misc-habanalabs-next-2019-04-19
tagged this
19 Apr 09:25
The major changes are: - Add a new IOCTL for debug, profiling and trace operations on the device. This will allow the user to perform profiling and debugging of the deep learning topologies that are executing on the ASIC. - Add a shadow table for the ASIC's MMU page tables to avoid doing page table walks on the device's DRAM during map/unmap operations. - re-factor of ASIC-dependent code to be common code for all ASICs In addition, there are many small fixes and changes. The notable ones are: - Allow accessing the DRAM using virtual address through the debugFS interface. Until now, only physical addresses were valid, but that is useless for debugging when working with MMU. - Allow the user to modify the TPC clock relaxation value to better control TPC power consumption during topology execution. - Allow the user to inquire about the device's status (operational/Malfunction/in-reset) in the INFO IOCTL. - Improvements to the device's removal function, to prevent crash in case of force removal by the OS. - Prevent PTE read/write during hard-reset. This will improve stability of the device during hard-reset.