vkd3d: Attempt workaround to redirect NULL sparse tiles to dummy page.

Signed-off-by: Hans-Kristian Arntzen <[email protected]>

include/vkd3d.h

@@ -113,6 +113,7 @@ extern "C" {
 #define VKD3D_CONFIG_FLAG_NO_GPU_UPLOAD_HEAP (1ull << 55)
 #define VKD3D_CONFIG_FLAG_ONE_TIME_SUBMIT (1ull << 56)
 #define VKD3D_CONFIG_FLAG_SKIP_NULL_SPARSE_TILES (1ull << 57)
+#define VKD3D_CONFIG_FLAG_DUMMY_NULL_SPARSE_TILES (1ull << 58)
 
 struct vkd3d_instance;
 

libs/vkd3d/command.c

@@ -17514,8 +17514,21 @@ static void STDMETHODCALLTYPE d3d12_command_queue_UpdateTileMappings(ID3D12Comma
 
                 if (range_flag == D3D12_TILE_RANGE_FLAG_NULL)
                 {
-                    bind->vk_memory = VK_NULL_HANDLE;
-                    bind->vk_offset = 0;
+                    if ((vkd3d_config_flags & VKD3D_CONFIG_FLAG_DUMMY_NULL_SPARSE_TILES) &&
+                            command_queue->device->workarounds.amdgpu_broken_null_tile_mapping)
+                    {
+                        /* We have a convenient dummy BO already. It will also be allocated in 32-bit VA range,
+                         * so it's easy to detect binds to it in RADV bo history.
+                         * If we didn't use descriptor buffers for whatever reason we'll get the default NULL mapping behavior
+                         * which is fine, since this is such a targeted workaround. */
+                        bind->vk_memory = command_queue->device->global_descriptor_buffer.resource.device_allocation.vk_memory;
+                        bind->vk_offset = 0;
+                    }
+                    else
+                    {
+                        bind->vk_memory = VK_NULL_HANDLE;
+                        bind->vk_offset = 0;
+                    }
                 }
                 else
                 {

libs/vkd3d/device.c

@@ -608,9 +608,10 @@ static const struct vkd3d_instance_application_meta application_override[] = {
     /* FFVII Rebirth (2909400).
      * Game can destroy PSOs while they are in-flight.
      * Also, add no-staggered since this is a UE title without the common workaround,
-     * although that only seems to matter when FSR/DLSS injectors are used. */
+     * although that only seems to matter when FSR/DLSS injectors are used.
+     * Also, attempt to workaround a painful amdgpu kernel bug with null sparse tiles. */
     { VKD3D_STRING_COMPARE_EXACT, "ff7rebirth_.exe",
-            VKD3D_CONFIG_FLAG_RETAIN_PSOS | VKD3D_CONFIG_FLAG_NO_STAGGERED_SUBMIT, 0 },
+            VKD3D_CONFIG_FLAG_RETAIN_PSOS | VKD3D_CONFIG_FLAG_NO_STAGGERED_SUBMIT | VKD3D_CONFIG_FLAG_DUMMY_NULL_SPARSE_TILES, 0 },
     /* Unreal Engine catch-all. ReBAR is a massive uplift on RX 7600 for example in Wukong.
      * AMD windows drivers also seem to have some kind of general app-opt for UE titles.
      * Use no-staggered-submit by default on UE. We've only observed issues in Wukong here, but
@@ -1016,6 +1017,7 @@ static const struct vkd3d_debug_option vkd3d_config_options[] =
     {"no_gpu_upload_heap", VKD3D_CONFIG_FLAG_NO_GPU_UPLOAD_HEAP},
     {"one_time_submit", VKD3D_CONFIG_FLAG_ONE_TIME_SUBMIT},
     {"skip_null_sparse_tiles", VKD3D_CONFIG_FLAG_SKIP_NULL_SPARSE_TILES},
+    {"dummy_null_sparse_tiles", VKD3D_CONFIG_FLAG_DUMMY_NULL_SPARSE_TILES},
 };
 
 static void vkd3d_config_flags_init_once(void)

libs/vkd3d/resource.c

@@ -9109,6 +9109,7 @@ HRESULT vkd3d_global_descriptor_buffer_init(struct vkd3d_global_descriptor_buffe
         struct d3d12_device *device)
 {
     VkBufferUsageFlags2KHR vk_usage_flags;
+    VkDeviceSize size;
     HRESULT hr;
 
     bool requires_offset_buffer = device->device_info.properties2.properties.limits.minStorageBufferOffsetAlignment > 4;
@@ -9164,8 +9165,14 @@ HRESULT vkd3d_global_descriptor_buffer_init(struct vkd3d_global_descriptor_buffe
     /* Creates a default descriptor buffer we can use if the application does not bind anything.
      * This might happen if a meta shader is used without any prior descriptor heap bound,
      * and we need to use push descriptors with bufferlessPushDescriptors == VK_FALSE. */
+
+    /* The global descriptor buffer can be used as a DUMMY_NULL_TILE when a workaround is applied,
+     * so make sure it's at least 64K. In cases where it's relevant, we can rely on zerovram behavior too. */
+    size = (vkd3d_config_flags & VKD3D_CONFIG_FLAG_DUMMY_NULL_SPARSE_TILES) ?
+            VKD3D_TILE_SIZE : 4 * 1024;
+
     if (FAILED(hr = vkd3d_create_buffer_explicit_usage(device, vk_usage_flags,
-            4 * 1024, "descriptor-buffer", &global_descriptor_buffer->resource.vk_buffer)))
+            size, "descriptor-buffer", &global_descriptor_buffer->resource.vk_buffer)))
     {
         return hr;
     }