Re-land^2 WebVR compositor bypass via BrowserMain context + mailbox

chrome/browser/android/vr_shell/vr_shell_gl.cc

Index: chrome/browser/android/vr_shell/vr_shell_gl.cc
diff --git a/chrome/browser/android/vr_shell/vr_shell_gl.cc b/chrome/browser/android/vr_shell/vr_shell_gl.cc
index 336d2b0472f4ce8c03bafe98d7629b20023598a8..389fa4b536d5b92faf7df3a61f16f6eea7238065 100644
--- a/chrome/browser/android/vr_shell/vr_shell_gl.cc
+++ b/chrome/browser/android/vr_shell/vr_shell_gl.cc
@@ -18,6 +18,7 @@
 #include "chrome/browser/android/vr_shell/vr_gl_util.h"
 #include "chrome/browser/android/vr_shell/vr_math.h"
 #include "chrome/browser/android/vr_shell/vr_shell.h"
+#include "chrome/browser/android/vr_shell/vr_shell_command_buffer_gl.h"
 #include "chrome/browser/android/vr_shell/vr_shell_delegate.h"
 #include "chrome/browser/android/vr_shell/vr_shell_renderer.h"
 #include "device/vr/android/gvr/gvr_device.h"
@@ -83,9 +84,12 @@ static constexpr int kViewportListHeadlockedOffset = 2;
 // 2-3 frames.
 static constexpr unsigned kPoseRingBufferSize = 8;
 
-// Magic numbers used to mark valid pose index values encoded in frame
-// data. Must match the magic numbers used in blink's VRDisplay.cpp.
-static constexpr std::array<uint8_t, 2> kWebVrPosePixelMagicNumbers{{42, 142}};
+// Default downscale percentage for computing the recommended WebVR
+// renderWidth/Height from the 1:1 pixel mapped size. Using a rather
+// aggressive downscale due to the high overhead of copying pixels
+// twice before handing off to GVR. For comparison, the polyfill
+// uses approximately 55% on a Pixel XL.
+static constexpr unsigned kWebVrRecommendedResolutionPercent = 50;

[bajones, 2017/03/07 00:48:07]

Nit: I'd recommend making this float kWebVrRecommendedResolutionScale = 0.5,
since it's still pretty readable and avoids the / 100 below.

[klausw, 2017/03/07 02:55:55]

Done.

 
 float Distance(const gvr::Vec3f& vec1, const gvr::Vec3f& vec2) {
   float xdiff = (vec1.x - vec2.x);
@@ -235,24 +239,50 @@ void VrShellGl::InitializeGl(gfx::AcceleratedWidget window) {
     return;
   }
 
-  unsigned int textures[2];
-  glGenTextures(2, textures);
+  unsigned int textures[3];
+  glGenTextures(3, textures);
   ui_texture_id_ = textures[0];
   content_texture_id_ = textures[1];
+  webvr_texture_id_ = textures[2];
   ui_surface_texture_ = gl::SurfaceTexture::Create(ui_texture_id_);
   content_surface_texture_ = gl::SurfaceTexture::Create(content_texture_id_);
+  webvr_surface_texture_ = gl::SurfaceTexture::Create(webvr_texture_id_);
   CreateUiSurface();
   CreateContentSurface();
+  // Don't create the WebVR surface yet, wait until we need it.

[bajones, 2017/03/07 00:48:07]

Comment seems to imply WebVR surface will be created on demand at another time,
but we're calling CreateOrResizeWebVRSurface() 30 lines down in the same
function?

[klausw, 2017/03/07 02:55:55]

Changed to: WebVR surface is created below.

   ui_surface_texture_->SetFrameAvailableCallback(base::Bind(
       &VrShellGl::OnUIFrameAvailable, weak_ptr_factory_.GetWeakPtr()));
   content_surface_texture_->SetFrameAvailableCallback(base::Bind(
       &VrShellGl::OnContentFrameAvailable, weak_ptr_factory_.GetWeakPtr()));
-  content_surface_texture_->SetDefaultBufferSize(
-      content_tex_physical_size_.width, content_tex_physical_size_.height);
+  webvr_surface_texture_->SetFrameAvailableCallback(base::Bind(
+      &VrShellGl::OnWebVRFrameAvailable, weak_ptr_factory_.GetWeakPtr()));
   ui_surface_texture_->SetDefaultBufferSize(ui_tex_physical_size_.width,
                                             ui_tex_physical_size_.height);
+  content_surface_texture_->SetDefaultBufferSize(
+      content_tex_physical_size_.width, content_tex_physical_size_.height);
   InitializeRenderer();
 
+  // Pick a size for the WebVR transfer surface based on a downscaled
+  // recommended render resolution, and also use that size as the
+  // client-recommended renderWidth/renderHeight and for the GVR
+  // framebuffer. If the client chooses a different size or resizes
+  // it while presenting, we'll resize the transfer surface and GVR
+  // framebuffer to match.
+
+  auto render_target_size = gvr_api_->GetMaximumEffectiveRenderTargetSize();
+
+  // There are some rendering artifacts in the right eye where straight lines
+  // have small dents in them. Round size to multiple of 16 pixel per eye
+  // to compensate?
+  gvr::Sizei size;
+  size.width =
+      render_target_size.width * kWebVrRecommendedResolutionPercent / 100;
+  size.height =
+      render_target_size.height * kWebVrRecommendedResolutionPercent / 100;
+  size.width &= ~0x1f;

[bajones, 2017/03/07 00:48:07]

Is this supposed to be 0xf instead of 0x1f? Also, comment above ends with a "?"
as though it's something to be done in a future CL, but you're doing the
rounding anyway right here. Maybe replace the "?" with a "."

[klausw, 2017/03/07 02:55:55]

I've removed the rounding and replaced with a TODO comment to try this. It
didn't seem to help with the artifacts. FWIW, 0x1f was intentional since I
wanted the overall width to be a multiple of 32 so that each eye is a multiple
of 16.

+  size.height &= ~0xf;
+  CreateOrResizeWebVRSurface(size);
+
   vsync_task_.Reset(base::Bind(&VrShellGl::OnVSync, base::Unretained(this)));
   OnVSync();
 
@@ -275,6 +305,57 @@ void VrShellGl::CreateUiSurface() {
                             ui_surface_->j_surface().obj()));
 }
 
+void VrShellGl::CreateOrResizeWebVRSurface(const gvr::Sizei& size) {
+  if (!webvr_surface_texture_) {
+    LOG(ERROR) << "No WebVR surface texture available";
+    return;
+  }
+
+  // ContentPhysicalBoundsChanged is getting called twice with
+  // identical sizes? Avoid thrashing the existing context.
+  if (size == webvr_surface_size_) {
+    return;
+  }
+
+  if (!size.width || !size.height) {
+    // Invalid size, defer until a new size arrives on a future bounds update.
+    return;
+  }
+
+  webvr_surface_texture_->SetDefaultBufferSize(size.width, size.height);
+  webvr_surface_size_ = size;
+
+  if (command_buffer_gl_) {
+    command_buffer_gl_->ResizeSurface(size.width, size.height);
+  } else {
+    command_buffer_gl_ = base::MakeUnique<VrShellCommandBufferGl>();
+    webvr_surface_ = command_buffer_gl_->CreateSurface(webvr_surface_texture_);
+  }
+}
+
+void VrShellGl::SubmitWebVRFrame(int16_t frame_index,
+                                 gpu::Mailbox mailbox,
+                                 gpu::SyncToken sync_token) {
+  TRACE_EVENT0("gpu", "VrShellGl::SubmitWebVRFrame");
+
+  bool swapped = command_buffer_gl_->CopyFrameToSurface(
+      frame_index, mailbox, sync_token, !pending_frames_.empty());
+  // Expect a new frame on the surface queue if draw was successful.
+  if (swapped) {
+    main_thread_task_runner_->PostTask(
+        FROM_HERE, base::Bind(&VrShell::OnSubmitWebVRFrameTransferred,
+                              weak_vr_shell_, frame_index));
+    pending_frames_.emplace(frame_index);
+  }
+
+  TRACE_EVENT0("gpu", "VrShellGl::glFinish");
+  // This is a load-bearing glFinish, please don't remove it without
+  // before/after timing comparisons. Goal is to clear the GPU queue
+  // of the native GL context to avoid stalls later in GVR frame
+  // acquire/submit.
+  glFinish();
+}
+
 void VrShellGl::OnUIFrameAvailable() {
   ui_surface_texture_->UpdateTexImage();
 }
@@ -284,42 +365,31 @@ void VrShellGl::OnContentFrameAvailable() {
   received_frame_ = true;
 }
 
-bool VrShellGl::GetPixelEncodedFrameIndex(uint16_t* frame_index) {
-  TRACE_EVENT0("gpu", "VrShellGl::GetPixelEncodedFrameIndex");
-  if (!received_frame_) {
-    if (last_frame_index_ == (uint16_t)-1)
-      return false;
-    *frame_index = last_frame_index_;
-    return true;
+void VrShellGl::OnWebVRFrameAvailable() {
+  // A "while" loop here is a bad idea. It's legal to call
+  // UpdateTexImage repeatedly even if no frames are available, but
+  // that does *not* wait for a new frame, it just reuses the most
+  // recent one. That would mess up the count.
+  if (pending_frames_.empty()) {
+    // We're expecting a frame, but it's not here yet. Retry in OnVsync.
+    ++premature_received_frames_;
+    return;
   }
-  received_frame_ = false;
 
-  // Read the pose index encoded in a bottom left pixel as color values.
-  // See also third_party/WebKit/Source/modules/vr/VRDisplay.cpp which
-  // encodes the pose index, and device/vr/android/gvr/gvr_device.cc
-  // which tracks poses. Returns the low byte (0..255) if valid, or -1
-  // if not valid due to bad magic number.
-  uint8_t pixels[4];
-  // Assume we're reading from the framebuffer we just wrote to.
-  // That's true currently, we may need to use glReadBuffer(GL_BACK)
-  // or equivalent if the rendering setup changes in the future.
-  glReadPixels(0, 0, 1, 1, GL_RGBA, GL_UNSIGNED_BYTE, pixels);
+  webvr_surface_texture_->UpdateTexImage();
+  int frame_index = pending_frames_.front();
+  TRACE_EVENT1("gpu", "VrShellGl::OnWebVRFrameAvailable", "frame", frame_index);
+  pending_frames_.pop();
 
-  // Check for the magic number written by VRDevice.cpp on submit.
-  // This helps avoid glitches from garbage data in the render
-  // buffer that can appear during initialization or resizing. These
-  // often appear as flashes of all-black or all-white pixels.
-  if (pixels[1] == kWebVrPosePixelMagicNumbers[0] &&
-      pixels[2] == kWebVrPosePixelMagicNumbers[1]) {
-    // Pose is good.
-    *frame_index = pixels[0];
-    last_frame_index_ = pixels[0];
-    return true;
-  }
-  VLOG(1) << "WebVR: reject decoded pose index " << static_cast<int>(pixels[0])
-          << ", bad magic number " << static_cast<int>(pixels[1]) << ", "
-          << static_cast<int>(pixels[2]);
-  return false;
+  // It is be legal for the WebVR client to submit a new frame now,
+  // since we've consumed the image. TODO(klausw): would timing be
+  // better to move the rendered notification after draw?
+
+  main_thread_task_runner_->PostTask(
+      FROM_HERE, base::Bind(&VrShell::OnSubmitWebVRFrameRendered,
+                            weak_vr_shell_, frame_index));
+
+  DrawFrame(frame_index);
 }
 
 void VrShellGl::GvrInit(gvr_context* gvr_api) {
@@ -344,12 +414,6 @@ void VrShellGl::GvrInit(gvr_context* gvr_api) {
 }
 
 void VrShellGl::InitializeRenderer() {
-  // While WebVR is going through the compositor path, it shares
-  // the same texture ID. This will change once it gets its own
-  // surface, but store it separately to avoid future confusion.
-  // TODO(klausw,crbug.com/655722): remove this.
-  webvr_texture_id_ = content_texture_id_;
-
   gvr_api_->InitializeGl();
   webvr_head_pose_.assign(kPoseRingBufferSize,
                           gvr_api_->GetHeadSpaceFromStartSpaceRotation(
@@ -359,6 +423,7 @@ void VrShellGl::InitializeRenderer() {
   // For kFramePrimaryBuffer (primary VrShell and WebVR content)
   specs.push_back(gvr_api_->CreateBufferSpec());
   render_size_primary_ = specs[kFramePrimaryBuffer].GetSize();
+  render_size_vrshell_ = render_size_primary_;
 
   // For kFrameHeadlockedBuffer (for WebVR insecure content warning).
   // Set this up at fixed resolution, the (smaller) FOV gets set below.
@@ -630,15 +695,33 @@ void VrShellGl::SendGesture(InputTarget input_target,
       base::Bind(target, weak_vr_shell_, base::Passed(std::move(event))));
 }
 
-void VrShellGl::DrawFrame() {
-  TRACE_EVENT0("gpu", "VrShellGl::DrawFrame");
+void VrShellGl::DrawFrame(int frame_index) {
+  TRACE_EVENT1("gpu", "VrShellGl::DrawFrame", "frame", frame_index);
 
   // Reset the viewport list to just the pair of viewports for the
   // primary buffer each frame. Head-locked viewports get added by
   // DrawVrShell if needed.
   buffer_viewport_list_->SetToRecommendedBufferViewports();
 
+  // If needed, resize the primary buffer for use with WebVR.
+  if (web_vr_mode_) {
+    if (render_size_primary_ != webvr_surface_size_) {
+      if (!webvr_surface_size_.width) {
+        return;
+      }
+      render_size_primary_ = webvr_surface_size_;
+      swap_chain_->ResizeBuffer(kFramePrimaryBuffer, render_size_primary_);
+    }
+  } else {
+    if (render_size_primary_ != render_size_vrshell_) {
+      render_size_primary_ = render_size_vrshell_;
+      swap_chain_->ResizeBuffer(kFramePrimaryBuffer, render_size_primary_);
+    }
+  }
+
+  TRACE_EVENT_BEGIN0("gpu", "VrShellGl::AcquireFrame");
   gvr::Frame frame = swap_chain_->AcquireFrame();
+  TRACE_EVENT_END0("gpu", "VrShellGl::AcquireFrame");
   if (!frame.is_valid()) {
     return;
   }
@@ -647,18 +730,12 @@ void VrShellGl::DrawFrame() {
     DrawWebVr();
   }
 
-  uint16_t frame_index;
   gvr::Mat4f head_pose;
 
   // When using async reprojection, we need to know which pose was used in
-  // the WebVR app for drawing this frame. Due to unknown amounts of
-  // buffering in the compositor and SurfaceTexture, we read the pose number
-  // from a corner pixel. There's no point in doing this for legacy
-  // distortion rendering since that doesn't need a pose, and reading back
-  // pixels is an expensive operation. TODO(klausw,crbug.com/655722): stop
-  // doing this once we have working no-compositor rendering for WebVR.
-  if (web_vr_mode_ && gvr_api_->GetAsyncReprojectionEnabled() &&
-      GetPixelEncodedFrameIndex(&frame_index)) {
+  // the WebVR app for drawing this frame. Only needed if reprojection is
+  // in use.
+  if (web_vr_mode_ && gvr_api_->GetAsyncReprojectionEnabled()) {
     static_assert(!((kPoseRingBufferSize - 1) & kPoseRingBufferSize),
                   "kPoseRingBufferSize must be a power of 2");
     head_pose = webvr_head_pose_[frame_index % kPoseRingBufferSize];
@@ -682,9 +759,10 @@ void VrShellGl::DrawFrame() {
       if (index > frame_index && index <= frame_index + kPoseRingBufferSize)
         break;
 
-      const BoundsPair& bounds = pending_bounds_.front().second;
-      webvr_left_viewport_->SetSourceUv(bounds.first);
-      webvr_right_viewport_->SetSourceUv(bounds.second);
+      const WebVrBounds& bounds = pending_bounds_.front().second;
+      webvr_left_viewport_->SetSourceUv(bounds.left_bounds);
+      webvr_right_viewport_->SetSourceUv(bounds.right_bounds);
+      CreateOrResizeWebVRSurface(bounds.source_size);
       pending_bounds_.pop();
     }
     buffer_viewport_list_->SetBufferViewport(GVR_LEFT_EYE,
@@ -711,21 +789,31 @@ void VrShellGl::DrawFrame() {
   const float screen_tilt = kDesktopScreenTiltDefault * M_PI / 180.0f;
   scene_->UpdateTransforms(screen_tilt, TimeInMicroseconds());
 
-  UpdateController(GetForwardVector(head_pose));
+  {
+    TRACE_EVENT0("gpu", "VrShellGl::UpdateController");
+    UpdateController(GetForwardVector(head_pose));
+  }
 
-  DrawVrShell(head_pose, frame);
+  // Finish drawing in the primary buffer, and draw the headlocked buffer
+  // if needed. This must be the last drawing call, this method will
+  // return with no frame being bound.
+  DrawVrShellAndUnbind(head_pose, frame);
 
-  frame.Unbind();
-  frame.Submit(*buffer_viewport_list_, head_pose);
+  {
+    TRACE_EVENT0("gpu", "VrShellGl::Submit");
+    frame.Submit(*buffer_viewport_list_, head_pose);
+  }
 
   // No need to swap buffers for surfaceless rendering.
   if (!surfaceless_rendering_) {
     // TODO(mthiesse): Support asynchronous SwapBuffers.
+    TRACE_EVENT0("gpu", "VrShellGl::SwapBuffers");
     surface_->SwapBuffers();
   }
 }
 
-void VrShellGl::DrawVrShell(const gvr::Mat4f& head_pose, gvr::Frame& frame) {
+void VrShellGl::DrawVrShellAndUnbind(const gvr::Mat4f& head_pose,
+                                     gvr::Frame& frame) {
   TRACE_EVENT0("gpu", "VrShellGl::DrawVrShell");
   std::vector<const ContentRectangle*> head_locked_elements;
   std::vector<const ContentRectangle*> world_elements;
@@ -762,6 +850,7 @@ void VrShellGl::DrawVrShell(const gvr::Mat4f& head_pose, gvr::Frame& frame) {
     DrawUiView(&head_pose, world_elements, render_size_primary_,
                kViewportListPrimaryOffset);
   }
+  frame.Unbind();  // Done with the primary buffer.
 
   if (!head_locked_elements.empty()) {
     // Add head-locked viewports. The list gets reset to just
@@ -774,23 +863,15 @@ void VrShellGl::DrawVrShell(const gvr::Mat4f& head_pose, gvr::Frame& frame) {
         *headlocked_right_viewport_);
 
     // Bind the headlocked framebuffer.
-    // TODO(mthiesse): We don't unbind this? Maybe some cleanup is in order
-    // here.
     frame.BindBuffer(kFrameHeadlockedBuffer);
     glClearColor(0.0f, 0.0f, 0.0f, 0.0f);
     glClear(GL_COLOR_BUFFER_BIT | GL_DEPTH_BUFFER_BIT);
     DrawUiView(nullptr, head_locked_elements, render_size_headlocked_,
                kViewportListHeadlockedOffset);
+    frame.Unbind();  // Done with the headlocked buffer.
   }
 }
 
-gvr::Sizei VrShellGl::GetWebVRCompositorSurfaceSize() {
-  // This is a stopgap while we're using the WebVR compositor rendering path.
-  // TODO(klausw,crbug.com/655722): Remove this method and member once we're
-  // using a separate WebVR render surface.
-  return content_tex_physical_size_;
-}
-
 void VrShellGl::DrawUiView(const gvr::Mat4f* head_pose,
                            const std::vector<const ContentRectangle*>& elements,
                            const gvr::Sizei& render_size,
@@ -997,7 +1078,7 @@ void VrShellGl::DrawWebVr() {
   glDisable(GL_BLEND);
   glDisable(GL_POLYGON_OFFSET_FILL);
 
-  glViewport(0, 0, render_size_primary_.width, render_size_primary_.height);
+  glViewport(0, 0, webvr_surface_size_.width, webvr_surface_size_.height);
   vr_shell_renderer_->GetWebVrRenderer()->Draw(webvr_texture_id_);
 }
 
@@ -1028,13 +1109,14 @@ void VrShellGl::SetWebVrMode(bool enabled) {
 
 void VrShellGl::UpdateWebVRTextureBounds(int16_t frame_index,
                                          const gvr::Rectf& left_bounds,
-                                         const gvr::Rectf& right_bounds) {
+                                         const gvr::Rectf& right_bounds,
+                                         const gvr::Sizei& source_size) {
   if (frame_index < 0) {
     webvr_left_viewport_->SetSourceUv(left_bounds);
     webvr_right_viewport_->SetSourceUv(right_bounds);
   } else {
-    pending_bounds_.emplace(
-        std::make_pair(frame_index, std::make_pair(left_bounds, right_bounds)));
+    pending_bounds_.emplace(std::make_pair(
+        frame_index, WebVrBounds(left_bounds, right_bounds, source_size)));
   }
 }
 
@@ -1068,6 +1150,12 @@ base::WeakPtr<VrShellGl> VrShellGl::GetWeakPtr() {
 }
 
 void VrShellGl::OnVSync() {
+  while (premature_received_frames_ > 0) {
+    TRACE_EVENT0("gpu", "VrShellGl::OnWebVRFrameAvailableRetry");
+    --premature_received_frames_;
+    OnWebVRFrameAvailable();
+  }
+
   base::TimeTicks now = base::TimeTicks::Now();
   base::TimeTicks target;
 
@@ -1087,7 +1175,9 @@ void VrShellGl::OnVSync() {
     pending_vsync_ = true;
     pending_time_ = time;
   }
-  DrawFrame();
+  if (!web_vr_mode_) {
+    DrawFrame(-1);
+  }
 }
 
 void VrShellGl::OnRequest(device::mojom::VRVSyncProviderRequest request) {
@@ -1157,8 +1247,11 @@ void VrShellGl::ResetPose() {
 void VrShellGl::CreateVRDisplayInfo(
     const base::Callback<void(device::mojom::VRDisplayInfoPtr)>& callback,
     uint32_t device_id) {
+  // This assumes that the initial webvr_surface_size_ was set to the
+  // appropriate recommended render resolution as the default size during
+  // InitializeGl. Revisit if the initialization order changes.
   device::mojom::VRDisplayInfoPtr info = VrShell::CreateVRDisplayInfo(
-      gvr_api_.get(), content_tex_physical_size_, device_id);
+      gvr_api_.get(), webvr_surface_size_, device_id);
   main_thread_task_runner_->PostTask(
       FROM_HERE,
       base::Bind(&RunVRDisplayInfoCallback, callback, base::Passed(&info)));