gfx: Add `ITransientResourceHeap::finish()` to avoid `Signal` after every queue submit. (#2158)

Co-authored-by: Yong He <yhe@nvidia.com>

8 files changed, 47 insertions, 20 deletions

diff --git a/examples/example-base/example-base.cpp b/examples/example-base/example-base.cpp
index 68a30fd60..9330dcf04 100644
--- a/examples/example-base/example-base.cpp
+++ b/examples/example-base/example-base.cpp
@@ -105,6 +105,7 @@ void WindowedAppBase::mainLoop()
 
     gTransientHeaps[frameBufferIndex]->synchronizeAndReset();
     renderFrame(frameBufferIndex);
+    gTransientHeaps[frameBufferIndex]->finish();
 }
 
 void WindowedAppBase::createSwapchainFramebuffers()
diff --git a/slang-gfx.h b/slang-gfx.h
index e9d33bc98..5350444e9 100644
--- a/slang-gfx.h
+++ b/slang-gfx.h
@@ -1890,8 +1890,17 @@ public:
         uint32_t constantBufferDescriptorCount;
         uint32_t accelerationStructureDescriptorCount;
     };
+
+    // Waits until GPU commands issued before last call to `finish()` has been completed, and resets
+    // all transient resources holds by the heap.
+    // This method must be called before using the transient heap to issue new GPU commands.
+    // In most situations this method should be called at the beginning of each frame.
     virtual SLANG_NO_THROW Result SLANG_MCALL synchronizeAndReset() = 0;
 
+    // Must be called when the application has done using this heap to issue commands. In most situations
+    // this method should be called at the end of each frame.
+    virtual SLANG_NO_THROW Result SLANG_MCALL finish() = 0;
+
     // Command buffers are one-time use. Once it is submitted to the queue via
     // `executeCommandBuffers` a command buffer is no longer valid to be used any more. Command
     // buffers must be closed before submission. The current D3D12 implementation has a limitation
diff --git a/tools/gfx/d3d12/render-d3d12.cpp b/tools/gfx/d3d12/render-d3d12.cpp
index 67c31158b..de3081f40 100644
--- a/tools/gfx/d3d12/render-d3d12.cpp
+++ b/tools/gfx/d3d12/render-d3d12.cpp
@@ -2378,6 +2378,7 @@ void DeviceImpl::submitResourceCommandsAndWait(const DeviceImpl::ResourceCommand
 {
     info.commandBuffer->close();
     m_resourceCommandQueue->executeCommandBuffer(info.commandBuffer);
+    m_resourceCommandTransientHeap->finish();
     m_resourceCommandTransientHeap->synchronizeAndReset();
 }
 
@@ -2785,18 +2786,9 @@ Result AccelerationStructureImpl::getNativeHandle(InteropHandle* outHandle)
 
 Result TransientResourceHeapImpl::synchronizeAndReset()
 {
-    Array<HANDLE, 16> waitHandles;
-    for (auto& waitInfo : m_waitInfos)
-    {
-        if (waitInfo.waitValue == 0)
-            continue;
-        if (waitInfo.fence)
-        {
-            waitInfo.fence->SetEventOnCompletion(waitInfo.waitValue, waitInfo.fenceEvent);
-            waitHandles.add(waitInfo.fenceEvent);
-        }
-    }
-    WaitForMultipleObjects((DWORD)waitHandles.getCount(), waitHandles.getBuffer(), TRUE, INFINITE);
+    WaitForMultipleObjects(
+        (DWORD)m_waitHandles.getCount(), m_waitHandles.getBuffer(), TRUE, INFINITE);
+    m_waitHandles.clear();
     m_currentViewHeapIndex = -1;
     m_currentSamplerHeapIndex = -1;
     allocateNewViewDescriptorHeap(m_device);
@@ -2809,6 +2801,22 @@ Result TransientResourceHeapImpl::synchronizeAndReset()
     return SLANG_OK;
 }
 
+Result TransientResourceHeapImpl::finish()
+{
+    for (auto& waitInfo : m_waitInfos)
+    {
+        if (waitInfo.waitValue == 0)
+            continue;
+        if (waitInfo.fence)
+        {
+            waitInfo.queue->Signal(waitInfo.fence, waitInfo.waitValue);
+            waitInfo.fence->SetEventOnCompletion(waitInfo.waitValue, waitInfo.fenceEvent);
+            m_waitHandles.add(waitInfo.fenceEvent);
+        }
+    }
+    return SLANG_OK;
+}
+
 TransientResourceHeapImpl::QueueWaitInfo& TransientResourceHeapImpl::getQueueWaitInfo(
     uint32_t queueIndex)
 {
@@ -3462,10 +3470,8 @@ void RayTracingCommandEncoderImpl::dispatchRays(
 
     auto shaderTableImpl = static_cast<ShaderTableImpl*>(shaderTable);
 
-    ResourceCommandEncoderImpl resourceCopyEncoder;
-    resourceCopyEncoder.init(m_commandBuffer);
     auto shaderTableBuffer =
-        shaderTableImpl->getOrCreateBuffer(pipelineImpl, m_transientHeap, &resourceCopyEncoder);
+        shaderTableImpl->getOrCreateBuffer(pipelineImpl, m_transientHeap, static_cast<ResourceCommandEncoderImpl*>(this));
     auto shaderTableAddr = shaderTableBuffer->getDeviceAddress();
 
     D3D12_DISPATCH_RAYS_DESC dispatchDesc = {};
@@ -5015,8 +5021,8 @@ void CommandQueueImpl::executeCommandBuffers(
             auto& waitInfo = transientHeap->getQueueWaitInfo(m_queueIndex);
             waitInfo.waitValue = m_fenceValue;
             waitInfo.fence = m_fence;
+            waitInfo.queue = m_d3dQueue;
         }
-        m_d3dQueue->Signal(m_fence, m_fenceValue);
     }
 
     if (fence)
@@ -6556,10 +6562,7 @@ void ResourceCommandEncoderImpl::textureBarrier(
 void ResourceCommandEncoderImpl::bufferBarrier(
     size_t count, IBufferResource* const* buffers, ResourceState src, ResourceState dst)
 {
-
-    List<D3D12_RESOURCE_BARRIER> barriers;
-    barriers.reserve(count);
-
+    ShortList<D3D12_RESOURCE_BARRIER, 16> barriers;
     for (size_t i = 0; i < count; i++)
     {
         auto bufferImpl = static_cast<BufferResourceImpl*>(buffers[i]);
diff --git a/tools/gfx/d3d12/render-d3d12.h b/tools/gfx/d3d12/render-d3d12.h
index 3ef4a84a7..2cb9a3fd3 100644
--- a/tools/gfx/d3d12/render-d3d12.h
+++ b/tools/gfx/d3d12/render-d3d12.h
@@ -490,9 +490,11 @@ public:
     {
         uint64_t waitValue;
         HANDLE fenceEvent;
+        ComPtr<ID3D12CommandQueue> queue;
         ComPtr<ID3D12Fence> fence = nullptr;
     };
     ShortList<QueueWaitInfo, 4> m_waitInfos;
+    Array<HANDLE, 16> m_waitHandles;
 
     QueueWaitInfo& getQueueWaitInfo(uint32_t queueIndex);
     // During command submission, we need all the descriptor tables that get
@@ -545,6 +547,8 @@ public:
         createCommandBuffer(ICommandBuffer** outCommandBuffer) override;
 
     virtual SLANG_NO_THROW Result SLANG_MCALL synchronizeAndReset() override;
+
+    virtual SLANG_NO_THROW Result SLANG_MCALL finish() override;
 };
 
 struct Submitter
diff --git a/tools/gfx/debug-layer.cpp b/tools/gfx/debug-layer.cpp
index 341f0034a..414434da6 100644
--- a/tools/gfx/debug-layer.cpp
+++ b/tools/gfx/debug-layer.cpp
@@ -1619,6 +1619,12 @@ Result DebugTransientResourceHeap::synchronizeAndReset()
     return baseObject->synchronizeAndReset();
 }
 
+Result DebugTransientResourceHeap::finish()
+{
+    SLANG_GFX_API_FUNC;
+    return baseObject->finish();
+}
+
 Result DebugTransientResourceHeap::createCommandBuffer(ICommandBuffer** outCommandBuffer)
 {
     SLANG_GFX_API_FUNC;
diff --git a/tools/gfx/debug-layer.h b/tools/gfx/debug-layer.h
index 84ae33c2a..de8ccfe8e 100644
--- a/tools/gfx/debug-layer.h
+++ b/tools/gfx/debug-layer.h
@@ -724,6 +724,7 @@ public:
 public:
     ITransientResourceHeap* getInterface(const Slang::Guid& guid);
     virtual SLANG_NO_THROW Result SLANG_MCALL synchronizeAndReset() override;
+    virtual SLANG_NO_THROW Result SLANG_MCALL finish() override;
     virtual SLANG_NO_THROW Result SLANG_MCALL
         createCommandBuffer(ICommandBuffer** outCommandBuffer) override;
 };
diff --git a/tools/gfx/renderer-shared.h b/tools/gfx/renderer-shared.h
index 24d51c87a..a9d9e19c4 100644
--- a/tools/gfx/renderer-shared.h
+++ b/tools/gfx/renderer-shared.h
@@ -1150,6 +1150,8 @@ public:
             return static_cast<ITransientResourceHeap*>(this);
         return nullptr;
     }
+
+    virtual SLANG_NO_THROW Result SLANG_MCALL finish() override { return SLANG_OK; }
 };
 
 class ShaderTableBase
diff --git a/tools/render-test/render-test-main.cpp b/tools/render-test/render-test-main.cpp
index 7046e2fda..81e8cd38a 100644
--- a/tools/render-test/render-test-main.cpp
+++ b/tools/render-test/render-test-main.cpp
@@ -939,6 +939,7 @@ Result RenderTestApp::writeBindingOutput(const String& fileName)
 
                 commandBuffer->close();
                 m_queue->executeCommandBuffer(commandBuffer);
+                m_transientHeap->finish();
                 m_transientHeap->synchronizeAndReset();
 
                 m_device->readBufferResource(stagingBuffer, 0, bufferSize, blob.writeRef());