EEVEE Film accumulation workaround for Metal/Intel iGPUs.
On Metal the Intel iGPUs do not support image read write
on array textures. However this limitation doesn't show
any artifacts when using the compute shader.
This PR is a work around that uses the film_comp shader
to process the film samples, but uses a separate film_copy_frag
shader to read the result and copy them to the frame buffer.
I deliberately didn't include the fix to the film_frag shader
as that would change the read/write resources and could lead
to performance issues for other platforms. Writable resources
are typically slower compared to read only resources.
Some code needed to be duplicated (and not added to `*_lib.glsl`)
as compilers would still raise compilation errors due to imageStore/Load
on incompatible resource access.
The Metal/Intel iGPU is also marked to have limited support as
raytracing and probes still produces big artifacts.
This workaround can be tested on any platform just by setting
`use_compute_ = true` in `Film::sync`
Related to #122361
Pull Request: https://projects.blender.org/blender/blender/pulls/123330
The previous code was using matrix multiplication to
get the local thickness to world thickness.
The correct way is to multiply the local thickness
by the scale of the object (length of each columns
of the object_to_world matrix).
When using particle system the generated coordinates where not set.
The reason is that a loose resource handle was created for particle
system that didn't contain enough information to calculate orcos.
This PR will create the object info from its parent object.
Pull Request: https://projects.blender.org/blender/blender/pulls/123307
This bias was added when the pixel radius estimation
computation was still broken. This resulted in a shadow
bias way too big in general.
Also fix shadow pixel size being clampped to the wrong
LOD level.
Fix#123295
A strip was wrongly deemed to be opaque (and thus strips below it
were skipped from rendering), if the strip content was opaque, but
it gained transparency via presence of a Mask modifier.
While at it, I also noticed that a strip could have been wrongly deemed
opaque when it had a Multiplier < 1.0 with "multiply alpha" option
checked under strip Color settings. So fixed that too.
The whole logic of that factored out into is_opaque_alpha_over function
to make the call site clearer.
Pull Request: https://projects.blender.org/blender/blender/pulls/123303
For very low frequency sounds (the ones where waveform drawing goes into
"line mode"), either of min or max can go outside of -1..+1 range.
The code was assuming that only max could be above +1, and only min
could be below -1. Fix that, and also make "line mode" also apply
red "clipped" color when clipping happens.
Pull Request: https://projects.blender.org/blender/blender/pulls/123246
EEVEE has a transparent render pass that renders materials where
the render mode is set to blended. This was introduced when EEVEE-Next
was still in development, but was never fully backported.
This PR adds the missing pieces.
Pull Request: https://projects.blender.org/blender/blender/pulls/123298
In `sound_bake_animation_exec`, the camera needs to be ensured each
frame so that marker camera switch can still give correct stereo/surround
effect during render.
Pull Request: https://projects.blender.org/blender/blender/pulls/123294
While implementing preview snapping, I noticed an oddity with snapping
videos to view bounds, and found that the preview window's total v2d has
half a pixel of extra padded width.
The bug was introduced in fb5e2f5610, which added `roundf` calls to
numerous areas in the blender code to replace instances of adding `0.5`
and then casting to an integer.
One mistake in the patch was fixed shortly after in d122911d10, but the
other mistake has been unfixed until now. This fixes the glitchy
behavior with preview snapping and is also small enough to add to 4.2
LTS.
Pull Request: https://projects.blender.org/blender/blender/pulls/123279
The issue is that timeline snapping code does not set the transform
`values` to a rounded frame delta, instead, it adds the rounded distance
between current source and target points to the unrounded `values`
float, leaving it unrounded.
Since snapping code is only run on some `transform_modal` calls due to
"time base quirky code" in `transform_snap_mixed_apply`, sometimes
frames briefly show mouse input added to `values` causing the strip to
get transformed away from the snap target before the snap code is run
once again.
To fix this, set transform `values` to the exact rounded snap target
frame diff (like in node snapping) instead of expecting it to get
properly rounded for every call. Do this by changing
`tsnap->snap_source[0]` to the original snap source on `transformInit`,
not the updated snap source.
Pull Request: https://projects.blender.org/blender/blender/pulls/123217
The fix is to support interleaving the drawing of zones and frames.
The draw order is determined by the width of the zones/frames.
Larger areas are drawn before smaller ones. This makes sure that
everything is as visible as possible.
The zone border outlines are still drawn on top of all frames.
Solving that is technically more challenging, because we don't want
to use transparency for zone backgrounds because that results in
many possible mixed colors which we want to avoid. Fortunately,
drawing the outlines on top seems to be quite useful anyway.
The issue was that we used `from_orthonormal_axes` which obviously expects that
the axes are orthonormal. However, with custom curve normals this is not always the
case. Math wise, an additional normalization is necessary because the cross product
is not automatically normalized anymore.
This change also means that `point_matrix` may have a shearing component. But I
think that's fine here because the matrix is only immediately applied on vertex positions.
This shouldn't affect any case where the normal and tangent are orthonormal.
Pull Request: https://projects.blender.org/blender/blender/pulls/123238
Do this by:
- Removing depth and depth test.
- Scale the sphere to the covered pixels size.
- Use infinite orthographic matrix
We can still improve this by scaling the sphere
bigger depending on the accuracy at its position.
Fix#123165
`booleans_mix_calc` was incorrect. While it computed the correct result in
many common cases, under some specific circumstances, at was wrong.
Whether it was wrong also depended on how the range was split up for
multi-threading which is not deterministic.
The function was used in `Mesh::normals_domain` which then also returned
the wrong domain in some cases.
Avoid warning the user that there is a geometry, if it does not have any points/data.
Such geometry components shouldn't just be removed generally. because they
still contain information (such as the set of existing attributes).
The threshold is way too high as the light power gets very small.
It still make sense to include the power here because
some lights might be disabled for some BSDF types, and
skipping the LTC eval code can be beneficial.
So reducing the threshold to very low value near FLT_MIN
to fix the issue.
Fix#123193
The issue was a floating point precision issue
between the number of the cycle and the time
within the cycle (`cycle` and `cyct`).
Due to that the `cycle` would advance to the
next whole number while the `cycle time`
is still slightly under that.
This is fixed by calculating the `cycle` with double precision.
This has a measurable performance
impact (for `fcm_cycles_time`)
| Before | After |
| - | - |
| 39ns | 44ns |
For fcurves with a cycle modifier,
this code is run at every evaluate call.
The only time when that would be an issue is
in the graph editor, where there is an evaluate
call for roughly every pixel for curves that have a modifier.
However even in that scenario the performance
is the same within run to run variance (for `graph_draw_curves`)
| Before | After |
| - | - |
| 91565ns | 91430ns |
Pull Request: https://projects.blender.org/blender/blender/pulls/123222
