diff --git a/patches/patchinstall.sh b/patches/patchinstall.sh
index bb019d95..74ec06fe 100755
--- a/patches/patchinstall.sh
+++ b/patches/patchinstall.sh
@@ -8493,6 +8493,8 @@ if test "$enable_wined3d_CSMT_Main" -eq 1; then
 		printf '%s\n' '+    { "Sebastian Lackner", "wined3d: Do not immediately submit stateblock updates.", 1 },';
 		printf '%s\n' '+    { "Sebastian Lackner", "wined3d: Get rid of TLS for command stream.", 1 },';
 		printf '%s\n' '+    { "Sebastian Lackner", "wined3d: Clean up cs lists on shutdown.", 1 },';
+		printf '%s\n' '+    { "Michael Müller", "wined3d: Use a separate lock for each CS list.", 1 },';
+		printf '%s\n' '+    { "Sebastian Lackner", "wined3d: Don'\''t crash in context_release when device was not fully created.", 1 },';
 	) >> "$patchlist"
 fi
 
diff --git a/patches/wined3d-CSMT_Main/0052-wined3d-Use-a-separate-lock-for-each-CS-list.patch b/patches/wined3d-CSMT_Main/0052-wined3d-Use-a-separate-lock-for-each-CS-list.patch
new file mode 100644
index 00000000..b4a8abd2
--- /dev/null
+++ b/patches/wined3d-CSMT_Main/0052-wined3d-Use-a-separate-lock-for-each-CS-list.patch
@@ -0,0 +1,96 @@
+From 2f24beca349dd4f2aa5a408e2da5d09e6425583d Mon Sep 17 00:00:00 2001
+From: =?UTF-8?q?Michael=20M=C3=BCller?= <michael@fds-team.de>
+Date: Thu, 9 Feb 2017 21:11:25 +0100
+Subject: wined3d: Use a separate lock for each CS list.
+
+---
+ dlls/wined3d/cs.c              | 26 +++++++++++---------------
+ dlls/wined3d/wined3d_private.h |  1 +
+ 2 files changed, 12 insertions(+), 15 deletions(-)
+
+diff --git a/dlls/wined3d/cs.c b/dlls/wined3d/cs.c
+index f68bbd55663..09936328cf4 100644
+--- a/dlls/wined3d/cs.c
++++ b/dlls/wined3d/cs.c
+@@ -25,16 +25,6 @@ WINE_DEFAULT_DEBUG_CHANNEL(d3d);
+ 
+ #define WINED3D_INITIAL_CS_SIZE 4096
+ 
+-static CRITICAL_SECTION wined3d_cs_list_mutex;
+-static CRITICAL_SECTION_DEBUG wined3d_cs_list_mutex_debug =
+-{
+-    0, 0, &wined3d_cs_list_mutex,
+-    {&wined3d_cs_list_mutex_debug.ProcessLocksList,
+-    &wined3d_cs_list_mutex_debug.ProcessLocksList},
+-    0, 0, {(DWORD_PTR)(__FILE__ ": wined3d_cs_list_mutex")}
+-};
+-static CRITICAL_SECTION wined3d_cs_list_mutex = {&wined3d_cs_list_mutex_debug, -1, 0, 0, 0, 0};
+-
+ enum wined3d_cs_op
+ {
+     WINED3D_CS_OP_SYNC,
+@@ -539,9 +529,9 @@ static void wined3d_cs_mt_submit(struct wined3d_cs *cs);
+ /* FIXME: The list synchronization probably isn't particularly fast. */
+ static void wined3d_cs_list_enqueue(struct wined3d_cs_list *list, struct wined3d_cs_block *block)
+ {
+-    EnterCriticalSection(&wined3d_cs_list_mutex);
++    EnterCriticalSection(&list->lock);
+     list_add_tail(&list->blocks, &block->entry);
+-    LeaveCriticalSection(&wined3d_cs_list_mutex);
++    LeaveCriticalSection(&list->lock);
+     InterlockedIncrement(&list->count);
+ }
+ 
+@@ -550,14 +540,14 @@ static struct wined3d_cs_block *wined3d_cs_list_dequeue(struct wined3d_cs_list *
+     struct list *head;
+ 
+     if (!list->count) return NULL;
+-    EnterCriticalSection(&wined3d_cs_list_mutex);
++    EnterCriticalSection(&list->lock);
+     if (!(head = list_head(&list->blocks)))
+     {
+-        LeaveCriticalSection(&wined3d_cs_list_mutex);
++        LeaveCriticalSection(&list->lock);
+         return NULL;
+     }
+     list_remove(head);
+-    LeaveCriticalSection(&wined3d_cs_list_mutex);
++    LeaveCriticalSection(&list->lock);
+     InterlockedDecrement(&list->count);
+ 
+     return LIST_ENTRY(head, struct wined3d_cs_block, entry);
+@@ -610,6 +600,9 @@ static struct wined3d_cs_block *wined3d_cs_dequeue_block(struct wined3d_cs *cs)
+ 
+ static void wined3d_cs_list_init(struct wined3d_cs_list *list)
+ {
++    InitializeCriticalSection(&list->lock);
++    list->lock.DebugInfo->Spare[0] = (DWORD_PTR)(__FILE__ ": wined3d_cs_list_lock");
++
+     list_init(&list->blocks);
+ }
+ 
+@@ -622,6 +615,9 @@ static void wined3d_cs_list_cleanup(struct wined3d_cs_list *list)
+         list_remove(&block->entry);
+         HeapFree(GetProcessHeap(), 0, block);
+     }
++
++    list->lock.DebugInfo->Spare[0] = 0;
++    DeleteCriticalSection(&list->lock);
+ }
+ 
+ static struct wined3d_cs_block *wined3d_cs_get_block(struct wined3d_cs *cs, struct wined3d_cs_list *list)
+diff --git a/dlls/wined3d/wined3d_private.h b/dlls/wined3d/wined3d_private.h
+index e6b0d363cdd..d4f05b204fd 100644
+--- a/dlls/wined3d/wined3d_private.h
++++ b/dlls/wined3d/wined3d_private.h
+@@ -3207,6 +3207,7 @@ enum wined3d_push_constants
+ 
+ struct wined3d_cs_list
+ {
++    CRITICAL_SECTION lock;
+     struct list blocks;
+     LONG count;
+ };
+-- 
+2.11.0
+
diff --git a/patches/wined3d-CSMT_Main/0053-wined3d-Don-t-crash-in-context_release-when-device-w.patch b/patches/wined3d-CSMT_Main/0053-wined3d-Don-t-crash-in-context_release-when-device-w.patch
new file mode 100644
index 00000000..c4e8e621
--- /dev/null
+++ b/patches/wined3d-CSMT_Main/0053-wined3d-Don-t-crash-in-context_release-when-device-w.patch
@@ -0,0 +1,30 @@
+From 9c3420b9d0fd92a9421e0b15177d4fb8b5a55601 Mon Sep 17 00:00:00 2001
+From: Sebastian Lackner <sebastian@fds-team.de>
+Date: Fri, 10 Feb 2017 20:31:21 +0100
+Subject: wined3d: Don't crash in context_release when device was not fully
+ created.
+
+---
+ dlls/wined3d/context.c | 5 ++++-
+ 1 file changed, 4 insertions(+), 1 deletion(-)
+
+diff --git a/dlls/wined3d/context.c b/dlls/wined3d/context.c
+index 1979a36d875..be5194daaf0 100644
+--- a/dlls/wined3d/context.c
++++ b/dlls/wined3d/context.c
+@@ -1353,8 +1353,11 @@ void context_release(struct wined3d_context *context)
+             WARN("Context %p is not the current context.\n", context);
+     }
+ 
+-    if (wined3d_settings.cs_multithreaded && context->device->cs->thread_id != GetCurrentThreadId())
++    if (wined3d_settings.cs_multithreaded && context->device &&
++        context->device->cs->thread_id != GetCurrentThreadId())
++    {
+         context->gl_info->gl_ops.gl.p_glFinish();
++    }
+ 
+     if (!--context->level)
+     {
+-- 
+2.11.0
+
diff --git a/patches/wined3d-CSMT_Main/9999-IfDefined.patch b/patches/wined3d-CSMT_Main/9999-IfDefined.patch
index 8afed4ab..fdee660a 100644
--- a/patches/wined3d-CSMT_Main/9999-IfDefined.patch
+++ b/patches/wined3d-CSMT_Main/9999-IfDefined.patch
@@ -53,19 +53,22 @@ diff --git a/dlls/wined3d/arb_program_shader.c b/dlls/wined3d/arb_program_shader
 diff --git a/dlls/wined3d/context.c b/dlls/wined3d/context.c
 --- a/dlls/wined3d/context.c
 +++ b/dlls/wined3d/context.c
-@@ -1353,6 +1353,11 @@ void context_release(struct wined3d_context *context)
+@@ -1353,6 +1353,14 @@ void context_release(struct wined3d_context *context)
              WARN("Context %p is not the current context.\n", context);
      }
  
 +#if defined(STAGING_CSMT)
-+    if (wined3d_settings.cs_multithreaded && context->device->cs->thread_id != GetCurrentThreadId())
++    if (wined3d_settings.cs_multithreaded && context->device &&
++        context->device->cs->thread_id != GetCurrentThreadId())
++    {
 +        context->gl_info->gl_ops.gl.p_glFinish();
++    }
 +
 +#endif /* STAGING_CSMT */
      if (!--context->level)
      {
          if (context_restore_pixel_format(context))
-@@ -1723,6 +1728,13 @@ struct wined3d_context *context_create(struct wined3d_swapchain *swapchain,
+@@ -1723,6 +1731,13 @@ struct wined3d_context *context_create(struct wined3d_swapchain *swapchain,
          goto out;
      }
  
@@ -79,7 +82,7 @@ diff --git a/dlls/wined3d/context.c b/dlls/wined3d/context.c
      /* Initialize the texture unit mapping to a 1:1 mapping */
      for (s = 0; s < MAX_COMBINED_SAMPLERS; ++s)
      {
-@@ -2044,6 +2056,9 @@ out:
+@@ -2044,6 +2059,9 @@ out:
      if (hdc) wined3d_release_dc(swapchain->win_handle, hdc);
      device->shader_backend->shader_free_context_data(ret);
      device->adapter->fragment_pipe->free_context_data(ret);
@@ -89,7 +92,7 @@ diff --git a/dlls/wined3d/context.c b/dlls/wined3d/context.c
      HeapFree(GetProcessHeap(), 0, ret->free_event_queries);
      HeapFree(GetProcessHeap(), 0, ret->free_occlusion_queries);
      HeapFree(GetProcessHeap(), 0, ret->free_timestamp_queries);
-@@ -2091,6 +2106,9 @@ void context_destroy(struct wined3d_device *device, struct wined3d_context *cont
+@@ -2091,6 +2109,9 @@ void context_destroy(struct wined3d_device *device, struct wined3d_context *cont
      device->shader_backend->shader_free_context_data(context);
      device->adapter->fragment_pipe->free_context_data(context);
      HeapFree(GetProcessHeap(), 0, context->fbo_key);
@@ -99,7 +102,7 @@ diff --git a/dlls/wined3d/context.c b/dlls/wined3d/context.c
      HeapFree(GetProcessHeap(), 0, context->draw_buffers);
      HeapFree(GetProcessHeap(), 0, context->blit_targets);
      device_context_remove(device, context);
-@@ -2640,7 +2658,11 @@ static BOOL context_validate_rt_config(UINT rt_count, struct wined3d_rendertarge
+@@ -2640,7 +2661,11 @@ static BOOL context_validate_rt_config(UINT rt_count, struct wined3d_rendertarge
  }
  
  /* Context activation is done by the caller. */
@@ -111,7 +114,7 @@ diff --git a/dlls/wined3d/context.c b/dlls/wined3d/context.c
          UINT rt_count, const struct wined3d_fb_state *fb)
  {
      struct wined3d_rendertarget_view **rts = fb->render_targets;
-@@ -2649,7 +2671,11 @@ BOOL context_apply_clear_state(struct wined3d_context *context, const struct win
+@@ -2649,7 +2674,11 @@ BOOL context_apply_clear_state(struct wined3d_context *context, const struct win
      DWORD rt_mask = 0, *cur_mask;
      UINT i;
  
@@ -123,7 +126,7 @@ diff --git a/dlls/wined3d/context.c b/dlls/wined3d/context.c
              || rt_count != gl_info->limits.buffers)
      {
          if (!context_validate_rt_config(rt_count, rts, dsv))
-@@ -2694,6 +2720,10 @@ BOOL context_apply_clear_state(struct wined3d_context *context, const struct win
+@@ -2694,6 +2723,10 @@ BOOL context_apply_clear_state(struct wined3d_context *context, const struct win
              rt_mask = context_generate_rt_mask_no_fbo(context,
                      rt_count ? wined3d_rendertarget_view_get_surface(rts[0])->container : NULL);
          }
@@ -134,7 +137,7 @@ diff --git a/dlls/wined3d/context.c b/dlls/wined3d/context.c
      }
      else if (wined3d_settings.offscreen_rendering_mode == ORM_FBO
              && (!rt_count || wined3d_resource_is_offscreen(rts[0]->resource)))
-@@ -2733,7 +2763,12 @@ BOOL context_apply_clear_state(struct wined3d_context *context, const struct win
+@@ -2733,7 +2766,12 @@ BOOL context_apply_clear_state(struct wined3d_context *context, const struct win
      gl_info->gl_ops.gl.p_glEnable(GL_SCISSOR_TEST);
      if (rt_count && gl_info->supported[ARB_FRAMEBUFFER_SRGB])
      {
@@ -147,7 +150,7 @@ diff --git a/dlls/wined3d/context.c b/dlls/wined3d/context.c
              gl_info->gl_ops.gl.p_glEnable(GL_FRAMEBUFFER_SRGB);
          else
              gl_info->gl_ops.gl.p_glDisable(GL_FRAMEBUFFER_SRGB);
-@@ -2750,7 +2785,11 @@ BOOL context_apply_clear_state(struct wined3d_context *context, const struct win
+@@ -2750,7 +2788,11 @@ BOOL context_apply_clear_state(struct wined3d_context *context, const struct win
  
  static DWORD find_draw_buffers_mask(const struct wined3d_context *context, const struct wined3d_state *state)
  {
@@ -159,7 +162,7 @@ diff --git a/dlls/wined3d/context.c b/dlls/wined3d/context.c
      struct wined3d_shader *ps = state->shader[WINED3D_SHADER_TYPE_PIXEL];
      DWORD rt_mask, rt_mask_bits;
      unsigned int i;
-@@ -2780,7 +2819,11 @@ static DWORD find_draw_buffers_mask(const struct wined3d_context *context, const
+@@ -2780,7 +2822,11 @@ static DWORD find_draw_buffers_mask(const struct wined3d_context *context, const
  void context_state_fb(struct wined3d_context *context, const struct wined3d_state *state, DWORD state_id)
  {
      DWORD rt_mask = find_draw_buffers_mask(context, state);
@@ -171,7 +174,7 @@ diff --git a/dlls/wined3d/context.c b/dlls/wined3d/context.c
      DWORD *cur_mask;
  
      if (wined3d_settings.offscreen_rendering_mode == ORM_FBO)
-@@ -3069,6 +3112,10 @@ void context_state_drawbuf(struct wined3d_context *context, const struct wined3d
+@@ -3069,6 +3115,10 @@ void context_state_drawbuf(struct wined3d_context *context, const struct wined3d
          context_apply_draw_buffers(context, rt_mask);
          *cur_mask = rt_mask;
      }
@@ -182,7 +185,7 @@ diff --git a/dlls/wined3d/context.c b/dlls/wined3d/context.c
  }
  
  static BOOL fixed_get_input(BYTE usage, BYTE usage_idx, unsigned int *regnum)
-@@ -3503,7 +3550,11 @@ BOOL context_apply_draw_state(struct wined3d_context *context,
+@@ -3503,7 +3553,11 @@ BOOL context_apply_draw_state(struct wined3d_context *context,
          const struct wined3d_device *device, const struct wined3d_state *state)
  {
      const struct StateEntry *state_table = context->state_table;
@@ -194,7 +197,7 @@ diff --git a/dlls/wined3d/context.c b/dlls/wined3d/context.c
      unsigned int i;
      WORD map;
  
-@@ -3705,6 +3756,15 @@ struct wined3d_context *context_acquire(const struct wined3d_device *device, str
+@@ -3705,6 +3759,15 @@ struct wined3d_context *context_acquire(const struct wined3d_device *device, str
  
      TRACE("device %p, target %p.\n", device, target);
  
@@ -213,7 +216,7 @@ diff --git a/dlls/wined3d/context.c b/dlls/wined3d/context.c
 diff --git a/dlls/wined3d/cs.c b/dlls/wined3d/cs.c
 --- a/dlls/wined3d/cs.c
 +++ b/dlls/wined3d/cs.c
-@@ -18,14 +18,33 @@
+@@ -18,6 +18,9 @@
  
  #include "config.h"
  #include "wine/port.h"
@@ -223,31 +226,17 @@ diff --git a/dlls/wined3d/cs.c b/dlls/wined3d/cs.c
  #include "wined3d_private.h"
  
  WINE_DEFAULT_DEBUG_CHANNEL(d3d);
+@@ -26,6 +29,9 @@ WINE_DEFAULT_DEBUG_CHANNEL(d3d);
  
- #define WINED3D_INITIAL_CS_SIZE 4096
- 
-+#if !defined(STAGING_CSMT)
  enum wined3d_cs_op
  {
-+#else  /* STAGING_CSMT */
-+static CRITICAL_SECTION wined3d_cs_list_mutex;
-+static CRITICAL_SECTION_DEBUG wined3d_cs_list_mutex_debug =
-+{
-+    0, 0, &wined3d_cs_list_mutex,
-+    {&wined3d_cs_list_mutex_debug.ProcessLocksList,
-+    &wined3d_cs_list_mutex_debug.ProcessLocksList},
-+    0, 0, {(DWORD_PTR)(__FILE__ ": wined3d_cs_list_mutex")}
-+};
-+static CRITICAL_SECTION wined3d_cs_list_mutex = {&wined3d_cs_list_mutex_debug, -1, 0, 0, 0, 0};
-+
-+enum wined3d_cs_op
-+{
++#if defined(STAGING_CSMT)
 +    WINED3D_CS_OP_SYNC,
 +#endif /* STAGING_CSMT */
      WINED3D_CS_OP_PRESENT,
      WINED3D_CS_OP_CLEAR,
      WINED3D_CS_OP_DISPATCH,
-@@ -57,10 +76,37 @@ enum wined3d_cs_op
+@@ -57,10 +63,37 @@ enum wined3d_cs_op
      WINED3D_CS_OP_RESET_STATE,
      WINED3D_CS_OP_DESTROY_OBJECT,
      WINED3D_CS_OP_QUERY_ISSUE,
@@ -285,7 +274,7 @@ diff --git a/dlls/wined3d/cs.c b/dlls/wined3d/cs.c
  };
  
  struct wined3d_cs_present
-@@ -297,6 +343,16 @@ struct wined3d_cs_query_issue
+@@ -297,6 +330,16 @@ struct wined3d_cs_query_issue
      DWORD flags;
  };
  
@@ -302,7 +291,7 @@ diff --git a/dlls/wined3d/cs.c b/dlls/wined3d/cs.c
  struct wined3d_cs_preload_resource
  {
      enum wined3d_cs_op opcode;
-@@ -328,7 +384,299 @@ struct wined3d_cs_unmap
+@@ -328,7 +371,305 @@ struct wined3d_cs_unmap
      HRESULT *hr;
  };
  
@@ -477,9 +466,9 @@ diff --git a/dlls/wined3d/cs.c b/dlls/wined3d/cs.c
 +/* FIXME: The list synchronization probably isn't particularly fast. */
 +static void wined3d_cs_list_enqueue(struct wined3d_cs_list *list, struct wined3d_cs_block *block)
 +{
-+    EnterCriticalSection(&wined3d_cs_list_mutex);
++    EnterCriticalSection(&list->lock);
 +    list_add_tail(&list->blocks, &block->entry);
-+    LeaveCriticalSection(&wined3d_cs_list_mutex);
++    LeaveCriticalSection(&list->lock);
 +    InterlockedIncrement(&list->count);
 +}
 +
@@ -488,14 +477,14 @@ diff --git a/dlls/wined3d/cs.c b/dlls/wined3d/cs.c
 +    struct list *head;
 +
 +    if (!list->count) return NULL;
-+    EnterCriticalSection(&wined3d_cs_list_mutex);
++    EnterCriticalSection(&list->lock);
 +    if (!(head = list_head(&list->blocks)))
 +    {
-+        LeaveCriticalSection(&wined3d_cs_list_mutex);
++        LeaveCriticalSection(&list->lock);
 +        return NULL;
 +    }
 +    list_remove(head);
-+    LeaveCriticalSection(&wined3d_cs_list_mutex);
++    LeaveCriticalSection(&list->lock);
 +    InterlockedDecrement(&list->count);
 +
 +    return LIST_ENTRY(head, struct wined3d_cs_block, entry);
@@ -548,6 +537,9 @@ diff --git a/dlls/wined3d/cs.c b/dlls/wined3d/cs.c
 +
 +static void wined3d_cs_list_init(struct wined3d_cs_list *list)
 +{
++    InitializeCriticalSection(&list->lock);
++    list->lock.DebugInfo->Spare[0] = (DWORD_PTR)(__FILE__ ": wined3d_cs_list_lock");
++
 +    list_init(&list->blocks);
 +}
 +
@@ -560,6 +552,9 @@ diff --git a/dlls/wined3d/cs.c b/dlls/wined3d/cs.c
 +        list_remove(&block->entry);
 +        HeapFree(GetProcessHeap(), 0, block);
 +    }
++
++    list->lock.DebugInfo->Spare[0] = 0;
++    DeleteCriticalSection(&list->lock);
 +}
 +
 +static struct wined3d_cs_block *wined3d_cs_get_block(struct wined3d_cs *cs, struct wined3d_cs_list *list)
@@ -602,7 +597,7 @@ diff --git a/dlls/wined3d/cs.c b/dlls/wined3d/cs.c
  {
      const struct wined3d_cs_present *op = data;
      struct wined3d_swapchain *swapchain;
-@@ -337,13 +685,24 @@ static void wined3d_cs_exec_present(struct wined3d_cs *cs, const void *data)
+@@ -337,13 +678,24 @@ static void wined3d_cs_exec_present(struct wined3d_cs *cs, const void *data)
      swapchain = op->swapchain;
      wined3d_swapchain_set_window(swapchain, op->dst_window_override);
  
@@ -627,7 +622,7 @@ diff --git a/dlls/wined3d/cs.c b/dlls/wined3d/cs.c
  }
  
  void wined3d_cs_emit_present(struct wined3d_cs *cs, struct wined3d_swapchain *swapchain,
-@@ -351,8 +710,14 @@ void wined3d_cs_emit_present(struct wined3d_cs *cs, struct wined3d_swapchain *sw
+@@ -351,8 +703,14 @@ void wined3d_cs_emit_present(struct wined3d_cs *cs, struct wined3d_swapchain *sw
  {
      struct wined3d_cs_present *op;
      unsigned int i;
@@ -642,7 +637,7 @@ diff --git a/dlls/wined3d/cs.c b/dlls/wined3d/cs.c
      op->opcode = WINED3D_CS_OP_PRESENT;
      op->dst_window_override = dst_window_override;
      op->swapchain = swapchain;
-@@ -366,34 +731,75 @@ void wined3d_cs_emit_present(struct wined3d_cs *cs, struct wined3d_swapchain *sw
+@@ -366,34 +724,75 @@ void wined3d_cs_emit_present(struct wined3d_cs *cs, struct wined3d_swapchain *sw
          wined3d_resource_acquire(&swapchain->back_buffers[i]->resource);
      }
  
@@ -718,7 +713,7 @@ diff --git a/dlls/wined3d/cs.c b/dlls/wined3d/cs.c
  }
  
  void wined3d_cs_emit_clear(struct wined3d_cs *cs, DWORD rect_count, const RECT *rects,
-@@ -403,7 +809,11 @@ void wined3d_cs_emit_clear(struct wined3d_cs *cs, DWORD rect_count, const RECT *
+@@ -403,7 +802,11 @@ void wined3d_cs_emit_clear(struct wined3d_cs *cs, DWORD rect_count, const RECT *
      struct wined3d_cs_clear *op;
      unsigned int i;
  
@@ -730,7 +725,7 @@ diff --git a/dlls/wined3d/cs.c b/dlls/wined3d/cs.c
      op->opcode = WINED3D_CS_OP_CLEAR;
      op->flags = flags;
      op->color = *color;
-@@ -416,12 +826,21 @@ void wined3d_cs_emit_clear(struct wined3d_cs *cs, DWORD rect_count, const RECT *
+@@ -416,12 +819,21 @@ void wined3d_cs_emit_clear(struct wined3d_cs *cs, DWORD rect_count, const RECT *
      {
          for (i = 0; i < cs->device->adapter->gl_info.limits.buffers; ++i)
          {
@@ -752,7 +747,7 @@ diff --git a/dlls/wined3d/cs.c b/dlls/wined3d/cs.c
  
      cs->ops->submit(cs);
  }
-@@ -532,9 +951,15 @@ static void release_unordered_access_resources(const struct wined3d_shader *shad
+@@ -532,9 +944,15 @@ static void release_unordered_access_resources(const struct wined3d_shader *shad
      }
  }
  
@@ -768,7 +763,7 @@ diff --git a/dlls/wined3d/cs.c b/dlls/wined3d/cs.c
      const struct wined3d_cs_dispatch *op = data;
  
      dispatch_compute(cs->device, state,
-@@ -543,6 +968,10 @@ static void wined3d_cs_exec_dispatch(struct wined3d_cs *cs, const void *data)
+@@ -543,6 +961,10 @@ static void wined3d_cs_exec_dispatch(struct wined3d_cs *cs, const void *data)
      release_shader_resources(state, 1u << WINED3D_SHADER_TYPE_COMPUTE);
      release_unordered_access_resources(state->shader[WINED3D_SHADER_TYPE_COMPUTE],
              state->unordered_access_view[WINED3D_PIPELINE_COMPUTE]);
@@ -779,7 +774,7 @@ diff --git a/dlls/wined3d/cs.c b/dlls/wined3d/cs.c
  }
  
  void wined3d_cs_emit_dispatch(struct wined3d_cs *cs,
-@@ -551,7 +980,11 @@ void wined3d_cs_emit_dispatch(struct wined3d_cs *cs,
+@@ -551,7 +973,11 @@ void wined3d_cs_emit_dispatch(struct wined3d_cs *cs,
      const struct wined3d_state *state = &cs->device->state;
      struct wined3d_cs_dispatch *op;
  
@@ -791,7 +786,7 @@ diff --git a/dlls/wined3d/cs.c b/dlls/wined3d/cs.c
      op->opcode = WINED3D_CS_OP_DISPATCH;
      op->group_count_x = group_count_x;
      op->group_count_y = group_count_y;
-@@ -564,9 +997,15 @@ void wined3d_cs_emit_dispatch(struct wined3d_cs *cs,
+@@ -564,9 +990,15 @@ void wined3d_cs_emit_dispatch(struct wined3d_cs *cs,
      cs->ops->submit(cs);
  }
  
@@ -807,7 +802,7 @@ diff --git a/dlls/wined3d/cs.c b/dlls/wined3d/cs.c
      const struct wined3d_cs_draw *op = data;
      unsigned int i;
  
-@@ -594,14 +1033,26 @@ static void wined3d_cs_exec_draw(struct wined3d_cs *cs, const void *data)
+@@ -594,14 +1026,26 @@ static void wined3d_cs_exec_draw(struct wined3d_cs *cs, const void *data)
      }
      for (i = 0; i < cs->device->adapter->gl_info.limits.buffers; ++i)
      {
@@ -834,7 +829,7 @@ diff --git a/dlls/wined3d/cs.c b/dlls/wined3d/cs.c
  }
  
  void wined3d_cs_emit_draw(struct wined3d_cs *cs, int base_vertex_idx, unsigned int start_idx,
-@@ -611,7 +1062,11 @@ void wined3d_cs_emit_draw(struct wined3d_cs *cs, int base_vertex_idx, unsigned i
+@@ -611,7 +1055,11 @@ void wined3d_cs_emit_draw(struct wined3d_cs *cs, int base_vertex_idx, unsigned i
      struct wined3d_cs_draw *op;
      unsigned int i;
  
@@ -846,7 +841,7 @@ diff --git a/dlls/wined3d/cs.c b/dlls/wined3d/cs.c
      op->opcode = WINED3D_CS_OP_DRAW;
      op->base_vertex_idx = base_vertex_idx;
      op->start_idx = start_idx;
-@@ -634,11 +1089,19 @@ void wined3d_cs_emit_draw(struct wined3d_cs *cs, int base_vertex_idx, unsigned i
+@@ -634,11 +1082,19 @@ void wined3d_cs_emit_draw(struct wined3d_cs *cs, int base_vertex_idx, unsigned i
      }
      for (i = 0; i < cs->device->adapter->gl_info.limits.buffers; ++i)
      {
@@ -866,7 +861,7 @@ diff --git a/dlls/wined3d/cs.c b/dlls/wined3d/cs.c
      acquire_shader_resources(state, ~(1u << WINED3D_SHADER_TYPE_COMPUTE));
      acquire_unordered_access_resources(state->shader[WINED3D_SHADER_TYPE_PIXEL],
              state->unordered_access_view[WINED3D_PIPELINE_GRAPHICS]);
-@@ -646,38 +1109,61 @@ void wined3d_cs_emit_draw(struct wined3d_cs *cs, int base_vertex_idx, unsigned i
+@@ -646,38 +1102,61 @@ void wined3d_cs_emit_draw(struct wined3d_cs *cs, int base_vertex_idx, unsigned i
      cs->ops->submit(cs);
  }
  
@@ -928,7 +923,7 @@ diff --git a/dlls/wined3d/cs.c b/dlls/wined3d/cs.c
      op = cs->ops->require_space(cs, sizeof(*op));
      op->opcode = WINED3D_CS_OP_SET_VIEWPORT;
      op->viewport = *viewport;
-@@ -686,17 +1172,30 @@ void wined3d_cs_emit_set_viewport(struct wined3d_cs *cs, const struct wined3d_vi
+@@ -686,17 +1165,30 @@ void wined3d_cs_emit_set_viewport(struct wined3d_cs *cs, const struct wined3d_vi
  }
  
  static void wined3d_cs_exec_set_scissor_rect(struct wined3d_cs *cs, const void *data)
@@ -959,7 +954,7 @@ diff --git a/dlls/wined3d/cs.c b/dlls/wined3d/cs.c
      op = cs->ops->require_space(cs, sizeof(*op));
      op->opcode = WINED3D_CS_OP_SET_SCISSOR_RECT;
      op->rect = *rect;
-@@ -705,11 +1204,26 @@ void wined3d_cs_emit_set_scissor_rect(struct wined3d_cs *cs, const RECT *rect)
+@@ -705,11 +1197,26 @@ void wined3d_cs_emit_set_scissor_rect(struct wined3d_cs *cs, const RECT *rect)
  }
  
  static void wined3d_cs_exec_set_rendertarget_view(struct wined3d_cs *cs, const void *data)
@@ -986,7 +981,7 @@ diff --git a/dlls/wined3d/cs.c b/dlls/wined3d/cs.c
  }
  
  void wined3d_cs_emit_set_rendertarget_view(struct wined3d_cs *cs, unsigned int view_idx,
-@@ -717,21 +1231,35 @@ void wined3d_cs_emit_set_rendertarget_view(struct wined3d_cs *cs, unsigned int v
+@@ -717,21 +1224,35 @@ void wined3d_cs_emit_set_rendertarget_view(struct wined3d_cs *cs, unsigned int v
  {
      struct wined3d_cs_set_rendertarget_view *op;
  
@@ -1022,7 +1017,7 @@ diff --git a/dlls/wined3d/cs.c b/dlls/wined3d/cs.c
      {
          struct wined3d_surface *prev_surface = wined3d_rendertarget_view_get_surface(prev);
  
-@@ -743,7 +1271,11 @@ static void wined3d_cs_exec_set_depth_stencil_view(struct wined3d_cs *cs, const
+@@ -743,7 +1264,11 @@ static void wined3d_cs_exec_set_depth_stencil_view(struct wined3d_cs *cs, const
          }
      }
  
@@ -1034,7 +1029,7 @@ diff --git a/dlls/wined3d/cs.c b/dlls/wined3d/cs.c
  
      if (!prev != !op->view)
      {
-@@ -760,12 +1292,17 @@ static void wined3d_cs_exec_set_depth_stencil_view(struct wined3d_cs *cs, const
+@@ -760,12 +1285,17 @@ static void wined3d_cs_exec_set_depth_stencil_view(struct wined3d_cs *cs, const
      }
  
      device_invalidate_state(device, STATE_FRAMEBUFFER);
@@ -1052,7 +1047,7 @@ diff --git a/dlls/wined3d/cs.c b/dlls/wined3d/cs.c
      op = cs->ops->require_space(cs, sizeof(*op));
      op->opcode = WINED3D_CS_OP_SET_DEPTH_STENCIL_VIEW;
      op->view = view;
-@@ -774,17 +1311,30 @@ void wined3d_cs_emit_set_depth_stencil_view(struct wined3d_cs *cs, struct wined3
+@@ -774,17 +1304,30 @@ void wined3d_cs_emit_set_depth_stencil_view(struct wined3d_cs *cs, struct wined3
  }
  
  static void wined3d_cs_exec_set_vertex_declaration(struct wined3d_cs *cs, const void *data)
@@ -1083,7 +1078,7 @@ diff --git a/dlls/wined3d/cs.c b/dlls/wined3d/cs.c
      op = cs->ops->require_space(cs, sizeof(*op));
      op->opcode = WINED3D_CS_OP_SET_VERTEX_DECLARATION;
      op->declaration = declaration;
-@@ -793,6 +1343,14 @@ void wined3d_cs_emit_set_vertex_declaration(struct wined3d_cs *cs, struct wined3
+@@ -793,6 +1336,14 @@ void wined3d_cs_emit_set_vertex_declaration(struct wined3d_cs *cs, struct wined3
  }
  
  static void wined3d_cs_exec_set_stream_source(struct wined3d_cs *cs, const void *data)
@@ -1098,7 +1093,7 @@ diff --git a/dlls/wined3d/cs.c b/dlls/wined3d/cs.c
  {
      const struct wined3d_cs_set_stream_source *op = data;
      struct wined3d_stream_state *stream;
-@@ -810,6 +1368,10 @@ static void wined3d_cs_exec_set_stream_source(struct wined3d_cs *cs, const void
+@@ -810,6 +1361,10 @@ static void wined3d_cs_exec_set_stream_source(struct wined3d_cs *cs, const void
          InterlockedDecrement(&prev->resource.bind_count);
  
      device_invalidate_state(cs->device, STATE_STREAMSRC);
@@ -1109,7 +1104,7 @@ diff --git a/dlls/wined3d/cs.c b/dlls/wined3d/cs.c
  }
  
  void wined3d_cs_emit_set_stream_source(struct wined3d_cs *cs, UINT stream_idx,
-@@ -817,17 +1379,27 @@ void wined3d_cs_emit_set_stream_source(struct wined3d_cs *cs, UINT stream_idx,
+@@ -817,17 +1372,27 @@ void wined3d_cs_emit_set_stream_source(struct wined3d_cs *cs, UINT stream_idx,
  {
      struct wined3d_cs_set_stream_source *op;
  
@@ -1137,7 +1132,7 @@ diff --git a/dlls/wined3d/cs.c b/dlls/wined3d/cs.c
  {
      const struct wined3d_cs_set_stream_source_freq *op = data;
      struct wined3d_stream_state *stream;
-@@ -837,22 +1409,36 @@ static void wined3d_cs_exec_set_stream_source_freq(struct wined3d_cs *cs, const
+@@ -837,22 +1402,36 @@ static void wined3d_cs_exec_set_stream_source_freq(struct wined3d_cs *cs, const
      stream->flags = op->flags;
  
      device_invalidate_state(cs->device, STATE_STREAMSRC);
@@ -1174,7 +1169,7 @@ diff --git a/dlls/wined3d/cs.c b/dlls/wined3d/cs.c
  {
      const struct wined3d_cs_set_stream_output *op = data;
      struct wined3d_stream_output *stream;
-@@ -867,6 +1453,10 @@ static void wined3d_cs_exec_set_stream_output(struct wined3d_cs *cs, const void
+@@ -867,6 +1446,10 @@ static void wined3d_cs_exec_set_stream_output(struct wined3d_cs *cs, const void
          InterlockedIncrement(&op->buffer->resource.bind_count);
      if (prev)
          InterlockedDecrement(&prev->resource.bind_count);
@@ -1185,7 +1180,7 @@ diff --git a/dlls/wined3d/cs.c b/dlls/wined3d/cs.c
  }
  
  void wined3d_cs_emit_set_stream_output(struct wined3d_cs *cs, UINT stream_idx,
-@@ -874,16 +1464,26 @@ void wined3d_cs_emit_set_stream_output(struct wined3d_cs *cs, UINT stream_idx,
+@@ -874,16 +1457,26 @@ void wined3d_cs_emit_set_stream_output(struct wined3d_cs *cs, UINT stream_idx,
  {
      struct wined3d_cs_set_stream_output *op;
  
@@ -1212,7 +1207,7 @@ diff --git a/dlls/wined3d/cs.c b/dlls/wined3d/cs.c
  {
      const struct wined3d_cs_set_index_buffer *op = data;
      struct wined3d_buffer *prev;
-@@ -899,6 +1499,10 @@ static void wined3d_cs_exec_set_index_buffer(struct wined3d_cs *cs, const void *
+@@ -899,6 +1492,10 @@ static void wined3d_cs_exec_set_index_buffer(struct wined3d_cs *cs, const void *
          InterlockedDecrement(&prev->resource.bind_count);
  
      device_invalidate_state(cs->device, STATE_INDEXBUFFER);
@@ -1223,7 +1218,7 @@ diff --git a/dlls/wined3d/cs.c b/dlls/wined3d/cs.c
  }
  
  void wined3d_cs_emit_set_index_buffer(struct wined3d_cs *cs, struct wined3d_buffer *buffer,
-@@ -906,16 +1510,26 @@ void wined3d_cs_emit_set_index_buffer(struct wined3d_cs *cs, struct wined3d_buff
+@@ -906,16 +1503,26 @@ void wined3d_cs_emit_set_index_buffer(struct wined3d_cs *cs, struct wined3d_buff
  {
      struct wined3d_cs_set_index_buffer *op;
  
@@ -1250,7 +1245,7 @@ diff --git a/dlls/wined3d/cs.c b/dlls/wined3d/cs.c
  {
      const struct wined3d_cs_set_constant_buffer *op = data;
      struct wined3d_buffer *prev;
-@@ -929,6 +1543,9 @@ static void wined3d_cs_exec_set_constant_buffer(struct wined3d_cs *cs, const voi
+@@ -929,6 +1536,9 @@ static void wined3d_cs_exec_set_constant_buffer(struct wined3d_cs *cs, const voi
          InterlockedDecrement(&prev->resource.bind_count);
  
      device_invalidate_state(cs->device, STATE_CONSTANT_BUFFER(op->type));
@@ -1260,7 +1255,7 @@ diff --git a/dlls/wined3d/cs.c b/dlls/wined3d/cs.c
  }
  
  void wined3d_cs_emit_set_constant_buffer(struct wined3d_cs *cs, enum wined3d_shader_type type,
-@@ -936,16 +1553,26 @@ void wined3d_cs_emit_set_constant_buffer(struct wined3d_cs *cs, enum wined3d_sha
+@@ -936,16 +1546,26 @@ void wined3d_cs_emit_set_constant_buffer(struct wined3d_cs *cs, enum wined3d_sha
  {
      struct wined3d_cs_set_constant_buffer *op;
  
@@ -1287,7 +1282,7 @@ diff --git a/dlls/wined3d/cs.c b/dlls/wined3d/cs.c
  {
      const struct wined3d_gl_info *gl_info = &cs->device->adapter->gl_info;
      const struct wined3d_d3d_info *d3d_info = &cs->device->adapter->d3d_info;
-@@ -1022,27 +1649,45 @@ static void wined3d_cs_exec_set_texture(struct wined3d_cs *cs, const void *data)
+@@ -1022,27 +1642,45 @@ static void wined3d_cs_exec_set_texture(struct wined3d_cs *cs, const void *data)
  
      if (new_use_color_key)
          device_invalidate_state(cs->device, STATE_COLOR_KEY);
@@ -1333,7 +1328,7 @@ diff --git a/dlls/wined3d/cs.c b/dlls/wined3d/cs.c
  }
  
  void wined3d_cs_emit_set_shader_resource_view(struct wined3d_cs *cs, enum wined3d_shader_type type,
-@@ -1050,16 +1695,26 @@ void wined3d_cs_emit_set_shader_resource_view(struct wined3d_cs *cs, enum wined3
+@@ -1050,16 +1688,26 @@ void wined3d_cs_emit_set_shader_resource_view(struct wined3d_cs *cs, enum wined3
  {
      struct wined3d_cs_set_shader_resource_view *op;
  
@@ -1360,7 +1355,7 @@ diff --git a/dlls/wined3d/cs.c b/dlls/wined3d/cs.c
  {
      const struct wined3d_cs_set_unordered_access_view *op = data;
      struct wined3d_unordered_access_view *prev;
-@@ -1073,6 +1728,10 @@ static void wined3d_cs_exec_set_unordered_access_view(struct wined3d_cs *cs, con
+@@ -1073,6 +1721,10 @@ static void wined3d_cs_exec_set_unordered_access_view(struct wined3d_cs *cs, con
          InterlockedDecrement(&prev->resource->bind_count);
  
      device_invalidate_state(cs->device, STATE_UNORDERED_ACCESS_VIEW_BINDING(op->pipeline));
@@ -1371,7 +1366,7 @@ diff --git a/dlls/wined3d/cs.c b/dlls/wined3d/cs.c
  }
  
  void wined3d_cs_emit_set_unordered_access_view(struct wined3d_cs *cs, enum wined3d_pipeline pipeline,
-@@ -1080,21 +1739,35 @@ void wined3d_cs_emit_set_unordered_access_view(struct wined3d_cs *cs, enum wined
+@@ -1080,21 +1732,35 @@ void wined3d_cs_emit_set_unordered_access_view(struct wined3d_cs *cs, enum wined
  {
      struct wined3d_cs_set_unordered_access_view *op;
  
@@ -1407,7 +1402,7 @@ diff --git a/dlls/wined3d/cs.c b/dlls/wined3d/cs.c
  }
  
  void wined3d_cs_emit_set_sampler(struct wined3d_cs *cs, enum wined3d_shader_type type,
-@@ -1102,16 +1775,26 @@ void wined3d_cs_emit_set_sampler(struct wined3d_cs *cs, enum wined3d_shader_type
+@@ -1102,16 +1768,26 @@ void wined3d_cs_emit_set_sampler(struct wined3d_cs *cs, enum wined3d_shader_type
  {
      struct wined3d_cs_set_sampler *op;
  
@@ -1434,7 +1429,7 @@ diff --git a/dlls/wined3d/cs.c b/dlls/wined3d/cs.c
  {
      const struct wined3d_cs_set_shader *op = data;
  
-@@ -1119,26 +1802,44 @@ static void wined3d_cs_exec_set_shader(struct wined3d_cs *cs, const void *data)
+@@ -1119,26 +1795,44 @@ static void wined3d_cs_exec_set_shader(struct wined3d_cs *cs, const void *data)
      device_invalidate_state(cs->device, STATE_SHADER(op->type));
      if (op->type != WINED3D_SHADER_TYPE_COMPUTE)
          device_invalidate_state(cs->device, STATE_SHADER_RESOURCE_BINDING);
@@ -1479,7 +1474,7 @@ diff --git a/dlls/wined3d/cs.c b/dlls/wined3d/cs.c
  }
  
  void wined3d_cs_emit_set_rasterizer_state(struct wined3d_cs *cs,
-@@ -1146,6 +1847,7 @@ void wined3d_cs_emit_set_rasterizer_state(struct wined3d_cs *cs,
+@@ -1146,6 +1840,7 @@ void wined3d_cs_emit_set_rasterizer_state(struct wined3d_cs *cs,
  {
      struct wined3d_cs_set_rasterizer_state *op;
  
@@ -1487,7 +1482,7 @@ diff --git a/dlls/wined3d/cs.c b/dlls/wined3d/cs.c
      op = cs->ops->require_space(cs, sizeof(*op));
      op->opcode = WINED3D_CS_OP_SET_RASTERIZER_STATE;
      op->state = rasterizer_state;
-@@ -1154,31 +1856,57 @@ void wined3d_cs_emit_set_rasterizer_state(struct wined3d_cs *cs,
+@@ -1154,31 +1849,57 @@ void wined3d_cs_emit_set_rasterizer_state(struct wined3d_cs *cs,
  }
  
  static void wined3d_cs_exec_set_render_state(struct wined3d_cs *cs, const void *data)
@@ -1545,7 +1540,7 @@ diff --git a/dlls/wined3d/cs.c b/dlls/wined3d/cs.c
  }
  
  void wined3d_cs_emit_set_texture_state(struct wined3d_cs *cs, UINT stage,
-@@ -1186,21 +1914,35 @@ void wined3d_cs_emit_set_texture_state(struct wined3d_cs *cs, UINT stage,
+@@ -1186,21 +1907,35 @@ void wined3d_cs_emit_set_texture_state(struct wined3d_cs *cs, UINT stage,
  {
      struct wined3d_cs_set_texture_state *op;
  
@@ -1581,7 +1576,7 @@ diff --git a/dlls/wined3d/cs.c b/dlls/wined3d/cs.c
  }
  
  void wined3d_cs_emit_set_sampler_state(struct wined3d_cs *cs, UINT sampler_idx,
-@@ -1208,22 +1950,36 @@ void wined3d_cs_emit_set_sampler_state(struct wined3d_cs *cs, UINT sampler_idx,
+@@ -1208,22 +1943,36 @@ void wined3d_cs_emit_set_sampler_state(struct wined3d_cs *cs, UINT sampler_idx,
  {
      struct wined3d_cs_set_sampler_state *op;
  
@@ -1618,7 +1613,7 @@ diff --git a/dlls/wined3d/cs.c b/dlls/wined3d/cs.c
  }
  
  void wined3d_cs_emit_set_transform(struct wined3d_cs *cs, enum wined3d_transform_state state,
-@@ -1231,35 +1987,59 @@ void wined3d_cs_emit_set_transform(struct wined3d_cs *cs, enum wined3d_transform
+@@ -1231,35 +1980,59 @@ void wined3d_cs_emit_set_transform(struct wined3d_cs *cs, enum wined3d_transform
  {
      struct wined3d_cs_set_transform *op;
  
@@ -1678,7 +1673,7 @@ diff --git a/dlls/wined3d/cs.c b/dlls/wined3d/cs.c
  {
      const struct wined3d_cs_set_color_key *op = data;
      struct wined3d_texture *texture = op->texture;
-@@ -1320,6 +2100,10 @@ static void wined3d_cs_exec_set_color_key(struct wined3d_cs *cs, const void *dat
+@@ -1320,6 +2093,10 @@ static void wined3d_cs_exec_set_color_key(struct wined3d_cs *cs, const void *dat
                  break;
          }
      }
@@ -1689,7 +1684,7 @@ diff --git a/dlls/wined3d/cs.c b/dlls/wined3d/cs.c
  }
  
  void wined3d_cs_emit_set_color_key(struct wined3d_cs *cs, struct wined3d_texture *texture,
-@@ -1327,7 +2111,11 @@ void wined3d_cs_emit_set_color_key(struct wined3d_cs *cs, struct wined3d_texture
+@@ -1327,7 +2104,11 @@ void wined3d_cs_emit_set_color_key(struct wined3d_cs *cs, struct wined3d_texture
  {
      struct wined3d_cs_set_color_key *op;
  
@@ -1701,7 +1696,7 @@ diff --git a/dlls/wined3d/cs.c b/dlls/wined3d/cs.c
      op->opcode = WINED3D_CS_OP_SET_COLOR_KEY;
      op->texture = texture;
      op->flags = flags;
-@@ -1338,22 +2126,33 @@ void wined3d_cs_emit_set_color_key(struct wined3d_cs *cs, struct wined3d_texture
+@@ -1338,22 +2119,33 @@ void wined3d_cs_emit_set_color_key(struct wined3d_cs *cs, struct wined3d_texture
      }
      else
          op->set = 0;
@@ -1735,7 +1730,7 @@ diff --git a/dlls/wined3d/cs.c b/dlls/wined3d/cs.c
      op = cs->ops->require_space(cs, sizeof(*op));
      op->opcode = WINED3D_CS_OP_SET_MATERIAL;
      op->material = *material;
-@@ -1364,17 +2163,37 @@ void wined3d_cs_emit_set_material(struct wined3d_cs *cs, const struct wined3d_ma
+@@ -1364,17 +2156,37 @@ void wined3d_cs_emit_set_material(struct wined3d_cs *cs, const struct wined3d_ma
  static void wined3d_cs_exec_reset_state(struct wined3d_cs *cs, const void *data)
  {
      struct wined3d_adapter *adapter = cs->device->adapter;
@@ -1773,7 +1768,7 @@ diff --git a/dlls/wined3d/cs.c b/dlls/wined3d/cs.c
      op = cs->ops->require_space(cs, sizeof(*op));
      op->opcode = WINED3D_CS_OP_RESET_STATE;
  
-@@ -1382,17 +2201,32 @@ void wined3d_cs_emit_reset_state(struct wined3d_cs *cs)
+@@ -1382,17 +2194,32 @@ void wined3d_cs_emit_reset_state(struct wined3d_cs *cs)
  }
  
  static void wined3d_cs_exec_destroy_object(struct wined3d_cs *cs, const void *data)
@@ -1806,7 +1801,7 @@ diff --git a/dlls/wined3d/cs.c b/dlls/wined3d/cs.c
      op->opcode = WINED3D_CS_OP_DESTROY_OBJECT;
      op->callback = callback;
      op->object = object;
-@@ -1400,40 +2234,96 @@ void wined3d_cs_emit_destroy_object(struct wined3d_cs *cs, void (*callback)(void
+@@ -1400,40 +2227,96 @@ void wined3d_cs_emit_destroy_object(struct wined3d_cs *cs, void (*callback)(void
      cs->ops->submit(cs);
  }
  
@@ -1903,7 +1898,7 @@ diff --git a/dlls/wined3d/cs.c b/dlls/wined3d/cs.c
      op->opcode = WINED3D_CS_OP_PRELOAD_RESOURCE;
      op->resource = resource;
  
-@@ -1442,20 +2332,32 @@ void wined3d_cs_emit_preload_resource(struct wined3d_cs *cs, struct wined3d_reso
+@@ -1442,20 +2325,32 @@ void wined3d_cs_emit_preload_resource(struct wined3d_cs *cs, struct wined3d_reso
      cs->ops->submit(cs);
  }
  
@@ -1936,7 +1931,7 @@ diff --git a/dlls/wined3d/cs.c b/dlls/wined3d/cs.c
      op->opcode = WINED3D_CS_OP_UNLOAD_RESOURCE;
      op->resource = resource;
  
-@@ -1464,13 +2366,21 @@ void wined3d_cs_emit_unload_resource(struct wined3d_cs *cs, struct wined3d_resou
+@@ -1464,13 +2359,21 @@ void wined3d_cs_emit_unload_resource(struct wined3d_cs *cs, struct wined3d_resou
      cs->ops->submit(cs);
  }
  
@@ -1958,7 +1953,7 @@ diff --git a/dlls/wined3d/cs.c b/dlls/wined3d/cs.c
  }
  
  HRESULT wined3d_cs_map(struct wined3d_cs *cs, struct wined3d_resource *resource, unsigned int sub_resource_idx,
-@@ -1479,7 +2389,11 @@ HRESULT wined3d_cs_map(struct wined3d_cs *cs, struct wined3d_resource *resource,
+@@ -1479,7 +2382,11 @@ HRESULT wined3d_cs_map(struct wined3d_cs *cs, struct wined3d_resource *resource,
      struct wined3d_cs_map *op;
      HRESULT hr;
  
@@ -1970,7 +1965,7 @@ diff --git a/dlls/wined3d/cs.c b/dlls/wined3d/cs.c
      op->opcode = WINED3D_CS_OP_MAP;
      op->resource = resource;
      op->sub_resource_idx = sub_resource_idx;
-@@ -1488,17 +2402,29 @@ HRESULT wined3d_cs_map(struct wined3d_cs *cs, struct wined3d_resource *resource,
+@@ -1488,17 +2395,29 @@ HRESULT wined3d_cs_map(struct wined3d_cs *cs, struct wined3d_resource *resource,
      op->flags = flags;
      op->hr = &hr;
  
@@ -2000,7 +1995,7 @@ diff --git a/dlls/wined3d/cs.c b/dlls/wined3d/cs.c
  }
  
  HRESULT wined3d_cs_unmap(struct wined3d_cs *cs, struct wined3d_resource *resource, unsigned int sub_resource_idx)
-@@ -1506,19 +2432,805 @@ HRESULT wined3d_cs_unmap(struct wined3d_cs *cs, struct wined3d_resource *resourc
+@@ -1506,19 +2425,805 @@ HRESULT wined3d_cs_unmap(struct wined3d_cs *cs, struct wined3d_resource *resourc
      struct wined3d_cs_unmap *op;
      HRESULT hr;
  
@@ -2015,13 +2010,13 @@ diff --git a/dlls/wined3d/cs.c b/dlls/wined3d/cs.c
      op->hr = &hr;
  
 +#if !defined(STAGING_CSMT)
-+    cs->ops->submit(cs);
+     cs->ops->submit(cs);
 +#else  /* STAGING_CSMT */
 +    cs->ops->submit_and_wait(cs);
-+
-+    return hr;
-+}
-+
+ 
+     return hr;
+ }
+ 
 +static UINT wined3d_cs_exec_glfinish(struct wined3d_cs *cs, const void *data)
 +{
 +    const struct wined3d_cs_finish *op = data;
@@ -2309,7 +2304,7 @@ diff --git a/dlls/wined3d/cs.c b/dlls/wined3d/cs.c
 +    if (src_surface && src_surface != dst_surface)
 +        wined3d_resource_acquire(&src_surface->container->resource);
 +
-     cs->ops->submit(cs);
++    cs->ops->submit(cs);
 +}
 +
 +static UINT wined3d_cs_exec_clear_rtv(struct wined3d_cs *cs, const void *data)
@@ -2498,10 +2493,10 @@ diff --git a/dlls/wined3d/cs.c b/dlls/wined3d/cs.c
 +
 +    cs->ops->submit_and_wait(cs);
 +#endif /* STAGING_CSMT */
- 
-     return hr;
- }
- 
++
++    return hr;
++}
++
 +#if !defined(STAGING_CSMT)
  static void (* const wined3d_cs_op_handlers[])(struct wined3d_cs *cs, const void *data) =
  {
@@ -2806,7 +2801,7 @@ diff --git a/dlls/wined3d/cs.c b/dlls/wined3d/cs.c
      /* WINED3D_CS_OP_PRESENT                    */ wined3d_cs_exec_present,
      /* WINED3D_CS_OP_CLEAR                      */ wined3d_cs_exec_clear,
      /* WINED3D_CS_OP_DISPATCH                   */ wined3d_cs_exec_dispatch,
-@@ -1550,13 +3262,43 @@ static void (* const wined3d_cs_op_handlers[])(struct wined3d_cs *cs, const void
+@@ -1550,13 +3255,43 @@ static void (* const wined3d_cs_op_handlers[])(struct wined3d_cs *cs, const void
      /* WINED3D_CS_OP_RESET_STATE                */ wined3d_cs_exec_reset_state,
      /* WINED3D_CS_OP_DESTROY_OBJECT             */ wined3d_cs_exec_destroy_object,
      /* WINED3D_CS_OP_QUERY_ISSUE                */ wined3d_cs_exec_query_issue,
@@ -2850,7 +2845,7 @@ diff --git a/dlls/wined3d/cs.c b/dlls/wined3d/cs.c
  {
      if (size > cs->data_size)
      {
-@@ -1580,6 +3322,7 @@ static void wined3d_cs_st_submit(struct wined3d_cs *cs)
+@@ -1580,6 +3315,7 @@ static void wined3d_cs_st_submit(struct wined3d_cs *cs)
      wined3d_cs_op_handlers[opcode](cs, cs->data);
  }
  
@@ -2858,7 +2853,7 @@ diff --git a/dlls/wined3d/cs.c b/dlls/wined3d/cs.c
  static void wined3d_cs_st_push_constants(struct wined3d_cs *cs, enum wined3d_push_constants p,
          unsigned int start_idx, unsigned int count, const void *constants)
  {
-@@ -1620,15 +3363,161 @@ static void wined3d_cs_st_push_constants(struct wined3d_cs *cs, enum wined3d_pus
+@@ -1620,15 +3356,161 @@ static void wined3d_cs_st_push_constants(struct wined3d_cs *cs, enum wined3d_pus
      for (i = 0, context_count = device->context_count; i < context_count; ++i)
      {
          device->contexts[i]->constant_update_mask |= push_constant_info[p].mask;
@@ -3020,7 +3015,7 @@ diff --git a/dlls/wined3d/cs.c b/dlls/wined3d/cs.c
  
  struct wined3d_cs *wined3d_cs_create(struct wined3d_device *device)
  {
-@@ -1638,34 +3527,97 @@ struct wined3d_cs *wined3d_cs_create(struct wined3d_device *device)
+@@ -1638,34 +3520,97 @@ struct wined3d_cs *wined3d_cs_create(struct wined3d_device *device)
      if (!(cs = HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, sizeof(*cs))))
          return NULL;
  
@@ -6366,13 +6361,14 @@ diff --git a/dlls/wined3d/wined3d_private.h b/dlls/wined3d/wined3d_private.h
  struct wined3d_cs_ops
  {
      void *(*require_space)(struct wined3d_cs *cs, size_t size);
-@@ -3165,34 +3274,128 @@ struct wined3d_cs_ops
+@@ -3165,34 +3274,129 @@ struct wined3d_cs_ops
      void (*push_constants)(struct wined3d_cs *cs, enum wined3d_push_constants p,
              unsigned int start_idx, unsigned int count, const void *constants);
  };
 +#else  /* STAGING_CSMT */
 +struct wined3d_cs_list
 +{
++    CRITICAL_SECTION lock;
 +    struct list blocks;
 +    LONG count;
 +};
@@ -6495,7 +6491,7 @@ diff --git a/dlls/wined3d/wined3d_private.h b/dlls/wined3d/wined3d_private.h
  void wined3d_cs_emit_set_clip_plane(struct wined3d_cs *cs, UINT plane_idx,
          const struct wined3d_vec4 *plane) DECLSPEC_HIDDEN;
  void wined3d_cs_emit_set_color_key(struct wined3d_cs *cs, struct wined3d_texture *texture,
-@@ -3203,9 +3406,17 @@ void wined3d_cs_emit_set_depth_stencil_view(struct wined3d_cs *cs,
+@@ -3203,9 +3407,17 @@ void wined3d_cs_emit_set_depth_stencil_view(struct wined3d_cs *cs,
          struct wined3d_rendertarget_view *view) DECLSPEC_HIDDEN;
  void wined3d_cs_emit_set_index_buffer(struct wined3d_cs *cs, struct wined3d_buffer *buffer,
          enum wined3d_format_id format_id, unsigned int offset) DECLSPEC_HIDDEN;
@@ -6513,7 +6509,7 @@ diff --git a/dlls/wined3d/wined3d_private.h b/dlls/wined3d/wined3d_private.h
  void wined3d_cs_emit_set_rasterizer_state(struct wined3d_cs *cs,
          struct wined3d_rasterizer_state *rasterizer_state) DECLSPEC_HIDDEN;
  void wined3d_cs_emit_set_render_state(struct wined3d_cs *cs,
-@@ -3237,17 +3448,35 @@ void wined3d_cs_emit_set_unordered_access_view(struct wined3d_cs *cs, enum wined
+@@ -3237,17 +3449,35 @@ void wined3d_cs_emit_set_unordered_access_view(struct wined3d_cs *cs, enum wined
  void wined3d_cs_emit_set_vertex_declaration(struct wined3d_cs *cs,
          struct wined3d_vertex_declaration *declaration) DECLSPEC_HIDDEN;
  void wined3d_cs_emit_set_viewport(struct wined3d_cs *cs, const struct wined3d_viewport *viewport) DECLSPEC_HIDDEN;
@@ -6549,7 +6545,7 @@ diff --git a/dlls/wined3d/wined3d_private.h b/dlls/wined3d/wined3d_private.h
  
  /* TODO: Add tests and support for FLOAT16_4 POSITIONT, D3DCOLOR position, other
   * fixed function semantics as D3DCOLOR or FLOAT16 */
-@@ -3348,6 +3577,14 @@ struct wined3d_gl_view
+@@ -3348,6 +3578,14 @@ struct wined3d_gl_view
      GLuint name;
  };
  
@@ -6564,7 +6560,7 @@ diff --git a/dlls/wined3d/wined3d_private.h b/dlls/wined3d/wined3d_private.h
  struct wined3d_shader_resource_view
  {
      LONG refcount;
-@@ -3383,7 +3620,12 @@ void wined3d_unordered_access_view_invalidate_location(struct wined3d_unordered_
+@@ -3383,7 +3621,12 @@ void wined3d_unordered_access_view_invalidate_location(struct wined3d_unordered_
  struct wined3d_swapchain_ops
  {
      void (*swapchain_present)(struct wined3d_swapchain *swapchain,
@@ -6577,7 +6573,7 @@ diff --git a/dlls/wined3d/wined3d_private.h b/dlls/wined3d/wined3d_private.h
      void (*swapchain_frontbuffer_updated)(struct wined3d_swapchain *swapchain);
  };
  
-@@ -3420,6 +3662,10 @@ struct wined3d_swapchain
+@@ -3420,6 +3663,10 @@ struct wined3d_swapchain
  
  void wined3d_swapchain_activate(struct wined3d_swapchain *swapchain, BOOL activate) DECLSPEC_HIDDEN;
  struct wined3d_context *swapchain_get_context(struct wined3d_swapchain *swapchain) DECLSPEC_HIDDEN;