video_core: Reimplement the buffer cache

Reimplement the buffer cache using cached bindings and page level
granularity for modification tracking. This also drops the usage of
shared pointers and virtual functions from the cache.

- Bindings are cached, allowing to skip work when the game changes few
  bits between draws.
- OpenGL Assembly shaders no longer copy when a region has been modified
  from the GPU to emulate constant buffers, instead GL_EXT_memory_object
  is used to alias sub-buffers within the same allocation.
- OpenGL Assembly shaders stream constant buffer data using
  glProgramBufferParametersIuivNV, from NV_parameter_buffer_object. In
  theory this should save one hash table resolve inside the driver
  compared to glBufferSubData.
- A new OpenGL stream buffer is implemented based on fences for drivers
  that are not Nvidia's proprietary, due to their low performance on
  partial glBufferSubData calls synchronized with 3D rendering (that
  some games use a lot).
- Most optimizations are shared between APIs now, allowing Vulkan to
  cache more bindings than before, skipping unnecesarry work.

This commit adds the necessary infrastructure to use Vulkan object from
OpenGL. Overall, it improves performance and fixes some bugs present on
the old cache. There are still some edge cases hit by some games that
harm performance on some vendors, this are planned to be fixed in later
commits.
This commit is contained in:
ReinUsesLisp 2021-01-16 20:48:58 -03:00
commit 2dfce2fca6
67 changed files with 2514 additions and 2607 deletions

View file

@ -2,10 +2,8 @@ add_subdirectory(host_shaders)
add_library(video_core STATIC
buffer_cache/buffer_base.h
buffer_cache/buffer_block.h
buffer_cache/buffer_cache.cpp
buffer_cache/buffer_cache.h
buffer_cache/map_interval.cpp
buffer_cache/map_interval.h
cdma_pusher.cpp
cdma_pusher.h
command_classes/codecs/codec.cpp
@ -152,8 +150,6 @@ add_library(video_core STATIC
renderer_vulkan/vk_staging_buffer_pool.h
renderer_vulkan/vk_state_tracker.cpp
renderer_vulkan/vk_state_tracker.h
renderer_vulkan/vk_stream_buffer.cpp
renderer_vulkan/vk_stream_buffer.h
renderer_vulkan/vk_swapchain.cpp
renderer_vulkan/vk_swapchain.h
renderer_vulkan/vk_texture_cache.cpp

View file

@ -1,62 +0,0 @@
// Copyright 2019 yuzu Emulator Project
// Licensed under GPLv2 or any later version
// Refer to the license.txt file included.
#pragma once
#include "common/common_types.h"
namespace VideoCommon {
class BufferBlock {
public:
[[nodiscard]] bool Overlaps(VAddr start, VAddr end) const {
return (cpu_addr < end) && (cpu_addr_end > start);
}
[[nodiscard]] bool IsInside(VAddr other_start, VAddr other_end) const {
return cpu_addr <= other_start && other_end <= cpu_addr_end;
}
[[nodiscard]] std::size_t Offset(VAddr in_addr) const {
return static_cast<std::size_t>(in_addr - cpu_addr);
}
[[nodiscard]] VAddr CpuAddr() const {
return cpu_addr;
}
[[nodiscard]] VAddr CpuAddrEnd() const {
return cpu_addr_end;
}
void SetCpuAddr(VAddr new_addr) {
cpu_addr = new_addr;
cpu_addr_end = new_addr + size;
}
[[nodiscard]] std::size_t Size() const {
return size;
}
[[nodiscard]] u64 Epoch() const {
return epoch;
}
void SetEpoch(u64 new_epoch) {
epoch = new_epoch;
}
protected:
explicit BufferBlock(VAddr cpu_addr_, std::size_t size_) : size{size_} {
SetCpuAddr(cpu_addr_);
}
private:
VAddr cpu_addr{};
VAddr cpu_addr_end{};
std::size_t size{};
u64 epoch{};
};
} // namespace VideoCommon

View file

@ -0,0 +1,13 @@
// Copyright 2021 yuzu Emulator Project
// Licensed under GPLv2 or any later version
// Refer to the license.txt file included.
#include "common/microprofile.h"
namespace VideoCommon {
MICROPROFILE_DEFINE(GPU_PrepareBuffers, "GPU", "Prepare buffers", MP_RGB(224, 128, 128));
MICROPROFILE_DEFINE(GPU_BindUploadBuffers, "GPU", "Bind and upload buffers", MP_RGB(224, 128, 128));
MICROPROFILE_DEFINE(GPU_DownloadMemory, "GPU", "Download buffers", MP_RGB(224, 128, 128));
} // namespace VideoCommon

File diff suppressed because it is too large Load diff

View file

@ -1,33 +0,0 @@
// Copyright 2020 yuzu Emulator Project
// Licensed under GPLv2 or any later version
// Refer to the license.txt file included.
#include <algorithm>
#include <array>
#include <cstddef>
#include <memory>
#include "video_core/buffer_cache/map_interval.h"
namespace VideoCommon {
MapIntervalAllocator::MapIntervalAllocator() {
FillFreeList(first_chunk);
}
MapIntervalAllocator::~MapIntervalAllocator() = default;
void MapIntervalAllocator::AllocateNewChunk() {
*new_chunk = std::make_unique<Chunk>();
FillFreeList(**new_chunk);
new_chunk = &(*new_chunk)->next;
}
void MapIntervalAllocator::FillFreeList(Chunk& chunk) {
const std::size_t old_size = free_list.size();
free_list.resize(old_size + chunk.data.size());
std::transform(chunk.data.rbegin(), chunk.data.rend(), free_list.begin() + old_size,
[](MapInterval& interval) { return &interval; });
}
} // namespace VideoCommon

View file

@ -1,93 +0,0 @@
// Copyright 2019 yuzu Emulator Project
// Licensed under GPLv2 or any later version
// Refer to the license.txt file included.
#pragma once
#include <array>
#include <cstddef>
#include <memory>
#include <vector>
#include <boost/intrusive/set_hook.hpp>
#include "common/common_types.h"
#include "video_core/gpu.h"
namespace VideoCommon {
struct MapInterval : public boost::intrusive::set_base_hook<boost::intrusive::optimize_size<true>> {
MapInterval() = default;
/*implicit*/ MapInterval(VAddr start_) noexcept : start{start_} {}
explicit MapInterval(VAddr start_, VAddr end_, GPUVAddr gpu_addr_) noexcept
: start{start_}, end{end_}, gpu_addr{gpu_addr_} {}
bool IsInside(VAddr other_start, VAddr other_end) const noexcept {
return start <= other_start && other_end <= end;
}
bool Overlaps(VAddr other_start, VAddr other_end) const noexcept {
return start < other_end && other_start < end;
}
void MarkAsModified(bool is_modified_, u64 ticks_) noexcept {
is_modified = is_modified_;
ticks = ticks_;
}
boost::intrusive::set_member_hook<> member_hook_;
VAddr start = 0;
VAddr end = 0;
GPUVAddr gpu_addr = 0;
u64 ticks = 0;
bool is_written = false;
bool is_modified = false;
bool is_registered = false;
bool is_memory_marked = false;
bool is_sync_pending = false;
};
struct MapIntervalCompare {
constexpr bool operator()(const MapInterval& lhs, const MapInterval& rhs) const noexcept {
return lhs.start < rhs.start;
}
};
class MapIntervalAllocator {
public:
MapIntervalAllocator();
~MapIntervalAllocator();
MapInterval* Allocate() {
if (free_list.empty()) {
AllocateNewChunk();
}
MapInterval* const interval = free_list.back();
free_list.pop_back();
return interval;
}
void Release(MapInterval* interval) {
free_list.push_back(interval);
}
private:
struct Chunk {
std::unique_ptr<Chunk> next;
std::array<MapInterval, 0x8000> data;
};
void AllocateNewChunk();
void FillFreeList(Chunk& chunk);
std::vector<MapInterval*> free_list;
Chunk first_chunk;
std::unique_ptr<Chunk>* new_chunk = &first_chunk.next;
};
} // namespace VideoCommon

View file

@ -110,12 +110,10 @@ void Vic::Execute() {
converted_frame_buffer.get(), block_height, 0, 0);
gpu.MemoryManager().WriteBlock(output_surface_luma_address, swizzled_data.data(), size);
gpu.Maxwell3D().OnMemoryWrite();
} else {
// send pitch linear frame
gpu.MemoryManager().WriteBlock(output_surface_luma_address, converted_frame_buf_addr,
linear_size);
gpu.Maxwell3D().OnMemoryWrite();
}
break;
}
@ -163,7 +161,6 @@ void Vic::Execute() {
}
gpu.MemoryManager().WriteBlock(output_surface_chroma_u_address, chroma_buffer.data(),
chroma_buffer.size());
gpu.Maxwell3D().OnMemoryWrite();
break;
}
default:

View file

@ -12,13 +12,30 @@
#define NUM(field_name) (sizeof(::Tegra::Engines::Maxwell3D::Regs::field_name) / (sizeof(u32)))
namespace VideoCommon::Dirty {
namespace {
using Tegra::Engines::Maxwell3D;
void SetupDirtyRenderTargets(Tegra::Engines::Maxwell3D::DirtyState::Tables& tables) {
void SetupDirtyVertexBuffers(Maxwell3D::DirtyState::Tables& tables) {
static constexpr std::size_t num_array = 3;
for (std::size_t i = 0; i < Maxwell3D::Regs::NumVertexArrays; ++i) {
const std::size_t array_offset = OFF(vertex_array) + i * NUM(vertex_array[0]);
const std::size_t limit_offset = OFF(vertex_array_limit) + i * NUM(vertex_array_limit[0]);
FillBlock(tables, array_offset, num_array, VertexBuffer0 + i, VertexBuffers);
FillBlock(tables, limit_offset, NUM(vertex_array_limit), VertexBuffer0 + i, VertexBuffers);
}
}
void SetupIndexBuffer(Maxwell3D::DirtyState::Tables& tables) {
FillBlock(tables[0], OFF(index_array), NUM(index_array), IndexBuffer);
}
void SetupDirtyDescriptors(Maxwell3D::DirtyState::Tables& tables) {
FillBlock(tables[0], OFF(tic), NUM(tic), Descriptors);
FillBlock(tables[0], OFF(tsc), NUM(tsc), Descriptors);
}
void SetupDirtyRenderTargets(Maxwell3D::DirtyState::Tables& tables) {
static constexpr std::size_t num_per_rt = NUM(rt[0]);
static constexpr std::size_t begin = OFF(rt);
static constexpr std::size_t num = num_per_rt * Maxwell3D::Regs::NumRenderTargets;
@ -41,5 +58,13 @@ void SetupDirtyRenderTargets(Tegra::Engines::Maxwell3D::DirtyState::Tables& tabl
FillBlock(table, OFF(zeta), NUM(zeta), flag);
}
}
} // Anonymous namespace
void SetupDirtyFlags(Maxwell3D::DirtyState::Tables& tables) {
SetupDirtyVertexBuffers(tables);
SetupIndexBuffer(tables);
SetupDirtyDescriptors(tables);
SetupDirtyRenderTargets(tables);
}
} // namespace VideoCommon::Dirty

View file

@ -30,6 +30,12 @@ enum : u8 {
ColorBuffer7,
ZetaBuffer,
VertexBuffers,
VertexBuffer0,
VertexBuffer31 = VertexBuffer0 + 31,
IndexBuffer,
LastCommonEntry,
};
@ -47,6 +53,6 @@ void FillBlock(Tegra::Engines::Maxwell3D::DirtyState::Tables& tables, std::size_
FillBlock(tables[1], begin, num, index_b);
}
void SetupDirtyRenderTargets(Tegra::Engines::Maxwell3D::DirtyState::Tables& tables);
void SetupDirtyFlags(Tegra::Engines::Maxwell3D::DirtyState::Tables& tables);
} // namespace VideoCommon::Dirty

View file

@ -23,8 +23,6 @@ void DmaPusher::DispatchCalls() {
MICROPROFILE_SCOPE(DispatchCalls);
gpu.SyncGuestHost();
// On entering GPU code, assume all memory may be touched by the ARM core.
gpu.Maxwell3D().OnMemoryWrite();
dma_pushbuffer_subindex = 0;

View file

@ -39,7 +39,6 @@ void KeplerCompute::CallMethod(u32 method, u32 method_argument, bool is_last_cal
case KEPLER_COMPUTE_REG_INDEX(data_upload): {
upload_state.ProcessData(method_argument, is_last_call);
if (is_last_call) {
system.GPU().Maxwell3D().OnMemoryWrite();
}
break;
}

View file

@ -33,7 +33,6 @@ void KeplerMemory::CallMethod(u32 method, u32 method_argument, bool is_last_call
case KEPLERMEMORY_REG_INDEX(data): {
upload_state.ProcessData(method_argument, is_last_call);
if (is_last_call) {
system.GPU().Maxwell3D().OnMemoryWrite();
}
break;
}

View file

@ -223,7 +223,6 @@ void Maxwell3D::ProcessMethodCall(u32 method, u32 argument, u32 nonshadow_argume
case MAXWELL3D_REG_INDEX(data_upload):
upload_state.ProcessData(argument, is_last_call);
if (is_last_call) {
OnMemoryWrite();
}
return;
case MAXWELL3D_REG_INDEX(fragment_barrier):
@ -570,17 +569,18 @@ std::optional<u64> Maxwell3D::GetQueryResult() {
}
}
void Maxwell3D::ProcessCBBind(std::size_t stage_index) {
void Maxwell3D::ProcessCBBind(size_t stage_index) {
// Bind the buffer currently in CB_ADDRESS to the specified index in the desired shader stage.
auto& shader = state.shader_stages[stage_index];
auto& bind_data = regs.cb_bind[stage_index];
ASSERT(bind_data.index < Regs::MaxConstBuffers);
auto& buffer = shader.const_buffers[bind_data.index];
const auto& bind_data = regs.cb_bind[stage_index];
auto& buffer = state.shader_stages[stage_index].const_buffers[bind_data.index];
buffer.enabled = bind_data.valid.Value() != 0;
buffer.address = regs.const_buffer.BufferAddress();
buffer.size = regs.const_buffer.cb_size;
const bool is_enabled = bind_data.valid.Value() != 0;
const GPUVAddr gpu_addr = is_enabled ? regs.const_buffer.BufferAddress() : 0;
const u32 size = is_enabled ? regs.const_buffer.cb_size : 0;
rasterizer->BindGraphicsUniformBuffer(stage_index, bind_data.index, gpu_addr, size);
}
void Maxwell3D::ProcessCBData(u32 value) {
@ -635,7 +635,6 @@ void Maxwell3D::FinishCBData() {
const u32 id = cb_data_state.id;
memory_manager.WriteBlock(address, cb_data_state.buffer[id].data(), size);
OnMemoryWrite();
cb_data_state.id = null_cb_data;
cb_data_state.current = null_cb_data;

View file

@ -1314,8 +1314,7 @@ public:
GPUVAddr LimitAddress() const {
return static_cast<GPUVAddr>((static_cast<GPUVAddr>(limit_high) << 32) |
limit_low) +
1;
limit_low);
}
} vertex_array_limit[NumVertexArrays];
@ -1403,6 +1402,7 @@ public:
};
std::array<ShaderStageInfo, Regs::MaxShaderStage> shader_stages;
u32 current_instance = 0; ///< Current instance to be used to simulate instanced rendering.
};
@ -1452,11 +1452,6 @@ public:
return *rasterizer;
}
/// Notify a memory write has happened.
void OnMemoryWrite() {
dirty.flags |= dirty.on_write_stores;
}
enum class MMEDrawMode : u32 {
Undefined,
Array,
@ -1478,7 +1473,6 @@ public:
using Tables = std::array<Table, 2>;
Flags flags;
Flags on_write_stores;
Tables tables{};
} dirty;
@ -1541,7 +1535,7 @@ private:
void FinishCBData();
/// Handles a write to the CB_BIND register.
void ProcessCBBind(std::size_t stage_index);
void ProcessCBBind(size_t stage_index);
/// Handles a write to the VERTEX_END_GL register, triggering a draw.
void DrawArrays();

View file

@ -60,9 +60,6 @@ void MaxwellDMA::Launch() {
return;
}
// All copies here update the main memory, so mark all rasterizer states as invalid.
system.GPU().Maxwell3D().OnMemoryWrite();
if (is_src_pitch && is_dst_pitch) {
CopyPitchToPitch();
} else {

View file

@ -143,22 +143,26 @@ private:
}
bool ShouldWait() const {
std::scoped_lock lock{buffer_cache.mutex, texture_cache.mutex};
return texture_cache.ShouldWaitAsyncFlushes() || buffer_cache.ShouldWaitAsyncFlushes() ||
query_cache.ShouldWaitAsyncFlushes();
}
bool ShouldFlush() const {
std::scoped_lock lock{buffer_cache.mutex, texture_cache.mutex};
return texture_cache.HasUncommittedFlushes() || buffer_cache.HasUncommittedFlushes() ||
query_cache.HasUncommittedFlushes();
}
void PopAsyncFlushes() {
std::scoped_lock lock{buffer_cache.mutex, texture_cache.mutex};
texture_cache.PopAsyncFlushes();
buffer_cache.PopAsyncFlushes();
query_cache.PopAsyncFlushes();
}
void CommitAsyncFlushes() {
std::scoped_lock lock{buffer_cache.mutex, texture_cache.mutex};
texture_cache.CommitAsyncFlushes();
buffer_cache.CommitAsyncFlushes();
query_cache.CommitAsyncFlushes();

View file

@ -12,7 +12,6 @@ set(SHADER_FILES
vulkan_blit_depth_stencil.frag
vulkan_present.frag
vulkan_present.vert
vulkan_quad_array.comp
vulkan_quad_indexed.comp
vulkan_uint8.comp
)

View file

@ -1,28 +0,0 @@
// Copyright 2019 yuzu Emulator Project
// Licensed under GPLv2 or any later version
// Refer to the license.txt file included.
#version 460 core
layout (local_size_x = 1024) in;
layout (std430, set = 0, binding = 0) buffer OutputBuffer {
uint output_indexes[];
};
layout (push_constant) uniform PushConstants {
uint first;
};
void main() {
uint primitive = gl_GlobalInvocationID.x;
if (primitive * 6 >= output_indexes.length()) {
return;
}
const uint quad_map[6] = uint[](0, 1, 2, 0, 2, 3);
for (uint vertex = 0; vertex < 6; ++vertex) {
uint index = first + primitive * 4 + quad_map[vertex];
output_indexes[primitive * 6 + vertex] = index;
}
}

View file

@ -16,9 +16,16 @@ layout (std430, set = 0, binding = 1) writeonly buffer OutputBuffer {
uint16_t output_indexes[];
};
uint AssembleIndex(uint id) {
// Most primitive restart indices are 0xFF
// Hardcode this to 0xFF for now
uint index = uint(input_indexes[id]);
return index == 0xFF ? 0xFFFF : index;
}
void main() {
uint id = gl_GlobalInvocationID.x;
if (id < input_indexes.length()) {
output_indexes[id] = uint16_t(input_indexes[id]);
output_indexes[id] = uint16_t(AssembleIndex(id));
}
}

View file

@ -7,6 +7,7 @@
#include <atomic>
#include <functional>
#include <optional>
#include <span>
#include "common/common_types.h"
#include "video_core/engines/fermi_2d.h"
#include "video_core/gpu.h"
@ -49,6 +50,10 @@ public:
/// Records a GPU query and caches it
virtual void Query(GPUVAddr gpu_addr, QueryType type, std::optional<u64> timestamp) = 0;
/// Signal an uniform buffer binding
virtual void BindGraphicsUniformBuffer(size_t stage, u32 index, GPUVAddr gpu_addr,
u32 size) = 0;
/// Signal a GPU based semaphore as a fence
virtual void SignalSemaphore(GPUVAddr addr, u32 value) = 0;

Some files were not shown because too many files have changed in this diff Show more