dolphin/Source/Core/VideoCommon/Fifo.cpp

Ignoring revisions in .git-blame-ignore-revs. Click here to bypass and see the normal blame view.

604 lines
19 KiB
C++
Raw Normal View History

// Copyright 2008 Dolphin Emulator Project
// SPDX-License-Identifier: GPL-2.0-or-later
#include "VideoCommon/Fifo.h"
#include <atomic>
#include <cstring>
#include "Common/Assert.h"
#include "Common/BlockingLoop.h"
#include "Common/ChunkFile.h"
#include "Common/Event.h"
#include "Common/FPURoundMode.h"
#include "Common/MemoryUtil.h"
#include "Common/MsgHandler.h"
#include "Core/Config/MainSettings.h"
2014-09-09 06:24:49 +02:00
#include "Core/ConfigManager.h"
#include "Core/CoreTiming.h"
#include "Core/HW/GPFifo.h"
#include "Core/HW/Memmap.h"
#include "Core/Host.h"
#include "Core/System.h"
#include "VideoCommon/AsyncRequests.h"
#include "VideoCommon/CPMemory.h"
#include "VideoCommon/CommandProcessor.h"
2014-07-08 16:49:33 +02:00
#include "VideoCommon/DataReader.h"
#include "VideoCommon/FramebufferManager.h"
#include "VideoCommon/OpcodeDecoding.h"
#include "VideoCommon/VertexLoaderManager.h"
#include "VideoCommon/VertexManagerBase.h"
#include "VideoCommon/VideoBackendBase.h"
2016-01-12 22:44:58 +01:00
namespace Fifo
{
static constexpr int GPU_TIME_SLOT_SIZE = 1000;
FifoManager::FifoManager(Core::System& system) : m_system{system}
{
}
FifoManager::~FifoManager() = default;
void FifoManager::RefreshConfig()
{
m_config_sync_gpu = Config::Get(Config::MAIN_SYNC_GPU);
m_config_sync_gpu_max_distance = Config::Get(Config::MAIN_SYNC_GPU_MAX_DISTANCE);
m_config_sync_gpu_min_distance = Config::Get(Config::MAIN_SYNC_GPU_MIN_DISTANCE);
m_config_sync_gpu_overclock = Config::Get(Config::MAIN_SYNC_GPU_OVERCLOCK);
}
void FifoManager::DoState(PointerWrap& p)
{
p.DoArray(m_video_buffer, FIFO_SIZE);
u8* write_ptr = m_video_buffer_write_ptr;
p.DoPointer(write_ptr, m_video_buffer);
m_video_buffer_write_ptr = write_ptr;
p.DoPointer(m_video_buffer_read_ptr, m_video_buffer);
if (p.IsReadMode() && m_use_deterministic_gpu_thread)
Add the 'desynced GPU thread' mode. It's a relatively big commit (less big with -w), but it's hard to test any of this separately... The basic problem is that in netplay or movies, the state of the CPU must be deterministic, including when the game receives notification that the GPU has processed FIFO data. Dual core mode notifies the game whenever the GPU thread actually gets around to doing the work, so it isn't deterministic. Single core mode is because it notifies the game 'instantly' (after processing the data synchronously), but it's too slow for many systems and games. My old dc-netplay branch worked as follows: everything worked as normal except the state of the CP registers was a lie, and the CPU thread only delivered results when idle detection triggered (waiting for the GPU if they weren't ready at that point). Usually, a game is idle iff all the work for the frame has been done, except for a small amount of work depending on the GPU result, so neither the CPU or the GPU waiting on the other affected performance much. However, it's possible that the game could be waiting for some earlier interrupt, and any of several games which, for whatever reason, never went into a detectable idle (even when I tried to improve the detection) would never receive results at all. (The current method should have better compatibility, but it also has slightly higher overhead and breaks some other things, so I want to reimplement this, hopefully with less impact on the code, in the future.) With this commit, the basic idea is that the CPU thread acts as if the work has been done instantly, like single core mode, but actually hands it off asynchronously to the GPU thread (after backing up some data that the game might change in memory before it's actually done). Since the work isn't done, any feedback from the GPU to the CPU, such as real XFB/EFB copies (virtual are OK), EFB pokes, performance queries, etc. is broken; but most games work with these options disabled, and there is no need to try to detect what the CPU thread is doing. Technically: when the flag g_use_deterministic_gpu_thread (currently stuck on) is on, the CPU thread calls RunGpu like in single core mode. This function synchronously copies the data from the FIFO to the internal video buffer and updates the CP registers, interrupts, etc. However, instead of the regular ReadDataFromFifo followed by running the opcode decoder, it runs ReadDataFromFifoOnCPU -> OpcodeDecoder_Preprocess, which relatively quickly scans through the FIFO data, detects SetFinish calls etc., which are immediately fired, and saves certain associated data from memory (e.g. display lists) in AuxBuffers (a parallel stream to the main FIFO, which is a bit slow at the moment), before handing the data off to the GPU thread to actually render. That makes up the bulk of this commit. In various circumstances, including the aforementioned EFB pokes and performance queries as well as swap requests (i.e. the end of a frame - we don't want the CPU potentially pumping out frames too quickly and the GPU falling behind*), SyncGPU is called to wait for actual completion. The overhead mainly comes from OpcodeDecoder_Preprocess (which is, again, synchronous), as well as the actual copying. Currently, display lists and such are escrowed from main memory even though they usually won't change over the course of a frame, and textures are not even though they might, resulting in a small chance of graphical glitches. When the texture locking (i.e. fault on write) code lands, I can make this all correct and maybe a little faster. * This suggests an alternate determinism method of just delaying results until a short time before the end of each frame. For all I know this might mostly work - I haven't tried it - but if any significant work hinges on the competion of render to texture etc., the frame will be missed.
2014-08-28 04:56:19 +02:00
{
// We're good and paused, right?
m_video_buffer_seen_ptr = m_video_buffer_pp_read_ptr = m_video_buffer_read_ptr;
Add the 'desynced GPU thread' mode. It's a relatively big commit (less big with -w), but it's hard to test any of this separately... The basic problem is that in netplay or movies, the state of the CPU must be deterministic, including when the game receives notification that the GPU has processed FIFO data. Dual core mode notifies the game whenever the GPU thread actually gets around to doing the work, so it isn't deterministic. Single core mode is because it notifies the game 'instantly' (after processing the data synchronously), but it's too slow for many systems and games. My old dc-netplay branch worked as follows: everything worked as normal except the state of the CP registers was a lie, and the CPU thread only delivered results when idle detection triggered (waiting for the GPU if they weren't ready at that point). Usually, a game is idle iff all the work for the frame has been done, except for a small amount of work depending on the GPU result, so neither the CPU or the GPU waiting on the other affected performance much. However, it's possible that the game could be waiting for some earlier interrupt, and any of several games which, for whatever reason, never went into a detectable idle (even when I tried to improve the detection) would never receive results at all. (The current method should have better compatibility, but it also has slightly higher overhead and breaks some other things, so I want to reimplement this, hopefully with less impact on the code, in the future.) With this commit, the basic idea is that the CPU thread acts as if the work has been done instantly, like single core mode, but actually hands it off asynchronously to the GPU thread (after backing up some data that the game might change in memory before it's actually done). Since the work isn't done, any feedback from the GPU to the CPU, such as real XFB/EFB copies (virtual are OK), EFB pokes, performance queries, etc. is broken; but most games work with these options disabled, and there is no need to try to detect what the CPU thread is doing. Technically: when the flag g_use_deterministic_gpu_thread (currently stuck on) is on, the CPU thread calls RunGpu like in single core mode. This function synchronously copies the data from the FIFO to the internal video buffer and updates the CP registers, interrupts, etc. However, instead of the regular ReadDataFromFifo followed by running the opcode decoder, it runs ReadDataFromFifoOnCPU -> OpcodeDecoder_Preprocess, which relatively quickly scans through the FIFO data, detects SetFinish calls etc., which are immediately fired, and saves certain associated data from memory (e.g. display lists) in AuxBuffers (a parallel stream to the main FIFO, which is a bit slow at the moment), before handing the data off to the GPU thread to actually render. That makes up the bulk of this commit. In various circumstances, including the aforementioned EFB pokes and performance queries as well as swap requests (i.e. the end of a frame - we don't want the CPU potentially pumping out frames too quickly and the GPU falling behind*), SyncGPU is called to wait for actual completion. The overhead mainly comes from OpcodeDecoder_Preprocess (which is, again, synchronous), as well as the actual copying. Currently, display lists and such are escrowed from main memory even though they usually won't change over the course of a frame, and textures are not even though they might, resulting in a small chance of graphical glitches. When the texture locking (i.e. fault on write) code lands, I can make this all correct and maybe a little faster. * This suggests an alternate determinism method of just delaying results until a short time before the end of each frame. For all I know this might mostly work - I haven't tried it - but if any significant work hinges on the competion of render to texture etc., the frame will be missed.
2014-08-28 04:56:19 +02:00
}
p.Do(m_sync_ticks);
p.Do(m_syncing_suspended);
}
void FifoManager::PauseAndLock(bool do_lock, bool unpause_on_unlock)
{
if (do_lock)
{
2016-08-19 04:35:58 +02:00
SyncGPU(SyncGPUReason::Other);
EmulatorState(false);
if (!m_system.IsDualCoreMode() || m_use_deterministic_gpu_thread)
return;
m_gpu_mainloop.WaitYield(std::chrono::milliseconds(100), Host_YieldToUI);
}
else
{
if (unpause_on_unlock)
EmulatorState(true);
}
}
void FifoManager::Init()
{
if (!m_config_callback_id)
m_config_callback_id = Config::AddConfigChangedCallback([this] { RefreshConfig(); });
RefreshConfig();
// Padded so that SIMD overreads in the vertex loader are safe
m_video_buffer = static_cast<u8*>(Common::AllocateMemoryPages(FIFO_SIZE + 4));
Add the 'desynced GPU thread' mode. It's a relatively big commit (less big with -w), but it's hard to test any of this separately... The basic problem is that in netplay or movies, the state of the CPU must be deterministic, including when the game receives notification that the GPU has processed FIFO data. Dual core mode notifies the game whenever the GPU thread actually gets around to doing the work, so it isn't deterministic. Single core mode is because it notifies the game 'instantly' (after processing the data synchronously), but it's too slow for many systems and games. My old dc-netplay branch worked as follows: everything worked as normal except the state of the CP registers was a lie, and the CPU thread only delivered results when idle detection triggered (waiting for the GPU if they weren't ready at that point). Usually, a game is idle iff all the work for the frame has been done, except for a small amount of work depending on the GPU result, so neither the CPU or the GPU waiting on the other affected performance much. However, it's possible that the game could be waiting for some earlier interrupt, and any of several games which, for whatever reason, never went into a detectable idle (even when I tried to improve the detection) would never receive results at all. (The current method should have better compatibility, but it also has slightly higher overhead and breaks some other things, so I want to reimplement this, hopefully with less impact on the code, in the future.) With this commit, the basic idea is that the CPU thread acts as if the work has been done instantly, like single core mode, but actually hands it off asynchronously to the GPU thread (after backing up some data that the game might change in memory before it's actually done). Since the work isn't done, any feedback from the GPU to the CPU, such as real XFB/EFB copies (virtual are OK), EFB pokes, performance queries, etc. is broken; but most games work with these options disabled, and there is no need to try to detect what the CPU thread is doing. Technically: when the flag g_use_deterministic_gpu_thread (currently stuck on) is on, the CPU thread calls RunGpu like in single core mode. This function synchronously copies the data from the FIFO to the internal video buffer and updates the CP registers, interrupts, etc. However, instead of the regular ReadDataFromFifo followed by running the opcode decoder, it runs ReadDataFromFifoOnCPU -> OpcodeDecoder_Preprocess, which relatively quickly scans through the FIFO data, detects SetFinish calls etc., which are immediately fired, and saves certain associated data from memory (e.g. display lists) in AuxBuffers (a parallel stream to the main FIFO, which is a bit slow at the moment), before handing the data off to the GPU thread to actually render. That makes up the bulk of this commit. In various circumstances, including the aforementioned EFB pokes and performance queries as well as swap requests (i.e. the end of a frame - we don't want the CPU potentially pumping out frames too quickly and the GPU falling behind*), SyncGPU is called to wait for actual completion. The overhead mainly comes from OpcodeDecoder_Preprocess (which is, again, synchronous), as well as the actual copying. Currently, display lists and such are escrowed from main memory even though they usually won't change over the course of a frame, and textures are not even though they might, resulting in a small chance of graphical glitches. When the texture locking (i.e. fault on write) code lands, I can make this all correct and maybe a little faster. * This suggests an alternate determinism method of just delaying results until a short time before the end of each frame. For all I know this might mostly work - I haven't tried it - but if any significant work hinges on the competion of render to texture etc., the frame will be missed.
2014-08-28 04:56:19 +02:00
ResetVideoBuffer();
if (m_system.IsDualCoreMode())
m_gpu_mainloop.Prepare();
m_sync_ticks.store(0);
}
void FifoManager::Shutdown()
{
if (m_gpu_mainloop.IsRunning())
PanicAlertFmt("FIFO shutting down while active");
Common::FreeMemoryPages(m_video_buffer, FIFO_SIZE + 4);
m_video_buffer = nullptr;
m_video_buffer_write_ptr = nullptr;
m_video_buffer_pp_read_ptr = nullptr;
m_video_buffer_read_ptr = nullptr;
m_video_buffer_seen_ptr = nullptr;
m_fifo_aux_write_ptr = nullptr;
m_fifo_aux_read_ptr = nullptr;
if (m_config_callback_id)
{
Config::RemoveConfigChangedCallback(*m_config_callback_id);
m_config_callback_id = std::nullopt;
}
}
// May be executed from any thread, even the graphics thread.
// Created to allow for self shutdown.
void FifoManager::ExitGpuLoop()
{
auto& command_processor = m_system.GetCommandProcessor();
auto& fifo = command_processor.GetFifo();
// This should break the wait loop in CPU thread
fifo.bFF_GPReadEnable.store(0, std::memory_order_relaxed);
FlushGpu();
// Terminate GPU thread loop
m_emu_running_state.Set();
m_gpu_mainloop.Stop(Common::BlockingLoop::StopMode::NonBlock);
}
void FifoManager::EmulatorState(bool running)
{
m_emu_running_state.Set(running);
if (running)
m_gpu_mainloop.Wakeup();
else
m_gpu_mainloop.AllowSleep();
}
void FifoManager::SyncGPU(SyncGPUReason reason, bool may_move_read_ptr)
Add the 'desynced GPU thread' mode. It's a relatively big commit (less big with -w), but it's hard to test any of this separately... The basic problem is that in netplay or movies, the state of the CPU must be deterministic, including when the game receives notification that the GPU has processed FIFO data. Dual core mode notifies the game whenever the GPU thread actually gets around to doing the work, so it isn't deterministic. Single core mode is because it notifies the game 'instantly' (after processing the data synchronously), but it's too slow for many systems and games. My old dc-netplay branch worked as follows: everything worked as normal except the state of the CP registers was a lie, and the CPU thread only delivered results when idle detection triggered (waiting for the GPU if they weren't ready at that point). Usually, a game is idle iff all the work for the frame has been done, except for a small amount of work depending on the GPU result, so neither the CPU or the GPU waiting on the other affected performance much. However, it's possible that the game could be waiting for some earlier interrupt, and any of several games which, for whatever reason, never went into a detectable idle (even when I tried to improve the detection) would never receive results at all. (The current method should have better compatibility, but it also has slightly higher overhead and breaks some other things, so I want to reimplement this, hopefully with less impact on the code, in the future.) With this commit, the basic idea is that the CPU thread acts as if the work has been done instantly, like single core mode, but actually hands it off asynchronously to the GPU thread (after backing up some data that the game might change in memory before it's actually done). Since the work isn't done, any feedback from the GPU to the CPU, such as real XFB/EFB copies (virtual are OK), EFB pokes, performance queries, etc. is broken; but most games work with these options disabled, and there is no need to try to detect what the CPU thread is doing. Technically: when the flag g_use_deterministic_gpu_thread (currently stuck on) is on, the CPU thread calls RunGpu like in single core mode. This function synchronously copies the data from the FIFO to the internal video buffer and updates the CP registers, interrupts, etc. However, instead of the regular ReadDataFromFifo followed by running the opcode decoder, it runs ReadDataFromFifoOnCPU -> OpcodeDecoder_Preprocess, which relatively quickly scans through the FIFO data, detects SetFinish calls etc., which are immediately fired, and saves certain associated data from memory (e.g. display lists) in AuxBuffers (a parallel stream to the main FIFO, which is a bit slow at the moment), before handing the data off to the GPU thread to actually render. That makes up the bulk of this commit. In various circumstances, including the aforementioned EFB pokes and performance queries as well as swap requests (i.e. the end of a frame - we don't want the CPU potentially pumping out frames too quickly and the GPU falling behind*), SyncGPU is called to wait for actual completion. The overhead mainly comes from OpcodeDecoder_Preprocess (which is, again, synchronous), as well as the actual copying. Currently, display lists and such are escrowed from main memory even though they usually won't change over the course of a frame, and textures are not even though they might, resulting in a small chance of graphical glitches. When the texture locking (i.e. fault on write) code lands, I can make this all correct and maybe a little faster. * This suggests an alternate determinism method of just delaying results until a short time before the end of each frame. For all I know this might mostly work - I haven't tried it - but if any significant work hinges on the competion of render to texture etc., the frame will be missed.
2014-08-28 04:56:19 +02:00
{
if (m_use_deterministic_gpu_thread)
Add the 'desynced GPU thread' mode. It's a relatively big commit (less big with -w), but it's hard to test any of this separately... The basic problem is that in netplay or movies, the state of the CPU must be deterministic, including when the game receives notification that the GPU has processed FIFO data. Dual core mode notifies the game whenever the GPU thread actually gets around to doing the work, so it isn't deterministic. Single core mode is because it notifies the game 'instantly' (after processing the data synchronously), but it's too slow for many systems and games. My old dc-netplay branch worked as follows: everything worked as normal except the state of the CP registers was a lie, and the CPU thread only delivered results when idle detection triggered (waiting for the GPU if they weren't ready at that point). Usually, a game is idle iff all the work for the frame has been done, except for a small amount of work depending on the GPU result, so neither the CPU or the GPU waiting on the other affected performance much. However, it's possible that the game could be waiting for some earlier interrupt, and any of several games which, for whatever reason, never went into a detectable idle (even when I tried to improve the detection) would never receive results at all. (The current method should have better compatibility, but it also has slightly higher overhead and breaks some other things, so I want to reimplement this, hopefully with less impact on the code, in the future.) With this commit, the basic idea is that the CPU thread acts as if the work has been done instantly, like single core mode, but actually hands it off asynchronously to the GPU thread (after backing up some data that the game might change in memory before it's actually done). Since the work isn't done, any feedback from the GPU to the CPU, such as real XFB/EFB copies (virtual are OK), EFB pokes, performance queries, etc. is broken; but most games work with these options disabled, and there is no need to try to detect what the CPU thread is doing. Technically: when the flag g_use_deterministic_gpu_thread (currently stuck on) is on, the CPU thread calls RunGpu like in single core mode. This function synchronously copies the data from the FIFO to the internal video buffer and updates the CP registers, interrupts, etc. However, instead of the regular ReadDataFromFifo followed by running the opcode decoder, it runs ReadDataFromFifoOnCPU -> OpcodeDecoder_Preprocess, which relatively quickly scans through the FIFO data, detects SetFinish calls etc., which are immediately fired, and saves certain associated data from memory (e.g. display lists) in AuxBuffers (a parallel stream to the main FIFO, which is a bit slow at the moment), before handing the data off to the GPU thread to actually render. That makes up the bulk of this commit. In various circumstances, including the aforementioned EFB pokes and performance queries as well as swap requests (i.e. the end of a frame - we don't want the CPU potentially pumping out frames too quickly and the GPU falling behind*), SyncGPU is called to wait for actual completion. The overhead mainly comes from OpcodeDecoder_Preprocess (which is, again, synchronous), as well as the actual copying. Currently, display lists and such are escrowed from main memory even though they usually won't change over the course of a frame, and textures are not even though they might, resulting in a small chance of graphical glitches. When the texture locking (i.e. fault on write) code lands, I can make this all correct and maybe a little faster. * This suggests an alternate determinism method of just delaying results until a short time before the end of each frame. For all I know this might mostly work - I haven't tried it - but if any significant work hinges on the competion of render to texture etc., the frame will be missed.
2014-08-28 04:56:19 +02:00
{
m_gpu_mainloop.Wait();
if (!m_gpu_mainloop.IsRunning())
Add the 'desynced GPU thread' mode. It's a relatively big commit (less big with -w), but it's hard to test any of this separately... The basic problem is that in netplay or movies, the state of the CPU must be deterministic, including when the game receives notification that the GPU has processed FIFO data. Dual core mode notifies the game whenever the GPU thread actually gets around to doing the work, so it isn't deterministic. Single core mode is because it notifies the game 'instantly' (after processing the data synchronously), but it's too slow for many systems and games. My old dc-netplay branch worked as follows: everything worked as normal except the state of the CP registers was a lie, and the CPU thread only delivered results when idle detection triggered (waiting for the GPU if they weren't ready at that point). Usually, a game is idle iff all the work for the frame has been done, except for a small amount of work depending on the GPU result, so neither the CPU or the GPU waiting on the other affected performance much. However, it's possible that the game could be waiting for some earlier interrupt, and any of several games which, for whatever reason, never went into a detectable idle (even when I tried to improve the detection) would never receive results at all. (The current method should have better compatibility, but it also has slightly higher overhead and breaks some other things, so I want to reimplement this, hopefully with less impact on the code, in the future.) With this commit, the basic idea is that the CPU thread acts as if the work has been done instantly, like single core mode, but actually hands it off asynchronously to the GPU thread (after backing up some data that the game might change in memory before it's actually done). Since the work isn't done, any feedback from the GPU to the CPU, such as real XFB/EFB copies (virtual are OK), EFB pokes, performance queries, etc. is broken; but most games work with these options disabled, and there is no need to try to detect what the CPU thread is doing. Technically: when the flag g_use_deterministic_gpu_thread (currently stuck on) is on, the CPU thread calls RunGpu like in single core mode. This function synchronously copies the data from the FIFO to the internal video buffer and updates the CP registers, interrupts, etc. However, instead of the regular ReadDataFromFifo followed by running the opcode decoder, it runs ReadDataFromFifoOnCPU -> OpcodeDecoder_Preprocess, which relatively quickly scans through the FIFO data, detects SetFinish calls etc., which are immediately fired, and saves certain associated data from memory (e.g. display lists) in AuxBuffers (a parallel stream to the main FIFO, which is a bit slow at the moment), before handing the data off to the GPU thread to actually render. That makes up the bulk of this commit. In various circumstances, including the aforementioned EFB pokes and performance queries as well as swap requests (i.e. the end of a frame - we don't want the CPU potentially pumping out frames too quickly and the GPU falling behind*), SyncGPU is called to wait for actual completion. The overhead mainly comes from OpcodeDecoder_Preprocess (which is, again, synchronous), as well as the actual copying. Currently, display lists and such are escrowed from main memory even though they usually won't change over the course of a frame, and textures are not even though they might, resulting in a small chance of graphical glitches. When the texture locking (i.e. fault on write) code lands, I can make this all correct and maybe a little faster. * This suggests an alternate determinism method of just delaying results until a short time before the end of each frame. For all I know this might mostly work - I haven't tried it - but if any significant work hinges on the competion of render to texture etc., the frame will be missed.
2014-08-28 04:56:19 +02:00
return;
Add the 'desynced GPU thread' mode. It's a relatively big commit (less big with -w), but it's hard to test any of this separately... The basic problem is that in netplay or movies, the state of the CPU must be deterministic, including when the game receives notification that the GPU has processed FIFO data. Dual core mode notifies the game whenever the GPU thread actually gets around to doing the work, so it isn't deterministic. Single core mode is because it notifies the game 'instantly' (after processing the data synchronously), but it's too slow for many systems and games. My old dc-netplay branch worked as follows: everything worked as normal except the state of the CP registers was a lie, and the CPU thread only delivered results when idle detection triggered (waiting for the GPU if they weren't ready at that point). Usually, a game is idle iff all the work for the frame has been done, except for a small amount of work depending on the GPU result, so neither the CPU or the GPU waiting on the other affected performance much. However, it's possible that the game could be waiting for some earlier interrupt, and any of several games which, for whatever reason, never went into a detectable idle (even when I tried to improve the detection) would never receive results at all. (The current method should have better compatibility, but it also has slightly higher overhead and breaks some other things, so I want to reimplement this, hopefully with less impact on the code, in the future.) With this commit, the basic idea is that the CPU thread acts as if the work has been done instantly, like single core mode, but actually hands it off asynchronously to the GPU thread (after backing up some data that the game might change in memory before it's actually done). Since the work isn't done, any feedback from the GPU to the CPU, such as real XFB/EFB copies (virtual are OK), EFB pokes, performance queries, etc. is broken; but most games work with these options disabled, and there is no need to try to detect what the CPU thread is doing. Technically: when the flag g_use_deterministic_gpu_thread (currently stuck on) is on, the CPU thread calls RunGpu like in single core mode. This function synchronously copies the data from the FIFO to the internal video buffer and updates the CP registers, interrupts, etc. However, instead of the regular ReadDataFromFifo followed by running the opcode decoder, it runs ReadDataFromFifoOnCPU -> OpcodeDecoder_Preprocess, which relatively quickly scans through the FIFO data, detects SetFinish calls etc., which are immediately fired, and saves certain associated data from memory (e.g. display lists) in AuxBuffers (a parallel stream to the main FIFO, which is a bit slow at the moment), before handing the data off to the GPU thread to actually render. That makes up the bulk of this commit. In various circumstances, including the aforementioned EFB pokes and performance queries as well as swap requests (i.e. the end of a frame - we don't want the CPU potentially pumping out frames too quickly and the GPU falling behind*), SyncGPU is called to wait for actual completion. The overhead mainly comes from OpcodeDecoder_Preprocess (which is, again, synchronous), as well as the actual copying. Currently, display lists and such are escrowed from main memory even though they usually won't change over the course of a frame, and textures are not even though they might, resulting in a small chance of graphical glitches. When the texture locking (i.e. fault on write) code lands, I can make this all correct and maybe a little faster. * This suggests an alternate determinism method of just delaying results until a short time before the end of each frame. For all I know this might mostly work - I haven't tried it - but if any significant work hinges on the competion of render to texture etc., the frame will be missed.
2014-08-28 04:56:19 +02:00
// Opportunistically reset FIFOs so we don't wrap around.
if (may_move_read_ptr && m_fifo_aux_write_ptr != m_fifo_aux_read_ptr)
{
PanicAlertFmt("Aux FIFO not synced ({}, {})", fmt::ptr(m_fifo_aux_write_ptr),
fmt::ptr(m_fifo_aux_read_ptr));
}
memmove(m_fifo_aux_data, m_fifo_aux_read_ptr, m_fifo_aux_write_ptr - m_fifo_aux_read_ptr);
m_fifo_aux_write_ptr -= (m_fifo_aux_read_ptr - m_fifo_aux_data);
m_fifo_aux_read_ptr = m_fifo_aux_data;
Add the 'desynced GPU thread' mode. It's a relatively big commit (less big with -w), but it's hard to test any of this separately... The basic problem is that in netplay or movies, the state of the CPU must be deterministic, including when the game receives notification that the GPU has processed FIFO data. Dual core mode notifies the game whenever the GPU thread actually gets around to doing the work, so it isn't deterministic. Single core mode is because it notifies the game 'instantly' (after processing the data synchronously), but it's too slow for many systems and games. My old dc-netplay branch worked as follows: everything worked as normal except the state of the CP registers was a lie, and the CPU thread only delivered results when idle detection triggered (waiting for the GPU if they weren't ready at that point). Usually, a game is idle iff all the work for the frame has been done, except for a small amount of work depending on the GPU result, so neither the CPU or the GPU waiting on the other affected performance much. However, it's possible that the game could be waiting for some earlier interrupt, and any of several games which, for whatever reason, never went into a detectable idle (even when I tried to improve the detection) would never receive results at all. (The current method should have better compatibility, but it also has slightly higher overhead and breaks some other things, so I want to reimplement this, hopefully with less impact on the code, in the future.) With this commit, the basic idea is that the CPU thread acts as if the work has been done instantly, like single core mode, but actually hands it off asynchronously to the GPU thread (after backing up some data that the game might change in memory before it's actually done). Since the work isn't done, any feedback from the GPU to the CPU, such as real XFB/EFB copies (virtual are OK), EFB pokes, performance queries, etc. is broken; but most games work with these options disabled, and there is no need to try to detect what the CPU thread is doing. Technically: when the flag g_use_deterministic_gpu_thread (currently stuck on) is on, the CPU thread calls RunGpu like in single core mode. This function synchronously copies the data from the FIFO to the internal video buffer and updates the CP registers, interrupts, etc. However, instead of the regular ReadDataFromFifo followed by running the opcode decoder, it runs ReadDataFromFifoOnCPU -> OpcodeDecoder_Preprocess, which relatively quickly scans through the FIFO data, detects SetFinish calls etc., which are immediately fired, and saves certain associated data from memory (e.g. display lists) in AuxBuffers (a parallel stream to the main FIFO, which is a bit slow at the moment), before handing the data off to the GPU thread to actually render. That makes up the bulk of this commit. In various circumstances, including the aforementioned EFB pokes and performance queries as well as swap requests (i.e. the end of a frame - we don't want the CPU potentially pumping out frames too quickly and the GPU falling behind*), SyncGPU is called to wait for actual completion. The overhead mainly comes from OpcodeDecoder_Preprocess (which is, again, synchronous), as well as the actual copying. Currently, display lists and such are escrowed from main memory even though they usually won't change over the course of a frame, and textures are not even though they might, resulting in a small chance of graphical glitches. When the texture locking (i.e. fault on write) code lands, I can make this all correct and maybe a little faster. * This suggests an alternate determinism method of just delaying results until a short time before the end of each frame. For all I know this might mostly work - I haven't tried it - but if any significant work hinges on the competion of render to texture etc., the frame will be missed.
2014-08-28 04:56:19 +02:00
if (may_move_read_ptr)
{
u8* write_ptr = m_video_buffer_write_ptr;
Add the 'desynced GPU thread' mode. It's a relatively big commit (less big with -w), but it's hard to test any of this separately... The basic problem is that in netplay or movies, the state of the CPU must be deterministic, including when the game receives notification that the GPU has processed FIFO data. Dual core mode notifies the game whenever the GPU thread actually gets around to doing the work, so it isn't deterministic. Single core mode is because it notifies the game 'instantly' (after processing the data synchronously), but it's too slow for many systems and games. My old dc-netplay branch worked as follows: everything worked as normal except the state of the CP registers was a lie, and the CPU thread only delivered results when idle detection triggered (waiting for the GPU if they weren't ready at that point). Usually, a game is idle iff all the work for the frame has been done, except for a small amount of work depending on the GPU result, so neither the CPU or the GPU waiting on the other affected performance much. However, it's possible that the game could be waiting for some earlier interrupt, and any of several games which, for whatever reason, never went into a detectable idle (even when I tried to improve the detection) would never receive results at all. (The current method should have better compatibility, but it also has slightly higher overhead and breaks some other things, so I want to reimplement this, hopefully with less impact on the code, in the future.) With this commit, the basic idea is that the CPU thread acts as if the work has been done instantly, like single core mode, but actually hands it off asynchronously to the GPU thread (after backing up some data that the game might change in memory before it's actually done). Since the work isn't done, any feedback from the GPU to the CPU, such as real XFB/EFB copies (virtual are OK), EFB pokes, performance queries, etc. is broken; but most games work with these options disabled, and there is no need to try to detect what the CPU thread is doing. Technically: when the flag g_use_deterministic_gpu_thread (currently stuck on) is on, the CPU thread calls RunGpu like in single core mode. This function synchronously copies the data from the FIFO to the internal video buffer and updates the CP registers, interrupts, etc. However, instead of the regular ReadDataFromFifo followed by running the opcode decoder, it runs ReadDataFromFifoOnCPU -> OpcodeDecoder_Preprocess, which relatively quickly scans through the FIFO data, detects SetFinish calls etc., which are immediately fired, and saves certain associated data from memory (e.g. display lists) in AuxBuffers (a parallel stream to the main FIFO, which is a bit slow at the moment), before handing the data off to the GPU thread to actually render. That makes up the bulk of this commit. In various circumstances, including the aforementioned EFB pokes and performance queries as well as swap requests (i.e. the end of a frame - we don't want the CPU potentially pumping out frames too quickly and the GPU falling behind*), SyncGPU is called to wait for actual completion. The overhead mainly comes from OpcodeDecoder_Preprocess (which is, again, synchronous), as well as the actual copying. Currently, display lists and such are escrowed from main memory even though they usually won't change over the course of a frame, and textures are not even though they might, resulting in a small chance of graphical glitches. When the texture locking (i.e. fault on write) code lands, I can make this all correct and maybe a little faster. * This suggests an alternate determinism method of just delaying results until a short time before the end of each frame. For all I know this might mostly work - I haven't tried it - but if any significant work hinges on the competion of render to texture etc., the frame will be missed.
2014-08-28 04:56:19 +02:00
// what's left over in the buffer
size_t size = write_ptr - m_video_buffer_pp_read_ptr;
memmove(m_video_buffer, m_video_buffer_pp_read_ptr, size);
Add the 'desynced GPU thread' mode. It's a relatively big commit (less big with -w), but it's hard to test any of this separately... The basic problem is that in netplay or movies, the state of the CPU must be deterministic, including when the game receives notification that the GPU has processed FIFO data. Dual core mode notifies the game whenever the GPU thread actually gets around to doing the work, so it isn't deterministic. Single core mode is because it notifies the game 'instantly' (after processing the data synchronously), but it's too slow for many systems and games. My old dc-netplay branch worked as follows: everything worked as normal except the state of the CP registers was a lie, and the CPU thread only delivered results when idle detection triggered (waiting for the GPU if they weren't ready at that point). Usually, a game is idle iff all the work for the frame has been done, except for a small amount of work depending on the GPU result, so neither the CPU or the GPU waiting on the other affected performance much. However, it's possible that the game could be waiting for some earlier interrupt, and any of several games which, for whatever reason, never went into a detectable idle (even when I tried to improve the detection) would never receive results at all. (The current method should have better compatibility, but it also has slightly higher overhead and breaks some other things, so I want to reimplement this, hopefully with less impact on the code, in the future.) With this commit, the basic idea is that the CPU thread acts as if the work has been done instantly, like single core mode, but actually hands it off asynchronously to the GPU thread (after backing up some data that the game might change in memory before it's actually done). Since the work isn't done, any feedback from the GPU to the CPU, such as real XFB/EFB copies (virtual are OK), EFB pokes, performance queries, etc. is broken; but most games work with these options disabled, and there is no need to try to detect what the CPU thread is doing. Technically: when the flag g_use_deterministic_gpu_thread (currently stuck on) is on, the CPU thread calls RunGpu like in single core mode. This function synchronously copies the data from the FIFO to the internal video buffer and updates the CP registers, interrupts, etc. However, instead of the regular ReadDataFromFifo followed by running the opcode decoder, it runs ReadDataFromFifoOnCPU -> OpcodeDecoder_Preprocess, which relatively quickly scans through the FIFO data, detects SetFinish calls etc., which are immediately fired, and saves certain associated data from memory (e.g. display lists) in AuxBuffers (a parallel stream to the main FIFO, which is a bit slow at the moment), before handing the data off to the GPU thread to actually render. That makes up the bulk of this commit. In various circumstances, including the aforementioned EFB pokes and performance queries as well as swap requests (i.e. the end of a frame - we don't want the CPU potentially pumping out frames too quickly and the GPU falling behind*), SyncGPU is called to wait for actual completion. The overhead mainly comes from OpcodeDecoder_Preprocess (which is, again, synchronous), as well as the actual copying. Currently, display lists and such are escrowed from main memory even though they usually won't change over the course of a frame, and textures are not even though they might, resulting in a small chance of graphical glitches. When the texture locking (i.e. fault on write) code lands, I can make this all correct and maybe a little faster. * This suggests an alternate determinism method of just delaying results until a short time before the end of each frame. For all I know this might mostly work - I haven't tried it - but if any significant work hinges on the competion of render to texture etc., the frame will be missed.
2014-08-28 04:56:19 +02:00
// This change always decreases the pointers. We write seen_ptr
// after write_ptr here, and read it before in RunGpuLoop, so
// 'write_ptr > seen_ptr' there cannot become spuriously true.
m_video_buffer_write_ptr = write_ptr = m_video_buffer + size;
m_video_buffer_pp_read_ptr = m_video_buffer;
m_video_buffer_read_ptr = m_video_buffer;
m_video_buffer_seen_ptr = write_ptr;
Add the 'desynced GPU thread' mode. It's a relatively big commit (less big with -w), but it's hard to test any of this separately... The basic problem is that in netplay or movies, the state of the CPU must be deterministic, including when the game receives notification that the GPU has processed FIFO data. Dual core mode notifies the game whenever the GPU thread actually gets around to doing the work, so it isn't deterministic. Single core mode is because it notifies the game 'instantly' (after processing the data synchronously), but it's too slow for many systems and games. My old dc-netplay branch worked as follows: everything worked as normal except the state of the CP registers was a lie, and the CPU thread only delivered results when idle detection triggered (waiting for the GPU if they weren't ready at that point). Usually, a game is idle iff all the work for the frame has been done, except for a small amount of work depending on the GPU result, so neither the CPU or the GPU waiting on the other affected performance much. However, it's possible that the game could be waiting for some earlier interrupt, and any of several games which, for whatever reason, never went into a detectable idle (even when I tried to improve the detection) would never receive results at all. (The current method should have better compatibility, but it also has slightly higher overhead and breaks some other things, so I want to reimplement this, hopefully with less impact on the code, in the future.) With this commit, the basic idea is that the CPU thread acts as if the work has been done instantly, like single core mode, but actually hands it off asynchronously to the GPU thread (after backing up some data that the game might change in memory before it's actually done). Since the work isn't done, any feedback from the GPU to the CPU, such as real XFB/EFB copies (virtual are OK), EFB pokes, performance queries, etc. is broken; but most games work with these options disabled, and there is no need to try to detect what the CPU thread is doing. Technically: when the flag g_use_deterministic_gpu_thread (currently stuck on) is on, the CPU thread calls RunGpu like in single core mode. This function synchronously copies the data from the FIFO to the internal video buffer and updates the CP registers, interrupts, etc. However, instead of the regular ReadDataFromFifo followed by running the opcode decoder, it runs ReadDataFromFifoOnCPU -> OpcodeDecoder_Preprocess, which relatively quickly scans through the FIFO data, detects SetFinish calls etc., which are immediately fired, and saves certain associated data from memory (e.g. display lists) in AuxBuffers (a parallel stream to the main FIFO, which is a bit slow at the moment), before handing the data off to the GPU thread to actually render. That makes up the bulk of this commit. In various circumstances, including the aforementioned EFB pokes and performance queries as well as swap requests (i.e. the end of a frame - we don't want the CPU potentially pumping out frames too quickly and the GPU falling behind*), SyncGPU is called to wait for actual completion. The overhead mainly comes from OpcodeDecoder_Preprocess (which is, again, synchronous), as well as the actual copying. Currently, display lists and such are escrowed from main memory even though they usually won't change over the course of a frame, and textures are not even though they might, resulting in a small chance of graphical glitches. When the texture locking (i.e. fault on write) code lands, I can make this all correct and maybe a little faster. * This suggests an alternate determinism method of just delaying results until a short time before the end of each frame. For all I know this might mostly work - I haven't tried it - but if any significant work hinges on the competion of render to texture etc., the frame will be missed.
2014-08-28 04:56:19 +02:00
}
}
}
void FifoManager::PushFifoAuxBuffer(const void* ptr, size_t size)
Add the 'desynced GPU thread' mode. It's a relatively big commit (less big with -w), but it's hard to test any of this separately... The basic problem is that in netplay or movies, the state of the CPU must be deterministic, including when the game receives notification that the GPU has processed FIFO data. Dual core mode notifies the game whenever the GPU thread actually gets around to doing the work, so it isn't deterministic. Single core mode is because it notifies the game 'instantly' (after processing the data synchronously), but it's too slow for many systems and games. My old dc-netplay branch worked as follows: everything worked as normal except the state of the CP registers was a lie, and the CPU thread only delivered results when idle detection triggered (waiting for the GPU if they weren't ready at that point). Usually, a game is idle iff all the work for the frame has been done, except for a small amount of work depending on the GPU result, so neither the CPU or the GPU waiting on the other affected performance much. However, it's possible that the game could be waiting for some earlier interrupt, and any of several games which, for whatever reason, never went into a detectable idle (even when I tried to improve the detection) would never receive results at all. (The current method should have better compatibility, but it also has slightly higher overhead and breaks some other things, so I want to reimplement this, hopefully with less impact on the code, in the future.) With this commit, the basic idea is that the CPU thread acts as if the work has been done instantly, like single core mode, but actually hands it off asynchronously to the GPU thread (after backing up some data that the game might change in memory before it's actually done). Since the work isn't done, any feedback from the GPU to the CPU, such as real XFB/EFB copies (virtual are OK), EFB pokes, performance queries, etc. is broken; but most games work with these options disabled, and there is no need to try to detect what the CPU thread is doing. Technically: when the flag g_use_deterministic_gpu_thread (currently stuck on) is on, the CPU thread calls RunGpu like in single core mode. This function synchronously copies the data from the FIFO to the internal video buffer and updates the CP registers, interrupts, etc. However, instead of the regular ReadDataFromFifo followed by running the opcode decoder, it runs ReadDataFromFifoOnCPU -> OpcodeDecoder_Preprocess, which relatively quickly scans through the FIFO data, detects SetFinish calls etc., which are immediately fired, and saves certain associated data from memory (e.g. display lists) in AuxBuffers (a parallel stream to the main FIFO, which is a bit slow at the moment), before handing the data off to the GPU thread to actually render. That makes up the bulk of this commit. In various circumstances, including the aforementioned EFB pokes and performance queries as well as swap requests (i.e. the end of a frame - we don't want the CPU potentially pumping out frames too quickly and the GPU falling behind*), SyncGPU is called to wait for actual completion. The overhead mainly comes from OpcodeDecoder_Preprocess (which is, again, synchronous), as well as the actual copying. Currently, display lists and such are escrowed from main memory even though they usually won't change over the course of a frame, and textures are not even though they might, resulting in a small chance of graphical glitches. When the texture locking (i.e. fault on write) code lands, I can make this all correct and maybe a little faster. * This suggests an alternate determinism method of just delaying results until a short time before the end of each frame. For all I know this might mostly work - I haven't tried it - but if any significant work hinges on the competion of render to texture etc., the frame will be missed.
2014-08-28 04:56:19 +02:00
{
if (size > (size_t)(m_fifo_aux_data + FIFO_SIZE - m_fifo_aux_write_ptr))
Add the 'desynced GPU thread' mode. It's a relatively big commit (less big with -w), but it's hard to test any of this separately... The basic problem is that in netplay or movies, the state of the CPU must be deterministic, including when the game receives notification that the GPU has processed FIFO data. Dual core mode notifies the game whenever the GPU thread actually gets around to doing the work, so it isn't deterministic. Single core mode is because it notifies the game 'instantly' (after processing the data synchronously), but it's too slow for many systems and games. My old dc-netplay branch worked as follows: everything worked as normal except the state of the CP registers was a lie, and the CPU thread only delivered results when idle detection triggered (waiting for the GPU if they weren't ready at that point). Usually, a game is idle iff all the work for the frame has been done, except for a small amount of work depending on the GPU result, so neither the CPU or the GPU waiting on the other affected performance much. However, it's possible that the game could be waiting for some earlier interrupt, and any of several games which, for whatever reason, never went into a detectable idle (even when I tried to improve the detection) would never receive results at all. (The current method should have better compatibility, but it also has slightly higher overhead and breaks some other things, so I want to reimplement this, hopefully with less impact on the code, in the future.) With this commit, the basic idea is that the CPU thread acts as if the work has been done instantly, like single core mode, but actually hands it off asynchronously to the GPU thread (after backing up some data that the game might change in memory before it's actually done). Since the work isn't done, any feedback from the GPU to the CPU, such as real XFB/EFB copies (virtual are OK), EFB pokes, performance queries, etc. is broken; but most games work with these options disabled, and there is no need to try to detect what the CPU thread is doing. Technically: when the flag g_use_deterministic_gpu_thread (currently stuck on) is on, the CPU thread calls RunGpu like in single core mode. This function synchronously copies the data from the FIFO to the internal video buffer and updates the CP registers, interrupts, etc. However, instead of the regular ReadDataFromFifo followed by running the opcode decoder, it runs ReadDataFromFifoOnCPU -> OpcodeDecoder_Preprocess, which relatively quickly scans through the FIFO data, detects SetFinish calls etc., which are immediately fired, and saves certain associated data from memory (e.g. display lists) in AuxBuffers (a parallel stream to the main FIFO, which is a bit slow at the moment), before handing the data off to the GPU thread to actually render. That makes up the bulk of this commit. In various circumstances, including the aforementioned EFB pokes and performance queries as well as swap requests (i.e. the end of a frame - we don't want the CPU potentially pumping out frames too quickly and the GPU falling behind*), SyncGPU is called to wait for actual completion. The overhead mainly comes from OpcodeDecoder_Preprocess (which is, again, synchronous), as well as the actual copying. Currently, display lists and such are escrowed from main memory even though they usually won't change over the course of a frame, and textures are not even though they might, resulting in a small chance of graphical glitches. When the texture locking (i.e. fault on write) code lands, I can make this all correct and maybe a little faster. * This suggests an alternate determinism method of just delaying results until a short time before the end of each frame. For all I know this might mostly work - I haven't tried it - but if any significant work hinges on the competion of render to texture etc., the frame will be missed.
2014-08-28 04:56:19 +02:00
{
2016-08-19 04:35:58 +02:00
SyncGPU(SyncGPUReason::AuxSpace, /* may_move_read_ptr */ false);
if (!m_gpu_mainloop.IsRunning())
{
// GPU is shutting down
return;
}
if (size > (size_t)(m_fifo_aux_data + FIFO_SIZE - m_fifo_aux_write_ptr))
Add the 'desynced GPU thread' mode. It's a relatively big commit (less big with -w), but it's hard to test any of this separately... The basic problem is that in netplay or movies, the state of the CPU must be deterministic, including when the game receives notification that the GPU has processed FIFO data. Dual core mode notifies the game whenever the GPU thread actually gets around to doing the work, so it isn't deterministic. Single core mode is because it notifies the game 'instantly' (after processing the data synchronously), but it's too slow for many systems and games. My old dc-netplay branch worked as follows: everything worked as normal except the state of the CP registers was a lie, and the CPU thread only delivered results when idle detection triggered (waiting for the GPU if they weren't ready at that point). Usually, a game is idle iff all the work for the frame has been done, except for a small amount of work depending on the GPU result, so neither the CPU or the GPU waiting on the other affected performance much. However, it's possible that the game could be waiting for some earlier interrupt, and any of several games which, for whatever reason, never went into a detectable idle (even when I tried to improve the detection) would never receive results at all. (The current method should have better compatibility, but it also has slightly higher overhead and breaks some other things, so I want to reimplement this, hopefully with less impact on the code, in the future.) With this commit, the basic idea is that the CPU thread acts as if the work has been done instantly, like single core mode, but actually hands it off asynchronously to the GPU thread (after backing up some data that the game might change in memory before it's actually done). Since the work isn't done, any feedback from the GPU to the CPU, such as real XFB/EFB copies (virtual are OK), EFB pokes, performance queries, etc. is broken; but most games work with these options disabled, and there is no need to try to detect what the CPU thread is doing. Technically: when the flag g_use_deterministic_gpu_thread (currently stuck on) is on, the CPU thread calls RunGpu like in single core mode. This function synchronously copies the data from the FIFO to the internal video buffer and updates the CP registers, interrupts, etc. However, instead of the regular ReadDataFromFifo followed by running the opcode decoder, it runs ReadDataFromFifoOnCPU -> OpcodeDecoder_Preprocess, which relatively quickly scans through the FIFO data, detects SetFinish calls etc., which are immediately fired, and saves certain associated data from memory (e.g. display lists) in AuxBuffers (a parallel stream to the main FIFO, which is a bit slow at the moment), before handing the data off to the GPU thread to actually render. That makes up the bulk of this commit. In various circumstances, including the aforementioned EFB pokes and performance queries as well as swap requests (i.e. the end of a frame - we don't want the CPU potentially pumping out frames too quickly and the GPU falling behind*), SyncGPU is called to wait for actual completion. The overhead mainly comes from OpcodeDecoder_Preprocess (which is, again, synchronous), as well as the actual copying. Currently, display lists and such are escrowed from main memory even though they usually won't change over the course of a frame, and textures are not even though they might, resulting in a small chance of graphical glitches. When the texture locking (i.e. fault on write) code lands, I can make this all correct and maybe a little faster. * This suggests an alternate determinism method of just delaying results until a short time before the end of each frame. For all I know this might mostly work - I haven't tried it - but if any significant work hinges on the competion of render to texture etc., the frame will be missed.
2014-08-28 04:56:19 +02:00
{
// That will sync us up to the last 32 bytes, so this short region
// of FIFO would have to point to a 2MB display list or something.
PanicAlertFmt("Absurdly large aux buffer");
Add the 'desynced GPU thread' mode. It's a relatively big commit (less big with -w), but it's hard to test any of this separately... The basic problem is that in netplay or movies, the state of the CPU must be deterministic, including when the game receives notification that the GPU has processed FIFO data. Dual core mode notifies the game whenever the GPU thread actually gets around to doing the work, so it isn't deterministic. Single core mode is because it notifies the game 'instantly' (after processing the data synchronously), but it's too slow for many systems and games. My old dc-netplay branch worked as follows: everything worked as normal except the state of the CP registers was a lie, and the CPU thread only delivered results when idle detection triggered (waiting for the GPU if they weren't ready at that point). Usually, a game is idle iff all the work for the frame has been done, except for a small amount of work depending on the GPU result, so neither the CPU or the GPU waiting on the other affected performance much. However, it's possible that the game could be waiting for some earlier interrupt, and any of several games which, for whatever reason, never went into a detectable idle (even when I tried to improve the detection) would never receive results at all. (The current method should have better compatibility, but it also has slightly higher overhead and breaks some other things, so I want to reimplement this, hopefully with less impact on the code, in the future.) With this commit, the basic idea is that the CPU thread acts as if the work has been done instantly, like single core mode, but actually hands it off asynchronously to the GPU thread (after backing up some data that the game might change in memory before it's actually done). Since the work isn't done, any feedback from the GPU to the CPU, such as real XFB/EFB copies (virtual are OK), EFB pokes, performance queries, etc. is broken; but most games work with these options disabled, and there is no need to try to detect what the CPU thread is doing. Technically: when the flag g_use_deterministic_gpu_thread (currently stuck on) is on, the CPU thread calls RunGpu like in single core mode. This function synchronously copies the data from the FIFO to the internal video buffer and updates the CP registers, interrupts, etc. However, instead of the regular ReadDataFromFifo followed by running the opcode decoder, it runs ReadDataFromFifoOnCPU -> OpcodeDecoder_Preprocess, which relatively quickly scans through the FIFO data, detects SetFinish calls etc., which are immediately fired, and saves certain associated data from memory (e.g. display lists) in AuxBuffers (a parallel stream to the main FIFO, which is a bit slow at the moment), before handing the data off to the GPU thread to actually render. That makes up the bulk of this commit. In various circumstances, including the aforementioned EFB pokes and performance queries as well as swap requests (i.e. the end of a frame - we don't want the CPU potentially pumping out frames too quickly and the GPU falling behind*), SyncGPU is called to wait for actual completion. The overhead mainly comes from OpcodeDecoder_Preprocess (which is, again, synchronous), as well as the actual copying. Currently, display lists and such are escrowed from main memory even though they usually won't change over the course of a frame, and textures are not even though they might, resulting in a small chance of graphical glitches. When the texture locking (i.e. fault on write) code lands, I can make this all correct and maybe a little faster. * This suggests an alternate determinism method of just delaying results until a short time before the end of each frame. For all I know this might mostly work - I haven't tried it - but if any significant work hinges on the competion of render to texture etc., the frame will be missed.
2014-08-28 04:56:19 +02:00
return;
}
}
memcpy(m_fifo_aux_write_ptr, ptr, size);
m_fifo_aux_write_ptr += size;
Add the 'desynced GPU thread' mode. It's a relatively big commit (less big with -w), but it's hard to test any of this separately... The basic problem is that in netplay or movies, the state of the CPU must be deterministic, including when the game receives notification that the GPU has processed FIFO data. Dual core mode notifies the game whenever the GPU thread actually gets around to doing the work, so it isn't deterministic. Single core mode is because it notifies the game 'instantly' (after processing the data synchronously), but it's too slow for many systems and games. My old dc-netplay branch worked as follows: everything worked as normal except the state of the CP registers was a lie, and the CPU thread only delivered results when idle detection triggered (waiting for the GPU if they weren't ready at that point). Usually, a game is idle iff all the work for the frame has been done, except for a small amount of work depending on the GPU result, so neither the CPU or the GPU waiting on the other affected performance much. However, it's possible that the game could be waiting for some earlier interrupt, and any of several games which, for whatever reason, never went into a detectable idle (even when I tried to improve the detection) would never receive results at all. (The current method should have better compatibility, but it also has slightly higher overhead and breaks some other things, so I want to reimplement this, hopefully with less impact on the code, in the future.) With this commit, the basic idea is that the CPU thread acts as if the work has been done instantly, like single core mode, but actually hands it off asynchronously to the GPU thread (after backing up some data that the game might change in memory before it's actually done). Since the work isn't done, any feedback from the GPU to the CPU, such as real XFB/EFB copies (virtual are OK), EFB pokes, performance queries, etc. is broken; but most games work with these options disabled, and there is no need to try to detect what the CPU thread is doing. Technically: when the flag g_use_deterministic_gpu_thread (currently stuck on) is on, the CPU thread calls RunGpu like in single core mode. This function synchronously copies the data from the FIFO to the internal video buffer and updates the CP registers, interrupts, etc. However, instead of the regular ReadDataFromFifo followed by running the opcode decoder, it runs ReadDataFromFifoOnCPU -> OpcodeDecoder_Preprocess, which relatively quickly scans through the FIFO data, detects SetFinish calls etc., which are immediately fired, and saves certain associated data from memory (e.g. display lists) in AuxBuffers (a parallel stream to the main FIFO, which is a bit slow at the moment), before handing the data off to the GPU thread to actually render. That makes up the bulk of this commit. In various circumstances, including the aforementioned EFB pokes and performance queries as well as swap requests (i.e. the end of a frame - we don't want the CPU potentially pumping out frames too quickly and the GPU falling behind*), SyncGPU is called to wait for actual completion. The overhead mainly comes from OpcodeDecoder_Preprocess (which is, again, synchronous), as well as the actual copying. Currently, display lists and such are escrowed from main memory even though they usually won't change over the course of a frame, and textures are not even though they might, resulting in a small chance of graphical glitches. When the texture locking (i.e. fault on write) code lands, I can make this all correct and maybe a little faster. * This suggests an alternate determinism method of just delaying results until a short time before the end of each frame. For all I know this might mostly work - I haven't tried it - but if any significant work hinges on the competion of render to texture etc., the frame will be missed.
2014-08-28 04:56:19 +02:00
}
void* FifoManager::PopFifoAuxBuffer(size_t size)
Add the 'desynced GPU thread' mode. It's a relatively big commit (less big with -w), but it's hard to test any of this separately... The basic problem is that in netplay or movies, the state of the CPU must be deterministic, including when the game receives notification that the GPU has processed FIFO data. Dual core mode notifies the game whenever the GPU thread actually gets around to doing the work, so it isn't deterministic. Single core mode is because it notifies the game 'instantly' (after processing the data synchronously), but it's too slow for many systems and games. My old dc-netplay branch worked as follows: everything worked as normal except the state of the CP registers was a lie, and the CPU thread only delivered results when idle detection triggered (waiting for the GPU if they weren't ready at that point). Usually, a game is idle iff all the work for the frame has been done, except for a small amount of work depending on the GPU result, so neither the CPU or the GPU waiting on the other affected performance much. However, it's possible that the game could be waiting for some earlier interrupt, and any of several games which, for whatever reason, never went into a detectable idle (even when I tried to improve the detection) would never receive results at all. (The current method should have better compatibility, but it also has slightly higher overhead and breaks some other things, so I want to reimplement this, hopefully with less impact on the code, in the future.) With this commit, the basic idea is that the CPU thread acts as if the work has been done instantly, like single core mode, but actually hands it off asynchronously to the GPU thread (after backing up some data that the game might change in memory before it's actually done). Since the work isn't done, any feedback from the GPU to the CPU, such as real XFB/EFB copies (virtual are OK), EFB pokes, performance queries, etc. is broken; but most games work with these options disabled, and there is no need to try to detect what the CPU thread is doing. Technically: when the flag g_use_deterministic_gpu_thread (currently stuck on) is on, the CPU thread calls RunGpu like in single core mode. This function synchronously copies the data from the FIFO to the internal video buffer and updates the CP registers, interrupts, etc. However, instead of the regular ReadDataFromFifo followed by running the opcode decoder, it runs ReadDataFromFifoOnCPU -> OpcodeDecoder_Preprocess, which relatively quickly scans through the FIFO data, detects SetFinish calls etc., which are immediately fired, and saves certain associated data from memory (e.g. display lists) in AuxBuffers (a parallel stream to the main FIFO, which is a bit slow at the moment), before handing the data off to the GPU thread to actually render. That makes up the bulk of this commit. In various circumstances, including the aforementioned EFB pokes and performance queries as well as swap requests (i.e. the end of a frame - we don't want the CPU potentially pumping out frames too quickly and the GPU falling behind*), SyncGPU is called to wait for actual completion. The overhead mainly comes from OpcodeDecoder_Preprocess (which is, again, synchronous), as well as the actual copying. Currently, display lists and such are escrowed from main memory even though they usually won't change over the course of a frame, and textures are not even though they might, resulting in a small chance of graphical glitches. When the texture locking (i.e. fault on write) code lands, I can make this all correct and maybe a little faster. * This suggests an alternate determinism method of just delaying results until a short time before the end of each frame. For all I know this might mostly work - I haven't tried it - but if any significant work hinges on the competion of render to texture etc., the frame will be missed.
2014-08-28 04:56:19 +02:00
{
void* ret = m_fifo_aux_read_ptr;
m_fifo_aux_read_ptr += size;
Add the 'desynced GPU thread' mode. It's a relatively big commit (less big with -w), but it's hard to test any of this separately... The basic problem is that in netplay or movies, the state of the CPU must be deterministic, including when the game receives notification that the GPU has processed FIFO data. Dual core mode notifies the game whenever the GPU thread actually gets around to doing the work, so it isn't deterministic. Single core mode is because it notifies the game 'instantly' (after processing the data synchronously), but it's too slow for many systems and games. My old dc-netplay branch worked as follows: everything worked as normal except the state of the CP registers was a lie, and the CPU thread only delivered results when idle detection triggered (waiting for the GPU if they weren't ready at that point). Usually, a game is idle iff all the work for the frame has been done, except for a small amount of work depending on the GPU result, so neither the CPU or the GPU waiting on the other affected performance much. However, it's possible that the game could be waiting for some earlier interrupt, and any of several games which, for whatever reason, never went into a detectable idle (even when I tried to improve the detection) would never receive results at all. (The current method should have better compatibility, but it also has slightly higher overhead and breaks some other things, so I want to reimplement this, hopefully with less impact on the code, in the future.) With this commit, the basic idea is that the CPU thread acts as if the work has been done instantly, like single core mode, but actually hands it off asynchronously to the GPU thread (after backing up some data that the game might change in memory before it's actually done). Since the work isn't done, any feedback from the GPU to the CPU, such as real XFB/EFB copies (virtual are OK), EFB pokes, performance queries, etc. is broken; but most games work with these options disabled, and there is no need to try to detect what the CPU thread is doing. Technically: when the flag g_use_deterministic_gpu_thread (currently stuck on) is on, the CPU thread calls RunGpu like in single core mode. This function synchronously copies the data from the FIFO to the internal video buffer and updates the CP registers, interrupts, etc. However, instead of the regular ReadDataFromFifo followed by running the opcode decoder, it runs ReadDataFromFifoOnCPU -> OpcodeDecoder_Preprocess, which relatively quickly scans through the FIFO data, detects SetFinish calls etc., which are immediately fired, and saves certain associated data from memory (e.g. display lists) in AuxBuffers (a parallel stream to the main FIFO, which is a bit slow at the moment), before handing the data off to the GPU thread to actually render. That makes up the bulk of this commit. In various circumstances, including the aforementioned EFB pokes and performance queries as well as swap requests (i.e. the end of a frame - we don't want the CPU potentially pumping out frames too quickly and the GPU falling behind*), SyncGPU is called to wait for actual completion. The overhead mainly comes from OpcodeDecoder_Preprocess (which is, again, synchronous), as well as the actual copying. Currently, display lists and such are escrowed from main memory even though they usually won't change over the course of a frame, and textures are not even though they might, resulting in a small chance of graphical glitches. When the texture locking (i.e. fault on write) code lands, I can make this all correct and maybe a little faster. * This suggests an alternate determinism method of just delaying results until a short time before the end of each frame. For all I know this might mostly work - I haven't tried it - but if any significant work hinges on the competion of render to texture etc., the frame will be missed.
2014-08-28 04:56:19 +02:00
return ret;
}
// Description: RunGpuLoop() sends data through this function.
void FifoManager::ReadDataFromFifo(u32 read_ptr)
{
if (GPFifo::GATHER_PIPE_SIZE >
static_cast<size_t>(m_video_buffer + FIFO_SIZE - m_video_buffer_write_ptr))
{
const size_t existing_len = m_video_buffer_write_ptr - m_video_buffer_read_ptr;
if (GPFifo::GATHER_PIPE_SIZE > static_cast<size_t>(FIFO_SIZE - existing_len))
{
PanicAlertFmt("FIFO out of bounds (existing {} + new {} > {})", existing_len,
GPFifo::GATHER_PIPE_SIZE, FIFO_SIZE);
return;
}
memmove(m_video_buffer, m_video_buffer_read_ptr, existing_len);
m_video_buffer_write_ptr = m_video_buffer + existing_len;
m_video_buffer_read_ptr = m_video_buffer;
}
// Copy new video instructions to m_video_buffer for future use in rendering the new picture
auto& memory = m_system.GetMemory();
memory.CopyFromEmu(m_video_buffer_write_ptr, read_ptr, GPFifo::GATHER_PIPE_SIZE);
m_video_buffer_write_ptr += GPFifo::GATHER_PIPE_SIZE;
}
Add the 'desynced GPU thread' mode. It's a relatively big commit (less big with -w), but it's hard to test any of this separately... The basic problem is that in netplay or movies, the state of the CPU must be deterministic, including when the game receives notification that the GPU has processed FIFO data. Dual core mode notifies the game whenever the GPU thread actually gets around to doing the work, so it isn't deterministic. Single core mode is because it notifies the game 'instantly' (after processing the data synchronously), but it's too slow for many systems and games. My old dc-netplay branch worked as follows: everything worked as normal except the state of the CP registers was a lie, and the CPU thread only delivered results when idle detection triggered (waiting for the GPU if they weren't ready at that point). Usually, a game is idle iff all the work for the frame has been done, except for a small amount of work depending on the GPU result, so neither the CPU or the GPU waiting on the other affected performance much. However, it's possible that the game could be waiting for some earlier interrupt, and any of several games which, for whatever reason, never went into a detectable idle (even when I tried to improve the detection) would never receive results at all. (The current method should have better compatibility, but it also has slightly higher overhead and breaks some other things, so I want to reimplement this, hopefully with less impact on the code, in the future.) With this commit, the basic idea is that the CPU thread acts as if the work has been done instantly, like single core mode, but actually hands it off asynchronously to the GPU thread (after backing up some data that the game might change in memory before it's actually done). Since the work isn't done, any feedback from the GPU to the CPU, such as real XFB/EFB copies (virtual are OK), EFB pokes, performance queries, etc. is broken; but most games work with these options disabled, and there is no need to try to detect what the CPU thread is doing. Technically: when the flag g_use_deterministic_gpu_thread (currently stuck on) is on, the CPU thread calls RunGpu like in single core mode. This function synchronously copies the data from the FIFO to the internal video buffer and updates the CP registers, interrupts, etc. However, instead of the regular ReadDataFromFifo followed by running the opcode decoder, it runs ReadDataFromFifoOnCPU -> OpcodeDecoder_Preprocess, which relatively quickly scans through the FIFO data, detects SetFinish calls etc., which are immediately fired, and saves certain associated data from memory (e.g. display lists) in AuxBuffers (a parallel stream to the main FIFO, which is a bit slow at the moment), before handing the data off to the GPU thread to actually render. That makes up the bulk of this commit. In various circumstances, including the aforementioned EFB pokes and performance queries as well as swap requests (i.e. the end of a frame - we don't want the CPU potentially pumping out frames too quickly and the GPU falling behind*), SyncGPU is called to wait for actual completion. The overhead mainly comes from OpcodeDecoder_Preprocess (which is, again, synchronous), as well as the actual copying. Currently, display lists and such are escrowed from main memory even though they usually won't change over the course of a frame, and textures are not even though they might, resulting in a small chance of graphical glitches. When the texture locking (i.e. fault on write) code lands, I can make this all correct and maybe a little faster. * This suggests an alternate determinism method of just delaying results until a short time before the end of each frame. For all I know this might mostly work - I haven't tried it - but if any significant work hinges on the competion of render to texture etc., the frame will be missed.
2014-08-28 04:56:19 +02:00
// The deterministic_gpu_thread version.
void FifoManager::ReadDataFromFifoOnCPU(u32 read_ptr)
Add the 'desynced GPU thread' mode. It's a relatively big commit (less big with -w), but it's hard to test any of this separately... The basic problem is that in netplay or movies, the state of the CPU must be deterministic, including when the game receives notification that the GPU has processed FIFO data. Dual core mode notifies the game whenever the GPU thread actually gets around to doing the work, so it isn't deterministic. Single core mode is because it notifies the game 'instantly' (after processing the data synchronously), but it's too slow for many systems and games. My old dc-netplay branch worked as follows: everything worked as normal except the state of the CP registers was a lie, and the CPU thread only delivered results when idle detection triggered (waiting for the GPU if they weren't ready at that point). Usually, a game is idle iff all the work for the frame has been done, except for a small amount of work depending on the GPU result, so neither the CPU or the GPU waiting on the other affected performance much. However, it's possible that the game could be waiting for some earlier interrupt, and any of several games which, for whatever reason, never went into a detectable idle (even when I tried to improve the detection) would never receive results at all. (The current method should have better compatibility, but it also has slightly higher overhead and breaks some other things, so I want to reimplement this, hopefully with less impact on the code, in the future.) With this commit, the basic idea is that the CPU thread acts as if the work has been done instantly, like single core mode, but actually hands it off asynchronously to the GPU thread (after backing up some data that the game might change in memory before it's actually done). Since the work isn't done, any feedback from the GPU to the CPU, such as real XFB/EFB copies (virtual are OK), EFB pokes, performance queries, etc. is broken; but most games work with these options disabled, and there is no need to try to detect what the CPU thread is doing. Technically: when the flag g_use_deterministic_gpu_thread (currently stuck on) is on, the CPU thread calls RunGpu like in single core mode. This function synchronously copies the data from the FIFO to the internal video buffer and updates the CP registers, interrupts, etc. However, instead of the regular ReadDataFromFifo followed by running the opcode decoder, it runs ReadDataFromFifoOnCPU -> OpcodeDecoder_Preprocess, which relatively quickly scans through the FIFO data, detects SetFinish calls etc., which are immediately fired, and saves certain associated data from memory (e.g. display lists) in AuxBuffers (a parallel stream to the main FIFO, which is a bit slow at the moment), before handing the data off to the GPU thread to actually render. That makes up the bulk of this commit. In various circumstances, including the aforementioned EFB pokes and performance queries as well as swap requests (i.e. the end of a frame - we don't want the CPU potentially pumping out frames too quickly and the GPU falling behind*), SyncGPU is called to wait for actual completion. The overhead mainly comes from OpcodeDecoder_Preprocess (which is, again, synchronous), as well as the actual copying. Currently, display lists and such are escrowed from main memory even though they usually won't change over the course of a frame, and textures are not even though they might, resulting in a small chance of graphical glitches. When the texture locking (i.e. fault on write) code lands, I can make this all correct and maybe a little faster. * This suggests an alternate determinism method of just delaying results until a short time before the end of each frame. For all I know this might mostly work - I haven't tried it - but if any significant work hinges on the competion of render to texture etc., the frame will be missed.
2014-08-28 04:56:19 +02:00
{
u8* write_ptr = m_video_buffer_write_ptr;
if (GPFifo::GATHER_PIPE_SIZE > static_cast<size_t>(m_video_buffer + FIFO_SIZE - write_ptr))
Add the 'desynced GPU thread' mode. It's a relatively big commit (less big with -w), but it's hard to test any of this separately... The basic problem is that in netplay or movies, the state of the CPU must be deterministic, including when the game receives notification that the GPU has processed FIFO data. Dual core mode notifies the game whenever the GPU thread actually gets around to doing the work, so it isn't deterministic. Single core mode is because it notifies the game 'instantly' (after processing the data synchronously), but it's too slow for many systems and games. My old dc-netplay branch worked as follows: everything worked as normal except the state of the CP registers was a lie, and the CPU thread only delivered results when idle detection triggered (waiting for the GPU if they weren't ready at that point). Usually, a game is idle iff all the work for the frame has been done, except for a small amount of work depending on the GPU result, so neither the CPU or the GPU waiting on the other affected performance much. However, it's possible that the game could be waiting for some earlier interrupt, and any of several games which, for whatever reason, never went into a detectable idle (even when I tried to improve the detection) would never receive results at all. (The current method should have better compatibility, but it also has slightly higher overhead and breaks some other things, so I want to reimplement this, hopefully with less impact on the code, in the future.) With this commit, the basic idea is that the CPU thread acts as if the work has been done instantly, like single core mode, but actually hands it off asynchronously to the GPU thread (after backing up some data that the game might change in memory before it's actually done). Since the work isn't done, any feedback from the GPU to the CPU, such as real XFB/EFB copies (virtual are OK), EFB pokes, performance queries, etc. is broken; but most games work with these options disabled, and there is no need to try to detect what the CPU thread is doing. Technically: when the flag g_use_deterministic_gpu_thread (currently stuck on) is on, the CPU thread calls RunGpu like in single core mode. This function synchronously copies the data from the FIFO to the internal video buffer and updates the CP registers, interrupts, etc. However, instead of the regular ReadDataFromFifo followed by running the opcode decoder, it runs ReadDataFromFifoOnCPU -> OpcodeDecoder_Preprocess, which relatively quickly scans through the FIFO data, detects SetFinish calls etc., which are immediately fired, and saves certain associated data from memory (e.g. display lists) in AuxBuffers (a parallel stream to the main FIFO, which is a bit slow at the moment), before handing the data off to the GPU thread to actually render. That makes up the bulk of this commit. In various circumstances, including the aforementioned EFB pokes and performance queries as well as swap requests (i.e. the end of a frame - we don't want the CPU potentially pumping out frames too quickly and the GPU falling behind*), SyncGPU is called to wait for actual completion. The overhead mainly comes from OpcodeDecoder_Preprocess (which is, again, synchronous), as well as the actual copying. Currently, display lists and such are escrowed from main memory even though they usually won't change over the course of a frame, and textures are not even though they might, resulting in a small chance of graphical glitches. When the texture locking (i.e. fault on write) code lands, I can make this all correct and maybe a little faster. * This suggests an alternate determinism method of just delaying results until a short time before the end of each frame. For all I know this might mostly work - I haven't tried it - but if any significant work hinges on the competion of render to texture etc., the frame will be missed.
2014-08-28 04:56:19 +02:00
{
// We can't wrap around while the GPU is working on the data.
// This should be very rare due to the reset in SyncGPU.
2016-08-19 04:35:58 +02:00
SyncGPU(SyncGPUReason::Wraparound);
if (!m_gpu_mainloop.IsRunning())
{
// GPU is shutting down, so the next asserts may fail
return;
}
if (m_video_buffer_pp_read_ptr != m_video_buffer_read_ptr)
Add the 'desynced GPU thread' mode. It's a relatively big commit (less big with -w), but it's hard to test any of this separately... The basic problem is that in netplay or movies, the state of the CPU must be deterministic, including when the game receives notification that the GPU has processed FIFO data. Dual core mode notifies the game whenever the GPU thread actually gets around to doing the work, so it isn't deterministic. Single core mode is because it notifies the game 'instantly' (after processing the data synchronously), but it's too slow for many systems and games. My old dc-netplay branch worked as follows: everything worked as normal except the state of the CP registers was a lie, and the CPU thread only delivered results when idle detection triggered (waiting for the GPU if they weren't ready at that point). Usually, a game is idle iff all the work for the frame has been done, except for a small amount of work depending on the GPU result, so neither the CPU or the GPU waiting on the other affected performance much. However, it's possible that the game could be waiting for some earlier interrupt, and any of several games which, for whatever reason, never went into a detectable idle (even when I tried to improve the detection) would never receive results at all. (The current method should have better compatibility, but it also has slightly higher overhead and breaks some other things, so I want to reimplement this, hopefully with less impact on the code, in the future.) With this commit, the basic idea is that the CPU thread acts as if the work has been done instantly, like single core mode, but actually hands it off asynchronously to the GPU thread (after backing up some data that the game might change in memory before it's actually done). Since the work isn't done, any feedback from the GPU to the CPU, such as real XFB/EFB copies (virtual are OK), EFB pokes, performance queries, etc. is broken; but most games work with these options disabled, and there is no need to try to detect what the CPU thread is doing. Technically: when the flag g_use_deterministic_gpu_thread (currently stuck on) is on, the CPU thread calls RunGpu like in single core mode. This function synchronously copies the data from the FIFO to the internal video buffer and updates the CP registers, interrupts, etc. However, instead of the regular ReadDataFromFifo followed by running the opcode decoder, it runs ReadDataFromFifoOnCPU -> OpcodeDecoder_Preprocess, which relatively quickly scans through the FIFO data, detects SetFinish calls etc., which are immediately fired, and saves certain associated data from memory (e.g. display lists) in AuxBuffers (a parallel stream to the main FIFO, which is a bit slow at the moment), before handing the data off to the GPU thread to actually render. That makes up the bulk of this commit. In various circumstances, including the aforementioned EFB pokes and performance queries as well as swap requests (i.e. the end of a frame - we don't want the CPU potentially pumping out frames too quickly and the GPU falling behind*), SyncGPU is called to wait for actual completion. The overhead mainly comes from OpcodeDecoder_Preprocess (which is, again, synchronous), as well as the actual copying. Currently, display lists and such are escrowed from main memory even though they usually won't change over the course of a frame, and textures are not even though they might, resulting in a small chance of graphical glitches. When the texture locking (i.e. fault on write) code lands, I can make this all correct and maybe a little faster. * This suggests an alternate determinism method of just delaying results until a short time before the end of each frame. For all I know this might mostly work - I haven't tried it - but if any significant work hinges on the competion of render to texture etc., the frame will be missed.
2014-08-28 04:56:19 +02:00
{
PanicAlertFmt("Desynced read pointers");
Add the 'desynced GPU thread' mode. It's a relatively big commit (less big with -w), but it's hard to test any of this separately... The basic problem is that in netplay or movies, the state of the CPU must be deterministic, including when the game receives notification that the GPU has processed FIFO data. Dual core mode notifies the game whenever the GPU thread actually gets around to doing the work, so it isn't deterministic. Single core mode is because it notifies the game 'instantly' (after processing the data synchronously), but it's too slow for many systems and games. My old dc-netplay branch worked as follows: everything worked as normal except the state of the CP registers was a lie, and the CPU thread only delivered results when idle detection triggered (waiting for the GPU if they weren't ready at that point). Usually, a game is idle iff all the work for the frame has been done, except for a small amount of work depending on the GPU result, so neither the CPU or the GPU waiting on the other affected performance much. However, it's possible that the game could be waiting for some earlier interrupt, and any of several games which, for whatever reason, never went into a detectable idle (even when I tried to improve the detection) would never receive results at all. (The current method should have better compatibility, but it also has slightly higher overhead and breaks some other things, so I want to reimplement this, hopefully with less impact on the code, in the future.) With this commit, the basic idea is that the CPU thread acts as if the work has been done instantly, like single core mode, but actually hands it off asynchronously to the GPU thread (after backing up some data that the game might change in memory before it's actually done). Since the work isn't done, any feedback from the GPU to the CPU, such as real XFB/EFB copies (virtual are OK), EFB pokes, performance queries, etc. is broken; but most games work with these options disabled, and there is no need to try to detect what the CPU thread is doing. Technically: when the flag g_use_deterministic_gpu_thread (currently stuck on) is on, the CPU thread calls RunGpu like in single core mode. This function synchronously copies the data from the FIFO to the internal video buffer and updates the CP registers, interrupts, etc. However, instead of the regular ReadDataFromFifo followed by running the opcode decoder, it runs ReadDataFromFifoOnCPU -> OpcodeDecoder_Preprocess, which relatively quickly scans through the FIFO data, detects SetFinish calls etc., which are immediately fired, and saves certain associated data from memory (e.g. display lists) in AuxBuffers (a parallel stream to the main FIFO, which is a bit slow at the moment), before handing the data off to the GPU thread to actually render. That makes up the bulk of this commit. In various circumstances, including the aforementioned EFB pokes and performance queries as well as swap requests (i.e. the end of a frame - we don't want the CPU potentially pumping out frames too quickly and the GPU falling behind*), SyncGPU is called to wait for actual completion. The overhead mainly comes from OpcodeDecoder_Preprocess (which is, again, synchronous), as well as the actual copying. Currently, display lists and such are escrowed from main memory even though they usually won't change over the course of a frame, and textures are not even though they might, resulting in a small chance of graphical glitches. When the texture locking (i.e. fault on write) code lands, I can make this all correct and maybe a little faster. * This suggests an alternate determinism method of just delaying results until a short time before the end of each frame. For all I know this might mostly work - I haven't tried it - but if any significant work hinges on the competion of render to texture etc., the frame will be missed.
2014-08-28 04:56:19 +02:00
return;
}
write_ptr = m_video_buffer_write_ptr;
const size_t existing_len = write_ptr - m_video_buffer_pp_read_ptr;
if (GPFifo::GATHER_PIPE_SIZE > static_cast<size_t>(FIFO_SIZE - existing_len))
Add the 'desynced GPU thread' mode. It's a relatively big commit (less big with -w), but it's hard to test any of this separately... The basic problem is that in netplay or movies, the state of the CPU must be deterministic, including when the game receives notification that the GPU has processed FIFO data. Dual core mode notifies the game whenever the GPU thread actually gets around to doing the work, so it isn't deterministic. Single core mode is because it notifies the game 'instantly' (after processing the data synchronously), but it's too slow for many systems and games. My old dc-netplay branch worked as follows: everything worked as normal except the state of the CP registers was a lie, and the CPU thread only delivered results when idle detection triggered (waiting for the GPU if they weren't ready at that point). Usually, a game is idle iff all the work for the frame has been done, except for a small amount of work depending on the GPU result, so neither the CPU or the GPU waiting on the other affected performance much. However, it's possible that the game could be waiting for some earlier interrupt, and any of several games which, for whatever reason, never went into a detectable idle (even when I tried to improve the detection) would never receive results at all. (The current method should have better compatibility, but it also has slightly higher overhead and breaks some other things, so I want to reimplement this, hopefully with less impact on the code, in the future.) With this commit, the basic idea is that the CPU thread acts as if the work has been done instantly, like single core mode, but actually hands it off asynchronously to the GPU thread (after backing up some data that the game might change in memory before it's actually done). Since the work isn't done, any feedback from the GPU to the CPU, such as real XFB/EFB copies (virtual are OK), EFB pokes, performance queries, etc. is broken; but most games work with these options disabled, and there is no need to try to detect what the CPU thread is doing. Technically: when the flag g_use_deterministic_gpu_thread (currently stuck on) is on, the CPU thread calls RunGpu like in single core mode. This function synchronously copies the data from the FIFO to the internal video buffer and updates the CP registers, interrupts, etc. However, instead of the regular ReadDataFromFifo followed by running the opcode decoder, it runs ReadDataFromFifoOnCPU -> OpcodeDecoder_Preprocess, which relatively quickly scans through the FIFO data, detects SetFinish calls etc., which are immediately fired, and saves certain associated data from memory (e.g. display lists) in AuxBuffers (a parallel stream to the main FIFO, which is a bit slow at the moment), before handing the data off to the GPU thread to actually render. That makes up the bulk of this commit. In various circumstances, including the aforementioned EFB pokes and performance queries as well as swap requests (i.e. the end of a frame - we don't want the CPU potentially pumping out frames too quickly and the GPU falling behind*), SyncGPU is called to wait for actual completion. The overhead mainly comes from OpcodeDecoder_Preprocess (which is, again, synchronous), as well as the actual copying. Currently, display lists and such are escrowed from main memory even though they usually won't change over the course of a frame, and textures are not even though they might, resulting in a small chance of graphical glitches. When the texture locking (i.e. fault on write) code lands, I can make this all correct and maybe a little faster. * This suggests an alternate determinism method of just delaying results until a short time before the end of each frame. For all I know this might mostly work - I haven't tried it - but if any significant work hinges on the competion of render to texture etc., the frame will be missed.
2014-08-28 04:56:19 +02:00
{
PanicAlertFmt("FIFO out of bounds (existing {} + new {} > {})", existing_len,
GPFifo::GATHER_PIPE_SIZE, FIFO_SIZE);
Add the 'desynced GPU thread' mode. It's a relatively big commit (less big with -w), but it's hard to test any of this separately... The basic problem is that in netplay or movies, the state of the CPU must be deterministic, including when the game receives notification that the GPU has processed FIFO data. Dual core mode notifies the game whenever the GPU thread actually gets around to doing the work, so it isn't deterministic. Single core mode is because it notifies the game 'instantly' (after processing the data synchronously), but it's too slow for many systems and games. My old dc-netplay branch worked as follows: everything worked as normal except the state of the CP registers was a lie, and the CPU thread only delivered results when idle detection triggered (waiting for the GPU if they weren't ready at that point). Usually, a game is idle iff all the work for the frame has been done, except for a small amount of work depending on the GPU result, so neither the CPU or the GPU waiting on the other affected performance much. However, it's possible that the game could be waiting for some earlier interrupt, and any of several games which, for whatever reason, never went into a detectable idle (even when I tried to improve the detection) would never receive results at all. (The current method should have better compatibility, but it also has slightly higher overhead and breaks some other things, so I want to reimplement this, hopefully with less impact on the code, in the future.) With this commit, the basic idea is that the CPU thread acts as if the work has been done instantly, like single core mode, but actually hands it off asynchronously to the GPU thread (after backing up some data that the game might change in memory before it's actually done). Since the work isn't done, any feedback from the GPU to the CPU, such as real XFB/EFB copies (virtual are OK), EFB pokes, performance queries, etc. is broken; but most games work with these options disabled, and there is no need to try to detect what the CPU thread is doing. Technically: when the flag g_use_deterministic_gpu_thread (currently stuck on) is on, the CPU thread calls RunGpu like in single core mode. This function synchronously copies the data from the FIFO to the internal video buffer and updates the CP registers, interrupts, etc. However, instead of the regular ReadDataFromFifo followed by running the opcode decoder, it runs ReadDataFromFifoOnCPU -> OpcodeDecoder_Preprocess, which relatively quickly scans through the FIFO data, detects SetFinish calls etc., which are immediately fired, and saves certain associated data from memory (e.g. display lists) in AuxBuffers (a parallel stream to the main FIFO, which is a bit slow at the moment), before handing the data off to the GPU thread to actually render. That makes up the bulk of this commit. In various circumstances, including the aforementioned EFB pokes and performance queries as well as swap requests (i.e. the end of a frame - we don't want the CPU potentially pumping out frames too quickly and the GPU falling behind*), SyncGPU is called to wait for actual completion. The overhead mainly comes from OpcodeDecoder_Preprocess (which is, again, synchronous), as well as the actual copying. Currently, display lists and such are escrowed from main memory even though they usually won't change over the course of a frame, and textures are not even though they might, resulting in a small chance of graphical glitches. When the texture locking (i.e. fault on write) code lands, I can make this all correct and maybe a little faster. * This suggests an alternate determinism method of just delaying results until a short time before the end of each frame. For all I know this might mostly work - I haven't tried it - but if any significant work hinges on the competion of render to texture etc., the frame will be missed.
2014-08-28 04:56:19 +02:00
return;
}
Add the 'desynced GPU thread' mode. It's a relatively big commit (less big with -w), but it's hard to test any of this separately... The basic problem is that in netplay or movies, the state of the CPU must be deterministic, including when the game receives notification that the GPU has processed FIFO data. Dual core mode notifies the game whenever the GPU thread actually gets around to doing the work, so it isn't deterministic. Single core mode is because it notifies the game 'instantly' (after processing the data synchronously), but it's too slow for many systems and games. My old dc-netplay branch worked as follows: everything worked as normal except the state of the CP registers was a lie, and the CPU thread only delivered results when idle detection triggered (waiting for the GPU if they weren't ready at that point). Usually, a game is idle iff all the work for the frame has been done, except for a small amount of work depending on the GPU result, so neither the CPU or the GPU waiting on the other affected performance much. However, it's possible that the game could be waiting for some earlier interrupt, and any of several games which, for whatever reason, never went into a detectable idle (even when I tried to improve the detection) would never receive results at all. (The current method should have better compatibility, but it also has slightly higher overhead and breaks some other things, so I want to reimplement this, hopefully with less impact on the code, in the future.) With this commit, the basic idea is that the CPU thread acts as if the work has been done instantly, like single core mode, but actually hands it off asynchronously to the GPU thread (after backing up some data that the game might change in memory before it's actually done). Since the work isn't done, any feedback from the GPU to the CPU, such as real XFB/EFB copies (virtual are OK), EFB pokes, performance queries, etc. is broken; but most games work with these options disabled, and there is no need to try to detect what the CPU thread is doing. Technically: when the flag g_use_deterministic_gpu_thread (currently stuck on) is on, the CPU thread calls RunGpu like in single core mode. This function synchronously copies the data from the FIFO to the internal video buffer and updates the CP registers, interrupts, etc. However, instead of the regular ReadDataFromFifo followed by running the opcode decoder, it runs ReadDataFromFifoOnCPU -> OpcodeDecoder_Preprocess, which relatively quickly scans through the FIFO data, detects SetFinish calls etc., which are immediately fired, and saves certain associated data from memory (e.g. display lists) in AuxBuffers (a parallel stream to the main FIFO, which is a bit slow at the moment), before handing the data off to the GPU thread to actually render. That makes up the bulk of this commit. In various circumstances, including the aforementioned EFB pokes and performance queries as well as swap requests (i.e. the end of a frame - we don't want the CPU potentially pumping out frames too quickly and the GPU falling behind*), SyncGPU is called to wait for actual completion. The overhead mainly comes from OpcodeDecoder_Preprocess (which is, again, synchronous), as well as the actual copying. Currently, display lists and such are escrowed from main memory even though they usually won't change over the course of a frame, and textures are not even though they might, resulting in a small chance of graphical glitches. When the texture locking (i.e. fault on write) code lands, I can make this all correct and maybe a little faster. * This suggests an alternate determinism method of just delaying results until a short time before the end of each frame. For all I know this might mostly work - I haven't tried it - but if any significant work hinges on the competion of render to texture etc., the frame will be missed.
2014-08-28 04:56:19 +02:00
}
auto& memory = m_system.GetMemory();
memory.CopyFromEmu(m_video_buffer_write_ptr, read_ptr, GPFifo::GATHER_PIPE_SIZE);
m_video_buffer_pp_read_ptr = OpcodeDecoder::RunFifo<true>(
DataReader(m_video_buffer_pp_read_ptr, write_ptr + GPFifo::GATHER_PIPE_SIZE), nullptr);
Add the 'desynced GPU thread' mode. It's a relatively big commit (less big with -w), but it's hard to test any of this separately... The basic problem is that in netplay or movies, the state of the CPU must be deterministic, including when the game receives notification that the GPU has processed FIFO data. Dual core mode notifies the game whenever the GPU thread actually gets around to doing the work, so it isn't deterministic. Single core mode is because it notifies the game 'instantly' (after processing the data synchronously), but it's too slow for many systems and games. My old dc-netplay branch worked as follows: everything worked as normal except the state of the CP registers was a lie, and the CPU thread only delivered results when idle detection triggered (waiting for the GPU if they weren't ready at that point). Usually, a game is idle iff all the work for the frame has been done, except for a small amount of work depending on the GPU result, so neither the CPU or the GPU waiting on the other affected performance much. However, it's possible that the game could be waiting for some earlier interrupt, and any of several games which, for whatever reason, never went into a detectable idle (even when I tried to improve the detection) would never receive results at all. (The current method should have better compatibility, but it also has slightly higher overhead and breaks some other things, so I want to reimplement this, hopefully with less impact on the code, in the future.) With this commit, the basic idea is that the CPU thread acts as if the work has been done instantly, like single core mode, but actually hands it off asynchronously to the GPU thread (after backing up some data that the game might change in memory before it's actually done). Since the work isn't done, any feedback from the GPU to the CPU, such as real XFB/EFB copies (virtual are OK), EFB pokes, performance queries, etc. is broken; but most games work with these options disabled, and there is no need to try to detect what the CPU thread is doing. Technically: when the flag g_use_deterministic_gpu_thread (currently stuck on) is on, the CPU thread calls RunGpu like in single core mode. This function synchronously copies the data from the FIFO to the internal video buffer and updates the CP registers, interrupts, etc. However, instead of the regular ReadDataFromFifo followed by running the opcode decoder, it runs ReadDataFromFifoOnCPU -> OpcodeDecoder_Preprocess, which relatively quickly scans through the FIFO data, detects SetFinish calls etc., which are immediately fired, and saves certain associated data from memory (e.g. display lists) in AuxBuffers (a parallel stream to the main FIFO, which is a bit slow at the moment), before handing the data off to the GPU thread to actually render. That makes up the bulk of this commit. In various circumstances, including the aforementioned EFB pokes and performance queries as well as swap requests (i.e. the end of a frame - we don't want the CPU potentially pumping out frames too quickly and the GPU falling behind*), SyncGPU is called to wait for actual completion. The overhead mainly comes from OpcodeDecoder_Preprocess (which is, again, synchronous), as well as the actual copying. Currently, display lists and such are escrowed from main memory even though they usually won't change over the course of a frame, and textures are not even though they might, resulting in a small chance of graphical glitches. When the texture locking (i.e. fault on write) code lands, I can make this all correct and maybe a little faster. * This suggests an alternate determinism method of just delaying results until a short time before the end of each frame. For all I know this might mostly work - I haven't tried it - but if any significant work hinges on the competion of render to texture etc., the frame will be missed.
2014-08-28 04:56:19 +02:00
// This would have to be locked if the GPU thread didn't spin.
m_video_buffer_write_ptr = write_ptr + GPFifo::GATHER_PIPE_SIZE;
Add the 'desynced GPU thread' mode. It's a relatively big commit (less big with -w), but it's hard to test any of this separately... The basic problem is that in netplay or movies, the state of the CPU must be deterministic, including when the game receives notification that the GPU has processed FIFO data. Dual core mode notifies the game whenever the GPU thread actually gets around to doing the work, so it isn't deterministic. Single core mode is because it notifies the game 'instantly' (after processing the data synchronously), but it's too slow for many systems and games. My old dc-netplay branch worked as follows: everything worked as normal except the state of the CP registers was a lie, and the CPU thread only delivered results when idle detection triggered (waiting for the GPU if they weren't ready at that point). Usually, a game is idle iff all the work for the frame has been done, except for a small amount of work depending on the GPU result, so neither the CPU or the GPU waiting on the other affected performance much. However, it's possible that the game could be waiting for some earlier interrupt, and any of several games which, for whatever reason, never went into a detectable idle (even when I tried to improve the detection) would never receive results at all. (The current method should have better compatibility, but it also has slightly higher overhead and breaks some other things, so I want to reimplement this, hopefully with less impact on the code, in the future.) With this commit, the basic idea is that the CPU thread acts as if the work has been done instantly, like single core mode, but actually hands it off asynchronously to the GPU thread (after backing up some data that the game might change in memory before it's actually done). Since the work isn't done, any feedback from the GPU to the CPU, such as real XFB/EFB copies (virtual are OK), EFB pokes, performance queries, etc. is broken; but most games work with these options disabled, and there is no need to try to detect what the CPU thread is doing. Technically: when the flag g_use_deterministic_gpu_thread (currently stuck on) is on, the CPU thread calls RunGpu like in single core mode. This function synchronously copies the data from the FIFO to the internal video buffer and updates the CP registers, interrupts, etc. However, instead of the regular ReadDataFromFifo followed by running the opcode decoder, it runs ReadDataFromFifoOnCPU -> OpcodeDecoder_Preprocess, which relatively quickly scans through the FIFO data, detects SetFinish calls etc., which are immediately fired, and saves certain associated data from memory (e.g. display lists) in AuxBuffers (a parallel stream to the main FIFO, which is a bit slow at the moment), before handing the data off to the GPU thread to actually render. That makes up the bulk of this commit. In various circumstances, including the aforementioned EFB pokes and performance queries as well as swap requests (i.e. the end of a frame - we don't want the CPU potentially pumping out frames too quickly and the GPU falling behind*), SyncGPU is called to wait for actual completion. The overhead mainly comes from OpcodeDecoder_Preprocess (which is, again, synchronous), as well as the actual copying. Currently, display lists and such are escrowed from main memory even though they usually won't change over the course of a frame, and textures are not even though they might, resulting in a small chance of graphical glitches. When the texture locking (i.e. fault on write) code lands, I can make this all correct and maybe a little faster. * This suggests an alternate determinism method of just delaying results until a short time before the end of each frame. For all I know this might mostly work - I haven't tried it - but if any significant work hinges on the competion of render to texture etc., the frame will be missed.
2014-08-28 04:56:19 +02:00
}
void FifoManager::ResetVideoBuffer()
{
m_video_buffer_read_ptr = m_video_buffer;
m_video_buffer_write_ptr = m_video_buffer;
m_video_buffer_seen_ptr = m_video_buffer;
m_video_buffer_pp_read_ptr = m_video_buffer;
m_fifo_aux_write_ptr = m_fifo_aux_data;
m_fifo_aux_read_ptr = m_fifo_aux_data;
}
// Description: Main FIFO update loop
// Purpose: Keep the Core HW updated about the CPU-GPU distance
void FifoManager::RunGpuLoop()
{
AsyncRequests::GetInstance()->SetEnable(true);
AsyncRequests::GetInstance()->SetPassthrough(false);
m_gpu_mainloop.Run(
[this] {
// Run events from the CPU thread.
AsyncRequests::GetInstance()->PullEvents();
// Do nothing while paused
if (!m_emu_running_state.IsSet())
return;
if (m_use_deterministic_gpu_thread)
Add the 'desynced GPU thread' mode. It's a relatively big commit (less big with -w), but it's hard to test any of this separately... The basic problem is that in netplay or movies, the state of the CPU must be deterministic, including when the game receives notification that the GPU has processed FIFO data. Dual core mode notifies the game whenever the GPU thread actually gets around to doing the work, so it isn't deterministic. Single core mode is because it notifies the game 'instantly' (after processing the data synchronously), but it's too slow for many systems and games. My old dc-netplay branch worked as follows: everything worked as normal except the state of the CP registers was a lie, and the CPU thread only delivered results when idle detection triggered (waiting for the GPU if they weren't ready at that point). Usually, a game is idle iff all the work for the frame has been done, except for a small amount of work depending on the GPU result, so neither the CPU or the GPU waiting on the other affected performance much. However, it's possible that the game could be waiting for some earlier interrupt, and any of several games which, for whatever reason, never went into a detectable idle (even when I tried to improve the detection) would never receive results at all. (The current method should have better compatibility, but it also has slightly higher overhead and breaks some other things, so I want to reimplement this, hopefully with less impact on the code, in the future.) With this commit, the basic idea is that the CPU thread acts as if the work has been done instantly, like single core mode, but actually hands it off asynchronously to the GPU thread (after backing up some data that the game might change in memory before it's actually done). Since the work isn't done, any feedback from the GPU to the CPU, such as real XFB/EFB copies (virtual are OK), EFB pokes, performance queries, etc. is broken; but most games work with these options disabled, and there is no need to try to detect what the CPU thread is doing. Technically: when the flag g_use_deterministic_gpu_thread (currently stuck on) is on, the CPU thread calls RunGpu like in single core mode. This function synchronously copies the data from the FIFO to the internal video buffer and updates the CP registers, interrupts, etc. However, instead of the regular ReadDataFromFifo followed by running the opcode decoder, it runs ReadDataFromFifoOnCPU -> OpcodeDecoder_Preprocess, which relatively quickly scans through the FIFO data, detects SetFinish calls etc., which are immediately fired, and saves certain associated data from memory (e.g. display lists) in AuxBuffers (a parallel stream to the main FIFO, which is a bit slow at the moment), before handing the data off to the GPU thread to actually render. That makes up the bulk of this commit. In various circumstances, including the aforementioned EFB pokes and performance queries as well as swap requests (i.e. the end of a frame - we don't want the CPU potentially pumping out frames too quickly and the GPU falling behind*), SyncGPU is called to wait for actual completion. The overhead mainly comes from OpcodeDecoder_Preprocess (which is, again, synchronous), as well as the actual copying. Currently, display lists and such are escrowed from main memory even though they usually won't change over the course of a frame, and textures are not even though they might, resulting in a small chance of graphical glitches. When the texture locking (i.e. fault on write) code lands, I can make this all correct and maybe a little faster. * This suggests an alternate determinism method of just delaying results until a short time before the end of each frame. For all I know this might mostly work - I haven't tried it - but if any significant work hinges on the competion of render to texture etc., the frame will be missed.
2014-08-28 04:56:19 +02:00
{
// All the fifo/CP stuff is on the CPU. We just need to run the opcode decoder.
u8* seen_ptr = m_video_buffer_seen_ptr;
u8* write_ptr = m_video_buffer_write_ptr;
Add the 'desynced GPU thread' mode. It's a relatively big commit (less big with -w), but it's hard to test any of this separately... The basic problem is that in netplay or movies, the state of the CPU must be deterministic, including when the game receives notification that the GPU has processed FIFO data. Dual core mode notifies the game whenever the GPU thread actually gets around to doing the work, so it isn't deterministic. Single core mode is because it notifies the game 'instantly' (after processing the data synchronously), but it's too slow for many systems and games. My old dc-netplay branch worked as follows: everything worked as normal except the state of the CP registers was a lie, and the CPU thread only delivered results when idle detection triggered (waiting for the GPU if they weren't ready at that point). Usually, a game is idle iff all the work for the frame has been done, except for a small amount of work depending on the GPU result, so neither the CPU or the GPU waiting on the other affected performance much. However, it's possible that the game could be waiting for some earlier interrupt, and any of several games which, for whatever reason, never went into a detectable idle (even when I tried to improve the detection) would never receive results at all. (The current method should have better compatibility, but it also has slightly higher overhead and breaks some other things, so I want to reimplement this, hopefully with less impact on the code, in the future.) With this commit, the basic idea is that the CPU thread acts as if the work has been done instantly, like single core mode, but actually hands it off asynchronously to the GPU thread (after backing up some data that the game might change in memory before it's actually done). Since the work isn't done, any feedback from the GPU to the CPU, such as real XFB/EFB copies (virtual are OK), EFB pokes, performance queries, etc. is broken; but most games work with these options disabled, and there is no need to try to detect what the CPU thread is doing. Technically: when the flag g_use_deterministic_gpu_thread (currently stuck on) is on, the CPU thread calls RunGpu like in single core mode. This function synchronously copies the data from the FIFO to the internal video buffer and updates the CP registers, interrupts, etc. However, instead of the regular ReadDataFromFifo followed by running the opcode decoder, it runs ReadDataFromFifoOnCPU -> OpcodeDecoder_Preprocess, which relatively quickly scans through the FIFO data, detects SetFinish calls etc., which are immediately fired, and saves certain associated data from memory (e.g. display lists) in AuxBuffers (a parallel stream to the main FIFO, which is a bit slow at the moment), before handing the data off to the GPU thread to actually render. That makes up the bulk of this commit. In various circumstances, including the aforementioned EFB pokes and performance queries as well as swap requests (i.e. the end of a frame - we don't want the CPU potentially pumping out frames too quickly and the GPU falling behind*), SyncGPU is called to wait for actual completion. The overhead mainly comes from OpcodeDecoder_Preprocess (which is, again, synchronous), as well as the actual copying. Currently, display lists and such are escrowed from main memory even though they usually won't change over the course of a frame, and textures are not even though they might, resulting in a small chance of graphical glitches. When the texture locking (i.e. fault on write) code lands, I can make this all correct and maybe a little faster. * This suggests an alternate determinism method of just delaying results until a short time before the end of each frame. For all I know this might mostly work - I haven't tried it - but if any significant work hinges on the competion of render to texture etc., the frame will be missed.
2014-08-28 04:56:19 +02:00
// See comment in SyncGPU
if (write_ptr > seen_ptr)
{
m_video_buffer_read_ptr =
OpcodeDecoder::RunFifo(DataReader(m_video_buffer_read_ptr, write_ptr), nullptr);
m_video_buffer_seen_ptr = write_ptr;
}
Add the 'desynced GPU thread' mode. It's a relatively big commit (less big with -w), but it's hard to test any of this separately... The basic problem is that in netplay or movies, the state of the CPU must be deterministic, including when the game receives notification that the GPU has processed FIFO data. Dual core mode notifies the game whenever the GPU thread actually gets around to doing the work, so it isn't deterministic. Single core mode is because it notifies the game 'instantly' (after processing the data synchronously), but it's too slow for many systems and games. My old dc-netplay branch worked as follows: everything worked as normal except the state of the CP registers was a lie, and the CPU thread only delivered results when idle detection triggered (waiting for the GPU if they weren't ready at that point). Usually, a game is idle iff all the work for the frame has been done, except for a small amount of work depending on the GPU result, so neither the CPU or the GPU waiting on the other affected performance much. However, it's possible that the game could be waiting for some earlier interrupt, and any of several games which, for whatever reason, never went into a detectable idle (even when I tried to improve the detection) would never receive results at all. (The current method should have better compatibility, but it also has slightly higher overhead and breaks some other things, so I want to reimplement this, hopefully with less impact on the code, in the future.) With this commit, the basic idea is that the CPU thread acts as if the work has been done instantly, like single core mode, but actually hands it off asynchronously to the GPU thread (after backing up some data that the game might change in memory before it's actually done). Since the work isn't done, any feedback from the GPU to the CPU, such as real XFB/EFB copies (virtual are OK), EFB pokes, performance queries, etc. is broken; but most games work with these options disabled, and there is no need to try to detect what the CPU thread is doing. Technically: when the flag g_use_deterministic_gpu_thread (currently stuck on) is on, the CPU thread calls RunGpu like in single core mode. This function synchronously copies the data from the FIFO to the internal video buffer and updates the CP registers, interrupts, etc. However, instead of the regular ReadDataFromFifo followed by running the opcode decoder, it runs ReadDataFromFifoOnCPU -> OpcodeDecoder_Preprocess, which relatively quickly scans through the FIFO data, detects SetFinish calls etc., which are immediately fired, and saves certain associated data from memory (e.g. display lists) in AuxBuffers (a parallel stream to the main FIFO, which is a bit slow at the moment), before handing the data off to the GPU thread to actually render. That makes up the bulk of this commit. In various circumstances, including the aforementioned EFB pokes and performance queries as well as swap requests (i.e. the end of a frame - we don't want the CPU potentially pumping out frames too quickly and the GPU falling behind*), SyncGPU is called to wait for actual completion. The overhead mainly comes from OpcodeDecoder_Preprocess (which is, again, synchronous), as well as the actual copying. Currently, display lists and such are escrowed from main memory even though they usually won't change over the course of a frame, and textures are not even though they might, resulting in a small chance of graphical glitches. When the texture locking (i.e. fault on write) code lands, I can make this all correct and maybe a little faster. * This suggests an alternate determinism method of just delaying results until a short time before the end of each frame. For all I know this might mostly work - I haven't tried it - but if any significant work hinges on the competion of render to texture etc., the frame will be missed.
2014-08-28 04:56:19 +02:00
}
else
{
auto& command_processor = m_system.GetCommandProcessor();
auto& fifo = command_processor.GetFifo();
command_processor.SetCPStatusFromGPU();
Add the 'desynced GPU thread' mode. It's a relatively big commit (less big with -w), but it's hard to test any of this separately... The basic problem is that in netplay or movies, the state of the CPU must be deterministic, including when the game receives notification that the GPU has processed FIFO data. Dual core mode notifies the game whenever the GPU thread actually gets around to doing the work, so it isn't deterministic. Single core mode is because it notifies the game 'instantly' (after processing the data synchronously), but it's too slow for many systems and games. My old dc-netplay branch worked as follows: everything worked as normal except the state of the CP registers was a lie, and the CPU thread only delivered results when idle detection triggered (waiting for the GPU if they weren't ready at that point). Usually, a game is idle iff all the work for the frame has been done, except for a small amount of work depending on the GPU result, so neither the CPU or the GPU waiting on the other affected performance much. However, it's possible that the game could be waiting for some earlier interrupt, and any of several games which, for whatever reason, never went into a detectable idle (even when I tried to improve the detection) would never receive results at all. (The current method should have better compatibility, but it also has slightly higher overhead and breaks some other things, so I want to reimplement this, hopefully with less impact on the code, in the future.) With this commit, the basic idea is that the CPU thread acts as if the work has been done instantly, like single core mode, but actually hands it off asynchronously to the GPU thread (after backing up some data that the game might change in memory before it's actually done). Since the work isn't done, any feedback from the GPU to the CPU, such as real XFB/EFB copies (virtual are OK), EFB pokes, performance queries, etc. is broken; but most games work with these options disabled, and there is no need to try to detect what the CPU thread is doing. Technically: when the flag g_use_deterministic_gpu_thread (currently stuck on) is on, the CPU thread calls RunGpu like in single core mode. This function synchronously copies the data from the FIFO to the internal video buffer and updates the CP registers, interrupts, etc. However, instead of the regular ReadDataFromFifo followed by running the opcode decoder, it runs ReadDataFromFifoOnCPU -> OpcodeDecoder_Preprocess, which relatively quickly scans through the FIFO data, detects SetFinish calls etc., which are immediately fired, and saves certain associated data from memory (e.g. display lists) in AuxBuffers (a parallel stream to the main FIFO, which is a bit slow at the moment), before handing the data off to the GPU thread to actually render. That makes up the bulk of this commit. In various circumstances, including the aforementioned EFB pokes and performance queries as well as swap requests (i.e. the end of a frame - we don't want the CPU potentially pumping out frames too quickly and the GPU falling behind*), SyncGPU is called to wait for actual completion. The overhead mainly comes from OpcodeDecoder_Preprocess (which is, again, synchronous), as well as the actual copying. Currently, display lists and such are escrowed from main memory even though they usually won't change over the course of a frame, and textures are not even though they might, resulting in a small chance of graphical glitches. When the texture locking (i.e. fault on write) code lands, I can make this all correct and maybe a little faster. * This suggests an alternate determinism method of just delaying results until a short time before the end of each frame. For all I know this might mostly work - I haven't tried it - but if any significant work hinges on the competion of render to texture etc., the frame will be missed.
2014-08-28 04:56:19 +02:00
// check if we are able to run this buffer
while (!command_processor.IsInterruptWaiting() &&
fifo.bFF_GPReadEnable.load(std::memory_order_relaxed) &&
fifo.CPReadWriteDistance.load(std::memory_order_relaxed) &&
!AtBreakpoint(m_system))
{
if (m_config_sync_gpu && m_sync_ticks.load() < m_config_sync_gpu_min_distance)
break;
u32 cyclesExecuted = 0;
u32 readPtr = fifo.CPReadPointer.load(std::memory_order_relaxed);
ReadDataFromFifo(readPtr);
if (readPtr == fifo.CPEnd.load(std::memory_order_relaxed))
readPtr = fifo.CPBase.load(std::memory_order_relaxed);
else
readPtr += GPFifo::GATHER_PIPE_SIZE;
const s32 distance =
static_cast<s32>(fifo.CPReadWriteDistance.load(std::memory_order_relaxed)) -
GPFifo::GATHER_PIPE_SIZE;
ASSERT_MSG(COMMANDPROCESSOR, distance >= 0,
2021-11-10 23:42:49 +01:00
"Negative fifo.CPReadWriteDistance = {} in FIFO Loop !\nThat can produce "
"instability in the game. Please report it.",
distance);
u8* write_ptr = m_video_buffer_write_ptr;
m_video_buffer_read_ptr = OpcodeDecoder::RunFifo(
DataReader(m_video_buffer_read_ptr, write_ptr), &cyclesExecuted);
fifo.CPReadPointer.store(readPtr, std::memory_order_relaxed);
fifo.CPReadWriteDistance.fetch_sub(GPFifo::GATHER_PIPE_SIZE, std::memory_order_seq_cst);
if ((write_ptr - m_video_buffer_read_ptr) == 0)
{
fifo.SafeCPReadPointer.store(fifo.CPReadPointer.load(std::memory_order_relaxed),
std::memory_order_relaxed);
}
command_processor.SetCPStatusFromGPU();
if (m_config_sync_gpu)
2015-03-29 12:20:24 +02:00
{
cyclesExecuted = (int)(cyclesExecuted / m_config_sync_gpu_overclock);
int old = m_sync_ticks.fetch_sub(cyclesExecuted);
if (old >= m_config_sync_gpu_max_distance &&
old - (int)cyclesExecuted < m_config_sync_gpu_max_distance)
{
m_sync_wakeup_event.Set();
}
2015-03-29 12:20:24 +02:00
}
Add the 'desynced GPU thread' mode. It's a relatively big commit (less big with -w), but it's hard to test any of this separately... The basic problem is that in netplay or movies, the state of the CPU must be deterministic, including when the game receives notification that the GPU has processed FIFO data. Dual core mode notifies the game whenever the GPU thread actually gets around to doing the work, so it isn't deterministic. Single core mode is because it notifies the game 'instantly' (after processing the data synchronously), but it's too slow for many systems and games. My old dc-netplay branch worked as follows: everything worked as normal except the state of the CP registers was a lie, and the CPU thread only delivered results when idle detection triggered (waiting for the GPU if they weren't ready at that point). Usually, a game is idle iff all the work for the frame has been done, except for a small amount of work depending on the GPU result, so neither the CPU or the GPU waiting on the other affected performance much. However, it's possible that the game could be waiting for some earlier interrupt, and any of several games which, for whatever reason, never went into a detectable idle (even when I tried to improve the detection) would never receive results at all. (The current method should have better compatibility, but it also has slightly higher overhead and breaks some other things, so I want to reimplement this, hopefully with less impact on the code, in the future.) With this commit, the basic idea is that the CPU thread acts as if the work has been done instantly, like single core mode, but actually hands it off asynchronously to the GPU thread (after backing up some data that the game might change in memory before it's actually done). Since the work isn't done, any feedback from the GPU to the CPU, such as real XFB/EFB copies (virtual are OK), EFB pokes, performance queries, etc. is broken; but most games work with these options disabled, and there is no need to try to detect what the CPU thread is doing. Technically: when the flag g_use_deterministic_gpu_thread (currently stuck on) is on, the CPU thread calls RunGpu like in single core mode. This function synchronously copies the data from the FIFO to the internal video buffer and updates the CP registers, interrupts, etc. However, instead of the regular ReadDataFromFifo followed by running the opcode decoder, it runs ReadDataFromFifoOnCPU -> OpcodeDecoder_Preprocess, which relatively quickly scans through the FIFO data, detects SetFinish calls etc., which are immediately fired, and saves certain associated data from memory (e.g. display lists) in AuxBuffers (a parallel stream to the main FIFO, which is a bit slow at the moment), before handing the data off to the GPU thread to actually render. That makes up the bulk of this commit. In various circumstances, including the aforementioned EFB pokes and performance queries as well as swap requests (i.e. the end of a frame - we don't want the CPU potentially pumping out frames too quickly and the GPU falling behind*), SyncGPU is called to wait for actual completion. The overhead mainly comes from OpcodeDecoder_Preprocess (which is, again, synchronous), as well as the actual copying. Currently, display lists and such are escrowed from main memory even though they usually won't change over the course of a frame, and textures are not even though they might, resulting in a small chance of graphical glitches. When the texture locking (i.e. fault on write) code lands, I can make this all correct and maybe a little faster. * This suggests an alternate determinism method of just delaying results until a short time before the end of each frame. For all I know this might mostly work - I haven't tried it - but if any significant work hinges on the competion of render to texture etc., the frame will be missed.
2014-08-28 04:56:19 +02:00
// This call is pretty important in DualCore mode and must be called in the FIFO Loop.
// If we don't, s_swapRequested or s_efbAccessRequested won't be set to false
// leading the CPU thread to wait in Video_OutputXFB or Video_AccessEFB thus slowing
Add the 'desynced GPU thread' mode. It's a relatively big commit (less big with -w), but it's hard to test any of this separately... The basic problem is that in netplay or movies, the state of the CPU must be deterministic, including when the game receives notification that the GPU has processed FIFO data. Dual core mode notifies the game whenever the GPU thread actually gets around to doing the work, so it isn't deterministic. Single core mode is because it notifies the game 'instantly' (after processing the data synchronously), but it's too slow for many systems and games. My old dc-netplay branch worked as follows: everything worked as normal except the state of the CP registers was a lie, and the CPU thread only delivered results when idle detection triggered (waiting for the GPU if they weren't ready at that point). Usually, a game is idle iff all the work for the frame has been done, except for a small amount of work depending on the GPU result, so neither the CPU or the GPU waiting on the other affected performance much. However, it's possible that the game could be waiting for some earlier interrupt, and any of several games which, for whatever reason, never went into a detectable idle (even when I tried to improve the detection) would never receive results at all. (The current method should have better compatibility, but it also has slightly higher overhead and breaks some other things, so I want to reimplement this, hopefully with less impact on the code, in the future.) With this commit, the basic idea is that the CPU thread acts as if the work has been done instantly, like single core mode, but actually hands it off asynchronously to the GPU thread (after backing up some data that the game might change in memory before it's actually done). Since the work isn't done, any feedback from the GPU to the CPU, such as real XFB/EFB copies (virtual are OK), EFB pokes, performance queries, etc. is broken; but most games work with these options disabled, and there is no need to try to detect what the CPU thread is doing. Technically: when the flag g_use_deterministic_gpu_thread (currently stuck on) is on, the CPU thread calls RunGpu like in single core mode. This function synchronously copies the data from the FIFO to the internal video buffer and updates the CP registers, interrupts, etc. However, instead of the regular ReadDataFromFifo followed by running the opcode decoder, it runs ReadDataFromFifoOnCPU -> OpcodeDecoder_Preprocess, which relatively quickly scans through the FIFO data, detects SetFinish calls etc., which are immediately fired, and saves certain associated data from memory (e.g. display lists) in AuxBuffers (a parallel stream to the main FIFO, which is a bit slow at the moment), before handing the data off to the GPU thread to actually render. That makes up the bulk of this commit. In various circumstances, including the aforementioned EFB pokes and performance queries as well as swap requests (i.e. the end of a frame - we don't want the CPU potentially pumping out frames too quickly and the GPU falling behind*), SyncGPU is called to wait for actual completion. The overhead mainly comes from OpcodeDecoder_Preprocess (which is, again, synchronous), as well as the actual copying. Currently, display lists and such are escrowed from main memory even though they usually won't change over the course of a frame, and textures are not even though they might, resulting in a small chance of graphical glitches. When the texture locking (i.e. fault on write) code lands, I can make this all correct and maybe a little faster. * This suggests an alternate determinism method of just delaying results until a short time before the end of each frame. For all I know this might mostly work - I haven't tried it - but if any significant work hinges on the competion of render to texture etc., the frame will be missed.
2014-08-28 04:56:19 +02:00
// things down.
AsyncRequests::GetInstance()->PullEvents();
Add the 'desynced GPU thread' mode. It's a relatively big commit (less big with -w), but it's hard to test any of this separately... The basic problem is that in netplay or movies, the state of the CPU must be deterministic, including when the game receives notification that the GPU has processed FIFO data. Dual core mode notifies the game whenever the GPU thread actually gets around to doing the work, so it isn't deterministic. Single core mode is because it notifies the game 'instantly' (after processing the data synchronously), but it's too slow for many systems and games. My old dc-netplay branch worked as follows: everything worked as normal except the state of the CP registers was a lie, and the CPU thread only delivered results when idle detection triggered (waiting for the GPU if they weren't ready at that point). Usually, a game is idle iff all the work for the frame has been done, except for a small amount of work depending on the GPU result, so neither the CPU or the GPU waiting on the other affected performance much. However, it's possible that the game could be waiting for some earlier interrupt, and any of several games which, for whatever reason, never went into a detectable idle (even when I tried to improve the detection) would never receive results at all. (The current method should have better compatibility, but it also has slightly higher overhead and breaks some other things, so I want to reimplement this, hopefully with less impact on the code, in the future.) With this commit, the basic idea is that the CPU thread acts as if the work has been done instantly, like single core mode, but actually hands it off asynchronously to the GPU thread (after backing up some data that the game might change in memory before it's actually done). Since the work isn't done, any feedback from the GPU to the CPU, such as real XFB/EFB copies (virtual are OK), EFB pokes, performance queries, etc. is broken; but most games work with these options disabled, and there is no need to try to detect what the CPU thread is doing. Technically: when the flag g_use_deterministic_gpu_thread (currently stuck on) is on, the CPU thread calls RunGpu like in single core mode. This function synchronously copies the data from the FIFO to the internal video buffer and updates the CP registers, interrupts, etc. However, instead of the regular ReadDataFromFifo followed by running the opcode decoder, it runs ReadDataFromFifoOnCPU -> OpcodeDecoder_Preprocess, which relatively quickly scans through the FIFO data, detects SetFinish calls etc., which are immediately fired, and saves certain associated data from memory (e.g. display lists) in AuxBuffers (a parallel stream to the main FIFO, which is a bit slow at the moment), before handing the data off to the GPU thread to actually render. That makes up the bulk of this commit. In various circumstances, including the aforementioned EFB pokes and performance queries as well as swap requests (i.e. the end of a frame - we don't want the CPU potentially pumping out frames too quickly and the GPU falling behind*), SyncGPU is called to wait for actual completion. The overhead mainly comes from OpcodeDecoder_Preprocess (which is, again, synchronous), as well as the actual copying. Currently, display lists and such are escrowed from main memory even though they usually won't change over the course of a frame, and textures are not even though they might, resulting in a small chance of graphical glitches. When the texture locking (i.e. fault on write) code lands, I can make this all correct and maybe a little faster. * This suggests an alternate determinism method of just delaying results until a short time before the end of each frame. For all I know this might mostly work - I haven't tried it - but if any significant work hinges on the competion of render to texture etc., the frame will be missed.
2014-08-28 04:56:19 +02:00
}
// fast skip remaining GPU time if fifo is empty
if (m_sync_ticks.load() > 0)
{
int old = m_sync_ticks.exchange(0);
if (old >= m_config_sync_gpu_max_distance)
m_sync_wakeup_event.Set();
}
// The fifo is empty and it's unlikely we will get any more work in the near future.
// Make sure VertexManager finishes drawing any primitives it has stored in it's buffer.
g_vertex_manager->Flush();
g_framebuffer_manager->RefreshPeekCache();
Add the 'desynced GPU thread' mode. It's a relatively big commit (less big with -w), but it's hard to test any of this separately... The basic problem is that in netplay or movies, the state of the CPU must be deterministic, including when the game receives notification that the GPU has processed FIFO data. Dual core mode notifies the game whenever the GPU thread actually gets around to doing the work, so it isn't deterministic. Single core mode is because it notifies the game 'instantly' (after processing the data synchronously), but it's too slow for many systems and games. My old dc-netplay branch worked as follows: everything worked as normal except the state of the CP registers was a lie, and the CPU thread only delivered results when idle detection triggered (waiting for the GPU if they weren't ready at that point). Usually, a game is idle iff all the work for the frame has been done, except for a small amount of work depending on the GPU result, so neither the CPU or the GPU waiting on the other affected performance much. However, it's possible that the game could be waiting for some earlier interrupt, and any of several games which, for whatever reason, never went into a detectable idle (even when I tried to improve the detection) would never receive results at all. (The current method should have better compatibility, but it also has slightly higher overhead and breaks some other things, so I want to reimplement this, hopefully with less impact on the code, in the future.) With this commit, the basic idea is that the CPU thread acts as if the work has been done instantly, like single core mode, but actually hands it off asynchronously to the GPU thread (after backing up some data that the game might change in memory before it's actually done). Since the work isn't done, any feedback from the GPU to the CPU, such as real XFB/EFB copies (virtual are OK), EFB pokes, performance queries, etc. is broken; but most games work with these options disabled, and there is no need to try to detect what the CPU thread is doing. Technically: when the flag g_use_deterministic_gpu_thread (currently stuck on) is on, the CPU thread calls RunGpu like in single core mode. This function synchronously copies the data from the FIFO to the internal video buffer and updates the CP registers, interrupts, etc. However, instead of the regular ReadDataFromFifo followed by running the opcode decoder, it runs ReadDataFromFifoOnCPU -> OpcodeDecoder_Preprocess, which relatively quickly scans through the FIFO data, detects SetFinish calls etc., which are immediately fired, and saves certain associated data from memory (e.g. display lists) in AuxBuffers (a parallel stream to the main FIFO, which is a bit slow at the moment), before handing the data off to the GPU thread to actually render. That makes up the bulk of this commit. In various circumstances, including the aforementioned EFB pokes and performance queries as well as swap requests (i.e. the end of a frame - we don't want the CPU potentially pumping out frames too quickly and the GPU falling behind*), SyncGPU is called to wait for actual completion. The overhead mainly comes from OpcodeDecoder_Preprocess (which is, again, synchronous), as well as the actual copying. Currently, display lists and such are escrowed from main memory even though they usually won't change over the course of a frame, and textures are not even though they might, resulting in a small chance of graphical glitches. When the texture locking (i.e. fault on write) code lands, I can make this all correct and maybe a little faster. * This suggests an alternate determinism method of just delaying results until a short time before the end of each frame. For all I know this might mostly work - I haven't tried it - but if any significant work hinges on the competion of render to texture etc., the frame will be missed.
2014-08-28 04:56:19 +02:00
}
},
100);
AsyncRequests::GetInstance()->SetEnable(false);
AsyncRequests::GetInstance()->SetPassthrough(true);
}
void FifoManager::FlushGpu()
{
if (!m_system.IsDualCoreMode() || m_use_deterministic_gpu_thread)
return;
m_gpu_mainloop.Wait();
}
void FifoManager::GpuMaySleep()
{
m_gpu_mainloop.AllowSleep();
}
BIG FIFO Commit PART 1! Sometimes you need to write everything from 0, so 10 days later Ive rewrited some parts of the FIFO in Dual Core mode. Is pending use the same code for SC mode. - Improved the GP Register Status: now this is all the time from the fifo loop. - Improved the Interrupts manage: 1) Removed All UpdateInturrupts from CommandProcessor Writes and Read. 2) now the CP interrupts are schedule from the video thread and the fifo loop waiting until this happens 3) considering Inmediate mode for the CP interrupts 3) Implemented Interrupt CP Cache State 4) Implemented only Overflow interrupt in GatherPipeCheck because this need to be detected quickly. - Implemented Overflow handling like a real HW, when Hiwatermark interrupt happens this write ClearRegister with True en HI and False in LO (FifoIntReset) after that a Control Register is writed and the FIFO is processed to LO Watermark. - Removed all ugly code from LO and HI watermark manage - Removed all ugly code from BP manage - Change >= by == in the BP clauses - Removed speed hack (1024 chunk) for better GP Status Control. - Commented GXSetGPFifo very soon hack - Commented FackWatchDog hack - Commented FIFO_RW_DISTANCE = WritePointer hack This is the beginning and the base for the future., If this broke your favorite game my apologize, only report this and will try solve it. If you have a Overflown don't worry, I've implemented the real solution code using the Overflow Interruption only need continue working for a perfect protection. Why I did it? Because is preferable a accurate and clean fifo instead hack y fifo for improve that. Thanks to DONKO for you awesome Video Plug in and skid for the chatting. PD: I have 7-10 fps more in the star fox video. bye :P git-svn-id: https://dolphin-emu.googlecode.com/svn/trunk@6554 8ced0084-cf51-0410-be5f-012b33b47a6e
2010-12-11 13:42:55 +01:00
bool AtBreakpoint(Core::System& system)
BIG FIFO Commit PART 1! Sometimes you need to write everything from 0, so 10 days later Ive rewrited some parts of the FIFO in Dual Core mode. Is pending use the same code for SC mode. - Improved the GP Register Status: now this is all the time from the fifo loop. - Improved the Interrupts manage: 1) Removed All UpdateInturrupts from CommandProcessor Writes and Read. 2) now the CP interrupts are schedule from the video thread and the fifo loop waiting until this happens 3) considering Inmediate mode for the CP interrupts 3) Implemented Interrupt CP Cache State 4) Implemented only Overflow interrupt in GatherPipeCheck because this need to be detected quickly. - Implemented Overflow handling like a real HW, when Hiwatermark interrupt happens this write ClearRegister with True en HI and False in LO (FifoIntReset) after that a Control Register is writed and the FIFO is processed to LO Watermark. - Removed all ugly code from LO and HI watermark manage - Removed all ugly code from BP manage - Change >= by == in the BP clauses - Removed speed hack (1024 chunk) for better GP Status Control. - Commented GXSetGPFifo very soon hack - Commented FackWatchDog hack - Commented FIFO_RW_DISTANCE = WritePointer hack This is the beginning and the base for the future., If this broke your favorite game my apologize, only report this and will try solve it. If you have a Overflown don't worry, I've implemented the real solution code using the Overflow Interruption only need continue working for a perfect protection. Why I did it? Because is preferable a accurate and clean fifo instead hack y fifo for improve that. Thanks to DONKO for you awesome Video Plug in and skid for the chatting. PD: I have 7-10 fps more in the star fox video. bye :P git-svn-id: https://dolphin-emu.googlecode.com/svn/trunk@6554 8ced0084-cf51-0410-be5f-012b33b47a6e
2010-12-11 13:42:55 +01:00
{
auto& command_processor = system.GetCommandProcessor();
const auto& fifo = command_processor.GetFifo();
return fifo.bFF_BPEnable.load(std::memory_order_relaxed) &&
(fifo.CPReadPointer.load(std::memory_order_relaxed) ==
fifo.CPBreakpoint.load(std::memory_order_relaxed));
BIG FIFO Commit PART 1! Sometimes you need to write everything from 0, so 10 days later Ive rewrited some parts of the FIFO in Dual Core mode. Is pending use the same code for SC mode. - Improved the GP Register Status: now this is all the time from the fifo loop. - Improved the Interrupts manage: 1) Removed All UpdateInturrupts from CommandProcessor Writes and Read. 2) now the CP interrupts are schedule from the video thread and the fifo loop waiting until this happens 3) considering Inmediate mode for the CP interrupts 3) Implemented Interrupt CP Cache State 4) Implemented only Overflow interrupt in GatherPipeCheck because this need to be detected quickly. - Implemented Overflow handling like a real HW, when Hiwatermark interrupt happens this write ClearRegister with True en HI and False in LO (FifoIntReset) after that a Control Register is writed and the FIFO is processed to LO Watermark. - Removed all ugly code from LO and HI watermark manage - Removed all ugly code from BP manage - Change >= by == in the BP clauses - Removed speed hack (1024 chunk) for better GP Status Control. - Commented GXSetGPFifo very soon hack - Commented FackWatchDog hack - Commented FIFO_RW_DISTANCE = WritePointer hack This is the beginning and the base for the future., If this broke your favorite game my apologize, only report this and will try solve it. If you have a Overflown don't worry, I've implemented the real solution code using the Overflow Interruption only need continue working for a perfect protection. Why I did it? Because is preferable a accurate and clean fifo instead hack y fifo for improve that. Thanks to DONKO for you awesome Video Plug in and skid for the chatting. PD: I have 7-10 fps more in the star fox video. bye :P git-svn-id: https://dolphin-emu.googlecode.com/svn/trunk@6554 8ced0084-cf51-0410-be5f-012b33b47a6e
2010-12-11 13:42:55 +01:00
}
void FifoManager::RunGpu()
{
const bool is_dual_core = m_system.IsDualCoreMode();
// wake up GPU thread
if (is_dual_core && !m_use_deterministic_gpu_thread)
{
m_gpu_mainloop.Wakeup();
}
// if the sync GPU callback is suspended, wake it up.
if (!is_dual_core || m_use_deterministic_gpu_thread || m_config_sync_gpu)
{
if (m_syncing_suspended)
{
m_syncing_suspended = false;
m_system.GetCoreTiming().ScheduleEvent(GPU_TIME_SLOT_SIZE, m_event_sync_gpu,
GPU_TIME_SLOT_SIZE);
}
}
}
int FifoManager::RunGpuOnCpu(int ticks)
{
auto& command_processor = m_system.GetCommandProcessor();
auto& fifo = command_processor.GetFifo();
bool reset_simd_state = false;
int available_ticks = int(ticks * m_config_sync_gpu_overclock) + m_sync_ticks.load();
while (fifo.bFF_GPReadEnable.load(std::memory_order_relaxed) &&
fifo.CPReadWriteDistance.load(std::memory_order_relaxed) && !AtBreakpoint(m_system) &&
available_ticks >= 0)
{
if (m_use_deterministic_gpu_thread)
{
ReadDataFromFifoOnCPU(fifo.CPReadPointer.load(std::memory_order_relaxed));
m_gpu_mainloop.Wakeup();
}
else
{
if (!reset_simd_state)
{
Common::FPU::SaveSIMDState();
Common::FPU::LoadDefaultSIMDState();
reset_simd_state = true;
}
ReadDataFromFifo(fifo.CPReadPointer.load(std::memory_order_relaxed));
u32 cycles = 0;
m_video_buffer_read_ptr = OpcodeDecoder::RunFifo(
DataReader(m_video_buffer_read_ptr, m_video_buffer_write_ptr), &cycles);
available_ticks -= cycles;
}
if (fifo.CPReadPointer.load(std::memory_order_relaxed) ==
fifo.CPEnd.load(std::memory_order_relaxed))
{
fifo.CPReadPointer.store(fifo.CPBase.load(std::memory_order_relaxed),
std::memory_order_relaxed);
}
else
{
fifo.CPReadPointer.fetch_add(GPFifo::GATHER_PIPE_SIZE, std::memory_order_relaxed);
}
fifo.CPReadWriteDistance.fetch_sub(GPFifo::GATHER_PIPE_SIZE, std::memory_order_relaxed);
}
command_processor.SetCPStatusFromGPU();
if (reset_simd_state)
{
Common::FPU::LoadSIMDState();
}
// Discard all available ticks as there is nothing to do any more.
m_sync_ticks.store(std::min(available_ticks, 0));
// If the GPU is idle, drop the handler.
if (available_ticks >= 0)
return -1;
// Always wait at least for GPU_TIME_SLOT_SIZE cycles.
return -available_ticks + GPU_TIME_SLOT_SIZE;
}
void FifoManager::UpdateWantDeterminism(bool want)
{
// We are paused (or not running at all yet), so
// it should be safe to change this.
bool gpu_thread = false;
switch (Config::GetGPUDeterminismMode())
{
case Config::GPUDeterminismMode::Auto:
gpu_thread = want;
break;
case Config::GPUDeterminismMode::Disabled:
gpu_thread = false;
break;
case Config::GPUDeterminismMode::FakeCompletion:
gpu_thread = true;
break;
}
gpu_thread = gpu_thread && m_system.IsDualCoreMode();
if (m_use_deterministic_gpu_thread != gpu_thread)
{
m_use_deterministic_gpu_thread = gpu_thread;
if (gpu_thread)
{
// These haven't been updated in non-deterministic mode.
m_video_buffer_seen_ptr = m_video_buffer_pp_read_ptr = m_video_buffer_read_ptr;
CopyPreprocessCPStateFromMain();
VertexLoaderManager::MarkAllDirty();
}
}
}
/* This function checks the emulated CPU - GPU distance and may wake up the GPU,
* or block the CPU if required. It should be called by the CPU thread regularly.
* @ticks The gone emulated CPU time.
* @return A good time to call WaitForGpuThread() next.
*/
int FifoManager::WaitForGpuThread(int ticks)
{
int old = m_sync_ticks.fetch_add(ticks);
2016-10-01 13:16:50 +02:00
int now = old + ticks;
// GPU is idle, so stop polling.
if (old >= 0 && m_gpu_mainloop.IsDone())
2016-10-01 13:16:50 +02:00
return -1;
// Wakeup GPU
if (old < m_config_sync_gpu_min_distance && now >= m_config_sync_gpu_min_distance)
RunGpu();
2016-10-01 13:16:50 +02:00
// If the GPU is still sleeping, wait for a longer time
if (now < m_config_sync_gpu_min_distance)
return GPU_TIME_SLOT_SIZE + m_config_sync_gpu_min_distance - now;
2016-10-01 13:16:50 +02:00
// Wait for GPU
if (now >= m_config_sync_gpu_max_distance)
m_sync_wakeup_event.Wait();
2016-10-01 13:16:50 +02:00
return GPU_TIME_SLOT_SIZE;
}
2016-01-12 22:44:58 +01:00
void FifoManager::SyncGPUCallback(Core::System& system, u64 ticks, s64 cyclesLate)
{
ticks += cyclesLate;
int next = -1;
auto& fifo = system.GetFifo();
if (!system.IsDualCoreMode() || fifo.m_use_deterministic_gpu_thread)
{
next = fifo.RunGpuOnCpu(int(ticks));
}
else if (fifo.m_config_sync_gpu)
{
next = fifo.WaitForGpuThread(int(ticks));
}
fifo.m_syncing_suspended = next < 0;
if (!fifo.m_syncing_suspended)
system.GetCoreTiming().ScheduleEvent(next, fifo.m_event_sync_gpu, next);
}
void FifoManager::SyncGPUForRegisterAccess()
{
SyncGPU(SyncGPUReason::Other);
if (!m_system.IsDualCoreMode() || m_use_deterministic_gpu_thread)
RunGpuOnCpu(GPU_TIME_SLOT_SIZE);
else if (m_config_sync_gpu)
WaitForGpuThread(GPU_TIME_SLOT_SIZE);
}
// Initialize GPU - CPU thread syncing, this gives us a deterministic way to start the GPU thread.
void FifoManager::Prepare()
{
m_event_sync_gpu = m_system.GetCoreTiming().RegisterEvent("SyncGPUCallback", SyncGPUCallback);
m_syncing_suspended = true;
}
} // namespace Fifo