From: Steinar H. Gunderson <sgunderson@bigfoot.com>
Date: Wed, 28 Jun 2017 19:34:58 +0000 (+0200)
Subject: Yield instead of busy-waiting on fences for NVIDIA cards.
X-Git-Tag: 1.6.1~22
X-Git-Url: https://git.sesse.net/?p=nageru;a=commitdiff_plain;h=fdf765b229349a1d1a30414341957a4d10a0b6c9

Yield instead of busy-waiting on fences for NVIDIA cards.
---

diff --git a/decklink_output.cpp b/decklink_output.cpp
index a19d0a5..d684677 100644
--- a/decklink_output.cpp
+++ b/decklink_output.cpp
@@ -2,6 +2,7 @@
 #include <movit/util.h>
 #include <movit/resource_pool.h>  // Must be above the Xlib includes.
 #include <pthread.h>
+#include <unistd.h>
 
 #include <mutex>
 
@@ -543,7 +544,15 @@ void DeckLinkOutput::present_thread_func()
 			++metric_decklink_output_inflight_frames;
 		}
 
-		glClientWaitSync(frame->fence.get(), /*flags=*/0, GL_TIMEOUT_IGNORED);
+		for ( ;; ) {
+			int err = glClientWaitSync(frame->fence.get(), /*flags=*/0, 0);
+			if (err == GL_TIMEOUT_EXPIRED) {
+				// NVIDIA likes to busy-wait; yield instead.
+				this_thread::sleep_for(milliseconds(1));
+			} else {
+				break;
+			}
+		}
 		check_error();
 		frame->fence.reset();
 
diff --git a/quicksync_encoder.cpp b/quicksync_encoder.cpp
index 1b36aa5..6cab461 100644
--- a/quicksync_encoder.cpp
+++ b/quicksync_encoder.cpp
@@ -1981,8 +1981,12 @@ void QuickSyncEncoderImpl::pass_frame(QuickSyncEncoderImpl::PendingFrame frame,
 	// Wait for the GPU to be done with the frame.
 	GLenum sync_status;
 	do {
-		sync_status = glClientWaitSync(frame.fence.get(), 0, 1000000000);
+		sync_status = glClientWaitSync(frame.fence.get(), 0, 0);
 		check_error();
+		if (sync_status == GL_TIMEOUT_EXPIRED) {
+			// NVIDIA likes to busy-wait; yield instead.
+			this_thread::sleep_for(milliseconds(1));
+		}
 	} while (sync_status == GL_TIMEOUT_EXPIRED);
 	assert(sync_status != GL_WAIT_FAILED);