As I had mentioned in my past blog about the Mandelbrot Research Repo “I’ve refactored this piece of code like 6140275936420759697 times…”, so I guess that makes it “6140275936420759697” plus one.
This time around I finalized the implementation for a Canvas class that collects all the pixels you want and writes to a file you give it. By making it a class it is now much easier to work on the underlying byte buffer without having to get your hands dirty.
template <size_t width, size_t height> class Canvas;
The only downside to this implementation is the fact that this class will get allocated on the stack so for any render of 1.5k+ pixels in both dimensions ends up overflowing most computer’s stacks nowadays. The way to get around this is by declaring the canvas object as a static duration object so that it gets stored into storage and not the actual stack.
static auto canvas = Canvas<2000, 2000>();
One cool thing about this implementation is the way you can use lambdas to modify the pixels isnide the byte buffer.
For example, with the new construct(auto op) method you can simply pass a function pointer (though realistically, you would only pass a lambda) and the construct method applies that function to each pixel in the canvas.
auto mandelbrot = [](std::complex<long double> coordinate) {
const auto max_iteration = 20000ul;
auto x = 0.0L;
auto y = 0.0L;
auto x_squared = 0.0L;
auto y_squared = 0.0L;
auto iteration = 0u;
while (std::islessequal((x_squared + y_squared), 4) &&
(iteration < max_iteration)) {
y = fma(2 * x, y, coordinate.imag());
x = x_squared - y_squared + coordinate.real();
x_squared = x * x;
y_squared = y * y;
iteration++;
}
return Color(iteration);
};
canvas.construct(mandelbrot);
This can then be parallelized using OpenMP with a range-based for-loop.
auto construct(auto op) -> void
{
#pragma omp parallel for
for (size_t idx = 0; idx < width * height; idx++)
pixels[idx] = op(coordinate(idx));
}
One final fix that’s pending is being able to do histogram coloring. This can (probably) be easily implemented by just changing the return Color(iteration) part at the end of that lambda. But I haven’t found a convenient way of doing that yet. Furthermore, it would be good to support different scaled images. The rendered image for this implementation breaks if width != height.