IPSDK  4_1_0_2
IPSDK : Image Processing Software Development Kit
IsOutOfRangeReg.h
Go to the documentation of this file.
1 // IsOutOfRangeReg.h:
3 // -------------------
4 //
14 
15 #ifndef __IPSDKUTIL_INSTRUCTIONSET_COMPARISON_DETAIL_SSE2_ISOUTOFRANGEREG_H__
16 #define __IPSDKUTIL_INSTRUCTIONSET_COMPARISON_DETAIL_SSE2_ISOUTOFRANGEREG_H__
17 
18 #include <IPSDKUtil/IPSDKUtilExports.h>
19 #include <IPSDKUtil/InstructionSet/SSE2/Sse2Types.h>
20 #include <IPSDKUtil/InstructionSet/Comparison/isOutOfRange.h>
21 #include <IPSDKUtil/InstructionSet/assign.h>
22 #include <IPSDKUtil/InstructionSet/Logical/bitwiseXOr.h>
23 #include <IPSDKUtil/Tools/ForceInline.h>
24 
25 namespace ipsdk {
26 namespace simd {
27 namespace detail {
28 
31 
32 template <>
33 struct IsOutOfRangeReg<eInstructionSet::eIS_Sse2, ipInt8>
34 {
35  static IPSDK_FORCEINLINE
36  RegMaskType<eInstructionSet::eIS_Sse2, ipInt8>::Type
37  act(const Sse2Type<ipInt8>::Type& in,
38  const Sse2Type<ipInt8>::Type& regMin,
39  const Sse2Type<ipInt8>::Type& regMax)
40  {
41  return _mm_or_si128(_mm_cmplt_epi8(in, regMin),
42  _mm_cmpgt_epi8(in, regMax));
43  }
44 
45  static IPSDK_FORCEINLINE
46  void
47  act(const Sse2Type<ipInt8>::Type& in,
48  const Sse2Type<ipInt8>::Type& regMin,
49  const Sse2Type<ipInt8>::Type& regMax,
50  RegMaskType<eInstructionSet::eIS_Sse2, ipInt8>::Type& out)
51  {
52  out = _mm_or_si128(_mm_cmplt_epi8(in, regMin),
53  _mm_cmpgt_epi8(in, regMax));
54  }
55 };
56 
57 template <>
58 struct IsOutOfRangeReg<eInstructionSet::eIS_Sse2, ipUInt8>
59 {
60  static IPSDK_FORCEINLINE
61  RegMaskType<eInstructionSet::eIS_Sse2, ipUInt8>::Type
62  act(const Sse2Type<ipUInt8>::Type& in,
63  const Sse2Type<ipUInt8>::Type& regMin,
64  const Sse2Type<ipUInt8>::Type& regMax)
65  {
66  const Sse2Type<ipUInt8>::Type& c =
67  AssignReg<eInstructionSet::eIS_Sse2, ipUInt8>::act(0x80);
68  const Sse2Type<ipUInt8>::Type& inCvt =
69  BitwiseXOrReg<eInstructionSet::eIS_Sse2, ipUInt8>::act(in, c);
70  const Sse2Type<ipUInt8>::Type& minCvt =
71  BitwiseXOrReg<eInstructionSet::eIS_Sse2, ipUInt8>::act(regMin, c);
72  const Sse2Type<ipUInt8>::Type& maxCvt =
73  BitwiseXOrReg<eInstructionSet::eIS_Sse2, ipUInt8>::act(regMax, c);
74 
75  return _mm_or_si128(_mm_cmplt_epi8(inCvt, minCvt),
76  _mm_cmpgt_epi8(inCvt, maxCvt));
77  }
78 
79  static IPSDK_FORCEINLINE
80  void
81  act(const Sse2Type<ipUInt8>::Type& in,
82  const Sse2Type<ipUInt8>::Type& regMin,
83  const Sse2Type<ipUInt8>::Type& regMax,
84  RegMaskType<eInstructionSet::eIS_Sse2, ipUInt8>::Type& out)
85  {
86  Sse2Type<ipUInt8>::Type c;
87  AssignReg<eInstructionSet::eIS_Sse2, ipUInt8>::act(c, 0x80);
88  Sse2Type<ipUInt8>::Type inCvt;
89  BitwiseXOrReg<eInstructionSet::eIS_Sse2, ipUInt8>::act(in, c, inCvt);
90  Sse2Type<ipUInt8>::Type minCvt;
91  BitwiseXOrReg<eInstructionSet::eIS_Sse2, ipUInt8>::act(regMin, c, minCvt);
92  Sse2Type<ipUInt8>::Type maxCvt;
93  BitwiseXOrReg<eInstructionSet::eIS_Sse2, ipUInt8>::act(regMax, c, maxCvt);
94 
95  out = _mm_or_si128(_mm_cmplt_epi8(inCvt, minCvt),
96  _mm_cmpgt_epi8(inCvt, maxCvt));
97  }
98 };
99 
100 template <>
101 struct IsOutOfRangeReg<eInstructionSet::eIS_Sse2, ipInt16>
102 {
103  static IPSDK_FORCEINLINE
104  RegMaskType<eInstructionSet::eIS_Sse2, ipInt16>::Type
105  act(const Sse2Type<ipInt16>::Type& in,
106  const Sse2Type<ipInt16>::Type& regMin,
107  const Sse2Type<ipInt16>::Type& regMax)
108  {
109  return _mm_or_si128(_mm_cmplt_epi16(in, regMin),
110  _mm_cmpgt_epi16(in, regMax));
111  }
112 
113  static IPSDK_FORCEINLINE
114  void
115  act(const Sse2Type<ipInt16>::Type& in,
116  const Sse2Type<ipInt16>::Type& regMin,
117  const Sse2Type<ipInt16>::Type& regMax,
118  RegMaskType<eInstructionSet::eIS_Sse2, ipInt16>::Type& out)
119  {
120  out = _mm_or_si128(_mm_cmplt_epi16(in, regMin),
121  _mm_cmpgt_epi16(in, regMax));
122  }
123 };
124 
125 template <>
126 struct IsOutOfRangeReg<eInstructionSet::eIS_Sse2, ipUInt16>
127 {
128  static IPSDK_FORCEINLINE
129  RegMaskType<eInstructionSet::eIS_Sse2, ipUInt16>::Type
130  act(const Sse2Type<ipUInt16>::Type& in,
131  const Sse2Type<ipUInt16>::Type& regMin,
132  const Sse2Type<ipUInt16>::Type& regMax)
133  {
134  const Sse2Type<ipUInt16>::Type& c =
135  AssignReg<eInstructionSet::eIS_Sse2, ipUInt16>::act(0x8000);
136  const Sse2Type<ipUInt16>::Type& inCvt =
137  BitwiseXOrReg<eInstructionSet::eIS_Sse2, ipUInt16>::act(in, c);
138  const Sse2Type<ipUInt16>::Type& minCvt =
139  BitwiseXOrReg<eInstructionSet::eIS_Sse2, ipUInt16>::act(regMin, c);
140  const Sse2Type<ipUInt16>::Type& maxCvt =
141  BitwiseXOrReg<eInstructionSet::eIS_Sse2, ipUInt16>::act(regMax, c);
142 
143  return _mm_or_si128(_mm_cmplt_epi16(inCvt, minCvt),
144  _mm_cmpgt_epi16(inCvt, maxCvt));
145  }
146 
147  static IPSDK_FORCEINLINE
148  void
149  act(const Sse2Type<ipUInt16>::Type& in,
150  const Sse2Type<ipUInt16>::Type& regMin,
151  const Sse2Type<ipUInt16>::Type& regMax,
152  RegMaskType<eInstructionSet::eIS_Sse2, ipUInt16>::Type& out)
153  {
154  Sse2Type<ipUInt16>::Type c;
155  AssignReg<eInstructionSet::eIS_Sse2, ipUInt16>::act(c, 0x8000);
156  Sse2Type<ipUInt16>::Type inCvt;
157  BitwiseXOrReg<eInstructionSet::eIS_Sse2, ipUInt16>::act(in, c, inCvt);
158  Sse2Type<ipUInt16>::Type minCvt;
159  BitwiseXOrReg<eInstructionSet::eIS_Sse2, ipUInt16>::act(regMin, c, minCvt);
160  Sse2Type<ipUInt16>::Type maxCvt;
161  BitwiseXOrReg<eInstructionSet::eIS_Sse2, ipUInt16>::act(regMax, c, maxCvt);
162 
163  out = _mm_or_si128(_mm_cmplt_epi16(inCvt, minCvt),
164  _mm_cmpgt_epi16(inCvt, maxCvt));
165  }
166 };
167 
168 template <>
169 struct IsOutOfRangeReg<eInstructionSet::eIS_Sse2, ipInt32>
170 {
171  static IPSDK_FORCEINLINE
172  RegMaskType<eInstructionSet::eIS_Sse2, ipInt32>::Type
173  act(const Sse2Type<ipInt32>::Type& in,
174  const Sse2Type<ipInt32>::Type& regMin,
175  const Sse2Type<ipInt32>::Type& regMax)
176  {
177  return _mm_or_si128(_mm_cmplt_epi32(in, regMin),
178  _mm_cmpgt_epi32(in, regMax));
179  }
180 
181  static IPSDK_FORCEINLINE
182  void
183  act(const Sse2Type<ipInt32>::Type& in,
184  const Sse2Type<ipInt32>::Type& regMin,
185  const Sse2Type<ipInt32>::Type& regMax,
186  RegMaskType<eInstructionSet::eIS_Sse2, ipInt32>::Type& out)
187  {
188  out = _mm_or_si128(_mm_cmplt_epi32(in, regMin),
189  _mm_cmpgt_epi32(in, regMax));
190  }
191 };
192 
193 template <>
194 struct IsOutOfRangeReg<eInstructionSet::eIS_Sse2, ipUInt32>
195 {
196  static IPSDK_FORCEINLINE
197  RegMaskType<eInstructionSet::eIS_Sse2, ipUInt32>::Type
198  act(const Sse2Type<ipUInt32>::Type& in,
199  const Sse2Type<ipUInt32>::Type& regMin,
200  const Sse2Type<ipUInt32>::Type& regMax)
201  {
202  const Sse2Type<ipUInt32>::Type& c =
203  AssignReg<eInstructionSet::eIS_Sse2, ipUInt32>::act(0x80000000);
204  const Sse2Type<ipUInt32>::Type& inCvt =
205  BitwiseXOrReg<eInstructionSet::eIS_Sse2, ipUInt32>::act(in, c);
206  const Sse2Type<ipUInt32>::Type& minCvt =
207  BitwiseXOrReg<eInstructionSet::eIS_Sse2, ipUInt32>::act(regMin, c);
208  const Sse2Type<ipUInt32>::Type& maxCvt =
209  BitwiseXOrReg<eInstructionSet::eIS_Sse2, ipUInt32>::act(regMax, c);
210 
211  return _mm_or_si128(_mm_cmplt_epi32(inCvt, minCvt),
212  _mm_cmpgt_epi32(inCvt, maxCvt));
213  }
214 
215  static IPSDK_FORCEINLINE
216  void
217  act(const Sse2Type<ipUInt32>::Type& in,
218  const Sse2Type<ipUInt32>::Type& regMin,
219  const Sse2Type<ipUInt32>::Type& regMax,
220  RegMaskType<eInstructionSet::eIS_Sse2, ipUInt32>::Type& out)
221  {
222  Sse2Type<ipUInt32>::Type c;
223  AssignReg<eInstructionSet::eIS_Sse2, ipUInt32>::act(c, 0x80000000);
224  Sse2Type<ipUInt32>::Type inCvt;
225  BitwiseXOrReg<eInstructionSet::eIS_Sse2, ipUInt32>::act(in, c, inCvt);
226  Sse2Type<ipUInt32>::Type minCvt;
227  BitwiseXOrReg<eInstructionSet::eIS_Sse2, ipUInt32>::act(regMin, c, minCvt);
228  Sse2Type<ipUInt32>::Type maxCvt;
229  BitwiseXOrReg<eInstructionSet::eIS_Sse2, ipUInt32>::act(regMax, c, maxCvt);
230 
231  out = _mm_or_si128(_mm_cmplt_epi32(inCvt, minCvt),
232  _mm_cmpgt_epi32(inCvt, maxCvt));
233  }
234 };
235 
236 
237 template <>
238 struct IsOutOfRangeReg<eInstructionSet::eIS_Sse2, ipReal32>
239 {
240  static IPSDK_FORCEINLINE
241  RegMaskType<eInstructionSet::eIS_Sse2, ipReal32>::Type
242  act(const Sse2Type<ipReal32>::Type& in,
243  const Sse2Type<ipReal32>::Type& regMin,
244  const Sse2Type<ipReal32>::Type& regMax)
245  {
246  return _mm_or_ps(_mm_cmplt_ps(in, regMin),
247  _mm_cmpgt_ps(in, regMax));
248  }
249 
250  static IPSDK_FORCEINLINE
251  void
252  act(const Sse2Type<ipReal32>::Type& in,
253  const Sse2Type<ipReal32>::Type& regMin,
254  const Sse2Type<ipReal32>::Type& regMax,
255  RegMaskType<eInstructionSet::eIS_Sse2, ipReal32>::Type& out)
256  {
257  out = _mm_or_ps(_mm_cmplt_ps(in, regMin),
258  _mm_cmpgt_ps(in, regMax));
259  }
260 };
261 
262 template <>
263 struct IsOutOfRangeReg<eInstructionSet::eIS_Sse2, ipReal64>
264 {
265  static IPSDK_FORCEINLINE
266  RegMaskType<eInstructionSet::eIS_Sse2, ipReal64>::Type
267  act(const Sse2Type<ipReal64>::Type& in,
268  const Sse2Type<ipReal64>::Type& regMin,
269  const Sse2Type<ipReal64>::Type& regMax)
270  {
271  return _mm_or_pd(_mm_cmplt_pd(in, regMin),
272  _mm_cmpgt_pd(in, regMax));
273  }
274 
275  static IPSDK_FORCEINLINE
276  void
277  act(const Sse2Type<ipReal64>::Type& in,
278  const Sse2Type<ipReal64>::Type& regMin,
279  const Sse2Type<ipReal64>::Type& regMax,
280  RegMaskType<eInstructionSet::eIS_Sse2, ipReal64>::Type& out)
281  {
282  out = _mm_or_pd(_mm_cmplt_pd(in, regMin),
283  _mm_cmpgt_pd(in, regMax));
284  }
285 };
286 
289 
290 } // end of namespace detail
291 } // end of namespace simd
292 } // end of namespace ipsdk
293 
294 #endif // __IPSDKUTIL_INSTRUCTIONSET_COMPARISON_DETAIL_SSE2_ISOUTOFRANGEREG_H__
ipsdk::simd::detail::IsOutOfRangeReg
Definition: IsOutOfRangeReg.h:30
ipsdk::ipInt8
int8_t ipInt8
Base types definition.
Definition: BaseTypes.h:48
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::ipInt32
int32_t ipInt32
Base types definition.
Definition: BaseTypes.h:52
ipsdk::ipInt16
int16_t ipInt16
Base types definition.
Definition: BaseTypes.h:50
ipsdk::ipUInt8
uint8_t ipUInt8
Base types definition.
Definition: BaseTypes.h:49
ipsdk::eInstructionSet
eInstructionSet
Enumerate for processor instruction set description.
Definition: InstructionSetTypes.h:31
assign.h
function assigning a given value of type T to a given Pack<instructionSet, T>
Sse2Types.h
Predefined types for Sse2 instruction set management.
ipsdk::simd::detail::BitwiseXOrReg
Definition: BitwiseXOrReg.h:30
IPSDKUtilExports.h
Definition of import/export macro for library.
ipsdk::simd::RegMaskType
Definition: RegMaskType.h:29
ipsdk::eIS_Sse2
Streaming SIMD Extensions 2.
Definition: InstructionSetTypes.h:36
ipsdk::Sse2Type
structure used to retrieve SSE2 type associated to a base type
Definition: Sse2Types.h:32
ipsdk::ipUInt16
uint16_t ipUInt16
Base types definition.
Definition: BaseTypes.h:51
ipsdk::simd::detail::AssignReg
Definition: AssignRegDecl.h:31
ipsdk::ipReal32
float ipReal32
Base types definition.
Definition: BaseTypes.h:56
ipsdk::ipUInt32
uint32_t ipUInt32
Base types definition.
Definition: BaseTypes.h:53
Generated on Tue Apr 15 2025 16:20:41 for IPSDK by   doxygen 1.8.14