raphiaura/KytyPS5
1
0
Fork 0
mirror of https://github.com/KytyPS5/KytyPS5.git synced 2026-10-20 10:56:58 +02:00
Code Issues Packages Projects Releases Wiki Activity

Fix unnecessary queue drains in predication and suspend points

Browse Source
This commit is contained in:
nmzik
2026-09-30 16:51:40 +04:00
parent 05057c9441
commit 840b9f5753
5 changed files with 289 additions and 110 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/graphics/guest_gpu/command_processor/commandProcessor.h
-1
View File
@@ -62,7 +62,6 @@ public:
void BufferInit();
void BufferFlush();
void BufferFlushAndWait();
void BufferWait();
HW::Context& GetCtx() { return m_ctx; }
HW::UserConfig& GetUcfg() { return m_ucfg; }
src/graphics/guest_gpu/graphicsRun.cpp
+24 -78
View File
@@ -35,7 +35,6 @@ namespace Libs::Graphics {
static thread_local CommandProcessor* g_current_processor = nullptr;
static thread_local Pm4Execution* g_current_execution = nullptr;
static thread_local bool g_gpu_mutex_owned = false;
static thread_local bool g_gpu_thread = false;
static thread_local GuestGpu* g_gpu_state = nullptr;
@@ -54,27 +53,6 @@ struct DrawIndexedIndirectArgs {
uint32_t start_instance_location;
};
class GpuMutexLock final {
public:
explicit GpuMutexLock(Common::Mutex& mutex): m_mutex(mutex) {
if (g_gpu_mutex_owned) {
EXIT("recursive GPU mutex acquisition\n");
}
g_gpu_mutex_owned = true;
m_mutex.Lock();
}
~GpuMutexLock() {
if (!g_gpu_mutex_owned) {
EXIT("invalid GPU mutex release\n");
}
m_mutex.Unlock();
g_gpu_mutex_owned = false;
}
private:
Common::Mutex& m_mutex;
};
static bool GraphicsRunDebugDumpEnabled() {
return Config::GraphicsDebugDumpEnabled() &&
Config::GetPrintfDirection() != Config::LogDirection::Silent;
@@ -84,6 +62,7 @@ GuestGpu::GuestGpu(RenderContext& renderer): m_renderer(renderer) {
EXIT_NOT_IMPLEMENTED(!Common::Thread::IsMainThread());
GraphicsInitJmpTables();
m_gfx_cp = std::make_unique<CommandProcessor>(renderer, 0);
m_gfx_cp->Reset();
m_thread = std::jthread(ThreadRun, this);
}
@@ -160,20 +139,16 @@ void GuestGpu::Submit(std::span<const uint32_t> draw_commands,
if (draw_commands.empty()) {
return;
}
GpuMutexLock lock(m_submission_mutex);
Submission submission;
Submission submission;
submission.type = SubmissionType::Graphics;
submission.queue_id = 0;
submission.commands = draw_commands;
submission.constant_commands = constant_commands;
submission.reset_processor = m_graphics_done;
m_graphics_done = false;
Enqueue(std::move(submission));
}
void GuestGpu::SubmitCompute(uint32_t queue, std::span<const uint32_t> commands) {
EXIT_IF(commands.empty());
GpuMutexLock lock(m_submission_mutex);
EXIT_NOT_IMPLEMENTED(queue < ComputeQueueBase || queue >= ComputeQueueBase + ComputeQueueCount);
@@ -186,22 +161,21 @@ void GuestGpu::SubmitCompute(uint32_t queue, std::span<const uint32_t> commands)
}
void GuestGpu::SubmitFlipPreparation(uint64_t request_id) {
GpuMutexLock lock(m_submission_mutex);
Submission submission;
Submission submission;
submission.type = SubmissionType::FlipPreparation;
submission.queue_id = 0;
submission.reset_processor = m_graphics_done;
submission.flip_request_id = request_id;
m_graphics_done = false;
Enqueue(std::move(submission));
}
void GuestGpu::Done() {
GpuMutexLock lock(m_submission_mutex);
if (!IsGpuThread()) {
WaitForIdle();
}
m_graphics_done = true;
void GuestGpu::SuspendPoint() {
EXIT_IF(IsGpuThread() || CommandScheduler::InDeferredOperation());
// Do not hold a queue lock while waiting: asynchronous work may be needed to
// finish the preceding graphics frame. The first point returns immediately.
m_suspend_point_ready->acquire();
Submission submission;
submission.type = SubmissionType::SuspendPoint;
Enqueue(std::move(submission));
m_done_num++;
}
@@ -229,6 +203,9 @@ void CommandProcessor::Reset() {
m_context_state_pushed = false;
m_index_type_and_size = 0;
m_index_buffer_size = 0;
m_index_base_addr = 0;
m_num_instances = 1;
m_predicate_skip = false;
m_user_data_marker = HW::UserSgprType::Unknown;
m_draw_indirect_args_base_addr = 0;
m_dispatch_indirect_args_base_addr = 0;
@@ -268,10 +245,6 @@ void CommandProcessor::BufferFlush() {
GetScheduler().Flush();
}
void CommandProcessor::BufferFlushAndWait() {
GetScheduler().FlushAndWait();
}
void CommandProcessor::BufferWait() {
BufferInit();
GetScheduler().Finish();
@@ -459,6 +432,7 @@ void GuestGpu::Enqueue(Submission submission) {
}
void GuestGpu::WaitForIdle() {
EXIT_IF(IsGpuThread() || CommandScheduler::InDeferredOperation());
Common::LockGuard lock(m_queue_mutex);
while (m_processing || !m_commands.empty() || m_submission_count != 0) {
m_idle.Wait(&m_queue_mutex);
@@ -564,11 +538,7 @@ void GuestGpu::ThreadRun(void* data) {
bool GuestGpu::Process(Submission& submission) {
const bool first_slice = !submission.started;
auto& cp = GetProcessor(submission.queue_id);
if (first_slice && submission.reset_processor) {
cp.Reset();
}
auto& cp = GetProcessor(submission.queue_id);
if (first_slice) {
submission.started = true;
@@ -645,6 +615,13 @@ bool GuestGpu::Process(Submission& submission) {
m_renderer.RunGarbageCollector();
cp.PrepareCpuFlip(submission.flip_request_id);
break;
case SubmissionType::SuspendPoint:
cp.EmitGlobalBarrier();
m_renderer.GetCommandScheduler().DeferPriorityOperation(
[ready = m_suspend_point_ready] { ready->release(); });
cp.BufferFlush();
cp.Reset();
break;
}
return complete;
@@ -849,9 +826,7 @@ void CommandProcessor::SetPredication(uint32_t condition, uint32_t op, uint32_t
}
} break;
case 0x03:
if (wait_op != 0) {
BufferFlushAndWait();
}
// The wait selector applies only to Z-pass query readiness.
EXIT_NOT_IMPLEMENTED(address == nullptr);
value = *reinterpret_cast<const volatile uint64_t*>(address);
break;
@@ -1034,9 +1009,6 @@ void CommandProcessor::DispatchDirect(uint32_t thread_group_x, uint32_t thread_g
m_sh_ctx.SetCsWaveSize(Pm4::ComputeWaveSize(mode));
uint32_t frame_num = 0;
// uint32_t local_x = 1;
// uint32_t local_y = 1;
// uint32_t local_z = 1;
{
frame_num = m_renderer.GetGpu().GetFrameNum();
@@ -1058,35 +1030,9 @@ void CommandProcessor::DispatchDirect(uint32_t thread_group_x, uint32_t thread_g
}
}
const auto& cs = m_sh_ctx.GetCs().cs_regs;
// local_x = std::max(cs.num_thread_x, 1u);
// local_y = std::max(cs.num_thread_y, 1u);
// local_z = std::max(cs.num_thread_z, 1u);
m_renderer.GetRenderExecutor().DispatchDirect(m_submit_id, CurrentBuffer(), thread_group_x,
thread_group_y, thread_group_z, mode);
}
/*constexpr uint32_t DispatchInitiatorUseThreadDimensions = 1u << 5u;
auto group_count = [](uint32_t threads, uint32_t group_size) {
return (threads == 0
? 0u
: (threads + std::max(group_size, 1u) - 1u) / std::max(group_size, 1u));
};
auto groups_x = thread_group_x;
auto groups_y = thread_group_y;
auto groups_z = thread_group_z;
if ((mode & DispatchInitiatorUseThreadDimensions) != 0) {
groups_x = group_count(thread_group_x, local_x);
groups_y = group_count(thread_group_y, local_y);
groups_z = group_count(thread_group_z, local_z);
}
const uint64_t invocations =
static_cast<uint64_t>(groups_x) * groups_y * groups_z * local_x * local_y * local_z;
if (invocations != 0) {
BufferFlushAndWait();
}*/
}
void CommandProcessor::DispatchIndirect(uint64_t args_addr, uint32_t mode) {
src/graphics/guest_gpu/graphicsRun.h
+13 -10
View File
@@ -13,6 +13,7 @@
#include <deque>
#include <memory>
#include <mutex>
#include <semaphore>
#include <span>
#include <thread>
@@ -32,11 +33,14 @@ public:
void SendCommandSync(Common::UniqueFunction<void>&& command);
// Submitted command memory is borrowed and must remain valid until GPU execution completes.
void Submit(std::span<const uint32_t> draw_commands,
std::span<const uint32_t> constant_commands);
void SubmitCompute(uint32_t queue, std::span<const uint32_t> commands);
void SubmitFlipPreparation(uint64_t request_id);
void Done();
void Submit(std::span<const uint32_t> draw_commands,
std::span<const uint32_t> constant_commands);
void SubmitCompute(uint32_t queue, std::span<const uint32_t> commands);
void SubmitFlipPreparation(uint64_t request_id);
// Insert an ordered graphics drain; only a previous suspend point can block the caller.
void SuspendPoint();
// Wait for guest command processing, including all compute queues (not native GPU completion).
void WaitForIdle();
[[nodiscard]] int GetFrameNum() const;
[[nodiscard]] static bool IsGpuThread() noexcept;
@@ -48,7 +52,7 @@ private:
static constexpr uint32_t ComputeQueueBase = 0x20;
static constexpr uint32_t QueueCount = 1 + ComputeQueueCount;
enum class SubmissionType { Graphics, Compute, FlipPreparation };
enum class SubmissionType { Graphics, Compute, FlipPreparation, SuspendPoint };
struct Submission {
SubmissionType type = SubmissionType::Graphics;
@@ -57,7 +61,6 @@ private:
std::span<const uint32_t> constant_commands;
Pm4Execution command_execution;
Pm4Execution constant_execution;
bool reset_processor = false;
bool started = false;
bool command_complete = false;
bool constant_complete = false;
@@ -66,14 +69,12 @@ private:
};
void Enqueue(Submission submission);
void WaitForIdle();
void ProcessCommands();
bool Process(Submission& submission);
static void ThreadRun(void* data);
CommandProcessor& GetProcessor(uint32_t queue_id);
RenderContext& m_renderer;
Common::Mutex m_submission_mutex;
Common::Mutex m_queue_mutex;
std::mutex m_shutdown_mutex;
Common::CondVar m_work_available;
@@ -84,10 +85,12 @@ private:
uint32_t m_next_queue = 0;
uint32_t m_submission_count = 0;
bool m_processing = false;
bool m_graphics_done = true;
bool m_accepting = true;
bool m_stopping = false;
bool m_shutdown_complete = false;
// Completion callbacks can outlive GuestGpu during renderer shutdown.
std::shared_ptr<std::binary_semaphore> m_suspend_point_ready =
std::make_shared<std::binary_semaphore>(1);
std::unique_ptr<CommandProcessor> m_gfx_cp;
std::array<std::unique_ptr<CommandProcessor>, ComputeQueueCount> m_compute_cp;
src/libs/agc.cpp
+1 -1
View File
@@ -1552,7 +1552,7 @@ int KYTY_SYSV_ABI AgcSuspendPoint() {
PRINT_NAME();
EXIT_IF(g_renderer == nullptr);
g_renderer->GetGpu().Done();
g_renderer->GetGpu().SuspendPoint();
return OK;
}