1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
|
// SPDX-FileCopyrightText: Copyright 2026 Eden Emulator Project
// SPDX-License-Identifier: GPL-3.0-or-later
// SPDX-FileCopyrightText: Copyright 2019 yuzu Emulator Project
// SPDX-License-Identifier: GPL-2.0-or-later
#include <chrono>
#include <thread>
#include "common/assert.h"
#include "common/scope_exit.h"
#include "common/settings.h"
#include "common/thread.h"
#include "core/core.h"
#include "core/frontend/graphics_context.h"
#include "video_core/control/scheduler.h"
#include "video_core/dma_pusher.h"
#include "video_core/gpu.h"
#include "video_core/gpu_thread.h"
#include "video_core/gpu_split.h"
#include "video_core/nx2_profile.h"
#include "video_core/host1x/host1x.h"
#include "video_core/renderer_base.h"
namespace VideoCommon::GPUThread {
namespace {
// Tells a CPU thread waiting in PushCommand that its command ran.
void SignalCommand(SynchState& state, u64 fence, bool block) {
state.signaled_fence.store(fence);
if (block) {
// We have to lock the write_lock to ensure that the condition_variable wait not get a
// race between the check and the lock itself.
std::scoped_lock lk{state.write_lock};
state.cv.notify_all();
}
}
} // Anonymous namespace
ThreadManager::ThreadManager(Core::System& system_)
: system{system_}
{}
ThreadManager::~ThreadManager() = default;
void ThreadManager::StartThread(VideoCore::RendererBase& renderer, Core::Frontend::GraphicsContext& context, Tegra::Control::Scheduler& scheduler) {
rasterizer = renderer.ReadRasterizer();
thread = std::jthread([&](std::stop_token stop_token) {
Common::SetCurrentThreadName("GPU");
Common::SetCurrentThreadPriority(Common::ThreadPriority::Critical);
Common::SetCurrentThreadToPerformanceCores();
system.RegisterHostThread();
auto current_context = context.Acquire();
CommandDataContainer next;
// When a pending channel last moved or a command arrived (see the polling below).
auto last_progress = std::chrono::steady_clock::now();
namespace Profile = VideoCommon::Nx2Profile;
auto& split = VideoCommon::GpuSplit::Get();
while (!stop_token.stop_requested()) {
Profile::Report();
split.MaybeStart(system);
if (scheduler.HasPendingChannels()) {
// Suspended channels wait on semaphores the CPU releases: poll them meanwhile.
{
Profile::Scope scope{Profile::Zone::ResumePending};
scheduler.ResumePending(system.GPU());
}
if (!state.queue.TryPop(next)) {
// The CPU usually releases the semaphore soon after: poll for half a
// millisecond before sleeping, or every wait also waits out a host sleep
// (title 1 waits ~25 times a second).
Profile::Scope scope{Profile::Zone::PendingSleep};
if (std::chrono::steady_clock::now() - last_progress <
std::chrono::microseconds{500}) {
std::this_thread::yield();
} else {
std::this_thread::sleep_for(std::chrono::milliseconds{1});
}
continue;
}
} else {
Profile::Scope scope{Profile::Zone::Idle};
state.queue.PopWait(next, stop_token);
}
if (stop_token.stop_requested()) {
break;
}
last_progress = std::chrono::steady_clock::now();
if (split.Active()) {
// Only command lists are parsed here; everything else runs on the back thread
// in order, and so does the signal that the command ran.
auto* const rasterizer_ptr = renderer.ReadRasterizer();
if (auto* submit_list = std::get_if<SubmitListCommand>(&next.data)) {
Profile::Scope scope{Profile::Zone::Push};
scheduler.Push(system.GPU(), submit_list->channel, std::move(submit_list->entries));
} else if (std::holds_alternative<GPUTickCommand>(next.data)) {
split.Defer([gpu = &system.GPU()] { gpu->TickWork(); });
} else if (auto* run = std::get_if<RunCommand>(&next.data)) {
split.DeferFunction(std::move(run->func));
} else if (const auto* flush = std::get_if<FlushRegionCommand>(&next.data)) {
split.Defer([rasterizer_ptr, addr = flush->addr, size = flush->size] {
rasterizer_ptr->FlushRegion(addr, size);
});
} else if (const auto* invalidate = std::get_if<InvalidateRegionCommand>(&next.data)) {
split.Defer([rasterizer_ptr, addr = invalidate->addr, size = invalidate->size] {
rasterizer_ptr->OnCacheInvalidation(addr, size);
});
} else {
ASSERT(false);
}
split.Defer([state_ptr = &state, fence = next.fence, block = next.block] {
SignalCommand(*state_ptr, fence, block);
});
split.Publish();
continue;
}
if (auto* submit_list = std::get_if<SubmitListCommand>(&next.data)) {
Profile::Scope scope{Profile::Zone::Push};
scheduler.Push(system.GPU(), submit_list->channel, std::move(submit_list->entries));
} else if (std::holds_alternative<GPUTickCommand>(next.data)) {
Profile::Scope scope{Profile::Zone::Tick};
system.GPU().TickWork();
} else if (auto* run = std::get_if<RunCommand>(&next.data)) {
Profile::Scope scope{Profile::Zone::Run};
run->func();
} else if (const auto* flush = std::get_if<FlushRegionCommand>(&next.data)) {
Profile::Scope scope{Profile::Zone::FlushInvalidate};
renderer.ReadRasterizer()->FlushRegion(flush->addr, flush->size);
} else if (const auto* invalidate = std::get_if<InvalidateRegionCommand>(&next.data)) {
Profile::Scope scope{Profile::Zone::FlushInvalidate};
renderer.ReadRasterizer()->OnCacheInvalidation(invalidate->addr, invalidate->size);
} else {
ASSERT(false);
}
SignalCommand(state, next.fence, next.block);
}
});
}
void ThreadManager::SubmitList(s32 channel, Tegra::CommandList&& entries, bool is_async) {
PushCommand(SubmitListCommand(channel, std::move(entries)), false, is_async);
}
void ThreadManager::FlushRegion(DAddr addr, u64 size, bool is_async) {
if (!is_async) {
// Always flush with synchronous GPU mode
PushCommand(FlushRegionCommand(addr, size), false, is_async);
}
}
void ThreadManager::TickGPU(bool is_async) {
PushCommand(GPUTickCommand(), false, is_async);
}
void ThreadManager::Run(std::function<void()>&& func, bool is_async) {
PushCommand(RunCommand{std::move(func)}, false, is_async);
}
void ThreadManager::InvalidateRegion(DAddr addr, u64 size) {
rasterizer->OnCacheInvalidation(addr, size);
}
void ThreadManager::FlushAndInvalidateRegion(DAddr addr, u64 size, bool is_async) {
if (Settings::IsGPULevelHigh()) {
if (!is_async) {
PushCommand(FlushRegionCommand(addr, size), false, is_async);
} else {
auto& gpu = system.GPU();
const u64 fence = gpu.RequestFlush(addr, size);
TickGPU(is_async);
gpu.WaitForSyncOperation(fence);
}
}
rasterizer->OnCacheInvalidation(addr, size);
}
u64 ThreadManager::PushCommand(CommandData&& command_data, bool block, bool is_async) {
if (!is_async) {
// In synchronous GPU mode, block the caller until the command has executed
block = true;
}
std::unique_lock lk(state.write_lock);
const u64 fence{++state.last_fence};
state.queue.EmplaceWait(std::move(command_data), fence, block);
if (block) {
state.cv.wait(lk, thread.get_stop_token(), [this, fence] {
return fence <= state.signaled_fence.load(std::memory_order_relaxed);
});
}
return fence;
}
void ThreadManager::NotifyShutdown() {
if (thread.joinable()) {
thread.request_stop();
thread.join();
}
VideoCommon::GpuSplit::Get().Stop();
}
} // namespace VideoCommon::GPUThread
|
