Implement OpenCL kernel generator #1342

Description

Now we have separate kernels for simple operations (such as matrix addition or transposition). These kernels do low amount of computation, compared to amouunt of data transfered from global GPU memory. So vast majority of time spent executing such kernels is needed for data transfers, instead of actual computations.

Usually when multiple simple operations are executed one after another they could be combined into a single kernel so that data is loaded only once. Less data transfers result in faster execution.

We should combine multiple operations on matrix_cl into expressions in similar way as Eigen does it. When evaluated or assigned to another matrix_cl an expression generates OpenCL kernel source, compiles and executes it.

Envisioned pull requests:

• basic kernel generator with matrices, scalars and binary element-wise operations (+, -, *, /)
• unary functions
• submatrix blocks
• selction operator
• comparisons and logical operators
• reductions(sum, min, max)
• multiple results in a single kernel
• broadcasting
• transposition
• more?

Example

The code below should result in execution of a single kernel.

matrix_cl<double> a;
matrix_cl<double> b;
matrix_cl<double> c;
matrix_cl<double> d = a + b - 3 * c;

Expected Output

Expressions of simple matrix operations on GPUs are faster. Creating simple OpenCL kernels does not require writing OpenCL source and can be done by non-GPU developers.

Current Version:

v2.20.0