Move Shader JIT Multithreading to a Separate Branch

This commit is contained in:
jbm11208 2025-05-20 12:01:32 -04:00 committed by OpenSauce
parent 6e090f428c
commit f831d9ed47
2 changed files with 15 additions and 138 deletions

View file

@ -1,4 +1,4 @@
// Copyright Citra Emulator Project / Azahar Emulator Project
// Copyright 2016 Citra Emulator Project
// Licensed under GPLv2 or any later version
// Refer to the license.txt file included.
@ -16,79 +16,11 @@
#if CITRA_ARCH(x86_64)
#include "video_core/shader/shader_jit_x64_compiler.h"
#endif
#include <future>
namespace Pica::Shader {
JitEngine::JitEngine() {
stub_shader = std::make_unique<JitShader>();
// Optionally, compile a minimal stub shader here if needed
StartThreadPool(std::thread::hardware_concurrency());
}
JitEngine::~JitEngine() {
StopThreadPool();
}
void JitEngine::StartThreadPool(size_t num_threads) {
stop_threads = false;
for (size_t i = 0; i < num_threads; ++i) {
thread_pool.emplace_back([this]() { ThreadWorker(); });
}
}
void JitEngine::StopThreadPool() {
{
std::lock_guard<std::mutex> lock(queue_mutex);
stop_threads = true;
}
queue_cv.notify_all();
for (auto& t : thread_pool) {
if (t.joinable())
t.join();
}
thread_pool.clear();
}
void JitEngine::ThreadWorker() {
while (true) {
std::function<void()> job;
{
std::unique_lock<std::mutex> lock(queue_mutex);
queue_cv.wait(lock, [this]() { return stop_threads || !compile_queue.empty(); });
if (stop_threads && compile_queue.empty())
return;
job = std::move(compile_queue.front());
compile_queue.pop();
}
job();
}
}
void JitEngine::EnqueueCompilation(u64 cache_key, ShaderSetup setup_copy) {
// WARNING: Copying ShaderSetup across threads may be unsafe if it contains raw pointers or
// non-trivial resources. Consider refactoring to only copy the necessary data for compilation.
auto promise = std::make_shared<std::promise<std::unique_ptr<JitShader>>>();
{
std::lock_guard<std::mutex> lock(queue_mutex);
compile_queue.emplace([this, cache_key, setup_copy, promise]() mutable {
auto shader = std::make_unique<JitShader>();
shader->Compile(&setup_copy.program_code, &setup_copy.swizzle_data);
{
std::lock_guard<std::mutex> lock2(cache_mutex);
if (cache.size() >= MAX_CACHE_SIZE) {
EvictLRU();
}
promise->set_value(std::move(shader));
cache[cache_key] = promise->get_future().share();
lru_list.push_front(cache_key);
}
});
// Store the future in the cache immediately so SetupBatch can wait on it
cache[cache_key] = promise->get_future().share();
}
queue_cv.notify_one();
}
JitEngine::JitEngine() = default;
JitEngine::~JitEngine() = default;
void JitEngine::SetupBatch(ShaderSetup& setup, u32 entry_point) {
ASSERT(entry_point < MAX_PROGRAM_CODE_LENGTH);
@ -96,54 +28,23 @@ void JitEngine::SetupBatch(ShaderSetup& setup, u32 entry_point) {
const u64 code_hash = setup.GetProgramCodeHash();
const u64 swizzle_hash = setup.GetSwizzleDataHash();
const u64 cache_key = Common::HashCombine(code_hash, swizzle_hash);
std::shared_future<std::unique_ptr<JitShader>> shader_future;
{
std::lock_guard<std::mutex> lock(cache_mutex);
auto iter = cache.find(cache_key);
if (iter != cache.end()) {
shader_future = iter->second;
UpdateLRU(cache_key);
} else {
// Compile synchronously and store the result
auto shader = std::make_unique<JitShader>();
shader->Compile(&setup.program_code, &setup.swizzle_data);
auto ready_future = std::make_shared<std::promise<std::unique_ptr<JitShader>>>();
ready_future->set_value(std::move(shader));
shader_future = ready_future->get_future().share();
cache[cache_key] = shader_future;
lru_list.push_front(cache_key);
}
auto iter = cache.find(cache_key);
if (iter != cache.end()) {
setup.cached_shader = iter->second.get();
} else {
auto shader = std::make_unique<JitShader>();
shader->Compile(&setup.program_code, &setup.swizzle_data);
setup.cached_shader = shader.get();
cache.emplace_hint(iter, cache_key, std::move(shader));
}
// Wait for the shader to be ready (if compiling in background)
setup.cached_shader = shader_future.get().get();
}
void JitEngine::EvictLRU() {
if (lru_list.empty()) {
return;
}
const u64 key = lru_list.back();
lru_list.pop_back();
cache.erase(key);
}
void JitEngine::UpdateLRU(u64 key) {
auto it = std::find(lru_list.begin(), lru_list.end(), key);
if (it != lru_list.end()) {
lru_list.erase(it);
}
lru_list.push_front(key);
}
MICROPROFILE_DECLARE(GPU_Shader);
void JitEngine::Run(const ShaderSetup& setup, ShaderUnit& state) const {
// Null check: skip draw if shader is not ready
if (!setup.cached_shader) {
return;
}
ASSERT(setup.cached_shader != nullptr);
MICROPROFILE_SCOPE(GPU_Shader);

View file

@ -1,4 +1,4 @@
// Copyright Citra Emulator Project / Azahar Emulator Project
// Copyright 2016 Citra Emulator Project
// Licensed under GPLv2 or any later version
// Refer to the license.txt file included.
@ -7,14 +7,8 @@
#include "common/arch.h"
#if CITRA_ARCH(x86_64) || CITRA_ARCH(arm64)
#include <functional>
#include <future>
#include <list>
#include <memory>
#include <mutex>
#include <thread>
#include <unordered_map>
#include <queue>
#include "common/common_types.h"
#include "video_core/shader/shader.h"
@ -31,25 +25,7 @@ public:
void Run(const ShaderSetup& setup, ShaderUnit& state) const override;
private:
static constexpr size_t MAX_CACHE_SIZE = 1000; // Maximum number of shaders to cache
std::unordered_map<u64, std::shared_future<std::unique_ptr<JitShader>>> cache;
std::list<u64> lru_list; // Track LRU order of shaders
mutable std::mutex cache_mutex;
// Parallel compilation support
std::vector<std::thread> thread_pool;
std::queue<std::function<void()>> compile_queue;
std::mutex queue_mutex;
std::condition_variable queue_cv;
bool stop_threads = false;
std::unique_ptr<JitShader> stub_shader;
void EvictLRU();
void UpdateLRU(u64 key);
void ThreadWorker();
void EnqueueCompilation(u64 cache_key, ShaderSetup setup_copy);
void StartThreadPool(size_t num_threads);
void StopThreadPool();
std::unordered_map<u64, std::unique_ptr<JitShader>> cache;
};
} // namespace Pica::Shader