Add display-resolution caching to GPU IDC

cc/tiles/gpu_image_decode_controller.h

 // Copyright 2016 The Chromium Authors. All rights reserved.
 // Use of this source code is governed by a BSD-style license that can be
 // found in the LICENSE file.
 
 #ifndef CC_TILES_GPU_IMAGE_DECODE_CONTROLLER_H_
 #define CC_TILES_GPU_IMAGE_DECODE_CONTROLLER_H_
 
 #include <memory>
 #include <unordered_map>
 #include <vector>
 
 #include "base/containers/mru_cache.h"
 #include "base/memory/discardable_memory.h"
 #include "base/synchronization/lock.h"
 #include "base/trace_event/memory_dump_provider.h"
 #include "cc/base/cc_export.h"
 #include "cc/resources/resource_format.h"
 #include "cc/tiles/image_decode_controller.h"
 #include "third_party/skia/include/core/SkRefCnt.h"
 
 class SkImageTextureData;
 
 namespace cc {
 
 class ContextProvider;
 
+// OVERVIEW:
+//
 // GpuImageDecodeController handles the decode and upload of images that will
 // be used by Skia's GPU raster path. It also maintains a cache of these
 // decoded/uploaded images for later re-use.
 //
 // Generally, when an image is required for raster, GpuImageDecodeController
 // creates two tasks, one to decode the image, and one to upload the image to
 // the GPU. These tasks are completed before the raster task which depends on
 // the image. We need to seperate decode and upload tasks, as decode can occur
 // simultaneously on multiple threads, while upload requires the GL context
 // lock must happen on our non-concurrent raster thread.
 //
 // Decoded and Uploaded image data share a single cache entry. Depending on how
 // far we've progressed, this cache entry may contain CPU-side decoded data,
 // GPU-side uploaded data, or both. Because CPU-side decoded data is stored in
 // discardable memory, and is only locked for short periods of time (until the
 // upload completes), this memory is not counted against our sized cache
 // limits. Uploaded GPU memory, being non-discardable, always counts against
 // our limits.
 //
 // In cases where the number of images needed exceeds our cache limits, we
 // operate in an "at-raster" mode. In this mode, there are no decode/upload
 // tasks, and images are decoded/uploaded as needed, immediately before being
 // used in raster. Cache entries for at-raster tasks are marked as such, which
 // prevents future tasks from taking a dependency on them and extending their
 // lifetime longer than is necessary.
+//
+// RASTER-SCALE CACHING:

[vmpstr, 2016/06/21 20:07:08]

Can you move this to a doc? Can you add a section to it about ref counts? Some
things are ref counted in the "scoped_refptr" sense and some are ref counted in
the "int ref_count" sense, which makes it a bit unclear.

[ericrk, 2016/06/22 18:56:38]

Added a ref-counting section. Will move to a doc once review is complete and
we're happy with the contents (rather than reviewing in two places).

+//
+// In order to save memory, images which are going to be scaled may be uploaded
+// at lower than original resolution. In these cases, we may later need to
+// re-upload the image at a higher resolution. To handle multiple images of
+// different scales being in use at the same time, we have a two-part caching
+// system.
+//
+// The first cache, |persistent_cache_|, stores one ImageData per image id.
+// These ImageDatas are not necessarily associated with a given DrawImage, and
+// are saved (persisted) even when their ref-count reaches zero (assuming they
+// fit in the current memory budget). This allows for future re-use of image
+// resources.
+//
+// The second cache, |in_use_cache_|, stores one image data per DrawImage -
+// this may be the same ImageData that is in the persistent_cache_.  These
+// cache entries are more transient and are deleted as soon as all refs to the
+// given DrawImage are released (the image is no longer in-use).
+//
+// Consider the following two cases, which show the cache-usage patterns.
+//
+// Case 1: Larger (1x) scale used before smaller (0.5x) scale.
+//
+//  Step 1: Create/Find Tasks and ImageData for 1x Scale (GetTaskForImageAndRef)
+//    1a: No existing cache data exists, 1x ImageData created
+//    1b: 1x ImageData cached in |persistent_cache_| with image's unique ID as
+//        key.
+//    1c: 1x ImageData cached in |in_use_cache_| with 1x Scale key.
+//    1d: 1x Decode task created, cached in 1x ImageData.
+//    1e: 1x Upload task created, cached in 1x ImageData.
+//
+//    ╔═══════════════════╗  ╔══════════════════════╗
+//    ║ Persistent Cache  ║  ║    In-Use Cache      ║
+//    ╟───────┬───────────╢  ╟───────────┬──────────╢
+//    ║ image │ 1x        ║  ║ 1x Scale  │ 1x       ║
+//    ║ ID Key│ ImageData ║  ║ Key       │ ImageData║
+//    ╚═══════╧═══════════╝  ╚═══════════╧══════════╝
+//
+//  Step 2: Create/Find Tasks and ImageData for 0.5x Scale
+//    2a: Find 1x ImageData in |persistent_cache_|. 0.5x scale can use this.
+//    2b: 1x ImageData cached in |in_use_cache_| with 0.5x scale key.
+//
+//    ╔═══════════════════╗  ╔══════════════════════╗
+//    ║ Persistent Cache  ║  ║    In-Use Cache      ║
+//    ╟───────┬───────────╢  ╟───────────┬──────────╢
+//    ║ image │ 1x        ║  ║ 1x Scale  │ 1x       ║
+//    ║ ID Key│ ImageData ║  ║ Key       │ ImageData║
+//    ╚═══════╧═══════════╝  ╟───────────┼──────────╢
+//                           ║ 0.5x Scale│ 1x       ║
+//                           ║ Key       │ ImageData║
+//                           ╚═══════════╧══════════╝
+//
+//  Step 4: Draw 1x Scale
+//          1x scale looked up in |in_use_cache_| via 1x key.
+//
+//  Step 5: Draw 0.5x Scale
+//          1x scale looked up in |in_use_cache_| via 0.5x key.
+//
+//  Step 6: Finished, refs released
+//          Tasks are deleted from ImageData.
+//          Entries are deleted from |in_use_cache_|.
+//          1x scale preserved in |persistent_cache_|.
+//
+//    ╔═══════════════════╗  ╔════════════════════╗
+//    ║ Persistent Cache  ║  ║   In-Use Cache     ║
+//    ╟───────┬───────────╢  ╚════════════════════╝
+//    ║ image │ 1x        ║
+//    ║ ID Key│ ImageData ║
+//    ╚═══════╧═══════════╝
+//
+// Case 2: Smaller (0.5x) scale used before larger (1x) scale.
+//
+//  Step 1: Create/Find Tasks and ImageData for 0.5x Scale
+//  (GetTaskForImageAndRef)
+//    1a: No existing cache data exists, 0.5x ImageData created
+//    1b: 0.5x ImageData cached in |persistent_cache_| with image's unique ID
+//        as key.
+//    1c: 0.5x ImageData cached in |in_use_cache_| with 0.5x Scale key.
+//    1d: 0.5x Decode task created, cached in 0.5x ImageData.
+//    1e: 0.5x Upload task created, cached in 0.5x ImageData.
+//
+//    ╔═══════════════════╗  ╔══════════════════════╗
+//    ║ Persistent Cache  ║  ║     In-Use Cache     ║
+//    ╟───────┬───────────╢  ╟───────────┬──────────╢
+//    ║ image │ 0.5       ║  ║ 0.5x Scale│ 0.5x     ║
+//    ║ ID Key│ ImageData ║  ║ Key       │ ImageData║
+//    ╚═══════╧═══════════╝  ╚═══════════╧══════════╝
+//
+//  Step 2: Create/Find Tasks and ImageData for 1x Scale
+//    2a: Find 0.5x ImageData in |persistent_cache_|. 1x scale can't use this.
+//    2b: 0.5x scale is about to be replaced by 1x scale, remove it from the

[vmpstr, 2016/06/21 20:07:08]

When does this happen? What about the case where the 0.5x raster is going to run
before 1x decode has a chance to run?

[ericrk, 2016/06/22 18:56:38]

The 0.5x scale is only removed from the persistent cache, not the in use cache -
the 0.5x draw will use the entry from the in use cache. Updated the comment to
reflect this.

+//        persistent_cache_.
+//    2c: 1x ImageData created
+//    2d: 1x ImageData cached in |persistent_cache_| with image's unique ID as
+//        key.
+//    2e: 1x ImageData cached in |in_use_cache_| with 1x Scale key.
+//    2f: 1x Decode task created, cached in 1x ImageData.
+//    2g: 1x Upload task created, cached in 1x ImageData.
+//
+//    ╔═══════════════════╗  ╔══════════════════════╗
+//    ║ Persistent Cache  ║  ║   In-Use Cache       ║
+//    ╟───────┬───────────╢  ╟───────────┬──────────╢
+//    ║ image │ 1x        ║  ║ 1x Scale  │ 1x       ║
+//    ║ ID Key│ ImageData ║  ║ Key       │ ImageData║
+//    ╚═══════╧═══════════╝  ╟───────────┼──────────╢
+//                           ║ 0.5x Scale│ 0.5x     ║
+//                           ║ Key       │ ImageData║
+//                           ╚═══════════╧══════════╝
+//
+//  Step 4: Draw 0.5x Scale
+//          0.5x scale looked up in |in_use_cache_| via 0.5x key.
+//
+//  Step 5: Draw 1x Scale
+//          1x scale looked up in |in_use_cache_| via 1x key.
+//
+//  Step 6: Finished, refs released
+//          Tasks are deleted from ImageDatas.
+//          Entries are deleted from |in_use_cache_|.
+//          1x scale preserved in |persistent_cache_|.
+//
+//    ╔═══════════════════╗  ╔════════════════════╗
+//    ║ Persistent Cache  ║  ║   In-Use Cache     ║
+//    ╟───────┬───────────╢  ╚════════════════════╝
+//    ║ image │ 1x        ║
+//    ║ ID Key│ ImageData ║
+//    ╚═══════╧═══════════╝
+//
 class CC_EXPORT GpuImageDecodeController
     : public ImageDecodeController,
       public base::trace_event::MemoryDumpProvider {
  public:
   explicit GpuImageDecodeController(ContextProvider* context,
                                     ResourceFormat decode_format,
                                     size_t max_gpu_image_bytes);
   ~GpuImageDecodeController() override;
 
   // ImageDecodeController overrides.
@@ skipping 24 matching lines @@
   void OnImageUploadTaskCompleted(const DrawImage& image);
 
   // For testing only.
   void SetCachedItemLimitForTesting(size_t limit) {
     cached_items_limit_ = limit;
   }
   void SetCachedBytesLimitForTesting(size_t limit) {
     cached_bytes_limit_ = limit;
   }
   size_t GetBytesUsedForTesting() const { return bytes_used_; }
+  size_t GetDrawImageSizeForTesting(const DrawImage& image);
   void SetImageDecodingFailedForTesting(const DrawImage& image);
   bool DiscardableIsLockedForTesting(const DrawImage& image);
 
  private:
   enum class DecodedDataMode { GPU, CPU };
 
   // Stores the CPU-side decoded bits of an image and supporting fields.
   struct DecodedImageData {
     DecodedImageData();
     ~DecodedImageData();
 
     bool is_locked() const { return is_locked_; }
     bool Lock();
     void Unlock();
     void SetLockedData(std::unique_ptr<base::DiscardableMemory> data);
     void ResetData();
     base::DiscardableMemory* data() const { return data_.get(); }
-
     void mark_used() { usage_stats_.used = true; }
 
-    // May be null if image not yet decoded.
     uint32_t ref_count = 0;
     // Set to true if the image was corrupt and could not be decoded.
     bool decode_failure = false;
+    // If non-null, this is the pending decode task for this image.
+    scoped_refptr<TileTask> task;
 
    private:
     struct UsageStats {
       int lock_count = 1;
       bool used = false;
       bool first_lock_wasted = false;
     };
 
     void ReportUsageStats() const;
 
@@ skipping 12 matching lines @@
 
     void mark_used() { usage_stats_.used = true; }
     void notify_ref_reached_zero() {
       if (++usage_stats_.ref_reached_zero_count == 1)
         usage_stats_.first_ref_wasted = !usage_stats_.used;
     }
 
     // True if the image is counting against our memory limits.
     bool budgeted = false;
     uint32_t ref_count = 0;
+    // If non-null, this is the pending upload task for this image.
+    scoped_refptr<TileTask> task;
 
    private:
     struct UsageStats {
       bool used = false;
       bool first_ref_wasted = false;
       int ref_reached_zero_count = 0;
     };
 
     void ReportUsageStats() const;
 
     // May be null if image not yet uploaded / prepared.
     sk_sp<SkImage> image_;
     UsageStats usage_stats_;
   };
 
-  struct ImageData {
-    ImageData(DecodedDataMode mode, size_t size);
-    ~ImageData();
+  struct ImageData : public base::RefCounted<ImageData> {
+    ImageData(DecodedDataMode mode,
+              size_t size,
+              int pre_scale_mip_level,
+              SkFilterQuality pre_scale_filter_quality);
 
     const DecodedDataMode mode;
     const size_t size;
     bool is_at_raster = false;
 
+    // Variables used to identify/track multiple scale levels of a single image.
+    int pre_scale_mip_level;
+    SkFilterQuality pre_scale_filter_quality;
+    bool is_orphaned = false;

[vmpstr, 2016/06/21 20:07:08]

Set the above variables as well (even though they're overridden by the ctor)

[ericrk, 2016/06/22 18:56:38]

Done.

+
     DecodedImageData decode;
     UploadedImageData upload;
+
+   private:
+    friend class base::RefCounted<ImageData>;
+    ~ImageData();
   };
 
-  using ImageDataMRUCache =
-      base::MRUCache<uint32_t, std::unique_ptr<ImageData>>;
+  // A ref-count and ImageData, used to associate the ImageData with a specific
+  // DrawImage in the |in_use_cache_|.
+  struct InUseCacheEntry {
+    explicit InUseCacheEntry(scoped_refptr<ImageData> image_data);
+    InUseCacheEntry(const InUseCacheEntry& other);
+    InUseCacheEntry(InUseCacheEntry&& other);
+    ~InUseCacheEntry();
+
+    uint32_t ref_count = 0;
+    scoped_refptr<ImageData> image_data;
+  };
+
+  // Uniquely identifies (without collisions) a specific DrawImage for use in
+  // the |in_use_cache_|.
+  using InUseCacheKey = uint64_t;
 
   // All private functions should only be called while holding |lock_|. Some
   // functions also require the |context_| lock. These are indicated by
   // additional comments.
 
   // Similar to GetTaskForImageAndRef, but gets the dependent decode task
   // rather than the upload task, if necessary.
   scoped_refptr<TileTask> GetImageDecodeTaskAndRef(
       const DrawImage& image,
       const TracingInfo& tracing_info);
 
   void RefImageDecode(const DrawImage& draw_image);
   void UnrefImageDecode(const DrawImage& draw_image);
   void RefImage(const DrawImage& draw_image);
   void UnrefImageInternal(const DrawImage& draw_image);
   void RefCountChanged(ImageData* image_data);
 
   // Ensures that the cache can hold an element of |required_size|, freeing
   // unreferenced cache entries if necessary to make room.
   bool EnsureCapacity(size_t required_size);
   bool CanFitSize(size_t size) const;
   bool ExceedsPreferredCount() const;
 
   void DecodeImageIfNecessary(const DrawImage& draw_image,
                               ImageData* image_data);
 
-  std::unique_ptr<GpuImageDecodeController::ImageData> CreateImageData(
+  scoped_refptr<GpuImageDecodeController::ImageData> CreateImageData(
       const DrawImage& image);
-  SkImageInfo CreateImageInfoForDrawImage(const DrawImage& draw_image) const;
+  SkImageInfo CreateImageInfoForDrawImage(const DrawImage& draw_image,
+                                          int pre_scale_mip_level) const;
+
+  // Finds the ImageData that should be used for the given DrawImage. Looks
+  // first in the |in_use_cache_|, and then in the |persistent_cache_|.

[vmpstr, 2016/06/21 20:07:08]

When would it check the persistent_cache? From your comment, I'm under the
impression that there would always be a persistent_cache entry? Or do you mean
this function can be used to populate persistent_cache?

[ericrk, 2016/06/22 18:56:38]

If the image has previously been used, but is not currently in-use, it won't
have an in_use_cache_ entry, but will have a persistent_cache_ entry. The
in_use_cache entry is generated the first time we AddRef an ImageData (if it
isn't already in the in_use_cache_)

+  ImageData* GetImageDataForDrawImage(const DrawImage& image);
+
+  // Returns true if the given ImageData can be used to draw the specified
+  // DrawImage.
+  bool IsCompatibleWithDrawImage(const ImageData* image_data,
+                                 const DrawImage& draw_image) const;
 
   // The following two functions also require the |context_| lock to be held.
   void UploadImageIfNecessary(const DrawImage& draw_image,
                               ImageData* image_data);
   void DeletePendingImages();
 
   const ResourceFormat format_;
   ContextProvider* context_;
   sk_sp<GrContextThreadSafeProxy> context_threadsafe_proxy_;
 
   // All members below this point must only be accessed while holding |lock_|.
   base::Lock lock_;
 
-  std::unordered_map<uint32_t, scoped_refptr<TileTask>>
-      pending_image_upload_tasks_;
-  std::unordered_map<uint32_t, scoped_refptr<TileTask>>
-      pending_image_decode_tasks_;
+  // |persistent_cache_| represents the long-lived cache, keeping a certain
+  // budget of ImageDatas alive even when their ref count reaches zero.
+  using ImageDataMRUCache = base::MRUCache<uint32_t, scoped_refptr<ImageData>>;
+  ImageDataMRUCache persistent_cache_;
 
-  ImageDataMRUCache image_data_;
+  // |in_use_cache_| represents the in-use (short-lived) cache. Entries are
+  // cleaned up as soon as their ref count reaches zero.
+  std::unordered_map<InUseCacheKey, InUseCacheEntry> in_use_cache_;
 
   size_t cached_items_limit_;
   size_t cached_bytes_limit_;
   size_t bytes_used_;
   const size_t max_gpu_image_bytes_;
 
   // We can't release GPU backed SkImages without holding the context lock,
   // so we add them to this list and defer deletion until the next time the lock
   // is held.
   std::vector<sk_sp<SkImage>> images_pending_deletion_;
 };
 
 }  // namespace cc
 
 #endif  // CC_TILES_GPU_IMAGE_DECODE_CONTROLLER_H_