rpcs3/rpcs3/Emu/RSX/Common/surface_store.h

1466 lines
43 KiB
C++

#pragma once
#include "surface_utils.h"
#include "simple_array.hpp"
#include "ranged_map.hpp"
#include "surface_cache_dma.hpp"
#include "../gcm_enums.h"
#include "../rsx_utils.h"
#include <list>
#include "util/asm.hpp"
namespace rsx
{
namespace utility
{
std::vector<u8> get_rtt_indexes(surface_target color_target);
u8 get_mrt_buffers_count(surface_target color_target);
usz get_aligned_pitch(surface_color_format format, u32 width);
usz get_packed_pitch(surface_color_format format, u32 width);
}
template <typename Traits>
struct surface_store
{
static constexpr u32 get_aa_factor_u(surface_antialiasing aa_mode)
{
return (aa_mode == surface_antialiasing::center_1_sample)? 1 : 2;
}
static constexpr u32 get_aa_factor_v(surface_antialiasing aa_mode)
{
switch (aa_mode)
{
case surface_antialiasing::center_1_sample:
case surface_antialiasing::diagonal_centered_2_samples:
return 1;
default:
return 2;
}
}
public:
using surface_storage_type = typename Traits::surface_storage_type;
using surface_type = typename Traits::surface_type;
using command_list_type = typename Traits::command_list_type;
using surface_overlap_info = surface_overlap_info_t<surface_type>;
using surface_ranged_map = ranged_map<surface_storage_type, 0x400000>;
using surface_cache_dma_map = surface_cache_dma<Traits, 0x400000>;
protected:
surface_ranged_map m_render_targets_storage = {};
surface_ranged_map m_depth_stencil_storage = {};
rsx::address_range32 m_render_targets_memory_range;
rsx::address_range32 m_depth_stencil_memory_range;
surface_cache_dma_map m_dma_block;
bool m_invalidate_on_write = false;
rsx::surface_raster_type m_active_raster_type = rsx::surface_raster_type::linear;
public:
rsx::simple_array<u8> m_bound_render_target_ids = {};
std::array<std::pair<u32, surface_type>, 4> m_bound_render_targets = {};
std::pair<u32, surface_type> m_bound_depth_stencil = {};
// List of sections derived from a section that has been split and invalidated
std::vector<std::pair<u32, surface_type>> orphaned_surfaces;
// List of sections that have been wholly inherited and invalidated
std::vector<surface_type> superseded_surfaces;
std::list<surface_storage_type> invalidated_resources;
const u64 max_invalidated_resources_count = 256ull;
u64 cache_tag = 1ull; // Use 1 as the start since 0 is default tag on new surfaces
u64 write_tag = 1ull;
// Amount of virtual PS3 memory tied to allocated textures
u64 m_active_memory_used = 0;
surface_store() = default;
~surface_store() = default;
surface_store(const surface_store&) = delete;
private:
template <bool is_depth_surface>
void split_surface_region(command_list_type cmd, u32 address, surface_type prev_surface, u16 width, u16 height, u8 bpp, rsx::surface_antialiasing aa)
{
auto insert_new_surface = [&](
u32 new_address,
deferred_clipped_region<surface_type>& region,
surface_ranged_map& data)
{
surface_storage_type sink;
surface_type invalidated = 0;
if (const auto found = data.find(new_address);
found != data.end())
{
if (Traits::is_compatible_surface(Traits::get(found->second), region.source, region.width, region.height, 1))
{
if (found->second->last_use_tag >= prev_surface->last_use_tag)
{
// If memory in this block is newer, do not overwrite with stale data
return;
}
// There is no need to erase due to the reinsertion below
sink = std::move(found->second);
}
else
{
invalidate(found->second);
data.erase(new_address);
auto &old = invalidated_resources.back();
if (Traits::surface_is_pitch_compatible(old, prev_surface->get_rsx_pitch()))
{
if (old->last_use_tag >= prev_surface->last_use_tag) [[unlikely]]
{
invalidated = Traits::get(old);
}
}
}
}
if (sink)
{
// Memory requirements can be altered when cloning
free_rsx_memory(Traits::get(sink));
}
Traits::clone_surface(cmd, sink, region.source, new_address, region);
allocate_rsx_memory(Traits::get(sink));
if (invalidated) [[unlikely]]
{
// Halfplement the merge by crude inheritance. Should recursively split the memory blocks instead.
if (sink->old_contents.empty()) [[likely]]
{
sink->set_old_contents(invalidated);
}
else
{
const auto existing = sink->get_normalized_memory_area();
const auto incoming = invalidated->get_normalized_memory_area();
deferred_clipped_region<surface_type> region{};
region.source = invalidated;
region.target = Traits::get(sink);
region.width = std::min(existing.x2, incoming.x2);
region.height = std::min(existing.y2, incoming.y2);
sink->set_old_contents_region(region, true);
}
}
ensure(region.target == Traits::get(sink));
orphaned_surfaces.push_back({ address, region.target });
data.emplace(region.target->get_memory_range(), std::move(sink));
};
// Define incoming region
size2u old, _new;
const auto prev_area = prev_surface->get_normalized_memory_area();
const auto prev_bpp = prev_surface->get_bpp();
old.width = prev_area.x2;
old.height = prev_area.y2;
_new.width = width * bpp * get_aa_factor_u(aa);
_new.height = height * get_aa_factor_v(aa);
if (old.width <= _new.width && old.height <= _new.height)
{
// No extra memory to be preserved
return;
}
// One-time data validity test
ensure(prev_surface);
if (prev_surface->read_barrier(cmd); !prev_surface->test())
{
return;
}
if (old.width > _new.width)
{
// Split in X
const u32 baseaddr = address + _new.width;
const u32 bytes_to_texels_x = (prev_bpp * prev_surface->samples_x);
deferred_clipped_region<surface_type> copy;
copy.src_x = _new.width / bytes_to_texels_x;
copy.src_y = 0;
copy.dst_x = 0;
copy.dst_y = 0;
copy.width = std::max<u16>((old.width - _new.width) / bytes_to_texels_x, 1);
copy.height = prev_surface->template get_surface_height<>();
copy.transfer_scale_x = 1.f;
copy.transfer_scale_y = 1.f;
copy.target = nullptr;
copy.source = prev_surface;
if constexpr (is_depth_surface)
{
insert_new_surface(baseaddr, copy, m_depth_stencil_storage);
}
else
{
insert_new_surface(baseaddr, copy, m_render_targets_storage);
}
}
if (old.height > _new.height)
{
// Split in Y
const u32 baseaddr = address + (_new.height * prev_surface->get_rsx_pitch());
const u32 bytes_to_texels_x = (prev_bpp * prev_surface->samples_x);
deferred_clipped_region<surface_type> copy;
copy.src_x = 0;
copy.src_y = _new.height / prev_surface->samples_y;
copy.dst_x = 0;
copy.dst_y = 0;
copy.width = std::max<u16>(std::min(_new.width, old.width) / bytes_to_texels_x, 1);
copy.height = std::max<u16>((old.height - _new.height) / prev_surface->samples_y, 1);
copy.transfer_scale_x = 1.f;
copy.transfer_scale_y = 1.f;
copy.target = nullptr;
copy.source = prev_surface;
if constexpr (is_depth_surface)
{
insert_new_surface(baseaddr, copy, m_depth_stencil_storage);
}
else
{
insert_new_surface(baseaddr, copy, m_render_targets_storage);
}
}
}
template <bool is_depth_surface>
void intersect_surface_region(command_list_type cmd, u32 address, surface_type new_surface, surface_type prev_surface)
{
auto scan_list = [&new_surface, address](const rsx::address_range32& mem_range,
surface_ranged_map& data) -> std::vector<std::pair<u32, surface_type>>
{
std::vector<std::pair<u32, surface_type>> result;
for (auto it = data.begin_range(mem_range); it != data.end(); ++it)
{
auto surface = Traits::get(it->second);
if (new_surface->last_use_tag >= surface->last_use_tag ||
new_surface == surface ||
address == it->first)
{
// Do not bother synchronizing with uninitialized data
continue;
}
// Memory partition check
if (mem_range.start >= constants::local_mem_base)
{
if (it->first < constants::local_mem_base) continue;
}
else
{
if (it->first >= constants::local_mem_base) continue;
}
// Pitch check
if (!rsx::pitch_compatible(surface, new_surface))
{
continue;
}
// Range check
const rsx::address_range32 this_range = surface->get_memory_range();
if (!this_range.overlaps(mem_range))
{
continue;
}
result.push_back({ it->first, surface });
ensure(it->first == surface->base_addr);
}
return result;
};
const rsx::address_range32 mem_range = new_surface->get_memory_range();
auto list1 = scan_list(mem_range, m_render_targets_storage);
auto list2 = scan_list(mem_range, m_depth_stencil_storage);
if (prev_surface)
{
// Append the previous removed surface to the intersection list
if constexpr (is_depth_surface)
{
list2.push_back({ address, prev_surface });
}
else
{
list1.push_back({ address, prev_surface });
}
}
else
{
if (list1.empty() && list2.empty())
{
return;
}
}
std::vector<std::pair<u32, surface_type>> surface_info;
if (list1.empty())
{
surface_info = std::move(list2);
}
else if (list2.empty())
{
surface_info = std::move(list1);
}
else
{
const auto reserve = list1.size() + list2.size();
surface_info = std::move(list1);
surface_info.reserve(reserve);
for (const auto& e : list2) surface_info.push_back(e);
}
for (const auto &e: surface_info)
{
auto this_address = e.first;
auto surface = e.second;
if (surface->old_contents.size() == 1) [[unlikely]]
{
// Dirty zombies are possible with unused pixel storage subslices and are valid
// Avoid double transfer if possible
// This is an optional optimization that can be safely disabled
surface = static_cast<decltype(surface)>(surface->old_contents[0].source);
// Ignore self-reference
if (new_surface == surface)
{
continue;
}
// If this surface has already been added via another descendant, just ignore it
bool ignore = false;
for (const auto& slice : new_surface->old_contents)
{
if (slice.source == surface)
{
ignore = true;
break;
}
}
if (ignore) continue;
this_address = surface->base_addr;
ensure(this_address);
}
if (new_surface->inherit_surface_contents(surface) == surface_inheritance_result::full &&
surface->memory_usage_flags == surface_usage_flags::storage &&
surface != prev_surface &&
surface == e.second)
{
// This has been 'swallowed' by the new surface and can be safely freed
auto& storage = surface->is_depth_surface() ? m_depth_stencil_storage : m_render_targets_storage;
auto& object = ::at32(storage, e.first);
ensure(object);
if (!surface->old_contents.empty()) [[unlikely]]
{
surface->read_barrier(cmd);
}
invalidate(object);
storage.erase(e.first);
superseded_surfaces.push_back(surface);
}
}
}
template <bool depth, typename format_type, typename ...Args>
surface_type bind_surface_address(
command_list_type command_list,
u32 address,
format_type format,
surface_antialiasing antialias,
usz width, usz height, usz pitch,
u8 bpp,
Args&&... extra_params)
{
surface_storage_type old_surface_storage;
surface_storage_type new_surface_storage;
surface_type old_surface = nullptr;
surface_type new_surface = nullptr;
bool do_intersection_test = true;
bool store = true;
// Workaround. Preserve new surface tag value because pitch convert is unimplemented
u64 new_content_tag = 0;
address_range32* storage_bounds;
surface_ranged_map* primary_storage;
surface_ranged_map* secondary_storage;
if constexpr (depth)
{
primary_storage = &m_depth_stencil_storage;
secondary_storage = &m_render_targets_storage;
storage_bounds = &m_depth_stencil_memory_range;
}
else
{
primary_storage = &m_render_targets_storage;
secondary_storage = &m_depth_stencil_storage;
storage_bounds = &m_render_targets_memory_range;
}
// Check if render target already exists
auto It = primary_storage->find(address);
if (It != primary_storage->end())
{
surface_storage_type& surface = It->second;
const bool pitch_compatible = Traits::surface_is_pitch_compatible(surface, pitch);
if (!pitch_compatible)
{
// This object should be pitch-converted and re-intersected with
if (old_surface_storage = Traits::convert_pitch(command_list, surface, pitch); old_surface_storage)
{
old_surface = Traits::get(old_surface_storage);
}
else
{
// Preserve content age. This is hacky, but matches previous behavior
// TODO: Remove when pitch convert is implemented
new_content_tag = Traits::get(surface)->last_use_tag;
}
}
if (Traits::surface_matches_properties(surface, format, width, height, antialias))
{
if (!pitch_compatible)
{
Traits::invalidate_surface_contents(command_list, Traits::get(surface), format, address, pitch);
}
Traits::notify_surface_persist(surface);
Traits::prepare_surface_for_drawing(command_list, Traits::get(surface));
new_surface = Traits::get(surface);
store = false;
}
else
{
if (pitch_compatible)
{
// Preserve memory outside the area to be inherited if needed
split_surface_region<depth>(command_list, address, Traits::get(surface), static_cast<u16>(width), static_cast<u16>(height), bpp, antialias);
old_surface = Traits::get(surface);
}
// This will be unconditionally moved to invalidated list shortly
free_rsx_memory(Traits::get(surface));
Traits::notify_surface_invalidated(surface);
if (old_surface_storage)
{
// Pitch-converted data. Send to invalidated pool immediately.
invalidated_resources.push_back(std::move(old_surface_storage));
}
old_surface_storage = std::move(surface);
primary_storage->erase(It);
}
}
if (!new_surface)
{
// Range test
const auto aa_factor_v = get_aa_factor_v(antialias);
rsx::address_range32 range = rsx::address_range32::start_length(address, static_cast<u32>(pitch * height * aa_factor_v));
*storage_bounds = range.get_min_max(*storage_bounds);
// Search invalidated resources for a suitable surface
for (auto It = invalidated_resources.begin(); It != invalidated_resources.end(); It++)
{
auto &surface = *It;
if (Traits::surface_matches_properties(surface, format, width, height, antialias, true))
{
new_surface_storage = std::move(surface);
Traits::notify_surface_reused(new_surface_storage);
if (old_surface_storage)
{
// Exchange this surface with the invalidated one
surface = std::move(old_surface_storage);
old_surface_storage = {};
}
else
{
// Iterator is now empty - erase it
invalidated_resources.erase(It);
}
new_surface = Traits::get(new_surface_storage);
Traits::invalidate_surface_contents(command_list, new_surface, format, address, pitch);
Traits::prepare_surface_for_drawing(command_list, new_surface);
allocate_rsx_memory(new_surface);
break;
}
}
}
// Check for stale storage
if (old_surface_storage)
{
// This was already determined to be invalid and is excluded from testing above
invalidated_resources.push_back(std::move(old_surface_storage));
}
if (!new_surface)
{
ensure(store);
new_surface_storage = Traits::create_new_surface(address, format, width, height, pitch, antialias, std::forward<Args>(extra_params)...);
new_surface = Traits::get(new_surface_storage);
Traits::prepare_surface_for_drawing(command_list, new_surface);
allocate_rsx_memory(new_surface);
}
// Remove and preserve if possible any overlapping/replaced surface from the other pool
auto aliased_surface = secondary_storage->find(address);
if (aliased_surface != secondary_storage->end())
{
if (Traits::surface_is_pitch_compatible(aliased_surface->second, pitch))
{
auto surface = Traits::get(aliased_surface->second);
split_surface_region<!depth>(command_list, address, surface, static_cast<u16>(width), static_cast<u16>(height), bpp, antialias);
if (!old_surface || old_surface->last_use_tag < surface->last_use_tag)
{
// TODO: This can leak data outside inherited bounds
old_surface = surface;
}
}
invalidate(aliased_surface->second);
secondary_storage->erase(aliased_surface);
}
// Check if old_surface is 'new' and hopefully avoid intersection
if (old_surface)
{
if (old_surface->last_use_tag < new_surface->last_use_tag)
{
// Can happen if aliasing occurs; a probable condition due to memory splitting
// This is highly unlikely but is possible in theory
old_surface = nullptr;
}
else if (old_surface->last_use_tag >= write_tag)
{
const auto new_area = new_surface->get_normalized_memory_area();
const auto old_area = old_surface->get_normalized_memory_area();
if (new_area.x2 <= old_area.x2 && new_area.y2 <= old_area.y2)
{
do_intersection_test = false;
new_surface->set_old_contents(old_surface);
}
}
}
if (do_intersection_test)
{
if (new_content_tag)
{
new_surface->last_use_tag = new_content_tag;
}
intersect_surface_region<depth>(command_list, address, new_surface, old_surface);
}
if (store)
{
// New surface was found among invalidated surfaces or created from scratch
primary_storage->emplace(new_surface->get_memory_range(), std::move(new_surface_storage));
}
ensure(!old_surface_storage);
ensure(new_surface->get_spp() == get_format_sample_count(antialias));
return new_surface;
}
void allocate_rsx_memory(surface_type surface)
{
const auto memory_size = surface->get_memory_range().length();
m_active_memory_used += memory_size;
}
void free_rsx_memory(surface_type surface)
{
ensure(surface->has_refs()); // "Surface memory double free"
if (const auto memory_size = surface->get_memory_range().length();
m_active_memory_used >= memory_size) [[likely]]
{
m_active_memory_used -= memory_size;
}
else
{
rsx_log.error("Memory allocation underflow!");
m_active_memory_used = 0;
}
}
inline void invalidate(surface_storage_type& storage)
{
free_rsx_memory(Traits::get(storage));
Traits::notify_surface_invalidated(storage);
invalidated_resources.push_back(std::move(storage));
}
int remove_duplicates_fast_impl(std::vector<surface_overlap_info>& sections, const rsx::address_range32& range)
{
// Range tests to check for gaps
std::list<utils::address_range32> m_ranges;
bool invalidate_sections = false;
int removed_count = 0;
for (auto it = sections.crbegin(); it != sections.crend(); ++it)
{
auto this_range = it->surface->get_memory_range();
if (invalidate_sections)
{
if (this_range.inside(range))
{
invalidate_surface_address(it->base_address, it->is_depth);
removed_count++;
}
continue;
}
if (it->surface->get_rsx_pitch() != it->surface->get_native_pitch() &&
it->surface->template get_surface_height<>() != 1)
{
// Memory gap in descriptor
continue;
}
// Insert the range, respecting sort order
bool inserted = false;
for (auto iter = m_ranges.begin(); iter != m_ranges.end(); ++iter)
{
if (this_range.start < iter->start)
{
// This range slots in here. Test ranges after this one to find the end position
auto pos = iter;
for (auto _p = ++iter; _p != m_ranges.end();)
{
if (_p->start > (this_range.end + 1))
{
// Gap
break;
}
// Consume
this_range.end = std::max(this_range.end, _p->end);
_p = m_ranges.erase(_p);
}
m_ranges.insert(pos, this_range);
break;
}
}
if (!inserted)
{
m_ranges.push_back(this_range);
}
else if (m_ranges.size() == 1 && range.inside(m_ranges.front()))
{
invalidate_sections = true;
}
}
rsx_log.notice("rsx::surface_cache::check_for_duplicates_fast analysed %u overlapping sections and removed %u", ::size32(sections), removed_count);
return removed_count;
}
void remove_duplicates_fallback_impl(std::vector<surface_overlap_info>& sections, const rsx::address_range32& range)
{
// Originally used to debug crashes but this function breaks often enough that I'll leave the checks in for now.
// Safe to remove after some time if no asserts are reported.
constexpr u32 overrun_cookie_value = 0xCAFEBABEu;
// Generic painter's algorithm to detect obsolete sections
ensure(range.length() < 64 * 0x100000);
std::vector<u8> marker(range.length() + sizeof(overrun_cookie_value), 0);
// Tag end
write_to_ptr(marker, range.length(), overrun_cookie_value);
u32 removed_count = 0;
auto compare_and_tag_row = [&](const u32 offset, u32 length) -> bool
{
u64 mask = 0;
u8* dst_ptr = marker.data() + offset;
while (length >= 8)
{
const u64 value = read_from_ptr<u64>(dst_ptr);
const u64 block_mask = ~value; // If the value is not all 1s, set valid to true
mask |= block_mask;
write_to_ptr<u64>(dst_ptr, umax);
dst_ptr += 8;
length -= 8;
}
if (length >= 4)
{
const u32 value = read_from_ptr<u32>(dst_ptr);
const u32 block_mask = ~value;
mask |= block_mask;
write_to_ptr<u32>(dst_ptr, umax);
dst_ptr += 4;
length -= 4;
}
if (length >= 2)
{
const u16 value = read_from_ptr<u16>(dst_ptr);
const u16 block_mask = ~value;
mask |= block_mask;
write_to_ptr<u16>(dst_ptr, umax);
dst_ptr += 2;
length -= 2;
}
if (length)
{
const u8 value = *dst_ptr;
const u8 block_mask = ~value;
mask |= block_mask;
*dst_ptr = umax;
}
return !!mask;
};
for (auto it = sections.crbegin(); it != sections.crend(); ++it)
{
auto this_range = it->surface->get_memory_range();
ensure(this_range.overlaps(range));
const auto native_pitch = it->surface->template get_surface_width<rsx::surface_metrics::bytes>();
const auto rsx_pitch = it->surface->get_rsx_pitch();
auto num_rows = it->surface->template get_surface_height<rsx::surface_metrics::samples>();
bool valid = false;
if (this_range.start < range.start)
{
// Starts outside bounds
const auto internal_offset = (range.start - this_range.start);
const auto row_num = internal_offset / rsx_pitch;
const auto row_offset = internal_offset % rsx_pitch;
// This section is unconditionally valid
valid = true;
if (row_offset < native_pitch)
{
compare_and_tag_row(0, std::min(native_pitch - row_offset, range.length()));
}
// Jump to next row...
this_range.start = this_range.start + (row_num + 1) * rsx_pitch;
}
if (this_range.end > range.end)
{
// Unconditionally valid
valid = true;
this_range.end = range.end;
}
if (valid)
{
if (this_range.start >= this_range.end)
{
continue;
}
num_rows = utils::aligned_div(this_range.length(), rsx_pitch);
}
for (u32 row = 0, offset = (this_range.start - range.start), section_len = (this_range.end - range.start + 1);
row < num_rows;
++row, offset += rsx_pitch)
{
valid |= compare_and_tag_row(offset, std::min<u32>(native_pitch, (section_len - offset)));
}
if (!valid)
{
removed_count++;
rsx_log.warning("Stale surface at address 0x%x will be deleted", it->base_address);
invalidate_surface_address(it->base_address, it->is_depth);
}
}
// Notify
rsx_log.notice("rsx::surface_cache::check_for_duplicates_fallback analysed %u overlapping sections and removed %u", ::size32(sections), removed_count);
// Verify no OOB
ensure(read_from_ptr<u32>(marker, range.length()) == overrun_cookie_value);
}
protected:
/**
* If render target already exists at address, issue state change operation on cmdList.
* Otherwise create one with width, height, clearColor info.
* returns the corresponding render target resource.
*/
template <typename ...Args>
surface_type bind_address_as_render_targets(
command_list_type command_list,
u32 address,
surface_color_format color_format,
surface_antialiasing antialias,
usz width, usz height, usz pitch,
Args&&... extra_params)
{
return bind_surface_address<false>(
command_list, address, color_format, antialias,
width, height, pitch, get_format_block_size_in_bytes(color_format),
std::forward<Args>(extra_params)...);
}
template <typename ...Args>
surface_type bind_address_as_depth_stencil(
command_list_type command_list,
u32 address,
surface_depth_format2 depth_format,
surface_antialiasing antialias,
usz width, usz height, usz pitch,
Args&&... extra_params)
{
return bind_surface_address<true>(
command_list, address, depth_format, antialias,
width, height, pitch,
get_format_block_size_in_bytes(depth_format),
std::forward<Args>(extra_params)...);
}
std::tuple<std::vector<surface_type>, std::vector<surface_type>>
find_overlapping_set(const utils::address_range32& range) const
{
std::vector<surface_type> color_result, depth_result;
utils::address_range32 result_range;
if (m_render_targets_memory_range.valid() &&
range.overlaps(m_render_targets_memory_range))
{
for (auto it = m_render_targets_storage.begin_range(range); it != m_render_targets_storage.end(); ++it)
{
auto surface = Traits::get(it->second);
const auto surface_range = surface->get_memory_range();
if (!range.overlaps(surface_range))
continue;
color_result.push_back(surface);
}
}
if (m_depth_stencil_memory_range.valid() &&
range.overlaps(m_depth_stencil_memory_range))
{
for (auto it = m_depth_stencil_storage.begin_range(range); it != m_depth_stencil_storage.end(); ++it)
{
auto surface = Traits::get(it->second);
const auto surface_range = surface->get_memory_range();
if (!range.overlaps(surface_range))
continue;
depth_result.push_back(surface);
}
}
return { color_result, depth_result, result_range };
}
void write_to_dma_buffers(
command_list_type command_list,
const utils::address_range32& range)
{
auto block_range = m_dma_block.to_block_range(range);
auto [color_data, depth_stencil_data] = find_overlapping_set(block_range);
auto [bo, offset, bo_timestamp] = m_dma_block
.with_range(command_list, block_range)
.get(block_range.start);
u64 src_offset, dst_offset, write_length;
auto block_length = block_range.length();
auto all_data = std::move(color_data);
all_data.insert(all_data.end(), depth_stencil_data.begin(), depth_stencil_data.end());
if (all_data.size() > 1)
{
std::sort(all_data.begin(), all_data.end(), [](const auto& a, const auto& b)
{
return a->last_use_tag < b->last_use_tag;
});
}
for (const auto& surface : all_data)
{
if (surface->last_use_tag <= bo_timestamp)
{
continue;
}
const auto this_range = surface->get_memory_range();
const auto max_length = this_range.length();
if (this_range.start < block_range.start)
{
src_offset = block_range.start - this_range.start;
dst_offset = 0;
}
else
{
src_offset = 0;
dst_offset = this_range.start - block_range.start;
}
write_length = std::min(max_length, block_length - dst_offset);
Traits::write_render_target_to_memory(command_list, bo, surface, dst_offset, src_offset, write_length);
}
m_dma_block.touch(block_range);
}
public:
/**
* Update bound color and depth surface.
* Must be called everytime surface format, clip, or addresses changes.
*/
template <typename ...Args>
void prepare_render_target(
command_list_type command_list,
surface_color_format color_format, surface_depth_format2 depth_format,
u32 clip_horizontal_reg, u32 clip_vertical_reg,
surface_target set_surface_target,
surface_antialiasing antialias,
surface_raster_type raster_type,
const std::array<u32, 4> &surface_addresses, u32 address_z,
const std::array<u32, 4> &surface_pitch, u32 zeta_pitch,
Args&&... extra_params)
{
u32 clip_width = clip_horizontal_reg;
u32 clip_height = clip_vertical_reg;
cache_tag = rsx::get_shared_tag();
m_invalidate_on_write = (antialias != rsx::surface_antialiasing::center_1_sample);
m_active_raster_type = raster_type;
// Make previous RTTs sampleable
for (const auto& i : m_bound_render_target_ids)
{
auto &rtt = m_bound_render_targets[i];
Traits::prepare_surface_for_sampling(command_list, std::get<1>(rtt));
rtt = std::make_pair(0, nullptr);
}
m_bound_render_target_ids.clear();
if (const auto rtt_indices = utility::get_rtt_indexes(set_surface_target);
!rtt_indices.empty()) [[likely]]
{
// Create/Reuse requested rtts
for (u8 surface_index : rtt_indices)
{
if (surface_addresses[surface_index] == 0)
continue;
m_bound_render_targets[surface_index] = std::make_pair(surface_addresses[surface_index],
bind_address_as_render_targets(command_list, surface_addresses[surface_index], color_format, antialias,
clip_width, clip_height, surface_pitch[surface_index], std::forward<Args>(extra_params)...));
m_bound_render_target_ids.push_back(surface_index);
}
}
// Same for depth buffer
if (std::get<1>(m_bound_depth_stencil) != nullptr)
{
Traits::prepare_surface_for_sampling(command_list, std::get<1>(m_bound_depth_stencil));
}
if (address_z) [[likely]]
{
m_bound_depth_stencil = std::make_pair(address_z,
bind_address_as_depth_stencil(command_list, address_z, depth_format, antialias,
clip_width, clip_height, zeta_pitch, std::forward<Args>(extra_params)...));
}
else
{
m_bound_depth_stencil = std::make_pair(0, nullptr);
}
}
u8 get_color_surface_count() const
{
return static_cast<u8>(m_bound_render_target_ids.size());
}
surface_type get_surface_at(u32 address)
{
auto It = m_render_targets_storage.find(address);
if (It != m_render_targets_storage.end())
return Traits::get(It->second);
auto _It = m_depth_stencil_storage.find(address);
if (_It != m_depth_stencil_storage.end())
return Traits::get(_It->second);
return nullptr;
}
/**
* Invalidates surface that exists at an address
*/
void invalidate_surface_address(u32 addr, bool depth)
{
if (address_is_bound(addr))
{
rsx_log.error("Cannot invalidate a currently bound render target!");
return;
}
if (!depth)
{
auto It = m_render_targets_storage.find(addr);
if (It != m_render_targets_storage.end())
{
invalidate(It->second);
m_render_targets_storage.erase(It);
return;
}
}
else
{
auto It = m_depth_stencil_storage.find(addr);
if (It != m_depth_stencil_storage.end())
{
invalidate(It->second);
m_depth_stencil_storage.erase(It);
return;
}
}
}
inline bool address_is_bound(u32 address) const
{
ensure(address);
for (int i = 0; i < 4; ++i)
{
if (m_bound_render_targets[i].first == address)
{
return true;
}
}
return (m_bound_depth_stencil.first == address);
}
template <typename commandbuffer_type>
std::vector<surface_overlap_info> get_merged_texture_memory_region(commandbuffer_type& cmd, u32 texaddr, u32 required_width, u32 required_height, u32 required_pitch, u8 required_bpp, rsx::surface_access access)
{
std::vector<surface_overlap_info> result;
std::vector<std::pair<u32, bool>> dirty;
const auto surface_internal_pitch = (required_width * required_bpp);
// Sanity check
if (surface_internal_pitch > required_pitch) [[unlikely]]
{
rsx_log.warning("Invalid 2D region descriptor. w=%d, h=%d, bpp=%d, pitch=%d",
required_width, required_height, required_bpp, required_pitch);
return {};
}
const auto test_range = utils::address_range32::start_length(texaddr, (required_pitch * required_height) - (required_pitch - surface_internal_pitch));
auto process_list_function = [&](surface_ranged_map& data, bool is_depth)
{
for (auto it = data.begin_range(test_range); it != data.end(); ++it)
{
const auto range = it->second->get_memory_range();
if (!range.overlaps(test_range))
continue;
auto surface = it->second.get();
if (access.is_transfer() && access.is_read() && surface->write_through())
{
// The surface has no data other than what can be loaded from CPU
continue;
}
if (!rsx::pitch_compatible(surface, required_pitch, required_height))
continue;
surface_overlap_info info;
u32 width, height;
info.surface = surface;
info.base_address = range.start;
info.is_depth = is_depth;
const u32 normalized_surface_width = surface->template get_surface_width<rsx::surface_metrics::bytes>() / required_bpp;
const u32 normalized_surface_height = surface->template get_surface_height<rsx::surface_metrics::samples>();
if (range.start >= texaddr) [[likely]]
{
const auto offset = range.start - texaddr;
info.dst_area.y = (offset / required_pitch);
info.dst_area.x = (offset % required_pitch) / required_bpp;
if (info.dst_area.x >= required_width || info.dst_area.y >= required_height) [[unlikely]]
{
// Out of bounds
continue;
}
info.src_area.x = 0;
info.src_area.y = 0;
width = std::min<u32>(normalized_surface_width, required_width - info.dst_area.x);
height = std::min<u32>(normalized_surface_height, required_height - info.dst_area.y);
}
else
{
const auto pitch = surface->get_rsx_pitch();
const auto offset = texaddr - range.start;
info.src_area.y = (offset / pitch);
info.src_area.x = (offset % pitch) / required_bpp;
if (info.src_area.x >= normalized_surface_width || info.src_area.y >= normalized_surface_height) [[unlikely]]
{
// Region lies outside the actual texture area, but inside the 'tile'
// In this case, a small region lies to the top-left corner, partially occupying the target
continue;
}
info.dst_area.x = 0;
info.dst_area.y = 0;
width = std::min<u32>(required_width, normalized_surface_width - info.src_area.x);
height = std::min<u32>(required_height, normalized_surface_height - info.src_area.y);
}
// Delay this as much as possible to avoid side-effects of spamming barrier
if (surface->memory_barrier(cmd, access); !surface->test())
{
dirty.emplace_back(range.start, is_depth);
continue;
}
info.is_clipped = (width < required_width || height < required_height);
info.src_area.height = info.dst_area.height = height;
info.dst_area.width = width;
if (auto surface_bpp = surface->get_bpp(); surface_bpp != required_bpp) [[unlikely]]
{
// Width is calculated in the coordinate-space of the requester; normalize
info.src_area.x = (info.src_area.x * required_bpp) / surface_bpp;
info.src_area.width = utils::align(width * required_bpp, surface_bpp) / surface_bpp;
}
else
{
info.src_area.width = width;
}
result.push_back(info);
}
};
// Range test helper to quickly discard blocks
// Fortunately, render targets tend to be clustered anyway
if (m_render_targets_memory_range.valid() &&
test_range.overlaps(m_render_targets_memory_range))
{
process_list_function(m_render_targets_storage, false);
}
if (m_depth_stencil_memory_range.valid() &&
test_range.overlaps(m_depth_stencil_memory_range))
{
process_list_function(m_depth_stencil_storage, true);
}
if (!dirty.empty())
{
for (const auto& p : dirty)
{
invalidate_surface_address(p.first, p.second);
}
}
if (result.size() > 1)
{
std::sort(result.begin(), result.end(), [](const auto &a, const auto &b)
{
if (a.surface->last_use_tag == b.surface->last_use_tag)
{
const auto area_a = a.dst_area.width * a.dst_area.height;
const auto area_b = b.dst_area.width * b.dst_area.height;
return area_a < area_b;
}
return a.surface->last_use_tag < b.surface->last_use_tag;
});
}
return result;
}
void check_for_duplicates(std::vector<surface_overlap_info>& sections)
{
utils::address_range32 test_range;
for (const auto& section : sections)
{
const auto range = section.surface->get_memory_range();
test_range.start = std::min(test_range.start, range.start);
test_range.end = std::max(test_range.end, range.end);
}
if (!remove_duplicates_fast_impl(sections, test_range))
{
remove_duplicates_fallback_impl(sections, test_range);
}
}
void on_write(const std::array<bool, 4>& color_mrt_writes_enabled, const bool depth_stencil_writes_enabled)
{
if (write_tag >= cache_tag && !m_invalidate_on_write)
{
return;
}
// TODO: Take WCB/WDB into account. Should speed this up a bit by skipping sync_tag calls
write_tag = rsx::get_shared_tag();
for (const auto& i : m_bound_render_target_ids)
{
if (color_mrt_writes_enabled[i])
{
auto surface = m_bound_render_targets[i].second;
if (surface->last_use_tag > cache_tag) [[ likely ]]
{
surface->on_write_fast(write_tag);
}
else
{
surface->on_write(write_tag, rsx::surface_state_flags::require_resolve, m_active_raster_type);
}
}
}
if (auto zsurface = m_bound_depth_stencil.second;
zsurface && depth_stencil_writes_enabled)
{
if (zsurface->last_use_tag > cache_tag) [[ likely ]]
{
zsurface->on_write_fast(write_tag);
}
else
{
zsurface->on_write(write_tag, rsx::surface_state_flags::require_resolve, m_active_raster_type);
}
}
}
void invalidate_all()
{
// Unbind and invalidate all resources
auto free_resource_list = [&](auto &data, const utils::address_range32& range)
{
for (auto it = data.begin_range(range); it != data.end(); ++it)
{
invalidate(it->second);
}
data.clear();
};
free_resource_list(m_render_targets_storage, m_render_targets_memory_range);
free_resource_list(m_depth_stencil_storage, m_depth_stencil_memory_range);
ensure(m_active_memory_used == 0);
m_bound_depth_stencil = std::make_pair(0, nullptr);
m_bound_render_target_ids.clear();
for (auto &rtt : m_bound_render_targets)
{
rtt = std::make_pair(0, nullptr);
}
}
void invalidate_range(const rsx::address_range32& range)
{
for (auto it = m_render_targets_storage.begin_range(range); it != m_render_targets_storage.end(); ++it)
{
auto& rtt = it->second;
if (range.overlaps(rtt->get_memory_range()))
{
rtt->clear_rw_barrier();
rtt->state_flags |= rsx::surface_state_flags::erase_bkgnd;
}
}
for (auto it = m_depth_stencil_storage.begin_range(range); it != m_depth_stencil_storage.end(); ++it)
{
auto& ds = it->second;
if (range.overlaps(ds->get_memory_range()))
{
ds->clear_rw_barrier();
ds->state_flags |= rsx::surface_state_flags::erase_bkgnd;
}
}
}
bool check_memory_usage(u64 max_safe_memory) const
{
if (m_active_memory_used <= max_safe_memory) [[likely]]
{
return false;
}
else if (m_active_memory_used > (max_safe_memory * 3) / 2)
{
rsx_log.warning("Surface cache is using too much memory! (%dM)", m_active_memory_used / 0x100000);
}
else
{
rsx_log.trace("Surface cache is using too much memory! (%dM)", m_active_memory_used / 0x100000);
}
return true;
}
virtual bool can_collapse_surface(const surface_storage_type&, problem_severity)
{
return true;
}
void trim_invalidated_resources(command_list_type cmd, problem_severity severity)
{
// It is possible to have stale invalidated resources holding references to other invalidated resources.
// This can bloat the VRAM usage significantly especially if the references are never collapsed.
for (auto& surface : invalidated_resources)
{
if (!surface->has_refs() || surface->old_contents.empty())
{
continue;
}
if (can_collapse_surface(surface, severity))
{
surface->memory_barrier(cmd, rsx::surface_access::transfer_read);
}
}
}
void collapse_dirty_surfaces(command_list_type cmd, problem_severity severity)
{
auto process_list_function = [&](surface_ranged_map& data, const utils::address_range32& range)
{
for (auto It = data.begin_range(range); It != data.end();)
{
auto surface = Traits::get(It->second);
if (surface->dirty())
{
// Force memory barrier to release some resources
if (can_collapse_surface(It->second, severity))
{
// NOTE: Do not call memory_barrier under fatal conditions as it can create allocations!
// It would be safer to leave the resources hanging around and spill them instead
surface->memory_barrier(cmd, rsx::surface_access::memory_read);
}
}
else if (!surface->test())
{
// Remove this
invalidate(It->second);
It = data.erase(It);
continue;
}
++It;
}
};
process_list_function(m_render_targets_storage, m_render_targets_memory_range);
process_list_function(m_depth_stencil_storage, m_depth_stencil_memory_range);
}
virtual bool handle_memory_pressure(command_list_type cmd, problem_severity severity)
{
ensure(severity >= rsx::problem_severity::moderate);
const auto old_usage = m_active_memory_used;
// Try and find old surfaces to remove
collapse_dirty_surfaces(cmd, severity);
// Check invalidated resources as they can have long dependency chains
if (invalidated_resources.size() > max_invalidated_resources_count ||
severity >= rsx::problem_severity::severe)
{
trim_invalidated_resources(cmd, severity);
}
return (m_active_memory_used < old_usage);
}
void run_cleanup_internal(
command_list_type cmd,
rsx::problem_severity memory_pressure,
u32 max_surface_store_memory_mb,
std::function<void(command_list_type)> pre_task_callback)
{
if (check_memory_usage(max_surface_store_memory_mb * 0x100000))
{
pre_task_callback(cmd);
const auto severity = std::max(memory_pressure, rsx::problem_severity::moderate);
handle_memory_pressure(cmd, severity);
}
else if (invalidated_resources.size() > max_invalidated_resources_count)
{
pre_task_callback(cmd);
rsx_log.warning("[PERFORMANCE WARNING] Invalidated resource pool has exceeded the desired limit. A trim will now be attempted. Current=%u, Limit=%u",
invalidated_resources.size(), max_invalidated_resources_count);
// Check invalidated resources as they can have long dependency chains
trim_invalidated_resources(cmd, memory_pressure);
if ((invalidated_resources.size() + 16u) > max_invalidated_resources_count)
{
// We didn't release enough resources, scan the active RTTs as well
collapse_dirty_surfaces(cmd, memory_pressure);
}
}
}
};
}