PGI User Forum

[karim]

Hello

Is it possible to compile SCALAPACK in quadruple precision using the PG compiler ?

Thanks

Karim.

[paulv]

Hello,

As far as I can tell, pgf90 v5.2-1 doesn't support quad precision. The result of my Test_Type_Kinds program for PGI on a linux x86 system is:

Kind types:
Double float kind type: 8
Quad float kind type: 8

(If the Quad kind type doesn't exist, i.e. = -1, then my code defaults to the next "largest" type, in this case the Double float kind which == 8.)

Using Lahey v6.2 shows that quad precision is available on x86,

Kind types:
Double float kind type: 8
Quad float kind type: 16

but my experience has been that, when available, Quad precision is extremely slow since the calcs aren't done in hardware, but in software. But, I know nothing about SCALAPACK or your hardware, so my observations may not apply.

cheers,

paulv

[mkcolg]

Hi Karim,

PGI does not support quad precision because of the lack of hardware support. As Paulv points out, software support of real*16 would have very poor performance.

Thanks for the question.
Mat

[karim]

Hi

Thanks for your answers,
For the problem of performance with quad precision, I though to use opteron processors.

Cheers

Karim.

[mwolfe]

Hi Karim
Opteron extends the x86 processor by enlarging the integer registers from 32bit to 64bits, adding 8 new 64bit integer registers, and adding 8 new SSE registers. However, it does not have hardware support for 128bit floating bit arithmetic. The 128bit SSE registers are used in four modes:
- single 32bit floating point, ignoring the other 96 bits (SSE1)
- packed 4x32bit floating point, using all 128 bits (SSE1)
- single 64bit floating point, ignore the other 64 bits (SSE2)
- packed 2x64bit floating point, using all 128 bits (SSE2)
None of the x86 processors have hardware for 128bit floating point arithmetic.
I'm not sure any mainstream microprocessor does.
-