb3)
• Fixed a correctness issue in 3D grouped convolution weight gradient on Intel GPUs (8a7996b)
• Fixed a page fault issue in f32 SDPA subgraph on Intel GPUs (98845e5)
• Fixed a performance regression in bf16 matmul on x64 CPUs with Intel AMX instruction set support (5b886e8, f3a79e7, 52cc900, cf9a11e)
• Fixed a segmentation fault in matmul on x64 processors with Intel AVX 10.2 and Intel AMX instruction set support (98aea2f)
• Fixed correctness issue in SDPA subgraph with non-trivial strides for mask on Intel GPUs (0ccdfba)

v3.11

Performance Optimizations

Intel 64/AMD64 Processors

• Improved fp32 matmul performance with fp4 compressed weights.
• Improved fp32 matmul performance for cases when one of the tensors has a trivial dimension on processors with Intel AVX-512 instruction set support.

Intel Graphics

• Improved fp16/bf16 matmul performance for large tensor cases on Intel Graphics for Intel Core Ultra processor Series 3 (formerly Panther Lake).
• Improved matmul performance for cases with 4-byte alignment on Intel GPUs based on Xe2 architecture.
• Improved performance of fp16/bf16 matmul with mxfp4 weights.
• Improved convolution performance with host-side scalar scales and zero points.
• Improved matmul performance for LLM inference workloads on Intel GPUs based on Xe2/Xe3 architectures.
• Improved f32 SDPA performance for small head sizes.

AArch64 Processors

• Improved performance of bf16 matmul.
• Improved performance of bf16/int8 convolutions.
• Improved matmul performance for cases when one of the tensor has a trivial dimension.
• Improved performance of s8/u8 eltwise post-ops on Arm(R) Neoverse(TM) V1 processors.
• Improved f16 and bf16 eltwise performance with abs, relu, square, sqrt, clip, and clip_v2 algorithms.
• Improved eltwise exp algorithm performance on Arm(R) Neoverse(TM) N1 processors.
• Improved reorder primitive performance.

RISC-V Processors

• Improved f32 matmul, inner product, convolution, softmax, batch normalization, layer normalization, and group normalization primitives performance.
• Improved eltwise and binary primitives performance.
• Improved f32 and fp16 pooling primitive performance.
• Improved fp32 to u8 reorder primitive performance.

Functionality

Functional API

• Introduced destination tensor dynamic quantization in matmul primitive following Open Compute Microscaling (MX) formats specification. See MXFP8 matmul tutorial for quick introduction into MX-capabilities in oneDNN.
• Introduced support for NVFP4 quantization scheme. The changes include support for fp8_e4m3 grouped scales and dynamic quantization support for destination tensor with NVFP4-specific formula for scales computation.
• Introduced support for dropout as a primitive attribute for matmul, softmax and eltwise primitives.

Graph API

• Introduced support for RMS Normalization operation.
• Introduced support for output gradient of attention mask for SDPA and GQA training.

Intel Graphics

• Introduced support for convolution with u8 weights.
• Introduced support for 2D grouped scales in fp8 and dual zero points in matmul.
• Extended support for 5D and 6D tensors in matmul with post-ops.

Intel 64/AMD64 Processors

• Introduced support for different data types of source and destination in pooling forward propagation.

AArch64 Processors

• Added limited support for the BRGEMM Microkernel API
• Added limited support for Windows on Arm builds with MSVC

Usability

Common

• Extended quantization attributes documentation to cover all quantization schemes supported by the library.
• Added matmul fp8 quantization example demonstrating use of matmul primitive with fp8 source, destination, and weights.
• Enabled ONEDNN_ENABLE_GRAPH_DUMP knob by default.

Intel 64/AMD64 Processors

• Extended oneDNN threadpool runtime with an option to support asynchronous execution and updated all CPU implementations accordingly. This extension makes oneDNN compatible with OpenXLA "thunk" runtime.
• Introduced ONEDNN_SAFE_RBP build knob that instructs x64 implementations to preserve value of rbp register for tools that rely on stack unwinding. This option may have visible performance impact on some workloads.

AArch64 Processors

• Fixed a potential overflow on AArch64 builds with Arm Compute Library.
• Significantly reduced memory consumption of convolution primitive with large spatial filters during primitive creation.

Intel Graphics

• Removed build time dependency on OpenCL runtime in SYCL build configuration.

Validation

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 Andrei Hutu @Anndrey24, Anna Sztukowska @asztukow, Arseniy Obolenskiy @aobolensk, Avanish Tiwari @Tiwari-Avanish, czekun @ZackyLake, Deeksha Kasture @kasturedeeksha, Fadi Arafeh @fadara01, Gassan Salama @gassan-arm, Henry Gardiner @henry-gar, @jstachowintel, Keanu Czirjak @keanucz, Krishna Sai @krishnasai-mcw, Murray Steele @murste01, Narendra Bagria @narenbagria, Joseph Kuo @PershingSquare, @pmanczak, @vishwascm, Yejing Lai @Yejing-Lai, 夏卓昭 @xiazhuozhao

v3.11-rc

Performance Optimizations

Intel 64/AMD64 Processors

• Improved fp32 matmul performance with fp4 compressed weights.
• Improved fp32 matmul performance for cases when one of the tensors has a trivial dimension on processors with Intel AVX-512 instruction set support.

Intel Graphics

• Improved fp16/bf16 matmul performance for large tensor cases on Intel Arc graphics for Intel Core Ultra processor series 3 (formerly Panther Lake).
• Improved matmul performance for cases with 4-byte alignment on Intel GPUs based on Xe2 architecture.
• Improved performance of fp16/bf16 matmul with mxfp4 weights.
• Improved convolution performance with host-side scalar scales and zero points.

AArch64 Processors

• Improved performance of s8/u8 eltwise post-ops on Arm(R) Neoverse(TM) V1 processors.
• Improved f16 and bf16 eltwise performance for abs, relu, square, sqrt, clip, and clip_v2.
• Improved exp eltwise performance on Arm(R) Neoverse(TM) N1 processors.
• Improved reorder primitive performance.
• Added matmul optimizations for GEMVs.
• Improved performance of bf16 matmul.
• Improved performance of bf16/int8 convolutions.
• Convolutions with large spatial filters now consume much less memory during primitive setup.

RISC-V Processors

• Improved eltwise and binary primitives performance.
• Improved f32 GEMM performance.
• Improved f32 matmul, softmax, convolution and inner product primitives performance.
• Improved f32 batch, group and layer normalization primitives performance.
• Improved f32 and fp16 pooling primitive performance.
• Improved reorder(fp32 to u8) primitive performance.

Functionality

Functional API

• Introduced destination tensor dynamic quantization in matmul primitive following Open Compute Microscaling (MX) formats specification. See MXFP8 matmul tutorial for quick introduction into MX-capabilities in oneDNN.
• Introduced support for NVFP4 quantization scheme. The changes include support for fp8_e4m3 grouped scales and dynamic quantization support for destination tensor with NVFP4-specific formula for scales computation.
• Introduced support for dropout as a primitive attribute for matmul, softmax and eltwise primitives.

Graph API

• Introduced support for RMS Normalization operation.
• Introduced support for output gradient of attention mask for SDPA and GQA training.

Intel Graphics

• Introduced support for convolution with u8 weights.
• Introduced support for 2D grouped scales in fp8 matmul.

Intel 64/AMD64 Processors

• Introduced support for different data types of source and destination in pooling forward propagation.

AArch64 Processors

• Added limited support for the BRGEMM Microkernel API.
• Added limited support for Windows on Arm builds with MSVC.

Usability

• Extended quantization attributes documentation to cover all quantization schemes supported by the library.
• Added matmul fp8 quantization example demonstrating use of matmul primitive with fp8 source, destination, and weights.
• Extended oneDNN threadpool runtime with an option to support asynchronous execution and updated all CPU implementations accordingly. This extension makes oneDNN compatible with OpenXLA "thunk" runtime.
• Extended information about primitive execution available in VTune(TM) Profiler with the same level of detail as reported by oneDNN verbose mode. This feature requires VTune Profiler 2025.7 or later.
• Introduced ONEDNN_SAFE_RBP build knob that instructs x64 implementations to preserve value of rbp register for tools that rely on stack unwinding. This option may have visible performance impact on some workloads.
• Removed build time dependency on OpenCL runtime in SYCL build configuration.
• ONEDNN_ENABLE_GRAPH_DUMP build knob is enabled by default.
• Fixed a potential overflow on AArch64 builds with Arm Compute Library.

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 Andrei Hutu @Anndrey24, Anna Sztukowska @asztukow, Arseniy Obolenskiy @aobolensk, Avanish Tiwari @Tiwari-Avanish, czekun @ZackyLake, Deeksha Kasture @kasturedeeksha, Fadi Arafeh @fadara01, Gassan Salama @gassan-arm, Henry Gardiner @henry-gar, @jstachowintel, Keanu Czirjak @keanucz, Krishna Sai @krishnasai-mcw, Murray Steele @murste01, Narendra Bagria @narenbagria, Joseph Kuo @PershingSquare, @pmanczak, @vishwascm, Yejing Lai @Yejing-Lai, 夏卓昭 @xiazhuozhao.

v3.10.2

This is a patch release containing the following changes to v3.10.1:

• Fixed a memory leak in Graph API related to host scalars use (0441245)
• Fixed f16 matmul performance regression with int4 weights on Intel Arc graphics for Intel Core Ultra processors (Series 3) (789711c, a160247)
• Fixed bf16 matmul performance regression on Intel Xeon processors with Intel AMX instruction set support (c29ec26)
• Changed register allocation in BRGEMM kernel to avoid register conflicts and improve code safety (95d651b)
• Fixed a crash related to incorrect caching of int8 convolution primitive on Intel GPUs (28ccca4, 0bc8060)
• Fixed a bug preventing correct detection of Intel AVX 10.2 instruction set on Intel Xeon processors (568171c)

v3.10.1

This is a patch release containing the following changes to v3.10:

• Fixed an issue with reorder primitive returning unimplemented for cases when only one scale mask is defined on AArch64 processors (be92457)
• Fixed sporadic correctness issue in fp32 matmul on Intel GPUs based on Xe2 architecture (b4a761c)
• Fixed correctness issue in fp16/bf16 matmul on Intel GPUs based on Xe3 architecture (48c114b)
• Fixed performance regression in bf16 convolution weight gradient on Intel Arc Graphics B-series (3b6665b)
• Improved convolution performance on AArch64 processors with SVE128 support (808227d)
• Fixed regression in matmul primitive creation time on Intel GPUs (599ecb5)
• Fixed potential overflow for matmul, convolution and inner product primitives with Arm Compute Library (be12d8c)
• Fixed convolution performance regression on Intel Arc Graphics B-series (7e27159)

v3.10

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 Alchemist) 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 reorder primitive performance.
• Improved bf16 convolutions performance.
• Improved convolutions performance on CPUs with 128-bit SVE support.
• Improved eltwise primitive performance on Arm(R) Neoverse(TM) N1 processor.

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 weight scales and zero-points with group size 16 in matmul with compressed weights.

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.

Known Limitations

• Convolution primitive may require excessive amount of scratchpad memory for shapes with large input width value on Intel CPUs.
• bf16 convolution primitive has a performance regression on Intel Arc B-series graphics.
• Reduction primitive may produce incorrect results for tensors exceeding 4 GB on Intel Arc graphics (formerly DG2) and Intel Arc Graphics for Intel Core Ultra processors (formerly Arrow Lake-H).
• Concat primitive may produce incorrect results for certain shapes on Intel Arc A-series GPUs.
• fp16 matmul primitive has a performance regression on Intel GPUs based on Xe2 architecture.
• f32 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 Center GPU Max series.
• bf16 layer normalization backpropagation may produce incorrect results on Intel Data Center 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.

Breaking Changes

AArch64-based Processors

• Bumped the minimum required Arm(R) Compute Library version to 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.