The issue was that if the evaluated object doesn't have an active
layer, the canvas is offset by twice the objects position.
The reason was that the code scaled the whole transformation
matrix by two to match the code in 4.2, but then had to overwrite
the location of the transformation again to counteract the scaling.
The fix is to not counteract the scaling and just scale the 3x3 part
of the matrix instead. This way we can remove the part afterwards
that writes to the location of the transform.
Pull Request: https://projects.blender.org/blender/blender/pulls/130454
NOTE: This also required some changes to Cycles code itself, who is now
directly including `BKE_image.hh` instead of declaring a few prototypes
of these functions in its `blender/utils.h` header (due to C++ functions
names mangling, this was not working anymore).
Pull Request: https://projects.blender.org/blender/blender/pulls/130174
Turns out the sampling module was not initialized when
computing lighting which resulted in undefined shadow location.
In some cases even producing NaN position which would make the
shadow disappear.
Basically, anything returned by `sampling_rng_nD_get` was returning
garbage, and on some occasion, it could be a negative number and
produce NaN if passed into `sample_disk` or any `sqrt`.
ASAN did not catch it because the uninitialized data load
was done on the GPU.
I ran the render tests and, minor difference aside, they all match.
#### Follow Up:
This raises the concern that we have no assertion for correctly
using some modules. At the very least, we should have some
mechanism to catch uninitialized buffer. Investigating this
was a pain au chocolat.
Pull Request: https://projects.blender.org/blender/blender/pulls/130150
The crash was caused by the texture wrapper being out of date.
This happened after workbench finished rendering AntiAliasing and
going into blit-result-only mode. Doing so was causing a use after
free caught by ASAN.
But this only partially fix the bug as workbench render then becomes
uninitialized (black most likely) after dragging the window to the
second monitor. This is because the TAA texture gets recreated since
the new viewport resolution doesn't match the old one. But this
recreation does not tag as a viewport update so the TAA stays in
blit-result-only mode.
Detecting the viewport resolution change at the engine init level
seems the safest option for now, and it fixes the remaining issue.
To be backported to 4.2.
Pull Request: https://projects.blender.org/blender/blender/pulls/130043
When the light volume probe cannot be allocated a pool of a different
size is used. A message is displayed to the user. This PR improves the
message by adding the previous size as well.
Ref #129889
Pull Request: https://projects.blender.org/blender/blender/pulls/129963
First this make the wire batch use the same set of drawing that the main
gpencil batch uses (as they share the same VBO data, they need to use
the same drawings).
Then we skip drawing the onion frames by nullifying the indices in the
index buffer. A better way would be to skip these strokes or drawing
but that can be done later and seems more bug prone.
Pull Request: https://projects.blender.org/blender/blender/pulls/129920
The issue was that the masks for the current frame would be
rendered for a different frame.
Unfortunately, we can't easily render masks for the onion
skinned frames correctly at the moment.
The fix makes it so that we render the mask only for the current frame.
Pull Request: https://projects.blender.org/blender/blender/pulls/129878
When using a lot of instances the requested and needed attributes
are merged. This process uses a lock even when no work needs to be
done.
By early exiting the merging process when no work needs to be done
the performance of navigating 60k cubes went from 17.5 fps to 18.3 fps.
Detected when researching #126391.
Pull Request: https://projects.blender.org/blender/blender/pulls/129791
This would happen if the viewport with large enough
and a tile resolution of 1:1 was choosen.
This changes the fallback behavior from simply clamping
the resolution (which did break a lot of math downstream),
to simply finding the next largest tile size that fits
the hardware requirements.
Adding a dummy storage buffer to the classification shader
seems to fix the issue on Qualcomm drivers (WoA).
The workaround is added to the force workaround option to
allow other platforms to test the fix.
Rel #122837
Pull Request: https://projects.blender.org/blender/blender/pulls/129857
The issue is caused by the new way to compute the projection
matrix that tries to recreate it from view3d properties.
This is needed for proper overscan support.
However, this breaks for VR as the view3d is only partially
setup.
Skip the setup in VR and remove overscan for VR to avoid any
possible issues with it.
This fix avoid changing the behavior of the
`ED_view3d_draw_offscreen` which is exposed from our pyGPU
API. The pyGPU API is kind of broken if using custom
projection matrix + overscan, but that's a different issue.
Fix#125456
Pull Request: https://projects.blender.org/blender/blender/pulls/129813
This was caused by uninitialized values at border. The correct
number of bricks were reserved but only the unpadded volume
was uploaded. Which was not touching the data of the border
bricks if the size was divisible by `IRRADIANCE_GRID_BRICK_SIZE - 1`.
Fix#127215
Pull Request: https://projects.blender.org/blender/blender/pulls/129810
For C/C++ doc-strings should be located in headers,
move function comments into the headers, in some cases merging
with existing doc-strings, in other cases, moving implementation
notes into the function body.
Introduced with 5fff95f519.
There were several issues:
* The indices into `edit_points_vflag` are incorrect. They need to be offset
by the start offset of the drawing.
* The code was writing to `edit_points_vflag` even if the layer is locked.
But the size of the `edit_points_vflag` buffer is not counting locked layers.
Pull Request: https://projects.blender.org/blender/blender/pulls/129625
Introduced by 5fff95f519.
The issue was that the `edit_points_vflag` buffer is of size `total_points_num`
which is based on the number of control points not evaluated positions.
The fix is to use `points_by_curve` instead of `points_by_curve_eval`
to get the right start and end indices.
Pull Request: https://projects.blender.org/blender/blender/pulls/129526
Since GPv3 shader currently uses particle strand/points shader, we
need to modify that slightly to display the grease pencil overlays.
This adds the missing `vflag` attribute to the edit gpencil batch.
Co-authored-by: Clément Foucault <foucault.clem@gmail.com>
Pull Request: https://projects.blender.org/blender/blender/pulls/128116
Caused by 364d62e59b.
In draw mode, we want to make sure to show vertex colors by default
because users can draw with them. Otherwise this can lead to unexpected
behavior in solid shading mode when vertex colors used to not be
rendered in draw mode.
The previous fix tried to do this by enforcing the `V3D_SHADING_VERTEX_COLOR`
mode in draw mode. But this has the side effect that the material
stroke and fill color are overriden with pure white.
To fix the issue correctly, simply ensure that `vert_col_opacity` in
`grease_pencil_layer_cache_add` is set to 1.0f. This means that
vertex colors and material colors are shown just like in the render.
Also resolves#128680.
Pull Request: https://projects.blender.org/blender/blender/pulls/129054
If a `DrawGroup` contained both inverted and non-inverted scale
the command generate shader would output the `resource_id`
content at conflicting indices. This is because the number of
instances stored inside the `DrawGroup` are the original
count before multiview. Actually, only `start` was taking the
multi-view count into account.
We cannot modify the value on CPU otherwise it would increase
the instance count for each submission. So the fix is to
pass the view count to the command generate shader and
multiply the instance count where needed.
Fix#128085
Pull Request: https://projects.blender.org/blender/blender/pulls/128854
This is likely caused by local_ray_up being degenerate with
very small radii. This is a temporary fix and should be revisited
later.
The issue is that the real fix is likely to have a higher
performance cost.
Fix#124636
When using OpenGL on Intel ARC the driver reports a max 3d allowed size
of 2048. The volume probe will create a texture that doesn't fit in this
dimension when selecting a probe size of 512 or 1024 MB.
This PR will reshape the volume pool atlas texture until it found a shape
that is optimal and fit on the device. When reshaping selects a different
pool size a warning message will be displayed as it might change the
visual quality.
When reshaping the smallest row size will be selected in order
to improve the occupancy. Reshaping will only happens when a
different setting is set in the `Performance->Memory->Light Probes Volume Pool`.
NOTE: Needs to be backported to 4.2
Pull Request: https://projects.blender.org/blender/blender/pulls/128877
The error message related to shadow updates in `eevee_shadow.cc`
currently contains a typo:
`"Error: Too many shadow updates, some shadow might be incorrect."`
This sentence should use the plural form of "shadows" to correctly
describe the context.
Fixing this typo ensures clarity and correctness in the error message,
providing developers and users with the appropriate feedback when
encountering shadow update issues.
Pull Request: https://projects.blender.org/blender/blender/pulls/128865
Object without lightprobe visibility should still
cast shadows during baking. They should only not
bounce indirect lighting.
This is more visible now that shadow linking is supported.
Fix#128812
When using OpenGL on Intel ARC the driver reports a max 3d allowed size
of 2048. The volume probe will create a texture that doesn't fit in this
dimension when selecting a probe size of 512 or 1024 MB.
This PR will reshape the volume pool atlas texture until it found a shape
that is optimal and fit on the device. When reshaping selects a different
pool size a warning message will be displayed as it might change the
visual quality.
When reshaping the smallest row size will be selected in order
to improve the occupancy. Reshaping will only happens when a
different setting is set in the `Performance->Memory->Light Probes Volume Pool`.
Pull Request: https://projects.blender.org/blender/blender/pulls/128518
Removes unused GPv2 functions in blenkernel.
Notes:
- Functions for layer masks are still in use, but annotations never
have layer masks in the first place. Would be good to remove the data
structures so we can remove the functions too.
- Some multi-frame edit functions are also still nominally used, but
multi-frame editing is not an active feature for annotations. This
should also be removed.
Pull Request: https://projects.blender.org/blender/blender/pulls/128709
Between 0bfd5e3536
and b1cbd9c889
the main branch is incorrectly processing the file
`draw_debug_info.hh` as GLSL and does some string
preprocessing on it. But the output filename matches
the name of the header source file used for compiling
the gpu module. This file not having been updated
since a long time doesn't get copied from the source
folder when switching to other branch and make compilation
fail.
In order to avoid breaking the buildbot longer, we
rename the incriminating file to force recreate it
when building the release branch.
This renames the mode identifiers to be consistent with e.g. the context mode identifiers and other names used for the new Grease Pencil.
For `object.mode`:
* `PAINT_GPENCIL` -> `PAINT_GREASE_PENCIL`
* `SCULPT_GPENCIL` -> `SCULPT_GREASE_PENCIL`
* `VERTEX_GPENCIL` -> `VERTEX_GREASE_PENCIL`
* `WEIGHT_GPENCIL` -> `WEIGHT_GREASE_PENCIL`
For the internal `ob->mode` flag:
* `OB_MODE_PAINT_GPENCIL_LEGACY` -> `OB_MODE_PAINT_GREASE_PENCIL`
* `OB_MODE_SCULPT_GPENCIL_LEGACY` -> `OB_MODE_SCULPT_GREASE_PENCIL`
* `OB_MODE_VERTEX_GPENCIL_LEGACY` -> `OB_MODE_VERTEX_GREASE_PENCIL`
* `OB_MODE_WEIGHT_GPENCIL_LEGACY` -> `OB_MODE_WEIGHT_GREASE_PENCIL`
Resolves#127374.
Pull Request: https://projects.blender.org/blender/blender/pulls/128604
The Legacy Cryptomatte node doesn't work in GPU execution mode if
Precision is set to Auto. That's because the colors picked from the Pick
layer might be in half precision and thus will not match the colors in
the Cryptomatte layers. This is due to the compositor using the
context's precision for Viewer outputs as opposed to the precision of
the image that actually needs to be viewed in the Viewer node.
To fix this, we set the Viewer node precision to be the precision of its
input, that way, the Cryptomatte pick layer will be output in full
precision as intended.
Pull Request: https://projects.blender.org/blender/blender/pulls/128495
There was a bug where vertex colors were not shown in solid view in
draw or vertex paint mode. The check to use `V3D_SHADING_VERTEX_COLOR`
was outdated and checked the wrong mode flags.
The fix makes sure that we use `V3D_SHADING_VERTEX_COLOR` in
draw and vertex paint mode.
Pull Request: https://projects.blender.org/blender/blender/pulls/128581
The armature shape drawing was broken on NVIDIA. The reason is that not
all the elements in of the geometry shader out stages are written to. On
NVIDIA platforms this leads to not storing all the output data.
This could be that geometry shader on those plaforms write directly to
the final location in GPU memory. Other platforms might keep the data in
a local registry and copy it to the final location when emitting the
vertex.
Blender 4.4 might not have this issue as overlay next will remove the
need of the geometry shader.
From Khronos GLSL spec:
```
GS code writes all of the output values for a vertex, then calls EmitVertex().
This tells the system to write those output values to where ever it is that
output vertices get written. After calling this function, all output variables
contain undefined values. So you will need to write to them all again before
emitting the next vertex (if there is a next vertex).
```
Pull Request: https://projects.blender.org/blender/blender/pulls/128578
On NVIDIA drivers it is not allowed to pass buffers directly from the
interface to a function as the qualifiers do not match. This is commonly
fixed by first storing it in a local variable to remove the qualifiers
and then pass it to the function that need the object.
Pull Request: https://projects.blender.org/blender/blender/pulls/128571
This adds feature parity with Cycles regarding light and shadow liking.
Technically, this extends the GBuffer header to 32 bits, and uses
the top bits to store the object's light set membership index.
The same index is also added to `ObjectInfo` in place of padding bytes.
For shadow linking, the shadow blocker sets bitmask is stored per
tilemap. It is then used during the GPU culling phase to cull objects
that do not belong to the shadow's sets.
Co-authored-by: Clément Foucault <foucault.clem@gmail.com>
Pull Request: https://projects.blender.org/blender/blender/pulls/127514
