Is it possible to infer double data rate registers from VHDL code?

[bwilson79]

I couldn't really find anything in both the Xilinx XST user guide or
the Synplicity user guide. I tried several different approaches and
synthesized with Synplify Pro 8.8.0.4 and it's not working.

 

[Kai Harrekilde-Petersen]

I couldn't really find anything in both the Xilinx XST user guide or
the Synplicity user guide. I tried several different approaches and
synthesized with Synplify Pro 8.8.0.4 and it's not working.

Check the Xilinx website - IIRC the way to do this is to instanciate
specific cells from a xilinx technology library.


Kai

 

[Brad Smallridge]

I believe that for Xilinx parts, any double data
rate is handled at the IO ports with ISERDES and
OSERDES primitives. To gain speed, an input ISERDES
will have two pairs of shift registers, one clocked
at the positive edge, one at the negative. The logic
values are then latched into an output register on
a clock that represents a "framing" rate.

In the fabric, data is moved along on either negative
or positive edges, but I think not both. You might
get your simulator tool to handle the inference of
a dual edge data transfer, but the Xilinx synthesis
tool will give you an error.

Brad Smallridge
Ai Vision

 

[Andy]

I believe that for Xilinx parts, any double data
rate is handled at the IO ports with ISERDES and
OSERDES primitives. To gain speed, an input ISERDES
will have two pairs of shift registers, one clocked
at the positive edge, one at the negative. The logic
values are then latched into an output register on
a clock that represents a "framing" rate.

In the fabric, data is moved along on either negative
or positive edges, but I think not both. You might
get your simulator tool to handle the inference of
a dual edge data transfer, but the Xilinx synthesis
tool will give you an error.

Brad Smallridge
Ai Vision

In Xilinx and Altera, DDR is handled in the IOB, but not necessarily
with ISERDES/OSERDES (i.e. DDR is supported in devices with no ISERDES/
OSERDES capability). There are two connections to the fabric,
representing the data from/for both edges. The data streams are
typically handled separately in the fabric, or can be handled as one
with a 2x clock.

There is a separate discussion going on about the desired ability to
infer DDR circuits within the fabric that consist of only SDR flops
and combinatorial logic (and are described in RTL as being DDR). I
don't know why it is not used more often, but in practical terms, if
you have a DLL/PLL clock manager, creating a 2x clock and processing
the data internally at that rate is easier and more efficient in most
applications. Still, there is the odd case of a non-uniform clock,
etc. where doubling the clock is not feasible, and such circuits would
be desirable.

Andy

 

[beckjer]

While not a strict DDR implementation, I have found in my designs
(Altera, I haven't tried on Xilinx) that using 1 process per edge
seems to work. Granted it is more a read on falling edge, write on
rising edge system, but so far it's working well. My system is a
parallel interface, ~30nS clock cycle.

I use
process (clk)
begin
wait until rising_edge(clk);
....
end process;

process (clk)
begin
wait until falling_edge(clk);
....
end process;

It simulates in Modelsim with no problems (that I have seen), and the
protoboard appears to work when attached to the actual hardware.

 

[Jim Lewis]

I couldn't really find anything in both the Xilinx XST user guide or
the Synplicity user guide. I tried several different approaches and
synthesized with Synplify Pro 8.8.0.4 and it's not working.

In the VHDL RTL Synthesis Standard (1076.6-2004), the coding style is:

DualEdge_Proc: process (Clk, Reset) is
begin
if Reset = '1' then
Q <= (others => '0');
elsif rising_edge(Clk) then
Q <= D4Rise;
elsif falling_edge(Clk) then
Q <= D4Fall;
end if;
end process DualEdge_Proc;


For different clocks:
process( reset, clk1, clk2 )
begin
if reset = '1' then
Q <= '0' ;
elsif rising_edge(clk1) then
Q <= data1 ;
elsif rising_edge(clk2) then
Q <= data2 ;
end if ;
end process ;

While this may not currently work, it does give you a standard
to reference when you file a bug report with your vendors.

I do note there is a coding style in Xilinx that uses shared
variables, however, their usage of shared variables is illegal
in VHDL-2002.

Cheers,
Jim
SynthWorks VHDL Training