Merge pull request #2413 from FernandoS27/opt-gpu

Rasterizer Cache: refactor flushing & optimize memory usage of surfaces
This commit is contained in:
Rodrigo Locatti 2019-05-13 23:01:59 -03:00 committed by GitHub
commit 940a71089d
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
7 changed files with 54 additions and 33 deletions

View file

@ -37,9 +37,6 @@ public:
/// Gets the size of the shader in guest memory, required for cache management /// Gets the size of the shader in guest memory, required for cache management
virtual std::size_t GetSizeInBytes() const = 0; virtual std::size_t GetSizeInBytes() const = 0;
/// Wriets any cached resources back to memory
virtual void Flush() = 0;
/// Sets whether the cached object should be considered registered /// Sets whether the cached object should be considered registered
void SetIsRegistered(bool registered) { void SetIsRegistered(bool registered) {
is_registered = registered; is_registered = registered;
@ -158,6 +155,8 @@ protected:
return ++modified_ticks; return ++modified_ticks;
} }
virtual void FlushObjectInner(const T& object) = 0;
/// Flushes the specified object, updating appropriate cache state as needed /// Flushes the specified object, updating appropriate cache state as needed
void FlushObject(const T& object) { void FlushObject(const T& object) {
std::lock_guard lock{mutex}; std::lock_guard lock{mutex};
@ -165,7 +164,7 @@ protected:
if (!object->IsDirty()) { if (!object->IsDirty()) {
return; return;
} }
object->Flush(); FlushObjectInner(object);
object->MarkAsModified(false, *this); object->MarkAsModified(false, *this);
} }

View file

@ -42,9 +42,6 @@ public:
return alignment; return alignment;
} }
// We do not have to flush this cache as things in it are never modified by us.
void Flush() override {}
private: private:
VAddr cpu_addr{}; VAddr cpu_addr{};
std::size_t size{}; std::size_t size{};
@ -75,6 +72,9 @@ public:
protected: protected:
void AlignBuffer(std::size_t alignment); void AlignBuffer(std::size_t alignment);
// We do not have to flush this cache as things in it are never modified by us.
void FlushObjectInner(const std::shared_ptr<CachedBufferEntry>& object) override {}
private: private:
OGLStreamBuffer stream_buffer; OGLStreamBuffer stream_buffer;

View file

@ -46,7 +46,7 @@ public:
/// Reloads the global region from guest memory /// Reloads the global region from guest memory
void Reload(u32 size_); void Reload(u32 size_);
void Flush() override; void Flush();
private: private:
VAddr cpu_addr{}; VAddr cpu_addr{};
@ -65,6 +65,11 @@ public:
GlobalRegion GetGlobalRegion(const GLShader::GlobalMemoryEntry& descriptor, GlobalRegion GetGlobalRegion(const GLShader::GlobalMemoryEntry& descriptor,
Tegra::Engines::Maxwell3D::Regs::ShaderStage stage); Tegra::Engines::Maxwell3D::Regs::ShaderStage stage);
protected:
void FlushObjectInner(const GlobalRegion& object) override {
object->Flush();
}
private: private:
GlobalRegion TryGetReservedGlobalRegion(CacheAddr addr, u32 size) const; GlobalRegion TryGetReservedGlobalRegion(CacheAddr addr, u32 size) const;
GlobalRegion GetUncachedGlobalRegion(GPUVAddr addr, u8* host_ptr, u32 size); GlobalRegion GetUncachedGlobalRegion(GPUVAddr addr, u8* host_ptr, u32 size);

View file

@ -628,8 +628,10 @@ CachedSurface::CachedSurface(const SurfaceParams& params)
} }
MICROPROFILE_DEFINE(OpenGL_SurfaceLoad, "OpenGL", "Surface Load", MP_RGB(128, 192, 64)); MICROPROFILE_DEFINE(OpenGL_SurfaceLoad, "OpenGL", "Surface Load", MP_RGB(128, 192, 64));
void CachedSurface::LoadGLBuffer() { void CachedSurface::LoadGLBuffer(RasterizerTemporaryMemory& res_cache_tmp_mem) {
MICROPROFILE_SCOPE(OpenGL_SurfaceLoad); MICROPROFILE_SCOPE(OpenGL_SurfaceLoad);
auto& gl_buffer = res_cache_tmp_mem.gl_buffer;
if (gl_buffer.size() < params.max_mip_level)
gl_buffer.resize(params.max_mip_level); gl_buffer.resize(params.max_mip_level);
for (u32 i = 0; i < params.max_mip_level; i++) for (u32 i = 0; i < params.max_mip_level; i++)
gl_buffer[i].resize(params.GetMipmapSizeGL(i)); gl_buffer[i].resize(params.GetMipmapSizeGL(i));
@ -671,13 +673,13 @@ void CachedSurface::LoadGLBuffer() {
} }
MICROPROFILE_DEFINE(OpenGL_SurfaceFlush, "OpenGL", "Surface Flush", MP_RGB(128, 192, 64)); MICROPROFILE_DEFINE(OpenGL_SurfaceFlush, "OpenGL", "Surface Flush", MP_RGB(128, 192, 64));
void CachedSurface::FlushGLBuffer() { void CachedSurface::FlushGLBuffer(RasterizerTemporaryMemory& res_cache_tmp_mem) {
MICROPROFILE_SCOPE(OpenGL_SurfaceFlush); MICROPROFILE_SCOPE(OpenGL_SurfaceFlush);
ASSERT_MSG(!IsPixelFormatASTC(params.pixel_format), "Unimplemented"); ASSERT_MSG(!IsPixelFormatASTC(params.pixel_format), "Unimplemented");
auto& gl_buffer = res_cache_tmp_mem.gl_buffer;
// OpenGL temporary buffer needs to be big enough to store raw texture size // OpenGL temporary buffer needs to be big enough to store raw texture size
gl_buffer.resize(1);
gl_buffer[0].resize(GetSizeInBytes()); gl_buffer[0].resize(GetSizeInBytes());
const FormatTuple& tuple = GetFormatTuple(params.pixel_format, params.component_type); const FormatTuple& tuple = GetFormatTuple(params.pixel_format, params.component_type);
@ -713,10 +715,12 @@ void CachedSurface::FlushGLBuffer() {
} }
} }
void CachedSurface::UploadGLMipmapTexture(u32 mip_map, GLuint read_fb_handle, void CachedSurface::UploadGLMipmapTexture(RasterizerTemporaryMemory& res_cache_tmp_mem, u32 mip_map,
GLuint draw_fb_handle) { GLuint read_fb_handle, GLuint draw_fb_handle) {
const auto& rect{params.GetRect(mip_map)}; const auto& rect{params.GetRect(mip_map)};
auto& gl_buffer = res_cache_tmp_mem.gl_buffer;
// Load data from memory to the surface // Load data from memory to the surface
const auto x0 = static_cast<GLint>(rect.left); const auto x0 = static_cast<GLint>(rect.left);
const auto y0 = static_cast<GLint>(rect.bottom); const auto y0 = static_cast<GLint>(rect.bottom);
@ -845,11 +849,12 @@ void CachedSurface::EnsureTextureDiscrepantView() {
} }
MICROPROFILE_DEFINE(OpenGL_TextureUL, "OpenGL", "Texture Upload", MP_RGB(128, 192, 64)); MICROPROFILE_DEFINE(OpenGL_TextureUL, "OpenGL", "Texture Upload", MP_RGB(128, 192, 64));
void CachedSurface::UploadGLTexture(GLuint read_fb_handle, GLuint draw_fb_handle) { void CachedSurface::UploadGLTexture(RasterizerTemporaryMemory& res_cache_tmp_mem,
GLuint read_fb_handle, GLuint draw_fb_handle) {
MICROPROFILE_SCOPE(OpenGL_TextureUL); MICROPROFILE_SCOPE(OpenGL_TextureUL);
for (u32 i = 0; i < params.max_mip_level; i++) for (u32 i = 0; i < params.max_mip_level; i++)
UploadGLMipmapTexture(i, read_fb_handle, draw_fb_handle); UploadGLMipmapTexture(res_cache_tmp_mem, i, read_fb_handle, draw_fb_handle);
} }
void CachedSurface::UpdateSwizzle(Tegra::Texture::SwizzleSource swizzle_x, void CachedSurface::UpdateSwizzle(Tegra::Texture::SwizzleSource swizzle_x,
@ -929,8 +934,8 @@ Surface RasterizerCacheOpenGL::GetColorBufferSurface(std::size_t index, bool pre
} }
void RasterizerCacheOpenGL::LoadSurface(const Surface& surface) { void RasterizerCacheOpenGL::LoadSurface(const Surface& surface) {
surface->LoadGLBuffer(); surface->LoadGLBuffer(temporal_memory);
surface->UploadGLTexture(read_framebuffer.handle, draw_framebuffer.handle); surface->UploadGLTexture(temporal_memory, read_framebuffer.handle, draw_framebuffer.handle);
surface->MarkAsModified(false, *this); surface->MarkAsModified(false, *this);
surface->MarkForReload(false); surface->MarkForReload(false);
} }

View file

@ -355,6 +355,12 @@ namespace OpenGL {
class RasterizerOpenGL; class RasterizerOpenGL;
// This is used to store temporary big buffers,
// instead of creating/destroying all the time
struct RasterizerTemporaryMemory {
std::vector<std::vector<u8>> gl_buffer;
};
class CachedSurface final : public RasterizerCacheObject { class CachedSurface final : public RasterizerCacheObject {
public: public:
explicit CachedSurface(const SurfaceParams& params); explicit CachedSurface(const SurfaceParams& params);
@ -371,10 +377,6 @@ public:
return memory_size; return memory_size;
} }
void Flush() override {
FlushGLBuffer();
}
const OGLTexture& Texture() const { const OGLTexture& Texture() const {
return texture; return texture;
} }
@ -397,11 +399,12 @@ public:
} }
// Read/Write data in Switch memory to/from gl_buffer // Read/Write data in Switch memory to/from gl_buffer
void LoadGLBuffer(); void LoadGLBuffer(RasterizerTemporaryMemory& res_cache_tmp_mem);
void FlushGLBuffer(); void FlushGLBuffer(RasterizerTemporaryMemory& res_cache_tmp_mem);
// Upload data in gl_buffer to this surface's texture // Upload data in gl_buffer to this surface's texture
void UploadGLTexture(GLuint read_fb_handle, GLuint draw_fb_handle); void UploadGLTexture(RasterizerTemporaryMemory& res_cache_tmp_mem, GLuint read_fb_handle,
GLuint draw_fb_handle);
void UpdateSwizzle(Tegra::Texture::SwizzleSource swizzle_x, void UpdateSwizzle(Tegra::Texture::SwizzleSource swizzle_x,
Tegra::Texture::SwizzleSource swizzle_y, Tegra::Texture::SwizzleSource swizzle_y,
@ -429,13 +432,13 @@ public:
} }
private: private:
void UploadGLMipmapTexture(u32 mip_map, GLuint read_fb_handle, GLuint draw_fb_handle); void UploadGLMipmapTexture(RasterizerTemporaryMemory& res_cache_tmp_mem, u32 mip_map,
GLuint read_fb_handle, GLuint draw_fb_handle);
void EnsureTextureDiscrepantView(); void EnsureTextureDiscrepantView();
OGLTexture texture; OGLTexture texture;
OGLTexture discrepant_view; OGLTexture discrepant_view;
std::vector<std::vector<u8>> gl_buffer;
SurfaceParams params{}; SurfaceParams params{};
GLenum gl_target{}; GLenum gl_target{};
GLenum gl_internal_format{}; GLenum gl_internal_format{};
@ -473,6 +476,11 @@ public:
void SignalPreDrawCall(); void SignalPreDrawCall();
void SignalPostDrawCall(); void SignalPostDrawCall();
protected:
void FlushObjectInner(const Surface& object) override {
object->FlushGLBuffer(temporal_memory);
}
private: private:
void LoadSurface(const Surface& surface); void LoadSurface(const Surface& surface);
Surface GetSurface(const SurfaceParams& params, bool preserve_contents = true); Surface GetSurface(const SurfaceParams& params, bool preserve_contents = true);
@ -519,6 +527,8 @@ private:
std::array<Surface, Maxwell::NumRenderTargets> current_color_buffers; std::array<Surface, Maxwell::NumRenderTargets> current_color_buffers;
Surface last_depth_buffer; Surface last_depth_buffer;
RasterizerTemporaryMemory temporal_memory;
using SurfaceIntervalCache = boost::icl::interval_map<CacheAddr, Surface>; using SurfaceIntervalCache = boost::icl::interval_map<CacheAddr, Surface>;
using SurfaceInterval = typename SurfaceIntervalCache::interval_type; using SurfaceInterval = typename SurfaceIntervalCache::interval_type;

View file

@ -57,9 +57,6 @@ public:
return shader_length; return shader_length;
} }
// We do not have to flush this cache as things in it are never modified by us.
void Flush() override {}
/// Gets the shader entries for the shader /// Gets the shader entries for the shader
const GLShader::ShaderEntries& GetShaderEntries() const { const GLShader::ShaderEntries& GetShaderEntries() const {
return entries; return entries;
@ -123,6 +120,10 @@ public:
/// Gets the current specified shader stage program /// Gets the current specified shader stage program
Shader GetStageProgram(Maxwell::ShaderProgram program); Shader GetStageProgram(Maxwell::ShaderProgram program);
protected:
// We do not have to flush this cache as things in it are never modified by us.
void FlushObjectInner(const Shader& object) override {}
private: private:
std::unordered_map<u64, UnspecializedShader> GenerateUnspecializedShaders( std::unordered_map<u64, UnspecializedShader> GenerateUnspecializedShaders(
const std::atomic_bool& stop_loading, const VideoCore::DiskResourceLoadCallback& callback, const std::atomic_bool& stop_loading, const VideoCore::DiskResourceLoadCallback& callback,

View file

@ -49,9 +49,6 @@ public:
return alignment; return alignment;
} }
// We do not have to flush this cache as things in it are never modified by us.
void Flush() override {}
private: private:
VAddr cpu_addr{}; VAddr cpu_addr{};
std::size_t size{}; std::size_t size{};
@ -87,6 +84,10 @@ public:
return buffer_handle; return buffer_handle;
} }
protected:
// We do not have to flush this cache as things in it are never modified by us.
void FlushObjectInner(const std::shared_ptr<CachedBufferEntry>& object) override {}
private: private:
void AlignBuffer(std::size_t alignment); void AlignBuffer(std::size_t alignment);