Reaper: Change memory restrictions on TC depending on host memory on VK.

master
Fernando Sahmkow 2021-06-17 00:29:48 +07:00
parent 0dd98842bf
commit ca6f47c686
10 changed files with 90 additions and 41 deletions

@ -106,6 +106,8 @@ public:
void TickFrame();
void RunGarbageCollector();
void WriteMemory(VAddr cpu_addr, u64 size);
void CachedWriteMemory(VAddr cpu_addr, u64 size);
@ -350,29 +352,7 @@ BufferCache<P>::BufferCache(VideoCore::RasterizerInterface& rasterizer_,
}
template <class P>
void BufferCache<P>::TickFrame() {
const bool enabled_gc = Settings::values.use_caches_gc.GetValue();
SCOPE_EXIT({
++frame_tick;
delayed_destruction_ring.Tick();
});
// 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 shots = std::reduce(uniform_cache_shots.begin(), uniform_cache_shots.end());
std::copy_n(uniform_cache_hits.begin(), uniform_cache_hits.size() - 1,
uniform_cache_hits.begin() + 1);
std::copy_n(uniform_cache_shots.begin(), uniform_cache_shots.size() - 1,
uniform_cache_shots.begin() + 1);
uniform_cache_hits[0] = 0;
uniform_cache_shots[0] = 0;
const bool skip_preferred = hits * 256 < shots * 251;
uniform_buffer_skip_cache_size = skip_preferred ? DEFAULT_SKIP_CACHE_SIZE : 0;
const bool activate_gc = enabled_gc && total_used_memory >= EXPECTED_MEMORY;
if (!activate_gc) {
return;
}
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;
@ -392,6 +372,28 @@ void BufferCache<P>::TickFrame() {
}
}
template <class P>
void BufferCache<P>::TickFrame() {
// 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 shots = std::reduce(uniform_cache_shots.begin(), uniform_cache_shots.end());
std::copy_n(uniform_cache_hits.begin(), uniform_cache_hits.size() - 1,
uniform_cache_hits.begin() + 1);
std::copy_n(uniform_cache_shots.begin(), uniform_cache_shots.size() - 1,
uniform_cache_shots.begin() + 1);
uniform_cache_hits[0] = 0;
uniform_cache_shots[0] = 0;
const bool skip_preferred = hits * 256 < shots * 251;
uniform_buffer_skip_cache_size = skip_preferred ? DEFAULT_SKIP_CACHE_SIZE : 0;
if (Settings::values.use_caches_gc.GetValue() && total_used_memory >= EXPECTED_MEMORY) {
RunGarbageCollector();
}
++frame_tick;
delayed_destruction_ring.Tick();
}
template <class P>
void BufferCache<P>::WriteMemory(VAddr cpu_addr, u64 size) {
ForEachBufferInRange(cpu_addr, size, [&](BufferId, Buffer& buffer) {

@ -235,6 +235,7 @@ struct TextureCacheParams {
static constexpr bool ENABLE_VALIDATION = true;
static constexpr bool FRAMEBUFFER_BLITS = true;
static constexpr bool HAS_EMULATED_COPIES = true;
static constexpr bool HAS_DEVICE_MEMORY_INFO = false;
using Runtime = OpenGL::TextureCacheRuntime;
using Image = OpenGL::Image;

@ -818,6 +818,10 @@ void TextureCacheRuntime::CopyImage(Image& dst, Image& src,
});
}
u64 TextureCacheRuntime::GetDeviceLocalMemory() const {
return device.GetDeviceLocalMemory();
}
Image::Image(TextureCacheRuntime& runtime, const ImageInfo& info_, GPUVAddr gpu_addr_,
VAddr cpu_addr_)
: VideoCommon::ImageBase(info_, gpu_addr_, cpu_addr_), scheduler{&runtime.scheduler},

@ -97,6 +97,8 @@ struct TextureCacheRuntime {
// All known Vulkan drivers can natively handle BGR textures
return true;
}
u64 GetDeviceLocalMemory() const;
};
class Image : public VideoCommon::ImageBase {
@ -257,6 +259,7 @@ struct TextureCacheParams {
static constexpr bool ENABLE_VALIDATION = true;
static constexpr bool FRAMEBUFFER_BLITS = false;
static constexpr bool HAS_EMULATED_COPIES = false;
static constexpr bool HAS_DEVICE_MEMORY_INFO = true;
using Runtime = Vulkan::TextureCacheRuntime;
using Image = Vulkan::Image;

@ -79,7 +79,7 @@ public:
Iterator(SlotVector<T>* slot_vector_, SlotId id_) noexcept
: slot_vector{slot_vector_}, id{id_} {}
bool IsValid(const u64* bitset) noexcept {
bool IsValid(const u64* bitset) const noexcept {
return ((bitset[id.index / 64] >> (id.index % 64)) & 1) != 0;
}

@ -71,14 +71,16 @@ class TextureCache {
static constexpr bool FRAMEBUFFER_BLITS = P::FRAMEBUFFER_BLITS;
/// True when some copies have to be emulated
static constexpr bool HAS_EMULATED_COPIES = P::HAS_EMULATED_COPIES;
/// 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;
/// Image view ID for null descriptors
static constexpr ImageViewId NULL_IMAGE_VIEW_ID{0};
/// Sampler ID for bugged sampler ids
static constexpr SamplerId NULL_SAMPLER_ID{0};
static constexpr u64 EXPECTED_MEMORY = Common::Size_1_GB;
static constexpr u64 CRITICAL_MEMORY = Common::Size_2_GB;
static constexpr u64 DEFAULT_EXPECTED_MEMORY = Common::Size_1_GB;
static constexpr u64 DEFAULT_CRITICAL_MEMORY = Common::Size_2_GB;
using Runtime = typename P::Runtime;
using Image = typename P::Image;
@ -108,6 +110,9 @@ public:
/// Notify the cache that a new frame has been queued
void TickFrame();
/// Runs the Garbage Collector.
void RunGarbageCollector();
/// Return a constant reference to the given image view id
[[nodiscard]] const ImageView& GetImageView(ImageViewId id) const noexcept;
@ -339,6 +344,8 @@ private:
bool has_deleted_images = false;
u64 total_used_memory = 0;
u64 expected_memory;
u64 critical_memory;
SlotVector<Image> slot_images;
SlotVector<ImageView> slot_image_views;
@ -382,21 +389,23 @@ TextureCache<P>::TextureCache(Runtime& runtime_, VideoCore::RasterizerInterface&
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;
const u64 possible_critical_memory = (device_memory * 6) / 10;
expected_memory = std::max(possible_expected_memory, DEFAULT_EXPECTED_MEMORY);
critical_memory = std::max(possible_critical_memory, DEFAULT_CRITICAL_MEMORY);
} else {
expected_memory = DEFAULT_EXPECTED_MEMORY;
critical_memory = DEFAULT_CRITICAL_MEMORY;
}
}
template <class P>
void TextureCache<P>::TickFrame() {
const bool enabled_gc = Settings::values.use_caches_gc.GetValue();
if (!enabled_gc) {
// @Note(Blinkhawk): compile error with SCOPE_EXIT on msvc.
sentenced_images.Tick();
sentenced_framebuffers.Tick();
sentenced_image_view.Tick();
++frame_tick;
return;
}
const bool high_priority_mode = total_used_memory >= EXPECTED_MEMORY;
const bool aggressive_mode = total_used_memory >= CRITICAL_MEMORY;
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) {
@ -451,11 +460,18 @@ void TextureCache<P>::TickFrame() {
UnregisterImage(image_id);
DeleteImage(image_id);
if (is_bad_overlap) {
num_iterations++;
++num_iterations;
}
}
++deletion_iterator;
}
}
template <class P>
void TextureCache<P>::TickFrame() {
if (Settings::values.use_caches_gc.GetValue()) {
RunGarbageCollector();
}
sentenced_images.Tick();
sentenced_framebuffers.Tick();
sentenced_image_view.Tick();

@ -408,6 +408,7 @@ Device::Device(VkInstance instance_, vk::PhysicalDevice physical_, VkSurfaceKHR
}
logical = vk::Device::Create(physical, queue_cis, extensions, first_next, dld);
CollectPhysicalMemoryInfo();
CollectTelemetryParameters();
CollectToolingInfo();
@ -818,6 +819,19 @@ void Device::CollectTelemetryParameters() {
}
}
void Device::CollectPhysicalMemoryInfo() {
const auto mem_properties = physical.GetMemoryProperties();
const std::size_t num_properties = mem_properties.memoryTypeCount;
device_access_memory = 0;
for (std::size_t element = 0; element < num_properties; element++) {
if ((mem_properties.memoryTypes[element].propertyFlags &
VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT) != 0) {
const std::size_t heap_index = mem_properties.memoryTypes[element].heapIndex;
device_access_memory += mem_properties.memoryHeaps[heap_index].size;
}
}
}
void Device::CollectToolingInfo() {
if (!ext_tooling_info) {
return;

@ -225,6 +225,10 @@ public:
return use_asynchronous_shaders;
}
u64 GetDeviceLocalMemory() const {
return device_access_memory;
}
private:
/// Checks if the physical device is suitable.
void CheckSuitability(bool requires_swapchain) const;
@ -244,6 +248,9 @@ private:
/// Collects information about attached tools.
void CollectToolingInfo();
/// Collects information about the device's local memory.
void CollectPhysicalMemoryInfo();
/// Returns a list of queue initialization descriptors.
std::vector<VkDeviceQueueCreateInfo> GetDeviceQueueCreateInfos() const;
@ -302,6 +309,8 @@ private:
/// Nsight Aftermath GPU crash tracker
std::unique_ptr<NsightAftermathTracker> nsight_aftermath_tracker;
u64 device_access_memory;
};
} // namespace Vulkan

@ -109,7 +109,7 @@
<string>Enables garbage collection for the GPU caches, this will try to keep VRAM within 3-4 GB by flushing the least used textures/buffers. May cause issues in a few games.</string>
</property>
<property name="text">
<string>Enable GPU caches garbage collection (unsafe)</string>
<string>Enable GPU cache garbage collection (unsafe)</string>
</property>
</widget>
</item>

@ -227,7 +227,7 @@ use_asynchronous_gpu_emulation =
# 0: Off, 1 (default): On
use_vsync =
# Whether to use garbage collection or not.
# Whether to use garbage collection or not for GPU caches.
# 0 (default): Off, 1: On
use_caches_gc =