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
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
|
//------------------------------------------------------------------------------
// GB_AxB_rowscale: C = D*B, row scale with diagonal matrix D
//------------------------------------------------------------------------------
// SuiteSparse:GraphBLAS, Timothy A. Davis, (c) 2017-2022, All Rights Reserved.
// SPDX-License-Identifier: Apache-2.0
//------------------------------------------------------------------------------
#include "GB_mxm.h"
#include "GB_binop.h"
#include "GB_apply.h"
#include "GB_stringify.h"
#ifndef GBCUDA_DEV
#include "GB_binop__include.h"
#endif
#define GB_FREE_ALL GB_phybix_free (C) ;
GrB_Info GB_AxB_rowscale // C = D*B, row scale with diagonal D
(
GrB_Matrix C, // output matrix, static header
const GrB_Matrix D, // diagonal input matrix
const GrB_Matrix B, // input matrix
const GrB_Semiring semiring, // semiring that defines C=D*B
const bool flipxy, // if true, do z=fmult(b,a) vs fmult(a,b)
GB_Context Context
)
{
//--------------------------------------------------------------------------
// check inputs
//--------------------------------------------------------------------------
GrB_Info info ;
ASSERT (C != NULL && (C->static_header || GBNSTATIC)) ;
ASSERT_MATRIX_OK (D, "D for rowscale D*B", GB0) ;
ASSERT_MATRIX_OK (B, "B for rowscale D*B", GB0) ;
ASSERT (!GB_ZOMBIES (D)) ;
ASSERT (!GB_JUMBLED (D)) ;
ASSERT (!GB_PENDING (D)) ;
ASSERT (!GB_ZOMBIES (B)) ;
ASSERT (GB_JUMBLED_OK (B)) ;
ASSERT (!GB_PENDING (B)) ;
ASSERT_SEMIRING_OK (semiring, "semiring for numeric D*B", GB0) ;
ASSERT (D->vdim == B->vlen) ;
ASSERT (GB_is_diagonal (D, Context)) ;
ASSERT (!GB_IS_BITMAP (D)) ; // bitmap or full: not needed
ASSERT (!GB_IS_BITMAP (B)) ;
ASSERT (!GB_IS_FULL (D)) ;
GBURBLE ("(%s=%s*%s) ",
GB_sparsity_char_matrix (B), // C has the sparsity structure of B
GB_sparsity_char_matrix (D),
GB_sparsity_char_matrix (B)) ;
//--------------------------------------------------------------------------
// get the semiring operators
//--------------------------------------------------------------------------
GrB_BinaryOp mult = semiring->multiply ;
GrB_Type ztype = mult->ztype ;
ASSERT (ztype == semiring->add->op->ztype) ;
GB_Opcode opcode = mult->opcode ;
// GB_reduce_to_vector does not use GB_AxB_rowscale:
ASSERT (!(mult->binop_function == NULL &&
(opcode == GB_FIRST_binop_code || opcode == GB_SECOND_binop_code))) ;
//--------------------------------------------------------------------------
// determine if C is iso (ignore the monoid since it isn't used)
//--------------------------------------------------------------------------
size_t zsize = ztype->size ;
GB_void cscalar [GB_VLA(zsize)] ;
bool C_iso = GB_iso_AxB (cscalar, D, B, D->vdim, semiring, flipxy, true) ;
#ifdef GB_DEBUGIFY_DEFN
GB_debugify_mxm (C_iso, GB_sparsity (B), ztype, NULL, false, false,
semiring, flipxy, D, B) ;
#endif
//--------------------------------------------------------------------------
// copy the pattern of B into C
//--------------------------------------------------------------------------
// allocate C->x but do not initialize it
// set C->iso = C_iso OK
GB_OK (GB_dup_worker (&C, C_iso, B, false, ztype, Context)) ;
GB_void *restrict Cx = (GB_void *) C->x ;
//--------------------------------------------------------------------------
// C = D*B, row scale, compute numerical values
//--------------------------------------------------------------------------
if (GB_OPCODE_IS_POSITIONAL (opcode))
{
//----------------------------------------------------------------------
// apply a positional operator: convert C=D*B to C=op(B)
//----------------------------------------------------------------------
// determine unary operator to compute C=D*B
ASSERT (!flipxy) ;
GrB_UnaryOp op = NULL ;
if (ztype == GrB_INT64)
{
switch (opcode)
{
// first_op(D,B) becomes position_i(B)
case GB_FIRSTI_binop_code :
case GB_FIRSTJ_binop_code : op = GxB_POSITIONI_INT64 ; break;
case GB_FIRSTI1_binop_code :
case GB_FIRSTJ1_binop_code : op = GxB_POSITIONI1_INT64; break;
// second_op(D,B) becomes position_op(B)
case GB_SECONDI_binop_code : op = GxB_POSITIONI_INT64 ; break;
case GB_SECONDJ_binop_code : op = GxB_POSITIONJ_INT64 ; break;
case GB_SECONDI1_binop_code : op = GxB_POSITIONI1_INT64; break;
case GB_SECONDJ1_binop_code : op = GxB_POSITIONJ1_INT64; break;
default: ;
}
}
else
{
switch (opcode)
{
// first_op(D,B) becomes position_i(B)
case GB_FIRSTI_binop_code :
case GB_FIRSTJ_binop_code : op = GxB_POSITIONI_INT32 ; break;
case GB_FIRSTI1_binop_code :
case GB_FIRSTJ1_binop_code : op = GxB_POSITIONI1_INT32; break;
// second_op(D,B) becomes position_op(B)
case GB_SECONDI_binop_code : op = GxB_POSITIONI_INT32 ; break;
case GB_SECONDJ_binop_code : op = GxB_POSITIONJ_INT32 ; break;
case GB_SECONDI1_binop_code : op = GxB_POSITIONI1_INT32; break;
case GB_SECONDJ1_binop_code : op = GxB_POSITIONJ1_INT32; break;
default: ;
}
}
GB_OK (GB_apply_op (Cx, C->type, GB_NON_ISO,
(GB_Operator) op, // positional op
NULL, false, false, B, Context)) ;
ASSERT_MATRIX_OK (C, "rowscale positional: C = D*B output", GB0) ;
}
else if (C_iso)
{
//----------------------------------------------------------------------
// C is iso; pattern already computed above
//----------------------------------------------------------------------
GBURBLE ("(iso rowscale) ") ;
memcpy (Cx, cscalar, zsize) ;
}
else
{
//----------------------------------------------------------------------
// C is non iso
//----------------------------------------------------------------------
//----------------------------------------------------------------------
// determine if the values are accessed
//----------------------------------------------------------------------
bool op_is_first = (opcode == GB_FIRST_binop_code) ;
bool op_is_second = (opcode == GB_SECOND_binop_code) ;
bool op_is_pair = (opcode == GB_PAIR_binop_code) ;
bool D_is_pattern = false ;
bool B_is_pattern = false ;
ASSERT (!op_is_pair) ;
if (flipxy)
{
// z = fmult (b,a) will be computed
D_is_pattern = op_is_first || op_is_pair ;
B_is_pattern = op_is_second || op_is_pair ;
ASSERT (GB_IMPLIES (!D_is_pattern,
GB_Type_compatible (D->type, mult->ytype))) ;
ASSERT (GB_IMPLIES (!B_is_pattern,
GB_Type_compatible (B->type, mult->xtype))) ;
}
else
{
// z = fmult (a,b) will be computed
D_is_pattern = op_is_second || op_is_pair ;
B_is_pattern = op_is_first || op_is_pair ;
ASSERT (GB_IMPLIES (!D_is_pattern,
GB_Type_compatible (D->type, mult->xtype))) ;
ASSERT (GB_IMPLIES (!B_is_pattern,
GB_Type_compatible (B->type, mult->ytype))) ;
}
//----------------------------------------------------------------------
// determine the number of threads to use
//----------------------------------------------------------------------
GB_GET_NTHREADS_MAX (nthreads_max, chunk, Context) ;
int nthreads = GB_nthreads (GB_nnz_held (B) + B->nvec, chunk,
nthreads_max) ;
bool done = false ;
#ifndef GBCUDA_DEV
//------------------------------------------------------------------
// define the worker for the switch factory
//------------------------------------------------------------------
#define GB_DxB(mult,xname) GB (_DxB_ ## mult ## xname)
#define GB_BINOP_WORKER(mult,xname) \
{ \
info = GB_DxB(mult,xname) (C, D, B, nthreads) ; \
done = (info != GrB_NO_VALUE) ; \
} \
break ;
//------------------------------------------------------------------
// launch the switch factory
//------------------------------------------------------------------
GB_Type_code xcode, ycode, zcode ;
if (GB_binop_builtin (D->type, D_is_pattern, B->type, B_is_pattern,
mult, flipxy, &opcode, &xcode, &ycode, &zcode))
{
// C=D*B, rowscale with built-in operator
#define GB_BINOP_IS_SEMIRING_MULTIPLIER
#define GB_NO_PAIR
#include "GB_binop_factory.c"
#undef GB_BINOP_IS_SEMIRING_MULTIPLIER
}
#endif
if (!done)
{
//------------------------------------------------------------------
// C = D*B, row scale, with typecasting or user-defined operator
//------------------------------------------------------------------
//------------------------------------------------------------------
// get operators, functions, workspace, contents of D, B, and C
//------------------------------------------------------------------
GB_BURBLE_MATRIX (C, "(generic C=D*B rowscale) ") ;
GxB_binary_function fmult = mult->binop_function ;
size_t csize = C->type->size ;
size_t dsize = D_is_pattern ? 0 : D->type->size ;
size_t bsize = B_is_pattern ? 0 : B->type->size ;
size_t xsize = mult->xtype->size ;
size_t ysize = mult->ytype->size ;
// scalar workspace: because of typecasting, the x/y types need not
// be the same as the size of the D and B types.
// flipxy false: dii = (xtype) D(i,i) and bij = (ytype) B(i,j)
// flipxy true: dii = (ytype) D(i,i) and bij = (xtype) B(i,j)
size_t dii_size = flipxy ? ysize : xsize ;
size_t bij_size = flipxy ? xsize : ysize ;
GB_cast_function cast_D, cast_B ;
if (flipxy)
{
// D is typecasted to y, and B is typecasted to x
cast_D = D_is_pattern ? NULL :
GB_cast_factory (mult->ytype->code, D->type->code) ;
cast_B = B_is_pattern ? NULL :
GB_cast_factory (mult->xtype->code, B->type->code) ;
}
else
{
// D is typecasted to x, and B is typecasted to y
cast_D = D_is_pattern ? NULL :
GB_cast_factory (mult->xtype->code, D->type->code) ;
cast_B = B_is_pattern ? NULL :
GB_cast_factory (mult->ytype->code, B->type->code) ;
}
//------------------------------------------------------------------
// C = D*B via function pointers, and typecasting
//------------------------------------------------------------------
// dii = D(i,i), located in Dx [i]
#define GB_A_IS_PATTERN 0
#define GB_GETA(dii,Dx,i,D_iso) \
GB_void dii [GB_VLA(dii_size)] ; \
if (!D_is_pattern) \
{ \
cast_D (dii, Dx +(D_iso ? 0:(i)*dsize), dsize) ; \
}
// bij = B(i,j), located in Bx [pB]
#define GB_B_IS_PATTERN 0
#define GB_GETB(bij,Bx,pB,B_iso) \
GB_void bij [GB_VLA(bij_size)] ; \
if (!B_is_pattern) \
{ \
cast_B (bij, Bx +(B_iso ? 0:(pB)*bsize), bsize) ; \
}
// address of Cx [p]
#define GB_CX(p) Cx +((p)*csize)
#define GB_ATYPE GB_void
#define GB_BTYPE GB_void
#define GB_CTYPE GB_void
// no vectorization
#define GB_PRAGMA_SIMD_VECTORIZE ;
if (flipxy)
{
#define GB_BINOP(z,x,y,i,j) fmult (z,y,x)
#include "GB_AxB_rowscale_template.c"
#undef GB_BINOP
}
else
{
#define GB_BINOP(z,x,y,i,j) fmult (z,x,y)
#include "GB_AxB_rowscale_template.c"
#undef GB_BINOP
}
}
}
//--------------------------------------------------------------------------
// return result
//--------------------------------------------------------------------------
ASSERT_MATRIX_OK (C, "rowscale: C = D*B output", GB0) ;
return (GrB_SUCCESS) ;
}
|
