1
2 #include <iostream>
3 #include <Eigen/Core>
4 #include <bench/BenchTimer.h>
5 using namespace Eigen;
6
7 #ifndef SIZE
8 #define SIZE 50
9 #endif
10
11 #ifndef REPEAT
12 #define REPEAT 10000
13 #endif
14
15 typedef float Scalar;
16
17 __attribute__ ((noinline)) void benchVec(Scalar* a, Scalar* b, Scalar* c, int size);
18 __attribute__ ((noinline)) void benchVec(MatrixXf& a, MatrixXf& b, MatrixXf& c);
19 __attribute__ ((noinline)) void benchVec(VectorXf& a, VectorXf& b, VectorXf& c);
20
main(int argc,char * argv[])21 int main(int argc, char* argv[])
22 {
23 int size = SIZE * 8;
24 int size2 = size * size;
25 Scalar* a = internal::aligned_new<Scalar>(size2);
26 Scalar* b = internal::aligned_new<Scalar>(size2+4)+1;
27 Scalar* c = internal::aligned_new<Scalar>(size2);
28
29 for (int i=0; i<size; ++i)
30 {
31 a[i] = b[i] = c[i] = 0;
32 }
33
34 BenchTimer timer;
35
36 timer.reset();
37 for (int k=0; k<10; ++k)
38 {
39 timer.start();
40 benchVec(a, b, c, size2);
41 timer.stop();
42 }
43 std::cout << timer.value() << "s " << (double(size2*REPEAT)/timer.value())/(1024.*1024.*1024.) << " GFlops\n";
44 return 0;
45 for (int innersize = size; innersize>2 ; --innersize)
46 {
47 if (size2%innersize==0)
48 {
49 int outersize = size2/innersize;
50 MatrixXf ma = Map<MatrixXf>(a, innersize, outersize );
51 MatrixXf mb = Map<MatrixXf>(b, innersize, outersize );
52 MatrixXf mc = Map<MatrixXf>(c, innersize, outersize );
53 timer.reset();
54 for (int k=0; k<3; ++k)
55 {
56 timer.start();
57 benchVec(ma, mb, mc);
58 timer.stop();
59 }
60 std::cout << innersize << " x " << outersize << " " << timer.value() << "s " << (double(size2*REPEAT)/timer.value())/(1024.*1024.*1024.) << " GFlops\n";
61 }
62 }
63
64 VectorXf va = Map<VectorXf>(a, size2);
65 VectorXf vb = Map<VectorXf>(b, size2);
66 VectorXf vc = Map<VectorXf>(c, size2);
67 timer.reset();
68 for (int k=0; k<3; ++k)
69 {
70 timer.start();
71 benchVec(va, vb, vc);
72 timer.stop();
73 }
74 std::cout << timer.value() << "s " << (double(size2*REPEAT)/timer.value())/(1024.*1024.*1024.) << " GFlops\n";
75
76 return 0;
77 }
78
benchVec(MatrixXf & a,MatrixXf & b,MatrixXf & c)79 void benchVec(MatrixXf& a, MatrixXf& b, MatrixXf& c)
80 {
81 for (int k=0; k<REPEAT; ++k)
82 a = a + b;
83 }
84
benchVec(VectorXf & a,VectorXf & b,VectorXf & c)85 void benchVec(VectorXf& a, VectorXf& b, VectorXf& c)
86 {
87 for (int k=0; k<REPEAT; ++k)
88 a = a + b;
89 }
90
benchVec(Scalar * a,Scalar * b,Scalar * c,int size)91 void benchVec(Scalar* a, Scalar* b, Scalar* c, int size)
92 {
93 typedef internal::packet_traits<Scalar>::type PacketScalar;
94 const int PacketSize = internal::packet_traits<Scalar>::size;
95 PacketScalar a0, a1, a2, a3, b0, b1, b2, b3;
96 for (int k=0; k<REPEAT; ++k)
97 for (int i=0; i<size; i+=PacketSize*8)
98 {
99 // a0 = internal::pload(&a[i]);
100 // b0 = internal::pload(&b[i]);
101 // a1 = internal::pload(&a[i+1*PacketSize]);
102 // b1 = internal::pload(&b[i+1*PacketSize]);
103 // a2 = internal::pload(&a[i+2*PacketSize]);
104 // b2 = internal::pload(&b[i+2*PacketSize]);
105 // a3 = internal::pload(&a[i+3*PacketSize]);
106 // b3 = internal::pload(&b[i+3*PacketSize]);
107 // internal::pstore(&a[i], internal::padd(a0, b0));
108 // a0 = internal::pload(&a[i+4*PacketSize]);
109 // b0 = internal::pload(&b[i+4*PacketSize]);
110 //
111 // internal::pstore(&a[i+1*PacketSize], internal::padd(a1, b1));
112 // a1 = internal::pload(&a[i+5*PacketSize]);
113 // b1 = internal::pload(&b[i+5*PacketSize]);
114 //
115 // internal::pstore(&a[i+2*PacketSize], internal::padd(a2, b2));
116 // a2 = internal::pload(&a[i+6*PacketSize]);
117 // b2 = internal::pload(&b[i+6*PacketSize]);
118 //
119 // internal::pstore(&a[i+3*PacketSize], internal::padd(a3, b3));
120 // a3 = internal::pload(&a[i+7*PacketSize]);
121 // b3 = internal::pload(&b[i+7*PacketSize]);
122 //
123 // internal::pstore(&a[i+4*PacketSize], internal::padd(a0, b0));
124 // internal::pstore(&a[i+5*PacketSize], internal::padd(a1, b1));
125 // internal::pstore(&a[i+6*PacketSize], internal::padd(a2, b2));
126 // internal::pstore(&a[i+7*PacketSize], internal::padd(a3, b3));
127
128 internal::pstore(&a[i+2*PacketSize], internal::padd(internal::ploadu(&a[i+2*PacketSize]), internal::ploadu(&b[i+2*PacketSize])));
129 internal::pstore(&a[i+3*PacketSize], internal::padd(internal::ploadu(&a[i+3*PacketSize]), internal::ploadu(&b[i+3*PacketSize])));
130 internal::pstore(&a[i+4*PacketSize], internal::padd(internal::ploadu(&a[i+4*PacketSize]), internal::ploadu(&b[i+4*PacketSize])));
131 internal::pstore(&a[i+5*PacketSize], internal::padd(internal::ploadu(&a[i+5*PacketSize]), internal::ploadu(&b[i+5*PacketSize])));
132 internal::pstore(&a[i+6*PacketSize], internal::padd(internal::ploadu(&a[i+6*PacketSize]), internal::ploadu(&b[i+6*PacketSize])));
133 internal::pstore(&a[i+7*PacketSize], internal::padd(internal::ploadu(&a[i+7*PacketSize]), internal::ploadu(&b[i+7*PacketSize])));
134 }
135 }
136