Extend support for JIT Backward Convolution Operators with ARM SVE 128bit #2165
Labels
enhancement
A feature or an optimization request
platform:cpu-aarch64
Codeowner: @oneapi-src/onednn-cpu-aarch64
Summary
On aarch64 platform, Convolution backward operators are supported via jitted SVE kernels. Today the support exists only for SVE 512 and SVE 256bit width, but not for SVE 128bit processors like AWS Graviton4. The request is to extend the existing SVE jitted kernels to support 128bit width.
Problem statement
resnet50 model training requires backward convolution operators, and these currently executed with reference 'c' kernels. Extending the following oneDNN operators for SVE 128bit accelerates these operators with SIMD and improved the performance by several orders.
Here are the details on the existing oneDNN jitted kernels for backward pass operators:
The kernel sources are here:
Preferred solution
Document your thoughts on what solution may look like.
The text was updated successfully, but these errors were encountered: