raphiaura/KytyPS5
1
0
Fork 0
mirror of https://github.com/KytyPS5/KytyPS5.git synced 2026-10-20 10:56:58 +02:00
Code Issues Packages Projects Releases Wiki Activity

Fix DCC clears after metadata writes

Browse Source
This commit is contained in:
nmzik
2026-09-14 14:32:57 +04:00
parent 968cf3c03f
commit 01df42a5f3
5 changed files with 438 additions and 156 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/graphics/host_gpu/renderer/cache/textureCache.cpp
Vendored
+141 -67
View File
@@ -33,12 +33,23 @@ namespace {
constexpr uint64_t NumFramesBeforeRemoval = 32;
[[nodiscard]] bool IsDccClearCode(uint32_t fill) {
switch (fill) {
case 0x00000000u:
case 0x20202020u:
case 0x40404040u:
case 0x80808080u:
case 0xc0c0c0c0u: return true;
default: return false;
}
}
[[nodiscard]] bool DecodeDccClear(const TextureCache::ImageDesc& desc, vk::Format format,
uint32_t fill, vk::ClearColorValue& clear) {
const auto code = static_cast<uint8_t>(fill);
if (fill != static_cast<uint32_t>(code) * 0x01010101u) {
if (!IsDccClearCode(fill)) {
return false;
}
const auto code = static_cast<uint8_t>(fill);
const auto& metadata = desc.info.metadata;
if (code == 0x20) {
// Clear-to-register is a color-buffer operation; the texture pipe cannot decode it.
@@ -46,9 +57,6 @@ constexpr uint64_t NumFramesBeforeRemoval = 32;
metadata.dcc_clear_register_valid &&
DecodePackedColorClear(format, metadata.dcc_clear_word, clear);
}
if (code != 0x00 && code != 0x40 && code != 0x80 && code != 0xc0) {
return false;
}
clear = {};
if (code == 0x00) {
return true;
@@ -188,6 +196,9 @@ TextureCache::TextureCache(GraphicContext& graphics, CommandScheduler& scheduler
}
TextureCache::~TextureCache() {
for (auto& [address, metadata]: m_surface_metas) {
UntrackMetadata(address, metadata);
}
m_slot_images.ForEach([&](ImageId id, const Image& image) {
if (image.registered) {
UnregisterImage(id);
@@ -323,10 +334,8 @@ void TextureCache::DeleteImage(ImageId id) {
if (image->info.HasMetadata()) {
const auto metadata = m_surface_metas.find(image->info.metadata.range.address);
if (metadata != m_surface_metas.end() &&
((image->info.metadata.kind == ImageMetadataKind::Dcc &&
metadata->second.type == MetaDataInfo::Type::Dcc) ||
(image->info.metadata.kind == ImageMetadataKind::Htile &&
metadata->second.type == MetaDataInfo::Type::HTile))) {
image->info.metadata.kind == ImageMetadataKind::Htile &&
metadata->second.type == MetaDataInfo::Type::HTile) {
// A later binding may have reused this address for another metadata type.
m_surface_metas.erase(metadata);
}
@@ -1124,22 +1133,99 @@ void TextureCache::InitializeImage(ImageId id) {
}
}
void TextureCache::PrepareDccClear(ImageId id, const ImageDesc& desc) {
void TextureCache::UntrackMetadata(uint64_t address, MetaDataInfo& metadata) {
if (metadata.tracked) {
m_page_manager.UpdatePageWatchers<false>(address, metadata.size);
metadata.tracked = false;
}
}
void TextureCache::InvalidateDccMetadata(uint64_t address, uint64_t size) {
for (auto& [base, metadata]: m_surface_metas) {
if (metadata.type != MetaDataInfo::Type::Dcc ||
metadata.size == 0 || base >= address + size || address >= base + metadata.size) {
continue;
}
UntrackMetadata(base, metadata);
metadata.dirty = true;
metadata.clear_mask = UINT32_MAX;
metadata.revision = ++m_metadata_revision;
}
}
void TextureCache::RefreshDccMetadata(const ImageDesc& desc) {
const auto range = desc.info.metadata.range;
if (desc.info.metadata.kind != ImageMetadataKind::Dcc || range.size == 0 ||
desc.info.resources.levels != 1) {
return;
}
// DCC allocations use complete 4 KiB metadata blocks. Besides validating
// the descriptor, this ensures that a CPU page fault identifies the metadata it modifies.
if (!range.Valid() || range.address % TRACKER_PAGE_SIZE != 0 ||
range.size % TRACKER_PAGE_SIZE != 0) {
EXIT("TextureCache: DCC allocation must contain aligned 4 KiB blocks\n");
}
for (;;) {
uint64_t revision = 0;
{
std::scoped_lock lock {m_lock};
auto [entry, inserted] = m_surface_metas.try_emplace(range.address);
auto& metadata = entry->second;
if (inserted || metadata.type != MetaDataInfo::Type::Dcc) {
UntrackMetadata(range.address, metadata);
metadata = {.type = MetaDataInfo::Type::Dcc, .clear_mask = UINT32_MAX};
}
if (metadata.size != range.size) {
UntrackMetadata(range.address, metadata);
metadata.size = range.size;
metadata.dirty = true;
metadata.revision = ++m_metadata_revision;
}
if (!metadata.dirty) {
return;
}
if (!metadata.tracked) {
m_page_manager.UpdatePageWatchers<true>(range.address, range.size);
metadata.tracked = true;
}
revision = metadata.revision;
}
// Discovery precedes descriptor/vertex uploads. Drain metadata writes before taking
// the texture lock; completion can retire buffers and run memory-tracking callbacks.
m_buffer_cache.ReadMemory(range.address, range.size, false);
std::vector<uint8_t> bytes(range.size);
std::scoped_lock lock {m_lock};
const auto entry = m_surface_metas.find(range.address);
if (entry == m_surface_metas.end() || entry->second.revision != revision) {
continue;
}
if (!LibKernel::Memory::TryReadBacking(range.address, bytes.data(), range.size)) {
EXIT("TextureCache: failed to read DCC metadata backing\n");
}
auto& metadata = entry->second;
metadata.fill_value = UINT32_MAX;
if (std::all_of(bytes.begin(), bytes.end(), [&](uint8_t byte) { return byte == bytes[0]; })) {
metadata.fill_value = static_cast<uint32_t>(bytes[0]) * 0x01010101u;
}
metadata.dirty = false;
return;
}
}
void TextureCache::PrepareDccClear(ImageId id, const ImageDesc& desc,
uint32_t metadata_base_layer) {
if (desc.info.metadata.kind != ImageMetadataKind::Dcc) {
return;
}
auto& image = m_slot_images[id];
image.info.metadata = desc.info.metadata;
auto [entry, inserted] = m_surface_metas.try_emplace(
desc.info.metadata.range.address, MetaDataInfo {.type = MetaDataInfo::Type::Dcc});
auto& metadata = entry->second;
if (!inserted && metadata.type == MetaDataInfo::Type::PendingDcc) {
metadata.type = MetaDataInfo::Type::Dcc;
} else if (metadata.type != MetaDataInfo::Type::Dcc) {
EXIT("TextureCache: image reuses non-DCC metadata\n");
const auto entry = m_surface_metas.find(desc.info.metadata.range.address);
if (entry == m_surface_metas.end() || entry->second.type != MetaDataInfo::Type::Dcc) {
return;
}
if (metadata.clear_mask == 0 || image.info.resources.levels != 1 ||
desc.info.metadata.range.size == 0 || metadata.fill_size < desc.info.metadata.range.size) {
auto& metadata = entry->second;
if (metadata.dirty || metadata.clear_mask == 0 || image.info.resources.levels != 1 ||
metadata.size != desc.info.metadata.range.size) {
return;
}
vk::ClearValue clear {};
@@ -1148,12 +1234,14 @@ void TextureCache::PrepareDccClear(ImageId id, const ImageDesc& desc) {
}
const auto& view = desc.view_info;
const bool volume_texture = image.info.IsVolume() && view.type == vk::ImageViewType::e3D;
const auto first = volume_texture ? 0u : view.base_layer;
const auto count = volume_texture ? std::max(image.info.extent.depth >> view.base_level, 1u)
const auto first = volume_texture ? 0u : metadata_base_layer;
const auto image_first = volume_texture ? 0u : view.base_layer;
const auto count = volume_texture ? std::max(desc.info.extent.depth >> view.base_level, 1u)
: view.layer_count;
if (first >= 32 || count > 32 - first) {
return;
}
const auto metadata_slice_size = metadata.size / desc.info.TransferLayers();
// The metadata fill covers the complete allocation. Consume each layer only after its
// native image contents exist; already materialized layers may have been rendered since.
for (uint32_t layer = first; layer < first + count;) {
@@ -1167,9 +1255,16 @@ void TextureCache::PrepareDccClear(ImageId id, const ImageDesc& desc) {
mask |= 1u << layer++;
} while (layer < first + count && (metadata.clear_mask & (1u << layer)) != 0);
ClearImage(m_scheduler.Current(), id,
{vk::ImageAspectFlagBits::eColor, view.base_level, view.level_count, start,
{vk::ImageAspectFlagBits::eColor, view.base_level, view.level_count,
image_first + start - first,
layer - start},
clear);
// The host now owns expanded color data. Publish uncompressed (0xff)
// keys for these slices, so a later partial write cannot reuse stale clear codes.
const auto address = desc.info.metadata.range.address + metadata_slice_size * start;
const auto size = metadata_slice_size * (layer - start);
auto [buffer, offset] = m_buffer_cache.ObtainBuffer(address, size, true, false);
buffer->Fill(offset, size, UINT32_MAX);
metadata.clear_mask &= ~mask;
}
}
@@ -1235,6 +1330,8 @@ ImageId TextureCache::FindImage(ImageDesc& desc, bool exact_format) {
std::scoped_lock lock {m_lock};
return GetNullImage(desc);
}
RefreshDccMetadata(desc);
const auto metadata_base_layer = desc.view_info.base_layer;
ImageId result {};
{
@@ -1299,6 +1396,7 @@ ImageId TextureCache::FindImage(ImageDesc& desc, bool exact_format) {
if (view_layer >= 0) {
desc.view_info.base_layer = static_cast<uint32_t>(view_layer);
}
PrepareDccClear(result, desc, metadata_base_layer);
image.tick_accessed_last = m_scheduler.CurrentTick();
TouchImage(image);
}
@@ -1365,7 +1463,6 @@ vk::ImageView TextureCache::FindTexture(ImageId id, const ImageDesc& desc) {
image.MarkGpuModified();
}
if (!image.info.data.Empty()) {
PrepareDccClear(id, desc);
RefreshImage(id);
}
switch (desc.type) {
@@ -1395,7 +1492,6 @@ vk::ImageView TextureCache::FindRenderTarget(ImageId id, const ImageDesc& desc)
TouchImage(image);
image.MarkGpuModified();
image.usage.render_target = true;
PrepareDccClear(id, desc);
RefreshImage(id);
CommitGpuWrite(image);
TrackImageDownload(id, image);
@@ -1426,6 +1522,7 @@ vk::ImageView TextureCache::FindDepthTarget(ImageId id, const ImageDesc& desc) {
if (!inserted && metadata->second.type != MetaDataInfo::Type::HTile) {
// PS5 allocations can reuse DCC storage as HTile while the old color image is cached.
// The depth binding defines the new type; incompatible fill state cannot carry over.
UntrackMetadata(metadata->first, metadata->second);
metadata->second = {.type = MetaDataInfo::Type::HTile,
.clear_mask = image.info.htile_clear_mask};
}
@@ -1599,6 +1696,7 @@ void TextureCache::InvalidateMemory(uint64_t address, uint64_t size) {
EXIT("TextureCache: invalid memory-invalidation range\n");
}
std::scoped_lock lock {m_lock};
InvalidateDccMetadata(address, size);
InvalidateCpuAliases(address, size);
}
@@ -1807,6 +1905,7 @@ void TextureCache::InvalidateMemoryFromGPU(uint64_t address, uint64_t size) {
return;
}
std::scoped_lock lock {m_lock};
InvalidateDccMetadata(address, size);
for (const auto id: FindImagesInRegion(address, size, true)) {
auto& image = m_slot_images[id];
if (image.depth_id || !image.Overlaps(address, size)) {
@@ -1861,27 +1960,29 @@ void TextureCache::InvalidateCpuAliases(uint64_t address, uint64_t size) {
bool TextureCache::IsMeta(uint64_t address) {
std::scoped_lock lock {m_lock};
const auto found = m_surface_metas.find(address);
return found != m_surface_metas.end() && found->second.type != MetaDataInfo::Type::PendingDcc;
return found != m_surface_metas.end();
}
bool TextureCache::IsMetaCleared(uint64_t address, uint32_t slice, uint32_t* fill_value) {
std::scoped_lock lock {m_lock};
const auto found = m_surface_metas.find(address);
if (found == m_surface_metas.end() || found->second.type == MetaDataInfo::Type::PendingDcc ||
slice >= 32) {
if (found == m_surface_metas.end() || slice >= 32) {
return false;
}
if (fill_value != nullptr) {
*fill_value = found->second.fill_value;
}
if (found->second.type == MetaDataInfo::Type::Dcc &&
(found->second.dirty || !IsDccClearCode(found->second.fill_value))) {
return false;
}
return (found->second.clear_mask & (1u << slice)) != 0;
}
bool TextureCache::ClearMeta(uint64_t address) {
std::scoped_lock lock {m_lock};
const auto found = m_surface_metas.find(address);
if (found == m_surface_metas.end() || found->second.type == MetaDataInfo::Type::PendingDcc ||
found->second.type == MetaDataInfo::Type::Dcc) {
if (found == m_surface_metas.end() || found->second.type == MetaDataInfo::Type::Dcc) {
// Preserve the broad metadata-clear operation for CMask/FMask/HTile. DCC requires a
// validated fill value, so an arbitrary compute write must not clear it.
return false;
@@ -1890,43 +1991,10 @@ bool TextureCache::ClearMeta(uint64_t address) {
return true;
}
void TextureCache::TrackDccFill(uint64_t address, uint64_t size, uint32_t fill_value) {
if (!GuestRange {address, size}.Valid()) {
EXIT("TextureCache: invalid DCC fill range\n");
}
// DCC fills use a repeated byte code. Require all four bytes of the detected dword to agree,
// and mark only recognized deferred-clear encodings as logically clear.
const auto dcc_clear_mask = [fill_value] {
const auto code = static_cast<uint8_t>(fill_value);
if (fill_value != static_cast<uint32_t>(code) * 0x01010101u) {
return 0u;
}
switch (code) {
case 0x00:
case 0x20:
case 0x40:
case 0x80:
case 0xc0: return UINT32_MAX;
default: return 0u;
}
}();
std::scoped_lock lock {m_lock};
// The guest dispatch still writes metadata. An unknown address remains PendingDcc until an
// image descriptor confirms its role; never reinterpret CMask/FMask/HTile as DCC.
const auto found = m_surface_metas.try_emplace(address).first;
if (found->second.type == MetaDataInfo::Type::PendingDcc ||
found->second.type == MetaDataInfo::Type::Dcc) {
found->second.clear_mask = dcc_clear_mask;
found->second.fill_value = fill_value;
found->second.fill_size = size;
}
}
bool TextureCache::TouchMeta(uint64_t address, uint32_t slice, bool is_clear) {
std::scoped_lock lock {m_lock};
const auto found = m_surface_metas.find(address);
if (found == m_surface_metas.end() || found->second.type == MetaDataInfo::Type::PendingDcc ||
slice >= 32) {
if (found == m_surface_metas.end() || slice >= 32) {
return false;
}
if (is_clear) {
@@ -1942,10 +2010,16 @@ void TextureCache::UnmapMemory(uint64_t address, uint64_t size) {
EXIT("TextureCache: invalid unmap range\n");
}
std::scoped_lock lock {m_lock};
const auto end = address + size;
for (auto metadata = m_surface_metas.lower_bound(address);
metadata != m_surface_metas.end() && metadata->first < end;) {
metadata = m_surface_metas.erase(metadata);
for (auto metadata = m_surface_metas.begin(); metadata != m_surface_metas.end();) {
const auto base = metadata->first;
if ((base >= address && base < address + size) ||
(metadata->second.size != 0 && base < address + size &&
address < base + metadata->second.size)) {
UntrackMetadata(base, metadata->second);
metadata = m_surface_metas.erase(metadata);
} else {
++metadata;
}
}
auto images = FindImagesInRegion(address, size, false);
for (const auto id: images) {
src/graphics/host_gpu/renderer/cache/textureCache.h
Vendored
+13 -11
View File
@@ -67,8 +67,6 @@ public:
[[nodiscard]] bool IsMetaCleared(uint64_t address, uint32_t slice,
uint32_t* fill_value = nullptr);
[[nodiscard]] bool ClearMeta(uint64_t address);
// Record deferred DCC state while the original guest dispatch writes the metadata.
void TrackDccFill(uint64_t address, uint64_t size, uint32_t fill_value);
[[nodiscard]] bool TouchMeta(uint64_t address, uint32_t slice, bool is_clear);
void UnmapMemory(uint64_t address, uint64_t size);
@@ -81,17 +79,16 @@ private:
struct ImageDownload;
struct MetaDataInfo {
// A guest metadata-fill dispatch may initialize DCC before its render target is bound.
// PendingDcc retains that exact fill until an image binding classifies the address,
// without exposing an unconfirmed buffer address to the normal metadata heuristics.
// Keep all surface metadata in one entry so CMask/FMask can be
// registered beside HTile and DCC without introducing parallel tracking paths.
enum class Type : uint8_t { PendingDcc, CMask, FMask, HTile, Dcc };
enum class Type : uint8_t { CMask, FMask, HTile, Dcc };
Type type = Type::PendingDcc;
Type type = Type::Dcc;
uint32_t clear_mask = 0;
uint32_t fill_value = 0xffffffffu;
uint64_t fill_size = 0;
// PS5 DCC bytes and layer consumption belong to the metadata allocation, not an image.
uint64_t size = 0;
uint64_t revision = 0;
bool dirty = true;
bool tracked = false;
};
struct OverlapResult {
@@ -149,7 +146,11 @@ private:
ImageId cached);
[[nodiscard]] ImageId ExpandImage(const ImageInfo& info, ImageId source);
void RefreshImage(ImageId id);
void PrepareDccClear(ImageId id, const ImageDesc& desc);
void RefreshDccMetadata(const ImageDesc& desc);
void PrepareDccClear(ImageId id, const ImageDesc& desc,
uint32_t metadata_base_layer);
void UntrackMetadata(uint64_t address, MetaDataInfo& metadata);
void InvalidateDccMetadata(uint64_t address, uint64_t size);
void InitializeImage(ImageId id);
[[nodiscard]] TextureTransfer
BuildTextureTransfer(const Image& image, BindingType binding, TransferDirection direction) const;
@@ -186,6 +187,7 @@ private:
Common::LeastRecentlyUsedCache<ImageId, uint64_t> m_lru_cache;
std::unordered_set<ImageId> m_download_images;
std::map<uint64_t, MetaDataInfo> m_surface_metas;
uint64_t m_metadata_revision = 0;
uint64_t m_total_used_memory = 0;
uint64_t m_trigger_gc_memory = 0;
uint64_t m_pressure_gc_memory = 1536ull * 1024 * 1024;
src/graphics/host_gpu/renderer/pipeline/descriptors.cpp
+1 -1
View File
@@ -147,7 +147,7 @@ NativeStorageBuffer(RenderContext& context, const PreparedBindings::BufferSource
}
buffer_offset = static_cast<uint32_t>(adjustment);
const vk::DescriptorBufferInfo result {buffer->Handle(), aligned_offset, size + adjustment};
if (resource.formatted && resource.written) {
if (resource.written) {
context.GetTextureCache().InvalidateMemoryFromGPU(address, size);
}
const char* access = "Read";
src/graphics/host_gpu/renderer/renderCompute.cpp
-8
View File
@@ -183,14 +183,6 @@ bool RenderExecutor::TryConsumeComputeImageClear(const ShaderComputeInputInfo& i
return false;
}
if (!cache.ClearImageFromBuffer(command, descriptor.Base48(), size, packed_clear)) {
// Track deferred DCC state while the original dispatch writes the metadata allocation.
cache.TrackDccFill(descriptor.Base48(), size, packed_clear);
static std::atomic<uint32_t> logged_metadata_clears {0};
if (logged_metadata_clears.fetch_add(1, std::memory_order_relaxed) < 32) {
LOGF("GraphicsRenderDispatchDirect: metadata fill shader=0x%016" PRIx64
" addr=0x%016" PRIx64 " size=0x%016" PRIx64 " value=0x%08" PRIx32 "\n",
input.stage.program->shader_hash, descriptor.Base48(), size, packed_clear);
}
return false;
}
static std::atomic<uint32_t> logged_clears {0};