It appears clMAGMA v1.0 is released, but it does not compile with ACML
First of all, the make.inc.acml contains mistakes. The LIBDIR lines are missing backslashes!
And tests do not compile with following: (it compiles with ATLAS instead of ACML)
g++ -O0 -DADD_ -g -Wall -DHAVE_clAmdBlas -DHAVE_CBLAS -fPIC -Xlinker -zmuldefs testing_zhetrd.o -o testing_zhetrd lin/liblapacktest.a \
-L../lib -lclmagma -lclmagmablas -L/export/modules/devel/ACML/4.4.0/gnu/gfortran64/gfortran64/lib -L/export/modules/devel/ACML/4.4.0/gnu/gfortran64/gfortran64_mp/lib -L/export/modules/devel/clAmdBlas/1.8.291/lib64 -lacml -lacml_mv -lclAmdBlas -lOpenCL
../lib/libclmagma.a(zlatrd.o): In function `magma_zlatrd':
/home/eyurtese/tmp/temp/clmagma-1.0.0/src/zlatrd.cpp:252: undefined reference to `cblas_zdotc_sub'
/home/eyurtese/tmp/temp/clmagma-1.0.0/src/zlatrd.cpp:326: undefined reference to `cblas_zdotc_sub'
collect2: error: ld returned 1 exit status
make: *** [testing_zhetrd] Error 1
make: Leaving directory `/home/eyurtese/tmp/temp/clmagma-1.0.0/testing'
make: *** [test] Error 2
Any ideas on where to report these? or would AMD contact the developers and get these problems fixed?
(in addition, I know this is not clMagma forum but I am not able to post there, it says my message is spam eh!)
It appears that the following files :
are missing the required pragma:
#pragma OPENCL EXTENSION cl_khr_fp64: enable
Therefore they are not getting built at compile time? Once clMAGMA is built, it can only run on the GPU it was built with? (I am trying to make a generic library build which can work on several different GPUs, efficiency is not an issue at this stage)
Also, I am new to clMAGMA so I wonder, what is the reason for building these at compile time? I am also getting errors like these when I try to run the examples.
Initializing clMAGMA runtime ...
clCreateProgramWithBinary failed at 833 in CL_MAGMA_RT.cpp