IPSDK  4_1_0_2
IPSDK : Image Processing Software Development Kit
ToBoolPack.h
Go to the documentation of this file.
1 // ToBoolPack.h:
3 // -------------------
4 //
14 
15 #ifndef __IPSDKUTIL_INSTRUCTIONSET_LOGICAL_DETAIL_AVX512_TOBOOLPACK_H__
16 #define __IPSDKUTIL_INSTRUCTIONSET_LOGICAL_DETAIL_AVX512_TOBOOLPACK_H__
17 
18 #include <IPSDKUtil/IPSDKUtilExports.h>
19 #include <IPSDKUtil/InstructionSet/AVX512/Avx512Types.h>
20 #include <IPSDKUtil/InstructionSet/detail/AVX512/BasePack.h>
21 #include <IPSDKUtil/InstructionSet/Logical/detail/ToBoolPack.h>
22 #include <IPSDKUtil/Tools/ForceInline.h>
23 
24 namespace ipsdk {
25 namespace simd {
26 namespace detail {
27 
30 
31 template <>
32 struct ToBoolPack<eInstructionSet::eIS_Avx512, ipUInt8>
33 {
34  static IPSDK_FORCEINLINE
35  BasePack<ePackType::ePT_Avx512, ipUInt8>
36  act(const BasePack<ePackType::ePT_Avx512, ipUInt8>& in)
37  {
38  BasePack<ePackType::ePT_Avx512, ipUInt8> out;
39  act(in, out);
40  return out;
41  }
42 
43  static IPSDK_FORCEINLINE
44  void
45  act(const BasePack<ePackType::ePT_Avx512, ipUInt8>& in,
46  BasePack<ePackType::ePT_Avx512, ipUInt8>& out)
47  {
48  BasePack<ePackType::ePT_Avx512, ipUInt8> zero, one;
49  BaseMaskPack<ePackType::ePT_Avx512, ipUInt8> nez;
50  assign<eInstructionSet::eIS_Avx512>(zero, 0);
51  assign<eInstructionSet::eIS_Avx512>(one, 1);
52  isNotEqual<eInstructionSet::eIS_Avx512>(in, zero, nez);
53  cast<eInstructionSet::eIS_Avx512>(nez, out);
54  bitwiseAnd<eInstructionSet::eIS_Avx512>(out, one, out);
55  }
56 
57  static IPSDK_FORCEINLINE
58  void
59  act(const BaseMaskPack<ePackType::ePT_Avx512, ipUInt8>& inMask,
60  BasePack<ePackType::ePT_Avx512, ipUInt8>& out)
61  {
62  BasePack<ePackType::ePT_Avx512, ipUInt8> in;
63  cast<eInstructionSet::eIS_Avx512>(inMask, in);
64  act(in, out);
65  }
66 };
67 
68 template <>
69 struct ToBoolPack<eInstructionSet::eIS_Avx512, ipInt8>
70 {
71  static IPSDK_FORCEINLINE
72  BasePack<ePackType::ePT_Avx512, ipUInt8>
73  act(const BasePack<ePackType::ePT_Avx512, ipInt8>& in)
74  {
75  BasePack<ePackType::ePT_Avx512, ipUInt8> out;
76  act(in, out);
77  return out;
78  }
79 
80  static IPSDK_FORCEINLINE
81  void
82  act(const BasePack<ePackType::ePT_Avx512, ipInt8>& in,
83  BasePack<ePackType::ePT_Avx512, ipUInt8>& out)
84  {
85  BasePack<ePackType::ePT_Avx512, ipInt8> zero, one;
86  BaseMaskPack<ePackType::ePT_Avx512, ipInt8> nez;
87  assign<eInstructionSet::eIS_Avx512>(zero, 0);
88  assign<eInstructionSet::eIS_Avx512>(one, 1);
89  BasePack<ePackType::ePT_Avx512, ipInt8> packInt8;
90  isNotEqual<eInstructionSet::eIS_Avx512>(in, zero, nez);
91  cast<eInstructionSet::eIS_Avx512>(nez, packInt8);
92  bitwiseAnd<eInstructionSet::eIS_Avx512>(packInt8, one, packInt8);
93 
94  cast<eInstructionSet::eIS_Avx512>(packInt8, out);
95  }
96 
97  static IPSDK_FORCEINLINE
98  void
99  act(const BaseMaskPack<ePackType::ePT_Avx512, ipInt8>& inMask,
100  BasePack<ePackType::ePT_Avx512, ipUInt8>& out)
101  {
102  BasePack<ePackType::ePT_Avx512, ipInt8> in;
103  cast<eInstructionSet::eIS_Avx512>(inMask, in);
104  act(in, out);
105  }
106 };
107 
108 template <typename T>
109 struct ToBoolPack<eInstructionSet::eIS_Avx512, T,
110  typename boost::enable_if_c<
111  boost::is_integral<T>::value && sizeof(T) != 1>::type>
112 {
113  static IPSDK_FORCEINLINE
114  BasePack<ePackType::ePT_Avx512, ipUInt8>
115  act(const BasePack<ePackType::ePT_Avx512, T>& in)
116  {
117  BasePack<ePackType::ePT_Avx512, ipUInt8> out;
118  act(in, out);
119  }
120 
121  static IPSDK_FORCEINLINE
122  void
123  act(const BasePack<ePackType::ePT_Avx512, T>& in,
124  BasePack<ePackType::ePT_Avx512, ipUInt8>& out)
125  {
126  BasePack<ePackType::ePT_Avx512, T> zero, one;
127  BaseMaskPack<ePackType::ePT_Avx512, T> nez;
128  assign<eInstructionSet::eIS_Avx512>(zero, 0);
129  assign<eInstructionSet::eIS_Avx512>(one, 1);
130 
131  BasePack<ePackType::ePT_Avx512, T> packT;
132  isNotEqual<eInstructionSet::eIS_Avx512>(in, zero, nez);
133  cast<eInstructionSet::eIS_Avx512>(nez, packT);
134  bitwiseAnd<eInstructionSet::eIS_Avx512>(packT, one, packT);
135 
136  cast<eInstructionSet::eIS_Avx512>(packT, out);
137  }
138 
139  static IPSDK_FORCEINLINE
140  void
141  act(const BaseMaskPack<ePackType::ePT_Avx512, T>& inMask,
142  BasePack<ePackType::ePT_Avx512, ipUInt8>& out)
143  {
144  BasePack<ePackType::ePT_Avx512, T> in;
145  cast<eInstructionSet::eIS_Avx512>(inMask, in);
146  act(in, out);
147  }
148 };
149 
150 template <>
151 struct ToBoolPack<eInstructionSet::eIS_Avx512, ipReal32>
152 {
153  static IPSDK_FORCEINLINE
154  BasePack<ePackType::ePT_Avx512, ipUInt8>
155  act(const BasePack<ePackType::ePT_Avx512, ipReal32>& in)
156  {
157  BasePack<ePackType::ePT_Avx512, ipUInt8> out;
158  act(in, out);
159  }
160 
161  static IPSDK_FORCEINLINE
162  void
163  act(const BasePack<ePackType::ePT_Avx512, ipReal32>& in,
164  BasePack<ePackType::ePT_Avx512, ipUInt8>& out)
165  {
166  BasePack<ePackType::ePT_Avx512, ipInt32> inCvt;
167  inCvt._val[0] = _mm512_castps_si512(in._val[0]);
168  inCvt._val[1] = _mm512_castps_si512(in._val[1]);
169  inCvt._val[2] = _mm512_castps_si512(in._val[2]);
170  inCvt._val[3] = _mm512_castps_si512(in._val[3]);
171 
172  ToBoolPack<eInstructionSet::eIS_Avx512, ipInt32>::act(inCvt, out);
173  }
174 
175  static IPSDK_FORCEINLINE
176  void
177  act(const BaseMaskPack<ePackType::ePT_Avx512, ipReal32>& inMask,
178  BasePack<ePackType::ePT_Avx512, ipUInt8>& out)
179  {
180  BasePack<ePackType::ePT_Avx512, ipReal32> in;
181  cast<eInstructionSet::eIS_Avx512>(inMask, in);
182  act(in, out);
183  }
184 };
185 
188 
189 } // end of namespace detail
190 } // end of namespace simd
191 } // end of namespace ipsdk
192 
193 #endif // __IPSDKUTIL_INSTRUCTIONSET_LOGICAL_DETAIL_AVX2_TOBOOLPACK_H__
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::simd::BaseMaskPack
structure containing set of masks for vectorized operations
Definition: BaseMaskPackDecl.h:29
ipsdk::eIS_Avx512
(including fundation and byte and word instructions)
Definition: InstructionSetTypes.h:51
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
Avx512Types.h
Predefined types for Avx512 instruction set management.
ipsdk::simd::detail::ToBoolPack
Definition: ToBoolPack.h:30
IPSDKUtilExports.h
Definition of import/export macro for library.
ipsdk::ipReal32
float ipReal32
Base types definition.
Definition: BaseTypes.h:56
ipsdk::simd::BasePack
structure containing intrinsic registers used to store vectorized data
Definition: BasePackDecl.h:29
Generated on Tue Apr 15 2025 16:20:45 for IPSDK by   doxygen 1.8.14