There are three types of AI Engine hardware I/O access: read, poll, and write. The AI Engine partition allows the
application to mmap all the I/O registers to read only to the
application so that the application can directly read and poll registers from the
userspace to save context switch time between
userspace and kernel space. The AI Engine partition driver limits register
write, and mask write can only be done with the AI Engine partition driver ioctl
syscall.
The reason that the AI Engine partition driver limits the application to write only
to AI Engine registers through
ioctl is because if userspace is the one to change the configuration without notifying the
Linux kernel, the Linux kernel loses track of which configuration has been changed, and
userspace might change which kernel Linux set.
Also, there are some registers which should be not written from userspace such as AI Engine interface tiles DMA registers. The AI Engine partition kernel driver writes to AI Engine registers. Be aware
of the AI Engine resource usage
because the AI Engine partition
kernel driver is responsible for events backtracking, tracing, and profiling. To use
ioctl to write to the register creates context
switch overhead. However, with the PDI configuration flow, writing to registers from
userspace is not frequent at runtime. Also, if
you want to speed up the configuration, consider allowing configuration files with
ioctl to execute the configuration requests in
batch mode with zero-copy.
sysfs interface can be used for userspace
to inquire the kernel module information. You can use it for events and errors
statistics and core DMA status dumps. See AI Engine Error and Events Handling for more
information about errors and events runtime handling.