video_core: Add per-image anisotropy heuristics (format & mip count)

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<const Sampler*, 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;
                 }
             }
         }
@@ -352,7 +351,7 @@ void GraphicsPipeline::ConfigureImpl(bool is_indexed) {
                 views[views_index++] = {handle.first};
 
                 Sampler* const sampler{texture_cache.GetGraphicsSampler(handle.second)};
-                samplers[sampler_binding++] = sampler->Handle();
+                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 Sampler** 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{**(samplers_it++)};
+                const bool use_fallback_sampler{sampler.HasAddedAnisotropy() &&
+                                                !image_view.SupportsAnisotropy()};
+                gl_samplers[sampler_binding++] =
+                    use_fallback_sampler ? sampler.HandleWithoutAnisotropy() : 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());

src/video_core/renderer_opengl/gl_texture_cache.cpp

@@ -1268,6 +1268,10 @@ Sampler::Sampler(TextureCacheRuntime& runtime, const TSCEntry& config) {
 
     UNIMPLEMENTED_IF(config.cubemap_anisotropy != 1);
 
+    const f32 max_anisotropy = std::clamp(config.MaxAnisotropy(), 1.0f, 16.0f);
+
+    const auto create_sampler = [&](const f32 max_anisotropy) {
+        OGLSampler sampler;
         sampler.Create();
         const GLuint handle = sampler.handle;
         glSamplerParameteri(handle, GL_TEXTURE_WRAP_S, MaxwellToGL::WrapMode(config.wrap_u));
@@ -1283,7 +1287,6 @@ Sampler::Sampler(TextureCacheRuntime& runtime, const TSCEntry& config) {
         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");
@@ -1299,6 +1302,13 @@ Sampler::Sampler(TextureCacheRuntime& runtime, const TSCEntry& config) {
             // We default to false because it's more common
             LOG_WARNING(Render_OpenGL, "GL_ARB_seamless_cubemap_per_texture is required");
         }
+        return sampler;
+    };
+
+    sampler = create_sampler(max_anisotropy);
+    if (Settings::values.max_anisotropy.GetValue() > 0 && max_anisotropy > 1.0f) {
+        sampler_without_anisotropy = create_sampler(1.0f);
+    }
 }
 
 Framebuffer::Framebuffer(TextureCacheRuntime& runtime, std::span<ImageView*, NUM_RT> color_buffers,

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 HandleWithoutAnisotropy() const noexcept {
+        return sampler_without_anisotropy.handle;
+    }
+
+    [[nodiscard]] bool HasAddedAnisotropy() const noexcept {
+        return static_cast<bool>(sampler_without_anisotropy.handle);
+    }
+
 private:
     OGLSampler sampler;
+    OGLSampler sampler_without_anisotropy;
 };
 
 class Framebuffer {

src/video_core/renderer_vulkan/pipeline_helper.h

@@ -178,7 +178,7 @@ public:
 inline void PushImageDescriptors(TextureCache& texture_cache,
                                  GuestDescriptorQueue& guest_descriptor_queue,
                                  const Shader::Info& info, RescalingPushConstant& rescaling,
-                                 const VkSampler*& samplers,
+                                 const Sampler**& samplers,
                                  const VideoCommon::ImageViewInOut*& views) {
     const u32 num_texture_buffers = Shader::NumDescriptors(info.texture_buffer_descriptors);
     const u32 num_image_buffers = Shader::NumDescriptors(info.image_buffer_descriptors);
@@ -187,10 +187,14 @@ inline void PushImageDescriptors(TextureCache& texture_cache,
     for (const auto& desc : info.texture_descriptors) {
         for (u32 index = 0; index < desc.count; ++index) {
             const VideoCommon::ImageViewId image_view_id{(views++)->id};
-            const VkSampler sampler{*(samplers++)};
             ImageView& image_view{texture_cache.GetImageView(image_view_id)};
             const VkImageView vk_image_view{image_view.Handle(desc.type)};
-            guest_descriptor_queue.AddSampledImage(vk_image_view, sampler);
+            const Sampler& sampler{**(samplers++)};
+            const bool use_fallback_sampler{sampler.HasAddedAnisotropy() &&
+                                            !image_view.SupportsAnisotropy()};
+            const VkSampler vk_sampler{use_fallback_sampler ? sampler.HandleWithoutAnisotropy()
+                                                            : sampler.Handle()};
+            guest_descriptor_queue.AddSampledImage(vk_image_view, vk_sampler);
             rescaling.PushTexture(texture_cache.IsRescaling(image_view));
         }
     }

src/video_core/renderer_vulkan/vk_compute_pipeline.cpp

@@ -115,7 +115,7 @@ void ComputePipeline::Configure(Tegra::Engines::KeplerCompute& kepler_compute,
 
     static constexpr size_t max_elements = 64;
     boost::container::static_vector<VideoCommon::ImageViewInOut, max_elements> views;
-    boost::container::static_vector<VkSampler, max_elements> samplers;
+    boost::container::static_vector<const Sampler*, max_elements> samplers;
 
     const auto& qmd{kepler_compute.launch_description};
     const auto& cbufs{qmd.const_buffer_config};
@@ -161,7 +161,7 @@ void ComputePipeline::Configure(Tegra::Engines::KeplerCompute& kepler_compute,
             views.push_back({handle.first});
 
             Sampler* const sampler = texture_cache.GetComputeSampler(handle.second);
-            samplers.push_back(sampler->Handle());
+            samplers.push_back(sampler);
         }
     }
     for (const auto& desc : info.image_descriptors) {
@@ -192,7 +192,7 @@ void ComputePipeline::Configure(Tegra::Engines::KeplerCompute& kepler_compute,
     buffer_cache.BindHostComputeBuffers();
 
     RescalingPushConstant rescaling;
-    const VkSampler* samplers_it{samplers.data()};
+    const Sampler** samplers_it{samplers.data()};
     const VideoCommon::ImageViewInOut* views_it{views.data()};
     PushImageDescriptors(texture_cache, guest_descriptor_queue, info, rescaling, samplers_it,
                          views_it);

src/video_core/renderer_vulkan/vk_graphics_pipeline.cpp

@@ -298,7 +298,7 @@ void GraphicsPipeline::AddTransition(GraphicsPipeline* transition) {
 template <typename Spec>
 void GraphicsPipeline::ConfigureImpl(bool is_indexed) {
     std::array<VideoCommon::ImageViewInOut, MAX_IMAGE_ELEMENTS> views;
-    std::array<VkSampler, MAX_IMAGE_ELEMENTS> samplers;
+    std::array<const Sampler*, MAX_IMAGE_ELEMENTS> samplers;
     size_t sampler_index{};
     size_t view_index{};
 
@@ -368,7 +368,7 @@ void GraphicsPipeline::ConfigureImpl(bool is_indexed) {
                 views[view_index++] = {handle.first};
 
                 Sampler* const sampler{texture_cache.GetGraphicsSampler(handle.second)};
-                samplers[sampler_index++] = sampler->Handle();
+                samplers[sampler_index++] = sampler;
             }
         }
         if constexpr (Spec::has_images) {
@@ -453,7 +453,7 @@ void GraphicsPipeline::ConfigureImpl(bool is_indexed) {
 
     RescalingPushConstant rescaling;
     RenderAreaPushConstant render_area;
-    const VkSampler* samplers_it{samplers.data()};
+    const Sampler** samplers_it{samplers.data()};
     const VideoCommon::ImageViewInOut* views_it{views.data()};
     const auto prepare_stage{[&](size_t stage) LAMBDA_FORCEINLINE {
         buffer_cache.BindHostStageBuffers(stage);

src/video_core/renderer_vulkan/vk_texture_cache.cpp

@@ -1802,7 +1802,8 @@ Sampler::Sampler(TextureCacheRuntime& runtime, const Tegra::Texture::TSCEntry& t
     // Some games have samplers with garbage. Sanitize them here.
     const f32 max_anisotropy = std::clamp(tsc.MaxAnisotropy(), 1.0f, 16.0f);
 
-    sampler = device.GetLogical().CreateSampler(VkSamplerCreateInfo{
+    const auto create_sampler = [&](const f32 max_anisotropy) {
+        return device.GetLogical().CreateSampler(VkSamplerCreateInfo{
             .sType = VK_STRUCTURE_TYPE_SAMPLER_CREATE_INFO,
             .pNext = pnext,
             .flags = 0,
@@ -1823,6 +1824,12 @@ Sampler::Sampler(TextureCacheRuntime& runtime, const Tegra::Texture::TSCEntry& t
                 arbitrary_borders ? VK_BORDER_COLOR_FLOAT_CUSTOM_EXT : ConvertBorderColor(color),
             .unnormalizedCoordinates = VK_FALSE,
         });
+    };
+
+    sampler = create_sampler(max_anisotropy);
+    if (Settings::values.max_anisotropy.GetValue() > 0 && max_anisotropy > 1.0f) {
+        sampler_without_anisotropy = create_sampler(1.0f);
+    }
 }
 
 Framebuffer::Framebuffer(TextureCacheRuntime& runtime, std::span<ImageView*, NUM_RT> color_buffers,

src/video_core/renderer_vulkan/vk_texture_cache.h

@@ -279,8 +279,17 @@ public:
         return *sampler;
     }
 
+    [[nodiscard]] VkSampler HandleWithoutAnisotropy() const noexcept {
+        return *sampler_without_anisotropy;
+    }
+
+    [[nodiscard]] bool HasAddedAnisotropy() const noexcept {
+        return static_cast<bool>(sampler_without_anisotropy);
+    }
+
 private:
     vk::Sampler sampler;
+    vk::Sampler sampler_without_anisotropy;
 };
 
 class Framebuffer {

src/video_core/texture_cache/image_view_base.cpp

@@ -45,4 +45,43 @@ ImageViewBase::ImageViewBase(const ImageInfo& info, const ImageViewInfo& view_in
 
 ImageViewBase::ImageViewBase(const NullImageViewParams&) : image_id{NULL_IMAGE_ID} {}
 
+bool ImageViewBase::SupportsAnisotropy() const noexcept {
+    using namespace VideoCommon;
+    switch (format) {
+    case PixelFormat::R8_UNORM:
+    case PixelFormat::R8_SNORM:
+    case PixelFormat::R8_SINT:
+    case PixelFormat::R8_UINT:
+    case PixelFormat::BC4_UNORM:
+    case PixelFormat::BC4_SNORM:
+    case PixelFormat::BC5_UNORM:
+    case PixelFormat::BC5_SNORM:
+    case PixelFormat::R32G32_FLOAT:
+    case PixelFormat::R32G32_SINT:
+    case PixelFormat::R32_FLOAT:
+    case PixelFormat::R16_FLOAT:
+    case PixelFormat::R16_UNORM:
+    case PixelFormat::R16_SNORM:
+    case PixelFormat::R16_UINT:
+    case PixelFormat::R16_SINT:
+    case PixelFormat::R16G16_UNORM:
+    case PixelFormat::R16G16_FLOAT:
+    case PixelFormat::R16G16_UINT:
+    case PixelFormat::R16G16_SINT:
+    case PixelFormat::R16G16_SNORM:
+    case PixelFormat::R8G8_UNORM:
+    case PixelFormat::R8G8_SNORM:
+    case PixelFormat::R8G8_SINT:
+    case PixelFormat::R8G8_UINT:
+    case PixelFormat::R32G32_UINT:
+    case PixelFormat::R32_UINT:
+    case PixelFormat::R32_SINT:
+    case PixelFormat::G4R4_UNORM:
+        return false;
+    default:
+        break;
+    }
+    return range.extent.levels > 1;
+}
+
 } // namespace VideoCommon

src/video_core/texture_cache/image_view_base.h

@@ -33,6 +33,8 @@ struct ImageViewBase {
         return type == ImageViewType::Buffer;
     }
 
+    [[nodiscard]] bool SupportsAnisotropy() const noexcept;
+
     ImageId image_id{};
     GPUVAddr gpu_addr = 0;
     PixelFormat format{};

src/video_core/textures/texture.cpp

@@ -62,12 +62,14 @@ std::array<float, 4> TSCEntry::BorderColor() const noexcept {
 }
 
 float TSCEntry::MaxAnisotropy() const noexcept {
-    const bool is_unsupported_mipmap_filter = Settings::values.use_aggressive_anisotropic_filtering
+    const bool is_suitable_mipmap_filter = Settings::values.use_aggressive_anisotropic_filtering
-                                                  ? mipmap_filter == TextureMipmapFilter::None
+                                               ? mipmap_filter != TextureMipmapFilter::None
-                                                  : mipmap_filter != TextureMipmapFilter::Linear;
+                                               : mipmap_filter == TextureMipmapFilter::Linear;
     const bool has_regular_lods = min_lod_clamp == 0 && max_lod_clamp >= 256;
-    if (max_anisotropy == 0 &&
+    const bool is_bilinear_filter = min_filter == TextureFilter::Linear &&
-        (depth_compare_enabled.Value() || !has_regular_lods || is_unsupported_mipmap_filter)) {
+                                    reduction_filter == SamplerReduction::WeightedAverage;
+    if (max_anisotropy == 0 && (depth_compare_enabled.Value() || !has_regular_lods ||
+                                !is_bilinear_filter || !is_suitable_mipmap_filter)) {
         return 1.0f;
     }
     const auto anisotropic_settings = Settings::values.max_anisotropy.GetValue();