Traces: --offscreen syncs on N-2 frame GPU completion

This avoids the issue described on the bug where an app's assumption
about double buffering can create a race.

This prevents the replay from getting too far ahead rendering to
offscreen framebuffers.

Also correct mOffscreenFramebuffers to use mTotalFrameCount as the use
of mOffscreenFrameCount appears accidental.

Bug: angleproject:370089935
Change-Id: I30f438eb66201fe77d61710cbe6e90d47e839dd8
Reviewed-on: https://chromium-review.googlesource.com/c/angle/angle/+/5895974
Reviewed-by: Cody Northrop <[email protected]>
Reviewed-by: Charlie Lao <[email protected]>
Commit-Queue: Roman Lavrov <[email protected]>

src/tests/perf_tests/TracePerfTest.cpp

@@ -223,9 +223,12 @@ class TracePerfTest : public ANGLERenderTest
     std::vector<TimeSample> mTimeline;
 
     bool mUseTimestampQueries                                           = false;
+    // Note: more than 2 offscreen buffers can cause races, surface is double buffered so real-world
+    // apps can rely on (now broken) assumptions about GPU completion of a previous frame
     static constexpr int mMaxOffscreenBufferCount                       = 2;
     std::array<GLuint, mMaxOffscreenBufferCount> mOffscreenFramebuffers = {0, 0};
     std::array<GLuint, mMaxOffscreenBufferCount> mOffscreenTextures     = {0, 0};
+    std::array<GLsync, mMaxOffscreenBufferCount> mOffscreenSyncs        = {0, 0};
     GLuint mOffscreenDepthStencil                                       = 0;
     int mWindowWidth                                                    = 0;
     int mWindowHeight                                                   = 0;
@@ -2102,6 +2105,19 @@ void TracePerfTest::drawBenchmark()
         // ping pong between them for each frame.
         glBindFramebuffer(GL_FRAMEBUFFER,
                           mOffscreenFramebuffers[mTotalFrameCount % mMaxOffscreenBufferCount]);
+
+        GLsync sync = mOffscreenSyncs[mTotalFrameCount % mMaxOffscreenBufferCount];
+        if (sync)
+        {
+            constexpr GLuint64 kTimeout = 2'000'000'000;  // 2 seconds
+            GLenum result = glClientWaitSync(sync, GL_SYNC_FLUSH_COMMANDS_BIT, kTimeout);
+            if (result != GL_CONDITION_SATISFIED && result != GL_ALREADY_SIGNALED)
+            {
+                failTest(std::string("glClientWaitSync unexpected result: ") +
+                         std::to_string(result));
+            }
+            glDeleteSync(sync);
+        }
     }
 
     char frameName[32];
@@ -2130,9 +2146,8 @@ void TracePerfTest::drawBenchmark()
             glGetIntegerv(GL_READ_FRAMEBUFFER_BINDING, &currentReadFBO);
 
             glBindFramebuffer(GL_DRAW_FRAMEBUFFER, 0);
-            glBindFramebuffer(
-                GL_READ_FRAMEBUFFER,
-                mOffscreenFramebuffers[mOffscreenFrameCount % mMaxOffscreenBufferCount]);
+            glBindFramebuffer(GL_READ_FRAMEBUFFER,
+                              mOffscreenFramebuffers[mTotalFrameCount % mMaxOffscreenBufferCount]);
 
             uint32_t frameX  = (mOffscreenFrameCount % kFramesPerXY) % kFramesPerX;
             uint32_t frameY  = (mOffscreenFrameCount % kFramesPerXY) / kFramesPerX;
@@ -2147,6 +2162,9 @@ void TracePerfTest::drawBenchmark()
                 glDisable(GL_SCISSOR_TEST);
             }
 
+            mOffscreenSyncs[mTotalFrameCount % mMaxOffscreenBufferCount] =
+                glFenceSync(GL_SYNC_GPU_COMMANDS_COMPLETE, 0);
+
             glBlitFramebuffer(0, 0, mWindowWidth, mWindowHeight, windowX, windowY,
                               windowX + kOffscreenFrameWidth, windowY + kOffscreenFrameHeight,
                               GL_COLOR_BUFFER_BIT, GL_NEAREST);