IPSDK  4_1_0_2
IPSDK : Image Processing Software Development Kit
ConditionalAddReg.h
Go to the documentation of this file.
1 // ConditionalAddReg.h:
3 // -------------------
4 //
13 
14 #ifndef __IPSDKUTIL_INSTRUCTIONSET_ARITHMETIC_DETAIL_AVX512_CONDITIONALADDREG_H__
15 #define __IPSDKUTIL_INSTRUCTIONSET_ARITHMETIC_DETAIL_AVX512_CONDITIONALADDREG_H__
16 
17 #include <IPSDKUtil/IPSDKUtilExports.h>
18 #include <IPSDKUtil/InstructionSet/Arithmetic/detail/ConditionalAddReg.h>
19 #include <IPSDKUtil/InstructionSet/RegType.h>
20 #include <IPSDKUtil/Tools/ForceInline.h>
21 
22 #include "immintrin.h"
23 
24 namespace ipsdk {
25 namespace simd {
26 namespace detail {
27 
30 
33 template <typename T>
34 struct ConditionalAddReg<
35  eInstructionSet::eIS_Avx512,
36  T,
37  typename boost::enable_if<
38  typename boost::mpl::and_<
39  typename boost::is_integral<T>::type,
40  typename boost::mpl::equal_to<
41  boost::mpl::sizeof_<T>,
42  boost::mpl::int_<1>
43  >::type
44  >::type
45  >::type
46 >
47 {
48  static IPSDK_FORCEINLINE
49  typename RegType<eInstructionSet::eIS_Avx512, T>::Type
50  act(const typename RegMaskType<eInstructionSet::eIS_Avx512, T>::Type& mask,
51  const typename RegType<eInstructionSet::eIS_Avx512, T>::Type& in1,
52  const typename RegType<eInstructionSet::eIS_Avx512, T>::Type& in2)
53  {
54  typename RegType<eInstructionSet::eIS_Avx512, T>::Type out;
55  act(mask, in1, in2, out);
56  return out;
57  }
58 
59  static IPSDK_FORCEINLINE
60  typename RegType<eInstructionSet::eIS_Avx512, T>::Type
61  act(const typename RegMaskType<eInstructionSet::eIS_Avx512, T>::Type& mask,
62  const typename RegType<eInstructionSet::eIS_Avx512, T>::Type& in1,
63  const T& in2)
64  {
65  typename RegType<eInstructionSet::eIS_Avx512, T>::Type out;
66  act(mask, in1, in2, out);
67  return out;
68  }
69 
70  static IPSDK_FORCEINLINE
71  void
72  act(const typename RegMaskType<eInstructionSet::eIS_Avx512, T>::Type& mask,
73  const typename RegType<eInstructionSet::eIS_Avx512, T>::Type& in1,
74  const typename RegType<eInstructionSet::eIS_Avx512, T>::Type& in2,
75  typename RegType<eInstructionSet::eIS_Avx512, T>::Type& out)
76  {
77  out = _mm512_mask_add_epi8(in1, mask, in1, in2);
78  }
79 
80  static IPSDK_FORCEINLINE
81  void
82  act(const typename RegMaskType<eInstructionSet::eIS_Avx512, T>::Type& mask,
83  const typename RegType<eInstructionSet::eIS_Avx512, T>::Type& in1,
84  const T& in2,
85  typename RegType<eInstructionSet::eIS_Avx512, T>::Type& out)
86  {
87  const typename RegType<eInstructionSet::eIS_Avx512, T>::Type regIn2 =
88  AssignReg<eInstructionSet::eIS_Avx512, T>::act(in2);
89  act(mask, in1, regIn2, out);
90  }
91 };
92 
95 template <typename T>
96 struct ConditionalAddReg<
97  eInstructionSet::eIS_Avx512,
98  T,
99  typename boost::enable_if<
100  typename boost::mpl::and_<
101  typename boost::is_integral<T>::type,
102  typename boost::mpl::equal_to<
103  boost::mpl::sizeof_<T>,
104  boost::mpl::int_<2>
105  >::type
106  >::type
107  >::type
108 >
109 {
110  static IPSDK_FORCEINLINE
111  typename RegType<eInstructionSet::eIS_Avx512, T>::Type
112  act(const typename RegMaskType<eInstructionSet::eIS_Avx512, T>::Type& mask,
113  const typename RegType<eInstructionSet::eIS_Avx512, T>::Type& in1,
114  const typename RegType<eInstructionSet::eIS_Avx512, T>::Type& in2)
115  {
116  typename RegType<eInstructionSet::eIS_Avx512, T>::Type out;
117  act(mask, in1, in2, out);
118  return out;
119  }
120 
121  static IPSDK_FORCEINLINE
122  typename RegType<eInstructionSet::eIS_Avx512, T>::Type
123  act(const typename RegMaskType<eInstructionSet::eIS_Avx512, T>::Type& mask,
124  const typename RegType<eInstructionSet::eIS_Avx512, T>::Type& in1,
125  const T& in2)
126  {
127  typename RegType<eInstructionSet::eIS_Avx512, T>::Type out;
128  act(mask, in1, in2, out);
129  return out;
130  }
131 
132  static IPSDK_FORCEINLINE
133  void
134  act(const typename RegMaskType<eInstructionSet::eIS_Avx512, T>::Type& mask,
135  const typename RegType<eInstructionSet::eIS_Avx512, T>::Type& in1,
136  const typename RegType<eInstructionSet::eIS_Avx512, T>::Type& in2,
137  typename RegType<eInstructionSet::eIS_Avx512, T>::Type& out)
138  {
139  out = _mm512_mask_add_epi16(in1, mask, in1, in2);
140  }
141 
142  static IPSDK_FORCEINLINE
143  void
144  act(const typename RegMaskType<eInstructionSet::eIS_Avx512, T>::Type& mask,
145  const typename RegType<eInstructionSet::eIS_Avx512, T>::Type& in1,
146  const T& in2,
147  typename RegType<eInstructionSet::eIS_Avx512, T>::Type& out)
148  {
149  const typename RegType<eInstructionSet::eIS_Avx512, T>::Type regIn2 =
150  AssignReg<eInstructionSet::eIS_Avx512, T>::act(in2);
151  act(mask, in1, regIn2, out);
152  }
153 };
154 
157 template <typename T>
158 struct ConditionalAddReg<
159  eInstructionSet::eIS_Avx512,
160  T,
161  typename boost::enable_if<
162  typename boost::mpl::and_<
163  typename boost::is_integral<T>::type,
164  typename boost::mpl::equal_to<
165  boost::mpl::sizeof_<T>,
166  boost::mpl::int_<4>
167  >::type
168  >::type
169  >::type
170 >
171 {
172  static IPSDK_FORCEINLINE
173  typename RegType<eInstructionSet::eIS_Avx512, T>::Type
174  act(const typename RegMaskType<eInstructionSet::eIS_Avx512, T>::Type& mask,
175  const typename RegType<eInstructionSet::eIS_Avx512, T>::Type& in1,
176  const typename RegType<eInstructionSet::eIS_Avx512, T>::Type& in2)
177  {
178  typename RegType<eInstructionSet::eIS_Avx512, T>::Type out;
179  act(mask, in1, in2, out);
180  return out;
181  }
182 
183  static IPSDK_FORCEINLINE
184  typename RegType<eInstructionSet::eIS_Avx512, T>::Type
185  act(const typename RegMaskType<eInstructionSet::eIS_Avx512, T>::Type& mask,
186  const typename RegType<eInstructionSet::eIS_Avx512, T>::Type& in1,
187  const T& in2)
188  {
189  typename RegType<eInstructionSet::eIS_Avx512, T>::Type out;
190  act(mask, in1, in2, out);
191  return out;
192  }
193 
194  static IPSDK_FORCEINLINE
195  void
196  act(const typename RegMaskType<eInstructionSet::eIS_Avx512, T>::Type& mask,
197  const typename RegType<eInstructionSet::eIS_Avx512, T>::Type& in1,
198  const typename RegType<eInstructionSet::eIS_Avx512, T>::Type& in2,
199  typename RegType<eInstructionSet::eIS_Avx512, T>::Type& out)
200  {
201  out = _mm512_mask_add_epi32(in1, mask, in1, in2);
202  }
203 
204  static IPSDK_FORCEINLINE
205  void
206  act(const typename RegMaskType<eInstructionSet::eIS_Avx512, T>::Type& mask,
207  const typename RegType<eInstructionSet::eIS_Avx512, T>::Type& in1,
208  const T& in2,
209  typename RegType<eInstructionSet::eIS_Avx512, T>::Type& out)
210  {
211  const typename RegType<eInstructionSet::eIS_Avx512, T>::Type regIn2 =
212  AssignReg<eInstructionSet::eIS_Avx512, T>::act(in2);
213  act(mask, in1, regIn2, out);
214  }
215 };
216 
219 template <>
220 struct ConditionalAddReg<eInstructionSet::eIS_Avx512, ipReal32>
221 {
222  static IPSDK_FORCEINLINE
223  RegType<eInstructionSet::eIS_Avx512, ipReal32>::Type
224  act(const RegMaskType<eInstructionSet::eIS_Avx512, ipReal32>::Type& mask,
225  const RegType<eInstructionSet::eIS_Avx512, ipReal32>::Type& in1,
226  const RegType<eInstructionSet::eIS_Avx512, ipReal32>::Type& in2)
227  {
228  RegType<eInstructionSet::eIS_Avx512, ipReal32>::Type out;
229  act(mask, in1, in2, out);
230  return out;
231  }
232 
233  static IPSDK_FORCEINLINE
234  RegType<eInstructionSet::eIS_Avx512, ipReal32>::Type
235  act(const RegMaskType<eInstructionSet::eIS_Avx512, ipReal32>::Type& mask,
236  const RegType<eInstructionSet::eIS_Avx512, ipReal32>::Type& in1,
237  const ipReal32& in2)
238  {
239  RegType<eInstructionSet::eIS_Avx512, ipReal32>::Type out;
240  act(mask, in1, in2, out);
241  return out;
242  }
243 
244  static IPSDK_FORCEINLINE
245  void
246  act(const RegMaskType<eInstructionSet::eIS_Avx512, ipReal32>::Type& mask,
247  const RegType<eInstructionSet::eIS_Avx512, ipReal32>::Type& in1,
248  const RegType<eInstructionSet::eIS_Avx512, ipReal32>::Type& in2,
249  RegType<eInstructionSet::eIS_Avx512, ipReal32>::Type& out)
250  {
251  out = _mm512_mask_add_ps(in1, mask, in1, in2);
252  }
253 
254  static IPSDK_FORCEINLINE
255  void
256  act(const RegMaskType<eInstructionSet::eIS_Avx512, ipReal32>::Type& mask,
257  const RegType<eInstructionSet::eIS_Avx512, ipReal32>::Type& in1,
258  const ipReal32& in2,
259  RegType<eInstructionSet::eIS_Avx512, ipReal32>::Type& out)
260  {
261  const RegType<eInstructionSet::eIS_Avx512, ipReal32>::Type regIn2 =
262  AssignReg<eInstructionSet::eIS_Avx512, ipReal32>::act(in2);
263  act(mask, in1, regIn2, out);
264  }
265 };
266 
269 template <>
270 struct ConditionalAddReg<eInstructionSet::eIS_Avx512, ipReal64>
271 {
272  static IPSDK_FORCEINLINE
273  RegType<eInstructionSet::eIS_Avx512, ipReal64>::Type
274  act(const RegMaskType<eInstructionSet::eIS_Avx512, ipReal64>::Type& mask,
275  const RegType<eInstructionSet::eIS_Avx512, ipReal64>::Type& in1,
276  const RegType<eInstructionSet::eIS_Avx512, ipReal64>::Type& in2)
277  {
278  RegType<eInstructionSet::eIS_Avx512, ipReal64>::Type out;
279  act(mask, in1, in2, out);
280  return out;
281  }
282 
283  static IPSDK_FORCEINLINE
284  RegType<eInstructionSet::eIS_Avx512, ipReal64>::Type
285  act(const RegMaskType<eInstructionSet::eIS_Avx512, ipReal64>::Type& mask,
286  const RegType<eInstructionSet::eIS_Avx512, ipReal64>::Type& in1,
287  const ipReal64& in2)
288  {
289  RegType<eInstructionSet::eIS_Avx512, ipReal64>::Type out;
290  act(mask, in1, in2, out);
291  return out;
292  }
293 
294  static IPSDK_FORCEINLINE
295  void
296  act(const RegMaskType<eInstructionSet::eIS_Avx512, ipReal64>::Type& mask,
297  const RegType<eInstructionSet::eIS_Avx512, ipReal64>::Type& in1,
298  const RegType<eInstructionSet::eIS_Avx512, ipReal64>::Type& in2,
299  RegType<eInstructionSet::eIS_Avx512, ipReal64>::Type& out)
300  {
301  out = _mm512_mask_add_pd(in1, mask, in1, in2);
302  }
303 
304  static IPSDK_FORCEINLINE
305  void
306  act(const RegMaskType<eInstructionSet::eIS_Avx512, ipReal64>::Type& mask,
307  const RegType<eInstructionSet::eIS_Avx512, ipReal64>::Type& in1,
308  const ipReal64& in2,
309  RegType<eInstructionSet::eIS_Avx512, ipReal64>::Type& out)
310  {
311  const RegType<eInstructionSet::eIS_Avx512, ipReal64>::Type regIn2 =
312  AssignReg<eInstructionSet::eIS_Avx512, ipReal64>::act(in2);
313  act(mask, in1, regIn2, out);
314  }
315 };
318 
319 } // end of namespace detail
320 } // end of namespace simd
321 } // end of namespace ipsdk
322 
323 #endif // __IPSDKUTIL_INSTRUCTIONSET_ARITHMETIC_DETAIL_AVX512_CONDITIONALADDREG_H__
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::eIS_Avx512
(including fundation and byte and word instructions)
Definition: InstructionSetTypes.h:51
RegType.h
RegType class.
ipsdk::eInstructionSet
eInstructionSet
Enumerate for processor instruction set description.
Definition: InstructionSetTypes.h:31
ipsdk::simd::detail::ConditionalAddReg
Definition: ConditionalAddReg.h:36
IPSDKUtilExports.h
Definition of import/export macro for library.
ipsdk::simd::RegMaskType
Definition: RegMaskType.h:29
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 16:20:39 for IPSDK by   doxygen 1.8.14