Intel Prepares AMX-COMPLEX Support For GCC & LLVM Compilers

[phoronix]

Phoronix: Intel Prepares AMX-COMPLEX Support For GCC & LLVM Compilers

Intel compiler engineers have sent out the initial GCC and LLVM/Clang compiler patches for enabling the newly-disclosed AMX-COMPLEX extension with next year's Xeon Scalable "Granite Rapids" processors...

https://www.phoronix.com/news/Intel-AMX-COMPLEX-GCC

[Setif]

AFAIK, complex numbers are only used in scientific computing where higher precision is required, but "complex numbers + FP16" does not make sense for me.
The only language (high-level, non-scripted) I know that have a first class support for Complex numbers is Fortran, where FP16 was only introduced recently by some compilers. (supporting Complex is optional in C11. Working with C++ complex doesn't feel native).

Maybe some other people use them for other purposes, who knows?

[brucethemoose]

AFAIK, complex numbers are only used in scientific computing where higher precision is required, but "complex numbers + FP16" does not make sense for me.
The only language (high-level, non-scripted) I know that have a first class support for Complex numbers is Fortran, where FP16 was only introduced recently by some compilers. (supporting Complex is optional in C11. Working with C++ complex doesn't feel native).

Maybe some other people use them for other purposes, who knows?

"Operations involving complex numbers in PyTorch are optimized to use vectorized assembly instructions and specialized kernels (e.g. LAPACK, cuBlas)."

https://pytorch.org/docs/stable/complex_numbers.html

[brucethemoose]

AFAIK, complex numbers are only used in scientific computing where higher precision is required, but "complex numbers + FP16" does not make sense for me.
The only language (high-level, non-scripted) I know that have a first class support for Complex numbers is Fortran, where FP16 was only introduced recently by some compilers. (supporting Complex is optional in C11. Working with C++ complex doesn't feel native).

Maybe some other people use them for other purposes, who knows?

Oh but nevermind, you are right, they only support float and double. For some reason I thought fp16 was in there too.

[jayN]

Interesting in that same document is that User Interrupts chapter has been removed.

There is now an RAO chapter in its place.

Are User Interrupts somehow being merged into Remote Atomic Operations? as in 11.8.2?