[vsipl++] ATLAS undefines

[Jules Bergmann]

Assem,

To follow up on our IRC discussion, I was mistaken in my earlier email 
below.

The BLAS provided in clapack/BLAS/SRC is a C implementation of the 
Fortran BLAS API.

I.e. the cdotu it provides looks like this from C/C++

	void cdotu(complex<float>* return_value,
                    int* n,
                    complex<float>* cx, int* incx,
                    complex<float>* cy, int* incy);

versus cblas_cdotu_sub, which would look like:

	void cblas_cdotu_sub(
		int n,
		complex<float>* cx, int incx,
		complex<float>* cy, int incy,
		complex<float>* return_value);

(or more accurately, void* instead of complex<float>* because C doesn't 
have complex<float> or complex when the API was created).


What we want to do is add a new way to configure VSIPL++ so that it uses 
the Fortran BLAS provided in clapack/BLAS/SRC and the Lapack provided in 
clapack/SRC.

The easiest way to do this is to have configure define 
VSIP_IMPL_USE_CBLAS to 0 when using CLAPACK's BLAS.

We don't want to break the way VSIPL++ works when it gets configured to 
use ATLAS (or MKL or ACML for that matter).  I.e. when using ATLAS, we 
should continue to have configure define VSIP_IMPL_USE_CBLAS to 1. 
Moreover, we don't want to try to use clapack/BLAS/SRC's blaswrap.h to 
abstract the difference between the Fortran and C BLAS APIs.

				-- Jules


Jules Bergmann wrote:
> Assem,
> 
> Thanks for posting this.
> 
> It looks like we're trying to use the CBLAS bindings for 
> CLAPACK/SRC/BLAS.  Unfortunately, looking at the source, it is a Fortran 
> API, with a few variances (the complex dot-product Fortran functions 
> have been converted to C "subroutines" that return the result by 
> reference).  I suspect if you tried to build other tests you would see 
> linker errors for functions like cblas_trsm, etc.
> 
> For this, we should take an approach similar to how we handled ACML:
> 
>  - Have configure define VSIP_IMPL_USE_CBLAS = 4 when using
>    CLAPACK/SRC/BLAS
> 
>  - In lapack.hpp, when VSIP_IMPL_USE_CBLAS == 4,
>     - wrap the dot-product functions to have a CBLAS interface and
>       define VSIP_IMPL_USE_CBLAS_DOT = 1.
> 
>       This should be done in a separate header file, similar to
>       acml_cblas.hpp.
> 
>     - Use Fotran API for other BLAS functions
>       (VSIP_IMPL_USE_CBLAS_OTHER = 0).
> 
> Does that sound OK?
> 
>                 -- Jules
> 
> Assem Salama wrote:
>> Everyone,
>>  As per Jule's request, this is the output of make when trying to 
>> compile convolution.cpp in the tests dir. The BLAS that I got with 
>> CLAPACK has functions similar to these but without the cblas prepended 
>> and without _sub.
>>
>> Thanks,
>> Assem Salama
>>
>>
>> ------------------------------------------------------------------------
>>
>> g++ -g -O2 -I../src -I/drive2/assem/work/checkout/vpp/tests/../src 
>> -I/include/atlas -I/include/fftw3  
>> -I/drive2/assem/work/checkout/vpp/vendor/atlas/include 
>> -I/drive2/assem/work/build/vpp_temp2/vendor/fftw/include  -o 
>> convolution.exe convolution.o -L/lib/atlas -L/lib/fftw3  
>> -L/drive2/assem/work/build/vpp_temp2/vendor/atlas/lib 
>> -L/drive2/assem/work/build/vpp_temp2/vendor/fftw/lib 
>> -L/drive2/assem/work/build/vpp_temp2/vendor/clapack 
>> -L/drive2/assem/work/build/vpp_temp2/lib -L../src/vsip -lvsip -llapack 
>> -lF77 -lcblas  -lfftw3f -lfftw3 -lfftw3l   || rm -f convolution.exe
>> convolution.o: In function `dot':
>> /drive2/assem/work/checkout/vpp/tests/../src/vsip/impl/lapack.hpp:180: 
>> undefined reference to `cblas_ddot'
>> /drive2/assem/work/checkout/vpp/tests/../src/vsip/impl/lapack.hpp:217: 
>> undefined reference to `cblas_zdotu_sub'
>> /drive2/assem/work/checkout/vpp/tests/../src/vsip/impl/lapack.hpp:179: 
>> undefined reference to `cblas_sdot'
>> /drive2/assem/work/checkout/vpp/tests/../src/vsip/impl/lapack.hpp:216: 
>> undefined reference to `cblas_cdotu_sub'
>> collect2: ld returned 1 exit status
> 
> 


-- 
Jules Bergmann
CodeSourcery
jules at codesourcery.com
(650) 331-3385 x705