• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // This file is part of Eigen, a lightweight C++ template library
2 // for linear algebra.
3 //
4 // Copyright (C) 2012 Gael Guennebaud <gael.guennebaud@inria.fr>
5 //
6 // This Source Code Form is subject to the terms of the Mozilla
7 // Public License v. 2.0. If a copy of the MPL was not distributed
8 // with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
9 
10 #ifndef EIGEN_SPARSELU_GEMM_KERNEL_H
11 #define EIGEN_SPARSELU_GEMM_KERNEL_H
12 
13 namespace Eigen {
14 
15 namespace internal {
16 
17 
18 /** \internal
19   * A general matrix-matrix product kernel optimized for the SparseLU factorization.
20   *  - A, B, and C must be column major
21   *  - lda and ldc must be multiples of the respective packet size
22   *  - C must have the same alignment as A
23   */
24 template<typename Scalar>
25 EIGEN_DONT_INLINE
sparselu_gemm(Index m,Index n,Index d,const Scalar * A,Index lda,const Scalar * B,Index ldb,Scalar * C,Index ldc)26 void sparselu_gemm(Index m, Index n, Index d, const Scalar* A, Index lda, const Scalar* B, Index ldb, Scalar* C, Index ldc)
27 {
28   using namespace Eigen::internal;
29 
30   typedef typename packet_traits<Scalar>::type Packet;
31   enum {
32     NumberOfRegisters = EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS,
33     PacketSize = packet_traits<Scalar>::size,
34     PM = 8,                             // peeling in M
35     RN = 2,                             // register blocking
36     RK = NumberOfRegisters>=16 ? 4 : 2, // register blocking
37     BM = 4096/sizeof(Scalar),           // number of rows of A-C per chunk
38     SM = PM*PacketSize                  // step along M
39   };
40   Index d_end = (d/RK)*RK;    // number of columns of A (rows of B) suitable for full register blocking
41   Index n_end = (n/RN)*RN;    // number of columns of B-C suitable for processing RN columns at once
42   Index i0 = internal::first_default_aligned(A,m);
43 
44   eigen_internal_assert(((lda%PacketSize)==0) && ((ldc%PacketSize)==0) && (i0==internal::first_default_aligned(C,m)));
45 
46   // handle the non aligned rows of A and C without any optimization:
47   for(Index i=0; i<i0; ++i)
48   {
49     for(Index j=0; j<n; ++j)
50     {
51       Scalar c = C[i+j*ldc];
52       for(Index k=0; k<d; ++k)
53         c += B[k+j*ldb] * A[i+k*lda];
54       C[i+j*ldc] = c;
55     }
56   }
57   // process the remaining rows per chunk of BM rows
58   for(Index ib=i0; ib<m; ib+=BM)
59   {
60     Index actual_b = std::min<Index>(BM, m-ib);                 // actual number of rows
61     Index actual_b_end1 = (actual_b/SM)*SM;                   // actual number of rows suitable for peeling
62     Index actual_b_end2 = (actual_b/PacketSize)*PacketSize;   // actual number of rows suitable for vectorization
63 
64     // Let's process two columns of B-C at once
65     for(Index j=0; j<n_end; j+=RN)
66     {
67       const Scalar* Bc0 = B+(j+0)*ldb;
68       const Scalar* Bc1 = B+(j+1)*ldb;
69 
70       for(Index k=0; k<d_end; k+=RK)
71       {
72 
73         // load and expand a RN x RK block of B
74         Packet b00, b10, b20, b30, b01, b11, b21, b31;
75                   { b00 = pset1<Packet>(Bc0[0]); }
76                   { b10 = pset1<Packet>(Bc0[1]); }
77         if(RK==4) { b20 = pset1<Packet>(Bc0[2]); }
78         if(RK==4) { b30 = pset1<Packet>(Bc0[3]); }
79                   { b01 = pset1<Packet>(Bc1[0]); }
80                   { b11 = pset1<Packet>(Bc1[1]); }
81         if(RK==4) { b21 = pset1<Packet>(Bc1[2]); }
82         if(RK==4) { b31 = pset1<Packet>(Bc1[3]); }
83 
84         Packet a0, a1, a2, a3, c0, c1, t0, t1;
85 
86         const Scalar* A0 = A+ib+(k+0)*lda;
87         const Scalar* A1 = A+ib+(k+1)*lda;
88         const Scalar* A2 = A+ib+(k+2)*lda;
89         const Scalar* A3 = A+ib+(k+3)*lda;
90 
91         Scalar* C0 = C+ib+(j+0)*ldc;
92         Scalar* C1 = C+ib+(j+1)*ldc;
93 
94                   a0 = pload<Packet>(A0);
95                   a1 = pload<Packet>(A1);
96         if(RK==4)
97         {
98           a2 = pload<Packet>(A2);
99           a3 = pload<Packet>(A3);
100         }
101         else
102         {
103           // workaround "may be used uninitialized in this function" warning
104           a2 = a3 = a0;
105         }
106 
107 #define KMADD(c, a, b, tmp) {tmp = b; tmp = pmul(a,tmp); c = padd(c,tmp);}
108 #define WORK(I)  \
109                      c0 = pload<Packet>(C0+i+(I)*PacketSize);    \
110                      c1 = pload<Packet>(C1+i+(I)*PacketSize);    \
111                      KMADD(c0, a0, b00, t0)                      \
112                      KMADD(c1, a0, b01, t1)                      \
113                      a0 = pload<Packet>(A0+i+(I+1)*PacketSize);  \
114                      KMADD(c0, a1, b10, t0)                      \
115                      KMADD(c1, a1, b11, t1)                      \
116                      a1 = pload<Packet>(A1+i+(I+1)*PacketSize);  \
117           if(RK==4){ KMADD(c0, a2, b20, t0)                     }\
118           if(RK==4){ KMADD(c1, a2, b21, t1)                     }\
119           if(RK==4){ a2 = pload<Packet>(A2+i+(I+1)*PacketSize); }\
120           if(RK==4){ KMADD(c0, a3, b30, t0)                     }\
121           if(RK==4){ KMADD(c1, a3, b31, t1)                     }\
122           if(RK==4){ a3 = pload<Packet>(A3+i+(I+1)*PacketSize); }\
123                      pstore(C0+i+(I)*PacketSize, c0);            \
124                      pstore(C1+i+(I)*PacketSize, c1)
125 
126         // process rows of A' - C' with aggressive vectorization and peeling
127         for(Index i=0; i<actual_b_end1; i+=PacketSize*8)
128         {
129           EIGEN_ASM_COMMENT("SPARSELU_GEMML_KERNEL1");
130                     prefetch((A0+i+(5)*PacketSize));
131                     prefetch((A1+i+(5)*PacketSize));
132           if(RK==4) prefetch((A2+i+(5)*PacketSize));
133           if(RK==4) prefetch((A3+i+(5)*PacketSize));
134 
135           WORK(0);
136           WORK(1);
137           WORK(2);
138           WORK(3);
139           WORK(4);
140           WORK(5);
141           WORK(6);
142           WORK(7);
143         }
144         // process the remaining rows with vectorization only
145         for(Index i=actual_b_end1; i<actual_b_end2; i+=PacketSize)
146         {
147           WORK(0);
148         }
149 #undef WORK
150         // process the remaining rows without vectorization
151         for(Index i=actual_b_end2; i<actual_b; ++i)
152         {
153           if(RK==4)
154           {
155             C0[i] += A0[i]*Bc0[0]+A1[i]*Bc0[1]+A2[i]*Bc0[2]+A3[i]*Bc0[3];
156             C1[i] += A0[i]*Bc1[0]+A1[i]*Bc1[1]+A2[i]*Bc1[2]+A3[i]*Bc1[3];
157           }
158           else
159           {
160             C0[i] += A0[i]*Bc0[0]+A1[i]*Bc0[1];
161             C1[i] += A0[i]*Bc1[0]+A1[i]*Bc1[1];
162           }
163         }
164 
165         Bc0 += RK;
166         Bc1 += RK;
167       } // peeled loop on k
168     } // peeled loop on the columns j
169     // process the last column (we now perform a matrix-vector product)
170     if((n-n_end)>0)
171     {
172       const Scalar* Bc0 = B+(n-1)*ldb;
173 
174       for(Index k=0; k<d_end; k+=RK)
175       {
176 
177         // load and expand a 1 x RK block of B
178         Packet b00, b10, b20, b30;
179                   b00 = pset1<Packet>(Bc0[0]);
180                   b10 = pset1<Packet>(Bc0[1]);
181         if(RK==4) b20 = pset1<Packet>(Bc0[2]);
182         if(RK==4) b30 = pset1<Packet>(Bc0[3]);
183 
184         Packet a0, a1, a2, a3, c0, t0/*, t1*/;
185 
186         const Scalar* A0 = A+ib+(k+0)*lda;
187         const Scalar* A1 = A+ib+(k+1)*lda;
188         const Scalar* A2 = A+ib+(k+2)*lda;
189         const Scalar* A3 = A+ib+(k+3)*lda;
190 
191         Scalar* C0 = C+ib+(n_end)*ldc;
192 
193                   a0 = pload<Packet>(A0);
194                   a1 = pload<Packet>(A1);
195         if(RK==4)
196         {
197           a2 = pload<Packet>(A2);
198           a3 = pload<Packet>(A3);
199         }
200         else
201         {
202           // workaround "may be used uninitialized in this function" warning
203           a2 = a3 = a0;
204         }
205 
206 #define WORK(I) \
207                    c0 = pload<Packet>(C0+i+(I)*PacketSize);     \
208                    KMADD(c0, a0, b00, t0)                       \
209                    a0 = pload<Packet>(A0+i+(I+1)*PacketSize);   \
210                    KMADD(c0, a1, b10, t0)                       \
211                    a1 = pload<Packet>(A1+i+(I+1)*PacketSize);   \
212         if(RK==4){ KMADD(c0, a2, b20, t0)                      }\
213         if(RK==4){ a2 = pload<Packet>(A2+i+(I+1)*PacketSize);  }\
214         if(RK==4){ KMADD(c0, a3, b30, t0)                      }\
215         if(RK==4){ a3 = pload<Packet>(A3+i+(I+1)*PacketSize);  }\
216                    pstore(C0+i+(I)*PacketSize, c0);
217 
218         // agressive vectorization and peeling
219         for(Index i=0; i<actual_b_end1; i+=PacketSize*8)
220         {
221           EIGEN_ASM_COMMENT("SPARSELU_GEMML_KERNEL2");
222           WORK(0);
223           WORK(1);
224           WORK(2);
225           WORK(3);
226           WORK(4);
227           WORK(5);
228           WORK(6);
229           WORK(7);
230         }
231         // vectorization only
232         for(Index i=actual_b_end1; i<actual_b_end2; i+=PacketSize)
233         {
234           WORK(0);
235         }
236         // remaining scalars
237         for(Index i=actual_b_end2; i<actual_b; ++i)
238         {
239           if(RK==4)
240             C0[i] += A0[i]*Bc0[0]+A1[i]*Bc0[1]+A2[i]*Bc0[2]+A3[i]*Bc0[3];
241           else
242             C0[i] += A0[i]*Bc0[0]+A1[i]*Bc0[1];
243         }
244 
245         Bc0 += RK;
246 #undef WORK
247       }
248     }
249 
250     // process the last columns of A, corresponding to the last rows of B
251     Index rd = d-d_end;
252     if(rd>0)
253     {
254       for(Index j=0; j<n; ++j)
255       {
256         enum {
257           Alignment = PacketSize>1 ? Aligned : 0
258         };
259         typedef Map<Matrix<Scalar,Dynamic,1>, Alignment > MapVector;
260         typedef Map<const Matrix<Scalar,Dynamic,1>, Alignment > ConstMapVector;
261         if(rd==1)       MapVector(C+j*ldc+ib,actual_b) += B[0+d_end+j*ldb] * ConstMapVector(A+(d_end+0)*lda+ib, actual_b);
262 
263         else if(rd==2)  MapVector(C+j*ldc+ib,actual_b) += B[0+d_end+j*ldb] * ConstMapVector(A+(d_end+0)*lda+ib, actual_b)
264                                                         + B[1+d_end+j*ldb] * ConstMapVector(A+(d_end+1)*lda+ib, actual_b);
265 
266         else            MapVector(C+j*ldc+ib,actual_b) += B[0+d_end+j*ldb] * ConstMapVector(A+(d_end+0)*lda+ib, actual_b)
267                                                         + B[1+d_end+j*ldb] * ConstMapVector(A+(d_end+1)*lda+ib, actual_b)
268                                                         + B[2+d_end+j*ldb] * ConstMapVector(A+(d_end+2)*lda+ib, actual_b);
269       }
270     }
271 
272   } // blocking on the rows of A and C
273 }
274 #undef KMADD
275 
276 } // namespace internal
277 
278 } // namespace Eigen
279 
280 #endif // EIGEN_SPARSELU_GEMM_KERNEL_H
281