7 files changed, 229 insertions, 66 deletions

diff --git a/src/video_core/renderer_opengl/gl_rasterizer.cpp b/src/video_core/renderer_opengl/gl_rasterizer.cpp
index 516e1b50f..8bfa75b84 100644
--- a/src/video_core/renderer_opengl/gl_rasterizer.cpp
+++ b/src/video_core/renderer_opengl/gl_rasterizer.cpp
@@ -14,6 +14,7 @@
 #include "common/logging/log.h"
 #include "common/math_util.h"
 #include "common/microprofile.h"
+#include "common/scope_exit.h"
 #include "core/core.h"
 #include "core/frontend/emu_window.h"
 #include "core/hle/kernel/process.h"
@@ -315,16 +316,14 @@ std::pair<Surface, Surface> RasterizerOpenGL::ConfigureFramebuffers(bool using_c
         using_color_fb = false;
     }
 
-    // TODO(bunnei): Implement this
-    const bool has_stencil = false;
-
+    const bool has_stencil = regs.stencil_enable;
     const bool write_color_fb =
         state.color_mask.red_enabled == GL_TRUE || state.color_mask.green_enabled == GL_TRUE ||
         state.color_mask.blue_enabled == GL_TRUE || state.color_mask.alpha_enabled == GL_TRUE;
 
     const bool write_depth_fb =
         (state.depth.test_enabled && state.depth.write_mask == GL_TRUE) ||
-        (has_stencil && state.stencil.test_enabled && state.stencil.write_mask != 0);
+        (has_stencil && (state.stencil.front.write_mask || state.stencil.back.write_mask));
 
     Surface color_surface;
     Surface depth_surface;
@@ -364,41 +363,70 @@ std::pair<Surface, Surface> RasterizerOpenGL::ConfigureFramebuffers(bool using_c
 }
 
 void RasterizerOpenGL::Clear() {
-    const auto& regs = Core::System::GetInstance().GPU().Maxwell3D().regs;
+    const auto prev_state{state};
+    SCOPE_EXIT({ prev_state.Apply(); });
 
+    const auto& regs = Core::System::GetInstance().GPU().Maxwell3D().regs;
     bool use_color_fb = false;
     bool use_depth_fb = false;
 
-    GLbitfield clear_mask = 0;
-    if (regs.clear_buffers.R && regs.clear_buffers.G && regs.clear_buffers.B &&
+    OpenGLState clear_state;
+    clear_state.draw.draw_framebuffer = state.draw.draw_framebuffer;
+    clear_state.color_mask.red_enabled = regs.clear_buffers.R ? GL_TRUE : GL_FALSE;
+    clear_state.color_mask.green_enabled = regs.clear_buffers.G ? GL_TRUE : GL_FALSE;
+    clear_state.color_mask.blue_enabled = regs.clear_buffers.B ? GL_TRUE : GL_FALSE;
+    clear_state.color_mask.alpha_enabled = regs.clear_buffers.A ? GL_TRUE : GL_FALSE;
+
+    GLbitfield clear_mask{};
+    if (regs.clear_buffers.R || regs.clear_buffers.G || regs.clear_buffers.B ||
         regs.clear_buffers.A) {
-        clear_mask |= GL_COLOR_BUFFER_BIT;
-        use_color_fb = true;
+        if (regs.clear_buffers.RT == 0) {
+            // We only support clearing the first color attachment for now
+            clear_mask |= GL_COLOR_BUFFER_BIT;
+            use_color_fb = true;
+        } else {
+            // TODO(subv): Add support for the other color attachments
+            LOG_CRITICAL(HW_GPU, "Clear unimplemented for RT {}", regs.clear_buffers.RT);
+        }
     }
     if (regs.clear_buffers.Z) {
+        ASSERT_MSG(regs.zeta_enable != 0, "Tried to clear Z but buffer is not enabled!");
+        use_depth_fb = true;
         clear_mask |= GL_DEPTH_BUFFER_BIT;
-        use_depth_fb = regs.zeta_enable != 0;
 
         // Always enable the depth write when clearing the depth buffer. The depth write mask is
         // ignored when clearing the buffer in the Switch, but OpenGL obeys it so we set it to true.
-        state.depth.test_enabled = true;
-        state.depth.write_mask = GL_TRUE;
-        state.depth.test_func = GL_ALWAYS;
-        state.Apply();
+        clear_state.depth.test_enabled = true;
+        clear_state.depth.test_func = GL_ALWAYS;
+    }
+    if (regs.clear_buffers.S) {
+        ASSERT_MSG(regs.zeta_enable != 0, "Tried to clear stencil but buffer is not enabled!");
+        use_depth_fb = true;
+        clear_mask |= GL_STENCIL_BUFFER_BIT;
+        clear_state.stencil.test_enabled = true;
     }
 
-    if (clear_mask == 0)
+    if (!use_color_fb && !use_depth_fb) {
+        // No color surface nor depth/stencil surface are enabled
         return;
+    }
+
+    if (clear_mask == 0) {
+        // No clear mask is enabled
+        return;
+    }
 
     ScopeAcquireGLContext acquire_context{emu_window};
 
     auto [dirty_color_surface, dirty_depth_surface] =
         ConfigureFramebuffers(use_color_fb, use_depth_fb, false);
 
-    // TODO(Subv): Support clearing only partial colors.
+    clear_state.Apply();
+
     glClearColor(regs.clear_color[0], regs.clear_color[1], regs.clear_color[2],
                  regs.clear_color[3]);
     glClearDepth(regs.clear_depth);
+    glClearStencil(regs.clear_stencil);
 
     glClear(clear_mask);
 
@@ -451,6 +479,7 @@ void RasterizerOpenGL::DrawArrays() {
         ConfigureFramebuffers(true, regs.zeta.Address() != 0 && regs.zeta_enable != 0, true);
 
     SyncDepthTestState();
+    SyncStencilTestState();
     SyncBlendState();
     SyncLogicOpState();
     SyncCullMode();
@@ -841,6 +870,34 @@ void RasterizerOpenGL::SyncDepthTestState() {
     state.depth.test_func = MaxwellToGL::ComparisonOp(regs.depth_test_func);
 }
 
+void RasterizerOpenGL::SyncStencilTestState() {
+    const auto& regs = Core::System::GetInstance().GPU().Maxwell3D().regs;
+    state.stencil.test_enabled = regs.stencil_enable != 0;
+
+    if (!regs.stencil_enable) {
+        return;
+    }
+
+    // TODO(bunnei): Verify behavior when this is not set
+    ASSERT(regs.stencil_two_side_enable);
+
+    state.stencil.front.test_func = MaxwellToGL::ComparisonOp(regs.stencil_front_func_func);
+    state.stencil.front.test_ref = regs.stencil_front_func_ref;
+    state.stencil.front.test_mask = regs.stencil_front_func_mask;
+    state.stencil.front.action_stencil_fail = MaxwellToGL::StencilOp(regs.stencil_front_op_fail);
+    state.stencil.front.action_depth_fail = MaxwellToGL::StencilOp(regs.stencil_front_op_zfail);
+    state.stencil.front.action_depth_pass = MaxwellToGL::StencilOp(regs.stencil_front_op_zpass);
+    state.stencil.front.write_mask = regs.stencil_front_mask;
+
+    state.stencil.back.test_func = MaxwellToGL::ComparisonOp(regs.stencil_back_func_func);
+    state.stencil.back.test_ref = regs.stencil_back_func_ref;
+    state.stencil.back.test_mask = regs.stencil_back_func_mask;
+    state.stencil.back.action_stencil_fail = MaxwellToGL::StencilOp(regs.stencil_back_op_fail);
+    state.stencil.back.action_depth_fail = MaxwellToGL::StencilOp(regs.stencil_back_op_zfail);
+    state.stencil.back.action_depth_pass = MaxwellToGL::StencilOp(regs.stencil_back_op_zpass);
+    state.stencil.back.write_mask = regs.stencil_back_mask;
+}
+
 void RasterizerOpenGL::SyncBlendState() {
     const auto& regs = Core::System::GetInstance().GPU().Maxwell3D().regs;
 
diff --git a/src/video_core/renderer_opengl/gl_rasterizer.h b/src/video_core/renderer_opengl/gl_rasterizer.h
index 59b727de0..531b04046 100644
--- a/src/video_core/renderer_opengl/gl_rasterizer.h
+++ b/src/video_core/renderer_opengl/gl_rasterizer.h
@@ -141,6 +141,9 @@ private:
     /// Syncs the depth test state to match the guest state
     void SyncDepthTestState();
 
+    /// Syncs the stencil test state to match the guest state
+    void SyncStencilTestState();
+
     /// Syncs the blend state to match the guest state
     void SyncBlendState();
 
diff --git a/src/video_core/renderer_opengl/gl_rasterizer_cache.cpp b/src/video_core/renderer_opengl/gl_rasterizer_cache.cpp
index b1769c99b..83d8d3d94 100644
--- a/src/video_core/renderer_opengl/gl_rasterizer_cache.cpp
+++ b/src/video_core/renderer_opengl/gl_rasterizer_cache.cpp
@@ -780,17 +780,30 @@ Surface RasterizerCacheOpenGL::GetSurface(const SurfaceParams& params, bool pres
         } else if (preserve_contents) {
             // If surface parameters changed and we care about keeping the previous data, recreate
             // the surface from the old one
-            return RecreateSurface(surface, params);
+            UnregisterSurface(surface);
+            Surface new_surface{RecreateSurface(surface, params)};
+            RegisterSurface(new_surface);
+            return new_surface;
         } else {
             // Delete the old surface before creating a new one to prevent collisions.
             UnregisterSurface(surface);
         }
     }
 
+    // Try to get a previously reserved surface
+    surface = TryGetReservedSurface(params);
+
     // No surface found - create a new one
-    surface = std::make_shared<CachedSurface>(params);
-    RegisterSurface(surface);
-    LoadSurface(surface);
+    if (!surface) {
+        surface = std::make_shared<CachedSurface>(params);
+        ReserveSurface(surface);
+        RegisterSurface(surface);
+    }
+
+    // Only load surface from memory if we care about the contents
+    if (preserve_contents) {
+        LoadSurface(surface);
+    }
 
     return surface;
 }
@@ -799,13 +812,18 @@ Surface RasterizerCacheOpenGL::RecreateSurface(const Surface& surface,
                                                const SurfaceParams& new_params) {
     // Verify surface is compatible for blitting
     const auto& params{surface->GetSurfaceParams()};
-    ASSERT(params.type == new_params.type);
-    ASSERT_MSG(params.GetCompressionFactor(params.pixel_format) == 1,
-               "Compressed texture reinterpretation is not supported");
 
     // Create a new surface with the new parameters, and blit the previous surface to it
     Surface new_surface{std::make_shared<CachedSurface>(new_params)};
 
+    // If format is unchanged, we can do a faster blit without reinterpreting pixel data
+    if (params.pixel_format == new_params.pixel_format) {
+        BlitTextures(surface->Texture().handle, params.GetRect(), new_surface->Texture().handle,
+                     new_surface->GetSurfaceParams().GetRect(), params.type,
+                     read_framebuffer.handle, draw_framebuffer.handle);
+        return new_surface;
+    }
+
     auto source_format = GetFormatTuple(params.pixel_format, params.component_type);
     auto dest_format = GetFormatTuple(new_params.pixel_format, new_params.component_type);
 
@@ -818,9 +836,13 @@ Surface RasterizerCacheOpenGL::RecreateSurface(const Surface& surface,
 
     glBindBuffer(GL_PIXEL_PACK_BUFFER, pbo.handle);
     glBufferData(GL_PIXEL_PACK_BUFFER, buffer_size, nullptr, GL_STREAM_DRAW_ARB);
-    glGetTextureImage(surface->Texture().handle, 0, source_format.format, source_format.type,
-                      params.SizeInBytes(), nullptr);
-
+    if (source_format.compressed) {
+        glGetCompressedTextureImage(surface->Texture().handle, 0,
+                                    static_cast<GLsizei>(params.SizeInBytes()), nullptr);
+    } else {
+        glGetTextureImage(surface->Texture().handle, 0, source_format.format, source_format.type,
+                          static_cast<GLsizei>(params.SizeInBytes()), nullptr);
+    }
     // If the new texture is bigger than the previous one, we need to fill in the rest with data
     // from the CPU.
     if (params.SizeInBytes() < new_params.SizeInBytes()) {
@@ -846,17 +868,21 @@ Surface RasterizerCacheOpenGL::RecreateSurface(const Surface& surface,
     const auto& dest_rect{new_params.GetRect()};
 
     glBindBuffer(GL_PIXEL_UNPACK_BUFFER, pbo.handle);
-    glTextureSubImage2D(
-        new_surface->Texture().handle, 0, 0, 0, static_cast<GLsizei>(dest_rect.GetWidth()),
-        static_cast<GLsizei>(dest_rect.GetHeight()), dest_format.format, dest_format.type, nullptr);
+    if (dest_format.compressed) {
+        glCompressedTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0,
+                                  static_cast<GLsizei>(dest_rect.GetWidth()),
+                                  static_cast<GLsizei>(dest_rect.GetHeight()), dest_format.format,
+                                  static_cast<GLsizei>(new_params.SizeInBytes()), nullptr);
+    } else {
+        glTextureSubImage2D(new_surface->Texture().handle, 0, 0, 0,
+                            static_cast<GLsizei>(dest_rect.GetWidth()),
+                            static_cast<GLsizei>(dest_rect.GetHeight()), dest_format.format,
+                            dest_format.type, nullptr);
+    }
     glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0);
 
     pbo.Release();
 
-    // Update cache accordingly
-    UnregisterSurface(surface);
-    RegisterSurface(new_surface);
-
     return new_surface;
 }
 
@@ -931,6 +957,21 @@ void RasterizerCacheOpenGL::UnregisterSurface(const Surface& surface) {
     surface_cache.erase(search);
 }
 
+void RasterizerCacheOpenGL::ReserveSurface(const Surface& surface) {
+    const auto& surface_reserve_key{SurfaceReserveKey::Create(surface->GetSurfaceParams())};
+    surface_reserve[surface_reserve_key] = surface;
+}
+
+Surface RasterizerCacheOpenGL::TryGetReservedSurface(const SurfaceParams& params) {
+    const auto& surface_reserve_key{SurfaceReserveKey::Create(params)};
+    auto search{surface_reserve.find(surface_reserve_key)};
+    if (search != surface_reserve.end()) {
+        RegisterSurface(search->second);
+        return search->second;
+    }
+    return {};
+}
+
 template <typename Map, typename Interval>
 constexpr auto RangeFromInterval(Map& map, const Interval& interval) {
     return boost::make_iterator_range(map.equal_range(interval));
diff --git a/src/video_core/renderer_opengl/gl_rasterizer_cache.h b/src/video_core/renderer_opengl/gl_rasterizer_cache.h
index f273152a2..c8c615df2 100644
--- a/src/video_core/renderer_opengl/gl_rasterizer_cache.h
+++ b/src/video_core/renderer_opengl/gl_rasterizer_cache.h
@@ -11,6 +11,7 @@
 #include <boost/icl/interval_map.hpp>
 
 #include "common/common_types.h"
+#include "common/hash.h"
 #include "common/math_util.h"
 #include "video_core/engines/maxwell_3d.h"
 #include "video_core/renderer_opengl/gl_resource_manager.h"
@@ -682,6 +683,27 @@ struct SurfaceParams {
     u32 cache_height;
 };
 
+}; // namespace OpenGL
+
+/// Hashable variation of SurfaceParams, used for a key in the surface cache
+struct SurfaceReserveKey : Common::HashableStruct<OpenGL::SurfaceParams> {
+    static SurfaceReserveKey Create(const OpenGL::SurfaceParams& params) {
+        SurfaceReserveKey res;
+        res.state = params;
+        return res;
+    }
+};
+namespace std {
+template <>
+struct hash<SurfaceReserveKey> {
+    size_t operator()(const SurfaceReserveKey& k) const {
+        return k.Hash();
+    }
+};
+} // namespace std
+
+namespace OpenGL {
+
 class CachedSurface final {
 public:
     CachedSurface(const SurfaceParams& params);
@@ -752,12 +774,23 @@ private:
     /// Remove surface from the cache
     void UnregisterSurface(const Surface& surface);
 
+    /// Reserves a unique surface that can be reused later
+    void ReserveSurface(const Surface& surface);
+
+    /// Tries to get a reserved surface for the specified parameters
+    Surface TryGetReservedSurface(const SurfaceParams& params);
+
     /// Increase/decrease the number of surface in pages touching the specified region
     void UpdatePagesCachedCount(Tegra::GPUVAddr addr, u64 size, int delta);
 
     std::unordered_map<Tegra::GPUVAddr, Surface> surface_cache;
     PageMap cached_pages;
 
+    /// The surface reserve is a "backup" cache, this is where we put unique surfaces that have
+    /// previously been used. This is to prevent surfaces from being constantly created and
+    /// destroyed when used with different surface parameters.
+    std::unordered_map<SurfaceReserveKey, Surface> surface_reserve;
+
     OGLFramebuffer read_framebuffer;
     OGLFramebuffer draw_framebuffer;
 };
diff --git a/src/video_core/renderer_opengl/gl_state.cpp b/src/video_core/renderer_opengl/gl_state.cpp
index e1a887d67..60a4defd1 100644
--- a/src/video_core/renderer_opengl/gl_state.cpp
+++ b/src/video_core/renderer_opengl/gl_state.cpp
@@ -27,13 +27,17 @@ OpenGLState::OpenGLState() {
     color_mask.alpha_enabled = GL_TRUE;
 
     stencil.test_enabled = false;
-    stencil.test_func = GL_ALWAYS;
-    stencil.test_ref = 0;
-    stencil.test_mask = 0xFF;
-    stencil.write_mask = 0xFF;
-    stencil.action_depth_fail = GL_KEEP;
-    stencil.action_depth_pass = GL_KEEP;
-    stencil.action_stencil_fail = GL_KEEP;
+    auto reset_stencil = [](auto& config) {
+        config.test_func = GL_ALWAYS;
+        config.test_ref = 0;
+        config.test_mask = 0xFFFFFFFF;
+        config.write_mask = 0xFFFFFFFF;
+        config.action_depth_fail = GL_KEEP;
+        config.action_depth_pass = GL_KEEP;
+        config.action_stencil_fail = GL_KEEP;
+    };
+    reset_stencil(stencil.front);
+    reset_stencil(stencil.back);
 
     blend.enabled = true;
     blend.rgb_equation = GL_FUNC_ADD;
@@ -129,24 +133,23 @@ void OpenGLState::Apply() const {
             glDisable(GL_STENCIL_TEST);
         }
     }
-
-    if (stencil.test_func != cur_state.stencil.test_func ||
-        stencil.test_ref != cur_state.stencil.test_ref ||
-        stencil.test_mask != cur_state.stencil.test_mask) {
-        glStencilFunc(stencil.test_func, stencil.test_ref, stencil.test_mask);
-    }
-
-    if (stencil.action_depth_fail != cur_state.stencil.action_depth_fail ||
-        stencil.action_depth_pass != cur_state.stencil.action_depth_pass ||
-        stencil.action_stencil_fail != cur_state.stencil.action_stencil_fail) {
-        glStencilOp(stencil.action_stencil_fail, stencil.action_depth_fail,
-                    stencil.action_depth_pass);
-    }
-
-    // Stencil mask
-    if (stencil.write_mask != cur_state.stencil.write_mask) {
-        glStencilMask(stencil.write_mask);
-    }
+    auto config_stencil = [](GLenum face, const auto& config, const auto& prev_config) {
+        if (config.test_func != prev_config.test_func || config.test_ref != prev_config.test_ref ||
+            config.test_mask != prev_config.test_mask) {
+            glStencilFuncSeparate(face, config.test_func, config.test_ref, config.test_mask);
+        }
+        if (config.action_depth_fail != prev_config.action_depth_fail ||
+            config.action_depth_pass != prev_config.action_depth_pass ||
+            config.action_stencil_fail != prev_config.action_stencil_fail) {
+            glStencilOpSeparate(face, config.action_stencil_fail, config.action_depth_fail,
+                                config.action_depth_pass);
+        }
+        if (config.write_mask != prev_config.write_mask) {
+            glStencilMaskSeparate(face, config.write_mask);
+        }
+    };
+    config_stencil(GL_FRONT, stencil.front, cur_state.stencil.front);
+    config_stencil(GL_BACK, stencil.back, cur_state.stencil.back);
 
     // Blending
     if (blend.enabled != cur_state.blend.enabled) {
diff --git a/src/video_core/renderer_opengl/gl_state.h b/src/video_core/renderer_opengl/gl_state.h
index 22b0b1e41..46e96a97d 100644
--- a/src/video_core/renderer_opengl/gl_state.h
+++ b/src/video_core/renderer_opengl/gl_state.h
@@ -58,14 +58,16 @@ public:
     } color_mask; // GL_COLOR_WRITEMASK
 
     struct {
-        bool test_enabled;          // GL_STENCIL_TEST
-        GLenum test_func;           // GL_STENCIL_FUNC
-        GLint test_ref;             // GL_STENCIL_REF
-        GLuint test_mask;           // GL_STENCIL_VALUE_MASK
-        GLuint write_mask;          // GL_STENCIL_WRITEMASK
-        GLenum action_stencil_fail; // GL_STENCIL_FAIL
-        GLenum action_depth_fail;   // GL_STENCIL_PASS_DEPTH_FAIL
-        GLenum action_depth_pass;   // GL_STENCIL_PASS_DEPTH_PASS
+        bool test_enabled; // GL_STENCIL_TEST
+        struct {
+            GLenum test_func;           // GL_STENCIL_FUNC
+            GLint test_ref;             // GL_STENCIL_REF
+            GLuint test_mask;           // GL_STENCIL_VALUE_MASK
+            GLuint write_mask;          // GL_STENCIL_WRITEMASK
+            GLenum action_stencil_fail; // GL_STENCIL_FAIL
+            GLenum action_depth_fail;   // GL_STENCIL_PASS_DEPTH_FAIL
+            GLenum action_depth_pass;   // GL_STENCIL_PASS_DEPTH_PASS
+        } front, back;
     } stencil;
 
     struct {
diff --git a/src/video_core/renderer_opengl/maxwell_to_gl.h b/src/video_core/renderer_opengl/maxwell_to_gl.h
index 0343759a6..67273e164 100644
--- a/src/video_core/renderer_opengl/maxwell_to_gl.h
+++ b/src/video_core/renderer_opengl/maxwell_to_gl.h
@@ -295,6 +295,30 @@ inline GLenum ComparisonOp(Maxwell::ComparisonOp comparison) {
     return {};
 }
 
+inline GLenum StencilOp(Maxwell::StencilOp stencil) {
+    switch (stencil) {
+    case Maxwell::StencilOp::Keep:
+        return GL_KEEP;
+    case Maxwell::StencilOp::Zero:
+        return GL_ZERO;
+    case Maxwell::StencilOp::Replace:
+        return GL_REPLACE;
+    case Maxwell::StencilOp::Incr:
+        return GL_INCR;
+    case Maxwell::StencilOp::Decr:
+        return GL_DECR;
+    case Maxwell::StencilOp::Invert:
+        return GL_INVERT;
+    case Maxwell::StencilOp::IncrWrap:
+        return GL_INCR_WRAP;
+    case Maxwell::StencilOp::DecrWrap:
+        return GL_DECR_WRAP;
+    }
+    LOG_CRITICAL(Render_OpenGL, "Unimplemented stencil op={}", static_cast<u32>(stencil));
+    UNREACHABLE();
+    return {};
+}
+
 inline GLenum FrontFace(Maxwell::Cull::FrontFace front_face) {
     switch (front_face) {
     case Maxwell::Cull::FrontFace::ClockWise: