CCF_100/yuzu-mirror
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

VideoCore: Rework Garbage Collection.

Browse Source
This commit is contained in:
Fernando Sahmkow
2021-08-15 15:35:53 +02:00
parent c20ea89390
commit d540d284b5
6 changed files with 213 additions and 101 deletions
Download Patch File Download Diff File Expand all files Collapse all files

13
src/video_core/buffer_cache/buffer_base.h
View File

@@ -261,16 +261,6 @@ public:
stream_score += score;
}
/// Sets the new frame tick
void SetFrameTick(u64 new_frame_tick) noexcept {
frame_tick = new_frame_tick;
}
/// Returns the new frame tick
[[nodiscard]] u64 FrameTick() const noexcept {
return frame_tick;
}
/// Returns the likeliness of this being a stream buffer
[[nodiscard]] int StreamScore() const noexcept {
return stream_score;
@@ -307,6 +297,8 @@ public:
return words.size_bytes;
}
size_t lru_id;
private:
template <Type type>
u64* Array() noexcept {
@@ -603,7 +595,6 @@ private:
RasterizerInterface* rasterizer = nullptr;
VAddr cpu_addr = 0;
Words words;
u64 frame_tick = 0;
BufferFlagBits flags{};
int stream_score = 0;
};

61
src/video_core/buffer_cache/buffer_cache.h
View File

@@ -20,6 +20,7 @@
#include "common/common_types.h"
#include "common/div_ceil.h"
#include "common/literals.h"
#include "common/lru_cache.h"
#include "common/microprofile.h"
#include "common/scope_exit.h"
#include "common/settings.h"
@@ -77,7 +78,7 @@ class BufferCache {
static constexpr BufferId NULL_BUFFER_ID{0};
static constexpr u64 EXPECTED_MEMORY = 512_MiB;
static constexpr u64 EXPECTED_MEMORY = 256_MiB;
static constexpr u64 CRITICAL_MEMORY = 1_GiB;
using Maxwell = Tegra::Engines::Maxwell3D::Regs;
@@ -330,7 +331,7 @@ private:
template <bool insert>
void ChangeRegister(BufferId buffer_id);
void TouchBuffer(Buffer& buffer) const noexcept;
void TouchBuffer(Buffer& buffer, BufferId buffer_id) noexcept;
bool SynchronizeBuffer(Buffer& buffer, VAddr cpu_addr, u32 size);
@@ -428,7 +429,11 @@ private:
size_t immediate_buffer_capacity = 0;
std::unique_ptr<u8[]> immediate_buffer_alloc;
typename SlotVector<Buffer>::Iterator deletion_iterator;
struct LRUItemParams {
using ObjectType = BufferId;
using TickType = u64;
};
Common::LeastRecentlyUsedCache<LRUItemParams> lru_cache;
u64 frame_tick = 0;
u64 total_used_memory = 0;
@@ -445,7 +450,6 @@ BufferCache<P>::BufferCache(VideoCore::RasterizerInterface& rasterizer_,
kepler_compute{kepler_compute_}, gpu_memory{gpu_memory_}, cpu_memory{cpu_memory_} {
// Ensure the first slot is used for the null buffer
void(slot_buffers.insert(runtime, NullBufferParams{}));
deletion_iterator = slot_buffers.end();
common_ranges.clear();
}
@@ -454,20 +458,17 @@ void BufferCache<P>::RunGarbageCollector() {
const bool aggressive_gc = total_used_memory >= CRITICAL_MEMORY;
const u64 ticks_to_destroy = aggressive_gc ? 60 : 120;
int num_iterations = aggressive_gc ? 64 : 32;
for (; num_iterations > 0; --num_iterations) {
if (deletion_iterator == slot_buffers.end()) {
deletion_iterator = slot_buffers.begin();
const auto clean_up = [this, &num_iterations](BufferId buffer_id) {
if (num_iterations == 0) {
return true;
}
++deletion_iterator;
if (deletion_iterator == slot_buffers.end()) {
break;
}
const auto [buffer_id, buffer] = *deletion_iterator;
if (buffer->FrameTick() + ticks_to_destroy < frame_tick) {
DownloadBufferMemory(*buffer);
DeleteBuffer(buffer_id);
}
}
--num_iterations;
auto& buffer = slot_buffers[buffer_id];
DownloadBufferMemory(buffer);
DeleteBuffer(buffer_id);
return false;
};
lru_cache.ForEachItemBelow(frame_tick - ticks_to_destroy, clean_up);
}
template <class P>
@@ -954,7 +955,7 @@ bool BufferCache<P>::IsRegionCpuModified(VAddr addr, size_t size) {
template <class P>
void BufferCache<P>::BindHostIndexBuffer() {
Buffer& buffer = slot_buffers[index_buffer.buffer_id];
TouchBuffer(buffer);
TouchBuffer(buffer, index_buffer.buffer_id);
const u32 offset = buffer.Offset(index_buffer.cpu_addr);
const u32 size = index_buffer.size;
SynchronizeBuffer(buffer, index_buffer.cpu_addr, size);
@@ -975,7 +976,7 @@ void BufferCache<P>::BindHostVertexBuffers() {
for (u32 index = 0; index < NUM_VERTEX_BUFFERS; ++index) {
const Binding& binding = vertex_buffers[index];
Buffer& buffer = slot_buffers[binding.buffer_id];
TouchBuffer(buffer);
TouchBuffer(buffer, binding.buffer_id);
SynchronizeBuffer(buffer, binding.cpu_addr, binding.size);
if (!flags[Dirty::VertexBuffer0 + index]) {
continue;
@@ -1011,7 +1012,7 @@ void BufferCache<P>::BindHostGraphicsUniformBuffer(size_t stage, u32 index, u32
const VAddr cpu_addr = binding.cpu_addr;
const u32 size = std::min(binding.size, (*uniform_buffer_sizes)[stage][index]);
Buffer& buffer = slot_buffers[binding.buffer_id];
TouchBuffer(buffer);
TouchBuffer(buffer, binding.buffer_id);
const bool use_fast_buffer = binding.buffer_id != NULL_BUFFER_ID &&
size <= uniform_buffer_skip_cache_size &&
!buffer.IsRegionGpuModified(cpu_addr, size);
@@ -1083,7 +1084,7 @@ void BufferCache<P>::BindHostGraphicsStorageBuffers(size_t stage) {
ForEachEnabledBit(enabled_storage_buffers[stage], [&](u32 index) {
const Binding& binding = storage_buffers[stage][index];
Buffer& buffer = slot_buffers[binding.buffer_id];
TouchBuffer(buffer);
TouchBuffer(buffer, binding.buffer_id);
const u32 size = binding.size;
SynchronizeBuffer(buffer, binding.cpu_addr, size);
@@ -1128,7 +1129,7 @@ void BufferCache<P>::BindHostTransformFeedbackBuffers() {
for (u32 index = 0; index < NUM_TRANSFORM_FEEDBACK_BUFFERS; ++index) {
const Binding& binding = transform_feedback_buffers[index];
Buffer& buffer = slot_buffers[binding.buffer_id];
TouchBuffer(buffer);
TouchBuffer(buffer, binding.buffer_id);
const u32 size = binding.size;
SynchronizeBuffer(buffer, binding.cpu_addr, size);
@@ -1148,7 +1149,7 @@ void BufferCache<P>::BindHostComputeUniformBuffers() {
ForEachEnabledBit(enabled_compute_uniform_buffer_mask, [&](u32 index) {
const Binding& binding = compute_uniform_buffers[index];
Buffer& buffer = slot_buffers[binding.buffer_id];
TouchBuffer(buffer);
TouchBuffer(buffer, binding.buffer_id);
const u32 size = std::min(binding.size, (*compute_uniform_buffer_sizes)[index]);
SynchronizeBuffer(buffer, binding.cpu_addr, size);
@@ -1168,7 +1169,7 @@ void BufferCache<P>::BindHostComputeStorageBuffers() {
ForEachEnabledBit(enabled_compute_storage_buffers, [&](u32 index) {
const Binding& binding = compute_storage_buffers[index];
Buffer& buffer = slot_buffers[binding.buffer_id];
TouchBuffer(buffer);
TouchBuffer(buffer, binding.buffer_id);
const u32 size = binding.size;
SynchronizeBuffer(buffer, binding.cpu_addr, size);
@@ -1513,11 +1514,11 @@ BufferId BufferCache<P>::CreateBuffer(VAddr cpu_addr, u32 wanted_size) {
const OverlapResult overlap = ResolveOverlaps(cpu_addr, wanted_size);
const u32 size = static_cast<u32>(overlap.end - overlap.begin);
const BufferId new_buffer_id = slot_buffers.insert(runtime, rasterizer, overlap.begin, size);
TouchBuffer(slot_buffers[new_buffer_id]);
for (const BufferId overlap_id : overlap.ids) {
JoinOverlap(new_buffer_id, overlap_id, !overlap.has_stream_leap);
}
Register(new_buffer_id);
TouchBuffer(slot_buffers[new_buffer_id], new_buffer_id);
return new_buffer_id;
}
@@ -1534,12 +1535,14 @@ void BufferCache<P>::Unregister(BufferId buffer_id) {
template <class P>
template <bool insert>
void BufferCache<P>::ChangeRegister(BufferId buffer_id) {
const Buffer& buffer = slot_buffers[buffer_id];
Buffer& buffer = slot_buffers[buffer_id];
const auto size = buffer.SizeBytes();
if (insert) {
total_used_memory += Common::AlignUp(size, 1024);
buffer.lru_id = lru_cache.Insert(buffer_id, frame_tick);
} else {
total_used_memory -= Common::AlignUp(size, 1024);
lru_cache.Free(buffer.lru_id);
}
const VAddr cpu_addr_begin = buffer.CpuAddr();
const VAddr cpu_addr_end = cpu_addr_begin + size;
@@ -1555,8 +1558,10 @@ void BufferCache<P>::ChangeRegister(BufferId buffer_id) {
}
template <class P>
void BufferCache<P>::TouchBuffer(Buffer& buffer) const noexcept {
buffer.SetFrameTick(frame_tick);
void BufferCache<P>::TouchBuffer(Buffer& buffer, BufferId buffer_id) noexcept {
if (buffer_id != NULL_BUFFER_ID) {
lru_cache.Touch(buffer.lru_id, frame_tick);
}
}
template <class P>

2
src/video_core/texture_cache/image_base.h
View File

@@ -80,7 +80,7 @@ struct ImageBase {
VAddr cpu_addr_end = 0;
u64 modification_tick = 0;
u64 frame_tick = 0;
size_t lru_index = ~0;
std::array<u32, MAX_MIP_LEVELS> mip_level_offsets{};

89
src/video_core/texture_cache/texture_cache.h
View File

@@ -43,8 +43,6 @@ TextureCache<P>::TextureCache(Runtime& runtime_, VideoCore::RasterizerInterface&
void(slot_image_views.insert(runtime, NullImageParams{}));
void(slot_samplers.insert(runtime, sampler_descriptor));
deletion_iterator = slot_images.begin();
if constexpr (HAS_DEVICE_MEMORY_INFO) {
const auto device_memory = runtime.GetDeviceLocalMemory();
const u64 possible_expected_memory = (device_memory * 3) / 10;
@@ -64,65 +62,33 @@ template <class P>
void TextureCache<P>::RunGarbageCollector() {
const bool high_priority_mode = total_used_memory >= expected_memory;
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) {
if (deletion_iterator == slot_images.end()) {
deletion_iterator = slot_images.begin();
if (deletion_iterator == slot_images.end()) {
break;
}
const u64 ticks_to_destroy = aggressive_mode ? 10ULL : high_priority_mode ? 50ULL : 100ULL;
size_t num_iterations = aggressive_mode ? 10000 : (high_priority_mode ? 50 : 5);
const auto clean_up = [this, &num_iterations, high_priority_mode](ImageId image_id) {
if (num_iterations == 0) {
return true;
}
auto [image_id, image_tmp] = *deletion_iterator;
Image* image = image_tmp; // fix clang error.
const bool is_alias = True(image->flags & ImageFlagBits::Alias);
const bool is_bad_overlap = True(image->flags & ImageFlagBits::BadOverlap);
const bool must_download = image->IsSafeDownload();
bool should_care = is_bad_overlap || is_alias || (high_priority_mode && !must_download);
const u64 ticks_needed =
is_bad_overlap
? ticks_to_destroy >> 4
: ((should_care && aggressive_mode) ? ticks_to_destroy >> 1 : ticks_to_destroy);
should_care |= aggressive_mode;
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, [&, image](const ImageId& overlap_id) {
auto& overlap = slot_images[overlap_id];
return overlap.frame_tick >= image->frame_tick;
});
if (!overlap_check) {
++deletion_iterator;
continue;
}
}
if (!is_bad_overlap && must_download) {
const bool alias_check = std::ranges::none_of(
image->aliased_images, [&, image](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) {
auto map = runtime.DownloadStagingBuffer(image->unswizzled_size_bytes);
const auto copies = FullDownloadCopies(image->info);
image->DownloadMemory(map, copies);
runtime.Finish();
SwizzleImage(gpu_memory, image->gpu_addr, image->info, copies, map.mapped_span);
}
}
if (True(image->flags & ImageFlagBits::Tracked)) {
UntrackImage(*image, image_id);
}
UnregisterImage(image_id);
DeleteImage(image_id);
if (is_bad_overlap) {
++num_iterations;
}
--num_iterations;
auto& image = slot_images[image_id];
const bool must_download = image.IsSafeDownload();
if (!high_priority_mode && must_download) {
return false;
}
++deletion_iterator;
}
if (must_download) {
auto map = runtime.DownloadStagingBuffer(image.unswizzled_size_bytes);
const auto copies = FullDownloadCopies(image.info);
image.DownloadMemory(map, copies);
runtime.Finish();
SwizzleImage(gpu_memory, image.gpu_addr, image.info, copies, map.mapped_span);
}
if (True(image.flags & ImageFlagBits::Tracked)) {
UntrackImage(image, image_id);
}
UnregisterImage(image_id);
DeleteImage(image_id);
return false;
};
lru_cache.ForEachItemBelow(frame_tick - ticks_to_destroy, clean_up);
}
template <class P>
@@ -1078,6 +1044,8 @@ void TextureCache<P>::RegisterImage(ImageId image_id) {
tentative_size = EstimatedDecompressedSize(tentative_size, image.info.format);
}
total_used_memory += Common::AlignUp(tentative_size, 1024);
image.lru_index = lru_cache.Insert(image_id, frame_tick);
ForEachGPUPage(image.gpu_addr, image.guest_size_bytes,
[this, image_id](u64 page) { gpu_page_table[page].push_back(image_id); });
if (False(image.flags & ImageFlagBits::Sparse)) {
@@ -1115,6 +1083,7 @@ void TextureCache<P>::UnregisterImage(ImageId image_id) {
tentative_size = EstimatedDecompressedSize(tentative_size, image.info.format);
}
total_used_memory -= Common::AlignUp(tentative_size, 1024);
lru_cache.Free(image.lru_index);
const auto& clear_page_table =
[this, image_id](
u64 page,
@@ -1384,7 +1353,7 @@ void TextureCache<P>::PrepareImage(ImageId image_id, bool is_modification, bool
if (is_modification) {
MarkModification(image);
}
image.frame_tick = frame_tick;
lru_cache.Touch(image.lru_index, frame_tick);
}
template <class P>

8
src/video_core/texture_cache/texture_cache_base.h
View File

@@ -14,6 +14,7 @@
#include "common/common_types.h"
#include "common/literals.h"
#include "common/lru_cache.h"
#include "video_core/compatible_formats.h"
#include "video_core/delayed_destruction_ring.h"
#include "video_core/engines/fermi_2d.h"
@@ -370,6 +371,12 @@ private:
std::vector<ImageId> uncommitted_downloads;
std::queue<std::vector<ImageId>> committed_downloads;
struct LRUItemParams {
using ObjectType = ImageId;
using TickType = u64;
};
Common::LeastRecentlyUsedCache<LRUItemParams> lru_cache;
static constexpr size_t TICKS_TO_DESTROY = 6;
DelayedDestructionRing<Image, TICKS_TO_DESTROY> sentenced_images;
DelayedDestructionRing<ImageView, TICKS_TO_DESTROY> sentenced_image_view;
@@ -379,7 +386,6 @@ private:
u64 modification_tick = 0;
u64 frame_tick = 0;
typename SlotVector<Image>::Iterator deletion_iterator;
};
} // namespace VideoCommon