hdkv/yuzu
1
Fork 0
Code Issues Pull Requests Releases 1 Activity

early-access version 3552

Browse Source
This commit is contained in:
pineappleEA
2023-05-04 16:44:44 +02:00
parent 87bd2b0bdd
commit bab2c14888
48 changed files with 508 additions and 482 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
src/video_core/buffer_cache/buffer_base.h
View File

@@ -18,6 +18,7 @@ namespace VideoCommon {
enum class BufferFlagBits {
Picked = 1 << 0,
CachedWrites = 1 << 1,
PreemtiveDownload = 1 << 2,
};
DECLARE_ENUM_FLAG_OPERATORS(BufferFlagBits)
@@ -54,6 +55,10 @@ public:
flags |= BufferFlagBits::Picked;
}
void MarkPreemtiveDownload() noexcept {
flags |= BufferFlagBits::PreemtiveDownload;
}
/// Unmark buffer as picked
void Unpick() noexcept {
flags &= ~BufferFlagBits::Picked;
@@ -84,6 +89,10 @@ public:
return True(flags & BufferFlagBits::CachedWrites);
}
bool IsPreemtiveDownload() const noexcept {
return True(flags & BufferFlagBits::PreemtiveDownload);
}
/// Returns the base CPU address of the buffer
[[nodiscard]] VAddr CpuAddr() const noexcept {
return cpu_addr;

152
src/video_core/buffer_cache/buffer_cache.h
View File

@@ -23,8 +23,6 @@ BufferCache<P>::BufferCache(VideoCore::RasterizerInterface& rasterizer_,
common_ranges.clear();
inline_buffer_id = NULL_BUFFER_ID;
active_async_buffers = !Settings::IsGPULevelHigh();
if (!runtime.CanReportMemoryUsage()) {
minimum_memory = DEFAULT_EXPECTED_MEMORY;
critical_memory = DEFAULT_CRITICAL_MEMORY;
@@ -75,8 +73,6 @@ void BufferCache<P>::TickFrame() {
uniform_cache_hits[0] = 0;
uniform_cache_shots[0] = 0;
active_async_buffers = !Settings::IsGPULevelHigh();
const bool skip_preferred = hits * 256 < shots * 251;
uniform_buffer_skip_cache_size = skip_preferred ? DEFAULT_SKIP_CACHE_SIZE : 0;
@@ -111,9 +107,25 @@ void BufferCache<P>::WriteMemory(VAddr cpu_addr, u64 size) {
template <class P>
void BufferCache<P>::CachedWriteMemory(VAddr cpu_addr, u64 size) {
memory_tracker.CachedCpuWrite(cpu_addr, size);
const IntervalType add_interval{Common::AlignDown(cpu_addr, YUZU_PAGESIZE),
Common::AlignUp(cpu_addr + size, YUZU_PAGESIZE)};
cached_ranges.add(add_interval);
}
template <class P>
std::optional<VideoCore::RasterizerDownloadArea> BufferCache<P>::GetFlushArea(VAddr cpu_addr,
u64 size) {
std::optional<VideoCore::RasterizerDownloadArea> area{};
area.emplace();
VAddr cpu_addr_start_aligned = Common::AlignDown(cpu_addr, Core::Memory::YUZU_PAGESIZE);
VAddr cpu_addr_end_aligned = Common::AlignUp(cpu_addr + size, Core::Memory::YUZU_PAGESIZE);
area->start_address = cpu_addr_start_aligned;
area->end_address = cpu_addr_end_aligned;
if (memory_tracker.IsRegionPreflushable(cpu_addr, size)) {
area->preemtive = true;
return area;
};
memory_tracker.MarkRegionAsPreflushable(cpu_addr_start_aligned,
cpu_addr_end_aligned - cpu_addr_start_aligned);
area->preemtive = !IsRegionGpuModified(cpu_addr, size);
return area;
}
template <class P>
@@ -191,8 +203,11 @@ bool BufferCache<P>::DMACopy(GPUVAddr src_address, GPUVAddr dest_address, u64 am
const VAddr new_base_address = *cpu_dest_address + diff;
const IntervalType add_interval{new_base_address, new_base_address + size};
tmp_intervals.push_back(add_interval);
uncommitted_ranges.add(add_interval);
pending_ranges.add(add_interval);
if (!Settings::values.use_reactive_flushing.GetValue() ||
memory_tracker.IsRegionPreflushable(new_base_address, new_base_address + size)) {
uncommitted_ranges.add(add_interval);
pending_ranges.add(add_interval);
}
};
ForEachInRangeSet(common_ranges, *cpu_src_address, amount, mirror);
// This subtraction in this order is important for overlapping copies.
@@ -205,7 +220,7 @@ bool BufferCache<P>::DMACopy(GPUVAddr src_address, GPUVAddr dest_address, u64 am
if (has_new_downloads) {
memory_tracker.MarkRegionAsGpuModified(*cpu_dest_address, amount);
}
std::vector<u8> tmp_buffer(amount);
tmp_buffer.resize(amount);
cpu_memory.ReadBlockUnsafe(*cpu_src_address, tmp_buffer.data(), amount);
cpu_memory.WriteBlockUnsafe(*cpu_dest_address, tmp_buffer.data(), amount);
return true;
@@ -441,9 +456,7 @@ void BufferCache<P>::BindComputeTextureBuffer(size_t tbo_index, GPUVAddr gpu_add
template <class P>
void BufferCache<P>::FlushCachedWrites() {
cached_write_buffer_ids.clear();
memory_tracker.FlushCachedWrites();
cached_ranges.clear();
}
template <class P>
@@ -474,9 +487,8 @@ void BufferCache<P>::CommitAsyncFlushesHigh() {
if (committed_ranges.empty()) {
if constexpr (IMPLEMENTS_ASYNC_DOWNLOADS) {
if (active_async_buffers) {
async_buffers.emplace_back(std::optional<Async_Buffer>{});
}
async_buffers.emplace_back(std::optional<Async_Buffer>{});
}
return;
}
@@ -537,64 +549,65 @@ void BufferCache<P>::CommitAsyncFlushesHigh() {
committed_ranges.clear();
if (downloads.empty()) {
if constexpr (IMPLEMENTS_ASYNC_DOWNLOADS) {
if (active_async_buffers) {
async_buffers.emplace_back(std::optional<Async_Buffer>{});
}
async_buffers.emplace_back(std::optional<Async_Buffer>{});
}
return;
}
if (active_async_buffers) {
if constexpr (IMPLEMENTS_ASYNC_DOWNLOADS) {
auto download_staging = runtime.DownloadStagingBuffer(total_size_bytes, true);
boost::container::small_vector<BufferCopy, 4> normalized_copies;
IntervalSet new_async_range{};
runtime.PreCopyBarrier();
for (auto& [copy, buffer_id] : downloads) {
copy.dst_offset += download_staging.offset;
const std::array copies{copy};
BufferCopy second_copy{copy};
Buffer& buffer = slot_buffers[buffer_id];
second_copy.src_offset = static_cast<size_t>(buffer.CpuAddr()) + copy.src_offset;
VAddr orig_cpu_addr = static_cast<VAddr>(second_copy.src_offset);
const IntervalType base_interval{orig_cpu_addr, orig_cpu_addr + copy.size};
async_downloads += std::make_pair(base_interval, 1);
runtime.CopyBuffer(download_staging.buffer, buffer, copies, false);
normalized_copies.push_back(second_copy);
}
runtime.PostCopyBarrier();
pending_downloads.emplace_back(std::move(normalized_copies));
async_buffers.emplace_back(download_staging);
} else {
if constexpr (IMPLEMENTS_ASYNC_DOWNLOADS) {
auto download_staging = runtime.DownloadStagingBuffer(total_size_bytes, true);
boost::container::small_vector<BufferCopy, 4> normalized_copies;
IntervalSet new_async_range{};
runtime.PreCopyBarrier();
for (auto& [copy, buffer_id] : downloads) {
copy.dst_offset += download_staging.offset;
const std::array copies{copy};
BufferCopy second_copy{copy};
Buffer& buffer = slot_buffers[buffer_id];
second_copy.src_offset = static_cast<size_t>(buffer.CpuAddr()) + copy.src_offset;
VAddr orig_cpu_addr = static_cast<VAddr>(second_copy.src_offset);
const IntervalType base_interval{orig_cpu_addr, orig_cpu_addr + copy.size};
async_downloads += std::make_pair(base_interval, 1);
runtime.CopyBuffer(download_staging.buffer, buffer, copies, false);
normalized_copies.push_back(second_copy);
}
runtime.PostCopyBarrier();
pending_downloads.emplace_back(std::move(normalized_copies));
async_buffers.emplace_back(download_staging);
} else {
if (!Settings::IsGPULevelHigh()) {
committed_ranges.clear();
uncommitted_ranges.clear();
}
} else {
if constexpr (USE_MEMORY_MAPS) {
auto download_staging = runtime.DownloadStagingBuffer(total_size_bytes);
runtime.PreCopyBarrier();
for (auto& [copy, buffer_id] : downloads) {
// Have in mind the staging buffer offset for the copy
copy.dst_offset += download_staging.offset;
const std::array copies{copy};
runtime.CopyBuffer(download_staging.buffer, slot_buffers[buffer_id], copies, false);
}
runtime.PostCopyBarrier();
runtime.Finish();
for (const auto& [copy, buffer_id] : downloads) {
const Buffer& buffer = slot_buffers[buffer_id];
const VAddr cpu_addr = buffer.CpuAddr() + copy.src_offset;
// Undo the modified offset
const u64 dst_offset = copy.dst_offset - download_staging.offset;
const u8* read_mapped_memory = download_staging.mapped_span.data() + dst_offset;
cpu_memory.WriteBlockUnsafe(cpu_addr, read_mapped_memory, copy.size);
}
} else {
const std::span<u8> immediate_buffer = ImmediateBuffer(largest_copy);
for (const auto& [copy, buffer_id] : downloads) {
Buffer& buffer = slot_buffers[buffer_id];
buffer.ImmediateDownload(copy.src_offset, immediate_buffer.subspan(0, copy.size));
const VAddr cpu_addr = buffer.CpuAddr() + copy.src_offset;
cpu_memory.WriteBlockUnsafe(cpu_addr, immediate_buffer.data(), copy.size);
if constexpr (USE_MEMORY_MAPS) {
auto download_staging = runtime.DownloadStagingBuffer(total_size_bytes);
runtime.PreCopyBarrier();
for (auto& [copy, buffer_id] : downloads) {
// Have in mind the staging buffer offset for the copy
copy.dst_offset += download_staging.offset;
const std::array copies{copy};
runtime.CopyBuffer(download_staging.buffer, slot_buffers[buffer_id], copies,
false);
}
runtime.PostCopyBarrier();
runtime.Finish();
for (const auto& [copy, buffer_id] : downloads) {
const Buffer& buffer = slot_buffers[buffer_id];
const VAddr cpu_addr = buffer.CpuAddr() + copy.src_offset;
// Undo the modified offset
const u64 dst_offset = copy.dst_offset - download_staging.offset;
const u8* read_mapped_memory = download_staging.mapped_span.data() + dst_offset;
cpu_memory.WriteBlockUnsafe(cpu_addr, read_mapped_memory, copy.size);
}
} else {
const std::span<u8> immediate_buffer = ImmediateBuffer(largest_copy);
for (const auto& [copy, buffer_id] : downloads) {
Buffer& buffer = slot_buffers[buffer_id];
buffer.ImmediateDownload(copy.src_offset,
immediate_buffer.subspan(0, copy.size));
const VAddr cpu_addr = buffer.CpuAddr() + copy.src_offset;
cpu_memory.WriteBlockUnsafe(cpu_addr, immediate_buffer.data(), copy.size);
}
}
}
}
@@ -1221,6 +1234,10 @@ void BufferCache<P>::MarkWrittenBuffer(BufferId buffer_id, VAddr cpu_addr, u32 s
const IntervalType base_interval{cpu_addr, cpu_addr + size};
common_ranges.add(base_interval);
if (Settings::values.use_reactive_flushing.GetValue() &&
!memory_tracker.IsRegionPreflushable(cpu_addr, cpu_addr + size)) {
return;
}
uncommitted_ranges.add(base_interval);
pending_ranges.add(base_interval);
}
@@ -1629,7 +1646,6 @@ void BufferCache<P>::DeleteBuffer(BufferId buffer_id, bool do_not_mark) {
replace(transform_feedback_buffers);
replace(compute_uniform_buffers);
replace(compute_storage_buffers);
std::erase(cached_write_buffer_ids, buffer_id);
// Mark the whole buffer as CPU written to stop tracking CPU writes
if (!do_not_mark) {

7
src/video_core/buffer_cache/buffer_cache_base.h
View File

@@ -188,6 +188,8 @@ public:
void DownloadMemory(VAddr cpu_addr, u64 size);
std::optional<VideoCore::RasterizerDownloadArea> GetFlushArea(VAddr cpu_addr, u64 size);
bool InlineMemory(VAddr dest_address, size_t copy_size, std::span<const u8> inlined_buffer);
void BindGraphicsUniformBuffer(size_t stage, u32 index, GPUVAddr gpu_addr, u32 size);
@@ -541,8 +543,6 @@ private:
std::array<std::array<u32, NUM_GRAPHICS_UNIFORM_BUFFERS>, NUM_STAGES>, Empty>
uniform_buffer_binding_sizes{};
std::vector<BufferId> cached_write_buffer_ids;
MemoryTracker memory_tracker;
IntervalSet uncommitted_ranges;
IntervalSet common_ranges;
@@ -572,9 +572,8 @@ private:
u64 critical_memory = 0;
BufferId inline_buffer_id;
bool active_async_buffers = false;
std::array<BufferId, ((1ULL << 39) >> CACHING_PAGEBITS)> page_table;
std::vector<u8> tmp_buffer;
};
} // namespace VideoCommon

26
src/video_core/buffer_cache/memory_tracker_base.h
View File

@@ -66,6 +66,14 @@ public:
});
}
/// Returns true if a region has been marked as Preflushable
[[nodiscard]] bool IsRegionPreflushable(VAddr query_cpu_addr, u64 query_size) noexcept {
return IteratePages<false>(
query_cpu_addr, query_size, [](Manager* manager, u64 offset, size_t size) {
return manager->template IsRegionModified<Type::Preflushable>(offset, size);
});
}
/// Mark region as CPU modified, notifying the rasterizer about this change
void MarkRegionAsCpuModified(VAddr dirty_cpu_addr, u64 query_size) {
IteratePages<true>(dirty_cpu_addr, query_size,
@@ -93,6 +101,15 @@ public:
});
}
/// Mark region as modified from the host GPU
void MarkRegionAsPreflushable(VAddr dirty_cpu_addr, u64 query_size) noexcept {
IteratePages<true>(dirty_cpu_addr, query_size,
[](Manager* manager, u64 offset, size_t size) {
manager->template ChangeRegionState<Type::Preflushable, true>(
manager->GetCpuAddr() + offset, size);
});
}
/// Unmark region as modified from the host GPU
void UnmarkRegionAsGpuModified(VAddr dirty_cpu_addr, u64 query_size) noexcept {
IteratePages<true>(dirty_cpu_addr, query_size,
@@ -102,6 +119,15 @@ public:
});
}
/// Unmark region as modified from the host GPU
void UnmarkRegionAsPreflushable(VAddr dirty_cpu_addr, u64 query_size) noexcept {
IteratePages<true>(dirty_cpu_addr, query_size,
[](Manager* manager, u64 offset, size_t size) {
manager->template ChangeRegionState<Type::Preflushable, false>(
manager->GetCpuAddr() + offset, size);
});
}
/// Mark region as modified from the CPU
/// but don't mark it as modified until FlusHCachedWrites is called.
void CachedCpuWrite(VAddr dirty_cpu_addr, u64 query_size) {

27
src/video_core/buffer_cache/word_manager.h
View File

@@ -26,6 +26,7 @@ enum class Type {
GPU,
CachedCPU,
Untracked,
Preflushable,
};
/// Vector tracking modified pages tightly packed with small vector optimization
@@ -55,17 +56,20 @@ struct Words {
gpu.stack.fill(0);
cached_cpu.stack.fill(0);
untracked.stack.fill(~u64{0});
preflushable.stack.fill(0);
} else {
// Share allocation between CPU and GPU pages and set their default values
u64* const alloc = new u64[num_words * 4];
u64* const alloc = new u64[num_words * 5];
cpu.heap = alloc;
gpu.heap = alloc + num_words;
cached_cpu.heap = alloc + num_words * 2;
untracked.heap = alloc + num_words * 3;
preflushable.heap = alloc + num_words * 4;
std::fill_n(cpu.heap, num_words, ~u64{0});
std::fill_n(gpu.heap, num_words, 0);
std::fill_n(cached_cpu.heap, num_words, 0);
std::fill_n(untracked.heap, num_words, ~u64{0});
std::fill_n(preflushable.heap, num_words, 0);
}
// Clean up tailing bits
const u64 last_word_size = size_bytes % BYTES_PER_WORD;
@@ -88,13 +92,14 @@ struct Words {
gpu = rhs.gpu;
cached_cpu = rhs.cached_cpu;
untracked = rhs.untracked;
preflushable = rhs.preflushable;
rhs.cpu.heap = nullptr;
return *this;
}
Words(Words&& rhs) noexcept
: size_bytes{rhs.size_bytes}, num_words{rhs.num_words}, cpu{rhs.cpu}, gpu{rhs.gpu},
cached_cpu{rhs.cached_cpu}, untracked{rhs.untracked} {
cached_cpu{rhs.cached_cpu}, untracked{rhs.untracked}, preflushable{rhs.preflushable} {
rhs.cpu.heap = nullptr;
}
@@ -129,6 +134,8 @@ struct Words {
return std::span<u64>(cached_cpu.Pointer(IsShort()), num_words);
} else if constexpr (type == Type::Untracked) {
return std::span<u64>(untracked.Pointer(IsShort()), num_words);
} else if constexpr (type == Type::Preflushable) {
return std::span<u64>(preflushable.Pointer(IsShort()), num_words);
}
}
@@ -142,6 +149,8 @@ struct Words {
return std::span<const u64>(cached_cpu.Pointer(IsShort()), num_words);
} else if constexpr (type == Type::Untracked) {
return std::span<const u64>(untracked.Pointer(IsShort()), num_words);
} else if constexpr (type == Type::Preflushable) {
return std::span<const u64>(preflushable.Pointer(IsShort()), num_words);
}
}
@@ -151,6 +160,7 @@ struct Words {
WordsArray<stack_words> gpu;
WordsArray<stack_words> cached_cpu;
WordsArray<stack_words> untracked;
WordsArray<stack_words> preflushable;
};
template <class RasterizerInterface, size_t stack_words = 1>
@@ -292,6 +302,9 @@ public:
(pending_pointer - pending_offset) * BYTES_PER_PAGE);
};
IterateWords(offset, size, [&](size_t index, u64 mask) {
if constexpr (type == Type::GPU) {
mask &= ~untracked_words[index];
}
const u64 word = state_words[index] & mask;
if constexpr (clear) {
if constexpr (type == Type::CPU || type == Type::CachedCPU) {
@@ -340,8 +353,13 @@ public:
static_assert(type != Type::Untracked);
const std::span<const u64> state_words = words.template Span<type>();
[[maybe_unused]] const std::span<const u64> untracked_words =
words.template Span<Type::Untracked>();
bool result = false;
IterateWords(offset, size, [&](size_t index, u64 mask) {
if constexpr (type == Type::GPU) {
mask &= ~untracked_words[index];
}
const u64 word = state_words[index] & mask;
if (word != 0) {
result = true;
@@ -362,9 +380,14 @@ public:
[[nodiscard]] std::pair<u64, u64> ModifiedRegion(u64 offset, u64 size) const noexcept {
static_assert(type != Type::Untracked);
const std::span<const u64> state_words = words.template Span<type>();
[[maybe_unused]] const std::span<const u64> untracked_words =
words.template Span<Type::Untracked>();
u64 begin = std::numeric_limits<u64>::max();
u64 end = 0;
IterateWords(offset, size, [&](size_t index, u64 mask) {
if constexpr (type == Type::GPU) {
mask &= ~untracked_words[index];
}
const u64 word = state_words[index] & mask;
if (word == 0) {
return;

8
src/video_core/engines/maxwell_dma.cpp
View File

@@ -223,7 +223,7 @@ void MaxwellDMA::CopyBlockLinearToPitch() {
write_buffer.resize_destructive(dst_size);
memory_manager.ReadBlock(src_operand.address, read_buffer.data(), src_size);
memory_manager.ReadBlockUnsafe(dst_operand.address, write_buffer.data(), dst_size);
memory_manager.ReadBlock(dst_operand.address, write_buffer.data(), dst_size);
UnswizzleSubrect(write_buffer, read_buffer, bytes_per_pixel, width, height, depth, x_offset,
src_params.origin.y, x_elements, regs.line_count, block_height, block_depth,
@@ -288,11 +288,7 @@ void MaxwellDMA::CopyPitchToBlockLinear() {
write_buffer.resize_destructive(dst_size);
memory_manager.ReadBlock(regs.offset_in, read_buffer.data(), src_size);
if (Settings::IsGPULevelExtreme()) {
memory_manager.ReadBlock(regs.offset_out, write_buffer.data(), dst_size);
} else {
memory_manager.ReadBlockUnsafe(regs.offset_out, write_buffer.data(), dst_size);
}
memory_manager.ReadBlock(regs.offset_out, write_buffer.data(), dst_size);
// If the input is linear and the output is tiled, swizzle the input and copy it over.
SwizzleSubrect(write_buffer, read_buffer, bytes_per_pixel, width, height, depth, x_offset,

5
src/video_core/fence_manager.h
View File

@@ -59,6 +59,11 @@ public:
buffer_cache.AccumulateFlushes();
}
void SignalReference() {
std::function<void()> do_nothing([] {});
SignalFence(std::move(do_nothing));
}
void SyncOperation(std::function<void()>&& func) {
uncommitted_operations.emplace_back(std::move(func));
}

19
src/video_core/gpu.cpp
View File

@@ -283,6 +283,21 @@ struct GPU::Impl {
gpu_thread.FlushRegion(addr, size);
}
VideoCore::RasterizerDownloadArea OnCPURead(VAddr addr, u64 size) {
auto raster_area = rasterizer->GetFlushArea(addr, size);
if (raster_area.preemtive) {
return raster_area;
}
raster_area.preemtive = true;
const u64 fence = RequestSyncOperation([this, &raster_area]() {
rasterizer->FlushRegion(raster_area.start_address,
raster_area.end_address - raster_area.start_address);
});
gpu_thread.TickGPU();
WaitForSyncOperation(fence);
return raster_area;
}
/// Notify rasterizer that any caches of the specified region should be invalidated
void InvalidateRegion(VAddr addr, u64 size) {
gpu_thread.InvalidateRegion(addr, size);
@@ -538,6 +553,10 @@ void GPU::SwapBuffers(const Tegra::FramebufferConfig* framebuffer) {
impl->SwapBuffers(framebuffer);
}
VideoCore::RasterizerDownloadArea GPU::OnCPURead(VAddr addr, u64 size) {
return impl->OnCPURead(addr, size);
}
void GPU::FlushRegion(VAddr addr, u64 size) {
impl->FlushRegion(addr, size);
}

4
src/video_core/gpu.h
View File

@@ -10,6 +10,7 @@
#include "core/hle/service/nvdrv/nvdata.h"
#include "video_core/cdma_pusher.h"
#include "video_core/framebuffer_config.h"
#include "video_core/rasterizer_download_area.h"
namespace Core {
class System;
@@ -240,6 +241,9 @@ public:
/// Swap buffers (render frame)
void SwapBuffers(const Tegra::FramebufferConfig* framebuffer);
/// Notify rasterizer that any caches of the specified region should be flushed to Switch memory
[[nodiscard]] VideoCore::RasterizerDownloadArea OnCPURead(VAddr addr, u64 size);
/// Notify rasterizer that any caches of the specified region should be flushed to Switch memory
void FlushRegion(VAddr addr, u64 size);

2
src/video_core/query_cache.h
View File

@@ -255,7 +255,6 @@ private:
if (!in_range(query)) {
continue;
}
rasterizer.UpdatePagesCachedCount(query.GetCpuAddr(), query.SizeInBytes(), -1);
AsyncJobId async_job_id = query.GetAsyncJob();
auto flush_result = query.Flush(async);
if (async_job_id == NULL_ASYNC_JOB_ID) {
@@ -273,7 +272,6 @@ private:
/// Registers the passed parameters as cached and returns a pointer to the stored cached query.
CachedQuery* Register(VideoCore::QueryType type, VAddr cpu_addr, u8* host_ptr, bool timestamp) {
rasterizer.UpdatePagesCachedCount(cpu_addr, CachedQuery::SizeInBytes(timestamp), 1);
const u64 page = static_cast<u64>(cpu_addr) >> YUZU_PAGEBITS;
return &cached_queries[page].emplace_back(static_cast<QueryCache&>(*this), type, cpu_addr,
host_ptr);

16
src/video_core/rasterizer_download_area.h Executable file
View File

@@ -0,0 +1,16 @@
// SPDX-FileCopyrightText: Copyright 2023 yuzu Emulator Project
// SPDX-License-Identifier: GPL-3.0-or-later
#pragma once
#include "common/common_types.h"
namespace VideoCore {
struct RasterizerDownloadArea {
VAddr start_address;
VAddr end_address;
bool preemtive;
};
} // namespace VideoCore

3
src/video_core/rasterizer_interface.h
View File

@@ -12,6 +12,7 @@
#include "video_core/cache_types.h"
#include "video_core/engines/fermi_2d.h"
#include "video_core/gpu.h"
#include "video_core/rasterizer_download_area.h"
namespace Tegra {
class MemoryManager;
@@ -95,6 +96,8 @@ public:
virtual bool MustFlushRegion(VAddr addr, u64 size,
VideoCommon::CacheType which = VideoCommon::CacheType::All) = 0;
virtual RasterizerDownloadArea GetFlushArea(VAddr addr, u64 size) = 0;
/// Notify rasterizer that any caches of the specified region should be invalidated
virtual void InvalidateRegion(VAddr addr, u64 size,
VideoCommon::CacheType which = VideoCommon::CacheType::All) = 0;

10
src/video_core/renderer_null/null_rasterizer.cpp
View File

@@ -1,6 +1,8 @@
// SPDX-FileCopyrightText: Copyright 2022 yuzu Emulator Project
// SPDX-License-Identifier: GPL-2.0-or-later
#include "common/alignment.h"
#include "core/memory.h"
#include "video_core/host1x/host1x.h"
#include "video_core/memory_manager.h"
#include "video_core/renderer_null/null_rasterizer.h"
@@ -46,6 +48,14 @@ bool RasterizerNull::MustFlushRegion(VAddr addr, u64 size, VideoCommon::CacheTyp
}
void RasterizerNull::InvalidateRegion(VAddr addr, u64 size, VideoCommon::CacheType) {}
void RasterizerNull::OnCPUWrite(VAddr addr, u64 size) {}
VideoCore::RasterizerDownloadArea RasterizerNull::GetFlushArea(VAddr addr, u64 size) {
VideoCore::RasterizerDownloadArea new_area{
.start_address = Common::AlignDown(addr, Core::Memory::YUZU_PAGESIZE),
.end_address = Common::AlignUp(addr + size, Core::Memory::YUZU_PAGESIZE),
.preemtive = true,
};
return new_area;
}
void RasterizerNull::InvalidateGPUCache() {}
void RasterizerNull::UnmapMemory(VAddr addr, u64 size) {}
void RasterizerNull::ModifyGPUMemory(size_t as_id, GPUVAddr addr, u64 size) {}

1
src/video_core/renderer_null/null_rasterizer.h
View File

@@ -54,6 +54,7 @@ public:
void InvalidateRegion(VAddr addr, u64 size,
VideoCommon::CacheType which = VideoCommon::CacheType::All) override;
void OnCPUWrite(VAddr addr, u64 size) override;
VideoCore::RasterizerDownloadArea GetFlushArea(VAddr addr, u64 size) override;
void InvalidateGPUCache() override;
void UnmapMemory(VAddr addr, u64 size) override;
void ModifyGPUMemory(size_t as_id, GPUVAddr addr, u64 size) override;

23
src/video_core/renderer_opengl/gl_rasterizer.cpp
View File

@@ -433,6 +433,29 @@ bool RasterizerOpenGL::MustFlushRegion(VAddr addr, u64 size, VideoCommon::CacheT
return false;
}
VideoCore::RasterizerDownloadArea RasterizerOpenGL::GetFlushArea(VAddr addr, u64 size) {
{
std::scoped_lock lock{texture_cache.mutex};
auto area = texture_cache.GetFlushArea(addr, size);
if (area) {
return *area;
}
}
{
std::scoped_lock lock{buffer_cache.mutex};
auto area = buffer_cache.GetFlushArea(addr, size);
if (area) {
return *area;
}
}
VideoCore::RasterizerDownloadArea new_area{
.start_address = Common::AlignDown(addr, Core::Memory::YUZU_PAGESIZE),
.end_address = Common::AlignUp(addr + size, Core::Memory::YUZU_PAGESIZE),
.preemtive = true,
};
return new_area;
}
void RasterizerOpenGL::InvalidateRegion(VAddr addr, u64 size, VideoCommon::CacheType which) {
MICROPROFILE_SCOPE(OpenGL_CacheManagement);
if (addr == 0 || size == 0) {

1
src/video_core/renderer_opengl/gl_rasterizer.h
View File

@@ -95,6 +95,7 @@ public:
VideoCommon::CacheType which = VideoCommon::CacheType::All) override;
bool MustFlushRegion(VAddr addr, u64 size,
VideoCommon::CacheType which = VideoCommon::CacheType::All) override;
VideoCore::RasterizerDownloadArea GetFlushArea(VAddr addr, u64 size) override;
void InvalidateRegion(VAddr addr, u64 size,
VideoCommon::CacheType which = VideoCommon::CacheType::All) override;
void OnCPUWrite(VAddr addr, u64 size) override;

2
src/video_core/renderer_vulkan/renderer_vulkan.cpp
View File

@@ -88,7 +88,7 @@ RendererVulkan::RendererVulkan(Core::TelemetrySession& telemetry_session_,
instance(CreateInstance(library, dld, VK_API_VERSION_1_1, render_window.GetWindowInfo().type,
Settings::values.renderer_debug.GetValue())),
debug_callback(Settings::values.renderer_debug ? CreateDebugCallback(instance) : nullptr),
surface(CreateSurface(instance, render_window.GetWindowInfo())),
surface(CreateSurface(instance, render_window)),
device(CreateDevice(instance, dld, *surface)), memory_allocator(device, false),
state_tracker(), scheduler(device, state_tracker),
swapchain(*surface, device, scheduler, render_window.GetFramebufferLayout().width,

27
src/video_core/renderer_vulkan/vk_rasterizer.cpp
View File

@@ -502,6 +502,29 @@ bool RasterizerVulkan::MustFlushRegion(VAddr addr, u64 size, VideoCommon::CacheT
return false;
}
VideoCore::RasterizerDownloadArea RasterizerVulkan::GetFlushArea(VAddr addr, u64 size) {
{
std::scoped_lock lock{texture_cache.mutex};
auto area = texture_cache.GetFlushArea(addr, size);
if (area) {
return *area;
}
}
{
std::scoped_lock lock{buffer_cache.mutex};
auto area = buffer_cache.GetFlushArea(addr, size);
if (area) {
return *area;
}
}
VideoCore::RasterizerDownloadArea new_area{
.start_address = Common::AlignDown(addr, Core::Memory::YUZU_PAGESIZE),
.end_address = Common::AlignUp(addr + size, Core::Memory::YUZU_PAGESIZE),
.preemtive = true,
};
return new_area;
}
void RasterizerVulkan::InvalidateRegion(VAddr addr, u64 size, VideoCommon::CacheType which) {
if (addr == 0 || size == 0) {
return;
@@ -598,7 +621,7 @@ void RasterizerVulkan::SignalSyncPoint(u32 value) {
}
void RasterizerVulkan::SignalReference() {
fence_manager.SignalOrdering();
fence_manager.SignalReference();
}
void RasterizerVulkan::ReleaseFences() {
@@ -631,7 +654,7 @@ void RasterizerVulkan::WaitForIdle() {
cmdbuf.SetEvent(event, flags);
cmdbuf.WaitEvents(event, flags, VK_PIPELINE_STAGE_TOP_OF_PIPE_BIT, {}, {}, {});
});
SignalReference();
fence_manager.SignalOrdering();
}
void RasterizerVulkan::FragmentBarrier() {

1
src/video_core/renderer_vulkan/vk_rasterizer.h
View File

@@ -92,6 +92,7 @@ public:
VideoCommon::CacheType which = VideoCommon::CacheType::All) override;
bool MustFlushRegion(VAddr addr, u64 size,
VideoCommon::CacheType which = VideoCommon::CacheType::All) override;
VideoCore::RasterizerDownloadArea GetFlushArea(VAddr addr, u64 size) override;
void InvalidateRegion(VAddr addr, u64 size,
VideoCommon::CacheType which = VideoCommon::CacheType::All) override;
void InnerInvalidation(std::span<const std::pair<VAddr, std::size_t>> sequences) override;

75
src/video_core/renderer_vulkan/vk_swapchain.cpp
View File

@@ -14,7 +14,6 @@
#include "video_core/renderer_vulkan/vk_swapchain.h"
#include "video_core/vulkan_common/vulkan_device.h"
#include "video_core/vulkan_common/vulkan_wrapper.h"
#include "vulkan/vulkan_core.h"
namespace Vulkan {
@@ -34,47 +33,23 @@ VkSurfaceFormatKHR ChooseSwapSurfaceFormat(vk::Span<VkSurfaceFormatKHR> formats)
return found != formats.end() ? *found : formats[0];
}
static constexpr VkPresentModeKHR ChooseSwapPresentMode(bool has_imm, bool has_mailbox,
bool has_fifo_relaxed) {
// Mailbox doesn't lock the application like FIFO (vsync)
// FIFO present mode locks the framerate to the monitor's refresh rate
Settings::VSyncMode setting = [has_imm, has_mailbox]() {
// Choose Mailbox or Immediate if unlocked and those modes are supported
const auto mode = Settings::values.vsync_mode.GetValue();
if (Settings::values.use_speed_limit.GetValue()) {
return mode;
}
switch (mode) {
case Settings::VSyncMode::FIFO:
case Settings::VSyncMode::FIFORelaxed:
if (has_mailbox) {
return Settings::VSyncMode::Mailbox;
} else if (has_imm) {
return Settings::VSyncMode::Immediate;
}
[[fallthrough]];
default:
return mode;
}
}();
if ((setting == Settings::VSyncMode::Mailbox && !has_mailbox) ||
(setting == Settings::VSyncMode::Immediate && !has_imm) ||
(setting == Settings::VSyncMode::FIFORelaxed && !has_fifo_relaxed)) {
setting = Settings::VSyncMode::FIFO;
}
switch (setting) {
case Settings::VSyncMode::Immediate:
return VK_PRESENT_MODE_IMMEDIATE_KHR;
case Settings::VSyncMode::Mailbox:
VkPresentModeKHR ChooseSwapPresentMode(vk::Span<VkPresentModeKHR> modes) {
// Mailbox (triple buffering) doesn't lock the application like fifo (vsync),
// prefer it if vsync option is not selected
const auto found_mailbox = std::find(modes.begin(), modes.end(), VK_PRESENT_MODE_MAILBOX_KHR);
if (Settings::values.fullscreen_mode.GetValue() == Settings::FullscreenMode::Borderless &&
found_mailbox != modes.end() && !Settings::values.use_vsync.GetValue()) {
return VK_PRESENT_MODE_MAILBOX_KHR;
case Settings::VSyncMode::FIFO:
return VK_PRESENT_MODE_FIFO_KHR;
case Settings::VSyncMode::FIFORelaxed:
return VK_PRESENT_MODE_FIFO_RELAXED_KHR;
default:
return VK_PRESENT_MODE_FIFO_KHR;
}
if (!Settings::values.use_speed_limit.GetValue()) {
// FIFO present mode locks the framerate to the monitor's refresh rate,
// Find an alternative to surpass this limitation if FPS is unlocked.
const auto found_imm = std::find(modes.begin(), modes.end(), VK_PRESENT_MODE_IMMEDIATE_KHR);
if (found_imm != modes.end()) {
return VK_PRESENT_MODE_IMMEDIATE_KHR;
}
}
return VK_PRESENT_MODE_FIFO_KHR;
}
VkExtent2D ChooseSwapExtent(const VkSurfaceCapabilitiesKHR& capabilities, u32 width, u32 height) {
@@ -192,17 +167,11 @@ void Swapchain::Present(VkSemaphore render_semaphore) {
void Swapchain::CreateSwapchain(const VkSurfaceCapabilitiesKHR& capabilities, bool srgb) {
const auto physical_device{device.GetPhysical()};
const auto formats{physical_device.GetSurfaceFormatsKHR(surface)};
const auto present_modes = physical_device.GetSurfacePresentModesKHR(surface);
has_mailbox = std::find(present_modes.begin(), present_modes.end(),
VK_PRESENT_MODE_MAILBOX_KHR) != present_modes.end();
has_imm = std::find(present_modes.begin(), present_modes.end(),
VK_PRESENT_MODE_IMMEDIATE_KHR) != present_modes.end();
has_fifo_relaxed = std::find(present_modes.begin(), present_modes.end(),
VK_PRESENT_MODE_FIFO_RELAXED_KHR) != present_modes.end();
const auto present_modes{physical_device.GetSurfacePresentModesKHR(surface)};
const VkCompositeAlphaFlagBitsKHR alpha_flags{ChooseAlphaFlags(capabilities)};
surface_format = ChooseSwapSurfaceFormat(formats);
present_mode = ChooseSwapPresentMode(has_imm, has_mailbox, has_fifo_relaxed);
present_mode = ChooseSwapPresentMode(present_modes);
u32 requested_image_count{capabilities.minImageCount + 1};
// Ensure Triple buffering if possible.
@@ -263,6 +232,7 @@ void Swapchain::CreateSwapchain(const VkSurfaceCapabilitiesKHR& capabilities, bo
extent = swapchain_ci.imageExtent;
current_srgb = srgb;
current_fps_unlocked = !Settings::values.use_speed_limit.GetValue();
images = swapchain.GetImages();
image_count = static_cast<u32>(images.size());
@@ -284,9 +254,14 @@ void Swapchain::Destroy() {
swapchain.reset();
}
bool Swapchain::HasFpsUnlockChanged() const {
return current_fps_unlocked != !Settings::values.use_speed_limit.GetValue();
}
bool Swapchain::NeedsPresentModeUpdate() const {
const auto requested_mode = ChooseSwapPresentMode(has_imm, has_mailbox, has_fifo_relaxed);
return present_mode != requested_mode;
// Mailbox present mode is the ideal for all scenarios. If it is not available,
// A different present mode is needed to support unlocked FPS above the monitor's refresh rate.
return present_mode != VK_PRESENT_MODE_MAILBOX_KHR && HasFpsUnlockChanged();
}
} // namespace Vulkan

6
src/video_core/renderer_vulkan/vk_swapchain.h
View File

@@ -116,6 +116,8 @@ private:
void Destroy();
bool HasFpsUnlockChanged() const;
bool NeedsPresentModeUpdate() const;
const VkSurfaceKHR surface;
@@ -140,11 +142,9 @@ private:
VkExtent2D extent{};
VkPresentModeKHR present_mode{};
VkSurfaceFormatKHR surface_format{};
bool has_imm{false};
bool has_mailbox{false};
bool has_fifo_relaxed{false};
bool current_srgb{};
bool current_fps_unlocked{};
bool is_outdated{};
bool is_suboptimal{};
};

2
src/video_core/texture_cache/image_info.h
View File

@@ -39,6 +39,8 @@ struct ImageInfo {
u32 tile_width_spacing = 0;
bool rescaleable = false;
bool downscaleable = false;
bool forced_flushed = false;
bool dma_downloaded = false;
};
} // namespace VideoCommon

5
src/video_core/texture_cache/image_view_base.cpp
View File

@@ -26,8 +26,9 @@ ImageViewBase::ImageViewBase(const ImageViewInfo& info, const ImageInfo& image_i
ASSERT_MSG(VideoCore::Surface::IsViewCompatible(image_info.format, info.format, false, true),
"Image view format {} is incompatible with image format {}", info.format,
image_info.format);
const bool is_async = Settings::values.use_asynchronous_gpu_emulation.GetValue();
if (image_info.type == ImageType::Linear && is_async) {
const bool preemptive =
!Settings::values.use_reactive_flushing.GetValue() && image_info.type == ImageType::Linear;
if (image_info.forced_flushed || preemptive) {
flags |= ImageViewFlagBits::PreemtiveDownload;
}
if (image_info.type == ImageType::e3D && info.type != ImageViewType::e3D) {

60
src/video_core/texture_cache/texture_cache.h
View File

@@ -490,6 +490,32 @@ void TextureCache<P>::DownloadMemory(VAddr cpu_addr, size_t size) {
}
}
template <class P>
std::optional<VideoCore::RasterizerDownloadArea> TextureCache<P>::GetFlushArea(VAddr cpu_addr,
u64 size) {
std::optional<VideoCore::RasterizerDownloadArea> area{};
ForEachImageInRegion(cpu_addr, size, [&](ImageId, ImageBase& image) {
if (False(image.flags & ImageFlagBits::GpuModified)) {
return;
}
if (!area) {
area.emplace();
area->start_address = cpu_addr;
area->end_address = cpu_addr + size;
area->preemtive = true;
}
area->start_address = std::min(area->start_address, image.cpu_addr);
area->end_address = std::max(area->end_address, image.cpu_addr_end);
for (auto image_view_id : image.image_view_ids) {
auto& image_view = slot_image_views[image_view_id];
image_view.flags |= ImageViewFlagBits::PreemtiveDownload;
}
area->preemtive &= image.info.forced_flushed;
image.info.forced_flushed = true;
});
return area;
}
template <class P>
void TextureCache<P>::UnmapMemory(VAddr cpu_addr, size_t size) {
std::vector<ImageId> deleted_images;
@@ -683,18 +709,43 @@ void TextureCache<P>::CommitAsyncFlushes() {
download_info.async_buffer_id = last_async_buffer_id;
}
}
if (any_none_dma) {
bool all_pre_sync = true;
auto download_map = runtime.DownloadStagingBuffer(total_size_bytes, true);
for (const PendingDownload& download_info : download_ids) {
if (download_info.is_swizzle) {
Image& image = slot_images[download_info.object_id];
all_pre_sync &= image.info.dma_downloaded;
image.info.dma_downloaded = true;
const auto copies = FullDownloadCopies(image.info);
image.DownloadMemory(download_map, copies);
download_map.offset += Common::AlignUp(image.unswizzled_size_bytes, 64);
}
}
uncommitted_async_buffers.emplace_back(download_map);
if (!all_pre_sync) {
runtime.Finish();
auto it = download_ids.begin();
while (it != download_ids.end()) {
const PendingDownload& download_info = *it;
if (download_info.is_swizzle) {
const ImageBase& image = slot_images[download_info.object_id];
const auto copies = FullDownloadCopies(image.info);
download_map.offset -= Common::AlignUp(image.unswizzled_size_bytes, 64);
std::span<u8> download_span =
download_map.mapped_span.subspan(download_map.offset);
SwizzleImage(*gpu_memory, image.gpu_addr, image.info, copies, download_span,
swizzle_data_buffer);
it = download_ids.erase(it);
} else {
it++;
}
}
} else {
uncommitted_async_buffers.emplace_back(download_map);
}
}
async_buffers.emplace_back(std::move(uncommitted_async_buffers));
uncommitted_async_buffers.clear();
}
@@ -789,11 +840,16 @@ ImageId TextureCache<P>::DmaImageId(const Tegra::DMA::ImageOperand& operand) {
if (!dst_id) {
return NULL_IMAGE_ID;
}
const auto& image = slot_images[dst_id];
auto& image = slot_images[dst_id];
if (False(image.flags & ImageFlagBits::GpuModified)) {
// No need to waste time on an image that's synced with guest
return NULL_IMAGE_ID;
}
if (!image.info.dma_downloaded) {
// Force a full sync.
image.info.dma_downloaded = true;
return NULL_IMAGE_ID;
}
const auto base = image.TryFindBase(operand.address);
if (!base) {
return NULL_IMAGE_ID;

2
src/video_core/texture_cache/texture_cache_base.h
View File

@@ -179,6 +179,8 @@ public:
/// Download contents of host images to guest memory in a region
void DownloadMemory(VAddr cpu_addr, size_t size);
std::optional<VideoCore::RasterizerDownloadArea> GetFlushArea(VAddr cpu_addr, u64 size);
/// Remove images in a region
void UnmapMemory(VAddr cpu_addr, size_t size);

6
src/video_core/vulkan_common/vulkan_surface.cpp
View File

@@ -23,10 +23,10 @@
namespace Vulkan {
vk::SurfaceKHR CreateSurface(
const vk::Instance& instance,
[[maybe_unused]] const Core::Frontend::EmuWindow::WindowSystemInfo& window_info) {
vk::SurfaceKHR CreateSurface(const vk::Instance& instance,
const Core::Frontend::EmuWindow& emu_window) {
[[maybe_unused]] const vk::InstanceDispatch& dld = instance.Dispatch();
[[maybe_unused]] const auto& window_info = emu_window.GetWindowInfo();
VkSurfaceKHR unsafe_surface = nullptr;
#ifdef _WIN32

9
src/video_core/vulkan_common/vulkan_surface.h
View File

@@ -3,12 +3,15 @@
#pragma once
#include "core/frontend/emu_window.h"
#include "video_core/vulkan_common/vulkan_wrapper.h"
namespace Core::Frontend {
class EmuWindow;
}
namespace Vulkan {
[[nodiscard]] vk::SurfaceKHR CreateSurface(
const vk::Instance& instance, const Core::Frontend::EmuWindow::WindowSystemInfo& window_info);
[[nodiscard]] vk::SurfaceKHR CreateSurface(const vk::Instance& instance,
const Core::Frontend::EmuWindow& emu_window);
} // namespace Vulkan