Re: [eigen] PGI compilation error

[Benoit Jacob <jacob.benoit.1@xxxxxxxxx>]

OK so it's looking like getting SSE support on PGI is going to require
a bit of work.

Meanwhile you can test Eigen without SSE and without any 16-byte
alignment stuff by defining EIGEN_DONT_ALIGN before you #include
Eigen.

Another solution is to do the EIGEN_HAS_MM_MALLOC change and then
compile with EIGEN_DONT_VECTORIZE.

EIGEN_DONT_VECTORIZE disables the use of SSE intrinsics but still
tries to do 16-byte aligned malloc's to preserve ABI compatibility.
EIGEN_DONT_ALIGN breaks ABI compatibility.

Benoit


2010/2/23 Philippe Marti <philippe.marti@xxxxxxxxx>:
> Well changing the EIGEN_HAS_MM_MALLOC, removed the first error but
> then fails on   _mm_cvtss_f32  and _mm_cvtsd_f64 :
>
> "/users/pmarti/tmp/ei3/eigen/Eigen/src/Core/arch/SSE/PacketMath.h",
> line 247: error:
>          identifier "_mm_cvtss_f32" is undefined
>  template<> EIGEN_STRONG_INLINE float  ei_pfirst<Packet4f>(const
> Packet4f& a) { return _mm_cvtss_f32(a); }
>
>                 ^
>
> "/users/pmarti/tmp/ei3/eigen/Eigen/src/Core/arch/SSE/PacketMath.h",
> line 248: error:
>          identifier "_mm_cvtsd_f64" is undefined
>  template<> EIGEN_STRONG_INLINE double ei_pfirst<Packet2d>(const
> Packet2d& a) { return _mm_cvtsd_f64(a); }
>
>                 ^
>
> "/users/pmarti/tmp/ei3/eigen/Eigen/src/Core/DiagonalProduct.h", line 40: error:
>          expression must be an lvalue or a function designator
>            | (PacketAccessBit & (unsigned int)(MatrixType::Flags) &
> (unsigned int)(DiagonalType::DiagonalVectorType::Flags)),
>
>
> As I'm concerned, I do not absolutely need this to work. I merely
> wanted to compare code performance between different compilers.
>  And as I've now been told, GCC compiled code seems to be faster anyway..
>
> So I'll be glad to provide some more information if required, but else
> I'm sure you have many more important things to do ;)
>
> Regards,
> Philippe
>
>
> On Tue, Feb 23, 2010 at 3:07 PM, Benoit Jacob <jacob.benoit.1@xxxxxxxxx> wrote:
>> Sorry, I'm not sure I'm following correctly.
>>
>> Is the current problem that _mm_malloc is not found?
>>
>> If yes, then the easy fix is to tell Eigen to not try using it. There
>> are other solutions for the same need.
>>
>> Just edit Eigen/src/Core/util/Memory.h, find this:
>>
>> #ifdef EIGEN_VECTORIZE_SSE
>>  #define EIGEN_HAS_MM_MALLOC 1
>> #else
>>  #define EIGEN_HAS_MM_MALLOC 0
>> #endif
>>
>> and replace the '1' by '0' to tell Eigen not to use it.
>>
>> In the worst case Eigen can always use ei_handmade_aligned_malloc.
>>
>> Cheers,
>> Benoit
>>
>> 2010/2/23 Hauke Heibel <hauke.heibel@xxxxxxxxxxxxxx>:
>>> The revision number is correct - you are in the development branch
>>> where I applied the fix.
>>>
>>> Regarding PGI, I don't have the compiler myself and right now I am
>>> unable to find out on the web which methods are used for aligned
>>> memory management (src\Core\util\Memory.h,
>>> free/_mm_free/_aligned_free,...). As soon as you can provide us with
>>> the information I can easily patch Eigen.
>>>
>>> Until then, the only chance to compile Eigen is disabling
>>> vectorization by defining EIGEN_DONT_VECTORIZE. Well, assuming there
>>> don't occur any other errors... ;)
>>>
>>> - Hauke
>>>
>>> On Tue, Feb 23, 2010 at 2:51 PM, Philippe Marti
>>> <philippe.marti@xxxxxxxxx> wrote:
>>>> Hello,
>>>>
>>>> I hope that's the right information, I'm not used to mercurial:
>>>>
>>>> hg identify:
>>>>    a2598077ebca tip
>>>>
>>>>
>>>> For the PGI compiler:
>>>> pgCC 9.0-4 64-bit
>>>>
>>>> Regards,
>>>> Philippe
>>>> On Tue, Feb 23, 2010 at 1:49 PM, Aron Ahmadia <aja2111@xxxxxxxxxxxx> wrote:
>>>>> Hi Philippe,
>>>>>
>>>>> Can you verify the hg version of Eigen that you are working from and
>>>>> the error again?  I don't have access to a PGI compiler here but you
>>>>> may as well note the version number of the compiler :)
>>>>>
>>>>> A
>>>>>
>>>>> On Tue, Feb 23, 2010 at 1:07 PM, Philippe Marti
>>>>> <philippe.marti@xxxxxxxxx> wrote:
>>>>>> Well ok, compiling the tests works fine with and the first tests
>>>>>> passed (interrupted afterwards as I'm not supposed to run to heavy
>>>>>> stuff on the login nodes)
>>>>>>
>>>>>> And I get the error posted previously for PGI.
>>>>>>
>>>>>> Oh and it also works for Pathscale. Just PGI is complaining.
>>>>>>
>>>>>> Thanks,
>>>>>> Philippe
>>>>>>
>>>>>> On Tue, Feb 23, 2010 at 12:41 PM, Philippe Marti
>>>>>> <philippe.marti@xxxxxxxxx> wrote:
>>>>>>> Oh, Wait before you start digging to deep. It's a little bit messy
>>>>>>> here to get to load the right compilers/library etc on the cluster. I
>>>>>>> might have forgotten something.
>>>>>>>
>>>>>>> I'll check that the gcc version is working
>>>>>>>
>>>>>>> regards,
>>>>>>> Philippe
>>>>>>>
>>>>>>> On Tue, Feb 23, 2010 at 12:38 PM, Hauke Heibel
>>>>>>> <hauke.heibel@xxxxxxxxxxxxxx> wrote:
>>>>>>>> This error is a little bit odd. It means that you have defined
>>>>>>>> EIGEN_VECTORIZE_SSE (see src/core/util/Memory.h) and thus you should
>>>>>>>> have included 'xmmintrin.h' (see Core) where (at least AFAIK) _mm_free
>>>>>>>> and _mm_malloc are declared as external or included via 'malloc.h'..
>>>>>>>>
>>>>>>>> As I said, I am a little bit at a loss...
>>>>>>>>
>>>>>>>> - Hauke
>>>>>>>>
>>>>>>>> On Tue, Feb 23, 2010 at 12:21 PM, Philippe Marti
>>>>>>>> <philippe.marti@xxxxxxxxx> wrote:
>>>>>>>>> The default branch is the 3.0 one right?
>>>>>>>>>
>>>>>>>>> I tried the to compile the tests with PGI, here are the first errors
>>>>>>>>> (not sure about what you need):
>>>>>>>>>
>>>>>>>>> "/users/pmarti/tmp/ei3/eigen/Eigen/src/Core/util/Memory.h", line 94: error:
>>>>>>>>>          identifier "_mm_malloc" is undefined
>>>>>>>>>      result = _mm_malloc(size, 16);
>>>>>>>>>               ^
>>>>>>>>>
>>>>>>>>> "/users/pmarti/tmp/ei3/eigen/Eigen/src/Core/util/Memory.h", line 165: error:
>>>>>>>>>          identifier "_mm_free" is undefined
>>>>>>>>>      _mm_free(ptr);
>>>>>>>>>      ^
>>>>>>>>>
>>>>>>>>> "/users/pmarti/tmp/ei3/eigen/Eigen/src/Core/arch/SSE/PacketMath.h",
>>>>>>>>> line 142: warning:
>>>>>>>>>          controlling expression is constant
>>>>>>>>>  { ei_assert(false && "packet integer division are not supported by SSE");
>>>>>>>>>    ^
>>>>>>>>>
>>>>>>>>> "/users/pmarti/tmp/ei3/eigen/Eigen/src/Core/arch/SSE/PacketMath.h",
>>>>>>>>> line 247: error:
>>>>>>>>>          identifier "_mm_cvtss_f32" is undefined
>>>>>>>>>  template<> EIGEN_STRONG_INLINE float  ei_pfirst<Packet4f>(const
>>>>>>>>> Packet4f& a) { return _mm_cvtss_f32(a); }
>>>>>>>>>
>>>>>>>>>                 ^
>>>>>>>>>
>>>>>>>>> "/users/pmarti/tmp/ei3/eigen/Eigen/src/Core/arch/SSE/PacketMath.h",
>>>>>>>>> line 248: error:
>>>>>>>>>          identifier "_mm_cvtsd_f64" is undefined
>>>>>>>>>  template<> EIGEN_STRONG_INLINE double ei_pfirst<Packet2d>(const
>>>>>>>>> Packet2d& a) { return _mm_cvtsd_f64(a); }
>>>>>>>>>
>>>>>>>>> Philippe
>>>>>>>>>
>>>>>>>>> On Tue, Feb 23, 2010 at 11:38 AM, Hauke Heibel
>>>>>>>>> <hauke.heibel@xxxxxxxxxxxxxx> wrote:
>>>>>>>>>> I tried to fix the aligned issue in the default branch. You may try to
>>>>>>>>>> grab that version and give it a shot.
>>>>>>>>>>
>>>>>>>>>> - Hauke
>>>>>>>>>>
>>>>>>>>>> On Tue, Feb 23, 2010 at 11:23 AM, Philippe Marti
>>>>>>>>>> <philippe.marti@xxxxxxxxx> wrote:
>>>>>>>>>>> PGI: pgCC 9.0-4 64-bit target on x86-64 Linux
>>>>>>>>>>>
>>>>>>>>>>>
>>>>>>>>>>> "/users/pmarti/Documents/Code/git/FSDynamoCode/External/eigen/Eigen/src/Core/util/Macros.h",
>>>>>>>>>>> line 193: catastrophic error:
>>>>>>>>>>>          #error directive: Please tell me what is the equivalent of
>>>>>>>>>>>          __attribute__((aligned(16))) for your compiler
>>>>>>>>>>>  #error Please tell me what is the equivalent of
>>>>>>>>>>> __attribute__((aligned(16))) for your compiler
>>>>>>>>>>>   ^
>>>>>>>>>>>
>>>>>>>>>>> 1 catastrophic error detected in the compilation of
>>>>>>>>>>> "/users/pmarti/Documents/Code/git/FSDynamoCode/Programs/ThermalConvectionRun.cpp".
>>>>>>>>>>> Compilation terminated.
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>
>>>>>>>>>
>>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>>
>>>>>
>>>>>
>>>>
>>>>
>>>>
>>>
>>>
>>>
>>
>>
>>
>
>
>