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_cache.h | 35 ++++++++++++++++++++++-------- 1 file changed, 26 insertions(+), 9 deletions(-) (limited to 'src/video_core/buffer_cache/buffer_cache.h') 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) { -- 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/video_core/buffer_cache/buffer_cache.h | 9 ++++++--- 1 file changed, 6 insertions(+), 3 deletions(-) (limited to 'src/video_core/buffer_cache/buffer_cache.h') 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); -- 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/video_core/buffer_cache/buffer_cache.h | 114 ++++++++++++++--------------- 1 file changed, 55 insertions(+), 59 deletions(-) (limited to 'src/video_core/buffer_cache/buffer_cache.h') 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); + } } } } -- 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 ++--------- 1 file changed, 2 insertions(+), 9 deletions(-) (limited to 'src/video_core/buffer_cache/buffer_cache.h') 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); } -- cgit v1.2.3