devtools/eigen
1
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2026-04-10 11:34:33 +08:00
Code Issues Packages Projects Releases Wiki Activity

* add bench/benchVecAdd.cpp by Gael, fix crash (ei_pload on non-aligned)

Browse Source 
* introduce packet(int), make use of it in linear vectorized paths
  --> completely fixes the slowdown noticed in benchVecAdd.
* generalize coeff(int) to linear-access xprs
* clarify the access flag bits
* rework api dox in Coeffs.h and util/Constants.h
* improve certain expressions's flags, allowing more vectorization
* fix bug in Block: start(int) and end(int) returned dyn*dyn size
* fix bug in Block: just because the Eval type has packet access
  doesn't imply the block xpr should have it too.
This commit is contained in:
Benoit Jacob
2008-06-26 16:06:41 +00:00
parent 5b0da4b778
commit 25ba9f377c
23 changed files with 558 additions and 264 deletions
Download Patch File Download Diff File Expand all files Collapse all files

134
bench/benchVecAdd.cpp Normal file
View File

@@ -0,0 +1,134 @@
#include <Eigen/Core>
#include <bench/BenchTimer.h>
using namespace Eigen;
#ifndef SIZE
#define SIZE 50
#endif
#ifndef REPEAT
#define REPEAT 10000
#endif
typedef float Scalar;
__attribute__ ((noinline)) void benchVec(Scalar* a, Scalar* b, Scalar* c, int size);
__attribute__ ((noinline)) void benchVec(MatrixXf& a, MatrixXf& b, MatrixXf& c);
__attribute__ ((noinline)) void benchVec(VectorXf& a, VectorXf& b, VectorXf& c);
int main(int argc, char* argv[])
{
int size = SIZE * 8;
int size2 = size * size;
Scalar* a = ei_aligned_malloc<Scalar>(size2);
Scalar* b = ei_aligned_malloc<Scalar>(size2);
Scalar* c = ei_aligned_malloc<Scalar>(size2);
for (int i=0; i<size; ++i)
{
a[i] = b[i] = c[i] = 0;
}
BenchTimer timer;
timer.reset();
for (int k=0; k<3; ++k)
{
timer.start();
benchVec(a, b, c, size2);
timer.stop();
}
std::cout << timer.value() << "s " << (double(size2*REPEAT)/timer.value())/(1024.*1024.*1024.) << " GFlops\n";
for (int innersize = size; innersize>2 ; --innersize)
{
if (size2%innersize==0)
{
int outersize = size2/innersize;
MatrixXf ma = MatrixXf::map(a, innersize, outersize );
MatrixXf mb = MatrixXf::map(b, innersize, outersize );
MatrixXf mc = MatrixXf::map(c, innersize, outersize );
timer.reset();
for (int k=0; k<3; ++k)
{
timer.start();
benchVec(ma, mb, mc);
timer.stop();
}
std::cout << innersize << " x " << outersize << " " << timer.value() << "s " << (double(size2*REPEAT)/timer.value())/(1024.*1024.*1024.) << " GFlops\n";
}
}
VectorXf va = VectorXf::map(a, size2);
VectorXf vb = VectorXf::map(b, size2);
VectorXf vc = VectorXf::map(c, size2);
timer.reset();
for (int k=0; k<3; ++k)
{
timer.start();
benchVec(va, vb, vc);
timer.stop();
}
std::cout << timer.value() << "s " << (double(size2*REPEAT)/timer.value())/(1024.*1024.*1024.) << " GFlops\n";
return 0;
}
void benchVec(MatrixXf& a, MatrixXf& b, MatrixXf& c)
{
for (int k=0; k<REPEAT; ++k)
a = a + b;
}
void benchVec(VectorXf& a, VectorXf& b, VectorXf& c)
{
for (int k=0; k<REPEAT; ++k)
a = a + b;
}
void benchVec(Scalar* a, Scalar* b, Scalar* c, int size)
{
typedef ei_packet_traits<Scalar>::type PacketScalar;
const int PacketSize = ei_packet_traits<Scalar>::size;
PacketScalar a0, a1, a2, a3, b0, b1, b2, b3;
for (int k=0; k<REPEAT; ++k)
for (int i=0; i<size; i+=PacketSize*8)
{
a0 = ei_pload(&a[i]);
b0 = ei_pload(&b[i]);
a1 = ei_pload(&a[i+1*PacketSize]);
b1 = ei_pload(&b[i+1*PacketSize]);
a2 = ei_pload(&a[i+2*PacketSize]);
b2 = ei_pload(&b[i+2*PacketSize]);
a3 = ei_pload(&a[i+3*PacketSize]);
b3 = ei_pload(&b[i+3*PacketSize]);
ei_pstore(&a[i], ei_padd(a0, b0));
a0 = ei_pload(&a[i+4*PacketSize]);
b0 = ei_pload(&b[i+4*PacketSize]);
ei_pstore(&a[i+1*PacketSize], ei_padd(a1, b1));
a1 = ei_pload(&a[i+5*PacketSize]);
b1 = ei_pload(&b[i+5*PacketSize]);
ei_pstore(&a[i+2*PacketSize], ei_padd(a2, b2));
a2 = ei_pload(&a[i+6*PacketSize]);
b2 = ei_pload(&b[i+6*PacketSize]);
ei_pstore(&a[i+3*PacketSize], ei_padd(a3, b3));
a3 = ei_pload(&a[i+7*PacketSize]);
b3 = ei_pload(&b[i+7*PacketSize]);
ei_pstore(&a[i+4*PacketSize], ei_padd(a0, b0));
ei_pstore(&a[i+5*PacketSize], ei_padd(a1, b1));
ei_pstore(&a[i+6*PacketSize], ei_padd(a2, b2));
ei_pstore(&a[i+7*PacketSize], ei_padd(a3, b3));
// ei_pstore(&a[i+2*PacketSize], ei_padd(ei_pload(&a[i+2*PacketSize]), ei_pload(&b[i+2*PacketSize])));
// ei_pstore(&a[i+3*PacketSize], ei_padd(ei_pload(&a[i+3*PacketSize]), ei_pload(&b[i+3*PacketSize])));
// ei_pstore(&a[i+4*PacketSize], ei_padd(ei_pload(&a[i+4*PacketSize]), ei_pload(&b[i+4*PacketSize])));
// ei_pstore(&a[i+5*PacketSize], ei_padd(ei_pload(&a[i+5*PacketSize]), ei_pload(&b[i+5*PacketSize])));
// ei_pstore(&a[i+6*PacketSize], ei_padd(ei_pload(&a[i+6*PacketSize]), ei_pload(&b[i+6*PacketSize])));
// ei_pstore(&a[i+7*PacketSize], ei_padd(ei_pload(&a[i+7*PacketSize]), ei_pload(&b[i+7*PacketSize])));
}
}