GitLab

a preprocessor flag was missing

In case of single precision calculations the stripe_width needs to
be a multiple, which differs from the double precision by a factor of 2
since one needs 32 bytes alignment and the sizeof(float) and sizeof(double)
is different by a factor of two

This commit closes issue #18

The sub-kernels _8_ and _4_ were wrong
This also solves problems with single precision SSE Block 6 kernel,
since it also uses the Block 4 kernel

The correct type is "float complex" for single precision, not
"complex". Double precision should be "double complex"

This closes issue #17

It turned out that if a CPU supports SSE the already existing
test for SSE assembly instructions always passes.
However, the compilation of gcc SSE intrinic instructions might
nevertheless fail if gcc is not called with one of the options
"-msse3", "-msse4" , "-msse4.1", "-msse4.2", "-mavx", or "-mavx2"!

Obviously gcc does still not consider SSE as a standard on X86_64
Intel CPUs.

An additional configure test has been introduced, which test for
gcc intrinsic sse instructions. If this test fails, the corresponding
kernels are switched off.

The C++ kernels can be written as C kernels, which simplifies the build
procedure

In order to increase type safty all ELPA2 kernels provide now
an interface. The interfaces for the C/C++ kernels are automatically
generated during the configure step

Quite likely the FMA4 (AMD) implementation has never been tested.
There is a fishy intrinsic call, which is most likely an typo
Abort with #error at compile time

The SSE kernels with blocking of 2,4,6 (real case) and 1,2 (complex)
case are now available by default

Thus the following changes have been done
- introduce new macros in configure.ac and Makefile.am
- renmae the AVX kernels in AVX_AVX2 (they also support AVX2)
- introduce new files with SSE kernel
- introduce new kernel parameters !
- make the SSE kernels callable

The results are identical with previous kernels

- The SSE part will be available in different files.
- Specify whether AVX or AVX2 was used to build

The single precision version of the SSE assembly kernel is about 1.8
times faster than the double precision version

library

It the configure option "--enable-single-precision" is specified,
ELPA will also be build for single precision usage. The double precision
and single precision will be available at the same time with names
"solve_evp_real_1stage_double" or "solve_evp_real_1stage_single" and
so on...

This change immplied some major refactoring of the ELPA code:
1.) functions/procedures had to be renamed with suffix "_double"

2.) If necessary the same functions have to be available with suffix
"_single"

3.) Variable kind definitions have to be consistent with the
intented use

To avoid uneccessary code duplication this is done (most of the time)
with preprocessor string substitution.

The documentation has been updated.

NOT SUPPORTED are at the moment:

- single precision usage of ELPA2 with kernels, others than "generic"
  and "generic_simple"

- single precision usage of GPU

The configure flag "--enable-shared-memory-only" triggers a build
of ELPA without MPI support:

- all MPI calls are skipped (or overloaded)
- all calls to scalapack functions are replaced by the corresponding
  lapack calls
- all calls to blacs are skipped

Using ELPA without MPI gives the same results as using ELPA with 1 MPI
task!

This version is not yet optimized for performance, here and there some
unecessary copies are done.

Ths version is intended for users, who do not have MPI in their
application but still would like to use ELPA on one compute node

This commit is performance critical and has to be timed carefully.
Thus one can switch back to the old implementation. The new one,
however is more safe and better to debug

This commit might be performance critical, it has to be timed
carefully. Thus one can switch back to the old implementation.
The new one, however, is more safe and better to debug