PATCH: Enable Back Face Culling (bfc)

[rsbx]

EDIT: I've been using the patch for a while without issues so I think it's ready. See later comments for the latest version.

Original comment follows: WARNING: The attached patch is incomplete in that LeoCAD does not track, or even parse, BFC CLIP and BFC NOCLIP meta commands and this patch does not change that. It's not an issue for most parts in the current LDraw parts library but a quick grep shows that there are still some with a BFC NOCLIP meta command and 3 with a BFC NOCERTIFY meta command.

[rsbx]

Attached is a patch that enables working Back Face Culling in LeoCAD.

bfc.patch.txt

I didn't (and still don't) really know how all of the mesh related code works, but it seems to work.

Limitations:

• No method to turn off without recompiling.
• Lsynth objects don't get culling. I wasn't able to follow the code enough to be comfortable making it work with Lsynth objects.
• I'm not a C++ programmer so I hope the code doesn't melt your eyes.

Fun extra: The studs texture on the base grid were actually on the bottom of the base grid plane so I inverted the base grid plane and enabled BFC on it. Now the studs texture is only visible from above.

[rsbx]

Attached is an updated patch to fix a normal inversion issue in the previous patch. Normals should now be the same as LeoCAD without this BFC patch.

bfc.patch.txt

[rsbx]

@leozide Bump

[rsbx]

Updated patch: bfc.patch.txt

• Ability to enable/disable BFC while LeoCAD is running by changing the conditional lines preference item.
• Now applied to faded and transparent parts also.

[leozide]

I've looked at this today:

• Looks like you removed the mirror matrix check, I don't see how that's correct now. How did you check that the code is working?
• The opaque meshes should be sorted by BFC state after texture.
• 3626cpx3 isn't drawing correctly.

[rsbx]

• Mirror matrix check: lc_meshloader.cpp lines 1703-1722. Examining the rendering of a sampling of parts; particularly ones with BFC NOCERTIFY, BFC NOCLIP, BFC INVERTNEXT, and left/right wedges & wings for mirroring checks.
• I don't understand enough of the meshes data structure & use to know where/how to do this. Is this for correctness or efficiency? Either way, point me at what I need to change. For context: to get it to the current state, I worked backwards from lcscene::DrawOpaqueMeshes
• Doesn't appear to be an OFFICIAL part. Does it render properly with LDView with the BFC colors options enabled?

[rsbx]

Updated patch:

bfc.patch.txt

[leozide]

I've checked in a change to print draw times (see lcViewWidget::paintGL), do you mind turning it on and running with and without this patch a couple of times?

On both my computers it was about 5% slower on the CPU and made no difference on the GPU as I expected (I tried BFC in the past and had the same results so I never looked into it again). My desktop was GPU bound so it didn't matter much but my laptop is CPU bound so it's slightly slower.

I tested with the tower bridge model from the OMR.

[rsbx]

I'll be a couple of hours before I can do any repeatable testing but a quick test of commit c613ff6, "Added simple draw timer", with the timer enabled is continously consuming both CPUs in the VM even with no model loaded. It also segfaults when I quit LeoCAD.

I run LeoCAD from an xterm so a simple printf() without the QT machinery might be sufficient and/or better for this profiling task.

[rsbx]

To be clear, the rendering time improvement I see with the BFC patch is when I rotating an already loaded model via the mouse. It does not appear to affect (perceptively) the time to load a model.

[rsbx]

Here are the the master and master+BFC timings for 3 different models on both the VM (llvm based 3D) and the host (Intel integrated graphics based 3D):

vm_timings.txt host_timings.txt

The last several updates in each test are due to just hitting the H (home view) shortcut.

I may not have been "twisting" the c11-color model fast enough during the host master+bfc timing run.

I used a simpler timing patch that did not try to consume all my CPUs:

`#ifndef LC_PROFILE_DRAW mView->OnDraw();

else

    struct timespec ts0, ts1;
    unsigned long elapsed;

    clock_gettime(CLOCK_MONOTONIC, &ts0);
    mView->OnDraw();
    clock_gettime(CLOCK_MONOTONIC, &ts1);

    elapsed = (ts1.tv_sec - ts0.tv_sec) * 1000000000 + ts1.tv_nsec - ts0.tv_nsec;
    printf("%lu\n", elapsed); 

endif

`

[rsbx]

Playing with commit b6acab71 and the on-screen timer display, the GPU values, of the 3 large models I tested, with the BFC patch is about 75% of the GPU values without the patch. This is true for both the VM and the host.

The CPU value of the largest model with the BFC patch is about 75% of the CPU value without the patch on the VM; on the HOST it's about 93%. The other models appeared similar but I wasn't watching the CPU value too closely.

[rsbx]

Rebased onto commit 1a64a787:

Back-Face-Culling.patch.txt

[rsbx]

Rebased onto commit 613bc91c:

0001-Back-Face-Culling.patch.txt

[rsbx]

Rebased onto commit 35ba24fb:

0001-Back-Face-Culling.patch.txt

[rsbx]

This updated patch, still relative to 35ba24f, now also enables back face culling for synthesized parts (lc_synth.cpp):

0001-Back-Face-Culling.patch.txt

[rsbx]

Rebased onto f24317e

0001-Back-Face-Culling.patch.txt