Guard acquire-semaphore reuse before re-acquire (VUID-01779)

Secondary swapchains presented via vhPresentSwapchain reused an acquire
semaphore slot whose prior GPU wait could still be in flight, tripping
VUID-vkAcquireNextImageKHR-semaphore-01779. Track the graphics-queue
timeline instance per acquire slot (acquireInstances, parallel to
acquireSemaphores) and vkWaitSemaphores on the previous instance before
reusing the slot. Mirrors the primary frame-instance wait but lives on
each swapchain, so it covers primary and secondaries alike. The global
frame-instance pacing is kept: it bounds in-flight flushes to
g_vhFramesInFlight, which the transient uniform ring (sized to
VRHI_MAX_FRAMES_INFLIGHT) depends on.

Also drain the backend (vhFlush(true)) at the top of vhPresentSwapchain so
the present path does not submit before the frame's recorded draws,
making it symmetric with vhFrame.

Author: hypernewbie

src/vrhi_device.cpp

@@ -1608,6 +1608,7 @@ void vhSwapchainCreate_Internal( vhSwapchain& sc, int width, int height )
     int newImageCount = ( int ) sc.images.size();
     int acquireSemaphoreCount = std::max( newImageCount, ( int ) g_vhFramesInFlight );
     sc.acquireSemaphores.resize( acquireSemaphoreCount );
+    sc.acquireInstances.assign( acquireSemaphoreCount, 0 );
     sc.presentSemaphores.resize( newImageCount );
 
     VkSemaphoreCreateInfo sci = { VK_STRUCTURE_TYPE_SEMAPHORE_CREATE_INFO };
@@ -1630,6 +1631,7 @@ void vhSwapchainDestroy_Internal( vhSwapchain& sc )
             vkDestroySemaphore( g_vulkanDevice, sem, nullptr );
     }
     sc.acquireSemaphores.clear();
+    sc.acquireInstances.clear();
     sc.presentSemaphores.clear();
 
     for ( vhTexture texture : sc.textures )
@@ -1675,6 +1677,7 @@ bool vhSwapchainPresentAndAcquire_Internal( vhSwapchain& sc, uint64_t& outInstan
     }
 
     outInstance = vhCmdListFlush( nvrhi::CommandQueue::Graphics );
+    sc.acquireInstances[sc.acquireSemaphoreIndex] = outInstance;
 
     VkPresentInfoKHR presentInfo = { VK_STRUCTURE_TYPE_PRESENT_INFO_KHR };
     presentInfo.waitSemaphoreCount = 1;
@@ -1699,6 +1702,24 @@ bool vhSwapchainPresentAndAcquire_Internal( vhSwapchain& sc, uint64_t& outInstan
 
     sc.acquireSemaphoreIndex = ( sc.acquireSemaphoreIndex + 1 ) % sc.acquireSemaphores.size();
 
+    // VUID-vkAcquireNextImageKHR-semaphore-01779: the acquire semaphore being reused must be idle.
+    uint64_t prevInstance = sc.acquireInstances[sc.acquireSemaphoreIndex];
+    if ( prevInstance )
+    {
+        VkSemaphore queueSem;
+        {
+            std::lock_guard< std::mutex > lock( g_nvRHIStateMutex );
+            queueSem = nvrhiDevice->getQueueSemaphore( nvrhi::CommandQueue::Graphics );
+        }
+        VkSemaphoreWaitInfo waitInfo = { VK_STRUCTURE_TYPE_SEMAPHORE_WAIT_INFO };
+        waitInfo.semaphoreCount = 1;
+        waitInfo.pSemaphores = &queueSem;
+        waitInfo.pValues = &prevInstance;
+        vhProfile( "vhFrame_WaitSemaphore", true );
+        vkWaitSemaphores( g_vulkanDevice, &waitInfo, UINT64_MAX );
+        vhProfile( "vhFrame_WaitSemaphore", false );
+    }
+
     // Acquire next image for this swapchain.
     uint32_t idx = 0;
     vhProfile( "vhFrame_AcquireNextImage", true );
@@ -1909,6 +1930,8 @@ bool vhPresentSwapchain( vhSwapchainID swapchain )
 {
     auto it = g_vhSwapchains.find( swapchain );
     if ( it == g_vhSwapchains.end() ) return false;
+    // Drain backend so the main thread can use the same nvrhi command list for swapchain semaphore ops below.
+    vhFlush( true );
     uint64_t instance;
     return vhSwapchainPresentAndAcquire_Internal( it->second, instance );
 }

src/vrhi_internal.h

@@ -272,6 +272,7 @@ struct vhSwapchain
     std::vector< vhTexture >            textures;
     std::vector< nvrhi::TextureHandle > nvrhiHandles;
     std::vector< VkSemaphore >          acquireSemaphores;
+    std::vector< uint64_t >             acquireInstances;
     std::vector< VkSemaphore >          presentSemaphores;
     uint32_t                            currentSwapchainIndex = 0;
     uint32_t                            acquireSemaphoreIndex = 0;

test/test_window.cpp

@@ -1003,6 +1003,9 @@ UTEST( Swapchain, SemaphoreArraysCoverFrameSlots )
         EXPECT_TRUE( vhFrame() );
     }
 
+    for ( uint64_t inst : g_vhSwapchains[sc].acquireInstances )
+        EXPECT_GT( inst, 0ull );
+
     vhDestroySwapchain( sc );
     TestEnsureShutdown();
     RGFW_window_close( secondary );