The AXI4 Lite bus Xilinx uses addresses the bytes in the 32 bit data word.
Since the config registers used in by the Linux system only cares about the 32
bit registers the lower two address bits get ignored. These changes makes it
easier to integrate the module when using it in an IP Package.
Some projects will name the bitstream files a standard name and thus by
renaming them we can ensure that the bitstream files are easily identified
after syntheses.
This allows for the the 'pkg-module' script to have bash tab completion when
invoked in the repos root directory. Its the same as the 'sim-module' auto
completion in that it looks in the 'hdl' directory for module file names.
Invoking this script with a modules filename as argument will try and create a
Xilinx IP Package. Once crated the package is moved to a 'ip/ip_repo' directory
and can be used in a Vivado project.
The method of package creation is simple and for the most part automatic and
thus we assume that Xilinx naming conventions etc are observed.
IP Packages are going to be created using HDL files in the repo. These files
need to be moved into the package directory for use in the package. This change
ensure that the files are made available. Its not the best as it copies ALL
files into the package directory and not just the needed files.
The functionality of the axis_memory_offset function can be replicated by using
mem_alloc_state while at the same times provides a more powerful interface to
the memory allocation system.
The 1st argument passed to the new mem_alloc_state functions is a pointer to a
void pointer (start) whose value will be overwritten to be the start pointer of
the CMA memory area. The 2nd argument is a pointer to a unsigned int whose
values will be overwritten to be the current 'offset' from that start pointer
and thus represents how many bytes has been allocated.
This function takes as argument an offset indexed from zero and returns a usage
pointer corresponding to the offset from the start of the CMA region. Thus an
argument offset of zero would give the pointer to the start of the CMA region.
This change prepends the CROSS_COMPILE environmental variable to the compilers
definition. Thus adding the possibility of cross compiling when the cross
compiling tool chain has been setup correctly.
By explicitly pairing the clock running the interface with the interface
in Vivado it allows for better timing analysis and remove some 'CRITICAL
WARNINGS'.
The memory is addressable using the AXIS ports in 64 bit (8 byte) chunks. Thus
when allocating memory from CMA region, the start address needs to be 64 bit
aligned, not page aligned. This change will thus reduce the 'padding' at the
end of arrays and thus RAM.
This devicetree code is an altered version of the
'arch/arm/boot/dts/zynq-7000.dtsi' file found in the linux-xlnx Xilinx
repo, master branch commit (da2d296bb6b89f7bc7644f6b552b9766ac1c17d5).
The only thing that was added was the following axis node:-
```
axis: axis@43C00000 {
compatible = "xlnx,axis-1.00";
reg = < 0x43C00000 0x10000 >;
xlnx,num-mem = <0x1>;
xlnx,num-reg = <0x20>;
xlnx,s-axi-min-size = <0x1ff>;
xlnx,slv-awidth = <0x20>;
xlnx,slv-dwidth = <0x20>;
};
```
Similar functions could be constructed using calls to the already existing
functions, however, it seems more efficient to add them to the library in case
the application cannot optimizes external library calls.
The uio driver exports files to the /sys directory, one of which gives the RAM
memory offset for the *mem* CMA array. This will be needed when giving address
to the axis ports, it is thus read from the file during the init function call.
Arrays that will be used to pass data between the host and fpga need to be page
aligned and be slices of the DMA array the driver has reserved. These new
functions allocate slices from the DMA array and ensures that the starting
address is page aligned.
The driver sets ups a DMA contiguous array that will be used as a memory area
to share data between the PS and PL. The library memory maps this DMA array to
make it available to the host applications. The array is cache coherent and
thus data written or read from it bypass the CPU cache and read/write directly
from main memory. This ensures that the data is always the same between the
host and fpga but at a cost of long data accesses from the host CPU.
Add a High Performance port (HP0) to the zedboard_axis project and
connect a AXIS module that can read and write to the main memory using
HP0. Add two small applications that test writing and reading from
memory.
The application takes one argument, the number (or length) of data to be
written to the memory.
The physical address in main memory that is going to be written to is the
address of the CMA memory region on my currently running system.
Because the axis modules are minimalist, it assumes that all
transactions are performed without error and that transaction processing
is performed in-order. Thus the transaction IDs can safely be set to
zero. Doing so also fixes a bug in the write path that had prevented
multiple burst of data being sent. This was due to the fact that the
write data burst ID needs to have the same ID as the write address. But
the write address module was incrementing the ID while the write data ID
was not being set at all.