synthesize floating point

[sanborne]

Sorry to be posting so frequently, but the information I have gotten
so far has been fantastic.

So one more question: Is it possible to synthesize floating point
data? I have heard that it is not, but some of the things I have seen
in my research of fixed point seems to indicate that it is possible to
synthesize VHDL REAL data. And it really seems that if floating point
data can never be synthesized, this is a huge gap that someone out
there must be working on.

Is there any documentation related to this question that someone could
point me to?

 

[Mike Treseler]

Is it possible to synthesize floating point
data?

http://www.vhdl.org/fphdl/vhdl.html

I have heard that it is not, but some of the things I have seen

in my research of fixed point seems to indicate that it is possible to
synthesize VHDL REAL data.

I use real constants in synthesis
to calculate integer constants.

 

[Thomas Stanka]

Hello,

So one more question: Is it possible to synthesize floating point
data? I have heard that it is not, but some of the things I have seen

It is possible. But not as easy as writting y = a*b and every tool did
a good job for you.
You need either a tool that mange this code (behavioral synthesis
tool) or you need to rewrite this code in a way that the result can be
processed by every tool (e.g. using slv and doing the math by shifting
mantisse, multipliying it and adding exponents than normalize result).

bye Thomas

 

[Tricky]

Sorry to be posting so frequently, but the information I have gotten
so far has been fantastic.

So one more question: Is it possible to synthesize floating point
data? I have heard that it is not, but some of the things I have seen
in my research of fixed point seems to indicate that it is possible to
synthesize VHDL REAL data. And it really seems that if floating point
data can never be synthesized, this is a huge gap that someone out
there must be working on.

Is there any documentation related to this question that someone could
point me to?

It is, but probably not in the way you are expecting.

The VHDL Real type is NOT synthesisable. It can only be used for
testbenching or setup constants/generics only.

But floating point IS possible. an IEEE FP number is just a 32 bit
number, but the arithmatic is fairly complicated compared to fixed
point. Both Xilinx and Altera have libraries of FP components that
will pipeline floating point arithmatic (latencies of 3-4 clock cycles
is fairly standard, but they can be parameterised), but they are
expensive in terms of logic compared to fixed point.

Personally, I would exhaust all avenues of fixed point before moving
to floating point. The IEEE float_pkg is a great addition to the
standard, but it will not produce a useful pipelined architectures at
the moment. The implementations of the float package for the different
vendors appear to recommend shortened FP numbes (like a 15bit mantissa
instead of 23).

If you have Quartus then info on the Altera FP components can be found
via Help->Megafunctions/LPM->Arithemtic megafunctions. All floating
point blocks are named altfp_...

The only downside to using the vendor libraries is that they are
vendor dependent, on purpose.

 

[Tricky]

Actually, the latencies vary alot.

In Altera, the default for a fp adder is 11 clocks, and a divider is
41 clocks.

 

[ajjc]

Sorry to be posting so frequently, but the information I have gotten
so far has been fantastic.

So one more question: Is it possible to synthesize floating point
data? I have heard that it is not, but some of the things I have seen
in my research of fixed point seems to indicate that it is possible to
synthesize VHDL REAL data. And it really seems that if floating point
data can never be synthesized, this is a huge gap that someone out
there must be working on.

Is there any documentation related to this question that someone could
point me to?

You can also use the FPGA coprocessor floating point package:

http://libhdlfltp.sourceforge.net

In the next release(soon), Altera qualification is added,
as are pipelined sin,cos, some bug fixes, and a working board example.

alan