Perform glReadPixels with PBOs in the gpu, if PBOs are available.

Make GL_ASYNC_PIXEL_TRANSFERS_COMPLETED_CHROMIUM wait for readpixel transfers.
Add signalQuery to get a callback when the transfer is done.
PLEASE NOTE: glMapBuffer does not wait for the readpixels transfer to complete anymore.
Nobody is currently relying on that behaviour.
Update gl_helper.cc and gl_renderer.cc to use queries.

BUG=249925

Review URL: https://chromiumcodereview.appspot.com/16831004

git-svn-id: svn://svn.chromium.org/chrome/trunk/src@209625 0039d316-1c4b-4281-b951-d872f2087c98

16 files changed, 301 insertions, 97 deletions

diff --git a/gpu/command_buffer/build_gles2_cmd_buffer.py b/gpu/command_buffer/build_gles2_cmd_buffer.py
index d3fa65a..af49a93 100755
--- a/gpu/command_buffer/build_gles2_cmd_buffer.py
+++ b/gpu/command_buffer/build_gles2_cmd_buffer.py
@@ -826,6 +826,7 @@ _ENUM_LISTS = {
       'GL_COMMANDS_ISSUED_CHROMIUM',
       'GL_LATENCY_QUERY_CHROMIUM',
       'GL_ASYNC_PIXEL_TRANSFERS_COMPLETED_CHROMIUM',
+      'GL_ASYNC_READ_PIXELS_COMPLETED_CHROMIUM',
     ],
   },
   'RenderBufferParameter': {
@@ -1901,7 +1902,8 @@ _FUNCTION_INFO = {
         'GLint x, GLint y, GLsizei width, GLsizei height, '
         'GLenumReadPixelFormat format, GLenumReadPixelType type, '
         'uint32 pixels_shm_id, uint32 pixels_shm_offset, '
-        'uint32 result_shm_id, uint32 result_shm_offset',
+        'uint32 result_shm_id, uint32 result_shm_offset, '
+        'GLboolean async',
     'result': ['uint32'],
     'defer_reads': True,
   },
diff --git a/gpu/command_buffer/client/gles2_cmd_helper_autogen.h b/gpu/command_buffer/client/gles2_cmd_helper_autogen.h
index a290365..9c2898e 100644
--- a/gpu/command_buffer/client/gles2_cmd_helper_autogen.h
+++ b/gpu/command_buffer/client/gles2_cmd_helper_autogen.h
@@ -923,12 +923,12 @@
   void ReadPixels(
       GLint x, GLint y, GLsizei width, GLsizei height, GLenum format,
       GLenum type, uint32 pixels_shm_id, uint32 pixels_shm_offset,
-      uint32 result_shm_id, uint32 result_shm_offset) {
+      uint32 result_shm_id, uint32 result_shm_offset, GLboolean async) {
     gles2::cmds::ReadPixels* c = GetCmdSpace<gles2::cmds::ReadPixels>();
     if (c) {
       c->Init(
           x, y, width, height, format, type, pixels_shm_id, pixels_shm_offset,
-          result_shm_id, result_shm_offset);
+          result_shm_id, result_shm_offset, async);
     }
   }
 
diff --git a/gpu/command_buffer/client/gles2_implementation.cc b/gpu/command_buffer/client/gles2_implementation.cc
index fbc92d5..57dd4e0 100644
--- a/gpu/command_buffer/client/gles2_implementation.cc
+++ b/gpu/command_buffer/client/gles2_implementation.cc
@@ -2236,8 +2236,7 @@ void GLES2Implementation::ReadPixels(
     if (buffer && buffer->shm_id() != -1) {
       helper_->ReadPixels(xoffset, yoffset, width, height, format, type,
                           buffer->shm_id(), buffer->shm_offset(),
-                          0, 0);
-      buffer->set_transfer_ready_token(helper_->InsertToken());
+                          0, 0, true);
       CheckGLError();
     }
     return;
@@ -2269,7 +2268,8 @@ void GLES2Implementation::ReadPixels(
     helper_->ReadPixels(
         xoffset, yoffset, width, num_rows, format, type,
         buffer.shm_id(), buffer.offset(),
-        GetResultShmId(), GetResultShmOffset());
+        GetResultShmId(), GetResultShmOffset(),
+        false);
     WaitForCmd();
     if (*result != 0) {
       // when doing a y-flip we have to iterate through top-to-bottom chunks
@@ -3175,36 +3175,31 @@ void GLES2Implementation::DeleteQueriesEXTHelper(
     return;
   }
   // When you delete a query you can't mark its memory as unused until it's
-  // completed.
+  // either completed, or deleted in the gpu process.
   // Note: If you don't do this you won't mess up the service but you will mess
   // up yourself.
 
-  // TODO(gman): Consider making this faster by putting pending quereies
-  // on some queue to be removed when they are finished.
   bool query_pending = false;
   for (GLsizei ii = 0; ii < n; ++ii) {
     QueryTracker::Query* query = query_tracker_->GetQuery(queries[ii]);
-    if (query && query->Pending()) {
+    if (query && !query->CheckResultsAvailable(helper_)) {
       query_pending = true;
       break;
     }
   }
 
+  helper_->DeleteQueriesEXTImmediate(n, queries);
+
   if (query_pending) {
+    // This should make sure that the GPU process have deleted the queries
+    // and given up any claim on the shared memory that goes along with
+    // those queries so that we can safely re-use the shared memory.
     WaitForCmd();
   }
 
   for (GLsizei ii = 0; ii < n; ++ii) {
-    QueryTracker::Query* query = query_tracker_->GetQuery(queries[ii]);
-    if (query && query->Pending()) {
-      if (!query->CheckResultsAvailable(helper_)) {
-        // Should only get here on context lost.
-        MustBeContextLost();
-      }
-    }
     query_tracker_->RemoveQuery(queries[ii], helper_->IsContextLost());
   }
-  helper_->DeleteQueriesEXTImmediate(n, queries);
 }
 
 // TODO(gman): Remove this. Queries are not shared resources.
diff --git a/gpu/command_buffer/client/gles2_implementation_unittest.cc b/gpu/command_buffer/client/gles2_implementation_unittest.cc
index 7208425..bdbde8f 100644
--- a/gpu/command_buffer/client/gles2_implementation_unittest.cc
+++ b/gpu/command_buffer/client/gles2_implementation_unittest.cc
@@ -1438,11 +1438,12 @@ TEST_F(GLES2ImplementationTest, ReadPixels2Reads) {
   Cmds expected;
   expected.read1.Init(
       0, 0, kWidth, kHeight / 2, kFormat, kType,
-      mem1.id, mem1.offset, result1.id, result1.offset);
+      mem1.id, mem1.offset, result1.id, result1.offset,
+      false);
   expected.set_token1.Init(GetNextToken());
   expected.read2.Init(
       0, kHeight / 2, kWidth, kHeight / 2, kFormat, kType,
-      mem2.id, mem2.offset, result2.id, result2.offset);
+      mem2.id, mem2.offset, result2.id, result2.offset, false);
   expected.set_token2.Init(GetNextToken());
   scoped_ptr<int8[]> buffer(new int8[kWidth * kHeight * kBytesPerPixel]);
 
@@ -1474,7 +1475,7 @@ TEST_F(GLES2ImplementationTest, ReadPixelsBadFormatType) {
   Cmds expected;
   expected.read.Init(
       0, 0, kWidth, kHeight, kFormat, kType,
-      mem1.id, mem1.offset, result1.id, result1.offset);
+      mem1.id, mem1.offset, result1.id, result1.offset, false);
   expected.set_token.Init(GetNextToken());
   scoped_ptr<int8[]> buffer(new int8[kWidth * kHeight * kBytesPerPixel]);
 
diff --git a/gpu/command_buffer/client/query_tracker.cc b/gpu/command_buffer/client/query_tracker.cc
index 76ebf72..7cd1ae8 100644
--- a/gpu/command_buffer/client/query_tracker.cc
+++ b/gpu/command_buffer/client/query_tracker.cc
@@ -113,9 +113,7 @@ void QueryTracker::Query::Begin(GLES2Implementation* gl) {
       gl->helper()->BeginQueryEXT(target(), id(), shm_id(), shm_offset());
       break;
     case GL_ASYNC_PIXEL_TRANSFERS_COMPLETED_CHROMIUM:
-      // tell service about id, shared memory and count
-      gl->helper()->BeginQueryEXT(target(), id(), shm_id(), shm_offset());
-      break;
+    case GL_ASYNC_READ_PIXELS_COMPLETED_CHROMIUM:
     default:
       // tell service about id, shared memory and count
       gl->helper()->BeginQueryEXT(target(), id(), shm_id(), shm_offset());
@@ -165,8 +163,7 @@ bool QueryTracker::Query::CheckResultsAvailable(
                              static_cast<uint64>(0xFFFFFFFFL));
           break;
         case GL_ASYNC_PIXEL_TRANSFERS_COMPLETED_CHROMIUM:
-          result_ = info_.sync->result;
-          break;
+        case GL_ASYNC_READ_PIXELS_COMPLETED_CHROMIUM:
         default:
           result_ = info_.sync->result;
           break;
diff --git a/gpu/command_buffer/common/gles2_cmd_format_autogen.h b/gpu/command_buffer/common/gles2_cmd_format_autogen.h
index 12fc743..504c445 100644
--- a/gpu/command_buffer/common/gles2_cmd_format_autogen.h
+++ b/gpu/command_buffer/common/gles2_cmd_format_autogen.h
@@ -4887,7 +4887,7 @@ struct ReadPixels {
   void Init(
       GLint _x, GLint _y, GLsizei _width, GLsizei _height, GLenum _format,
       GLenum _type, uint32 _pixels_shm_id, uint32 _pixels_shm_offset,
-      uint32 _result_shm_id, uint32 _result_shm_offset) {
+      uint32 _result_shm_id, uint32 _result_shm_offset, GLboolean _async) {
     SetHeader();
     x = _x;
     y = _y;
@@ -4899,17 +4899,18 @@ struct ReadPixels {
     pixels_shm_offset = _pixels_shm_offset;
     result_shm_id = _result_shm_id;
     result_shm_offset = _result_shm_offset;
+    async = _async;
   }
 
   void* Set(
       void* cmd, GLint _x, GLint _y, GLsizei _width, GLsizei _height,
       GLenum _format, GLenum _type, uint32 _pixels_shm_id,
       uint32 _pixels_shm_offset, uint32 _result_shm_id,
-      uint32 _result_shm_offset) {
+      uint32 _result_shm_offset, GLboolean _async) {
     static_cast<ValueType*>(
         cmd)->Init(
             _x, _y, _width, _height, _format, _type, _pixels_shm_id,
-            _pixels_shm_offset, _result_shm_id, _result_shm_offset);
+            _pixels_shm_offset, _result_shm_id, _result_shm_offset, _async);
     return NextCmdAddress<ValueType>(cmd);
   }
 
@@ -4924,10 +4925,11 @@ struct ReadPixels {
   uint32 pixels_shm_offset;
   uint32 result_shm_id;
   uint32 result_shm_offset;
+  uint32 async;
 };
 
-COMPILE_ASSERT(sizeof(ReadPixels) == 44,
-               Sizeof_ReadPixels_is_not_44);
+COMPILE_ASSERT(sizeof(ReadPixels) == 48,
+               Sizeof_ReadPixels_is_not_48);
 COMPILE_ASSERT(offsetof(ReadPixels, header) == 0,
                OffsetOf_ReadPixels_header_not_0);
 COMPILE_ASSERT(offsetof(ReadPixels, x) == 4,
@@ -4950,6 +4952,8 @@ COMPILE_ASSERT(offsetof(ReadPixels, result_shm_id) == 36,
                OffsetOf_ReadPixels_result_shm_id_not_36);
 COMPILE_ASSERT(offsetof(ReadPixels, result_shm_offset) == 40,
                OffsetOf_ReadPixels_result_shm_offset_not_40);
+COMPILE_ASSERT(offsetof(ReadPixels, async) == 44,
+               OffsetOf_ReadPixels_async_not_44);
 
 struct ReleaseShaderCompiler {
   typedef ReleaseShaderCompiler ValueType;
diff --git a/gpu/command_buffer/common/gles2_cmd_format_test_autogen.h b/gpu/command_buffer/common/gles2_cmd_format_test_autogen.h
index b23b569..8bb1ec0 100644
--- a/gpu/command_buffer/common/gles2_cmd_format_test_autogen.h
+++ b/gpu/command_buffer/common/gles2_cmd_format_test_autogen.h
@@ -1803,7 +1803,8 @@ TEST_F(GLES2FormatTest, ReadPixels) {
       static_cast<uint32>(17),
       static_cast<uint32>(18),
       static_cast<uint32>(19),
-      static_cast<uint32>(20));
+      static_cast<uint32>(20),
+      static_cast<GLboolean>(21));
   EXPECT_EQ(static_cast<uint32>(cmds::ReadPixels::kCmdId),
             cmd.header.command);
   EXPECT_EQ(sizeof(cmd), cmd.header.size * 4u);
@@ -1817,6 +1818,7 @@ TEST_F(GLES2FormatTest, ReadPixels) {
   EXPECT_EQ(static_cast<uint32>(18), cmd.pixels_shm_offset);
   EXPECT_EQ(static_cast<uint32>(19), cmd.result_shm_id);
   EXPECT_EQ(static_cast<uint32>(20), cmd.result_shm_offset);
+  EXPECT_EQ(static_cast<GLboolean>(21), cmd.async);
   CheckBytesWrittenMatchesExpectedSize(
       next_cmd, sizeof(cmd));
 }
diff --git a/gpu/command_buffer/common/gles2_cmd_utils_implementation_autogen.h b/gpu/command_buffer/common/gles2_cmd_utils_implementation_autogen.h
index 36975dd..3c3fb33 100644
--- a/gpu/command_buffer/common/gles2_cmd_utils_implementation_autogen.h
+++ b/gpu/command_buffer/common/gles2_cmd_utils_implementation_autogen.h
@@ -86,6 +86,7 @@ static GLES2Util::EnumToString enum_to_string_table[] = {
   { 0x2601, "GL_LINEAR", },
   { 0x8C03, "GL_COMPRESSED_RGBA_PVRTC_2BPPV1_IMG", },
   { 0x9242, "GL_UNPACK_UNPREMULTIPLY_ALPHA_CHROMIUM", },
+  { 0x88BA, "GL_READ_WRITE", },
   { 0x88BB, "GL_BUFFER_ACCESS_OES", },
   { 0x88BC, "GL_BUFFER_MAPPED_OES", },
   { 0x88BD, "GL_BUFFER_MAP_POINTER_OES", },
@@ -227,7 +228,7 @@ static GLES2Util::EnumToString enum_to_string_table[] = {
   { 0x84F5, "GL_ASYNC_PIXEL_TRANSFERS_COMPLETED_CHROMIUM", },
   { 0x882A, "GL_DRAW_BUFFER5_NV", },
   { 0x80AA, "GL_SAMPLE_COVERAGE_VALUE", },
-  { 0x84F6, "GL_TEXTURE_BINDING_RECTANGLE_ARB", },
+  { 0x84F6, "GL_ASYNC_READ_PIXELS_COMPLETED_CHROMIUM", },
   { 0x80AB, "GL_SAMPLE_COVERAGE_INVERT", },
   { 0x8FC4, "GL_SHADER_BINARY_VIV", },
   { 0x882B, "GL_DRAW_BUFFER6_NV", },
@@ -1079,6 +1080,8 @@ std::string GLES2Util::GetStringQueryTarget(uint32 value) {
     { GL_LATENCY_QUERY_CHROMIUM, "GL_LATENCY_QUERY_CHROMIUM" },
     { GL_ASYNC_PIXEL_TRANSFERS_COMPLETED_CHROMIUM,
     "GL_ASYNC_PIXEL_TRANSFERS_COMPLETED_CHROMIUM" },
+    { GL_ASYNC_READ_PIXELS_COMPLETED_CHROMIUM,
+    "GL_ASYNC_READ_PIXELS_COMPLETED_CHROMIUM" },
   };
   return GLES2Util::GetQualifiedEnumString(
       string_table, arraysize(string_table), value);
diff --git a/gpu/command_buffer/service/feature_info.cc b/gpu/command_buffer/service/feature_info.cc
index 5d95848..396311c 100644
--- a/gpu/command_buffer/service/feature_info.cc
+++ b/gpu/command_buffer/service/feature_info.cc
@@ -117,7 +117,8 @@ FeatureInfo::FeatureFlags::FeatureFlags()
       enable_shader_name_hashing(false),
       enable_samplers(false),
       ext_draw_buffers(false),
-      ext_frag_depth(false) {
+      ext_frag_depth(false),
+      use_async_readpixels(false) {
 }
 
 FeatureInfo::Workarounds::Workarounds() :
@@ -623,6 +624,16 @@ void FeatureInfo::AddFeatures(const CommandLine& command_line) {
     feature_flags_.ext_frag_depth = true;
   }
 
+  bool ui_gl_fence_works =
+      extensions.Contains("GL_NV_fence") ||
+      extensions.Contains("GL_ARB_sync");
+
+  if (ui_gl_fence_works &&
+      extensions.Contains("GL_ARB_pixel_buffer_object") &&
+      !workarounds_.disable_async_readpixels) {
+    feature_flags_.use_async_readpixels = true;
+  }
+
   if (!disallowed_features_.swap_buffer_complete_callback)
     AddExtensionString("GL_CHROMIUM_swapbuffers_complete_callback");
 
diff --git a/gpu/command_buffer/service/feature_info.h b/gpu/command_buffer/service/feature_info.h
index 3754a3c..87abfba 100644
--- a/gpu/command_buffer/service/feature_info.h
+++ b/gpu/command_buffer/service/feature_info.h
@@ -45,6 +45,7 @@ class GPU_EXPORT FeatureInfo : public base::RefCounted<FeatureInfo> {
     bool enable_samplers;
     bool ext_draw_buffers;
     bool ext_frag_depth;
+    bool use_async_readpixels;
   };
 
   struct Workarounds {
diff --git a/gpu/command_buffer/service/gles2_cmd_decoder.cc b/gpu/command_buffer/service/gles2_cmd_decoder.cc
index 6ed619c..70837bf 100644
--- a/gpu/command_buffer/service/gles2_cmd_decoder.cc
+++ b/gpu/command_buffer/service/gles2_cmd_decoder.cc
@@ -59,6 +59,7 @@
 #include "gpu/command_buffer/service/vertex_array_manager.h"
 #include "gpu/command_buffer/service/vertex_attrib_manager.h"
 #include "ui/gl/gl_bindings.h"
+#include "ui/gl/gl_fence.h"
 #include "ui/gl/gl_image.h"
 #include "ui/gl/gl_implementation.h"
 #include "ui/gl/gl_surface.h"
@@ -477,6 +478,19 @@ class BackFramebuffer {
   DISALLOW_COPY_AND_ASSIGN(BackFramebuffer);
 };
 
+struct FenceCallback {
+  explicit FenceCallback()
+      : fence(gfx::GLFence::Create()) {
+    DCHECK(fence);
+  }
+  void AddCallback(base::Closure cb) {
+    callbacks.push_back(cb);
+  }
+  std::vector<base::Closure> callbacks;
+  scoped_ptr<gfx::GLFence> fence;
+};
+
+
 // }  // anonymous namespace.
 
 bool GLES2Decoder::GetServiceTextureId(uint32 client_texture_id,
@@ -587,6 +601,8 @@ class GLES2DecoderImpl : public GLES2Decoder {
   virtual bool HasMoreIdleWork() OVERRIDE;
   virtual void PerformIdleWork() OVERRIDE;
 
+  virtual void WaitForReadPixels(base::Closure callback) OVERRIDE;
+
   virtual void SetResizeCallback(
       const base::Callback<void(gfx::Size, float)>& callback) OVERRIDE;
 
@@ -1559,6 +1575,9 @@ class GLES2DecoderImpl : public GLES2Decoder {
            surface_->DeferDraws();
   }
 
+  void ProcessPendingReadPixels();
+  void FinishReadPixels(const cmds::ReadPixels& c, GLuint buffer);
+
   void ForceCompileShaderIfPending(Shader* shader);
 
   // Generate a member function prototype for each command in an automated and
@@ -1726,6 +1745,8 @@ class GLES2DecoderImpl : public GLES2Decoder {
 
   scoped_ptr<GPUTracer> gpu_tracer_;
 
+  std::queue<linked_ptr<FenceCallback> > pending_readpixel_fences_;
+
   DISALLOW_COPY_AND_ASSIGN(GLES2DecoderImpl);
 };
 
@@ -2805,6 +2826,7 @@ bool GLES2DecoderImpl::MakeCurrent() {
 }
 
 void GLES2DecoderImpl::ProcessFinishedAsyncTransfers() {
+  ProcessPendingReadPixels();
   if (engine() && query_manager_.get())
     query_manager_->ProcessPendingTransferQueries();
 
@@ -3533,6 +3555,7 @@ bool GLES2DecoderImpl::CreateShaderHelper(GLenum type, GLuint client_id) {
 
 void GLES2DecoderImpl::DoFinish() {
   glFinish();
+  ProcessPendingReadPixels();
   ProcessPendingQueries();
 }
 
@@ -6695,6 +6718,96 @@ error::Error GLES2DecoderImpl::HandleVertexAttribDivisorANGLE(
   return error::kNoError;
 }
 
+void GLES2DecoderImpl::FinishReadPixels(
+    const cmds::ReadPixels& c,
+    GLuint buffer) {
+  TRACE_EVENT0("gpu", "GLES2DecoderImpl::FinishReadPixels");
+  GLsizei width = c.width;
+  GLsizei height = c.height;
+  GLenum format = c.format;
+  GLenum type = c.type;
+  typedef cmds::ReadPixels::Result Result;
+  uint32 pixels_size;
+  Result* result = NULL;
+  if (c.result_shm_id != 0) {
+    result = GetSharedMemoryAs<Result*>(
+        c.result_shm_id, c.result_shm_offset, sizeof(*result));
+    if (!result) {
+      if (buffer != 0) {
+        glDeleteBuffersARB(1, &buffer);
+      }
+      return;
+    }
+  }
+  GLES2Util::ComputeImageDataSizes(
+      width, height, format, type, state_.pack_alignment, &pixels_size,
+      NULL, NULL);
+  void* pixels = GetSharedMemoryAs<void*>(
+      c.pixels_shm_id, c.pixels_shm_offset, pixels_size);
+  if (!pixels) {
+    if (buffer != 0) {
+      glDeleteBuffersARB(1, &buffer);
+    }
+    return;
+  }
+
+  if (buffer != 0) {
+    glBindBuffer(GL_PIXEL_PACK_BUFFER_ARB, buffer);
+    void* data = glMapBuffer(GL_PIXEL_PACK_BUFFER_ARB, GL_READ_ONLY);
+    memcpy(pixels, data, pixels_size);
+    // GL_PIXEL_PACK_BUFFER_ARB is currently unused, so we don't
+    // have to restore the state.
+    glUnmapBuffer(GL_PIXEL_PACK_BUFFER_ARB);
+    glBindBuffer(GL_PIXEL_PACK_BUFFER_ARB, 0);
+    glDeleteBuffersARB(1, &buffer);
+  }
+
+  if (result != NULL) {
+    *result = true;
+  }
+
+  GLenum read_format = GetBoundReadFrameBufferInternalFormat();
+  uint32 channels_exist = GLES2Util::GetChannelsForFormat(read_format);
+  if ((channels_exist & 0x0008) == 0 &&
+      workarounds().clear_alpha_in_readpixels) {
+    // Set the alpha to 255 because some drivers are buggy in this regard.
+    uint32 temp_size;
+
+    uint32 unpadded_row_size;
+    uint32 padded_row_size;
+    if (!GLES2Util::ComputeImageDataSizes(
+            width, 2, format, type, state_.pack_alignment, &temp_size,
+            &unpadded_row_size, &padded_row_size)) {
+      return;
+    }
+    // NOTE: Assumes the type is GL_UNSIGNED_BYTE which was true at the time
+    // of this implementation.
+    if (type != GL_UNSIGNED_BYTE) {
+      return;
+    }
+    switch (format) {
+      case GL_RGBA:
+      case GL_BGRA_EXT:
+      case GL_ALPHA: {
+        int offset = (format == GL_ALPHA) ? 0 : 3;
+        int step = (format == GL_ALPHA) ? 1 : 4;
+        uint8* dst = static_cast<uint8*>(pixels) + offset;
+        for (GLint yy = 0; yy < height; ++yy) {
+          uint8* end = dst + unpadded_row_size;
+          for (uint8* d = dst; d < end; d += step) {
+            *d = 255;
+          }
+          dst += padded_row_size;
+        }
+        break;
+      }
+      default:
+        break;
+    }
+  }
+}
+
+
 error::Error GLES2DecoderImpl::HandleReadPixels(
     uint32 immediate_data_size, const cmds::ReadPixels& c) {
   if (ShouldDeferReads())
@@ -6705,6 +6818,7 @@ error::Error GLES2DecoderImpl::HandleReadPixels(
   GLsizei height = c.height;
   GLenum format = c.format;
   GLenum type = c.type;
+  GLboolean async = c.async;
   if (width < 0 || height < 0) {
     LOCAL_SET_GL_ERROR(GL_INVALID_VALUE, "glReadPixels", "dimensions < 0");
     return error::kNoError;
@@ -6804,6 +6918,25 @@ error::Error GLES2DecoderImpl::HandleReadPixels(
       dst += padded_row_size;
     }
   } else {
+    if (async && features().use_async_readpixels) {
+      GLuint buffer;
+      glGenBuffersARB(1, &buffer);
+      glBindBuffer(GL_PIXEL_PACK_BUFFER_ARB, buffer);
+      glBufferData(GL_PIXEL_PACK_BUFFER_ARB, pixels_size, NULL, GL_STREAM_READ);
+      GLenum error = glGetError();
+      if (error == GL_NO_ERROR) {
+        glReadPixels(x, y, width, height, format, type, 0);
+        pending_readpixel_fences_.push(linked_ptr<FenceCallback>(
+            new FenceCallback()));
+        WaitForReadPixels(base::Bind(
+            &GLES2DecoderImpl::FinishReadPixels,
+            base::internal::SupportsWeakPtrBase::StaticAsWeakPtr
+            <GLES2DecoderImpl>(this),
+            c, buffer));
+        glBindBuffer(GL_PIXEL_PACK_BUFFER_ARB, 0);
+        return error::kNoError;
+      }
+    }
     glReadPixels(x, y, width, height, format, type, pixels);
   }
   GLenum error = LOCAL_PEEK_GL_ERROR("glReadPixels");
@@ -6811,51 +6944,7 @@ error::Error GLES2DecoderImpl::HandleReadPixels(
     if (result != NULL) {
       *result = true;
     }
-
-    GLenum read_format = GetBoundReadFrameBufferInternalFormat();
-    uint32 channels_exist = GLES2Util::GetChannelsForFormat(read_format);
-    if ((channels_exist & 0x0008) == 0 &&
-        workarounds().clear_alpha_in_readpixels) {
-      // Set the alpha to 255 because some drivers are buggy in this regard.
-      uint32 temp_size;
-
-      uint32 unpadded_row_size;
-      uint32 padded_row_size;
-      if (!GLES2Util::ComputeImageDataSizes(
-          width, 2, format, type, state_.pack_alignment, &temp_size,
-          &unpadded_row_size, &padded_row_size)) {
-        LOCAL_SET_GL_ERROR(
-            GL_INVALID_VALUE, "glReadPixels", "dimensions out of range");
-        return error::kNoError;
-      }
-      // NOTE: Assumes the type is GL_UNSIGNED_BYTE which was true at the time
-      // of this implementation.
-      if (type != GL_UNSIGNED_BYTE) {
-        LOCAL_SET_GL_ERROR(
-            GL_INVALID_OPERATION, "glReadPixels",
-            "unsupported readPixel format");
-        return error::kNoError;
-      }
-      switch (format) {
-        case GL_RGBA:
-        case GL_BGRA_EXT:
-        case GL_ALPHA: {
-          int offset = (format == GL_ALPHA) ? 0 : 3;
-          int step = (format == GL_ALPHA) ? 1 : 4;
-          uint8* dst = static_cast<uint8*>(pixels) + offset;
-          for (GLint yy = 0; yy < height; ++yy) {
-            uint8* end = dst + unpadded_row_size;
-            for (uint8* d = dst; d < end; d += step) {
-              *d = 255;
-            }
-            dst += padded_row_size;
-          }
-          break;
-        }
-        default:
-          break;
-      }
-    }
+    FinishReadPixels(c, 0);
   }
 
   return error::kNoError;
@@ -9085,11 +9174,35 @@ bool GLES2DecoderImpl::ProcessPendingQueries() {
   return query_manager_->HavePendingQueries();
 }
 
+// Note that if there are no pending readpixels right now,
+// this function will call the callback immediately.
+void GLES2DecoderImpl::WaitForReadPixels(base::Closure callback) {
+  if (features().use_async_readpixels && !pending_readpixel_fences_.empty()) {
+    pending_readpixel_fences_.back()->callbacks.push_back(callback);
+  } else {
+    callback.Run();
+  }
+}
+
+void GLES2DecoderImpl::ProcessPendingReadPixels() {
+  while (!pending_readpixel_fences_.empty() &&
+         pending_readpixel_fences_.front()->fence->HasCompleted()) {
+    std::vector<base::Closure> callbacks =
+        pending_readpixel_fences_.front()->callbacks;
+    pending_readpixel_fences_.pop();
+    for (size_t i = 0; i < callbacks.size(); i++) {
+      callbacks[i].Run();
+    }
+  }
+}
+
 bool GLES2DecoderImpl::HasMoreIdleWork() {
-  return async_pixel_transfer_manager_->NeedsProcessMorePendingTransfers();
+  return !pending_readpixel_fences_.empty() ||
+      async_pixel_transfer_manager_->NeedsProcessMorePendingTransfers();
 }
 
 void GLES2DecoderImpl::PerformIdleWork() {
+  ProcessPendingReadPixels();
   if (!async_pixel_transfer_manager_->NeedsProcessMorePendingTransfers())
     return;
   async_pixel_transfer_manager_->ProcessMorePendingTransfers();
@@ -9107,6 +9220,7 @@ error::Error GLES2DecoderImpl::HandleBeginQueryEXT(
     case GL_COMMANDS_ISSUED_CHROMIUM:
     case GL_LATENCY_QUERY_CHROMIUM:
     case GL_ASYNC_PIXEL_TRANSFERS_COMPLETED_CHROMIUM:
+    case GL_ASYNC_READ_PIXELS_COMPLETED_CHROMIUM:
     case GL_GET_ERROR_QUERY_CHROMIUM:
       break;
     default:
@@ -9119,6 +9233,8 @@ error::Error GLES2DecoderImpl::HandleBeginQueryEXT(
       break;
   }
 
+  // TODO(hubbe): Make it possible to have one query per type running at the
+  // same time.
   if (state_.current_query.get()) {
     LOCAL_SET_GL_ERROR(
         GL_INVALID_OPERATION, "glBeginQueryEXT", "query already in progress");
diff --git a/gpu/command_buffer/service/gles2_cmd_decoder.h b/gpu/command_buffer/service/gles2_cmd_decoder.h
index 97c8f58..76d61ea 100644
--- a/gpu/command_buffer/service/gles2_cmd_decoder.h
+++ b/gpu/command_buffer/service/gles2_cmd_decoder.h
@@ -213,6 +213,7 @@ class GPU_EXPORT GLES2Decoder : public base::SupportsWeakPtr<GLES2Decoder>,
   virtual void SetWaitSyncPointCallback(
       const WaitSyncPointCallback& callback) = 0;
 
+  virtual void WaitForReadPixels(base::Closure callback) = 0;
   virtual uint32 GetTextureUploadCount() = 0;
   virtual base::TimeDelta GetTotalTextureUploadTime() = 0;
   virtual base::TimeDelta GetTotalProcessingCommandsTime() = 0;
diff --git a/gpu/command_buffer/service/gles2_cmd_decoder_mock.h b/gpu/command_buffer/service/gles2_cmd_decoder_mock.h
index abac629..5050518 100644
--- a/gpu/command_buffer/service/gles2_cmd_decoder_mock.h
+++ b/gpu/command_buffer/service/gles2_cmd_decoder_mock.h
@@ -102,6 +102,8 @@ class MockGLES2Decoder : public GLES2Decoder {
                void(const ShaderCacheCallback& callback));
   MOCK_METHOD1(SetWaitSyncPointCallback,
                void(const WaitSyncPointCallback& callback));
+  MOCK_METHOD1(WaitForReadPixels,
+               void(base::Closure callback));
   MOCK_METHOD0(GetTextureUploadCount, uint32());
   MOCK_METHOD0(GetTotalTextureUploadTime, base::TimeDelta());
   MOCK_METHOD0(GetTotalProcessingCommandsTime, base::TimeDelta());
diff --git a/gpu/command_buffer/service/gles2_cmd_decoder_unittest.cc b/gpu/command_buffer/service/gles2_cmd_decoder_unittest.cc
index 09ed184..b426286 100644
--- a/gpu/command_buffer/service/gles2_cmd_decoder_unittest.cc
+++ b/gpu/command_buffer/service/gles2_cmd_decoder_unittest.cc
@@ -2316,7 +2316,8 @@ void GLES2DecoderTest::CheckReadPixelsOutOfRange(
   cmd.Init(in_read_x, in_read_y, in_read_width, in_read_height,
            kFormat, GL_UNSIGNED_BYTE,
            pixels_shm_id, pixels_shm_offset,
-           result_shm_id, result_shm_offset);
+           result_shm_id, result_shm_offset,
+           false);
   EXPECT_EQ(error::kNoError, ExecuteCmd(cmd));
 
   GLint unpadded_row_size = emu.ComputeImageDataSize(in_read_width, 1);
@@ -2393,7 +2394,8 @@ TEST_F(GLES2DecoderTest, ReadPixels) {
   ReadPixels cmd;
   cmd.Init(0, 0, kWidth, kHeight, GL_RGB, GL_UNSIGNED_BYTE,
            pixels_shm_id, pixels_shm_offset,
-           result_shm_id, result_shm_offset);
+           result_shm_id, result_shm_offset,
+           false);
   EXPECT_EQ(error::kNoError, ExecuteCmd(cmd));
   for (GLint yy = 0; yy < kHeight; ++yy) {
     EXPECT_TRUE(emu.CompareRowSegment(
@@ -2440,7 +2442,8 @@ TEST_F(GLES2DecoderRGBBackbufferTest, ReadPixelsNoAlphaBackbuffer) {
   ReadPixels cmd;
   cmd.Init(0, 0, kWidth, kHeight, GL_RGBA, GL_UNSIGNED_BYTE,
            pixels_shm_id, pixels_shm_offset,
-           result_shm_id, result_shm_offset);
+           result_shm_id, result_shm_offset,
+           false);
   EXPECT_EQ(error::kNoError, ExecuteCmd(cmd));
   for (GLint yy = 0; yy < kHeight; ++yy) {
     EXPECT_TRUE(emu.CompareRowSegment(
@@ -2477,34 +2480,41 @@ TEST_F(GLES2DecoderTest, ReadPixelsInvalidArgs) {
   ReadPixels cmd;
   cmd.Init(0, 0, -1, 1, GL_RGB, GL_UNSIGNED_BYTE,
            pixels_shm_id, pixels_shm_offset,
-           result_shm_id, result_shm_offset);
+           result_shm_id, result_shm_offset,
+           false);
   EXPECT_EQ(error::kNoError, ExecuteCmd(cmd));
   EXPECT_EQ(GL_INVALID_VALUE, GetGLError());
   cmd.Init(0, 0, 1, -1, GL_RGB, GL_UNSIGNED_BYTE,
            pixels_shm_id, pixels_shm_offset,
-           result_shm_id, result_shm_offset);
+           result_shm_id, result_shm_offset,
+           false);
   EXPECT_EQ(error::kNoError, ExecuteCmd(cmd));
   EXPECT_EQ(GL_INVALID_VALUE, GetGLError());
   cmd.Init(0, 0, 1, 1, GL_RGB, GL_INT,
            pixels_shm_id, pixels_shm_offset,
-           result_shm_id, result_shm_offset);
+           result_shm_id, result_shm_offset,
+           false);
   EXPECT_EQ(error::kNoError, ExecuteCmd(cmd));
   EXPECT_EQ(GL_INVALID_ENUM, GetGLError());
   cmd.Init(0, 0, 1, 1, GL_RGB, GL_UNSIGNED_BYTE,
            kInvalidSharedMemoryId, pixels_shm_offset,
-           result_shm_id, result_shm_offset);
+           result_shm_id, result_shm_offset,
+           false);
   EXPECT_NE(error::kNoError, ExecuteCmd(cmd));
   cmd.Init(0, 0, 1, 1, GL_RGB, GL_UNSIGNED_BYTE,
            pixels_shm_id, kInvalidSharedMemoryOffset,
-           result_shm_id, result_shm_offset);
+           result_shm_id, result_shm_offset,
+           false);
   EXPECT_NE(error::kNoError, ExecuteCmd(cmd));
   cmd.Init(0, 0, 1, 1, GL_RGB, GL_UNSIGNED_BYTE,
            pixels_shm_id, pixels_shm_offset,
-           kInvalidSharedMemoryId, result_shm_offset);
+           kInvalidSharedMemoryId, result_shm_offset,
+           false);
   EXPECT_NE(error::kNoError, ExecuteCmd(cmd));
   cmd.Init(0, 0, 1, 1, GL_RGB, GL_UNSIGNED_BYTE,
            pixels_shm_id, pixels_shm_offset,
-           result_shm_id, kInvalidSharedMemoryOffset);
+           result_shm_id, kInvalidSharedMemoryOffset,
+           false);
   EXPECT_NE(error::kNoError, ExecuteCmd(cmd));
 }
 
@@ -4932,7 +4942,8 @@ TEST_F(GLES2DecoderTest, ReadPixelsGLError) {
   ReadPixels cmd;
   cmd.Init(x, y, width, height, kFormat, GL_UNSIGNED_BYTE,
            pixels_shm_id, pixels_shm_offset,
-           result_shm_id, result_shm_offset);
+           result_shm_id, result_shm_offset,
+           false);
   EXPECT_EQ(error::kNoError, ExecuteCmd(cmd));
   EXPECT_EQ(GL_OUT_OF_MEMORY, GetGLError());
 }
@@ -6664,8 +6675,9 @@ TEST_F(GLES2DecoderWithShaderTest, UnClearedAttachmentsGetClearedOnReadPixels) {
   uint32 pixels_shm_offset = kSharedMemoryOffset + sizeof(*result);
   ReadPixels cmd;
   cmd.Init(0, 0, 1, 1, GL_RGBA, GL_UNSIGNED_BYTE,
-         pixels_shm_id, pixels_shm_offset,
-         result_shm_id, result_shm_offset);
+           pixels_shm_id, pixels_shm_offset,
+           result_shm_id, result_shm_offset,
+           false);
   EXPECT_EQ(error::kNoError, ExecuteCmd(cmd));
   EXPECT_EQ(GL_NO_ERROR, GetGLError());
 }
@@ -6724,8 +6736,9 @@ TEST_F(GLES2DecoderManualInitTest,
   uint32 pixels_shm_offset = kSharedMemoryOffset + sizeof(Result);
   ReadPixels cmd;
   cmd.Init(0, 0, 1, 1, GL_RGBA, GL_UNSIGNED_BYTE,
-         pixels_shm_id, pixels_shm_offset,
-         result_shm_id, result_shm_offset);
+           pixels_shm_id, pixels_shm_offset,
+           result_shm_id, result_shm_offset,
+           false);
   EXPECT_EQ(error::kNoError, ExecuteCmd(cmd));
   EXPECT_EQ(GL_NO_ERROR, GetGLError());
 }
diff --git a/gpu/command_buffer/service/gles2_cmd_validation_implementation_autogen.h b/gpu/command_buffer/service/gles2_cmd_validation_implementation_autogen.h
index ebe7e07..db9725d 100644
--- a/gpu/command_buffer/service/gles2_cmd_validation_implementation_autogen.h
+++ b/gpu/command_buffer/service/gles2_cmd_validation_implementation_autogen.h
@@ -300,6 +300,7 @@ static GLenum valid_query_target_table[] = {
   GL_COMMANDS_ISSUED_CHROMIUM,
   GL_LATENCY_QUERY_CHROMIUM,
   GL_ASYNC_PIXEL_TRANSFERS_COMPLETED_CHROMIUM,
+  GL_ASYNC_READ_PIXELS_COMPLETED_CHROMIUM,
 };
 
 static GLenum valid_read_pixel_format_table[] = {
diff --git a/gpu/command_buffer/service/query_manager.cc b/gpu/command_buffer/service/query_manager.cc
index e66478a..6d2f84a 100644
--- a/gpu/command_buffer/service/query_manager.cc
+++ b/gpu/command_buffer/service/query_manager.cc
@@ -223,9 +223,8 @@ bool AsyncPixelTransfersCompletedQuery::End(uint32 submit_count) {
   // on the current thread.
   manager()->decoder()->GetAsyncPixelTransferManager()->AsyncNotifyCompletion(
       mem_params,
-      base::Bind(AsyncPixelTransfersCompletedQuery::MarkAsCompletedThreadSafe,
+      base::Bind(&AsyncPixelTransfersCompletedQuery::MarkAsCompletedThreadSafe,
                  submit_count));
-
   return AddToPendingTransferQueue(submit_count);
 }
 
@@ -254,6 +253,58 @@ void AsyncPixelTransfersCompletedQuery::Destroy(bool /* have_context */) {
 AsyncPixelTransfersCompletedQuery::~AsyncPixelTransfersCompletedQuery() {
 }
 
+class AsyncReadPixelsCompletedQuery
+    : public QueryManager::Query,
+      public base::SupportsWeakPtr<AsyncReadPixelsCompletedQuery> {
+ public:
+  AsyncReadPixelsCompletedQuery(
+      QueryManager* manager, GLenum target, int32 shm_id, uint32 shm_offset);
+
+  virtual bool Begin() OVERRIDE;
+  virtual bool End(uint32 submit_count) OVERRIDE;
+  virtual bool Process() OVERRIDE;
+  virtual void Destroy(bool have_context) OVERRIDE;
+
+ protected:
+  void Complete();
+  virtual ~AsyncReadPixelsCompletedQuery();
+};
+
+AsyncReadPixelsCompletedQuery::AsyncReadPixelsCompletedQuery(
+    QueryManager* manager, GLenum target, int32 shm_id, uint32 shm_offset)
+    : Query(manager, target, shm_id, shm_offset) {
+}
+
+bool AsyncReadPixelsCompletedQuery::Begin() {
+  return true;
+}
+
+bool AsyncReadPixelsCompletedQuery::End(uint32 submit_count) {
+  manager()->decoder()->WaitForReadPixels(
+      base::Bind(&AsyncReadPixelsCompletedQuery::Complete,
+                 AsWeakPtr()));
+
+  return AddToPendingTransferQueue(submit_count);
+}
+
+void AsyncReadPixelsCompletedQuery::Complete() {
+  MarkAsCompleted(1);
+}
+
+bool AsyncReadPixelsCompletedQuery::Process() {
+  return !pending();
+}
+
+void AsyncReadPixelsCompletedQuery::Destroy(bool /* have_context */) {
+  if (!IsDeleted()) {
+    MarkAsDeleted();
+  }
+}
+
+AsyncReadPixelsCompletedQuery::~AsyncReadPixelsCompletedQuery() {
+}
+
+
 class GetErrorQuery : public QueryManager::Query {
  public:
   GetErrorQuery(
@@ -345,6 +396,10 @@ QueryManager::Query* QueryManager::CreateQuery(
       query = new AsyncPixelTransfersCompletedQuery(
           this, target, shm_id, shm_offset);
       break;
+    case GL_ASYNC_READ_PIXELS_COMPLETED_CHROMIUM:
+      query = new AsyncReadPixelsCompletedQuery(
+          this, target, shm_id, shm_offset);
+      break;
     case GL_GET_ERROR_QUERY_CHROMIUM:
       query = new GetErrorQuery(this, target, shm_id, shm_offset);
       break;