mirror/rpcs3
1
0
Fork 0
mirror of https://github.com/RPCS3/rpcs3.git synced 2025-03-14 10:21:21 +00:00
Code Issues Packages Projects Releases Wiki Activity

vk: Use image hot-cache for faster allocation times

Browse Source 
- Creating new images is expensive.
- We can keep around a set of images that have been recently discarded and use them instead of creating new ones from scratch each time.
This commit is contained in:
kd-11 2022-02-05 21:11:56 +03:00 committed by kd-11
parent 02cdf8ac63
commit dca3d477c9
5 changed files with 143 additions and 127 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
rpcs3/Emu/RSX/Common/texture_cache.h
View File

@ -1021,6 +1021,12 @@ namespace rsx
virtual void on_frame_end()
{
// Must manually release each cached entry
for (auto& entry : m_temporary_subresource_cache)
{
release_temporary_subresource(entry.second.second);
}
m_temporary_subresource_cache.clear();
m_predictor.on_frame_end();
reset_frame_statistics();

7
rpcs3/Emu/RSX/VK/VKOverlays.cpp
View File

@ -589,6 +589,13 @@ namespace vk
for (const auto& _key : keys_to_remove)
{
auto& img_data = temp_image_cache[_key];
auto& view_data = temp_view_cache[_key];
auto gc = vk::get_resource_manager();
gc->dispose(img_data.second);
gc->dispose(view_data);
temp_image_cache.erase(_key);
temp_view_cache.erase(_key);
}

17
rpcs3/Emu/RSX/VK/VKResourceManager.h
View File

@ -14,6 +14,11 @@ namespace vk
u64 last_completed_event_id();
void on_event_completed(u64 event_id, bool flush = false);
struct disposable_t
{
virtual void dispose() = 0;
};
struct eid_scope_t
{
u64 eid;
@ -24,6 +29,7 @@ namespace vk
std::vector<std::unique_ptr<vk::event>> m_disposed_events;
std::vector<std::unique_ptr<vk::query_pool>> m_disposed_query_pools;
std::vector<std::unique_ptr<vk::sampler>> m_disposed_samplers;
std::vector<std::unique_ptr<vk::disposable_t>> m_disposables;
eid_scope_t(u64 _eid):
eid(_eid), m_device(g_render_device)
@ -42,6 +48,12 @@ namespace vk
m_disposed_images.clear();
m_disposed_query_pools.clear();
m_disposed_samplers.clear();
for (auto& disposable : m_disposables)
{
disposable->dispose();
}
m_disposables.clear();
}
};
@ -185,6 +197,11 @@ namespace vk
get_current_eid_scope().m_disposed_samplers.emplace_back(std::move(sampler));
}
void dispose(std::unique_ptr<vk::disposable_t>& disposable)
{
get_current_eid_scope().m_disposables.emplace_back(std::move(disposable));
}
void eid_completed(u64 eid)
{
while (!m_eid_map.empty())

177
rpcs3/Emu/RSX/VK/VKTextureCache.cpp
View File

@ -7,6 +7,23 @@
namespace vk
{
texture_cache::cached_image_reference_t::cached_image_reference_t(texture_cache* parent, std::unique_ptr<vk::viewable_image>& previous)
{
this->parent = parent;
this->data = std::move(previous);
}
void texture_cache::cached_image_reference_t::dispose()
{
// Erase layout information to force TOP_OF_PIPE transition next time.
data->current_layout = VK_IMAGE_LAYOUT_UNDEFINED;
// Move this object to the cached image pool
std::lock_guard lock(parent->m_cached_pool_lock);
parent->m_cached_memory_size += data->memory->size();
parent->m_cached_images.emplace_front(std::move(data));
}
void cached_texture_section::dma_transfer(vk::command_buffer& cmd, vk::image* src, const areai& src_area, const utils::address_range& valid_range, u32 pitch)
{
ensure(src->samples() == 1);
@ -167,7 +184,8 @@ namespace vk
{
if (tex.is_managed())
{
vk::get_resource_manager()->dispose(tex.get_texture());
auto disposable = std::unique_ptr<vk::disposable_t>(new cached_image_reference_t(this, tex.get_texture()));
vk::get_resource_manager()->dispose(disposable);
}
}
@ -175,8 +193,8 @@ namespace vk
{
baseclass::clear();
m_temporary_storage.clear();
m_temporary_memory_size = 0;
m_cached_images.clear();
m_cached_memory_size = 0;
}
void texture_cache::copy_transfer_regions_impl(vk::command_buffer& cmd, vk::image* dst, const std::vector<copy_region_descriptor>& sections_to_transfer) const
@ -449,36 +467,47 @@ namespace vk
return result;
}
std::unique_ptr<vk::viewable_image> texture_cache::find_temporary_image(VkFormat format, u16 w, u16 h, u16 d, u8 mipmaps)
std::unique_ptr<vk::viewable_image> texture_cache::find_cached_image(VkFormat format, u16 w, u16 h, u16 d, u8 mipmaps, VkFlags flags)
{
for (auto& e : m_temporary_storage)
auto hash_properties = [](VkFormat format, u16 w, u16 h, u16 d, u8 mipmaps, VkFlags flags)
{
if (e.can_reuse && e.matches(format, w, h, d, mipmaps, 0))
ensure(static_cast<u32>(format) < 0xFF);
return (static_cast<u64>(format) & 0xFF) |
(static_cast<u64>(w) << 8) |
(static_cast<u64>(h) << 24) |
(static_cast<u64>(d) << 40) |
(static_cast<u64>(mipmaps) << 48) |
(static_cast<u64>(flags) << 56);
};
reader_lock lock(m_cached_pool_lock);
if (!m_cached_images.empty())
{
const u64 desired_key = hash_properties(format, w, h, d, mipmaps, flags);
lock.upgrade();
for (auto it = m_cached_images.begin(); it != m_cached_images.end(); ++it)
{
m_temporary_memory_size -= e.block_size;
e.block_size = 0;
e.can_reuse = false;
return std::move(e.combined_image);
const auto& info = (*it)->info;
const u64 this_key = hash_properties(info.format, info.extent.width, info.extent.height, info.extent.depth, info.mipLevels, info.flags);
if (this_key == desired_key)
{
auto ret = std::move(*it);
m_cached_images.erase(it);
m_cached_memory_size -= ret->memory->size();
return ret;
}
}
}
return {};
}
std::unique_ptr<vk::viewable_image> texture_cache::find_temporary_cubemap(VkFormat format, u16 size)
std::unique_ptr<vk::viewable_image> texture_cache::find_cached_cubemap(VkFormat format, u16 size)
{
for (auto& e : m_temporary_storage)
{
if (e.can_reuse && e.matches(format, size, size, 1, 1, VK_IMAGE_CREATE_CUBE_COMPATIBLE_BIT))
{
m_temporary_memory_size -= e.block_size;
e.block_size = 0;
e.can_reuse = false;
return std::move(e.combined_image);
}
}
return {};
return find_cached_image(format, size, size, 1, 1, VK_IMAGE_CREATE_CUBE_COMPATIBLE_BIT);
}
vk::image_view* texture_cache::create_temporary_subresource_view_impl(vk::command_buffer& cmd, vk::image* source, VkImageType image_type, VkImageViewType view_type,
@ -492,11 +521,11 @@ namespace vk
if (!image_flags) [[likely]]
{
image = find_temporary_image(dst_format, w, h, 1, mips);
image = find_cached_image(dst_format, w, h, 1, mips, 0);
}
else
{
image = find_temporary_cubemap(dst_format, w);
image = find_cached_cubemap(dst_format, w);
layers = 6;
}
@ -551,11 +580,8 @@ namespace vk
vk::change_image_layout(cmd, image.get(), VK_IMAGE_LAYOUT_SHADER_READ_ONLY_OPTIMAL);
}
const u32 resource_memory = w * h * 4; //Rough approximate
m_temporary_storage.emplace_back(image);
m_temporary_storage.back().block_size = resource_memory;
m_temporary_memory_size += resource_memory;
// TODO: Floating reference. We can do better with some restructuring.
image.release();
return view;
}
@ -715,15 +741,12 @@ namespace vk
void texture_cache::release_temporary_subresource(vk::image_view* view)
{
auto handle = dynamic_cast<vk::viewable_image*>(view->image());
for (auto& e : m_temporary_storage)
{
if (e.combined_image.get() == handle)
{
e.can_reuse = true;
return;
}
}
auto resource = dynamic_cast<vk::viewable_image*>(view->image());
ensure(resource);
auto image = std::unique_ptr<vk::viewable_image>(resource);
auto disposable = std::unique_ptr<vk::disposable_t>(new cached_image_reference_t(this, image));
vk::get_resource_manager()->dispose(disposable);
}
void texture_cache::update_image_contents(vk::command_buffer& cmd, vk::image_view* dst_view, vk::image* src, u16 width, u16 height)
@ -826,13 +849,21 @@ namespace vk
{
const bool is_cubemap = type == rsx::texture_dimension_extended::texture_dimension_cubemap;
const VkFormat vk_format = get_compatible_sampler_format(m_formats_support, gcm_format);
const VkFlags flags = is_cubemap ? VK_IMAGE_CREATE_CUBE_COMPATIBLE_BIT : 0;
image = new vk::viewable_image(*m_device,
m_memory_types.device_local, VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT,
image_type, vk_format,
width, height, depth, mipmaps, layer, VK_SAMPLE_COUNT_1_BIT, VK_IMAGE_LAYOUT_UNDEFINED,
VK_IMAGE_TILING_OPTIMAL, usage_flags, is_cubemap ? VK_IMAGE_CREATE_CUBE_COMPATIBLE_BIT : 0,
VMM_ALLOCATION_POOL_TEXTURE_CACHE, rsx::classify_format(gcm_format));
if (auto found = find_cached_image(vk_format, width, height, depth, mipmaps, flags))
{
image = found.release();
}
else
{
image = new vk::viewable_image(*m_device,
m_memory_types.device_local, VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT,
image_type, vk_format,
width, height, depth, mipmaps, layer, VK_SAMPLE_COUNT_1_BIT, VK_IMAGE_LAYOUT_UNDEFINED,
VK_IMAGE_TILING_OPTIMAL, usage_flags, is_cubemap ? VK_IMAGE_CREATE_CUBE_COMPATIBLE_BIT : 0,
VMM_ALLOCATION_POOL_TEXTURE_CACHE, rsx::classify_format(gcm_format));
}
// New section, we must prepare it
region.reset(rsx_range);
@ -1135,14 +1166,14 @@ namespace vk
auto any_released = baseclass::handle_memory_pressure(severity);
// TODO: This can cause invalidation of in-flight resources
if (severity <= rsx::problem_severity::low || !m_temporary_memory_size)
if (severity <= rsx::problem_severity::low || !m_cached_memory_size)
{
// Nothing left to do
return any_released;
}
constexpr u64 _1M = 0x100000;
if (severity <= rsx::problem_severity::moderate && m_temporary_memory_size < (64 * _1M))
if (severity <= rsx::problem_severity::moderate && m_cached_memory_size < (64 * _1M))
{
// Some memory is consumed by the temporary resources, but no need to panic just yet
return any_released;
@ -1159,20 +1190,9 @@ namespace vk
auto gc = vk::get_resource_manager();
u64 actual_released_memory = 0;
for (auto& entry : m_temporary_storage)
{
if (!entry.combined_image)
{
continue;
}
m_cached_images.clear();
m_cached_memory_size = 0;
actual_released_memory += entry.combined_image->memory->size();
gc->dispose(entry.combined_image);
m_temporary_memory_size -= entry.block_size;
}
ensure(m_temporary_memory_size == 0);
m_temporary_storage.clear();
m_temporary_subresource_cache.clear();
rsx_log.warning("Texture cache released %lluM of temporary resources.", (actual_released_memory / _1M));
@ -1183,27 +1203,27 @@ namespace vk
{
trim_sections();
if (m_storage.m_unreleased_texture_objects >= m_max_zombie_objects ||
m_temporary_memory_size > 0x4000000) //If already holding over 64M in discardable memory, be frugal with memory resources
if (m_storage.m_unreleased_texture_objects >= m_max_zombie_objects)
{
purge_unreleased_sections();
}
const u64 last_complete_frame = vk::get_last_completed_frame_id();
m_temporary_storage.remove_if([&](const temporary_storage& o)
if (m_cached_images.size() > max_cached_image_pool_size ||
m_cached_memory_size > 256 * 0x100000)
{
if (!o.block_size || o.test(last_complete_frame))
{
m_temporary_memory_size -= o.block_size;
return true;
}
return false;
});
std::lock_guard lock(m_cached_pool_lock);
m_temporary_subresource_cache.clear();
reset_frame_statistics();
const auto new_size = m_cached_images.size() / 2;
for (usz i = new_size; i < m_cached_images.size(); ++i)
{
m_cached_memory_size -= m_cached_images[i]->memory->size();
}
m_cached_images.resize(new_size);
}
baseclass::on_frame_end();
reset_frame_statistics();
}
vk::viewable_image* texture_cache::upload_image_simple(vk::command_buffer& cmd, VkFormat format, u32 address, u32 width, u32 height, u32 pitch)
@ -1266,10 +1286,9 @@ namespace vk
vk::change_image_layout(cmd, image.get(), VK_IMAGE_LAYOUT_TRANSFER_SRC_OPTIMAL);
auto result = image.get();
const u32 resource_memory = width * height * 4; //Rough approximate
m_temporary_storage.emplace_back(image);
m_temporary_storage.back().block_size = resource_memory;
m_temporary_memory_size += resource_memory;
const u32 resource_memory = image->memory->size();
auto disposable = std::unique_ptr<vk::disposable_t>(new cached_image_reference_t(this, image));
vk::get_resource_manager()->dispose(disposable);
return result;
}
@ -1294,12 +1313,12 @@ namespace vk
u32 texture_cache::get_unreleased_textures_count() const
{
return baseclass::get_unreleased_textures_count() + ::size32(m_temporary_storage);
return baseclass::get_unreleased_textures_count() + ::size32(m_cached_images);
}
u32 texture_cache::get_temporary_memory_in_use() const
{
return m_temporary_memory_size;
return m_cached_memory_size;
}
bool texture_cache::is_overallocated() const

63
rpcs3/Emu/RSX/VK/VKTextureCache.h
View File

@ -350,56 +350,21 @@ namespace vk
}
};
struct temporary_storage
{
std::unique_ptr<vk::viewable_image> combined_image;
bool can_reuse = false;
// Memory held by this temp storage object
u32 block_size = 0;
// Frame id tag
const u64 frame_tag = vk::get_current_frame_id();
temporary_storage(std::unique_ptr<vk::viewable_image>& _img)
{
combined_image = std::move(_img);
}
temporary_storage(vk::cached_texture_section& tex)
{
combined_image = std::move(tex.get_texture());
block_size = tex.get_section_size();
}
bool test(u64 ref_frame) const
{
return ref_frame > 0 && frame_tag <= ref_frame;
}
bool matches(VkFormat format, u16 w, u16 h, u16 d, u16 mipmaps, VkFlags flags) const
{
if (combined_image &&
combined_image->info.flags == flags &&
combined_image->format() == format &&
combined_image->width() == w &&
combined_image->height() == h &&
combined_image->depth() == d &&
combined_image->mipmaps() == mipmaps)
{
return true;
}
return false;
}
};
class texture_cache : public rsx::texture_cache<vk::texture_cache, vk::texture_cache_traits>
{
private:
using baseclass = rsx::texture_cache<vk::texture_cache, vk::texture_cache_traits>;
friend baseclass;
struct cached_image_reference_t : vk::disposable_t
{
std::unique_ptr<vk::viewable_image> data;
texture_cache* parent;
cached_image_reference_t(texture_cache* parent, std::unique_ptr<vk::viewable_image>& previous);
void dispose() override;
};
public:
enum texture_create_flags : u32
{
@ -418,8 +383,10 @@ namespace vk
vk::data_heap* m_texture_upload_heap;
//Stuff that has been dereferenced goes into these
std::list<temporary_storage> m_temporary_storage;
atomic_t<u32> m_temporary_memory_size = { 0 };
const u32 max_cached_image_pool_size = 256;
std::deque<std::unique_ptr<vk::viewable_image>> m_cached_images;
atomic_t<u32> m_cached_memory_size = { 0 };
shared_mutex m_cached_pool_lock;
void clear();
@ -429,9 +396,9 @@ namespace vk
vk::image* get_template_from_collection_impl(const std::vector<copy_region_descriptor>& sections_to_transfer) const;
std::unique_ptr<vk::viewable_image> find_temporary_image(VkFormat format, u16 w, u16 h, u16 d, u8 mipmaps);
std::unique_ptr<vk::viewable_image> find_cached_image(VkFormat format, u16 w, u16 h, u16 d, u8 mipmaps, VkFlags flags);
std::unique_ptr<vk::viewable_image> find_temporary_cubemap(VkFormat format, u16 size);
std::unique_ptr<vk::viewable_image> find_cached_cubemap(VkFormat format, u16 size);
protected:
vk::image_view* create_temporary_subresource_view_impl(vk::command_buffer& cmd, vk::image* source, VkImageType image_type, VkImageViewType view_type,