vkd3d: Add workaround to explicitly clear memory on 6.10+ amdgpu.

It seems like there is a bug with the new zerovram code in kernel. Manifests as test failures in specific circumstances. Signed-off-by: Hans-Kristian Arntzen <[email protected]>
HansKristian-Work · Dec 4, 2024 · 5645ca1 · 5645ca1
1 parent f97a7f6
commit 5645ca1
Show file tree

Hide file tree

Showing 3 changed files with 34 additions and 0 deletions.
diff --git a/libs/vkd3d/device.c b/libs/vkd3d/device.c
@@ -2998,6 +2998,32 @@ static HRESULT vkd3d_select_queues(const struct d3d12_device *device,
     return S_OK;
 }
 
+static void d3d12_device_init_workarounds(struct d3d12_device *device)
+{
+    uint32_t major, minor, patch;
+
+    if (vkd3d_config_flags & VKD3D_CONFIG_FLAG_SKIP_DRIVER_WORKAROUNDS)
+        return;
+
+    if (device->device_info.properties2.properties.vendorID != 0x1002)
+        return;
+
+    if (vkd3d_get_linux_kernel_version(&major, &minor, &patch))
+    {
+        /* 6.10 amdgpu kernel changes the clear vram code to do background clears instead
+         * of on-demand clearing. This seems to have bugs, and we have been able to observe
+         * non-zeroed VRAM coming from the affected kernels.
+         * This workaround needs to be in place until we have confirmed a fix in upstream kernel. */
+        INFO("Detected Linux kernel version %u.%u.%u\n", major, minor, patch);
+
+        if (major > 6 || (major == 6 && minor >= 10))
+        {
+            INFO("AMDGPU broken kernel detected. Enabling manual memory clearing path.\n");
+            device->workarounds.amdgpu_broken_clearvram = true;
+        }
+    }
+}
+
 static HRESULT vkd3d_create_vk_device(struct d3d12_device *device,
         const struct vkd3d_device_create_info *create_info)
 {
@@ -3136,6 +3162,8 @@ static HRESULT vkd3d_create_vk_device(struct d3d12_device *device,
     device->vk_info.extension_count = device_info.enabledExtensionCount;
     device->vk_info.extension_names = extensions;
 
+    d3d12_device_init_workarounds(device);
+
     TRACE("Created Vulkan device %p.\n", vk_device);
 
     return hr;

diff --git a/libs/vkd3d/memory.c b/libs/vkd3d/memory.c
@@ -1892,6 +1892,7 @@ HRESULT vkd3d_allocate_memory(struct d3d12_device *device, struct vkd3d_memory_a
      * RADV definitely does this, and it seems like NV also does it.
      * TODO: an extension for this would be nice. */
     implementation_implicitly_clears =
+            !device->workarounds.amdgpu_broken_clearvram &&
             vkd3d_driver_implicitly_clears(device->device_info.vulkan_1_2_properties.driverID) &&
             !suballocate;
 

diff --git a/libs/vkd3d/vkd3d_private.h b/libs/vkd3d/vkd3d_private.h
@@ -4980,6 +4980,11 @@ struct d3d12_device
 
     struct vkd3d_device_swapchain_info swapchain_info;
     struct vkd3d_device_frame_markers frame_markers;
+
+    struct
+    {
+        bool amdgpu_broken_clearvram;
+    } workarounds;
 };
 
 HRESULT d3d12_device_create(struct vkd3d_instance *instance,