108 KiB
108 KiB
ATRIP: An MPI-asynchronous implementation of CCSD(T)
- Implementation
Implementation
The algorithm uses two main data types, the Slice and the
SliceUnion as a container and resource manager of the Slice.
The slice
The following section introduces the idea of a slice.
Introduction
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:
- the object \( \mathsf{T}(a)^b_{ij} \) which for every \( a \) gets assigned the tensor \( T^{ab}_{ij} \) of size \( N_\mathrm{o}^2 \times N_\mathrm{v} \)
- the object \( \mathsf{T}(a,b)_{ij} \) which for every pair of \( a, b \) corresponds the \( N_\mathrm{o}^2 \)-sized tensor \( T^{ab}_{ij} \).
Location
Every slice set, for instance,
\( S_k = \left\{
a \mapsto \mathsf{T}(a)^{b}_{ij}
\mid
a \in A_k
\right\} \)
where \( A_k \) is some subset of
\( \mathsf{N}_\mathrm{v} \),
gets stored in some rank \( k \).
In general however, the number of elements in \( A_k \) can be bigger
than the number of processes \( n_p \). Therefore in order to uniquely
indentify a given slice in \( S_k \) we need two identifiers,
the rank \( k \), which tells us in which core's memory the slice is
allocated, and an additional tag which we will call source.
The datatype that simply models this state of affairs is therefore a simple structure:
struct Location { size_t rank; size_t source; };Type
Due to the permutation operators in the equations it is noticeable that for every one dimensional slice and triple \( (a,b,c) \)
\begin{equation*} a \mapsto \mathsf{t}(a) \end{equation*}one needs at the same time \( \mathsf{t}(a) \), \( \mathsf{t}(b) \) and \( \mathsf{t}(c) \). For two dimensional slices, i.e., slices of the form
\begin{equation*} (a,b) \mapsto \mathsf{t}(a,b) \end{equation*}one needs in the equations the slices \( \mathsf{t}(a,b) \), \( \mathsf{t}(b,c) \) and \( \mathsf{t}(a,c) \). In addition, in the case of diagrams where the integral \( V^{ab}_{ci} \) appears, we additionaly need the permuted slices from before, i.e. \( \mathsf{t}(b,a) \), \( \mathsf{t}(c,b) \) and \( \mathsf{t}(c,a) \).
This means, every slice has associated with it a type which denotes which permutation it is.
enum Type
{ A = 10
, B
, C
// Two-parameter slices
, AB = 20
, BC
, AC
// for abci and the doubles
, CB
, BA
, CA
// The non-typed slice
, Blank = 404
};State
Every slice can be in different states and every state denotes which function the slice is going to provide and which relations they have between themselves.
- Fetch
-
A slice is in state
Fetchwhen it has a valid data pointer that must be written to. AFetchslice should not live very long, this means that after the database send and receive phase,Fetchslices should be changed intoDispatchedin order to start the process of writing to the data pointer from some other rank. - Dispatched
-
A
Dispatchedslice indicates that at some point send and receive MPI calls have been dispatched in order to get the data. However, the calls have just been dispatched and there is no warranty for the data to be there, for that, the slice must be unwrapped. - Ready
-
Readymeans that the data pointer can be read from directly. - SelfSufficient
-
A slice is
SelfSufficientwhen its contents are located in the same rank that it lives, so that it does not have to fetch from no other rank. This is important in order to handle the data pointers correctly and in order to save calls to MPI receive and send functions. - Recycled
-
Recycledmeans that this slice gets its data pointer from another slice, so it should not be written to - Acceptor
-
Acceptormeans that the slice can accept a new slice, it is the counterpart of theBlanktype, but for states
Again the implementation is a simple enum type.
enum State {
Fetch = 0,
Dispatched = 2,
Ready = 1,
SelfSufficient = 911,
Recycled = 123,
Acceptor = 405
};The Info structure
Every slice has an information structure associated with it that keeps track of the variable type, state and so on.
struct Info {
// which part of a,b,c the slice holds
PartialTuple tuple;
// The type of slice for the user to retrieve the correct one
Type type;
// What is the state of the slice
State state;
// Where the slice is to be retrieved
Location from;
// If the data are actually to be found in this other slice
Type recycling;
Info() : tuple{0,0}
, type{Blank}
, state{Acceptor}
, from{0,0}
, recycling{Blank}
{}
};
using Ty_x_Tu = std::pair< Type, PartialTuple >;Name
CCSD(T) needs in this algorithm 5 types of tensor slices, namely \( V^{ij}_{ka} \), \( V^{ab}_{ci} \), \( V^{ab}_{ij} \) and two times \( T^{ab}_{ij} \). The reason why we need two times the doubles amplitudes is because in the doubles contribution to the energy, the \( T \) amplidutes will be sliced through one parameter for the particle contribution and through two parameters for the hole contribution.
enum Name
{ TA = 100
, VIJKA = 101
, VABCI = 200
, TABIJ = 201
, VABIJ = 202
};Database
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 represents and the internal information structure.
struct LocalDatabaseElement {
Slice::Name name;
Slice::Info info;
};A local database (of a given rank) and the global database is thus simply a vector of these elements.
using LocalDatabase = std::vector<LocalDatabaseElement>;
using Database = LocalDatabase;MPI Types
struct mpi {
static MPI_Datatype vector(size_t n, MPI_Datatype const& DT) {
MPI_Datatype dt;
MPI_Type_vector(n, 1, 1, DT, &dt);
MPI_Type_commit(&dt);
return dt;
}
static MPI_Datatype sliceLocation () {
constexpr int n = 2;
// create a sliceLocation to measure in the current architecture
// the packing of the struct
Slice::Location measure;
MPI_Datatype dt;
const std::vector<int> lengths(n, 1);
const MPI_Datatype types[n] = {usizeDt(), usizeDt()};
// measure the displacements in the struct
size_t j = 0;
MPI_Aint displacements[n];
MPI_Get_address(&measure.rank, &displacements[j++]);
MPI_Get_address(&measure.source, &displacements[j++]);
for (size_t i = 1; i < n; i++) displacements[i] -= displacements[0];
displacements[0] = 0;
MPI_Type_create_struct(n, lengths.data(), displacements, types, &dt);
MPI_Type_commit(&dt);
return dt;
}
static MPI_Datatype enumDt() { return MPI_INT; }
static MPI_Datatype usizeDt() { return MPI_UINT64_T; }
static MPI_Datatype sliceInfo () {
constexpr int n = 5;
MPI_Datatype dt;
Slice::Info measure;
const std::vector<int> lengths(n, 1);
const MPI_Datatype types[n]
= { vector(2, usizeDt())
, enumDt()
, enumDt()
, sliceLocation()
, enumDt()
};
// create the displacements from the info measurement struct
size_t j = 0;
MPI_Aint displacements[n];
MPI_Get_address(measure.tuple.data(), &displacements[j++]);
MPI_Get_address(&measure.type, &displacements[j++]);
MPI_Get_address(&measure.state, &displacements[j++]);
MPI_Get_address(&measure.from, &displacements[j++]);
MPI_Get_address(&measure.recycling, &displacements[j++]);
for (size_t i = 1; i < n; i++) displacements[i] -= displacements[0];
displacements[0] = 0;
MPI_Type_create_struct(n, lengths.data(), displacements, types, &dt);
MPI_Type_commit(&dt);
return dt;
}
static MPI_Datatype localDatabaseElement () {
constexpr int n = 2;
MPI_Datatype dt;
LocalDatabaseElement measure;
const std::vector<int> lengths(n, 1);
const MPI_Datatype types[n]
= { enumDt()
, sliceInfo()
};
// measure the displacements in the struct
size_t j = 0;
MPI_Aint displacements[n];
MPI_Get_address(&measure.name, &displacements[j++]);
MPI_Get_address(&measure.info, &displacements[j++]);
for (size_t i = 1; i < n; i++) displacements[i] -= displacements[0];
displacements[0] = 0;
MPI_Type_create_struct(n, lengths.data(), displacements, types, &dt);
MPI_Type_commit(&dt);
return dt;
}
};Static utilities
This section presents some functions which are useful to work with slices and are inside the namespace created by the slice struct.
The function subtupleBySlice gives to every Slice::Type
its meaning in terms of the triples \( (a,b,c) \).
Notice that since in general the relation
\( a < b < c \) holds (in our implementation), the case
of one-dimensional parametrizations A, B and C is well
defined.
The function should only throw if there is an implementation
error where the Slice::Type enum has been expanded and this
function has not been updated accordingly.
static
PartialTuple subtupleBySlice(ABCTuple abc, Type sliceType) {
switch (sliceType) {
case AB: return {abc[0], abc[1]};
case BC: return {abc[1], abc[2]};
case AC: return {abc[0], abc[2]};
case CB: return {abc[2], abc[1]};
case BA: return {abc[1], abc[0]};
case CA: return {abc[2], abc[0]};
case A: return {abc[0], 0};
case B: return {abc[1], 0};
case C: return {abc[2], 0};
default: throw "Switch statement not exhaustive!";
}
}In the context of cleaning up slices during the main loop, it is important to check if a given slice has some slices referencing to it in quality of recycled slices.
This function should therefore return a vector of pointers of slices referencing to the given slice's info, when the length of the vector is zero, then there are no dangling links.
static std::vector<Slice*> hasRecycledReferencingToIt
( std::vector<Slice> &slices
, Info const& info
) {
std::vector<Slice*> result;
for (auto& s: slices)
if ( s.info.recycling == info.type
&& s.info.tuple == info.tuple
&& s.info.state == Recycled
) result.push_back(&s);
return result;
}
The rest of the coming functions are utilities in order to find in a vector
of slices a given slice by reference. Mostly they are merely convenience
wrappers to the standard library function std::find_if.
They are named as find<...>, where <...> represents some condition
and must always return a reference to the found slice, i.e., Slice&.
Atrip relies on these functions to find the sought for slices,
therefore these functions will throw a std::domain_error if the
given slice could not be found.
static Slice& findOneByType(std::vector<Slice> &slices, Slice::Type type) {
const auto sliceIt
= std::find_if(slices.begin(), slices.end(),
[&type](Slice const& s) {
return type == s.info.type;
});
WITH_CRAZY_DEBUG
WITH_RANK
<< "__slice__:find:looking for " << type << "\n";
if (sliceIt == slices.end())
throw std::domain_error("Slice by type not found!");
return *sliceIt;
}static Slice&
findRecycledSource (std::vector<Slice> &slices, Slice::Info info) {
const auto sliceIt
= std::find_if(slices.begin(), slices.end(),
[&info](Slice const& s) {
return info.recycling == s.info.type
&& info.tuple == s.info.tuple
&& State::Recycled != s.info.state
;
});
WITH_CRAZY_DEBUG
WITH_RANK << "__slice__:find: recycling source of "
<< pretty_print(info) << "\n";
if (sliceIt == slices.end())
throw std::domain_error( "Slice not found: "
+ pretty_print(info)
+ " rank: "
+ pretty_print(Atrip::rank)
);
WITH_RANK << "__slice__:find: " << pretty_print(sliceIt->info) << "\n";
return *sliceIt;
}static Slice& findByTypeAbc
( std::vector<Slice> &slices
, Slice::Type type
, ABCTuple const& abc
) {
const auto tuple = Slice::subtupleBySlice(abc, type);
const auto sliceIt
= std::find_if(slices.begin(), slices.end(),
[&type, &tuple](Slice const& s) {
return type == s.info.type
&& tuple == s.info.tuple
;
});
WITH_CRAZY_DEBUG
WITH_RANK << "__slice__:find:" << type << " and tuple "
<< pretty_print(tuple)
<< "\n";
if (sliceIt == slices.end())
throw std::domain_error( "Slice not found: "
+ pretty_print(tuple)
+ ", "
+ pretty_print(type)
+ " rank: "
+ pretty_print(Atrip::rank)
);
return *sliceIt;
}static Slice& findByInfo(std::vector<Slice> &slices,
Slice::Info const& info) {
const auto sliceIt
= std::find_if(slices.begin(), slices.end(),
[&info](Slice const& s) {
// TODO: maybe implement comparison in Info struct
return info.type == s.info.type
&& info.state == s.info.state
&& info.tuple == s.info.tuple
&& info.from.rank == s.info.from.rank
&& info.from.source == s.info.from.source
;
});
WITH_CRAZY_DEBUG
WITH_RANK << "__slice__:find:looking for " << pretty_print(info) << "\n";
if (sliceIt == slices.end())
throw std::domain_error( "Slice by info not found: "
+ pretty_print(info));
return *sliceIt;
}Attributes
A slice object does not own data, it is just a container or a pointer to data together with additional bookkeeping facilities.
It includes an info structure with the information about the slice,
Type, State etc, which will be later communicated to other ranks.
Info info;
A pointer to data is also necessary for the Slice but not necessary
to be communicated to other ranks. The Slice should never allocate
or deallocate itself the pointer.
F *data;
An MPI_Request handle is also included so that the slices that are
to receive data through MPI can know which request they belong to.
MPI_Request request;
For practical purposes in MPI calls, the number of elements in data is also included.
const size_t size;Member functions
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.
void markReady() noexcept {
info.state = Ready;
info.recycling = Blank;
}
The following function asks wether or not
the slice has effectively been unwrapped or not,
i.e., wether or not the data are accessible and already
there. This can only happen in two ways, either
is the slice Ready or it is SelfSufficient,
i.e., the data pointed to was pre-distributed to the current node.
bool isUnwrapped() const noexcept {
return info.state == Ready
|| info.state == SelfSufficient
;
}
The function isUnwrappable answers which slices can be unwrapped
potentially. Unwrapped slices can be unwrapped again idempotentially.
Also Recycled slices can be unwrapped, i.e. the slices pointed to by them
will be unwrapped.
The only other possibility is that the slice has been dispatched
in the past and can be unwrapped. The case where the state
is Dispatched is the canonical intuitive case where a real process
of unwrapping, i.e. waiting for the data to get through the network,
is done.
bool isUnwrappable() const noexcept {
return isUnwrapped()
|| info.state == Recycled
|| info.state == Dispatched
;
}
inline bool isDirectlyFetchable() const noexcept {
return info.state == Ready || info.state == Dispatched;
}
void free() noexcept {
info.tuple = {0, 0};
info.type = Blank;
info.state = Acceptor;
info.from = {0, 0};
info.recycling = Blank;
data = nullptr;
}
inline bool isFree() const noexcept {
return info.tuple == PartialTuple{0, 0}
&& info.type == Blank
&& info.state == Acceptor
&& info.from.rank == 0
&& info.from.source == 0
&& info.recycling == Blank
&& data == nullptr
;
}
The function isRecylable answers the question, which slices can be recycled.
A slice can only be recycled if it is Fetch or Ready and has a valid datapointer.
In particular, SelfSufficient are not recyclable, since it is easier just to create a SelfSufficient slice than deal with data dependencies.
Furthermore, a recycled slice is not recyclable, if this is the case then it is either bad design or a bug.
inline bool isRecyclable() const noexcept {
return ( info.state == Dispatched
|| info.state == Ready
|| info.state == Fetch
)
&& hasValidDataPointer()
;
}
The function hasValidDataPointer describes if a slice has a valid
data pointer.
This is important to know if the slice has some data to it, also
some structural checks are done, so that it should not be Acceptor
or Blank, if this is the case then it is a bug.
inline bool hasValidDataPointer() const noexcept {
return data != nullptr
&& info.state != Acceptor
&& info.type != Blank
;
}
The function
unwrapAndMarkReady
calls the low-level MPI functions
in order to wait whenever the state of the slice is correct.
The main behaviour of the function should
- return if state is
Ready, since then there is nothing to be done. - throw if the state is not
Dispatched, only a dispatched slice can be unwrapped through MPI. - throw if an MPI error happens.
void unwrapAndMarkReady() {
if (info.state == Ready) return;
if (info.state != Dispatched)
throw
std::domain_error("Can't unwrap a non-ready, non-dispatched slice!");
markReady();
MPI_Status status;
#ifdef HAVE_OCD
WITH_RANK << "__slice__:mpi: waiting " << "\n";
#endif
const int errorCode = MPI_Wait(&request, &status);
if (errorCode != MPI_SUCCESS)
throw "MPI ERROR HAPPENED....";
#ifdef HAVE_OCD
char errorString[MPI_MAX_ERROR_STRING];
int errorSize;
MPI_Error_string(errorCode, errorString, &errorSize);
WITH_RANK << "__slice__:mpi: status "
<< "{ .source=" << status.MPI_SOURCE
<< ", .tag=" << status.MPI_TAG
<< ", .error=" << status.MPI_ERROR
<< ", .errCode=" << errorCode
<< ", .err=" << errorString
<< " }"
<< "\n";
#endif
}Utils
This section presents some utilities
Pretty printing
The pretty printing uses the dbg-macro package.
template <typename T>
std::string pretty_print(T&& value) {
std::stringstream stream;
#if ATRIP_DEBUG > 1
dbg::pretty_print(stream, std::forward<T>(value));
#endif
return stream.str();
}Chrono
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.
#define WITH_CHRONO(__chrono_name, ...) \
Atrip::chrono[__chrono_name].start(); \
__VA_ARGS__ \
Atrip::chrono[__chrono_name].stop();
struct Timer {
using Clock = std::chrono::high_resolution_clock;
using Event = std::chrono::time_point<Clock>;
std::chrono::duration<double> duration;
Event _start;
inline void start() noexcept { _start = Clock::now(); }
inline void stop() noexcept { duration += Clock::now() - _start; }
inline void clear() noexcept { duration *= 0; }
inline double count() const noexcept { return duration.count(); }
};
using Timings = std::map<std::string, Timer>;The rank mapping
This section introduces the concept of rank mapping, which defines how slices will be allocated to every rank.
#pragma once
#include <vector>
#include <algorithm>
#include <atrip/Slice.hpp>
#include <atrip/Tuples.hpp>
namespace atrip {
struct RankMap {
static bool RANK_ROUND_ROBIN;
std::vector<size_t> const lengths;
size_t const np, size;
ClusterInfo const clusterInfo;
RankMap(std::vector<size_t> lens, size_t np_, MPI_Comm comm)
: lengths(lens)
, np(np_)
, size(std::accumulate(lengths.begin(), lengths.end(),
1UL, std::multiplies<size_t>()))
, clusterInfo(getClusterInfo(comm))
{ assert(lengths.size() <= 2); }
size_t find(Slice::Location const& p) const noexcept {
if (RANK_ROUND_ROBIN) {
return p.source * np + p.rank;
} else {
const size_t
rankPosition = p.source * clusterInfo.ranksPerNode
+ clusterInfo.rankInfos[p.rank].localRank
;
return rankPosition * clusterInfo.nNodes
+ clusterInfo.rankInfos[p.rank].nodeId
;
}
}
size_t nSources() const noexcept {
return size / np + size_t(size % np != 0);
}
bool isPaddingRank(size_t rank) const noexcept {
return size % np == 0
? false
: rank > (size % np - 1)
;
}
bool isSourcePadding(size_t rank, size_t source) const noexcept {
return source == nSources() && isPaddingRank(rank);
}
Slice::Location
find(ABCTuple const& abc, Slice::Type sliceType) const {
// tuple = {11, 8} when abc = {11, 8, 9} and sliceType = AB
// tuple = {11, 0} when abc = {11, 8, 9} and sliceType = A
const auto tuple = Slice::subtupleBySlice(abc, sliceType);
const size_t index
= tuple[0]
+ tuple[1] * (lengths.size() > 1 ? lengths[0] : 0)
;
size_t rank, source;
if (RANK_ROUND_ROBIN) {
rank = index % np;
source = index / np;
} else {
size_t const
// the node that will be assigned to
nodeId = index % clusterInfo.nNodes
// how many times it has been assigned to the node
, s_n = index / clusterInfo.nNodes
// which local rank in the node should be
, localRank = s_n % clusterInfo.ranksPerNode
// and the local source (how many times we chose this local rank)
, localSource = s_n / clusterInfo.ranksPerNode
;
// find the localRank-th entry in clusterInfo
auto const& it =
std::find_if(clusterInfo.rankInfos.begin(),
clusterInfo.rankInfos.end(),
[nodeId, localRank](RankInfo const& ri) {
return ri.nodeId == nodeId
&& ri.localRank == localRank
;
});
if (it == clusterInfo.rankInfos.end()) {
throw "FATAL! Error in node distribution of the slices";
}
rank = (*it).globalRank;
source = localSource;
}
return
{ rank
, source
};
}
};
}The slice union
#pragma once
#include <atrip/Debug.hpp>
#include <atrip/Slice.hpp>
#include <atrip/RankMap.hpp>
namespace atrip {
struct SliceUnion {
using F = double;
using Tensor = CTF::Tensor<F>;
virtual void
sliceIntoBuffer(size_t iteration, Tensor &to, Tensor const& from) = 0;
/*
* This function should enforce an important property of a SliceUnion.
* Namely, there can be no two Slices of the same nature.
*
* This means that there can be at most one slice with a given Ty_x_Tu.
*/
void checkForDuplicates() const {
std::vector<Slice::Ty_x_Tu> tytus;
for (auto const& s: slices) {
if (s.isFree()) continue;
tytus.push_back({s.info.type, s.info.tuple});
}
for (auto const& tytu: tytus) {
if (std::count(tytus.begin(), tytus.end(), tytu) > 1)
throw "Invariance violated, more than one slice with same Ty_x_Tu";
}
}
std::vector<Slice::Ty_x_Tu> neededSlices(ABCTuple const& abc) {
std::vector<Slice::Ty_x_Tu> needed(sliceTypes.size());
// build the needed vector
std::transform(sliceTypes.begin(), sliceTypes.end(),
needed.begin(),
[&abc](Slice::Type const type) {
auto tuple = Slice::subtupleBySlice(abc, type);
return std::make_pair(type, tuple);
});
return needed;
}
/* buildLocalDatabase
*
* It should build a database of slices so that we know what is needed
* to fetch in the next iteration represented by the tuple 'abc'.
*
* 1. The algorithm works as follows, we build a database of the all
* the slice types that we need together with their tuple.
*
* 2. Look in the SliceUnion if we already have this tuple,
* if we already have it mark it (TODO)
*
* 3. If we don't have the tuple, look for a (state=acceptor, type=blank)
* slice and mark this slice as type=Fetch with the corresponding type
* and tuple.
*
* NOTE: The algorithm should certify that we always have enough blank
* slices.
*
*/
Slice::LocalDatabase buildLocalDatabase(ABCTuple const& abc) {
Slice::LocalDatabase result;
auto const needed = neededSlices(abc);
WITH_RANK << "__db__:needed:" << pretty_print(needed) << "\n";
// BUILD THE DATABASE
// we need to loop over all sliceTypes that this TensorUnion
// is representing and find out how we will get the corresponding
// slice for the abc we are considering right now.
for (auto const& pair: needed) {
auto const type = pair.first;
auto const tuple = pair.second;
auto const from = rankMap.find(abc, type);
#ifdef HAVE_OCD
WITH_RANK << "__db__:want:" << pretty_print(pair) << "\n";
for (auto const& s: slices)
WITH_RANK << "__db__:guts:ocd "
<< s.info << " pt " << s.data
<< "\n";
#endif
#ifdef HAVE_OCD
WITH_RANK << "__db__: checking if exact match" << "\n";
#endif
{
// FIRST: look up if there is already a *Ready* slice matching what we
// need
auto const& it
= std::find_if(slices.begin(), slices.end(),
[&tuple, &type](Slice const& other) {
return other.info.tuple == tuple
&& other.info.type == type
// we only want another slice when it
// has already ready-to-use data
&& other.isUnwrappable()
;
});
if (it != slices.end()) {
// if we find this slice, it means that we don't have to do anything
WITH_RANK << "__db__: EXACT: found EXACT in name=" << name
<< " for tuple " << tuple[0] << ", " << tuple[1]
<< " ptr " << it->data
<< "\n";
result.push_back({name, it->info});
continue;
}
}
#ifdef HAVE_OCD
WITH_RANK << "__db__: checking if recycle" << "\n";
#endif
// Try to find a recyling possibility ie. find a slice with the same
// tuple and that has a valid data pointer.
auto const& recycleIt
= std::find_if(slices.begin(), slices.end(),
[&tuple, &type](Slice const& other) {
return other.info.tuple == tuple
&& other.info.type != type
&& other.isRecyclable()
;
});
// if we find this recylce, then we find a Blank slice
// (which should exist by construction :THINK)
//
if (recycleIt != slices.end()) {
auto& blank = Slice::findOneByType(slices, Slice::Blank);
// TODO: formalize this through a method to copy information
// from another slice
blank.data = recycleIt->data;
blank.info.type = type;
blank.info.tuple = tuple;
blank.info.state = Slice::Recycled;
blank.info.from = from;
blank.info.recycling = recycleIt->info.type;
result.push_back({name, blank.info});
WITH_RANK << "__db__: RECYCLING: n" << name
<< " " << pretty_print(abc)
<< " get " << pretty_print(blank.info)
<< " from " << pretty_print(recycleIt->info)
<< " ptr " << recycleIt->data
<< "\n"
;
continue;
}
// in this case we have to create a new slice
// this means that we should have a blank slice at our disposal
// and also the freePointers should have some elements inside,
// so we pop a data pointer from the freePointers container
#ifdef HAVE_OCD
WITH_RANK << "__db__: none work, doing new" << "\n";
#endif
{
WITH_RANK << "__db__: NEW: finding blank in " << name
<< " for type " << type
<< " for tuple " << tuple[0] << ", " << tuple[1]
<< "\n"
;
auto& blank = Slice::findOneByType(slices, Slice::Blank);
blank.info.type = type;
blank.info.tuple = tuple;
blank.info.from = from;
// Handle self sufficiency
blank.info.state = Atrip::rank == from.rank
? Slice::SelfSufficient
: Slice::Fetch
;
if (blank.info.state == Slice::SelfSufficient) {
blank.data = sources[from.source].data();
} else {
if (freePointers.size() == 0) {
std::stringstream stream;
stream << "No more free pointers "
<< "for type " << type
<< " and name " << name
;
throw std::domain_error(stream.str());
}
auto dataPointer = freePointers.begin();
freePointers.erase(dataPointer);
blank.data = *dataPointer;
}
result.push_back({name, blank.info});
continue;
}
}
#ifdef HAVE_OCD
for (auto const& s: slices)
WITH_RANK << "__db__:guts:ocd:__end__ " << s.info << "\n";
#endif
return result;
}
/*
* Garbage collect slices not needed for the next iteration.
*
* It will throw if it tries to gc a slice that has not been
* previously unwrapped, as a safety mechanism.
*/
void clearUnusedSlicesForNext(ABCTuple const& abc) {
auto const needed = neededSlices(abc);
// CLEAN UP SLICES, FREE THE ONES THAT ARE NOT NEEDED ANYMORE
for (auto& slice: slices) {
// if the slice is free, then it was not used anyways
if (slice.isFree()) continue;
// try to find the slice in the needed slices list
auto const found
= std::find_if(needed.begin(), needed.end(),
[&slice] (Slice::Ty_x_Tu const& tytu) {
return slice.info.tuple == tytu.second
&& slice.info.type == tytu.first
;
});
// if we did not find slice in needed, then erase it
if (found == needed.end()) {
// We have to be careful about the data pointer,
// for SelfSufficient, the data pointer is a source pointer
// of the slice, so we should just wipe it.
//
// For Ready slices, we have to be careful if there are some
// recycled slices depending on it.
bool freeSlicePointer = true;
// allow to gc unwrapped and recycled, never Fetch,
// if we have a Fetch slice then something has gone very wrong.
if (!slice.isUnwrapped() && slice.info.state != Slice::Recycled)
throw
std::domain_error("Trying to garbage collect "
" a non-unwrapped slice! "
+ pretty_print(&slice)
+ pretty_print(slice.info));
// it can be that our slice is ready, but it has some hanging
// references lying around in the form of a recycled slice.
// Of course if we need the recycled slice the next iteration
// this would be fatal, because we would then free the pointer
// of the slice and at some point in the future we would
// overwrite it. Therefore, we must check if slice has some
// references in slices and if so then
//
// - we should mark those references as the original (since the data
// pointer should be the same)
//
// - we should make sure that the data pointer of slice
// does not get freed.
//
if (slice.info.state == Slice::Ready) {
WITH_OCD WITH_RANK
<< "__gc__:" << "checking for data recycled dependencies\n";
auto recycled
= Slice::hasRecycledReferencingToIt(slices, slice.info);
if (recycled.size()) {
Slice* newReady = recycled[0];
WITH_OCD WITH_RANK
<< "__gc__:" << "swaping recycled "
<< pretty_print(newReady->info)
<< " and "
<< pretty_print(slice.info)
<< "\n";
newReady->markReady();
assert(newReady->data == slice.data);
freeSlicePointer = false;
for (size_t i = 1; i < recycled.size(); i++) {
auto newRecyled = recycled[i];
newRecyled->info.recycling = newReady->info.type;
WITH_OCD WITH_RANK
<< "__gc__:" << "updating recycled "
<< pretty_print(newRecyled->info)
<< "\n";
}
}
}
// if the slice is self sufficient, do not dare touching the
// pointer, since it is a pointer to our sources in our rank.
if ( slice.info.state == Slice::SelfSufficient
|| slice.info.state == Slice::Recycled
) {
freeSlicePointer = false;
}
// make sure we get its data pointer to be used later
// only for non-recycled, since it can be that we need
// for next iteration the data of the slice that the recycled points
// to
if (freeSlicePointer) {
freePointers.insert(slice.data);
WITH_RANK << "~~~:cl(" << name << ")"
<< " added to freePointer "
<< pretty_print(freePointers)
<< "\n";
} else {
WITH_OCD WITH_RANK << "__gc__:not touching the free Pointer\n";
}
// at this point, let us blank the slice
WITH_RANK << "~~~:cl(" << name << ")"
<< " freeing up slice "
<< " info " << slice.info
<< "\n";
slice.free();
}
}
}
// CONSTRUCTOR
SliceUnion( Tensor const& sourceTensor
, std::vector<Slice::Type> sliceTypes_
, std::vector<size_t> sliceLength_
, std::vector<size_t> paramLength
, size_t np
, MPI_Comm child_world
, MPI_Comm global_world
, Slice::Name name_
, size_t nSliceBuffers = 4
)
: rankMap(paramLength, np, global_world)
, world(child_world)
, universe(global_world)
, sliceLength(sliceLength_)
, sources(rankMap.nSources(),
std::vector<F>
(std::accumulate(sliceLength.begin(),
sliceLength.end(),
1UL, std::multiplies<size_t>())))
, name(name_)
, sliceTypes(sliceTypes_)
, sliceBuffers(nSliceBuffers, sources[0])
//, slices(2 * sliceTypes.size(), Slice{ sources[0].size() })
{ // constructor begin
LOG(0,"Atrip") << "INIT SliceUnion: " << name << "\n";
slices
= std::vector<Slice>(2 * sliceTypes.size(), { sources[0].size() });
// TODO: think exactly ^------------------- about this number
// initialize the freePointers with the pointers to the buffers
std::transform(sliceBuffers.begin(), sliceBuffers.end(),
std::inserter(freePointers, freePointers.begin()),
[](std::vector<F> &vec) { return vec.data(); });
LOG(1,"Atrip") << "rankMap.nSources "
<< rankMap.nSources() << "\n";
LOG(1,"Atrip") << "#slices "
<< slices.size() << "\n";
LOG(1,"Atrip") << "#slices[0] "
<< slices[0].size << "\n";
LOG(1,"Atrip") << "#sources "
<< sources.size() << "\n";
LOG(1,"Atrip") << "#sources[0] "
<< sources[0].size() << "\n";
LOG(1,"Atrip") << "#freePointers "
<< freePointers.size() << "\n";
LOG(1,"Atrip") << "#sliceBuffers "
<< sliceBuffers.size() << "\n";
LOG(1,"Atrip") << "#sliceBuffers[0] "
<< sliceBuffers[0].size() << "\n";
LOG(1,"Atrip") << "#sliceLength "
<< sliceLength.size() << "\n";
LOG(1,"Atrip") << "#paramLength "
<< paramLength.size() << "\n";
LOG(1,"Atrip") << "GB*" << np << " "
<< double(sources.size() + sliceBuffers.size())
* sources[0].size()
* 8 * np
/ 1073741824.0
<< "\n";
} // constructor ends
void init(Tensor const& sourceTensor) {
CTF::World w(world);
const int rank = Atrip::rank
, order = sliceLength.size()
;
std::vector<int> const syms(order, NS);
std::vector<int> __sliceLength(sliceLength.begin(), sliceLength.end());
Tensor toSliceInto(order,
__sliceLength.data(),
syms.data(),
w);
LOG(1,"Atrip") << "slicing... \n";
// setUp sources
for (size_t it(0); it < rankMap.nSources(); ++it) {
const size_t
source = rankMap.isSourcePadding(rank, source) ? 0 : it;
WITH_OCD
WITH_RANK
<< "Init:toSliceInto it-" << it
<< " :: source " << source << "\n";
sliceIntoBuffer(source, toSliceInto, sourceTensor);
}
}
/**
* \brief Send asynchronously only if the state is Fetch
*/
void send( size_t otherRank
, Slice::LocalDatabaseElement const& el
, size_t tag) const noexcept {
MPI_Request request;
bool sendData_p = false;
auto const& info = el.info;
if (info.state == Slice::Fetch) sendData_p = true;
// TODO: remove this because I have SelfSufficient
if (otherRank == info.from.rank) sendData_p = false;
if (!sendData_p) return;
switch (el.name) {
case Slice::Name::TA:
case Slice::Name::VIJKA:
if (otherRank / 48 == Atrip::rank / 48) {
Atrip::localSend++;
} else {
Atrip::networkSend++;
}
}
MPI_Isend( sources[info.from.source].data()
, sources[info.from.source].size()
, MPI_DOUBLE /* TODO: adapt this with traits */
, otherRank
, tag
, universe
, &request
);
WITH_CRAZY_DEBUG
WITH_RANK << "sent to " << otherRank << "\n";
}
/**
* \brief Receive asynchronously only if the state is Fetch
*/
void receive(Slice::Info const& info, size_t tag) noexcept {
auto& slice = Slice::findByInfo(slices, info);
if (Atrip::rank == info.from.rank) return;
if (slice.info.state == Slice::Fetch) {
// TODO: do it through the slice class
slice.info.state = Slice::Dispatched;
MPI_Request request;
slice.request = request;
MPI_Irecv( slice.data
, slice.size
, MPI_DOUBLE // TODO: Adapt this with traits
, info.from.rank
, tag
, universe
, &slice.request
//, MPI_STATUS_IGNORE
);
}
}
void unwrapAll(ABCTuple const& abc) {
for (auto type: sliceTypes) unwrapSlice(type, abc);
}
F* unwrapSlice(Slice::Type type, ABCTuple const& abc) {
WITH_CRAZY_DEBUG
WITH_RANK << "__unwrap__:slice " << type << " w n "
<< name
<< " abc" << pretty_print(abc)
<< "\n";
auto& slice = Slice::findByTypeAbc(slices, type, abc);
WITH_RANK << "__unwrap__:info " << slice.info << "\n";
switch (slice.info.state) {
case Slice::Dispatched:
WITH_RANK << "__unwrap__:Fetch: " << &slice
<< " info " << pretty_print(slice.info)
<< "\n";
slice.unwrapAndMarkReady();
return slice.data;
break;
case Slice::SelfSufficient:
WITH_RANK << "__unwrap__:SelfSufficient: " << &slice
<< " info " << pretty_print(slice.info)
<< "\n";
return slice.data;
break;
case Slice::Ready:
WITH_RANK << "__unwrap__:READY: UNWRAPPED ALREADY" << &slice
<< " info " << pretty_print(slice.info)
<< "\n";
return slice.data;
break;
case Slice::Recycled:
WITH_RANK << "__unwrap__:RECYCLED " << &slice
<< " info " << pretty_print(slice.info)
<< "\n";
return unwrapSlice(slice.info.recycling, abc);
break;
case Slice::Fetch:
case Slice::Acceptor:
throw std::domain_error("Can't unwrap an acceptor or fetch slice!");
break;
default:
throw std::domain_error("Unknown error unwrapping slice!");
}
return slice.data;
}
const RankMap rankMap;
const MPI_Comm world;
const MPI_Comm universe;
const std::vector<size_t> sliceLength;
std::vector< std::vector<F> > sources;
std::vector< Slice > slices;
Slice::Name name;
const std::vector<Slice::Type> sliceTypes;
std::vector< std::vector<F> > sliceBuffers;
std::set<F*> freePointers;
};
SliceUnion&
unionByName(std::vector<SliceUnion*> const& unions, Slice::Name name) {
const auto sliceUnionIt
= std::find_if(unions.begin(), unions.end(),
[&name](SliceUnion const* s) {
return name == s->name;
});
if (sliceUnionIt == unions.end()) {
std::stringstream stream;
stream << "SliceUnion(" << name << ") not found!";
throw std::domain_error(stream.str());
}
return **sliceUnionIt;
}
}Tuples
This section introduces the types for tuples \( (a,b,c) \) as well as their distribution to nodes and cores.
Tuples types
The main tuple types are simple type aliases for finite-size arrays. A tuple is thus simply 3 natural numbers \( (a,b,c) \) whereas a partial tuple is a two dimensional subset of these three.
using ABCTuple = std::array<size_t, 3>;
using PartialTuple = std::array<size_t, 2>;
using ABCTuples = std::vector<ABCTuple>;
constexpr ABCTuple FAKE_TUPLE = {0, 0, 0};
constexpr ABCTuple INVALID_TUPLE = {1, 1, 1};Distributing the tuples
In general it is our task to distribute all the tuples
\( (a,b,c) \) among the ranks. Every distribution should
make sure to allocate the same amount of tuples to every rank,
padding the list with FAKE_TUPLE elements as necessary.
The interface that we propose for this is simplye
struct TuplesDistribution {
virtual ABCTuples getTuples(size_t Nv, MPI_Comm universe) = 0;
virtual bool tupleIsFake(ABCTuple const& t) { return t == FAKE_TUPLE; }
};Node information
- nodeList
- List of hostnames of size \( N_n \)
- nodeInfos
- List of (hostname, local rank Id) of size \( N_p \), i.e., size of ranks where local rank id goes from 0 to 48.
getNodeNames gets the names of the nodes used,
i.e., the size of the resulting vector gives the
number of nodes.
std::vector<std::string> getNodeNames(MPI_Comm comm){
int rank, np;
MPI_Comm_rank(comm, &rank);
MPI_Comm_size(comm, &np);
std::vector<std::string> nodeList(np);
char nodeName[MPI_MAX_PROCESSOR_NAME]
, nodeNames[np*MPI_MAX_PROCESSOR_NAME]
;
std::vector<int> nameLengths(np)
, off(np)
;
int nameLength;
MPI_Get_processor_name(nodeName, &nameLength);
MPI_Allgather(&nameLength,
1,
MPI_INT,
nameLengths.data(),
1,
MPI_INT,
comm);
for (int i(1); i < np; i++)
off[i] = off[i-1] + nameLengths[i-1];
MPI_Allgatherv(nodeName,
nameLengths[rank],
MPI_BYTE,
nodeNames,
nameLengths.data(),
off.data(),
MPI_BYTE,
comm);
for (int i(0); i < np; i++) {
std::string const s(&nodeNames[off[i]], nameLengths[i]);
nodeList[i] = s;
}
return nodeList;
}
getNodeInfos
struct RankInfo {
const std::string name;
const size_t nodeId;
const size_t globalRank;
const size_t localRank;
const size_t ranksPerNode;
};
template <typename A>
A unique(A const &xs) {
auto result = xs;
std::sort(std::begin(result), std::end(result));
auto const& last = std::unique(std::begin(result), std::end(result));
result.erase(last, std::end(result));
return result;
}
std::vector<RankInfo>
getNodeInfos(std::vector<string> const& nodeNames) {
std::vector<RankInfo> result;
auto const uniqueNames = unique(nodeNames);
auto const index = [&uniqueNames](std::string const& s) {
auto const& it = std::find(uniqueNames.begin(), uniqueNames.end(), s);
return std::distance(uniqueNames.begin(), it);
};
std::vector<size_t> localRanks(uniqueNames.size(), 0);
size_t globalRank = 0;
for (auto const& name: nodeNames) {
const size_t nodeId = index(name);
result.push_back({name,
nodeId,
globalRank++,
localRanks[nodeId]++,
std::count(nodeNames.begin(),
nodeNames.end(),
name)
});
}
return result;
}
struct ClusterInfo {
const size_t nNodes, np, ranksPerNode;
const std::vector<RankInfo> rankInfos;
};
ClusterInfo
getClusterInfo(MPI_Comm comm) {
auto const names = getNodeNames(comm);
auto const rankInfos = getNodeInfos(names);
return ClusterInfo {
unique(names).size(),
names.size(),
rankInfos[0].ranksPerNode,
rankInfos
};
}Naive list
The naive implementation of the global tuples list is simple three for loops creating tuples of the sort \( (a,b,c) \) where the following conditions are met at the same time:
- \( a \leq b \leq c \)
- \( a \neq b \land b \neq c \)
This means, \( (1, 2, 3) , (1, 1, 3) , (1, 2, 2) \) are acceptable tuples wherease \( (2, 1, 1) \) and \( (1, 1, 1) \) are not.
ABCTuples getTuplesList(size_t Nv, size_t rank, size_t np) {
const size_t
// total number of tuples for the problem
n = Nv * (Nv + 1) * (Nv + 2) / 6 - Nv
// all ranks should have the same number of tuples_per_rank
, tuples_per_rank = n / np + size_t(n % np != 0)
// start index for the global tuples list
, start = tuples_per_rank * rank
// end index for the global tuples list
, end = tuples_per_rank * (rank + 1)
;
LOG(1,"Atrip") << "tuples_per_rank = " << tuples_per_rank << "\n";
WITH_RANK << "start, end = " << start << ", " << end << "\n";
ABCTuples result(tuples_per_rank, FAKE_TUPLE);
for (size_t a(0), r(0), g(0); a < Nv; a++)
for (size_t b(a); b < Nv; b++)
for (size_t c(b); c < Nv; c++){
if ( a == b && b == c ) continue;
if ( start <= g && g < end) result[r++] = {a, b, c};
g++;
}
return result;
}and all tuples would simply be
ABCTuples getAllTuplesList(const size_t Nv) {
const size_t n = Nv * (Nv + 1) * (Nv + 2) / 6 - Nv;
ABCTuples result(n);
for (size_t a(0), u(0); a < Nv; a++)
for (size_t b(a); b < Nv; b++)
for (size_t c(b); c < Nv; c++){
if ( a == b && b == c ) continue;
result[u++] = {a, b, c};
}
return result;
}
With getTupleList we can easily define a tuple distribution like
struct NaiveDistribution : public TuplesDistribution {
ABCTuples getTuples(size_t Nv, MPI_Comm universe) override {
int rank, np;
MPI_Comm_rank(universe, &rank);
MPI_Comm_size(universe, &np);
return getTuplesList(Nv, (size_t)rank, (size_t)np);
}
};Group and sort list
Utils
// Provides the node on which the slice-element is found
// Right now we distribute the slices in a round robin fashion
// over the different nodes (NOTE: not mpi ranks but nodes)
inline
size_t isOnNode(size_t tuple, size_t nNodes) { return tuple % nNodes; }
// return the node (or all nodes) where the elements of this
// tuple are located
std::vector<size_t> getTupleNodes(ABCTuple const& t, size_t nNodes) {
std::vector<size_t>
nTuple = { isOnNode(t[0], nNodes)
, isOnNode(t[1], nNodes)
, isOnNode(t[2], nNodes)
};
return unique(nTuple);
}
struct Info {
size_t nNodes;
size_t nodeId;
};Distribution
wording: home element = element which is located on the given node
- we distribute the tuples such that each tuple has at least one 'home element'
- we sort each tuple in a way that the 'home element' are the fastest indices
- we sort the list of tuples on every node
- we resort the tuples that for every tuple abc the following holds: a<b<c
ABCTuples specialDistribution(Info const& info, ABCTuples const& allTuples) {
ABCTuples nodeTuples;
size_t const nNodes(info.nNodes);
std::vector<ABCTuples>
container1d(nNodes)
, container2d(nNodes * nNodes)
, container3d(nNodes * nNodes * nNodes)
;
if (info.nodeId == 0)
std::cout << "\tGoing through all "
<< allTuples.size()
<< " tuples in "
<< nNodes
<< " nodes\n";
// build container-n-d's
for (auto const& t: allTuples) {
// one which node(s) are the tuple elements located...
// put them into the right container
auto const _nodes = getTupleNodes(t, nNodes);
switch (_nodes.size()) {
case 1:
container1d[_nodes[0]].push_back(t);
break;
case 2:
container2d[ _nodes[0]
+ _nodes[1] * nNodes
].push_back(t);
break;
case 3:
container3d[ _nodes[0]
+ _nodes[1] * nNodes
+ _nodes[2] * nNodes * nNodes
].push_back(t);
break;
}
}
if (info.nodeId == 0)
std::cout << "\tBuilding 1-d containers\n";
// DISTRIBUTE 1-d containers
// every tuple which is only located at one node belongs to this node
{
auto const& _tuples = container1d[info.nodeId];
nodeTuples.resize(_tuples.size(), INVALID_TUPLE);
std::copy(_tuples.begin(), _tuples.end(), nodeTuples.begin());
}
if (info.nodeId == 0)
std::cout << "\tBuilding 2-d containers\n";
// DISTRIBUTE 2-d containers
//the tuples which are located at two nodes are half/half given to these nodes
for (size_t yx = 0; yx < container2d.size(); yx++) {
auto const& _tuples = container2d[yx];
const
size_t idx = yx % nNodes
// remeber: yx = idy * nNodes + idx
, idy = yx / nNodes
, n_half = _tuples.size() / 2
, size = nodeTuples.size()
;
size_t nbeg, nend;
if (info.nodeId == idx) {
nbeg = 0 * n_half;
nend = n_half;
} else if (info.nodeId == idy) {
nbeg = 1 * n_half;
nend = _tuples.size();
} else {
// either idx or idy is my node
continue;
}
size_t const nextra = nend - nbeg;
nodeTuples.resize(size + nextra, INVALID_TUPLE);
std::copy(_tuples.begin() + nbeg,
_tuples.begin() + nend,
nodeTuples.begin() + size);
}
if (info.nodeId == 0)
std::cout << "\tBuilding 3-d containers\n";
// DISTRIBUTE 3-d containers
for (size_t zyx = 0; zyx < container3d.size(); zyx++) {
auto const& _tuples = container3d[zyx];
const
size_t idx = zyx % nNodes
, idy = (zyx / nNodes) % nNodes
// remember: zyx = idx + idy * nNodes + idz * nNodes^2
, idz = zyx / nNodes / nNodes
, n_third = _tuples.size() / 3
, size = nodeTuples.size()
;
size_t nbeg, nend;
if (info.nodeId == idx) {
nbeg = 0 * n_third;
nend = 1 * n_third;
} else if (info.nodeId == idy) {
nbeg = 1 * n_third;
nend = 2 * n_third;
} else if (info.nodeId == idz) {
nbeg = 2 * n_third;
nend = _tuples.size();
} else {
// either idx or idy or idz is my node
continue;
}
size_t const nextra = nend - nbeg;
nodeTuples.resize(size + nextra, INVALID_TUPLE);
std::copy(_tuples.begin() + nbeg,
_tuples.begin() + nend,
nodeTuples.begin() + size);
}
if (info.nodeId == 0) std::cout << "\tswapping tuples...\n";
/*
* sort part of group-and-sort algorithm
* every tuple on a given node is sorted in a way that
* the 'home elements' are the fastest index.
* 1:yyy 2:yyn(x) 3:yny(x) 4:ynn(x) 5:nyy 6:nyn(x) 7:nny 8:nnn
*/
for (auto &nt: nodeTuples){
if ( isOnNode(nt[0], nNodes) == info.nodeId ){ // 1234
if ( isOnNode(nt[2], nNodes) != info.nodeId ){ // 24
size_t const x(nt[0]);
nt[0] = nt[2]; // switch first and last
nt[2] = x;
}
else if ( isOnNode(nt[1], nNodes) != info.nodeId){ // 3
size_t const x(nt[0]);
nt[0] = nt[1]; // switch first two
nt[1] = x;
}
} else {
if ( isOnNode(nt[1], nNodes) == info.nodeId // 56
&& isOnNode(nt[2], nNodes) != info.nodeId
) { // 6
size_t const x(nt[1]);
nt[1] = nt[2]; // switch last two
nt[2] = x;
}
}
}
if (info.nodeId == 0) std::cout << "\tsorting list of tuples...\n";
//now we sort the list of tuples
std::sort(nodeTuples.begin(), nodeTuples.end());
if (info.nodeId == 0) std::cout << "\trestoring tuples...\n";
// we bring the tuples abc back in the order a<b<c
for (auto &t: nodeTuples) std::sort(t.begin(), t.end());
#if ATRIP_DEBUG > 1
if (info.nodeId == 0)
std::cout << "checking for validity of " << nodeTuples.size() << std::endl;
const bool anyInvalid
= std::any_of(nodeTuples.begin(),
nodeTuples.end(),
[](ABCTuple const& t) { return t == INVALID_TUPLE; });
if (anyInvalid) throw "Some tuple is invalid in group-and-sort algorithm";
#endif
return nodeTuples;
}Main
The main routine should return the list of tuples to be handled by the current rank.
Let \( N_p \) be the number of ranks or processes. Let \( N_n \) be the number of nodes or sockets.
Then we have the following
Global rank | 0 1 2 3 4 5 6 7 8 key | global rank nodeId | 0 1 0 1 1 0 2 2 2 Local rank | 0 0 1 1 2 2 0 1 2 intra color | 0 1 0 1 1 0 2 2 2
std::vector<ABCTuple> main(MPI_Comm universe, size_t Nv) {
int rank, np;
MPI_Comm_rank(universe, &rank);
MPI_Comm_size(universe, &np);
std::vector<ABCTuple> result;
auto const nodeNames(getNodeNames(universe));
size_t const nNodes = unique(nodeNames).size();
auto const nodeInfos = getNodeInfos(nodeNames);
// We want to construct a communicator which only contains of one
// element per node
bool const computeDistribution
= nodeInfos[rank].localRank == 0;
std::vector<ABCTuple>
nodeTuples
= computeDistribution
? specialDistribution(Info{nNodes, nodeInfos[rank].nodeId},
getAllTuplesList(Nv))
: std::vector<ABCTuple>()
;
LOG(1,"Atrip") << "got nodeTuples\n";
// now we have to send the data from **one** rank on each node
// to all others ranks of this node
const
int color = nodeInfos[rank].nodeId
, key = nodeInfos[rank].localRank
;
MPI_Comm INTRA_COMM;
MPI_Comm_split(universe, color, key, &INTRA_COMM);
Every node has to distribute at least
nodeTuples.size() / nodeInfos[rank].ranksPerNode
tuples among the ranks.
We have to communicate this quantity among all nodes.
size_t const
tuplesPerRankLocal
= nodeTuples.size() / nodeInfos[rank].ranksPerNode
+ size_t(nodeTuples.size() % nodeInfos[rank].ranksPerNode != 0)
;
size_t tuplesPerRankGlobal;
MPI_Reduce(&tuplesPerRankLocal,
&tuplesPerRankGlobal,
1,
MPI_UINT64_T,
MPI_MAX,
0,
universe);
MPI_Bcast(&tuplesPerRankGlobal,
1,
MPI_UINT64_T,
0,
universe);
LOG(1,"Atrip") << "Tuples per rank: " << tuplesPerRankGlobal << "\n";
LOG(1,"Atrip") << "ranks per node " << nodeInfos[rank].ranksPerNode << "\n";
LOG(1,"Atrip") << "#nodes " << nNodes << "\n";
Now we have the tuples that every rank has to have, i.e.,
tuplesPerRankGlobal.
However before this,
the tuples in nodeTuples now have to be sent from the local rank
in every node to all the ranks in the given node,
and we have to make sure that every rank inside a given node
gets the same amount of tuples, in this case it should be
tuplesPerRankLocal, and in our node the total number
of tuples should be tuplesPerRankLocal * nodeInfos[rank].ranksPerNode,
however this might not be the case up to now due to divisibility issues.
Up to now we have exactly nodeTuples.size() tuples, we have to make sure by
resizing that the condition above is met, i.e., so we can resize
and add some fake tuples at the end as padding.
size_t const totalTuples
= tuplesPerRankGlobal * nodeInfos[rank].ranksPerNode;
if (computeDistribution) {
// pad with FAKE_TUPLEs
nodeTuples.insert(nodeTuples.end(),
totalTuples - nodeTuples.size(),
FAKE_TUPLE);
}
And now we can simply scatter the tuples in nodeTuples and send
tuplesPerRankGlobal to the different ranks in the node,
{
// construct mpi type for abctuple
MPI_Datatype MPI_ABCTUPLE;
MPI_Type_vector(nodeTuples[0].size(), 1, 1, MPI_UINT64_T, &MPI_ABCTUPLE);
MPI_Type_commit(&MPI_ABCTUPLE);
LOG(1,"Atrip") << "scattering tuples \n";
result.resize(tuplesPerRankGlobal);
MPI_Scatter(nodeTuples.data(),
tuplesPerRankGlobal,
MPI_ABCTUPLE,
result.data(),
tuplesPerRankGlobal,
MPI_ABCTUPLE,
0,
INTRA_COMM);
MPI_Type_free(&MPI_ABCTUPLE);
}
The next step is sending the tuples in the local root rank
to the other ranks in the node, this we do with the MPI function
MPI_Scatterv.
Every rank gets tuplesPerRankLocal tuples and
the nodeTuples vector is now homogeneous and divisible by the number
of ranks per node in our node.
Therefore, the displacements are simply the vector
and the sendCounts vector is simply the constant vector
tuplesPerRankLocal of size ranksPerNode.
return result;
}Interface
The distribution interface will then simply be
struct Distribution : public TuplesDistribution {
ABCTuples getTuples(size_t Nv, MPI_Comm universe) override {
return main(universe, Nv);
}
};Unions
Every slice pertaining to every different tensor is sliced differently.
#pragma once
#include <atrip/SliceUnion.hpp>
namespace atrip {
void sliceIntoVector
( std::vector<double> &v
, CTF::Tensor<double> &toSlice
, std::vector<int64_t> const low
, std::vector<int64_t> const up
, CTF::Tensor<double> const& origin
, std::vector<int64_t> const originLow
, std::vector<int64_t> const originUp
) {
// Thank you CTF for forcing me to do this
struct { std::vector<int> up, low; }
toSlice_ = { {up.begin(), up.end()}
, {low.begin(), low.end()} }
, origin_ = { {originUp.begin(), originUp.end()}
, {originLow.begin(), originLow.end()} }
;
WITH_OCD
WITH_RANK << "slicing into " << pretty_print(toSlice_.up)
<< "," << pretty_print(toSlice_.low)
<< " from " << pretty_print(origin_.up)
<< "," << pretty_print(origin_.low)
<< "\n";
#ifndef ATRIP_DONT_SLICE
toSlice.slice( toSlice_.low.data()
, toSlice_.up.data()
, 0.0
, origin
, origin_.low.data()
, origin_.up.data()
, 1.0);
memcpy(v.data(), toSlice.data, sizeof(double) * v.size());
#endif
}
struct TAPHH : public SliceUnion {
TAPHH( Tensor const& sourceTensor
, size_t No
, size_t Nv
, size_t np
, MPI_Comm child_world
, MPI_Comm global_world
) : SliceUnion( sourceTensor
, {Slice::A, Slice::B, Slice::C}
, {Nv, No, No} // size of the slices
, {Nv}
, np
, child_world
, global_world
, Slice::TA
, 6) {
init(sourceTensor);
}
void sliceIntoBuffer(size_t it, Tensor &to, Tensor const& from) override
{
const int Nv = sliceLength[0]
, No = sliceLength[1]
, a = rankMap.find({static_cast<size_t>(Atrip::rank), it});
;
sliceIntoVector( sources[it]
, to, {0, 0, 0}, {Nv, No, No}
, from, {a, 0, 0, 0}, {a+1, Nv, No, No}
);
}
};
struct HHHA : public SliceUnion {
HHHA( Tensor const& sourceTensor
, size_t No
, size_t Nv
, size_t np
, MPI_Comm child_world
, MPI_Comm global_world
) : SliceUnion( sourceTensor
, {Slice::A, Slice::B, Slice::C}
, {No, No, No} // size of the slices
, {Nv} // size of the parametrization
, np
, child_world
, global_world
, Slice::VIJKA
, 6) {
init(sourceTensor);
}
void sliceIntoBuffer(size_t it, Tensor &to, Tensor const& from) override
{
const int No = sliceLength[0]
, a = rankMap.find({static_cast<size_t>(Atrip::rank), it})
;
sliceIntoVector( sources[it]
, to, {0, 0, 0}, {No, No, No}
, from, {0, 0, 0, a}, {No, No, No, a+1}
);
}
};
struct ABPH : public SliceUnion {
ABPH( Tensor const& sourceTensor
, size_t No
, size_t Nv
, size_t np
, MPI_Comm child_world
, MPI_Comm global_world
) : SliceUnion( sourceTensor
, { Slice::AB, Slice::BC, Slice::AC
, Slice::BA, Slice::CB, Slice::CA
}
, {Nv, No} // size of the slices
, {Nv, Nv} // size of the parametrization
, np
, child_world
, global_world
, Slice::VABCI
, 2*6) {
init(sourceTensor);
}
void sliceIntoBuffer(size_t it, Tensor &to, Tensor const& from) override {
const int Nv = sliceLength[0]
, No = sliceLength[1]
, el = rankMap.find({static_cast<size_t>(Atrip::rank), it})
, a = el % Nv
, b = el / Nv
;
sliceIntoVector( sources[it]
, to, {0, 0}, {Nv, No}
, from, {a, b, 0, 0}, {a+1, b+1, Nv, No}
);
}
};
struct ABHH : public SliceUnion {
ABHH( Tensor const& sourceTensor
, size_t No
, size_t Nv
, size_t np
, MPI_Comm child_world
, MPI_Comm global_world
) : SliceUnion( sourceTensor
, {Slice::AB, Slice::BC, Slice::AC}
, {No, No} // size of the slices
, {Nv, Nv} // size of the parametrization
, np
, child_world
, global_world
, Slice::VABIJ
, 6) {
init(sourceTensor);
}
void sliceIntoBuffer(size_t it, Tensor &to, Tensor const& from) override {
const int Nv = from.lens[0]
, No = sliceLength[1]
, el = rankMap.find({static_cast<size_t>(Atrip::rank), it})
, a = el % Nv
, b = el / Nv
;
sliceIntoVector( sources[it]
, to, {0, 0}, {No, No}
, from, {a, b, 0, 0}, {a+1, b+1, No, No}
);
}
};
struct TABHH : public SliceUnion {
TABHH( Tensor const& sourceTensor
, size_t No
, size_t Nv
, size_t np
, MPI_Comm child_world
, MPI_Comm global_world
) : SliceUnion( sourceTensor
, {Slice::AB, Slice::BC, Slice::AC}
, {No, No} // size of the slices
, {Nv, Nv} // size of the parametrization
, np
, child_world
, global_world
, Slice::TABIJ
, 6) {
init(sourceTensor);
}
void sliceIntoBuffer(size_t it, Tensor &to, Tensor const& from) override {
// TODO: maybe generalize this with ABHH
const int Nv = from.lens[0]
, No = sliceLength[1]
, el = rankMap.find({static_cast<size_t>(Atrip::rank), it})
, a = el % Nv
, b = el / Nv
;
sliceIntoVector( sources[it]
, to, {0, 0}, {No, No}
, from, {a, b, 0, 0}, {a+1, b+1, No, No}
);
}
};
}Equations
#pragma once
#include<atrip/Slice.hpp>
#include<atrip/Blas.hpp>
namespace atrip {
double getEnergyDistinct
( const double epsabc
, std::vector<double> const& epsi
, std::vector<double> const& Tijk_
, std::vector<double> const& Zijk_
) {
constexpr size_t blockSize=16;
double energy(0.);
const size_t No = epsi.size();
for (size_t kk=0; kk<No; kk+=blockSize){
const size_t kend( std::min(No, kk+blockSize) );
for (size_t jj(kk); jj<No; jj+=blockSize){
const size_t jend( std::min( No, jj+blockSize) );
for (size_t ii(jj); ii<No; ii+=blockSize){
const size_t iend( std::min( No, ii+blockSize) );
for (size_t k(kk); k < kend; k++){
const double ek(epsi[k]);
const size_t jstart = jj > k ? jj : k;
for (size_t j(jstart); j < jend; j++){
const double ej(epsi[j]);
double facjk( j == k ? 0.5 : 1.0);
size_t istart = ii > j ? ii : j;
for (size_t i(istart); i < iend; i++){
const double ei(epsi[i]);
double facij ( i==j ? 0.5 : 1.0);
double denominator(epsabc - ei - ej - ek);
double U(Zijk_[i + No*j + No*No*k]);
double V(Zijk_[i + No*k + No*No*j]);
double W(Zijk_[j + No*i + No*No*k]);
double X(Zijk_[j + No*k + No*No*i]);
double Y(Zijk_[k + No*i + No*No*j]);
double Z(Zijk_[k + No*j + No*No*i]);
double A(Tijk_[i + No*j + No*No*k]);
double B(Tijk_[i + No*k + No*No*j]);
double C(Tijk_[j + No*i + No*No*k]);
double D(Tijk_[j + No*k + No*No*i]);
double E(Tijk_[k + No*i + No*No*j]);
double F(Tijk_[k + No*j + No*No*i]);
double value(3.0*(A*U+B*V+C*W+D*X+E*Y+F*Z)
+((U+X+Y)-2.0*(V+W+Z))*(A+D+E)
+((V+W+Z)-2.0*(U+X+Y))*(B+C+F));
energy += 2.0*value / denominator * facjk * facij;
} // i
} // j
} // k
} // ii
} // jj
} // kk
return energy;
}
double getEnergySame
( const double epsabc
, std::vector<double> const& epsi
, std::vector<double> const& Tijk_
, std::vector<double> const& Zijk_
) {
constexpr size_t blockSize = 16;
const size_t No = epsi.size();
double energy(0.);
for (size_t kk=0; kk<No; kk+=blockSize){
const size_t kend( std::min( kk+blockSize, No) );
for (size_t jj(kk); jj<No; jj+=blockSize){
const size_t jend( std::min( jj+blockSize, No) );
for (size_t ii(jj); ii<No; ii+=blockSize){
const size_t iend( std::min( ii+blockSize, No) );
for (size_t k(kk); k < kend; k++){
const double ek(epsi[k]);
const size_t jstart = jj > k ? jj : k;
for(size_t j(jstart); j < jend; j++){
const double facjk( j == k ? 0.5 : 1.0);
const double ej(epsi[j]);
const size_t istart = ii > j ? ii : j;
for(size_t i(istart); i < iend; i++){
double ei(epsi[i]);
double facij ( i==j ? 0.5 : 1.0);
double denominator(epsabc - ei - ej - ek);
double U(Zijk_[i + No*j + No*No*k]);
double V(Zijk_[j + No*k + No*No*i]);
double W(Zijk_[k + No*i + No*No*j]);
double A(Tijk_[i + No*j + No*No*k]);
double B(Tijk_[j + No*k + No*No*i]);
double C(Tijk_[k + No*i + No*No*j]);
double value(3.0*( A*U + B*V + C*W) - (A+B+C)*(U+V+W));
energy += 2.0*value / denominator * facjk * facij;
} // i
} // j
} // k
} // ii
} // jj
} // kk
return energy;
}
void singlesContribution
( size_t No
, size_t Nv
, const ABCTuple &abc
, double const* Tph
, double const* VABij
, double const* VACij
, double const* VBCij
, double *Zijk
) {
const size_t a(abc[0]), b(abc[1]), c(abc[2]);
for (size_t k=0; k < No; k++)
for (size_t i=0; i < No; i++)
for (size_t j=0; j < No; j++) {
const size_t ijk = i + j*No + k*No*No
, jk = j + No * k
;
Zijk[ijk] += Tph[ a + i * Nv ] * VBCij[ j + k * No ];
Zijk[ijk] += Tph[ b + j * Nv ] * VACij[ i + k * No ];
Zijk[ijk] += Tph[ c + k * Nv ] * VABij[ i + j * No ];
}
}
void doublesContribution
( const ABCTuple &abc
, size_t const No
, size_t const Nv
// -- VABCI
, double const* VABph
, double const* VACph
, double const* VBCph
, double const* VBAph
, double const* VCAph
, double const* VCBph
// -- VHHHA
, double const* VhhhA
, double const* VhhhB
, double const* VhhhC
// -- TA
, double const* TAphh
, double const* TBphh
, double const* TCphh
// -- TABIJ
, double const* TABhh
, double const* TAChh
, double const* TBChh
// -- TIJK
, double *Tijk
) {
const size_t a = abc[0], b = abc[1], c = abc[2]
, NoNo = No*No, NoNv = No*Nv
;
#if defined(ATRIP_USE_DGEMM)
#define _IJK_(i, j, k) i + j*No + k*NoNo
#define REORDER(__II, __JJ, __KK) \
WITH_CHRONO("double:reorder", \
for (size_t k = 0; k < No; k++) \
for (size_t j = 0; j < No; j++) \
for (size_t i = 0; i < No; i++) { \
Tijk[_IJK_(i, j, k)] \
+= _t_buffer[_IJK_(__II, __JJ, __KK)]; \
} \
)
#define DGEMM_PARTICLES(__A, __B) \
atrip::dgemm_("T", \
"N", \
(int const*)&NoNo, \
(int const*)&No, \
(int const*)&Nv, \
&one, \
__A, \
(int const*)&Nv, \
__B, \
(int const*)&Nv, \
&zero, \
_t_buffer.data(), \
(int const*)&NoNo);
#define DGEMM_HOLES(__A, __B, __TRANSB) \
atrip::dgemm_("N", \
__TRANSB, \
(int const*)&NoNo, \
(int const*)&No, \
(int const*)&No, \
&m_one, \
__A, \
(int const*)&NoNo, \
__B, \
(int const*)&No, \
&zero, \
_t_buffer.data(), \
(int const*)&NoNo);
using F = double;
const size_t NoNoNo = No*NoNo;
std::vector<double> _t_buffer;
_t_buffer.reserve(NoNoNo);
F one{1.0}, m_one{-1.0}, zero{0.0};
WITH_CHRONO("double:reorder",
for (size_t k = 0; k < NoNoNo; k++) {
Tijk[k] = 0.0;
})
WITH_CHRONO("doubles:holes",
{ // Holes part ================================================
// VhhhC[i + k*No + L*NoNo] * TABhh[L + j*No]; H1
WITH_CHRONO("doubles:holes:1",
DGEMM_HOLES(VhhhC, TABhh, "N")
REORDER(i, k, j)
)
// VhhhC[j + k*No + L*NoNo] * TABhh[i + L*No]; H0
WITH_CHRONO("doubles:holes:2",
DGEMM_HOLES(VhhhC, TABhh, "T")
REORDER(j, k, i)
)
// VhhhB[i + j*No + L*NoNo] * TAChh[L + k*No]; H5
WITH_CHRONO("doubles:holes:3",
DGEMM_HOLES(VhhhB, TAChh, "N")
REORDER(i, j, k)
)
// VhhhB[k + j*No + L*NoNo] * TAChh[i + L*No]; H3
WITH_CHRONO("doubles:holes:4",
DGEMM_HOLES(VhhhB, TAChh, "T")
REORDER(k, j, i)
)
// VhhhA[j + i*No + L*NoNo] * TBChh[L + k*No]; H1
WITH_CHRONO("doubles:holes:5",
DGEMM_HOLES(VhhhA, TBChh, "N")
REORDER(j, i, k)
)
// VhhhA[k + i*No + L*NoNo] * TBChh[j + L*No]; H4
WITH_CHRONO("doubles:holes:6",
DGEMM_HOLES(VhhhA, TBChh, "T")
REORDER(k, i, j)
)
}
)
WITH_CHRONO("doubles:particles",
{ // Particle part ===========================================
// TAphh[E + i*Nv + j*NoNv] * VBCph[E + k*Nv]; P0
WITH_CHRONO("doubles:particles:1",
DGEMM_PARTICLES(TAphh, VBCph)
REORDER(i, j, k)
)
// TAphh[E + i*Nv + k*NoNv] * VCBph[E + j*Nv]; P3
WITH_CHRONO("doubles:particles:2",
DGEMM_PARTICLES(TAphh, VCBph)
REORDER(i, k, j)
)
// TCphh[E + k*Nv + i*NoNv] * VABph[E + j*Nv]; P5
WITH_CHRONO("doubles:particles:3",
DGEMM_PARTICLES(TCphh, VABph)
REORDER(k, i, j)
)
// TCphh[E + k*Nv + j*NoNv] * VBAph[E + i*Nv]; P2
WITH_CHRONO("doubles:particles:4",
DGEMM_PARTICLES(TCphh, VBAph)
REORDER(k, j, i)
)
// TBphh[E + j*Nv + i*NoNv] * VACph[E + k*Nv]; P1
WITH_CHRONO("doubles:particles:5",
DGEMM_PARTICLES(TBphh, VACph)
REORDER(j, i, k)
)
// TBphh[E + j*Nv + k*NoNv] * VCAph[E + i*Nv]; P4
WITH_CHRONO("doubles:particles:6",
DGEMM_PARTICLES(TBphh, VCAph)
REORDER(j, k, i)
)
}
)
#undef REORDER
#undef DGEMM_HOLES
#undef DGEMM_PARTICLES
#undef _IJK_
#else
for (size_t k = 0; k < No; k++)
for (size_t j = 0; j < No; j++)
for (size_t i = 0; i < No; i++){
const size_t ijk = i + j*No + k*NoNo
, jk = j + k*No
;
Tijk[ijk] = 0.0; // :important
// HOLE DIAGRAMS: TABHH and VHHHA
for (size_t L = 0; L < No; L++){
// t[abLj] * V[Lcik] H1
// t[baLi] * V[Lcjk] H0 TODO: conjugate T for complex
Tijk[ijk] -= TABhh[L + j*No] * VhhhC[i + k*No + L*NoNo];
Tijk[ijk] -= TABhh[i + L*No] * VhhhC[j + k*No + L*NoNo];
// t[acLk] * V[Lbij] H5
// t[caLi] * V[Lbkj] H3
Tijk[ijk] -= TAChh[L + k*No] * VhhhB[i + j*No + L*NoNo];
Tijk[ijk] -= TAChh[i + L*No] * VhhhB[k + j*No + L*NoNo];
// t[bcLk] * V[Laji] H2
// t[cbLj] * V[Laki] H4
Tijk[ijk] -= TBChh[L + k*No] * VhhhA[j + i*No + L*NoNo];
Tijk[ijk] -= TBChh[j + L*No] * VhhhA[k + i*No + L*NoNo];
}
// PARTILCE DIAGRAMS: TAPHH and VABPH
for (size_t E = 0; E < Nv; E++) {
// t[aEij] * V[bcEk] P0
// t[aEik] * V[cbEj] P3 // TODO: CHECK THIS ONE, I DONT KNOW
Tijk[ijk] += TAphh[E + i*Nv + j*NoNv] * VBCph[E + k*Nv];
Tijk[ijk] += TAphh[E + i*Nv + k*NoNv] * VCBph[E + j*Nv];
// t[cEki] * V[abEj] P5
// t[cEkj] * V[baEi] P2
Tijk[ijk] += TCphh[E + k*Nv + i*NoNv] * VABph[E + j*Nv];
Tijk[ijk] += TCphh[E + k*Nv + j*NoNv] * VBAph[E + i*Nv];
// t[bEji] * V[acEk] P1
// t[bEjk] * V[caEi] P4
Tijk[ijk] += TBphh[E + j*Nv + i*NoNv] * VACph[E + k*Nv];
Tijk[ijk] += TBphh[E + j*Nv + k*NoNv] * VCAph[E + i*Nv];
}
}
#endif
}
}Blas
The main matrix-matrix multiplication method used in this algorithm
is mainly using the DGEMM function, which we declare as
extern since it should be known only at link-time.
#pragma once
namespace atrip {
extern "C" {
void dgemm_(
const char *transa,
const char *transb,
const int *m,
const int *n,
const int *k,
double *alpha,
const double *A,
const int *lda,
const double *B,
const int *ldb,
double *beta,
double *C,
const int *ldc
);
}
}Atrip
Header
#pragma once
#include <sstream>
#include <string>
#include <map>
#include <ctf.hpp>
#include <atrip/Utils.hpp>
#define ADD_ATTRIBUTE(_type, _name, _default) \
_type _name = _default; \
Input& with_ ## _name(_type i) { \
_name = i; \
return *this; \
}
namespace atrip {
struct Atrip {
static int rank;
static int np;
static Timings chrono;
static size_t networkSend;
static size_t localSend;
static void init();
struct Input {
CTF::Tensor<double> *ei = nullptr
, *ea = nullptr
, *Tph = nullptr
, *Tpphh = nullptr
, *Vpphh = nullptr
, *Vhhhp = nullptr
, *Vppph = nullptr
;
Input& with_epsilon_i(CTF::Tensor<double> * t) { ei = t; return *this; }
Input& with_epsilon_a(CTF::Tensor<double> * t) { ea = t; return *this; }
Input& with_Tai(CTF::Tensor<double> * t) { Tph = t; return *this; }
Input& with_Tabij(CTF::Tensor<double> * t) { Tpphh = t; return *this; }
Input& with_Vabij(CTF::Tensor<double> * t) { Vpphh = t; return *this; }
Input& with_Vijka(CTF::Tensor<double> * t) { Vhhhp = t; return *this; }
Input& with_Vabci(CTF::Tensor<double> * t) { Vppph = t; return *this; }
enum TuplesDistribution {
NAIVE,
GROUP_AND_SORT,
};
ADD_ATTRIBUTE(bool, rankRoundRobin, false)
ADD_ATTRIBUTE(bool, chrono, false)
ADD_ATTRIBUTE(bool, barrier, false)
ADD_ATTRIBUTE(int, maxIterations, 0)
ADD_ATTRIBUTE(int, iterationMod, -1)
ADD_ATTRIBUTE(int, percentageMod, -1)
ADD_ATTRIBUTE(TuplesDistribution, tuplesDistribution, NAIVE)
};
struct Output {
double energy;
};
static Output run(Input const& in);
};
}
#undef ADD_ATTRIBUTEMain
#include <iomanip>
#include <atrip/Atrip.hpp>
#include <atrip/Utils.hpp>
#include <atrip/Equations.hpp>
#include <atrip/SliceUnion.hpp>
#include <atrip/Unions.hpp>
using namespace atrip;
bool RankMap::RANK_ROUND_ROBIN;
int Atrip::rank;
int Atrip::np;
Timings Atrip::chrono;
size_t Atrip::networkSend;
size_t Atrip::localSend;
// user printing block
IterationDescriptor IterationDescription::descriptor;
void atrip::registerIterationDescriptor(IterationDescriptor d) {
IterationDescription::descriptor = d;
}
void Atrip::init() {
MPI_Comm_rank(MPI_COMM_WORLD, &Atrip::rank);
MPI_Comm_size(MPI_COMM_WORLD, &Atrip::np);
Atrip::networkSend = 0;
Atrip::localSend = 0;
}
Atrip::Output Atrip::run(Atrip::Input const& in) {
const int np = Atrip::np;
const int rank = Atrip::rank;
MPI_Comm universe = in.ei->wrld->comm;
const size_t No = in.ei->lens[0];
const size_t Nv = in.ea->lens[0];
LOG(0,"Atrip") << "No: " << No << "\n";
LOG(0,"Atrip") << "Nv: " << Nv << "\n";
LOG(0,"Atrip") << "np: " << np << "\n";
// allocate the three scratches, see piecuch
std::vector<double> Tijk(No*No*No) // doubles only (see piecuch)
, Zijk(No*No*No) // singles + doubles (see piecuch)
// we need local copies of the following tensors on every
// rank
, epsi(No)
, epsa(Nv)
, Tai(No * Nv)
;
in.ei->read_all(epsi.data());
in.ea->read_all(epsa.data());
in.Tph->read_all(Tai.data());
RankMap::RANK_ROUND_ROBIN = in.rankRoundRobin;
if (RankMap::RANK_ROUND_ROBIN) {
LOG(0,"Atrip") << "Doing rank round robin slices distribution" << "\n";
} else {
LOG(0,"Atrip")
<< "Doing node > local rank round robin slices distribution" << "\n";
}
// COMMUNICATOR CONSTRUCTION ========================================={{{1
//
// Construct a new communicator living only on a single rank
int child_size = 1
, child_rank
;
const
int color = rank / child_size
, crank = rank % child_size
;
MPI_Comm child_comm;
if (np == 1) {
child_comm = universe;
} else {
MPI_Comm_split(universe, color, crank, &child_comm);
MPI_Comm_rank(child_comm, &child_rank);
MPI_Comm_size(child_comm, &child_size);
}
// BUILD SLICES PARAMETRIZED BY NV ==================================={{{1
WITH_CHRONO("nv-slices",
LOG(0,"Atrip") << "BUILD NV-SLICES\n";
TAPHH taphh(*in.Tpphh, (size_t)No, (size_t)Nv, (size_t)np, child_comm, universe);
HHHA hhha(*in.Vhhhp, (size_t)No, (size_t)Nv, (size_t)np, child_comm, universe);
)
// BUILD SLICES PARAMETRIZED BY NV x NV =============================={{{1
WITH_CHRONO("nv-nv-slices",
LOG(0,"Atrip") << "BUILD NV x NV-SLICES\n";
ABPH abph(*in.Vppph, (size_t)No, (size_t)Nv, (size_t)np, child_comm, universe);
ABHH abhh(*in.Vpphh, (size_t)No, (size_t)Nv, (size_t)np, child_comm, universe);
TABHH tabhh(*in.Tpphh, (size_t)No, (size_t)Nv, (size_t)np, child_comm, universe);
)
// all tensors
std::vector< SliceUnion* > unions = {&taphh, &hhha, &abph, &abhh, &tabhh};
// get tuples for the current rank
TuplesDistribution *distribution;
if (in.tuplesDistribution == Atrip::Input::TuplesDistribution::NAIVE) {
LOG(0,"Atrip") << "Using the naive distribution\n";
distribution = new NaiveDistribution();
} else {
LOG(0,"Atrip") << "Using the group-and-sort distribution\n";
distribution = new group_and_sort::Distribution();
}
LOG(0,"Atrip") << "BUILDING TUPLE LIST\n";
WITH_CHRONO("tuples:build",
auto const tuplesList = distribution->getTuples(Nv, universe);
)
const size_t nIterations = tuplesList.size();
{
const size_t _all_tuples = Nv * (Nv + 1) * (Nv + 2) / 6 - Nv;
LOG(0,"Atrip") << "#iterations: "
<< nIterations
<< "/"
<< nIterations * np
<< "\n";
}
const size_t
iterationMod = (in.percentageMod > 0)
? nIterations * in.percentageMod / 100
: in.iterationMod
, iteration1Percent = nIterations * 0.01
;
auto const isFakeTuple
= [&tuplesList, distribution](size_t const i) {
return distribution->tupleIsFake(tuplesList[i]);
};
auto communicateDatabase
= [ &unions
, np
] (ABCTuple const& abc, MPI_Comm const& c) -> Slice::Database {
WITH_CHRONO("db:comm:type:do",
auto MPI_LDB_ELEMENT = Slice::mpi::localDatabaseElement();
)
WITH_CHRONO("db:comm:ldb",
Slice::LocalDatabase ldb;
for (auto const& tensor: unions) {
auto const& tensorDb = tensor->buildLocalDatabase(abc);
ldb.insert(ldb.end(), tensorDb.begin(), tensorDb.end());
}
)
Slice::Database db(np * ldb.size(), ldb[0]);
WITH_CHRONO("oneshot-db:comm:allgather",
WITH_CHRONO("db:comm:allgather",
MPI_Allgather( ldb.data()
, ldb.size()
, MPI_LDB_ELEMENT
, db.data()
, ldb.size()
, MPI_LDB_ELEMENT
, c);
))
WITH_CHRONO("db:comm:type:free",
MPI_Type_free(&MPI_LDB_ELEMENT);
)
return db;
};
auto doIOPhase
= [&unions, &rank, &np, &universe] (Slice::Database const& db) {
const size_t localDBLength = db.size() / np;
size_t sendTag = 0
, recvTag = rank * localDBLength
;
// RECIEVE PHASE ======================================================
{
// At this point, we have already send to everyone that fits
auto const& begin = &db[rank * localDBLength]
, end = begin + localDBLength
;
for (auto it = begin; it != end; ++it) {
recvTag++;
auto const& el = *it;
auto& u = unionByName(unions, el.name);
WITH_DBG std::cout
<< rank << ":r"
<< "♯" << recvTag << " =>"
<< " «n" << el.name
<< ", t" << el.info.type
<< ", s" << el.info.state
<< "»"
<< " ⊙ {" << rank << "⇐" << el.info.from.rank
<< ", "
<< el.info.from.source << "}"
<< " ∴ {" << el.info.tuple[0]
<< ", "
<< el.info.tuple[1]
<< "}"
<< "\n"
;
WITH_CHRONO("db:io:recv",
u.receive(el.info, recvTag);
)
} // recv
}
// SEND PHASE =========================================================
for (size_t otherRank = 0; otherRank<np; otherRank++) {
auto const& begin = &db[otherRank * localDBLength]
, end = begin + localDBLength
;
for (auto it = begin; it != end; ++it) {
sendTag++;
Slice::LocalDatabaseElement const& el = *it;
if (el.info.from.rank != rank) continue;
auto& u = unionByName(unions, el.name);
WITH_DBG std::cout
<< rank << ":s"
<< "♯" << sendTag << " =>"
<< " «n" << el.name
<< ", t" << el.info.type
<< ", s" << el.info.state
<< "»"
<< " ⊙ {" << el.info.from.rank << "⇒" << otherRank
<< ", "
<< el.info.from.source << "}"
<< " ∴ {" << el.info.tuple[0]
<< ", "
<< el.info.tuple[1]
<< "}"
<< "\n"
;
WITH_CHRONO("db:io:send",
u.send(otherRank, el, sendTag);
)
} // send phase
} // otherRank
};
#if defined(HAVE_OCD) || defined(ATRIP_PRINT_TUPLES)
std::map<ABCTuple, double> tupleEnergies;
#endif
const double doublesFlops
= double(No)
,* double(No)
,* double(No)
,* (double(No) + double(Nv))
,* 2
,* 6
/ 1e9
;
// START MAIN LOOP ======================================================{{{1
double energy(0.);
for ( size_t i = 0, iteration = 1
; i < tuplesList.size()
; i++, iteration++
) {
Atrip::chrono["iterations"].start();
// check overhead from chrono over all iterations
WITH_CHRONO("start:stop", {})
// check overhead of doing a barrier at the beginning
WITH_CHRONO("oneshot-mpi:barrier",
WITH_CHRONO("mpi:barrier",
if (in.barrier) MPI_Barrier(universe);
))
if (iteration % iterationMod == 0 || iteration == iteration1Percent) {
if (IterationDescription::descriptor) {
IterationDescription::descriptor({
iteration,
nIterations,
Atrip::chrono["iterations"].count()
});
}
size_t networkSend;
MPI_Reduce(&Atrip::networkSend,
&networkSend,
1,
MPI_UINT64_T,
MPI_SUM,
0,
universe);
size_t localSend;
MPI_Reduce(&Atrip::localSend,
&localSend,
1,
MPI_UINT64_T,
MPI_SUM,
0,
universe);
LOG(0,"Atrip")
<< "iteration " << iteration
<< " [" << 100 * iteration / nIterations << "%]"
<< " (" << doublesFlops * iteration / Atrip::chrono["doubles"].count()
<< "GF)"
<< " (" << doublesFlops * iteration / Atrip::chrono["iterations"].count()
<< "GF)"
<< " :net " << networkSend
<< " :loc " << localSend
<< " :loc/net " << (double(localSend) / double(networkSend))
//<< " ===========================\n"
<< "\n";
// PRINT TIMINGS
if (in.chrono)
for (auto const& pair: Atrip::chrono)
LOG(1, " ") << pair.first << " :: "
<< pair.second.count()
<< std::endl;
}
const ABCTuple abc = isFakeTuple(i)
? tuplesList[tuplesList.size() - 1]
: tuplesList[i]
, *abcNext = i == (tuplesList.size() - 1)
? nullptr
: &tuplesList[i + 1]
;
WITH_CHRONO("with_rank",
WITH_RANK << " :it " << iteration
<< " :abc " << pretty_print(abc)
<< " :abcN "
<< (abcNext ? pretty_print(*abcNext) : "None")
<< "\n";
)
// COMM FIRST DATABASE ================================================{{{1
if (i == 0) {
WITH_RANK << "__first__:first database ............ \n";
const auto db = communicateDatabase(abc, universe);
WITH_RANK << "__first__:first database communicated \n";
WITH_RANK << "__first__:first database io phase \n";
doIOPhase(db);
WITH_RANK << "__first__:first database io phase DONE\n";
WITH_RANK << "__first__::::Unwrapping all slices for first database\n";
for (auto& u: unions) u->unwrapAll(abc);
WITH_RANK << "__first__::::Unwrapping slices for first database DONE\n";
MPI_Barrier(universe);
}
// COMM NEXT DATABASE ================================================={{{1
if (abcNext) {
WITH_RANK << "__comm__:" << iteration << "th communicating database\n";
WITH_CHRONO("db:comm",
const auto db = communicateDatabase(*abcNext, universe);
)
WITH_CHRONO("db:io",
doIOPhase(db);
)
WITH_RANK << "__comm__:" << iteration << "th database io phase DONE\n";
}
// COMPUTE DOUBLES ===================================================={{{1
OCD_Barrier(universe);
if (!isFakeTuple(i)) {
WITH_RANK << iteration << "-th doubles\n";
WITH_CHRONO("oneshot-unwrap",
WITH_CHRONO("unwrap",
WITH_CHRONO("unwrap:doubles",
for (auto& u: decltype(unions){&abph, &hhha, &taphh, &tabhh}) {
u->unwrapAll(abc);
}
)))
WITH_CHRONO("oneshot-doubles",
WITH_CHRONO("doubles",
doublesContribution( abc, (size_t)No, (size_t)Nv
// -- VABCI
, abph.unwrapSlice(Slice::AB, abc)
, abph.unwrapSlice(Slice::AC, abc)
, abph.unwrapSlice(Slice::BC, abc)
, abph.unwrapSlice(Slice::BA, abc)
, abph.unwrapSlice(Slice::CA, abc)
, abph.unwrapSlice(Slice::CB, abc)
// -- VHHHA
, hhha.unwrapSlice(Slice::A, abc)
, hhha.unwrapSlice(Slice::B, abc)
, hhha.unwrapSlice(Slice::C, abc)
// -- TA
, taphh.unwrapSlice(Slice::A, abc)
, taphh.unwrapSlice(Slice::B, abc)
, taphh.unwrapSlice(Slice::C, abc)
// -- TABIJ
, tabhh.unwrapSlice(Slice::AB, abc)
, tabhh.unwrapSlice(Slice::AC, abc)
, tabhh.unwrapSlice(Slice::BC, abc)
// -- TIJK
, Tijk.data()
);
WITH_RANK << iteration << "-th doubles done\n";
))
}
// COMPUTE SINGLES =================================================== {{{1
OCD_Barrier(universe);
if (!isFakeTuple(i)) {
WITH_CHRONO("oneshot-unwrap",
WITH_CHRONO("unwrap",
WITH_CHRONO("unwrap:singles",
abhh.unwrapAll(abc);
)))
WITH_CHRONO("reorder",
for (size_t I(0); I < Zijk.size(); I++) Zijk[I] = Tijk[I];
)
WITH_CHRONO("singles",
singlesContribution( No, Nv, abc
, Tai.data()
, abhh.unwrapSlice(Slice::AB, abc)
, abhh.unwrapSlice(Slice::AC, abc)
, abhh.unwrapSlice(Slice::BC, abc)
, Zijk.data());
)
}
// COMPUTE ENERGY ==================================================== {{{1
if (!isFakeTuple(i)) {
double tupleEnergy(0.);
int distinct(0);
if (abc[0] == abc[1]) distinct++;
if (abc[1] == abc[2]) distinct--;
const double epsabc(epsa[abc[0]] + epsa[abc[1]] + epsa[abc[2]]);
WITH_CHRONO("energy",
if ( distinct == 0)
tupleEnergy = getEnergyDistinct(epsabc, epsi, Tijk, Zijk);
else
tupleEnergy = getEnergySame(epsabc, epsi, Tijk, Zijk);
)
#if defined(HAVE_OCD) || defined(ATRIP_PRINT_TUPLES)
tupleEnergies[abc] = tupleEnergy;
#endif
energy += tupleEnergy;
}
// TODO: remove this
if (isFakeTuple(i)) {
// fake iterations should also unwrap whatever they got
WITH_RANK << iteration
<< "th unwrapping because of fake in "
<< i << "\n";
for (auto& u: unions) u->unwrapAll(abc);
}
#ifdef HAVE_OCD
for (auto const& u: unions) {
WITH_RANK << "__dups__:"
<< iteration
<< "-th n" << u->name << " checking duplicates\n";
u->checkForDuplicates();
}
#endif
// CLEANUP UNIONS ===================================================={{{1
OCD_Barrier(universe);
if (abcNext) {
WITH_RANK << "__gc__:" << iteration << "-th cleaning up.......\n";
for (auto& u: unions) {
u->unwrapAll(abc);
WITH_RANK << "__gc__:n" << u->name << " :it " << iteration
<< " :abc " << pretty_print(abc)
<< " :abcN " << pretty_print(*abcNext)
<< "\n";
for (auto const& slice: u->slices)
WITH_RANK << "__gc__:guts:" << slice.info << "\n";
u->clearUnusedSlicesForNext(*abcNext);
WITH_RANK << "__gc__: checking validity\n";
#ifdef HAVE_OCD
// check for validity of the slices
for (auto type: u->sliceTypes) {
auto tuple = Slice::subtupleBySlice(abc, type);
for (auto& slice: u->slices) {
if ( slice.info.type == type
&& slice.info.tuple == tuple
&& slice.isDirectlyFetchable()
) {
if (slice.info.state == Slice::Dispatched)
throw std::domain_error( "This slice should not be undispatched! "
+ pretty_print(slice.info));
}
}
}
#endif
}
}
WITH_RANK << iteration << "-th cleaning up....... DONE\n";
Atrip::chrono["iterations"].stop();
// ITERATION END ====================================================={{{1
}
// END OF MAIN LOOP
MPI_Barrier(universe);
// PRINT TUPLES ========================================================={{{1
#if defined(HAVE_OCD) || defined(ATRIP_PRINT_TUPLES)
LOG(0,"Atrip") << "tuple energies" << "\n";
for (size_t i = 0; i < np; i++) {
MPI_Barrier(universe);
for (auto const& pair: tupleEnergies) {
if (i == rank)
std::cout << pair.first[0]
<< " " << pair.first[1]
<< " " << pair.first[2]
<< std::setprecision(15) << std::setw(23)
<< " tupleEnergy: " << pair.second
<< "\n"
;
}
}
#endif
// COMMUNICATE THE ENERGIES ============================================={{{1
LOG(0,"Atrip") << "COMMUNICATING ENERGIES \n";
double globalEnergy = 0;
MPI_Reduce(&energy, &globalEnergy, 1, MPI_DOUBLE, MPI_SUM, 0, universe);
WITH_RANK << "local energy " << energy << "\n";
LOG(0, "Atrip") << "Energy: "
<< std::setprecision(15) << std::setw(23)
<< (- globalEnergy) << std::endl;
// PRINT TIMINGS {{{1
if (in.chrono)
for (auto const& pair: Atrip::chrono)
LOG(0,"atrip:chrono") << pair.first << " "
<< pair.second.count() << std::endl;
LOG(0, "atrip:flops(doubles)")
<< nIterations * doublesFlops / Atrip::chrono["doubles"].count() << "\n";
LOG(0, "atrip:flops(iterations)")
<< nIterations * doublesFlops / Atrip::chrono["iterations"].count() << "\n";
// TODO: change the sign in the getEnergy routines
return { - globalEnergy };
}Debug and Logging
Macros
#pragma once
#include <functional>
#define ATRIP_BENCHMARK
//#define ATRIP_DONT_SLICE
//#define ATRIP_WORKLOAD_DUMP
#define ATRIP_USE_DGEMM
//#define ATRIP_PRINT_TUPLES
#ifndef ATRIP_DEBUG
#define ATRIP_DEBUG 1
#endif
#if ATRIP_DEBUG == 4
# pragma message("WARNING: You have OCD debugging ABC triples " \
"expect GB of output and consult your therapist")
# include <dbg.h>
# define HAVE_OCD
# define OCD_Barrier(com) MPI_Barrier(com)
# define WITH_OCD
# define WITH_ROOT if (atrip::Atrip::rank == 0)
# define WITH_SPECIAL(r) if (atrip::Atrip::rank == r)
# define WITH_RANK std::cout << atrip::Atrip::rank << ": "
# define WITH_CRAZY_DEBUG
# define WITH_DBG
# define DBG(...) dbg(__VA_ARGS__)
#elif ATRIP_DEBUG == 3
# pragma message("WARNING: You have crazy debugging ABC triples," \
" expect GB of output")
# include <dbg.h>
# define OCD_Barrier(com)
# define WITH_OCD if (false)
# define WITH_ROOT if (atrip::Atrip::rank == 0)
# define WITH_SPECIAL(r) if (atrip::Atrip::rank == r)
# define WITH_RANK std::cout << atrip::Atrip::rank << ": "
# define WITH_CRAZY_DEBUG
# define WITH_DBG
# define DBG(...) dbg(__VA_ARGS__)
#elif ATRIP_DEBUG == 2
# pragma message("WARNING: You have some debugging info for ABC triples")
# include <dbg.h>
# define OCD_Barrier(com)
# define WITH_OCD if (false)
# define WITH_ROOT if (atrip::Atrip::rank == 0)
# define WITH_SPECIAL(r) if (atrip::Atrip::rank == r)
# define WITH_RANK std::cout << atrip::Atrip::rank << ": "
# define WITH_CRAZY_DEBUG if (false)
# define WITH_DBG
# define DBG(...) dbg(__VA_ARGS__)
#else
# define OCD_Barrier(com)
# define WITH_OCD if (false)
# define WITH_ROOT if (false)
# define WITH_SPECIAL(r) if (false)
# define WITH_RANK if (false) std::cout << atrip::Atrip::rank << ": "
# define WITH_DBG if (false)
# define WITH_CRAZY_DEBUG if (false)
# define DBG(...)
#endif
And users of the library can redefine the LOG macro
which in case of not being defined is defined as follows:
#ifndef LOG
#define LOG(level, name) if (Atrip::rank == 0) std::cout << name << ": "
#endif
Furthermore, if you do not wish to see any output from ATRIP, simply
define ATRIP_NO_OUTPUT
#ifdef ATRIP_NO_OUTPUT
# undef LOG
# define LOG(level, name) if (false) std::cout << name << ": "
#endifIteration informer
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. The input of the function is an /gallo/atrip/src/commit/52f17015df6b61691a8185bc5c41acde938b8d5a/IterationDescriptor structure and the output should be nothing.
namespace atrip {
struct IterationDescription;
using IterationDescriptor = std::function<void(IterationDescription const&)>;
struct IterationDescription {
static IterationDescriptor descriptor;
size_t currentIteration;
size_t totalIterations;
double currentElapsedTime;
};
void registerIterationDescriptor(IterationDescriptor);
}Include header
#pragma once
#include <atrip/Atrip.hpp>