vkd3d: Get rid of default BufferCount latency optimization in swapchain.

There are various cases now observed in the wild where BufferCount = 2
was causing problems. The problem is obvious in games with reasonably
high CPU and GPU subscription. Since present wait events come in at a
different alignment than presentation, it can happen that the CPU is
delayed just long enough for a GPU frame to end up being falsely
dropped.

This is quite bad on Deck since GPU clocks will start dropping as a
result and then CPU clocks drop, and we have a really bad time.

The risk here is that some games will get more latency, but we have to
err on the side of caution here unfortunately. Ideally, the internal
latency handle should be more adaptive in nature. It's currently unclear
which system component should be responsible for implementing that
however.

Signed-off-by: Hans-Kristian Arntzen <post@arntzen-software.no>

libs/vkd3d/swapchain.c

@@ -112,7 +112,6 @@ struct dxgi_vk_swap_chain
 
     UINT frame_latency;
     UINT frame_latency_internal;
-    bool frame_latency_internal_is_static;
     VkSurfaceKHR vk_surface;
 
     struct low_latency_state requested_low_latency_state;
@@ -780,26 +779,6 @@ static HRESULT STDMETHODCALLTYPE dxgi_vk_swap_chain_ChangeProperties(IDXGIVkSwap
         return hr;
     }
 
-    /* If BufferCount changes, so does expectations about latency. */
-    if (vkd3d_native_sync_handle_is_valid(chain->frame_latency_event_internal) &&
-            !chain->frame_latency_internal_is_static)
-    {
-        if (chain->desc.BufferCount > chain->frame_latency_internal)
-        {
-            vkd3d_native_sync_handle_release(chain->frame_latency_event_internal,
-                    chain->desc.BufferCount - chain->frame_latency_internal);
-            chain->frame_latency_internal = chain->desc.BufferCount;
-        }
-        else
-        {
-            while (chain->frame_latency_internal > chain->desc.BufferCount)
-            {
-                vkd3d_native_sync_handle_acquire(chain->frame_latency_event_internal);
-                chain->frame_latency_internal--;
-            }
-        }
-    }
-
     if (chain->user.index >= chain->desc.BufferCount)
     {
         /* Need to reset the user index in case the buffer count is lowered.
@@ -1377,18 +1356,20 @@ static HRESULT dxgi_vk_swap_chain_init_sync_objects(struct dxgi_vk_swap_chain *c
          * If we don't have present wait, we will sync with present_request_done_event (below) instead.
          * Adding more sync against internal blit fences is meaningless
          * and can cause weird pumping issues in some cases since we're simultaneously syncing
-         * against two different timelines. Deduce latency based on BufferCount by default. */
-        chain->frame_latency_internal = chain->desc.BufferCount;
+         * against two different timelines.
+         * While we used to deduce latency based on BufferCount by default,
+         * it was causing various issues, especially on Deck.
+         * With 2 frame latency where CPU and GPU are decently subscribed,
+         * the present wait event will come in with VBlank-alignment, and that extra delay can be enough to nudge CPU timings to the point
+         * where GPU goes falsely idle. This ends up dropping GPU clocks, and we have bad feedback loop effects.
+         * This effect has been observed in enough games now that it's too risky to enable that behavior by default. */
+        chain->frame_latency_internal = DEFAULT_FRAME_LATENCY;
 
         if (vkd3d_get_env_var("VKD3D_SWAPCHAIN_LATENCY_FRAMES", env, sizeof(env)))
         {
             latency_override = strtoul(env, NULL, 0);
-
             if (latency_override >= 1 && latency_override <= DXGI_MAX_SWAP_CHAIN_BUFFERS)
-            {
                 chain->frame_latency_internal = latency_override;
-                chain->frame_latency_internal_is_static = true;
-            }
         }
 
         INFO("Ensure maximum latency of %u frames with KHR_present_wait.\n",