[ozone/wayland] rely on implicit sync for Broadcom and Intel GPUs.

The issue with fences was identified a long time ago, but it did not
concern Linux before. As long as we use surfaceless drawing and
insert fences between each buffer swap, we noticed some perf
regression.

On a powerful laptop (dell 9560) with Intel 630 graphics, gpu::AsyncSwap
took ~8ms with fences and ~1ms without fences.

On Raspberry Pi3 with Broadcom GPU, swap took ~105ms and
after the applied workaround it took 77ms. The FPS rate
became 16-17FPS instead of 11-12FPS in aquarium webgl demo.

https://docs.google.com/spreadsheets/d/1dKLIQ6dkql-X0ppdc5s5y-ws2hPM-RLgqgGzkFpgUe0/edit?usp=sharing

Bug: 938286
Change-Id: I9fc616976a6ee3f9b4544ad7d8effc41e3ad4a91
Reviewed-on: https://chromium-review.googlesource.com/c/chromium/src/+/1503576
Commit-Queue: Maksim Sisov <msisov@igalia.com>
Reviewed-by: Robert Kroeger <rjkroege@chromium.org>
Reviewed-by: Antoine Labour <piman@chromium.org>
Cr-Commit-Position: refs/heads/master@{#638037}

gpu/config/gpu_driver_bug_list.json

@@ -3183,6 +3183,30 @@
       "features": [
         "disable_direct_composition_layers"
       ]
+    },
+    {
+      "id": 295,
+      "description": "Avoid waiting on a egl fence before swapping buffers and rely on implicit sync on Intel GPUs",
+      "cr_bugs": [938286],
+      "os": {
+        "type": "linux"
+      },
+      "gl_vendor": "Intel.*",
+      "features": [
+        "rely_on_implicit_sync_for_swap_buffers"
+      ]
+    },
+    {
+      "id": 296,
+      "description": "Avoid waiting on a egl fence before swapping buffers and rely on implicit sync on Broadcom GPUs",
+      "cr_bugs": [938286],
+      "os": {
+        "type": "linux"
+      },
+      "gl_vendor": "Broadcom.*",
+      "features": [
+        "rely_on_implicit_sync_for_swap_buffers"
+      ]
     }
   ]
 }

ui/ozone/platform/wayland/gpu/gbm_surfaceless_wayland.cc

@@ -103,6 +103,12 @@ void GbmSurfacelessWayland::SwapBuffersAsync(
   frame->presentation_callback = std::move(presentation_callback);
   unsubmitted_frames_.push_back(std::make_unique<PendingFrame>());
 
+  if (!use_egl_fence_sync_) {
+    frame->ready = true;
+    SubmitFrame();
+    return;
+  }
+
   // TODO: the following should be replaced by a per surface flush as it gets
   // implemented in GL drivers.
   EGLSyncKHR fence = InsertFence(has_implicit_external_sync_);
@@ -156,6 +162,10 @@ EGLConfig GbmSurfacelessWayland::GetConfig() {
   return config_;
 }
 
+void GbmSurfacelessWayland::SetRelyOnImplicitSync() {
+  use_egl_fence_sync_ = false;
+}
+
 GbmSurfacelessWayland::~GbmSurfacelessWayland() {
   surface_factory_->UnregisterSurface(widget_);
 }

ui/ozone/platform/wayland/gpu/gbm_surfaceless_wayland.h

@@ -53,6 +53,7 @@ class GbmSurfacelessWayland : public gl::SurfacelessEGL {
                           SwapCompletionCallback completion_callback,
                           PresentationCallback presentation_callback) override;
   EGLConfig GetConfig() override;
+  void SetRelyOnImplicitSync() override;
 
  private:
   ~GbmSurfacelessWayland() override;
@@ -95,6 +96,7 @@ class GbmSurfacelessWayland : public gl::SurfacelessEGL {
   std::unique_ptr<PendingFrame> submitted_frame_;
   bool has_implicit_external_sync_;
   bool last_swap_buffers_result_ = true;
+  bool use_egl_fence_sync_ = true;
 
   base::WeakPtrFactory<GbmSurfacelessWayland> weak_factory_;