This PR replaces submissions with pipeline barriers when generating
mipmaps. This would reduce the amount of submissions and improve
performance during mipmap generation as dependencies will be tracked
between commands.
Pull Request: https://projects.blender.org/blender/blender/pulls/114306
Goal is to reduce the number of command buffer flushes by tracking what is
happening in the different command queues. This is an initial step towards
advanced queue-ing strategies.
The new (intermediate) strategy records commands to different command
buffers based on what they do. There is a command buffer for data transfers,
compute pipelines and graphics pipelines.
When a compute command is recorded it ensures that all graphic commands
are finished. When a graphic command is recorded it ensures all compute
commands are finished. When a graphic or compute command is scheduled
all recorded data transfer commands are scheduled as well.
Some improvements are expected as multiple compute and data transfers
commands can now be scheduled at the same time and don't need to unbind
and rebind render passes. Especially when using EEVEE-Next which is
compute centric the performance change is visible for the user.
Pull Request: https://projects.blender.org/blender/blender/pulls/114104
PR Introduces GPU_storagebuf_sync_to_host as an explicit routine to
flush GPU-resident storage buffer memory back to the host within the
GPU command stream.
The previous implmentation relied on implicit synchronization of
resources using OpenGL barriers which does not match the
paradigm of explicit APIs, where indiviaul resources may need
to be tracked.
This patch ensures GPU_storagebuf_read can be called without
stalling the GPU pipeline while work finishes executing. There are
two possible use cases:
1) If GPU_storagebuf_read is called AFTER an explicit call to
GPU_storagebuf_sync_to_host, the read will be synchronized.
If the dependent work is still executing on the GPU, the host
will stall until GPU work has completed and results are available.
2) If GPU_storagebuf_read is called WITHOUT an explicit call to
GPU_storagebuf_sync_to_host, the read will be asynchronous
and whatever memory is visible to the host at that time will be used.
(This is the same as assuming a sync event has already been signalled.)
This patch also addresses a gap in the Metal implementation where
there was missing read support for GPU-only storage buffers.
This routine now uses a staging buffer to copy results if no
host-visible buffer was available.
Reading from a GPU-only storage buffer will always stall
the host, as it is not possible to pre-flush results, as no
host-resident buffer is available.
Authored by Apple: Michael Parkin-White
Pull Request: https://projects.blender.org/blender/blender/pulls/113456
Allow binding of framebuffers without sRGB to linear transform.
`GPU_framebuffer_bind_no_srgb`. This Patch removes color transform
artifacts in node, image and sequence editor.
When the framebuffer is an srgb framebuffer and it is bound without
the transformation, the SRGB textures are bound as UNORM variants.
As framebuffer, render pass and subpass recreation is ensured by
`VKCommandBuffer` we don't need to mark the framebuffer dirty at
this time. Later on we can optimize this by adding a state changed
detection for framebuffers and render passes.
Pull Request: https://projects.blender.org/blender/blender/pulls/113838
With the shift to GPU-driven rendering pipeline,
the SSBO vertex fetch paradigm used to
implement workbench shadows on Metal
instead of utilising the geometry shader
path no longer worked correctly.
This is because the draw submission
required vertex amplification up-front,
based on the expected output geometry
amount for a given input geometry.
This patch aims to resolve this
issue through addition of API to
enable the features within the
GPU driven pipeline.
Co-authored-by: Michael Parkin-White <mparkinwhite@apple.com>
Pull Request: https://projects.blender.org/blender/blender/pulls/113498
When drawing node sockets an immediate mode buffer is created that can
contain all the node sockets of the node. Only visible node sockets will
then be added. Vulkan assumed that all elements in the buffer needed to
be drawn, resulting using uninitialzed memory for drawing node sockets.
This resulted in very colorful and big artifacts rendered in the node
editor. This was detected during drawing of node sockets, but would have
been visible in other places as well.
Pull Request: https://projects.blender.org/blender/blender/pulls/113830
When Vulkan is started with validation layers the console is flooded.
Somewhere in the Vulkan validation layers or mesa driver (or the
combination) there is an issue where maxBufferSize is reported by the
driver to be 4GB, but the validation layers are reporting any buffer
size to be larger than 4GB.
For now we skip this message to be logged.
Pull Request: https://projects.blender.org/blender/blender/pulls/113652
With the shift to GPU-driven rendering pipeline,
the SSBO vertex fetch paradigm used to
implement workbench shadows on Metal
instead of utilising the geometry shader
path no longer worked correctly.
This is because the draw submission
required vertex amplification up-front,
based on the expected output geometry
amount for a given input geometry.
This WIP patch aims to resolve this
issue through addition of API to
enable the features within the
GPU driven pipeline.
Co-authored-by: Michael Parkin-White <mparkinwhite@apple.com>
Pull Request: https://projects.blender.org/blender/blender/pulls/113498
This PR adds workarounds for platforms that don't support `shaderOutputLayer`
or `shaderOutputViewportIndex`. Some NVIDIA laptop GPUs and ARM GPUs don't
have those device features.
The workaround uses the same approach as OpenGL. A geometry shader is injected
to emulate the feature.
For testing the workarounds they have also been connected to the
`--debug-gpu-force-workarounds` command line argument.
Fixes#113475 by implementing #113529
Pull Request: https://projects.blender.org/blender/blender/pulls/113605
Vulkan backend doesn't support subpass transition yet. Currently
shader_builder fails as when compiling shaders that use subpass
transitions.
This PR adds a dummy global variable to the GLSL so the compilation
continues.
Development of subpass would take some time as the current API doesn't
fit Vulkan.
Pull Request: https://projects.blender.org/blender/blender/pulls/113604
When using Immediate mode emulation the data is stored in a ring buffer.
As data isn't overwritten it is safe to call multiple draw commands in
a single command buffer.
This improves drawing performance when immediate mode is being used, it
mostly improves the readability of the GPU traces in renderdoc.
Pull Request: https://projects.blender.org/blender/blender/pulls/113482
Wireframe mode wasn't working and viewport had flickering artifacts.
Reason was that the render pass creation failed for its framebuffer as
the input data was filled with garbage for attachments that were unused.
Vulkan requires every attachment to be filled upto the highest used attachment
slot. This PR fills missing attachments with a dummy texture.
Pull Request: https://projects.blender.org/blender/blender/pulls/113141
This adds basic emulation of the subpass input feature
of vulkan and to a lower extend Raster Order Group on Metal.
This help test paths that might use this feature in the future
(like shadow rendering) on all platform and or simplify higher
level code for supporting older hardware.
This add clear description to the load/store ops and to the
new `GPUAttachementState`.
The OpenGL backend will correctly mask un-writable
attachments and will bind as texture readable attachments.
Even if possible by the vulkan standard, the GPU API prohibit
the read and write to the same attachment inside the same
subpass.
In the GL backend, this is implemented using `glTextureBarrier`
and `texelFetch` as it is described in the ARB_texture_barrier
extension.
https://registry.khronos.org/OpenGL/extensions/ARB/ARB_texture_barrier.txt
Pull Request: https://projects.blender.org/blender/blender/pulls/112051
During Blitting it is not allowed to read/write outside
the texture bounds. This patch ensures that the bounds
(called extents in Vulkan) are inside the expected range.
Pull Request: https://projects.blender.org/blender/blender/pulls/113050
This PR adds support for `GPU_framebuffer_blit` When used with `GPU_DEPTH_BIT`.
The challenge with is that not all GPUs support using a depth texture as a blit destination.
An AMD GPU doesn't support a depth buffer with stencil buffer as blit destination.
> NOTE: AMD GPUs don't support 24 bit unsigned normalized depth textures at all.
In all cases when we use depth blitting we are blitting the whole texture
and in stead we can use a texture copy.
A negative effect is that we need to unbind the framebuffer when copying
depth textures, but a positive effect is that we can use a data transfer function what should
theoretically be faster.
This should be revisited when we are investigating in areas to improve the
performance of the Vulkan backend.
Pull Request: https://projects.blender.org/blender/blender/pulls/112674
When initialy adding textures it wasn't possible to allocate during
initialization as its usage wasn't known. Bot Metal and Vulkan would do
a delayed allocation adding more checks when using textures.
Recently the behavior of setting texture usage was changed and now is
known when initializing textures. This PR makes advantage of this to
give correct feedback to the calling code if the texture can be created
and remove all the delayed allocating of textures.
Pull Request: https://projects.blender.org/blender/blender/pulls/112997
In Vulkan (and Metal) it is not possible to use a global as point
size. It needs to be set for each vertex when drawing points. Blender
has specialized shaders for that, but not all code respect those
shaders.
This PR will add an assert inside the vulkan backend when incorrect
usage of shaders are detected.
Pull Request: https://projects.blender.org/blender/blender/pulls/112906
Blender 4.0 requires OpenGL 4.3 which always support SSBO's.
Platforms that don't support enough SSBO bind points will be marked
as unsupported.
Users who start Blender on those platforms will be informed via a
dialog. This PR also updates the `--debug-gpu-force-workarounds`
to match our minimum requirements. Note that some bugs are still
there that should be solved in other PRs:
* Workbench only renders the object using a unit matrix this is because
there is a bug in the workaround for shader_draw_parameters
* Navigating with middle mouse button is not working. Unsure what the
cause is, but might be a missing feature check in the OpenGL backend.
Related to #112224
Pull Request: https://projects.blender.org/blender/blender/pulls/112572
In Vulkan the grid overlay wasn't rendering correctly. The reason was
a mis alignment in the std430 layout. This PR fixes the std430 layout
so the grid overlay (and perhaps also other sections) would work.
Root cause is that the reserve size if an nVec3 is different when used
as an array versus a single attribute. This was not taken into account
where the overlay flag was read from uninitialized memory.
Pull Request: https://projects.blender.org/blender/blender/pulls/112564
Array textures stores the array length inside the size of the image. In
vulkan the size mustn't contain the array length, but it should be set
as a separate parameter. This logic wasn't implemented when updating
the mipmap chain.
Patch also includes some smaller fixes where textures would not yet be
allocated on the device, but was expected to be.
Pull Request: https://projects.blender.org/blender/blender/pulls/112518
Uniform buffers can have data provided during construction, or as
attached data. Attached data should be uploaded during bind.
This PR implements `GPU_uniformbuf_create_from_list`.
Pull Request: https://projects.blender.org/blender/blender/pulls/112517
Resources can still be in use inside the vulkan command queue, when
they are destroyed. Especially as Vulkan doesn't allow resetting
bindings to nothing.
This PR will collect the resources that needs to be destroyed inside
the VKDevice and actually destroy the resources when on frame end.
The resources currently include:
* VkBuffer
* VkImage
* VkImageView
* VkFramebuffer
* VkRenderPass
Pull Request: https://projects.blender.org/blender/blender/pulls/112514
Previous implemetation had some limitations that were not working
for complex shaders. This limitations included:
* All bindings should be bound to a location < 16.
* Collision could happen when image locations where read that
overlapped with the texture locations.
This PR replaces the fixed binding locations with a dynamic vector.
When allocating locations for images the image offset is determined
that it is always clear if the requested locations is already
offsetted or not. This is done by ensuring the offset to larger than
the max slot of samplers, but also of image slots.
When using Eevee-next materials the slot numbers can be very high
as the location of the first material texture is set to
`GPU_max_textures_frag()` On current NVIDIA drivers this number is
set to 1K. Images are stored after the textures slots upwards.
In the future driver could set this to MAX_INT which will not work
with the new implementation. This isn't the case at this moment
and perhaps also not realistic. We don't check for this situation yet.
Pull Request: https://projects.blender.org/blender/blender/pulls/112511
Component swizzling is part of an image view and localized in the code.
With component swizzling a different channel (or 0 or 1) can be read
when the component is accessed.
Pull Request: https://projects.blender.org/blender/blender/pulls/112273
Adding support for Texture View in Vulkan backend.
When a VKTexture is a texture view many options become
irrelevant, but isn't actually clear via its API. Added
asserts to make sure that the usage is supported.
This patch allows rendering simple workbench-next scenes
without any validation errors.
Pull Request: https://projects.blender.org/blender/blender/pulls/110887
When `COMP_U*` is used together with a `GPU_FETCH_INT` the backend
tried to convert them to floats. What is of course not what should
happen.
Issue was that these cases were never correctly tested. With this
change the `data` attribute of the `overlay_edit_mesh` shaders are
transferred correctly and can draw some pixels as intended.
Pull Request: https://projects.blender.org/blender/blender/pulls/112270
Shader draw parameter support is needed for Workbench-next. (Almost) All
devices that we support also supports shader draw parameters. This PR
doesn't include a workaround for devices that don't have support.
For end-users this PR will allow workbench-next to render the object on
the correct location. Before this patch they were always drawn on the origin.
The catch is that OpenGL uses gl_InstanceID which always starts counting at 0.
Vulkan has gl_InstanceIndex which holds the correct instance number. By
subtracting the gl_InstanceIndex from gl_BaseInstance will get the correct
gl_InstanceID.
This patch also activates some device extensions to perform in shader printf
statements inside a shader. Documentation for this will be added to the dev
docs how to use them.
Pull Request: https://projects.blender.org/blender/blender/pulls/111461
Enhance custom framebuffer binding state to allow
specification of clear color as part of the loadstore
state at framebuffer bind time.
This ensures all parameters controlling attachment
loading and storage behaviour can be explicitly
specified when binding a framebuffer.
This change enables optimizations which leverage
explicit framebuffer load store state to also specify
a clear color without prematurely triggering a
clear which may occur independently to
render work when using GPU_framebuffer_clear(..).
Authored by Apple: Michael Parkin-White.
Pull Request: https://projects.blender.org/blender/blender/pulls/111810
This is a rewrite of GHOST_ContextVK to align with Metal backend as described
in #111389 - solution 3 with the adaptation that GHOST is still responsible
for presenting the swap chain image and a post callback is still needed in
case the swapchain is recreated.
This PR also includes some smaller improvements in stability.
Technical documentation: https://developer.blender.org/docs/eevee_and_viewport/gpu/vulkan/swap_chain/
* Renderpasses and framebuffers are not owned anymore by GHOST_ContextVK
* VKFramebuffer doesn't contain a swap chain image.
* Swapchain images can only be used as a blit destination or present source.
Not as an attachment.
* GHOST_ContextVK::swapBuffers would call a callback with the image the
GPU module needs to blit the results to.
* Clearing of depth/stencil attachments when no depth write state is set.
* Enable VK_KHR_maintenance4 to relax the stage interface mapping.
* Removes most vulkan validation warnings/errors.
* Detection of frame buffer changes that needs to be applied before
performing a command requiring render pass (draw/clear attachment)
**Benefits**
* Late retrieval of a swap chain image results in better overall performance as
Blender doesn't need to wait until the image is presented on the screen.
* Easier API and clearer state (transitions)
* More control over Image layouts and command buffer states. (Better alignment with
Vulkan API)
Pull Request: https://projects.blender.org/blender/blender/pulls/111473
Vulkan device was deallocated, but when destroying the GHOST context
the GPU context is reactivated and want to allocate buffers lazilly.
This is solved by de-initializing the device on platform exit, resetting
buffer pointers so double free can be detected.
Pull Request: https://projects.blender.org/blender/blender/pulls/111462
