IPSDK  4_1_0_2
IPSDK : Image Processing Software Development Kit
DivPack.h
Go to the documentation of this file.
1 // DivPack.h:
3 // -----------------
4 //
14 
15 #ifndef __IPSDKUTIL_INSTRUCTIONSET_ARITHMETIC_DETAIL_AVX2_DIVPACK_H__
16 #define __IPSDKUTIL_INSTRUCTIONSET_ARITHMETIC_DETAIL_AVX2_DIVPACK_H__
17 
18 #include <IPSDKUtil/IPSDKUtilExports.h>
19 #include <IPSDKUtil/InstructionSet/cast.h>
20 #include <IPSDKUtil/InstructionSet/InstructionSetTypes.h>
21 #include <IPSDKUtil/InstructionSet/PromotedType.h>
22 #include <IPSDKUtil/InstructionSet/BasePack.h>
23 #include <IPSDKUtil/InstructionSet/Arithmetic/detail/DivPack.h>
24 #include <IPSDKUtil/InstructionSet/Arithmetic/detail/DivReg.h>
25 #include <IPSDKUtil/InstructionSet/detail/BinaryPackOp.h>
26 #include <IPSDKUtil/Tools/ForceInline.h>
27 
28 #include <boost/type_traits/is_same.hpp>
29 #include <boost/utility/enable_if.hpp>
30 
31 namespace ipsdk {
32 namespace simd {
33 namespace detail {
34 
37 
40 template <>
41 struct DivPack<eInstructionSet::eIS_Avx2, ipReal32, ipReal32>
42 {
43  static
44  IPSDK_FORCEINLINE
45  BasePack<ePackType::ePT_Avx, ipReal32>
46  act(const BasePack<ePackType::ePT_Avx, ipReal32>& in1,
47  const BasePack<ePackType::ePT_Avx, ipReal32>& in2)
48  {
49  return BinaryPackOp<eInstructionSet::eIS_Avx, ipReal32,
50  DivReg<eInstructionSet::eIS_Avx, ipReal32> >::act(in1, in2);
51  }
52 
53  static
54  IPSDK_FORCEINLINE
55  void
56  act(const BasePack<ePackType::ePT_Avx, ipReal32>& in1,
57  const BasePack<ePackType::ePT_Avx, ipReal32>& in2,
58  BasePack<ePackType::ePT_Avx, ipReal32>& out)
59  {
60  BinaryPackOp<eInstructionSet::eIS_Avx, ipReal32,
61  DivReg<eInstructionSet::eIS_Avx, ipReal32> >::act(in1, in2, out);
62  }
63 };
64 
67 template <>
68 struct DivPack<eInstructionSet::eIS_Avx2, ipReal64, ipReal64>
69 {
70  static
71  IPSDK_FORCEINLINE
72  BasePack<ePackType::ePT_Avx, ipReal64>
73  act(const BasePack<ePackType::ePT_Avx, ipReal64>& in1,
74  const BasePack<ePackType::ePT_Avx, ipReal64>& in2)
75  {
76  return BinaryPackOp<eInstructionSet::eIS_Avx, ipReal64,
77  DivReg<eInstructionSet::eIS_Avx, ipReal64> >::act(in1, in2);
78  }
79 
80  static
81  IPSDK_FORCEINLINE
82  void
83  act(const BasePack<ePackType::ePT_Avx, ipReal64>& in1,
84  const BasePack<ePackType::ePT_Avx, ipReal64>& in2,
85  BasePack<ePackType::ePT_Avx, ipReal64>& out)
86  {
87  BinaryPackOp<eInstructionSet::eIS_Avx, ipReal64,
88  DivReg<eInstructionSet::eIS_Avx, ipReal64> >::act(in1, in2, out);
89  }
90 };
91 
94 template <typename TIn1, typename TIn2>
95 struct DivPack<eInstructionSet::eIS_Avx2, TIn1, TIn2,
96  typename boost::enable_if_c<!boost::is_same<TIn1, TIn2>::value
97  && boost::is_same<TIn1, typename PromotedType<TIn1, TIn2>::Type>::value
98  >::type>
99 {
100  static
101  IPSDK_FORCEINLINE
102  BasePack<ePackType::ePT_Avx, TIn1>
103  act(const BasePack<ePackType::ePT_Avx, TIn1>& in1,
104  const BasePack<ePackType::ePT_Avx, TIn2>& in2)
105  {
106  const BasePack<ePackType::ePT_Avx, TIn1> in2Promoted =
107  cast<eInstructionSet::eIS_Avx2, TIn2, TIn1>(in2);
108  return DivPack<eInstructionSet::eIS_Avx2, TIn1, TIn1>::act(in1, in2Promoted);
109  }
110 
111  static
112  IPSDK_FORCEINLINE
113  void
114  act(const BasePack<ePackType::ePT_Avx, TIn1>& in1,
115  const BasePack<ePackType::ePT_Avx, TIn2>& in2,
116  BasePack<ePackType::ePT_Avx, TIn1>& out)
117  {
118  BasePack<ePackType::ePT_Avx, TIn1> in2Promoted;
119  cast<eInstructionSet::eIS_Avx2, TIn2, TIn1>(in2, in2Promoted);
120  DivPack<eInstructionSet::eIS_Avx2, TIn1, TIn1>::act(in1, in2Promoted, out);
121  }
122 };
123 
126 template <typename TIn1, typename TIn2>
127 struct DivPack<eInstructionSet::eIS_Avx2, TIn1, TIn2,
128  typename boost::enable_if_c<!boost::is_same<TIn1, TIn2>::value
129  && !boost::is_same<TIn1, typename PromotedType<TIn1, TIn2>::Type>::value
130  >::type>
131 {
132  static
133  IPSDK_FORCEINLINE
134  BasePack<ePackType::ePT_Avx, TIn2>
135  act(const BasePack<ePackType::ePT_Avx, TIn1>& in1,
136  const BasePack<ePackType::ePT_Avx, TIn2>& in2)
137  {
138  const BasePack<ePackType::ePT_Avx, TIn2> in1Promoted =
139  cast<eInstructionSet::eIS_Avx2, TIn1, TIn2>(in1);
140  return DivPack<eInstructionSet::eIS_Avx2, TIn2, TIn2>::act(
141  in1Promoted, in2);
142  }
143 
144  static
145  IPSDK_FORCEINLINE
146  void
147  act(const BasePack<ePackType::ePT_Avx, TIn1>& in1,
148  const BasePack<ePackType::ePT_Avx, TIn2>& in2,
149  BasePack<ePackType::ePT_Avx, TIn2>& out)
150  {
151  BasePack<ePackType::ePT_Avx, TIn2> in1Promoted;
152  cast<eInstructionSet::eIS_Avx2, TIn1, TIn2>(in1, in1Promoted);
153  DivPack<eInstructionSet::eIS_Avx2, TIn2, TIn2>::act(
154  in1Promoted, in2, out);
155  }
156 };
157 
160 template <typename TIn>
161 struct DivPack<eInstructionSet::eIS_Avx2, TIn, TIn,
162  typename boost::enable_if_c<boost::is_integral<TIn>::value>::type>
163 {
164  static
165  IPSDK_FORCEINLINE
166  BasePack<ePackType::ePT_Avx, TIn>
167  act(const BasePack<ePackType::ePT_Avx, TIn>& in1,
168  const BasePack<ePackType::ePT_Avx, TIn>& in2)
169  {
170  const BasePack<ePackType::ePT_Avx, ipReal32> in1Cast =
171  cast<eInstructionSet::eIS_Avx2, TIn, ipReal32>(in1);
172  const BasePack<ePackType::ePT_Avx, ipReal32> in2Cast =
173  cast<eInstructionSet::eIS_Avx2, TIn, ipReal32>(in2);
174  const BasePack<ePackType::ePT_Avx, ipReal32> outCast =
175  DivPack<eInstructionSet::eIS_Avx2, ipReal32, ipReal32>::act(in1Cast, in2Cast);
176 
177  return cast<eInstructionSet::eIS_Avx2, ipReal32, TIn>(outCast);
178  }
179 
180  static
181  IPSDK_FORCEINLINE
182  void
183  act(const BasePack<ePackType::ePT_Avx, TIn>& in1,
184  const BasePack<ePackType::ePT_Avx, TIn>& in2,
185  BasePack<ePackType::ePT_Avx, TIn>& out)
186  {
187  BasePack<ePackType::ePT_Avx, ipReal32> in1Cast;
188  cast<eInstructionSet::eIS_Avx2>(in1, in1Cast);
189  BasePack<ePackType::ePT_Avx, ipReal32> in2Cast;
190  cast<eInstructionSet::eIS_Avx2>(in2, in2Cast);
191  BasePack<ePackType::ePT_Avx, ipReal32> outCast;
192  DivPack<eInstructionSet::eIS_Avx, ipReal32, ipReal32>::act(in1Cast, in2Cast, outCast);
193  cast<eInstructionSet::eIS_Avx2>(outCast, out);
194  }
195 };
196 
200 
201 } // end of namespace detail
202 } // end of namespace simd
203 } // end of namespace ipsdk
204 
205 #endif // __IPSDKUTIL_INSTRUCTIONSET_ARITHMETIC_DETAIL_AVX2_DIVPACK_H__
ForceInline.h
Defines the IPSDK_FORCEINLINE.
ipsdk
Main namespace for IPSDK library.
Definition: AlgorithmFunctionEfficiency.h:22
ipsdk::simd::detail::DivReg< eInstructionSet::eIS_Avx, ipReal32 >
DivReg template specialization for instruction set AVX and for ipReal32.
Definition: DivReg.h:33
cast.h
cast function; casts a Pack<instructionSet, TIn> to a Pack<instructionSet, TOut>
ipsdk::simd::detail::DivPack
Definition: DivPack.h:40
ipsdk::ipReal64
double ipReal64
Base types definition.
Definition: BaseTypes.h:57
BasePack.h
BasePack class; defines a set of scalars (for instruction set "standard") or registers (for all other...
ipsdk::eInstructionSet
eInstructionSet
Enumerate for processor instruction set description.
Definition: InstructionSetTypes.h:31
ipsdk::eIS_Avx
Advanced Vector Extensions.
Definition: InstructionSetTypes.h:44
PromotedType.h
defines template structures PromotedType<T1, T2> and PromotedType3<T1, T2, T3>; their attribute Type ...
ipsdk::eIS_Avx2
Advanced Vector Extensions 2.
Definition: InstructionSetTypes.h:48
InstructionSetTypes.h
Predefined types associated to instruction set management.
IPSDKUtilExports.h
Definition of import/export macro for library.
ipsdk::simd::detail::DivReg< eInstructionSet::eIS_Avx, ipReal64 >
DivReg template specialization for instruction set AVX and for ipReal64.
Definition: DivReg.h:56
ipsdk::ipReal32
float ipReal32
Base types definition.
Definition: BaseTypes.h:56
ipsdk::simd::detail::BinaryPackOp
Definition: BinaryPackOp.h:31
DivPack.h
structure DivPack<typename TIn1, typename TIn2, typename Enable=void>
ipsdk::simd::BasePack
structure containing intrinsic registers used to store vectorized data
Definition: BasePackDecl.h:29
Generated on Tue Apr 15 2025 16:20:38 for IPSDK by   doxygen 1.8.14