| 19 | __attribute__ ((noinline)) void benchVec(VectorXf& a, VectorXf& b, VectorXf& c); |
| 20 | |
| 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 | |