Kirin devices, such as Kirin 9000, have a customized kernel driver that is reporting core masks incorrectly with multiple coherency group masks even though the coherency group count is 1. This results in the counter gen query API reporting an incorrect number of shader cores which propagates through to Streamline $MaliConstantsShaderCoreCount.
Workaround is to bitmask all coherency group masks together, irrespective of how many coherency groups are reported.
Fix implemented for libGPUInfo here: