Texture Cache: Fix downscaling and correct memory comsumption.

src/video_core/renderer_opengl/gl_texture_cache.cpp

@@ -876,7 +876,7 @@ void Image::CopyImageToBuffer(const VideoCommon::BufferImageCopy& copy, size_t b
     }
 }
 
-bool Image::Scale() {
+bool Image::Scale(bool up_scale) {
     const auto format_type = GetFormatType(info.format);
     const GLenum attachment = [format_type] {
         switch (format_type) {
@@ -944,14 +944,25 @@ bool Image::Scale() {
     const GLuint draw_fbo = runtime->rescale_draw_fbos[fbo_index].handle;
     for (s32 layer = 0; layer < info.resources.layers; ++layer) {
         for (s32 level = 0; level < info.resources.levels; ++level) {
-            const u32 src_level_width = std::max(1u, original_width >> level);
-            const u32 src_level_height = std::max(1u, original_height >> level);
-            const u32 dst_level_width = std::max(1u, scaled_width >> level);
-            const u32 dst_level_height = std::max(1u, scaled_height >> level);
+            const u32 src_level_width =
+                std::max(1u, (up_scale ? original_width : scaled_width) >> level);
+            const u32 src_level_height =
+                std::max(1u, (up_scale ? original_height : scaled_height) >> level);
+            const u32 dst_level_width =
+                std::max(1u, (up_scale ? scaled_width : original_width) >> level);
+            const u32 dst_level_height =
+                std::max(1u, (up_scale ? scaled_height : original_height) >> level);
 
+            if (up_scale) {
                 glNamedFramebufferTextureLayer(read_fbo, attachment, texture.handle, level, layer);
                 glNamedFramebufferTextureLayer(draw_fbo, attachment, upscaled_backup.handle, level,
                                                layer);
+            } else {
+                glNamedFramebufferTextureLayer(read_fbo, attachment, upscaled_backup.handle, level,
+                                               layer);
+                glNamedFramebufferTextureLayer(draw_fbo, attachment, texture.handle, level, layer);
+            }
+
             glBlitNamedFramebuffer(read_fbo, draw_fbo, 0, 0, src_level_width, src_level_height, 0,
                                    0, dst_level_width, dst_level_height, mask, filter);
         }
@@ -959,7 +970,12 @@ bool Image::Scale() {
     if (scissor_test != GL_FALSE) {
         glEnablei(GL_SCISSOR_TEST, 0);
     }
+    if (up_scale) {
         current_texture = upscaled_backup.handle;
+    } else {
+        current_texture = texture.handle;
+    }
+
     return true;
 }
 
@@ -981,6 +997,7 @@ bool Image::ScaleUp() {
         flags &= ~ImageFlagBits::Rescaled;
         return false;
     }
+    scale_count++;
     if (!Scale()) {
         flags &= ~ImageFlagBits::Rescaled;
         return false;
@@ -996,7 +1013,11 @@ bool Image::ScaleDown() {
     if (!runtime->resolution.active) {
         return false;
     }
-    current_texture = texture.handle;
+    scale_count++;
+    if (!Scale(false)) {
+        flags &= ~ImageFlagBits::Rescaled;
+        return false;
+    }
     return true;
 }
 

src/video_core/renderer_opengl/gl_texture_cache.h

@@ -205,7 +205,7 @@ private:
 
     void CopyImageToBuffer(const VideoCommon::BufferImageCopy& copy, size_t buffer_offset);
 
-    bool Scale();
+    bool Scale(bool up_scale = true);
 
     OGLTexture texture;
     OGLTexture upscaled_backup;

src/video_core/renderer_vulkan/vk_texture_cache.cpp

@@ -592,7 +592,8 @@ struct RangedBarrierRange {
 }
 
 void BlitScale(VKScheduler& scheduler, VkImage src_image, VkImage dst_image, const ImageInfo& info,
-               VkImageAspectFlags aspect_mask, const Settings::ResolutionScalingInfo& resolution) {
+               VkImageAspectFlags aspect_mask, const Settings::ResolutionScalingInfo& resolution,
+               bool up_scaling = true) {
     const bool is_2d = info.type == ImageType::e2D;
     const auto resources = info.resources;
     const VkExtent2D extent{
@@ -605,14 +606,16 @@ void BlitScale(VKScheduler& scheduler, VkImage src_image, VkImage dst_image, con
 
     scheduler.RequestOutsideRenderPassOperationContext();
     scheduler.Record([dst_image, src_image, extent, resources, aspect_mask, resolution, is_2d,
-                      vk_filter](vk::CommandBuffer cmdbuf) {
+                      vk_filter, up_scaling](vk::CommandBuffer cmdbuf) {
         const VkOffset2D src_size{
-            .x = static_cast<s32>(extent.width),
-            .y = static_cast<s32>(extent.height),
+            .x = static_cast<s32>(up_scaling ? extent.width : resolution.ScaleUp(extent.width)),
+            .y = static_cast<s32>(is_2d && up_scaling ? extent.height
+                                                      : resolution.ScaleUp(extent.height)),
         };
         const VkOffset2D dst_size{
-            .x = static_cast<s32>(resolution.ScaleUp(extent.width)),
-            .y = static_cast<s32>(is_2d ? resolution.ScaleUp(extent.height) : extent.height),
+            .x = static_cast<s32>(up_scaling ? resolution.ScaleUp(extent.width) : extent.width),
+            .y = static_cast<s32>(is_2d && up_scaling ? resolution.ScaleUp(extent.height)
+                                                      : extent.height),
         };
         boost::container::small_vector<VkImageBlit, 4> regions;
         regions.reserve(resources.levels);
@@ -1134,6 +1137,7 @@ bool Image::ScaleUp() {
     if (!resolution.active) {
         return false;
     }
+    scale_count++;
     const auto& device = runtime->device;
     const bool is_2d = info.type == ImageType::e2D;
     const u32 scaled_width = resolution.ScaleUp(info.size.width);
@@ -1161,8 +1165,10 @@ bool Image::ScaleUp() {
         using namespace VideoCommon;
         static constexpr auto BLIT_OPERATION = Tegra::Engines::Fermi2D::Operation::SrcCopy;
 
+        if (!scale_view) {
             const auto view_info = ImageViewInfo(ImageViewType::e2D, info.format);
             scale_view = std::make_unique<ImageView>(*runtime, view_info, NULL_IMAGE_ID, *this);
+        }
         auto* view_ptr = scale_view.get();
 
         const Region2D src_region{
@@ -1178,7 +1184,10 @@ bool Image::ScaleUp() {
             .height = scaled_height,
         };
         if (aspect_mask == VK_IMAGE_ASPECT_COLOR_BIT) {
-            scale_framebuffer = std::make_unique<Framebuffer>(*runtime, view_ptr, nullptr, extent);
+            if (!scale_framebuffer) {
+                scale_framebuffer =
+                    std::make_unique<Framebuffer>(*runtime, view_ptr, nullptr, extent);
+            }
             const auto color_view = scale_view->Handle(Shader::TextureType::Color2D);
 
             runtime->blit_image_helper.BlitColor(
@@ -1186,7 +1195,10 @@ bool Image::ScaleUp() {
                 Tegra::Engines::Fermi2D::Filter::Bilinear, BLIT_OPERATION);
         } else if (!runtime->device.IsBlitDepthStencilSupported() &&
                    aspect_mask == (VK_IMAGE_ASPECT_DEPTH_BIT | VK_IMAGE_ASPECT_STENCIL_BIT)) {
-            scale_framebuffer = std::make_unique<Framebuffer>(*runtime, nullptr, view_ptr, extent);
+            if (!scale_framebuffer) {
+                scale_framebuffer =
+                    std::make_unique<Framebuffer>(*runtime, view_ptr, nullptr, extent);
+            }
             runtime->blit_image_helper.BlitDepthStencil(
                 scale_framebuffer.get(), scale_view->DepthView(), scale_view->StencilView(),
                 dst_region, src_region, Tegra::Engines::Fermi2D::Filter::Point, BLIT_OPERATION);
@@ -1209,6 +1221,67 @@ bool Image::ScaleDown() {
     if (!resolution.active) {
         return false;
     }
+    const auto& device = runtime->device;
+    const bool is_2d = info.type == ImageType::e2D;
+    const u32 scaled_width = resolution.ScaleUp(info.size.width);
+    const u32 scaled_height = is_2d ? resolution.ScaleUp(info.size.height) : info.size.height;
+    if (aspect_mask == 0) {
+        aspect_mask = ImageAspectMask(info.format);
+    }
+    static constexpr auto OPTIMAL_FORMAT = FormatType::Optimal;
+    const PixelFormat format = StorageFormat(info.format);
+    const auto vk_format = MaxwellToVK::SurfaceFormat(device, OPTIMAL_FORMAT, false, format).format;
+    const auto blit_usage = VK_FORMAT_FEATURE_BLIT_SRC_BIT | VK_FORMAT_FEATURE_BLIT_DST_BIT;
+    if (device.IsFormatSupported(vk_format, blit_usage, OPTIMAL_FORMAT)) {
+        BlitScale(*scheduler, *scaled_image, *original_image, info, aspect_mask, resolution, false);
+    } else {
+        using namespace VideoCommon;
+        static constexpr auto BLIT_OPERATION = Tegra::Engines::Fermi2D::Operation::SrcCopy;
+
+        if (!normal_view) {
+            const auto view_info = ImageViewInfo(ImageViewType::e2D, info.format);
+            normal_view = std::make_unique<ImageView>(*runtime, view_info, NULL_IMAGE_ID, *this);
+        }
+        auto* view_ptr = normal_view.get();
+
+        const Region2D src_region{
+            .start = {0, 0},
+            .end = {static_cast<s32>(scaled_width), static_cast<s32>(scaled_height)},
+        };
+        const Region2D dst_region{
+            .start = {0, 0},
+            .end = {static_cast<s32>(info.size.width), static_cast<s32>(info.size.height)},
+        };
+        const VkExtent2D extent{
+            .width = scaled_width,
+            .height = scaled_height,
+        };
+        if (aspect_mask == VK_IMAGE_ASPECT_COLOR_BIT) {
+            if (!normal_framebuffer) {
+                normal_framebuffer =
+                    std::make_unique<Framebuffer>(*runtime, view_ptr, nullptr, extent);
+            }
+            const auto color_view = normal_view->Handle(Shader::TextureType::Color2D);
+
+            runtime->blit_image_helper.BlitColor(
+                normal_framebuffer.get(), color_view, dst_region, src_region,
+                Tegra::Engines::Fermi2D::Filter::Bilinear, BLIT_OPERATION);
+        } else if (!runtime->device.IsBlitDepthStencilSupported() &&
+                   aspect_mask == (VK_IMAGE_ASPECT_DEPTH_BIT | VK_IMAGE_ASPECT_STENCIL_BIT)) {
+            if (!normal_framebuffer) {
+                normal_framebuffer =
+                    std::make_unique<Framebuffer>(*runtime, view_ptr, nullptr, extent);
+            }
+            runtime->blit_image_helper.BlitDepthStencil(
+                normal_framebuffer.get(), normal_view->DepthView(), normal_view->StencilView(),
+                dst_region, src_region, Tegra::Engines::Fermi2D::Filter::Point, BLIT_OPERATION);
+        } else {
+            // TODO: Use helper blits where applicable
+            flags &= ~ImageFlagBits::Rescaled;
+            LOG_ERROR(Render_Vulkan, "Device does not support scaling format {}", format);
+            return false;
+        }
+    }
     ASSERT(info.type != ImageType::Linear);
     current_image = *original_image;
     return true;

src/video_core/renderer_vulkan/vk_texture_cache.h

@@ -148,6 +148,9 @@ private:
 
     std::unique_ptr<Framebuffer> scale_framebuffer;
     std::unique_ptr<ImageView> scale_view;
+
+    std::unique_ptr<Framebuffer> normal_framebuffer;
+    std::unique_ptr<ImageView> normal_view;
 };
 
 class ImageView : public VideoCommon::ImageViewBase {

src/video_core/texture_cache/image_base.cpp

@@ -60,8 +60,8 @@ namespace {
 ImageBase::ImageBase(const ImageInfo& info_, GPUVAddr gpu_addr_, VAddr cpu_addr_)
     : info{info_}, guest_size_bytes{CalculateGuestSizeInBytes(info)},
       unswizzled_size_bytes{CalculateUnswizzledSizeBytes(info)},
-      converted_size_bytes{CalculateConvertedSizeBytes(info)}, scale_rating{},
-      scale_tick{}, gpu_addr{gpu_addr_}, cpu_addr{cpu_addr_},
+      converted_size_bytes{CalculateConvertedSizeBytes(info)}, scale_rating{}, scale_tick{},
+      scale_count{}, gpu_addr{gpu_addr_}, cpu_addr{cpu_addr_},
       cpu_addr_end{cpu_addr + guest_size_bytes}, mip_level_offsets{CalculateMipLevelOffsets(info)} {
     if (info.type == ImageType::e3D) {
         slice_offsets = CalculateSliceOffsets(info);

src/video_core/texture_cache/image_base.h

@@ -77,6 +77,10 @@ struct ImageBase {
     void CheckBadOverlapState();
     void CheckAliasState();
 
+    bool HasScaled() {
+        return scale_count > 0;
+    }
+
     ImageInfo info;
 
     u32 guest_size_bytes = 0;
@@ -84,6 +88,7 @@ struct ImageBase {
     u32 converted_size_bytes = 0;
     u32 scale_rating = 0;
     u64 scale_tick = 0;
+    u32 scale_count = 0;
     ImageFlagBits flags = ImageFlagBits::CpuModified;
 
     GPUVAddr gpu_addr = 0;

src/video_core/texture_cache/texture_cache.h

@@ -854,8 +854,8 @@ void TextureCache<P>::InvalidateScale(Image& image) {
 }
 
 template <class P>
-u64 TextureCache<P>::GetScaledImageSizeBytes(Image& image) {
-    const f32 add_to_size = Settings::values.resolution_info.up_factor - 1.0f;
+u64 TextureCache<P>::GetScaledImageSizeBytes(ImageBase& image) {
+    const f32 add_to_size = Settings::values.resolution_info.up_factor;
     const bool sign = std::signbit(add_to_size);
     const u32 image_size_bytes = std::max(image.guest_size_bytes, image.unswizzled_size_bytes);
     const u64 tentative_size = image_size_bytes * static_cast<u64>(std::abs(add_to_size));
@@ -865,11 +865,14 @@ u64 TextureCache<P>::GetScaledImageSizeBytes(Image& image) {
 
 template <class P>
 bool TextureCache<P>::ScaleUp(Image& image) {
+    const bool has_copy = image.HasScaled();
     const bool rescaled = image.ScaleUp();
     if (!rescaled) {
         return false;
     }
+    if (!has_copy) {
         total_used_memory += GetScaledImageSizeBytes(image);
+    }
     InvalidateScale(image);
     return true;
 }
@@ -880,7 +883,10 @@ bool TextureCache<P>::ScaleDown(Image& image) {
     if (!rescaled) {
         return false;
     }
+    const bool has_copy = image.HasScaled();
+    if (!has_copy) {
         total_used_memory -= GetScaledImageSizeBytes(image);
+    }
     InvalidateScale(image);
     return true;
 }
@@ -1391,13 +1397,6 @@ void TextureCache<P>::UnregisterImage(ImageId image_id) {
                "Trying to unregister an already registered image");
     image.flags &= ~ImageFlagBits::Registered;
     image.flags &= ~ImageFlagBits::BadOverlap;
-    u64 tentative_size = std::max(image.guest_size_bytes, image.unswizzled_size_bytes);
-    if ((IsPixelFormatASTC(image.info.format) &&
-         True(image.flags & ImageFlagBits::AcceleratedUpload)) ||
-        True(image.flags & ImageFlagBits::Converted)) {
-        tentative_size = EstimatedDecompressedSize(tentative_size, image.info.format);
-    }
-    total_used_memory -= Common::AlignUp(tentative_size, 1024);
     lru_cache.Free(image.lru_index);
     const auto& clear_page_table =
         [this, image_id](
@@ -1478,6 +1477,16 @@ template <class P>
 void TextureCache<P>::TrackImage(ImageBase& image, ImageId image_id) {
     ASSERT(False(image.flags & ImageFlagBits::Tracked));
     image.flags |= ImageFlagBits::Tracked;
+    if (image.HasScaled()) {
+        total_used_memory -= GetScaledImageSizeBytes(image);
+    }
+    u64 tentative_size = std::max(image.guest_size_bytes, image.unswizzled_size_bytes);
+    if ((IsPixelFormatASTC(image.info.format) &&
+         True(image.flags & ImageFlagBits::AcceleratedUpload)) ||
+        True(image.flags & ImageFlagBits::Converted)) {
+        tentative_size = EstimatedDecompressedSize(tentative_size, image.info.format);
+    }
+    total_used_memory -= Common::AlignUp(tentative_size, 1024);
     if (False(image.flags & ImageFlagBits::Sparse)) {
         rasterizer.UpdatePagesCachedCount(image.cpu_addr, image.guest_size_bytes, 1);
         return;

src/video_core/texture_cache/texture_cache_base.h

@@ -331,7 +331,7 @@ private:
     void InvalidateScale(Image& image);
     bool ScaleUp(Image& image);
     bool ScaleDown(Image& image);
-    u64 GetScaledImageSizeBytes(Image& image);
+    u64 GetScaledImageSizeBytes(ImageBase& image);
 
     Runtime& runtime;
     VideoCore::RasterizerInterface& rasterizer;