-
Notifications
You must be signed in to change notification settings - Fork 15
/
psc_fields_cuda.h
200 lines (160 loc) · 4.92 KB
/
psc_fields_cuda.h
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
#ifndef PSC_FIELDS_CUDA_H
#define PSC_FIELDS_CUDA_H
#include <mpi.h>
#include "cuda_bits.h"
#include "fields3d.hxx"
#include "fields_traits.hxx"
#include "psc_fields_single.h"
#include "mrc_json.h"
// ======================================================================
// MfieldsCuda
namespace psc
{
template <typename T, gt::size_type N>
using gtensor_device = gt::gtensor_container<psc::device_vector<T>, N>;
}
struct MfieldsCuda;
template <>
struct MfieldsCRTPInnerTypes<MfieldsCuda>
{
using Storage = psc::gtensor_device<float, 5>;
};
struct MfieldsCuda
: MfieldsBase
, MfieldsCRTP<MfieldsCuda>
{
using real_t = float;
using Real = real_t;
using Base = MfieldsCRTP<MfieldsCuda>;
using Storage = typename Base::Storage;
using space = gt::space::device;
MfieldsCuda(const Grid_t& grid, int n_fields, Int3 ibn)
: MfieldsBase{grid, n_fields, ibn},
Base(n_fields, {-ibn, grid.ldims + 2 * ibn}, grid.n_patches()),
storage_(gt::shape(Base::box().im(0), Base::box().im(1),
Base::box().im(2), n_fields, Base::n_patches()))
{
thrust::fill(storage_.data(), storage_.data() + storage_.size(), real_t{});
}
const Grid_t& grid() const { return _grid(); }
void reset(const Grid_t& new_grid) override
{
*this = MfieldsCuda(new_grid, n_comps(), ibn());
}
auto gt()
{
return gt::adapt_device(storage_.data().get(), storage_.shape());
}
auto gt() const
{
return gt::adapt_device(storage_.data().get(), storage_.shape());
}
static const Convert convert_to_, convert_from_;
const Convert& convert_to() override { return convert_to_; }
const Convert& convert_from() override { return convert_from_; }
private:
Storage storage_;
Storage& storageImpl() { return storage_; }
const Storage& storageImpl() const { return storage_; }
friend class MfieldsCRTP<MfieldsCuda>;
};
inline MfieldsSingle hostMirror(MfieldsCuda& mflds)
{
return MfieldsSingle{mflds.grid(), mflds.n_comps(), mflds.ibn()};
}
inline MfieldsSingle hostMirror(const MfieldsCuda& mflds)
{
return MfieldsSingle{mflds.grid(), mflds.n_comps(), mflds.ibn()};
}
inline void copy(const MfieldsCuda& mflds, MfieldsSingle& hmflds)
{
static int pr;
if (!pr) {
pr = prof_register("mflds to host", 1., 0, 0);
}
prof_start(pr);
// gt::copy(mflds.gt(), hmflds.storage());
thrust::copy(mflds.gt().data(), mflds.gt().data() + mflds.gt().size(),
hmflds.storage().data());
prof_stop(pr);
}
inline void copy(const MfieldsSingle& hmflds, MfieldsCuda& mflds)
{
static int pr;
if (!pr) {
pr = prof_register("mflds from host", 1., 0, 0);
}
prof_start(pr);
// gt::copy(hmflds.storage(), mflds.gt());
thrust::copy(hmflds.storage().data(),
hmflds.storage().data() + hmflds.storage().size(),
mflds.gt().data());
prof_stop(pr);
}
// ======================================================================
// MfieldsStateCuda
struct MfieldsStateCuda : MfieldsStateBase
{
using real_t = MfieldsCuda::real_t;
using Storage = MfieldsCuda::Storage;
using space = gt::space::device;
MfieldsStateCuda(const Grid_t& grid)
: MfieldsStateBase{grid, NR_FIELDS, grid.ibn},
mflds_{grid, NR_FIELDS, grid.ibn}
{}
int n_patches() const { return mflds_.n_patches(); }
int n_comps() const { return mflds_.n_comps(); }
const Int3& ib() const { return mflds_.ib(); }
const Grid_t& grid() const { return *grid_; }
static const Convert convert_to_, convert_from_;
const Convert& convert_to() override { return convert_to_; }
const Convert& convert_from() override { return convert_from_; }
MfieldsCuda& mflds() { return mflds_; }
const MfieldsCuda& mflds() const { return mflds_; }
auto gt() { return mflds_.gt(); }
auto& storage() { return mflds_.storage(); }
auto& storage() const { return mflds_.storage(); }
private:
MfieldsCuda mflds_;
};
template <>
struct Mfields_traits<MfieldsCuda>
{
static constexpr const char* name = "cuda";
static MPI_Datatype mpi_dtype() { return MPI_FLOAT; }
};
namespace detail
{
template <typename T>
struct Mfields_from_type_space<T, gt::space::device>
{
static_assert(std::is_same<T, float>::value, "CUDA only supports float");
using type = MfieldsCuda;
};
} // namespace detail
inline MfieldsSingle hostMirror(MfieldsStateCuda& mflds)
{
return hostMirror(mflds.mflds());
}
inline MfieldsSingle hostMirror(const MfieldsStateCuda& mflds)
{
return hostMirror(mflds.mflds());
}
inline void copy(const MfieldsStateCuda& mflds, MfieldsSingle& hmflds)
{
copy(mflds.mflds(), hmflds);
}
inline void copy(const MfieldsSingle& hmflds, MfieldsStateCuda& mflds)
{
copy(hmflds, mflds.mflds());
}
// ----------------------------------------------------------------------
// FIXME hacky workaround for lack of gt::view on device
template <typename E>
GT_INLINE auto view_patch(E&& gt, int p)
{
return gt::adapt_device<4>(
(>(0, 0, 0, 0, p)).get(),
{gt.shape(0), gt.shape(1), gt.shape(2), gt.shape(3)});
}
#endif