mesytec-mnode/external/taskflow-3.8.0/doxygen/examples/matrix_multiplication.dox

namespace tf {

/** @page matrix_multiplication Matrix Multiplication

We study the classic problem, <em>2D matrix multiplication</em>. 
We will start with a short introduction about the problem and 
then discuss how to solve it parallel CPUs.

@tableofcontents

@section MatrixMultiplicationProblem Problem Formulation

We are multiplying two matrices, @c A (@c MxK) and @c B (@c KxN).
The numbers of columns of @c A must match the number of rows of @c B. 
The output matrix @c C has the shape of @c (MxN) 
where @c M is the rows of @c A and @c N the columns of @c B.
The following example multiplies a @c 3x3 matrix with a @c 3x2 matrix to derive
a @c 3x2 matrix.

@image html images/matrix_multiplication_1.png width=50%

As a general view, 
for each element of @c C we iterate a complete row of @c A 
and a complete column of @c B, 
multiplying each element and summing them.

@image html images/matrix_multiplication_2.png width=50%

We can implement matrix multiplication using three nested loops.

@code{.cpp}
for(int m=0; m<M; m++) {
  for(int n=0; n<N; n++) {
    C[m][n] = 0;
    for(int k=0; k<K; k++) {
      C[m][n] += A[m][k] * B[k][n];
    }
  }
}
@endcode

@section MatrixMultiplicationParallelPattern Parallel Patterns

At a fine-grained level, computing each element of @c C is independent of each other.
Similarly, computing each row of @c C or each column of @c C is also independent of one another.
With task parallelism, we prefer <em>coarse-grained</em> model 
to have each task perform rather large computation to amortize
the overhead of creating and scheduling tasks.
In this case, we avoid intensive tasks each working on only a single element.
by creating a task per row of @c C 
to multiply a row of @c A by every column of @c B.

@code{.cpp}
// C = A * B
// A is a MxK matrix, B is a KxN matrix, and C is a MxN matrix
void matrix_multiplication(int** A, int** B, int** C, int M, int K, int N) {
  tf::Taskflow taskflow;
  tf::Executor executor;
  for(int m=0; m<M; ++m) {
    taskflow.emplace([m, &] () {
      for(int n=0; n<N; n++) {
        for(int k=0; k<K; k++) {
          C[m][n] += A[m][k] * B[k][n];  // inner product
        }
      }
    });
  }
  executor.run(taskflow).wait();
}
@endcode

Instead of creating tasks one-by-one over a loop,
you can leverage Taskflow::for_each_index to create a <em>parallel-for</em> task.
A parallel-for task spawns a subflow to perform parallel iterations over the given range.

@code{.cpp}
// perform parallel iterations on the range [0, M) with the step size of 1
tf::Task task = taskflow.for_each_index(0, M, 1, [&] (int m) {
  for(int n=0; n<N; n++) {
    for(int k=0; k<K; k++) {
      C[m][n] += A[m][k] * B[k][n];
    }   
  }   
}); 
@endcode

Please visit @ref ParallelIterations for more details.

@section MatrixMultiplicationBenchmarking Benchmarking

Based on the discussion above, we compare the runtime of computing
various matrix sizes of @c A, @c B, and @c C between a sequential
CPU and parallel CPUs on a machine of 12 Intel i7-8700 CPUs at 3.2 GHz.

<div align="center">
| A         | B         | C         | CPU Sequential | CPU Parallel | 
| :-:       | :-:       | :-:       | :-:            | :-:          | 
| 10x10     | 10x10     | 10x10     | 0.142 ms       | 0.414 ms     | 
| 100x100   | 100x100   | 100x100   | 1.641 ms       | 0.733 ms     | 
| 1000x1000 | 1000x1000 | 1000x1000 | 1532  ms       | 504 ms       | 
| 2000x2000 | 2000x2000 | 2000x2000 | 25688 ms       | 4387 ms      | 
| 3000x3000 | 3000x3000 | 3000x3000 | 104838 ms      | 16170 ms     | 
| 4000x4000 | 4000x4000 | 4000x4000 | 250133 ms      | 39646 ms     | 
</div>

The speed-up of parallel execution becomes clean as we increase the problem
size. For example, at @c 4000x4000, the parallel runtime is 6.3 times faster
than the sequential runtime.

*/

}