SpeciaLUT

Run-time choosing of template specializations using compile-time lookup-tables (LUT). Simply put: compile all possible states of a template function, but execute the optimal one at run-time.

Use cases

Heavy functions can have many branching scenarios inside loops. Some branches may be dependent on conditions that do not change during execution of the function. Branching prediction works good on CPUs, but in any case, conditions and jumping takes time. Moreover, GPUs do not use branching prediction.

Therefore, the safest thing is that the non-used parts of the code are not even there -- that the compiler deletes them. This is possible by introducing compile-time conditions, if the conditions are immutable during the function execution.

How it works

User defines number of states that each condition (template parameter) has.
Compiler compiles all possible specializations of the function, and stores the function pointers in a lookup table (LUT).
When user wants to execute the function with run-time parameters, the optimal one is executed.

How to use it

Requirements: C++20 compiler (enabled with -std=c++20)

Test: Run CMake as usual, or open the project in an IDE.

Set up: Copy specialut.hpp into your project and include it.

Example:

Include the library:

#include "specialut.hpp"

A function run has both compile-time and run-time parameters:

template<bool condition, int state>
void run(double some_param) {

    while (loop_condition) {

        if constexpr (condition) {
            /* ... */
        } else {
            /* ... */
        }

        if constexpr (state == SOME_ENUM) {
            /* ... */
        }
    }

}

Make an instance of Chooser class that requires: the template function and number of states for each template parameter:

SpeciaLUT::Chooser<TABULATE(run), 2, 3> test;

Find the optimal function based on the run-time conditions (first brackets), pass other parameters and execute the function (second brackets).

test(runtime_bool, int_state)(double_parameter);

There is a construct for CUDA kernels as well (see main.cpp for an example), which is used as:

SpeciaLUT::CudaChooser<TABULATE(some_cuda_kernel), 2, 3> test;

Try online in Compiler Explorer

Be aware of ...

slow compilation of large functions. This thing compiles all possible specializations. E.g. if you have 3 parameters with 3 states, it will compile 3^3 = 27 functions.

Tested on

Clang >= 13.0
Clang 12 requires that auto table = TABULATE(run) is done, and table passed to Chooser
GCC >= 10.1
GCC 9.4 requires -std=c++2a instead of -std=c++20
MSVC >= 19.30 requires that constexpr auto table = TABULATE(run) is done, and table passed to Chooser

Roadmap

Non-member functions (DONE)
Member functions (DONE)
Lambdas and functors (DONE)
C++17 workarounds (DONE, checkout branch cxx17)
CUDA kernels (DONE)
HIP kernels (maybe)

Caball009/SpeciaLUT