Now the callbacks are setup for each debug context.
The formating has been reworked to be less verbose and make errors
and warnings stand out from the notifications.
Errors are most of the time sufficiently explicit in their message.
This also remove the support for AMD_debug_output which is 10 years old.
This is related to the Vulkan port T68990.
This is part of the Vulkan backend task T68990.
This is mostly a cleanup, however, there is a small change:
We don't use a special Vertex Array binding function for Immediate
anymore and just reuse the one for batches.
This might create a bit more state changes but this could be fixed
easily if it causes perf regression.
# Conflicts:
# source/blender/gpu/intern/gpu_context.cc
This was affecting Mesa drivers as well as AMD pro driver. But it
might have been noticeable on other config too.
This was introduced by rBa9f2ebb21508.
This way it is way clearer what each viewport state is. There is
no more save and reset. The scissor test is also saved per
framebuffer.
The only rule to remember is that the viewport state (size and
origin) is reset for both the viewport and scissor when a texture
is attached or detached from an attachment slot.
This is related to the Vulkan port T68990.
This is a full cleanup of the Framebuffer module and a separation
of OpenGL related functions.
There is some changes with how the default framebuffers are handled.
Now the default framebuffers are individually wrapped inside special
GLFrameBuffers. This make it easier to keep track of the currently bound
framebuffer state and have some specificity for operations on these
framebuffers.
Another change is dropping the optimisation of only configuring the
changed attachements during framebuffers update. This does not give
any benefits and add some complexity to the code. This might be brought
back if it has a performance impact on some systems.
This also adds support for naming framebuffers but it is currently not
used.
This is in preparation of the Framebuffer GL backend.
This is a just changing types and moving some code.
No logic is changed... almost... it just removes the context attach.
i.e: `gpu_context_add/remove_framebuffer()`
This is not needed for now and was even disabled in release.
This is part of T68990.
The CPP Shader class does not initialize the interface attribute.
What will crash when deleting the shader.
Reviewed By: Clément Foucault
Differential Revision: https://developer.blender.org/D8740
In recent refactoring {a9f2ebb21508} an issue was introduced that the
opengl rasterizer would be disabled when only writing to a stencil
buffer.
This fix adds stencil writing to the write mask and set it. This makes
the write map not evaluate to GPU_WRITE_NONE and the rasterizer will be
enabled in `GLStateManager::set_write_mask`.
Reviewed By: Clément Foucault
Differential Revision: https://developer.blender.org/D8743
The CPP Shader class does not initialize the interface attribute.
What will crash when deleting the shader.
Reviewed By: Clément Foucault
Differential Revision: https://developer.blender.org/D8740
This fix a GL_INVALID_VALUE error on startup due to 0.0f max line width.
Also moves the max anisotropy filter to the sampler creation.
This reduces code fragmentation.
This should not cause any problem since the depth test is required
in order to draw to the depth buffer and this is not altered by
this change.
To be on the safe side, we still restor the mask after altering it.
This is quite embarassing... it was returning the base instance instead of
the correct vao. No wonder that it was causing crash and at most drawing
issues.
This replace `GPU_clear()` by `GPU_clear_color()` and `GPU_clear_depth()`.
Since we always set the clear value before clearing, it is unecessary
to track the clear color state.
Moreover, it makes it clearer what we clear the framebuffer to.
This changes the state of occlusion queries quite a bit. Now scissors is
explicitely disabled and we enabled color write.
I still don't understand why we now need this. This patch is just trial and
error on an affected setup. Note that on the same computer, renderdoc
was not able to capture the regression (the regression did not manifest
during capture).
Regression likely introduced by rB5f414234ddea
