gpu: Implement idle async pixel transfers.

gpu/command_buffer/service/texture_manager.cc

 // Copyright (c) 2012 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.
 
 #include "gpu/command_buffer/service/texture_manager.h"
 #include "base/bits.h"
 #include "base/stringprintf.h"
 #include "gpu/command_buffer/common/gles2_cmd_utils.h"
 #include "gpu/command_buffer/service/feature_info.h"
 #include "gpu/command_buffer/service/gles2_cmd_decoder.h"
@@ skipping 1198 matching lines @@
 
 void TextureManager::BindFinishedAsyncPixelTransfers(
     bool* texture_dirty, bool* framebuffer_dirty) {
   DCHECK(texture_dirty);
   DCHECK(framebuffer_dirty);
   *texture_dirty = false;
   *framebuffer_dirty = false;
 
   // Remove finished transfers from the list, while
   // marking whether texture unit 0 or frame_buffer status is dirty.
-  while(!pending_async_transfers_.empty()) {
-    PendingAsyncTransfer state_info = pending_async_transfers_.front();
+  PendingAsyncTransferList::iterator iter = pending_async_transfers_.begin();
+  while (iter != pending_async_transfers_.end()) {

[epenner, 2013/03/08 21:44:57]

I think this makes this more expensive right? Is it due to uploads finishing in
an arbitrary order now?

I think this will benefit from my patch to move this logic into the delegate,
since the single-thread delegate can do the 'late bind' callback on the same
thread when the upload completes, and avoid this loop all together.

[reveman, 2013/03/12 00:31:42]

yes, but I'm pretty sure the number of pending uploads are never enough to have
a significant impact on perf.
Sounds like a good cleanup. How soon do you think we can land that change of
yours? I'd like to get this idle upload stuff landed asap so it's possible to
turn on impl-side painting on cros.

+    PendingAsyncTransfer state_info = *iter;
     if (!state_info.first.get()) {
       // The AsyncState is owned by the Texture. So if the
       // async state is deleted, so is the Texture.
-      pending_async_transfers_.pop_front();
+      pending_async_transfers_.erase(iter++);

[greggman, 2013/03/08 21:06:21]

this is not the correct way to erase in the stl is it?
it should be

  iter = pending_async_tranfers_.erase(iter);

[reveman, 2013/03/12 00:31:42]

changed to that. seem to be a mix in chromium but I like this better.

       continue;
     }
-    // Terminate early, as all transfers finish in order.
-    if (state_info.first->TransferIsInProgress())
-      break;
+    if (state_info.first->TransferIsInProgress()) {
+      iter++;

[greggman, 2013/03/08 21:06:21]

this should be ++iter no?

[reveman, 2013/03/12 00:31:42]

Done. I guess that's a bit more efficient.

+      continue;
+    }
     // If the transfer is finished, bind it to the texture,
     // update the Texture, and remove it from pending list.
     *texture_dirty = true;
     *framebuffer_dirty |= state_info.second->IsAttachedToFramebuffer();
     gfx::AsyncTexImage2DParams tex_define_params;
     state_info.second->
         GetAsyncTransferState()->BindTransfer(&tex_define_params);
     SetLevelInfo(
         state_info.second,
         tex_define_params.target,
         tex_define_params.level,
         tex_define_params.internal_format,
         tex_define_params.width,
         tex_define_params.height,
         1, // depth
         tex_define_params.border,
         tex_define_params.format,
         tex_define_params.type,
         true); // cleared
-    pending_async_transfers_.pop_front();
+    pending_async_transfers_.erase(iter++);

[greggman, 2013/03/08 21:06:21]

same as above

[reveman, 2013/03/12 00:31:42]

Done.

   }
 }
 
 }  // namespace gles2
 }  // namespace gpu