0001-wined3d-use-SwitchToThread-and-calls-to-select-in-bu.patch

From 5e89ce2476b3e516c19afc58b1b14a156acd2f43 Mon Sep 17 00:00:00 2001
From: Nils Kuhnhenn <kuhnhenn.nils@gmail.com>
Date: Sat, 6 Aug 2016 16:25:27 +0200
Subject: [PATCH] wined3d: use SwitchToThread() and calls to select() in
 busy-loops

---
 dlls/wined3d/cs.c              | 39 +++++++++++++++++++++++++++++++--------
 dlls/wined3d/wined3d_private.h | 42 ++++++++++++++++++++++++++++++++++++++++--
 2 files changed, 71 insertions(+), 10 deletions(-)

diff --git a/dlls/wined3d/cs.c b/dlls/wined3d/cs.c
index 505661d..a099c33 100644
--- a/dlls/wined3d/cs.c
+++ b/dlls/wined3d/cs.c
@@ -642,6 +642,8 @@ void wined3d_cs_emit_present(struct wined3d_cs *cs, struct wined3d_swapchain *sw
     unsigned int i;
 #if defined(STAGING_CSMT)
     LONG pending;
+    LARGE_INTEGER spin_start;
+    spin_start.QuadPart = 0;
 #endif /* STAGING_CSMT */
 
     op = cs->ops->require_space(cs, sizeof(*op));
@@ -677,7 +679,10 @@ 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(&spin_start);
         pending = InterlockedCompareExchange(&cs->pending_presents, 0, 0);
+    }
 }
 
 static UINT wined3d_cs_exec_clear(struct wined3d_cs *cs, const void *data)
@@ -3389,17 +3394,25 @@ static void wined3d_cs_emit_stop(struct wined3d_cs *cs)
 static void wined3d_cs_mt_finish(struct wined3d_cs *cs)
 {
     BOOL fence;
+
+    LARGE_INTEGER spin_start;
+    spin_start.QuadPart = 0;
+
     wined3d_cs_emit_fence(cs, &fence);
 
     /* 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(&spin_start);
 }
 
 static void wined3d_cs_mt_finish_prio(struct wined3d_cs *cs)
 {
     BOOL fence;
 
+    LARGE_INTEGER spin_start;
+    spin_start.QuadPart = 0;
+
     if (cs->thread_id == GetCurrentThreadId())
     {
         static BOOL once;
@@ -3415,7 +3428,8 @@ 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(&spin_start);
 }
 
 static const struct wined3d_cs_ops wined3d_cs_mt_ops =
@@ -3533,7 +3547,11 @@ 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_start;
+    spin_start.QuadPart = 0;
+
+    BOOL spin_time_exeeded = FALSE;
 
     TRACE("Started.\n");
 
@@ -3556,19 +3574,24 @@ 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))
+            if (!spin_time_exeeded || !list_empty(&cs->query_poll_list))
+                spin_time_exeeded = wined3d_cs_mt_yield(&spin_start).QuadPart > WINED3D_CS_SPIN_TIME;
+            else
                 wined3d_cs_wait_event(cs);
 
             continue;
         }
 
-        spin_count = 0;
+        spin_start.QuadPart = 0;
+        spin_time_exeeded = FALSE;
 
         tail = queue->tail;
         opcode = *(const enum wined3d_cs_op *)&queue->data[tail];
diff --git a/dlls/wined3d/wined3d_private.h b/dlls/wined3d/wined3d_private.h
index 4247449..b119f73 100644
--- a/dlls/wined3d/wined3d_private.h
+++ b/dlls/wined3d/wined3d_private.h
@@ -2678,9 +2678,47 @@ static inline void wined3d_resource_release(struct wined3d_resource *resource)
     InterlockedDecrement(&resource->access_count);
 }
 
+#if defined(STAGING_CSMT)
+#define WINED3D_YIELD_SPIN_TIME  1000 /* 1000 = 0.1 milliseconds = 100 microseconds */
+
+static LARGE_INTEGER wined3d_cs_mt_yield(LARGE_INTEGER *spin_start)
+{
+    struct timeval tv;
+    LARGE_INTEGER now, diff;
+
+    NtQuerySystemTime(&now);
+
+    if (spin_start->QuadPart <= 0)
+        spin_start->QuadPart = now.QuadPart;
+
+    diff.QuadPart = now.QuadPart - spin_start->QuadPart;
+
+    if (diff.QuadPart > WINED3D_YIELD_SPIN_TIME)
+    {
+        tv.tv_sec = 0;
+        tv.tv_usec = 9;
+        select( 0, NULL, NULL, NULL, &tv);
+    }
+    else
+        SwitchToThread();
+
+    return diff;
+}
+#endif /* STAGING_CSMT */
+
 static inline void wined3d_resource_wait_idle(struct wined3d_resource *resource)
 {
-    while (InterlockedCompareExchange(&resource->access_count, 0, 0));
+#if defined(STAGING_CSMT)
+    LARGE_INTEGER spin_start;
+    spin_start.QuadPart = 0;
+#endif /* STAGING_CSMT */
+
+    while (InterlockedCompareExchange(&resource->access_count, 0, 0))
+    {
+#if defined(STAGING_CSMT)
+        wined3d_cs_mt_yield(&spin_start);
+#endif /* STAGING_CSMT */
+    }
 }
 
 void resource_cleanup(struct wined3d_resource *resource) DECLSPEC_HIDDEN;
@@ -3169,7 +3207,7 @@ struct wined3d_cs_list
 };
 
 #define WINED3D_CS_QUEUE_SIZE 0x100000
-#define WINED3D_CS_SPIN_COUNT 10000000
+#define WINED3D_CS_SPIN_TIME  200000 /* 20000 = 20 milliseconds */
 
 struct wined3d_cs_queue
 {
-- 
2.9.2