From 33a05976036211754251958654a26106d954ae43 Mon Sep 17 00:00:00 2001 From: ReinUsesLisp Date: Sun, 24 Feb 2019 04:22:33 -0300 Subject: vk_stream_buffer: Implement a stream buffer This manages two kinds of streaming buffers: one for unified memory models and one for dedicated GPUs. The first one skips the copy from the staging buffer to the real buffer, since it creates an unified buffer. This implementation waits for all fences to finish their operation before "invalidating". This is suboptimal since it should allocate another buffer or start searching from the beginning. There is room for improvement here. This could also handle AMD's "pinned" memory (a heap with 256 MiB) that seems to be designed for buffer streaming. --- src/video_core/CMakeLists.txt | 4 +- .../renderer_vulkan/vk_stream_buffer.cpp | 124 +++++++++++++++++++++ src/video_core/renderer_vulkan/vk_stream_buffer.h | 73 ++++++++++++ 3 files changed, 200 insertions(+), 1 deletion(-) create mode 100644 src/video_core/renderer_vulkan/vk_stream_buffer.cpp create mode 100644 src/video_core/renderer_vulkan/vk_stream_buffer.h diff --git a/src/video_core/CMakeLists.txt b/src/video_core/CMakeLists.txt index 6036d6ed3..60529323e 100644 --- a/src/video_core/CMakeLists.txt +++ b/src/video_core/CMakeLists.txt @@ -111,7 +111,9 @@ if (ENABLE_VULKAN) renderer_vulkan/vk_resource_manager.cpp renderer_vulkan/vk_resource_manager.h renderer_vulkan/vk_scheduler.cpp - renderer_vulkan/vk_scheduler.h) + renderer_vulkan/vk_scheduler.h + renderer_vulkan/vk_stream_buffer.cpp + renderer_vulkan/vk_stream_buffer.h) target_include_directories(video_core PRIVATE ../../externals/Vulkan-Headers/include) target_compile_definitions(video_core PRIVATE HAS_VULKAN) diff --git a/src/video_core/renderer_vulkan/vk_stream_buffer.cpp b/src/video_core/renderer_vulkan/vk_stream_buffer.cpp new file mode 100644 index 000000000..1c5aefaec --- /dev/null +++ b/src/video_core/renderer_vulkan/vk_stream_buffer.cpp @@ -0,0 +1,124 @@ +// Copyright 2019 yuzu Emulator Project +// Licensed under GPLv2 or any later version +// Refer to the license.txt file included. + +#include +#include +#include +#include + +#include "common/assert.h" +#include "video_core/renderer_vulkan/declarations.h" +#include "video_core/renderer_vulkan/vk_device.h" +#include "video_core/renderer_vulkan/vk_memory_manager.h" +#include "video_core/renderer_vulkan/vk_resource_manager.h" +#include "video_core/renderer_vulkan/vk_scheduler.h" +#include "video_core/renderer_vulkan/vk_stream_buffer.h" + +namespace Vulkan { + +constexpr u64 WATCHES_INITIAL_RESERVE = 0x4000; +constexpr u64 WATCHES_RESERVE_CHUNK = 0x1000; + +VKStreamBuffer::VKStreamBuffer(const VKDevice& device, VKMemoryManager& memory_manager, + VKScheduler& scheduler, u64 size, vk::BufferUsageFlags usage, + vk::AccessFlags access, vk::PipelineStageFlags pipeline_stage) + : device{device}, scheduler{scheduler}, + has_device_exclusive_memory{!memory_manager.IsMemoryUnified()}, + buffer_size{size}, access{access}, pipeline_stage{pipeline_stage} { + CreateBuffers(memory_manager, usage); + ReserveWatches(WATCHES_INITIAL_RESERVE); +} + +VKStreamBuffer::~VKStreamBuffer() = default; + +std::tuple VKStreamBuffer::Reserve(u64 size, bool keep_in_host) { + ASSERT(size <= buffer_size); + mapped_size = size; + + if (offset + size > buffer_size) { + // The buffer would overflow, save the amount of used buffers, signal an invalidation and + // reset the state. + invalidation_mark = used_watches; + used_watches = 0; + offset = 0; + } + + use_device = has_device_exclusive_memory && !keep_in_host; + + const vk::Buffer buffer = use_device ? *device_buffer : *mappable_buffer; + return {mapped_pointer + offset, offset, buffer, invalidation_mark.has_value()}; +} + +VKExecutionContext VKStreamBuffer::Send(VKExecutionContext exctx, u64 size) { + ASSERT_MSG(size <= mapped_size, "Reserved size is too small"); + + if (invalidation_mark) { + // TODO(Rodrigo): Find a better way to invalidate than waiting for all watches to finish. + exctx = scheduler.Flush(); + std::for_each(watches.begin(), watches.begin() + *invalidation_mark, + [&](auto& resource) { resource->Wait(); }); + invalidation_mark = std::nullopt; + } + + // Only copy to VRAM when requested. + if (use_device) { + const auto& dld = device.GetDispatchLoader(); + const u32 graphics_family = device.GetGraphicsFamily(); + const auto cmdbuf = exctx.GetCommandBuffer(); + + // Buffers are mirrored, that's why the copy is done with the same offset on both buffers. + const vk::BufferCopy copy_region(offset, offset, size); + cmdbuf.copyBuffer(*mappable_buffer, *device_buffer, {copy_region}, dld); + + // Protect the buffer from GPU usage until the copy has finished. + const vk::BufferMemoryBarrier barrier(vk::AccessFlagBits::eTransferWrite, access, + graphics_family, graphics_family, *device_buffer, + offset, size); + cmdbuf.pipelineBarrier(vk::PipelineStageFlagBits::eTransfer, pipeline_stage, {}, {}, + {barrier}, {}, dld); + } + + if (used_watches + 1 >= watches.size()) { + // Ensure that there are enough watches. + ReserveWatches(WATCHES_RESERVE_CHUNK); + } + // Add a watch for this allocation. + watches[used_watches++]->Watch(exctx.GetFence()); + + offset += size; + + return exctx; +} + +void VKStreamBuffer::CreateBuffers(VKMemoryManager& memory_manager, vk::BufferUsageFlags usage) { + vk::BufferUsageFlags mappable_usage = usage; + if (has_device_exclusive_memory) { + mappable_usage |= vk::BufferUsageFlagBits::eTransferSrc; + } + const vk::BufferCreateInfo buffer_ci({}, buffer_size, mappable_usage, + vk::SharingMode::eExclusive, 0, nullptr); + + const auto dev = device.GetLogical(); + const auto& dld = device.GetDispatchLoader(); + mappable_buffer = dev.createBufferUnique(buffer_ci, nullptr, dld); + mappable_commit = memory_manager.Commit(*mappable_buffer, true); + mapped_pointer = mappable_commit->GetData(); + + if (has_device_exclusive_memory) { + const vk::BufferCreateInfo buffer_ci({}, buffer_size, + usage | vk::BufferUsageFlagBits::eTransferDst, + vk::SharingMode::eExclusive, 0, nullptr); + device_buffer = dev.createBufferUnique(buffer_ci, nullptr, dld); + device_commit = memory_manager.Commit(*device_buffer, false); + } +} + +void VKStreamBuffer::ReserveWatches(std::size_t grow_size) { + const std::size_t previous_size = watches.size(); + watches.resize(previous_size + grow_size); + std::generate(watches.begin() + previous_size, watches.end(), + []() { return std::make_unique(); }); +} + +} // namespace Vulkan diff --git a/src/video_core/renderer_vulkan/vk_stream_buffer.h b/src/video_core/renderer_vulkan/vk_stream_buffer.h new file mode 100644 index 000000000..8c00d383a --- /dev/null +++ b/src/video_core/renderer_vulkan/vk_stream_buffer.h @@ -0,0 +1,73 @@ +// Copyright 2019 yuzu Emulator Project +// Licensed under GPLv2 or any later version +// Refer to the license.txt file included. + +#pragma once + +#include +#include +#include +#include + +#include "common/common_types.h" +#include "video_core/renderer_vulkan/declarations.h" +#include "video_core/renderer_vulkan/vk_memory_manager.h" + +namespace Vulkan { + +class VKDevice; +class VKFence; +class VKFenceWatch; +class VKResourceManager; +class VKScheduler; + +class VKStreamBuffer { +public: + explicit VKStreamBuffer(const VKDevice& device, VKMemoryManager& memory_manager, + VKScheduler& scheduler, u64 size, vk::BufferUsageFlags usage, + vk::AccessFlags access, vk::PipelineStageFlags pipeline_stage); + ~VKStreamBuffer(); + + /** + * Reserves a region of memory from the stream buffer. + * @param size Size to reserve. + * @param keep_in_host Mapped buffer will be in host memory, skipping the copy to device local. + * @returns A tuple in the following order: Raw memory pointer (with offset added), buffer + * offset, Vulkan buffer handle, buffer has been invalited. + */ + std::tuple Reserve(u64 size, bool keep_in_host); + + /// Ensures that "size" bytes of memory are available to the GPU, potentially recording a copy. + [[nodiscard]] VKExecutionContext Send(VKExecutionContext exctx, u64 size); + +private: + /// Creates Vulkan buffer handles committing the required the required memory. + void CreateBuffers(VKMemoryManager& memory_manager, vk::BufferUsageFlags usage); + + /// Increases the amount of watches available. + void ReserveWatches(std::size_t grow_size); + + const VKDevice& device; ///< Vulkan device manager. + VKScheduler& scheduler; ///< Command scheduler. + const u64 buffer_size; ///< Total size of the stream buffer. + const bool has_device_exclusive_memory; ///< True if the streaming buffer will use VRAM. + const vk::AccessFlags access; ///< Access usage of this stream buffer. + const vk::PipelineStageFlags pipeline_stage; ///< Pipeline usage of this stream buffer. + + UniqueBuffer mappable_buffer; ///< Mapped buffer. + UniqueBuffer device_buffer; ///< Buffer exclusive to the GPU. + VKMemoryCommit mappable_commit; ///< Commit visible from the CPU. + VKMemoryCommit device_commit; ///< Commit stored in VRAM. + u8* mapped_pointer{}; ///< Pointer to the host visible commit + + u64 offset{}; ///< Buffer iterator. + u64 mapped_size{}; ///< Size reserved for the current copy. + bool use_device{}; ///< True if the current uses VRAM. + + std::vector> watches; ///< Total watches + std::size_t used_watches{}; ///< Count of watches, reset on invalidation. + std::optional + invalidation_mark{}; ///< Number of watches used in the current invalidation. +}; + +} // namespace Vulkan -- cgit v1.2.3