Merge pull request #4443 from ameerj/vk-async-shaders

vulkan_renderer: Async shader/graphics pipeline compilation
This commit is contained in:
David 2020-08-17 15:06:11 +10:00 committed by GitHub
commit cbaf1bc711
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
15 changed files with 210 additions and 88 deletions

View file

@ -2,7 +2,6 @@
// Licensed under GPLv2 or any later version
// Refer to the license.txt file included.
#include <chrono>
#include <condition_variable>
#include <mutex>
#include <thread>
@ -20,9 +19,18 @@ 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) {
void AsyncShaders::AllocateWorkers() {
// Max worker threads we should allow
constexpr u32 MAX_THREADS = 4;
// Deduce how many threads we can use
const u32 threads_used = std::thread::hardware_concurrency() / 4;
// Always allow at least 1 thread regardless of our settings
const auto max_worker_count = std::max(1U, threads_used);
// Don't use more than MAX_THREADS
const auto num_workers = std::min(max_worker_count, MAX_THREADS);
// If we already have workers queued, ignore
if (num_workers == worker_threads.size()) {
return;
}
@ -111,24 +119,50 @@ void AsyncShaders::QueueOpenGLShader(const OpenGL::Device& device,
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};
WorkerParams params{
.backend = device.UseAssemblyShaders() ? Backend::GLASM : Backend::OpenGL,
.device = &device,
.shader_type = shader_type,
.uid = uid,
.code = std::move(code),
.code_b = std::move(code_b),
.main_offset = main_offset,
.compiler_settings = compiler_settings,
.registry = registry,
.cpu_address = cpu_addr,
};
std::unique_lock lock(queue_mutex);
pending_queue.push_back(std::move(params));
pending_queue.push(std::move(params));
cv.notify_one();
}
void AsyncShaders::QueueVulkanShader(Vulkan::VKPipelineCache* pp_cache,
const Vulkan::VKDevice& device, Vulkan::VKScheduler& scheduler,
Vulkan::VKDescriptorPool& descriptor_pool,
Vulkan::VKUpdateDescriptorQueue& update_descriptor_queue,
Vulkan::VKRenderPassCache& renderpass_cache,
std::vector<VkDescriptorSetLayoutBinding> bindings,
Vulkan::SPIRVProgram program,
Vulkan::GraphicsPipelineCacheKey key) {
WorkerParams params{
.backend = Backend::Vulkan,
.pp_cache = pp_cache,
.vk_device = &device,
.scheduler = &scheduler,
.descriptor_pool = &descriptor_pool,
.update_descriptor_queue = &update_descriptor_queue,
.renderpass_cache = &renderpass_cache,
.bindings = bindings,
.program = program,
.key = key,
};
std::unique_lock lock(queue_mutex);
pending_queue.push(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; });
@ -144,18 +178,17 @@ void AsyncShaders::ShaderCompilerThread(Core::Frontend::GraphicsContext* context
if (pending_queue.empty()) {
continue;
}
// Pull work from queue
WorkerParams work = std::move(pending_queue.front());
pending_queue.pop_front();
pending_queue.pop();
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);
if (work.backend == Backend::OpenGL || work.backend == 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);
OpenGL::BuildShader(*work.device, work.shader_type, work.uid, ir, *work.registry);
Result result{};
result.backend = work.backend;
result.cpu_address = work.cpu_address;
@ -164,9 +197,9 @@ void AsyncShaders::ShaderCompilerThread(Core::Frontend::GraphicsContext* context
result.code_b = std::move(work.code_b);
result.shader_type = work.shader_type;
if (work.backend == AsyncShaders::Backend::OpenGL) {
if (work.backend == Backend::OpenGL) {
result.program.opengl = std::move(program->source_program);
} else if (work.backend == AsyncShaders::Backend::GLASM) {
} else if (work.backend == Backend::GLASM) {
result.program.glasm = std::move(program->assembly_program);
}
@ -174,6 +207,13 @@ void AsyncShaders::ShaderCompilerThread(Core::Frontend::GraphicsContext* context
std::unique_lock complete_lock(completed_mutex);
finished_work.push_back(std::move(result));
}
} else if (work.backend == Backend::Vulkan) {
auto pipeline = std::make_unique<Vulkan::VKGraphicsPipeline>(
*work.vk_device, *work.scheduler, *work.descriptor_pool,
*work.update_descriptor_queue, *work.renderpass_cache, work.key, work.bindings,
work.program);
work.pp_cache->EmplacePipeline(std::move(pipeline));
}
}
}

View file

@ -14,6 +14,10 @@
#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"
#include "video_core/renderer_vulkan/vk_device.h"
#include "video_core/renderer_vulkan/vk_pipeline_cache.h"
#include "video_core/renderer_vulkan/vk_scheduler.h"
#include "video_core/renderer_vulkan/vk_update_descriptor.h"
namespace Core::Frontend {
class EmuWindow;
@ -24,6 +28,10 @@ namespace Tegra {
class GPU;
}
namespace Vulkan {
class VKPipelineCache;
}
namespace VideoCommon::Shader {
class AsyncShaders {
@ -31,6 +39,7 @@ public:
enum class Backend {
OpenGL,
GLASM,
Vulkan,
};
struct ResultPrograms {
@ -52,7 +61,7 @@ public:
~AsyncShaders();
/// Start up shader worker threads
void AllocateWorkers(std::size_t num_workers);
void AllocateWorkers();
/// Clear the shader queue and kill all worker threads
void FreeWorkers();
@ -76,6 +85,14 @@ public:
VideoCommon::Shader::CompilerSettings compiler_settings,
const VideoCommon::Shader::Registry& registry, VAddr cpu_addr);
void QueueVulkanShader(Vulkan::VKPipelineCache* pp_cache, const Vulkan::VKDevice& device,
Vulkan::VKScheduler& scheduler,
Vulkan::VKDescriptorPool& descriptor_pool,
Vulkan::VKUpdateDescriptorQueue& update_descriptor_queue,
Vulkan::VKRenderPassCache& renderpass_cache,
std::vector<VkDescriptorSetLayoutBinding> bindings,
Vulkan::SPIRVProgram program, Vulkan::GraphicsPipelineCacheKey key);
private:
void ShaderCompilerThread(Core::Frontend::GraphicsContext* context);
@ -83,16 +100,28 @@ private:
bool HasWorkQueued();
struct WorkerParams {
AsyncShaders::Backend backend;
OpenGL::Device device;
Backend backend;
// For OGL
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;
VideoCommon::Shader::Registry registry;
std::optional<VideoCommon::Shader::Registry> registry;
VAddr cpu_address;
// For Vulkan
Vulkan::VKPipelineCache* pp_cache;
const Vulkan::VKDevice* vk_device;
Vulkan::VKScheduler* scheduler;
Vulkan::VKDescriptorPool* descriptor_pool;
Vulkan::VKUpdateDescriptorQueue* update_descriptor_queue;
Vulkan::VKRenderPassCache* renderpass_cache;
std::vector<VkDescriptorSetLayoutBinding> bindings;
Vulkan::SPIRVProgram program;
Vulkan::GraphicsPipelineCacheKey key;
};
std::condition_variable cv;
@ -101,7 +130,7 @@ private:
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::queue<WorkerParams> pending_queue;
std::vector<AsyncShaders::Result> finished_work;
Core::Frontend::EmuWindow& emu_window;
};