IPSDK  4_1_0_2
IPSDK : Image Processing Software Development Kit
TernaryPackOp.h
Go to the documentation of this file.
1 // TernaryPackOp.h:
3 // ------------
4 //
14 
15 #ifndef __IPSDKUTIL_INSTRUCTIONSET_DETAIL_FMA3_TERNARYPACKOP_H__
16 #define __IPSDKUTIL_INSTRUCTIONSET_DETAIL_FMA3_TERNARYPACKOP_H__
17 
18 #include <IPSDKUtil/IPSDKUtilExports.h>
19 #include <IPSDKUtil/InstructionSet/InstructionSetTypes.h>
20 #include <IPSDKUtil/Tools/ForceInline.h>
21 
22 namespace ipsdk {
23 namespace simd {
24 namespace detail {
25 
28 
29 // template structure associated to unload function
30 template <typename T,
31  class OpClass>
32 struct TernaryPackOp<eInstructionSet::eIS_Fma3, T, OpClass,
33  typename boost::enable_if_c<sizeof(T) == 1>::type>
34 {
35  static IPSDK_FORCEINLINE
36  BasePack<ePackType::ePT_Avx, T>
37  act(const BasePack<ePackType::ePT_Avx, T>& in1,
38  const BasePack<ePackType::ePT_Avx, T>& in2,
39  const BasePack<ePackType::ePT_Avx, T>& in3)
40  {
41  BasePack<ePackType::ePT_Avx, T> out;
42  out._val[0] = OpClass::act(in1._val[0], in2._val[0], in3._val[0]);
43  return out;
44  }
45 
46  static IPSDK_FORCEINLINE
47  void
48  act(const BasePack<ePackType::ePT_Avx, T>& in1,
49  const BasePack<ePackType::ePT_Avx, T>& in2,
50  const BasePack<ePackType::ePT_Avx, T>& in3,
51  BasePack<ePackType::ePT_Avx, T>& out)
52  {
53  OpClass::act(in1._val[0], in2._val[0], in3._val[0], out._val[0]);
54  }
55 };
56 
57 template <typename T, class OpClass>
58 struct TernaryPackOp<eInstructionSet::eIS_Fma3, T, OpClass,
59  typename boost::enable_if_c<sizeof(T) == 2>::type>
60 {
61  static IPSDK_FORCEINLINE
62  BasePack<ePackType::ePT_Avx, T>
63  act(const BasePack<ePackType::ePT_Avx, T>& in1,
64  const BasePack<ePackType::ePT_Avx, T>& in2,
65  const BasePack<ePackType::ePT_Avx, T>& in3)
66  {
67  BasePack<ePackType::ePT_Avx, T> out;
68  out._val[0] = OpClass::act(in1._val[0], in2._val[0], in3._val[0]);
69  out._val[1] = OpClass::act(in1._val[1], in2._val[1], in3._val[1]);
70  return out;
71  }
72 
73  static IPSDK_FORCEINLINE
74  void
75  act(const BasePack<ePackType::ePT_Avx, T>& in1,
76  const BasePack<ePackType::ePT_Avx, T>& in2,
77  const BasePack<ePackType::ePT_Avx, T>& in3,
78  BasePack<ePackType::ePT_Avx, T>& out)
79  {
80  OpClass::act(in1._val[0], in2._val[0], in3._val[0], out._val[0]);
81  OpClass::act(in1._val[1], in2._val[1], in3._val[1], out._val[1]);
82  }
83 };
84 
85 template <typename T, class OpClass>
86 struct TernaryPackOp<eInstructionSet::eIS_Fma3, T, OpClass,
87  typename boost::enable_if_c<sizeof(T) == 4>::type>
88 {
89  static IPSDK_FORCEINLINE
90  BasePack<ePackType::ePT_Avx, T>
91  act(const BasePack<ePackType::ePT_Avx, T>& in1,
92  const BasePack<ePackType::ePT_Avx, T>& in2,
93  const BasePack<ePackType::ePT_Avx, T>& in3)
94  {
95  BasePack<ePackType::ePT_Avx, T> out;
96  out._val[0] = OpClass::act(in1._val[0], in2._val[0], in3._val[0]);
97  out._val[1] = OpClass::act(in1._val[1], in2._val[1], in3._val[1]);
98  out._val[2] = OpClass::act(in1._val[2], in2._val[2], in3._val[2]);
99  out._val[3] = OpClass::act(in1._val[3], in2._val[3], in3._val[3]);
100  return out;
101  }
102 
103  static IPSDK_FORCEINLINE
104  void
105  act(const BasePack<ePackType::ePT_Avx, T>& in1,
106  const BasePack<ePackType::ePT_Avx, T>& in2,
107  const BasePack<ePackType::ePT_Avx, T>& in3,
108  BasePack<ePackType::ePT_Avx, T>& out)
109  {
110  OpClass::act(in1._val[0], in2._val[0], in3._val[0], out._val[0]);
111  OpClass::act(in1._val[1], in2._val[1], in3._val[1], out._val[1]);
112  OpClass::act(in1._val[2], in2._val[2], in3._val[2], out._val[2]);
113  OpClass::act(in1._val[3], in2._val[3], in3._val[3], out._val[3]);
114  }
115 };
116 
117 template <typename T, class OpClass>
118 struct TernaryPackOp<eInstructionSet::eIS_Fma3, T, OpClass,
119  typename boost::enable_if_c<sizeof(T) == 8>::type>
120 {
121  static IPSDK_FORCEINLINE
122  BasePack<ePackType::ePT_Avx, T>
123  act(const BasePack<ePackType::ePT_Avx, T>& in1,
124  const BasePack<ePackType::ePT_Avx, T>& in2,
125  const BasePack<ePackType::ePT_Avx, T>& in3)
126  {
127  BasePack<ePackType::ePT_Avx, T> out;
128  out._val[0] = OpClass::act(in1._val[0], in2._val[0], in3._val[0]);
129  out._val[1] = OpClass::act(in1._val[1], in2._val[1], in3._val[1]);
130  out._val[2] = OpClass::act(in1._val[2], in2._val[2], in3._val[2]);
131  out._val[3] = OpClass::act(in1._val[3], in2._val[3], in3._val[3]);
132  out._val[4] = OpClass::act(in1._val[4], in2._val[4], in3._val[4]);
133  out._val[5] = OpClass::act(in1._val[5], in2._val[5], in3._val[5]);
134  out._val[6] = OpClass::act(in1._val[6], in2._val[6], in3._val[6]);
135  out._val[7] = OpClass::act(in1._val[7], in2._val[7], in3._val[7]);
136  return out;
137  }
138 
139  static IPSDK_FORCEINLINE
140  void
141  act(const BasePack<ePackType::ePT_Avx, T>& in1,
142  const BasePack<ePackType::ePT_Avx, T>& in2,
143  const BasePack<ePackType::ePT_Avx, T>& in3,
144  BasePack<ePackType::ePT_Avx, T>& out)
145  {
146  OpClass::act(in1._val[0], in2._val[0], in3._val[0], out._val[0]);
147  OpClass::act(in1._val[1], in2._val[1], in3._val[1], out._val[1]);
148  OpClass::act(in1._val[2], in2._val[2], in3._val[2], out._val[2]);
149  OpClass::act(in1._val[3], in2._val[3], in3._val[3], out._val[3]);
150  OpClass::act(in1._val[4], in2._val[4], in3._val[4], out._val[4]);
151  OpClass::act(in1._val[5], in2._val[5], in3._val[5], out._val[5]);
152  OpClass::act(in1._val[6], in2._val[6], in3._val[6], out._val[6]);
153  OpClass::act(in1._val[7], in2._val[7], in3._val[7], out._val[7]);
154  }
155 };
156 
159 
160 } // end of namespace detail
161 } // end of namespace simd
162 } // end of namespace ipsdk
163 
164 #endif // __IPSDKUTIL_INSTRUCTIONSET_DETAIL_FMA3_TERNARYPACKOP_H__
ipsdk::eIS_Fma3
Fused multiply–add.
Definition: InstructionSetTypes.h:46
ForceInline.h
Defines the IPSDK_FORCEINLINE.
ipsdk
Main namespace for IPSDK library.
Definition: AlgorithmFunctionEfficiency.h:22
ipsdk::eInstructionSet
eInstructionSet
Enumerate for processor instruction set description.
Definition: InstructionSetTypes.h:31
ipsdk::simd::detail::TernaryPackOp
Definition: TernaryPackOp.h:31
InstructionSetTypes.h
Predefined types associated to instruction set management.
IPSDKUtilExports.h
Definition of import/export macro for library.
Generated on Tue Apr 15 2025 16:20:43 for IPSDK by   doxygen 1.8.14