AVX2: no such instruction

[txje]

Hey Jeff,
This looks like a phenomenal piece of work and I'd love to try it out. I am running into problems during the build. For reference, I'm trying to run this on a RHEL cluster with potentially heterogeneous architectures. I get the following error:

make
...
CCLD libparasail_sse41_table.la
CC src/libparasail_avx2_memory_la-memory_avx2.lo
/tmp/cc4hi1vp.s: Assembler messages:
/tmp/cc4hi1vp.s:46: Error: no such instruction: vmovdqa %ymm0,(%rdi,%rax)' /tmp/cc4hi1vp.s:56: Error: no such instruction:vzeroupper'

I have little to no experience with vector programming, so I'm afraid I don't know where to start troubleshooting. Any ideas you might have will be helpful.

Jeremy

[jeffdaily]

I'm looking at this right now.

I was trying to be flexible with the compilation and let some of the advanced compilers go ahead and generate code for AVX2 even if the build system did not support running such code. I was trying to consider an RPM maintainer wanting to produce a single RPM that supported all architectures.

If you could send me your config.log to my [email protected] email (I don't think github issues allow for attachments?) then I can see how the configure tests performed for the AVX2 portion. It is interesting that configure determined your compiler supported AVX2 -- I probably did not test thoroughly enough.

As a short-term work-around, you can disable the compilation of AVX2 code (or any unsupported instruction set) using the following configure parameters. They should be pretty much self-documenting. I use the parameter "choke", though any nonsense flag should work so long as a compiler will barf when trying to process it.

configure SSE2_FLAGS=choke SSE41_FLAGS=choke AVX2_CFLAGS=choke

The above parameters correspond to the following portion of the configure.ac file:

AC_MSG_CHECKING([for SSE2 flags])
AS_IF([test "x$SSE2_CFLAGS" = x],
      [AS_CASE(["$vendor"],
               [clang],    [SSE2_CFLAGS="-msse2"],
               [gnu],      [SSE2_CFLAGS="-march=core2"],
               [intel],    [SSE2_CFLAGS="-march=core2"])])
AC_MSG_RESULT([$SSE2_CFLAGS])
AC_SUBST([SSE2_CFLAGS])

AC_MSG_CHECKING([for SSE4.1 flags])
AS_IF([test "x$SSE41_CFLAGS" = x],
      [AS_CASE(["$vendor"],
               [clang],    [SSE41_CFLAGS="-msse4"],
               [gnu],      [SSE41_CFLAGS="-march=corei7"],
               [intel],    [SSE41_CFLAGS="-march=corei7"])])
AC_MSG_RESULT([$SSE41_CFLAGS])
AC_SUBST([SSE41_CFLAGS])

AC_MSG_CHECKING([for AVX2 flags])
AS_IF([test "x$AVX2_CFLAGS" = x],
      [AS_CASE(["$vendor"],
               [clang],    [AVX2_CFLAGS="-mavx2"],
               [gnu],      [AVX2_CFLAGS="-march=core-avx2"],
               [intel],    [AVX2_CFLAGS="-march=core-avx2"])])
AC_MSG_RESULT([$AVX2_CFLAGS])
AC_SUBST([AVX2_CFLAGS])

[jeffdaily]

I just pushed 45e2b22 which adds the offending code you came across to the configure test for AVX2 support. Without using my work-around above, please try this commit and rerun configure.

I checked it on my macbook using gcc 4.8 from macports and I got the configure test to fail with the same error message you reported -- so I think it works now, correctly detecting AVX2 as not available.

[txje]

Perfect. My build completed successfully with 45e2b22. I'll keep you posted as I try to run some sequences through.