Initial commit of in progress developments of RAJA and KOKKOS stream
This commit is contained in:
Matthew Martineau
parent
0a738efa54
commit
45381da0b2
123
KOKKOSStream.cpp
Normal file
123
KOKKOSStream.cpp
Normal file
@ -0,0 +1,123 @@
|
||||
// Copyright (c) 2015-16 Tom Deakin, Simon McIntosh-Smith,
|
||||
// University of Bristol HPC
|
||||
//
|
||||
// For full license terms please see the LICENSE file distributed with this
|
||||
// source code
|
||||
|
||||
|
||||
#include "KOKKOSStream.hpp"
|
||||
|
||||
using Kokkos::parallel_for;
|
||||
|
||||
template <class T>
|
||||
KOKKOSStream<T>::KOKKOSStream(
|
||||
const unsigned int ARRAY_SIZE, const int device_index)
|
||||
: array_size(ARRAY_SIZE)
|
||||
{
|
||||
Kokkos::initialize();
|
||||
|
||||
new(d_a) Kokkos::View<double*, DEVICE>("d_a", ARRAY_SIZE);
|
||||
new(d_b) Kokkos::View<double*, DEVICE>("d_b", ARRAY_SIZE);
|
||||
new(d_c) Kokkos::View<double*, DEVICE>("d_c", ARRAY_SIZE);
|
||||
new(hm_a) Kokkos::View<double*>::HostMirror();
|
||||
new(hm_b) Kokkos::View<double*>::HostMirror();
|
||||
new(hm_c) Kokkos::View<double*>::HostMirror();
|
||||
hm_a = Kokkos::create_mirror_view(d_a);
|
||||
hm_b = Kokkos::create_mirror_view(d_b);
|
||||
hm_c = Kokkos::create_mirror_view(d_c);
|
||||
}
|
||||
|
||||
template <class T>
|
||||
KOKKOSStream<T>::~KOKKOSStream()
|
||||
{
|
||||
Kokkos::finalize();
|
||||
}
|
||||
|
||||
template <class T>
|
||||
void KOKKOSStream<T>::write_arrays(
|
||||
const std::vector<T>& a, const std::vector<T>& b, const std::vector<T>& c)
|
||||
{
|
||||
for(int ii = 0; ii < array_size; ++ii)
|
||||
{
|
||||
hm_a(ii) = a[ii];
|
||||
hm_b(ii) = b[ii];
|
||||
hm_c(ii) = c[ii];
|
||||
}
|
||||
Kokkos::deep_copy(hm_a, d_a);
|
||||
Kokkos::deep_copy(hm_b, d_b);
|
||||
Kokkos::deep_copy(hm_c, d_c);
|
||||
}
|
||||
|
||||
template <class T>
|
||||
void KOKKOSStream<T>::read_arrays(
|
||||
std::vector<T>& a, std::vector<T>& b, std::vector<T>& c)
|
||||
{
|
||||
Kokkos::deep_copy(d_a, hm_a);
|
||||
Kokkos::deep_copy(d_a, hm_b);
|
||||
Kokkos::deep_copy(d_a, hm_c);
|
||||
for(int ii = 0; ii < array_size; ++ii)
|
||||
{
|
||||
a[ii] = hm_a(ii);
|
||||
b[ii] = hm_b(ii);
|
||||
c[ii] = hm_c(ii);
|
||||
}
|
||||
}
|
||||
|
||||
template <class T>
|
||||
void KOKKOSStream<T>::copy()
|
||||
{
|
||||
Kokkos::parallel_for(array_size, KOKKOS_LAMBDA (const int index)
|
||||
{
|
||||
d_c[index] = d_a[index];
|
||||
});
|
||||
}
|
||||
|
||||
template <class T>
|
||||
void KOKKOSStream<T>::mul()
|
||||
{
|
||||
const T scalar = 3.0;
|
||||
parallel_for(array_size, KOKKOS_LAMBDA (const int index)
|
||||
{
|
||||
d_b[index] = scalar*d_c[index];
|
||||
});
|
||||
}
|
||||
|
||||
template <class T>
|
||||
void KOKKOSStream<T>::add()
|
||||
{
|
||||
parallel_for(array_size, KOKKOS_LAMBDA (const int index)
|
||||
{
|
||||
d_c[index] = d_a[index] + d_b[index];
|
||||
});
|
||||
}
|
||||
|
||||
template <class T>
|
||||
void KOKKOSStream<T>::triad()
|
||||
{
|
||||
const T scalar = 3.0;
|
||||
parallel_for(array_size, KOKKOS_LAMBDA (const int index)
|
||||
{
|
||||
d_a[index] = d_b[index] + scalar*d_c[index];
|
||||
});
|
||||
}
|
||||
|
||||
void listDevices(void)
|
||||
{
|
||||
std::cout << "This is not the device you are looking for.";
|
||||
}
|
||||
|
||||
|
||||
std::string getDeviceName(const int device)
|
||||
{
|
||||
return "Kokkos";
|
||||
}
|
||||
|
||||
|
||||
std::string getDeviceDriver(const int device)
|
||||
{
|
||||
return "Kokkos";
|
||||
}
|
||||
|
||||
template class KOKKOSStream<float>;
|
||||
template class KOKKOSStream<double>;
|
||||
|
||||
53
KOKKOSStream.hpp
Normal file
53
KOKKOSStream.hpp
Normal file
@ -0,0 +1,53 @@
|
||||
// Copyright (c) 2015-16 Tom Deakin, Simon McIntosh-Smith,
|
||||
// University of Bristol HPC
|
||||
//
|
||||
// For full license terms please see the LICENSE file distributed with this
|
||||
// source code
|
||||
|
||||
#pragma once
|
||||
|
||||
#include <iostream>
|
||||
#include <stdexcept>
|
||||
|
||||
#include <Kokkos_Core.hpp>
|
||||
#include <Kokkos_Parallel.hpp>
|
||||
#include <Kokkos_View.hpp>
|
||||
|
||||
#include "Stream.h"
|
||||
|
||||
#define IMPLEMENTATION_STRING "KOKKOS"
|
||||
|
||||
#define DEVICE Kokkos::OpenMP
|
||||
|
||||
|
||||
template <class T>
|
||||
class KOKKOSStream : public Stream<T>
|
||||
{
|
||||
protected:
|
||||
// Size of arrays
|
||||
unsigned int array_size;
|
||||
|
||||
// Device side pointers to arrays
|
||||
Kokkos::View<double*, DEVICE> d_a;
|
||||
Kokkos::View<double*, DEVICE> d_b;
|
||||
Kokkos::View<double*, DEVICE> d_c;
|
||||
Kokkos::View<double*>::HostMirror hm_a;
|
||||
Kokkos::View<double*>::HostMirror hm_b;
|
||||
Kokkos::View<double*>::HostMirror hm_c;
|
||||
|
||||
public:
|
||||
|
||||
KOKKOSStream(const unsigned int, const int);
|
||||
~KOKKOSStream();
|
||||
|
||||
virtual void copy() override;
|
||||
virtual void add() override;
|
||||
virtual void mul() override;
|
||||
virtual void triad() override;
|
||||
|
||||
virtual void write_arrays(
|
||||
const std::vector<T>& a, const std::vector<T>& b, const std::vector<T>& c) override;
|
||||
virtual void read_arrays(
|
||||
std::vector<T>& a, std::vector<T>& b, std::vector<T>& c) override;
|
||||
};
|
||||
|
||||
105
RAJAStream.cpp
Normal file
105
RAJAStream.cpp
Normal file
@ -0,0 +1,105 @@
|
||||
|
||||
// Copyright (c) 2015-16 Tom Deakin, Simon McIntosh-Smith,
|
||||
// University of Bristol HPC
|
||||
//
|
||||
// For full license terms please see the LICENSE file distributed with this
|
||||
// source code
|
||||
|
||||
#include "RAJAStream.hpp"
|
||||
|
||||
using RAJA::forall;
|
||||
using RAJA::RangeSegment;
|
||||
|
||||
template <class T>
|
||||
RAJAStream<T>::RAJAStream(const unsigned int ARRAY_SIZE, const int device_index)
|
||||
: array_size(ARRAY_SIZE)
|
||||
{
|
||||
RangeSegment seg(0, ARRAY_SIZE);
|
||||
index_set.push_back(seg);
|
||||
d_a = new T[ARRAY_SIZE];
|
||||
d_b = new T[ARRAY_SIZE];
|
||||
d_c = new T[ARRAY_SIZE];
|
||||
}
|
||||
|
||||
template <class T>
|
||||
RAJAStream<T>::~RAJAStream()
|
||||
{
|
||||
delete[] d_a;
|
||||
delete[] d_b;
|
||||
delete[] d_c;
|
||||
}
|
||||
|
||||
template <class T>
|
||||
void RAJAStream<T>::write_arrays(const std::vector<T>& a, const std::vector<T>& b, const std::vector<T>& c)
|
||||
{
|
||||
std::copy(a.begin(), a.end(), d_a);
|
||||
std::copy(b.begin(), b.end(), d_b);
|
||||
std::copy(c.begin(), c.end(), d_c);
|
||||
}
|
||||
|
||||
template <class T>
|
||||
void RAJAStream<T>::read_arrays(std::vector<T>& a, std::vector<T>& b, std::vector<T>& c)
|
||||
{
|
||||
std::copy(d_a, d_a + array_size - 1, a.data());
|
||||
std::copy(d_b, d_b + array_size - 1, b.data());
|
||||
std::copy(d_c, d_c + array_size - 1, c.data());
|
||||
}
|
||||
|
||||
template <class T>
|
||||
void RAJAStream<T>::copy()
|
||||
{
|
||||
forall<policy>(index_set, [=] RAJA_DEVICE (int index)
|
||||
{
|
||||
d_c[index] = d_a[index];
|
||||
});
|
||||
}
|
||||
|
||||
template <class T>
|
||||
void RAJAStream<T>::mul()
|
||||
{
|
||||
const T scalar = 3.0;
|
||||
forall<policy>(index_set, [=] RAJA_DEVICE (int index)
|
||||
{
|
||||
d_b[index] = scalar*d_c[index];
|
||||
});
|
||||
}
|
||||
|
||||
template <class T>
|
||||
void RAJAStream<T>::add()
|
||||
{
|
||||
forall<policy>(index_set, [=] RAJA_DEVICE (int index)
|
||||
{
|
||||
d_c[index] = d_a[index] + d_b[index];
|
||||
});
|
||||
}
|
||||
|
||||
template <class T>
|
||||
void RAJAStream<T>::triad()
|
||||
{
|
||||
const T scalar = 3.0;
|
||||
forall<policy>(index_set, [=] RAJA_DEVICE (int index)
|
||||
{
|
||||
d_a[index] = d_b[index] + scalar*d_c[index];
|
||||
});
|
||||
}
|
||||
|
||||
void listDevices(void)
|
||||
{
|
||||
std::cout << "This is not the device you are looking for.";
|
||||
}
|
||||
|
||||
|
||||
std::string getDeviceName(const int device)
|
||||
{
|
||||
return "RAJA";
|
||||
}
|
||||
|
||||
|
||||
std::string getDeviceDriver(const int device)
|
||||
{
|
||||
return "RAJA";
|
||||
}
|
||||
|
||||
template class RAJAStream<float>;
|
||||
template class RAJAStream<double>;
|
||||
|
||||
58
RAJAStream.hpp
Normal file
58
RAJAStream.hpp
Normal file
@ -0,0 +1,58 @@
|
||||
// Copyright (c) 2015-16 Tom Deakin, Simon McIntosh-Smith,
|
||||
// University of Bristol HPC
|
||||
//
|
||||
// For full license terms please see the LICENSE file distributed with this
|
||||
// source code
|
||||
|
||||
#pragma once
|
||||
|
||||
#include <iostream>
|
||||
#include <stdexcept>
|
||||
#include "RAJA/RAJA.hxx"
|
||||
|
||||
#include "Stream.h"
|
||||
|
||||
#define IMPLEMENTATION_STRING "RAJA"
|
||||
|
||||
#ifdef RAJA_USE_CUDA
|
||||
const size_t block_size = 128;
|
||||
typedef RAJA::IndexSet::ExecPolicy<
|
||||
RAJA::seq_segit,
|
||||
RAJA::cuda_exec_async<block_size>> policy;
|
||||
#else
|
||||
typedef RAJA::IndexSet::ExecPolicy<
|
||||
RAJA::seq_segit,
|
||||
RAJA::omp_parallel_for_exec> policy;
|
||||
#endif
|
||||
|
||||
template <class T>
|
||||
class RAJAStream : public Stream<T>
|
||||
{
|
||||
protected:
|
||||
// Size of arrays
|
||||
unsigned int array_size;
|
||||
|
||||
// Contains iteration space
|
||||
RAJA::IndexSet index_set;
|
||||
|
||||
// Device side pointers to arrays
|
||||
T* d_a;
|
||||
T* d_b;
|
||||
T* d_c;
|
||||
|
||||
public:
|
||||
|
||||
RAJAStream(const unsigned int, const int);
|
||||
~RAJAStream();
|
||||
|
||||
virtual void copy() override;
|
||||
virtual void add() override;
|
||||
virtual void mul() override;
|
||||
virtual void triad() override;
|
||||
|
||||
virtual void write_arrays(
|
||||
const std::vector<T>& a, const std::vector<T>& b, const std::vector<T>& c) override;
|
||||
virtual void read_arrays(
|
||||
std::vector<T>& a, std::vector<T>& b, std::vector<T>& c) override;
|
||||
};
|
||||
|
||||
Loading…
Reference in New Issue
Block a user