GPU: Fix Warp race condition in CUDA with Turing and later GPUs

davidrohr · davidrohr · commit 09d91d249a85 · 2021-03-02T08:26:57.000+01:00
diff --git a/GPU/GPUTracking/DataCompression/GPUTPCCompressionKernels.cxx b/GPU/GPUTracking/DataCompression/GPUTPCCompressionKernels.cxx
@@ -481,16 +481,19 @@ GPUdi() void GPUTPCCompressionGatherKernels::compressorMemcpyBuffered(V* buf, T*
       compressorMemcpyBasic(bufT + shmPos, src + srcOffset + srcPos, size, nLanes, iLane);
       srcPos += size;
       shmPos += size;
+      GPUbarrierWarp();
 
       if (shmPos >= bufTSize) {
         compressorMemcpyBasic(dstAligned + dstOffset, buf, bufSize, nLanes, iLane);
         dstOffset += bufSize;
         shmPos = 0;
+        GPUbarrierWarp();
       }
     }
   }
 
   compressorMemcpyBasic(reinterpret_cast<T*>(dstAligned + dstOffset), bufT, shmPos, nLanes, iLane);
+  GPUbarrierWarp();
 }
 
 template <typename T>

Original file line number	Diff line number	Diff line change
`@@ -481,16 +481,19 @@ GPUdi() void GPUTPCCompressionGatherKernels::compressorMemcpyBuffered(V* buf, T*`
`481`	`481`	`compressorMemcpyBasic(bufT + shmPos, src + srcOffset + srcPos, size, nLanes, iLane);`
`482`	`482`	`srcPos += size;`
`483`	`483`	`shmPos += size;`
	`484`	`+ GPUbarrierWarp();`
`484`	`485`
`485`	`486`	`if (shmPos >= bufTSize) {`
`486`	`487`	`compressorMemcpyBasic(dstAligned + dstOffset, buf, bufSize, nLanes, iLane);`
`487`	`488`	`dstOffset += bufSize;`
`488`	`489`	`shmPos = 0;`
	`490`	`+ GPUbarrierWarp();`
`489`	`491`	`}`
`490`	`492`	`}`
`491`	`493`	`}`
`492`	`494`
`493`	`495`	`compressorMemcpyBasic(reinterpret_cast<T*>(dstAligned + dstOffset), bufT, shmPos, nLanes, iLane);`
	`496`	`+ GPUbarrierWarp();`
`494`	`497`	`}`
`495`	`498`
`496`	`499`	`template <typename T>`