raphiaura/KytyPS5
1
0
Fork 0
mirror of https://github.com/KytyPS5/KytyPS5.git synced 2026-10-24 21:06:55 +02:00
Code Issues Packages Projects Releases Wiki Activity

graphics: index live cache slots directly

Browse Source
This commit is contained in:
nmzik
2026-10-24 04:26:16 +02:00
parent 170c107779
commit 269e411ef2
5 changed files with 67 additions and 95 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/graphics/host_gpu/renderer/cache/bufferCache.h
Vendored
+1 -4
View File
@@ -95,10 +95,7 @@ public:
private:
friend struct BufferCacheTestAccess;
bool IsBufferInvalid(BufferId id) const {
const auto* buffer = m_slot_buffers.try_get(id);
return buffer == nullptr || buffer->is_deleted;
}
bool IsBufferInvalid(BufferId id) const { return !id || m_slot_buffers[id].is_deleted; }
using BufferMap = std::map<uint64_t, BufferId>;
struct OverlapResult {
src/graphics/host_gpu/renderer/cache/textureCache.cpp
Vendored
+45 -60
View File
@@ -251,24 +251,24 @@ void TextureCache::UnregisterImage(ImageId id) {
}
void TextureCache::DeleteImage(ImageId id) {
auto* image = m_slot_images.try_get(id);
if (image == nullptr || !image->registered) {
auto& image = m_slot_images[id];
if (!image.registered) {
return;
}
if (image->depth_id) {
std::erase(m_slot_images[image->depth_id].stencil_ids, id);
if (image.depth_id) {
std::erase(m_slot_images[image.depth_id].stencil_ids, id);
} else {
// Detach first: deleting each association also unlinks it from its depth owner.
for (const auto association: std::exchange(image->stencil_ids, {})) {
for (const auto association: std::exchange(image.stencil_ids, {})) {
FreeImage(association);
}
}
if (image->IsGpuModified()) {
if (image.IsGpuModified()) {
EXIT("TextureCache: deleting a GPU-modified image without resolving its contents\n");
}
m_download_images.erase(id);
if (image->info.metadata.kind == ImageMetadataKind::Htile) {
m_surface_metas.erase(image->info.metadata.range.address);
if (image.info.metadata.kind == ImageMetadataKind::Htile) {
m_surface_metas.erase(image.info.metadata.range.address);
}
UnregisterImage(id);
if (m_scheduler.Active()) {
@@ -779,15 +779,14 @@ void Resolve(TextureCache& cache, ImageId cached_id, ImageId merged_id) {
TextureCache::OverlapResult TextureCache::ResolveOverlap(const ImageInfo& requested,
BindingType binding, ImageId cached_id,
ImageId merged_id) {
auto owner = m_slot_images.try_get(cached_id);
if (owner == nullptr || !owner->registered) {
auto& cached = m_slot_images[cached_id];
if (!cached.registered) {
return {merged_id};
}
if (BcAlias::Matches(requested, owner->info) && !owner->IsCpuDirty()) {
if (BcAlias::Matches(requested, cached.info) && !cached.IsCpuDirty()) {
BcAlias::Resolve(*this, cached_id, merged_id);
return {merged_id};
}
auto& cached = *owner;
const auto current_tick = m_scheduler.CurrentTick();
const bool safe_to_delete =
current_tick - std::min(current_tick, cached.tick_accessed_last) > NumFramesBeforeRemoval &&
@@ -861,9 +860,8 @@ TextureCache::OverlapResult TextureCache::ResolveOverlap(const ImageInfo& reques
if (requested.pixel_format != cached.info.pixel_format ||
requested.data.size <= cached.info.data.size) {
const auto result_id = merged_id ? merged_id : cached_id;
const auto result = m_slot_images.try_get(result_id);
return {result != nullptr && ImageViewOps::FormatsCompatible(result->info.pixel_format,
requested.pixel_format)
return {ImageViewOps::FormatsCompatible(m_slot_images[result_id].info.pixel_format,
requested.pixel_format)
? result_id
: ImageId {}};
}
@@ -1210,9 +1208,8 @@ ImageId TextureCache::AssociateStencil(ImageId depth_id, GuestRange stencil) {
ImageId association {};
for (const auto id: FindImagesInRegion(stencil.address, stencil.size, false)) {
const auto owner = m_slot_images.try_get(id);
if (owner != nullptr && owner->info.data == stencil &&
owner->info.extent == depth.info.extent) {
const auto& owner = m_slot_images[id];
if (owner.info.data == stencil && owner.info.extent == depth.info.extent) {
association = id;
}
}
@@ -1307,10 +1304,9 @@ ImageId TextureCache::FindImage(ImageDesc& desc, bool exact_format) {
}
} else if (block_alias) {
for (const auto candidate: candidates) {
const auto* owner = m_slot_images.try_get(candidate);
if (candidate != result && owner != nullptr && owner->registered &&
!owner->IsCpuDirty() &&
BcAlias::Matches(m_slot_images[result].info, owner->info)) {
const auto& owner = m_slot_images[candidate];
if (candidate != result && owner.registered && !owner.IsCpuDirty() &&
BcAlias::Matches(m_slot_images[result].info, owner.info)) {
BcAlias::Resolve(*this, candidate, result);
}
}
@@ -1373,14 +1369,14 @@ ImageId TextureCache::FindImageFromRange(uint64_t address, uint64_t size, bool e
std::scoped_lock lock {m_lock};
ImageIds matches;
for (const auto id: FindImagesInRegion(address, size, false)) {
auto owner = m_slot_images.try_get(id);
if (owner == nullptr || owner->info.data.address != address) {
const auto* owner = &m_slot_images[id];
if (owner->info.data.address != address) {
continue;
}
if (ensure_valid && owner->depth_id) {
owner = m_slot_images.try_get(owner->depth_id);
owner = &m_slot_images[owner->depth_id];
}
if (owner == nullptr || (ensure_valid && !owner->SafeToDownload())) {
if (ensure_valid && !owner->SafeToDownload()) {
continue;
}
matches.push_back(id);
@@ -1397,11 +1393,8 @@ ImageId TextureCache::FindImageFromRange(uint64_t address, uint64_t size, bool e
}
}
}
if (selected && ensure_valid) {
const auto owner = m_slot_images.try_get(selected);
if (owner != nullptr && owner->depth_id) {
selected = owner->depth_id;
}
if (selected && ensure_valid && m_slot_images[selected].depth_id) {
selected = m_slot_images[selected].depth_id;
}
return selected;
}
@@ -1422,9 +1415,8 @@ vk::ImageView TextureCache::FindTexture(ImageId id, const ImageDesc& desc) {
desc.info.data.End() <= image.info.stencil.End()) {
for (const auto stencil_id:
FindImagesInRegion(image.info.stencil.address, image.info.stencil.size, false)) {
const auto* stencil = m_slot_images.try_get(stencil_id);
if (stencil != nullptr && stencil->depth_id == id &&
stencil->info.data == image.info.stencil) {
const auto& stencil = m_slot_images[stencil_id];
if (stencil.depth_id == id && stencil.info.data == image.info.stencil) {
RefreshImage(stencil_id);
break;
}
@@ -1535,27 +1527,24 @@ bool TextureCache::ClearImageFromBuffer(CommandBuffer& command, uint64_t address
vk::ImageAspectFlags aspect {};
ImageSubresourceRange subresources;
for (const auto id: FindImagesInRegion(address, size, false)) {
auto owner = m_slot_images.try_get(id);
if (owner == nullptr) {
continue;
}
const auto& owner = m_slot_images[id];
vk::ImageAspectFlags candidate {};
ImageId candidate_id = id;
ImageSubresourceRange candidate_subresources {
0, owner->info.resources.levels, 0, owner->info.TransferLayers()};
if (owner->depth_id && owner->info.data.address == address &&
owner->info.data.size == size) {
0, owner.info.resources.levels, 0, owner.info.TransferLayers()};
if (owner.depth_id && owner.info.data.address == address &&
owner.info.data.size == size) {
candidate = vk::ImageAspectFlagBits::eStencil;
candidate_id = owner->depth_id;
candidate_subresources = owner->stencil_subresources;
owner = m_slot_images.try_get(candidate_id);
if (owner == nullptr || owner->backing.image == nullptr || !owner->info.HasStencil()) {
candidate_id = owner.depth_id;
candidate_subresources = owner.stencil_subresources;
const auto& depth = m_slot_images[candidate_id];
if (depth.backing.image == nullptr || !depth.info.HasStencil()) {
continue;
}
} else if (!owner->depth_id && owner->info.data.address == address &&
owner->info.data.size == size) {
candidate = owner->info.IsDepth() ? vk::ImageAspectFlagBits::eDepth
: vk::ImageAspectFlagBits::eColor;
} else if (!owner.depth_id && owner.info.data.address == address &&
owner.info.data.size == size) {
candidate = owner.info.IsDepth() ? vk::ImageAspectFlagBits::eDepth
: vk::ImageAspectFlagBits::eColor;
}
if (!candidate) {
continue;
@@ -1896,23 +1885,20 @@ void TextureCache::InvalidateCpuAliases(uint64_t address, uint64_t size) {
const auto page_begin = Common::AlignDown(address, TRACKER_PAGE_SIZE);
const auto page_end = Common::AlignUp(address + size, TRACKER_PAGE_SIZE);
for (const auto id: FindImagesInRegion(address, size, true)) {
auto owner = m_slot_images.try_get(id);
if (owner == nullptr) {
continue;
}
if (owner->Overlaps(address, size)) {
owner->MarkCpuDirty();
auto& owner = m_slot_images[id];
if (owner.Overlaps(address, size)) {
owner.MarkCpuDirty();
UntrackImage(id);
continue;
}
const auto image_begin = owner->info.data.address;
const auto image_end = owner->info.data.End();
const auto image_begin = owner.info.data.address;
const auto image_end = owner.info.data.End();
if (page_end < image_end) {
UntrackImageHead(id);
} else if (image_begin < page_begin) {
UntrackImageTail(id);
} else {
MarkAsMaybeDirty(id, *owner);
MarkAsMaybeDirty(id, owner);
}
}
}
@@ -2052,8 +2038,7 @@ void TextureCache::ProcessDownloadImages() {
}
std::scoped_lock lock {m_lock};
for (const auto id: m_download_images) {
const auto owner = m_slot_images.try_get(id);
if (owner != nullptr && owner->registered && owner->IsGpuModified()) {
if (m_slot_images[id].IsGpuModified()) {
(void)DownloadImageMemory(id);
}
}
src/graphics/host_gpu/renderer/pipeline/descriptors.cpp
+11 -20
View File
@@ -672,9 +672,7 @@ void RenderExecutor::BindRenderTarget(ImageId id) {
void RenderExecutor::ResetBindings() {
for (const auto id: m_bound_images) {
if (auto* image = m_context.GetTextureCache().m_slot_images.try_get(id); image != nullptr) {
image->binding = {};
}
m_context.GetTextureCache().m_slot_images[id].binding = {};
}
m_bound_images.clear();
}
@@ -823,12 +821,9 @@ void RenderExecutor::RebindImages(PreparedBindings& prepared) {
// Acquire each view before a later overlapping descriptor can replace its image.
for (uint32_t i = 0; i < program.info.images.size(); i++) {
if (!images[i].image_id) continue;
const auto old_image = texture_cache.m_slot_images.try_get(images[i].image_id);
if (old_image == nullptr || !old_image->registered ||
old_image->binding.needs_rebind) {
if (old_image != nullptr) {
old_image->binding = {};
}
auto& old_image = texture_cache.m_slot_images[images[i].image_id];
if (!old_image.registered || old_image.binding.needs_rebind) {
old_image.binding = {};
images[i] = ResolveTexture(program.info.images[i], snapshot.images[i]);
BindImage(images[i].image_id,
images[i].desc.type == TextureCache::BindingType::Storage);
@@ -879,12 +874,9 @@ void RenderExecutor::PrepareGraphicsBindings(std::span<PreparedBindings* const>
auto& cache = m_context.GetTextureCache();
for (auto& target: colors) {
EXIT_IF(!target.image_id);
const auto old_image = cache.m_slot_images.try_get(target.image_id);
if (old_image == nullptr || !old_image->registered ||
old_image->binding.needs_rebind) {
if (old_image != nullptr) {
old_image->binding = {};
}
auto& old_image = cache.m_slot_images[target.image_id];
if (!old_image.registered || old_image.binding.needs_rebind) {
old_image.binding = {};
target.desc.view_info.base_level = target.guest_mip_level;
target.desc.view_info.base_layer = target.guest_array_layer;
target.image_id = cache.FindImage(target.desc);
@@ -944,11 +936,10 @@ void RenderExecutor::CommitBindings(CommandBuffer& buffer,
}
for (const auto [address, size]: reads) {
for (const auto id: m_bound_images) {
const auto* image = m_context.GetTextureCache().m_slot_images.try_get(id);
if (image == nullptr ||
(!image->binding.shader_write && !image->binding.is_target)) continue;
for (const auto written: {image->info.data, image->info.stencil,
image->info.metadata.range}) {
const auto& image = m_context.GetTextureCache().m_slot_images[id];
if (!image.binding.shader_write && !image.binding.is_target) continue;
for (const auto written: {image.info.data, image.info.stencil,
image.info.metadata.range}) {
if (written.size != 0 && ImageRangeOverlaps(address, size,
written.address, written.size)) {
EXIT("scalar resource reads overlap an image or attachment write\n");
src/graphics/host_gpu/renderer/renderDraw.cpp
+4 -5
View File
@@ -474,9 +474,8 @@ RenderState RenderExecutor::AcquireRenderTargets(CommandBuffer& buffer, RenderCo
for (uint32_t i = 0; i < color_count; i++) {
auto& target = colors[i];
EXIT_IF(!target.image_id);
const auto owner = cache.m_slot_images.try_get(target.image_id);
if (owner == nullptr || (!owner->registered && !owner->info.data.Empty()) ||
owner->binding.needs_rebind) {
const auto& owner = cache.m_slot_images[target.image_id];
if ((!owner.registered && !owner.info.data.Empty()) || owner.binding.needs_rebind) {
EXIT("color target changed after render-state discovery\n");
}
const auto image_view = cache.FindRenderTarget(target.image_id, target.desc);
@@ -502,8 +501,8 @@ RenderState RenderExecutor::AcquireRenderTargets(CommandBuffer& buffer, RenderCo
attachment.image_layout = layout;
}
if (depth.image_id) {
const auto owner = cache.m_slot_images.try_get(depth.image_id);
if (owner == nullptr || !owner->registered || owner->binding.needs_rebind) {
const auto& owner = cache.m_slot_images[depth.image_id];
if (!owner.registered || owner.binding.needs_rebind) {
EXIT("depth target changed after render-state discovery\n");
}
const auto image_view = cache.FindDepthTarget(depth.image_id, depth.desc);
tests/ShaderRecompilerComputeTests.cpp
+6 -6
View File
@@ -4317,12 +4317,6 @@ public:
BufferCacheTestAccess::PageOwner(
cache, index_begin + 2 * index_page) == index_bridge,
"bridging acquisition did not publish its merged Buffer range");
scheduler.Finish();
constexpr uint64_t recycled_offset = index_offset + 4 * index_page;
const auto recycled = cache.FindBuffer(base + recycled_offset + 0x100,
sizeof(uint32_t));
Require(name, "slot reuse", recycled == index_left || recycled == index_right,
"retired Buffer slot was not reused");
const auto [resolved_left, resolved_left_offset] = cache.ObtainBuffer(
index_begin + 0x100, sizeof(uint32_t), true, false, index_left);
const auto [resolved_right, resolved_right_offset] =
@@ -4333,6 +4327,12 @@ public:
resolved_right == resolved_left &&
resolved_left_offset != resolved_right_offset,
"saved descriptor IDs did not resolve through the merged owner");
scheduler.Finish();
constexpr uint64_t recycled_offset = index_offset + 4 * index_page;
const auto recycled = cache.FindBuffer(base + recycled_offset + 0x100,
sizeof(uint32_t));
Require(name, "slot reuse", recycled == index_left || recycled == index_right,
"retired Buffer slot was not reused");
constexpr uint64_t stream_page = BufferCache::CACHING_PAGESIZE;
constexpr uint64_t stream_leap = stream_page * 128;