xenia-project/xenia
1
0
Fork 0
mirror of https://github.com/xenia-project/xenia.git synced 2025-12-06 07:12:03 +01:00
Code Issues Actions Packages Projects Releases Wiki Activity

[GPU] Occlusion query faking extension

Browse source
This commit is contained in:
Gliniak 2025-05-14 20:53:47 +02:00 committed by Radosław Gliński
parent fb3c12d362
commit b447699a5b
3 changed files with 34 additions and 20 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

8
src/xenia/gpu/gpu_flags.cc
View file

@ -55,10 +55,14 @@ DEFINE_bool(
"when MSAA is used with fullscreen passes.", "when MSAA is used with fullscreen passes.",
"GPU"); "GPU");
DEFINE_int32(query_occlusion_fake_sample_count, 100, DEFINE_int32(query_occlusion_sample_lower_threshold, 80,
"If set to -1 no sample counts are written, games may hang. Else, " "If set to -1 no sample counts are written, games may hang. Else, "
"the sample count of every tile will be incremented on every " "the sample count of every tile will be incremented on every "
"EVENT_WRITE_ZPD by this number. Setting this to 0 means " "EVENT_WRITE_ZPD by this number. Setting this to 0 means "
"everything is reported as occluded.", "everything is reported as occluded.",
"GPU"); "GPU");
UPDATE_from_int32(query_occlusion_fake_sample_count, 2024, 9, 23, 9, 1000); DEFINE_int32(
query_occlusion_sample_upper_threshold, 100,
"Set to higher number than query_occlusion_sample_lower_threshold. This "
"value is ignored if query_occlusion_sample_lower_threshold is set to -1.",
"GPU");

4
src/xenia/gpu/gpu_flags.h
View file

@ -26,7 +26,9 @@ DECLARE_bool(non_seamless_cube_map);
DECLARE_bool(half_pixel_offset); DECLARE_bool(half_pixel_offset);
DECLARE_int32(query_occlusion_fake_sample_count); DECLARE_int32(query_occlusion_sample_lower_threshold);
DECLARE_int32(query_occlusion_sample_upper_threshold);
DECLARE_bool(disassemble_pm4); DECLARE_bool(disassemble_pm4);

42
src/xenia/gpu/pm4_command_processor_implement.h
View file

@ -953,6 +953,9 @@ bool COMMAND_PROCESSOR::ExecutePacketType3_EVENT_WRITE_EXT(
trace_writer_.WriteMemoryWrite(CpuToGpu(address), sizeof(extents)); trace_writer_.WriteMemoryWrite(CpuToGpu(address), sizeof(extents));
return true; return true;
} }
static uint32_t samples = cvars::query_occlusion_sample_upper_threshold;
XE_NOINLINE XE_NOINLINE
bool COMMAND_PROCESSOR::ExecutePacketType3_EVENT_WRITE_ZPD( bool COMMAND_PROCESSOR::ExecutePacketType3_EVENT_WRITE_ZPD(
uint32_t packet, uint32_t count) XE_RESTRICT { uint32_t packet, uint32_t count) XE_RESTRICT {
@ -963,28 +966,33 @@ bool COMMAND_PROCESSOR::ExecutePacketType3_EVENT_WRITE_ZPD(
// Writeback initiator. // Writeback initiator.
COMMAND_PROCESSOR::WriteEventInitiator(initiator & 0x3F); COMMAND_PROCESSOR::WriteEventInitiator(initiator & 0x3F);
if (cvars::query_occlusion_sample_lower_threshold < 0) {
return true;
}
// Occlusion queries: // Occlusion queries:
// This command is send on query begin and end. // This command is send on query begin and end.
// As a workaround report some fixed amount of passed samples. // As a workaround report some fixed amount of passed samples.
auto fake_sample_count = cvars::query_occlusion_fake_sample_count; auto* pSampleCounts = memory_->TranslatePhysical<xe_gpu_depth_sample_counts*>(
if (fake_sample_count >= 0) { register_file_->values[XE_GPU_REG_RB_SAMPLE_COUNT_ADDR]);
auto* pSampleCounts = // 0xFFFFFEED is written to this two locations by D3D only on D3DISSUE_END
memory_->TranslatePhysical<xe_gpu_depth_sample_counts*>( // and used to detect a finished query.
register_file_->values[XE_GPU_REG_RB_SAMPLE_COUNT_ADDR]); bool is_end_via_z_pass = pSampleCounts->ZPass_A == kQueryFinished &&
// 0xFFFFFEED is written to this two locations by D3D only on D3DISSUE_END pSampleCounts->ZPass_B == kQueryFinished;
// and used to detect a finished query. // Older versions of D3D also checks for ZFail (4D5307D5).
bool is_end_via_z_pass = pSampleCounts->ZPass_A == kQueryFinished && bool is_end_via_z_fail = pSampleCounts->ZFail_A == kQueryFinished &&
pSampleCounts->ZPass_B == kQueryFinished; pSampleCounts->ZFail_B == kQueryFinished;
// Older versions of D3D also checks for ZFail (4D5307D5). std::memset(pSampleCounts, 0, sizeof(xe_gpu_depth_sample_counts));
bool is_end_via_z_fail = pSampleCounts->ZFail_A == kQueryFinished && if (is_end_via_z_pass || is_end_via_z_fail) {
pSampleCounts->ZFail_B == kQueryFinished; pSampleCounts->ZPass_A = samples;
std::memset(pSampleCounts, 0, sizeof(xe_gpu_depth_sample_counts)); pSampleCounts->Total_A = samples;
if (is_end_via_z_pass || is_end_via_z_fail) {
pSampleCounts->ZPass_A = fake_sample_count;
pSampleCounts->Total_A = fake_sample_count;
}
} }
samples =
samples <= static_cast<uint32_t>(
cvars::query_occlusion_sample_lower_threshold)
? static_cast<uint32_t>(cvars::query_occlusion_sample_upper_threshold)
: samples - 1;
return true; return true;
} }