fix the fix in previous commit

Author: devsh

include/nbl/video/utilities/IUtilities.h

@@ -359,7 +359,8 @@ class NBL_API2 IUtilities : public core::IReferenceCounted
             if (manualFlush)
                 flushRanges.reserve((bufferRange.size-1)/m_defaultUploadBuffer.get()->max_size()+1);
 
-            const auto oldScratchStage = nextSubmit.scratchSemaphore.stageMask;
+            // for the signal to be useful for us to let go of memory, we need to signal after transfer is finished
+            const auto oldScratchStage = nextSubmit.scratchSemaphore.stageMask|=asset::PIPELINE_STAGE_FLAGS::COPY_BIT;
             //
             auto* uploadBuffer = m_defaultUploadBuffer.get()->getBuffer();
             // no pipeline barriers necessary because write and optional flush happens before submit, and memory allocation is reclaimed after fence signal
@@ -400,11 +401,9 @@ class NBL_API2 IUtilities : public core::IReferenceCounted
                         flushRanges.clear();
                     }
                     const auto completed = nextSubmit.getFutureScratchSemaphore();
-                    // for the signal to be useful for us to let go of memory, we need to signal after transfer is finished
-                    nextSubmit.scratchSemaphore.stageMask |= asset::PIPELINE_STAGE_FLAGS::COPY_BIT;
                     nextSubmit.overflowSubmit(scratch);
                     // first submit we respect whatever stages the user had (maybe they wanted to be notified of the completion of `nextSubmit.prevCommandBuffers`
-                    nextSubmit.scratchSemaphore.stageMask = {};
+                    nextSubmit.scratchSemaphore.stageMask = asset::PIPELINE_STAGE_FLAGS::COPY_BIT;
                     // overflowSubmit no longer blocks for the last submit to have completed, so we must do it ourselves here
                     // TODO: if we cleverly overflowed BEFORE completely running out of memory (better heuristics) then we wouldn't need to do this and some CPU-GPU overlap could be achieved
                     if (nextSubmit.overflowCallback)
@@ -607,7 +606,8 @@ class NBL_API2 IUtilities : public core::IReferenceCounted
             // TODO: Why did we settle on `/4` ? It definitely wasn't about the uint32_t size!
             const uint32_t optimalTransferAtom = core::min<uint32_t>(limits.maxResidentInvocations*OptimalCoalescedInvocationXferSize,m_defaultDownloadBuffer->get_total_size()/4);
 
-            const auto oldScratchStage = nextSubmit.scratchSemaphore.stageMask;
+            // for the signal to be useful for us to execute the data consumer callback, the signal must happen after the copy is done
+            const auto oldScratchStage = nextSubmit.scratchSemaphore.stageMask|=asset::PIPELINE_STAGE_FLAGS::COPY_BIT;
             // Basically downloadedSize is downloadRecordedIntoCommandBufferSize :D
             for (size_t downloadedSize=0ull; downloadedSize<srcBufferRange.size;)
             {
@@ -643,11 +643,9 @@ class NBL_API2 IUtilities : public core::IReferenceCounted
                 else // but first sumbit the already buffered up copies
                 {
                     const auto completed = nextSubmit.getFutureScratchSemaphore();
-                    // for the signal to be useful for us to execute the data consumer callback, the signal must happen after the copy is done
-                    nextSubmit.scratchSemaphore.stageMask |= asset::PIPELINE_STAGE_FLAGS::COPY_BIT;
                     nextSubmit.overflowSubmit(scratch);
                     // first submit we respect whatever stages the user had (maybe they wanted to be notified of the completion of `nextSubmit.prevCommandBuffers`
-                    nextSubmit.scratchSemaphore.stageMask = {};
+                    nextSubmit.scratchSemaphore.stageMask = asset::PIPELINE_STAGE_FLAGS::COPY_BIT;
                     // overflowSubmit no longer blocks for the last submit to have completed, so we must do it ourselves here
                     // TODO: if we cleverly overflowed BEFORE completely running out of memory (better heuristics) then we wouldn't need to do this and some CPU-GPU overlap could be achieved
                     if (nextSubmit.overflowCallback)

src/nbl/video/utilities/IUtilities.cpp

@@ -79,7 +79,8 @@ bool IUtilities::updateImageViaStagingBuffer(
         flushRanges.reserve(maxIterations);
 
     auto* uploadBuffer = m_defaultUploadBuffer.get()->getBuffer();
-    const auto oldScratchStage = intendedNextSubmit.scratchSemaphore.stageMask;
+    // for the signal to be useful for us to let go of memory, we need to signal after transfer is finished
+    const auto oldScratchStage = intendedNextSubmit.scratchSemaphore.stageMask|=asset::PIPELINE_STAGE_FLAGS::COPY_BIT;
     while (!regionIterator.isFinished())
     {
         size_t memoryNeededForRemainingRegions = regionIterator.getMemoryNeededForRemainingRegions();
@@ -108,11 +109,9 @@ bool IUtilities::updateImageViaStagingBuffer(
                 flushRanges.clear();
             }
             const auto completed = intendedNextSubmit.getFutureScratchSemaphore();
-            // for the signal to be useful for us to let go of memory, we need to signal after transfer is finished
-            intendedNextSubmit.scratchSemaphore.stageMask |= asset::PIPELINE_STAGE_FLAGS::COPY_BIT;
             intendedNextSubmit.overflowSubmit(scratch);
             // first submit we respect whatever stages the user had (maybe they wanted to be notified of the completion of `nextSubmit.prevCommandBuffers`
-            intendedNextSubmit.scratchSemaphore.stageMask = {};
+            intendedNextSubmit.scratchSemaphore.stageMask = asset::PIPELINE_STAGE_FLAGS::COPY_BIT;
             // overflowSubmit no longer blocks for the last submit to have completed, so we must do it ourselves here
             // TODO: if we cleverly overflowed BEFORE completely running out of memory (better heuristics) then we wouldn't need to do this and some CPU-GPU overlap could be achieved
             if (intendedNextSubmit.overflowCallback)