Implement a generic shader cache for fast lookups and invalidations.
Invalidations are cheap but expensive when a shader is invalidated.
Use two mutexes instead of one to avoid locking invalidations for
lookups and vice versa. When a shader has to be removed, lookups are
locked as expected.
Drop the std::list hack to allocate memory indefinitely.
Instead use a custom allocator that keeps references valid until
destruction. This allocates fixed chunks of memory and puts pointers in
a free list. When an allocation is no longer used put it back to the
free list, this doesn't heap allocate because std::vector doesn't change
the capacity. If the free list is empty, allocate a new chunk.
Deduplicate code shared between vk_pipeline_cache and gl_shader_cache as
well as shader decoder code.
While we are at it, fix a bug in gl_shader_cache where compute shaders
had an start offset of a stage shader.
Adds optional support for Nsight Aftermath. It is enabled through
ENABLE_NSIGHT_AFTERMATH in cmake. A path to the SDK has to be provided
by the environment variable NSIGHT_AFTERMATH_SDK.
Nsight Aftermath allows an application to generate "minidumps" of the
GPU state when a device loss happens. By analysing these on Nsight we
can know what a game was doing and why it triggered a device loss.
The dump is generated inside %APPDATA%\yuzu\log\gpucrash and this
directory is deleted every time a new instance is initialized with
Nsight enabled.
To enable it on yuzu there has a to be a driver and device capable of
running Nsight Aftermath on Vulkan. That means only Turing based GPUs
on the latest stable driver, beta drivers won't work for now.
It is manually enabled in Configuration>Debug>Enable Graphics Debugging
because when using all debugging capabilities there is a runtime cost.
This is a reversed look up table extracted from
https://gist.github.com/rygorous/2203834#file-gistfile1-cpp-L41-L62
that is used in
04d4e9e587/source/maxwell/tsc_generate.cpp (L38)
Games usually bind 0xFD expecting a float texture border of 1.0f.
The conversion previous to this commit was multiplying the uint8 sRGB
texture border color by 255. This is close to 1.0f but when that
difference matters, some graphical glitches appear.
This look up table is manually changed in the edges, clamping towards
0.0f and 1.0f.
While we are at it, move this logic to its own translation unit.
The intention behind a Vulkan wrapper is to drop Vulkan-Hpp.
The issues with Vulkan-Hpp are:
- Regular breaks of the API.
- Copy constructors that do the same as the aggregates (fixed recently)
- External dynamic dispatch that is hard to remove
- Alias KHR handles with non-KHR handles making it impossible to use
smart handles on Vulkan 1.0 instances with extensions that were included
on Vulkan 1.1.
- Dynamic dispatchers silently change size depending on preprocessor
definitions. Different files will have different dispatch definitions,
generating all kinds of hard to debug memory issues.
In other words, Vulkan-Hpp is not "production ready" for our needs and
this wrapper aims to replace it without losing RAII and exception
safety.
Abstract the current OpenGL implementation into the VideoCommon
namespace and reimplement it on top of that. Doing this avoids repeating
code and logic in the Vulkan implementation.
This currently only supports quad arrays and u8 indices.
In the future we can remove quad arrays with a table written from the
CPU, but this was used to bootstrap the other passes helpers and it
was left in the code.
The blob code is generated from the "shaders/" directory. Read the
instructions there to know how to generate the SPIR-V.
This abstractio represents the state of the 3D engine at a given draw.
Instead of changing individual bits of the pipeline how it's done in
APIs like D3D11, OpenGL and NVN; on Vulkan we are forced to put
everything together into a single, immutable object.
It takes advantage of the few dynamic states Vulkan offers.
The update descriptor is used to store in flat memory a large chunk of
staging data used to update descriptor sets through templates. It
provides a push interface to easily insert descriptors following the
current pipeline. The order used in the descriptor update template has
to be implicitly followed. We can catch bugs here using validation
layers.
Create a large descriptor pool where we allocate all our descriptors
from. It has to be wide enough to support any pipeline, hence its large
numbers.
If the descritor pool is filled, we allocate more memory at that moment.
This way we can take advantage of permissive drivers like Nvidia's that
allocate more descriptors than what the spec requires.
The job of this abstraction is to provide staging buffers for temporary
operations. Think of image uploads or buffer uploads to device memory.
It automatically deletes unused buffers.
This object's job is to contain an image and manage its transitions.
Since Nvidia hardware doesn't know what a transition is but Vulkan
requires them anyway, we have to state track image subresources
individually.
To avoid the overhead of tracking each subresource in images with many
subresources (think of cubemap arrays with several mipmaps), this commit
tracks when subresources have diverged. As long as this doesn't happen
we can check the state of the first subresource (that will be shared
with all subresources) and update accordingly.
Image transitions are deferred to the scheduler command buffer.
The intention behind this hasheable structure is to describe the state
of fixed function pipeline state that gets compiled to a single graphics
pipeline state object. This is all dynamic state in OpenGL but Vulkan
wants it in an immutable state, even if hardware can edit it freely.
In this commit the structure is defined in an optimized state (it uses
booleans, has paddings and many data entries that can be packed to
single integers). This is intentional as an initial implementation that
is easier to debug, implement and review. It will be optimized in later
stages, or it might change if Vulkan gets more dynamic states.
Use a large flat array to look up texture formats. This allows us to
properly implement formats with different component types. It should
also be faster.
Implement VOTE using Nvidia's intrinsics. Documentation about these can
be found here
https://developer.nvidia.com/reading-between-threads-shader-intrinsics
Instead of using portable ARB instructions I opted to use Nvidia
intrinsics because these are the closest we have to how Tegra X1
hardware renders.
To stub VOTE on non-Nvidia drivers (including nouveau) this commit
simulates a GPU with a warp size of one, returning what is meaningful
for the instruction being emulated:
* anyThreadNV(value) -> value
* allThreadsNV(value) -> value
* allThreadsEqualNV(value) -> true
ballotARB, also known as "uint64_t(activeThreadsNV())", emits
VOTE.ANY Rd, PT, PT;
on nouveau's compiler. This doesn't match exactly to Nvidia's code
VOTE.ALL Rd, PT, PT;
Which is emulated with activeThreadsNV() by this commit. In theory this
shouldn't really matter since .ANY, .ALL and .EQ affect the predicates
(set to PT on those cases) and not the registers.
Analysis passes do not have a good reason to depend on shader_ir.h to
work on top of nodes. This splits node-related declarations to their own
file and leaves the IR in shader_ir.h
Instead of having a vector of unique_ptr stored in a vector and
returning star pointers to this, use shared_ptr. While changing
initialization code, move it to a separate file when possible.
This is a first step to allow code analysis and node generation beyond
the ShaderIR class.
Implements an API agnostic texture view based texture cache. Classes
defined here are intended to be inherited by the API implementation and
used in API-specific code.
This implementation exposes protected virtual functions to be called
from the implementer.
Before executing any surface copies methods (defined in API-specific code)
it tries to detect if the overlapping surface is a superset and if it
is, it creates a view. Views are references of a subset of a surface, it
can be a superset view (the same as referencing the whole texture).
Current code manages 1D, 1D array, 2D, 2D array, cube maps and cube map
arrays with layer and mipmap level views. Texture 3D slices views are
not implemented.
If the view attempt fails, the fast path is invoked with the overlapping
textures (defined in the implementer). If that one fails (returning
nullptr) it will flush and reload the texture.
This manages two kinds of streaming buffers: one for unified memory
models and one for dedicated GPUs. The first one skips the copy from the
staging buffer to the real buffer, since it creates an unified buffer.
This implementation waits for all fences to finish their operation
before "invalidating". This is suboptimal since it should allocate
another buffer or start searching from the beginning. There is room for
improvement here.
This could also handle AMD's "pinned" memory (a heap with 256 MiB) that
seems to be designed for buffer streaming.
The scheduler abstracts command buffer and fence management with an
interface that's able to do OpenGL-like operations on Vulkan command
buffers.
It returns by value a command buffer and fence that have to be used for
subsequent operations until Flush or Finish is executed, after that the
current execution context (the pair of command buffers and fences) gets
invalidated a new one must be fetched. Thankfully validation layers will
quickly detect if this is skipped throwing an error due to modifications
to a sent command buffer.
VKDevice contains all the data required to manage and initialize a
physical device. Its intention is to be passed across Vulkan objects to
query device-specific data (for example the logical device and the
dispatch loader).
This file is intended to be included instead of vulkan/vulkan.hpp. It
includes declarations of unique handlers using a dynamic dispatcher
instead of a static one (which would require linking to a Vulkan
library).
When I originally added the compute assert I used the wrong
documentation. This addresses that.
The dispatch register was tested with homebrew against hardware and is
triggered by some games (e.g. Super Mario Odyssey). What exactly is
missing to get a valid program bound by this engine requires more
investigation.
Those implementations are quite costly, so there is no need to inline them to the caller.
Ressource deletion is often a performance bug, so in this way, we support to add breakpoints to them.
The idea of this cache is to avoid redundant uploads. So we are going
to cache the uploaded buffers within the stream_buffer and just reuse
the old pointers.
The next step is to implement a VBO cache on GPU memory, but for now,
I want to check the overhead of the cache management. Fetching the
buffer over PCI-E should be quite fast.
The Ryujinx macro interpreter and envydis were used as reference.
Macros are programs that are uploaded by the games during boot and can later be called by writing to their method id in a GPU command buffer.
The geometry pipeline manages data transfer between VS, GS and primitive assembler. It has known four modes:
- no GS mode: sends VS output directly to the primitive assembler (what citra currently does)
- GS mode 0: sends VS output to GS input registers, and sends GS output to primitive assembler
- GS mode 1: sends VS output to GS uniform registers, and sends GS output to primitive assembler. It also takes an index from the index buffer at the beginning of each primitive for determine the primitive size.
- GS mode 2: similar to mode 1, but doesn't take the index and uses a fixed primitive size.
hwtest shows that immediate mode also supports GS (at least for mode 0), so the geometry pipeline gets refactored into its own class for supporting both drawing mode.
In the immediate mode, some games don't set the pipeline registers to a valid value until the first attribute input, so a geometry pipeline reset flag is set in `pipeline.vs_default_attributes_setup.index` trigger, and the actual pipeline reconfigure is triggered in the first attribute input.
In the normal drawing mode with index buffer, the vertex cache is a little bit modified to support the geometry pipeline. Instead of OutputVertex, it now holds AttributeBuffer, which is the input to the geometry pipeline. The AttributeBuffer->OutputVertex conversion is done inside the pipeline vertex handler. The actual hardware vertex cache is believed to be implemented in a similar way (because this is the only way that makes sense).
Both geometry pipeline and GS unit rely on states preservation across drawing call, so they are put into the global state. In the future, the other three vertex shader units should be also placed in the global state, and a scheduler should be implemented on top of the four units. Note that the current gs_unit already allows running VS on it in the future.