Add display-resolution caching to GPU IDC

cc/tiles/gpu_image_decode_controller.cc

Index: cc/tiles/gpu_image_decode_controller.cc
diff --git a/cc/tiles/gpu_image_decode_controller.cc b/cc/tiles/gpu_image_decode_controller.cc
index 8a76daf9cd20278e89c9aed82e6aeb61f9fe6846..2e8318f5666a7f3905504a080ea84d377f4dc61d 100644
--- a/cc/tiles/gpu_image_decode_controller.cc
+++ b/cc/tiles/gpu_image_decode_controller.cc
@@ -14,6 +14,7 @@
 #include "cc/output/context_provider.h"
 #include "cc/raster/tile_task.h"
 #include "cc/resources/resource_format_utils.h"
+#include "cc/tiles/mipmap_util.h"
 #include "gpu/command_buffer/client/context_support.h"
 #include "gpu/command_buffer/client/gles2_interface.h"
 #include "gpu_image_decode_controller.h"
@@ -46,16 +47,75 @@ bool SkipImage(const DrawImage& draw_image) {
 }
 
 SkImage::DeferredTextureImageUsageParams ParamsFromDrawImage(
-    const DrawImage& draw_image) {
+    const DrawImage& draw_image,
+    int pre_scale_mip_level) {
   SkImage::DeferredTextureImageUsageParams params;
   params.fMatrix = draw_image.matrix();
   params.fQuality = draw_image.filter_quality();
+  params.fPreScaleMipLevel = pre_scale_mip_level;
 
   return params;
 }
 
+// Calculate the mp level to pre-scale the image to before uploading. We use mip

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

s/mp/mip/

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

Done.

+// levels rather than exact scales to increase re-use of scaled images.
+int CalculatePreScaleMipLevel(const DrawImage& draw_image) {

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

Can we rename "Pre" to "Upload"

[ericrk, 2016/06/22 18:56:37]

Done.

+  if (draw_image.src_rect() != draw_image.image()->bounds()) {
+    // Images which are being clipped will have color-bleeding if scaled.

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

Move the comment before the if, please

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

Done.

+    // TODO(ericrk): Investigate uploading clipped images to handle this case
+    // and provide further optimization. crbug.com/620899
+    return 0;
+  }
+
+  gfx::Size base_size(draw_image.image()->width(),
+                      draw_image.image()->height());
+  // Ceil our scaled size so that the mip map generated is guaranteed to be
+  // larger.
+  gfx::Size scaled_size = gfx::ScaleToCeiledSize(
+      base_size, draw_image.scale().width(), draw_image.scale().height());
+
+  return MipMapUtil::GetLevelForSize(base_size, scaled_size);
+}
+
+SkSize CalculatePreScaleFactor(const DrawImage& draw_image, int mip_level) {

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

Put a comment before the function please

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

Done.

+  gfx::Size base_size(draw_image.image()->width(),
+                      draw_image.image()->height());
+  return MipMapUtil::GetScaleAdjustmentForLevel(base_size, mip_level);
+}
+
+// Generates a uint64_t which uniquely identifies a DrawImage for the purposes
+// of the |in_use_cache_|. The key is generated as follows:
+//   ╔══════════════════════╤═══════════╤═══════════╗
+//   ║       image_id       │ mip_level │  quality  ║
+//   ╚════════32═bits═══════╧══16═bits══╧══16═bits══╝
+uint64_t GenerateInUseCacheKey(const DrawImage& draw_image) {
+  static_assert(
+      kLast_SkFilterQuality <= std::numeric_limits<uint16_t>::max(),
+      "InUseCacheKey depends on SkFilterQuality fitting in a uint16_t.");
+
+  SkFilterQuality filter_quality = draw_image.filter_quality();
+  DCHECK_LE(filter_quality, kLast_SkFilterQuality);
+
+  // An image has at most log_2(max(width, height)) mip levels, so given our
+  // usage of 32-bit sizes for images, key.mip_level is at most 31.
+  int32_t mip_level = CalculatePreScaleMipLevel(draw_image);
+  DCHECK_LT(mip_level, 32);
+
+  return (static_cast<uint64_t>(draw_image.image()->uniqueID()) << 32) |
+         (mip_level << 16) | draw_image.filter_quality();
+}
+
 }  // namespace
 
+GpuImageDecodeController::InUseCacheEntry::InUseCacheEntry(
+    scoped_refptr<ImageData> image_data)
+    : image_data(std::move(image_data)) {}
+GpuImageDecodeController::InUseCacheEntry::InUseCacheEntry(
+    const InUseCacheEntry&) = default;
+GpuImageDecodeController::InUseCacheEntry::InUseCacheEntry(InUseCacheEntry&&) =
+    default;
+GpuImageDecodeController::InUseCacheEntry::~InUseCacheEntry() = default;
+
 // Task which decodes an image and stores the result in discardable memory.
 // This task does not use GPU resources and can be run on any thread.
 class ImageDecodeTaskImpl : public TileTask {
@@ -175,12 +235,12 @@ void GpuImageDecodeController::DecodedImageData::ResetData() {
 }
 
 void GpuImageDecodeController::DecodedImageData::ReportUsageStats() const {
-  // lock_count | used  | result state
-  // ===========+=======+==================
-  //  1         | false | WASTED_ONCE
-  //  1         | true  | USED_ONCE
-  //  >1        | false | WASTED_RELOCKED
-  //  >1        | true  | USED_RELOCKED
+  // lock_count │ used  │ result state
+  // ═══════════╪═══════╪══════════════════

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

nice!

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

Done.

+  //  1         │ false │ WASTED_ONCE
+  //  1         │ true  │ USED_ONCE
+  //  >1        │ false │ WASTED_RELOCKED
+  //  >1        │ true  │ USED_RELOCKED
   // Note that it's important not to reorder the following enums, since the
   // numerical values are used in the histogram code.
   enum State : int {
@@ -231,18 +291,33 @@ void GpuImageDecodeController::UploadedImageData::ReportUsageStats() const {
                         usage_stats_.first_ref_wasted);
 }
 
-GpuImageDecodeController::ImageData::ImageData(DecodedDataMode mode,
-                                               size_t size)
-    : mode(mode), size(size) {}
-
-GpuImageDecodeController::ImageData::~ImageData() = default;
+GpuImageDecodeController::ImageData::ImageData(
+    DecodedDataMode mode,
+    size_t size,
+    int pre_scale_mip_level,
+    SkFilterQuality pre_scale_filter_quality)
+    : mode(mode),
+      size(size),
+      pre_scale_mip_level(pre_scale_mip_level),
+      pre_scale_filter_quality(pre_scale_filter_quality) {}
+
+GpuImageDecodeController::ImageData::~ImageData() {
+  // We should never delete ImageData while it is in use or before it has been
+  // cleaned up.
+  DCHECK_EQ(0u, upload.ref_count);
+  DCHECK_EQ(0u, decode.ref_count);
+  DCHECK_EQ(false, decode.is_locked());
+  // This should always be cleaned up before deleting the image, as it needs to
+  // be freed with the GL context lock held.
+  DCHECK(!upload.image());
+}
 
 GpuImageDecodeController::GpuImageDecodeController(ContextProvider* context,
                                                    ResourceFormat decode_format,
                                                    size_t max_gpu_image_bytes)
     : format_(decode_format),
       context_(context),
-      image_data_(ImageDataMRUCache::NO_AUTO_EVICT),
+      persistent_cache_(ImageDataMRUCache::NO_AUTO_EVICT),
       cached_items_limit_(kMaxDiscardableItems),
       cached_bytes_limit_(max_gpu_image_bytes),
       bytes_used_(0),
@@ -285,10 +360,8 @@ bool GpuImageDecodeController::GetTaskForImageAndRef(
 
   base::AutoLock lock(lock_);
   const auto image_id = draw_image.image()->uniqueID();
-
-  auto found = image_data_.Get(image_id);
-  if (found != image_data_.end()) {
-    ImageData* image_data = found->second.get();
+  ImageData* image_data = GetImageDataForDrawImage(draw_image);
+  if (image_data) {
     if (image_data->is_at_raster) {
       // Image is at-raster, just return, this usage will be at-raster as well.
       *task = nullptr;
@@ -307,32 +380,28 @@ bool GpuImageDecodeController::GetTaskForImageAndRef(
       *task = nullptr;
       return true;
     }
-  }
 
-  // We didn't have a pre-uploaded image, so we need an upload task. Try to find
-  // an existing one.
-  scoped_refptr<TileTask>& existing_task =
-      pending_image_upload_tasks_[image_id];
-  if (existing_task) {
-    // We had an existing upload task, ref the image and return the task.
-    RefImage(draw_image);
-    *task = existing_task;
-    return true;
+    // We didn't have a pre-uploaded image, so we need an upload task. Try to
+    // find an existing one.
+    if (image_data->upload.task) {
+      // We had an existing upload task, ref the image and return the task.
+      RefImage(draw_image);
+      *task = image_data->upload.task;
+      return true;
+    } else {
+    }
   }
 
   // We will be creating a new upload task. If necessary, create a placeholder
   // ImageData to hold the result.
-  std::unique_ptr<ImageData> new_data;
-  ImageData* data;
-  if (found == image_data_.end()) {
+  scoped_refptr<ImageData> new_data;
+  if (!image_data) {
     new_data = CreateImageData(draw_image);
-    data = new_data.get();
-  } else {
-    data = found->second.get();
+    image_data = new_data.get();
   }
 
   // Ensure that the image we're about to decode/upload will fit in memory.
-  if (!EnsureCapacity(data->size)) {
+  if (!EnsureCapacity(image_data->size)) {
     // Image will not fit, do an at-raster decode.
     *task = nullptr;
     return false;
@@ -341,19 +410,19 @@ bool GpuImageDecodeController::GetTaskForImageAndRef(
   // If we had to create new image data, add it to our map now that we know it
   // will fit.
   if (new_data)
-    found = image_data_.Put(image_id, std::move(new_data));
+    persistent_cache_.Put(image_id, std::move(new_data));
 
   // Ref image and create a upload and decode tasks. We will release this ref
   // in UploadTaskCompleted.
   RefImage(draw_image);
-  existing_task = make_scoped_refptr(new ImageUploadTaskImpl(
+  *task = make_scoped_refptr(new ImageUploadTaskImpl(
       this, draw_image, GetImageDecodeTaskAndRef(draw_image, tracing_info),
       tracing_info));
+  image_data->upload.task = *task;
 
   // Ref the image again - this ref is owned by the caller, and it is their
   // responsibility to release it by calling UnrefImage.
   RefImage(draw_image);
-  *task = existing_task;
   return true;
 }
 
@@ -374,16 +443,14 @@ DecodedDrawImage GpuImageDecodeController::GetDecodedImageForDraw(
   TRACE_EVENT0("cc", "GpuImageDecodeController::GetDecodedImageForDraw");
 
   base::AutoLock lock(lock_);
-  const uint32_t unique_id = draw_image.image()->uniqueID();
-  auto found = image_data_.Peek(unique_id);
-  if (found == image_data_.end()) {
+  ImageData* image_data = GetImageDataForDrawImage(draw_image);
+  if (!image_data) {
     // We didn't find the image, create a new entry.
     auto data = CreateImageData(draw_image);
-    found = image_data_.Put(unique_id, std::move(data));
+    image_data = data.get();
+    persistent_cache_.Put(draw_image.image()->uniqueID(), std::move(data));
   }
 
-  ImageData* image_data = found->second.get();
-
   if (!image_data->upload.budgeted) {
     // If image data is not budgeted by this point, it is at-raster.
     image_data->is_at_raster = true;
@@ -406,7 +473,9 @@ DecodedDrawImage GpuImageDecodeController::GetDecodedImageForDraw(
   image_data->upload.mark_used();
   DCHECK(image || image_data->decode.decode_failure);
 
-  DecodedDrawImage decoded_draw_image(std::move(image),
+  SkSize scale_factor =
+      CalculatePreScaleFactor(draw_image, image_data->pre_scale_mip_level);
+  DecodedDrawImage decoded_draw_image(std::move(image), SkSize(), scale_factor,
                                       draw_image.filter_quality());
   decoded_draw_image.set_at_raster_decode(image_data->is_at_raster);
   return decoded_draw_image;
@@ -423,6 +492,7 @@ void GpuImageDecodeController::DrawWithImageFinished(
     return;
 
   base::AutoLock lock(lock_);
+  TRACE_EVENT0("cc", "GpuImageDecodeController::DrawWithImageFinished");
   UnrefImageInternal(draw_image);
 
   // We are mid-draw and holding the context lock, ensure we clean up any
@@ -458,7 +528,7 @@ void GpuImageDecodeController::SetShouldAggressivelyFreeResources(
 bool GpuImageDecodeController::OnMemoryDump(
     const base::trace_event::MemoryDumpArgs& args,
     base::trace_event::ProcessMemoryDump* pmd) {
-  for (const auto& image_pair : image_data_) {
+  for (const auto& image_pair : persistent_cache_) {
     const ImageData* image_data = image_pair.second.get();
     const uint32_t image_id = image_pair.first;
 
@@ -469,7 +539,6 @@ bool GpuImageDecodeController::OnMemoryDump(
       base::trace_event::MemoryAllocatorDump* dump =
           image_data->decode.data()->CreateMemoryAllocatorDump(
               discardable_dump_name.c_str(), pmd);
-
       // If our image is locked, dump the "locked_size" as an additional column.
       // This lets us see the amount of discardable which is contributing to
       // memory pressure.
@@ -516,26 +585,29 @@ bool GpuImageDecodeController::OnMemoryDump(
 
 void GpuImageDecodeController::DecodeImage(const DrawImage& draw_image) {
   base::AutoLock lock(lock_);
-  auto found = image_data_.Peek(draw_image.image()->uniqueID());
-  DCHECK(found != image_data_.end());
-  DCHECK(!found->second->is_at_raster);
-  DecodeImageIfNecessary(draw_image, found->second.get());
+  ImageData* image_data = GetImageDataForDrawImage(draw_image);
+  DCHECK(image_data);
+  DCHECK(!image_data->is_at_raster);
+  DecodeImageIfNecessary(draw_image, image_data);
 }
 
 void GpuImageDecodeController::UploadImage(const DrawImage& draw_image) {
   ContextProvider::ScopedContextLock context_lock(context_);
   base::AutoLock lock(lock_);
-  auto found = image_data_.Peek(draw_image.image()->uniqueID());
-  DCHECK(found != image_data_.end());
-  DCHECK(!found->second->is_at_raster);
-  UploadImageIfNecessary(draw_image, found->second.get());
+  ImageData* image_data = GetImageDataForDrawImage(draw_image);
+  DCHECK(image_data);
+  DCHECK(!image_data->is_at_raster);
+  UploadImageIfNecessary(draw_image, image_data);
 }
 
 void GpuImageDecodeController::OnImageDecodeTaskCompleted(
     const DrawImage& draw_image) {
   base::AutoLock lock(lock_);
-  // Decode task is complete, remove it from our list of pending tasks.
-  pending_image_decode_tasks_.erase(draw_image.image()->uniqueID());
+  // Decode task is complete, remove our reference to it.
+  ImageData* image_data = GetImageDataForDrawImage(draw_image);
+  DCHECK(image_data);
+  DCHECK(image_data->decode.task);
+  image_data->decode.task = nullptr;
 
   // While the decode task is active, we keep a ref on the decoded data.
   // Release that ref now.
@@ -545,8 +617,11 @@ void GpuImageDecodeController::OnImageDecodeTaskCompleted(
 void GpuImageDecodeController::OnImageUploadTaskCompleted(
     const DrawImage& draw_image) {
   base::AutoLock lock(lock_);
-  // Upload task is complete, remove it from our list of pending tasks.
-  pending_image_upload_tasks_.erase(draw_image.image()->uniqueID());
+  // Upload task is complete, remove our referene to it.

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

s/ne/nce/

[ericrk, 2016/06/22 18:56:37]

Done.

+  ImageData* image_data = GetImageDataForDrawImage(draw_image);
+  DCHECK(image_data);
+  DCHECK(image_data->upload.task);
+  image_data->upload.task = nullptr;
 
   // While the upload task is active, we keep a ref on both the image it will be
   // populating, as well as the decode it needs to populate it. Release these
@@ -562,24 +637,22 @@ scoped_refptr<TileTask> GpuImageDecodeController::GetImageDecodeTaskAndRef(
     const TracingInfo& tracing_info) {
   lock_.AssertAcquired();
 
-  const uint32_t image_id = draw_image.image()->uniqueID();
-
   // This ref is kept alive while an upload task may need this decode. We
   // release this ref in UploadTaskCompleted.
   RefImageDecode(draw_image);
 
-  auto found = image_data_.Peek(image_id);
-  if (found != image_data_.end() && found->second->decode.is_locked()) {
+  ImageData* image_data = GetImageDataForDrawImage(draw_image);
+  DCHECK(image_data);
+  if (image_data->decode.is_locked()) {
     // We should never be creating a decode task for an at raster image.
-    DCHECK(!found->second->is_at_raster);
+    DCHECK(!image_data->is_at_raster);
     // We should never be creating a decode for an already-uploaded image.
-    DCHECK(!found->second->upload.image());
+    DCHECK(!image_data->upload.image());
     return nullptr;
   }
 
   // We didn't have an existing locked image, create a task to lock or decode.
-  scoped_refptr<TileTask>& existing_task =
-      pending_image_decode_tasks_[image_id];
+  scoped_refptr<TileTask>& existing_task = image_data->decode.task;
   if (!existing_task) {
     // Ref image decode and create a decode task. This ref will be released in
     // DecodeTaskCompleted.
@@ -592,38 +665,64 @@ scoped_refptr<TileTask> GpuImageDecodeController::GetImageDecodeTaskAndRef(
 
 void GpuImageDecodeController::RefImageDecode(const DrawImage& draw_image) {
   lock_.AssertAcquired();
-  auto found = image_data_.Peek(draw_image.image()->uniqueID());
-  DCHECK(found != image_data_.end());
-  ++found->second->decode.ref_count;
-  RefCountChanged(found->second.get());
+  auto found = in_use_cache_.find(GenerateInUseCacheKey(draw_image));
+  DCHECK(found != in_use_cache_.end());
+  ++found->second.ref_count;
+  ++found->second.image_data->decode.ref_count;
+  RefCountChanged(found->second.image_data.get());
 }
 
 void GpuImageDecodeController::UnrefImageDecode(const DrawImage& draw_image) {
   lock_.AssertAcquired();
-  auto found = image_data_.Peek(draw_image.image()->uniqueID());
-  DCHECK(found != image_data_.end());
-  DCHECK_GT(found->second->decode.ref_count, 0u);
-  --found->second->decode.ref_count;
-  RefCountChanged(found->second.get());
+  auto found = in_use_cache_.find(GenerateInUseCacheKey(draw_image));
+  DCHECK(found != in_use_cache_.end());
+  DCHECK_GT(found->second.image_data->decode.ref_count, 0u);
+  DCHECK_GT(found->second.ref_count, 0u);
+  --found->second.ref_count;
+  --found->second.image_data->decode.ref_count;
+  RefCountChanged(found->second.image_data.get());
+  if (found->second.ref_count == 0u) {
+    in_use_cache_.erase(found);
+  }
 }
 
 void GpuImageDecodeController::RefImage(const DrawImage& draw_image) {
   lock_.AssertAcquired();
-  auto found = image_data_.Peek(draw_image.image()->uniqueID());
-  DCHECK(found != image_data_.end());
-  ++found->second->upload.ref_count;
-  RefCountChanged(found->second.get());
+  InUseCacheKey key = GenerateInUseCacheKey(draw_image);
+  auto found = in_use_cache_.find(key);
+
+  // If no secondary cache entry was found for the given |draw_image|, then
+  // the draw_image only exists in the |persistent_cache_|. Create an in-use
+  // cache entry now.
+  if (found == in_use_cache_.end()) {
+    auto found_image = persistent_cache_.Peek(draw_image.image()->uniqueID());
+    DCHECK(found_image != persistent_cache_.end());
+    DCHECK(found_image->second->pre_scale_mip_level <=
+           CalculatePreScaleMipLevel(draw_image));
+    found =
+        in_use_cache_
+            .insert(std::make_pair(key, InUseCacheEntry(found_image->second)))

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

make_pair here will make a std::pair<InUseCacheKey, CacheEntry> and then insert
that into the map, which expects a std::pair<const InUseCacheKey, CacheEntry>
which makes a copy. 

Can we use emplace? I think it might work but might not compile on some systems,
try it out?

or decltype(in_use_cache_)::value_type(key, ...)

[ericrk, 2016/06/22 18:56:37]

Emplace works on windows - will update this again if it fails on other
platforms.

+            .first;
+  }
+
+  DCHECK(found != in_use_cache_.end());
+  ++found->second.ref_count;
+  ++found->second.image_data->upload.ref_count;
+  RefCountChanged(found->second.image_data.get());
 }
 
 void GpuImageDecodeController::UnrefImageInternal(const DrawImage& draw_image) {
   lock_.AssertAcquired();
-  auto found = image_data_.Peek(draw_image.image()->uniqueID());
-  DCHECK(found != image_data_.end());
-  DCHECK_GT(found->second->upload.ref_count, 0u);
-  --found->second->upload.ref_count;
-  if (found->second->upload.ref_count == 0)
-    found->second->upload.notify_ref_reached_zero();
-  RefCountChanged(found->second.get());
+  auto found = in_use_cache_.find(GenerateInUseCacheKey(draw_image));
+  DCHECK(found != in_use_cache_.end());
+  DCHECK_GT(found->second.image_data->upload.ref_count, 0u);
+  DCHECK_GT(found->second.ref_count, 0u);
+  --found->second.ref_count;
+  --found->second.image_data->upload.ref_count;
+  RefCountChanged(found->second.image_data.get());
+  if (found->second.ref_count == 0u) {
+    in_use_cache_.erase(found);
+  }
 }
 
 // Called any time an image or decode ref count changes. Takes care of any
@@ -634,6 +733,12 @@ void GpuImageDecodeController::RefCountChanged(ImageData* image_data) {
   bool has_any_refs =
       image_data->upload.ref_count > 0 || image_data->decode.ref_count > 0;
 
+  // Don't keep around orphaned images.
+  if (image_data->is_orphaned && !has_any_refs) {
+    images_pending_deletion_.push_back(std::move(image_data->upload.image()));
+    image_data->upload.SetImage(nullptr);
+  }
+
   // Don't keep CPU images if they are unused, these images can be recreated by
   // re-locking discardable (rather than requiring a full upload like GPU
   // images).
@@ -719,7 +824,7 @@ bool GpuImageDecodeController::EnsureCapacity(size_t required_size) {
   // things: 1) We can free the uploaded image, reducing the memory usage of
   // the cache and 2) we can remove the entry entirely, reducing the count of
   // elements in the cache.
-  for (auto it = image_data_.rbegin(); it != image_data_.rend();) {
+  for (auto it = persistent_cache_.rbegin(); it != persistent_cache_.rend();) {
     if (it->second->decode.ref_count != 0 ||
         it->second->upload.ref_count != 0) {
       ++it;
@@ -745,7 +850,7 @@ bool GpuImageDecodeController::EnsureCapacity(size_t required_size) {
 
     // Free the entire entry if necessary.
     if (ExceedsPreferredCount()) {
-      it = image_data_.Erase(it);
+      it = persistent_cache_.Erase(it);
     } else {
       ++it;
     }
@@ -770,7 +875,7 @@ bool GpuImageDecodeController::CanFitSize(size_t size) const {
 bool GpuImageDecodeController::ExceedsPreferredCount() const {
   lock_.AssertAcquired();
 
-  return image_data_.size() > cached_items_limit_;
+  return persistent_cache_.size() > cached_items_limit_;
 }
 
 void GpuImageDecodeController::DecodeImageIfNecessary(
@@ -807,10 +912,19 @@ void GpuImageDecodeController::DecodeImageIfNecessary(
         backing_memory =
             base::DiscardableMemoryAllocator::GetInstance()
                 ->AllocateLockedDiscardableMemory(image_data->size);
-        SkImageInfo image_info = CreateImageInfoForDrawImage(draw_image);
-        if (!draw_image.image()->readPixels(image_info, backing_memory->data(),
-                                            image_info.minRowBytes(), 0, 0,
-                                            SkImage::kDisallow_CachingHint)) {
+        SkImageInfo image_info = CreateImageInfoForDrawImage(
+            draw_image, image_data->pre_scale_mip_level);
+        // In order to match GPU scaling quality (which uses mip-maps at high
+        // quality), we want to use at most medium filter quality for the
+        // scale.
+        SkFilterQuality scale_quality =
+            std::max(kMedium_SkFilterQuality, draw_image.filter_quality());
+        SkPixmap image_pixmap(image_info, backing_memory->data(),
+                              image_info.minRowBytes());
+        // Note that scalePixels falls back to readPixels if the sale is 1x, so
+        // no need to special case that as an optimization.
+        if (!draw_image.image()->scalePixels(image_pixmap, scale_quality,
+                                             SkImage::kDisallow_CachingHint)) {
           backing_memory.reset();
         }
         break;
@@ -819,7 +933,8 @@ void GpuImageDecodeController::DecodeImageIfNecessary(
         backing_memory =
             base::DiscardableMemoryAllocator::GetInstance()
                 ->AllocateLockedDiscardableMemory(image_data->size);
-        auto params = ParamsFromDrawImage(draw_image);
+        auto params =
+            ParamsFromDrawImage(draw_image, image_data->pre_scale_mip_level);
         if (!draw_image.image()->getDeferredTextureImageData(
                 *context_threadsafe_proxy_.get(), &params, 1,
                 backing_memory->data())) {
@@ -875,7 +990,8 @@ void GpuImageDecodeController::UploadImageIfNecessary(
     base::AutoUnlock unlock(lock_);
     switch (image_data->mode) {
       case DecodedDataMode::CPU: {
-        SkImageInfo image_info = CreateImageInfoForDrawImage(draw_image);
+        SkImageInfo image_info = CreateImageInfoForDrawImage(
+            draw_image, image_data->pre_scale_mip_level);
         SkPixmap pixmap(image_info, image_data->decode.data()->data(),
                         image_info.minRowBytes());
         uploaded_image =
@@ -899,26 +1015,29 @@ void GpuImageDecodeController::UploadImageIfNecessary(
     image_data->upload.SetImage(std::move(uploaded_image));
 }
 
-std::unique_ptr<GpuImageDecodeController::ImageData>
+scoped_refptr<GpuImageDecodeController::ImageData>
 GpuImageDecodeController::CreateImageData(const DrawImage& draw_image) {
   lock_.AssertAcquired();
 
   DecodedDataMode mode;
-  SkImageInfo info = CreateImageInfoForDrawImage(draw_image);
+  int pre_scale_mip_level = CalculatePreScaleMipLevel(draw_image);
   SkImage::DeferredTextureImageUsageParams params =
-      ParamsFromDrawImage(draw_image);
+      ParamsFromDrawImage(draw_image, pre_scale_mip_level);
   size_t data_size = draw_image.image()->getDeferredTextureImageData(
       *context_threadsafe_proxy_.get(), &params, 1, nullptr);
 
   if (data_size == 0) {
     // Can't upload image, too large or other failure. Try to use SW fallback.
-    data_size = info.getSafeSize(info.minRowBytes());
+    SkImageInfo image_info =
+        CreateImageInfoForDrawImage(draw_image, pre_scale_mip_level);
+    data_size = image_info.getSafeSize(image_info.minRowBytes());
     mode = DecodedDataMode::CPU;
   } else {
     mode = DecodedDataMode::GPU;
   }
 
-  return base::WrapUnique(new ImageData(mode, data_size));
+  return make_scoped_refptr(new ImageData(mode, data_size, pre_scale_mip_level,
+                                          draw_image.filter_quality()));
 }
 
 void GpuImageDecodeController::DeletePendingImages() {
@@ -928,17 +1047,74 @@ void GpuImageDecodeController::DeletePendingImages() {
 }
 
 SkImageInfo GpuImageDecodeController::CreateImageInfoForDrawImage(
-    const DrawImage& draw_image) const {
+    const DrawImage& draw_image,
+    int pre_scale_mip_level) const {
+  DrawImage scaled_draw_image = draw_image.ApplyScale(
+      CalculatePreScaleFactor(draw_image, pre_scale_mip_level));
   return SkImageInfo::Make(
-      draw_image.image()->width(), draw_image.image()->height(),
+      scaled_draw_image.image()->width() * scaled_draw_image.scale().width(),
+      scaled_draw_image.image()->height() * scaled_draw_image.scale().height(),
       ResourceFormatToClosestSkColorType(format_), kPremul_SkAlphaType);
 }
 
+// Tries to find an ImageData that can be used to draw the provided
+// |draw_image|. First looks for an exact entry in our |in_use_cache_|. If one
+// cannot be found, it looks for a compatible entry in our |persistent_cache_|.
+GpuImageDecodeController::ImageData*
+GpuImageDecodeController::GetImageDataForDrawImage(
+    const DrawImage& draw_image) {
+  lock_.AssertAcquired();
+  {
+    auto found = in_use_cache_.find(GenerateInUseCacheKey(draw_image));

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

I prefer if you just name a variable a different thing and remove the extra
scopes

[ericrk, 2016/06/22 18:56:37]

Done.

+    if (found != in_use_cache_.end()) {

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

no braces

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

Done.

+      return found->second.image_data.get();
+    }
+  }
+  {
+    auto found = persistent_cache_.Get(draw_image.image()->uniqueID());
+    if (found != persistent_cache_.end()) {
+      if (IsCompatibleWithDrawImage(found->second.get(), draw_image)) {
+        return found->second.get();
+      } else {
+        found->second->is_orphaned = true;
+        // Call RefCountChanged before erasing the orphaned task to ensure
+        // that we clean up any SkImage safely.
+        RefCountChanged(found->second.get());

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

Where did we change the ref count? Can this be currently used during raster? I'm
not sure I understand the relationship here and why we can erase things.

[ericrk, 2016/06/22 18:56:37]

RefCountChanged really now means "ownership changed" or something like that...
updated the name. We just orphaned a task in the persistent cache - if it has 0
references, we will delete it in the line below. In order to ensure that we
clean up SkImages and byte counts, we need to call
RefCountChanged/OwnershipChanged before deleting it.

+        persistent_cache_.Erase(found);
+      }
+    }
+  }
+
+  return nullptr;
+}
+
+// Determines if we can draw the provided |draw_image| using the provided
+// |image_data|. This is true if the |image_data| is not scaled, or if it
+// is scaled at an equal or larger scale and equal or larger quality to
+// the provided |draw_image|.
+bool GpuImageDecodeController::IsCompatibleWithDrawImage(
+    const ImageData* image_data,
+    const DrawImage& draw_image) const {
+  bool not_scaled = image_data->pre_scale_mip_level == 0;

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

can you name bools in the positive? ie "is_scaled" and then the if check can be
!is_scaled

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

Done.

+  bool scale_is_compatible =
+      CalculatePreScaleMipLevel(draw_image) >= image_data->pre_scale_mip_level;
+  bool quality_is_compatible =
+      draw_image.filter_quality() <= image_data->pre_scale_filter_quality;
+  return not_scaled || (scale_is_compatible && quality_is_compatible);
+}
+
+size_t GpuImageDecodeController::GetDrawImageSizeForTesting(
+    const DrawImage& image) {
+  base::AutoLock lock(lock_);
+  scoped_refptr<ImageData> data = CreateImageData(image);
+  return data->size;
+}
+
 void GpuImageDecodeController::SetImageDecodingFailedForTesting(
     const DrawImage& image) {
   base::AutoLock lock(lock_);
-  auto found = image_data_.Peek(image.image()->uniqueID());
-  DCHECK(found != image_data_.end());
+  auto found = persistent_cache_.Peek(image.image()->uniqueID());
+  DCHECK(found != persistent_cache_.end());
   ImageData* image_data = found->second.get();
   image_data->decode.decode_failure = true;
 }
@@ -946,8 +1122,8 @@ void GpuImageDecodeController::SetImageDecodingFailedForTesting(
 bool GpuImageDecodeController::DiscardableIsLockedForTesting(
     const DrawImage& image) {
   base::AutoLock lock(lock_);
-  auto found = image_data_.Peek(image.image()->uniqueID());
-  DCHECK(found != image_data_.end());
+  auto found = persistent_cache_.Peek(image.image()->uniqueID());
+  DCHECK(found != persistent_cache_.end());
   ImageData* image_data = found->second.get();
   return image_data->decode.is_locked();
 }