From c6cac2ffaad4ac27f35cea25022d9c59c7ecfbf4 Mon Sep 17 00:00:00 2001 From: Fernando Sahmkow Date: Sun, 30 Apr 2023 17:14:06 +0200 Subject: GPU: Add Reactive flushing --- src/video_core/buffer_cache/buffer_base.h | 9 ++++++ src/video_core/buffer_cache/buffer_cache.h | 35 +++++++++++++++++------ src/video_core/buffer_cache/buffer_cache_base.h | 5 ++-- src/video_core/buffer_cache/memory_tracker_base.h | 26 +++++++++++++++++ src/video_core/buffer_cache/word_manager.h | 14 +++++++-- 5 files changed, 76 insertions(+), 13 deletions(-) (limited to 'src/video_core/buffer_cache') diff --git a/src/video_core/buffer_cache/buffer_base.h b/src/video_core/buffer_cache/buffer_base.h index 9cbd95c4b..0bb3bf8ae 100644 --- a/src/video_core/buffer_cache/buffer_base.h +++ b/src/video_core/buffer_cache/buffer_base.h @@ -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; diff --git a/src/video_core/buffer_cache/buffer_cache.h b/src/video_core/buffer_cache/buffer_cache.h index e534e1e9c..479a1a508 100644 --- a/src/video_core/buffer_cache/buffer_cache.h +++ b/src/video_core/buffer_cache/buffer_cache.h @@ -111,9 +111,24 @@ void BufferCache

::WriteMemory(VAddr cpu_addr, u64 size) { template void BufferCache

::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 +std::optional BufferCache

::GetFlushArea(VAddr cpu_addr, + u64 size) { + std::optional 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 @@ -191,8 +206,10 @@ bool BufferCache

::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 (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 +222,7 @@ bool BufferCache

::DMACopy(GPUVAddr src_address, GPUVAddr dest_address, u64 am if (has_new_downloads) { memory_tracker.MarkRegionAsGpuModified(*cpu_dest_address, amount); } - std::vector 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 +458,7 @@ void BufferCache

::BindComputeTextureBuffer(size_t tbo_index, GPUVAddr gpu_add template void BufferCache

::FlushCachedWrites() { - cached_write_buffer_ids.clear(); memory_tracker.FlushCachedWrites(); - cached_ranges.clear(); } template @@ -1221,6 +1236,9 @@ void BufferCache

::MarkWrittenBuffer(BufferId buffer_id, VAddr cpu_addr, u32 s const IntervalType base_interval{cpu_addr, cpu_addr + size}; common_ranges.add(base_interval); + if (!memory_tracker.IsRegionPreflushable(cpu_addr, cpu_addr + size)) { + return; + } uncommitted_ranges.add(base_interval); pending_ranges.add(base_interval); } @@ -1629,7 +1647,6 @@ void BufferCache

::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) { diff --git a/src/video_core/buffer_cache/buffer_cache_base.h b/src/video_core/buffer_cache/buffer_cache_base.h index 656baa550..e3914a53a 100644 --- a/src/video_core/buffer_cache/buffer_cache_base.h +++ b/src/video_core/buffer_cache/buffer_cache_base.h @@ -188,6 +188,8 @@ public: void DownloadMemory(VAddr cpu_addr, u64 size); + std::optional GetFlushArea(VAddr cpu_addr, u64 size); + bool InlineMemory(VAddr dest_address, size_t copy_size, std::span inlined_buffer); void BindGraphicsUniformBuffer(size_t stage, u32 index, GPUVAddr gpu_addr, u32 size); @@ -541,8 +543,6 @@ private: std::array, NUM_STAGES>, Empty> uniform_buffer_binding_sizes{}; - std::vector cached_write_buffer_ids; - MemoryTracker memory_tracker; IntervalSet uncommitted_ranges; IntervalSet common_ranges; @@ -575,6 +575,7 @@ private: bool active_async_buffers = false; std::array> CACHING_PAGEBITS)> page_table; + std::vector tmp_buffer; }; } // namespace VideoCommon diff --git a/src/video_core/buffer_cache/memory_tracker_base.h b/src/video_core/buffer_cache/memory_tracker_base.h index dc4ebfcaa..6036b21c9 100644 --- a/src/video_core/buffer_cache/memory_tracker_base.h +++ b/src/video_core/buffer_cache/memory_tracker_base.h @@ -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( + query_cpu_addr, query_size, [](Manager* manager, u64 offset, size_t size) { + return manager->template IsRegionModified(offset, size); + }); + } + /// Mark region as CPU modified, notifying the rasterizer about this change void MarkRegionAsCpuModified(VAddr dirty_cpu_addr, u64 query_size) { IteratePages(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(dirty_cpu_addr, query_size, + [](Manager* manager, u64 offset, size_t size) { + manager->template ChangeRegionState( + manager->GetCpuAddr() + offset, size); + }); + } + /// Unmark region as modified from the host GPU void UnmarkRegionAsGpuModified(VAddr dirty_cpu_addr, u64 query_size) noexcept { IteratePages(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(dirty_cpu_addr, query_size, + [](Manager* manager, u64 offset, size_t size) { + manager->template ChangeRegionState( + 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) { diff --git a/src/video_core/buffer_cache/word_manager.h b/src/video_core/buffer_cache/word_manager.h index a42455045..0fb199a54 100644 --- a/src/video_core/buffer_cache/word_manager.h +++ b/src/video_core/buffer_cache/word_manager.h @@ -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(cached_cpu.Pointer(IsShort()), num_words); } else if constexpr (type == Type::Untracked) { return std::span(untracked.Pointer(IsShort()), num_words); + } else if constexpr (type == Type::Preflushable) { + return std::span(preflushable.Pointer(IsShort()), num_words); } } @@ -142,6 +149,8 @@ struct Words { return std::span(cached_cpu.Pointer(IsShort()), num_words); } else if constexpr (type == Type::Untracked) { return std::span(untracked.Pointer(IsShort()), num_words); + } else if constexpr (type == Type::Preflushable) { + return std::span(preflushable.Pointer(IsShort()), num_words); } } @@ -151,6 +160,7 @@ struct Words { WordsArray gpu; WordsArray cached_cpu; WordsArray untracked; + WordsArray preflushable; }; template -- cgit v1.2.3 From 92da86290c5ea657ae918bfe36071bdf7ac15075 Mon Sep 17 00:00:00 2001 From: Fernando Sahmkow Date: Thu, 4 May 2023 02:34:49 +0200 Subject: Settings: add option to enable / disable reactive flushing --- src/common/settings.cpp | 2 ++ src/common/settings.h | 1 + src/core/memory.cpp | 3 ++- src/video_core/buffer_cache/buffer_cache.h | 9 ++++++--- src/video_core/texture_cache/image_view_base.cpp | 3 ++- src/yuzu/configuration/config.cpp | 2 ++ src/yuzu/configuration/configure_graphics_advanced.cpp | 7 +++++++ src/yuzu/configuration/configure_graphics_advanced.h | 1 + src/yuzu/configuration/configure_graphics_advanced.ui | 10 ++++++++++ src/yuzu_cmd/config.cpp | 1 + src/yuzu_cmd/default_ini.h | 4 ++++ 11 files changed, 38 insertions(+), 5 deletions(-) (limited to 'src/video_core/buffer_cache') diff --git a/src/common/settings.cpp b/src/common/settings.cpp index f1ee42ab2..db1774c71 100644 --- a/src/common/settings.cpp +++ b/src/common/settings.cpp @@ -62,6 +62,7 @@ void LogSettings() { log_setting("Renderer_AccelerateASTC", values.accelerate_astc.GetValue()); log_setting("Renderer_AsyncASTC", values.async_astc.GetValue()); log_setting("Renderer_UseVsync", values.vsync_mode.GetValue()); + log_setting("Renderer_UseReactiveFlushing", values.use_reactive_flushing.GetValue()); log_setting("Renderer_ShaderBackend", values.shader_backend.GetValue()); log_setting("Renderer_UseAsynchronousShaders", values.use_asynchronous_shaders.GetValue()); log_setting("Renderer_AnisotropicFilteringLevel", values.max_anisotropy.GetValue()); @@ -223,6 +224,7 @@ void RestoreGlobalState(bool is_powered_on) { values.nvdec_emulation.SetGlobal(true); values.accelerate_astc.SetGlobal(true); values.async_astc.SetGlobal(true); + values.use_reactive_flushing.SetGlobal(true); values.shader_backend.SetGlobal(true); values.use_asynchronous_shaders.SetGlobal(true); values.use_fast_gpu_time.SetGlobal(true); diff --git a/src/common/settings.h b/src/common/settings.h index 2bf191cef..f4eb4e3cd 100644 --- a/src/common/settings.h +++ b/src/common/settings.h @@ -465,6 +465,7 @@ struct Values { SwitchableSetting async_astc{false, "async_astc"}; Setting vsync_mode{VSyncMode::FIFO, VSyncMode::Immediate, VSyncMode::FIFORelaxed, "use_vsync"}; + SwitchableSetting use_reactive_flushing{true, "use_reactive_flushing"}; SwitchableSetting shader_backend{ShaderBackend::GLSL, ShaderBackend::GLSL, ShaderBackend::SPIRV, "shader_backend"}; SwitchableSetting use_asynchronous_shaders{false, "use_asynchronous_shaders"}; diff --git a/src/core/memory.cpp b/src/core/memory.cpp index 7b79cb8bc..549b64ac4 100644 --- a/src/core/memory.cpp +++ b/src/core/memory.cpp @@ -465,7 +465,8 @@ struct Memory::Impl { } if (Settings::IsFastmemEnabled()) { - system.DeviceMemory().buffer.Protect(vaddr, size, !cached, !cached); + const bool is_read_enable = !Settings::values.use_reactive_flushing.GetValue() || !cached; + system.DeviceMemory().buffer.Protect(vaddr, size, is_read_enable, !cached); } // Iterate over a contiguous CPU address space, which corresponds to the specified GPU diff --git a/src/video_core/buffer_cache/buffer_cache.h b/src/video_core/buffer_cache/buffer_cache.h index 479a1a508..474822354 100644 --- a/src/video_core/buffer_cache/buffer_cache.h +++ b/src/video_core/buffer_cache/buffer_cache.h @@ -126,7 +126,8 @@ std::optional BufferCache

::GetFlushArea(VA area->preemtive = true; return area; }; - memory_tracker.MarkRegionAsPreflushable(cpu_addr_start_aligned, cpu_addr_end_aligned - cpu_addr_start_aligned); + memory_tracker.MarkRegionAsPreflushable(cpu_addr_start_aligned, + cpu_addr_end_aligned - cpu_addr_start_aligned); area->preemtive = !IsRegionGpuModified(cpu_addr, size); return area; } @@ -206,7 +207,8 @@ bool BufferCache

::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); - if (memory_tracker.IsRegionPreflushable(new_base_address, new_base_address + size)) { + 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); } @@ -1236,7 +1238,8 @@ void BufferCache

::MarkWrittenBuffer(BufferId buffer_id, VAddr cpu_addr, u32 s const IntervalType base_interval{cpu_addr, cpu_addr + size}; common_ranges.add(base_interval); - if (!memory_tracker.IsRegionPreflushable(cpu_addr, cpu_addr + size)) { + if (Settings::values.use_reactive_flushing.GetValue() && + !memory_tracker.IsRegionPreflushable(cpu_addr, cpu_addr + size)) { return; } uncommitted_ranges.add(base_interval); diff --git a/src/video_core/texture_cache/image_view_base.cpp b/src/video_core/texture_cache/image_view_base.cpp index 8f28342d5..30a7c11f5 100644 --- a/src/video_core/texture_cache/image_view_base.cpp +++ b/src/video_core/texture_cache/image_view_base.cpp @@ -26,7 +26,8 @@ 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); - if (image_info.forced_flushed) { + 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) { diff --git a/src/yuzu/configuration/config.cpp b/src/yuzu/configuration/config.cpp index a85eb4687..a49d12266 100644 --- a/src/yuzu/configuration/config.cpp +++ b/src/yuzu/configuration/config.cpp @@ -710,6 +710,7 @@ void Config::ReadRendererValues() { ReadGlobalSetting(Settings::values.nvdec_emulation); ReadGlobalSetting(Settings::values.accelerate_astc); ReadGlobalSetting(Settings::values.async_astc); + ReadGlobalSetting(Settings::values.use_reactive_flushing); ReadGlobalSetting(Settings::values.shader_backend); ReadGlobalSetting(Settings::values.use_asynchronous_shaders); ReadGlobalSetting(Settings::values.use_fast_gpu_time); @@ -1355,6 +1356,7 @@ void Config::SaveRendererValues() { Settings::values.nvdec_emulation.UsingGlobal()); WriteGlobalSetting(Settings::values.accelerate_astc); WriteGlobalSetting(Settings::values.async_astc); + WriteGlobalSetting(Settings::values.use_reactive_flushing); WriteSetting(QString::fromStdString(Settings::values.shader_backend.GetLabel()), static_cast(Settings::values.shader_backend.GetValue(global)), static_cast(Settings::values.shader_backend.GetDefault()), diff --git a/src/yuzu/configuration/configure_graphics_advanced.cpp b/src/yuzu/configuration/configure_graphics_advanced.cpp index 005b022ca..627ed8b17 100644 --- a/src/yuzu/configuration/configure_graphics_advanced.cpp +++ b/src/yuzu/configuration/configure_graphics_advanced.cpp @@ -21,6 +21,7 @@ ConfigureGraphicsAdvanced::~ConfigureGraphicsAdvanced() = default; void ConfigureGraphicsAdvanced::SetConfiguration() { const bool runtime_lock = !system.IsPoweredOn(); + ui->use_reactive_flushing->setEnabled(runtime_lock); ui->async_present->setEnabled(runtime_lock); ui->renderer_force_max_clock->setEnabled(runtime_lock); ui->async_astc->setEnabled(runtime_lock); @@ -29,6 +30,7 @@ void ConfigureGraphicsAdvanced::SetConfiguration() { ui->async_present->setChecked(Settings::values.async_presentation.GetValue()); ui->renderer_force_max_clock->setChecked(Settings::values.renderer_force_max_clock.GetValue()); + ui->use_reactive_flushing->setChecked(Settings::values.use_reactive_flushing.GetValue()); ui->async_astc->setChecked(Settings::values.async_astc.GetValue()); ui->use_asynchronous_shaders->setChecked(Settings::values.use_asynchronous_shaders.GetValue()); ui->use_fast_gpu_time->setChecked(Settings::values.use_fast_gpu_time.GetValue()); @@ -60,6 +62,8 @@ void ConfigureGraphicsAdvanced::ApplyConfiguration() { renderer_force_max_clock); ConfigurationShared::ApplyPerGameSetting(&Settings::values.max_anisotropy, ui->anisotropic_filtering_combobox); + ConfigurationShared::ApplyPerGameSetting(&Settings::values.use_reactive_flushing, + ui->use_reactive_flushing, use_reactive_flushing); ConfigurationShared::ApplyPerGameSetting(&Settings::values.async_astc, ui->async_astc, async_astc); ConfigurationShared::ApplyPerGameSetting(&Settings::values.use_asynchronous_shaders, @@ -91,6 +95,7 @@ void ConfigureGraphicsAdvanced::SetupPerGameUI() { ui->async_present->setEnabled(Settings::values.async_presentation.UsingGlobal()); ui->renderer_force_max_clock->setEnabled( Settings::values.renderer_force_max_clock.UsingGlobal()); + ui->use_reactive_flushing->setEnabled(Settings::values.use_reactive_flushing.UsingGlobal()); ui->async_astc->setEnabled(Settings::values.async_astc.UsingGlobal()); ui->use_asynchronous_shaders->setEnabled( Settings::values.use_asynchronous_shaders.UsingGlobal()); @@ -108,6 +113,8 @@ void ConfigureGraphicsAdvanced::SetupPerGameUI() { ConfigurationShared::SetColoredTristate(ui->renderer_force_max_clock, Settings::values.renderer_force_max_clock, renderer_force_max_clock); + ConfigurationShared::SetColoredTristate( + ui->use_reactive_flushing, Settings::values.use_reactive_flushing, use_reactive_flushing); ConfigurationShared::SetColoredTristate(ui->async_astc, Settings::values.async_astc, async_astc); ConfigurationShared::SetColoredTristate(ui->use_asynchronous_shaders, diff --git a/src/yuzu/configuration/configure_graphics_advanced.h b/src/yuzu/configuration/configure_graphics_advanced.h index ff5060957..ae3c10946 100644 --- a/src/yuzu/configuration/configure_graphics_advanced.h +++ b/src/yuzu/configuration/configure_graphics_advanced.h @@ -40,6 +40,7 @@ private: ConfigurationShared::CheckState renderer_force_max_clock; ConfigurationShared::CheckState use_vsync; ConfigurationShared::CheckState async_astc; + ConfigurationShared::CheckState use_reactive_flushing; ConfigurationShared::CheckState use_asynchronous_shaders; ConfigurationShared::CheckState use_fast_gpu_time; ConfigurationShared::CheckState use_vulkan_driver_pipeline_cache; diff --git a/src/yuzu/configuration/configure_graphics_advanced.ui b/src/yuzu/configuration/configure_graphics_advanced.ui index d073fe9b1..9d8cbea09 100644 --- a/src/yuzu/configuration/configure_graphics_advanced.ui +++ b/src/yuzu/configuration/configure_graphics_advanced.ui @@ -96,6 +96,16 @@ + + + + Uses reactive flushing instead of predictive flushing. Allowing a more accurate syncing of memory. + + + Enable Reactive Flushing + + + diff --git a/src/yuzu_cmd/config.cpp b/src/yuzu_cmd/config.cpp index a6418e693..abe7092fc 100644 --- a/src/yuzu_cmd/config.cpp +++ b/src/yuzu_cmd/config.cpp @@ -312,6 +312,7 @@ void Config::ReadValues() { ReadSetting("Renderer", Settings::values.use_asynchronous_gpu_emulation); ReadSetting("Renderer", Settings::values.vsync_mode); ReadSetting("Renderer", Settings::values.shader_backend); + ReadSetting("Renderer", Settings::values.use_reactive_flushing); ReadSetting("Renderer", Settings::values.use_asynchronous_shaders); ReadSetting("Renderer", Settings::values.nvdec_emulation); ReadSetting("Renderer", Settings::values.accelerate_astc); diff --git a/src/yuzu_cmd/default_ini.h b/src/yuzu_cmd/default_ini.h index 086ed4cfa..5e7c3ac04 100644 --- a/src/yuzu_cmd/default_ini.h +++ b/src/yuzu_cmd/default_ini.h @@ -340,6 +340,10 @@ use_vsync = # 0: GLSL, 1 (default): GLASM, 2: SPIR-V shader_backend = +# Uses reactive flushing instead of predictive flushing. Allowing a more accurate syncing of memory. +# 0: Off, 1 (default): On +use_reactive_flushing = + # Whether to allow asynchronous shader building. # 0 (default): Off, 1: On use_asynchronous_shaders = -- cgit v1.2.3 From 36c302fa32b475abb1b211934eab14fe0cad936a Mon Sep 17 00:00:00 2001 From: Fernando Sahmkow Date: Thu, 4 May 2023 13:23:36 +0200 Subject: Buffer cache: always use async buffer downloads and fix regression. --- src/tests/video_core/memory_tracker.cpp | 4 +- src/video_core/buffer_cache/buffer_cache.h | 114 ++++++++++++------------ src/video_core/buffer_cache/buffer_cache_base.h | 2 - src/video_core/buffer_cache/word_manager.h | 13 +++ 4 files changed, 70 insertions(+), 63 deletions(-) (limited to 'src/video_core/buffer_cache') diff --git a/src/tests/video_core/memory_tracker.cpp b/src/tests/video_core/memory_tracker.cpp index 3981907a2..618793668 100644 --- a/src/tests/video_core/memory_tracker.cpp +++ b/src/tests/video_core/memory_tracker.cpp @@ -535,12 +535,12 @@ TEST_CASE("MemoryTracker: Cached write downloads") { memory_track->MarkRegionAsGpuModified(c + PAGE, PAGE); int num = 0; memory_track->ForEachDownloadRangeAndClear(c, WORD, [&](u64 offset, u64 size) { ++num; }); - REQUIRE(num == 1); + REQUIRE(num == 0); num = 0; memory_track->ForEachUploadRange(c, WORD, [&](u64 offset, u64 size) { ++num; }); REQUIRE(num == 0); REQUIRE(!memory_track->IsRegionCpuModified(c + PAGE, PAGE)); - REQUIRE(!memory_track->IsRegionGpuModified(c + PAGE, PAGE)); + REQUIRE(memory_track->IsRegionGpuModified(c + PAGE, PAGE)); memory_track->FlushCachedWrites(); REQUIRE(memory_track->IsRegionCpuModified(c + PAGE, PAGE)); REQUIRE(!memory_track->IsRegionGpuModified(c + PAGE, PAGE)); diff --git a/src/video_core/buffer_cache/buffer_cache.h b/src/video_core/buffer_cache/buffer_cache.h index 474822354..0b15944d6 100644 --- a/src/video_core/buffer_cache/buffer_cache.h +++ b/src/video_core/buffer_cache/buffer_cache.h @@ -23,8 +23,6 @@ BufferCache

::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

::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; @@ -491,9 +487,8 @@ void BufferCache

::CommitAsyncFlushesHigh() { if (committed_ranges.empty()) { if constexpr (IMPLEMENTS_ASYNC_DOWNLOADS) { - if (active_async_buffers) { - async_buffers.emplace_back(std::optional{}); - } + + async_buffers.emplace_back(std::optional{}); } return; } @@ -554,64 +549,65 @@ void BufferCache

::CommitAsyncFlushesHigh() { committed_ranges.clear(); if (downloads.empty()) { if constexpr (IMPLEMENTS_ASYNC_DOWNLOADS) { - if (active_async_buffers) { - async_buffers.emplace_back(std::optional{}); - } + + async_buffers.emplace_back(std::optional{}); } return; } - if (active_async_buffers) { - if constexpr (IMPLEMENTS_ASYNC_DOWNLOADS) { - auto download_staging = runtime.DownloadStagingBuffer(total_size_bytes, true); - boost::container::small_vector 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(buffer.CpuAddr()) + copy.src_offset; - VAddr orig_cpu_addr = static_cast(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 { - committed_ranges.clear(); - uncommitted_ranges.clear(); + if constexpr (IMPLEMENTS_ASYNC_DOWNLOADS) { + auto download_staging = runtime.DownloadStagingBuffer(total_size_bytes, true); + boost::container::small_vector 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(buffer.CpuAddr()) + copy.src_offset; + VAddr orig_cpu_addr = static_cast(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 (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); - } + if (!Settings::IsGPULevelHigh()) { + committed_ranges.clear(); + uncommitted_ranges.clear(); } else { - const std::span 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 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); + } } } } diff --git a/src/video_core/buffer_cache/buffer_cache_base.h b/src/video_core/buffer_cache/buffer_cache_base.h index e3914a53a..0445ec47f 100644 --- a/src/video_core/buffer_cache/buffer_cache_base.h +++ b/src/video_core/buffer_cache/buffer_cache_base.h @@ -572,8 +572,6 @@ private: u64 critical_memory = 0; BufferId inline_buffer_id; - bool active_async_buffers = false; - std::array> CACHING_PAGEBITS)> page_table; std::vector tmp_buffer; }; diff --git a/src/video_core/buffer_cache/word_manager.h b/src/video_core/buffer_cache/word_manager.h index 0fb199a54..a336bde41 100644 --- a/src/video_core/buffer_cache/word_manager.h +++ b/src/video_core/buffer_cache/word_manager.h @@ -302,6 +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) { @@ -350,8 +353,13 @@ public: static_assert(type != Type::Untracked); const std::span state_words = words.template Span(); + [[maybe_unused]] const std::span untracked_words = + words.template Span(); 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; @@ -372,9 +380,14 @@ public: [[nodiscard]] std::pair ModifiedRegion(u64 offset, u64 size) const noexcept { static_assert(type != Type::Untracked); const std::span state_words = words.template Span(); + [[maybe_unused]] const std::span untracked_words = + words.template Span(); u64 begin = std::numeric_limits::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; -- cgit v1.2.3 From 2df19ef0fd5a91ca87e2c2cf201166a40c9d44dc Mon Sep 17 00:00:00 2001 From: Fernando Sahmkow Date: Sun, 7 May 2023 23:25:34 +0200 Subject: Buffer Cache: disable reactive flushing in it. --- src/video_core/buffer_cache/buffer_cache.h | 11 ++--------- src/video_core/renderer_vulkan/vk_rasterizer.cpp | 7 ------- src/video_core/texture_cache/texture_cache.h | 8 ++++++-- 3 files changed, 8 insertions(+), 18 deletions(-) (limited to 'src/video_core/buffer_cache') diff --git a/src/video_core/buffer_cache/buffer_cache.h b/src/video_core/buffer_cache/buffer_cache.h index 0b15944d6..6624919a4 100644 --- a/src/video_core/buffer_cache/buffer_cache.h +++ b/src/video_core/buffer_cache/buffer_cache.h @@ -203,11 +203,8 @@ bool BufferCache

::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); - 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); - } + 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. @@ -1234,10 +1231,6 @@ void BufferCache

::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); } diff --git a/src/video_core/renderer_vulkan/vk_rasterizer.cpp b/src/video_core/renderer_vulkan/vk_rasterizer.cpp index 2beb0efea..0bcc20544 100644 --- a/src/video_core/renderer_vulkan/vk_rasterizer.cpp +++ b/src/video_core/renderer_vulkan/vk_rasterizer.cpp @@ -510,13 +510,6 @@ VideoCore::RasterizerDownloadArea RasterizerVulkan::GetFlushArea(VAddr addr, u64 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), diff --git a/src/video_core/texture_cache/texture_cache.h b/src/video_core/texture_cache/texture_cache.h index 29ac01eb4..d49f3a7a0 100644 --- a/src/video_core/texture_cache/texture_cache.h +++ b/src/video_core/texture_cache/texture_cache.h @@ -1323,7 +1323,6 @@ ImageId TextureCache

::JoinImages(const ImageInfo& info, GPUVAddr gpu_addr, VA all_siblings.push_back(overlap_id); } else { bad_overlap_ids.push_back(overlap_id); - overlap.flags |= ImageFlagBits::BadOverlap; } }; ForEachImageInRegion(cpu_addr, size_bytes, region_check); @@ -1434,7 +1433,12 @@ ImageId TextureCache

::JoinImages(const ImageInfo& info, GPUVAddr gpu_addr, VA ImageBase& aliased = slot_images[aliased_id]; aliased.overlapping_images.push_back(new_image_id); new_image.overlapping_images.push_back(aliased_id); - new_image.flags |= ImageFlagBits::BadOverlap; + if (aliased.info.resources.levels == 1 && aliased.overlapping_images.size() > 1) { + aliased.flags |= ImageFlagBits::BadOverlap; + } + if (new_image.info.resources.levels == 1 && new_image.overlapping_images.size() > 1) { + new_image.flags |= ImageFlagBits::BadOverlap; + } } RegisterImage(new_image_id); return new_image_id; -- cgit v1.2.3