diff --git a/dlls/wined3d/cs.c b/dlls/wined3d/cs.c
index 85594e45ac4..6d413d49483 100644
--- a/dlls/wined3d/cs.c
+++ b/dlls/wined3d/cs.c
@@ -30,6 +30,7 @@ enum wined3d_cs_op
     WINED3D_CS_OP_CLEAR,
     WINED3D_CS_OP_DISPATCH,
     WINED3D_CS_OP_DRAW,
+    WINED3D_CS_OP_FLUSH,
     WINED3D_CS_OP_SET_PREDICATION,
     WINED3D_CS_OP_SET_VIEWPORT,
     WINED3D_CS_OP_SET_SCISSOR_RECT,
@@ -112,6 +113,11 @@ struct wined3d_cs_draw
     BOOL indexed;
 };
 
+struct wined3d_cs_flush
+{
+    enum wined3d_cs_op opcode;
+};
+
 struct wined3d_cs_set_predication
 {
     enum wined3d_cs_op opcode;
@@ -759,6 +765,25 @@ void wined3d_cs_emit_draw(struct wined3d_cs *cs, GLenum primitive_type, int base
     cs->ops->submit(cs);
 }
 
+static void wined3d_cs_exec_flush(struct wined3d_cs *cs, const void *data)
+{
+    struct wined3d_context *context;
+
+    context = context_acquire(cs->device, NULL, 0);
+    context->gl_info->gl_ops.gl.p_glFlush();
+    context_release(context);
+}
+
+void wined3d_cs_emit_flush(struct wined3d_cs *cs)
+{
+    struct wined3d_cs_flush *op;
+
+    op = cs->ops->require_space(cs, sizeof(*op));
+    op->opcode = WINED3D_CS_OP_FLUSH;
+
+    cs->ops->submit(cs);
+}
+
 static void wined3d_cs_exec_set_predication(struct wined3d_cs *cs, const void *data)
 {
     const struct wined3d_cs_set_predication *op = data;
@@ -2007,6 +2032,7 @@ static void (* const wined3d_cs_op_handlers[])(struct wined3d_cs *cs, const void
     /* WINED3D_CS_OP_CLEAR                      */ wined3d_cs_exec_clear,
     /* WINED3D_CS_OP_DISPATCH                   */ wined3d_cs_exec_dispatch,
     /* WINED3D_CS_OP_DRAW                       */ wined3d_cs_exec_draw,
+    /* WINED3D_CS_OP_FLUSH                      */ wined3d_cs_exec_flush,
     /* WINED3D_CS_OP_SET_PREDICATION            */ wined3d_cs_exec_set_predication,
     /* WINED3D_CS_OP_SET_VIEWPORT               */ wined3d_cs_exec_set_viewport,
     /* WINED3D_CS_OP_SET_SCISSOR_RECT           */ wined3d_cs_exec_set_scissor_rect,
diff --git a/dlls/wined3d/device.c b/dlls/wined3d/device.c
index a43e080c121..83965e7869b 100644
--- a/dlls/wined3d/device.c
+++ b/dlls/wined3d/device.c
@@ -3374,8 +3374,6 @@ HRESULT CDECL wined3d_device_begin_scene(struct wined3d_device *device)
 
 HRESULT CDECL wined3d_device_end_scene(struct wined3d_device *device)
 {
-    struct wined3d_context *context;
-
     TRACE("device %p.\n", device);
 
     if (!device->inScene)
@@ -3384,12 +3382,7 @@ HRESULT CDECL wined3d_device_end_scene(struct wined3d_device *device)
         return WINED3DERR_INVALIDCALL;
     }
 
-    context = context_acquire(device, NULL, 0);
-    /* We only have to do this if we need to read the, swapbuffers performs a flush for us */
-    context->gl_info->gl_ops.gl.p_glFlush();
-    /* No checkGLcall here to avoid locking the lock just for checking a call that hardly ever
-     * fails. */
-    context_release(context);
+    wined3d_cs_emit_flush(device->cs);
 
     device->inScene = FALSE;
     return WINED3D_OK;
diff --git a/dlls/wined3d/wined3d_private.h b/dlls/wined3d/wined3d_private.h
index b39bc2fc0e2..74b61fb839c 100644
--- a/dlls/wined3d/wined3d_private.h
+++ b/dlls/wined3d/wined3d_private.h
@@ -3258,6 +3258,7 @@ void wined3d_cs_emit_dispatch(struct wined3d_cs *cs,
 void wined3d_cs_emit_draw(struct wined3d_cs *cs, GLenum primitive_type, int base_vertex_idx,
         unsigned int start_idx, unsigned int index_count, unsigned int start_instance,
         unsigned int instance_count, BOOL indexed) DECLSPEC_HIDDEN;
+void wined3d_cs_emit_flush(struct wined3d_cs *cs) DECLSPEC_HIDDEN;
 void wined3d_cs_emit_preload_resource(struct wined3d_cs *cs, struct wined3d_resource *resource) DECLSPEC_HIDDEN;
 void wined3d_cs_emit_present(struct wined3d_cs *cs, struct wined3d_swapchain *swapchain,
         const RECT *src_rect, const RECT *dst_rect, HWND dst_window_override, DWORD flags) DECLSPEC_HIDDEN;