mirror of https://github.com/yuzu-mirror/yuzu
Merge pull request #4273 from ogniK5377/async-shaders-prod
video_core: Add asynchronous shader decompilation and compilationpull/8/head
bunnei
5 years ago
committed by
GitHub
commit
90cbcaa44a
@ -0,0 +1,181 @@
|
||||
// Copyright 2020 yuzu Emulator Project
|
||||
// Licensed under GPLv2 or any later version
|
||||
// Refer to the license.txt file included.
|
||||
|
||||
#include <chrono>
|
||||
#include <condition_variable>
|
||||
#include <mutex>
|
||||
#include <thread>
|
||||
#include <vector>
|
||||
#include "video_core/engines/maxwell_3d.h"
|
||||
#include "video_core/renderer_base.h"
|
||||
#include "video_core/renderer_opengl/gl_shader_cache.h"
|
||||
#include "video_core/shader/async_shaders.h"
|
||||
|
||||
namespace VideoCommon::Shader {
|
||||
|
||||
AsyncShaders::AsyncShaders(Core::Frontend::EmuWindow& emu_window) : emu_window(emu_window) {}
|
||||
|
||||
AsyncShaders::~AsyncShaders() {
|
||||
KillWorkers();
|
||||
}
|
||||
|
||||
void AsyncShaders::AllocateWorkers(std::size_t num_workers) {
|
||||
// If we're already have workers queued or don't want to queue workers, ignore
|
||||
if (num_workers == worker_threads.size() || num_workers == 0) {
|
||||
return;
|
||||
}
|
||||
|
||||
// If workers already exist, clear them
|
||||
if (!worker_threads.empty()) {
|
||||
FreeWorkers();
|
||||
}
|
||||
|
||||
// Create workers
|
||||
for (std::size_t i = 0; i < num_workers; i++) {
|
||||
context_list.push_back(emu_window.CreateSharedContext());
|
||||
worker_threads.push_back(std::move(
|
||||
std::thread(&AsyncShaders::ShaderCompilerThread, this, context_list[i].get())));
|
||||
}
|
||||
}
|
||||
|
||||
void AsyncShaders::FreeWorkers() {
|
||||
// Mark all threads to quit
|
||||
is_thread_exiting.store(true);
|
||||
cv.notify_all();
|
||||
for (auto& thread : worker_threads) {
|
||||
thread.join();
|
||||
}
|
||||
// Clear our shared contexts
|
||||
context_list.clear();
|
||||
|
||||
// Clear our worker threads
|
||||
worker_threads.clear();
|
||||
}
|
||||
|
||||
void AsyncShaders::KillWorkers() {
|
||||
is_thread_exiting.store(true);
|
||||
for (auto& thread : worker_threads) {
|
||||
thread.detach();
|
||||
}
|
||||
// Clear our shared contexts
|
||||
context_list.clear();
|
||||
|
||||
// Clear our worker threads
|
||||
worker_threads.clear();
|
||||
}
|
||||
|
||||
bool AsyncShaders::HasWorkQueued() {
|
||||
return !pending_queue.empty();
|
||||
}
|
||||
|
||||
bool AsyncShaders::HasCompletedWork() {
|
||||
std::shared_lock lock{completed_mutex};
|
||||
return !finished_work.empty();
|
||||
}
|
||||
|
||||
bool AsyncShaders::IsShaderAsync(const Tegra::GPU& gpu) const {
|
||||
const auto& regs = gpu.Maxwell3D().regs;
|
||||
|
||||
// If something is using depth, we can assume that games are not rendering anything which will
|
||||
// be used one time.
|
||||
if (regs.zeta_enable) {
|
||||
return true;
|
||||
}
|
||||
|
||||
// If games are using a small index count, we can assume these are full screen quads. Usually
|
||||
// these shaders are only used once for building textures so we can assume they can't be built
|
||||
// async
|
||||
if (regs.index_array.count <= 6 || regs.vertex_buffer.count <= 6) {
|
||||
return false;
|
||||
}
|
||||
|
||||
return true;
|
||||
}
|
||||
|
||||
std::vector<AsyncShaders::Result> AsyncShaders::GetCompletedWork() {
|
||||
std::vector<AsyncShaders::Result> results;
|
||||
{
|
||||
std::unique_lock lock{completed_mutex};
|
||||
results.assign(std::make_move_iterator(finished_work.begin()),
|
||||
std::make_move_iterator(finished_work.end()));
|
||||
finished_work.clear();
|
||||
}
|
||||
return results;
|
||||
}
|
||||
|
||||
void AsyncShaders::QueueOpenGLShader(const OpenGL::Device& device,
|
||||
Tegra::Engines::ShaderType shader_type, u64 uid,
|
||||
std::vector<u64> code, std::vector<u64> code_b,
|
||||
u32 main_offset,
|
||||
VideoCommon::Shader::CompilerSettings compiler_settings,
|
||||
const VideoCommon::Shader::Registry& registry,
|
||||
VAddr cpu_addr) {
|
||||
WorkerParams params{device.UseAssemblyShaders() ? AsyncShaders::Backend::GLASM
|
||||
: AsyncShaders::Backend::OpenGL,
|
||||
device,
|
||||
shader_type,
|
||||
uid,
|
||||
std::move(code),
|
||||
std::move(code_b),
|
||||
main_offset,
|
||||
compiler_settings,
|
||||
registry,
|
||||
cpu_addr};
|
||||
std::unique_lock lock(queue_mutex);
|
||||
pending_queue.push_back(std::move(params));
|
||||
cv.notify_one();
|
||||
}
|
||||
|
||||
void AsyncShaders::ShaderCompilerThread(Core::Frontend::GraphicsContext* context) {
|
||||
using namespace std::chrono_literals;
|
||||
while (!is_thread_exiting.load(std::memory_order_relaxed)) {
|
||||
std::unique_lock lock{queue_mutex};
|
||||
cv.wait(lock, [this] { return HasWorkQueued() || is_thread_exiting; });
|
||||
if (is_thread_exiting) {
|
||||
return;
|
||||
}
|
||||
|
||||
// Partial lock to allow all threads to read at the same time
|
||||
if (!HasWorkQueued()) {
|
||||
continue;
|
||||
}
|
||||
// Another thread beat us, just unlock and wait for the next load
|
||||
if (pending_queue.empty()) {
|
||||
continue;
|
||||
}
|
||||
// Pull work from queue
|
||||
WorkerParams work = std::move(pending_queue.front());
|
||||
pending_queue.pop_front();
|
||||
|
||||
lock.unlock();
|
||||
|
||||
if (work.backend == AsyncShaders::Backend::OpenGL ||
|
||||
work.backend == AsyncShaders::Backend::GLASM) {
|
||||
const ShaderIR ir(work.code, work.main_offset, work.compiler_settings, work.registry);
|
||||
const auto scope = context->Acquire();
|
||||
auto program =
|
||||
OpenGL::BuildShader(work.device, work.shader_type, work.uid, ir, work.registry);
|
||||
Result result{};
|
||||
result.backend = work.backend;
|
||||
result.cpu_address = work.cpu_address;
|
||||
result.uid = work.uid;
|
||||
result.code = std::move(work.code);
|
||||
result.code_b = std::move(work.code_b);
|
||||
result.shader_type = work.shader_type;
|
||||
|
||||
if (work.backend == AsyncShaders::Backend::OpenGL) {
|
||||
result.program.opengl = std::move(program->source_program);
|
||||
} else if (work.backend == AsyncShaders::Backend::GLASM) {
|
||||
result.program.glasm = std::move(program->assembly_program);
|
||||
}
|
||||
|
||||
{
|
||||
std::unique_lock complete_lock(completed_mutex);
|
||||
finished_work.push_back(std::move(result));
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
} // namespace VideoCommon::Shader
|
||||
@ -0,0 +1,109 @@
|
||||
// Copyright 2020 yuzu Emulator Project
|
||||
// Licensed under GPLv2 or any later version
|
||||
// Refer to the license.txt file included.
|
||||
|
||||
#pragma once
|
||||
|
||||
#include <condition_variable>
|
||||
#include <deque>
|
||||
#include <memory>
|
||||
#include <shared_mutex>
|
||||
#include <thread>
|
||||
#include "common/bit_field.h"
|
||||
#include "common/common_types.h"
|
||||
#include "video_core/renderer_opengl/gl_device.h"
|
||||
#include "video_core/renderer_opengl/gl_resource_manager.h"
|
||||
#include "video_core/renderer_opengl/gl_shader_decompiler.h"
|
||||
|
||||
namespace Core::Frontend {
|
||||
class EmuWindow;
|
||||
class GraphicsContext;
|
||||
} // namespace Core::Frontend
|
||||
|
||||
namespace Tegra {
|
||||
class GPU;
|
||||
}
|
||||
|
||||
namespace VideoCommon::Shader {
|
||||
|
||||
class AsyncShaders {
|
||||
public:
|
||||
enum class Backend {
|
||||
OpenGL,
|
||||
GLASM,
|
||||
};
|
||||
|
||||
struct ResultPrograms {
|
||||
OpenGL::OGLProgram opengl;
|
||||
OpenGL::OGLAssemblyProgram glasm;
|
||||
};
|
||||
|
||||
struct Result {
|
||||
u64 uid;
|
||||
VAddr cpu_address;
|
||||
Backend backend;
|
||||
ResultPrograms program;
|
||||
std::vector<u64> code;
|
||||
std::vector<u64> code_b;
|
||||
Tegra::Engines::ShaderType shader_type;
|
||||
};
|
||||
|
||||
explicit AsyncShaders(Core::Frontend::EmuWindow& emu_window);
|
||||
~AsyncShaders();
|
||||
|
||||
/// Start up shader worker threads
|
||||
void AllocateWorkers(std::size_t num_workers);
|
||||
|
||||
/// Clear the shader queue and kill all worker threads
|
||||
void FreeWorkers();
|
||||
|
||||
// Force end all threads
|
||||
void KillWorkers();
|
||||
|
||||
/// Check to see if any shaders have actually been compiled
|
||||
bool HasCompletedWork();
|
||||
|
||||
/// Deduce if a shader can be build on another thread of MUST be built in sync. We cannot build
|
||||
/// every shader async as some shaders are only built and executed once. We try to "guess" which
|
||||
/// shader would be used only once
|
||||
bool IsShaderAsync(const Tegra::GPU& gpu) const;
|
||||
|
||||
/// Pulls completed compiled shaders
|
||||
std::vector<Result> GetCompletedWork();
|
||||
|
||||
void QueueOpenGLShader(const OpenGL::Device& device, Tegra::Engines::ShaderType shader_type,
|
||||
u64 uid, std::vector<u64> code, std::vector<u64> code_b, u32 main_offset,
|
||||
VideoCommon::Shader::CompilerSettings compiler_settings,
|
||||
const VideoCommon::Shader::Registry& registry, VAddr cpu_addr);
|
||||
|
||||
private:
|
||||
void ShaderCompilerThread(Core::Frontend::GraphicsContext* context);
|
||||
|
||||
/// Check our worker queue to see if we have any work queued already
|
||||
bool HasWorkQueued();
|
||||
|
||||
struct WorkerParams {
|
||||
AsyncShaders::Backend backend;
|
||||
OpenGL::Device device;
|
||||
Tegra::Engines::ShaderType shader_type;
|
||||
u64 uid;
|
||||
std::vector<u64> code;
|
||||
std::vector<u64> code_b;
|
||||
u32 main_offset;
|
||||
VideoCommon::Shader::CompilerSettings compiler_settings;
|
||||
VideoCommon::Shader::Registry registry;
|
||||
VAddr cpu_address;
|
||||
};
|
||||
|
||||
std::condition_variable cv;
|
||||
std::mutex queue_mutex;
|
||||
std::shared_mutex completed_mutex;
|
||||
std::atomic<bool> is_thread_exiting{};
|
||||
std::vector<std::unique_ptr<Core::Frontend::GraphicsContext>> context_list;
|
||||
std::vector<std::thread> worker_threads;
|
||||
std::deque<WorkerParams> pending_queue;
|
||||
std::vector<AsyncShaders::Result> finished_work;
|
||||
Core::Frontend::EmuWindow& emu_window;
|
||||
};
|
||||
|
||||
} // namespace VideoCommon::Shader
|
||||
@ -0,0 +1,42 @@
|
||||
// Copyright 2020 yuzu Emulator Project
|
||||
// Licensed under GPLv2 or any later version
|
||||
// Refer to the license.txt file included.
|
||||
|
||||
#include "video_core/shader_notify.h"
|
||||
|
||||
using namespace std::chrono_literals;
|
||||
|
||||
namespace VideoCore {
|
||||
namespace {
|
||||
constexpr auto UPDATE_TICK = 32ms;
|
||||
}
|
||||
|
||||
ShaderNotify::ShaderNotify() = default;
|
||||
ShaderNotify::~ShaderNotify() = default;
|
||||
|
||||
std::size_t ShaderNotify::GetShadersBuilding() {
|
||||
const auto now = std::chrono::high_resolution_clock::now();
|
||||
const auto diff = now - last_update;
|
||||
if (diff > UPDATE_TICK) {
|
||||
std::shared_lock lock(mutex);
|
||||
last_updated_count = accurate_count;
|
||||
}
|
||||
return last_updated_count;
|
||||
}
|
||||
|
||||
std::size_t ShaderNotify::GetShadersBuildingAccurate() {
|
||||
std::shared_lock lock{mutex};
|
||||
return accurate_count;
|
||||
}
|
||||
|
||||
void ShaderNotify::MarkShaderComplete() {
|
||||
std::unique_lock lock{mutex};
|
||||
accurate_count--;
|
||||
}
|
||||
|
||||
void ShaderNotify::MarkSharderBuilding() {
|
||||
std::unique_lock lock{mutex};
|
||||
accurate_count++;
|
||||
}
|
||||
|
||||
} // namespace VideoCore
|
||||
@ -0,0 +1,29 @@
|
||||
// Copyright 2020 yuzu Emulator Project
|
||||
// Licensed under GPLv2 or any later version
|
||||
// Refer to the license.txt file included.
|
||||
|
||||
#pragma once
|
||||
|
||||
#include <chrono>
|
||||
#include <shared_mutex>
|
||||
#include "common/common_types.h"
|
||||
|
||||
namespace VideoCore {
|
||||
class ShaderNotify {
|
||||
public:
|
||||
ShaderNotify();
|
||||
~ShaderNotify();
|
||||
|
||||
std::size_t GetShadersBuilding();
|
||||
std::size_t GetShadersBuildingAccurate();
|
||||
|
||||
void MarkShaderComplete();
|
||||
void MarkSharderBuilding();
|
||||
|
||||
private:
|
||||
std::size_t last_updated_count{};
|
||||
std::size_t accurate_count{};
|
||||
std::shared_mutex mutex;
|
||||
std::chrono::high_resolution_clock::time_point last_update{};
|
||||
};
|
||||
} // namespace VideoCore
|
||||
Loading…
Reference in New Issue