diff --git a/patches/wined3d-CSMT_Main/9999-IfDefined.patch b/patches/wined3d-CSMT_Main/9999-IfDefined.patch
index 739b8d93..c5ff0132 100644
--- a/patches/wined3d-CSMT_Main/9999-IfDefined.patch
+++ b/patches/wined3d-CSMT_Main/9999-IfDefined.patch
@@ -6,10 +6,39 @@ Based on patches by:
     Sebastian Lackner <sebastian@fds-team.de>
     Stefan Dösinger <stefan@codeweavers.com>
 
+diff --git a/dlls/wined3d/buffer.c b/dlls/wined3d/buffer.c
+--- a/dlls/wined3d/buffer.c
++++ b/dlls/wined3d/buffer.c
+@@ -1219,6 +1219,9 @@ HRESULT wined3d_buffer_upload_data(struct wined3d_buffer *buffer,
+         const struct wined3d_box *box, const void *data)
+ {
+     UINT offset, size;
++#if defined(STAGING_CSMT)
++    DWORD flags = 0;
++#endif /* STAGING_CSMT */
+     HRESULT hr;
+     BYTE *ptr;
+ 
+@@ -1233,7 +1236,14 @@ HRESULT wined3d_buffer_upload_data(struct wined3d_buffer *buffer,
+         size = buffer->resource.size;
+     }
+ 
++#if !defined(STAGING_CSMT)
+     if (FAILED(hr = wined3d_buffer_map(buffer, offset, size, &ptr, 0)))
++#else  /* STAGING_CSMT */
++    if (offset == 0 && size == buffer->resource.size)
++        flags = WINED3D_MAP_DISCARD;
++
++    if (FAILED(hr = wined3d_buffer_map(buffer, offset, size, &ptr, flags)))
++#endif /* STAGING_CSMT */
+         return hr;
+ 
+     memcpy(ptr, data, size);
 diff --git a/dlls/wined3d/cs.c b/dlls/wined3d/cs.c
+index 2f77c7b0b7..e8bb63ef3b 100644
 --- a/dlls/wined3d/cs.c
 +++ b/dlls/wined3d/cs.c
-@@ -400,6 +400,9 @@ struct wined3d_cs_update_sub_resource
+@@ -405,6 +405,9 @@ struct wined3d_cs_update_sub_resource
      unsigned int sub_resource_idx;
      struct wined3d_box box;
      struct wined3d_sub_resource_data data;
@@ -19,7 +48,7 @@ diff --git a/dlls/wined3d/cs.c b/dlls/wined3d/cs.c
  };
  
  struct wined3d_cs_add_dirty_texture_region
-@@ -2247,6 +2250,53 @@ void wined3d_cs_emit_update_sub_resource(struct wined3d_cs *cs, struct wined3d_r
+@@ -2255,6 +2258,53 @@ void wined3d_cs_emit_update_sub_resource(struct wined3d_cs *cs, struct wined3d_r
          unsigned int slice_pitch)
  {
      struct wined3d_cs_update_sub_resource *op;
@@ -73,19 +102,21 @@ diff --git a/dlls/wined3d/cs.c b/dlls/wined3d/cs.c
  
      op = cs->ops->require_space(cs, sizeof(*op), WINED3D_CS_QUEUE_MAP);
      op->opcode = WINED3D_CS_OP_UPDATE_SUB_RESOURCE;
-@@ -2260,8 +2310,10 @@ void wined3d_cs_emit_update_sub_resource(struct wined3d_cs *cs, struct wined3d_r
+@@ -2268,8 +2318,10 @@ void wined3d_cs_emit_update_sub_resource(struct wined3d_cs *cs, struct wined3d_r
      wined3d_resource_acquire(resource);
  
      cs->ops->submit(cs, WINED3D_CS_QUEUE_MAP);
+-    /* The data pointer may go away, so we need to wait until it is read.
+-     * Copying the data may be faster if it's small. */
 +#if !defined(STAGING_CSMT)
-     /* The data pointer may go away, so we need to wait until it is read.
-      * Copying the data may be faster if it's small. */
++     /* The data pointer may go away, so we need to wait until it is read.
++      * Copying the data may be faster if it's small. */
 +#endif /* STAGING_CSMT */
      cs->ops->finish(cs, WINED3D_CS_QUEUE_MAP);
  }
  
-@@ -2640,6 +2692,13 @@ static void (* const wined3d_cs_op_handlers[])(struct wined3d_cs *cs, const void
-     /* WINED3D_CS_OP_GENERATE_MIPS               */ wined3d_cs_exec_generate_mips,
+@@ -2450,6 +2504,14 @@ static void (* const wined3d_cs_op_handlers[])(struct wined3d_cs *cs, const void
+     /* WINED3D_CS_OP_GENERATE_MIPMAPS            */ wined3d_cs_exec_generate_mipmaps,
  };
  
 +#if defined(STAGING_CSMT)
@@ -95,10 +126,11 @@ diff --git a/dlls/wined3d/cs.c b/dlls/wined3d/cs.c
 +}
 +
 +#endif /* STAGING_CSMT */
++
  static void *wined3d_cs_st_require_space(struct wined3d_cs *cs, size_t size, enum wined3d_cs_queue_id queue_id)
  {
      if (size > (cs->data_size - cs->end))
-@@ -2693,6 +2752,9 @@ static void wined3d_cs_st_finish(struct wined3d_cs *cs, enum wined3d_cs_queue_id
+@@ -2503,6 +2565,9 @@ static void wined3d_cs_st_finish(struct wined3d_cs *cs, enum wined3d_cs_queue_id
  
  static const struct wined3d_cs_ops wined3d_cs_st_ops =
  {
@@ -108,7 +140,7 @@ diff --git a/dlls/wined3d/cs.c b/dlls/wined3d/cs.c
      wined3d_cs_st_require_space,
      wined3d_cs_st_submit,
      wined3d_cs_st_finish,
-@@ -2726,6 +2788,21 @@ static void wined3d_cs_mt_submit(struct wined3d_cs *cs, enum wined3d_cs_queue_id
+@@ -2536,6 +2601,22 @@ static void wined3d_cs_mt_submit(struct wined3d_cs *cs, enum wined3d_cs_queue_id
      wined3d_cs_queue_submit(&cs->queue[queue_id], cs);
  }
  
@@ -127,10 +159,11 @@ diff --git a/dlls/wined3d/cs.c b/dlls/wined3d/cs.c
 +}
 +
 +#endif /* STAGING_CSMT */
++
  static void *wined3d_cs_queue_require_space(struct wined3d_cs_queue *queue, size_t size, struct wined3d_cs *cs)
  {
      size_t queue_size = ARRAY_SIZE(queue->data);
-@@ -2787,6 +2864,16 @@ static void *wined3d_cs_queue_require_space(struct wined3d_cs_queue *queue, size
+@@ -2597,6 +2678,16 @@ static void *wined3d_cs_queue_require_space(struct wined3d_cs_queue *queue, size
      return packet->data;
  }
  
@@ -147,7 +180,7 @@ diff --git a/dlls/wined3d/cs.c b/dlls/wined3d/cs.c
  static void *wined3d_cs_mt_require_space(struct wined3d_cs *cs, size_t size, enum wined3d_cs_queue_id queue_id)
  {
      if (cs->thread_id == GetCurrentThreadId())
-@@ -2806,6 +2893,9 @@ static void wined3d_cs_mt_finish(struct wined3d_cs *cs, enum wined3d_cs_queue_id
+@@ -2616,6 +2707,9 @@ static void wined3d_cs_mt_finish(struct wined3d_cs *cs, enum wined3d_cs_queue_id
  
  static const struct wined3d_cs_ops wined3d_cs_mt_ops =
  {
@@ -271,14 +304,14 @@ diff --git a/dlls/wined3d/device.c b/dlls/wined3d/device.c
 diff --git a/dlls/wined3d/surface.c b/dlls/wined3d/surface.c
 --- a/dlls/wined3d/surface.c
 +++ b/dlls/wined3d/surface.c
-@@ -2461,7 +2461,11 @@ static BOOL surface_load_texture(struct
+@@ -2410,7 +2410,11 @@ static BOOL surface_load_texture(struct wined3d_surface *surface,
      /* Don't use PBOs for converted surfaces. During PBO conversion we look at
       * WINED3D_TEXTURE_CONVERTED but it isn't set (yet) in all cases it is
       * getting called. */
 +#if !defined(STAGING_CSMT)
-     if ((format.conv_byte_count || conversion) && texture->sub_resources[sub_resource_idx].buffer_object)
+     if ((format.convert || conversion) && texture->sub_resources[sub_resource_idx].buffer_object)
 +#else  /* STAGING_CSMT */
-+    if ((format.conv_byte_count || conversion) && texture->sub_resources[sub_resource_idx].buffer)
++    if ((format.convert || conversion) && texture->sub_resources[sub_resource_idx].buffer)
 +#endif /* STAGING_CSMT */
      {
          TRACE("Removing the pbo attached to surface %p.\n", surface);
@@ -286,9 +319,9 @@ diff --git a/dlls/wined3d/surface.c b/dlls/wined3d/surface.c
 diff --git a/dlls/wined3d/swapchain.c b/dlls/wined3d/swapchain.c
 --- a/dlls/wined3d/swapchain.c
 +++ b/dlls/wined3d/swapchain.c
-@@ -476,7 +476,11 @@ static void swapchain_gl_present(struct
-     if (swapchain->render_to_fbo)
+@@ -483,7 +483,11 @@ static void swapchain_gl_present(struct wined3d_swapchain *swapchain,
          swapchain_blit(swapchain, context, src_rect, dst_rect);
+     }
  
 +#if !defined(STAGING_CSMT)
      if (swapchain->num_contexts > 1)
@@ -573,12 +606,27 @@ diff --git a/dlls/wined3d/view.c b/dlls/wined3d/view.c
      wined3d_cs_init_object(resource->device->cs, wined3d_unordered_access_view_cs_init, view);
  
      return WINED3D_OK;
+diff --git a/dlls/wined3d/wined3d_main.c b/dlls/wined3d/wined3d_main.c
+--- a/dlls/wined3d/wined3d_main.c
++++ b/dlls/wined3d/wined3d_main.c
+@@ -72,7 +72,11 @@ static CRITICAL_SECTION wined3d_wndproc_cs = {&wined3d_wndproc_cs_debug, -1, 0,
+  * where appropriate. */
+ struct wined3d_settings wined3d_settings =
+ {
++#if !defined(STAGING_CSMT)
+     FALSE,          /* No multithreaded CS by default. */
++#else  /* STAGING_CSMT */
++    TRUE,           /* Multithreaded CS by default. */
++#endif /* STAGING_CSMT */
+     MAKEDWORD_VERSION(1, 0), /* Default to legacy OpenGL */
+     TRUE,           /* Use of GLSL enabled by default */
+     ORM_FBO,        /* Use FBOs to do offscreen rendering */
 diff --git a/dlls/wined3d/wined3d_private.h b/dlls/wined3d/wined3d_private.h
 --- a/dlls/wined3d/wined3d_private.h
 +++ b/dlls/wined3d/wined3d_private.h
-@@ -2898,6 +2898,16 @@ struct wined3d_dummy_textures
-     GLuint tex_2d_ms_array;
- };
+@@ -2860,6 +2860,16 @@ static inline BOOL wined3d_dualblend_enabled(const struct wined3d_state *state,
+     return FALSE;
+ }
  
 +#if defined(STAGING_CSMT)
 +struct wined3d_gl_bo
@@ -593,7 +641,7 @@ diff --git a/dlls/wined3d/wined3d_private.h b/dlls/wined3d/wined3d_private.h
  #define WINED3D_UNMAPPED_STAGE ~0u
  
  /* Multithreaded flag. Removed from the public header to signal that
-@@ -2998,6 +3008,12 @@ LRESULT device_process_message(struct wi
+@@ -2971,6 +2981,12 @@ LRESULT device_process_message(struct wined3d_device *device, HWND window, BOOL
  void device_resource_add(struct wined3d_device *device, struct wined3d_resource *resource) DECLSPEC_HIDDEN;
  void device_resource_released(struct wined3d_device *device, struct wined3d_resource *resource) DECLSPEC_HIDDEN;
  void device_invalidate_state(const struct wined3d_device *device, DWORD state) DECLSPEC_HIDDEN;
@@ -606,7 +654,7 @@ diff --git a/dlls/wined3d/wined3d_private.h b/dlls/wined3d/wined3d_private.h
  
  static inline BOOL isStateDirty(const struct wined3d_context *context, DWORD state)
  {
-@@ -3188,7 +3204,11 @@ struct wined3d_texture
+@@ -3158,7 +3174,11 @@ struct wined3d_texture
  
          unsigned int map_count;
          DWORD locations;
@@ -618,7 +666,7 @@ diff --git a/dlls/wined3d/wined3d_private.h b/dlls/wined3d/wined3d_private.h
      } sub_resources[1];
  };
  
-@@ -3504,6 +3524,9 @@ struct wined3d_cs_queue
+@@ -3474,6 +3494,9 @@ struct wined3d_cs_queue
  
  struct wined3d_cs_ops
  {