Merge pull request #3700 from wwylele/texcache-watcher

gl_rasterizer_cache: cache texture cube

src/video_core/renderer_opengl/gl_rasterizer.cpp

@@ -45,7 +45,6 @@ RasterizerOpenGL::RasterizerOpenGL()
     // Create cubemap texture and sampler objects
     texture_cube_sampler.Create();
     state.texture_cube_unit.sampler = texture_cube_sampler.sampler.handle;
-    texture_cube.Create();
 
     // Generate VBO, VAO and UBO
     vertex_array.Create();
@@ -393,19 +392,18 @@ void RasterizerOpenGL::DrawTriangles() {
                 switch (texture.config.type.Value()) {
                 case TextureType::TextureCube:
                     using CubeFace = Pica::TexturingRegs::CubeFace;
-                    if (res_cache.FillTextureCube(
+                    TextureCubeConfig config;
-                            texture_cube.handle, texture,
+                    config.px = regs.texturing.GetCubePhysicalAddress(CubeFace::PositiveX);
-                            regs.texturing.GetCubePhysicalAddress(CubeFace::PositiveX),
+                    config.nx = regs.texturing.GetCubePhysicalAddress(CubeFace::NegativeX);
-                            regs.texturing.GetCubePhysicalAddress(CubeFace::NegativeX),
+                    config.py = regs.texturing.GetCubePhysicalAddress(CubeFace::PositiveY);
-                            regs.texturing.GetCubePhysicalAddress(CubeFace::PositiveY),
+                    config.ny = regs.texturing.GetCubePhysicalAddress(CubeFace::NegativeY);
-                            regs.texturing.GetCubePhysicalAddress(CubeFace::NegativeY),
+                    config.pz = regs.texturing.GetCubePhysicalAddress(CubeFace::PositiveZ);
-                            regs.texturing.GetCubePhysicalAddress(CubeFace::PositiveZ),
+                    config.nz = regs.texturing.GetCubePhysicalAddress(CubeFace::NegativeZ);
-                            regs.texturing.GetCubePhysicalAddress(CubeFace::NegativeZ))) {
+                    config.width = texture.config.width;
-                        state.texture_cube_unit.texture_cube = texture_cube.handle;
+                    config.format = texture.format;
-                    } else {
+                    state.texture_cube_unit.texture_cube =
-                        // Can occur when texture addr is null or its memory is unmapped/invalid
+                        res_cache.GetTextureCube(config).texture.handle;
-                        state.texture_cube_unit.texture_cube = 0;
+
-                    }
                     texture_cube_sampler.SyncWithConfig(texture.config);
                     state.texture_units[texture_index].texture_2d = 0;
                     continue; // Texture unit 0 setup finished. Continue to next unit

src/video_core/renderer_opengl/gl_rasterizer.h

@@ -252,8 +252,6 @@ private:
     GLint uniform_buffer_alignment;
     size_t uniform_size_aligned_fs;
 
-    // TODO (wwylele): consider caching texture cube in the rasterizer cache
-    OGLTexture texture_cube;
     SamplerInfo texture_cube_sampler;
 
     OGLBuffer lighting_lut_buffer;

src/video_core/renderer_opengl/gl_rasterizer_cache.cpp

@@ -28,7 +28,6 @@
 #include "video_core/pica_state.h"
 #include "video_core/renderer_opengl/gl_rasterizer_cache.h"
 #include "video_core/renderer_opengl/gl_state.h"
-#include "video_core/texture/texture_decode.h"
 #include "video_core/utils.h"
 #include "video_core/video_core.h"
 
@@ -231,6 +230,32 @@ static void AllocateSurfaceTexture(GLuint texture, const FormatTuple& format_tup
     cur_state.Apply();
 }
 
+static void AllocateTextureCube(GLuint texture, const FormatTuple& format_tuple, u32 width) {
+    OpenGLState cur_state = OpenGLState::GetCurState();
+
+    // Keep track of previous texture bindings
+    GLuint old_tex = cur_state.texture_cube_unit.texture_cube;
+    cur_state.texture_cube_unit.texture_cube = texture;
+    cur_state.Apply();
+    glActiveTexture(TextureUnits::TextureCube.Enum());
+
+    for (auto faces : {
+             GL_TEXTURE_CUBE_MAP_POSITIVE_X,
+             GL_TEXTURE_CUBE_MAP_POSITIVE_Y,
+             GL_TEXTURE_CUBE_MAP_POSITIVE_Z,
+             GL_TEXTURE_CUBE_MAP_NEGATIVE_X,
+             GL_TEXTURE_CUBE_MAP_NEGATIVE_Y,
+             GL_TEXTURE_CUBE_MAP_NEGATIVE_Z,
+         }) {
+        glTexImage2D(faces, 0, format_tuple.internal_format, width, width, 0, format_tuple.format,
+                     format_tuple.type, nullptr);
+    }
+
+    // Restore previous texture bindings
+    cur_state.texture_cube_unit.texture_cube = old_tex;
+    cur_state.Apply();
+}
+
 static bool BlitTextures(GLuint src_tex, const MathUtil::Rectangle<u32>& src_rect, GLuint dst_tex,
                          const MathUtil::Rectangle<u32>& dst_rect, SurfaceType type,
                          GLuint read_fb_handle, GLuint draw_fb_handle) {
@@ -761,6 +786,8 @@ void CachedSurface::UploadGLTexture(const MathUtil::Rectangle<u32>& rect, GLuint
         BlitTextures(unscaled_tex.handle, {0, rect.GetHeight(), rect.GetWidth(), 0}, texture.handle,
                      scaled_rect, type, read_fb_handle, draw_fb_handle);
     }
+
+    InvalidateAllWatcher();
 }
 
 MICROPROFILE_DEFINE(OpenGL_TextureDL, "OpenGL", "Texture Download", MP_RGB(128, 192, 64));
@@ -1193,14 +1220,13 @@ SurfaceRect_Tuple RasterizerCacheOpenGL::GetSurfaceSubRect(const SurfaceParams&
 }
 
 Surface RasterizerCacheOpenGL::GetTextureSurface(
-    const Pica::TexturingRegs::FullTextureConfig& config, PAddr addr_override) {
+    const Pica::TexturingRegs::FullTextureConfig& config) {
     Pica::Texture::TextureInfo info =
         Pica::Texture::TextureInfo::FromPicaRegister(config.config, config.format);
+    return GetTextureSurface(info);
+}
 
-    if (addr_override != 0) {
+Surface RasterizerCacheOpenGL::GetTextureSurface(const Pica::Texture::TextureInfo& info) {
-        info.physical_address = addr_override;
-    }
-
     SurfaceParams params;
     params.addr = info.physical_address;
     params.width = info.width;
@@ -1228,80 +1254,94 @@ Surface RasterizerCacheOpenGL::GetTextureSurface(
     return GetSurface(params, ScaleMatch::Ignore, true);
 }
 
-bool RasterizerCacheOpenGL::FillTextureCube(GLuint dest_handle,
+const CachedTextureCube& RasterizerCacheOpenGL::GetTextureCube(const TextureCubeConfig& config) {
-                                            const Pica::TexturingRegs::FullTextureConfig& config,
+    auto& cube = texture_cube_cache[config];
-                                            PAddr px, PAddr nx, PAddr py, PAddr ny, PAddr pz,
+
-                                            PAddr nz) {
+    struct Face {
-    ASSERT(config.config.width == config.config.height);
+        Face(std::shared_ptr<SurfaceWatcher>& watcher, PAddr address, GLenum gl_face)
-    struct FaceTuple {
+            : watcher(watcher), address(address), gl_face(gl_face) {}
+        std::shared_ptr<SurfaceWatcher>& watcher;
         PAddr address;
         GLenum gl_face;
-        Surface surface;
     };
-    std::array<FaceTuple, 6> faces{{
+
-        {px, GL_TEXTURE_CUBE_MAP_POSITIVE_X},
+    const std::array<Face, 6> faces{{
-        {nx, GL_TEXTURE_CUBE_MAP_NEGATIVE_X},
+        {cube.px, config.px, GL_TEXTURE_CUBE_MAP_POSITIVE_X},
-        {py, GL_TEXTURE_CUBE_MAP_POSITIVE_Y},
+        {cube.nx, config.nx, GL_TEXTURE_CUBE_MAP_NEGATIVE_X},
-        {ny, GL_TEXTURE_CUBE_MAP_NEGATIVE_Y},
+        {cube.py, config.py, GL_TEXTURE_CUBE_MAP_POSITIVE_Y},
-        {pz, GL_TEXTURE_CUBE_MAP_POSITIVE_Z},
+        {cube.ny, config.ny, GL_TEXTURE_CUBE_MAP_NEGATIVE_Y},
-        {nz, GL_TEXTURE_CUBE_MAP_NEGATIVE_Z},
+        {cube.pz, config.pz, GL_TEXTURE_CUBE_MAP_POSITIVE_Z},
+        {cube.nz, config.nz, GL_TEXTURE_CUBE_MAP_NEGATIVE_Z},
     }};
 
-    u16 res_scale = 1;
+    for (const Face& face : faces) {
-    for (auto& face : faces) {
+        if (!face.watcher || !face.watcher->Get()) {
-        face.surface = GetTextureSurface(config, face.address);
+            Pica::Texture::TextureInfo info;
-        if (face.surface == nullptr)
+            info.physical_address = face.address;
-            return false;
+            info.height = info.width = config.width;
-        res_scale = std::max(res_scale, face.surface->res_scale);
+            info.format = config.format;
+            info.SetDefaultStride();
+            auto surface = GetTextureSurface(info);
+            if (surface) {
+                face.watcher = surface->CreateWatcher();
+            } else {
+                // Can occur when texture address is invalid. We mark the watcher with nullptr in
+                // this case and the content of the face wouldn't get updated. These are usually
+                // leftover setup in the texture unit and games are not supposed to draw using them.
+                face.watcher = nullptr;
+            }
+        }
     }
 
-    u32 scaled_size = res_scale * config.config.width;
+    if (cube.texture.handle == 0) {
+        for (const Face& face : faces) {
+            if (face.watcher) {
+                auto surface = face.watcher->Get();
+                cube.res_scale = std::max(cube.res_scale, surface->res_scale);
+            }
+        }
+
+        cube.texture.Create();
+        AllocateTextureCube(
+            cube.texture.handle,
+            GetFormatTuple(CachedSurface::PixelFormatFromTextureFormat(config.format)),
+            cube.res_scale * config.width);
+    }
+
+    u32 scaled_size = cube.res_scale * config.width;
 
     OpenGLState state = OpenGLState::GetCurState();
 
     OpenGLState prev_state = state;
     SCOPE_EXIT({ prev_state.Apply(); });
 
-    state.texture_cube_unit.texture_cube = dest_handle;
-    state.Apply();
-    glActiveTexture(TextureUnits::TextureCube.Enum());
-    FormatTuple format_tuple = GetFormatTuple(faces[0].surface->pixel_format);
-
-    GLint cur_size, cur_format;
-    glGetTexLevelParameteriv(GL_TEXTURE_CUBE_MAP_POSITIVE_X, 0, GL_TEXTURE_WIDTH, &cur_size);
-    glGetTexLevelParameteriv(GL_TEXTURE_CUBE_MAP_POSITIVE_X, 0, GL_TEXTURE_INTERNAL_FORMAT,
-                             &cur_format);
-
-    if (cur_size != scaled_size || cur_format != format_tuple.internal_format) {
-        for (auto& face : faces) {
-            glTexImage2D(face.gl_face, 0, format_tuple.internal_format, scaled_size, scaled_size, 0,
-                         format_tuple.format, format_tuple.type, nullptr);
-        }
-    }
-
     state.draw.read_framebuffer = read_framebuffer.handle;
     state.draw.draw_framebuffer = draw_framebuffer.handle;
-    state.ResetTexture(dest_handle);
+    state.ResetTexture(cube.texture.handle);
 
-    for (auto& face : faces) {
+    for (const Face& face : faces) {
-        state.ResetTexture(face.surface->texture.handle);
+        if (face.watcher && !face.watcher->IsValid()) {
+            auto surface = face.watcher->Get();
+            state.ResetTexture(surface->texture.handle);
             state.Apply();
             glFramebufferTexture2D(GL_READ_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, GL_TEXTURE_2D,
-                               face.surface->texture.handle, 0);
+                                   surface->texture.handle, 0);
-        glFramebufferTexture2D(GL_READ_FRAMEBUFFER, GL_DEPTH_STENCIL_ATTACHMENT, GL_TEXTURE_2D, 0,
+            glFramebufferTexture2D(GL_READ_FRAMEBUFFER, GL_DEPTH_STENCIL_ATTACHMENT, GL_TEXTURE_2D,
-                               0);
+                                   0, 0);
 
-        glFramebufferTexture2D(GL_DRAW_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, face.gl_face, dest_handle,
+            glFramebufferTexture2D(GL_DRAW_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, face.gl_face,
-                               0);
+                                   cube.texture.handle, 0);
-        glFramebufferTexture2D(GL_DRAW_FRAMEBUFFER, GL_DEPTH_STENCIL_ATTACHMENT, GL_TEXTURE_2D, 0,
+            glFramebufferTexture2D(GL_DRAW_FRAMEBUFFER, GL_DEPTH_STENCIL_ATTACHMENT, GL_TEXTURE_2D,
-                               0);
+                                   0, 0);
 
-        auto src_rect = face.surface->GetScaledRect();
+            auto src_rect = surface->GetScaledRect();
             glBlitFramebuffer(src_rect.left, src_rect.bottom, src_rect.right, src_rect.top, 0, 0,
                               scaled_size, scaled_size, GL_COLOR_BUFFER_BIT, GL_LINEAR);
+            face.watcher->Validate();
+        }
     }
 
-    return true;
+    return cube;
 }
 
 SurfaceSurfaceRect_Tuple RasterizerCacheOpenGL::GetFramebufferSurfaces(
@@ -1316,6 +1356,7 @@ SurfaceSurfaceRect_Tuple RasterizerCacheOpenGL::GetFramebufferSurfaces(
         FlushAll();
         while (!surface_cache.empty())
             UnregisterSurface(*surface_cache.begin()->second.begin());
+        texture_cube_cache.clear();
     }
 
     MathUtil::Rectangle<u32> viewport_clamped{

src/video_core/renderer_opengl/gl_rasterizer_cache.h

@@ -5,6 +5,7 @@
 #pragma once
 
 #include <array>
+#include <list>
 #include <memory>
 #include <set>
 #include <tuple>
@@ -17,6 +18,8 @@
 #ifdef __GNUC__
 #pragma GCC diagnostic pop
 #endif
+#include <unordered_map>
+#include <boost/functional/hash.hpp>
 #include <glad/glad.h>
 #include "common/assert.h"
 #include "common/common_funcs.h"
@@ -26,6 +29,7 @@
 #include "video_core/regs_framebuffer.h"
 #include "video_core/regs_texturing.h"
 #include "video_core/renderer_opengl/gl_resource_manager.h"
+#include "video_core/texture/texture_decode.h"
 
 struct CachedSurface;
 using Surface = std::shared_ptr<CachedSurface>;
@@ -253,7 +257,41 @@ struct SurfaceParams {
     SurfaceType type = SurfaceType::Invalid;
 };
 
-struct CachedSurface : SurfaceParams {
+/**
+ * A watcher that notifies whether a cached surface has been changed. This is useful for caching
+ * surface collection objects, including texture cube and mipmap.
+ */
+struct SurfaceWatcher {
+public:
+    explicit SurfaceWatcher(std::weak_ptr<CachedSurface>&& surface) : surface(std::move(surface)) {}
+
+    /**
+     * Checks whether the surface has been changed.
+     * @return false if the surface content has been changed since last Validate() call or has been
+     * destroyed; otherwise true
+     */
+    bool IsValid() const {
+        return !surface.expired() && valid;
+    }
+
+    /// Marks that the content of the referencing surface has been updated to the watcher user.
+    void Validate() {
+        ASSERT(!surface.expired());
+        valid = true;
+    }
+
+    /// Gets the referencing surface. Returns null if the surface has been destroyed
+    Surface Get() const {
+        return surface.lock();
+    }
+
+private:
+    friend struct CachedSurface;
+    std::weak_ptr<CachedSurface> surface;
+    bool valid = false;
+};
+
+struct CachedSurface : SurfaceParams, std::enable_shared_from_this<CachedSurface> {
     bool CanFill(const SurfaceParams& dest_surface, SurfaceInterval fill_interval) const;
     bool CanCopy(const SurfaceParams& dest_surface, SurfaceInterval copy_interval) const;
 
@@ -294,6 +332,72 @@ struct CachedSurface : SurfaceParams {
                          GLuint draw_fb_handle);
     void DownloadGLTexture(const MathUtil::Rectangle<u32>& rect, GLuint read_fb_handle,
                            GLuint draw_fb_handle);
+
+    std::shared_ptr<SurfaceWatcher> CreateWatcher() {
+        auto watcher = std::make_shared<SurfaceWatcher>(weak_from_this());
+        watchers.push_front(watcher);
+        return watcher;
+    }
+
+    void InvalidateAllWatcher() {
+        for (const auto& watcher : watchers) {
+            if (auto locked = watcher.lock()) {
+                locked->valid = false;
+            }
+        }
+    }
+
+private:
+    std::list<std::weak_ptr<SurfaceWatcher>> watchers;
+};
+
+struct TextureCubeConfig {
+    PAddr px;
+    PAddr nx;
+    PAddr py;
+    PAddr ny;
+    PAddr pz;
+    PAddr nz;
+    u32 width;
+    Pica::TexturingRegs::TextureFormat format;
+
+    bool operator==(const TextureCubeConfig& rhs) const {
+        return std::tie(px, nx, py, ny, pz, nz, width, format) ==
+               std::tie(rhs.px, rhs.nx, rhs.py, rhs.ny, rhs.pz, rhs.nz, rhs.width, rhs.format);
+    }
+
+    bool operator!=(const TextureCubeConfig& rhs) const {
+        return !(*this == rhs);
+    }
+};
+
+namespace std {
+template <>
+struct hash<TextureCubeConfig> {
+    size_t operator()(const TextureCubeConfig& config) const {
+        std::size_t hash = 0;
+        boost::hash_combine(hash, config.px);
+        boost::hash_combine(hash, config.nx);
+        boost::hash_combine(hash, config.py);
+        boost::hash_combine(hash, config.ny);
+        boost::hash_combine(hash, config.pz);
+        boost::hash_combine(hash, config.nz);
+        boost::hash_combine(hash, config.width);
+        boost::hash_combine(hash, static_cast<u32>(config.format));
+        return hash;
+    }
+};
+} // namespace std
+
+struct CachedTextureCube {
+    OGLTexture texture;
+    u16 res_scale = 1;
+    std::shared_ptr<SurfaceWatcher> px;
+    std::shared_ptr<SurfaceWatcher> nx;
+    std::shared_ptr<SurfaceWatcher> py;
+    std::shared_ptr<SurfaceWatcher> ny;
+    std::shared_ptr<SurfaceWatcher> pz;
+    std::shared_ptr<SurfaceWatcher> nz;
 };
 
 class RasterizerCacheOpenGL : NonCopyable {
@@ -322,12 +426,11 @@ public:
                                         bool load_if_create);
 
     /// Get a surface based on the texture configuration
-    Surface GetTextureSurface(const Pica::TexturingRegs::FullTextureConfig& config,
+    Surface GetTextureSurface(const Pica::TexturingRegs::FullTextureConfig& config);
-                              PAddr addr_override = 0);
+    Surface GetTextureSurface(const Pica::Texture::TextureInfo& info);
 
-    /// Copy surfaces to a cubemap texture based on the texture configuration
+    /// Get a texture cube based on the texture configuration
-    bool FillTextureCube(GLuint dest_handle, const Pica::TexturingRegs::FullTextureConfig& config,
+    const CachedTextureCube& GetTextureCube(const TextureCubeConfig& config);
-                         PAddr px, PAddr nx, PAddr py, PAddr ny, PAddr pz, PAddr nz);
 
     /// Get the color and depth surfaces based on the framebuffer configuration
     SurfaceSurfaceRect_Tuple GetFramebufferSurfaces(bool using_color_fb, bool using_depth_fb,
@@ -380,4 +483,6 @@ private:
     OGLProgram d24s8_abgr_shader;
     GLint d24s8_abgr_tbo_size_u_id;
     GLint d24s8_abgr_viewport_u_id;
+
+    std::unordered_map<TextureCubeConfig, CachedTextureCube> texture_cube_cache;
 };