Node group inputs should behave like cycles now.
----
We create dummy nodes that contains the default values for the nodegroup
inputs and link them to the sockets. This way the uniform buffer gathering
function can read them.
But we also need to disconnect all the internal links to sockets that have
hidden values. Theses sockets must not get the values from the nodegroup
input sockets if there is no link to them otherwise we get broken results
(i.e.: normals for a bsdf node).
This implementation is a bit hacky but match cycles pretty close.
If pixel size is not enabled, it will use the geom shader to
compute distances between vertices. This will have a cost.
Implementation is a bit hacky in gpu_codegen to make the geom shader works
in an optional manner.
Terms get/set don't make much sense when casting values.
Name macros so the conversion is obvious,
use common prefix for easier completion.
- GET_INT_FROM_POINTER -> POINTER_AS_INT
- SET_INT_IN_POINTER -> POINTER_FROM_INT
- GET_UINT_FROM_POINTER -> POINTER_AS_UINT
- SET_UINT_IN_POINTER -> POINTER_FROM_UINT
This happens on NVidia GPUs, using more textures than the maximum allowed
by the gl will NOT trigger a linking issue (maybe because of bindless
texture implementation?).
So in this case we manually count the number of samplers per shader stage
and compare it against the GL limit. We discard the shader if the sampler
count is too high. This shows the user something is wrong with the shader.
This supports meshes and hairs too. Matches cycles output.
This adds barycentric coords to the GPUBuiltin enum which will also be used
for the wireframe node.
A lot of code in codegen was used by previous GLSL system.
Now most of it is not used due to all the things being written from scratch
in 2.8 and we can clean most of it.
As a side not this make the system a bit less flexible (no support for
cubemaps, preview image, etc...) but can be extended again.
* Remove GPU_link_changed which is unused.
* Remove all GPU link function that are not used anymore.
* GPU_uniform_buffer is now GPU_uniform.
* GPU_texture_ramp is now GPU_color_band.
* GPU_uniform is now GPU_constant.
This lower the use of texture samplers slots and let users use more real
textures in their shaders.
This patch also make the ramp texture 16 bit floating point. Meaning you
can now use value greater than one in your color ramps.
With the limit of 128 colorband per shader (a color band being either a
color ramp, a wavelength node or a curve node (and maybe wavelength node in
the future)).
Only drawback with the current implementation is that it does not remove
colorband from pruned GPUNodes but it shouldn't really matter in practice.
This should fix T56010
This was likely caused by some sort of race condition where the drawing
thread would request the state of the shader before the shader has been
compiled.
Authored by Pavel Rudko (PavelRudko) on Thu, Jun 21, 10:41 AM.
Original description:
I've encounterd into a problem with blender 2.8.
See the example screenshot (scene, shader setup). For some materials it's not enough to calculate hash based on fragment code and defines. In some cases attribute names may change. And if we change uv set name, for example, vertex shader is not recompiled, and another attribute name inside it is used.
This is part of the work needed to refactor the material parameters update.
Now the gpupass cache is polled before adding the gpumaterial to the
deferred compilation queue.
We store gpupasses in a single linked list grouped based on their hashes.
This is not the most efficient way but it can be improved upon later.
Now they are properly converted to Linear space before interpolation.
Since the only way to get vertex color in eevee and cycles is via the
attribute node with the CD_AUTO_FROM_NAME flag, we have to know at binding
time which type of buffer will be connected to this auto input.
We store this information inside the batch cache (together with the according
uniform name) and pass it as uniform to the shader which does conversion if
needed. The same shader can then be reused to draw another mesh with
different auto layers configuration.
Because:
- Less redundancy.
- Better suffixes.
Also a few modification to GPU_texture_create_* to simplify the API:
- make the format explicit to the texture creation process.
- remove the component count as it's specified in the GPUTextureFormat.
Brecht authored this commit, but he gave me the honours to actually
do it. Here it goes; Blender Internal. Bye bye, you did great!
* Point density, voxel data, ocean, environment map textures were removed,
as these only worked within BI rendering. Note that the ocean modifier
and the Cycles point density shader node continue to work.
* Dynamic paint using material shading was removed, as this only worked
with BI. If we ever wanted to support this again probably it should go
through the baking API.
* GPU shader export through the Python API was removed. This only worked
for the old BI GLSL shaders, which no longer exists. Doing something
similar for Eevee would be significantly more complicated because it
uses a lot of multiplass rendering and logic outside the shader, it's
probably impractical.
* Collada material import / export code is mostly gone, as it only worked
for BI materials. We need to add Cycles / Eevee material support at some
point.
* The mesh noise operator was removed since it only worked with BI
material texture slots. A displacement modifier can be used instead.
* The delete texture paint slot operator was removed since it only worked
for BI material texture slots. Could be added back with node support.
* Not all legacy viewport features are supported in the new viewport, but
their code was removed. If we need to bring anything back we can look at
older git revisions.
* There is some legacy viewport code that I could not remove yet, and some
that I probably missed.
* Shader node execution code was left mostly intact, even though it is not
used anywhere now. We may eventually use this to replace the texture
nodes with Cycles / Eevee shader nodes.
* The Cycles Bake panel now includes settings for baking multires normal
and displacement maps. The underlying code needs to be merged properly,
and we plan to add back support for multires AO baking and add support
to Cycles baking for features like vertex color, displacement, and other
missing baking features.
* This commit removes DNA and the Python API for BI material, lamp, world
and scene settings. This breaks a lot of addons.
* There is more DNA that can be removed or renamed, where Cycles or Eevee
are reusing some old BI properties but the names are not really correct
anymore.
* Texture slots for materials, lamps and world were removed. They remain
for brushes, particles and freestyle linestyles.
* 'BLENDER_RENDER' remains in the COMPAT_ENGINES of UI panels. Cycles and
other renderers use this to find all panels to show, minus a few panels
that they have their own replacement for.
This is mostly to avoid re-compilation when using undo/redo operators.
This also has the benefit to reuse the same GPUShader for multiple materials using the same nodetree configuration.
The cache stores GPUPasses that already contains the shader code and a hash to test for matches.
We use refcounts to know when a GPUPass is not used anymore.
I had to move the GPUInput list from GPUPass to GPUMaterial because it's containing references to the material nodetree and cannot be reused.
A garbage collection is hardcoded to run every 60 seconds to free every unused GPUPass.
ntree_shader_relink_displacement is creating a transient node that does not have a correct original to point to.
In this case we revert to constant uniform.
It should behave like cycles.
Even if not efficient at all, we still do the same create - draw - free process that was done in the old viewport to save vram (maybe not really the case now) and not care about simulation's GPU texture state sync.
For users that means you can tweak shaders in the nodetree and things
are way faster. This is a huge improvement, particularly in
systems that have no shader cache.
From the code perspective it means we are no longer re-compiling the
shader every time a value is tweaked in the UI. We are using uniforms
for those values.
It would be slow to add that many uniforms for all the shaders. So
instead we are using UBO (Uniform Buffer Objects).
This fixes the main issue of T51467. However GWN_shaderinterface_create() still
needs to be improvedi. When opening a .blend all shaders are compiled once, so
optimizing it will bring a measurable impact.
========================================================================
NOTE: This breaks update of Cycles material upon nodetree nodes
tweaking. It will be fixed separately by depsgraph, once tackling T51925
(Animated Eevee values slowdown).
The idea is to make Depsgraph update more granular. The XXX TODO in
rna_nodetree.c will be tackled at that time as well.
========================================================================
Reviewers: sergey, brecht, fclem
Differential Revision: https://developer.blender.org/D2739
This allow specialized shaders to redefine the closure interface to fit their needs.
For instance, Volumetric closure needs to pass more than one vec4 (absorption vec3, scattering vec3, anisotropy float).
