Merge pull request #3986 from ReinUsesLisp/shader-cache

shader_cache: Implement a generic runtime shader cache
merge-requests/60/head
bunnei 2020-06-12 23:14:48 +07:00 committed by GitHub
commit 5633887569
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
14 changed files with 364 additions and 417 deletions

@ -49,8 +49,6 @@ add_library(video_core STATIC
query_cache.h
rasterizer_accelerated.cpp
rasterizer_accelerated.h
rasterizer_cache.cpp
rasterizer_cache.h
rasterizer_interface.h
renderer_base.cpp
renderer_base.h
@ -93,6 +91,7 @@ add_library(video_core STATIC
renderer_opengl/utils.h
sampler_cache.cpp
sampler_cache.h
shader_cache.h
shader/decode/arithmetic.cpp
shader/decode/arithmetic_immediate.cpp
shader/decode/bfe.cpp

@ -1,7 +0,0 @@
// Copyright 2018 yuzu Emulator Project
// Licensed under GPLv2 or any later version
// Refer to the license.txt file included.
#include "video_core/rasterizer_cache.h"
RasterizerCacheObject::~RasterizerCacheObject() = default;

@ -1,253 +0,0 @@
// Copyright 2018 yuzu Emulator Project
// Licensed under GPLv2 or any later version
// Refer to the license.txt file included.
#pragma once
#include <mutex>
#include <set>
#include <unordered_map>
#include <boost/icl/interval_map.hpp>
#include <boost/range/iterator_range_core.hpp>
#include "common/common_types.h"
#include "core/settings.h"
#include "video_core/gpu.h"
#include "video_core/rasterizer_interface.h"
class RasterizerCacheObject {
public:
explicit RasterizerCacheObject(const VAddr cpu_addr) : cpu_addr{cpu_addr} {}
virtual ~RasterizerCacheObject();
VAddr GetCpuAddr() const {
return cpu_addr;
}
/// Gets the size of the shader in guest memory, required for cache management
virtual std::size_t GetSizeInBytes() const = 0;
/// Sets whether the cached object should be considered registered
void SetIsRegistered(bool registered) {
is_registered = registered;
}
/// Returns true if the cached object is registered
bool IsRegistered() const {
return is_registered;
}
/// Returns true if the cached object is dirty
bool IsDirty() const {
return is_dirty;
}
/// Returns ticks from when this cached object was last modified
u64 GetLastModifiedTicks() const {
return last_modified_ticks;
}
/// Marks an object as recently modified, used to specify whether it is clean or dirty
template <class T>
void MarkAsModified(bool dirty, T& cache) {
is_dirty = dirty;
last_modified_ticks = cache.GetModifiedTicks();
}
void SetMemoryMarked(bool is_memory_marked_) {
is_memory_marked = is_memory_marked_;
}
bool IsMemoryMarked() const {
return is_memory_marked;
}
void SetSyncPending(bool is_sync_pending_) {
is_sync_pending = is_sync_pending_;
}
bool IsSyncPending() const {
return is_sync_pending;
}
private:
bool is_registered{}; ///< Whether the object is currently registered with the cache
bool is_dirty{}; ///< Whether the object is dirty (out of sync with guest memory)
bool is_memory_marked{}; ///< Whether the object is marking rasterizer memory.
bool is_sync_pending{}; ///< Whether the object is pending deletion.
u64 last_modified_ticks{}; ///< When the object was last modified, used for in-order flushing
VAddr cpu_addr{}; ///< Cpu address memory, unique from emulated virtual address space
};
template <class T>
class RasterizerCache : NonCopyable {
friend class RasterizerCacheObject;
public:
explicit RasterizerCache(VideoCore::RasterizerInterface& rasterizer) : rasterizer{rasterizer} {}
/// Write any cached resources overlapping the specified region back to memory
void FlushRegion(VAddr addr, std::size_t size) {
std::lock_guard lock{mutex};
const auto& objects{GetSortedObjectsFromRegion(addr, size)};
for (auto& object : objects) {
FlushObject(object);
}
}
/// Mark the specified region as being invalidated
void InvalidateRegion(VAddr addr, u64 size) {
std::lock_guard lock{mutex};
const auto& objects{GetSortedObjectsFromRegion(addr, size)};
for (auto& object : objects) {
if (!object->IsRegistered()) {
// Skip duplicates
continue;
}
Unregister(object);
}
}
void OnCPUWrite(VAddr addr, std::size_t size) {
std::lock_guard lock{mutex};
for (const auto& object : GetSortedObjectsFromRegion(addr, size)) {
if (object->IsRegistered()) {
UnmarkMemory(object);
object->SetSyncPending(true);
marked_for_unregister.emplace_back(object);
}
}
}
void SyncGuestHost() {
std::lock_guard lock{mutex};
for (const auto& object : marked_for_unregister) {
if (object->IsRegistered()) {
object->SetSyncPending(false);
Unregister(object);
}
}
marked_for_unregister.clear();
}
/// Invalidates everything in the cache
void InvalidateAll() {
std::lock_guard lock{mutex};
while (interval_cache.begin() != interval_cache.end()) {
Unregister(*interval_cache.begin()->second.begin());
}
}
protected:
/// Tries to get an object from the cache with the specified cache address
T TryGet(VAddr addr) const {
const auto iter = map_cache.find(addr);
if (iter != map_cache.end())
return iter->second;
return nullptr;
}
/// Register an object into the cache
virtual void Register(const T& object) {
std::lock_guard lock{mutex};
object->SetIsRegistered(true);
interval_cache.add({GetInterval(object), ObjectSet{object}});
map_cache.insert({object->GetCpuAddr(), object});
rasterizer.UpdatePagesCachedCount(object->GetCpuAddr(), object->GetSizeInBytes(), 1);
object->SetMemoryMarked(true);
}
/// Unregisters an object from the cache
virtual void Unregister(const T& object) {
std::lock_guard lock{mutex};
UnmarkMemory(object);
object->SetIsRegistered(false);
if (object->IsSyncPending()) {
marked_for_unregister.remove(object);
object->SetSyncPending(false);
}
const VAddr addr = object->GetCpuAddr();
interval_cache.subtract({GetInterval(object), ObjectSet{object}});
map_cache.erase(addr);
}
void UnmarkMemory(const T& object) {
if (!object->IsMemoryMarked()) {
return;
}
rasterizer.UpdatePagesCachedCount(object->GetCpuAddr(), object->GetSizeInBytes(), -1);
object->SetMemoryMarked(false);
}
/// Returns a ticks counter used for tracking when cached objects were last modified
u64 GetModifiedTicks() {
std::lock_guard lock{mutex};
return ++modified_ticks;
}
virtual void FlushObjectInner(const T& object) = 0;
/// Flushes the specified object, updating appropriate cache state as needed
void FlushObject(const T& object) {
std::lock_guard lock{mutex};
if (!object->IsDirty()) {
return;
}
FlushObjectInner(object);
object->MarkAsModified(false, *this);
}
std::recursive_mutex mutex;
private:
/// Returns a list of cached objects from the specified memory region, ordered by access time
std::vector<T> GetSortedObjectsFromRegion(VAddr addr, u64 size) {
if (size == 0) {
return {};
}
std::vector<T> objects;
const ObjectInterval interval{addr, addr + size};
for (auto& pair : boost::make_iterator_range(interval_cache.equal_range(interval))) {
for (auto& cached_object : pair.second) {
if (!cached_object) {
continue;
}
objects.push_back(cached_object);
}
}
std::sort(objects.begin(), objects.end(), [](const T& a, const T& b) -> bool {
return a->GetLastModifiedTicks() < b->GetLastModifiedTicks();
});
return objects;
}
using ObjectSet = std::set<T>;
using ObjectCache = std::unordered_map<VAddr, T>;
using IntervalCache = boost::icl::interval_map<VAddr, ObjectSet>;
using ObjectInterval = typename IntervalCache::interval_type;
static auto GetInterval(const T& object) {
return ObjectInterval::right_open(object->GetCpuAddr(),
object->GetCpuAddr() + object->GetSizeInBytes());
}
ObjectCache map_cache;
IntervalCache interval_cache; ///< Cache of objects
u64 modified_ticks{}; ///< Counter of cache state ticks, used for in-order flushing
VideoCore::RasterizerInterface& rasterizer;
std::list<T> marked_for_unregister;
};

@ -10,7 +10,6 @@
#include "common/common_types.h"
#include "video_core/buffer_cache/buffer_cache.h"
#include "video_core/engines/maxwell_3d.h"
#include "video_core/rasterizer_cache.h"
#include "video_core/renderer_opengl/gl_resource_manager.h"
#include "video_core/renderer_opengl/gl_stream_buffer.h"

@ -30,6 +30,7 @@
#include "video_core/renderer_opengl/gl_shader_cache.h"
#include "video_core/renderer_opengl/maxwell_to_gl.h"
#include "video_core/renderer_opengl/renderer_opengl.h"
#include "video_core/shader_cache.h"
namespace OpenGL {
@ -310,7 +311,7 @@ void RasterizerOpenGL::SetupShaders(GLenum primitive_mode) {
continue;
}
Shader shader{shader_cache.GetStageProgram(program)};
Shader* const shader = shader_cache.GetStageProgram(program);
if (device.UseAssemblyShaders()) {
// Check for ARB limitation. We only have 16 SSBOs per context state. To workaround this
@ -870,7 +871,7 @@ bool RasterizerOpenGL::AccelerateDisplay(const Tegra::FramebufferConfig& config,
return true;
}
void RasterizerOpenGL::SetupDrawConstBuffers(std::size_t stage_index, const Shader& shader) {
void RasterizerOpenGL::SetupDrawConstBuffers(std::size_t stage_index, Shader* shader) {
static constexpr std::array PARAMETER_LUT = {
GL_VERTEX_PROGRAM_PARAMETER_BUFFER_NV, GL_TESS_CONTROL_PROGRAM_PARAMETER_BUFFER_NV,
GL_TESS_EVALUATION_PROGRAM_PARAMETER_BUFFER_NV, GL_GEOMETRY_PROGRAM_PARAMETER_BUFFER_NV,
@ -900,7 +901,7 @@ void RasterizerOpenGL::SetupDrawConstBuffers(std::size_t stage_index, const Shad
}
}
void RasterizerOpenGL::SetupComputeConstBuffers(const Shader& kernel) {
void RasterizerOpenGL::SetupComputeConstBuffers(Shader* kernel) {
MICROPROFILE_SCOPE(OpenGL_UBO);
const auto& launch_desc = system.GPU().KeplerCompute().launch_description;
const auto& entries = kernel->GetEntries();
@ -969,7 +970,7 @@ void RasterizerOpenGL::SetupConstBuffer(GLenum stage, u32 binding,
}
}
void RasterizerOpenGL::SetupDrawGlobalMemory(std::size_t stage_index, const Shader& shader) {
void RasterizerOpenGL::SetupDrawGlobalMemory(std::size_t stage_index, Shader* shader) {
auto& gpu{system.GPU()};
auto& memory_manager{gpu.MemoryManager()};
const auto cbufs{gpu.Maxwell3D().state.shader_stages[stage_index]};
@ -984,7 +985,7 @@ void RasterizerOpenGL::SetupDrawGlobalMemory(std::size_t stage_index, const Shad
}
}
void RasterizerOpenGL::SetupComputeGlobalMemory(const Shader& kernel) {
void RasterizerOpenGL::SetupComputeGlobalMemory(Shader* kernel) {
auto& gpu{system.GPU()};
auto& memory_manager{gpu.MemoryManager()};
const auto cbufs{gpu.KeplerCompute().launch_description.const_buffer_config};
@ -1007,7 +1008,7 @@ void RasterizerOpenGL::SetupGlobalMemory(u32 binding, const GlobalMemoryEntry& e
static_cast<GLsizeiptr>(size));
}
void RasterizerOpenGL::SetupDrawTextures(std::size_t stage_index, const Shader& shader) {
void RasterizerOpenGL::SetupDrawTextures(std::size_t stage_index, Shader* shader) {
MICROPROFILE_SCOPE(OpenGL_Texture);
const auto& maxwell3d = system.GPU().Maxwell3D();
u32 binding = device.GetBaseBindings(stage_index).sampler;
@ -1020,7 +1021,7 @@ void RasterizerOpenGL::SetupDrawTextures(std::size_t stage_index, const Shader&
}
}
void RasterizerOpenGL::SetupComputeTextures(const Shader& kernel) {
void RasterizerOpenGL::SetupComputeTextures(Shader* kernel) {
MICROPROFILE_SCOPE(OpenGL_Texture);
const auto& compute = system.GPU().KeplerCompute();
u32 binding = 0;
@ -1049,7 +1050,7 @@ void RasterizerOpenGL::SetupTexture(u32 binding, const Tegra::Texture::FullTextu
}
}
void RasterizerOpenGL::SetupDrawImages(std::size_t stage_index, const Shader& shader) {
void RasterizerOpenGL::SetupDrawImages(std::size_t stage_index, Shader* shader) {
const auto& maxwell3d = system.GPU().Maxwell3D();
u32 binding = device.GetBaseBindings(stage_index).image;
for (const auto& entry : shader->GetEntries().images) {
@ -1059,7 +1060,7 @@ void RasterizerOpenGL::SetupDrawImages(std::size_t stage_index, const Shader& sh
}
}
void RasterizerOpenGL::SetupComputeImages(const Shader& shader) {
void RasterizerOpenGL::SetupComputeImages(Shader* shader) {
const auto& compute = system.GPU().KeplerCompute();
u32 binding = 0;
for (const auto& entry : shader->GetEntries().images) {

@ -19,7 +19,6 @@
#include "video_core/engines/const_buffer_info.h"
#include "video_core/engines/maxwell_3d.h"
#include "video_core/rasterizer_accelerated.h"
#include "video_core/rasterizer_cache.h"
#include "video_core/rasterizer_interface.h"
#include "video_core/renderer_opengl/gl_buffer_cache.h"
#include "video_core/renderer_opengl/gl_device.h"
@ -100,10 +99,10 @@ private:
void ConfigureClearFramebuffer(bool using_color, bool using_depth_stencil);
/// Configures the current constbuffers to use for the draw command.
void SetupDrawConstBuffers(std::size_t stage_index, const Shader& shader);
void SetupDrawConstBuffers(std::size_t stage_index, Shader* shader);
/// Configures the current constbuffers to use for the kernel invocation.
void SetupComputeConstBuffers(const Shader& kernel);
void SetupComputeConstBuffers(Shader* kernel);
/// Configures a constant buffer.
void SetupConstBuffer(GLenum stage, u32 binding, const Tegra::Engines::ConstBufferInfo& buffer,
@ -111,30 +110,30 @@ private:
std::size_t unified_offset);
/// Configures the current global memory entries to use for the draw command.
void SetupDrawGlobalMemory(std::size_t stage_index, const Shader& shader);
void SetupDrawGlobalMemory(std::size_t stage_index, Shader* shader);
/// Configures the current global memory entries to use for the kernel invocation.
void SetupComputeGlobalMemory(const Shader& kernel);
void SetupComputeGlobalMemory(Shader* kernel);
/// Configures a constant buffer.
void SetupGlobalMemory(u32 binding, const GlobalMemoryEntry& entry, GPUVAddr gpu_addr,
std::size_t size);
/// Configures the current textures to use for the draw command.
void SetupDrawTextures(std::size_t stage_index, const Shader& shader);
void SetupDrawTextures(std::size_t stage_index, Shader* shader);
/// Configures the textures used in a compute shader.
void SetupComputeTextures(const Shader& kernel);
void SetupComputeTextures(Shader* kernel);
/// Configures a texture.
void SetupTexture(u32 binding, const Tegra::Texture::FullTextureInfo& texture,
const SamplerEntry& entry);
/// Configures images in a graphics shader.
void SetupDrawImages(std::size_t stage_index, const Shader& shader);
void SetupDrawImages(std::size_t stage_index, Shader* shader);
/// Configures images in a compute shader.
void SetupComputeImages(const Shader& shader);
void SetupComputeImages(Shader* shader);
/// Configures an image.
void SetupImage(u32 binding, const Tegra::Texture::TICEntry& tic, const ImageEntry& entry);

@ -29,6 +29,7 @@
#include "video_core/shader/memory_util.h"
#include "video_core/shader/registry.h"
#include "video_core/shader/shader_ir.h"
#include "video_core/shader_cache.h"
namespace OpenGL {
@ -194,12 +195,9 @@ std::unordered_set<GLenum> GetSupportedFormats() {
} // Anonymous namespace
CachedShader::CachedShader(VAddr cpu_addr, std::size_t size_in_bytes,
std::shared_ptr<VideoCommon::Shader::Registry> registry,
ShaderEntries entries, ProgramSharedPtr program_)
: RasterizerCacheObject{cpu_addr}, registry{std::move(registry)}, entries{std::move(entries)},
size_in_bytes{size_in_bytes}, program{std::move(program_)} {
// Assign either the assembly program or source program. We can't have both.
Shader::Shader(std::shared_ptr<VideoCommon::Shader::Registry> registry_, ShaderEntries entries_,
ProgramSharedPtr program_)
: registry{std::move(registry_)}, entries{std::move(entries_)}, program{std::move(program_)} {
handle = program->assembly_program.handle;
if (handle == 0) {
handle = program->source_program.handle;
@ -207,16 +205,16 @@ CachedShader::CachedShader(VAddr cpu_addr, std::size_t size_in_bytes,
ASSERT(handle != 0);
}
CachedShader::~CachedShader() = default;
Shader::~Shader() = default;
GLuint CachedShader::GetHandle() const {
GLuint Shader::GetHandle() const {
DEBUG_ASSERT(registry->IsConsistent());
return handle;
}
Shader CachedShader::CreateStageFromMemory(const ShaderParameters& params,
Maxwell::ShaderProgram program_type, ProgramCode code,
ProgramCode code_b) {
std::unique_ptr<Shader> Shader::CreateStageFromMemory(const ShaderParameters& params,
Maxwell::ShaderProgram program_type,
ProgramCode code, ProgramCode code_b) {
const auto shader_type = GetShaderType(program_type);
const std::size_t size_in_bytes = code.size() * sizeof(u64);
@ -241,12 +239,12 @@ Shader CachedShader::CreateStageFromMemory(const ShaderParameters& params,
entry.bindless_samplers = registry->GetBindlessSamplers();
params.disk_cache.SaveEntry(std::move(entry));
return std::shared_ptr<CachedShader>(
new CachedShader(params.cpu_addr, size_in_bytes, std::move(registry),
MakeEntries(params.device, ir, shader_type), std::move(program)));
return std::unique_ptr<Shader>(new Shader(
std::move(registry), MakeEntries(params.device, ir, shader_type), std::move(program)));
}
Shader CachedShader::CreateKernelFromMemory(const ShaderParameters& params, ProgramCode code) {
std::unique_ptr<Shader> Shader::CreateKernelFromMemory(const ShaderParameters& params,
ProgramCode code) {
const std::size_t size_in_bytes = code.size() * sizeof(u64);
auto& engine = params.system.GPU().KeplerCompute();
@ -266,23 +264,23 @@ Shader CachedShader::CreateKernelFromMemory(const ShaderParameters& params, Prog
entry.bindless_samplers = registry->GetBindlessSamplers();
params.disk_cache.SaveEntry(std::move(entry));
return std::shared_ptr<CachedShader>(
new CachedShader(params.cpu_addr, size_in_bytes, std::move(registry),
MakeEntries(params.device, ir, ShaderType::Compute), std::move(program)));
return std::unique_ptr<Shader>(new Shader(std::move(registry),
MakeEntries(params.device, ir, ShaderType::Compute),
std::move(program)));
}
Shader CachedShader::CreateFromCache(const ShaderParameters& params,
const PrecompiledShader& precompiled_shader,
std::size_t size_in_bytes) {
return std::shared_ptr<CachedShader>(
new CachedShader(params.cpu_addr, size_in_bytes, precompiled_shader.registry,
precompiled_shader.entries, precompiled_shader.program));
std::unique_ptr<Shader> Shader::CreateFromCache(const ShaderParameters& params,
const PrecompiledShader& precompiled_shader) {
return std::unique_ptr<Shader>(new Shader(
precompiled_shader.registry, precompiled_shader.entries, precompiled_shader.program));
}
ShaderCacheOpenGL::ShaderCacheOpenGL(RasterizerOpenGL& rasterizer, Core::System& system,
Core::Frontend::EmuWindow& emu_window, const Device& device)
: RasterizerCache{rasterizer}, system{system}, emu_window{emu_window}, device{device},
disk_cache{system} {}
: VideoCommon::ShaderCache<Shader>{rasterizer}, system{system},
emu_window{emu_window}, device{device}, disk_cache{system} {}
ShaderCacheOpenGL::~ShaderCacheOpenGL() = default;
void ShaderCacheOpenGL::LoadDiskCache(const std::atomic_bool& stop_loading,
const VideoCore::DiskResourceLoadCallback& callback) {
@ -436,7 +434,7 @@ ProgramSharedPtr ShaderCacheOpenGL::GeneratePrecompiledProgram(
return program;
}
Shader ShaderCacheOpenGL::GetStageProgram(Maxwell::ShaderProgram program) {
Shader* ShaderCacheOpenGL::GetStageProgram(Maxwell::ShaderProgram program) {
if (!system.GPU().Maxwell3D().dirty.flags[Dirty::Shaders]) {
return last_shaders[static_cast<std::size_t>(program)];
}
@ -446,8 +444,7 @@ Shader ShaderCacheOpenGL::GetStageProgram(Maxwell::ShaderProgram program) {
// Look up shader in the cache based on address
const auto cpu_addr{memory_manager.GpuToCpuAddress(address)};
Shader shader{cpu_addr ? TryGet(*cpu_addr) : null_shader};
if (shader) {
if (Shader* const shader{cpu_addr ? TryGet(*cpu_addr) : null_shader.get()}) {
return last_shaders[static_cast<std::size_t>(program)] = shader;
}
@ -468,30 +465,29 @@ Shader ShaderCacheOpenGL::GetStageProgram(Maxwell::ShaderProgram program) {
const ShaderParameters params{system, disk_cache, device,
*cpu_addr, host_ptr, unique_identifier};
std::unique_ptr<Shader> shader;
const auto found = runtime_cache.find(unique_identifier);
if (found == runtime_cache.end()) {
shader = CachedShader::CreateStageFromMemory(params, program, std::move(code),
std::move(code_b));
shader = Shader::CreateStageFromMemory(params, program, std::move(code), std::move(code_b));
} else {
const std::size_t size_in_bytes = code.size() * sizeof(u64);
shader = CachedShader::CreateFromCache(params, found->second, size_in_bytes);
shader = Shader::CreateFromCache(params, found->second);
}
Shader* const result = shader.get();
if (cpu_addr) {
Register(shader);
Register(std::move(shader), *cpu_addr, code.size() * sizeof(u64));
} else {
null_shader = shader;
null_shader = std::move(shader);
}
return last_shaders[static_cast<std::size_t>(program)] = shader;
return last_shaders[static_cast<std::size_t>(program)] = result;
}
Shader ShaderCacheOpenGL::GetComputeKernel(GPUVAddr code_addr) {
Shader* ShaderCacheOpenGL::GetComputeKernel(GPUVAddr code_addr) {
auto& memory_manager{system.GPU().MemoryManager()};
const auto cpu_addr{memory_manager.GpuToCpuAddress(code_addr)};
auto kernel = cpu_addr ? TryGet(*cpu_addr) : null_kernel;
if (kernel) {
if (Shader* const kernel = cpu_addr ? TryGet(*cpu_addr) : null_kernel.get()) {
return kernel;
}
@ -503,20 +499,21 @@ Shader ShaderCacheOpenGL::GetComputeKernel(GPUVAddr code_addr) {
const ShaderParameters params{system, disk_cache, device,
*cpu_addr, host_ptr, unique_identifier};
std::unique_ptr<Shader> kernel;
const auto found = runtime_cache.find(unique_identifier);
if (found == runtime_cache.end()) {
kernel = CachedShader::CreateKernelFromMemory(params, std::move(code));
kernel = Shader::CreateKernelFromMemory(params, std::move(code));
} else {
const std::size_t size_in_bytes = code.size() * sizeof(u64);
kernel = CachedShader::CreateFromCache(params, found->second, size_in_bytes);
kernel = Shader::CreateFromCache(params, found->second);
}
Shader* const result = kernel.get();
if (cpu_addr) {
Register(kernel);
Register(std::move(kernel), *cpu_addr, code.size() * sizeof(u64));
} else {
null_kernel = kernel;
null_kernel = std::move(kernel);
}
return kernel;
return result;
}
} // namespace OpenGL

@ -18,12 +18,12 @@
#include "common/common_types.h"
#include "video_core/engines/shader_type.h"
#include "video_core/rasterizer_cache.h"
#include "video_core/renderer_opengl/gl_resource_manager.h"
#include "video_core/renderer_opengl/gl_shader_decompiler.h"
#include "video_core/renderer_opengl/gl_shader_disk_cache.h"
#include "video_core/shader/registry.h"
#include "video_core/shader/shader_ir.h"
#include "video_core/shader_cache.h"
namespace Core {
class System;
@ -35,12 +35,10 @@ class EmuWindow;
namespace OpenGL {
class CachedShader;
class Device;
class RasterizerOpenGL;
struct UnspecializedShader;
using Shader = std::shared_ptr<CachedShader>;
using Maxwell = Tegra::Engines::Maxwell3D::Regs;
struct ProgramHandle {
@ -64,62 +62,53 @@ struct ShaderParameters {
u64 unique_identifier;
};
class CachedShader final : public RasterizerCacheObject {
class Shader final {
public:
~CachedShader();
~Shader();
/// Gets the GL program handle for the shader
GLuint GetHandle() const;
/// Returns the size in bytes of the shader
std::size_t GetSizeInBytes() const override {
return size_in_bytes;
}
/// Gets the shader entries for the shader
const ShaderEntries& GetEntries() const {
return entries;
}
static Shader CreateStageFromMemory(const ShaderParameters& params,
static std::unique_ptr<Shader> CreateStageFromMemory(const ShaderParameters& params,
Maxwell::ShaderProgram program_type,
ProgramCode program_code, ProgramCode program_code_b);
static Shader CreateKernelFromMemory(const ShaderParameters& params, ProgramCode code);
ProgramCode program_code,
ProgramCode program_code_b);
static std::unique_ptr<Shader> CreateKernelFromMemory(const ShaderParameters& params,
ProgramCode code);
static Shader CreateFromCache(const ShaderParameters& params,
const PrecompiledShader& precompiled_shader,
std::size_t size_in_bytes);
static std::unique_ptr<Shader> CreateFromCache(const ShaderParameters& params,
const PrecompiledShader& precompiled_shader);
private:
explicit CachedShader(VAddr cpu_addr, std::size_t size_in_bytes,
std::shared_ptr<VideoCommon::Shader::Registry> registry,
ShaderEntries entries, ProgramSharedPtr program);
explicit Shader(std::shared_ptr<VideoCommon::Shader::Registry> registry, ShaderEntries entries,
ProgramSharedPtr program);
std::shared_ptr<VideoCommon::Shader::Registry> registry;
ShaderEntries entries;
std::size_t size_in_bytes = 0;
ProgramSharedPtr program;
GLuint handle = 0;
};
class ShaderCacheOpenGL final : public RasterizerCache<Shader> {
class ShaderCacheOpenGL final : public VideoCommon::ShaderCache<Shader> {
public:
explicit ShaderCacheOpenGL(RasterizerOpenGL& rasterizer, Core::System& system,
Core::Frontend::EmuWindow& emu_window, const Device& device);
~ShaderCacheOpenGL() override;
/// Loads disk cache for the current game
void LoadDiskCache(const std::atomic_bool& stop_loading,
const VideoCore::DiskResourceLoadCallback& callback);
/// Gets the current specified shader stage program
Shader GetStageProgram(Maxwell::ShaderProgram program);
Shader* GetStageProgram(Maxwell::ShaderProgram program);
/// Gets a compute kernel in the passed address
Shader GetComputeKernel(GPUVAddr code_addr);
protected:
// We do not have to flush this cache as things in it are never modified by us.
void FlushObjectInner(const Shader& object) override {}
Shader* GetComputeKernel(GPUVAddr code_addr);
private:
ProgramSharedPtr GeneratePrecompiledProgram(
@ -132,10 +121,10 @@ private:
ShaderDiskCacheOpenGL disk_cache;
std::unordered_map<u64, PrecompiledShader> runtime_cache;
Shader null_shader{};
Shader null_kernel{};
std::unique_ptr<Shader> null_shader;
std::unique_ptr<Shader> null_kernel;
std::array<Shader, Maxwell::MaxShaderProgram> last_shaders;
std::array<Shader*, Maxwell::MaxShaderProgram> last_shaders{};
};
} // namespace OpenGL

@ -8,7 +8,6 @@
#include "common/common_types.h"
#include "video_core/buffer_cache/buffer_cache.h"
#include "video_core/rasterizer_cache.h"
#include "video_core/renderer_vulkan/vk_memory_manager.h"
#include "video_core/renderer_vulkan/vk_staging_buffer_pool.h"
#include "video_core/renderer_vulkan/vk_stream_buffer.h"

@ -27,6 +27,7 @@
#include "video_core/renderer_vulkan/wrapper.h"
#include "video_core/shader/compiler_settings.h"
#include "video_core/shader/memory_util.h"
#include "video_core/shader_cache.h"
namespace Vulkan {
@ -132,19 +133,18 @@ bool ComputePipelineCacheKey::operator==(const ComputePipelineCacheKey& rhs) con
return std::memcmp(&rhs, this, sizeof *this) == 0;
}
CachedShader::CachedShader(Core::System& system, Tegra::Engines::ShaderType stage,
GPUVAddr gpu_addr, VAddr cpu_addr, ProgramCode program_code,
u32 main_offset)
: RasterizerCacheObject{cpu_addr}, gpu_addr{gpu_addr}, program_code{std::move(program_code)},
Shader::Shader(Core::System& system, Tegra::Engines::ShaderType stage, GPUVAddr gpu_addr,
VideoCommon::Shader::ProgramCode program_code, u32 main_offset)
: gpu_addr{gpu_addr}, program_code{std::move(program_code)},
registry{stage, GetEngine(system, stage)}, shader_ir{this->program_code, main_offset,
compiler_settings, registry},
entries{GenerateShaderEntries(shader_ir)} {}
CachedShader::~CachedShader() = default;
Shader::~Shader() = default;
Tegra::Engines::ConstBufferEngineInterface& CachedShader::GetEngine(
Core::System& system, Tegra::Engines::ShaderType stage) {
if (stage == Tegra::Engines::ShaderType::Compute) {
Tegra::Engines::ConstBufferEngineInterface& Shader::GetEngine(Core::System& system,
Tegra::Engines::ShaderType stage) {
if (stage == ShaderType::Compute) {
return system.GPU().KeplerCompute();
} else {
return system.GPU().Maxwell3D();
@ -156,16 +156,16 @@ VKPipelineCache::VKPipelineCache(Core::System& system, RasterizerVulkan& rasteri
VKDescriptorPool& descriptor_pool,
VKUpdateDescriptorQueue& update_descriptor_queue,
VKRenderPassCache& renderpass_cache)
: RasterizerCache{rasterizer}, system{system}, device{device}, scheduler{scheduler},
descriptor_pool{descriptor_pool}, update_descriptor_queue{update_descriptor_queue},
renderpass_cache{renderpass_cache} {}
: VideoCommon::ShaderCache<Shader>{rasterizer}, system{system}, device{device},
scheduler{scheduler}, descriptor_pool{descriptor_pool},
update_descriptor_queue{update_descriptor_queue}, renderpass_cache{renderpass_cache} {}
VKPipelineCache::~VKPipelineCache() = default;
std::array<Shader, Maxwell::MaxShaderProgram> VKPipelineCache::GetShaders() {
std::array<Shader*, Maxwell::MaxShaderProgram> VKPipelineCache::GetShaders() {
const auto& gpu = system.GPU().Maxwell3D();
std::array<Shader, Maxwell::MaxShaderProgram> shaders;
std::array<Shader*, Maxwell::MaxShaderProgram> shaders{};
for (std::size_t index = 0; index < Maxwell::MaxShaderProgram; ++index) {
const auto program{static_cast<Maxwell::ShaderProgram>(index)};
@ -178,24 +178,28 @@ std::array<Shader, Maxwell::MaxShaderProgram> VKPipelineCache::GetShaders() {
const GPUVAddr program_addr{GetShaderAddress(system, program)};
const std::optional cpu_addr = memory_manager.GpuToCpuAddress(program_addr);
ASSERT(cpu_addr);
auto shader = cpu_addr ? TryGet(*cpu_addr) : null_shader;
if (!shader) {
Shader* result = cpu_addr ? TryGet(*cpu_addr) : null_shader.get();
if (!result) {
const auto host_ptr{memory_manager.GetPointer(program_addr)};
// No shader found - create a new one
constexpr u32 stage_offset = STAGE_MAIN_OFFSET;
const auto stage = static_cast<Tegra::Engines::ShaderType>(index == 0 ? 0 : index - 1);
const auto stage = static_cast<ShaderType>(index == 0 ? 0 : index - 1);
ProgramCode code = GetShaderCode(memory_manager, program_addr, host_ptr, false);
const std::size_t size_in_bytes = code.size() * sizeof(u64);
auto shader = std::make_unique<Shader>(system, stage, program_addr, std::move(code),
stage_offset);
result = shader.get();
shader = std::make_shared<CachedShader>(system, stage, program_addr, *cpu_addr,
std::move(code), stage_offset);
if (cpu_addr) {
Register(shader);
Register(std::move(shader), *cpu_addr, size_in_bytes);
} else {
null_shader = shader;
null_shader = std::move(shader);
}
}
shaders[index] = std::move(shader);
shaders[index] = result;
}
return last_shaders = shaders;
}
@ -236,19 +240,22 @@ VKComputePipeline& VKPipelineCache::GetComputePipeline(const ComputePipelineCach
const auto cpu_addr = memory_manager.GpuToCpuAddress(program_addr);
ASSERT(cpu_addr);
auto shader = cpu_addr ? TryGet(*cpu_addr) : null_kernel;
Shader* shader = cpu_addr ? TryGet(*cpu_addr) : null_kernel.get();
if (!shader) {
// No shader found - create a new one
const auto host_ptr = memory_manager.GetPointer(program_addr);
ProgramCode code = GetShaderCode(memory_manager, program_addr, host_ptr, true);
shader = std::make_shared<CachedShader>(system, Tegra::Engines::ShaderType::Compute,
program_addr, *cpu_addr, std::move(code),
KERNEL_MAIN_OFFSET);
const std::size_t size_in_bytes = code.size() * sizeof(u64);
auto shader_info = std::make_unique<Shader>(system, ShaderType::Compute, program_addr,
std::move(code), KERNEL_MAIN_OFFSET);
shader = shader_info.get();
if (cpu_addr) {
Register(shader);
Register(std::move(shader_info), *cpu_addr, size_in_bytes);
} else {
null_kernel = shader;
null_kernel = std::move(shader_info);
}
}
@ -264,7 +271,7 @@ VKComputePipeline& VKPipelineCache::GetComputePipeline(const ComputePipelineCach
return *entry;
}
void VKPipelineCache::Unregister(const Shader& shader) {
void VKPipelineCache::OnShaderRemoval(Shader* shader) {
bool finished = false;
const auto Finish = [&] {
// TODO(Rodrigo): Instead of finishing here, wait for the fences that use this pipeline and
@ -296,8 +303,6 @@ void VKPipelineCache::Unregister(const Shader& shader) {
Finish();
it = compute_cache.erase(it);
}
RasterizerCache::Unregister(shader);
}
std::pair<SPIRVProgram, std::vector<VkDescriptorSetLayoutBinding>>
@ -332,12 +337,11 @@ VKPipelineCache::DecompileShaders(const GraphicsPipelineCacheKey& key) {
}
const GPUVAddr gpu_addr = GetShaderAddress(system, program_enum);
const auto cpu_addr = memory_manager.GpuToCpuAddress(gpu_addr);
const auto shader = cpu_addr ? TryGet(*cpu_addr) : null_shader;
ASSERT(shader);
const std::optional<VAddr> cpu_addr = memory_manager.GpuToCpuAddress(gpu_addr);
Shader* const shader = cpu_addr ? TryGet(*cpu_addr) : null_shader.get();
const std::size_t stage = index == 0 ? 0 : index - 1; // Stage indices are 0 - 5
const auto program_type = GetShaderType(program_enum);
const ShaderType program_type = GetShaderType(program_enum);
const auto& entries = shader->GetEntries();
program[stage] = {
Decompile(device, shader->GetIR(), program_type, shader->GetRegistry(), specialization),

@ -17,7 +17,6 @@
#include "common/common_types.h"
#include "video_core/engines/const_buffer_engine_interface.h"
#include "video_core/engines/maxwell_3d.h"
#include "video_core/rasterizer_cache.h"
#include "video_core/renderer_vulkan/fixed_pipeline_state.h"
#include "video_core/renderer_vulkan/vk_graphics_pipeline.h"
#include "video_core/renderer_vulkan/vk_renderpass_cache.h"
@ -26,6 +25,7 @@
#include "video_core/shader/memory_util.h"
#include "video_core/shader/registry.h"
#include "video_core/shader/shader_ir.h"
#include "video_core/shader_cache.h"
namespace Core {
class System;
@ -41,8 +41,6 @@ class VKFence;
class VKScheduler;
class VKUpdateDescriptorQueue;
class CachedShader;
using Shader = std::shared_ptr<CachedShader>;
using Maxwell = Tegra::Engines::Maxwell3D::Regs;
struct GraphicsPipelineCacheKey {
@ -102,21 +100,16 @@ struct hash<Vulkan::ComputePipelineCacheKey> {
namespace Vulkan {
class CachedShader final : public RasterizerCacheObject {
class Shader {
public:
explicit CachedShader(Core::System& system, Tegra::Engines::ShaderType stage, GPUVAddr gpu_addr,
VAddr cpu_addr, VideoCommon::Shader::ProgramCode program_code,
u32 main_offset);
~CachedShader();
explicit Shader(Core::System& system, Tegra::Engines::ShaderType stage, GPUVAddr gpu_addr,
VideoCommon::Shader::ProgramCode program_code, u32 main_offset);
~Shader();
GPUVAddr GetGpuAddr() const {
return gpu_addr;
}
std::size_t GetSizeInBytes() const override {
return program_code.size() * sizeof(u64);
}
VideoCommon::Shader::ShaderIR& GetIR() {
return shader_ir;
}
@ -144,25 +137,23 @@ private:
ShaderEntries entries;
};
class VKPipelineCache final : public RasterizerCache<Shader> {
class VKPipelineCache final : public VideoCommon::ShaderCache<Shader> {
public:
explicit VKPipelineCache(Core::System& system, RasterizerVulkan& rasterizer,
const VKDevice& device, VKScheduler& scheduler,
VKDescriptorPool& descriptor_pool,
VKUpdateDescriptorQueue& update_descriptor_queue,
VKRenderPassCache& renderpass_cache);
~VKPipelineCache();
~VKPipelineCache() override;
std::array<Shader, Maxwell::MaxShaderProgram> GetShaders();
std::array<Shader*, Maxwell::MaxShaderProgram> GetShaders();
VKGraphicsPipeline& GetGraphicsPipeline(const GraphicsPipelineCacheKey& key);
VKComputePipeline& GetComputePipeline(const ComputePipelineCacheKey& key);
protected:
void Unregister(const Shader& shader) override;
void FlushObjectInner(const Shader& object) override {}
void OnShaderRemoval(Shader* shader) final;
private:
std::pair<SPIRVProgram, std::vector<VkDescriptorSetLayoutBinding>> DecompileShaders(
@ -175,10 +166,10 @@ private:
VKUpdateDescriptorQueue& update_descriptor_queue;
VKRenderPassCache& renderpass_cache;
Shader null_shader{};
Shader null_kernel{};
std::unique_ptr<Shader> null_shader;
std::unique_ptr<Shader> null_kernel;
std::array<Shader, Maxwell::MaxShaderProgram> last_shaders;
std::array<Shader*, Maxwell::MaxShaderProgram> last_shaders{};
GraphicsPipelineCacheKey last_graphics_key;
VKGraphicsPipeline* last_graphics_pipeline = nullptr;

@ -38,6 +38,7 @@
#include "video_core/renderer_vulkan/vk_texture_cache.h"
#include "video_core/renderer_vulkan/vk_update_descriptor.h"
#include "video_core/renderer_vulkan/wrapper.h"
#include "video_core/shader_cache.h"
namespace Vulkan {
@ -98,7 +99,7 @@ VkRect2D GetScissorState(const Maxwell& regs, std::size_t index) {
}
std::array<GPUVAddr, Maxwell::MaxShaderProgram> GetShaderAddresses(
const std::array<Shader, Maxwell::MaxShaderProgram>& shaders) {
const std::array<Shader*, Maxwell::MaxShaderProgram>& shaders) {
std::array<GPUVAddr, Maxwell::MaxShaderProgram> addresses;
for (std::size_t i = 0; i < std::size(addresses); ++i) {
addresses[i] = shaders[i] ? shaders[i]->GetGpuAddr() : 0;
@ -776,12 +777,12 @@ RasterizerVulkan::DrawParameters RasterizerVulkan::SetupGeometry(FixedPipelineSt
}
void RasterizerVulkan::SetupShaderDescriptors(
const std::array<Shader, Maxwell::MaxShaderProgram>& shaders) {
const std::array<Shader*, Maxwell::MaxShaderProgram>& shaders) {
texture_cache.GuardSamplers(true);
for (std::size_t stage = 0; stage < Maxwell::MaxShaderStage; ++stage) {
// Skip VertexA stage
const auto& shader = shaders[stage + 1];
Shader* const shader = shaders[stage + 1];
if (!shader) {
continue;
}

@ -168,7 +168,7 @@ private:
bool is_indexed, bool is_instanced);
/// Setup descriptors in the graphics pipeline.
void SetupShaderDescriptors(const std::array<Shader, Maxwell::MaxShaderProgram>& shaders);
void SetupShaderDescriptors(const std::array<Shader*, Maxwell::MaxShaderProgram>& shaders);
void SetupImageTransitions(Texceptions texceptions,
const std::array<View, Maxwell::NumRenderTargets>& color_attachments,

@ -0,0 +1,228 @@
// Copyright 2020 yuzu Emulator Project
// Licensed under GPLv2 or any later version
// Refer to the license.txt file included.
#pragma once
#include <algorithm>
#include <memory>
#include <mutex>
#include <unordered_map>
#include <utility>
#include <vector>
#include "common/assert.h"
#include "common/common_types.h"
#include "video_core/rasterizer_interface.h"
namespace VideoCommon {
template <class T>
class ShaderCache {
static constexpr u64 PAGE_SHIFT = 14;
struct Entry {
VAddr addr_start;
VAddr addr_end;
T* data;
bool is_memory_marked = true;
constexpr bool Overlaps(VAddr start, VAddr end) const noexcept {
return start < addr_end && addr_start < end;
}
};
public:
virtual ~ShaderCache() = default;
/// @brief Removes shaders inside a given region
/// @note Checks for ranges
/// @param addr Start address of the invalidation
/// @param size Number of bytes of the invalidation
void InvalidateRegion(VAddr addr, std::size_t size) {
std::scoped_lock lock{invalidation_mutex};
InvalidatePagesInRegion(addr, size);
RemovePendingShaders();
}
/// @brief Unmarks a memory region as cached and marks it for removal
/// @param addr Start address of the CPU write operation
/// @param size Number of bytes of the CPU write operation
void OnCPUWrite(VAddr addr, std::size_t size) {
std::lock_guard lock{invalidation_mutex};
InvalidatePagesInRegion(addr, size);
}
/// @brief Flushes delayed removal operations
void SyncGuestHost() {
std::scoped_lock lock{invalidation_mutex};
RemovePendingShaders();
}
/// @brief Tries to obtain a cached shader starting in a given address
/// @note Doesn't check for ranges, the given address has to be the start of the shader
/// @param addr Start address of the shader, this doesn't cache for region
/// @return Pointer to a valid shader, nullptr when nothing is found
T* TryGet(VAddr addr) const {
std::scoped_lock lock{lookup_mutex};
const auto it = lookup_cache.find(addr);
if (it == lookup_cache.end()) {
return nullptr;
}
return it->second->data;
}
protected:
explicit ShaderCache(VideoCore::RasterizerInterface& rasterizer_) : rasterizer{rasterizer_} {}
/// @brief Register in the cache a given entry
/// @param data Shader to store in the cache
/// @param addr Start address of the shader that will be registered
/// @param size Size in bytes of the shader
void Register(std::unique_ptr<T> data, VAddr addr, std::size_t size) {
std::scoped_lock lock{invalidation_mutex, lookup_mutex};
const VAddr addr_end = addr + size;
Entry* const entry = NewEntry(addr, addr_end, data.get());
const u64 page_end = addr_end >> PAGE_SHIFT;
for (u64 page = addr >> PAGE_SHIFT; page <= page_end; ++page) {
invalidation_cache[page].push_back(entry);
}
storage.push_back(std::move(data));
rasterizer.UpdatePagesCachedCount(addr, size, 1);
}
/// @brief Called when a shader is going to be removed
/// @param shader Shader that will be removed
/// @pre invalidation_cache is locked
/// @pre lookup_mutex is locked
virtual void OnShaderRemoval([[maybe_unused]] T* shader) {}
private:
/// @brief Invalidate pages in a given region
/// @pre invalidation_mutex is locked
void InvalidatePagesInRegion(VAddr addr, std::size_t size) {
const VAddr addr_end = addr + size;
const u64 page_end = addr_end >> PAGE_SHIFT;
for (u64 page = addr >> PAGE_SHIFT; page <= page_end; ++page) {
const auto it = invalidation_cache.find(page);
if (it == invalidation_cache.end()) {
continue;
}
std::vector<Entry*>& entries = it->second;
InvalidatePageEntries(entries, addr, addr_end);
// If there's nothing else in this page, remove it to avoid overpopulating the hash map.
if (entries.empty()) {
invalidation_cache.erase(it);
}
}
}
/// @brief Remove shaders marked for deletion
/// @pre invalidation_mutex is locked
void RemovePendingShaders() {
if (marked_for_removal.empty()) {
return;
}
std::scoped_lock lock{lookup_mutex};
std::vector<T*> removed_shaders;
removed_shaders.reserve(marked_for_removal.size());
for (Entry* const entry : marked_for_removal) {
if (lookup_cache.erase(entry->addr_start) > 0) {
removed_shaders.push_back(entry->data);
}
}
marked_for_removal.clear();
if (!removed_shaders.empty()) {
RemoveShadersFromStorage(std::move(removed_shaders));
}
}
/// @brief Invalidates entries in a given range for the passed page
/// @param entries Vector of entries in the page, it will be modified on overlaps
/// @param addr Start address of the invalidation
/// @param addr_end Non-inclusive end address of the invalidation
/// @pre invalidation_mutex is locked
void InvalidatePageEntries(std::vector<Entry*>& entries, VAddr addr, VAddr addr_end) {
auto it = entries.begin();
while (it != entries.end()) {
Entry* const entry = *it;
if (!entry->Overlaps(addr, addr_end)) {
++it;
continue;
}
UnmarkMemory(entry);
marked_for_removal.push_back(entry);
it = entries.erase(it);
}
}
/// @brief Unmarks an entry from the rasterizer cache
/// @param entry Entry to unmark from memory
void UnmarkMemory(Entry* entry) {
if (!entry->is_memory_marked) {
return;
}
entry->is_memory_marked = false;
const VAddr addr = entry->addr_start;
const std::size_t size = entry->addr_end - addr;
rasterizer.UpdatePagesCachedCount(addr, size, -1);
}
/// @brief Removes a vector of shaders from a list
/// @param removed_shaders Shaders to be removed from the storage, it can contain duplicates
/// @pre invalidation_mutex is locked
/// @pre lookup_mutex is locked
void RemoveShadersFromStorage(std::vector<T*> removed_shaders) {
// Remove duplicates
std::sort(removed_shaders.begin(), removed_shaders.end());
removed_shaders.erase(std::unique(removed_shaders.begin(), removed_shaders.end()),
removed_shaders.end());
// Now that there are no duplicates, we can notify removals
for (T* const shader : removed_shaders) {
OnShaderRemoval(shader);
}
// Remove them from the cache
const auto is_removed = [&removed_shaders](std::unique_ptr<T>& shader) {
return std::find(removed_shaders.begin(), removed_shaders.end(), shader.get()) !=
removed_shaders.end();
};
storage.erase(std::remove_if(storage.begin(), storage.end(), is_removed), storage.end());
}
/// @brief Creates a new entry in the lookup cache and returns its pointer
/// @pre lookup_mutex is locked
Entry* NewEntry(VAddr addr, VAddr addr_end, T* data) {
auto entry = std::make_unique<Entry>(Entry{addr, addr_end, data});
Entry* const entry_pointer = entry.get();
lookup_cache.emplace(addr, std::move(entry));
return entry_pointer;
}
VideoCore::RasterizerInterface& rasterizer;
mutable std::mutex lookup_mutex;
std::mutex invalidation_mutex;
std::unordered_map<u64, std::unique_ptr<Entry>> lookup_cache;
std::unordered_map<u64, std::vector<Entry*>> invalidation_cache;
std::vector<std::unique_ptr<T>> storage;
std::vector<Entry*> marked_for_removal;
};
} // namespace VideoCommon