gl_rasterizer: Upload constant buffers with glNamedBufferSubData

Nvidia's OpenGL driver maps gl(Named)BufferSubData with some requirements
to a fast. This path has an extra memcpy but updates the buffer without
orphaning or waiting for previous calls. It can be seen as a better
model for "push constants" that can upload a whole UBO instead of 256
bytes.

This path has some requirements established here:
http://on-demand.gputechconf.com/gtc/2014/presentations/S4379-opengl-44-scene-rendering-techniques.pdf#page=24

Instead of using the stream buffer, this commits moves constant buffers
uploads to calls of glNamedBufferSubData and from my testing it brings a
performance improvement. This is disabled when the vendor is not Nvidia
since it brings performance regressions.
This commit is contained in:
ReinUsesLisp 2019-11-02 04:08:31 -03:00
parent 11e39da02b
commit 76ca2a5f82
No known key found for this signature in database
GPG Key ID: 2DFC508897B39CFE
6 changed files with 84 additions and 19 deletions

View File

@ -30,7 +30,7 @@ public:
using BufferInfo = std::pair<const TBufferType*, u64>; using BufferInfo = std::pair<const TBufferType*, u64>;
BufferInfo UploadMemory(GPUVAddr gpu_addr, std::size_t size, std::size_t alignment = 4, BufferInfo UploadMemory(GPUVAddr gpu_addr, std::size_t size, std::size_t alignment = 4,
bool is_written = false) { bool is_written = false, bool use_fast_cbuf = false) {
std::lock_guard lock{mutex}; std::lock_guard lock{mutex};
auto& memory_manager = system.GPU().MemoryManager(); auto& memory_manager = system.GPU().MemoryManager();
@ -43,9 +43,13 @@ public:
// Cache management is a big overhead, so only cache entries with a given size. // Cache management is a big overhead, so only cache entries with a given size.
// TODO: Figure out which size is the best for given games. // TODO: Figure out which size is the best for given games.
constexpr std::size_t max_stream_size = 0x800; constexpr std::size_t max_stream_size = 0x800;
if (size < max_stream_size) { if (use_fast_cbuf || size < max_stream_size) {
if (!is_written && !IsRegionWritten(cache_addr, cache_addr + size - 1)) { if (!is_written && !IsRegionWritten(cache_addr, cache_addr + size - 1)) {
return StreamBufferUpload(host_ptr, size, alignment); if (use_fast_cbuf) {
return ConstBufferUpload(host_ptr, size);
} else {
return StreamBufferUpload(host_ptr, size, alignment);
}
} }
} }
@ -152,6 +156,10 @@ protected:
virtual void CopyBlock(const TBuffer& src, const TBuffer& dst, std::size_t src_offset, virtual void CopyBlock(const TBuffer& src, const TBuffer& dst, std::size_t src_offset,
std::size_t dst_offset, std::size_t size) = 0; std::size_t dst_offset, std::size_t size) = 0;
virtual BufferInfo ConstBufferUpload(const void* raw_pointer, std::size_t size) {
return {};
}
/// Register an object into the cache /// Register an object into the cache
void Register(const MapInterval& new_map, bool inherit_written = false) { void Register(const MapInterval& new_map, bool inherit_written = false) {
const CacheAddr cache_ptr = new_map->GetStart(); const CacheAddr cache_ptr = new_map->GetStart();

View File

@ -8,13 +8,17 @@
#include "common/assert.h" #include "common/assert.h"
#include "common/microprofile.h" #include "common/microprofile.h"
#include "video_core/engines/maxwell_3d.h"
#include "video_core/rasterizer_interface.h" #include "video_core/rasterizer_interface.h"
#include "video_core/renderer_opengl/gl_buffer_cache.h" #include "video_core/renderer_opengl/gl_buffer_cache.h"
#include "video_core/renderer_opengl/gl_device.h"
#include "video_core/renderer_opengl/gl_rasterizer.h" #include "video_core/renderer_opengl/gl_rasterizer.h"
#include "video_core/renderer_opengl/gl_resource_manager.h" #include "video_core/renderer_opengl/gl_resource_manager.h"
namespace OpenGL { namespace OpenGL {
using Maxwell = Tegra::Engines::Maxwell3D::Regs;
MICROPROFILE_DEFINE(OpenGL_Buffer_Download, "OpenGL", "Buffer Download", MP_RGB(192, 192, 128)); MICROPROFILE_DEFINE(OpenGL_Buffer_Download, "OpenGL", "Buffer Download", MP_RGB(192, 192, 128));
CachedBufferBlock::CachedBufferBlock(CacheAddr cache_addr, const std::size_t size) CachedBufferBlock::CachedBufferBlock(CacheAddr cache_addr, const std::size_t size)
@ -26,11 +30,22 @@ CachedBufferBlock::CachedBufferBlock(CacheAddr cache_addr, const std::size_t siz
CachedBufferBlock::~CachedBufferBlock() = default; CachedBufferBlock::~CachedBufferBlock() = default;
OGLBufferCache::OGLBufferCache(RasterizerOpenGL& rasterizer, Core::System& system, OGLBufferCache::OGLBufferCache(RasterizerOpenGL& rasterizer, Core::System& system,
std::size_t stream_size) const Device& device, std::size_t stream_size)
: VideoCommon::BufferCache<Buffer, GLuint, OGLStreamBuffer>{ : GenericBufferCache{rasterizer, system, std::make_unique<OGLStreamBuffer>(stream_size, true)} {
rasterizer, system, std::make_unique<OGLStreamBuffer>(stream_size, true)} {} if (!device.HasFastBufferSubData()) {
return;
}
OGLBufferCache::~OGLBufferCache() = default; static constexpr auto size = static_cast<GLsizeiptr>(Maxwell::MaxConstBufferSize);
glCreateBuffers(static_cast<GLsizei>(std::size(cbufs)), std::data(cbufs));
for (const GLuint cbuf : cbufs) {
glNamedBufferData(cbuf, size, nullptr, GL_STREAM_DRAW);
}
}
OGLBufferCache::~OGLBufferCache() {
glDeleteBuffers(static_cast<GLsizei>(std::size(cbufs)), std::data(cbufs));
}
Buffer OGLBufferCache::CreateBlock(CacheAddr cache_addr, std::size_t size) { Buffer OGLBufferCache::CreateBlock(CacheAddr cache_addr, std::size_t size) {
return std::make_shared<CachedBufferBlock>(cache_addr, size); return std::make_shared<CachedBufferBlock>(cache_addr, size);
@ -69,4 +84,12 @@ void OGLBufferCache::CopyBlock(const Buffer& src, const Buffer& dst, std::size_t
static_cast<GLsizeiptr>(size)); static_cast<GLsizeiptr>(size));
} }
OGLBufferCache::BufferInfo OGLBufferCache::ConstBufferUpload(const void* raw_pointer,
std::size_t size) {
DEBUG_ASSERT(cbuf_cursor < std::size(cbufs));
const GLuint& cbuf = cbufs[cbuf_cursor++];
glNamedBufferSubData(cbuf, 0, static_cast<GLsizeiptr>(size), raw_pointer);
return {&cbuf, 0};
}
} // namespace OpenGL } // namespace OpenGL

View File

@ -4,10 +4,12 @@
#pragma once #pragma once
#include <array>
#include <memory> #include <memory>
#include "common/common_types.h" #include "common/common_types.h"
#include "video_core/buffer_cache/buffer_cache.h" #include "video_core/buffer_cache/buffer_cache.h"
#include "video_core/engines/maxwell_3d.h"
#include "video_core/rasterizer_cache.h" #include "video_core/rasterizer_cache.h"
#include "video_core/renderer_opengl/gl_resource_manager.h" #include "video_core/renderer_opengl/gl_resource_manager.h"
#include "video_core/renderer_opengl/gl_stream_buffer.h" #include "video_core/renderer_opengl/gl_stream_buffer.h"
@ -18,12 +20,14 @@ class System;
namespace OpenGL { namespace OpenGL {
class Device;
class OGLStreamBuffer; class OGLStreamBuffer;
class RasterizerOpenGL; class RasterizerOpenGL;
class CachedBufferBlock; class CachedBufferBlock;
using Buffer = std::shared_ptr<CachedBufferBlock>; using Buffer = std::shared_ptr<CachedBufferBlock>;
using GenericBufferCache = VideoCommon::BufferCache<Buffer, GLuint, OGLStreamBuffer>;
class CachedBufferBlock : public VideoCommon::BufferBlock { class CachedBufferBlock : public VideoCommon::BufferBlock {
public: public:
@ -38,14 +42,18 @@ private:
OGLBuffer gl_buffer{}; OGLBuffer gl_buffer{};
}; };
class OGLBufferCache final : public VideoCommon::BufferCache<Buffer, GLuint, OGLStreamBuffer> { class OGLBufferCache final : public GenericBufferCache {
public: public:
explicit OGLBufferCache(RasterizerOpenGL& rasterizer, Core::System& system, explicit OGLBufferCache(RasterizerOpenGL& rasterizer, Core::System& system,
std::size_t stream_size); const Device& device, std::size_t stream_size);
~OGLBufferCache(); ~OGLBufferCache();
const GLuint* GetEmptyBuffer(std::size_t) override; const GLuint* GetEmptyBuffer(std::size_t) override;
void Acquire() noexcept {
cbuf_cursor = 0;
}
protected: protected:
Buffer CreateBlock(CacheAddr cache_addr, std::size_t size) override; Buffer CreateBlock(CacheAddr cache_addr, std::size_t size) override;
@ -61,6 +69,14 @@ protected:
void CopyBlock(const Buffer& src, const Buffer& dst, std::size_t src_offset, void CopyBlock(const Buffer& src, const Buffer& dst, std::size_t src_offset,
std::size_t dst_offset, std::size_t size) override; std::size_t dst_offset, std::size_t size) override;
BufferInfo ConstBufferUpload(const void* raw_pointer, std::size_t size) override;
private:
std::size_t cbuf_cursor = 0;
std::array<GLuint, Tegra::Engines::Maxwell3D::Regs::MaxConstBuffers *
Tegra::Engines::Maxwell3D::Regs::MaxShaderProgram>
cbufs;
}; };
} // namespace OpenGL } // namespace OpenGL

View File

@ -51,8 +51,11 @@ bool HasExtension(const std::vector<std::string_view>& images, std::string_view
} // Anonymous namespace } // Anonymous namespace
Device::Device() { Device::Device() {
const std::string_view vendor = reinterpret_cast<const char*>(glGetString(GL_VENDOR));
const std::vector extensions = GetExtensions(); const std::vector extensions = GetExtensions();
const bool is_nvidia = vendor == "NVIDIA Corporation";
uniform_buffer_alignment = GetInteger<std::size_t>(GL_UNIFORM_BUFFER_OFFSET_ALIGNMENT); uniform_buffer_alignment = GetInteger<std::size_t>(GL_UNIFORM_BUFFER_OFFSET_ALIGNMENT);
shader_storage_alignment = GetInteger<std::size_t>(GL_SHADER_STORAGE_BUFFER_OFFSET_ALIGNMENT); shader_storage_alignment = GetInteger<std::size_t>(GL_SHADER_STORAGE_BUFFER_OFFSET_ALIGNMENT);
max_vertex_attributes = GetInteger<u32>(GL_MAX_VERTEX_ATTRIBS); max_vertex_attributes = GetInteger<u32>(GL_MAX_VERTEX_ATTRIBS);
@ -64,6 +67,7 @@ Device::Device() {
has_variable_aoffi = TestVariableAoffi(); has_variable_aoffi = TestVariableAoffi();
has_component_indexing_bug = TestComponentIndexingBug(); has_component_indexing_bug = TestComponentIndexingBug();
has_precise_bug = TestPreciseBug(); has_precise_bug = TestPreciseBug();
has_fast_buffer_sub_data = is_nvidia;
LOG_INFO(Render_OpenGL, "Renderer_VariableAOFFI: {}", has_variable_aoffi); LOG_INFO(Render_OpenGL, "Renderer_VariableAOFFI: {}", has_variable_aoffi);
LOG_INFO(Render_OpenGL, "Renderer_ComponentIndexingBug: {}", has_component_indexing_bug); LOG_INFO(Render_OpenGL, "Renderer_ComponentIndexingBug: {}", has_component_indexing_bug);

View File

@ -54,6 +54,10 @@ public:
return has_precise_bug; return has_precise_bug;
} }
bool HasFastBufferSubData() const {
return has_fast_buffer_sub_data;
}
private: private:
static bool TestVariableAoffi(); static bool TestVariableAoffi();
static bool TestComponentIndexingBug(); static bool TestComponentIndexingBug();
@ -69,6 +73,7 @@ private:
bool has_variable_aoffi{}; bool has_variable_aoffi{};
bool has_component_indexing_bug{}; bool has_component_indexing_bug{};
bool has_precise_bug{}; bool has_precise_bug{};
bool has_fast_buffer_sub_data{};
}; };
} // namespace OpenGL } // namespace OpenGL

View File

@ -67,7 +67,7 @@ static std::size_t GetConstBufferSize(const Tegra::Engines::ConstBufferInfo& buf
RasterizerOpenGL::RasterizerOpenGL(Core::System& system, Core::Frontend::EmuWindow& emu_window, RasterizerOpenGL::RasterizerOpenGL(Core::System& system, Core::Frontend::EmuWindow& emu_window,
ScreenInfo& info) ScreenInfo& info)
: texture_cache{system, *this, device}, shader_cache{*this, system, emu_window, device}, : texture_cache{system, *this, device}, shader_cache{*this, system, emu_window, device},
system{system}, screen_info{info}, buffer_cache{*this, system, STREAM_BUFFER_SIZE} { system{system}, screen_info{info}, buffer_cache{*this, system, device, STREAM_BUFFER_SIZE} {
shader_program_manager = std::make_unique<GLShader::ProgramManager>(); shader_program_manager = std::make_unique<GLShader::ProgramManager>();
state.draw.shader_program = 0; state.draw.shader_program = 0;
state.Apply(); state.Apply();
@ -558,6 +558,8 @@ void RasterizerOpenGL::DrawPrelude() {
SyncPolygonOffset(); SyncPolygonOffset();
SyncAlphaTest(); SyncAlphaTest();
buffer_cache.Acquire();
// Draw the vertex batch // Draw the vertex batch
const bool is_indexed = accelerate_draw == AccelDraw::Indexed; const bool is_indexed = accelerate_draw == AccelDraw::Indexed;
@ -573,9 +575,11 @@ void RasterizerOpenGL::DrawPrelude() {
(sizeof(GLShader::MaxwellUniformData) + device.GetUniformBufferAlignment()) * (sizeof(GLShader::MaxwellUniformData) + device.GetUniformBufferAlignment()) *
Maxwell::MaxShaderStage; Maxwell::MaxShaderStage;
// Add space for at least 18 constant buffers if (!device.HasFastBufferSubData()) {
buffer_size += Maxwell::MaxConstBuffers * // Add space for at least 18 constant buffers
(Maxwell::MaxConstBufferSize + device.GetUniformBufferAlignment()); buffer_size += Maxwell::MaxConstBuffers *
(Maxwell::MaxConstBufferSize + device.GetUniformBufferAlignment());
}
// Prepare the vertex array. // Prepare the vertex array.
buffer_cache.Map(buffer_size); buffer_cache.Map(buffer_size);
@ -739,10 +743,12 @@ void RasterizerOpenGL::DispatchCompute(GPUVAddr code_addr) {
state.draw.shader_program = program; state.draw.shader_program = program;
state.draw.program_pipeline = 0; state.draw.program_pipeline = 0;
const std::size_t buffer_size = if (!device.HasFastBufferSubData()) {
Tegra::Engines::KeplerCompute::NumConstBuffers * const std::size_t buffer_size =
(Maxwell::MaxConstBufferSize + device.GetUniformBufferAlignment()); Tegra::Engines::KeplerCompute::NumConstBuffers *
buffer_cache.Map(buffer_size); (Maxwell::MaxConstBufferSize + device.GetUniformBufferAlignment());
buffer_cache.Map(buffer_size);
}
bind_ubo_pushbuffer.Setup(0); bind_ubo_pushbuffer.Setup(0);
bind_ssbo_pushbuffer.Setup(0); bind_ssbo_pushbuffer.Setup(0);
@ -750,7 +756,9 @@ void RasterizerOpenGL::DispatchCompute(GPUVAddr code_addr) {
SetupComputeConstBuffers(kernel); SetupComputeConstBuffers(kernel);
SetupComputeGlobalMemory(kernel); SetupComputeGlobalMemory(kernel);
buffer_cache.Unmap(); if (!device.HasFastBufferSubData()) {
buffer_cache.Unmap();
}
bind_ubo_pushbuffer.Bind(); bind_ubo_pushbuffer.Bind();
bind_ssbo_pushbuffer.Bind(); bind_ssbo_pushbuffer.Bind();
@ -879,7 +887,8 @@ void RasterizerOpenGL::SetupConstBuffer(const Tegra::Engines::ConstBufferInfo& b
const std::size_t size = Common::AlignUp(GetConstBufferSize(buffer, entry), sizeof(GLvec4)); const std::size_t size = Common::AlignUp(GetConstBufferSize(buffer, entry), sizeof(GLvec4));
const auto alignment = device.GetUniformBufferAlignment(); const auto alignment = device.GetUniformBufferAlignment();
const auto [cbuf, offset] = buffer_cache.UploadMemory(buffer.address, size, alignment); const auto [cbuf, offset] = buffer_cache.UploadMemory(buffer.address, size, alignment, false,
device.HasFastBufferSubData());
bind_ubo_pushbuffer.Push(cbuf, offset, size); bind_ubo_pushbuffer.Push(cbuf, offset, size);
} }