Newer
Older
// This file is part of Eigen, a lightweight C++ template library
// for linear algebra.
//
// Copyright (C) 2012 Gael Guennebaud <gael.guennebaud@inria.fr>
//
// This Source Code Form is subject to the terms of the Mozilla
// Public License v. 2.0. If a copy of the MPL was not distributed
// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
#ifndef EIGEN_SPARSELU_GEMM_KERNEL_H
#define EIGEN_SPARSELU_GEMM_KERNEL_H
namespace Eigen {
namespace internal {
/** \internal
* A general matrix-matrix product kernel optimized for the SparseLU factorization.
* - A, B, and C must be column major
* - lda and ldc must be multiples of the respective packet size
* - C must have the same alignment as A
*/
EIGEN_DONT_INLINE
void sparselu_gemm(Index m, Index n, Index d, const Scalar* A, Index lda, const Scalar* B, Index ldb, Scalar* C, Index ldc)
{
using namespace Eigen::internal;
typedef typename packet_traits<Scalar>::type Packet;
enum {
NumberOfRegisters = EIGEN_ARCH_DEFAULT_NUMBER_OF_REGISTERS,
PacketSize = packet_traits<Scalar>::size,
PM = 8, // peeling in M
RN = 2, // register blocking
RK = NumberOfRegisters>=16 ? 4 : 2, // register blocking
BM = 4096/sizeof(Scalar), // number of rows of A-C per chunk
SM = PM*PacketSize // step along M
};
Index d_end = (d/RK)*RK; // number of columns of A (rows of B) suitable for full register blocking
Index n_end = (n/RN)*RN; // number of columns of B-C suitable for processing RN columns at once
eigen_internal_assert(((lda%PacketSize)==0) && ((ldc%PacketSize)==0) && (i0==internal::first_default_aligned(C,m)));
// handle the non aligned rows of A and C without any optimization:
for(Index i=0; i<i0; ++i)
{
for(Index j=0; j<n; ++j)
{
Scalar c = C[i+j*ldc];
for(Index k=0; k<d; ++k)
c += B[k+j*ldb] * A[i+k*lda];
C[i+j*ldc] = c;
}
}
// process the remaining rows per chunk of BM rows
for(Index ib=i0; ib<m; ib+=BM)
{
Index actual_b = std::min<Index>(BM, m-ib); // actual number of rows
Index actual_b_end1 = (actual_b/SM)*SM; // actual number of rows suitable for peeling
Index actual_b_end2 = (actual_b/PacketSize)*PacketSize; // actual number of rows suitable for vectorization
// Let's process two columns of B-C at once
for(Index j=0; j<n_end; j+=RN)
{
const Scalar* Bc0 = B+(j+0)*ldb;
const Scalar* Bc1 = B+(j+1)*ldb;
for(Index k=0; k<d_end; k+=RK)
{
// load and expand a RN x RK block of B
Packet b00, b10, b20, b30, b01, b11, b21, b31;
{ b00 = pset1<Packet>(Bc0[0]); }
{ b10 = pset1<Packet>(Bc0[1]); }
if(RK==4) { b20 = pset1<Packet>(Bc0[2]); }
if(RK==4) { b30 = pset1<Packet>(Bc0[3]); }
{ b01 = pset1<Packet>(Bc1[0]); }
{ b11 = pset1<Packet>(Bc1[1]); }
if(RK==4) { b21 = pset1<Packet>(Bc1[2]); }
if(RK==4) { b31 = pset1<Packet>(Bc1[3]); }
Packet a0, a1, a2, a3, c0, c1, t0, t1;
const Scalar* A0 = A+ib+(k+0)*lda;
const Scalar* A1 = A+ib+(k+1)*lda;
const Scalar* A2 = A+ib+(k+2)*lda;
const Scalar* A3 = A+ib+(k+3)*lda;
Scalar* C0 = C+ib+(j+0)*ldc;
Scalar* C1 = C+ib+(j+1)*ldc;
a0 = pload<Packet>(A0);
a1 = pload<Packet>(A1);
if(RK==4)
{
a2 = pload<Packet>(A2);
a3 = pload<Packet>(A3);
}
else
{
// workaround "may be used uninitialized in this function" warning
a2 = a3 = a0;
}
#define KMADD(c, a, b, tmp) {tmp = b; tmp = pmul(a,tmp); c = padd(c,tmp);}
#define WORK(I) \
c0 = pload<Packet>(C0+i+(I)*PacketSize); \
c1 = pload<Packet>(C1+i+(I)*PacketSize); \
KMADD(c0, a0, b00, t0) \
KMADD(c1, a0, b01, t1) \
a0 = pload<Packet>(A0+i+(I+1)*PacketSize); \
KMADD(c0, a1, b10, t0) \
KMADD(c1, a1, b11, t1) \
a1 = pload<Packet>(A1+i+(I+1)*PacketSize); \
if(RK==4){ KMADD(c0, a2, b20, t0) }\
if(RK==4){ KMADD(c1, a2, b21, t1) }\
if(RK==4){ a2 = pload<Packet>(A2+i+(I+1)*PacketSize); }\
if(RK==4){ KMADD(c0, a3, b30, t0) }\
if(RK==4){ KMADD(c1, a3, b31, t1) }\
if(RK==4){ a3 = pload<Packet>(A3+i+(I+1)*PacketSize); }\
pstore(C0+i+(I)*PacketSize, c0); \
pstore(C1+i+(I)*PacketSize, c1)
// process rows of A' - C' with aggressive vectorization and peeling
for(Index i=0; i<actual_b_end1; i+=PacketSize*8)
{
EIGEN_ASM_COMMENT("SPARSELU_GEMML_KERNEL1");
prefetch((A0+i+(5)*PacketSize));
prefetch((A1+i+(5)*PacketSize));
if(RK==4) prefetch((A2+i+(5)*PacketSize));
if(RK==4) prefetch((A3+i+(5)*PacketSize));
WORK(0);
WORK(1);
WORK(2);
WORK(3);
WORK(4);
WORK(5);
WORK(6);
WORK(7);
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
}
// process the remaining rows with vectorization only
for(Index i=actual_b_end1; i<actual_b_end2; i+=PacketSize)
{
WORK(0);
}
#undef WORK
// process the remaining rows without vectorization
for(Index i=actual_b_end2; i<actual_b; ++i)
{
if(RK==4)
{
C0[i] += A0[i]*Bc0[0]+A1[i]*Bc0[1]+A2[i]*Bc0[2]+A3[i]*Bc0[3];
C1[i] += A0[i]*Bc1[0]+A1[i]*Bc1[1]+A2[i]*Bc1[2]+A3[i]*Bc1[3];
}
else
{
C0[i] += A0[i]*Bc0[0]+A1[i]*Bc0[1];
C1[i] += A0[i]*Bc1[0]+A1[i]*Bc1[1];
}
}
Bc0 += RK;
Bc1 += RK;
} // peeled loop on k
} // peeled loop on the columns j
// process the last column (we now perform a matrix-vector product)
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
if((n-n_end)>0)
{
const Scalar* Bc0 = B+(n-1)*ldb;
for(Index k=0; k<d_end; k+=RK)
{
// load and expand a 1 x RK block of B
Packet b00, b10, b20, b30;
b00 = pset1<Packet>(Bc0[0]);
b10 = pset1<Packet>(Bc0[1]);
if(RK==4) b20 = pset1<Packet>(Bc0[2]);
if(RK==4) b30 = pset1<Packet>(Bc0[3]);
Packet a0, a1, a2, a3, c0, t0/*, t1*/;
const Scalar* A0 = A+ib+(k+0)*lda;
const Scalar* A1 = A+ib+(k+1)*lda;
const Scalar* A2 = A+ib+(k+2)*lda;
const Scalar* A3 = A+ib+(k+3)*lda;
Scalar* C0 = C+ib+(n_end)*ldc;
a0 = pload<Packet>(A0);
a1 = pload<Packet>(A1);
if(RK==4)
{
a2 = pload<Packet>(A2);
a3 = pload<Packet>(A3);
}
else
{
// workaround "may be used uninitialized in this function" warning
a2 = a3 = a0;
}
#define WORK(I) \
c0 = pload<Packet>(C0+i+(I)*PacketSize); \
KMADD(c0, a0, b00, t0) \
a0 = pload<Packet>(A0+i+(I+1)*PacketSize); \
KMADD(c0, a1, b10, t0) \
a1 = pload<Packet>(A1+i+(I+1)*PacketSize); \
if(RK==4){ KMADD(c0, a2, b20, t0) }\
if(RK==4){ a2 = pload<Packet>(A2+i+(I+1)*PacketSize); }\
if(RK==4){ KMADD(c0, a3, b30, t0) }\
if(RK==4){ a3 = pload<Packet>(A3+i+(I+1)*PacketSize); }\
pstore(C0+i+(I)*PacketSize, c0);
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
// agressive vectorization and peeling
for(Index i=0; i<actual_b_end1; i+=PacketSize*8)
{
EIGEN_ASM_COMMENT("SPARSELU_GEMML_KERNEL2");
WORK(0);
WORK(1);
WORK(2);
WORK(3);
WORK(4);
WORK(5);
WORK(6);
WORK(7);
}
// vectorization only
for(Index i=actual_b_end1; i<actual_b_end2; i+=PacketSize)
{
WORK(0);
}
// remaining scalars
for(Index i=actual_b_end2; i<actual_b; ++i)
{
if(RK==4)
C0[i] += A0[i]*Bc0[0]+A1[i]*Bc0[1]+A2[i]*Bc0[2]+A3[i]*Bc0[3];
else
C0[i] += A0[i]*Bc0[0]+A1[i]*Bc0[1];
}
Bc0 += RK;
#undef WORK
}
}
// process the last columns of A, corresponding to the last rows of B
Index rd = d-d_end;
if(rd>0)
{
for(Index j=0; j<n; ++j)
{
enum {
Alignment = PacketSize>1 ? Aligned : 0
};
typedef Map<Matrix<Scalar,Dynamic,1>, Alignment > MapVector;
typedef Map<const Matrix<Scalar,Dynamic,1>, Alignment > ConstMapVector;
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);
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)
+ B[1+d_end+j*ldb] * ConstMapVector(A+(d_end+1)*lda+ib, actual_b);
else MapVector(C+j*ldc+ib,actual_b) += B[0+d_end+j*ldb] * ConstMapVector(A+(d_end+0)*lda+ib, actual_b)
+ B[1+d_end+j*ldb] * ConstMapVector(A+(d_end+1)*lda+ib, actual_b)
+ B[2+d_end+j*ldb] * ConstMapVector(A+(d_end+2)*lda+ib, actual_b);
}
}
} // blocking on the rows of A and C
}
#undef KMADD
} // namespace internal
} // namespace Eigen
#endif // EIGEN_SPARSELU_GEMM_KERNEL_H