Please vote about upgrading MASM

[rudi_m]

Originally Posted by Prime95

I dread converting all those FFT macros. It's probably months of work for zero reward. I'd rather MS just did their job, but they've invested little effort in MASM over the last decade.

That said, you are right -- I should investigate the alternatives.

Beside AVX512 support and more portability there would be another benefit:

1. Supporting new CPU features as early as possible would maybe help to avoid issues like the skylake bug. Who knows, I guess Intel has added prime95 to their test-suite lately. ;)

2. IMO we have a licensing issue with masm anyways. AFAIR masm EULA only allows generated binaries it to be run in Windows operating systems.

Originally Posted by Mark Rose

I bet someone has written a compiler to convert masm to nasm.

JWASM and PoAsm seem to support masm syntax.

[bgbeuning]

I made a small (70 lines of C++ and 80 lines of lex) program
that converts MASM macro like this

fft2048K MACRO clm
pass1levels9pfa8 pass2_12_levels, 2048, clm
ENDM

fft2048Kp MACRO
pass1levels9complex4 pass2_12_levels, 2048, 2
ENDM

to an NASM macro like this

%macro fft2048K 1
pass1levels9pfa8 pass2_12_levels, 2048, %1
%endmacro

%macro fft2048Kp 0
pass1levels9complex4 pass2_12_levels, 2048, 2
%endmacro

I will hold off on doing more until George decides on a path forward.

[Anonuser]

It seems that Microsoft has implemented partial AVX-512 support in MASM. Better late than never.

[petrw1]

+3

[Batalov]

Yup. Email came:

Status update

An idea you support has been updated:

226 votes
started
+3 votes

Add support for AVX-512 in MASM
(No Status) → Started

Visual Studio Team
Visual Studio Team
Product Team, Microsoft Visual Studio

[ewmayer]

Originally Posted by Anonuser

It seems that Microsoft has implemented partial AVX-512 support in MASM. Better late than never.

Not necessarily - see my note in the Hardware/Let's-buy-GIMPS-a-KNL thread.

[tServo]

I just received an email from them about George's request to have people vote to get
it included in MASM:

=======================================================

An idea you supported has been closed. Thank you for your feedback.

3 votes have been returned to you. Give feedback on other ideas to help us decide what to build next.

265
votes
completed
+3 votes
Add support for AVX-512 in MASM
Started → Completed

As Andrew mentioned, we have implemented partial AVX-512 support in MASM and the compiler. Right now, it’s in the form of intrinsic functions and full support will come in a future VS release.

For more details check out this blog: https://blogs.msdn.microsoft.com/vcb...intel-avx-512/

We are marking this suggestion as complete and returning your votes. Please put them to good use again!

— Visual C++ Team

Visual Studio Team
Visual Studio Team
Product Team, Microsoft Visual Studio

==============================================================

Interestingly, the blog post referred to is from July !

[Mysticial]

Originally Posted by tServo

I just received an email from them about George's request to have people vote to get
it included in MASM:

=======================================================

An idea you supported has been closed. Thank you for your feedback.

3 votes have been returned to you. Give feedback on other ideas to help us decide what to build next.

265
votes
completed
+3 votes
Add support for AVX-512 in MASM
Started → Completed

As Andrew mentioned, we have implemented partial AVX-512 support in MASM and the compiler. Right now, it’s in the form of intrinsic functions and full support will come in a future VS release.

For more details check out this blog: https://blogs.msdn.microsoft.com/vcb...intel-avx-512/

We are marking this suggestion as complete and returning your votes. Please put them to good use again!

— Visual C++ Team

Visual Studio Team
Visual Studio Team
Product Team, Microsoft Visual Studio

==============================================================

Interestingly, the blog post referred to is from July !

I'll note that the AVX512 support in the compiler is uselessly broken at the moment. Within the first hour of trying it, I ran into about 5 different bugs in their implementation and optimizer.

• AVX512 register types aren't aligned to 64 bytes. So code will crash - let alone the performance penalty.
• The compiler never uses (x/y/z)mm16 - 31. It only knows the bottom 16 registers.
• For moves with blend-masking, the optimizer fails to treat the destination register as an input - thereby generating bad code.
• Some of their permute intrinsics have the wrong prototype. So stuff that compiles on GCC/ICC won't compile on MSVC.
• Using their AVX512 intrinsics leads to many spurious warnings in their header files.

I'm sure there are more. But since most of these were blockers, I couldn't test any further. Some of these require interface-breaking changes to fix. So I don't expect them to be fixed until VS2018 or VS2019. IOW, give them a few more years to get it right. But some of these *should* be fixed in 15.5 or shortly after.

These bugs are pretty egregious. And it's pretty clear that MS did not test their AVX512 intrinsics support much (if at all).

There's also an annoying bug that carries over from AVX2/FMA where the compiler never generates 213 or 132 FMA instructions when all the operands are registers. So you get tons of unnecessary reg-reg moves.

I can't speak much for the assembler though. And I assume that's the more important one for most of the people on this forum. But if you want to use AVX512 intrinsics on Windows, your only option is the Intel Compiler. AVX512 is uselessly broken on MSVC and last time I checked, MinGW is still uselessly broken for all AVX and up.

[Prime95]

I've switched to Uasm (formerly HJWasm). While it has had many bugs, the developers fixed them quickly. Right now it is pretty solid, but I don't use all the possible AVX-512 instructions.