csmt_yield.diff

diff --git a/dlls/wined3d/cs.c b/dlls/wined3d/cs.c
index 7731a82..301e395 100644
--- a/dlls/wined3d/cs.c
+++ b/dlls/wined3d/cs.c
@@ -662,7 +662,10 @@ static void wined3d_cs_exec_clear(struct wined3d_cs *cs, const void *data)
      * In Counter-Strike: Source a frame difference of 3 causes noticable
      * input delay that makes the game unplayable. */
     while (pending > 1)
+    {
+        wined3d_cs_mt_yield();
         pending = InterlockedCompareExchange(&cs->pending_presents, 0, 0);
+    }
 }
 
 static UINT wined3d_cs_exec_clear(struct wined3d_cs *cs, const void *data)
@@ -3285,7 +3288,8 @@ static void wined3d_cs_mt_finish(struct wined3d_cs *cs)
 
     /* A busy wait should be fine, we're not supposed to have to wait very
      * long. */
-    while (!InterlockedCompareExchange(&fence, TRUE, TRUE));
+    while (!InterlockedCompareExchange(&fence, TRUE, TRUE))
+        wined3d_cs_mt_yield();
 }
 
 static void wined3d_cs_mt_finish_prio(struct wined3d_cs *cs)
@@ -3307,7 +3311,8 @@ static void wined3d_cs_mt_finish_prio(struct wined3d_cs *cs)
 
     /* A busy wait should be fine, we're not supposed to have to wait very
      * long. */
-    while (!InterlockedCompareExchange(&fence, TRUE, TRUE));
+    while (!InterlockedCompareExchange(&fence, TRUE, TRUE))
+        wined3d_cs_mt_yield();
 }
 
 static const struct wined3d_cs_ops wined3d_cs_mt_ops =
@@ -3425,7 +3430,9 @@ static DWORD WINAPI wined3d_cs_run(void *thread_param)
     LONG tail;
     char poll = 0;
     struct wined3d_cs_queue *queue;
-    unsigned int spin_count = 0;
+
+    LARGE_INTEGER spin_time, now, diff;
+    spin_time.QuadPart = 0;
 
     TRACE("Started.\n");
 
@@ -3448,19 +3455,37 @@ static DWORD WINAPI wined3d_cs_run(void *thread_param)
         else if (!queue_is_empty(&cs->queue))
         {
             queue = &cs->queue;
-            if (!queue_is_empty(&cs->prio_queue))
-                queue = &cs->prio_queue;
+        }
+        else if (poll != 0)
+        {
+            poll = 10;
+            continue;
         }
         else
         {
-            spin_count++;
-            if (spin_count >= WINED3D_CS_SPIN_COUNT && list_empty(&cs->query_poll_list))
+            NtQuerySystemTime(&now);
+
+            if (spin_time.QuadPart == 0)
+            {
+                spin_time.QuadPart = now.QuadPart;
+            }
+
+            diff.QuadPart = now.QuadPart - spin_time.QuadPart;
+
+            if (diff.QuadPart < WINED3D_CS_SPIN_TIME ||
+                !list_empty(&cs->query_poll_list))
+            {
+                wined3d_cs_mt_yield();
+            }
+            else
+            {
                 wined3d_cs_wait_event(cs);
+            }
 
             continue;
         }
 
-        spin_count = 0;
+        spin_time.QuadPart = 0;
 
         tail = queue->tail;
         opcode = *(const enum wined3d_cs_op *)&queue->data[tail];
diff --git a/dlls/wined3d/wined3d_main.c b/dlls/wined3d/wined3d_main.c
index ac6aa6e..f262fca 100644
--- a/dlls/wined3d/wined3d_main.c
+++ b/dlls/wined3d/wined3d_main.c
@@ -93,6 +93,7 @@ struct wined3d_settings wined3d_settings =
     FALSE,          /* 3D support enabled by default. */
 #if defined(STAGING_CSMT)
     TRUE,           /* Multithreaded CS by default. */
+    2,              /* 0 = Busy wait, 1 = usleep, 2 = SwitchToThread */
 #endif /* STAGING_CSMT */
 };
 
@@ -346,6 +347,22 @@ static BOOL wined3d_dll_init(HINSTANCE hInstDLL)
             TRACE("Disabling multithreaded command stream.\n");
             wined3d_settings.cs_multithreaded = FALSE;
         }
+        if (!get_config_key(hkey, appkey, "CSMTYield", buffer, size))
+        {
+            if (!strcmp(buffer, "nothing"))
+            {
+                TRACE("CSMT will use busy-wait loops.\n");
+                wined3d_settings.cs_multithreaded_yield = 0;
+            } else if (!strcmp(buffer, "usleep"))
+            {
+                TRACE("CSMT will use usleep in wait loops.\n");
+                wined3d_settings.cs_multithreaded_yield = 1;
+            } else if (!strcmp(buffer, "SwitchToThread"))
+            {
+                TRACE("CSMT will use SwitchToThread in wait loops.\n");
+                wined3d_settings.cs_multithreaded_yield = 2;
+            }
+        }
     }
 
     FIXME_(winediag)("Experimental wined3d CSMT feature is currently %s.\n",
diff --git a/dlls/wined3d/wined3d_private.h b/dlls/wined3d/wined3d_private.h
index 199731d..dee3a78 100644
--- a/dlls/wined3d/wined3d_private.h
+++ b/dlls/wined3d/wined3d_private.h
@@ -315,6 +315,7 @@ struct wined3d_settings
     BOOL no_3d;
 #if defined(STAGING_CSMT)
     BOOL cs_multithreaded;
+    int cs_multithreaded_yield;
 #endif /* STAGING_CSMT */
 };
 
@@ -2667,9 +2668,26 @@ static inline void wined3d_resource_release(struct wined3d_resource *resource)
     InterlockedDecrement(&resource->access_count);
 }
 
+#if defined(STAGING_CSMT)
+static inline void wined3d_cs_mt_yield(void)
+{
+    int yield_strategy = wined3d_settings.cs_multithreaded_yield;
+
+    if (yield_strategy == 1)
+        usleep(0);
+    else if (yield_strategy == 2)
+        SwitchToThread();
+}
+#endif /* STAGING_CSMT */
+
 static inline void wined3d_resource_wait_idle(struct wined3d_resource *resource)
 {
-    while (InterlockedCompareExchange(&resource->access_count, 0, 0));
+    while (InterlockedCompareExchange(&resource->access_count, 0, 0))
+    {
+#if defined(STAGING_CSMT)
+        wined3d_cs_mt_yield();
+#endif /* STAGING_CSMT */
+    }
 }
 
 void resource_cleanup(struct wined3d_resource *resource) DECLSPEC_HIDDEN;
@@ -3167,7 +3185,7 @@ struct wined3d_cs_list
 };
 
 #define WINED3D_CS_QUEUE_SIZE 0x100000
-#define WINED3D_CS_SPIN_COUNT 10000000
+#define WINED3D_CS_SPIN_TIME  100000
 
 struct wined3d_cs_queue
 {

I also added a yield to the main loop, while making sure that we call "poll_queries" at once before calling yield.

The default is now "SwitchToThread", which is basically sched_yield and the same thing the nvidia driver does. On laptops and in other situations where there are limiting factors like TDP, "usleep" might yield better overall performance.

HKCU/Software/Wine/Direct3D/CSMTYield can be set to one of the following values to control behavior:
nothing - this is the old behavior, use busy-waits
usleep - use usleep(0), recommended for machines that hit their TDP
SwitchToThread - this is basically sched_yield() on Linux