IPSDK  4_1_0_2
IPSDK : Image Processing Software Development Kit
MaskCastReg.h
Go to the documentation of this file.
1 // MaskCastReg.h:
3 // ------------
4 //
14 
15 #ifndef __IPSDKUTIL_INSTRUCTIONSET_DETAIL_SSE2_MASKCASTREG_H__
16 #define __IPSDKUTIL_INSTRUCTIONSET_DETAIL_SSE2_MASKCASTREG_H__
17 
18 #include <IPSDKUtil/IPSDKUtilExports.h>
19 #include <IPSDKUtil/InstructionSet/detail/CastReg.h>
20 #include <IPSDKUtil/InstructionSet/detail/AssignReg.h>
21 #include <IPSDKUtil/InstructionSet/RegMaskType.h>
22 #include <IPSDKUtil/InstructionSet/Comparison/detail/IsNotEqualReg.h>
23 #include <IPSDKUtil/Tools/ForceInline.h>
24 
25 #include <boost/mpl/and.hpp>
26 #include <boost/mpl/equal_to.hpp>
27 #include <boost/mpl/int.hpp>
28 #include <boost/mpl/not.hpp>
29 #include <boost/mpl/or.hpp>
30 #include <boost/mpl/sizeof.hpp>
31 #include <boost/type_traits/is_signed.hpp>
32 
33 namespace ipsdk {
34 namespace simd {
35 namespace detail {
36 
39 
42 template <typename TIn, typename TOut>
43 struct MaskCastReg<
44  eInstructionSet::eIS_Sse2,
45  TIn,
46  TOut,
47  typename boost::enable_if<
48  typename boost::mpl::or_<
49  typename boost::is_same<TIn, TOut>::type,
50  typename boost::mpl::and_<
51  typename boost::is_integral<TIn>::type,
52  typename boost::mpl::and_<
53  typename boost::is_integral<TOut>::type,
54  typename boost::mpl::equal_to<
55  boost::mpl::sizeof_<TIn>,
56  boost::mpl::sizeof_<TOut>
57  >::type
58  >::type
59  >::type
60  >::type
61  >::type
62 >
63 {
64  static IPSDK_FORCEINLINE
65  void act(const typename RegMaskType<eInstructionSet::eIS_Sse2, TIn>::Type& in,
66  typename RegMaskType<eInstructionSet::eIS_Sse2, TOut>::Type& out)
67  {
68  out = in;
69  }
70 };
71 
74 template <typename TIn, typename TOut>
75 struct MaskCastReg<
76  eInstructionSet::eIS_Sse2,
77  TIn,
78  TOut,
79  typename boost::enable_if<
80  typename boost::mpl::and_<
81  typename boost::mpl::equal_to<
82  boost::mpl::sizeof_<TIn>,
83  boost::mpl::sizeof_<TOut>
84  >::type,
85  typename boost::mpl::or_<
86  typename boost::mpl::and_<
87  typename boost::mpl::not_<typename boost::is_integral<TIn>::type>::type,
88  typename boost::is_integral<TOut>::type
89  >::type,
90  typename boost::mpl::and_<
91  typename boost::mpl::not_<typename boost::is_integral<TOut>::type>::type,
92  typename boost::is_integral<TIn>::type
93  >::type
94  >::type
95  >::type
96  >::type
97 >
98 {
99  static IPSDK_FORCEINLINE
100  void act(
101  const typename RegMaskType<eInstructionSet::eIS_Sse2, TIn>::Type& in,
102  typename RegMaskType<eInstructionSet::eIS_Sse2, TOut>::Type& out)
103  {
104  BitwiseCastReg<eInstructionSet::eIS_Sse2, TIn, TOut>::act(in, out);
105  }
106 };
107 
110 /*template <typename TOut>
111 struct MaskCastReg<
112  eInstructionSet::eIS_Sse2,
113  ipReal64,
114  TOut,
115  typename boost::enable_if<
116  typename boost::mpl::and_<
117  typename boost::mpl::equal_to<
118  boost::mpl::sizeof_<TOut>,
119  boost::mpl::int_<8>
120  >::type,
121  typename boost::is_integral<TOut>::type
122  >::type
123  >::type
124 >
125 {
126  static IPSDK_FORCEINLINE
127  void act(
128  const RegMaskType<eInstructionSet::eIS_Sse2, ipReal64>::Type& in,
129  typename RegMaskType<eInstructionSet::eIS_Sse2, TOut>::Type& out)
130  {
131  BitwiseCastReg<eInstructionSet::eIS_Sse2, ipReal64, TOut>::act(in, out);
132  }
133 };*/
134 
137 template <typename TIn, typename TOut>
138 struct MaskCastReg<
139  eInstructionSet::eIS_Sse2,
140  TIn,
141  TOut,
142  typename boost::enable_if<
143  typename boost::mpl::and_<
144  typename boost::mpl::equal_to<
145  boost::mpl::sizeof_<TIn>,
146  boost::mpl::int_<1>
147  >::type,
148  typename boost::mpl::equal_to<
149  boost::mpl::sizeof_<TOut>,
150  boost::mpl::int_<2>
151  >::type
152  >::type
153  >::type
154 >
155 {
156  static IPSDK_FORCEINLINE
157  void act(
158  const typename RegMaskType<eInstructionSet::eIS_Sse2, TIn>::Type& in,
159  typename RegMaskType<eInstructionSet::eIS_Sse2, TOut>::Type& out0,
160  typename RegMaskType<eInstructionSet::eIS_Sse2, TOut>::Type& out1)
161  {
162  typename RegType<eInstructionSet::eIS_Sse2, TOut>::Type zero;
163  AssignReg<eInstructionSet::eIS_Sse2, TOut>::act(zero, 0);
164  CastReg<eInstructionSet::eIS_Sse2, TIn, TOut>::act(in, out0, out1);
165  IsNotEqualReg<eInstructionSet::eIS_Sse2, TOut>::act(out0, zero, out0);
166  IsNotEqualReg<eInstructionSet::eIS_Sse2, TOut>::act(out1, zero, out1);
167  }
168 };
169 
172 template <typename TIn, typename TOut>
173 struct MaskCastReg<
174  eInstructionSet::eIS_Sse2,
175  TIn,
176  TOut,
177  typename boost::enable_if<
178  typename boost::mpl::and_<
179  typename boost::mpl::equal_to<
180  boost::mpl::sizeof_<TIn>,
181  boost::mpl::int_<1>
182  >::type,
183  typename boost::mpl::equal_to<
184  boost::mpl::sizeof_<TOut>,
185  boost::mpl::int_<4>
186  >::type
187  >::type
188  >::type
189 >
190 {
191  static IPSDK_FORCEINLINE
192  void act(
193  const typename RegMaskType<eInstructionSet::eIS_Sse2, TIn>::Type& in,
194  typename RegMaskType<eInstructionSet::eIS_Sse2, TOut>::Type& out0,
195  typename RegMaskType<eInstructionSet::eIS_Sse2, TOut>::Type& out1,
196  typename RegMaskType<eInstructionSet::eIS_Sse2, TOut>::Type& out2,
197  typename RegMaskType<eInstructionSet::eIS_Sse2, TOut>::Type& out3)
198  {
199  typedef typename UpperType<TIn>::Type T16;
200  typename RegMaskType<eInstructionSet::eIS_Sse2, T16>::Type in0, in1;
201  CastReg<eInstructionSet::eIS_Sse2, TIn, T16>::act(in, in0, in1);
202  MaskCastReg<eInstructionSet::eIS_Sse2, T16, TOut>::act(in0, out0, out1);
203  MaskCastReg<eInstructionSet::eIS_Sse2, T16, TOut>::act(in1, out2, out3);
204  }
205 };
206 
209 template <typename TIn, typename TOut>
210 struct MaskCastReg<
211  eInstructionSet::eIS_Sse2,
212  TIn,
213  TOut,
214  typename boost::enable_if<
215  typename boost::mpl::and_<
216  typename boost::mpl::equal_to<
217  boost::mpl::sizeof_<TIn>,
218  boost::mpl::int_<1>
219  >::type,
220  typename boost::mpl::equal_to<
221  boost::mpl::sizeof_<TOut>,
222  boost::mpl::int_<8>
223  >::type
224  >::type
225  >::type
226 >
227 {
228  static IPSDK_FORCEINLINE
229  void act(
230  const typename RegMaskType<eInstructionSet::eIS_Sse2, TIn>::Type& in,
231  typename RegMaskType<eInstructionSet::eIS_Sse2, TOut>::Type& out0,
232  typename RegMaskType<eInstructionSet::eIS_Sse2, TOut>::Type& out1,
233  typename RegMaskType<eInstructionSet::eIS_Sse2, TOut>::Type& out2,
234  typename RegMaskType<eInstructionSet::eIS_Sse2, TOut>::Type& out3,
235  typename RegMaskType<eInstructionSet::eIS_Sse2, TOut>::Type& out4,
236  typename RegMaskType<eInstructionSet::eIS_Sse2, TOut>::Type& out5,
237  typename RegMaskType<eInstructionSet::eIS_Sse2, TOut>::Type& out6,
238  typename RegMaskType<eInstructionSet::eIS_Sse2, TOut>::Type& out7)
239  {
240  typedef typename UpperType<TIn>::Type T16;
241  typename RegMaskType<eInstructionSet::eIS_Sse2, T16>::Type in16_0, in16_1;
242  CastReg<eInstructionSet::eIS_Sse2, TIn, T16>::act(in, in16_0, in16_1);
243 
244  typedef typename UpperType<T16>::Type T32;
245  typename RegMaskType<eInstructionSet::eIS_Sse2, T32>::Type in32_0, in32_1, in32_2, in32_3;
246  CastReg<eInstructionSet::eIS_Sse2, T16, T32>::act(in16_0, in32_0, in32_1);
247  CastReg<eInstructionSet::eIS_Sse2, T16, T32>::act(in16_1, in32_2, in32_3);
248 
249  MaskCastReg<eInstructionSet::eIS_Sse2, T32, TOut>::act(in32_0, out0, out1);
250  MaskCastReg<eInstructionSet::eIS_Sse2, T32, TOut>::act(in32_1, out2, out3);
251  MaskCastReg<eInstructionSet::eIS_Sse2, T32, TOut>::act(in32_2, out4, out5);
252  MaskCastReg<eInstructionSet::eIS_Sse2, T32, TOut>::act(in32_3, out6, out7);
253  }
254 };
255 
258 template <typename TIn, typename TOut>
259 struct MaskCastReg<
260  eInstructionSet::eIS_Sse2,
261  TIn,
262  TOut,
263  typename boost::enable_if<
264  typename boost::mpl::and_<
265  typename boost::mpl::equal_to<
266  boost::mpl::sizeof_<TIn>,
267  boost::mpl::int_<2>
268  >::type,
269  typename boost::mpl::equal_to<
270  boost::mpl::sizeof_<TOut>,
271  boost::mpl::int_<4>
272  >::type
273  >::type
274  >::type
275 >
276 {
277  static IPSDK_FORCEINLINE
278  void act(
279  const typename RegMaskType<eInstructionSet::eIS_Sse2, TIn>::Type& in,
280  typename RegMaskType<eInstructionSet::eIS_Sse2, TOut>::Type& out0,
281  typename RegMaskType<eInstructionSet::eIS_Sse2, TOut>::Type& out1)
282  {
283  typename RegType<eInstructionSet::eIS_Sse2, TOut>::Type zero;
284  AssignReg<eInstructionSet::eIS_Sse2, TOut>::act(zero, 0);
285  CastReg<eInstructionSet::eIS_Sse2, TIn, TOut>::act(in, out0, out1);
286  IsNotEqualReg<eInstructionSet::eIS_Sse2, TOut>::act(out0, zero, out0);
287  IsNotEqualReg<eInstructionSet::eIS_Sse2, TOut>::act(out1, zero, out1);
288  }
289 };
290 
293 template <typename TIn, typename TOut>
294 struct MaskCastReg<
295  eInstructionSet::eIS_Sse2,
296  TIn,
297  TOut,
298  typename boost::enable_if<
299  typename boost::mpl::and_<
300  typename boost::mpl::equal_to<
301  boost::mpl::sizeof_<TIn>,
302  boost::mpl::int_<2>
303  >::type,
304  typename boost::mpl::equal_to<
305  boost::mpl::sizeof_<TOut>,
306  boost::mpl::int_<8>
307  >::type
308  >::type
309  >::type
310 >
311 {
312  static IPSDK_FORCEINLINE
313  void act(
314  const typename RegMaskType<eInstructionSet::eIS_Sse2, TIn>::Type& in,
315  typename RegMaskType<eInstructionSet::eIS_Sse2, TOut>::Type& out0,
316  typename RegMaskType<eInstructionSet::eIS_Sse2, TOut>::Type& out1,
317  typename RegMaskType<eInstructionSet::eIS_Sse2, TOut>::Type& out2,
318  typename RegMaskType<eInstructionSet::eIS_Sse2, TOut>::Type& out3)
319  {
320  typedef typename UpperType<TIn>::Type T32;
321  typename RegMaskType<eInstructionSet::eIS_Sse2, T32>::Type in0, in1;
322  CastReg<eInstructionSet::eIS_Sse2, TIn, T32>::act(in, in0, in1);
323  MaskCastReg<eInstructionSet::eIS_Sse2, T32, TOut>::act(in0, out0, out1);
324  MaskCastReg<eInstructionSet::eIS_Sse2, T32, TOut>::act(in1, out2, out3);
325  }
326 };
327 
330 template <typename TIn, typename TOut>
331 struct MaskCastReg<
332  eInstructionSet::eIS_Sse2,
333  TIn,
334  TOut,
335  typename boost::enable_if<
336  typename boost::mpl::and_<
337  typename boost::mpl::and_<
338  typename boost::mpl::equal_to<
339  boost::mpl::sizeof_<TIn>,
340  boost::mpl::int_<4>
341  >::type,
342  typename boost::is_integral<TIn>::type
343  >::type,
344  typename boost::mpl::equal_to<
345  boost::mpl::sizeof_<TOut>,
346  boost::mpl::int_<8>
347  >::type
348  >::type
349  >::type
350 >
351 {
352  static IPSDK_FORCEINLINE
353  void act(
354  const typename RegMaskType<eInstructionSet::eIS_Sse2, TIn>::Type& in,
355  typename RegMaskType<eInstructionSet::eIS_Sse2, TOut>::Type& out0,
356  typename RegMaskType<eInstructionSet::eIS_Sse2, TOut>::Type& out1)
357  {
358  typename RegType<eInstructionSet::eIS_Sse2, TOut>::Type zero;
359  AssignReg<eInstructionSet::eIS_Sse2, TOut>::act(zero, 0);
360  CastReg<eInstructionSet::eIS_Sse2, TIn, TOut>::act(in, out0, out1);
361  IsNotEqualReg<eInstructionSet::eIS_Sse2, TOut>::act(out0, zero, out0);
362  IsNotEqualReg<eInstructionSet::eIS_Sse2, TOut>::act(out1, zero, out1);
363  }
364 };
365 
368 template <typename TOut>
369 struct MaskCastReg<
370  eInstructionSet::eIS_Sse2,
371  ipReal32,
372  TOut,
373  typename boost::enable_if<
374  typename boost::mpl::equal_to<
375  boost::mpl::sizeof_<TOut>,
376  boost::mpl::int_<8>
377  >::type
378  >::type
379 >
380 {
381  static IPSDK_FORCEINLINE
382  void act(
383  const RegMaskType<eInstructionSet::eIS_Sse2, ipReal32>::Type& in,
384  typename RegMaskType<eInstructionSet::eIS_Sse2, TOut>::Type& out0,
385  typename RegMaskType<eInstructionSet::eIS_Sse2, TOut>::Type& out1)
386  {
387  RegMaskType<eInstructionSet::eIS_Sse2, ipUInt32>::Type inUI;
388  MaskCastReg<eInstructionSet::eIS_Sse2, ipReal32, ipUInt32>::act(in, inUI);
389  MaskCastReg<eInstructionSet::eIS_Sse2, ipUInt32, TOut>::act(inUI, out0, out1);
390  }
391 };
392 
395 template <typename TIn, typename TOut>
396 struct MaskCastReg<
397  eInstructionSet::eIS_Sse2,
398  TIn,
399  TOut,
400  typename boost::enable_if<
401  typename boost::mpl::and_<
402  typename boost::mpl::equal_to<
403  boost::mpl::sizeof_<TIn>,
404  boost::mpl::int_<2>
405  >::type,
406  typename boost::mpl::equal_to<
407  boost::mpl::sizeof_<TOut>,
408  boost::mpl::int_<1>
409  >::type
410  >::type
411  >::type
412 >
413 {
414  static IPSDK_FORCEINLINE
415  void act(
416  const typename RegMaskType<eInstructionSet::eIS_Sse2, TIn>::Type& in0,
417  const typename RegMaskType<eInstructionSet::eIS_Sse2, TIn>::Type& in1,
418  typename RegMaskType<eInstructionSet::eIS_Sse2, TOut>::Type& out)
419  {
420  typename RegType<eInstructionSet::eIS_Sse2, TOut>::Type zero;
421  AssignReg<eInstructionSet::eIS_Sse2, TOut>::act(zero, 0);
422  CastReg<eInstructionSet::eIS_Sse2, ipInt16, ipInt8>::act(in0, in1, out);
423  IsNotEqualReg<eInstructionSet::eIS_Sse2, TOut>::act(out, zero, out);
424  }
425 };
426 
429 template <typename TIn, typename TOut>
430 struct MaskCastReg<
431  eInstructionSet::eIS_Sse2,
432  TIn,
433  TOut,
434  typename boost::enable_if<
435  typename boost::mpl::and_<
436  typename boost::mpl::and_<
437  typename boost::mpl::equal_to<
438  boost::mpl::sizeof_<TIn>,
439  boost::mpl::int_<4>
440  >::type,
441  typename boost::is_integral<TIn>::type
442  >::type,
443  typename boost::mpl::equal_to<
444  boost::mpl::sizeof_<TOut>,
445  boost::mpl::int_<2>
446  >::type
447  >::type
448  >::type
449 >
450 {
451  static IPSDK_FORCEINLINE
452  void act(
453  const typename RegMaskType<eInstructionSet::eIS_Sse2, TIn>::Type& in0,
454  const typename RegMaskType<eInstructionSet::eIS_Sse2, TIn>::Type& in1,
455  typename RegMaskType<eInstructionSet::eIS_Sse2, TOut>::Type& out)
456  {
457  typename RegType<eInstructionSet::eIS_Sse2, TOut>::Type zero;
458  AssignReg<eInstructionSet::eIS_Sse2, TOut>::act(zero, 0);
459  CastReg<eInstructionSet::eIS_Sse2, ipInt32, ipInt16>::act(in0, in1, out);
460  IsNotEqualReg<eInstructionSet::eIS_Sse2, TOut>::act(out, zero, out);
461  }
462 };
463 
466 template <typename TOut>
467 struct MaskCastReg<
468  eInstructionSet::eIS_Sse2,
469  ipReal32,
470  TOut,
471  typename boost::enable_if<
472  typename boost::mpl::equal_to<
473  boost::mpl::sizeof_<TOut>,
474  boost::mpl::int_<2>
475  >::type
476  >::type
477 >
478 {
479  static IPSDK_FORCEINLINE
480  void act(
481  const RegMaskType<eInstructionSet::eIS_Sse2, ipReal32>::Type& in0,
482  const typename RegMaskType<eInstructionSet::eIS_Sse2, ipReal32>::Type& in1,
483  typename RegMaskType<eInstructionSet::eIS_Sse2, TOut>::Type& out)
484  {
485  RegMaskType<eInstructionSet::eIS_Sse2, ipUInt32>::Type inUI0, inUI1;
486  MaskCastReg<eInstructionSet::eIS_Sse2, ipReal32, ipUInt32>::act(in0, inUI0);
487  MaskCastReg<eInstructionSet::eIS_Sse2, ipReal32, ipUInt32>::act(in1, inUI1);
488  MaskCastReg<eInstructionSet::eIS_Sse2, ipUInt32, TOut>::act(inUI0, inUI1, out);
489  }
490 };
491 
494 template <typename TIn, typename TOut>
495 struct MaskCastReg<
496  eInstructionSet::eIS_Sse2,
497  TIn,
498  TOut,
499  typename boost::enable_if<
500  typename boost::mpl::and_<
501  typename boost::mpl::and_<
502  typename boost::mpl::equal_to<
503  boost::mpl::sizeof_<TIn>,
504  boost::mpl::int_<4>
505  >::type,
506  typename boost::is_integral<TIn>::type
507  >::type,
508  typename boost::mpl::equal_to<
509  boost::mpl::sizeof_<TOut>,
510  boost::mpl::int_<1>
511  >::type
512  >::type
513  >::type
514 >
515 {
516  static IPSDK_FORCEINLINE
517  void act(
518  const typename RegMaskType<eInstructionSet::eIS_Sse2, TIn>::Type& in0,
519  const typename RegMaskType<eInstructionSet::eIS_Sse2, TIn>::Type& in1,
520  const typename RegMaskType<eInstructionSet::eIS_Sse2, TIn>::Type& in2,
521  const typename RegMaskType<eInstructionSet::eIS_Sse2, TIn>::Type& in3,
522  typename RegMaskType<eInstructionSet::eIS_Sse2, TOut>::Type& out)
523  {
524  RegMaskType<eInstructionSet::eIS_Sse2, ipInt16>::Type out0, out1;
525  CastReg<eInstructionSet::eIS_Sse2, ipInt32, ipInt16>::act(in0, in1, out0);
526  CastReg<eInstructionSet::eIS_Sse2, ipInt32, ipInt16>::act(in2, in3, out1);
527  MaskCastReg<eInstructionSet::eIS_Sse2, ipInt16, TOut>::act(out0, out1, out);
528  }
529 };
530 
533 template <typename TOut>
534 struct MaskCastReg<
535  eInstructionSet::eIS_Sse2,
536  ipReal32,
537  TOut,
538  typename boost::enable_if<
539  typename boost::mpl::equal_to<
540  boost::mpl::sizeof_<TOut>,
541  boost::mpl::int_<1>
542  >::type
543  >::type
544 >
545 {
546  static IPSDK_FORCEINLINE
547  void act(
548  const RegMaskType<eInstructionSet::eIS_Sse2, ipReal32>::Type& in0,
549  const RegMaskType<eInstructionSet::eIS_Sse2, ipReal32>::Type& in1,
550  const RegMaskType<eInstructionSet::eIS_Sse2, ipReal32>::Type& in2,
551  const RegMaskType<eInstructionSet::eIS_Sse2, ipReal32>::Type& in3,
552  typename RegMaskType<eInstructionSet::eIS_Sse2, TOut>::Type& out)
553  {
554  RegMaskType<eInstructionSet::eIS_Sse2, ipUInt32>::Type inUI0, inUI1, inUI2, inUI3;
555  MaskCastReg<eInstructionSet::eIS_Sse2, ipReal32, ipUInt32>::act(in0, inUI0);
556  MaskCastReg<eInstructionSet::eIS_Sse2, ipReal32, ipUInt32>::act(in1, inUI1);
557  MaskCastReg<eInstructionSet::eIS_Sse2, ipReal32, ipUInt32>::act(in2, inUI2);
558  MaskCastReg<eInstructionSet::eIS_Sse2, ipReal32, ipUInt32>::act(in3, inUI3);
559  MaskCastReg<eInstructionSet::eIS_Sse2, ipUInt32, TOut>::act(inUI0, inUI1, inUI2, inUI3, out);
560  }
561 };
562 
565 /*template <typename TIn, typename TOut>
566 struct MaskCastReg<
567  eInstructionSet::eIS_Sse2,
568  TIn,
569  TOut,
570  typename boost::enable_if<
571  typename boost::mpl::and_<
572  typename boost::mpl::and_<
573  typename boost::mpl::equal_to<
574  boost::mpl::sizeof_<TIn>,
575  boost::mpl::int_<8>
576  >::type,
577  typename boost::is_integral<TIn>::type
578  >::type,
579  typename boost::mpl::and_<
580  typename boost::mpl::equal_to<
581  boost::mpl::sizeof_<TOut>,
582  boost::mpl::int_<4>
583  >::type,
584  typename boost::is_integral<TOut>::type
585  >::type
586  >::type
587  >::type
588 >
589 {
590  static IPSDK_FORCEINLINE
591  void act(
592  const typename RegMaskType<eInstructionSet::eIS_Sse2, TIn>::Type& in0,
593  const typename RegMaskType<eInstructionSet::eIS_Sse2, TIn>::Type& in1,
594  typename RegMaskType<eInstructionSet::eIS_Sse2, TOut>::Type& out)
595  {
596  typename RegType<eInstructionSet::eIS_Sse2, TOut>::Type zero;
597  AssignReg<eInstructionSet::eIS_Sse2, TOut>::act(zero, 0);
598  CastReg<eInstructionSet::eIS_Sse2, ipInt64, ipInt32>::act(in0, in1, out);
599  IsNotEqualReg<eInstructionSet::eIS_Sse2, TOut>::act(out, zero, out);
600  }
601 };*/
602 
605 template <typename TOut>
606 struct MaskCastReg<
607  eInstructionSet::eIS_Sse2,
608  ipReal64,
609  TOut,
610  typename boost::enable_if<
611  typename boost::mpl::and_<
612  typename boost::mpl::equal_to<
613  boost::mpl::sizeof_<TOut>,
614  boost::mpl::int_<4>
615  >::type,
616  typename boost::is_integral<TOut>::type
617  >::type
618  >::type
619 >
620 {
621  static IPSDK_FORCEINLINE
622  void act(
623  const RegMaskType<eInstructionSet::eIS_Sse2, ipReal64>::Type& in0,
624  const RegMaskType<eInstructionSet::eIS_Sse2, ipReal64>::Type& in1,
625  typename RegMaskType<eInstructionSet::eIS_Sse2, TOut>::Type& out)
626  {
627  typename RegType<eInstructionSet::eIS_Sse2, ipReal32>::Type outI;
628  CastReg<eInstructionSet::eIS_Sse2, ipReal64, ipReal32>::act(in0, in1, outI);
629  MaskCastReg<eInstructionSet::eIS_Sse2, ipReal32, TOut>::act(outI, out);
630  }
631 };
632 
635 template <typename TIn>
636 struct MaskCastReg<
637  eInstructionSet::eIS_Sse2,
638  TIn,
639  ipReal32,
640  typename boost::enable_if<
641  typename boost::mpl::and_<
642  typename boost::mpl::equal_to<
643  boost::mpl::sizeof_<TIn>,
644  boost::mpl::int_<8>
645  >::type,
646  typename boost::is_integral<TIn>::type
647  >::type
648  >::type
649 >
650 {
651  static IPSDK_FORCEINLINE
652  void act(
653  const typename RegMaskType<eInstructionSet::eIS_Sse2, TIn>::Type& in0,
654  const typename RegMaskType<eInstructionSet::eIS_Sse2, TIn>::Type& in1,
655  RegMaskType<eInstructionSet::eIS_Sse2, ipReal32>::Type& out)
656  {
657  RegMaskType<eInstructionSet::eIS_Sse2, ipInt32>::Type outI;
658  MaskCastReg<eInstructionSet::eIS_Sse2, TIn, ipInt32>::act(in0, in1, outI);
659  BitwiseCastReg<eInstructionSet::eIS_Sse2, ipInt32, ipReal32>::act(outI, out);
660  }
661 };
662 
665 template <typename TIn, typename TOut>
666 struct MaskCastReg<
667  eInstructionSet::eIS_Sse2,
668  TIn,
669  TOut,
670  typename boost::enable_if<
671  typename boost::mpl::and_<
672  typename boost::mpl::and_<
673  typename boost::mpl::equal_to<
674  boost::mpl::sizeof_<TIn>,
675  boost::mpl::int_<8>
676  >::type,
677  typename boost::is_integral<TIn>::type
678  >::type,
679  typename boost::mpl::and_<
680  typename boost::mpl::equal_to<
681  boost::mpl::sizeof_<TOut>,
682  boost::mpl::int_<2>
683  >::type,
684  typename boost::is_integral<TOut>::type
685  >::type
686  >::type
687  >::type
688 >
689 {
690  static IPSDK_FORCEINLINE
691  void act(
692  const typename RegMaskType<eInstructionSet::eIS_Sse2, TIn>::Type& in0,
693  const typename RegMaskType<eInstructionSet::eIS_Sse2, TIn>::Type& in1,
694  const typename RegMaskType<eInstructionSet::eIS_Sse2, TIn>::Type& in2,
695  const typename RegMaskType<eInstructionSet::eIS_Sse2, TIn>::Type& in3,
696  typename RegMaskType<eInstructionSet::eIS_Sse2, TOut>::Type& out)
697  {
698  RegMaskType<eInstructionSet::eIS_Sse2, ipInt32>::Type out0, out1;
699  CastReg<eInstructionSet::eIS_Sse2, TIn, ipInt32>::act(in0, in1, out0);
700  CastReg<eInstructionSet::eIS_Sse2, TIn, ipInt32>::act(in2, in3, out1);
701  CastReg<eInstructionSet::eIS_Sse2, ipInt32, ipInt16>::act(out0, out1, out);
702  }
703 };
704 
707 template <typename TOut>
708 struct MaskCastReg<
709  eInstructionSet::eIS_Sse2,
710  ipReal64,
711  TOut,
712  typename boost::enable_if<
713  typename boost::mpl::and_<
714  typename boost::mpl::equal_to<
715  boost::mpl::sizeof_<TOut>,
716  boost::mpl::int_<2>
717  >::type,
718  typename boost::is_integral<TOut>::type
719  >::type
720  >::type
721 >
722 {
723  static IPSDK_FORCEINLINE
724  void act(
725  const RegMaskType<eInstructionSet::eIS_Sse2, ipReal64>::Type& in0,
726  const RegMaskType<eInstructionSet::eIS_Sse2, ipReal64>::Type& in1,
727  const RegMaskType<eInstructionSet::eIS_Sse2, ipReal64>::Type& in2,
728  const RegMaskType<eInstructionSet::eIS_Sse2, ipReal64>::Type& in3,
729  typename RegMaskType<eInstructionSet::eIS_Sse2, TOut>::Type& out)
730  {
731 
732  RegMaskType<eInstructionSet::eIS_Sse2, ipReal32>::Type in0F32, in1F32;
733  CastReg<eInstructionSet::eIS_Sse2, ipReal64, ipReal32>::act(in0, in1, in0F32);
734  CastReg<eInstructionSet::eIS_Sse2, ipReal64, ipReal32>::act(in2, in3, in1F32);
735 
736  MaskCastReg<eInstructionSet::eIS_Sse2, ipReal32, TOut>::act(
737  in0F32, in1F32, out);
738  }
739 };
740 
743 template <typename TIn, typename TOut>
744 struct MaskCastReg<
745  eInstructionSet::eIS_Sse2,
746  TIn,
747  TOut,
748  typename boost::enable_if<
749  typename boost::mpl::and_<
750  typename boost::mpl::and_<
751  typename boost::mpl::equal_to<
752  boost::mpl::sizeof_<TIn>,
753  boost::mpl::int_<8>
754  >::type,
755  typename boost::is_integral<TIn>::type
756  >::type,
757  typename boost::mpl::and_<
758  typename boost::mpl::equal_to<
759  boost::mpl::sizeof_<TOut>,
760  boost::mpl::int_<1>
761  >::type,
762  typename boost::is_integral<TOut>::type
763  >::type
764  >::type
765  >::type
766 >
767 {
768  static IPSDK_FORCEINLINE
769  void act(
770  const typename RegMaskType<eInstructionSet::eIS_Sse2, TIn>::Type& in0,
771  const typename RegMaskType<eInstructionSet::eIS_Sse2, TIn>::Type& in1,
772  const typename RegMaskType<eInstructionSet::eIS_Sse2, TIn>::Type& in2,
773  const typename RegMaskType<eInstructionSet::eIS_Sse2, TIn>::Type& in3,
774  const typename RegMaskType<eInstructionSet::eIS_Sse2, TIn>::Type& in4,
775  const typename RegMaskType<eInstructionSet::eIS_Sse2, TIn>::Type& in5,
776  const typename RegMaskType<eInstructionSet::eIS_Sse2, TIn>::Type& in6,
777  const typename RegMaskType<eInstructionSet::eIS_Sse2, TIn>::Type& in7,
778  typename RegMaskType<eInstructionSet::eIS_Sse2, TOut>::Type& out)
779  {
780  RegMaskType<eInstructionSet::eIS_Sse2, ipInt32>::Type out0, out1, out2, out3;
781  CastReg<eInstructionSet::eIS_Sse2, TIn, ipInt32>::act(in0, in1, out0);
782  CastReg<eInstructionSet::eIS_Sse2, TIn, ipInt32>::act(in2, in3, out1);
783  CastReg<eInstructionSet::eIS_Sse2, TIn, ipInt32>::act(in4, in5, out2);
784  CastReg<eInstructionSet::eIS_Sse2, TIn, ipInt32>::act(in6, in7, out3);
785  MaskCastReg<eInstructionSet::eIS_Sse2, ipInt32, ipInt8>::act(out0, out1, out2, out3, out);
786  }
787 };
788 
791 template <typename TOut>
792 struct MaskCastReg<
793  eInstructionSet::eIS_Sse2,
794  ipReal64,
795  TOut,
796  typename boost::enable_if<
797  typename boost::mpl::and_<
798  typename boost::mpl::equal_to<
799  boost::mpl::sizeof_<TOut>,
800  boost::mpl::int_<1>
801  >::type,
802  typename boost::is_integral<TOut>::type
803  >::type
804  >::type
805 >
806 {
807  static IPSDK_FORCEINLINE
808  void act(
809  const RegMaskType<eInstructionSet::eIS_Sse2, ipReal64>::Type& in0,
810  const RegMaskType<eInstructionSet::eIS_Sse2, ipReal64>::Type& in1,
811  const RegMaskType<eInstructionSet::eIS_Sse2, ipReal64>::Type& in2,
812  const RegMaskType<eInstructionSet::eIS_Sse2, ipReal64>::Type& in3,
813  const RegMaskType<eInstructionSet::eIS_Sse2, ipReal64>::Type& in4,
814  const RegMaskType<eInstructionSet::eIS_Sse2, ipReal64>::Type& in5,
815  const RegMaskType<eInstructionSet::eIS_Sse2, ipReal64>::Type& in6,
816  const RegMaskType<eInstructionSet::eIS_Sse2, ipReal64>::Type& in7,
817  typename RegMaskType<eInstructionSet::eIS_Sse2, TOut>::Type& out)
818  {
819 
820  RegMaskType<eInstructionSet::eIS_Sse2, ipInt64>::Type in0I, in1I, in2I, in3I, in4I, in5I, in6I, in7I;
821  BitwiseCastReg<eInstructionSet::eIS_Sse2, ipReal64, ipInt64>::act(in0, in0I);
822  BitwiseCastReg<eInstructionSet::eIS_Sse2, ipReal64, ipInt64>::act(in1, in1I);
823  BitwiseCastReg<eInstructionSet::eIS_Sse2, ipReal64, ipInt64>::act(in2, in2I);
824  BitwiseCastReg<eInstructionSet::eIS_Sse2, ipReal64, ipInt64>::act(in3, in3I);
825  BitwiseCastReg<eInstructionSet::eIS_Sse2, ipReal64, ipInt64>::act(in4, in4I);
826  BitwiseCastReg<eInstructionSet::eIS_Sse2, ipReal64, ipInt64>::act(in5, in5I);
827  BitwiseCastReg<eInstructionSet::eIS_Sse2, ipReal64, ipInt64>::act(in6, in6I);
828  BitwiseCastReg<eInstructionSet::eIS_Sse2, ipReal64, ipInt64>::act(in7, in7I);
829 
830  MaskCastReg<eInstructionSet::eIS_Sse2, ipInt64, TOut>::act(
831  in0I, in1I, in2I, in3I, in4I, in5I, in6I, in7I, out);
832  }
833 };
836 
837 } // end of namespace detail
838 } // end of namespace simd
839 } // end of namespace ipsdk
840 
841 #endif // __IPSDKUTIL_INSTRUCTIONSET_DETAIL_SSE2_MASKCASTREG_H__
ipsdk::simd::detail::MaskCastReg
Definition: MaskCastReg.h:30
ForceInline.h
Defines the IPSDK_FORCEINLINE.
ipsdk
Main namespace for IPSDK library.
Definition: AlgorithmFunctionEfficiency.h:22
ipsdk::ipReal64
double ipReal64
Base types definition.
Definition: BaseTypes.h:57
ipsdk::simd::detail::CastReg
Definition: CastReg.h:30
ipsdk::eInstructionSet
eInstructionSet
Enumerate for processor instruction set description.
Definition: InstructionSetTypes.h:31
ipsdk::UpperType
template structure UpperType<typename T>; its typedef Type gives the type just upper to T...
Definition: UpperType.h:42
RegMaskType.h
RegMaskType class.
IPSDKUtilExports.h
Definition of import/export macro for library.
ipsdk::simd::detail::IsNotEqualReg
Definition: IsNotEqualRegDecl.h:30
ipsdk::simd::RegMaskType
Definition: RegMaskType.h:29
ipsdk::simd::detail::BitwiseCastReg
Definition: BitwiseCastReg.h:29
ipsdk::eIS_Sse2
Streaming SIMD Extensions 2.
Definition: InstructionSetTypes.h:36
ipsdk::simd::RegType
Definition: RegType.h:29
ipsdk::simd::detail::AssignReg
Definition: AssignRegDecl.h:31
ipsdk::ipReal32
float ipReal32
Base types definition.
Definition: BaseTypes.h:56
Generated on Tue Apr 15 2025 16:20:43 for IPSDK by   doxygen 1.8.14