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 (at) 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,typename Index> 25 EIGEN_DONT_INLINE 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_aligned(A,m); 43 44 eigen_internal_assert(((lda%PacketSize)==0) && ((ldc%PacketSize)==0) && (i0==internal::first_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 WORK(0); 135 WORK(1); 136 WORK(2); 137 WORK(3); 138 WORK(4); 139 WORK(5); 140 WORK(6); 141 WORK(7); 142 } 143 // process the remaining rows with vectorization only 144 for(Index i=actual_b_end1; i<actual_b_end2; i+=PacketSize) 145 { 146 WORK(0); 147 } 148 #undef WORK 149 // process the remaining rows without vectorization 150 for(Index i=actual_b_end2; i<actual_b; ++i) 151 { 152 if(RK==4) 153 { 154 C0[i] += A0[i]*Bc0[0]+A1[i]*Bc0[1]+A2[i]*Bc0[2]+A3[i]*Bc0[3]; 155 C1[i] += A0[i]*Bc1[0]+A1[i]*Bc1[1]+A2[i]*Bc1[2]+A3[i]*Bc1[3]; 156 } 157 else 158 { 159 C0[i] += A0[i]*Bc0[0]+A1[i]*Bc0[1]; 160 C1[i] += A0[i]*Bc1[0]+A1[i]*Bc1[1]; 161 } 162 } 163 164 Bc0 += RK; 165 Bc1 += RK; 166 } // peeled loop on k 167 } // peeled loop on the columns j 168 // process the last column (we now perform a matrux-vector product) 169 if((n-n_end)>0) 170 { 171 const Scalar* Bc0 = B+(n-1)*ldb; 172 173 for(Index k=0; k<d_end; k+=RK) 174 { 175 176 // load and expand a 1 x RK block of B 177 Packet b00, b10, b20, b30; 178 b00 = pset1<Packet>(Bc0[0]); 179 b10 = pset1<Packet>(Bc0[1]); 180 if(RK==4) b20 = pset1<Packet>(Bc0[2]); 181 if(RK==4) b30 = pset1<Packet>(Bc0[3]); 182 183 Packet a0, a1, a2, a3, c0, t0/*, t1*/; 184 185 const Scalar* A0 = A+ib+(k+0)*lda; 186 const Scalar* A1 = A+ib+(k+1)*lda; 187 const Scalar* A2 = A+ib+(k+2)*lda; 188 const Scalar* A3 = A+ib+(k+3)*lda; 189 190 Scalar* C0 = C+ib+(n_end)*ldc; 191 192 a0 = pload<Packet>(A0); 193 a1 = pload<Packet>(A1); 194 if(RK==4) 195 { 196 a2 = pload<Packet>(A2); 197 a3 = pload<Packet>(A3); 198 } 199 else 200 { 201 // workaround "may be used uninitialized in this function" warning 202 a2 = a3 = a0; 203 } 204 205 #define WORK(I) \ 206 c0 = pload<Packet>(C0+i+(I)*PacketSize); \ 207 KMADD(c0, a0, b00, t0) \ 208 a0 = pload<Packet>(A0+i+(I+1)*PacketSize); \ 209 KMADD(c0, a1, b10, t0) \ 210 a1 = pload<Packet>(A1+i+(I+1)*PacketSize); \ 211 if(RK==4) KMADD(c0, a2, b20, t0) \ 212 if(RK==4) a2 = pload<Packet>(A2+i+(I+1)*PacketSize); \ 213 if(RK==4) KMADD(c0, a3, b30, t0) \ 214 if(RK==4) a3 = pload<Packet>(A3+i+(I+1)*PacketSize); \ 215 pstore(C0+i+(I)*PacketSize, c0); 216 217 // agressive vectorization and peeling 218 for(Index i=0; i<actual_b_end1; i+=PacketSize*8) 219 { 220 EIGEN_ASM_COMMENT("SPARSELU_GEMML_KERNEL2"); 221 WORK(0); 222 WORK(1); 223 WORK(2); 224 WORK(3); 225 WORK(4); 226 WORK(5); 227 WORK(6); 228 WORK(7); 229 } 230 // vectorization only 231 for(Index i=actual_b_end1; i<actual_b_end2; i+=PacketSize) 232 { 233 WORK(0); 234 } 235 // remaining scalars 236 for(Index i=actual_b_end2; i<actual_b; ++i) 237 { 238 if(RK==4) 239 C0[i] += A0[i]*Bc0[0]+A1[i]*Bc0[1]+A2[i]*Bc0[2]+A3[i]*Bc0[3]; 240 else 241 C0[i] += A0[i]*Bc0[0]+A1[i]*Bc0[1]; 242 } 243 244 Bc0 += RK; 245 #undef WORK 246 } 247 } 248 249 // process the last columns of A, corresponding to the last rows of B 250 Index rd = d-d_end; 251 if(rd>0) 252 { 253 for(Index j=0; j<n; ++j) 254 { 255 enum { 256 Alignment = PacketSize>1 ? Aligned : 0 257 }; 258 typedef Map<Matrix<Scalar,Dynamic,1>, Alignment > MapVector; 259 typedef Map<const Matrix<Scalar,Dynamic,1>, Alignment > ConstMapVector; 260 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); 261 262 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) 263 + B[1+d_end+j*ldb] * ConstMapVector(A+(d_end+1)*lda+ib, actual_b); 264 265 else MapVector(C+j*ldc+ib,actual_b) += B[0+d_end+j*ldb] * ConstMapVector(A+(d_end+0)*lda+ib, actual_b) 266 + B[1+d_end+j*ldb] * ConstMapVector(A+(d_end+1)*lda+ib, actual_b) 267 + B[2+d_end+j*ldb] * ConstMapVector(A+(d_end+2)*lda+ib, actual_b); 268 } 269 } 270 271 } // blocking on the rows of A and C 272 } 273 #undef KMADD 274 275 } // namespace internal 276 277 } // namespace Eigen 278 279 #endif // EIGEN_SPARSELU_GEMM_KERNEL_H 280