1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
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
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
|
// g++-4.4 bench_gemm.cpp -I .. -O2 -DNDEBUG -lrt -fopenmp && OMP_NUM_THREADS=2 ./a.out
// icpc bench_gemm.cpp -I .. -O3 -DNDEBUG -lrt -openmp && OMP_NUM_THREADS=2 ./a.out
#include <iostream>
#include <Eigen/Core>
#include <bench/BenchTimer.h>
using namespace std;
using namespace Eigen;
#ifndef SCALAR
// #define SCALAR std::complex<double>
#define SCALAR float
#endif
typedef SCALAR Scalar;
typedef Matrix<Scalar,Dynamic,Dynamic> M;
#ifdef HAVE_BLAS
extern "C" {
#include <bench/btl/libs/C_BLAS/blas.h>
}
static float fone = 1;
static float fzero = 0;
static double done = 1;
static double szero = 0;
static std::complex<float> cfone = 1;
static std::complex<float> cfzero = 0;
static std::complex<double> cdone = 1;
static std::complex<double> cdzero = 0;
static char notrans = 'N';
static char trans = 'T';
static char nonunit = 'N';
static char lower = 'L';
static char right = 'R';
static int intone = 1;
void blas_gemm(const MatrixXf& a, const MatrixXf& b, MatrixXf& c)
{
int M = c.rows(); int N = c.cols(); int K = a.cols();
int lda = a.rows(); int ldb = b.rows(); int ldc = c.rows();
sgemm_(¬rans,¬rans,&M,&N,&K,&fone,
const_cast<float*>(a.data()),&lda,
const_cast<float*>(b.data()),&ldb,&fone,
c.data(),&ldc);
}
void blas_gemm(const MatrixXcf& a, const MatrixXcf& b, MatrixXcf& c)
{
int M = c.rows(); int N = c.cols(); int K = a.cols();
int lda = a.rows(); int ldb = b.rows(); int ldc = c.rows();
cgemm_(¬rans,¬rans,&M,&N,&K,(float*)&cfone,
const_cast<float*>((const float*)a.data()),&lda,
const_cast<float*>((const float*)b.data()),&ldb,(float*)&cfone,
(float*)c.data(),&ldc);
}
void blas_gemm(const MatrixXcd& a, const MatrixXcd& b, MatrixXcd& c)
{
int M = c.rows(); int N = c.cols(); int K = a.cols();
int lda = a.rows(); int ldb = b.rows(); int ldc = c.rows();
zgemm_(¬rans,¬rans,&M,&N,&K,(double*)&cdone,
const_cast<double*>((const double*)a.data()),&lda,
const_cast<double*>((const double*)b.data()),&ldb,(double*)&cdone,
(double*)c.data(),&ldc);
}
void blas_gemm(const MatrixXd& a, const MatrixXd& b, MatrixXd& c)
{
int M = c.rows(); int N = c.cols(); int K = a.cols();
int lda = a.rows(); int ldb = b.rows(); int ldc = c.rows();
dgemm_(¬rans,¬rans,&M,&N,&K,&done,
const_cast<double*>(a.data()),&lda,
const_cast<double*>(b.data()),&ldb,&done,
c.data(),&ldc);
}
#endif
template<typename M>
EIGEN_DONT_INLINE void gemm(const M& a, const M& b, M& c)
{
//c.noalias() += a * b;
for(int i=0;i<b.cols();++i)
c.col(i).noalias() += a * b.col(i);
}
int main(int argc, char ** argv)
{
std::ptrdiff_t l1 = ei_queryL1CacheSize();
std::ptrdiff_t l2 = ei_queryTopLevelCacheSize();
std::cout << "L1 cache size = " << (l1>0 ? l1/1024 : -1) << " KB\n";
std::cout << "L2/L3 cache size = " << (l2>0 ? l2/1024 : -1) << " KB\n";
typedef ei_product_blocking_traits<Scalar> Blocking;
std::cout << "Register blocking = " << Blocking::mr << " x " << Blocking::nr << "\n";
int rep = 1; // number of repetitions per try
int tries = 2; // number of tries, we keep the best
int s = 2048;
int cache_size = -1;
bool need_help = false;
for (int i=1; i<argc; ++i)
{
if(argv[i][0]=='s')
s = atoi(argv[i]+1);
else if(argv[i][0]=='c')
cache_size = atoi(argv[i]+1);
else if(argv[i][0]=='t')
tries = atoi(argv[i]+1);
else if(argv[i][0]=='p')
rep = atoi(argv[i]+1);
else
need_help = true;
}
if(need_help)
{
std::cout << argv[0] << " s<matrix size> c<cache size> t<nb tries> p<nb repeats>\n";
return 1;
}
if(cache_size>0)
setCpuCacheSizes(cache_size,96*cache_size);
int m = s;
int n = s;
int p = s;
M a(m,n); a.setRandom();
M b(n,p); b.setRandom();
M c(m,p); c.setOnes();
std::cout << "Matrix sizes = " << m << "x" << p << " * " << p << "x" << n << "\n";
std::ptrdiff_t cm(m), cn(n), ck(p);
computeProductBlockingSizes<Scalar,Scalar>(ck, cm, cn);
std::cout << "blocking size = " << cm << " x " << ck << "\n";
M r = c;
// check the parallel product is correct
#ifdef EIGEN_HAS_OPENMP
int procs = omp_get_max_threads();
if(procs>1)
{
#ifdef HAVE_BLAS
blas_gemm(a,b,r);
#else
omp_set_num_threads(1);
r.noalias() += a * b;
omp_set_num_threads(procs);
#endif
c.noalias() += a * b;
if(!r.isApprox(c)) std::cerr << "Warning, your parallel product is crap!\n\n";
}
#endif
#ifdef HAVE_BLAS
BenchTimer tblas;
BENCH(tblas, tries, rep, blas_gemm(a,b,c));
std::cout << "blas cpu " << tblas.best(CPU_TIMER)/rep << "s \t" << (double(m)*n*p*rep*2/tblas.best(CPU_TIMER))*1e-9 << " GFLOPS \t(" << tblas.total(CPU_TIMER) << "s)\n";
std::cout << "blas real " << tblas.best(REAL_TIMER)/rep << "s \t" << (double(m)*n*p*rep*2/tblas.best(REAL_TIMER))*1e-9 << " GFLOPS \t(" << tblas.total(REAL_TIMER) << "s)\n";
#endif
BenchTimer tmt;
BENCH(tmt, tries, rep, gemm(a,b,c));
std::cout << "eigen cpu " << tmt.best(CPU_TIMER)/rep << "s \t" << (double(m)*n*p*rep*2/tmt.best(CPU_TIMER))*1e-9 << " GFLOPS \t(" << tmt.total(CPU_TIMER) << "s)\n";
std::cout << "eigen real " << tmt.best(REAL_TIMER)/rep << "s \t" << (double(m)*n*p*rep*2/tmt.best(REAL_TIMER))*1e-9 << " GFLOPS \t(" << tmt.total(REAL_TIMER) << "s)\n";
#ifdef EIGEN_HAS_OPENMP
if(procs>1)
{
BenchTimer tmono;
//omp_set_num_threads(1);
Eigen::setNbThreads(1);
BENCH(tmono, tries, rep, gemm(a,b,c));
std::cout << "eigen mono cpu " << tmono.best(CPU_TIMER)/rep << "s \t" << (double(m)*n*p*rep*2/tmono.best(CPU_TIMER))*1e-9 << " GFLOPS \t(" << tmono.total(CPU_TIMER) << "s)\n";
std::cout << "eigen mono real " << tmono.best(REAL_TIMER)/rep << "s \t" << (double(m)*n*p*rep*2/tmono.best(REAL_TIMER))*1e-9 << " GFLOPS \t(" << tmono.total(REAL_TIMER) << "s)\n";
std::cout << "mt speed up x" << tmono.best(CPU_TIMER) / tmt.best(REAL_TIMER) << " => " << (100.0*tmono.best(CPU_TIMER) / tmt.best(REAL_TIMER))/procs << "%\n";
}
#endif
return 0;
}
|
