Merge pull request #7720 from FernandoS27/yfc-gc
First Nugget: Reworked Garbage Collection to be smarter [originally from Project YFC]
This commit is contained in:
commit
ab6a5784fa
20 changed files with 259 additions and 43 deletions
|
@ -76,8 +76,9 @@ class BufferCache {
|
|||
|
||||
static constexpr BufferId NULL_BUFFER_ID{0};
|
||||
|
||||
static constexpr u64 EXPECTED_MEMORY = 512_MiB;
|
||||
static constexpr u64 CRITICAL_MEMORY = 1_GiB;
|
||||
static constexpr s64 DEFAULT_EXPECTED_MEMORY = 512_MiB;
|
||||
static constexpr s64 DEFAULT_CRITICAL_MEMORY = 1_GiB;
|
||||
static constexpr s64 TARGET_THRESHOLD = 4_GiB;
|
||||
|
||||
using Maxwell = Tegra::Engines::Maxwell3D::Regs;
|
||||
|
||||
|
@ -436,6 +437,8 @@ private:
|
|||
Common::LeastRecentlyUsedCache<LRUItemParams> lru_cache;
|
||||
u64 frame_tick = 0;
|
||||
u64 total_used_memory = 0;
|
||||
u64 minimum_memory = 0;
|
||||
u64 critical_memory = 0;
|
||||
|
||||
std::array<BufferId, ((1ULL << 39) >> PAGE_BITS)> page_table;
|
||||
};
|
||||
|
@ -451,11 +454,30 @@ BufferCache<P>::BufferCache(VideoCore::RasterizerInterface& rasterizer_,
|
|||
// Ensure the first slot is used for the null buffer
|
||||
void(slot_buffers.insert(runtime, NullBufferParams{}));
|
||||
common_ranges.clear();
|
||||
|
||||
if (!runtime.CanReportMemoryUsage()) {
|
||||
minimum_memory = DEFAULT_EXPECTED_MEMORY;
|
||||
critical_memory = DEFAULT_CRITICAL_MEMORY;
|
||||
return;
|
||||
}
|
||||
|
||||
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_threshold = std::min(device_memory, TARGET_THRESHOLD);
|
||||
const s64 min_vacancy_expected = (6 * mem_threshold) / 10;
|
||||
const s64 min_vacancy_critical = (3 * mem_threshold) / 10;
|
||||
minimum_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));
|
||||
}
|
||||
|
||||
template <class P>
|
||||
void BufferCache<P>::RunGarbageCollector() {
|
||||
const bool aggressive_gc = total_used_memory >= CRITICAL_MEMORY;
|
||||
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;
|
||||
const auto clean_up = [this, &num_iterations](BufferId buffer_id) {
|
||||
|
@ -486,7 +508,11 @@ void BufferCache<P>::TickFrame() {
|
|||
const bool skip_preferred = hits * 256 < shots * 251;
|
||||
uniform_buffer_skip_cache_size = skip_preferred ? DEFAULT_SKIP_CACHE_SIZE : 0;
|
||||
|
||||
if (total_used_memory >= EXPECTED_MEMORY) {
|
||||
// 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();
|
||||
}
|
||||
++frame_tick;
|
||||
|
|
|
@ -135,6 +135,20 @@ BufferCacheRuntime::BufferCacheRuntime(const Device& device_)
|
|||
buffer.Create();
|
||||
glNamedBufferData(buffer.handle, 0x10'000, nullptr, GL_STREAM_COPY);
|
||||
}
|
||||
|
||||
device_access_memory = [this]() -> u64 {
|
||||
if (device.CanReportMemoryUsage()) {
|
||||
return device.GetCurrentDedicatedVideoMemory() + 512_MiB;
|
||||
}
|
||||
return 2_GiB; // Return minimum requirements
|
||||
}();
|
||||
}
|
||||
|
||||
u64 BufferCacheRuntime::GetDeviceMemoryUsage() const {
|
||||
if (device.CanReportMemoryUsage()) {
|
||||
return device_access_memory - device.GetCurrentDedicatedVideoMemory();
|
||||
}
|
||||
return 2_GiB;
|
||||
}
|
||||
|
||||
void BufferCacheRuntime::CopyBuffer(Buffer& dst_buffer, Buffer& src_buffer,
|
||||
|
|
|
@ -89,6 +89,8 @@ public:
|
|||
void BindImageBuffer(Buffer& buffer, u32 offset, u32 size,
|
||||
VideoCore::Surface::PixelFormat format);
|
||||
|
||||
u64 GetDeviceMemoryUsage() const;
|
||||
|
||||
void BindFastUniformBuffer(size_t stage, u32 binding_index, u32 size) {
|
||||
const GLuint handle = fast_uniforms[stage][binding_index].handle;
|
||||
const GLsizeiptr gl_size = static_cast<GLsizeiptr>(size);
|
||||
|
@ -151,6 +153,14 @@ public:
|
|||
use_storage_buffers = use_storage_buffers_;
|
||||
}
|
||||
|
||||
u64 GetDeviceLocalMemory() const {
|
||||
return device_access_memory;
|
||||
}
|
||||
|
||||
bool CanReportMemoryUsage() const {
|
||||
return device.CanReportMemoryUsage();
|
||||
}
|
||||
|
||||
private:
|
||||
static constexpr std::array PABO_LUT{
|
||||
GL_VERTEX_PROGRAM_PARAMETER_BUFFER_NV, GL_TESS_CONTROL_PROGRAM_PARAMETER_BUFFER_NV,
|
||||
|
@ -184,6 +194,8 @@ private:
|
|||
std::array<OGLBuffer, VideoCommon::NUM_COMPUTE_UNIFORM_BUFFERS> copy_compute_uniforms;
|
||||
|
||||
u32 index_buffer_offset = 0;
|
||||
|
||||
u64 device_access_memory;
|
||||
};
|
||||
|
||||
struct BufferCacheParams {
|
||||
|
|
|
@ -13,12 +13,15 @@
|
|||
|
||||
#include <glad/glad.h>
|
||||
|
||||
#include "common/literals.h"
|
||||
#include "common/logging/log.h"
|
||||
#include "common/settings.h"
|
||||
#include "shader_recompiler/stage.h"
|
||||
#include "video_core/renderer_opengl/gl_device.h"
|
||||
#include "video_core/renderer_opengl/gl_resource_manager.h"
|
||||
|
||||
using namespace Common::Literals;
|
||||
|
||||
namespace OpenGL {
|
||||
namespace {
|
||||
constexpr std::array LIMIT_UBOS = {
|
||||
|
@ -165,6 +168,7 @@ Device::Device() {
|
|||
has_sparse_texture_2 = GLAD_GL_ARB_sparse_texture2;
|
||||
warp_size_potentially_larger_than_guest = !is_nvidia && !is_intel;
|
||||
need_fastmath_off = is_nvidia;
|
||||
can_report_memory = GLAD_GL_NVX_gpu_memory_info;
|
||||
|
||||
// At the moment of writing this, only Nvidia's driver optimizes BufferSubData on exclusive
|
||||
// uniform buffers as "push constants"
|
||||
|
@ -276,4 +280,10 @@ void main() {
|
|||
})");
|
||||
}
|
||||
|
||||
u64 Device::GetCurrentDedicatedVideoMemory() const {
|
||||
GLint cur_avail_mem_kb = 0;
|
||||
glGetIntegerv(GL_GPU_MEMORY_INFO_DEDICATED_VIDMEM_NVX, &cur_avail_mem_kb);
|
||||
return static_cast<u64>(cur_avail_mem_kb) * 1_KiB;
|
||||
}
|
||||
|
||||
} // namespace OpenGL
|
||||
|
|
|
@ -20,6 +20,8 @@ public:
|
|||
|
||||
[[nodiscard]] std::string GetVendorName() const;
|
||||
|
||||
u64 GetCurrentDedicatedVideoMemory() const;
|
||||
|
||||
u32 GetMaxUniformBuffers(Shader::Stage stage) const noexcept {
|
||||
return max_uniform_buffers[static_cast<size_t>(stage)];
|
||||
}
|
||||
|
@ -168,6 +170,10 @@ public:
|
|||
return vendor_name == "ATI Technologies Inc.";
|
||||
}
|
||||
|
||||
bool CanReportMemoryUsage() const {
|
||||
return can_report_memory;
|
||||
}
|
||||
|
||||
private:
|
||||
static bool TestVariableAoffi();
|
||||
static bool TestPreciseBug();
|
||||
|
@ -210,6 +216,7 @@ private:
|
|||
bool need_fastmath_off{};
|
||||
bool has_cbuf_ftou_bug{};
|
||||
bool has_bool_ref_bug{};
|
||||
bool can_report_memory{};
|
||||
|
||||
std::string vendor_name;
|
||||
};
|
||||
|
|
|
@ -484,6 +484,13 @@ TextureCacheRuntime::TextureCacheRuntime(const Device& device_, ProgramManager&
|
|||
rescale_read_fbos[i].Create();
|
||||
}
|
||||
}
|
||||
|
||||
device_access_memory = [this]() -> u64 {
|
||||
if (device.CanReportMemoryUsage()) {
|
||||
return device.GetCurrentDedicatedVideoMemory() + 512_MiB;
|
||||
}
|
||||
return 2_GiB; // Return minimum requirements
|
||||
}();
|
||||
}
|
||||
|
||||
TextureCacheRuntime::~TextureCacheRuntime() = default;
|
||||
|
@ -500,13 +507,11 @@ ImageBufferMap TextureCacheRuntime::DownloadStagingBuffer(size_t size) {
|
|||
return download_buffers.RequestMap(size, false);
|
||||
}
|
||||
|
||||
u64 TextureCacheRuntime::GetDeviceLocalMemory() 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;
|
||||
u64 TextureCacheRuntime::GetDeviceMemoryUsage() const {
|
||||
if (device.CanReportMemoryUsage()) {
|
||||
return device_access_memory - device.GetCurrentDedicatedVideoMemory();
|
||||
}
|
||||
return 2_GiB; // Return minimum requirements
|
||||
return 2_GiB;
|
||||
}
|
||||
|
||||
void TextureCacheRuntime::CopyImage(Image& dst_image, Image& src_image,
|
||||
|
@ -686,6 +691,7 @@ Image::Image(TextureCacheRuntime& runtime_, const VideoCommon::ImageInfo& info_,
|
|||
}
|
||||
if (IsConverted(runtime->device, info.format, info.type)) {
|
||||
flags |= ImageFlagBits::Converted;
|
||||
flags |= ImageFlagBits::CostlyLoad;
|
||||
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;
|
||||
|
|
|
@ -10,6 +10,7 @@
|
|||
#include <glad/glad.h>
|
||||
|
||||
#include "shader_recompiler/shader_info.h"
|
||||
#include "video_core/renderer_opengl/gl_device.h"
|
||||
#include "video_core/renderer_opengl/gl_resource_manager.h"
|
||||
#include "video_core/renderer_opengl/util_shaders.h"
|
||||
#include "video_core/texture_cache/image_view_base.h"
|
||||
|
@ -21,7 +22,6 @@ struct ResolutionScalingInfo;
|
|||
|
||||
namespace OpenGL {
|
||||
|
||||
class Device;
|
||||
class ProgramManager;
|
||||
class StateTracker;
|
||||
|
||||
|
@ -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 device.CanReportMemoryUsage();
|
||||
}
|
||||
|
||||
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 {
|
||||
|
|
|
@ -141,6 +141,18 @@ StagingBufferRef BufferCacheRuntime::DownloadStagingBuffer(size_t size) {
|
|||
return staging_pool.Request(size, MemoryUsage::Download);
|
||||
}
|
||||
|
||||
u64 BufferCacheRuntime::GetDeviceLocalMemory() const {
|
||||
return device.GetDeviceLocalMemory();
|
||||
}
|
||||
|
||||
u64 BufferCacheRuntime::GetDeviceMemoryUsage() const {
|
||||
return device.GetDeviceMemoryUsage();
|
||||
}
|
||||
|
||||
bool BufferCacheRuntime::CanReportMemoryUsage() const {
|
||||
return device.CanReportMemoryUsage();
|
||||
}
|
||||
|
||||
void BufferCacheRuntime::Finish() {
|
||||
scheduler.Finish();
|
||||
}
|
||||
|
|
|
@ -65,6 +65,12 @@ public:
|
|||
|
||||
void Finish();
|
||||
|
||||
u64 GetDeviceLocalMemory() const;
|
||||
|
||||
u64 GetDeviceMemoryUsage() const;
|
||||
|
||||
bool CanReportMemoryUsage() const;
|
||||
|
||||
[[nodiscard]] StagingBufferRef UploadStagingBuffer(size_t size);
|
||||
|
||||
[[nodiscard]] StagingBufferRef DownloadStagingBuffer(size_t size);
|
||||
|
|
|
@ -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,
|
||||
|
|
|
@ -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::CostlyLoad;
|
||||
}
|
||||
if (runtime->device.HasDebuggingToolAttached()) {
|
||||
original_image.SetObjectNameEXT(VideoCommon::Name(*this).c_str());
|
||||
|
|
|
@ -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,
|
||||
|
|
|
@ -33,11 +33,12 @@ enum class ImageFlagBits : u32 {
|
|||
///< garbage collection priority
|
||||
Alias = 1 << 11, ///< This image has aliases and has priority on garbage
|
||||
///< collection
|
||||
CostlyLoad = 1 << 12, ///< Protected from low-tier GC as it is costly 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)
|
||||
|
||||
|
|
|
@ -50,14 +50,20 @@ 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_threshold = std::min(device_memory, TARGET_THRESHOLD);
|
||||
const s64 min_vacancy_expected = (6 * mem_threshold) / 10;
|
||||
const s64 min_vacancy_critical = (3 * mem_threshold) / 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_threshold) / 2);
|
||||
} 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;
|
||||
|
@ -66,18 +72,21 @@ TextureCache<P>::TextureCache(Runtime& runtime_, VideoCore::RasterizerInterface&
|
|||
|
||||
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 = aggressive_mode ? 10ULL : high_priority_mode ? 25ULL : 100ULL;
|
||||
size_t num_iterations = aggressive_mode ? 300 : (high_priority_mode ? 50 : 10);
|
||||
const auto clean_up = [this, &num_iterations, high_priority_mode](ImageId image_id) {
|
||||
bool high_priority_mode = total_used_memory >= expected_memory;
|
||||
bool aggressive_mode = total_used_memory >= critical_memory;
|
||||
const u64 ticks_to_destroy = aggressive_mode ? 10ULL : high_priority_mode ? 25ULL : 50ULL;
|
||||
size_t num_iterations = aggressive_mode ? 40 : (high_priority_mode ? 20 : 10);
|
||||
const auto clean_up = [this, &num_iterations, &high_priority_mode,
|
||||
&aggressive_mode](ImageId image_id) {
|
||||
if (num_iterations == 0) {
|
||||
return true;
|
||||
}
|
||||
--num_iterations;
|
||||
auto& image = slot_images[image_id];
|
||||
const bool must_download = image.IsSafeDownload();
|
||||
if (!high_priority_mode && must_download) {
|
||||
const bool must_download =
|
||||
image.IsSafeDownload() && False(image.flags & ImageFlagBits::BadOverlap);
|
||||
if (!high_priority_mode &&
|
||||
(must_download || True(image.flags & ImageFlagBits::CostlyLoad))) {
|
||||
return false;
|
||||
}
|
||||
if (must_download) {
|
||||
|
@ -92,6 +101,18 @@ void TextureCache<P>::RunGarbageCollector() {
|
|||
}
|
||||
UnregisterImage(image_id);
|
||||
DeleteImage(image_id, image.scale_tick > frame_tick + 5);
|
||||
if (total_used_memory < critical_memory) {
|
||||
if (aggressive_mode) {
|
||||
// Sink the aggresiveness.
|
||||
num_iterations >>= 2;
|
||||
aggressive_mode = false;
|
||||
return false;
|
||||
}
|
||||
if (high_priority_mode && total_used_memory < expected_memory) {
|
||||
num_iterations >>= 1;
|
||||
high_priority_mode = false;
|
||||
}
|
||||
}
|
||||
return false;
|
||||
};
|
||||
lru_cache.ForEachItemBelow(frame_tick - ticks_to_destroy, clean_up);
|
||||
|
@ -99,6 +120,10 @@ 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();
|
||||
}
|
||||
|
@ -106,6 +131,7 @@ void TextureCache<P>::TickFrame() {
|
|||
sentenced_framebuffers.Tick();
|
||||
sentenced_image_view.Tick();
|
||||
runtime.TickFrame();
|
||||
critical_gc = 0;
|
||||
++frame_tick;
|
||||
}
|
||||
|
||||
|
@ -1052,6 +1078,11 @@ 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 +1445,10 @@ 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 +1739,9 @@ 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()) {
|
||||
|
|
|
@ -59,8 +59,10 @@ 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 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 +374,7 @@ private:
|
|||
u64 minimum_memory;
|
||||
u64 expected_memory;
|
||||
u64 critical_memory;
|
||||
size_t critical_gc;
|
||||
|
||||
SlotVector<Image> slot_images;
|
||||
SlotVector<ImageMapView> slot_map_views;
|
||||
|
|
|
@ -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;
|
||||
is_virtual = properties.deviceType == VK_PHYSICAL_DEVICE_TYPE_VIRTUAL_GPU;
|
||||
is_non_gpu = properties.deviceType == VK_PHYSICAL_DEVICE_TYPE_OTHER ||
|
||||
properties.deviceType == VK_PHYSICAL_DEVICE_TYPE_CPU;
|
||||
|
||||
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 +1275,50 @@ 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)));
|
||||
}
|
||||
|
||||
void Device::CollectToolingInfo() {
|
||||
|
|
|
@ -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;
|
||||
|
|
|
@ -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} {}
|
||||
|
|
|
@ -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;
|
||||
}
|
||||
|
||||
|
|
|
@ -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;
|
||||
|
|
Loading…
Reference in a new issue