Extremely poor VertexArray performance

[Flandan]

Hi, I am hitting a bit of a performance barrier using this library.

Using the sf.VertexArray I have done some quick and dirty benchmarks:
with N = 5000
sf.VertexArray of size N: ~15ms for a single window.draw() call
list of N sf.CircleShape objects: ~14ms for N window.draw() calls (for loop)

If I increase N, the time taken increases much faster for the list of CircleShapes as expected. For the vertex array the increase is marginal for up to N=50000. Great news!
Only problem is this weird baseline ~15ms that is used regardless. With any application we only have 16.7ms to maintain a constant 60 fps which leaves us with no budget for anything other than drawing. Given the VertexArray is really the only goto for batch-like drawing this poses a serious problem.

Is there some sort of workaround for this issue?

My quick (and very dirty) test:

from sfml import sf
import time

N = 5000

vertex_array = sf.VertexArray(sf.PrimitiveType.POINTS, N)
shape_list = [sf.CircleShape(5.0) for i in range(N)]

window = sf.RenderWindow(sf.VideoMode(1366, 768), "Benchmark Window")#, sf.Style.FULLSCREEN)

window.clear()

time.sleep(0.3)

time_initial = time.time()
window.draw(vertex_array)

time_va = time.time()
for shape in shape_list:
    window.draw(shape)
    
time_sa = time.time()

elapsedA = (time_va - time_initial)*1000.0
elapsedB = (time_sa - time_va)*1000.0

print("VertexArray: %f" % elapsedA, flush=True)
print("list of shapes: %f" % elapsedB, flush=True)

[Flandan]

Ok so after some more exploration of the examples in this repository I have come up with a workaround. By wrapping the VertexArray in a class that inherits from sf.Drawable and drawing that instead, the mystery 15ms seems to disappear and becomes < 1ms.

class VertDrawable(sf.Drawable):
    def __init__(self):
        sf.Drawable.__init__(self)
        self.points = sf.VertexArray(sf.PrimitiveType.POINTS)
        for i in range(0, LENGTH):
            self.points.append(sf.Vertex(sf.Vector2(0, 0), sf.Color(255, 255, 255)))

    def draw(self, target, states):
        target.draw(self.points, states)

This is the way it is done in the shader.py example. If anyone has a good explaination as to why it works this way I would love to understand more. I am guessing it is due to the way cython interacts with the c++ code but it seems odd, and inconsistent as it doesnt follow the way the c++ api behaves as far as I can tell.