Reaper: Tune it up to be an smart GC.

This commit is contained in:
Fernando Sahmkow 2021-06-13 15:47:54 +02:00
parent a11bc4a382
commit d8ad6aa187
5 changed files with 130 additions and 13 deletions

View File

@ -65,6 +65,9 @@ class BufferCache {
static constexpr BufferId NULL_BUFFER_ID{0}; static constexpr BufferId NULL_BUFFER_ID{0};
static constexpr u64 expected_memory = 512ULL * 1024ULL * 1024ULL;
static constexpr u64 critical_memory = 1024ULL * 1024ULL * 1024ULL;
using Maxwell = Tegra::Engines::Maxwell3D::Regs; using Maxwell = Tegra::Engines::Maxwell3D::Regs;
using Runtime = typename P::Runtime; using Runtime = typename P::Runtime;
@ -327,6 +330,7 @@ private:
typename SlotVector<Buffer>::Iterator deletion_iterator; typename SlotVector<Buffer>::Iterator deletion_iterator;
u64 frame_tick = 0; u64 frame_tick = 0;
u64 total_used_memory = 0;
std::array<BufferId, ((1ULL << 39) >> PAGE_BITS)> page_table; std::array<BufferId, ((1ULL << 39) >> PAGE_BITS)> page_table;
}; };
@ -346,6 +350,10 @@ BufferCache<P>::BufferCache(VideoCore::RasterizerInterface& rasterizer_,
template <class P> template <class P>
void BufferCache<P>::TickFrame() { void BufferCache<P>::TickFrame() {
SCOPE_EXIT({
++frame_tick;
delayed_destruction_ring.Tick();
});
// Calculate hits and shots and move hit bits to the right // Calculate hits and shots and move hit bits to the right
const u32 hits = std::reduce(uniform_cache_hits.begin(), uniform_cache_hits.end()); const u32 hits = std::reduce(uniform_cache_hits.begin(), uniform_cache_hits.end());
const u32 shots = std::reduce(uniform_cache_shots.begin(), uniform_cache_shots.end()); const u32 shots = std::reduce(uniform_cache_shots.begin(), uniform_cache_shots.end());
@ -359,8 +367,13 @@ void BufferCache<P>::TickFrame() {
const bool skip_preferred = hits * 256 < shots * 251; const bool skip_preferred = hits * 256 < shots * 251;
uniform_buffer_skip_cache_size = skip_preferred ? DEFAULT_SKIP_CACHE_SIZE : 0; uniform_buffer_skip_cache_size = skip_preferred ? DEFAULT_SKIP_CACHE_SIZE : 0;
static constexpr u64 ticks_to_destroy = 120; const bool activate_gc = total_used_memory >= expected_memory;
int num_iterations = 32; if (!activate_gc) {
return;
}
const bool agressive_gc = total_used_memory >= critical_memory;
const u64 ticks_to_destroy = agressive_gc ? 60 : 120;
int num_iterations = agressive_gc ? 64 : 32;
for (; num_iterations > 0; --num_iterations) { for (; num_iterations > 0; --num_iterations) {
if (deletion_iterator == slot_buffers.end()) { if (deletion_iterator == slot_buffers.end()) {
deletion_iterator = slot_buffers.begin(); deletion_iterator = slot_buffers.begin();
@ -375,8 +388,6 @@ void BufferCache<P>::TickFrame() {
DeleteBuffer(buffer_id); DeleteBuffer(buffer_id);
} }
} }
delayed_destruction_ring.Tick();
++frame_tick;
} }
template <class P> template <class P>
@ -1115,8 +1126,14 @@ template <class P>
template <bool insert> template <bool insert>
void BufferCache<P>::ChangeRegister(BufferId buffer_id) { void BufferCache<P>::ChangeRegister(BufferId buffer_id) {
const Buffer& buffer = slot_buffers[buffer_id]; const Buffer& buffer = slot_buffers[buffer_id];
const auto size = buffer.SizeBytes();
if (insert) {
total_used_memory += Common::AlignUp(size, 1024);
} else {
total_used_memory -= Common::AlignUp(size, 1024);
}
const VAddr cpu_addr_begin = buffer.CpuAddr(); const VAddr cpu_addr_begin = buffer.CpuAddr();
const VAddr cpu_addr_end = cpu_addr_begin + buffer.SizeBytes(); const VAddr cpu_addr_end = cpu_addr_begin + size;
const u64 page_begin = cpu_addr_begin / PAGE_SIZE; const u64 page_begin = cpu_addr_begin / PAGE_SIZE;
const u64 page_end = Common::DivCeil(cpu_addr_end, PAGE_SIZE); const u64 page_end = Common::DivCeil(cpu_addr_end, PAGE_SIZE);
for (u64 page = page_begin; page != page_end; ++page) { for (u64 page = page_begin; page != page_end; ++page) {

View File

@ -130,6 +130,26 @@ bool ImageBase::IsSafeDownload() const noexcept {
return true; return true;
} }
void ImageBase::CheckBadOverlapState() {
if (False(flags & ImageFlagBits::BadOverlap)) {
return;
}
if (!overlapping_images.empty()) {
return;
}
flags &= ~ImageFlagBits::BadOverlap;
}
void ImageBase::CheckAliasState() {
if (False(flags & ImageFlagBits::Alias)) {
return;
}
if (!aliased_images.empty()) {
return;
}
flags &= ~ImageFlagBits::Alias;
}
void AddImageAlias(ImageBase& lhs, ImageBase& rhs, ImageId lhs_id, ImageId rhs_id) { void AddImageAlias(ImageBase& lhs, ImageBase& rhs, ImageId lhs_id, ImageId rhs_id) {
static constexpr auto OPTIONS = RelaxedOptions::Size | RelaxedOptions::Format; static constexpr auto OPTIONS = RelaxedOptions::Size | RelaxedOptions::Format;
ASSERT(lhs.info.type == rhs.info.type); ASSERT(lhs.info.type == rhs.info.type);

View File

@ -25,6 +25,12 @@ enum class ImageFlagBits : u32 {
Strong = 1 << 5, ///< Exists in the image table, the dimensions are can be trusted Strong = 1 << 5, ///< Exists in the image table, the dimensions are can be trusted
Registered = 1 << 6, ///< True when the image is registered Registered = 1 << 6, ///< True when the image is registered
Picked = 1 << 7, ///< Temporary flag to mark the image as picked Picked = 1 << 7, ///< Temporary flag to mark the image as picked
// Garbage Collection Flags
BadOverlap = 1 << 8, ///< This image overlaps other but doesn't fit, has higher
///< garbage collection priority
Alias = 1 << 9, ///< This image has aliases and has priority on garbage
///< collection
}; };
DECLARE_ENUM_FLAG_OPERATORS(ImageFlagBits) DECLARE_ENUM_FLAG_OPERATORS(ImageFlagBits)
@ -51,6 +57,9 @@ struct ImageBase {
return cpu_addr < overlap_end && overlap_cpu_addr < cpu_addr_end; return cpu_addr < overlap_end && overlap_cpu_addr < cpu_addr_end;
} }
void CheckBadOverlapState();
void CheckAliasState();
ImageInfo info; ImageInfo info;
u32 guest_size_bytes = 0; u32 guest_size_bytes = 0;
@ -74,6 +83,7 @@ struct ImageBase {
std::vector<SubresourceBase> slice_subresources; std::vector<SubresourceBase> slice_subresources;
std::vector<AliasedImage> aliased_images; std::vector<AliasedImage> aliased_images;
std::vector<ImageId> overlapping_images;
}; };
struct ImageAllocBase { struct ImageAllocBase {

View File

@ -75,6 +75,9 @@ class TextureCache {
/// Sampler ID for bugged sampler ids /// Sampler ID for bugged sampler ids
static constexpr SamplerId NULL_SAMPLER_ID{0}; static constexpr SamplerId NULL_SAMPLER_ID{0};
static constexpr u64 expected_memory = 1024ULL * 1024ULL * 1024ULL;
static constexpr u64 critical_memory = 2 * 1024ULL * 1024ULL * 1024ULL;
using Runtime = typename P::Runtime; using Runtime = typename P::Runtime;
using Image = typename P::Image; using Image = typename P::Image;
using ImageAlloc = typename P::ImageAlloc; using ImageAlloc = typename P::ImageAlloc;
@ -333,6 +336,7 @@ private:
std::unordered_map<u64, std::vector<ImageId>, IdentityHash<u64>> page_table; std::unordered_map<u64, std::vector<ImageId>, IdentityHash<u64>> page_table;
bool has_deleted_images = false; bool has_deleted_images = false;
u64 total_used_memory = 0;
SlotVector<Image> slot_images; SlotVector<Image> slot_images;
SlotVector<ImageView> slot_image_views; SlotVector<ImageView> slot_image_views;
@ -380,8 +384,10 @@ TextureCache<P>::TextureCache(Runtime& runtime_, VideoCore::RasterizerInterface&
template <class P> template <class P>
void TextureCache<P>::TickFrame() { void TextureCache<P>::TickFrame() {
static constexpr u64 ticks_to_destroy = 120; const bool high_priority_mode = total_used_memory >= expected_memory;
int num_iterations = 32; const bool aggressive_mode = total_used_memory >= critical_memory;
const u64 ticks_to_destroy = high_priority_mode ? 60 : 100;
int num_iterations = aggressive_mode ? 256 : (high_priority_mode ? 128 : 64);
for (; num_iterations > 0; --num_iterations) { for (; num_iterations > 0; --num_iterations) {
if (deletion_iterator == slot_images.end()) { if (deletion_iterator == slot_images.end()) {
deletion_iterator = slot_images.begin(); deletion_iterator = slot_images.begin();
@ -390,11 +396,42 @@ void TextureCache<P>::TickFrame() {
} }
} }
const auto [image_id, image] = *deletion_iterator; const auto [image_id, image] = *deletion_iterator;
if (image->frame_tick + ticks_to_destroy < frame_tick) { const bool is_alias = True(image->flags & ImageFlagBits::Alias);
if (image->IsSafeDownload() && if (is_alias && image->aliased_images.size() <= 1) {
std::ranges::none_of(image->aliased_images, [&](const AliasedImage& alias) { ++deletion_iterator;
return slot_images[alias.id].modification_tick > image->modification_tick; continue;
})) { }
const bool is_bad_overlap = True(image->flags & ImageFlagBits::BadOverlap);
const bool must_download = image->IsSafeDownload();
const u64 ticks_needed = is_bad_overlap ? ticks_to_destroy >> 4 : ticks_to_destroy;
const bool should_care =
aggressive_mode || is_bad_overlap || is_alias || (high_priority_mode && !must_download);
if (should_care && image->frame_tick + ticks_needed < frame_tick) {
if (is_bad_overlap) {
const bool overlap_check =
std::ranges::all_of(image->overlapping_images, [&](const ImageId& overlap_id) {
auto& overlap = slot_images[overlap_id];
return (overlap.frame_tick >= image->frame_tick) &&
(overlap.modification_tick > image->modification_tick);
});
if (!overlap_check) {
++deletion_iterator;
continue;
}
}
if (!is_bad_overlap && must_download) {
if (is_alias) {
const bool alias_check =
std::ranges::all_of(image->aliased_images, [&](const AliasedImage& alias) {
auto& alias_image = slot_images[alias.id];
return (alias_image.frame_tick >= image->frame_tick) &&
(alias_image.modification_tick > image->modification_tick);
});
if (!alias_check) {
++deletion_iterator;
continue;
}
}
auto map = runtime.DownloadStagingBuffer(image->unswizzled_size_bytes); auto map = runtime.DownloadStagingBuffer(image->unswizzled_size_bytes);
const auto copies = FullDownloadCopies(image->info); const auto copies = FullDownloadCopies(image->info);
image->DownloadMemory(map, copies); image->DownloadMemory(map, copies);
@ -406,10 +443,12 @@ void TextureCache<P>::TickFrame() {
} }
UnregisterImage(image_id); UnregisterImage(image_id);
DeleteImage(image_id); DeleteImage(image_id);
if (is_bad_overlap) {
num_iterations++;
}
} }
++deletion_iterator; ++deletion_iterator;
} }
// Tick sentenced resources in this order to ensure they are destroyed in the right order
sentenced_images.Tick(); sentenced_images.Tick();
sentenced_framebuffers.Tick(); sentenced_framebuffers.Tick();
sentenced_image_view.Tick(); sentenced_image_view.Tick();
@ -989,6 +1028,7 @@ ImageId TextureCache<P>::JoinImages(const ImageInfo& info, GPUVAddr gpu_addr, VA
std::vector<ImageId> overlap_ids; std::vector<ImageId> overlap_ids;
std::vector<ImageId> left_aliased_ids; std::vector<ImageId> left_aliased_ids;
std::vector<ImageId> right_aliased_ids; std::vector<ImageId> right_aliased_ids;
std::vector<ImageId> bad_overlap_ids;
ForEachImageInRegion(cpu_addr, size_bytes, [&](ImageId overlap_id, ImageBase& overlap) { ForEachImageInRegion(cpu_addr, size_bytes, [&](ImageId overlap_id, ImageBase& overlap) {
if (info.type != overlap.info.type) { if (info.type != overlap.info.type) {
return; return;
@ -1014,9 +1054,14 @@ ImageId TextureCache<P>::JoinImages(const ImageInfo& info, GPUVAddr gpu_addr, VA
const ImageBase new_image_base(new_info, gpu_addr, cpu_addr); const ImageBase new_image_base(new_info, gpu_addr, cpu_addr);
if (IsSubresource(new_info, overlap, gpu_addr, options, broken_views, native_bgr)) { if (IsSubresource(new_info, overlap, gpu_addr, options, broken_views, native_bgr)) {
left_aliased_ids.push_back(overlap_id); left_aliased_ids.push_back(overlap_id);
overlap.flags |= ImageFlagBits::Alias;
} else if (IsSubresource(overlap.info, new_image_base, overlap.gpu_addr, options, } else if (IsSubresource(overlap.info, new_image_base, overlap.gpu_addr, options,
broken_views, native_bgr)) { broken_views, native_bgr)) {
right_aliased_ids.push_back(overlap_id); right_aliased_ids.push_back(overlap_id);
overlap.flags |= ImageFlagBits::Alias;
} else {
bad_overlap_ids.push_back(overlap_id);
overlap.flags |= ImageFlagBits::BadOverlap;
} }
}); });
const ImageId new_image_id = slot_images.insert(runtime, new_info, gpu_addr, cpu_addr); const ImageId new_image_id = slot_images.insert(runtime, new_info, gpu_addr, cpu_addr);
@ -1044,10 +1089,18 @@ ImageId TextureCache<P>::JoinImages(const ImageInfo& info, GPUVAddr gpu_addr, VA
for (const ImageId aliased_id : right_aliased_ids) { for (const ImageId aliased_id : right_aliased_ids) {
ImageBase& aliased = slot_images[aliased_id]; ImageBase& aliased = slot_images[aliased_id];
AddImageAlias(new_image_base, aliased, new_image_id, aliased_id); AddImageAlias(new_image_base, aliased, new_image_id, aliased_id);
new_image.flags |= ImageFlagBits::Alias;
} }
for (const ImageId aliased_id : left_aliased_ids) { for (const ImageId aliased_id : left_aliased_ids) {
ImageBase& aliased = slot_images[aliased_id]; ImageBase& aliased = slot_images[aliased_id];
AddImageAlias(aliased, new_image_base, aliased_id, new_image_id); AddImageAlias(aliased, new_image_base, aliased_id, new_image_id);
new_image.flags |= ImageFlagBits::Alias;
}
for (const ImageId aliased_id : bad_overlap_ids) {
ImageBase& aliased = slot_images[aliased_id];
aliased.overlapping_images.push_back(new_image_id);
new_image.overlapping_images.push_back(aliased_id);
new_image.flags |= ImageFlagBits::BadOverlap;
} }
RegisterImage(new_image_id); RegisterImage(new_image_id);
return new_image_id; return new_image_id;
@ -1217,6 +1270,8 @@ void TextureCache<P>::RegisterImage(ImageId image_id) {
image.flags |= ImageFlagBits::Registered; image.flags |= ImageFlagBits::Registered;
ForEachPage(image.cpu_addr, image.guest_size_bytes, ForEachPage(image.cpu_addr, image.guest_size_bytes,
[this, image_id](u64 page) { page_table[page].push_back(image_id); }); [this, image_id](u64 page) { page_table[page].push_back(image_id); });
total_used_memory +=
Common::AlignUp(std::max(image.guest_size_bytes, image.unswizzled_size_bytes), 1024);
} }
template <class P> template <class P>
@ -1225,6 +1280,9 @@ void TextureCache<P>::UnregisterImage(ImageId image_id) {
ASSERT_MSG(True(image.flags & ImageFlagBits::Registered), ASSERT_MSG(True(image.flags & ImageFlagBits::Registered),
"Trying to unregister an already registered image"); "Trying to unregister an already registered image");
image.flags &= ~ImageFlagBits::Registered; image.flags &= ~ImageFlagBits::Registered;
image.flags &= ~ImageFlagBits::BadOverlap;
total_used_memory -=
Common::AlignUp(std::max(image.guest_size_bytes, image.unswizzled_size_bytes), 1024);
ForEachPage(image.cpu_addr, image.guest_size_bytes, [this, image_id](u64 page) { ForEachPage(image.cpu_addr, image.guest_size_bytes, [this, image_id](u64 page) {
const auto page_it = page_table.find(page); const auto page_it = page_table.find(page);
if (page_it == page_table.end()) { if (page_it == page_table.end()) {
@ -1298,9 +1356,19 @@ void TextureCache<P>::DeleteImage(ImageId image_id) {
std::erase_if(other_image.aliased_images, [image_id](const AliasedImage& other_alias) { std::erase_if(other_image.aliased_images, [image_id](const AliasedImage& other_alias) {
return other_alias.id == image_id; return other_alias.id == image_id;
}); });
other_image.CheckAliasState();
ASSERT_MSG(num_removed_aliases == 1, "Invalid number of removed aliases: {}", ASSERT_MSG(num_removed_aliases == 1, "Invalid number of removed aliases: {}",
num_removed_aliases); num_removed_aliases);
} }
for (const ImageId overlap_id : image.overlapping_images) {
ImageBase& other_image = slot_images[overlap_id];
[[maybe_unused]] const size_t num_removed_overlaps = std::erase_if(
other_image.overlapping_images,
[image_id](const ImageId other_overlap_id) { return other_overlap_id == image_id; });
other_image.CheckBadOverlapState();
ASSERT_MSG(num_removed_overlaps == 1, "Invalid number of removed overlapps: {}",
num_removed_overlaps);
}
for (const ImageViewId image_view_id : image_view_ids) { for (const ImageViewId image_view_id : image_view_ids) {
sentenced_image_view.Push(std::move(slot_image_views[image_view_id])); sentenced_image_view.Push(std::move(slot_image_views[image_view_id]));
slot_image_views.erase(image_view_id); slot_image_views.erase(image_view_id);

View File

@ -581,6 +581,8 @@ void SwizzleBlockLinearImage(Tegra::MemoryManager& gpu_memory, GPUVAddr gpu_addr
for (s32 layer = 0; layer < info.resources.layers; ++layer) { for (s32 layer = 0; layer < info.resources.layers; ++layer) {
const std::span<const u8> src = input.subspan(host_offset); const std::span<const u8> src = input.subspan(host_offset);
gpu_memory.ReadBlockUnsafe(gpu_addr + guest_offset, dst.data(), dst.size_bytes());
SwizzleTexture(dst, src, bytes_per_block, num_tiles.width, num_tiles.height, SwizzleTexture(dst, src, bytes_per_block, num_tiles.width, num_tiles.height,
num_tiles.depth, block.height, block.depth); num_tiles.depth, block.height, block.depth);