Garbage Collection: Redesign the algorithm to do a better use of memory.

This commit is contained in:
Fernando Sahmkow 2022-01-16 04:43:06 +01:00
parent 8f2e5f5666
commit ecb3342145
13 changed files with 156 additions and 32 deletions

View file

@ -484,6 +484,15 @@ TextureCacheRuntime::TextureCacheRuntime(const Device& device_, ProgramManager&
rescale_read_fbos[i].Create();
}
}
device_access_memory = []() -> u64 {
if (GLAD_GL_NVX_gpu_memory_info) {
GLint cur_avail_mem_kb = 0;
glGetIntegerv(GL_GPU_MEMORY_INFO_TOTAL_AVAILABLE_MEMORY_NVX, &cur_avail_mem_kb);
return static_cast<u64>(cur_avail_mem_kb) * 1_KiB;
}
return 2_GiB; // Return minimum requirements
}();
}
TextureCacheRuntime::~TextureCacheRuntime() = default;
@ -500,13 +509,13 @@ ImageBufferMap TextureCacheRuntime::DownloadStagingBuffer(size_t size) {
return download_buffers.RequestMap(size, false);
}
u64 TextureCacheRuntime::GetDeviceLocalMemory() const {
u64 TextureCacheRuntime::GetDeviceMemoryUsage() const {
if (GLAD_GL_NVX_gpu_memory_info) {
GLint cur_avail_mem_kb = 0;
glGetIntegerv(GL_GPU_MEMORY_INFO_CURRENT_AVAILABLE_VIDMEM_NVX, &cur_avail_mem_kb);
return static_cast<u64>(cur_avail_mem_kb) * 1_KiB;
}
return 2_GiB; // Return minimum requirements
return 2_GiB;
}
void TextureCacheRuntime::CopyImage(Image& dst_image, Image& src_image,
@ -686,6 +695,7 @@ Image::Image(TextureCacheRuntime& runtime_, const VideoCommon::ImageInfo& info_,
}
if (IsConverted(runtime->device, info.format, info.type)) {
flags |= ImageFlagBits::Converted;
flags |= ImageFlagBits::GCProtected;
gl_internal_format = IsPixelFormatSRGB(info.format) ? GL_SRGB8_ALPHA8 : GL_RGBA8;
gl_format = GL_RGBA;
gl_type = GL_UNSIGNED_INT_8_8_8_8_REV;

View file

@ -83,7 +83,15 @@ public:
ImageBufferMap DownloadStagingBuffer(size_t size);
u64 GetDeviceLocalMemory() const;
u64 GetDeviceLocalMemory() const {
return device_access_memory;
}
u64 GetDeviceMemoryUsage() const;
bool CanReportMemoryUsage() const {
return GLAD_GL_NVX_gpu_memory_info;
}
bool ShouldReinterpret([[maybe_unused]] Image& dst, [[maybe_unused]] Image& src) {
return true;
@ -172,6 +180,7 @@ private:
std::array<OGLFramebuffer, 4> rescale_draw_fbos;
std::array<OGLFramebuffer, 4> rescale_read_fbos;
const Settings::ResolutionScalingInfo& resolution;
u64 device_access_memory;
};
class Image : public VideoCommon::ImageBase {

View file

@ -118,7 +118,7 @@ StagingBufferPool::StagingBufferPool(const Device& device_, MemoryAllocator& mem
.image = nullptr,
.buffer = *stream_buffer,
};
const auto memory_properties = device.GetPhysical().GetMemoryProperties();
const auto memory_properties = device.GetPhysical().GetMemoryProperties().memoryProperties;
VkMemoryAllocateInfo stream_memory_info{
.sType = VK_STRUCTURE_TYPE_MEMORY_ALLOCATE_INFO,
.pNext = make_dedicated ? &dedicated_info : nullptr,

View file

@ -1189,6 +1189,14 @@ u64 TextureCacheRuntime::GetDeviceLocalMemory() const {
return device.GetDeviceLocalMemory();
}
u64 TextureCacheRuntime::GetDeviceMemoryUsage() const {
return device.GetDeviceMemoryUsage();
}
bool TextureCacheRuntime::CanReportMemoryUsage() const {
return device.CanReportMemoryUsage();
}
void TextureCacheRuntime::TickFrame() {}
Image::Image(TextureCacheRuntime& runtime_, const ImageInfo& info_, GPUVAddr gpu_addr_,
@ -1203,6 +1211,7 @@ Image::Image(TextureCacheRuntime& runtime_, const ImageInfo& info_, GPUVAddr gpu
} else {
flags |= VideoCommon::ImageFlagBits::Converted;
}
flags |= VideoCommon::ImageFlagBits::GCProtected;
}
if (runtime->device.HasDebuggingToolAttached()) {
original_image.SetObjectNameEXT(VideoCommon::Name(*this).c_str());

View file

@ -55,6 +55,10 @@ public:
u64 GetDeviceLocalMemory() const;
u64 GetDeviceMemoryUsage() const;
bool CanReportMemoryUsage() const;
void BlitImage(Framebuffer* dst_framebuffer, ImageView& dst, ImageView& src,
const Region2D& dst_region, const Region2D& src_region,
Tegra::Engines::Fermi2D::Filter filter,

View file

@ -29,15 +29,16 @@ enum class ImageFlagBits : u32 {
Sparse = 1 << 9, ///< Image has non continous submemory.
// Garbage Collection Flags
BadOverlap = 1 << 10, ///< This image overlaps other but doesn't fit, has higher
///< garbage collection priority
Alias = 1 << 11, ///< This image has aliases and has priority on garbage
///< collection
BadOverlap = 1 << 10, ///< This image overlaps other but doesn't fit, has higher
///< garbage collection priority
Alias = 1 << 11, ///< This image has aliases and has priority on garbage
///< collection
GCProtected = 1 << 12, ///< Protected from low-tier GC as they are costy to load back.
// Rescaler
Rescaled = 1 << 12,
CheckingRescalable = 1 << 13,
IsRescalable = 1 << 14,
Rescaled = 1 << 13,
CheckingRescalable = 1 << 14,
IsRescalable = 1 << 15,
};
DECLARE_ENUM_FLAG_OPERATORS(ImageFlagBits)

View file

@ -50,14 +50,21 @@ TextureCache<P>::TextureCache(Runtime& runtime_, VideoCore::RasterizerInterface&
void(slot_samplers.insert(runtime, sampler_descriptor));
if constexpr (HAS_DEVICE_MEMORY_INFO) {
const auto device_memory = runtime.GetDeviceLocalMemory();
const u64 possible_expected_memory = (device_memory * 4) / 10;
const u64 possible_critical_memory = (device_memory * 7) / 10;
expected_memory = std::max(possible_expected_memory, DEFAULT_EXPECTED_MEMORY - 256_MiB);
critical_memory = std::max(possible_critical_memory, DEFAULT_CRITICAL_MEMORY - 512_MiB);
minimum_memory = 0;
const s64 device_memory = static_cast<s64>(runtime.GetDeviceLocalMemory());
const s64 min_spacing_expected = device_memory - 1_GiB - 512_MiB;
const s64 min_spacing_critical = device_memory - 1_GiB;
const s64 mem_tresshold = std::min(device_memory, TARGET_THRESHOLD);
const s64 min_vacancy_expected = (6 * mem_tresshold) / 10;
const s64 min_vacancy_critical = (3 * mem_tresshold) / 10;
expected_memory = static_cast<u64>(
std::max(std::min(device_memory - min_vacancy_expected, min_spacing_expected),
DEFAULT_EXPECTED_MEMORY));
critical_memory = static_cast<u64>(
std::max(std::min(device_memory - min_vacancy_critical, min_spacing_critical),
DEFAULT_CRITICAL_MEMORY));
minimum_memory = static_cast<u64>((device_memory - mem_tresshold) / 2);
LOG_CRITICAL(Debug, "Available Memory: {}", device_memory / 1_MiB);
} else {
// On OpenGL we can be more conservatives as the driver takes care.
expected_memory = DEFAULT_EXPECTED_MEMORY + 512_MiB;
critical_memory = DEFAULT_CRITICAL_MEMORY + 1_GiB;
minimum_memory = 0;
@ -76,7 +83,8 @@ void TextureCache<P>::RunGarbageCollector() {
}
--num_iterations;
auto& image = slot_images[image_id];
const bool must_download = image.IsSafeDownload();
const bool must_download =
image.IsSafeDownload() && False(image.flags & ImageFlagBits::BadOverlap);
if (!high_priority_mode && must_download) {
return false;
}
@ -99,6 +107,10 @@ void TextureCache<P>::RunGarbageCollector() {
template <class P>
void TextureCache<P>::TickFrame() {
// If we can obtain the memory info, use it instead of the estimate.
if (runtime.CanReportMemoryUsage()) {
total_used_memory = runtime.GetDeviceMemoryUsage();
}
if (total_used_memory > minimum_memory) {
RunGarbageCollector();
}
@ -106,7 +118,9 @@ void TextureCache<P>::TickFrame() {
sentenced_framebuffers.Tick();
sentenced_image_view.Tick();
runtime.TickFrame();
critical_gc = 0;
++frame_tick;
LOG_CRITICAL(Debug, "Current memory: {}", total_used_memory / 1_MiB);
}
template <class P>
@ -1052,6 +1066,11 @@ ImageId TextureCache<P>::JoinImages(const ImageInfo& info, GPUVAddr gpu_addr, VA
for (const ImageId overlap_id : overlap_ids) {
Image& overlap = slot_images[overlap_id];
if (True(overlap.flags & ImageFlagBits::GpuModified)) {
new_image.flags |= ImageFlagBits::GpuModified;
new_image.modification_tick =
std::max(overlap.modification_tick, new_image.modification_tick);
}
if (overlap.info.num_samples != new_image.info.num_samples) {
LOG_WARNING(HW_GPU, "Copying between images with different samples is not implemented");
} else {
@ -1414,6 +1433,10 @@ void TextureCache<P>::RegisterImage(ImageId image_id) {
tentative_size = EstimatedDecompressedSize(tentative_size, image.info.format);
}
total_used_memory += Common::AlignUp(tentative_size, 1024);
if (total_used_memory > critical_memory && critical_gc < GC_EMERGENCY_COUNTS) {
RunGarbageCollector();
critical_gc++;
}
image.lru_index = lru_cache.Insert(image_id, frame_tick);
ForEachGPUPage(image.gpu_addr, image.guest_size_bytes,
@ -1704,6 +1727,9 @@ void TextureCache<P>::SynchronizeAliases(ImageId image_id) {
most_recent_tick = std::max(most_recent_tick, aliased_image.modification_tick);
aliased_images.push_back(&aliased);
any_rescaled |= True(aliased_image.flags & ImageFlagBits::Rescaled);
if (True(aliased_image.flags & ImageFlagBits::GpuModified)) {
image.flags |= ImageFlagBits::GpuModified;
}
}
}
if (aliased_images.empty()) {

View file

@ -59,8 +59,12 @@ class TextureCache {
/// True when the API can provide info about the memory of the device.
static constexpr bool HAS_DEVICE_MEMORY_INFO = P::HAS_DEVICE_MEMORY_INFO;
static constexpr u64 DEFAULT_EXPECTED_MEMORY = 1_GiB;
static constexpr u64 DEFAULT_CRITICAL_MEMORY = 2_GiB;
static constexpr s64 TARGET_THRESHOLD = 4_GiB;
static constexpr s64 MIN_VACANCY_EXPECTED = (6 * TARGET_THRESHOLD) / 10;
static constexpr s64 MIN_VACANCY_CRITICAL = (3 * TARGET_THRESHOLD) / 10;
static constexpr s64 DEFAULT_EXPECTED_MEMORY = 1_GiB + 125_MiB;
static constexpr s64 DEFAULT_CRITICAL_MEMORY = 1_GiB + 625_MiB;
static constexpr size_t GC_EMERGENCY_COUNTS = 2;
using Runtime = typename P::Runtime;
using Image = typename P::Image;
@ -372,6 +376,7 @@ private:
u64 minimum_memory;
u64 expected_memory;
u64 critical_memory;
size_t critical_gc;
SlotVector<Image> slot_images;
SlotVector<ImageMapView> slot_map_views;

View file

@ -12,12 +12,14 @@
#include <vector>
#include "common/assert.h"
#include "common/literals.h"
#include "common/settings.h"
#include "video_core/vulkan_common/nsight_aftermath_tracker.h"
#include "video_core/vulkan_common/vulkan_device.h"
#include "video_core/vulkan_common/vulkan_wrapper.h"
namespace Vulkan {
using namespace Common::Literals;
namespace {
namespace Alternatives {
constexpr std::array STENCIL8_UINT{
@ -596,6 +598,11 @@ Device::Device(VkInstance instance_, vk::PhysicalDevice physical_, VkSurfaceKHR
}
logical = vk::Device::Create(physical, queue_cis, extensions, first_next, dld);
is_integrated = (properties.deviceType & VK_PHYSICAL_DEVICE_TYPE_INTEGRATED_GPU) != 0;
is_virtual = (properties.deviceType & VK_PHYSICAL_DEVICE_TYPE_VIRTUAL_GPU) != 0;
is_non_gpu = (properties.deviceType & VK_PHYSICAL_DEVICE_TYPE_OTHER) != 0 ||
(properties.deviceType & VK_PHYSICAL_DEVICE_TYPE_CPU) != 0;
CollectPhysicalMemoryInfo();
CollectTelemetryParameters();
CollectToolingInfo();
@ -985,6 +992,7 @@ std::vector<const char*> Device::LoadExtensions(bool requires_surface) {
test(has_khr_swapchain_mutable_format, VK_KHR_SWAPCHAIN_MUTABLE_FORMAT_EXTENSION_NAME,
false);
test(has_ext_line_rasterization, VK_EXT_LINE_RASTERIZATION_EXTENSION_NAME, false);
test(ext_memory_budget, VK_EXT_MEMORY_BUDGET_EXTENSION_NAME, true);
if (Settings::values.enable_nsight_aftermath) {
test(nv_device_diagnostics_config, VK_NV_DEVICE_DIAGNOSTICS_CONFIG_EXTENSION_NAME,
true);
@ -997,7 +1005,7 @@ std::vector<const char*> Device::LoadExtensions(bool requires_surface) {
VkPhysicalDeviceFeatures2KHR features{};
features.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_FEATURES_2_KHR;
VkPhysicalDeviceProperties2KHR physical_properties;
VkPhysicalDeviceProperties2KHR physical_properties{};
physical_properties.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PROPERTIES_2_KHR;
if (has_khr_shader_float16_int8) {
@ -1267,15 +1275,51 @@ void Device::CollectTelemetryParameters() {
vendor_name = driver.driverName;
}
u64 Device::GetDeviceMemoryUsage() const {
VkPhysicalDeviceMemoryBudgetPropertiesEXT budget;
budget.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MEMORY_BUDGET_PROPERTIES_EXT;
budget.pNext = nullptr;
physical.GetMemoryProperties(&budget);
u64 result{};
for (const size_t heap : valid_heap_memory) {
result += budget.heapUsage[heap];
}
return result;
}
void Device::CollectPhysicalMemoryInfo() {
const auto mem_properties = physical.GetMemoryProperties();
VkPhysicalDeviceMemoryBudgetPropertiesEXT budget{};
budget.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MEMORY_BUDGET_PROPERTIES_EXT;
const auto mem_info = physical.GetMemoryProperties(ext_memory_budget ? &budget : nullptr);
const auto& mem_properties = mem_info.memoryProperties;
const size_t num_properties = mem_properties.memoryHeapCount;
device_access_memory = 0;
u64 device_initial_usage = 0;
u64 local_memory = 0;
for (size_t element = 0; element < num_properties; ++element) {
if ((mem_properties.memoryHeaps[element].flags & VK_MEMORY_HEAP_DEVICE_LOCAL_BIT) != 0) {
device_access_memory += mem_properties.memoryHeaps[element].size;
const bool is_heap_local =
mem_properties.memoryHeaps[element].flags & VK_MEMORY_HEAP_DEVICE_LOCAL_BIT != 0;
if (!is_integrated && !is_heap_local) {
continue;
}
valid_heap_memory.push_back(element);
if (is_heap_local) {
local_memory += mem_properties.memoryHeaps[element].size;
}
if (ext_memory_budget) {
device_initial_usage += budget.heapUsage[element];
device_access_memory += budget.heapBudget[element];
continue;
}
device_access_memory += mem_properties.memoryHeaps[element].size;
}
if (!is_integrated) {
return;
}
const s64 available_memory = static_cast<s64>(device_access_memory - device_initial_usage);
device_access_memory = static_cast<u64>(std::max<s64>(
std::min<s64>(available_memory - 8_GiB, 4_GiB), static_cast<s64>(local_memory)));
device_initial_usage = 0;
}
void Device::CollectToolingInfo() {

View file

@ -341,6 +341,12 @@ public:
return device_access_memory;
}
bool CanReportMemoryUsage() const {
return ext_memory_budget;
}
u64 GetDeviceMemoryUsage() const;
u32 GetSetsPerPool() const {
return sets_per_pool;
}
@ -421,6 +427,9 @@ private:
bool is_topology_list_restart_supported{}; ///< Support for primitive restart with list
///< topologies.
bool is_patch_list_restart_supported{}; ///< Support for primitive restart with list patch.
bool is_integrated{}; ///< Is GPU an iGPU.
bool is_virtual{}; ///< Is GPU a virtual GPU.
bool is_non_gpu{}; ///< Is SoftwareRasterizer, FPGA, non-GPU device.
bool nv_viewport_swizzle{}; ///< Support for VK_NV_viewport_swizzle.
bool nv_viewport_array2{}; ///< Support for VK_NV_viewport_array2.
bool nv_geometry_shader_passthrough{}; ///< Support for VK_NV_geometry_shader_passthrough.
@ -445,6 +454,7 @@ private:
bool ext_shader_atomic_int64{}; ///< Support for VK_KHR_shader_atomic_int64.
bool ext_conservative_rasterization{}; ///< Support for VK_EXT_conservative_rasterization.
bool ext_provoking_vertex{}; ///< Support for VK_EXT_provoking_vertex.
bool ext_memory_budget{}; ///< Support for VK_EXT_memory_budget.
bool nv_device_diagnostics_config{}; ///< Support for VK_NV_device_diagnostics_config.
bool has_broken_cube_compatibility{}; ///< Has broken cube compatiblity bit
bool has_renderdoc{}; ///< Has RenderDoc attached
@ -456,6 +466,7 @@ private:
// Telemetry parameters
std::string vendor_name; ///< Device's driver name.
std::vector<std::string> supported_extensions; ///< Reported Vulkan extensions.
std::vector<size_t> valid_heap_memory; ///< Heaps used.
/// Format properties dictionary.
std::unordered_map<VkFormat, VkFormatProperties> format_properties;

View file

@ -227,7 +227,7 @@ void MemoryCommit::Release() {
}
MemoryAllocator::MemoryAllocator(const Device& device_, bool export_allocations_)
: device{device_}, properties{device_.GetPhysical().GetMemoryProperties()},
: device{device_}, properties{device_.GetPhysical().GetMemoryProperties().memoryProperties},
export_allocations{export_allocations_},
buffer_image_granularity{
device_.GetPhysical().GetProperties().limits.bufferImageGranularity} {}

View file

@ -237,8 +237,8 @@ bool Load(VkInstance instance, InstanceDispatch& dld) noexcept {
return X(vkCreateDevice) && X(vkDestroyDevice) && X(vkDestroyDevice) &&
X(vkEnumerateDeviceExtensionProperties) && X(vkEnumeratePhysicalDevices) &&
X(vkGetDeviceProcAddr) && X(vkGetPhysicalDeviceFormatProperties) &&
X(vkGetPhysicalDeviceMemoryProperties) && X(vkGetPhysicalDeviceProperties) &&
X(vkGetPhysicalDeviceQueueFamilyProperties);
X(vkGetPhysicalDeviceMemoryProperties) && X(vkGetPhysicalDeviceMemoryProperties2) &&
X(vkGetPhysicalDeviceProperties) && X(vkGetPhysicalDeviceQueueFamilyProperties);
#undef X
}
@ -926,9 +926,12 @@ std::vector<VkPresentModeKHR> PhysicalDevice::GetSurfacePresentModesKHR(
return modes;
}
VkPhysicalDeviceMemoryProperties PhysicalDevice::GetMemoryProperties() const noexcept {
VkPhysicalDeviceMemoryProperties properties;
dld->vkGetPhysicalDeviceMemoryProperties(physical_device, &properties);
VkPhysicalDeviceMemoryProperties2 PhysicalDevice::GetMemoryProperties(
void* next_structures) const noexcept {
VkPhysicalDeviceMemoryProperties2 properties{};
properties.sType = VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_MEMORY_PROPERTIES_2;
properties.pNext = next_structures;
dld->vkGetPhysicalDeviceMemoryProperties2(physical_device, &properties);
return properties;
}

View file

@ -172,6 +172,7 @@ struct InstanceDispatch {
PFN_vkGetPhysicalDeviceFeatures2KHR vkGetPhysicalDeviceFeatures2KHR{};
PFN_vkGetPhysicalDeviceFormatProperties vkGetPhysicalDeviceFormatProperties{};
PFN_vkGetPhysicalDeviceMemoryProperties vkGetPhysicalDeviceMemoryProperties{};
PFN_vkGetPhysicalDeviceMemoryProperties2 vkGetPhysicalDeviceMemoryProperties2{};
PFN_vkGetPhysicalDeviceProperties vkGetPhysicalDeviceProperties{};
PFN_vkGetPhysicalDeviceProperties2KHR vkGetPhysicalDeviceProperties2KHR{};
PFN_vkGetPhysicalDeviceQueueFamilyProperties vkGetPhysicalDeviceQueueFamilyProperties{};
@ -950,7 +951,8 @@ public:
std::vector<VkPresentModeKHR> GetSurfacePresentModesKHR(VkSurfaceKHR) const;
VkPhysicalDeviceMemoryProperties GetMemoryProperties() const noexcept;
VkPhysicalDeviceMemoryProperties2 GetMemoryProperties(
void* next_structures = nullptr) const noexcept;
private:
VkPhysicalDevice physical_device = nullptr;