gallo/atrip
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

Update org mode structure

Browse Source
This commit is contained in:
Alejandro Gallo 2021-12-29 13:42:31 +01:00
parent 52f17015df
commit d7699b3c9b
1 changed files with 47 additions and 47 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

94
atrip.org
View File

@ -1,16 +1,16 @@
#+title: ATRIP: An MPI-asynchronous implementation of CCSD(T) #+title: ATRIP: An MPI-asynchronous implementation of CCSD(T)
#+PROPERTY: header-args+ :noweb yes :comments noweb :mkdirp t #+PROPERTY: header-args+ :noweb yes :comments noweb :mkdirp t
* Implementation * Introduction
The algorithm uses two main data types, the =Slice= and the The algorithm uses two main data types, the =Slice= and the
=SliceUnion= as a container and resource manager of the =Slice=. =SliceUnion= as a container and resource manager of the =Slice=.
** The slice * The slice
The following section introduces the idea of a slice. The following section introduces the idea of a slice.
*** Prolog :noexport: ** Prolog :noexport:
#+begin_src c++ :tangle (atrip-slice-h) #+begin_src c++ :tangle (atrip-slice-h)
#pragma once #pragma once
#include <iostream> #include <iostream>
@ -28,7 +28,7 @@ struct Slice {
using F = double; using F = double;
#+end_src #+end_src
*** Introduction ** Introduction
A slice is the concept of a subset of values of a given tensor. A slice is the concept of a subset of values of a given tensor.
As an example, for the doubles amplitudes \( T^{ab}_{ij} \), one need two kinds of objects: As an example, for the doubles amplitudes \( T^{ab}_{ij} \), one need two kinds of objects:
@ -38,7 +38,7 @@ As an example, for the doubles amplitudes \( T^{ab}_{ij} \), one need two kinds
- the object \( \mathsf{T}(a,b)_{ij} \) which for every pair of \( a, b \) - the object \( \mathsf{T}(a,b)_{ij} \) which for every pair of \( a, b \)
corresponds the \( N_\mathrm{o}^2 \)-sized tensor \( T^{ab}_{ij} \). corresponds the \( N_\mathrm{o}^2 \)-sized tensor \( T^{ab}_{ij} \).
*** Location ** Location
Every slice set, for instance, Every slice set, for instance,
\( S_k = \left\{ \( S_k = \left\{
@ -62,7 +62,7 @@ is therefore a simple structure:
struct Location { size_t rank; size_t source; }; struct Location { size_t rank; size_t source; };
#+end_src #+end_src
*** Type ** Type
Due to the permutation operators in the equations Due to the permutation operators in the equations
it is noticeable that for every one dimensional it is noticeable that for every one dimensional
@ -112,7 +112,7 @@ a type which denotes which permutation it is.
}; };
#+end_src #+end_src
*** State ** State
Every slice can be in different states and every state Every slice can be in different states and every state
denotes which function the slice is going to provide denotes which function the slice is going to provide
@ -162,7 +162,7 @@ Again the implementation is a simple enum type.
}; };
#+end_src #+end_src
*** The Info structure ** The Info structure
Every slice has an information structure associated with it Every slice has an information structure associated with it
that keeps track of the **variable** type, state and so on. that keeps track of the **variable** type, state and so on.
@ -191,7 +191,7 @@ struct Info {
using Ty_x_Tu = std::pair< Type, PartialTuple >; using Ty_x_Tu = std::pair< Type, PartialTuple >;
#+end_src #+end_src
*** Name ** Name
CCSD(T) needs in this algorithm 5 types of tensor slices, CCSD(T) needs in this algorithm 5 types of tensor slices,
namely namely
@ -215,7 +215,7 @@ and through two parameters for the hole contribution.
}; };
#+end_src #+end_src
*** Database ** Database
The database is a simple representation of the slices of a slice union. The database is a simple representation of the slices of a slice union.
Every element of the database is given by the name of the tensor it Every element of the database is given by the name of the tensor it
@ -236,7 +236,7 @@ a vector of these elements.
using Database = LocalDatabase; using Database = LocalDatabase;
#+end_src #+end_src
*** MPI Types ** MPI Types
#+begin_src c++ :tangle (atrip-slice-h) #+begin_src c++ :tangle (atrip-slice-h)
struct mpi { struct mpi {
@ -327,7 +327,7 @@ struct mpi {
}; };
#+end_src #+end_src
*** Static utilities ** Static utilities
This section presents some functions which are useful to work with This section presents some functions which are useful to work with
slices and are inside the namespace created by the slice struct. slices and are inside the namespace created by the slice struct.
@ -493,7 +493,7 @@ static Slice& findByInfo(std::vector<Slice> &slices,
} }
#+end_src #+end_src
*** Attributes ** Attributes
A slice object does not own data, it is just a container A slice object does not own data, it is just a container
or a pointer to data together with additional bookkeeping facilities. or a pointer to data together with additional bookkeeping facilities.
@ -523,7 +523,7 @@ For practical purposes in MPI calls, the number of elements in =data= is also in
const size_t size; const size_t size;
#+end_src #+end_src
*** Member functions ** Member functions
It is important to note that a ready slice should not be recycled from It is important to note that a ready slice should not be recycled from
any other slice, so that it can have access by itself to the data. any other slice, so that it can have access by itself to the data.
@ -675,7 +675,7 @@ The main behaviour of the function should
} }
#+end_src #+end_src
*** Epilog :noexport: ** Epilog :noexport:
#+begin_src c++ :tangle (atrip-slice-h) #+begin_src c++ :tangle (atrip-slice-h)
Slice(size_t size_) Slice(size_t size_)
: info({}) : info({})
@ -688,7 +688,7 @@ The main behaviour of the function should
#+end_src #+end_src
*** Debug :noexport: ** Debug :noexport:
#+begin_src c++ :tangle (atrip-slice-h) #+begin_src c++ :tangle (atrip-slice-h)
std::ostream& operator<<(std::ostream& out, Slice::Location const& v) { std::ostream& operator<<(std::ostream& out, Slice::Location const& v) {
@ -709,10 +709,10 @@ std::ostream& operator<<(std::ostream& out, Slice::Info const& i) {
} // namespace atrip } // namespace atrip
#+end_src #+end_src
** Utils * Utils
This section presents some utilities This section presents some utilities
*** Prolog :noexport: ** Prolog :noexport:
#+begin_src c++ :tangle (atrip-utils-h) #+begin_src c++ :tangle (atrip-utils-h)
#pragma once #pragma once
#include <sstream> #include <sstream>
@ -726,7 +726,7 @@ This section presents some utilities
namespace atrip { namespace atrip {
#+end_src #+end_src
*** Pretty printing ** Pretty printing
The pretty printing uses the [[https://github.com/sharkdp/dbg-macro][dbg-macro]] package. The pretty printing uses the [[https://github.com/sharkdp/dbg-macro][dbg-macro]] package.
@ -742,7 +742,7 @@ The pretty printing uses the [[https://github.com/sharkdp/dbg-macro][dbg-macro]]
#+end_src #+end_src
*** Chrono ** Chrono
The chrono is just a simple wrapper for a high resolution clock The chrono is just a simple wrapper for a high resolution clock
that can be found in the =std::chrono= namespace of the standard library. that can be found in the =std::chrono= namespace of the standard library.
@ -768,12 +768,12 @@ using Timings = std::map<std::string, Timer>;
#+end_src #+end_src
*** Epilog :noexport: ** Epilog :noexport:
#+begin_src c++ :tangle (atrip-utils-h) #+begin_src c++ :tangle (atrip-utils-h)
} }
#+end_src #+end_src
** The rank mapping * The rank mapping
This section introduces the concept of rank mapping, This section introduces the concept of rank mapping,
which defines how slices will be allocated to every which defines how slices will be allocated to every
@ -898,7 +898,7 @@ namespace atrip {
} }
#+end_src #+end_src
** The slice union * The slice union
#+begin_src c++ :tangle (atrip-slice-union-h) #+begin_src c++ :tangle (atrip-slice-union-h)
#pragma once #pragma once
#include <atrip/Debug.hpp> #include <atrip/Debug.hpp>
@ -1464,13 +1464,13 @@ namespace atrip {
} }
#+end_src #+end_src
** Tuples * Tuples
This section introduces the types for tuples \( (a,b,c) \) This section introduces the types for tuples \( (a,b,c) \)
as well as their distribution to nodes and cores. as well as their distribution to nodes and cores.
*** Prolog :noexport: ** Prolog :noexport:
#+begin_src c++ :tangle (atrip-tuples-h) #+begin_src c++ :tangle (atrip-tuples-h)
#pragma once #pragma once
@ -1494,7 +1494,7 @@ as well as their distribution to nodes and cores.
namespace atrip { namespace atrip {
#+end_src #+end_src
*** Tuples types ** Tuples types
The main tuple types are simple type aliases for finite-size arrays. The main tuple types are simple type aliases for finite-size arrays.
A tuple is thus simply 3 natural numbers \( (a,b,c) \) A tuple is thus simply 3 natural numbers \( (a,b,c) \)
@ -1509,7 +1509,7 @@ constexpr ABCTuple FAKE_TUPLE = {0, 0, 0};
constexpr ABCTuple INVALID_TUPLE = {1, 1, 1}; constexpr ABCTuple INVALID_TUPLE = {1, 1, 1};
#+end_src #+end_src
*** Distributing the tuples ** Distributing the tuples
In general it is our task to distribute all the tuples In general it is our task to distribute all the tuples
\( (a,b,c) \) among the ranks. Every distribution should \( (a,b,c) \) among the ranks. Every distribution should
@ -1527,7 +1527,7 @@ struct TuplesDistribution {
*** Node information ** Node information
- nodeList :: - nodeList ::
List of hostnames of size \( N_n \) List of hostnames of size \( N_n \)
@ -1644,7 +1644,7 @@ getClusterInfo(MPI_Comm comm) {
} }
#+end_src #+end_src
*** Naive list ** Naive list
The naive implementation of the global tuples list is simple The naive implementation of the global tuples list is simple
three for loops creating tuples of the sort three for loops creating tuples of the sort
@ -1728,13 +1728,13 @@ struct NaiveDistribution : public TuplesDistribution {
#+end_src #+end_src
*** Group and sort list ** Group and sort list
**** Prolog :noexport: *** Prolog :noexport:
#+begin_src c++ :tangle (atrip-tuples-h) #+begin_src c++ :tangle (atrip-tuples-h)
namespace group_and_sort { namespace group_and_sort {
#+end_src #+end_src
**** Utils *** Utils
#+begin_src c++ :tangle (atrip-tuples-h) #+begin_src c++ :tangle (atrip-tuples-h)
@ -1763,7 +1763,7 @@ struct Info {
#+end_src #+end_src
**** Distribution *** Distribution
wording: home element = element which is located on the given node wording: home element = element which is located on the given node
1. we distribute the tuples such that each tuple has at least one 'home element' 1. we distribute the tuples such that each tuple has at least one 'home element'
@ -1953,7 +1953,7 @@ ABCTuples specialDistribution(Info const& info, ABCTuples const& allTuples) {
#+end_src #+end_src
**** Main *** Main
The main routine should return the list of tuples to be handled by the current rank. The main routine should return the list of tuples to be handled by the current rank.
@ -2134,7 +2134,7 @@ and the =sendCounts= vector is simply the constant vector
} }
#+end_src #+end_src
**** Interface *** Interface
The distribution interface will then simply be The distribution interface will then simply be
@ -2147,18 +2147,18 @@ struct Distribution : public TuplesDistribution {
#+end_src #+end_src
**** Epilog :noexport: *** Epilog :noexport:
#+begin_src c++ :tangle (atrip-tuples-h) #+begin_src c++ :tangle (atrip-tuples-h)
} // namespace group_and_sort } // namespace group_and_sort
#+end_src #+end_src
*** Epilog :noexport: ** Epilog :noexport:
#+begin_src c++ :tangle (atrip-tuples-h) #+begin_src c++ :tangle (atrip-tuples-h)
} }
#+end_src #+end_src
** Unions * Unions
Every slice pertaining to every different tensor Every slice pertaining to every different tensor
is sliced differently. is sliced differently.
@ -2401,7 +2401,7 @@ namespace atrip {
#+end_src #+end_src
** Equations * Equations
#+begin_src c++ :tangle (atrip-equations-h) #+begin_src c++ :tangle (atrip-equations-h)
#pragma once #pragma once
@ -2735,7 +2735,7 @@ namespace atrip {
} }
#+end_src #+end_src
** Blas * Blas
The main matrix-matrix multiplication method used in this algorithm The main matrix-matrix multiplication method used in this algorithm
is mainly using the =DGEMM= function, which we declare as is mainly using the =DGEMM= function, which we declare as
=extern= since it should be known only at link-time. =extern= since it should be known only at link-time.
@ -2762,8 +2762,8 @@ namespace atrip {
} }
#+end_src #+end_src
** Atrip * Atrip
*** Header ** Header
#+begin_src c++ :tangle (atrip-atrip-h) #+begin_src c++ :tangle (atrip-atrip-h)
#pragma once #pragma once
#include <sstream> #include <sstream>
@ -2836,7 +2836,7 @@ namespace atrip {
#undef ADD_ATTRIBUTE #undef ADD_ATTRIBUTE
#+end_src #+end_src
*** Main ** Main
#+begin_src c++ :tangle (atrip-atrip-cxx) #+begin_src c++ :tangle (atrip-atrip-cxx)
#include <iomanip> #include <iomanip>
@ -3431,8 +3431,8 @@ Atrip::Output Atrip::run(Atrip::Input const& in) {
#+end_src #+end_src
** Debug and Logging * Debug and Logging
*** Macros ** Macros
#+begin_src c++ :tangle (atrip-debug-h) #+begin_src c++ :tangle (atrip-debug-h)
#pragma once #pragma once
@ -3515,7 +3515,7 @@ define =ATRIP_NO_OUTPUT=
#endif #endif
#+end_src #+end_src
*** Iteration informer ** Iteration informer
In general a code writer will want to write some messages in every iteration. In general a code writer will want to write some messages in every iteration.
A developer then can register a function to be used in this sense. A developer then can register a function to be used in this sense.
@ -3540,7 +3540,7 @@ namespace atrip {
} }
#+end_src #+end_src
** Include header * Include header
#+begin_src c++ :tangle (atrip-main-h) #+begin_src c++ :tangle (atrip-main-h)
#pragma once #pragma once