gremling/KytyPS5
1
0 Fork 0
mirror of https://github.com/KytyPS5/KytyPS5.git synced 2026-10-27 16:02:37 +00:00
Code Issues Packages Projects Releases Wiki Activity

graphics: count occlusion queries on the host GPU behind --occlusion-queries

Browse Source
This commit is contained in:
nmzik committed 2026-10-23 19:48:53 +02:00
1 parent ba0e93daac
commit 246d511568
18 files changed
+288 -14
Download Patch File Download Diff File Expand all files Collapse all files

No files matched your search

src/graphics/guest_gpu/graphicsRun.cpp
+19 -13
View File
@@ -237,6 +237,8 @@ void CommandProcessor::BufferInit() {
void CommandProcessor::BufferFlush() {
GetScheduler().Flush();
// The occlusion counters reach a guest that polls them without waiting for a later fence.
m_renderer.GetOcclusionQueries().Publish();
}
void CommandProcessor::BufferWait() {
@@ -803,6 +805,7 @@ void CommandProcessor::SetPredication(uint32_t condition, uint32_t op, uint32_t
return;
case 0x01: {
EXIT_NOT_IMPLEMENTED(address == nullptr);
m_renderer.GetOcclusionQueries().Publish();
// One begin/end pair per DB; bit 63 marks each counter ready.
constexpr uint64_t ready_bit = 1ull << 63u;
const auto* results = reinterpret_cast<const volatile uint64_t*>(address);
@@ -1047,6 +1050,8 @@ void CommandProcessor::WriteAtEndOfPipe(uint32_t cache_policy, uint32_t event_wr
EXIT_NOT_IMPLEMENTED(cache_policy != 0x00000000);
EXIT_NOT_IMPLEMENTED(event_write_dest != 0x00000000);
// Dumped occlusion counters are written before any end-of-pipe signal reaches the guest.
m_renderer.GetOcclusionQueries().Publish();
bool with_interrupt = false;
switch (interrupt_selector) {
@@ -1213,6 +1218,7 @@ void CommandProcessor::WriteAtEndOfPipe64(uint32_t cache_policy, uint32_t event_
}
void CommandProcessor::TriggerEopEventAtEndOfPipe(uint32_t interrupt_context_id) {
m_renderer.GetOcclusionQueries().Publish();
(void)CurrentBuffer().Handle();
GetScheduler().DeferPriorityOperation(
[&renderer = m_renderer, event_id = m_interrupt_event_id, interrupt_context_id] {
@@ -1270,15 +1276,19 @@ void CommandProcessor::TriggerEvent(uint32_t event_type, uint32_t event_index,
"\n",
event_index, event_address);
}
if (auto& queries = m_renderer.GetOcclusionQueries(); queries.Enabled()) {
queries.Dump(event_address);
break;
}
static std::once_flag warning_once;
std::call_once(warning_once, [] {
std::printf("Warning: game uses occlusion queries, which are currently treated as "
"always visible; GPU usage may be higher and FPS may be lower.\n");
std::printf("Warning: game uses occlusion queries, which are treated as always "
"visible without --occlusion-queries; GPU usage may be higher and FPS "
"may be lower.\n");
});
// Until host occlusion queries are implemented, publish an always-visible result. The
// PS5 layout contains one interleaved begin/end pair per DB, and bit 63 marks a result
// ready.
// Without host occlusion queries, publish an always-visible result. The PS5 layout
// contains one interleaved begin/end pair per DB, and bit 63 marks a result ready.
constexpr uint64_t ready_bit = 1ull << 63u;
constexpr uint64_t counter_mask = ready_bit - 1u;
auto* results = reinterpret_cast<volatile uint64_t*>(event_address);
@@ -1298,6 +1308,7 @@ void CommandProcessor::Flip() {
if (GraphicsRunDebugDumpEnabled()) {
LOGF("CommandProcessor::Flip()\n");
}
m_renderer.GetOcclusionQueries().Publish();
auto& command = CurrentBuffer();
auto request = Sync::PrepareVideoOutFlip(command, m_flip.handle, m_flip.index, m_flip.flip_mode,
@@ -1309,8 +1320,6 @@ void CommandProcessor::Flip() {
}
void CommandProcessor::Flip(void* dst_gpu_addr, uint32_t value) {
auto& command = CurrentBuffer();
if (GraphicsRunDebugDumpEnabled()) {
LOGF("CommandProcessor::Flip()\n"
"\t dst_gpu_addr = 0x%016" PRIx64 "\n"
@@ -1319,13 +1328,9 @@ void CommandProcessor::Flip(void* dst_gpu_addr, uint32_t value) {
}
EXIT_IF(dst_gpu_addr == nullptr);
m_renderer.GetOcclusionQueries().Publish();
std::memcpy(dst_gpu_addr, &value, sizeof(value));
auto request = Sync::PrepareVideoOutFlip(command, m_flip.handle, m_flip.index, m_flip.flip_mode,
m_flip.flip_arg);
(void)command.Handle();
GetScheduler().DeferPriorityOperation(
[&renderer = m_renderer, request] { renderer.GetVideoOut().CompleteFlip(request); });
GetScheduler().Flush();
Flip();
}
void CommandProcessor::FlipWithInterrupt(uint32_t eop_event_type, uint32_t cache_action,
@@ -1345,6 +1350,7 @@ void CommandProcessor::FlipWithInterrupt(uint32_t eop_event_type, uint32_t cache
EXIT("unknown event type\n");
}
EXIT_IF(dst_gpu_addr == nullptr);
m_renderer.GetOcclusionQueries().Publish();
std::memcpy(dst_gpu_addr, &value, sizeof(value));
auto request = Sync::PrepareVideoOutFlip(command, m_flip.handle, m_flip.index, m_flip.flip_mode,
m_flip.flip_arg);