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
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
|
#include <Python.h>
#define PY_ARRAY_UNIQUE_SYMBOL GPAW_ARRAY_API
#define NO_IMPORT_ARRAY
#include <numpy/arrayobject.h>
#include <stdlib.h>
#include <pthread.h>
#include "../extensions.h"
#define __TRANSFORMERS_C
#include "../transformers.h"
#undef __TRANSFORMERS_C
#include "bmgs.h"
#include "gpu.h"
static double *transformer_buf_gpu = NULL;
static double *transformer_buf16_gpu = NULL;
static int transformer_buf_size = 0;
static int transformer_init_count = 0;
/*
* Increment reference count to register a new tranformer object.
*/
void transformer_init_gpu(TransformerObject *self)
{
transformer_init_count++;
}
/*
* Ensure buffer is allocated and is big enough. Reallocate only if
* size has increased.
*/
void transformer_init_buffers(TransformerObject *self, int blocks)
{
const boundary_conditions* bc = self->bc;
const int* size2 = bc->size2;
int ng2 = (bc->ndouble * size2[0] * size2[1] * size2[2]) * blocks;
if (ng2 > transformer_buf_size) {
gpuFree(transformer_buf_gpu);
gpuCheckLastError();
gpuMalloc(&transformer_buf_gpu, sizeof(double) * ng2);
gpuFree(transformer_buf16_gpu);
gpuCheckLastError();
gpuMalloc(&transformer_buf16_gpu, sizeof(double) * ng2 * 16);
transformer_buf_size = ng2;
}
}
/*
* Reset reference count and unset buffer.
*/
void transformer_init_buffers_gpu()
{
transformer_buf_gpu = NULL;
transformer_buf16_gpu = NULL;
transformer_buf_size = 0;
transformer_init_count = 0;
}
/*
* Deallocate buffer or decrease reference count.
*
* arguments:
* (int) force -- if true, force deallocation
*/
void transformer_dealloc_gpu(int force)
{
if (force)
transformer_init_count = 1;
if (transformer_init_count == 1) {
gpuFree(transformer_buf_gpu);
gpuCheckLastError();
transformer_init_buffers_gpu();
return;
}
if (transformer_init_count > 0)
transformer_init_count--;
}
/*
* Run the interpolate and restrict algorithm (see transapply_worker()
* in ../transformers.c) on the GPU.
*/
static void _transformer_apply_gpu(TransformerObject* self,
const double *in, double *out,
int nin, int blocks, bool real,
const double_complex *ph, bool stencil)
{
boundary_conditions* bc = self->bc;
const int* size1 = bc->size1;
int ng = bc->ndouble * size1[0] * size1[1] * size1[2];
int out_ng = bc->ndouble * self->size_out[0] * self->size_out[1]
* self->size_out[2];
int mpi_size = 1;
if ((bc->maxsend || bc->maxrecv) && bc->comm != MPI_COMM_NULL)
MPI_Comm_size(bc->comm, &mpi_size);
MPI_Request recvreq[3][2];
MPI_Request sendreq[3][2];
transformer_init_buffers(self, blocks);
double* buf = transformer_buf_gpu;
double* buf16 = transformer_buf16_gpu;
/* use stencil version, if no optimised kernel available */
if (self->k != 2) {
stencil = 1;
}
for (int n = 0; n < nin; n += blocks) {
const double* in2 = in + n * ng;
double* out2 = out + n * out_ng;
int myblocks = MIN(blocks, nin - n);
bc_unpack_paste_gpu(bc, in2, buf, recvreq, 0, myblocks);
for (int i=0; i < 3; i++) {
bc_unpack_gpu(bc, buf, i, recvreq, sendreq[i],
ph + 2 * i, 0, myblocks);
}
if (self->interpolate) {
if (stencil) {
if (real) {
bmgs_interpolate_stencil_gpu(self->k, self->skip, buf,
bc->size2, out2,
self->size_out, buf16,
myblocks);
} else {
bmgs_interpolate_stencil_gpuz(self->k, self->skip,
(gpuDoubleComplex*) (buf),
bc->size2,
(gpuDoubleComplex*) (out2),
self->size_out,
(gpuDoubleComplex*) (buf16),
myblocks);
}
} else {
if (real) {
bmgs_interpolate_gpu(self->k, self->skip, buf,
bc->size2, out2, self->size_out,
myblocks);
} else {
bmgs_interpolate_gpuz(self->k, self->skip,
(gpuDoubleComplex*) (buf),
bc->size2,
(gpuDoubleComplex*) (out2),
self->size_out, myblocks);
}
}
} else {
if (stencil) {
if (real) {
bmgs_restrict_stencil_gpu(self->k, buf, bc->size2,
out2, self->size_out, buf16,
myblocks);
} else {
bmgs_restrict_stencil_gpuz(self->k,
(gpuDoubleComplex*) (buf),
bc->size2,
(gpuDoubleComplex*) (out2),
self->size_out,
(gpuDoubleComplex*) (buf16),
myblocks);
}
} else {
if (real) {
bmgs_restrict_gpu(self->k, buf, bc->size2,
out2, self->size_out, myblocks);
} else {
bmgs_restrict_gpuz(self->k,
(gpuDoubleComplex*) (buf),
bc->size2,
(gpuDoubleComplex*) (out2),
self->size_out, myblocks);
}
}
}
}
}
/*
* Python interface for the GPU version of the interpolate and restrict
* algorithm (similar to Transformer_apply() for CPUs).
*
* arguments:
* input_gpu -- pointer to device memory (GPUArray.gpudata)
* output_gpu -- pointer to device memory (GPUArray.gpudata)
* shape -- shape of the array (tuple)
* type -- datatype of array elements
* phases -- phase (complex) (ignored if type is NPY_DOUBLE)
* stencil -- use stencil version of interpolate functions
*/
PyObject* Transformer_apply_gpu(TransformerObject *self, PyObject *args)
{
PyArrayObject* phases = 0;
void *input_gpu;
void *output_gpu;
PyObject *shape;
PyArray_Descr *type;
int stencil = 0;
if (!PyArg_ParseTuple(args, "nnOO|Oi", &input_gpu, &output_gpu, &shape,
&type, &phases, &stencil))
return NULL;
int nin = 1;
if (PyTuple_Size(shape) == 4)
nin = (int) PyLong_AsLong(PyTuple_GetItem(shape, 0));
const double* in = (double*) input_gpu;
double* out = (double*) output_gpu;
bool real = (type->type_num == NPY_DOUBLE);
const double_complex* ph = (real ? 0 : COMPLEXP(phases));
boundary_conditions* bc = self->bc;
int mpi_size = 1;
if ((bc->maxsend || bc->maxrecv) && bc->comm != MPI_COMM_NULL)
MPI_Comm_size(bc->comm, &mpi_size);
int blocks = MAX(1, MIN(nin, MIN((GPU_BLOCKS_MIN) * mpi_size,
(GPU_BLOCKS_MAX) / bc->ndouble)));
_transformer_apply_gpu(self, in, out, nin, blocks, real, ph, stencil);
if (PyErr_Occurred())
return NULL;
else
Py_RETURN_NONE;
}
|
