Deploy a Decimal FP co-processor with openJDK, is it possible?

[tarekeldeeb]

Hello community,

I work for SilMinds; a company that provides Decimal floating point IP cores. We are currently building up a pilot application for our customers as well. Our decimal floating point unit is compliant with IEEE 754-2008 standard. Executing decimal arithmetic routines on hardware shows a delay-energy boost of 500X.

I want to patch the openJDK so that all Decimal data types are executed on my hardware co-processor. Java applications shall be accelerated without recoding or even re-compiling.

 I have some basic questions:

1- What are all decimal datatypes? java.math.BigDecimal ? other string representations ?
2- Is there any java conversion functions to be IEEE 754-2008 compliant?
3- I know that not all the JDK is open, is this correct? Is the arithmetic part -including the BigDecimal- open ?
4- What documents shall I start with to override the default arithmetic routines and deploy my hardware ?
5- Do you have any personal hints or suggestions towards what I had said ?

I appreciate any clues or information.

Thanks in advance,

Tarek

[Volker Simonis]

As a starting point, you can have a look at the webrev for:

6622432: RFE: Performance improvements to java.math.BigDecimal
http://mail.openjdk.java.net/pipermail/hotspot-dev/2009-February/001226.html

which is a performance improvement for BigDecimal. This implies that
the BigDecimal implementation is in the OpenJDK.

In order to gain full advantage of your FP-coporocessor, you would
probably have to natively reimplement or intrinsify some of the
methods of BigDecimal. This will however probably only pay off for big
numbers.

Regards,
Volker


On 3/23/09, tarekeldeeb <tarekeldeeb@...> wrote:

>
>  Hello community,
>
>  I work for SilMinds; a company that provides Decimal floating point IP
>  cores. We are currently building up a pilot application for our customers as
>  well. Our decimal floating point unit is compliant with IEEE 754-2008
>  standard. Executing decimal arithmetic routines on hardware shows a
>  delay-energy boost of 500X.
>
>  I want to patch the openJDK so that all Decimal data types are executed on
>  my hardware co-processor. Java applications shall be accelerated without
>  recoding or even re-compiling.
>
>   I have some basic questions:
>
>  1- What are all decimal datatypes? java.math.BigDecimal ? other string
>  representations ?
>  2- Is there any java conversion functions to be IEEE 754-2008 compliant?
>  3- I know that not all the JDK is open, is this correct? Is the arithmetic
>  part -including the BigDecimal- open ?
>  4- What documents shall I start with to override the default arithmetic
>  routines and deploy my hardware ?
>  5- Do you have any personal hints or suggestions towards what I had said ?
>
>  I appreciate any clues or information.
>
>  Thanks in advance,
>
>  Tarek
>
> --
>  View this message in context: http://www.nabble.com/Deploy-a-Decimal-FP-co-processor-with-openJDK%2C-is-it-possible--tp22621754p22621754.html
>  Sent from the OpenJDK General discussion mailing list archive at Nabble.com.
>
>

[Joe Darcy]

tarekeldeeb wrote:

> Hello community,
>
> I work for SilMinds; a company that provides Decimal floating point IP
> cores. We are currently building up a pilot application for our customers as
> well. Our decimal floating point unit is compliant with IEEE 754-2008
> standard. Executing decimal arithmetic routines on hardware shows a
> delay-energy boost of 500X.
>
> I want to patch the openJDK so that all Decimal data types are executed on
> my hardware co-processor. Java applications shall be accelerated without
> recoding or even re-compiling.
>
>  I have some basic questions:
>
> 1- What are all decimal datatypes? java.math.BigDecimal ? other string
> representations ?
>  

That is the one.

> 2- Is there any java conversion functions to be IEEE 754-2008 compliant?
>  

The operations on BigDecimal with limited precision are similar but not
identical to those defined by IEEE 754-2008.  BigDecimal does not have
infinities, signed zero, or NaN values.  (Roughly) the exponent range of
BigDecimal is a 32-bit integer and the significand can be in effect
arbitrarily large.

> 3- I know that not all the JDK is open, is this correct? Is the arithmetic
> part -including the BigDecimal- open ?
>  

Yes, all the code necessary to implement java.math is in the open.

> 4- What documents shall I start with to override the default arithmetic
> routines and deploy my hardware ?
>  

As Volker indicates, the likely path is hacking HotSpot to intrinsify
certain operations.

> 5- Do you have any personal hints or suggestions towards what I had said ?
>  

The semantics of BigDecimal are defined by its specification, including
the rounding operations.  I'd imagine care would need to be taken
bridging between hardware decimal support and software support for out
of range exponents or significands.

Have fun,

-Joe