raphiaura/KytyPS5
1
0
Fork 0
mirror of https://github.com/KytyPS5/KytyPS5.git synced 2026-10-24 12:56:53 +02:00
Code Issues Packages Projects Releases Wiki Activity

texture cache: select color-metadata clears on the GPU with conditional rendering

Browse Source
This commit is contained in:
Mr. Big Britches
2026-10-22 11:50:35 +02:00
committed by nmzik
parent b6495a5b67
commit b54057942d
12 changed files with 311 additions and 20 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tests/ShaderRecompilerComputeTests.cpp
+72 -12
View File
@@ -10266,6 +10266,32 @@ public:
ReadCachedTexel(name, context, color.image_id, {0, 0, 31}) ==
std::vector<u32>{0x5a5a5a5a},
"clearing one volume slice changed untouched slice contents");
// A GPU writer clears slice 9's metadata; its bind selects the clear on the GPU.
auto [writer, writer_offset] = resources.GetBufferCache().ObtainBuffer(
dcc_address + 9 * metadata_slice_size, metadata_slice_size, true);
writer->Fill(writer_offset, metadata_slice_size, 0);
Common::DebugCounters::SetEnabled(true);
const auto readbacks =
Common::DebugCounters::Get(Common::DebugCounters::Counter::ColorMetadataReadbacks);
const auto checks =
Common::DebugCounters::Get(Common::DebugCounters::Counter::ColorMetadataChecks);
RenderExecutorTestAccess::ResetBindings(executor);
registers.SetColorView(0, {.base_array_slice_index = 9, .last_array_slice_index = 9});
RenderColorInfo gpu_slice{};
RenderExecutorTestAccess::ResolveRenderColorTarget(
executor, scheduler.Current(), gpu_slice, 0);
Common::DebugCounters::SetEnabled(false);
Require(name, "GPU-written volume metadata clears on the GPU",
gpu_slice.image_id == color.image_id &&
Common::DebugCounters::Get(
Common::DebugCounters::Counter::ColorMetadataReadbacks) == readbacks &&
Common::DebugCounters::Get(
Common::DebugCounters::Counter::ColorMetadataChecks) == checks + 1 &&
ReadCachedTexel(name, context, color.image_id, {0, 0, 9}) ==
std::vector<u32>{0} &&
ReadCachedTexel(name, context, color.image_id, {0, 0, 31}) ==
std::vector<u32>{0x5a5a5a5a},
"a volume slice's GPU-written metadata was read back or cleared other slices");
auto storage_desc = color.desc;
storage_desc.type = BindingType::Storage;
storage_desc.info.metadata = {};
@@ -12271,14 +12297,43 @@ public:
const auto readbacks =
Common::DebugCounters::Get(Common::DebugCounters::Counter::ColorMetadataReadbacks);
bind();
Common::DebugCounters::SetEnabled(false);
Require(name, "metadata readback counter",
Common::DebugCounters::Get(
Common::DebugCounters::Counter::ColorMetadataReadbacks) == readbacks + 1,
"binding GPU-written metadata did not count exactly one readback");
Require(name, "same-value raw native re-clear",
read_texel() == expected,
"a repeated metadata fill did not clear an already drawn target");
const auto checks =
Common::DebugCounters::Get(Common::DebugCounters::Counter::ColorMetadataChecks);
bind();
Require(name, "checked metadata skips rebinding",
Common::DebugCounters::Get(
Common::DebugCounters::Counter::ColorMetadataChecks) == checks,
"rebinding unwritten GPU-checked metadata repeated its clear check");
check_expanded_metadata();
paint();
fill_metadata(metadata_words - 1, true);
bind();
Require(name, "partial raw native metadata fill", retains_painted_texel(),
"a nonuniform GPU-written metadata slice cleared the whole target");
fill_metadata(metadata_words, true);
bind();
Require(name, "complete raw native fill after partial fill", read_texel() == expected,
"completing a GPU-written metadata fill did not restore its clear");
if (layout == Prospero::ChannelLayout::k16_16_16_16 && !fill_case.cmask &&
!fill_case.reuse_unorm) {
auto video = color.desc;
video.type = BindingType::VideoOut;
paint();
fill_metadata(metadata_words, true);
const bool cleared =
texture_cache.FindImage(video) == color.image_id && read_texel() == expected;
std::vector<uint32_t> keys(metadata_words);
context.GetBufferCache().ReadMemory(dcc_address, metadata_size, false);
Require(name, "video-out clear keeps its metadata keys",
cleared &&
LibKernel::Memory::TryReadBacking(dcc_address, keys.data(), metadata_size) &&
std::ranges::all_of(keys,
[&](uint32_t word) { return word == fill_case.fill; }),
"a video-out bind skipped its clear or expanded the flip's metadata keys");
}
paint();
WriteMetadata(context, dcc_address, metadata_size, UINT32_MAX);
fill_metadata(metadata_words - 1);
@@ -12287,15 +12342,15 @@ public:
retains_painted_texel(),
"a nonuniform metadata allocation cleared the whole target");
fill_metadata(metadata_words);
if (fill_case.goat) {
Require(name, "repeated Goat fill stays CPU-owned",
!resources.GetBufferCache().IsRegionGpuModified(dcc_address, metadata_size),
"re-clearing expanded metadata required another GPU readback");
}
bind();
Require(name, "complete fill after partial fill",
read_texel() == expected,
"completing the metadata overwrite did not restore its clear");
Common::DebugCounters::SetEnabled(false);
Require(name, "GPU-written metadata stays on the GPU",
Common::DebugCounters::Get(
Common::DebugCounters::Counter::ColorMetadataReadbacks) == readbacks,
"binding GPU-written color metadata read it back");
// Consuming a clear key updates metadata backing without invalidating
// a pooled image whose first texel precedes the metadata range.
@@ -20434,6 +20489,7 @@ private:
m_runtime_context.queue = m_queue;
m_runtime_context.shader_image_int64_atomics_enabled = true;
m_runtime_context.maintenance8_enabled = true;
m_runtime_context.conditional_rendering_enabled = true;
m_runtime_context.attachment_feedback_loop_enabled = m_rasterization_supported;
m_runtime_context.provoking_vertex_last_enabled = m_rasterization_supported;
const vk::PhysicalDeviceImageFormatInfo2 block_texel_view_info{
@@ -20717,7 +20773,10 @@ private:
vk::PhysicalDeviceRobustness2FeaturesEXT robustness2{};
robustness2.nullDescriptor = true;
robustness2.pNext = &maintenance8;
device_info.pNext = &robustness2;
vk::PhysicalDeviceConditionalRenderingFeaturesEXT conditional_rendering{};
conditional_rendering.conditionalRendering = true;
conditional_rendering.pNext = &robustness2;
device_info.pNext = &conditional_rendering;
vk::PhysicalDeviceFeatures device_features{};
device_features.shaderStorageImageWriteWithoutFormat = true;
device_features.shaderImageGatherExtended = true;
@@ -20736,7 +20795,8 @@ private:
VK_EXT_SHADER_IMAGE_ATOMIC_INT64_EXTENSION_NAME,
VK_EXT_ROBUSTNESS_2_EXTENSION_NAME,
VK_KHR_MAINTENANCE_8_EXTENSION_NAME,
VK_EXT_IMAGE_VIEW_MIN_LOD_EXTENSION_NAME};
VK_EXT_IMAGE_VIEW_MIN_LOD_EXTENSION_NAME,
VK_EXT_CONDITIONAL_RENDERING_EXTENSION_NAME};
if (m_rasterization_supported) {
device_extensions.push_back(VK_EXT_DEPTH_CLIP_ENABLE_EXTENSION_NAME);
device_extensions.push_back(VK_EXT_DEPTH_CLIP_CONTROL_EXTENSION_NAME);