Add more config-ext interface

Author: MaxwellGengYF

include/luisa/backends/ext/dx_config_ext.h

@@ -76,7 +76,10 @@ struct DirectXDeviceConfigExt : public DeviceConfigExt {
     virtual bool SignalFence(
         ID3D12CommandQueue *queue,
         ID3D12Fence *fence, uint64_t fenceIndex) noexcept { return false; }
-
+    virtual bool WaitFence(
+        ID3D12CommandQueue *queue,
+        ID3D12Fence *fence, uint64_t fenceIndex) noexcept { return false; }
+    virtual bool SyncFence(ID3D12Fence *fence, uint64_t fenceIndex) { return false; }
     virtual ~DirectXDeviceConfigExt() noexcept override = default;
 };
 

include/luisa/backends/ext/vk_config_ext.h

@@ -28,6 +28,9 @@ class VulkanDeviceConfigExt : public DeviceConfigExt {
     virtual VkCommandBuffer borrow_command_buffer(
         StreamTag stream_tag) noexcept { return nullptr; }
     virtual bool execute_command_buffer(VkCommandBuffer cmd_buffer) noexcept { return false; }
+    virtual bool signal_semaphore(VkQueue queue, VkSemaphore _semaphore, uint64_t index) { return false; }
+    virtual bool wait_semaphore(VkQueue queue, VkSemaphore _semaphore, uint64_t index) { return false; }
+    virtual bool sync_semaphore(VkSemaphore _semaphore, uint64_t index) { return false; }
     virtual void readback_vulkan_device(
         VkInstance instance,
         VkPhysicalDevice physical_device,

src/backends/dx/DXApi/LCEvent.cpp

@@ -19,7 +19,8 @@ void LCEvent::Sync(uint64_t fenceIdx) const {
 }
 void LCEvent::Signal(CommandQueue *queue, uint64 fenceIdx) const {
     std::lock_guard lck(eventMtx);
-    queue->Queue()->Signal(fence.Get(), fenceIdx);
+    if (!device->deviceSettings || !device->deviceSettings->SignalFence(queue->Queue(), fence.Get(), fenceIdx))
+        queue->Queue()->Signal(fence.Get(), fenceIdx);
     lastFence = std::max(lastFence, fenceIdx);
     queue->AddEvent(this, fenceIdx);
 }
@@ -31,7 +32,8 @@ void LCEvent::Signal(DStorageCommandQueue *queue, uint64 fenceIdx) const {
 }
 void LCEvent::Wait(CommandQueue *queue, uint64 fenceIdx) const {
     std::lock_guard lck(eventMtx);
-    queue->Queue()->Wait(fence.Get(), fenceIdx);
+    if (!device->deviceSettings || !device->deviceSettings->WaitFence(queue->Queue(), fence.Get(), fenceIdx))
+        queue->Queue()->Wait(fence.Get(), fenceIdx);
 }
 bool LCEvent::IsComplete(uint64 fenceIdx) const {
     std::lock_guard lck(eventMtx);

src/backends/dx/DXRuntime/Device.cpp

@@ -50,6 +50,7 @@ Device::~Device() {
 
 void Device::WaitFence(ID3D12Fence *fence, uint64 fenceIndex) {
     if (fenceIndex <= 0) return;
+    if (deviceSettings && deviceSettings->SyncFence(fence, fenceIndex)) return;
     HANDLE eventHandle = CreateEventEx(nullptr, nullptr, false, EVENT_ALL_ACCESS);
     auto d = vstd::scope_exit([&] {
         CloseHandle(eventHandle);

src/backends/vk/event.cpp

@@ -51,6 +51,8 @@ void Event::signal(Stream &stream, uint64_t value, VkCommandBuffer *cmdbuffer) {
         std::lock_guard lck(eventMtx);
         lastFence = std::max(lastFence, value);
     }
+    if (device()->config_ext() && device()->config_ext()->signal_semaphore(stream.queue(), _semaphore, value))
+        return;
     auto timelineInfo1 = get_timeline_submit(&value);
     VkSubmitInfo info1{};
     info1.sType = VK_STRUCTURE_TYPE_SUBMIT_INFO;
@@ -65,7 +67,9 @@ void Event::signal(Stream &stream, uint64_t value, VkCommandBuffer *cmdbuffer) {
 
     VK_CHECK_RESULT(vkQueueSubmit(stream.queue(), 1, &info1, VK_NULL_HANDLE));
 }
-void Event::wait(Stream &stream, uint64_t value, VkCommandBuffer *cmdbuffer) {
+void Event::wait(Stream &stream, uint64_t value) {
+    if (device()->config_ext() && device()->config_ext()->wait_semaphore(stream.queue(), _semaphore, value))
+        return;
     VkTimelineSemaphoreSubmitInfo timelineInfo1{};
     timelineInfo1.sType = VK_STRUCTURE_TYPE_TIMELINE_SEMAPHORE_SUBMIT_INFO;
     timelineInfo1.pNext = nullptr;
@@ -84,11 +88,13 @@ void Event::wait(Stream &stream, uint64_t value, VkCommandBuffer *cmdbuffer) {
     info1.signalSemaphoreCount = 0;
     info1.pSignalSemaphores = nullptr;
     // ... Enqueue initial device work here.
-    info1.commandBufferCount = cmdbuffer ? 1 : 0;
-    info1.pCommandBuffers = cmdbuffer;
+    info1.commandBufferCount = 0;
+    info1.pCommandBuffers = nullptr;
     VK_CHECK_RESULT(vkQueueSubmit(stream.queue(), 1, &info1, VK_NULL_HANDLE));
 }
 void Event::host_wait(uint64_t value) {
+    if (device()->config_ext() && device()->config_ext()->sync_semaphore(_semaphore, value))
+        return;
     VkSemaphoreWaitInfo info{
         .sType = VK_STRUCTURE_TYPE_SEMAPHORE_WAIT_INFO,
         .semaphoreCount = 1,

src/backends/vk/event.h

@@ -11,7 +11,7 @@ class Event : public Resource {
     void update_fence(uint64_t value);
     void signal(Stream &stream, uint64_t value, VkCommandBuffer *cmdbuffer = nullptr);
     void signal_sparse(Stream &stream, uint64_t const* value_ptr, VkBindSparseInfo *sparse_info, VkTimelineSemaphoreSubmitInfo* timeline_ptr);
-    void wait(Stream &stream, uint64_t value, VkCommandBuffer *cmdbuffer = nullptr);
+    void wait(Stream &stream, uint64_t value);
     void host_wait(uint64_t value);
     void notify(uint64_t value);
 public: