Performance issue

[artyom-beilis]

I run some benchmarks on this build on rx560

• Resnet50 train on 16 batch size takes 1.2s
• Old version of rocm (I tested several years ago) gave me 0.56s per same batch.

My current OpenCL backend gives on same setup gives 0.88s per same batch. Something similar to what I used to have (60% of rocm) - I almost never suppressed native rocm build.

Do you have an idea what may be wrong?