1+ #include < iostream>
2+ #include < chrono>
3+ #include < torch/torch.h>
4+ #include < Eigen/Dense>
5+
6+ int main () {
7+ // Define large, mid and small matrix dimensions
8+ int large_rows = 1000 ;
9+ int large_cols = 1000 ;
10+ int mid_rows = 100 ;
11+ int mid_cols = 100 ;
12+ int small_rows = 10 ;
13+ int small_cols = 10 ;
14+
15+ // Create large, mid and small torch::Tensors and Eigen matrices
16+ torch::Tensor large_torch_tensor = torch::rand ({large_rows, large_cols});
17+ torch::Tensor mid_torch_tensor = torch::rand ({mid_rows, mid_cols});
18+ torch::Tensor small_torch_tensor = torch::rand ({small_rows, small_cols});
19+ Eigen::MatrixXd large_eigen_matrix = Eigen::MatrixXd::Random (large_rows, large_cols);
20+ Eigen::MatrixXd mid_eigen_matrix = Eigen::MatrixXd::Random (mid_rows, mid_cols);
21+ Eigen::MatrixXd small_eigen_matrix = Eigen::MatrixXd::Random (small_rows, small_cols);
22+
23+ // Get the default device (CPU or CUDA if available)
24+ torch::Device device = torch::cuda::is_available () ? torch::kCUDA : torch::kCPU ;
25+ // Print the default device
26+ std::cout << " Default device: "
27+
28+ // Move torch::Tensors to the default device
29+ large_torch_tensor = large_torch_tensor.to (device);
30+ mid_torch_tensor = mid_torch_tensor.to (device);
31+ small_torch_tensor = small_torch_tensor.to (device);
32+
33+ // Benchmark large torch::Tensor matrix multiplication
34+ auto start_large_torch_matmul = std::chrono::high_resolution_clock::now ();
35+ // pre-allocate memory for the result
36+ torch::Tensor large_torch_result = torch::empty ({large_rows, large_rows}, device);
37+ large_torch_result = torch::matmul (large_torch_tensor, large_torch_tensor.transpose (0 , 1 ));
38+ auto end_large_torch_matmul = std::chrono::high_resolution_clock::now ();
39+ std::chrono::duration<double > elapsed_large_torch_matmul = end_large_torch_matmul - start_large_torch_matmul;
40+
41+ // Benchmark large Eigen matrix multiplication
42+ auto start_large_eigen_matmul = std::chrono::high_resolution_clock::now ();
43+ Eigen::MatrixXd large_eigen_result = large_eigen_matrix * large_eigen_matrix.transpose ();
44+ auto end_large_eigen_matmul = std::chrono::high_resolution_clock::now ();
45+ std::chrono::duration<double > elapsed_large_eigen_matmul = end_large_eigen_matmul - start_large_eigen_matmul;
46+
47+ // Benchmark mid torch::Tensor matrix multiplication
48+ auto start_mid_torch_matmul = std::chrono::high_resolution_clock::now ();
49+ // pre-allocate memory for the result
50+ torch::Tensor mid_torch_result = torch::empty ({mid_rows, mid_rows}, device);
51+ mid_torch_result = torch::matmul (mid_torch_tensor, mid_torch_tensor.transpose (0 , 1 ));
52+ auto end_mid_torch_matmul = std::chrono::high_resolution_clock::now ();
53+ std::chrono::duration<double > elapsed_mid_torch_matmul = end_mid_torch_matmul - start_mid_torch_matmul;
54+
55+ // Benchmark mid Eigen matrix multiplication
56+ auto start_mid_eigen_matmul = std::chrono::high_resolution_clock::now ();
57+ Eigen::MatrixXd mid_eigen_result = mid_eigen_matrix * mid_eigen_matrix.transpose ();
58+ auto end_mid_eigen_matmul = std::chrono::high_resolution_clock::now ();
59+ std::chrono::duration<double > elapsed_mid_eigen_matmul = end_mid_eigen_matmul - start_mid_eigen_matmul;
60+
61+ // Benchmark small torch::Tensor matrix multiplication
62+ auto start_small_torch_matmul = std::chrono::high_resolution_clock::now ();
63+ // pre-allocate memory for the result
64+ torch::Tensor small_torch_result = torch::empty ({small_rows, small_rows}, device);
65+ small_torch_result = torch::matmul (small_torch_tensor, small_torch_tensor.transpose (0 , 1 ));
66+ auto end_small_torch_matmul = std::chrono::high_resolution_clock::now ();
67+ std::chrono::duration<double > elapsed_small_torch_matmul = end_small_torch_matmul - start_small_torch_matmul;
68+
69+ // Benchmark small Eigen matrix multiplication
70+ auto start_small_eigen_matmul = std::chrono::high_resolution_clock::now ();
71+ Eigen::MatrixXd small_eigen_result = small_eigen_matrix * small_eigen_matrix.transpose ();
72+ auto end_small_eigen_matmul = std::chrono::high_resolution_clock::now ();
73+ std::chrono::duration<double > elapsed_small_eigen_matmul = end_small_eigen_matmul - start_small_eigen_matmul;
74+
75+ // Benchmark small torch::Tensor inverse
76+ auto start_small_torch_inverse = std::chrono::high_resolution_clock::now ();
77+ // pre-allocate memory for the result
78+ torch::Tensor small_torch_inverse = torch::empty ({small_rows, small_cols}, device);
79+ small_torch_inverse = torch::inverse (small_torch_tensor);
80+ auto end_small_torch_inverse = std::chrono::high_resolution_clock::now ();
81+ std::chrono::duration<double > elapsed_small_torch_inverse = end_small_torch_inverse - start_small_torch_inverse;
82+
83+ // Benchmark small Eigen matrix inverse
84+ auto start_small_eigen_inverse = std::chrono::high_resolution_clock::now ();
85+ Eigen::MatrixXd small_eigen_inverse = small_eigen_matrix.inverse ();
86+ auto end_small_eigen_inverse = std::chrono::high_resolution_clock::now ();
87+ std::chrono::duration<double > elapsed_small_eigen_inverse = end_small_eigen_inverse - start_small_eigen_inverse;
88+
89+ // Print results
90+ std::cout << " Large Torch::Tensor matrix multiplication time: " count () << " seconds"
91+ std::cout << " Large Eigen matrix multiplication time: " count () << " seconds"
92+ std::cout << " Mid Torch::Tensor matrix multiplication time: " count () << " seconds"
93+ std::cout << " Mid Eigen matrix multiplication time: " count () << " seconds"
94+ std::cout << " Small Torch::Tensor matrix multiplication time: " count () << " seconds"
95+ std::cout << " Small Eigen matrix multiplication time: " count () << " seconds"
96+ std::cout << " Small Torch::Tensor inverse time: " count () << " seconds"
97+ std::cout << " Small Eigen matrix inverse time: " count () << " seconds"
98+
99+ return 0 ;
100+ }
101+
102+ // PC specs:
103+ // - CPU: 13th Gen Intel® Core™ i7-13620H × 16
104+ // - GPU: NVIDIA GeForce RTX 4060
105+ // ```
106+ // $ ./test_torch_eigen
107+ // Default device: cuda
108+ // Large Torch::Tensor matrix multiplication time: 0.0187716 seconds
109+ // Large Eigen matrix multiplication time: 0.121851 seconds
110+ // Mid Torch::Tensor matrix multiplication time: 0.00302689 seconds
111+ // Mid Eigen matrix multiplication time: 0.000178552 seconds
112+ // Small Torch::Tensor matrix multiplication time: 0.00202544 seconds
113+ // Small Eigen matrix multiplication time: 2.229e-06 seconds
114+ // Small Torch::Tensor inverse time: 0.0824715 seconds
115+ // Small Eigen matrix inverse time: 7.489e-06 seconds
116+ // ```
0 commit comments