Skip to content

GitLab

Commit ee761dd8 authored by Andreas Marek's avatar Andreas Marek
Browse files

Update documentation; change some autotuning levels

parent 17f5d6b1
Hide whitespace changes
Inline Side-by-side
INSTALL.md
View file @ ee761dd8
......@@ -105,7 +105,7 @@ explicitely enabled.
Please note, that it is absolutely supported that both versions of the *ELPA* library are build
and installed in the same directory.
### A) Setting of MPI compiler and libraries ###
#### A) Setting of MPI compiler and libraries ####
In the standard case *ELPA* needs a MPI compiler and MPI libraries. The configure script
will try to set this by itself. If, however, on the build system the compiler wrapper
......@@ -125,7 +125,7 @@ to the configure call.
Please continue reading at "C) Enabling GPU support"
### B) Building *ELPA* without MPI support ###
#### B) Building *ELPA* without MPI support ####
If you want to build *ELPA* without MPI support, add
......@@ -145,7 +145,7 @@ Note, that the installed *ELPA* library files will be suffixed with
Please continue reading at "C) Enabling GPU support"
### C) Enabling GPU support ###
### Enabling GPU support ###
The *ELPA* library can be build with GPU support. If *ELPA* is build with GPU
support, users can choose at RUNTIME, whether to use the GPU version or not.
......@@ -165,7 +165,7 @@ It might be necessary to also set the options (please see configure --help)
Please continue reading at "D) Enabling OpenMP support".
### D) Enabling OpenMP support ###
### Enabling OpenMP support ###
The *ELPA* library can be build with OpenMP support. This can be support of hybrid
MPI/OpenMP parallelization, since *ELPA* is build with MPI support (see A ) or only
......@@ -185,7 +185,7 @@ However, the GPU choice at runtime is not compatible with OpenMP support.
Please continue reading at "E) Standard libraries in default installation paths".
### E) Standard libraries in default installation paths###
### Standard libraries in default installation paths ###
In order to build the *ELPA* library, some (depending on the settings during the
configure step) libraries are needed.
......@@ -207,7 +207,7 @@ If your configure steps finish succcessfully, please continue at "G) Choice of E
If your configure step aborts, or you want to use libraries in non standard paths please continue at
"F) Non standard paths or non standard libraries".
### F) Non standard paths or non standard libraries ###
### Non standard paths or non standard libraries ###
If standard libraries are on the build system either installed in non standard paths, or
special non standard libraries (e.g. *Intel's MKL*) should be used, it might be necessary
......@@ -236,18 +236,26 @@ Please, for the correct link-line refer to the documentation of the correspondig
suggest the [Intel Math Kernel Library Link Line Advisor] (https://software.intel.com/en-us/articles/intel-mkl-link-line-advisor).
### G) Choice of ELPA2 compute kernels ###
### Choice of ELPA2 compute kernels ###
ELPA 2stage can be used with different implementations of compute intensive kernels, which are architecture dependent.
Some kernels (all for x86_64 architectures) are enabled by default (and must be disabled if you do not want them),
others are disabled by default and must be enabled if they are wanted.
One can enable or disable "kernel classes" by setting e.g.
One can enable "kernel classes" by setting e.g.
--enable-avx2
--enable-avx2
This will try to build all the AVX2 kernels. Please see configure --help for all options
With
--disable-avx2
one chan choose not to build the AVX2 kernels.
During the configure step all possible kernels will be printed, and whether they will be enabled or not.
It is possible to build *ELPA* with as many kernels as desired, the user can then choose at runtime which
......@@ -257,6 +265,44 @@ It this is not desired, it is possible to build *ELPA* with only one (not necess
real and complex valued case, respectively. This can be done with the "--with-fixed-real-kernel=NAME" or
"--with-fixed-complex-kernel=NAME" configure options. For details please do a "configure --help"
#### Cross compilation ####
The ELPA library does _not_ supports cross-compilation by itself, i.e. compilation of the ELPA library on an architecture wich is not
identical than the architecture ELPA should be used on.
Whenever a cross-compilation situation might occur, great care has to be taken during the build process by the user.
At the moment we see two potential pitfalls:
1.) The "build architecure" is inferior to the "target" architecture (w.r.t. the instructions sets)
In this case, at the moment, the ELPA library can only be build with instructions sets supported on the build
system. All later instruction sets will _not_ be used in the compilation. This case might lead to less optimal
performance compared to the case that ELPA is build directly on the target system.
For example, if the "build architecture" consists of an HASWELL node (supporting up to Intel's AVX2 instruction set) and the
"target architecture" is a Skylake node (supporting Intel's AVX-512 instruction set) than the AVX-512 kernels can not be build
This will lead to a performance degradation on the Skylake nodes, but is otherwise harmless (no chrashes).
2.) The "build architecure" is superior to the "target" architecture (w.r.t. the instructions sets)
This case is a critical one, since ELPA will by default build with instructions sets which are not supported on the target
system. This will lead to crashes, if during build the user does not take care to solve this issue.
For example, if the "build architecture" supports Intels' AVX-2 instruction set and the
"target architecture" does only support Intel's AVX instruction set, then by default ELPA will be build with AVX-2 instruction set
and this will also be used at runtime (since it improves the performance). However, at the moment, since the target system does not support
AVX-2 instructions this will lead to a crash.
One can avoid this unfortunate situation by disabling instructions set which are _not_ supported on the target system.
In the case above, setting
--disable-avx2
during build, will remdy this problem.
### Doxygen documentation ###
A doxygen documentation can be created with the "--enable-doxygen-doc" configure option
......
USERS_GUIDE.md
View file @ ee761dd8
......@@ -98,6 +98,10 @@ Fortran synopsis
! USERS_GUIDE.md
call e%set("solver", ELPA_SOLVER_2STAGE, success)
! set the AVX BLOCK2 kernel, otherwise ELPA_2STAGE_REAL_DEFAULT will
! be used
call e%set("real_kernel", ELPA_2STAGE_REAL_AVX_BLOCK2, success)
! use method solve to solve the eigenvalue problem to obtain eigenvalues
! and eigenvectors
! other possible methods are desribed in USERS_GUIDE.md
......@@ -141,6 +145,10 @@ C Synopsis:
USERS_GUIDE.md */
elpa_set(handle, "solver", ELPA_SOLVER_2STAGE, &error);
// set the AVX BLOCK2 kernel, otherwise ELPA_2STAGE_REAL_DEFAULT will
// be used
elpa_set(handle, "real_kernel", ELPA_2STAGE_REAL_AVX_BLOCK2, &error)
/* use method solve to solve the eigenvalue problem */
/* other possible methods are desribed in USERS_GUIDE.md */
......@@ -181,6 +189,7 @@ The following compute routines are available in *ELPA*: Please have a look at th
| hermitian_multiply | do (real) a^T x b <br> (complex) a^H x b | 20170403 |
| cholesky | do cholesky factorisation | 20170403 |
| invert_triangular | invert a upper triangular matrix | 20170403 |
| solve_tridiagonal | solve EVP for a tridiagonal matrix | 20170403 |
## IV) Using OpenMP threading ##
......@@ -243,11 +252,12 @@ There are two ways, how the user can influence the autotuning steps:
Each level defines a different set of tunable parameter. The autouning option will be extended by future releases of the *ELPA* library, at the moment the following
sets are supported:
| AUTOTUNE LEVEL | Parameters |
| :------------------- | :-------------------------------------------------- |
| ELPA_AUTOTUNE_FAST | { solver, real_kernel, complex_kernel, omp_threads } |
| ELPA_AUTOTUNE_MEDIUM | { gpu } |
| AUTOTUNE LEVEL | Parameters |
| :---------------------- | :------------------------------------------------------ |
| ELPA_AUTOTUNE_FAST | { solver, real_kernel, complex_kernel, omp_threads } |
| ELPA_AUTOTUNE_MEDIUM | all of abvoe + { gpu, partly gpu } |
| ELPA_AUTOTUNE_EXTENSIVE | all of above + { various blocking factors, stripewidth, |
| | intermediate_bandwidth } |
2.) the user can **remove** tunable parameters from the list of autotuning possibilites by explicetly setting this parameter,
e.g. if the user sets in his code
......
src/elpa_index.c
View file @ ee761dd8
......@@ -220,14 +220,14 @@ static const elpa_index_int_entry_t int_entries[] = {
INT_ENTRY("intermediate_bandwidth", "Specifies the intermediate bandwidth in ELPA2 full->banded step. Must be a multiple of nblk", 0, ELPA_AUTOTUNE_NOT_TUNABLE, ELPA_AUTOTUNE_DOMAIN_ANY,
intermediate_bandwidth_cardinality, intermediate_bandwidth_enumerate, intermediate_bandwidth_is_valid, NULL, PRINT_YES),
INT_ENTRY("blocking_in_band_to_full", "Loop blocking, default 3", 3, ELPA_AUTOTUNE_MEDIUM, ELPA_AUTOTUNE_DOMAIN_ANY,
INT_ENTRY("blocking_in_band_to_full", "Loop blocking, default 3", 3, ELPA_AUTOTUNE_EXTENSIVE, ELPA_AUTOTUNE_DOMAIN_ANY,
band_to_full_cardinality, band_to_full_enumerate, band_to_full_is_valid, NULL, PRINT_YES),
INT_ENTRY("stripewidth_real", "Stripewidth_real, default 48. Must be a multiple of 4", 48, ELPA_AUTOTUNE_EXTENSIVE, ELPA_AUTOTUNE_DOMAIN_REAL,
stripewidth_real_cardinality, stripewidth_real_enumerate, stripewidth_real_is_valid, NULL, PRINT_YES),
INT_ENTRY("stripewidth_complex", "Stripewidth_complex, default 96. Must be a multiple of 8", 96, ELPA_AUTOTUNE_EXTENSIVE, ELPA_AUTOTUNE_DOMAIN_COMPLEX,
stripewidth_complex_cardinality, stripewidth_complex_enumerate, stripewidth_complex_is_valid, NULL, PRINT_YES),
INT_ENTRY("max_stored_rows", "Maximum number of stored rows used in ELPA 1 backtransformation, default 63", 63, ELPA_AUTOTUNE_MEDIUM, ELPA_AUTOTUNE_DOMAIN_ANY,
INT_ENTRY("max_stored_rows", "Maximum number of stored rows used in ELPA 1 backtransformation, default 63", 63, ELPA_AUTOTUNE_EXTENSIVE, ELPA_AUTOTUNE_DOMAIN_ANY,
max_stored_rows_cardinality, max_stored_rows_enumerate, max_stored_rows_is_valid, NULL, PRINT_YES),
#ifdef WITH_OPENMP
INT_ENTRY("omp_threads", "OpenMP threads used in ELPA, default 1", 1, ELPA_AUTOTUNE_FAST, ELPA_AUTOTUNE_DOMAIN_ANY,
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment