Texture Cache: fix memory managment and optimize scaled downloads, uploads.

src/video_core/renderer_opengl/gl_texture_cache.cpp

@@ -697,7 +697,7 @@ void Image::UploadMemory(const ImageBufferMap& map,
                          std::span<const VideoCommon::BufferImageCopy> copies) {
     const bool is_rescaled = True(flags & ImageFlagBits::Rescaled);
     if (is_rescaled) {
-        ScaleDown();
+        ScaleDown(true);
     }
     glBindBuffer(GL_PIXEL_UNPACK_BUFFER, map.buffer);
     glFlushMappedBufferRange(GL_PIXEL_UNPACK_BUFFER, map.offset, unswizzled_size_bytes);
@@ -725,6 +725,10 @@ void Image::UploadMemory(const ImageBufferMap& map,
 
 void Image::DownloadMemory(ImageBufferMap& map,
                            std::span<const VideoCommon::BufferImageCopy> copies) {
+    const bool is_rescaled = True(flags & ImageFlagBits::Rescaled);
+    if (is_rescaled) {
+        ScaleDown();
+    }
     glMemoryBarrier(GL_PIXEL_BUFFER_BARRIER_BIT); // TODO: Move this to its own API
     glBindBuffer(GL_PIXEL_PACK_BUFFER, map.buffer);
     glPixelStorei(GL_PACK_ALIGNMENT, 1);
@@ -743,6 +747,9 @@ void Image::DownloadMemory(ImageBufferMap& map,
         }
         CopyImageToBuffer(copy, map.offset);
     }
+    if (is_rescaled) {
+        ScaleUp(true);
+    }
 }
 
 GLuint Image::StorageHandle() noexcept {
@@ -979,7 +986,7 @@ bool Image::Scale(bool up_scale) {
     return true;
 }
 
-bool Image::ScaleUp() {
+bool Image::ScaleUp(bool ignore) {
     if (True(flags & ImageFlagBits::Rescaled)) {
         return false;
     }
@@ -997,7 +1004,11 @@ bool Image::ScaleUp() {
         flags &= ~ImageFlagBits::Rescaled;
         return false;
     }
-    scale_count++;
+    has_scaled = true;
+    if (ignore) {
+        current_texture = upscaled_backup.handle;
+        return true;
+    }
     if (!Scale()) {
         flags &= ~ImageFlagBits::Rescaled;
         return false;
@@ -1005,7 +1016,7 @@ bool Image::ScaleUp() {
     return true;
 }
 
-bool Image::ScaleDown() {
+bool Image::ScaleDown(bool ignore) {
     if (False(flags & ImageFlagBits::Rescaled)) {
         return false;
     }
@@ -1013,7 +1024,10 @@ bool Image::ScaleDown() {
     if (!runtime->resolution.active) {
         return false;
     }
-    scale_count++;
+    if (ignore) {
+        current_texture = texture.handle;
+        return true;
+    }
     if (!Scale(false)) {
         flags &= ~ImageFlagBits::Rescaled;
         return false;

src/video_core/renderer_opengl/gl_texture_cache.h

@@ -196,9 +196,9 @@ public:
         return gl_type;
     }
 
-    bool ScaleUp();
+    bool ScaleUp(bool ignore = false);
 
-    bool ScaleDown();
+    bool ScaleDown(bool ignore = false);
 
 private:
     void CopyBufferToImage(const VideoCommon::BufferImageCopy& copy, size_t buffer_offset);

src/video_core/renderer_vulkan/vk_texture_cache.cpp

@@ -1055,7 +1055,7 @@ void Image::UploadMemory(const StagingBufferRef& map, std::span<const BufferImag
     // TODO: Move this to another API
     const bool is_rescaled = True(flags & ImageFlagBits::Rescaled);
     if (is_rescaled) {
-        ScaleDown();
+        ScaleDown(true);
     }
     scheduler->RequestOutsideRenderPassOperationContext();
     std::vector vk_copies = TransformBufferImageCopies(copies, map.offset, aspect_mask);
@@ -1073,6 +1073,10 @@ void Image::UploadMemory(const StagingBufferRef& map, std::span<const BufferImag
 }
 
 void Image::DownloadMemory(const StagingBufferRef& map, std::span<const BufferImageCopy> copies) {
+    const bool is_rescaled = True(flags & ImageFlagBits::Rescaled);
+    if (is_rescaled) {
+        ScaleDown();
+    }
     std::vector vk_copies = TransformBufferImageCopies(copies, map.offset, aspect_mask);
     scheduler->RequestOutsideRenderPassOperationContext();
     scheduler->Record([buffer = map.buffer, image = *original_image, aspect_mask = aspect_mask,
@@ -1125,9 +1129,12 @@ void Image::DownloadMemory(const StagingBufferRef& map, std::span<const BufferIm
         cmdbuf.PipelineBarrier(VK_PIPELINE_STAGE_TRANSFER_BIT, VK_PIPELINE_STAGE_ALL_COMMANDS_BIT,
                                0, memory_write_barrier, nullptr, image_write_barrier);
     });
+    if (is_rescaled) {
+        ScaleUp(true);
+    }
 }
 
-bool Image::ScaleUp() {
+bool Image::ScaleUp(bool ignore) {
     if (True(flags & ImageFlagBits::Rescaled)) {
         return false;
     }
@@ -1137,7 +1144,7 @@ bool Image::ScaleUp() {
     if (!resolution.active) {
         return false;
     }
-    scale_count++;
+    has_scaled = true;
     const auto& device = runtime->device;
     const bool is_2d = info.type == ImageType::e2D;
     const u32 scaled_width = resolution.ScaleUp(info.size.width);
@@ -1149,8 +1156,12 @@ bool Image::ScaleUp() {
         scaled_image = MakeImage(device, scaled_info);
         auto& allocator = runtime->memory_allocator;
         scaled_commit = MemoryCommit(allocator.Commit(scaled_image, MemoryUsage::DeviceLocal));
+        ignore = false;
     }
     current_image = *scaled_image;
+    if (ignore) {
+      return true;
+    }
 
     if (aspect_mask == 0) {
         aspect_mask = ImageAspectMask(info.format);
@@ -1212,7 +1223,7 @@ bool Image::ScaleUp() {
     return true;
 }
 
-bool Image::ScaleDown() {
+bool Image::ScaleDown(bool ignore) {
     if (False(flags & ImageFlagBits::Rescaled)) {
         return false;
     }
@@ -1221,6 +1232,10 @@ bool Image::ScaleDown() {
     if (!resolution.active) {
         return false;
     }
+    if (ignore) {
+      current_image = *original_image;
+      return true;
+    }
     const auto& device = runtime->device;
     const bool is_2d = info.type == ImageType::e2D;
     const u32 scaled_width = resolution.ScaleUp(info.size.width);

src/video_core/renderer_vulkan/vk_texture_cache.h

@@ -129,9 +129,9 @@ public:
         return std::exchange(initialized, true);
     }
 
-    bool ScaleUp();
+    bool ScaleUp(bool ignore = false);
 
-    bool ScaleDown();
+    bool ScaleDown(bool ignore = false);
 
 private:
     VKScheduler* scheduler{};

src/video_core/texture_cache/image_base.cpp

@@ -61,7 +61,7 @@ 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{},
-      scale_count{}, gpu_addr{gpu_addr_}, cpu_addr{cpu_addr_},
+      has_scaled{}, 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

@@ -77,8 +77,8 @@ struct ImageBase {
     void CheckBadOverlapState();
     void CheckAliasState();
 
-    bool HasScaled() {
-        return scale_count > 0;
+    bool HasScaled() const {
+        return has_scaled;
     }
 
     ImageInfo info;
@@ -88,7 +88,7 @@ struct ImageBase {
     u32 converted_size_bytes = 0;
     u32 scale_rating = 0;
     u64 scale_tick = 0;
-    u32 scale_count = 0;
+    bool has_scaled = false;
     ImageFlagBits flags = ImageFlagBits::CpuModified;
 
     GPUVAddr gpu_addr = 0;

src/video_core/texture_cache/texture_cache.h

@@ -60,7 +60,7 @@ TextureCache<P>::TextureCache(Runtime& runtime_, VideoCore::RasterizerInterface&
         // On OpenGL we can be more conservatives as the driver takes care.
         expected_memory = DEFAULT_EXPECTED_MEMORY + 512_MiB;
         critical_memory = DEFAULT_CRITICAL_MEMORY + 1_GiB;
-        minimum_memory = expected_memory;
+        minimum_memory = 0;
     }
 }
 
@@ -1464,16 +1464,6 @@ 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;
@@ -1519,6 +1509,16 @@ void TextureCache<P>::UntrackImage(ImageBase& image, ImageId image_id) {
 template <class P>
 void TextureCache<P>::DeleteImage(ImageId image_id) {
     ImageBase& image = slot_images[image_id];
+    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);
     const GPUVAddr gpu_addr = image.gpu_addr;
     const auto alloc_it = image_allocs_table.find(gpu_addr);
     if (alloc_it == image_allocs_table.end()) {