Apropos forceinline.

yuriisig · ‎08-11-2012

If program and subroutine are in different files forceinline does not work. Why?

Steven_L_Intel1 · ‎08-11-2012

The compiler can't inline unless they are in the same source or if you use /Qipo

yuriisig · ‎08-11-2012

Steve,

Thanks.
Why Intel Fortran does not support double extended?

Steven_L_Intel1 · ‎08-12-2012

You mean 80-bit floating? That's an x87 feature not supported by the SSE instructions. It would be a lot of work to support for not much benefit, so we chose to do other things. We do support 128-bit floating.

yuriisig · ‎08-12-2012

>>You mean 80-bit floating?
Yes.
>>That's an x87 feature not supported by the SSE instructions.
I know it.

>> ...to support for not much benefit.
Not always
>>We do support 128-bit floating.
It very slowly.

yuriisig · ‎08-13-2012

Quoting yuriisig

>>We do support 128-bit floating.
It very slowly.

For example, speed of multiplication of matrixes decreases almost in 300 times in comparison with dgemm Intel MKL!!! It is absolutely unacceptable result. My algorithm concedes in speed of all in 13 times.

Steven_L_Intel1 · ‎08-13-2012

I understand that 128-bit floating is slow. But adding a new datatype, especially one that is not a power of two in byte size, is a major undertaking for the compiler and, while we've received a few requests for this over the years, other work has been given higher priority.

MKL DGEMM is very heavily optimized and uses multiple cores - not really a fair comparison.

yuriisig · ‎08-13-2012

Quoting Steve Lionel (Intel)

MKL DGEMM is very heavily optimized and uses multiple cores - not really a fair comparison.

I know it: http://redfort-software.intel.com/en-us/forums/showthread.php?t=76595&o=d&s=lr

My algorithm with accuracy more than 30 decimal signs concedes MKL DGEMM In total in 13 times (for x64).

yuriisig · ‎08-24-2012

Quoting Steve Lionel (Intel)

It would be a lot of work to support for not much benefit, so we chose to do other things.

You are wrong: the good compiler should support the main possibilities of the processor. For development of many algorithms with accuracy double support 80-bit floating is important.