[GPU] Implement ZPD occlusion queries
Implements EVENT_WRITE_ZPD handling across shared CommandProcessor, D3D12, and Vulkan backends. This adds a shared report lifecycle. Reports can span submissions/render passes, split into multiple segments, and asynchronously retire. Reports can also survive same-slot reuse, which is critical for titles that aggressively recycle against a small pool. RTV/FBO use their respective host occlusion queries, whereas ROV/FSI use counter buffers instead, with shader helpers that are called from the translated PS, accumulating surviving MSAA sample coverage into active slots. Regardless of approach, results are copied to small, dedicated readbacks and fed through shared retirement. Fast mode writes speculative values and patches in resolved results later. Strict mode waits for real results with a small retire backstop so it doesn't spin forever when a report gets weird. Fake mode's fundamentals are unchanged, it just now uses XenosZPDReport memory helpers like all the modes do. Also adds optional sample count saturation for instances when attenuation isn't a one-for-one with real hardware, an optional alternate fast behavior that preserves cached zeroes (that can improve flare-heavy titles but might regress occlusion culling), and normalization for upscaling. QueryBatch is still unsupported other than a fake path that's enabled when a range is set via cvar.
This commit is contained in:
bomabomabomaboma
committed by
Radosław Gliński
Radosław Gliński
parent
29311ddcdd
commit
fbd620c22b
@@ -954,65 +954,95 @@ bool COMMAND_PROCESSOR::ExecutePacketType3_EVENT_WRITE_EXT(
|
||||
return true;
|
||||
}
|
||||
|
||||
static uint32_t samples = cvars::query_occlusion_sample_upper_threshold;
|
||||
static uint32_t batched_samples = 0;
|
||||
|
||||
XE_NOINLINE
|
||||
bool COMMAND_PROCESSOR::ExecutePacketType3_EVENT_WRITE_ZPD(
|
||||
uint32_t packet, uint32_t count) XE_RESTRICT {
|
||||
// Set by D3D as BE but struct ABI is LE
|
||||
const uint32_t kQueryFinished = xe::byte_swap(0xFFFFFEED);
|
||||
assert_true(count == 1);
|
||||
uint32_t initiator = reader_.ReadAndSwap<uint32_t>();
|
||||
// Writeback initiator.
|
||||
COMMAND_PROCESSOR::WriteEventInitiator(initiator & 0x3F);
|
||||
|
||||
if (cvars::query_occlusion_sample_lower_threshold < 0) {
|
||||
uint32_t report_address =
|
||||
register_file_->values[XE_GPU_REG_RB_SAMPLE_COUNT_ADDR];
|
||||
uint32_t report_record_base = XenosZPDReport::GetRecordBase(report_address);
|
||||
bool is_begin_record = XenosZPDReport::IsBeginRecord(report_address);
|
||||
bool is_end_record = XenosZPDReport::IsEndRecord(report_address);
|
||||
|
||||
xe_gpu_depth_sample_counts* report =
|
||||
report_record_base
|
||||
? memory_->TranslatePhysical<xe_gpu_depth_sample_counts*>(
|
||||
report_record_base)
|
||||
: nullptr;
|
||||
|
||||
// True if the record has the pending D3D sentinel.
|
||||
// Useful as a hint, but not authoritative for report boundaries.
|
||||
// QueryBatch titles can have multiple pending sentinels in a row and don't
|
||||
// necessarily update in an order we currently observe.
|
||||
bool guest_marks_end = report && XenosZPDReport::HasPendingSentinel(report);
|
||||
bool logical_active = zpd_active_segment_.logical_active;
|
||||
|
||||
// QueryBatch fake fallback, which ignores record layout and just returns an
|
||||
// incrementing sample count on each event.
|
||||
if (cvars::occlusion_query_querybatch_range > 0) {
|
||||
uint32_t sample_count =
|
||||
XenosZPDReport::QueryBatchFakeSamples(querybatch_zpd_sample_count_);
|
||||
if (report) {
|
||||
// Both QueryBatch and conventional fake samples skip elective saturation.
|
||||
XenosZPDReport::WriteSampleCount(report, sample_count, false);
|
||||
}
|
||||
return true;
|
||||
}
|
||||
// Occlusion queries:
|
||||
// This command is send on query begin and end.
|
||||
// As a workaround report some fixed amount of passed samples.
|
||||
auto* pSampleCounts = memory_->TranslatePhysical<xe_gpu_depth_sample_counts*>(
|
||||
register_file_->values[XE_GPU_REG_RB_SAMPLE_COUNT_ADDR]);
|
||||
// 0xFFFFFEED is written to this two locations by D3D only on D3DISSUE_END
|
||||
// and used to detect a finished query.
|
||||
bool is_end_via_z_pass = pSampleCounts->ZPass_A == kQueryFinished ||
|
||||
pSampleCounts->ZPass_B == kQueryFinished;
|
||||
// Older versions of D3D also checks for ZFail (4D5307D5).
|
||||
bool is_end_via_z_fail = pSampleCounts->ZFail_A == kQueryFinished ||
|
||||
pSampleCounts->ZFail_B == kQueryFinished;
|
||||
std::memset(pSampleCounts, 0, sizeof(xe_gpu_depth_sample_counts));
|
||||
|
||||
// Titles that use QueryBatch (4D5309B1, 4E4D0801) don't use an END result for
|
||||
// every query. Each ISSUE snapshots the sample count into a new slot and
|
||||
// recovers the total by looking at differences between slots.
|
||||
// END detection isn't sufficient here.
|
||||
if (cvars::query_occlusion_querybatch_range > 0) {
|
||||
// Mimic batched behavior by reporting a running count and writing it on
|
||||
// every event.
|
||||
const uint32_t base =
|
||||
static_cast<uint32_t>(cvars::query_occlusion_sample_lower_threshold);
|
||||
const uint32_t range =
|
||||
static_cast<uint32_t>(cvars::query_occlusion_querybatch_range);
|
||||
if (batched_samples < base || batched_samples - base + 1 >= range) {
|
||||
batched_samples = base;
|
||||
} else {
|
||||
++batched_samples;
|
||||
if (GetZPDMode() != ZPDMode::kFake && !zpd_force_fake_fallback_) {
|
||||
if (logical_active && is_end_record) {
|
||||
COMMAND_PROCESSOR::EndZPDReport(report_address, false);
|
||||
return true;
|
||||
}
|
||||
pSampleCounts->ZPass_A = batched_samples;
|
||||
pSampleCounts->Total_A = batched_samples;
|
||||
} else if (is_end_via_z_pass || is_end_via_z_fail) {
|
||||
pSampleCounts->ZPass_A = samples;
|
||||
pSampleCounts->Total_A = samples;
|
||||
if (is_begin_record) {
|
||||
// Clear the record so the game knows the BEGIN was processed and
|
||||
// stale sentinel data from a prior query lifetime doesn't persist.
|
||||
if (report) {
|
||||
std::memset(report, 0, sizeof(xe_gpu_depth_sample_counts));
|
||||
}
|
||||
COMMAND_PROCESSOR::BeginZPDReport(report_address);
|
||||
return true;
|
||||
}
|
||||
if (!logical_active && is_end_record) {
|
||||
// No logical report is active for this slot, so this is likely an
|
||||
// orphaned END. In fast mode, replay the last cached delta so polling
|
||||
// code does not sit on the sentinel forever.
|
||||
if (GetZPDMode() == ZPDMode::kFast) {
|
||||
uint32_t cached_delta = 1;
|
||||
auto cache_it = fast_zpd_report_cached_values_.find(report_record_base);
|
||||
if (cache_it != fast_zpd_report_cached_values_.end()) {
|
||||
cached_delta = cache_it->second;
|
||||
}
|
||||
COMMAND_PROCESSOR::WriteZPDReport(0, report_record_base, 0,
|
||||
cached_delta, false);
|
||||
} else {
|
||||
// In strict mode, just pump in case a previous report has resolved.
|
||||
COMMAND_PROCESSOR::PumpQueryResolves();
|
||||
}
|
||||
return true;
|
||||
}
|
||||
// Address is neither BEGIN nor END (non-standard layout). Fall through
|
||||
// to the fake path so the guest at least gets a result written rather
|
||||
// than leaving the sentinel in place forever.
|
||||
}
|
||||
|
||||
samples =
|
||||
samples <= static_cast<uint32_t>(
|
||||
cvars::query_occlusion_sample_lower_threshold)
|
||||
? static_cast<uint32_t>(cvars::query_occlusion_sample_upper_threshold)
|
||||
: samples - 1;
|
||||
// Conventional fake fallback, which only touches records marked as pending.
|
||||
if (cvars::occlusion_query_fake_lower_threshold < 0 || !report_record_base ||
|
||||
!guest_marks_end) {
|
||||
return true;
|
||||
}
|
||||
|
||||
fake_zpd_sample_count_ =
|
||||
(fake_zpd_sample_count_ <=
|
||||
static_cast<uint32_t>(cvars::occlusion_query_fake_lower_threshold))
|
||||
? static_cast<uint32_t>(cvars::occlusion_query_fake_upper_threshold)
|
||||
: fake_zpd_sample_count_ - 1;
|
||||
|
||||
XenosZPDReport::WriteSampleCount(report, fake_zpd_sample_count_, false);
|
||||
return true;
|
||||
}
|
||||
|
||||
|
||||
Reference in New Issue
Block a user