Magma testing_dgemm program problem

[cobrabb]

Hi,
I am trying to work with Magma. I successfully installed it. At least, it reported no error messages when I installed it.

However, when I run testing_dgemm, I get magma error values of 100%. I would assume that this means that the magma array specified in the test code is zeroed out upon return.

Once again, I am getting no error messages, either in installation or in the running of the test file. Also, I don't believe that it is a problem with the machine, as one of my colleagues has it running perfectly on the same machine.

Thanks in advance,
Jacob

[mgates3]

Please copy & paste the complete input (i.e., command line) & output of your testing_dgemm run, as well as your make.inc file and any environment variables (e.g., GPU_TARGET) you may have set.

Do you get similar errors with testing_sgemm (single precision)?

-mark

[cobrabb]

Thanks for the help.

When you asked me to post my make.inc, I looked at it one more time, and I figured that I might as well try targeting all architectures.

When I did that, the problem disappeared. I thought that I had the correct architecture targeted, but apparently not.

[mgates3]

Okay.

Other testers (e.g., testing_dgetrf) would output info about your actual device and warn if it wasn't supported by the GPU_TARGET you chose. This check was inadvertently left out of the testing_*gemm files, which has been fixed for the next release.

-mark

[sokol]

Hi,

I've got the same problem.
Actually, there are two gpu cards in my machine: Tesla K20c and Quadro K2000.
For the first one the test of dgemm is ok for me (even if labeled as "failed") and for the
second one the cublas error is O(1) which is obviously a nonsens. Here the output exterps:

Code: Select all

# ./testing_dgemm -l --dev 0
MAGMA 1.5.0  compiled for CUDA capability >= 3.0
CUDA runtime 6000, driver 6000. OpenMP threads 8. 
device 0: Tesla K20c, 705.5 MHz clock, 4799.6 MB memory, capability 3.5
device 1: Quadro K2000, 954.0 MHz clock, 2047.7 MB memory, capability 3.0
Usage: ./testing_dgemm [options] [-h|--help]

If running lapack (option --lapack), MAGMA and CUBLAS error are both computed
relative to CPU BLAS result. Else, MAGMA error is computed relative to CUBLAS result.

transA = No transpose, transB = No transpose
    M     N     K   MAGMA Gflop/s (ms)  CUBLAS Gflop/s (ms)   CPU Gflop/s (ms)  MAGMA error  CUBLAS error
=========================================================================================================
 1088  1088  1088    531.11 (   4.85)     774.24 (   3.33)      4.62 ( 557.21)    5.95e-15     5.95e-15   failed
 2112  2112  2112    577.41 (  32.63)    1012.92 (  18.60)      4.68 (4028.40)    7.95e-15     7.95e-15   failed
 3136  3136  3136    579.34 ( 106.47)    1013.40 (  60.87)      4.70 (13127.65)    1.11e-14     1.11e-14   failed

(the above test is OK for both magma and cublas)

Code: Select all

# ./testing_dgemm -l --dev 1
MAGMA 1.5.0  compiled for CUDA capability >= 3.0
CUDA runtime 6000, driver 6000. OpenMP threads 8. 
device 0: Tesla K20c, 705.5 MHz clock, 4799.6 MB memory, capability 3.5
device 1: Quadro K2000, 954.0 MHz clock, 2047.7 MB memory, capability 3.0
Usage: ./testing_dgemm [options] [-h|--help]

If running lapack (option --lapack), MAGMA and CUBLAS error are both computed
relative to CPU BLAS result. Else, MAGMA error is computed relative to CUBLAS result.

transA = No transpose, transB = No transpose
    M     N     K   MAGMA Gflop/s (ms)  CUBLAS Gflop/s (ms)   CPU Gflop/s (ms)  MAGMA error  CUBLAS error
=========================================================================================================
 1088  1088  1088     27.56 (  93.45)    6362.66 (   0.40)      4.66 ( 553.25)    5.95e-15     9.98e-01   failed
 2112  2112  2112     27.85 ( 676.61)    401148.72 (   0.05)      4.70 (4008.75)    7.95e-15     9.99e-01   failed
 3136  3136  3136     27.90 (2210.83)    1340481.28 (   0.05)      4.71 (13093.26)    1.11e-14     9.99e-01   failed

(while the latter test fails for cublas).

My make.inc is

Code: Select all

#//////////////////////////////////////////////////////////////////////////////
#   -- MAGMA (version 1.5.0) --
#      Univ. of Tennessee, Knoxville
#      Univ. of California, Berkeley
#      Univ. of Colorado, Denver
#      @date September 2014
#//////////////////////////////////////////////////////////////////////////////

# GPU_TARGET contains one or more of Tesla, Fermi, or Kepler,
# to specify for which GPUs you want to compile MAGMA:
#     Tesla  - NVIDIA compute capability 1.x cards
#     Fermi  - NVIDIA compute capability 2.x cards
#     Kepler - NVIDIA compute capability 3.x cards
# The default is all, "Tesla Fermi Kepler".
# See http://developer.nvidia.com/cuda-gpus
#
#GPU_TARGET ?= Tesla Fermi Kepler
GPU_TARGET = Kepler

CC        = gcc
CXX       = g++
NVCC      = nvcc
FORT      = gfortran

ARCH      = ar
ARCHFLAGS = cr
RANLIB    = ranlib

CFLAGS    = -fPIC -O3 -fomit-frame-pointer -mavx -m64 -DADD_ -fopenmp -DMAGMA_SETAFFINITY
FFLAGS    = -fPIC -O3 -fomit-frame-pointer -mavx -m64 -DADD_
F90FLAGS  = -fPIC -O3 -fomit-frame-pointer -mavx -m64 -DADD_ -x f95-cpp-input
NVCCFLAGS = -O3 -DADD_ -Xcompiler "-fno-strict-aliasing -fPIC"
LDFLAGS   = -fPIC -fopenmp

# Depending on how ATLAS and LAPACK were compiled, you may need one or more of:
# -lifcore -ldl -lf2c -lgfortran
LIB       = -llapack -lf77blas -latlas -lcblas -lcublas -lcudart -lstdc++ -lm -lgfortran #-lifcore

# define library directories here or in your environment
LAPACKDIR ?= /usr/local/src/lapack-3.5.0
ATLASDIR  ?= /usr/local/atlas
CUDADIR   ?= /usr/local/cuda
-include make.check-atlas
-include make.check-cuda

LIBDIR    = -L$(LAPACKDIR) \
            -L$(ATLASDIR)/lib \
            -L$(CUDADIR)/lib64

INC       = -I$(CUDADIR)/include

Hoping that some could advice me how to avoid cublas errors.
Serguei.

[mgates3]

Quick answer: this is a bug in the tester, not in MAGMA itself. We'll fix it for the next release.

Details: the tester creates a cublas handle on device 0, inside TESTING_INIT(). After that, it parses the command line arguments and finds that you want to use device 1, and allocates memory there. But it still tries to run the cublasDgemm using the handle for device 0, which will obviously fail.

-mark

[sokol]

Thanks Marc. I confirm, adding :

Code: Select all

    cublasDestroy( handle );
    cudaSetDevice( opts.device );
    cublasCreate( &handle );

just after

Code: Select all

parse_opts( argc, argv, &opts );

has solved the problem:

Code: Select all

$ ./testing_dgemm -l --dev 1
MAGMA 1.5.0  compiled for CUDA capability >= 3.0
CUDA runtime 6000, driver 6000. OpenMP threads 8. 
device 0: Tesla K20c, 705.5 MHz clock, 4799.6 MB memory, capability 3.5
device 1: Quadro K2000, 954.0 MHz clock, 2047.7 MB memory, capability 3.0
Usage: ./testing_dgemm [options] [-h|--help]

If running lapack (option --lapack), MAGMA and CUBLAS error are both computed
relative to CPU BLAS result. Else, MAGMA error is computed relative to CUBLAS result.

transA = No transpose, transB = No transpose
    M     N     K   MAGMA Gflop/s (ms)  CUBLAS Gflop/s (ms)   CPU Gflop/s (ms)  MAGMA error  CUBLAS error
=========================================================================================================
 1088  1088  1088     27.56 (  93.47)      25.44 ( 101.26)      4.62 ( 557.20)    5.95e-15     5.95e-15   failed
 2112  2112  2112     28.21 ( 667.97)      28.11 ( 670.21)      4.68 (4025.88)    7.95e-15     7.95e-15   failed
 3136  3136  3136     28.01 (2202.12)      28.03 (2200.69)      4.69 (13139.79)    1.11e-14     1.11e-14   failed