Lazy-load the dcomp library (#8216)

Cargo.lock

@@ -5126,6 +5126,7 @@ dependencies = [
  "naga",
  "ndk-sys 0.6.0+11769913",
  "objc",
+ "once_cell",
  "ordered-float 5.0.0",
  "parking_lot",
  "portable-atomic",

wgpu-hal/Cargo.toml

@@ -136,6 +136,7 @@ dx12 = [
     "dep:libloading",
     "dep:log",
     "dep:ordered-float",
+    "dep:once_cell",
     "dep:parking_lot",
     "dep:profiling",
     "dep:range-alloc",
@@ -258,6 +259,7 @@ windows-core = { workspace = true, optional = true }
 bit-set = { workspace = true, optional = true }
 range-alloc = { workspace = true, optional = true }
 gpu-allocator = { workspace = true, optional = true }
+once_cell = { workspace = true, optional = true }
 # backend: GLES
 glutin_wgl_sys = { workspace = true, optional = true }
 

wgpu-hal/src/dx12/adapter.rs

@@ -17,7 +17,7 @@ use crate::{
         self,
         dxgi::{factory::DxgiAdapter, result::HResult},
     },
-    dx12::{shader_compilation, SurfaceTarget},
+    dx12::{dcomp::DCompLib, shader_compilation, SurfaceTarget},
 };
 
 impl Drop for super::Adapter {
@@ -55,6 +55,7 @@ impl super::Adapter {
     pub(super) fn expose(
         adapter: DxgiAdapter,
         library: &Arc<D3D12Lib>,
+        dcomp_lib: &Arc<DCompLib>,
         instance_flags: wgt::InstanceFlags,
         memory_budget_thresholds: wgt::MemoryBudgetThresholds,
         compiler_container: Arc<shader_compilation::CompilerContainer>,
@@ -559,6 +560,7 @@ impl super::Adapter {
                 raw: adapter,
                 device,
                 library: Arc::clone(library),
+                dcomp_lib: Arc::clone(dcomp_lib),
                 private_caps,
                 presentation_timer,
                 workarounds,
@@ -726,6 +728,7 @@ impl crate::Adapter for super::Adapter {
             memory_hints,
             self.private_caps,
             &self.library,
+            &self.dcomp_lib,
             self.memory_budget_thresholds,
             self.compiler_container.clone(),
             self.options.clone(),

wgpu-hal/src/dx12/dcomp.rs

@@ -1,4 +1,65 @@
+use alloc::sync::Arc;
+use core::{ffi, ptr};
+
+use once_cell::sync::Lazy;
 use windows::Win32::{Foundation::HWND, Graphics::DirectComposition};
+use windows_core::Interface;
+
+use super::DynLib;
+
+// Lazy-loaded DirectComposition library
+#[derive(Debug)]
+pub(crate) struct DCompLib {
+    lib: Lazy<Result<DynLib, crate::SurfaceError>>,
+}
+
+impl DCompLib {
+    pub(crate) fn new() -> Self {
+        Self {
+            lib: Lazy::new(|| unsafe {
+                DynLib::new("dcomp.dll").map_err(|err| {
+                    log::error!("Error loading dcomp.dll: {err}");
+                    crate::SurfaceError::Other("Error loading dcomp.dll")
+                })
+            }),
+        }
+    }
+
+    fn get_lib(&self) -> Result<&DynLib, crate::SurfaceError> {
+        match self.lib.as_ref() {
+            Ok(lib) => Ok(lib),
+            Err(err) => Err(err.clone()),
+        }
+    }
+
+    pub(crate) fn create_device(
+        &self,
+    ) -> Result<DirectComposition::IDCompositionDevice, crate::SurfaceError> {
+        let lib = self.get_lib()?;
+
+        // Calls windows::Win32::Graphics::DirectComposition::DCompositionCreateDevice2 on dcomp.dll
+        type Fun = extern "system" fn(
+            pdxdevice: *mut ffi::c_void,
+            riid: *const windows_core::GUID,
+            ppdcompdevice: *mut *mut ffi::c_void,
+        ) -> windows_core::HRESULT;
+        let func: libloading::Symbol<Fun> =
+            unsafe { lib.get(c"DCompositionCreateDevice2".to_bytes()) }?;
+
+        let mut res: Option<DirectComposition::IDCompositionDevice> = None;
+
+        (func)(
+            ptr::null_mut(),
+            &DirectComposition::IDCompositionDevice::IID,
+            <*mut _>::cast(&mut res),
+        )
+        .map(|| res.unwrap())
+        .map_err(|err| {
+            log::error!("DirectComposition::DCompositionCreateDevice2 failed: {err}");
+            crate::SurfaceError::Other("DirectComposition::DCompositionCreateDevice2")
+        })
+    }
+}
 
 #[derive(Default)]
 pub struct DCompState {
@@ -10,10 +71,11 @@ impl DCompState {
     /// If the device is already initialized, it will return the existing state.
     pub unsafe fn get_or_init(
         &mut self,
+        lib: &Arc<DCompLib>,
         hwnd: &HWND,
     ) -> Result<&mut InnerState, crate::SurfaceError> {
         if self.inner.is_none() {
-            self.inner = Some(unsafe { InnerState::init(hwnd) }?);
+            self.inner = Some(unsafe { InnerState::init(lib, hwnd) }?);
         }
         Ok(self.inner.as_mut().unwrap())
     }
@@ -28,14 +90,9 @@ pub struct InnerState {
 
 impl InnerState {
     /// Creates a DirectComposition device and a target for the given window handle.
-    pub unsafe fn init(hwnd: &HWND) -> Result<Self, crate::SurfaceError> {
+    pub unsafe fn init(lib: &Arc<DCompLib>, hwnd: &HWND) -> Result<Self, crate::SurfaceError> {
         profiling::scope!("DCompState::init");
-        let dcomp_device: DirectComposition::IDCompositionDevice = {
+        let dcomp_device = lib.create_device()?;
-            unsafe { DirectComposition::DCompositionCreateDevice2(None) }.map_err(|err| {
-                log::error!("DirectComposition::DCompositionCreateDevice failed: {err}");
-                crate::SurfaceError::Other("DirectComposition::DCompositionCreateDevice")
-            })?
-        };
 
         let target = unsafe { dcomp_device.CreateTargetForHwnd(*hwnd, false) }.map_err(|err| {
             log::error!("IDCompositionDevice::CreateTargetForHwnd failed: {err}");

wgpu-hal/src/dx12/device.rs

@@ -26,7 +26,7 @@ use crate::{
         dxgi::{name::ObjectExt, result::HResult},
     },
     dx12::{
-        borrow_optional_interface_temporarily, shader_compilation, suballocation,
+        borrow_optional_interface_temporarily, shader_compilation, suballocation, DCompLib,
         DynamicStorageBufferOffsets, Event, ShaderCacheKey, ShaderCacheValue,
     },
     AccelerationStructureEntries, TlasInstance,
@@ -46,6 +46,7 @@ impl super::Device {
         memory_hints: &wgt::MemoryHints,
         private_caps: super::PrivateCapabilities,
         library: &Arc<D3D12Lib>,
+        dcomp_lib: &Arc<DCompLib>,
         memory_budget_thresholds: wgt::MemoryBudgetThresholds,
         compiler_container: Arc<shader_compilation::CompilerContainer>,
         backend_options: wgt::Dx12BackendOptions,
@@ -201,6 +202,7 @@ impl super::Device {
             )),
             options: backend_options,
             library: Arc::clone(library),
+            dcomp_lib: Arc::clone(dcomp_lib),
             #[cfg(feature = "renderdoc")]
             render_doc: Default::default(),
             null_rtv_handle,

wgpu-hal/src/dx12/instance.rs

@@ -12,7 +12,7 @@ use windows::{
 use super::SurfaceTarget;
 use crate::{
     auxil,
-    dx12::{shader_compilation::CompilerContainer, D3D12Lib},
+    dx12::{shader_compilation::CompilerContainer, D3D12Lib, DCompLib},
 };
 
 impl crate::Instance for super::Instance {
@@ -104,6 +104,7 @@ impl crate::Instance for super::Instance {
             factory,
             factory_media,
             library: Arc::new(lib_main),
+            dcomp_lib: Arc::new(DCompLib::new()),
             presentation_system: desc.backend_options.dx12.presentation_system,
             _lib_dxgi: lib_dxgi,
             supports_allow_tearing,
@@ -158,6 +159,7 @@ impl crate::Instance for super::Instance {
                 super::Adapter::expose(
                     raw,
                     &self.library,
+                    &self.dcomp_lib,
                     self.flags,
                     self.memory_budget_thresholds,
                     self.compiler_container.clone(),

wgpu-hal/src/dx12/mod.rs

@@ -101,6 +101,7 @@ use windows::{
     },
 };
 
+use self::dcomp::DCompLib;
 use crate::auxil::{
     self,
     dxgi::{
@@ -460,6 +461,7 @@ pub struct Instance {
     factory: DxgiFactory,
     factory_media: Option<Dxgi::IDXGIFactoryMedia>,
     library: Arc<D3D12Lib>,
+    dcomp_lib: Arc<DCompLib>,
     supports_allow_tearing: bool,
     presentation_system: wgt::Dx12SwapchainKind,
     _lib_dxgi: DxgiLib,
@@ -612,6 +614,7 @@ pub struct Adapter {
     raw: DxgiAdapter,
     device: Direct3D12::ID3D12Device,
     library: Arc<D3D12Lib>,
+    dcomp_lib: Arc<DCompLib>,
     private_caps: PrivateCapabilities,
     presentation_timer: auxil::dxgi::time::PresentationTimer,
     // Note: this isn't used right now, but we'll need it later.
@@ -685,6 +688,7 @@ pub struct Device {
     srv_uav_pool: Mutex<descriptor::CpuPool>,
     // library
     library: Arc<D3D12Lib>,
+    dcomp_lib: Arc<DCompLib>,
     #[cfg(feature = "renderdoc")]
     render_doc: auxil::renderdoc::RenderDoc,
     null_rtv_handle: descriptor::Handle,
@@ -1358,7 +1362,8 @@ impl crate::Surface for Surface {
                         dcomp_state,
                     } => {
                         let mut dcomp_state = dcomp_state.lock();
-                        let dcomp_state = unsafe { dcomp_state.get_or_init(handle) }?;
+                        let dcomp_state =
+                            unsafe { dcomp_state.get_or_init(&device.dcomp_lib, handle) }?;
                         // Set the new swap chain as the content for the backing visual
                         // and commit the changes to the composition visual tree.
                         {