Jit uni reorders give incorrect results with certain 4d matrices and src zero point != 0

[Ryo-not-rio]

Jit:uni reorders for certain 4d matrices are returning incorrect results when the src zero point != 0. A guard has been added to skip jit uni reorders so the simple:any implementation will be used in #2207, however a fix for jit:uni should be created to improve performance.

From my initial investigation, it seems to happen for some shapes when dim[2] > 16 and dim[3] != multiple of 2 but I'm sure there are other shapes that would cause the same issue.

Version

oneDNN v3.7.0 (commit 2fc6ff9)

Environment

oneDNN includes hardware-specific optimizations and may behave
differently on depending on the compiler and build environment. Include
the following information to help reproduce the issue:

• CPU make and model: aarch64, graviton 3
• OS version: Linux 22.04.1-Ubuntu
• Compiler version: gcc 11.4.0
• CMake version: 3.30.2
• git hash: 1712f97

Steps to reproduce

./tests/benchdnn/benchdnn --reorder  --sdt=u8 --ddt=f32 --stag=adbc --attr-zero-points=src:common:1 1x32x128x33

Observed behavior

benchdnn test fails.

Expected behavior

benchdnn test passes.

[Ryo-not-rio]

@Shreyas-fuj I noticed you contributed to Jit:uni recently. Could you have a look at this or tag the appropriate person for this? Thank you

[Shreyas-fuj]

Hi @Ryo-not-rio,
I will have a look at this. Thanks for the temporary fix : #2207.