Narayan Kamath | c981c48 | 2012-11-02 10:59:05 +0000 | [diff] [blame] | 1 | /* |
| 2 | Copyright (c) 2011, Intel Corporation. All rights reserved. |
| 3 | |
| 4 | Redistribution and use in source and binary forms, with or without modification, |
| 5 | are permitted provided that the following conditions are met: |
| 6 | |
| 7 | * Redistributions of source code must retain the above copyright notice, this |
| 8 | list of conditions and the following disclaimer. |
| 9 | * Redistributions in binary form must reproduce the above copyright notice, |
| 10 | this list of conditions and the following disclaimer in the documentation |
| 11 | and/or other materials provided with the distribution. |
| 12 | * Neither the name of Intel Corporation nor the names of its contributors may |
| 13 | be used to endorse or promote products derived from this software without |
| 14 | specific prior written permission. |
| 15 | |
| 16 | THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND |
| 17 | ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED |
| 18 | WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE |
| 19 | DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR |
| 20 | ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES |
| 21 | (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; |
| 22 | LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON |
| 23 | ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT |
| 24 | (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS |
| 25 | SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. |
| 26 | |
| 27 | ******************************************************************************** |
| 28 | * Content : Eigen bindings to Intel(R) MKL |
| 29 | * MKL VML support for coefficient-wise unary Eigen expressions like a=b.sin() |
| 30 | ******************************************************************************** |
| 31 | */ |
| 32 | |
| 33 | #ifndef EIGEN_ASSIGN_VML_H |
| 34 | #define EIGEN_ASSIGN_VML_H |
| 35 | |
| 36 | namespace Eigen { |
| 37 | |
| 38 | namespace internal { |
| 39 | |
| 40 | template<typename Op> struct vml_call |
| 41 | { enum { IsSupported = 0 }; }; |
| 42 | |
| 43 | template<typename Dst, typename Src, typename UnaryOp> |
| 44 | class vml_assign_traits |
| 45 | { |
| 46 | private: |
| 47 | enum { |
| 48 | DstHasDirectAccess = Dst::Flags & DirectAccessBit, |
| 49 | SrcHasDirectAccess = Src::Flags & DirectAccessBit, |
| 50 | |
| 51 | StorageOrdersAgree = (int(Dst::IsRowMajor) == int(Src::IsRowMajor)), |
| 52 | InnerSize = int(Dst::IsVectorAtCompileTime) ? int(Dst::SizeAtCompileTime) |
| 53 | : int(Dst::Flags)&RowMajorBit ? int(Dst::ColsAtCompileTime) |
| 54 | : int(Dst::RowsAtCompileTime), |
| 55 | InnerMaxSize = int(Dst::IsVectorAtCompileTime) ? int(Dst::MaxSizeAtCompileTime) |
| 56 | : int(Dst::Flags)&RowMajorBit ? int(Dst::MaxColsAtCompileTime) |
| 57 | : int(Dst::MaxRowsAtCompileTime), |
| 58 | MaxSizeAtCompileTime = Dst::SizeAtCompileTime, |
| 59 | |
| 60 | MightEnableVml = vml_call<UnaryOp>::IsSupported && StorageOrdersAgree && DstHasDirectAccess && SrcHasDirectAccess |
| 61 | && Src::InnerStrideAtCompileTime==1 && Dst::InnerStrideAtCompileTime==1, |
| 62 | MightLinearize = MightEnableVml && (int(Dst::Flags) & int(Src::Flags) & LinearAccessBit), |
| 63 | VmlSize = MightLinearize ? MaxSizeAtCompileTime : InnerMaxSize, |
| 64 | LargeEnough = VmlSize==Dynamic || VmlSize>=EIGEN_MKL_VML_THRESHOLD, |
| 65 | MayEnableVml = MightEnableVml && LargeEnough, |
| 66 | MayLinearize = MayEnableVml && MightLinearize |
| 67 | }; |
| 68 | public: |
| 69 | enum { |
| 70 | Traversal = MayLinearize ? LinearVectorizedTraversal |
| 71 | : MayEnableVml ? InnerVectorizedTraversal |
| 72 | : DefaultTraversal |
| 73 | }; |
| 74 | }; |
| 75 | |
| 76 | template<typename Derived1, typename Derived2, typename UnaryOp, int Traversal, int Unrolling, |
| 77 | int VmlTraversal = vml_assign_traits<Derived1, Derived2, UnaryOp>::Traversal > |
| 78 | struct vml_assign_impl |
| 79 | : assign_impl<Derived1, Eigen::CwiseUnaryOp<UnaryOp, Derived2>,Traversal,Unrolling,BuiltIn> |
| 80 | { |
| 81 | }; |
| 82 | |
| 83 | template<typename Derived1, typename Derived2, typename UnaryOp, int Traversal, int Unrolling> |
| 84 | struct vml_assign_impl<Derived1, Derived2, UnaryOp, Traversal, Unrolling, InnerVectorizedTraversal> |
| 85 | { |
| 86 | typedef typename Derived1::Scalar Scalar; |
| 87 | typedef typename Derived1::Index Index; |
| 88 | static inline void run(Derived1& dst, const CwiseUnaryOp<UnaryOp, Derived2>& src) |
| 89 | { |
| 90 | // in case we want to (or have to) skip VML at runtime we can call: |
| 91 | // assign_impl<Derived1,Eigen::CwiseUnaryOp<UnaryOp, Derived2>,Traversal,Unrolling,BuiltIn>::run(dst,src); |
| 92 | const Index innerSize = dst.innerSize(); |
| 93 | const Index outerSize = dst.outerSize(); |
| 94 | for(Index outer = 0; outer < outerSize; ++outer) { |
| 95 | const Scalar *src_ptr = src.IsRowMajor ? &(src.nestedExpression().coeffRef(outer,0)) : |
| 96 | &(src.nestedExpression().coeffRef(0, outer)); |
| 97 | Scalar *dst_ptr = dst.IsRowMajor ? &(dst.coeffRef(outer,0)) : &(dst.coeffRef(0, outer)); |
| 98 | vml_call<UnaryOp>::run(src.functor(), innerSize, src_ptr, dst_ptr ); |
| 99 | } |
| 100 | } |
| 101 | }; |
| 102 | |
| 103 | template<typename Derived1, typename Derived2, typename UnaryOp, int Traversal, int Unrolling> |
| 104 | struct vml_assign_impl<Derived1, Derived2, UnaryOp, Traversal, Unrolling, LinearVectorizedTraversal> |
| 105 | { |
| 106 | static inline void run(Derived1& dst, const CwiseUnaryOp<UnaryOp, Derived2>& src) |
| 107 | { |
| 108 | // in case we want to (or have to) skip VML at runtime we can call: |
| 109 | // assign_impl<Derived1,Eigen::CwiseUnaryOp<UnaryOp, Derived2>,Traversal,Unrolling,BuiltIn>::run(dst,src); |
| 110 | vml_call<UnaryOp>::run(src.functor(), dst.size(), src.nestedExpression().data(), dst.data() ); |
| 111 | } |
| 112 | }; |
| 113 | |
| 114 | // Macroses |
| 115 | |
| 116 | #define EIGEN_MKL_VML_SPECIALIZE_ASSIGN(TRAVERSAL,UNROLLING) \ |
| 117 | template<typename Derived1, typename Derived2, typename UnaryOp> \ |
| 118 | struct assign_impl<Derived1, Eigen::CwiseUnaryOp<UnaryOp, Derived2>, TRAVERSAL, UNROLLING, Specialized> { \ |
| 119 | static inline void run(Derived1 &dst, const Eigen::CwiseUnaryOp<UnaryOp, Derived2> &src) { \ |
| 120 | vml_assign_impl<Derived1,Derived2,UnaryOp,TRAVERSAL,UNROLLING>::run(dst, src); \ |
| 121 | } \ |
| 122 | }; |
| 123 | |
| 124 | EIGEN_MKL_VML_SPECIALIZE_ASSIGN(DefaultTraversal,NoUnrolling) |
| 125 | EIGEN_MKL_VML_SPECIALIZE_ASSIGN(DefaultTraversal,CompleteUnrolling) |
| 126 | EIGEN_MKL_VML_SPECIALIZE_ASSIGN(DefaultTraversal,InnerUnrolling) |
| 127 | EIGEN_MKL_VML_SPECIALIZE_ASSIGN(LinearTraversal,NoUnrolling) |
| 128 | EIGEN_MKL_VML_SPECIALIZE_ASSIGN(LinearTraversal,CompleteUnrolling) |
| 129 | EIGEN_MKL_VML_SPECIALIZE_ASSIGN(InnerVectorizedTraversal,NoUnrolling) |
| 130 | EIGEN_MKL_VML_SPECIALIZE_ASSIGN(InnerVectorizedTraversal,CompleteUnrolling) |
| 131 | EIGEN_MKL_VML_SPECIALIZE_ASSIGN(InnerVectorizedTraversal,InnerUnrolling) |
| 132 | EIGEN_MKL_VML_SPECIALIZE_ASSIGN(LinearVectorizedTraversal,CompleteUnrolling) |
| 133 | EIGEN_MKL_VML_SPECIALIZE_ASSIGN(LinearVectorizedTraversal,NoUnrolling) |
| 134 | EIGEN_MKL_VML_SPECIALIZE_ASSIGN(SliceVectorizedTraversal,NoUnrolling) |
| 135 | |
| 136 | |
| 137 | #if !defined (EIGEN_FAST_MATH) || (EIGEN_FAST_MATH != 1) |
| 138 | #define EIGEN_MKL_VML_MODE VML_HA |
| 139 | #else |
| 140 | #define EIGEN_MKL_VML_MODE VML_LA |
| 141 | #endif |
| 142 | |
| 143 | #define EIGEN_MKL_VML_DECLARE_UNARY_CALL(EIGENOP, VMLOP, EIGENTYPE, VMLTYPE) \ |
| 144 | template<> struct vml_call< scalar_##EIGENOP##_op<EIGENTYPE> > { \ |
| 145 | enum { IsSupported = 1 }; \ |
| 146 | static inline void run( const scalar_##EIGENOP##_op<EIGENTYPE>& /*func*/, \ |
| 147 | int size, const EIGENTYPE* src, EIGENTYPE* dst) { \ |
| 148 | VMLOP(size, (const VMLTYPE*)src, (VMLTYPE*)dst); \ |
| 149 | } \ |
| 150 | }; |
| 151 | |
| 152 | #define EIGEN_MKL_VML_DECLARE_UNARY_CALL_LA(EIGENOP, VMLOP, EIGENTYPE, VMLTYPE) \ |
| 153 | template<> struct vml_call< scalar_##EIGENOP##_op<EIGENTYPE> > { \ |
| 154 | enum { IsSupported = 1 }; \ |
| 155 | static inline void run( const scalar_##EIGENOP##_op<EIGENTYPE>& /*func*/, \ |
| 156 | int size, const EIGENTYPE* src, EIGENTYPE* dst) { \ |
| 157 | MKL_INT64 vmlMode = EIGEN_MKL_VML_MODE; \ |
| 158 | VMLOP(size, (const VMLTYPE*)src, (VMLTYPE*)dst, vmlMode); \ |
| 159 | } \ |
| 160 | }; |
| 161 | |
| 162 | #define EIGEN_MKL_VML_DECLARE_POW_CALL(EIGENOP, VMLOP, EIGENTYPE, VMLTYPE) \ |
| 163 | template<> struct vml_call< scalar_##EIGENOP##_op<EIGENTYPE> > { \ |
| 164 | enum { IsSupported = 1 }; \ |
| 165 | static inline void run( const scalar_##EIGENOP##_op<EIGENTYPE>& func, \ |
| 166 | int size, const EIGENTYPE* src, EIGENTYPE* dst) { \ |
| 167 | EIGENTYPE exponent = func.m_exponent; \ |
| 168 | MKL_INT64 vmlMode = EIGEN_MKL_VML_MODE; \ |
| 169 | VMLOP(&size, (const VMLTYPE*)src, (const VMLTYPE*)&exponent, \ |
| 170 | (VMLTYPE*)dst, &vmlMode); \ |
| 171 | } \ |
| 172 | }; |
| 173 | |
| 174 | #define EIGEN_MKL_VML_DECLARE_UNARY_CALLS_REAL(EIGENOP, VMLOP) \ |
| 175 | EIGEN_MKL_VML_DECLARE_UNARY_CALL(EIGENOP, vs##VMLOP, float, float) \ |
| 176 | EIGEN_MKL_VML_DECLARE_UNARY_CALL(EIGENOP, vd##VMLOP, double, double) |
| 177 | |
| 178 | #define EIGEN_MKL_VML_DECLARE_UNARY_CALLS_COMPLEX(EIGENOP, VMLOP) \ |
| 179 | EIGEN_MKL_VML_DECLARE_UNARY_CALL(EIGENOP, vc##VMLOP, scomplex, MKL_Complex8) \ |
| 180 | EIGEN_MKL_VML_DECLARE_UNARY_CALL(EIGENOP, vz##VMLOP, dcomplex, MKL_Complex16) |
| 181 | |
| 182 | #define EIGEN_MKL_VML_DECLARE_UNARY_CALLS(EIGENOP, VMLOP) \ |
| 183 | EIGEN_MKL_VML_DECLARE_UNARY_CALLS_REAL(EIGENOP, VMLOP) \ |
| 184 | EIGEN_MKL_VML_DECLARE_UNARY_CALLS_COMPLEX(EIGENOP, VMLOP) |
| 185 | |
| 186 | |
| 187 | #define EIGEN_MKL_VML_DECLARE_UNARY_CALLS_REAL_LA(EIGENOP, VMLOP) \ |
| 188 | EIGEN_MKL_VML_DECLARE_UNARY_CALL_LA(EIGENOP, vms##VMLOP, float, float) \ |
| 189 | EIGEN_MKL_VML_DECLARE_UNARY_CALL_LA(EIGENOP, vmd##VMLOP, double, double) |
| 190 | |
| 191 | #define EIGEN_MKL_VML_DECLARE_UNARY_CALLS_COMPLEX_LA(EIGENOP, VMLOP) \ |
| 192 | EIGEN_MKL_VML_DECLARE_UNARY_CALL_LA(EIGENOP, vmc##VMLOP, scomplex, MKL_Complex8) \ |
| 193 | EIGEN_MKL_VML_DECLARE_UNARY_CALL_LA(EIGENOP, vmz##VMLOP, dcomplex, MKL_Complex16) |
| 194 | |
| 195 | #define EIGEN_MKL_VML_DECLARE_UNARY_CALLS_LA(EIGENOP, VMLOP) \ |
| 196 | EIGEN_MKL_VML_DECLARE_UNARY_CALLS_REAL_LA(EIGENOP, VMLOP) \ |
| 197 | EIGEN_MKL_VML_DECLARE_UNARY_CALLS_COMPLEX_LA(EIGENOP, VMLOP) |
| 198 | |
| 199 | |
| 200 | EIGEN_MKL_VML_DECLARE_UNARY_CALLS_LA(sin, Sin) |
| 201 | EIGEN_MKL_VML_DECLARE_UNARY_CALLS_LA(asin, Asin) |
| 202 | EIGEN_MKL_VML_DECLARE_UNARY_CALLS_LA(cos, Cos) |
| 203 | EIGEN_MKL_VML_DECLARE_UNARY_CALLS_LA(acos, Acos) |
| 204 | EIGEN_MKL_VML_DECLARE_UNARY_CALLS_LA(tan, Tan) |
| 205 | //EIGEN_MKL_VML_DECLARE_UNARY_CALLS(abs, Abs) |
| 206 | EIGEN_MKL_VML_DECLARE_UNARY_CALLS_LA(exp, Exp) |
| 207 | EIGEN_MKL_VML_DECLARE_UNARY_CALLS_LA(log, Ln) |
| 208 | EIGEN_MKL_VML_DECLARE_UNARY_CALLS_LA(sqrt, Sqrt) |
| 209 | |
| 210 | EIGEN_MKL_VML_DECLARE_UNARY_CALLS_REAL(square, Sqr) |
| 211 | |
| 212 | // The vm*powx functions are not avaibale in the windows version of MKL. |
Carlos Hernandez | 7faaa9f | 2014-08-05 17:53:32 -0700 | [diff] [blame] | 213 | #ifndef _WIN32 |
Narayan Kamath | c981c48 | 2012-11-02 10:59:05 +0000 | [diff] [blame] | 214 | EIGEN_MKL_VML_DECLARE_POW_CALL(pow, vmspowx_, float, float) |
| 215 | EIGEN_MKL_VML_DECLARE_POW_CALL(pow, vmdpowx_, double, double) |
| 216 | EIGEN_MKL_VML_DECLARE_POW_CALL(pow, vmcpowx_, scomplex, MKL_Complex8) |
| 217 | EIGEN_MKL_VML_DECLARE_POW_CALL(pow, vmzpowx_, dcomplex, MKL_Complex16) |
| 218 | #endif |
| 219 | |
| 220 | } // end namespace internal |
| 221 | |
| 222 | } // end namespace Eigen |
| 223 | |
| 224 | #endif // EIGEN_ASSIGN_VML_H |