Experimental patch for adapting adaptation to CPU count on Mac.

webrtc/video/overuse_frame_detector.h

Index: webrtc/video/overuse_frame_detector.h
diff --git a/webrtc/video/overuse_frame_detector.h b/webrtc/video/overuse_frame_detector.h
index 0ef2e43857d2601c800c48a09d0dc6e237692e53..f817ec9cacaf847c187e4b00abc8925ef2c74684 100644
--- a/webrtc/video/overuse_frame_detector.h
+++ b/webrtc/video/overuse_frame_detector.h
@@ -19,6 +19,10 @@
 #include "webrtc/base/thread_checker.h"
 #include "webrtc/modules/include/module.h"
 
+#if WEBRTC_MAC
+#include <mach/mach.h>
+#endif
+
 namespace webrtc {
 
 class Clock;
@@ -38,12 +42,34 @@ class CpuOveruseObserver {
 
 struct CpuOveruseOptions {
   CpuOveruseOptions()
-      : low_encode_usage_threshold_percent(55),
-        high_encode_usage_threshold_percent(85),
+      : low_encode_usage_threshold_percent(42),

[tommi, 2016/02/04 16:40:02]

Is 42 something that should be set for all platforms or only for mac?

[torbjorng (webrtc), 2016/02/04 16:54:58]

I decreased this to make it about 85/2; the interval should really reflect the
scaling factors made available by libyuv (or more accurately, these factors
squared!). Else we will presumably tend to oscillate.

I deny that I pulled 42 out of a hat; it is the result of a huge computation.

         frame_timeout_interval_ms(1500),
         min_frame_samples(120),
         min_process_count(3),
-        high_threshold_consecutive_count(2) {}
+        high_threshold_consecutive_count(2) {
+#if WEBRTC_MAC

[tommi, 2016/02/04 16:40:02]

nit: #if defined(WEBRTC_MAC)

[torbjorng (webrtc), 2016/02/04 16:54:59]

Done.

+    // This is a proof-of-concept hack for letting the physical core count
+    // affect the interval into which we attempt to scale.  For now, the code
+    // is Mac OS specific, since that's the platform were we saw the problem.
+    // Note that we make the interval 2x+epsilon wide, since scaling steps are
+    // close to that.
+    struct host_basic_info hbi;
+    mach_msg_type_number_t datasize;
+    // FIXME(torbjorng): Add syscall error checking, fallback code.

[tommi, 2016/02/04 16:40:02]

nit: s/FIXME/TODO

[torbjorng (webrtc), 2016/02/04 16:54:59]

Done.

+    host_info(mach_host_self(), HOST_BASIC_INFO, (host_info_t)&hbi, &datasize);

[tommi, 2016/02/04 16:40:02]

does this work as expected inside the sandbox?

[torbjorng (webrtc), 2016/02/04 16:54:58]

Good question. I will verify that.  I observed chrome://webrtc-internals, and it
looked as it was effective. But timing reproducibility could as we know be
better on this platform.

We should use SystemInfo (base/systeminfo.h) but I did not because I couldn't
get *physical* core count from it.

+
+    if (hbi.physical_cpu == 1) {
+      low_encode_usage_threshold_percent = 17;
+    } else if (hbi.physical_cpu == 2) {
+      low_encode_usage_threshold_percent = 27;
+    } else {
+      // Init list default.
+    }
+#endif
+    // Make high = low + epsilon.
+    high_encode_usage_threshold_percent
+      = 33 * low_encode_usage_threshold_percent / 16 + 1;

[tommi, 2016/02/04 16:40:02]

nit: assignment operator on previous line (could also run git cl format)

[torbjorng (webrtc), 2016/02/04 16:54:58]

Done.

+  }
 
   int low_encode_usage_threshold_percent;  // Threshold for triggering underuse.
   int high_encode_usage_threshold_percent;  // Threshold for triggering overuse.