cc: Use max_bytes_per_copy_operation setting to determine one-copy flush interval.

cc/raster/one_copy_tile_task_worker_pool.cc

Index: cc/raster/one_copy_tile_task_worker_pool.cc
diff --git a/cc/raster/one_copy_tile_task_worker_pool.cc b/cc/raster/one_copy_tile_task_worker_pool.cc
index f4db3993fd723d0f04c1f613ae564cb71f1721d1..ee513a117f43b4207699699c557f7699033d93f7 100644
--- a/cc/raster/one_copy_tile_task_worker_pool.cc
+++ b/cc/raster/one_copy_tile_task_worker_pool.cc
@@ -74,9 +74,6 @@ class RasterBufferImpl : public RasterBuffer {
   DISALLOW_COPY_AND_ASSIGN(RasterBufferImpl);
 };
 
-// Flush interval when performing copy operations.
-const int kCopyFlushPeriod = 4;
-
 // Number of in-flight copy operations to allow.
 const int kMaxCopyOperations = 32;
 
@@ -135,6 +132,7 @@ OneCopyTileTaskWorkerPool::OneCopyTileTaskWorkerPool(
       last_flushed_copy_operation_(0),
       lock_(),
       copy_operation_count_cv_(&lock_),
+      bytes_scheduled_since_last_flush_(0),
       issued_copy_operation_count_(0),
       next_copy_operation_sequence_(1),
       check_for_completed_copy_operations_pending_(false),
@@ -354,13 +352,19 @@ OneCopyTileTaskWorkerPool::PlaybackAndScheduleCopyOnWorkerThread(
     // Acquire a sequence number for this copy operation.
     sequence = next_copy_operation_sequence_++;
 
+    // Increment |bytes_scheduled_since_last_flush_| by the amount of memory
+    // used for this copy operation.
+    bytes_scheduled_since_last_flush_ += rows_to_copy * bytes_per_row;

[vmpstr, 2015/06/01 18:35:52]

So for big tiles (that are split up into several copies) this would cause us to
possibly post several tasks for the one tile (one per full copy)?

[reveman, 2015/06/04 01:19:51]

Yup, that's the behavior we want.

+
     // Post task that will advance last flushed copy operation to |sequence|
-    // if we have reached the flush period.
-    if ((sequence % kCopyFlushPeriod) == 0) {
+    // when |bytes_scheduled_since_last_flush_| has reached
+    // |max_bytes_per_copy_operation_|.
+    if (bytes_scheduled_since_last_flush_ >= max_bytes_per_copy_operation_) {

[vmpstr, 2015/06/02 23:23:27]

I think you need to DCHECK that max_bytes_per_copy_operation_ is less than
std::numeric_limits<size_t>::max(), since then we would never flush?

Either that or change the layer tree settings default to something smaller.

[reveman, 2015/06/04 01:19:51]

A flush here is not required for correct behavior. numeric_limits<size_t>::max()
should work fine. This flush is just to reduce the size of copy operation
batches sent to the GPU process.

       task_runner_->PostTask(
           FROM_HERE,
           base::Bind(&OneCopyTileTaskWorkerPool::AdvanceLastFlushedCopyTo,
                      weak_ptr_factory_.GetWeakPtr(), sequence));
+      bytes_scheduled_since_last_flush_ = 0;
     }
   }