Page 2 of 2

Re: FPU Register Accuracy

Posted: Tue Nov 15, 2016 10:24 pm
by Robert
Some interesting comments above.

I did wonder if a disassembler might be a useful tool to investigate this issue.

Re: FPU Register Accuracy

Posted: Wed Nov 16, 2016 2:51 am
by DGDanforth
Robert wrote:Some interesting comments above
Robert,
I have my post order set to "Descending" so the interesting comments are "below".

Re: FPU Register Accuracy

Posted: Wed Nov 16, 2016 7:50 am
by cfbsoftware
I'm not an expert on this subject but before somebody goes too far down the path of enhancing the 80-bit x87 capabilities of BlackBox it might be worth taking a leap ahead and consider using the 128-bit capable SIMD instructions instead:
The x87‭ ‬FPU instructions are executed by the so-called "math coprocessor"‬.‭ ‬These instructions operate on floating-point,‭ integer,‭ ‬and binary-coded decimal (BCD) ‬operands.‭ The main purpose of these instructions are to perform floating-point arithmetic.‭ ‬But nowadays we have SIMD instructions that are much faster than FPU.‭ ‬So,‭ please,‭ don't use the FPU in newly written code,‭ because it usage is marked as outdated in Linux ‬64‭ ‬ABI.‭ ‬Use SIMD instead.
http://linasm.sourceforge.net/docs/inst ... ompare_flt

However, that might have to wait for a 64-bit version of BlackBox?

Re: FPU Register Accuracy

Posted: Wed Nov 16, 2016 10:06 am
by Robert
cfbsoftware wrote:... it might be worth taking a leap ahead and consider using the 128-bit capable SIMD instructions instead:
I think this is rather a different direction.
1 - The 128-bit refers to the parallel multiple data stream path. It does not imply 128-bit accuracy, but rather doing 2 64-bit calculations in one memory fetch, or 4 32-bit calculations, or ...
2 - I think that direct language support of multiple data streams is a big change from where we are today, and would need to be convinced that it was worth while.
3 - Writing a library to do, for example, additions of Vectors comprising 64-bit, or 32-bit REALs which itself used SIMD internally may well have uses. Such a library Module could be written today using a combination of Component Pascal and Assembler without changing the current language.

Having an assembler well integrated into the source code editor would be helpful here, but that is a different topic.

Re: FPU Register Accuracy

Posted: Fri Nov 18, 2016 10:37 am
by luowy
@Josef, are you start the work for Real80?

Re: FPU Register Accuracy

Posted: Fri Nov 18, 2016 10:45 am
by Josef Templ
luowy wrote:@Josef, are you start the work for Real80?
No.