Upbge: Bug: Alpha sort slower in UPBGE than in Blender 2.78c

[Akira1San]

blend: http://pasteall.org/blend/index.php?id=45937

In the material tab, game setings switch to alpha sort.

[DCubix]

?

[Akira1San]

I tested in 0.1.0 and 011, they give big performance while 012 gives less.
Can anyone else test?

[panzergame]

I confirm the slow down, in 0.1.1 i have 0.75ms for the rasterizer whereas i have 1.21 in 0.1.2.

[panzergame]

Commit : 5718bff90f5158a66dcc9f3255855d2c6e6ebd10

[panzergame]

The issue is that we need to flush the index array of the sorted polygons every frame, this wasn't done before, anyway the way i designed it in 0.1.2 is not so good.
What i propose now is to replace:

    if (material->IsZSort() && rasty->GetDrawingMode() >= RAS_Rasterizer::RAS_SOLID) {
        m_mesh->SortPolygons(this, args.m_trans * MT_Transform(m_meshUser->GetMatrix()));
        m_displayArrayBucket->SetPolygonsModified(rasty);
    }

By:

    if (material->IsZSort() && rasty->GetDrawingMode() >= RAS_Rasterizer::RAS_SOLID) {
                int *index = m_displayArrayBucket->GetIndexMap()
        m_mesh->SortPolygons(this, args.m_trans * MT_Transform(m_meshUser->GetMatrix(), index));
        m_displayArrayBucket->FlushIndexMap();
    }

And use the glMap/UnmapBuffer function to avoid extra copy of the indices.

[youle31]

First bad commit: https://github.com/UPBGE/blender/commit/a147c7e7082548792b664ed7122fd0f619e2113c

(before 900 fps, after 300 fps)

[panzergame]

@Akira1San : Could you test with latest master please ?

[Akira1San]

Tested, the same low fps.

[youle31]

About this issue, I did tests in last master.
I have 500 fps with the test file. I remarked that the issue came from storage->GetIndexMap() called each frame.

So do you think we can cache the buffer returned by GetIndexMap() ?

When I cache the buffer in the test file I have 800+ fps

[Akira1San]

Made a test: http://pasteall.org/pic/index.php?id=117793

Looks great, even fixed the lamp slow down as can be seen from the images(I mean when switching the hemp to point) However, theres still a regression as can be seen, compared to bf bge. I think, that regression comes after .010 version.

[youle31]

In fact I don't know in what case alpha sort option is used, I don't have good test files in my files so I can't tell if I don't break the feature with the branch I pushed. This was just to show the impact on performances of GetIndexMap.

@Akira_san: In the image you shared, the shading looks not the same in BF bge/upbge. Is this normal?

Else is there is still a performance difference this can also be explained with NodeUpdateGS called each frame for ApplyRotation in the code.

[Akira1San]

I see.
The difference of light is from changing the lamp from hemi to point.

[panzergame]

The vertex memory model changed in 0.2.4 and many modifications happened.

I retested with this file:
ge_mesh_sort_test.zip

I get for UPBGE 0.2.4 (086255edf5c8abbf298a92d95fde71b4c4e79452) 26.0ms for rasterizer,
for blender 37.5ms.

[panzergame]

@Akira1San, @youle31 : Can you confirm the test please ?

[youle31]

I have these results in BF master and upbge master:

BF:

upbge:

[Akira1San]

heres my test; image 023 - 024: http://pasteall.org/pic/f730f61d5b7931c55a4be51e61a45dab
In bf bge it shows better result. - pic bf bge: http://pasteall.org/pic/167cba229ccef184efb46f1ce2f67e9a