The following revision refers to this bug:
  https://chromium.googlesource.com/chromium/src.git/+/3972370575795597962787da7449243bd1d5d57b

commit 3972370575795597962787da7449243bd1d5d57b
Author: yunchao.he <yunchao.he@intel.com>
Date: Thu Aug 11 07:32:02 2016

Command buffer: CopyTexSubImage3D should clear the 3D texture if it is uncleared

BUG= 631934 ,  636626 
CQ_INCLUDE_TRYBOTS=master.tryserver.chromium.linux:linux_optional_gpu_tests_rel;master.tryserver.chromium.mac:mac_optional_gpu_tests_rel;master.tryserver.chromium.win:win_optional_gpu_tests_rel

Review-Url: https://codereview.chromium.org/2232743002
Cr-Commit-Position: refs/heads/master@{#411283}

[modify] https://crrev.com/3972370575795597962787da7449243bd1d5d57b/content/test/gpu/gpu_tests/webgl2_conformance_expectations.py
[modify] https://crrev.com/3972370575795597962787da7449243bd1d5d57b/gpu/command_buffer/service/gles2_cmd_decoder.cc
[modify] https://crrev.com/3972370575795597962787da7449243bd1d5d57b/gpu/command_buffer/service/gles2_cmd_decoder_unittest_textures.cc

(Sorry, wrote comments to a wrong crbug).
Zhenyao and Ken, I have one question about clear the uncleared texture. 

As you know, on command buffer service side, we have a class TextureManager to manage the texture related information. In TextureManager, we have complicated code to set/get the cleared rect for a particular mipmap level for 2D texture. 
But for 3D texture, we just simply clear the entire texture. So I wonder do we also need to upgrade the code to clear a small rect for a specific level/layer for 3D texture? In this CL: https://codereview.chromium.org/2208733002/. I tried to follow the design and add LayerInfo to LevelInfo to clear a small rect for a specific layer of a mipmap level for 3D texture. 
But I am not sure whether this logic is necessary. 
Why it is necessary for 2D/3D texture to clear a small rect? For instance, If a specific level of the texture is 4*4, and we only read a 2*2 rect from fbo by CopyTex{Sub}Image2D to the texture, then we just clear the 2*2 rect. If we read another 2*2 next to the previous rect, we will clear this 2*2 rect, and combine the cleared rect to 2*4. why don't we clear the entire mipmap level (4*4) at the first time, and don't need to care about the cleared rect in following operations? In addition, we don't need so many complicated code to manage the cleared rect.
 
I guess that we can initialize as small part as possible for texture. So we can upload a small bunch of data and save the memory bandwidth and improve the perf for some cases. If this is true, should we clear a small part for 3D texture too, instead of clearing the entire 3D texture? You know, a 3D texture may be very big, need to upload lots of data (zeros).

Any comments?

add piman to the cc list

For 2D textures it's necessary for performance reasons.  For example, compositor might allocate a large texture and gradually fill it up.  So we could avoid the initialization-to-zero if we fill up the entire texture before using it.

This is not a use case for 3D textures. As you mentioned, it's complicated logic (even more complicated for 3D) so we just do simple lazy initialization for 3D textures.  I feel that's enough for now.

OK. Thanks for your explanation, Zhenyao. Then simply clear the entire 3D texture during lazy initialization is OK. We don't need to clear a small rect for a specific layer/level for 3D texture. 

So this crbug is fixed. I will close this change which try to clear a small rect for a particular layer/level for 3D texture at https://codereview.chromium.org/2208733002/.