This PR implements a new the threading model for building render graphs
based on tests performed last month. For out workload multithreaded
command building will block in the driver or device. So better to use a
single thread for command building.
Details of the internal working is documented at https://developer.blender.org/docs/features/gpu/vulkan/render_graph/
- When a context is activated on a thread the context asks for a
render graph it can use by calling `VKDevice::render_graph_new`.
- Parts of the GPU backend that requires GPU commands will add a
specific render graph node to the render graph. The nodes also
contains a reference to all resources it needs including the
access it needs and the image layout.
- When the context is flushed the render graph is submitted to the
device by calling `VKDevice::render_graph_submit`.
- The device puts the render graph in `VKDevice::submission_pool`.
- There is a single background thread that gets the next render
graph to send to the GPU (`VKDevice::submission_runner`).
- Reorder the commands of the render graph to comply with Vulkan
specific command order rules and reducing possible bottlenecks.
(`VKScheduler`)
- Generate the required barriers `VKCommandBuilder::groups_extract_barriers`.
This is a separate step to reduce resource locking giving other
threads access to the resource states when they are building
the render graph nodes.
- GPU commands and pipeline barriers are recorded to a VkCommandBuffer.
(`VKCommandBuilder::record_commands`)
- When completed the command buffer can be submitted to the device
queue. `vkQueueSubmit`
- Render graphs that have been submitted can be reused by a next
thread. This is done by pushing the render graph to the
`VKDevice::unused_render_graphs` queue.
Pull Request: https://projects.blender.org/blender/blender/pulls/132681
Memory areas was requested to be preferable host visible. On some
platforms this would fail to allocate. Best is to not add preferable
host visible for typically large allocations.
This PR also gives the caller the responsibility to set the allocation flags.
Pull Request: https://projects.blender.org/blender/blender/pulls/133528
This change adds the option to update a buffer via the render
graph via `vkCmdUpdateBuffer`. This is only enabled for
uniform buffers as they are small and aligned/sized correctly.
Pull Request: https://projects.blender.org/blender/blender/pulls/128416
Resource binding was over-complicated as I didn't understood the state
manager and vulkan to make the correct decisions at that time. This
refactor will remove a lot of the complexity and improves the performance.
**Performance**
The performance improvement is noticeable in complex grease pencil
scenes.
Grease pencil benchmark file picknick:
- `NVIDIA Quadro RTX 6000` 17 fps -> 24 fps
- `Intel(R) Arc(tm) A750 Graphics (DG2)` 6 -> 21 fps
**Bottle-neck**
The performance improvements originates from moving the update entry
point from state manager to shader interface. The previous implementation
(state manager) had to loop over all the bound resources and find in the
shader interface where it was located in the descriptor set. Ignoring
resources that were not used by the shader. But also making it hard
to determine if descriptor sets actually changed. Previous implementation
assumed descriptor sets always changed.
When descriptor set changed a new descriptor set needed to be allocated.
Most drivers this is a fast operation, but on Intel/Mesa this was measurable
slow. Using an allocation pool doesn't fit the Vulkan API as you are only
able to reuse when the layout matches exactly. Of course doable, but requires
another structure to keep track of the actual layouts.
**Solution**
By using the shader interface as entry point we can:
1. Keep track if there are any changes in the state manager. If not and the
layout is the same, the previous shader can be reused.
2. In stead of looping over each bound resource, we loop over bind points.
**Future extensions**
Bundle all descriptor set uploads just before use. This would be more
in line with how 'modern' Vulkan should be implemented. This PR already
separates the uploading from the updating and technically allows to upload
more than one descriptor set.
Instead of looking 1 set back we should measure if we can handle multiple
or keep track of the different layouts resources to improve the performance
even further.
Optional use `VK_KHR_descriptor_buffer` when available.
Pull Request: https://projects.blender.org/blender/blender/pulls/128068
Removes two levels of indirection when updating descriptor sets.
These are the easy ones to remove. Others will be removed in
a future PR.
This is part of reworking of how descriptor sets are used.
This PR Mostly reduces complexity.
Pull Request: https://projects.blender.org/blender/blender/pulls/127915
When building the resource access used when adding dispatch/draw commands
to the render graph, the access mask is required. This PR stores the
access mask in the shader interface. When binding the resources referenced
by the state manager, the resource access info struct is populated with
the access flags.
In the near future the resource access info will be passed when adding
a dispatch/draw node to the render graph to generate the links.
Pull Request: https://projects.blender.org/blender/blender/pulls/120908
A developer can switch `vk_common.hh#use_render_graph` to enable render graph.
When enabled the buffers and images are tracked by the device resource state
tracker. The storage buffer commands are recorded to the context render graph.
The next unit tests will pass:
- GPUVulkanTest.storage_buffer_create_update_read
- GPUVulkanTest.storage_buffer_clear_zero
- GPUVulkanTest.storage_buffer_clear
- GPUVulkanTest.storage_buffer_copy_from_vertex_buffer
The pattern to migrate to render graph is:
- always construct CreateInfo for class.
- based on `use_render_graph` call `context.command_buffers.something`
or `context.render_graph.add_node`.
- Hide calls to `context.flush` when `use_render_graph` is true.
Pull Request: https://projects.blender.org/blender/blender/pulls/120812
Previously a storage buffer was used to store draw list commands as it
matches already existing APIs. Unfortunately StorageBuffers prefers to
be stored on the GPU device and would reduce the benefit of a dynamic
draw list.
This PR replaces the storage buffer with a regular buffer, which keeps
more control where to store the buffer.
Pull Request: https://projects.blender.org/blender/blender/pulls/117712
A draw list bundles multiple draw commands for the same geometry
and sends the draw commands in a single command. This reduces
the overhead of pipeline checking, resource validation and can
keep the load higher on the gpu as more work needs to be done.
Previously the draw list didn't bundle any commands and would still
send each call separately to the GPU. This PR implements the bundling
of the commands.
Pull Request: https://projects.blender.org/blender/blender/pulls/117548
Currently all buffer types were stored in host memory, which is visible to the GPU as well.
This is typically slow as the data would be transferred over the PCI bus when used.
Most of the time Index and Vertex buffers are written once and read many times so it makes
more sense to locate them on the GPU. Storage buffers typically require quick access as they
are created for shading/compute purposes.
This PR will try to store vertex buffers, index buffers and storage buffers on device memory
to improve the performance.
Uniform buffers are still located on host memory as they can be uploaded during binding process.
This can (will) reset the graphics pipeline triggering draw calls using unattached resources.
In future this could be optimized further as in:
* using different pools for allocating specific buffers, with a fallback when buffers cannot be
stored on the GPU anymore.
* store uniform buffers in device memory
Pull Request: https://projects.blender.org/blender/blender/pulls/115343
In Blender a context should not be shared between threads. In Vulkan a
command pool must not be shared between threads. In the current
implementation the command pool are stored on device level and could
therefore be shared between multiple context which made the implementation
not matching these rules.
This PR moves the command pool from device to command buffers where it
would not conflict between other contexts. This PR doesn't make the Vulkan
backend fully multithreaded. The access to the queue is still missing.
Pull Request: https://projects.blender.org/blender/blender/pulls/114977
Vulkan API uses Flags and FlagBits for enumerations. The FlagBits
contains the options that can be hold with the Flags data type.
This wasn't well understood at the beginning of the project and
the FlagBits where used where Flags should have been used. This cleanup
fixes this, improving the readability of the code where bit
manipulations where used.
Pull Request: https://projects.blender.org/blender/blender/pulls/114459
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
Listing the "Blender Foundation" as copyright holder implied the Blender
Foundation holds copyright to files which may include work from many
developers.
While keeping copyright on headers makes sense for isolated libraries,
Blender's own code may be refactored or moved between files in a way
that makes the per file copyright holders less meaningful.
Copyright references to the "Blender Foundation" have been replaced with
"Blender Authors", with the exception of `./extern/` since these this
contains libraries which are more isolated, any changed to license
headers there can be handled on a case-by-case basis.
Some directories in `./intern/` have also been excluded:
- `./intern/cycles/` it's own `AUTHORS` file is planned.
- `./intern/opensubdiv/`.
An "AUTHORS" file has been added, using the chromium projects authors
file as a template.
Design task: #110784
Ref !110783.
The current Vulkan resource management has some issues as context that
are not active can still use resources that are freed via another
context.
When this happens incorrect data can be read on the GPU and even crash
Blender. When trying to bind something that now contains other memory
pointers.
This change introduces that contexts are tracked via the device.
Context will be registered/unregistered with the device instance.
Unbinding of resources must pass the device and the device will check
all registered contexts. Binding of resources will happen via the active
context only.
On user perspective this now allowes:
- Opening/switching files
- Switching workspaces
- Switching render engines
Pull Request: https://projects.blender.org/blender/blender/pulls/108968
This PR adds support for indirect compute.
Indirect compute is almost the same as regular compute. The
only difference is that the parameters for the compute dispatch
isn't passed as a parameter, but that these parameters are part
of a buffer.
Pull Request: https://projects.blender.org/blender/blender/pulls/108879
A lot of files were missing copyright field in the header and
the Blender Foundation contributed to them in a sense of bug
fixing and general maintenance.
This change makes it explicit that those files are at least
partially copyrighted by the Blender Foundation.
Note that this does not make it so the Blender Foundation is
the only holder of the copyright in those files, and developers
who do not have a signed contract with the foundation still
hold the copyright as well.
Another aspect of this change is using SPDX format for the
header. We already used it for the license specification,
and now we state it for the copyright as well, following the
FAQ:
https://reuse.software/faq/
ShaderInput lookup used to be explicit as in we know for sure that the
shader input exists. This was guarded by a assert. During development
of the graphics pipeline we came to the conclusion that this isn't
always the case due to its late bindings of resources.
This PR Makes the shader input lookup return an optional result.
This result can then be checked depending on the area where it is used
to validate the existence.
Pull Request: https://projects.blender.org/blender/blender/pulls/107735
Previous GHOST_ContextVK would create a logical device for each
context. Blender uses multiple contexts at the same time and wasn't able
to share resources between them as the logical device where different.
This patch will create a single logical device and share them between
multiple contexts. This allows sharing memory/shaders between contexts
and make sure that all memory allocations are freed from the device it
was allocated from.
Some allocations in Blender are freed when there isn't a context, this
was failing in the previous implementation. We didn't noticed it before
as we didn't test multiple contexts.
This patch also moves device specific data structures from VKContext to
VKDevice like the descriptor pools, debug layers etc.
Pull Request: https://projects.blender.org/blender/blender/pulls/107606
The goal is to solve confusion of the "All rights reserved" for licensing
code under an open-source license.
The phrase "All rights reserved" comes from a historical convention that
required this phrase for the copyright protection to apply. This convention
is no longer relevant.
However, even though the phrase has no meaning in establishing the copyright
it has not lost meaning in terms of licensing.
This change makes it so code under the Blender Foundation copyright does
not use "all rights reserved". This is also how the GPL license itself
states how to apply it to the source code:
<one line to give the program's name and a brief idea of what it does.>
Copyright (C) <year> <name of author>
This program is free software ...
This change does not change copyright notice in cases when the copyright
is dual (BF and an author), or just an author of the code. It also does
mot change copyright which is inherited from NaN Holding BV as it needs
some further investigation about what is the proper way to handle it.
For example
```
OIIOOutputDriver::~OIIOOutputDriver()
{
}
```
becomes
```
OIIOOutputDriver::~OIIOOutputDriver() {}
```
Saves quite some vertical space, which is especially handy for
constructors.
Pull Request: https://projects.blender.org/blender/blender/pulls/105594
The previous API for clearing storage buffers was following the OpenGL
api. OpenGL has many options to support for data conversions, striding
and sizzling. Metal and Vulkan don't have these features and we have to
deal it ourselves.
Blender internally only uses a tiny subset for what is possible in
OpenGL. Making the current API to difficult to implement on our future
platforms as we had to implement all cases, most even not used at all.
By changing the API we make future development easier as we only need
to implement what we are actually using.
**New API**
`GPU_storagebuf_clear(GPUStorageBuf* ssbo, uint32_t clear_value)`
Related issue: #105492
Pull Request: https://projects.blender.org/blender/blender/pulls/105521
It is recommended to map buffers once and not each time the
mapped memory is needed. This patch will map the buffer when
created and unmap the buffer when the buffer is freed.
This patch will reduce the overhead where the Vulkan driver or
the virtual memory manager needs to be accessed.
Pull Request: https://projects.blender.org/blender/blender/pulls/105588
Descriptor set locations are now determined in the
VKShaderInterface. Issues with the previous solution:
- Due to legacy code in GPU module the locations/bindings
must be the same. Using one for something else might
result in undesired lookups, incorrect resource
bindings.
- Images/Textures reuses the same namespace, that didn't
work as expected when looking up the resources via
its binding.
This refactoring is required for adding support for
push constants.
Pull Request #105073
This patch adds initial support for compute shaders to
the vulkan backend. As the development is oriented to the test-
cases we have the implementation is limited to what is used there.
It has been validated that with this patch that the following test
cases are running as expected
- `GPUVulkanTest.gpu_shader_compute_vbo`
- `GPUVulkanTest.gpu_shader_compute_ibo`
- `GPUVulkanTest.gpu_shader_compute_ssbo`
- `GPUVulkanTest.gpu_storage_buffer_create_update_read`
- `GPUVulkanTest.gpu_shader_compute_2d`
This patch includes:
- Allocating VkBuffer on device.
- Uploading data from CPU to VkBuffer.
- Binding VkBuffer as SSBO to a compute shader.
- Execute compute shader and altering VkBuffer.
- Download the VkBuffer to CPU ram.
- Validate that it worked.
- Use device only vertex buffer as SSBO
- Use device only index buffer as SSBO
- Use device only image buffers
GHOST API has been changed as the original design was created before
we even had support for compute shaders in blender. The function
`GHOST_getVulkanBackbuffer` has been separated to retrieve the command
buffer without a backbuffer (`GHOST_getVulkanCommandBuffer`). In order
to do correct command buffer processing we needed access to the queue
owned by GHOST. This is returned as part of the `GHOST_getVulkanHandles`
function.
Open topics (not considered part of this patch)
- Memory barriers & command buffer encoding
- Indirect compute dispatching
- Rest of the test cases
- Data conversions when requested data format is different than on device.
- GPUVulkanTest.gpu_shader_compute_1d is supported on AMD devices.
NVIDIA doesn't seem to support 1d textures.
Pull-request: #104518
This patch adds a placeholder for the vulkan backend.
When activated (`WITH_VULKAN_BACKEND=On` and `--gpu-backend vulkan`)
it might open a blender screen, but nothing should be visible as
none of the functions are implemented or otherwise crash on a nullptr.
This is expected as this is just a placeholder. The goal is to add shader compilation
+validation to this backend as one of the next steps so we can validate
changes to existing shaders on OpenGL, Metal and Vulkan at the same time.
Reviewed By: fclem
Differential Revision: https://developer.blender.org/D16338
