|
1
|
|
/*
|
|
2
|
|
* This file provides optimized sum of product implementations used internally
|
|
3
|
|
* by einsum.
|
|
4
|
|
*
|
|
5
|
|
* Copyright (c) 2011 by Mark Wiebe (mwwiebe@gmail.com)
|
|
6
|
|
* The University of British Columbia
|
|
7
|
|
*
|
|
8
|
|
* See LICENSE.txt for the license.
|
|
9
|
|
*/
|
|
10
|
|
|
|
11
|
|
#define NPY_NO_DEPRECATED_API NPY_API_VERSION
|
|
12
|
|
#define _MULTIARRAYMODULE
|
|
13
|
|
|
|
14
|
|
#include <numpy/npy_common.h>
|
|
15
|
|
#include <numpy/ndarraytypes.h> /* for NPY_NTYPES */
|
|
16
|
|
#include <numpy/halffloat.h>
|
|
17
|
|
|
|
18
|
|
#include "einsum_sumprod.h"
|
|
19
|
|
#include "einsum_debug.h"
|
|
20
|
|
|
|
21
|
|
|
|
22
|
|
#ifdef NPY_HAVE_SSE_INTRINSICS
|
|
23
|
|
#define EINSUM_USE_SSE1 1
|
|
24
|
|
#else
|
|
25
|
|
#define EINSUM_USE_SSE1 0
|
|
26
|
|
#endif
|
|
27
|
|
|
|
28
|
|
#ifdef NPY_HAVE_SSE2_INTRINSICS
|
|
29
|
|
#define EINSUM_USE_SSE2 1
|
|
30
|
|
#else
|
|
31
|
|
#define EINSUM_USE_SSE2 0
|
|
32
|
|
#endif
|
|
33
|
|
|
|
34
|
|
#if EINSUM_USE_SSE1
|
|
35
|
|
#include <xmmintrin.h>
|
|
36
|
|
#endif
|
|
37
|
|
|
|
38
|
|
#if EINSUM_USE_SSE2
|
|
39
|
|
#include <emmintrin.h>
|
|
40
|
|
#endif
|
|
41
|
|
|
|
42
|
|
#define EINSUM_IS_SSE_ALIGNED(x) ((((npy_intp)x)&0xf) == 0)
|
|
43
|
|
|
|
44
|
|
/**********************************************/
|
|
45
|
|
|
|
46
|
|
/**begin repeat
|
|
47
|
|
* #name = byte, short, int, long, longlong,
|
|
48
|
|
* ubyte, ushort, uint, ulong, ulonglong,
|
|
49
|
|
* half, float, double, longdouble,
|
|
50
|
|
* cfloat, cdouble, clongdouble#
|
|
51
|
|
* #type = npy_byte, npy_short, npy_int, npy_long, npy_longlong,
|
|
52
|
|
* npy_ubyte, npy_ushort, npy_uint, npy_ulong, npy_ulonglong,
|
|
53
|
|
* npy_half, npy_float, npy_double, npy_longdouble,
|
|
54
|
|
* npy_cfloat, npy_cdouble, npy_clongdouble#
|
|
55
|
|
* #temptype = npy_byte, npy_short, npy_int, npy_long, npy_longlong,
|
|
56
|
|
* npy_ubyte, npy_ushort, npy_uint, npy_ulong, npy_ulonglong,
|
|
57
|
|
* npy_float, npy_float, npy_double, npy_longdouble,
|
|
58
|
|
* npy_float, npy_double, npy_longdouble#
|
|
59
|
|
* #to = ,,,,,
|
|
60
|
|
* ,,,,,
|
|
61
|
|
* npy_float_to_half,,,,
|
|
62
|
|
* ,,#
|
|
63
|
|
* #from = ,,,,,
|
|
64
|
|
* ,,,,,
|
|
65
|
|
* npy_half_to_float,,,,
|
|
66
|
|
* ,,#
|
|
67
|
|
* #complex = 0*5,
|
|
68
|
|
* 0*5,
|
|
69
|
|
* 0*4,
|
|
70
|
|
* 1*3#
|
|
71
|
|
* #float32 = 0*5,
|
|
72
|
|
* 0*5,
|
|
73
|
|
* 0,1,0,0,
|
|
74
|
|
* 0*3#
|
|
75
|
|
* #float64 = 0*5,
|
|
76
|
|
* 0*5,
|
|
77
|
|
* 0,0,1,0,
|
|
78
|
|
* 0*3#
|
|
79
|
|
*/
|
|
80
|
|
|
|
81
|
|
/**begin repeat1
|
|
82
|
|
* #nop = 1, 2, 3, 1000#
|
|
83
|
|
* #noplabel = one, two, three, any#
|
|
84
|
|
*/
|
|
85
|
|
static void
|
|
86
|
1
|
@name@_sum_of_products_@noplabel@(int nop, char **dataptr,
|
|
87
|
|
npy_intp const *strides, npy_intp count)
|
|
88
|
|
{
|
|
89
|
|
#if (@nop@ == 1) || (@nop@ <= 3 && !@complex@)
|
|
90
|
1
|
char *data0 = dataptr[0];
|
|
91
|
1
|
npy_intp stride0 = strides[0];
|
|
92
|
|
#endif
|
|
93
|
|
#if (@nop@ == 2 || @nop@ == 3) && !@complex@
|
|
94
|
1
|
char *data1 = dataptr[1];
|
|
95
|
1
|
npy_intp stride1 = strides[1];
|
|
96
|
|
#endif
|
|
97
|
|
#if (@nop@ == 3) && !@complex@
|
|
98
|
1
|
char *data2 = dataptr[2];
|
|
99
|
1
|
npy_intp stride2 = strides[2];
|
|
100
|
|
#endif
|
|
101
|
|
#if (@nop@ == 1) || (@nop@ <= 3 && !@complex@)
|
|
102
|
1
|
char *data_out = dataptr[@nop@];
|
|
103
|
1
|
npy_intp stride_out = strides[@nop@];
|
|
104
|
|
#endif
|
|
105
|
|
|
|
106
|
|
NPY_EINSUM_DBG_PRINT1("@name@_sum_of_products_@noplabel@ (%d)\n", (int)count);
|
|
107
|
|
|
|
108
|
1
|
while (count--) {
|
|
109
|
|
#if !@complex@
|
|
110
|
|
# if @nop@ == 1
|
|
111
|
1
|
*(@type@ *)data_out = @to@(@from@(*(@type@ *)data0) +
|
|
112
|
1
|
@from@(*(@type@ *)data_out));
|
|
113
|
1
|
data0 += stride0;
|
|
114
|
1
|
data_out += stride_out;
|
|
115
|
|
# elif @nop@ == 2
|
|
116
|
1
|
*(@type@ *)data_out = @to@(@from@(*(@type@ *)data0) *
|
|
117
|
1
|
@from@(*(@type@ *)data1) +
|
|
118
|
1
|
@from@(*(@type@ *)data_out));
|
|
119
|
1
|
data0 += stride0;
|
|
120
|
1
|
data1 += stride1;
|
|
121
|
1
|
data_out += stride_out;
|
|
122
|
|
# elif @nop@ == 3
|
|
123
|
1
|
*(@type@ *)data_out = @to@(@from@(*(@type@ *)data0) *
|
|
124
|
1
|
@from@(*(@type@ *)data1) *
|
|
125
|
1
|
@from@(*(@type@ *)data2) +
|
|
126
|
1
|
@from@(*(@type@ *)data_out));
|
|
127
|
1
|
data0 += stride0;
|
|
128
|
1
|
data1 += stride1;
|
|
129
|
1
|
data2 += stride2;
|
|
130
|
1
|
data_out += stride_out;
|
|
131
|
|
# else
|
|
132
|
1
|
@temptype@ temp = @from@(*(@type@ *)dataptr[0]);
|
|
133
|
|
int i;
|
|
134
|
1
|
for (i = 1; i < nop; ++i) {
|
|
135
|
1
|
temp *= @from@(*(@type@ *)dataptr[i]);
|
|
136
|
|
}
|
|
137
|
1
|
*(@type@ *)dataptr[nop] = @to@(temp +
|
|
138
|
1
|
@from@(*(@type@ *)dataptr[i]));
|
|
139
|
1
|
for (i = 0; i <= nop; ++i) {
|
|
140
|
1
|
dataptr[i] += strides[i];
|
|
141
|
|
}
|
|
142
|
|
# endif
|
|
143
|
|
#else /* complex */
|
|
144
|
|
# if @nop@ == 1
|
|
145
|
0
|
((@temptype@ *)data_out)[0] = ((@temptype@ *)data0)[0] +
|
|
146
|
0
|
((@temptype@ *)data_out)[0];
|
|
147
|
0
|
((@temptype@ *)data_out)[1] = ((@temptype@ *)data0)[1] +
|
|
148
|
0
|
((@temptype@ *)data_out)[1];
|
|
149
|
0
|
data0 += stride0;
|
|
150
|
0
|
data_out += stride_out;
|
|
151
|
|
# else
|
|
152
|
|
# if @nop@ <= 3
|
|
153
|
|
#define _SUMPROD_NOP @nop@
|
|
154
|
|
# else
|
|
155
|
|
#define _SUMPROD_NOP nop
|
|
156
|
|
# endif
|
|
157
|
|
@temptype@ re, im, tmp;
|
|
158
|
|
int i;
|
|
159
|
1
|
re = ((@temptype@ *)dataptr[0])[0];
|
|
160
|
1
|
im = ((@temptype@ *)dataptr[0])[1];
|
|
161
|
1
|
for (i = 1; i < _SUMPROD_NOP; ++i) {
|
|
162
|
1
|
tmp = re * ((@temptype@ *)dataptr[i])[0] -
|
|
163
|
1
|
im * ((@temptype@ *)dataptr[i])[1];
|
|
164
|
1
|
im = re * ((@temptype@ *)dataptr[i])[1] +
|
|
165
|
1
|
im * ((@temptype@ *)dataptr[i])[0];
|
|
166
|
1
|
re = tmp;
|
|
167
|
|
}
|
|
168
|
1
|
((@temptype@ *)dataptr[_SUMPROD_NOP])[0] = re +
|
|
169
|
1
|
((@temptype@ *)dataptr[_SUMPROD_NOP])[0];
|
|
170
|
1
|
((@temptype@ *)dataptr[_SUMPROD_NOP])[1] = im +
|
|
171
|
1
|
((@temptype@ *)dataptr[_SUMPROD_NOP])[1];
|
|
172
|
|
|
|
173
|
1
|
for (i = 0; i <= _SUMPROD_NOP; ++i) {
|
|
174
|
1
|
dataptr[i] += strides[i];
|
|
175
|
|
}
|
|
176
|
|
#undef _SUMPROD_NOP
|
|
177
|
|
# endif
|
|
178
|
|
#endif
|
|
179
|
|
}
|
|
180
|
1
|
}
|
|
181
|
|
|
|
182
|
|
#if @nop@ == 1
|
|
183
|
|
|
|
184
|
|
static void
|
|
185
|
1
|
@name@_sum_of_products_contig_one(int nop, char **dataptr,
|
|
186
|
|
npy_intp const *NPY_UNUSED(strides), npy_intp count)
|
|
187
|
|
{
|
|
188
|
1
|
@type@ *data0 = (@type@ *)dataptr[0];
|
|
189
|
1
|
@type@ *data_out = (@type@ *)dataptr[1];
|
|
190
|
|
|
|
191
|
|
NPY_EINSUM_DBG_PRINT1("@name@_sum_of_products_contig_one (%d)\n",
|
|
192
|
|
(int)count);
|
|
193
|
|
|
|
194
|
|
/* This is placed before the main loop to make small counts faster */
|
|
195
|
1
|
finish_after_unrolled_loop:
|
|
196
|
1
|
switch (count) {
|
|
197
|
|
/**begin repeat2
|
|
198
|
|
* #i = 6, 5, 4, 3, 2, 1, 0#
|
|
199
|
|
*/
|
|
200
|
1
|
case @i@+1:
|
|
201
|
|
#if !@complex@
|
|
202
|
1
|
data_out[@i@] = @to@(@from@(data0[@i@]) +
|
|
203
|
1
|
@from@(data_out[@i@]));
|
|
204
|
|
#else
|
|
205
|
1
|
((@temptype@ *)data_out + 2*@i@)[0] =
|
|
206
|
1
|
((@temptype@ *)data0 + 2*@i@)[0] +
|
|
207
|
1
|
((@temptype@ *)data_out + 2*@i@)[0];
|
|
208
|
1
|
((@temptype@ *)data_out + 2*@i@)[1] =
|
|
209
|
1
|
((@temptype@ *)data0 + 2*@i@)[1] +
|
|
210
|
1
|
((@temptype@ *)data_out + 2*@i@)[1];
|
|
211
|
|
#endif
|
|
212
|
|
/**end repeat2**/
|
|
213
|
|
case 0:
|
|
214
|
1
|
return;
|
|
215
|
|
}
|
|
216
|
|
|
|
217
|
|
/* Unroll the loop by 8 */
|
|
218
|
1
|
while (count >= 8) {
|
|
219
|
1
|
count -= 8;
|
|
220
|
|
|
|
221
|
|
/**begin repeat2
|
|
222
|
|
* #i = 0, 1, 2, 3, 4, 5, 6, 7#
|
|
223
|
|
*/
|
|
224
|
|
#if !@complex@
|
|
225
|
1
|
data_out[@i@] = @to@(@from@(data0[@i@]) +
|
|
226
|
1
|
@from@(data_out[@i@]));
|
|
227
|
|
#else /* complex */
|
|
228
|
1
|
((@temptype@ *)data_out + 2*@i@)[0] =
|
|
229
|
1
|
((@temptype@ *)data0 + 2*@i@)[0] +
|
|
230
|
1
|
((@temptype@ *)data_out + 2*@i@)[0];
|
|
231
|
1
|
((@temptype@ *)data_out + 2*@i@)[1] =
|
|
232
|
1
|
((@temptype@ *)data0 + 2*@i@)[1] +
|
|
233
|
1
|
((@temptype@ *)data_out + 2*@i@)[1];
|
|
234
|
|
#endif
|
|
235
|
|
/**end repeat2**/
|
|
236
|
1
|
data0 += 8;
|
|
237
|
1
|
data_out += 8;
|
|
238
|
|
}
|
|
239
|
|
|
|
240
|
|
/* Finish off the loop */
|
|
241
|
|
goto finish_after_unrolled_loop;
|
|
242
|
|
}
|
|
243
|
|
|
|
244
|
|
#elif @nop@ == 2 && !@complex@
|
|
245
|
|
|
|
246
|
|
static void
|
|
247
|
|
@name@_sum_of_products_contig_two(int nop, char **dataptr,
|
|
248
|
|
npy_intp const *NPY_UNUSED(strides), npy_intp count)
|
|
249
|
|
{
|
|
250
|
|
@type@ *data0 = (@type@ *)dataptr[0];
|
|
251
|
|
@type@ *data1 = (@type@ *)dataptr[1];
|
|
252
|
|
@type@ *data_out = (@type@ *)dataptr[2];
|
|
253
|
|
|
|
254
|
|
#if EINSUM_USE_SSE1 && @float32@
|
|
255
|
|
__m128 a, b;
|
|
256
|
|
#elif EINSUM_USE_SSE2 && @float64@
|
|
257
|
|
__m128d a, b;
|
|
258
|
|
#endif
|
|
259
|
|
|
|
260
|
|
NPY_EINSUM_DBG_PRINT1("@name@_sum_of_products_contig_two (%d)\n",
|
|
261
|
|
(int)count);
|
|
262
|
|
|
|
263
|
|
/* This is placed before the main loop to make small counts faster */
|
|
264
|
|
finish_after_unrolled_loop:
|
|
265
|
|
switch (count) {
|
|
266
|
|
/**begin repeat2
|
|
267
|
|
* #i = 6, 5, 4, 3, 2, 1, 0#
|
|
268
|
|
*/
|
|
269
|
1
|
case @i@+1:
|
|
270
|
1
|
data_out[@i@] = @to@(@from@(data0[@i@]) *
|
|
271
|
1
|
@from@(data1[@i@]) +
|
|
272
|
1
|
@from@(data_out[@i@]));
|
|
273
|
|
/**end repeat2**/
|
|
274
|
|
case 0:
|
|
275
|
1
|
return;
|
|
276
|
|
}
|
|
277
|
|
|
|
278
|
|
#if EINSUM_USE_SSE1 && @float32@
|
|
279
|
|
/* Use aligned instructions if possible */
|
|
280
|
1
|
if (EINSUM_IS_SSE_ALIGNED(data0) && EINSUM_IS_SSE_ALIGNED(data1) &&
|
|
281
|
1
|
EINSUM_IS_SSE_ALIGNED(data_out)) {
|
|
282
|
|
/* Unroll the loop by 8 */
|
|
283
|
1
|
while (count >= 8) {
|
|
284
|
1
|
count -= 8;
|
|
285
|
|
|
|
286
|
|
/**begin repeat2
|
|
287
|
|
* #i = 0, 4#
|
|
288
|
|
*/
|
|
289
|
1
|
a = _mm_mul_ps(_mm_load_ps(data0+@i@), _mm_load_ps(data1+@i@));
|
|
290
|
1
|
b = _mm_add_ps(a, _mm_load_ps(data_out+@i@));
|
|
291
|
1
|
_mm_store_ps(data_out+@i@, b);
|
|
292
|
|
/**end repeat2**/
|
|
293
|
1
|
data0 += 8;
|
|
294
|
1
|
data1 += 8;
|
|
295
|
1
|
data_out += 8;
|
|
296
|
|
}
|
|
297
|
|
|
|
298
|
|
/* Finish off the loop */
|
|
299
|
|
goto finish_after_unrolled_loop;
|
|
300
|
|
}
|
|
301
|
|
#elif EINSUM_USE_SSE2 && @float64@
|
|
302
|
|
/* Use aligned instructions if possible */
|
|
303
|
|
if (EINSUM_IS_SSE_ALIGNED(data0) && EINSUM_IS_SSE_ALIGNED(data1) &&
|
|
304
|
|
EINSUM_IS_SSE_ALIGNED(data_out)) {
|
|
305
|
|
/* Unroll the loop by 8 */
|
|
306
|
1
|
while (count >= 8) {
|
|
307
|
1
|
count -= 8;
|
|
308
|
|
|
|
309
|
1
|
/**begin repeat2
|
|
310
|
1
|
* #i = 0, 2, 4, 6#
|
|
311
|
|
*/
|
|
312
|
1
|
a = _mm_mul_pd(_mm_load_pd(data0+@i@), _mm_load_pd(data1+@i@));
|
|
313
|
1
|
b = _mm_add_pd(a, _mm_load_pd(data_out+@i@));
|
|
314
|
1
|
_mm_store_pd(data_out+@i@, b);
|
|
315
|
|
/**end repeat2**/
|
|
316
|
1
|
data0 += 8;
|
|
317
|
1
|
data1 += 8;
|
|
318
|
1
|
data_out += 8;
|
|
319
|
|
}
|
|
320
|
|
|
|
321
|
|
/* Finish off the loop */
|
|
322
|
|
goto finish_after_unrolled_loop;
|
|
323
|
|
}
|
|
324
|
|
#endif
|
|
325
|
|
|
|
326
|
|
/* Unroll the loop by 8 */
|
|
327
|
1
|
while (count >= 8) {
|
|
328
|
1
|
count -= 8;
|
|
329
|
|
|
|
330
|
|
#if EINSUM_USE_SSE1 && @float32@
|
|
331
|
|
/**begin repeat2
|
|
332
|
|
* #i = 0, 4#
|
|
333
|
|
*/
|
|
334
|
1
|
a = _mm_mul_ps(_mm_loadu_ps(data0+@i@), _mm_loadu_ps(data1+@i@));
|
|
335
|
1
|
b = _mm_add_ps(a, _mm_loadu_ps(data_out+@i@));
|
|
336
|
1
|
_mm_storeu_ps(data_out+@i@, b);
|
|
337
|
|
/**end repeat2**/
|
|
338
|
|
#elif EINSUM_USE_SSE2 && @float64@
|
|
339
|
|
/**begin repeat2
|
|
340
|
1
|
* #i = 0, 2, 4, 6#
|
|
341
|
1
|
*/
|
|
342
|
1
|
a = _mm_mul_pd(_mm_loadu_pd(data0+@i@), _mm_loadu_pd(data1+@i@));
|
|
343
|
1
|
b = _mm_add_pd(a, _mm_loadu_pd(data_out+@i@));
|
|
344
|
1
|
_mm_storeu_pd(data_out+@i@, b);
|
|
345
|
|
/**end repeat2**/
|
|
346
|
|
#else
|
|
347
|
|
/**begin repeat2
|
|
348
|
|
* #i = 0, 1, 2, 3, 4, 5, 6, 7#
|
|
349
|
|
*/
|
|
350
|
1
|
data_out[@i@] = @to@(@from@(data0[@i@]) *
|
|
351
|
1
|
@from@(data1[@i@]) +
|
|
352
|
1
|
@from@(data_out[@i@]));
|
|
353
|
|
/**end repeat2**/
|
|
354
|
|
#endif
|
|
355
|
1
|
data0 += 8;
|
|
356
|
1
|
data1 += 8;
|
|
357
|
1
|
data_out += 8;
|
|
358
|
|
}
|
|
359
|
|
|
|
360
|
|
/* Finish off the loop */
|
|
361
|
|
goto finish_after_unrolled_loop;
|
|
362
|
|
}
|
|
363
|
|
|
|
364
|
|
/* Some extra specializations for the two operand case */
|
|
365
|
|
static void
|
|
366
|
1
|
@name@_sum_of_products_stride0_contig_outcontig_two(int nop, char **dataptr,
|
|
367
|
|
npy_intp const *NPY_UNUSED(strides), npy_intp count)
|
|
368
|
|
{
|
|
369
|
1
|
@temptype@ value0 = @from@(*(@type@ *)dataptr[0]);
|
|
370
|
1
|
@type@ *data1 = (@type@ *)dataptr[1];
|
|
371
|
1
|
@type@ *data_out = (@type@ *)dataptr[2];
|
|
372
|
|
|
|
373
|
|
#if EINSUM_USE_SSE1 && @float32@
|
|
374
|
|
__m128 a, b, value0_sse;
|
|
375
|
|
#elif EINSUM_USE_SSE2 && @float64@
|
|
376
|
|
__m128d a, b, value0_sse;
|
|
377
|
|
#endif
|
|
378
|
|
|
|
379
|
|
NPY_EINSUM_DBG_PRINT1("@name@_sum_of_products_stride0_contig_outcontig_two (%d)\n",
|
|
380
|
|
(int)count);
|
|
381
|
|
|
|
382
|
|
/* This is placed before the main loop to make small counts faster */
|
|
383
|
1
|
finish_after_unrolled_loop:
|
|
384
|
1
|
switch (count) {
|
|
385
|
|
/**begin repeat2
|
|
386
|
|
* #i = 6, 5, 4, 3, 2, 1, 0#
|
|
387
|
|
*/
|
|
388
|
1
|
case @i@+1:
|
|
389
|
1
|
data_out[@i@] = @to@(value0 *
|
|
390
|
1
|
@from@(data1[@i@]) +
|
|
391
|
1
|
@from@(data_out[@i@]));
|
|
392
|
|
/**end repeat2**/
|
|
393
|
|
case 0:
|
|
394
|
|
return;
|
|
395
|
|
}
|
|
396
|
|
|
|
397
|
|
#if EINSUM_USE_SSE1 && @float32@
|
|
398
|
1
|
value0_sse = _mm_set_ps1(value0);
|
|
399
|
1
|
|
|
400
|
|
/* Use aligned instructions if possible */
|
|
401
|
1
|
if (EINSUM_IS_SSE_ALIGNED(data1) && EINSUM_IS_SSE_ALIGNED(data_out)) {
|
|
402
|
1
|
/* Unroll the loop by 8 */
|
|
403
|
1
|
while (count >= 8) {
|
|
404
|
1
|
count -= 8;
|
|
405
|
|
|
|
406
|
|
/**begin repeat2
|
|
407
|
|
* #i = 0, 4#
|
|
408
|
|
*/
|
|
409
|
1
|
a = _mm_mul_ps(value0_sse, _mm_load_ps(data1+@i@));
|
|
410
|
1
|
b = _mm_add_ps(a, _mm_load_ps(data_out+@i@));
|
|
411
|
1
|
_mm_store_ps(data_out+@i@, b);
|
|
412
|
1
|
/**end repeat2**/
|
|
413
|
1
|
data1 += 8;
|
|
414
|
1
|
data_out += 8;
|
|
415
|
1
|
}
|
|
416
|
1
|
|
|
417
|
1
|
/* Finish off the loop */
|
|
418
|
1
|
if (count > 0) {
|
|
419
|
|
goto finish_after_unrolled_loop;
|
|
420
|
|
}
|
|
421
|
1
|
else {
|
|
422
|
1
|
return;
|
|
423
|
1
|
}
|
|
424
|
|
}
|
|
425
|
|
#elif EINSUM_USE_SSE2 && @float64@
|
|
426
|
|
value0_sse = _mm_set1_pd(value0);
|
|
427
|
|
|
|
428
|
|
/* Use aligned instructions if possible */
|
|
429
|
1
|
if (EINSUM_IS_SSE_ALIGNED(data1) && EINSUM_IS_SSE_ALIGNED(data_out)) {
|
|
430
|
1
|
/* Unroll the loop by 8 */
|
|
431
|
|
while (count >= 8) {
|
|
432
|
1
|
count -= 8;
|
|
433
|
|
|
|
434
|
1
|
/**begin repeat2
|
|
435
|
1
|
* #i = 0, 2, 4, 6#
|
|
436
|
1
|
*/
|
|
437
|
1
|
a = _mm_mul_pd(value0_sse, _mm_load_pd(data1+@i@));
|
|
438
|
1
|
b = _mm_add_pd(a, _mm_load_pd(data_out+@i@));
|
|
439
|
1
|
_mm_store_pd(data_out+@i@, b);
|
|
440
|
|
/**end repeat2**/
|
|
441
|
1
|
data1 += 8;
|
|
442
|
1
|
data_out += 8;
|
|
443
|
|
}
|
|
444
|
|
|
|
445
|
|
/* Finish off the loop */
|
|
446
|
1
|
if (count > 0) {
|
|
447
|
|
goto finish_after_unrolled_loop;
|
|
448
|
|
}
|
|
449
|
|
else {
|
|
450
|
|
return;
|
|
451
|
|
}
|
|
452
|
|
}
|
|
453
|
|
#endif
|
|
454
|
|
|
|
455
|
|
/* Unroll the loop by 8 */
|
|
456
|
1
|
while (count >= 8) {
|
|
457
|
1
|
count -= 8;
|
|
458
|
|
|
|
459
|
|
#if EINSUM_USE_SSE1 && @float32@
|
|
460
|
|
/**begin repeat2
|
|
461
|
|
* #i = 0, 4#
|
|
462
|
|
*/
|
|
463
|
1
|
a = _mm_mul_ps(value0_sse, _mm_loadu_ps(data1+@i@));
|
|
464
|
1
|
b = _mm_add_ps(a, _mm_loadu_ps(data_out+@i@));
|
|
465
|
1
|
_mm_storeu_ps(data_out+@i@, b);
|
|
466
|
|
/**end repeat2**/
|
|
467
|
|
#elif EINSUM_USE_SSE2 && @float64@
|
|
468
|
|
/**begin repeat2
|
|
469
|
1
|
* #i = 0, 2, 4, 6#
|
|
470
|
1
|
*/
|
|
471
|
1
|
a = _mm_mul_pd(value0_sse, _mm_loadu_pd(data1+@i@));
|
|
472
|
1
|
b = _mm_add_pd(a, _mm_loadu_pd(data_out+@i@));
|
|
473
|
1
|
_mm_storeu_pd(data_out+@i@, b);
|
|
474
|
|
/**end repeat2**/
|
|
475
|
|
#else
|
|
476
|
|
/**begin repeat2
|
|
477
|
|
* #i = 0, 1, 2, 3, 4, 5, 6, 7#
|
|
478
|
|
*/
|
|
479
|
1
|
data_out[@i@] = @to@(value0 *
|
|
480
|
1
|
@from@(data1[@i@]) +
|
|
481
|
1
|
@from@(data_out[@i@]));
|
|
482
|
|
/**end repeat2**/
|
|
483
|
|
#endif
|
|
484
|
1
|
data1 += 8;
|
|
485
|
1
|
data_out += 8;
|
|
486
|
|
}
|
|
487
|
|
|
|
488
|
|
/* Finish off the loop */
|
|
489
|
1
|
if (count > 0) {
|
|
490
|
|
goto finish_after_unrolled_loop;
|
|
491
|
|
}
|
|
492
|
|
}
|
|
493
|
|
|
|
494
|
|
static void
|
|
495
|
1
|
@name@_sum_of_products_contig_stride0_outcontig_two(int nop, char **dataptr,
|
|
496
|
|
npy_intp const *NPY_UNUSED(strides), npy_intp count)
|
|
497
|
|
{
|
|
498
|
1
|
@type@ *data0 = (@type@ *)dataptr[0];
|
|
499
|
1
|
@temptype@ value1 = @from@(*(@type@ *)dataptr[1]);
|
|
500
|
1
|
@type@ *data_out = (@type@ *)dataptr[2];
|
|
501
|
|
|
|
502
|
|
#if EINSUM_USE_SSE1 && @float32@
|
|
503
|
|
__m128 a, b, value1_sse;
|
|
504
|
|
#elif EINSUM_USE_SSE2 && @float64@
|
|
505
|
|
__m128d a, b, value1_sse;
|
|
506
|
|
#endif
|
|
507
|
|
|
|
508
|
|
NPY_EINSUM_DBG_PRINT1("@name@_sum_of_products_contig_stride0_outcontig_two (%d)\n",
|
|
509
|
|
(int)count);
|
|
510
|
|
|
|
511
|
|
/* This is placed before the main loop to make small counts faster */
|
|
512
|
1
|
finish_after_unrolled_loop:
|
|
513
|
1
|
switch (count) {
|
|
514
|
|
/**begin repeat2
|
|
515
|
|
* #i = 6, 5, 4, 3, 2, 1, 0#
|
|
516
|
|
*/
|
|
517
|
1
|
case @i@+1:
|
|
518
|
1
|
data_out[@i@] = @to@(@from@(data0[@i@])*
|
|
519
|
1
|
value1 +
|
|
520
|
1
|
@from@(data_out[@i@]));
|
|
521
|
|
/**end repeat2**/
|
|
522
|
1
|
case 0:
|
|
523
|
1
|
return;
|
|
524
|
|
}
|
|
525
|
|
|
|
526
|
|
#if EINSUM_USE_SSE1 && @float32@
|
|
527
|
1
|
value1_sse = _mm_set_ps1(value1);
|
|
528
|
1
|
|
|
529
|
|
/* Use aligned instructions if possible */
|
|
530
|
1
|
if (EINSUM_IS_SSE_ALIGNED(data0) && EINSUM_IS_SSE_ALIGNED(data_out)) {
|
|
531
|
1
|
/* Unroll the loop by 8 */
|
|
532
|
1
|
while (count >= 8) {
|
|
533
|
1
|
count -= 8;
|
|
534
|
|
|
|
535
|
1
|
/**begin repeat2
|
|
536
|
|
* #i = 0, 4#
|
|
537
|
|
*/
|
|
538
|
1
|
a = _mm_mul_ps(_mm_load_ps(data0+@i@), value1_sse);
|
|
539
|
1
|
b = _mm_add_ps(a, _mm_load_ps(data_out+@i@));
|
|
540
|
1
|
_mm_store_ps(data_out+@i@, b);
|
|
541
|
1
|
/**end repeat2**/
|
|
542
|
1
|
data0 += 8;
|
|
543
|
1
|
data_out += 8;
|
|
544
|
1
|
}
|
|
545
|
1
|
|
|
546
|
1
|
/* Finish off the loop */
|
|
547
|
|
goto finish_after_unrolled_loop;
|
|
548
|
|
}
|
|
549
|
|
#elif EINSUM_USE_SSE2 && @float64@
|
|
550
|
|
value1_sse = _mm_set1_pd(value1);
|
|
551
|
|
|
|
552
|
|
/* Use aligned instructions if possible */
|
|
553
|
1
|
if (EINSUM_IS_SSE_ALIGNED(data0) && EINSUM_IS_SSE_ALIGNED(data_out)) {
|
|
554
|
|
/* Unroll the loop by 8 */
|
|
555
|
|
while (count >= 8) {
|
|
556
|
1
|
count -= 8;
|
|
557
|
|
|
|
558
|
1
|
/**begin repeat2
|
|
559
|
1
|
* #i = 0, 2, 4, 6#
|
|
560
|
|
*/
|
|
561
|
1
|
a = _mm_mul_pd(_mm_load_pd(data0+@i@), value1_sse);
|
|
562
|
1
|
b = _mm_add_pd(a, _mm_load_pd(data_out+@i@));
|
|
563
|
1
|
_mm_store_pd(data_out+@i@, b);
|
|
564
|
|
/**end repeat2**/
|
|
565
|
1
|
data0 += 8;
|
|
566
|
1
|
data_out += 8;
|
|
567
|
|
}
|
|
568
|
|
|
|
569
|
|
/* Finish off the loop */
|
|
570
|
|
goto finish_after_unrolled_loop;
|
|
571
|
|
}
|
|
572
|
|
#endif
|
|
573
|
|
|
|
574
|
|
/* Unroll the loop by 8 */
|
|
575
|
1
|
while (count >= 8) {
|
|
576
|
1
|
count -= 8;
|
|
577
|
|
|
|
578
|
|
#if EINSUM_USE_SSE1 && @float32@
|
|
579
|
|
/**begin repeat2
|
|
580
|
|
* #i = 0, 4#
|
|
581
|
|
*/
|
|
582
|
1
|
a = _mm_mul_ps(_mm_loadu_ps(data0+@i@), value1_sse);
|
|
583
|
1
|
b = _mm_add_ps(a, _mm_loadu_ps(data_out+@i@));
|
|
584
|
1
|
_mm_storeu_ps(data_out+@i@, b);
|
|
585
|
|
/**end repeat2**/
|
|
586
|
|
#elif EINSUM_USE_SSE2 && @float64@
|
|
587
|
|
/**begin repeat2
|
|
588
|
1
|
* #i = 0, 2, 4, 6#
|
|
589
|
1
|
*/
|
|
590
|
1
|
a = _mm_mul_pd(_mm_loadu_pd(data0+@i@), value1_sse);
|
|
591
|
1
|
b = _mm_add_pd(a, _mm_loadu_pd(data_out+@i@));
|
|
592
|
1
|
_mm_storeu_pd(data_out+@i@, b);
|
|
593
|
|
/**end repeat2**/
|
|
594
|
|
#else
|
|
595
|
|
/**begin repeat2
|
|
596
|
|
* #i = 0, 1, 2, 3, 4, 5, 6, 7#
|
|
597
|
|
*/
|
|
598
|
1
|
data_out[@i@] = @to@(@from@(data0[@i@])*
|
|
599
|
1
|
value1 +
|
|
600
|
1
|
@from@(data_out[@i@]));
|
|
601
|
|
/**end repeat2**/
|
|
602
|
|
#endif
|
|
603
|
1
|
data0 += 8;
|
|
604
|
1
|
data_out += 8;
|
|
605
|
|
}
|
|
606
|
|
|
|
607
|
|
/* Finish off the loop */
|
|
608
|
|
goto finish_after_unrolled_loop;
|
|
609
|
|
}
|
|
610
|
|
|
|
611
|
|
static void
|
|
612
|
1
|
@name@_sum_of_products_contig_contig_outstride0_two(int nop, char **dataptr,
|
|
613
|
|
npy_intp const *NPY_UNUSED(strides), npy_intp count)
|
|
614
|
|
{
|
|
615
|
1
|
@type@ *data0 = (@type@ *)dataptr[0];
|
|
616
|
1
|
@type@ *data1 = (@type@ *)dataptr[1];
|
|
617
|
1
|
@temptype@ accum = 0;
|
|
618
|
|
|
|
619
|
|
#if EINSUM_USE_SSE1 && @float32@
|
|
620
|
|
__m128 a, accum_sse = _mm_setzero_ps();
|
|
621
|
|
#elif EINSUM_USE_SSE2 && @float64@
|
|
622
|
|
__m128d a, accum_sse = _mm_setzero_pd();
|
|
623
|
|
#endif
|
|
624
|
|
|
|
625
|
|
NPY_EINSUM_DBG_PRINT1("@name@_sum_of_products_contig_contig_outstride0_two (%d)\n",
|
|
626
|
|
(int)count);
|
|
627
|
|
|
|
628
|
|
/* This is placed before the main loop to make small counts faster */
|
|
629
|
1
|
finish_after_unrolled_loop:
|
|
630
|
1
|
switch (count) {
|
|
631
|
|
/**begin repeat2
|
|
632
|
|
* #i = 6, 5, 4, 3, 2, 1, 0#
|
|
633
|
|
*/
|
|
634
|
1
|
case @i@+1:
|
|
635
|
1
|
accum += @from@(data0[@i@]) * @from@(data1[@i@]);
|
|
636
|
1
|
/**end repeat2**/
|
|
637
|
1
|
case 0:
|
|
638
|
1
|
*(@type@ *)dataptr[2] = @to@(@from@(*(@type@ *)dataptr[2]) + accum);
|
|
639
|
1
|
return;
|
|
640
|
|
}
|
|
641
|
|
|
|
642
|
|
#if EINSUM_USE_SSE1 && @float32@
|
|
643
|
|
/* Use aligned instructions if possible */
|
|
644
|
1
|
if (EINSUM_IS_SSE_ALIGNED(data0) && EINSUM_IS_SSE_ALIGNED(data1)) {
|
|
645
|
1
|
/* Unroll the loop by 8 */
|
|
646
|
1
|
while (count >= 8) {
|
|
647
|
1
|
count -= 8;
|
|
648
|
1
|
|
|
649
|
1
|
_mm_prefetch(data0 + 512, _MM_HINT_T0);
|
|
650
|
1
|
_mm_prefetch(data1 + 512, _MM_HINT_T0);
|
|
651
|
|
|
|
652
|
|
/**begin repeat2
|
|
653
|
|
* #i = 0, 4#
|
|
654
|
|
*/
|
|
655
|
1
|
/*
|
|
656
|
|
* NOTE: This accumulation changes the order, so will likely
|
|
657
|
|
* produce slightly different results.
|
|
658
|
1
|
*/
|
|
659
|
1
|
a = _mm_mul_ps(_mm_load_ps(data0+@i@), _mm_load_ps(data1+@i@));
|
|
660
|
1
|
accum_sse = _mm_add_ps(accum_sse, a);
|
|
661
|
1
|
/**end repeat2**/
|
|
662
|
1
|
data0 += 8;
|
|
663
|
1
|
data1 += 8;
|
|
664
|
|
}
|
|
665
|
|
|
|
666
|
1
|
/* Add the four SSE values and put in accum */
|
|
667
|
1
|
a = _mm_shuffle_ps(accum_sse, accum_sse, _MM_SHUFFLE(2,3,0,1));
|
|
668
|
1
|
accum_sse = _mm_add_ps(a, accum_sse);
|
|
669
|
1
|
a = _mm_shuffle_ps(accum_sse, accum_sse, _MM_SHUFFLE(1,0,3,2));
|
|
670
|
1
|
accum_sse = _mm_add_ps(a, accum_sse);
|
|
671
|
|
_mm_store_ss(&accum, accum_sse);
|
|
672
|
|
|
|
673
|
|
/* Finish off the loop */
|
|
674
|
|
goto finish_after_unrolled_loop;
|
|
675
|
|
}
|
|
676
|
1
|
#elif EINSUM_USE_SSE2 && @float64@
|
|
677
|
|
/* Use aligned instructions if possible */
|
|
678
|
|
if (EINSUM_IS_SSE_ALIGNED(data0) && EINSUM_IS_SSE_ALIGNED(data1)) {
|
|
679
|
|
/* Unroll the loop by 8 */
|
|
680
|
|
while (count >= 8) {
|
|
681
|
|
count -= 8;
|
|
682
|
|
|
|
683
|
|
_mm_prefetch(data0 + 512, _MM_HINT_T0);
|
|
684
|
1
|
_mm_prefetch(data1 + 512, _MM_HINT_T0);
|
|
685
|
|
|
|
686
|
1
|
/**begin repeat2
|
|
687
|
1
|
* #i = 0, 2, 4, 6#
|
|
688
|
|
*/
|
|
689
|
1
|
/*
|
|
690
|
1
|
* NOTE: This accumulation changes the order, so will likely
|
|
691
|
|
* produce slightly different results.
|
|
692
|
|
*/
|
|
693
|
1
|
a = _mm_mul_pd(_mm_load_pd(data0+@i@), _mm_load_pd(data1+@i@));
|
|
694
|
1
|
accum_sse = _mm_add_pd(accum_sse, a);
|
|
695
|
|
/**end repeat2**/
|
|
696
|
1
|
data0 += 8;
|
|
697
|
1
|
data1 += 8;
|
|
698
|
|
}
|
|
699
|
|
|
|
700
|
|
/* Add the two SSE2 values and put in accum */
|
|
701
|
1
|
a = _mm_shuffle_pd(accum_sse, accum_sse, _MM_SHUFFLE2(0,1));
|
|
702
|
1
|
accum_sse = _mm_add_pd(a, accum_sse);
|
|
703
|
|
_mm_store_sd(&accum, accum_sse);
|
|
704
|
|
|
|
705
|
|
/* Finish off the loop */
|
|
706
|
|
goto finish_after_unrolled_loop;
|
|
707
|
|
}
|
|
708
|
|
#endif
|
|
709
|
|
|
|
710
|
|
/* Unroll the loop by 8 */
|
|
711
|
1
|
while (count >= 8) {
|
|
712
|
1
|
count -= 8;
|
|
713
|
|
|
|
714
|
|
#if EINSUM_USE_SSE1 && @float32@
|
|
715
|
|
_mm_prefetch(data0 + 512, _MM_HINT_T0);
|
|
716
|
|
_mm_prefetch(data1 + 512, _MM_HINT_T0);
|
|
717
|
|
|
|
718
|
|
/**begin repeat2
|
|
719
|
|
* #i = 0, 4#
|
|
720
|
|
*/
|
|
721
|
|
/*
|
|
722
|
|
* NOTE: This accumulation changes the order, so will likely
|
|
723
|
|
* produce slightly different results.
|
|
724
|
|
*/
|
|
725
|
1
|
a = _mm_mul_ps(_mm_loadu_ps(data0+@i@), _mm_loadu_ps(data1+@i@));
|
|
726
|
1
|
accum_sse = _mm_add_ps(accum_sse, a);
|
|
727
|
|
/**end repeat2**/
|
|
728
|
|
#elif EINSUM_USE_SSE2 && @float64@
|
|
729
|
|
_mm_prefetch(data0 + 512, _MM_HINT_T0);
|
|
730
|
|
_mm_prefetch(data1 + 512, _MM_HINT_T0);
|
|
731
|
|
|
|
732
|
|
/**begin repeat2
|
|
733
|
1
|
* #i = 0, 2, 4, 6#
|
|
734
|
1
|
*/
|
|
735
|
1
|
/*
|
|
736
|
1
|
* NOTE: This accumulation changes the order, so will likely
|
|
737
|
1
|
* produce slightly different results.
|
|
738
|
1
|
*/
|
|
739
|
1
|
a = _mm_mul_pd(_mm_loadu_pd(data0+@i@), _mm_loadu_pd(data1+@i@));
|
|
740
|
1
|
accum_sse = _mm_add_pd(accum_sse, a);
|
|
741
|
|
/**end repeat2**/
|
|
742
|
|
#else
|
|
743
|
|
/**begin repeat2
|
|
744
|
|
* #i = 0, 1, 2, 3, 4, 5, 6, 7#
|
|
745
|
|
*/
|
|
746
|
1
|
accum += @from@(data0[@i@]) * @from@(data1[@i@]);
|
|
747
|
|
/**end repeat2**/
|
|
748
|
|
#endif
|
|
749
|
1
|
data0 += 8;
|
|
750
|
1
|
data1 += 8;
|
|
751
|
|
}
|
|
752
|
|
|
|
753
|
|
#if EINSUM_USE_SSE1 && @float32@
|
|
754
|
|
/* Add the four SSE values and put in accum */
|
|
755
|
|
a = _mm_shuffle_ps(accum_sse, accum_sse, _MM_SHUFFLE(2,3,0,1));
|
|
756
|
|
accum_sse = _mm_add_ps(a, accum_sse);
|
|
757
|
|
a = _mm_shuffle_ps(accum_sse, accum_sse, _MM_SHUFFLE(1,0,3,2));
|
|
758
|
|
accum_sse = _mm_add_ps(a, accum_sse);
|
|
759
|
|
_mm_store_ss(&accum, accum_sse);
|
|
760
|
|
#elif EINSUM_USE_SSE2 && @float64@
|
|
761
|
|
/* Add the two SSE2 values and put in accum */
|
|
762
|
|
a = _mm_shuffle_pd(accum_sse, accum_sse, _MM_SHUFFLE2(0,1));
|
|
763
|
|
accum_sse = _mm_add_pd(a, accum_sse);
|
|
764
|
|
_mm_store_sd(&accum, accum_sse);
|
|
765
|
|
#endif
|
|
766
|
|
|
|
767
|
|
/* Finish off the loop */
|
|
768
|
1
|
goto finish_after_unrolled_loop;
|
|
769
|
1
|
}
|
|
770
|
|
|
|
771
|
|
static void
|
|
772
|
1
|
@name@_sum_of_products_stride0_contig_outstride0_two(int nop, char **dataptr,
|
|
773
|
|
npy_intp const *NPY_UNUSED(strides), npy_intp count)
|
|
774
|
|
{
|
|
775
|
1
|
@temptype@ value0 = @from@(*(@type@ *)dataptr[0]);
|
|
776
|
1
|
@type@ *data1 = (@type@ *)dataptr[1];
|
|
777
|
1
|
@temptype@ accum = 0;
|
|
778
|
|
|
|
779
|
|
#if EINSUM_USE_SSE1 && @float32@
|
|
780
|
|
__m128 a, accum_sse = _mm_setzero_ps();
|
|
781
|
1
|
#elif EINSUM_USE_SSE2 && @float64@
|
|
782
|
1
|
__m128d a, accum_sse = _mm_setzero_pd();
|
|
783
|
|
#endif
|
|
784
|
|
|
|
785
|
|
NPY_EINSUM_DBG_PRINT1("@name@_sum_of_products_stride0_contig_outstride0_two (%d)\n",
|
|
786
|
|
(int)count);
|
|
787
|
|
|
|
788
|
|
/* This is placed before the main loop to make small counts faster */
|
|
789
|
1
|
finish_after_unrolled_loop:
|
|
790
|
1
|
switch (count) {
|
|
791
|
1
|
/**begin repeat2
|
|
792
|
|
* #i = 6, 5, 4, 3, 2, 1, 0#
|
|
793
|
|
*/
|
|
794
|
1
|
case @i@+1:
|
|
795
|
1
|
accum += @from@(data1[@i@]);
|
|
796
|
1
|
/**end repeat2**/
|
|
797
|
1
|
case 0:
|
|
798
|
1
|
*(@type@ *)dataptr[2] = @to@(@from@(*(@type@ *)dataptr[2]) + value0 * accum);
|
|
799
|
1
|
return;
|
|
800
|
|
}
|
|
801
|
1
|
|
|
802
|
|
#if EINSUM_USE_SSE1 && @float32@
|
|
803
|
|
/* Use aligned instructions if possible */
|
|
804
|
1
|
if (EINSUM_IS_SSE_ALIGNED(data1)) {
|
|
805
|
|
/* Unroll the loop by 8 */
|
|
806
|
1
|
while (count >= 8) {
|
|
807
|
1
|
count -= 8;
|
|
808
|
|
|
|
809
|
1
|
/**begin repeat2
|
|
810
|
|
* #i = 0, 4#
|
|
811
|
|
*/
|
|
812
|
|
/*
|
|
813
|
1
|
* NOTE: This accumulation changes the order, so will likely
|
|
814
|
|
* produce slightly different results.
|
|
815
|
|
*/
|
|
816
|
1
|
accum_sse = _mm_add_ps(accum_sse, _mm_load_ps(data1+@i@));
|
|
817
|
1
|
/**end repeat2**/
|
|
818
|
1
|
data1 += 8;
|
|
819
|
|
}
|
|
820
|
|
/* Add the four SSE values and put in accum */
|
|
821
|
1
|
a = _mm_shuffle_ps(accum_sse, accum_sse, _MM_SHUFFLE(2,3,0,1));
|
|
822
|
1
|
accum_sse = _mm_add_ps(a, accum_sse);
|
|
823
|
1
|
a = _mm_shuffle_ps(accum_sse, accum_sse, _MM_SHUFFLE(1,0,3,2));
|
|
824
|
1
|
accum_sse = _mm_add_ps(a, accum_sse);
|
|
825
|
|
_mm_store_ss(&accum, accum_sse);
|
|
826
|
|
|
|
827
|
|
/* Finish off the loop */
|
|
828
|
|
goto finish_after_unrolled_loop;
|
|
829
|
|
}
|
|
830
|
|
#elif EINSUM_USE_SSE2 && @float64@
|
|
831
|
1
|
/* Use aligned instructions if possible */
|
|
832
|
|
if (EINSUM_IS_SSE_ALIGNED(data1)) {
|
|
833
|
|
/* Unroll the loop by 8 */
|
|
834
|
|
while (count >= 8) {
|
|
835
|
|
count -= 8;
|
|
836
|
|
|
|
837
|
1
|
/**begin repeat2
|
|
838
|
|
* #i = 0, 2, 4, 6#
|
|
839
|
1
|
*/
|
|
840
|
1
|
/*
|
|
841
|
|
* NOTE: This accumulation changes the order, so will likely
|
|
842
|
|
* produce slightly different results.
|
|
843
|
|
*/
|
|
844
|
1
|
accum_sse = _mm_add_pd(accum_sse, _mm_load_pd(data1+@i@));
|
|
845
|
|
/**end repeat2**/
|
|
846
|
1
|
data1 += 8;
|
|
847
|
|
}
|
|
848
|
|
/* Add the two SSE2 values and put in accum */
|
|
849
|
1
|
a = _mm_shuffle_pd(accum_sse, accum_sse, _MM_SHUFFLE2(0,1));
|
|
850
|
1
|
accum_sse = _mm_add_pd(a, accum_sse);
|
|
851
|
|
_mm_store_sd(&accum, accum_sse);
|
|
852
|
|
|
|
853
|
|
/* Finish off the loop */
|
|
854
|
|
goto finish_after_unrolled_loop;
|
|
855
|
|
}
|
|
856
|
|
#endif
|
|
857
|
|
|
|
858
|
|
/* Unroll the loop by 8 */
|
|
859
|
1
|
while (count >= 8) {
|
|
860
|
1
|
count -= 8;
|
|
861
|
|
|
|
862
|
|
#if EINSUM_USE_SSE1 && @float32@
|
|
863
|
|
/**begin repeat2
|
|
864
|
|
* #i = 0, 4#
|
|
865
|
|
*/
|
|
866
|
|
/*
|
|
867
|
|
* NOTE: This accumulation changes the order, so will likely
|
|
868
|
|
* produce slightly different results.
|
|
869
|
|
*/
|
|
870
|
1
|
accum_sse = _mm_add_ps(accum_sse, _mm_loadu_ps(data1+@i@));
|
|
871
|
|
/**end repeat2**/
|
|
872
|
|
#elif EINSUM_USE_SSE2 && @float64@
|
|
873
|
|
/**begin repeat2
|
|
874
|
|
* #i = 0, 2, 4, 6#
|
|
875
|
|
*/
|
|
876
|
|
/*
|
|
877
|
|
* NOTE: This accumulation changes the order, so will likely
|
|
878
|
1
|
* produce slightly different results.
|
|
879
|
1
|
*/
|
|
880
|
1
|
accum_sse = _mm_add_pd(accum_sse, _mm_loadu_pd(data1+@i@));
|
|
881
|
|
/**end repeat2**/
|
|
882
|
|
#else
|
|
883
|
1
|
/**begin repeat2
|
|
884
|
1
|
* #i = 0, 1, 2, 3, 4, 5, 6, 7#
|
|
885
|
|
*/
|
|
886
|
1
|
accum += @from@(data1[@i@]);
|
|
887
|
|
/**end repeat2**/
|
|
888
|
|
#endif
|
|
889
|
1
|
data1 += 8;
|
|
890
|
|
}
|
|
891
|
|
|
|
892
|
|
#if EINSUM_USE_SSE1 && @float32@
|
|
893
|
|
/* Add the four SSE values and put in accum */
|
|
894
|
|
a = _mm_shuffle_ps(accum_sse, accum_sse, _MM_SHUFFLE(2,3,0,1));
|
|
895
|
|
accum_sse = _mm_add_ps(a, accum_sse);
|
|
896
|
|
a = _mm_shuffle_ps(accum_sse, accum_sse, _MM_SHUFFLE(1,0,3,2));
|
|
897
|
|
accum_sse = _mm_add_ps(a, accum_sse);
|
|
898
|
|
_mm_store_ss(&accum, accum_sse);
|
|
899
|
|
#elif EINSUM_USE_SSE2 && @float64@
|
|
900
|
|
/* Add the two SSE2 values and put in accum */
|
|
901
|
|
a = _mm_shuffle_pd(accum_sse, accum_sse, _MM_SHUFFLE2(0,1));
|
|
902
|
|
accum_sse = _mm_add_pd(a, accum_sse);
|
|
903
|
|
_mm_store_sd(&accum, accum_sse);
|
|
904
|
|
#endif
|
|
905
|
|
|
|
906
|
|
/* Finish off the loop */
|
|
907
|
|
goto finish_after_unrolled_loop;
|
|
908
|
1
|
}
|
|
909
|
|
|
|
910
|
|
static void
|
|
911
|
1
|
@name@_sum_of_products_contig_stride0_outstride0_two(int nop, char **dataptr,
|
|
912
|
|
npy_intp const *NPY_UNUSED(strides), npy_intp count)
|
|
913
|
|
{
|
|
914
|
1
|
@type@ *data0 = (@type@ *)dataptr[0];
|
|
915
|
1
|
@temptype@ value1 = @from@(*(@type@ *)dataptr[1]);
|
|
916
|
1
|
@temptype@ accum = 0;
|
|
917
|
|
|
|
918
|
|
#if EINSUM_USE_SSE1 && @float32@
|
|
919
|
|
__m128 a, accum_sse = _mm_setzero_ps();
|
|
920
|
1
|
#elif EINSUM_USE_SSE2 && @float64@
|
|
921
|
1
|
__m128d a, accum_sse = _mm_setzero_pd();
|
|
922
|
|
#endif
|
|
923
|
|
|
|
924
|
|
NPY_EINSUM_DBG_PRINT1("@name@_sum_of_products_contig_stride0_outstride0_two (%d)\n",
|
|
925
|
|
(int)count);
|
|
926
|
|
|
|
927
|
1
|
/* This is placed before the main loop to make small counts faster */
|
|
928
|
1
|
finish_after_unrolled_loop:
|
|
929
|
1
|
switch (count) {
|
|
930
|
1
|
/**begin repeat2
|
|
931
|
|
* #i = 6, 5, 4, 3, 2, 1, 0#
|
|
932
|
1
|
*/
|
|
933
|
1
|
case @i@+1:
|
|
934
|
1
|
accum += @from@(data0[@i@]);
|
|
935
|
1
|
/**end repeat2**/
|
|
936
|
1
|
case 0:
|
|
937
|
1
|
*(@type@ *)dataptr[2] = @to@(@from@(*(@type@ *)dataptr[2]) + accum * value1);
|
|
938
|
1
|
return;
|
|
939
|
|
}
|
|
940
|
1
|
|
|
941
|
|
#if EINSUM_USE_SSE1 && @float32@
|
|
942
|
|
/* Use aligned instructions if possible */
|
|
943
|
1
|
if (EINSUM_IS_SSE_ALIGNED(data0)) {
|
|
944
|
|
/* Unroll the loop by 8 */
|
|
945
|
1
|
while (count >= 8) {
|
|
946
|
1
|
count -= 8;
|
|
947
|
|
|
|
948
|
1
|
/**begin repeat2
|
|
949
|
1
|
* #i = 0, 4#
|
|
950
|
|
*/
|
|
951
|
|
/*
|
|
952
|
1
|
* NOTE: This accumulation changes the order, so will likely
|
|
953
|
1
|
* produce slightly different results.
|
|
954
|
1
|
*/
|
|
955
|
1
|
accum_sse = _mm_add_ps(accum_sse, _mm_load_ps(data0+@i@));
|
|
956
|
|
/**end repeat2**/
|
|
957
|
1
|
data0 += 8;
|
|
958
|
|
}
|
|
959
|
|
/* Add the four SSE values and put in accum */
|
|
960
|
1
|
a = _mm_shuffle_ps(accum_sse, accum_sse, _MM_SHUFFLE(2,3,0,1));
|
|
961
|
1
|
accum_sse = _mm_add_ps(a, accum_sse);
|
|
962
|
1
|
a = _mm_shuffle_ps(accum_sse, accum_sse, _MM_SHUFFLE(1,0,3,2));
|
|
963
|
1
|
accum_sse = _mm_add_ps(a, accum_sse);
|
|
964
|
|
_mm_store_ss(&accum, accum_sse);
|
|
965
|
|
/* Finish off the loop */
|
|
966
|
|
goto finish_after_unrolled_loop;
|
|
967
|
1
|
}
|
|
968
|
|
#elif EINSUM_USE_SSE2 && @float64@
|
|
969
|
|
/* Use aligned instructions if possible */
|
|
970
|
|
if (EINSUM_IS_SSE_ALIGNED(data0)) {
|
|
971
|
|
/* Unroll the loop by 8 */
|
|
972
|
|
while (count >= 8) {
|
|
973
|
|
count -= 8;
|
|
974
|
|
|
|
975
|
1
|
/**begin repeat2
|
|
976
|
|
* #i = 0, 2, 4, 6#
|
|
977
|
1
|
*/
|
|
978
|
1
|
/*
|
|
979
|
|
* NOTE: This accumulation changes the order, so will likely
|
|
980
|
|
* produce slightly different results.
|
|
981
|
|
*/
|
|
982
|
1
|
accum_sse = _mm_add_pd(accum_sse, _mm_load_pd(data0+@i@));
|
|
983
|
|
/**end repeat2**/
|
|
984
|
1
|
data0 += 8;
|
|
985
|
|
}
|
|
986
|
|
/* Add the two SSE2 values and put in accum */
|
|
987
|
1
|
a = _mm_shuffle_pd(accum_sse, accum_sse, _MM_SHUFFLE2(0,1));
|
|
988
|
1
|
accum_sse = _mm_add_pd(a, accum_sse);
|
|
989
|
|
_mm_store_sd(&accum, accum_sse);
|
|
990
|
|
/* Finish off the loop */
|
|
991
|
|
goto finish_after_unrolled_loop;
|
|
992
|
|
}
|
|
993
|
|
#endif
|
|
994
|
|
|
|
995
|
|
/* Unroll the loop by 8 */
|
|
996
|
1
|
while (count >= 8) {
|
|
997
|
1
|
count -= 8;
|
|
998
|
|
|
|
999
|
|
#if EINSUM_USE_SSE1 && @float32@
|
|
1000
|
|
/**begin repeat2
|
|
1001
|
|
* #i = 0, 4#
|
|
1002
|
|
*/
|
|
1003
|
|
/*
|
|
1004
|
|
* NOTE: This accumulation changes the order, so will likely
|
|
1005
|
|
* produce slightly different results.
|
|
1006
|
|
*/
|
|
1007
|
1
|
accum_sse = _mm_add_ps(accum_sse, _mm_loadu_ps(data0+@i@));
|
|
1008
|
|
/**end repeat2**/
|
|
1009
|
|
#elif EINSUM_USE_SSE2 && @float64@
|
|
1010
|
|
/**begin repeat2
|
|
1011
|
|
* #i = 0, 2, 4, 6#
|
|
1012
|
|
*/
|
|
1013
|
|
/*
|
|
1014
|
|
* NOTE: This accumulation changes the order, so will likely
|
|
1015
|
1
|
* produce slightly different results.
|
|
1016
|
1
|
*/
|
|
1017
|
1
|
accum_sse = _mm_add_pd(accum_sse, _mm_loadu_pd(data0+@i@));
|
|
1018
|
|
/**end repeat2**/
|
|
1019
|
|
#else
|
|
1020
|
1
|
/**begin repeat2
|
|
1021
|
1
|
* #i = 0, 1, 2, 3, 4, 5, 6, 7#
|
|
1022
|
|
*/
|
|
1023
|
1
|
accum += @from@(data0[@i@]);
|
|
1024
|
|
/**end repeat2**/
|
|
1025
|
|
#endif
|
|
1026
|
1
|
data0 += 8;
|
|
1027
|
|
}
|
|
1028
|
|
|
|
1029
|
|
#if EINSUM_USE_SSE1 && @float32@
|
|
1030
|
|
/* Add the four SSE values and put in accum */
|
|
1031
|
|
a = _mm_shuffle_ps(accum_sse, accum_sse, _MM_SHUFFLE(2,3,0,1));
|
|
1032
|
|
accum_sse = _mm_add_ps(a, accum_sse);
|
|
1033
|
|
a = _mm_shuffle_ps(accum_sse, accum_sse, _MM_SHUFFLE(1,0,3,2));
|
|
1034
|
|
accum_sse = _mm_add_ps(a, accum_sse);
|
|
1035
|
|
_mm_store_ss(&accum, accum_sse);
|
|
1036
|
|
#elif EINSUM_USE_SSE2 && @float64@
|
|
1037
|
|
/* Add the two SSE2 values and put in accum */
|
|
1038
|
|
a = _mm_shuffle_pd(accum_sse, accum_sse, _MM_SHUFFLE2(0,1));
|
|
1039
|
|
accum_sse = _mm_add_pd(a, accum_sse);
|
|
1040
|
|
_mm_store_sd(&accum, accum_sse);
|
|
1041
|
|
#endif
|
|
1042
|
|
|
|
1043
|
|
/* Finish off the loop */
|
|
1044
|
|
goto finish_after_unrolled_loop;
|
|
1045
|
1
|
}
|
|
1046
|
|
|
|
1047
|
|
#elif @nop@ == 3 && !@complex@
|
|
1048
|
|
|
|
1049
|
|
static void
|
|
1050
|
|
@name@_sum_of_products_contig_three(int nop, char **dataptr,
|
|
1051
|
|
npy_intp const *NPY_UNUSED(strides), npy_intp count)
|
|
1052
|
|
{
|
|
1053
|
|
@type@ *data0 = (@type@ *)dataptr[0];
|
|
1054
|
|
@type@ *data1 = (@type@ *)dataptr[1];
|
|
1055
|
|
@type@ *data2 = (@type@ *)dataptr[2];
|
|
1056
|
|
@type@ *data_out = (@type@ *)dataptr[3];
|
|
1057
|
1
|
|
|
1058
|
1
|
/* Unroll the loop by 8 */
|
|
1059
|
|
while (count >= 8) {
|
|
1060
|
|
count -= 8;
|
|
1061
|
|
|
|
1062
|
|
/**begin repeat2
|
|
1063
|
|
* #i = 0, 1, 2, 3, 4, 5, 6, 7#
|
|
1064
|
1
|
*/
|
|
1065
|
1
|
data_out[@i@] = @to@(@from@(data0[@i@]) *
|
|
1066
|
1
|
@from@(data1[@i@]) *
|
|
1067
|
1
|
@from@(data2[@i@]) +
|
|
1068
|
1
|
@from@(data_out[@i@]));
|
|
1069
|
1
|
/**end repeat2**/
|
|
1070
|
1
|
data0 += 8;
|
|
1071
|
1
|
data1 += 8;
|
|
1072
|
1
|
data2 += 8;
|
|
1073
|
1
|
data_out += 8;
|
|
1074
|
|
}
|
|
1075
|
|
|
|
1076
|
|
/* Finish off the loop */
|
|
1077
|
|
|
|
1078
|
|
/**begin repeat2
|
|
1079
|
|
* #i = 0, 1, 2, 3, 4, 5, 6, 7#
|
|
1080
|
|
*/
|
|
1081
|
1
|
if (count-- == 0) {
|
|
1082
|
|
return;
|
|
1083
|
|
}
|
|
1084
|
1
|
data_out[@i@] = @to@(@from@(data0[@i@]) *
|
|
1085
|
1
|
@from@(data1[@i@]) *
|
|
1086
|
1
|
@from@(data2[@i@]) +
|
|
1087
|
1
|
@from@(data_out[@i@]));
|
|
1088
|
|
/**end repeat2**/
|
|
1089
|
|
}
|
|
1090
|
|
|
|
1091
|
|
#else /* @nop@ > 3 || @complex */
|
|
1092
|
|
|
|
1093
|
|
static void
|
|
1094
|
|
@name@_sum_of_products_contig_@noplabel@(int nop, char **dataptr,
|
|
1095
|
|
npy_intp const *NPY_UNUSED(strides), npy_intp count)
|
|
1096
|
|
{
|
|
1097
|
|
NPY_EINSUM_DBG_PRINT1("@name@_sum_of_products_contig_@noplabel@ (%d)\n",
|
|
1098
|
|
(int)count);
|
|
1099
|
|
|
|
1100
|
|
while (count--) {
|
|
1101
|
|
#if !@complex@
|
|
1102
|
|
@temptype@ temp = @from@(*(@type@ *)dataptr[0]);
|
|
1103
|
|
int i;
|
|
1104
|
|
for (i = 1; i < nop; ++i) {
|
|
1105
|
|
temp *= @from@(*(@type@ *)dataptr[i]);
|
|
1106
|
|
}
|
|
1107
|
|
*(@type@ *)dataptr[nop] = @to@(temp +
|
|
1108
|
|
@from@(*(@type@ *)dataptr[i]));
|
|
1109
|
|
for (i = 0; i <= nop; ++i) {
|
|
1110
|
|
dataptr[i] += sizeof(@type@);
|
|
1111
|
|
}
|
|
1112
|
|
#else /* complex */
|
|
1113
|
|
# if @nop@ <= 3
|
|
1114
|
|
# define _SUMPROD_NOP @nop@
|
|
1115
|
|
# else
|
|
1116
|
|
# define _SUMPROD_NOP nop
|
|
1117
|
|
# endif
|
|
1118
|
|
@temptype@ re, im, tmp;
|
|
1119
|
|
int i;
|
|
1120
|
|
re = ((@temptype@ *)dataptr[0])[0];
|
|
1121
|
|
im = ((@temptype@ *)dataptr[0])[1];
|
|
1122
|
|
for (i = 1; i < _SUMPROD_NOP; ++i) {
|
|
1123
|
|
tmp = re * ((@temptype@ *)dataptr[i])[0] -
|
|
1124
|
|
im * ((@temptype@ *)dataptr[i])[1];
|
|
1125
|
|
im = re * ((@temptype@ *)dataptr[i])[1] +
|
|
1126
|
|
im * ((@temptype@ *)dataptr[i])[0];
|
|
1127
|
|
re = tmp;
|
|
1128
|
|
}
|
|
1129
|
|
((@temptype@ *)dataptr[_SUMPROD_NOP])[0] = re +
|
|
1130
|
|
((@temptype@ *)dataptr[_SUMPROD_NOP])[0];
|
|
1131
|
|
((@temptype@ *)dataptr[_SUMPROD_NOP])[1] = im +
|
|
1132
|
|
((@temptype@ *)dataptr[_SUMPROD_NOP])[1];
|
|
1133
|
|
|
|
1134
|
|
for (i = 0; i <= _SUMPROD_NOP; ++i) {
|
|
1135
|
|
dataptr[i] += sizeof(@type@);
|
|
1136
|
|
}
|
|
1137
|
|
# undef _SUMPROD_NOP
|
|
1138
|
|
#endif
|
|
1139
|
|
}
|
|
1140
|
|
}
|
|
1141
|
|
|
|
1142
|
|
#endif /* functions for various @nop@ */
|
|
1143
|
|
|
|
1144
|
|
#if @nop@ == 1
|
|
1145
|
|
|
|
1146
|
|
static void
|
|
1147
|
|
@name@_sum_of_products_contig_outstride0_one(int nop, char **dataptr,
|
|
1148
|
|
npy_intp const *strides, npy_intp count)
|
|
1149
|
|
{
|
|
1150
|
|
#if @complex@
|
|
1151
|
|
@temptype@ accum_re = 0, accum_im = 0;
|
|
1152
|
|
@temptype@ *data0 = (@temptype@ *)dataptr[0];
|
|
1153
|
|
#else
|
|
1154
|
|
@temptype@ accum = 0;
|
|
1155
|
|
@type@ *data0 = (@type@ *)dataptr[0];
|
|
1156
|
|
#endif
|
|
1157
|
|
|
|
1158
|
|
#if EINSUM_USE_SSE1 && @float32@
|
|
1159
|
|
__m128 a, accum_sse = _mm_setzero_ps();
|
|
1160
|
|
#elif EINSUM_USE_SSE2 && @float64@
|
|
1161
|
|
__m128d a, accum_sse = _mm_setzero_pd();
|
|
1162
|
|
#endif
|
|
1163
|
|
|
|
1164
|
|
|
|
1165
|
|
NPY_EINSUM_DBG_PRINT1("@name@_sum_of_products_contig_outstride0_one (%d)\n",
|
|
1166
|
|
(int)count);
|
|
1167
|
|
|
|
1168
|
|
/* This is placed before the main loop to make small counts faster */
|
|
1169
|
|
finish_after_unrolled_loop:
|
|
1170
|
|
switch (count) {
|
|
1171
|
|
/**begin repeat2
|
|
1172
|
|
* #i = 6, 5, 4, 3, 2, 1, 0#
|
|
1173
|
|
*/
|
|
1174
|
1
|
case @i@+1:
|
|
1175
|
|
#if !@complex@
|
|
1176
|
1
|
accum += @from@(data0[@i@]);
|
|
1177
|
|
#else /* complex */
|
|
1178
|
1
|
accum_re += data0[2*@i@+0];
|
|
1179
|
1
|
accum_im += data0[2*@i@+1];
|
|
1180
|
|
#endif
|
|
1181
|
|
/**end repeat2**/
|
|
1182
|
1
|
case 0:
|
|
1183
|
|
#if @complex@
|
|
1184
|
1
|
((@temptype@ *)dataptr[1])[0] += accum_re;
|
|
1185
|
1
|
((@temptype@ *)dataptr[1])[1] += accum_im;
|
|
1186
|
|
#else
|
|
1187
|
1
|
*((@type@ *)dataptr[1]) = @to@(accum +
|
|
1188
|
1
|
@from@(*((@type@ *)dataptr[1])));
|
|
1189
|
|
#endif
|
|
1190
|
1
|
return;
|
|
1191
|
|
}
|
|
1192
|
|
|
|
1193
|
|
#if EINSUM_USE_SSE1 && @float32@
|
|
1194
|
|
/* Use aligned instructions if possible */
|
|
1195
|
1
|
if (EINSUM_IS_SSE_ALIGNED(data0)) {
|
|
1196
|
|
/* Unroll the loop by 8 */
|
|
1197
|
1
|
while (count >= 8) {
|
|
1198
|
1
|
count -= 8;
|
|
1199
|
|
|
|
1200
|
1
|
_mm_prefetch(data0 + 512, _MM_HINT_T0);
|
|
1201
|
|
|
|
1202
|
|
/**begin repeat2
|
|
1203
|
|
* #i = 0, 4#
|
|
1204
|
|
*/
|
|
1205
|
|
/*
|
|
1206
|
|
* NOTE: This accumulation changes the order, so will likely
|
|
1207
|
|
* produce slightly different results.
|
|
1208
|
|
*/
|
|
1209
|
1
|
accum_sse = _mm_add_ps(accum_sse, _mm_load_ps(data0+@i@));
|
|
1210
|
|
/**end repeat2**/
|
|
1211
|
1
|
data0 += 8;
|
|
1212
|
|
}
|
|
1213
|
|
|
|
1214
|
|
/* Add the four SSE values and put in accum */
|
|
1215
|
1
|
a = _mm_shuffle_ps(accum_sse, accum_sse, _MM_SHUFFLE(2,3,0,1));
|
|
1216
|
1
|
accum_sse = _mm_add_ps(a, accum_sse);
|
|
1217
|
1
|
a = _mm_shuffle_ps(accum_sse, accum_sse, _MM_SHUFFLE(1,0,3,2));
|
|
1218
|
1
|
accum_sse = _mm_add_ps(a, accum_sse);
|
|
1219
|
|
_mm_store_ss(&accum, accum_sse);
|
|
1220
|
|
|
|
1221
|
|
/* Finish off the loop */
|
|
1222
|
|
goto finish_after_unrolled_loop;
|
|
1223
|
|
}
|
|
1224
|
|
#elif EINSUM_USE_SSE2 && @float64@
|
|
1225
|
|
/* Use aligned instructions if possible */
|
|
1226
|
|
if (EINSUM_IS_SSE_ALIGNED(data0)) {
|
|
1227
|
|
/* Unroll the loop by 8 */
|
|
1228
|
|
while (count >= 8) {
|
|
1229
|
|
count -= 8;
|
|
1230
|
|
|
|
1231
|
1
|
_mm_prefetch(data0 + 512, _MM_HINT_T0);
|
|
1232
|
|
|
|
1233
|
1
|
/**begin repeat2
|
|
1234
|
1
|
* #i = 0, 2, 4, 6#
|
|
1235
|
|
*/
|
|
1236
|
1
|
/*
|
|
1237
|
|
* NOTE: This accumulation changes the order, so will likely
|
|
1238
|
|
* produce slightly different results.
|
|
1239
|
|
*/
|
|
1240
|
1
|
accum_sse = _mm_add_pd(accum_sse, _mm_load_pd(data0+@i@));
|
|
1241
|
|
/**end repeat2**/
|
|
1242
|
1
|
data0 += 8;
|
|
1243
|
|
}
|
|
1244
|
|
|
|
1245
|
|
/* Add the two SSE2 values and put in accum */
|
|
1246
|
1
|
a = _mm_shuffle_pd(accum_sse, accum_sse, _MM_SHUFFLE2(0,1));
|
|
1247
|
1
|
accum_sse = _mm_add_pd(a, accum_sse);
|
|
1248
|
|
_mm_store_sd(&accum, accum_sse);
|
|
1249
|
|
|
|
1250
|
|
/* Finish off the loop */
|
|
1251
|
|
goto finish_after_unrolled_loop;
|
|
1252
|
|
}
|
|
1253
|
|
#endif
|
|
1254
|
|
|
|
1255
|
|
/* Unroll the loop by 8 */
|
|
1256
|
1
|
while (count >= 8) {
|
|
1257
|
1
|
count -= 8;
|
|
1258
|
|
|
|
1259
|
|
#if EINSUM_USE_SSE1 && @float32@
|
|
1260
|
|
_mm_prefetch(data0 + 512, _MM_HINT_T0);
|
|
1261
|
|
|
|
1262
|
|
/**begin repeat2
|
|
1263
|
|
* #i = 0, 4#
|
|
1264
|
|
*/
|
|
1265
|
|
/*
|
|
1266
|
|
* NOTE: This accumulation changes the order, so will likely
|
|
1267
|
|
* produce slightly different results.
|
|
1268
|
|
*/
|
|
1269
|
1
|
accum_sse = _mm_add_ps(accum_sse, _mm_loadu_ps(data0+@i@));
|
|
1270
|
|
/**end repeat2**/
|
|
1271
|
|
#elif EINSUM_USE_SSE2 && @float64@
|
|
1272
|
|
_mm_prefetch(data0 + 512, _MM_HINT_T0);
|
|
1273
|
|
|
|
1274
|
|
/**begin repeat2
|
|
1275
|
1
|
* #i = 0, 2, 4, 6#
|
|
1276
|
1
|
*/
|
|
1277
|
1
|
/*
|
|
1278
|
|
* NOTE: This accumulation changes the order, so will likely
|
|
1279
|
1
|
* produce slightly different results.
|
|
1280
|
1
|
*/
|
|
1281
|
1
|
accum_sse = _mm_add_pd(accum_sse, _mm_loadu_pd(data0+@i@));
|
|
1282
|
|
/**end repeat2**/
|
|
1283
|
|
#else
|
|
1284
|
|
/**begin repeat2
|
|
1285
|
|
* #i = 0, 1, 2, 3, 4, 5, 6, 7#
|
|
1286
|
|
*/
|
|
1287
|
|
# if !@complex@
|
|
1288
|
1
|
accum += @from@(data0[@i@]);
|
|
1289
|
|
# else /* complex */
|
|
1290
|
1
|
accum_re += data0[2*@i@+0];
|
|
1291
|
1
|
accum_im += data0[2*@i@+1];
|
|
1292
|
|
# endif
|
|
1293
|
|
/**end repeat2**/
|
|
1294
|
|
#endif
|
|
1295
|
|
|
|
1296
|
|
#if !@complex@
|
|
1297
|
1
|
data0 += 8;
|
|
1298
|
|
#else
|
|
1299
|
1
|
data0 += 8*2;
|
|
1300
|
|
#endif
|
|
1301
|
|
}
|
|
1302
|
|
|
|
1303
|
|
#if EINSUM_USE_SSE1 && @float32@
|
|
1304
|
|
/* Add the four SSE values and put in accum */
|
|
1305
|
|
a = _mm_shuffle_ps(accum_sse, accum_sse, _MM_SHUFFLE(2,3,0,1));
|
|
1306
|
|
accum_sse = _mm_add_ps(a, accum_sse);
|
|
1307
|
|
a = _mm_shuffle_ps(accum_sse, accum_sse, _MM_SHUFFLE(1,0,3,2));
|
|
1308
|
|
accum_sse = _mm_add_ps(a, accum_sse);
|
|
1309
|
|
_mm_store_ss(&accum, accum_sse);
|
|
1310
|
|
#elif EINSUM_USE_SSE2 && @float64@
|
|
1311
|
|
/* Add the two SSE2 values and put in accum */
|
|
1312
|
|
a = _mm_shuffle_pd(accum_sse, accum_sse, _MM_SHUFFLE2(0,1));
|
|
1313
|
|
accum_sse = _mm_add_pd(a, accum_sse);
|
|
1314
|
|
_mm_store_sd(&accum, accum_sse);
|
|
1315
|
|
#endif
|
|
1316
|
|
|
|
1317
|
|
/* Finish off the loop */
|
|
1318
|
|
goto finish_after_unrolled_loop;
|
|
1319
|
|
}
|
|
1320
|
|
|
|
1321
|
|
#endif /* @nop@ == 1 */
|
|
1322
|
|
|
|
1323
|
|
static void
|
|
1324
|
1
|
@name@_sum_of_products_outstride0_@noplabel@(int nop, char **dataptr,
|
|
1325
|
|
npy_intp const *strides, npy_intp count)
|
|
1326
|
|
{
|
|
1327
|
|
#if @complex@
|
|
1328
|
1
|
@temptype@ accum_re = 0, accum_im = 0;
|
|
1329
|
|
#else
|
|
1330
|
1
|
@temptype@ accum = 0;
|
|
1331
|
|
#endif
|
|
1332
|
|
|
|
1333
|
|
#if (@nop@ == 1) || (@nop@ <= 3 && !@complex@)
|
|
1334
|
1
|
char *data0 = dataptr[0];
|
|
1335
|
1
|
npy_intp stride0 = strides[0];
|
|
1336
|
|
#endif
|
|
1337
|
|
#if (@nop@ == 2 || @nop@ == 3) && !@complex@
|
|
1338
|
|
char *data1 = dataptr[1];
|
|
1339
|
|
npy_intp stride1 = strides[1];
|
|
1340
|
|
#endif
|
|
1341
|
|
#if (@nop@ == 3) && !@complex@
|
|
1342
|
|
char *data2 = dataptr[2];
|
|
1343
|
|
npy_intp stride2 = strides[2];
|
|
1344
|
|
#endif
|
|
1345
|
|
|
|
1346
|
|
NPY_EINSUM_DBG_PRINT1("@name@_sum_of_products_outstride0_@noplabel@ (%d)\n",
|
|
1347
|
|
(int)count);
|
|
1348
|
|
|
|
1349
|
1
|
while (count--) {
|
|
1350
|
|
#if !@complex@
|
|
1351
|
1
|
# if @nop@ == 1
|
|
1352
|
1
|
accum += @from@(*(@type@ *)data0);
|
|
1353
|
1
|
data0 += stride0;
|
|
1354
|
|
# elif @nop@ == 2
|
|
1355
|
|
accum += @from@(*(@type@ *)data0) *
|
|
1356
|
|
@from@(*(@type@ *)data1);
|
|
1357
|
|
data0 += stride0;
|
|
1358
|
|
data1 += stride1;
|
|
1359
|
|
# elif @nop@ == 3
|
|
1360
|
|
accum += @from@(*(@type@ *)data0) *
|
|
1361
|
|
@from@(*(@type@ *)data1) *
|
|
1362
|
|
@from@(*(@type@ *)data2);
|
|
1363
|
|
data0 += stride0;
|
|
1364
|
|
data1 += stride1;
|
|
1365
|
|
data2 += stride2;
|
|
1366
|
1
|
# else
|
|
1367
|
1
|
@temptype@ temp = @from@(*(@type@ *)dataptr[0]);
|
|
1368
|
|
int i;
|
|
1369
|
|
for (i = 1; i < nop; ++i) {
|
|
1370
|
1
|
temp *= @from@(*(@type@ *)dataptr[i]);
|
|
1371
|
|
}
|
|
1372
|
|
accum += temp;
|
|
1373
|
|
for (i = 0; i < nop; ++i) {
|
|
1374
|
|
dataptr[i] += strides[i];
|
|
1375
|
|
}
|
|
1376
|
|
# endif
|
|
1377
|
|
#else /* complex */
|
|
1378
|
1
|
# if @nop@ == 1
|
|
1379
|
1
|
accum_re += ((@temptype@ *)data0)[0];
|
|
1380
|
1
|
accum_im += ((@temptype@ *)data0)[1];
|
|
1381
|
1
|
data0 += stride0;
|
|
1382
|
|
# else
|
|
1383
|
|
# if @nop@ <= 3
|
|
1384
|
1
|
#define _SUMPROD_NOP @nop@
|
|
1385
|
|
# else
|
|
1386
|
|
#define _SUMPROD_NOP nop
|
|
1387
|
|
# endif
|
|
1388
|
1
|
@temptype@ re, im, tmp;
|
|
1389
|
1
|
int i;
|
|
1390
|
|
re = ((@temptype@ *)dataptr[0])[0];
|
|
1391
|
|
im = ((@temptype@ *)dataptr[0])[1];
|
|
1392
|
|
for (i = 1; i < _SUMPROD_NOP; ++i) {
|
|
1393
|
|
tmp = re * ((@temptype@ *)dataptr[i])[0] -
|
|
1394
|
|
im * ((@temptype@ *)dataptr[i])[1];
|
|
1395
|
|
im = re * ((@temptype@ *)dataptr[i])[1] +
|
|
1396
|
|
im * ((@temptype@ *)dataptr[i])[0];
|
|
1397
|
1
|
re = tmp;
|
|
1398
|
|
}
|
|
1399
|
|
accum_re += re;
|
|
1400
|
|
accum_im += im;
|
|
1401
|
|
for (i = 0; i < _SUMPROD_NOP; ++i) {
|
|
1402
|
|
dataptr[i] += strides[i];
|
|
1403
|
1
|
}
|
|
1404
|
|
#undef _SUMPROD_NOP
|
|
1405
|
|
# endif
|
|
1406
|
1
|
#endif
|
|
1407
|
1
|
}
|
|
1408
|
1
|
|
|
1409
|
|
#if @complex@
|
|
1410
|
|
# if @nop@ <= 3
|
|
1411
|
1
|
((@temptype@ *)dataptr[@nop@])[0] += accum_re;
|
|
1412
|
1
|
((@temptype@ *)dataptr[@nop@])[1] += accum_im;
|
|
1413
|
|
# else
|
|
1414
|
|
((@temptype@ *)dataptr[nop])[0] += accum_re;
|
|
1415
|
|
((@temptype@ *)dataptr[nop])[1] += accum_im;
|
|
1416
|
|
# endif
|
|
1417
|
|
#else
|
|
1418
|
|
# if @nop@ <= 3
|
|
1419
|
1
|
*((@type@ *)dataptr[@nop@]) = @to@(accum +
|
|
1420
|
1
|
@from@(*((@type@ *)dataptr[@nop@])));
|
|
1421
|
|
# else
|
|
1422
|
1
|
*((@type@ *)dataptr[nop]) = @to@(accum +
|
|
1423
|
|
@from@(*((@type@ *)dataptr[nop])));
|
|
1424
|
|
# endif
|
|
1425
|
1
|
#endif
|
|
1426
|
1
|
|
|
1427
|
1
|
}
|
|
1428
|
|
|
|
1429
|
|
/**end repeat1**/
|
|
1430
|
|
|
|
1431
|
|
/**end repeat**/
|
|
1432
|
|
|
|
1433
|
|
|
|
1434
|
|
/* Do OR of ANDs for the boolean type */
|
|
1435
|
|
|
|
1436
|
|
/**begin repeat
|
|
1437
|
|
* #nop = 1, 2, 3, 1000#
|
|
1438
|
|
* #noplabel = one, two, three, any#
|
|
1439
|
|
*/
|
|
1440
|
|
|
|
1441
|
|
static void
|
|
1442
|
0
|
bool_sum_of_products_@noplabel@(int nop, char **dataptr,
|
|
1443
|
|
npy_intp const *strides, npy_intp count)
|
|
1444
|
|
{
|
|
1445
|
|
#if (@nop@ <= 3)
|
|
1446
|
0
|
char *data0 = dataptr[0];
|
|
1447
|
0
|
npy_intp stride0 = strides[0];
|
|
1448
|
|
#endif
|
|
1449
|
|
#if (@nop@ == 2 || @nop@ == 3)
|
|
1450
|
0
|
char *data1 = dataptr[1];
|
|
1451
|
0
|
npy_intp stride1 = strides[1];
|
|
1452
|
|
#endif
|
|
1453
|
|
#if (@nop@ == 3)
|
|
1454
|
0
|
char *data2 = dataptr[2];
|
|
1455
|
0
|
npy_intp stride2 = strides[2];
|
|
1456
|
|
#endif
|
|
1457
|
|
#if (@nop@ <= 3)
|
|
1458
|
0
|
char *data_out = dataptr[@nop@];
|
|
1459
|
0
|
npy_intp stride_out = strides[@nop@];
|
|
1460
|
|
#endif
|
|
1461
|
|
|
|
1462
|
0
|
while (count--) {
|
|
1463
|
|
#if @nop@ == 1
|
|
1464
|
0
|
*(npy_bool *)data_out = *(npy_bool *)data0 ||
|
|
1465
|
0
|
*(npy_bool *)data_out;
|
|
1466
|
0
|
data0 += stride0;
|
|
1467
|
0
|
data_out += stride_out;
|
|
1468
|
|
#elif @nop@ == 2
|
|
1469
|
0
|
*(npy_bool *)data_out = (*(npy_bool *)data0 &&
|
|
1470
|
0
|
*(npy_bool *)data1) ||
|
|
1471
|
0
|
*(npy_bool *)data_out;
|
|
1472
|
0
|
data0 += stride0;
|
|
1473
|
1
|
data1 += stride1;
|
|
1474
|
1
|
data_out += stride_out;
|
|
1475
|
|
#elif @nop@ == 3
|
|
1476
|
0
|
*(npy_bool *)data_out = (*(npy_bool *)data0 &&
|
|
1477
|
0
|
*(npy_bool *)data1 &&
|
|
1478
|
1
|
*(npy_bool *)data2) ||
|
|
1479
|
0
|
*(npy_bool *)data_out;
|
|
1480
|
0
|
data0 += stride0;
|
|
1481
|
1
|
data1 += stride1;
|
|
1482
|
0
|
data2 += stride2;
|
|
1483
|
0
|
data_out += stride_out;
|
|
1484
|
1
|
#else
|
|
1485
|
0
|
npy_bool temp = *(npy_bool *)dataptr[0];
|
|
1486
|
|
int i;
|
|
1487
|
0
|
for (i = 1; i < nop; ++i) {
|
|
1488
|
1
|
temp = temp && *(npy_bool *)dataptr[i];
|
|
1489
|
1
|
}
|
|
1490
|
0
|
*(npy_bool *)dataptr[nop] = temp || *(npy_bool *)dataptr[i];
|
|
1491
|
0
|
for (i = 0; i <= nop; ++i) {
|
|
1492
|
1
|
dataptr[i] += strides[i];
|
|
1493
|
1
|
}
|
|
1494
|
|
#endif
|
|
1495
|
|
}
|
|
1496
|
1
|
}
|
|
1497
|
1
|
|
|
1498
|
|
static void
|
|
1499
|
1
|
bool_sum_of_products_contig_@noplabel@(int nop, char **dataptr,
|
|
1500
|
1
|
npy_intp const *strides, npy_intp count)
|
|
1501
|
|
{
|
|
1502
|
|
#if (@nop@ <= 3)
|
|
1503
|
1
|
char *data0 = dataptr[0];
|
|
1504
|
|
#endif
|
|
1505
|
|
#if (@nop@ == 2 || @nop@ == 3)
|
|
1506
|
1
|
char *data1 = dataptr[1];
|
|
1507
|
|
#endif
|
|
1508
|
|
#if (@nop@ == 3)
|
|
1509
|
1
|
char *data2 = dataptr[2];
|
|
1510
|
|
#endif
|
|
1511
|
|
#if (@nop@ <= 3)
|
|
1512
|
1
|
char *data_out = dataptr[@nop@];
|
|
1513
|
|
#endif
|
|
1514
|
1
|
|
|
1515
|
1
|
#if (@nop@ <= 3)
|
|
1516
|
1
|
/* This is placed before the main loop to make small counts faster */
|
|
1517
|
1
|
finish_after_unrolled_loop:
|
|
1518
|
1
|
switch (count) {
|
|
1519
|
1
|
/**begin repeat1
|
|
1520
|
|
* #i = 6, 5, 4, 3, 2, 1, 0#
|
|
1521
|
|
*/
|
|
1522
|
1
|
case @i@+1:
|
|
1523
|
|
# if @nop@ == 1
|
|
1524
|
0
|
((npy_bool *)data_out)[@i@] = ((npy_bool *)data0)[@i@] ||
|
|
1525
|
0
|
((npy_bool *)data_out)[@i@];
|
|
1526
|
|
# elif @nop@ == 2
|
|
1527
|
1
|
((npy_bool *)data_out)[@i@] =
|
|
1528
|
1
|
(((npy_bool *)data0)[@i@] &&
|
|
1529
|
1
|
((npy_bool *)data1)[@i@]) ||
|
|
1530
|
0
|
((npy_bool *)data_out)[@i@];
|
|
1531
|
|
# elif @nop@ == 3
|
|
1532
|
0
|
((npy_bool *)data_out)[@i@] =
|
|
1533
|
0
|
(((npy_bool *)data0)[@i@] &&
|
|
1534
|
0
|
((npy_bool *)data1)[@i@] &&
|
|
1535
|
0
|
((npy_bool *)data2)[@i@]) ||
|
|
1536
|
0
|
((npy_bool *)data_out)[@i@];
|
|
1537
|
|
# endif
|
|
1538
|
|
/**end repeat1**/
|
|
1539
|
|
case 0:
|
|
1540
|
1
|
return;
|
|
1541
|
|
}
|
|
1542
|
|
#endif
|
|
1543
|
|
|
|
1544
|
|
/* Unroll the loop by 8 for fixed-size nop */
|
|
1545
|
|
#if (@nop@ <= 3)
|
|
1546
|
1
|
while (count >= 8) {
|
|
1547
|
1
|
count -= 8;
|
|
1548
|
|
#else
|
|
1549
|
|
while (count--) {
|
|
1550
|
|
#endif
|
|
1551
|
|
|
|
1552
|
|
# if @nop@ == 1
|
|
1553
|
|
/**begin repeat1
|
|
1554
|
|
* #i = 0, 1, 2, 3, 4, 5, 6, 7#
|
|
1555
|
|
*/
|
|
1556
|
0
|
*((npy_bool *)data_out + @i@) = (*((npy_bool *)data0 + @i@)) ||
|
|
1557
|
0
|
(*((npy_bool *)data_out + @i@));
|
|
1558
|
|
/**end repeat1**/
|
|
1559
|
0
|
data0 += 8*sizeof(npy_bool);
|
|
1560
|
0
|
data_out += 8*sizeof(npy_bool);
|
|
1561
|
|
# elif @nop@ == 2
|
|
1562
|
|
/**begin repeat1
|
|
1563
|
|
* #i = 0, 1, 2, 3, 4, 5, 6, 7#
|
|
1564
|
|
*/
|
|
1565
|
0
|
*((npy_bool *)data_out + @i@) =
|
|
1566
|
0
|
((*((npy_bool *)data0 + @i@)) &&
|
|
1567
|
0
|
(*((npy_bool *)data1 + @i@))) ||
|
|
1568
|
0
|
(*((npy_bool *)data_out + @i@));
|
|
1569
|
|
/**end repeat1**/
|
|
1570
|
0
|
data0 += 8*sizeof(npy_bool);
|
|
1571
|
0
|
data1 += 8*sizeof(npy_bool);
|
|
1572
|
0
|
data_out += 8*sizeof(npy_bool);
|
|
1573
|
1
|
# elif @nop@ == 3
|
|
1574
|
1
|
/**begin repeat1
|
|
1575
|
|
* #i = 0, 1, 2, 3, 4, 5, 6, 7#
|
|
1576
|
|
*/
|
|
1577
|
1
|
*((npy_bool *)data_out + @i@) =
|
|
1578
|
1
|
((*((npy_bool *)data0 + @i@)) &&
|
|
1579
|
1
|
(*((npy_bool *)data1 + @i@)) &&
|
|
1580
|
1
|
(*((npy_bool *)data2 + @i@))) ||
|
|
1581
|
1
|
(*((npy_bool *)data_out + @i@));
|
|
1582
|
|
/**end repeat1**/
|
|
1583
|
1
|
data0 += 8*sizeof(npy_bool);
|
|
1584
|
1
|
data1 += 8*sizeof(npy_bool);
|
|
1585
|
1
|
data2 += 8*sizeof(npy_bool);
|
|
1586
|
1
|
data_out += 8*sizeof(npy_bool);
|
|
1587
|
|
# else
|
|
1588
|
|
npy_bool temp = *(npy_bool *)dataptr[0];
|
|
1589
|
1
|
int i;
|
|
1590
|
1
|
for (i = 1; i < nop; ++i) {
|
|
1591
|
|
temp = temp && *(npy_bool *)dataptr[i];
|
|
1592
|
|
}
|
|
1593
|
1
|
*(npy_bool *)dataptr[nop] = temp || *(npy_bool *)dataptr[i];
|
|
1594
|
1
|
for (i = 0; i <= nop; ++i) {
|
|
1595
|
|
dataptr[i] += sizeof(npy_bool);
|
|
1596
|
|
}
|
|
1597
|
|
# endif
|
|
1598
|
1
|
}
|
|
1599
|
|
|
|
1600
|
|
/* If the loop was unrolled, we need to finish it off */
|
|
1601
|
|
#if (@nop@ <= 3)
|
|
1602
|
|
goto finish_after_unrolled_loop;
|
|
1603
|
|
#endif
|
|
1604
|
1
|
}
|
|
1605
|
|
|
|
1606
|
|
static void
|
|
1607
|
0
|
bool_sum_of_products_outstride0_@noplabel@(int nop, char **dataptr,
|
|
1608
|
1
|
npy_intp const *strides, npy_intp count)
|
|
1609
|
1
|
{
|
|
1610
|
1
|
npy_bool accum = 0;
|
|
1611
|
1
|
|
|
1612
|
1
|
#if (@nop@ <= 3)
|
|
1613
|
1
|
char *data0 = dataptr[0];
|
|
1614
|
0
|
npy_intp stride0 = strides[0];
|
|
1615
|
|
#endif
|
|
1616
|
|
#if (@nop@ == 2 || @nop@ == 3)
|
|
1617
|
1
|
char *data1 = dataptr[1];
|
|
1618
|
0
|
npy_intp stride1 = strides[1];
|
|
1619
|
|
#endif
|
|
1620
|
|
#if (@nop@ == 3)
|
|
1621
|
0
|
char *data2 = dataptr[2];
|
|
1622
|
0
|
npy_intp stride2 = strides[2];
|
|
1623
|
1
|
#endif
|
|
1624
|
|
|
|
1625
|
0
|
while (count--) {
|
|
1626
|
|
#if @nop@ == 1
|
|
1627
|
1
|
accum = *(npy_bool *)data0 || accum;
|
|
1628
|
1
|
data0 += stride0;
|
|
1629
|
1
|
#elif @nop@ == 2
|
|
1630
|
1
|
accum = (*(npy_bool *)data0 && *(npy_bool *)data1) || accum;
|
|
1631
|
1
|
data0 += stride0;
|
|
1632
|
1
|
data1 += stride1;
|
|
1633
|
|
#elif @nop@ == 3
|
|
1634
|
0
|
accum = (*(npy_bool *)data0 &&
|
|
1635
|
0
|
*(npy_bool *)data1 &&
|
|
1636
|
0
|
*(npy_bool *)data2) || accum;
|
|
1637
|
0
|
data0 += stride0;
|
|
1638
|
0
|
data1 += stride1;
|
|
1639
|
0
|
data2 += stride2;
|
|
1640
|
|
#else
|
|
1641
|
|
npy_bool temp = *(npy_bool *)dataptr[0];
|
|
1642
|
1
|
int i;
|
|
1643
|
|
for (i = 1; i < nop; ++i) {
|
|
1644
|
|
temp = temp && *(npy_bool *)dataptr[i];
|
|
1645
|
|
}
|
|
1646
|
|
accum = temp || accum;
|
|
1647
|
|
for (i = 0; i <= nop; ++i) {
|
|
1648
|
1
|
dataptr[i] += strides[i];
|
|
1649
|
1
|
}
|
|
1650
|
1
|
#endif
|
|
1651
|
1
|
}
|
|
1652
|
|
|
|
1653
|
|
# if @nop@ <= 3
|
|
1654
|
0
|
*((npy_bool *)dataptr[@nop@]) = accum || *((npy_bool *)dataptr[@nop@]);
|
|
1655
|
0
|
# else
|
|
1656
|
|
*((npy_bool *)dataptr[nop]) = accum || *((npy_bool *)dataptr[nop]);
|
|
1657
|
0
|
# endif
|
|
1658
|
0
|
}
|
|
1659
|
|
|
|
1660
|
0
|
/**end repeat**/
|
|
1661
|
0
|
|
|
1662
|
|
/* These tables need to match up with the type enum */
|
|
1663
|
|
static sum_of_products_fn
|
|
1664
|
0
|
_contig_outstride0_unary_specialization_table[NPY_NTYPES] = {
|
|
1665
|
0
|
/**begin repeat
|
|
1666
|
|
* #name = bool,
|
|
1667
|
|
* byte, ubyte,
|
|
1668
|
|
* short, ushort,
|
|
1669
|
|
* int, uint,
|
|
1670
|
|
* long, ulong,
|
|
1671
|
|
* longlong, ulonglong,
|
|
1672
|
0
|
* float, double, longdouble,
|
|
1673
|
|
* cfloat, cdouble, clongdouble,
|
|
1674
|
1
|
* object, string, unicode, void,
|
|
1675
|
1
|
* datetime, timedelta, half#
|
|
1676
|
|
* #use = 0,
|
|
1677
|
0
|
* 1, 1,
|
|
1678
|
0
|
* 1, 1,
|
|
1679
|
0
|
* 1, 1,
|
|
1680
|
|
* 1, 1,
|
|
1681
|
|
* 1, 1,
|
|
1682
|
1
|
* 1, 1, 1,
|
|
1683
|
|
* 1, 1, 1,
|
|
1684
|
|
* 0, 0, 0, 0,
|
|
1685
|
|
* 0, 0, 1#
|
|
1686
|
|
*/
|
|
1687
|
|
#if @use@
|
|
1688
|
|
&@name@_sum_of_products_contig_outstride0_one,
|
|
1689
|
|
#else
|
|
1690
|
|
NULL,
|
|
1691
|
|
#endif
|
|
1692
|
|
/**end repeat**/
|
|
1693
|
1
|
}; /* End of _contig_outstride0_unary_specialization_table */
|
|
1694
|
1
|
|
|
1695
|
|
static sum_of_products_fn _binary_specialization_table[NPY_NTYPES][5] = {
|
|
1696
|
|
/**begin repeat
|
|
1697
|
0
|
* #name = bool,
|
|
1698
|
|
* byte, ubyte,
|
|
1699
|
|
* short, ushort,
|
|
1700
|
0
|
* int, uint,
|
|
1701
|
1
|
* long, ulong,
|
|
1702
|
|
* longlong, ulonglong,
|
|
1703
|
|
* float, double, longdouble,
|
|
1704
|
|
* cfloat, cdouble, clongdouble,
|
|
1705
|
0
|
* object, string, unicode, void,
|
|
1706
|
|
* datetime, timedelta, half#
|
|
1707
|
|
* #use = 0,
|
|
1708
|
|
* 1, 1,
|
|
1709
|
|
* 1, 1,
|
|
1710
|
|
* 1, 1,
|
|
1711
|
|
* 1, 1,
|
|
1712
|
1
|
* 1, 1,
|
|
1713
|
1
|
* 1, 1, 1,
|
|
1714
|
0
|
* 0, 0, 0,
|
|
1715
|
0
|
* 0, 0, 0, 0,
|
|
1716
|
|
* 0, 0, 1#
|
|
1717
|
|
*/
|
|
1718
|
|
#if @use@
|
|
1719
|
|
{
|
|
1720
|
1
|
&@name@_sum_of_products_stride0_contig_outstride0_two,
|
|
1721
|
|
&@name@_sum_of_products_stride0_contig_outcontig_two,
|
|
1722
|
|
&@name@_sum_of_products_contig_stride0_outstride0_two,
|
|
1723
|
|
&@name@_sum_of_products_contig_stride0_outcontig_two,
|
|
1724
|
|
&@name@_sum_of_products_contig_contig_outstride0_two,
|
|
1725
|
|
},
|
|
1726
|
|
#else
|
|
1727
|
|
{NULL, NULL, NULL, NULL, NULL},
|
|
1728
|
|
#endif
|
|
1729
|
|
/**end repeat**/
|
|
1730
|
|
}; /* End of _binary_specialization_table */
|
|
1731
|
|
|
|
1732
|
|
static sum_of_products_fn _outstride0_specialized_table[NPY_NTYPES][4] = {
|
|
1733
|
|
/**begin repeat
|
|
1734
|
|
* #name = bool,
|
|
1735
|
|
* byte, ubyte,
|
|
1736
|
|
* short, ushort,
|
|
1737
|
|
* int, uint,
|
|
1738
|
|
* long, ulong,
|
|
1739
|
|
* longlong, ulonglong,
|
|
1740
|
|
* float, double, longdouble,
|
|
1741
|
0
|
* cfloat, cdouble, clongdouble,
|
|
1742
|
|
* object, string, unicode, void,
|
|
1743
|
|
* datetime, timedelta, half#
|
|
1744
|
|
* #use = 1,
|
|
1745
|
0
|
* 1, 1,
|
|
1746
|
|
* 1, 1,
|
|
1747
|
|
* 1, 1,
|
|
1748
|
|
* 1, 1,
|
|
1749
|
|
* 1, 1,
|
|
1750
|
|
* 1, 1, 1,
|
|
1751
|
|
* 1, 1, 1,
|
|
1752
|
|
* 0, 0, 0, 0,
|
|
1753
|
|
* 0, 0, 1#
|
|
1754
|
|
*/
|
|
1755
|
|
#if @use@
|
|
1756
|
|
{
|
|
1757
|
|
&@name@_sum_of_products_outstride0_any,
|
|
1758
|
|
&@name@_sum_of_products_outstride0_one,
|
|
1759
|
|
&@name@_sum_of_products_outstride0_two,
|
|
1760
|
|
&@name@_sum_of_products_outstride0_three
|
|
1761
|
|
},
|
|
1762
|
|
#else
|
|
1763
|
|
{NULL, NULL, NULL, NULL},
|
|
1764
|
|
#endif
|
|
1765
|
|
/**end repeat**/
|
|
1766
|
|
}; /* End of _outstride0_specialized_table */
|
|
1767
|
|
|
|
1768
|
|
static sum_of_products_fn _allcontig_specialized_table[NPY_NTYPES][4] = {
|
|
1769
|
|
/**begin repeat
|
|
1770
|
|
* #name = bool,
|
|
1771
|
|
* byte, ubyte,
|
|
1772
|
|
* short, ushort,
|
|
1773
|
|
* int, uint,
|
|
1774
|
|
* long, ulong,
|
|
1775
|
|
* longlong, ulonglong,
|
|
1776
|
|
* float, double, longdouble,
|
|
1777
|
|
* cfloat, cdouble, clongdouble,
|
|
1778
|
|
* object, string, unicode, void,
|
|
1779
|
|
* datetime, timedelta, half#
|
|
1780
|
|
* #use = 1,
|
|
1781
|
|
* 1, 1,
|
|
1782
|
|
* 1, 1,
|
|
1783
|
|
* 1, 1,
|
|
1784
|
|
* 1, 1,
|
|
1785
|
|
* 1, 1,
|
|
1786
|
|
* 1, 1, 1,
|
|
1787
|
|
* 1, 1, 1,
|
|
1788
|
|
* 0, 0, 0, 0,
|
|
1789
|
|
* 0, 0, 1#
|
|
1790
|
|
*/
|
|
1791
|
|
#if @use@
|
|
1792
|
|
{
|
|
1793
|
1
|
&@name@_sum_of_products_contig_any,
|
|
1794
|
|
&@name@_sum_of_products_contig_one,
|
|
1795
|
|
&@name@_sum_of_products_contig_two,
|
|
1796
|
1
|
&@name@_sum_of_products_contig_three
|
|
1797
|
1
|
},
|
|
1798
|
1
|
#else
|
|
1799
|
1
|
{NULL, NULL, NULL, NULL},
|
|
1800
|
|
#endif
|
|
1801
|
|
/**end repeat**/
|
|
1802
|
1
|
}; /* End of _allcontig_specialized_table */
|
|
1803
|
1
|
|
|
1804
|
|
static sum_of_products_fn _unspecialized_table[NPY_NTYPES][4] = {
|
|
1805
|
|
/**begin repeat
|
|
1806
|
|
* #name = bool,
|
|
1807
|
0
|
* byte, ubyte,
|
|
1808
|
|
* short, ushort,
|
|
1809
|
0
|
* int, uint,
|
|
1810
|
0
|
* long, ulong,
|
|
1811
|
|
* longlong, ulonglong,
|
|
1812
|
0
|
* float, double, longdouble,
|
|
1813
|
0
|
* cfloat, cdouble, clongdouble,
|
|
1814
|
0
|
* object, string, unicode, void,
|
|
1815
|
|
* datetime, timedelta, half#
|
|
1816
|
|
* #use = 1,
|
|
1817
|
|
* 1, 1,
|
|
1818
|
|
* 1, 1,
|
|
1819
|
|
* 1, 1,
|
|
1820
|
1
|
* 1, 1,
|
|
1821
|
|
* 1, 1,
|
|
1822
|
|
* 1, 1, 1,
|
|
1823
|
0
|
* 1, 1, 1,
|
|
1824
|
1
|
* 0, 0, 0, 0,
|
|
1825
|
1
|
* 0, 0, 1#
|
|
1826
|
0
|
*/
|
|
1827
|
1
|
#if @use@
|
|
1828
|
1
|
{
|
|
1829
|
0
|
&@name@_sum_of_products_any,
|
|
1830
|
|
&@name@_sum_of_products_one,
|
|
1831
|
|
&@name@_sum_of_products_two,
|
|
1832
|
1
|
&@name@_sum_of_products_three
|
|
1833
|
|
},
|
|
1834
|
1
|
#else
|
|
1835
|
|
{NULL, NULL, NULL, NULL},
|
|
1836
|
|
#endif
|
|
1837
|
|
/**end repeat**/
|
|
1838
|
|
}; /* End of _unnspecialized_table */
|
|
1839
|
|
|
|
1840
|
|
NPY_VISIBILITY_HIDDEN sum_of_products_fn
|
|
1841
|
1
|
get_sum_of_products_function(int nop, int type_num,
|
|
1842
|
1
|
npy_intp itemsize, npy_intp const *fixed_strides)
|
|
1843
|
1
|
{
|
|
1844
|
0
|
int iop;
|
|
1845
|
|
|
|
1846
|
1
|
if (type_num >= NPY_NTYPES) {
|
|
1847
|
|
return NULL;
|
|
1848
|
|
}
|
|
1849
|
|
|
|
1850
|
|
/* contiguous reduction */
|
|
1851
|
1
|
if (nop == 1 && fixed_strides[0] == itemsize && fixed_strides[1] == 0) {
|
|
1852
|
1
|
sum_of_products_fn ret =
|
|
1853
|
|
_contig_outstride0_unary_specialization_table[type_num];
|
|
1854
|
1
|
if (ret != NULL) {
|
|
1855
|
|
return ret;
|
|
1856
|
|
}
|
|
1857
|
|
}
|
|
1858
|
|
|
|
1859
|
|
/* nop of 2 has more specializations */
|
|
1860
|
1
|
if (nop == 2) {
|
|
1861
|
|
/* Encode the zero/contiguous strides */
|
|
1862
|
0
|
int code;
|
|
1863
|
1
|
code = (fixed_strides[0] == 0) ? 0 :
|
|
1864
|
1
|
(fixed_strides[0] == itemsize) ? 2*2*1 : 8;
|
|
1865
|
1
|
code += (fixed_strides[1] == 0) ? 0 :
|
|
1866
|
1
|
(fixed_strides[1] == itemsize) ? 2*1 : 8;
|
|
1867
|
1
|
code += (fixed_strides[2] == 0) ? 0 :
|
|
1868
|
1
|
(fixed_strides[2] == itemsize) ? 1 : 8;
|
|
1869
|
1
|
if (code >= 2 && code < 7) {
|
|
1870
|
1
|
sum_of_products_fn ret =
|
|
1871
|
|
_binary_specialization_table[type_num][code-2];
|
|
1872
|
1
|
if (ret != NULL) {
|
|
1873
|
|
return ret;
|
|
1874
|
|
}
|
|
1875
|
0
|
}
|
|
1876
|
|
}
|
|
1877
|
0
|
|
|
1878
|
|
/* Inner loop with an output stride of 0 */
|
|
1879
|
1
|
if (fixed_strides[nop] == 0) {
|
|
1880
|
1
|
return _outstride0_specialized_table[type_num][nop <= 3 ? nop : 0];
|
|
1881
|
|
}
|
|
1882
|
|
|
|
1883
|
|
/* Check for all contiguous */
|
|
1884
|
1
|
for (iop = 0; iop < nop + 1; ++iop) {
|
|
1885
|
1
|
if (fixed_strides[iop] != itemsize) {
|
|
1886
|
|
break;
|
|
1887
|
|
}
|
|
1888
|
|
}
|
|
1889
|
|
|
|
1890
|
|
/* Contiguous loop */
|
|
1891
|
1
|
if (iop == nop + 1) {
|
|
1892
|
1
|
return _allcontig_specialized_table[type_num][nop <= 3 ? nop : 0];
|
|
1893
|
|
}
|
|
1894
|
|
|
|
1895
|
|
/* None of the above specializations caught it, general loops */
|
|
1896
|
1
|
return _unspecialized_table[type_num][nop <= 3 ? nop : 0];
|
|
1897
|
|
}
|
BvB93