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
|
// =============================================================================
// === GPUQREngine/Source/GPUQREngine_ExpertDense.cpp ==========================
// =============================================================================
//
// This file contains the dense GPUQREngine wrapper that finds the staircase,
// makes a copy of the user's front data, then calls down into the Internal
// GPUQREngine factorization routine.
//
// Other functions include:
// - GPUQREngine_Cleanup: Cleans up relevant workspaces in the dense
// factorization depending on how we're exiting.
// - GPUQREngine_FindStaircase: Finds the staircase for a front and returns
// the staircase as an Int* list
// =============================================================================
#include "GPUQREngine_Internal.hpp"
QREngineResultCode GPUQREngine_Cleanup
(
QREngineResultCode code, // The result code that we're exiting with
Front *userFronts, // The user-provided list of fronts
Front *fronts, // The internal copy of the user's fronts
Int numFronts, // The number of fronts to be factorized
Workspace *wsMongoF, // Pointer to the total GPU Front workspace
Workspace *wsMongoR // Pointer to the total CPU R workspace
);
QREngineResultCode GPUQREngine
(
size_t gpuMemorySize, // The total available GPU memory size in bytes
Front *userFronts, // The list of fronts to factorize
Int numFronts, // The number of fronts to factorize
QREngineStats *stats // An optional parameter. If present, statistics
// are collected and passed back to the caller
// via this struct
)
{
/* Allocate workspaces */
Front *fronts = (Front*) SuiteSparse_calloc(numFronts, sizeof(Front));
if(!fronts)
{
return QRENGINE_OUTOFMEMORY;
}
size_t FSize, RSize;
FSize = RSize = 0;
for(int f=0; f<numFronts; f++)
{
/* Configure the front */
Front *userFront = &(userFronts[f]);
Int m = userFront->fm;
Int n = userFront->fn;
Front *front = new (&fronts[f]) Front(f, EMPTY, m, n);
FSize += front->getNumFrontValues();
RSize += front->getNumRValues();
}
// We have to allocate page-locked CPU-GPU space to leverage asynchronous
// memory transfers. This has to be done in a way that the CUDA driver is
// aware of, which unfortunately means making a copy of the user input.
// calloc pagelocked space on CPU, and calloc space on the GPU
Workspace *wsMongoF = Workspace::allocate(FSize, // CPU and GPU
sizeof(double), true, true, true, true);
// calloc pagelocked space on the CPU. Nothing on the GPU
Workspace *wsMongoR = Workspace::allocate(RSize, // CPU
sizeof(double), true, true, false, true);
/* Cleanup and return if we ran out of memory. */
if(!wsMongoF || !wsMongoR)
{
return GPUQREngine_Cleanup (QRENGINE_OUTOFMEMORY,
userFronts, fronts, numFronts, wsMongoF, wsMongoR);
}
/* Prepare the fronts for GPU execution. */
size_t FOffset, ROffset;
FOffset = ROffset = 0;
for(int f=0; f<numFronts; f++)
{
// Set the front pointers; make the copy from user data into front data.
Front *front = &(fronts[f]);
front->F = CPU_REFERENCE(wsMongoF, double*) + FOffset;
front->gpuF = GPU_REFERENCE(wsMongoF, double*) + FOffset;
front->cpuR = CPU_REFERENCE(wsMongoR, double*) + ROffset;
FOffset += front->getNumFrontValues();
ROffset += front->getNumRValues();
/* COPY USER DATA (user's F to our F) */
Front *userFront = &(userFronts[f]);
double *userF = userFront->F;
double *F = front->F;
Int m = userFront->fm;
Int n = userFront->fn;
bool isColMajor = userFront->isColMajor;
Int ldn = userFront->ldn;
for(Int i=0; i<m; i++)
{
for(Int j=0; j<n; j++)
{
F[i*n+j] = (isColMajor ? userF[j*ldn+i] : userF[i*ldn+j]);
}
}
/* Attach either the user-specified Stair, or compute it. */
front->Stair = userFront->Stair;
if(!front->Stair) front->Stair = GPUQREngine_FindStaircase(front);
/* Cleanup and return if we ran out of memory building the staircase */
if(!front->Stair)
{
return GPUQREngine_Cleanup (QRENGINE_OUTOFMEMORY,
userFronts, fronts, numFronts, wsMongoF, wsMongoR);
}
}
/* Transfer the fronts to the GPU. */
if(!wsMongoF->transfer(cudaMemcpyHostToDevice))
{
return GPUQREngine_Cleanup (QRENGINE_GPUERROR,
userFronts, fronts, numFronts, wsMongoF, wsMongoR);
}
/* Do the factorization for this set of fronts. */
QREngineResultCode result = GPUQREngine_Internal(gpuMemorySize, fronts,
numFronts, NULL, NULL, NULL, stats);
if(result != QRENGINE_SUCCESS)
{
return GPUQREngine_Cleanup (result,
userFronts, fronts, numFronts, wsMongoF, wsMongoR);
}
/* COPY USER DATA (our R back to user's R) */
for(int f=0; f<numFronts; f++)
{
Front *userFront = &(userFronts[f]);
double *R = (&fronts[f])->cpuR;
double *userR = userFront->cpuR;
Int m = userFront->fm;
Int n = userFront->fn;
Int rank = userFront->rank;
bool isColMajor = userFront->isColMajor;
Int ldn = userFront->ldn;
for(Int i=0; i<rank; i++)
{
for(Int j=0; j<n; j++)
{
userR[i*ldn+j] = (isColMajor ? R[j*n+i] : R[i*n+j]);
}
}
}
/* Return that the factorization was successful. */
return GPUQREngine_Cleanup (QRENGINE_SUCCESS,
userFronts, fronts, numFronts, wsMongoF, wsMongoR);
}
QREngineResultCode GPUQREngine_Cleanup
(
QREngineResultCode code, // The result code that we're exiting with
Front *userFronts, // The user-provided list of fronts
Front *fronts, // The internal copy of the user's fronts
Int numFronts, // The number of fronts to be factorized
Workspace *wsMongoF, // Pointer to the total GPU Front workspace
Workspace *wsMongoR // Pointer to the total CPU R workspace
)
{
/* Cleanup fronts. */
for(int f=0; f<numFronts; f++)
{
Front *userFront = (&userFronts[f]);
Front *front = &(fronts[f]);
if(front != NULL)
{
/* If we had to attach our own stair, clean it up. */
if(userFront->Stair == NULL && front->Stair != NULL)
{
front->Stair = (Int *) SuiteSparse_free(front->Stair);
}
/* Detach front data since it's managed by the mongo. */
front->F = NULL;
}
}
fronts = (Front *) SuiteSparse_free(fronts);
/* Free the mongo structures. Note that Workspace checks for NULL. */
wsMongoF = Workspace::destroy(wsMongoF);
wsMongoR = Workspace::destroy(wsMongoR);
return code;
}
Int *GPUQREngine_FindStaircase
(
Front *front // The front whose staircase we are computing
)
{
Int fm = front->fm;
Int fn = front->fn;
double *F = front->F;
Int *Stair = (Int*) SuiteSparse_malloc(fn, sizeof(Int));
if(!F || !Stair) return NULL;
Int lastStair = 0;
for(int j=0; j<fn; j++)
{
int i;
for(i=fm-1; i>lastStair && F[i*fn+j] == 0.0; i--);
Stair[j] = lastStair = i;
}
return Stair;
}
|
