Antialiasing doesn't appear to work for png output (2.12)

[durack1]

Using the code example in #266 I have been trying to generate clean *.png output, however it seems that antialiasing is not working for this output format.

Example output seems to show slightly different file sizes, but the line output in the examples doesn't appear to improve with a high antialiasing setting - there are jaggies in all of the png outputs, whereas the pdf output is much sharper.

antialiasing = 1 (44.7K)

antialiasing = 8 (45.5K)

antialiasing = 64 (45.5K)

pdf output
171102_antiAliasing_1_.pdf

[doutriaux1]

@aashish24 @danlipsa that's the bug I was just talking about with you.

[doutriaux1]

@danlipsa this is a big problem, I can confirm at least with mesa no antialiasing ever seems to be applied.

[scottwittenburg]

Here is the code snippet from the issue mentioned above:

import vcs

x = vcs.init(bg=False,geometry=(900,900))

for anti in [0,8,64]:
    x.setantialiasing(anti)
    x.plot([1,3,5,7,10,24,36,3,6,8,9,10])
    x.png('_'.join(['antialias',str(anti),'.png']))
    x.pdf('_'.join(['antialias',str(anti),'.pdf']))
    x.clear()

I have been able to confirm that this results in three pngs with no antialiasing. However, it seems to be related to the clearing of the canvas in the last line of the loop. If instead I change the code to for anti in [64, 8, 0]: (just reverse the order of the multisamples to be set), then I get three pngs with antialiasing.

Additionally, if I change the code as follows:

import vcs

for anti in [0,8,64]:
    x = vcs.init(bg=False,geometry=(900,900))
    x.setantialiasing(anti)
    x.plot([1,3,5,7,10,24,36,3,6,8,9,10])
    x.png('_'.join(['antialias',str(anti),'.png']))
    x.pdf('_'.join(['antialias',str(anti),'.pdf']))

Then I get one png which appears with no antialiasing, but two with it. For this reason, I believe there may be a bug with the clear method.

For now I may leave this alone, as once we update our pipelines to use Context2D (thus allowing us to use the OpenGL2 backend), the clear issue may go away on it's own. If not, it would make sense to debug the clear issue further at that point.

ping @doutriaux1 @danlipsa @aashish24

[scottwittenburg]

So now that I have a working debug build of VTK to test vtk-cdat issues, I have been able to confirm that although you can set MultiSamples on a vtkRenderWindow at any time, it is only used when the thing is actually initialized.

This explains why we can't seem to change the value in the loop, but only the first value chosen is used. It would also mean that there is probably not a bug in the canvas clear function.

Does this jive with your understanding of how multisample antialiasing works in VTK @danlipsa and @aashish24?

[danlipsa]

Maybe its a bug in VTK and you need to call Modified so that it can re-set the proper OpenGL functions?
I wonder if OpenGL2 works correctly. If that's the case we can focus on the OpenGL2 conversion.

[scottwittenburg]

Here's a vanilla vtkpython script to reproduce the issue outside of vcs with a plot that looks similar to the reproduction scripts listed in comments above:

import vtk

def buildPolyData():
    N = 12

    pts = [
        [0,   0.0,  0.0, 0.0],
        [1,   3.0,  3.0, 0.0],
        [2,   6.0,  6.0, 0.0],
        [3,   9.0,  7.0, 0.0],
        [4,  12.0, 10.0, 0.0],
        [5,  15.0, 24.0, 0.0],
        [6,  18.0, 36.0, 0.0],
        [7,  21.0,  3.0, 0.0],
        [8,  24.0,  6.5, 0.0],
        [9,  27.0,  8.0, 0.0],
        [10, 30.0,  9.0, 0.0],
        [11, 33.0, 10.0, 0.0]
    ]

    points = vtk.vtkPoints()
    points.SetNumberOfPoints(N)

    for i in range(N):
        points.SetPoint(*pts[i])

    lines = vtk.vtkCellArray()

    for i in range(N - 1):
        lines.InsertNextCell(2)
        lines.InsertCellPoint(i)
        lines.InsertCellPoint(i + 1)

    poly = vtk.vtkPolyData()
    poly.SetPoints(points)
    poly.SetLines(lines)

    return poly


pd = buildPolyData()

polyMapper = vtk.vtkPolyDataMapper()
polyMapper.SetInputData(pd)
polyMapper.Update()

polyActor1 = vtk.vtkActor()
polyActor1.SetMapper(polyMapper)

ren1 = vtk.vtkRenderer()
ren1.AddActor(polyActor1)
ren1.SetBackground(0.1, 0.2, 0.4)
ren1.ResetCamera()

renWin = vtk.vtkRenderWindow()
renWin.AddRenderer(ren1)
renWin.SetSize(800, 800)

for anti in [0,8,64]:
    renWin.SetMultiSamples(anti)
    print('RenderWindow multisamples is now: %d' % renWin.GetMultiSamples())
    renWin.Render()

    # Just gives us a chance to look at the render window
    stuff = raw_input('Please enter something:\n')
    print('You entered %s' % stuff)

    pd.Modified()

I have run this with both OpenGL and OpenGL2 builds and got the same results. That is, the first value of multisamples is the only one used in all three Render() calls. So that Modified() call upstream of the rendering pipeline doesn't seem to help.

I have noticed in the vtkXOpenGLRenderWindow code (at least under OpenGL2), that a call to WindowRemap should re-trigger the creation of the framebuffer and visual. I have tried this out, and with the OpenGL2 backend, it causes the new value of MultiSamples to be used and then behaves as we would like/expect. However, calling WindowRemap under the OpenGL backend results in X Error of failed request: GLXBadDrawable.

So you're probably right we should focus on the OpenGL2 conversion.