Search found 918 matches

by mgates3
Wed Sep 25, 2019 11:43 am
Forum: User discussion
Topic: MAGMA 2.5.1 compilation problem on windows
Replies: 5
Views: 1274

Re: MAGMA 2.5.1 compilation problem on windows

We'll look into it, though it might take a little while as we don't have any Windows machines readily setup for testing. All our development is on Linux and macOS.
-mark
by mgates3
Fri Sep 20, 2019 12:57 pm
Forum: User discussion
Topic: LOBPCG single precision fails to converge for most problems
Replies: 2
Views: 917

Re: LOBPCG single precision fails to converge for most problems

These problems have condition numbers >= 1e6, which mean they likely won't work well in single precision, which has a precision of only 1e-7. Roughly, single precision has 7 digits of accuracy, and a problem loses log10( cond ) digits of accuracy, in this case about 6, meaning there is at most 1 cor...
by mgates3
Wed Sep 18, 2019 12:01 pm
Forum: User discussion
Topic: axpy_batched
Replies: 1
Views: 800

Re: axpy_batched

So you really want scal_batched: y_i = alpha * y_i for i = 1, ..., batch_size Is that right? Are your vectors scattered in memory, or packed into a single array? If in a single array, would a matrix scale function meet your needs, i.e., A = alpha * A for an m-by-batch_size matrix A. -mark
by mgates3
Tue Sep 17, 2019 2:33 pm
Forum: User discussion
Topic: Compilation problem of magma on ibm power 9 for olcf summit system
Replies: 5
Views: 1750

Re: Compilation problem of magma on ibm power 9 for olcf summit system

I'm confused by your link line: lessl -L/sw/summit/essl/... Did you chop off the front part of that line that shows the actual linker (say, xlc++)? From the errors, the linking appears to be missing OpenMP functions. You need to specify the same OpenMP option when linking as when compiling, e.g., fo...
by mgates3
Mon Sep 16, 2019 8:54 pm
Forum: User discussion
Topic: Compilation problem of magma on ibm power 9 for olcf summit system
Replies: 5
Views: 1750

Re: Compilation problem of magma on ibm power 9 for olcf summit system

copysignf is part of the C standard library, in math.h, not BLAS or LAPACK. The make.inc file may need -lm added to the end of LIBS for the math library. Shouldn't need any recompiling, just finish linking.
-mark
by mgates3
Fri Sep 13, 2019 2:03 pm
Forum: User discussion
Topic: Is there a step by step tutorial on installing magma?
Replies: 2
Views: 1299

Re: Is there a step by step tutorial on installing magma?

Are you following the steps using CMake in README-Windows? Can you be more specific about what happened and when the error occurred?
-mark
by mgates3
Thu Sep 05, 2019 12:50 am
Forum: User discussion
Topic: MAGMA on MAC OS
Replies: 1
Views: 1230

Re: MAGMA on MAC OS

Yes, though since Apple has stopped supporting NVIDIA cards on MacOS Mojave (10.14), anything using CUDA is limited to MacOS 10.13 and older (as I understand it). Compiling MAGMA on MacOS is similar to on Linux. I use the gcc & gfortran compilers from HPC MacOS X: http://hpc.sourceforge.net/ Follow ...
by mgates3
Mon Sep 02, 2019 1:17 pm
Forum: User discussion
Topic: nvcc "command line is too long" on windows
Replies: 4
Views: 2242

Re: nvcc "command line is too long" on windows

From a magma/build directory, try:

Code: Select all

cmake -DCMAKE_WINDOWS_EXPORT_ALL_SYMBOLS=TRUE -DBUILD_SHARED_LIBS=TRUE ..
This assumes CMake >= 3.4.
See https://blog.kitware.com/create-dlls-on ... l-feature/

-mark
by mgates3
Wed Aug 28, 2019 10:19 am
Forum: User discussion
Topic: nvcc "command line is too long" on windows
Replies: 4
Views: 2242

Re: nvcc "command line is too long" on windows

I haven't seen this, but usually use the Makefile, not CMake. When I use CMake on Windows, I use MS Visual Studio (the free version works fine). I gather you run CMake to configure MAGMA. What build system do you use after configuring with CMake: Makefile, MS Visual Studio, ...?
-mark
by mgates3
Tue Aug 20, 2019 2:31 pm
Forum: User discussion
Topic: Pinned memory for diagonalization dsygvd (Divide and conquer)
Replies: 1
Views: 1348

Re: Pinned memory for diagonalization dsygvd (Divide and conquer)

Because of the complexity of managing an array distributed across multiple GPUs, we don't currently have a version of sygvdx_m where the matrix is given on the GPUs (i.e., sygvdx_mgpu).

I do recommend trying the 2-stage version, dsygvdx_2stage_m, which is often faster.

-mark