Merge pull request #10744 from Wollnashorn/af-for-all

video_core: Improved anisotropic filtering heuristics

4 files changed, 91 insertions, 42 deletions

diff --git a/src/video_core/renderer_opengl/gl_compute_pipeline.cpp b/src/video_core/renderer_opengl/gl_compute_pipeline.cpp
index 1a0cea9b7..3151c0db8 100644
--- a/src/video_core/renderer_opengl/gl_compute_pipeline.cpp
+++ b/src/video_core/renderer_opengl/gl_compute_pipeline.cpp
@@ -87,7 +87,8 @@ void ComputePipeline::Configure() {
     texture_cache.SynchronizeComputeDescriptors();
 
     boost::container::static_vector<VideoCommon::ImageViewInOut, MAX_TEXTURES + MAX_IMAGES> views;
-    std::array<GLuint, MAX_TEXTURES> samplers;
+    boost::container::static_vector<VideoCommon::SamplerId, MAX_TEXTURES> samplers;
+    std::array<GLuint, MAX_TEXTURES> gl_samplers;
     std::array<GLuint, MAX_TEXTURES> textures;
     std::array<GLuint, MAX_IMAGES> images;
     GLsizei sampler_binding{};
@@ -131,7 +132,6 @@ void ComputePipeline::Configure() {
         for (u32 index = 0; index < desc.count; ++index) {
             const auto handle{read_handle(desc, index)};
             views.push_back({handle.first});
-            samplers[sampler_binding++] = 0;
         }
     }
     for (const auto& desc : info.image_buffer_descriptors) {
@@ -142,8 +142,8 @@ void ComputePipeline::Configure() {
             const auto handle{read_handle(desc, index)};
             views.push_back({handle.first});
 
-            Sampler* const sampler = texture_cache.GetComputeSampler(handle.second);
-            samplers[sampler_binding++] = sampler->Handle();
+            VideoCommon::SamplerId sampler = texture_cache.GetComputeSamplerId(handle.second);
+            samplers.push_back(sampler);
         }
     }
     for (const auto& desc : info.image_descriptors) {
@@ -186,10 +186,17 @@ void ComputePipeline::Configure() {
 
     const VideoCommon::ImageViewInOut* views_it{views.data() + num_texture_buffers +
                                                 num_image_buffers};
+    const VideoCommon::SamplerId* samplers_it{samplers.data()};
     texture_binding += num_texture_buffers;
     image_binding += num_image_buffers;
 
     u32 texture_scaling_mask{};
+
+    for (const auto& desc : info.texture_buffer_descriptors) {
+        for (u32 index = 0; index < desc.count; ++index) {
+            gl_samplers[sampler_binding++] = 0;
+        }
+    }
     for (const auto& desc : info.texture_descriptors) {
         for (u32 index = 0; index < desc.count; ++index) {
             ImageView& image_view{texture_cache.GetImageView((views_it++)->id)};
@@ -198,6 +205,12 @@ void ComputePipeline::Configure() {
                 texture_scaling_mask |= 1u << texture_binding;
             }
             ++texture_binding;
+
+            const Sampler& sampler{texture_cache.GetSampler(*(samplers_it++))};
+            const bool use_fallback_sampler{sampler.HasAddedAnisotropy() &&
+                                            !image_view.SupportsAnisotropy()};
+            gl_samplers[sampler_binding++] =
+                use_fallback_sampler ? sampler.HandleWithDefaultAnisotropy() : sampler.Handle();
         }
     }
     u32 image_scaling_mask{};
@@ -228,7 +241,7 @@ void ComputePipeline::Configure() {
     if (texture_binding != 0) {
         ASSERT(texture_binding == sampler_binding);
         glBindTextures(0, texture_binding, textures.data());
-        glBindSamplers(0, sampler_binding, samplers.data());
+        glBindSamplers(0, sampler_binding, gl_samplers.data());
     }
     if (image_binding != 0) {
         glBindImageTextures(0, image_binding, images.data());
diff --git a/src/video_core/renderer_opengl/gl_graphics_pipeline.cpp b/src/video_core/renderer_opengl/gl_graphics_pipeline.cpp
index 89000d6e0..c58f760b8 100644
--- a/src/video_core/renderer_opengl/gl_graphics_pipeline.cpp
+++ b/src/video_core/renderer_opengl/gl_graphics_pipeline.cpp
@@ -275,9 +275,9 @@ GraphicsPipeline::GraphicsPipeline(const Device& device, TextureCache& texture_c
 template <typename Spec>
 void GraphicsPipeline::ConfigureImpl(bool is_indexed) {
     std::array<VideoCommon::ImageViewInOut, MAX_TEXTURES + MAX_IMAGES> views;
-    std::array<GLuint, MAX_TEXTURES> samplers;
+    std::array<VideoCommon::SamplerId, MAX_TEXTURES> samplers;
     size_t views_index{};
-    GLsizei sampler_binding{};
+    size_t samplers_index{};
 
     texture_cache.SynchronizeGraphicsDescriptors();
 
@@ -337,7 +337,6 @@ void GraphicsPipeline::ConfigureImpl(bool is_indexed) {
                 for (u32 index = 0; index < desc.count; ++index) {
                     const auto handle{read_handle(desc, index)};
                     views[views_index++] = {handle.first};
-                    samplers[sampler_binding++] = 0;
                 }
             }
         }
@@ -351,8 +350,8 @@ void GraphicsPipeline::ConfigureImpl(bool is_indexed) {
                 const auto handle{read_handle(desc, index)};
                 views[views_index++] = {handle.first};
 
-                Sampler* const sampler{texture_cache.GetGraphicsSampler(handle.second)};
-                samplers[sampler_binding++] = sampler->Handle();
+                VideoCommon::SamplerId sampler{texture_cache.GetGraphicsSamplerId(handle.second)};
+                samplers[samplers_index++] = sampler;
             }
         }
         if constexpr (Spec::has_images) {
@@ -445,10 +444,13 @@ void GraphicsPipeline::ConfigureImpl(bool is_indexed) {
         program_manager.BindSourcePrograms(source_programs);
     }
     const VideoCommon::ImageViewInOut* views_it{views.data()};
+    const VideoCommon::SamplerId* samplers_it{samplers.data()};
     GLsizei texture_binding = 0;
     GLsizei image_binding = 0;
+    GLsizei sampler_binding{};
     std::array<GLuint, MAX_TEXTURES> textures;
     std::array<GLuint, MAX_IMAGES> images;
+    std::array<GLuint, MAX_TEXTURES> gl_samplers;
     const auto prepare_stage{[&](size_t stage) {
         buffer_cache.runtime.SetImagePointers(&textures[texture_binding], &images[image_binding]);
         buffer_cache.BindHostStageBuffers(stage);
@@ -465,6 +467,13 @@ void GraphicsPipeline::ConfigureImpl(bool is_indexed) {
         u32 stage_image_binding{};
 
         const auto& info{stage_infos[stage]};
+        if constexpr (Spec::has_texture_buffers) {
+            for (const auto& desc : info.texture_buffer_descriptors) {
+                for (u32 index = 0; index < desc.count; ++index) {
+                    gl_samplers[sampler_binding++] = 0;
+                }
+            }
+        }
         for (const auto& desc : info.texture_descriptors) {
             for (u32 index = 0; index < desc.count; ++index) {
                 ImageView& image_view{texture_cache.GetImageView((views_it++)->id)};
@@ -474,6 +483,12 @@ void GraphicsPipeline::ConfigureImpl(bool is_indexed) {
                 }
                 ++texture_binding;
                 ++stage_texture_binding;
+
+                const Sampler& sampler{texture_cache.GetSampler(*(samplers_it++))};
+                const bool use_fallback_sampler{sampler.HasAddedAnisotropy() &&
+                                                !image_view.SupportsAnisotropy()};
+                gl_samplers[sampler_binding++] =
+                    use_fallback_sampler ? sampler.HandleWithDefaultAnisotropy() : sampler.Handle();
             }
         }
         for (const auto& desc : info.image_descriptors) {
@@ -534,7 +549,7 @@ void GraphicsPipeline::ConfigureImpl(bool is_indexed) {
     if (texture_binding != 0) {
         ASSERT(texture_binding == sampler_binding);
         glBindTextures(0, texture_binding, textures.data());
-        glBindSamplers(0, sampler_binding, samplers.data());
+        glBindSamplers(0, sampler_binding, gl_samplers.data());
     }
     if (image_binding != 0) {
         glBindImageTextures(0, image_binding, images.data());
diff --git a/src/video_core/renderer_opengl/gl_texture_cache.cpp b/src/video_core/renderer_opengl/gl_texture_cache.cpp
index 1c5dbcdd8..3b446be07 100644
--- a/src/video_core/renderer_opengl/gl_texture_cache.cpp
+++ b/src/video_core/renderer_opengl/gl_texture_cache.cpp
@@ -1268,36 +1268,48 @@ Sampler::Sampler(TextureCacheRuntime& runtime, const TSCEntry& config) {
 
     UNIMPLEMENTED_IF(config.cubemap_anisotropy != 1);
 
-    sampler.Create();
-    const GLuint handle = sampler.handle;
-    glSamplerParameteri(handle, GL_TEXTURE_WRAP_S, MaxwellToGL::WrapMode(config.wrap_u));
-    glSamplerParameteri(handle, GL_TEXTURE_WRAP_T, MaxwellToGL::WrapMode(config.wrap_v));
-    glSamplerParameteri(handle, GL_TEXTURE_WRAP_R, MaxwellToGL::WrapMode(config.wrap_p));
-    glSamplerParameteri(handle, GL_TEXTURE_COMPARE_MODE, compare_mode);
-    glSamplerParameteri(handle, GL_TEXTURE_COMPARE_FUNC, compare_func);
-    glSamplerParameteri(handle, GL_TEXTURE_MAG_FILTER, mag);
-    glSamplerParameteri(handle, GL_TEXTURE_MIN_FILTER, min);
-    glSamplerParameterf(handle, GL_TEXTURE_LOD_BIAS, config.LodBias());
-    glSamplerParameterf(handle, GL_TEXTURE_MIN_LOD, config.MinLod());
-    glSamplerParameterf(handle, GL_TEXTURE_MAX_LOD, config.MaxLod());
-    glSamplerParameterfv(handle, GL_TEXTURE_BORDER_COLOR, config.BorderColor().data());
-
-    if (GLAD_GL_ARB_texture_filter_anisotropic || GLAD_GL_EXT_texture_filter_anisotropic) {
-        const f32 max_anisotropy = std::clamp(config.MaxAnisotropy(), 1.0f, 16.0f);
-        glSamplerParameterf(handle, GL_TEXTURE_MAX_ANISOTROPY, max_anisotropy);
-    } else {
-        LOG_WARNING(Render_OpenGL, "GL_ARB_texture_filter_anisotropic is required");
-    }
-    if (GLAD_GL_ARB_texture_filter_minmax || GLAD_GL_EXT_texture_filter_minmax) {
-        glSamplerParameteri(handle, GL_TEXTURE_REDUCTION_MODE_ARB, reduction_filter);
-    } else if (reduction_filter != GL_WEIGHTED_AVERAGE_ARB) {
-        LOG_WARNING(Render_OpenGL, "GL_ARB_texture_filter_minmax is required");
-    }
-    if (GLAD_GL_ARB_seamless_cubemap_per_texture || GLAD_GL_AMD_seamless_cubemap_per_texture) {
-        glSamplerParameteri(handle, GL_TEXTURE_CUBE_MAP_SEAMLESS, seamless);
-    } else if (seamless == GL_FALSE) {
-        // We default to false because it's more common
-        LOG_WARNING(Render_OpenGL, "GL_ARB_seamless_cubemap_per_texture is required");
+    const f32 max_anisotropy = std::clamp(config.MaxAnisotropy(), 1.0f, 16.0f);
+
+    const auto create_sampler = [&](const f32 anisotropy) {
+        OGLSampler new_sampler;
+        new_sampler.Create();
+        const GLuint handle = new_sampler.handle;
+        glSamplerParameteri(handle, GL_TEXTURE_WRAP_S, MaxwellToGL::WrapMode(config.wrap_u));
+        glSamplerParameteri(handle, GL_TEXTURE_WRAP_T, MaxwellToGL::WrapMode(config.wrap_v));
+        glSamplerParameteri(handle, GL_TEXTURE_WRAP_R, MaxwellToGL::WrapMode(config.wrap_p));
+        glSamplerParameteri(handle, GL_TEXTURE_COMPARE_MODE, compare_mode);
+        glSamplerParameteri(handle, GL_TEXTURE_COMPARE_FUNC, compare_func);
+        glSamplerParameteri(handle, GL_TEXTURE_MAG_FILTER, mag);
+        glSamplerParameteri(handle, GL_TEXTURE_MIN_FILTER, min);
+        glSamplerParameterf(handle, GL_TEXTURE_LOD_BIAS, config.LodBias());
+        glSamplerParameterf(handle, GL_TEXTURE_MIN_LOD, config.MinLod());
+        glSamplerParameterf(handle, GL_TEXTURE_MAX_LOD, config.MaxLod());
+        glSamplerParameterfv(handle, GL_TEXTURE_BORDER_COLOR, config.BorderColor().data());
+
+        if (GLAD_GL_ARB_texture_filter_anisotropic || GLAD_GL_EXT_texture_filter_anisotropic) {
+            glSamplerParameterf(handle, GL_TEXTURE_MAX_ANISOTROPY, anisotropy);
+        } else {
+            LOG_WARNING(Render_OpenGL, "GL_ARB_texture_filter_anisotropic is required");
+        }
+        if (GLAD_GL_ARB_texture_filter_minmax || GLAD_GL_EXT_texture_filter_minmax) {
+            glSamplerParameteri(handle, GL_TEXTURE_REDUCTION_MODE_ARB, reduction_filter);
+        } else if (reduction_filter != GL_WEIGHTED_AVERAGE_ARB) {
+            LOG_WARNING(Render_OpenGL, "GL_ARB_texture_filter_minmax is required");
+        }
+        if (GLAD_GL_ARB_seamless_cubemap_per_texture || GLAD_GL_AMD_seamless_cubemap_per_texture) {
+            glSamplerParameteri(handle, GL_TEXTURE_CUBE_MAP_SEAMLESS, seamless);
+        } else if (seamless == GL_FALSE) {
+            // We default to false because it's more common
+            LOG_WARNING(Render_OpenGL, "GL_ARB_seamless_cubemap_per_texture is required");
+        }
+        return new_sampler;
+    };
+
+    sampler = create_sampler(max_anisotropy);
+
+    const f32 max_anisotropy_default = static_cast<f32>(1U << config.max_anisotropy);
+    if (max_anisotropy > max_anisotropy_default) {
+        sampler_default_anisotropy = create_sampler(max_anisotropy_default);
     }
 }
 
diff --git a/src/video_core/renderer_opengl/gl_texture_cache.h b/src/video_core/renderer_opengl/gl_texture_cache.h
index 1148b73d7..3676eaaa9 100644
--- a/src/video_core/renderer_opengl/gl_texture_cache.h
+++ b/src/video_core/renderer_opengl/gl_texture_cache.h
@@ -309,12 +309,21 @@ class Sampler {
 public:
     explicit Sampler(TextureCacheRuntime&, const Tegra::Texture::TSCEntry&);
 
-    GLuint Handle() const noexcept {
+    [[nodiscard]] GLuint Handle() const noexcept {
         return sampler.handle;
     }
 
+    [[nodiscard]] GLuint HandleWithDefaultAnisotropy() const noexcept {
+        return sampler_default_anisotropy.handle;
+    }
+
+    [[nodiscard]] bool HasAddedAnisotropy() const noexcept {
+        return static_cast<bool>(sampler_default_anisotropy.handle);
+    }
+
 private:
     OGLSampler sampler;
+    OGLSampler sampler_default_anisotropy;
 };
 
 class Framebuffer {