Download Latest Version v3.13 source code.zip (17.8 MB)
Email in envelope

Get an email when there's a new version of oneDNN

Home / v3.12.1
Name Modified Size InfoDownloads / Week
Parent folder
README.md 2026-06-05 1.2 kB
v3.12.1 source code.tar.gz 2026-06-05 14.0 MB
v3.12.1 source code.zip 2026-06-05 17.6 MB
Totals: 3 Items   31.6 MB 0

This is a patch release containing the following changes to v3.12:

  • Enabled SYCL Graph record/replay mode support in Graph API on Intel GPUs (f0147241f57870ea719310cb9355e1a84cd2ffee, 3e19991d38fa5a3fce68c463fdbc6afbc9e81ab2, 982650a485b1ed949adc16a749f26769eaba4306)
  • Fixed a performance regression in matmul with 4D shapes and N == 1 or M == 1 on x64 CPUs (de9498ff8862ab9e0c5e9b6fd8b3a8f8c6ab53ba, 540d53dbe9cdcec4cece82a9b7e3e474755ea8a5)
  • Fixed correctness issue in matmul primitive with ReLU post-op on RV64 CPUs (772ca13e27124aa477b18e043cd83d5ddefc7cea)
  • Fixed a segfault in s8/u8 depthwise convolution on x64 processors with Intel AVX10.2 instruction set support (7f413f9fe0bf78cdf9c1dd59a92d3c7b03a838af, dac73991dfe806e3feba63b60c08ef1c58dbfed7)
  • Fixed a correctness issue in s8s8 convolution on x64 processors with Intel AVX-512 and Intel DL Boost instructions support (0dc2ca853ccef716a3ba1702000adcc2bad40e96)
  • Fixed an issue with incorrect memory use estimation of layer normalization, group normalization, and batch normalization primitives in benchdnn (881e9b68d275a63979ebb6c82cea378959a487ec)
  • Fixed an assertion in benchdnn --graph driver (79b2593ac550bbe118335fad96649d087354b89f)
Source: README.md, updated 2026-06-05