Merge pull request #387 from SegmentLinking/dup_removal_fixes

Fix race conditions in duplicate removal
SegmentLinking · Apr 12, 2024 · a8bb417 · a8bb417
2 parents fc97a6b + 64f247a
commit a8bb417
Show file tree

Hide file tree

Showing 2 changed files with 7 additions and 26 deletions.
diff --git a/SDL/Event.cc b/SDL/Event.cc
@@ -930,11 +930,8 @@ void SDL::Event<SDL::Acc>::createPixelTriplets() {
       createWorkDiv(blocksPerGridDupPixTrip, threadsPerBlockDupPixTrip, elementsPerThread);
 
   SDL::removeDupPixelTripletsInGPUFromMap removeDupPixelTripletsInGPUFromMap_kernel;
-  auto const removeDupPixelTripletsInGPUFromMapTask(
-      alpaka::createTaskKernel<Acc>(removeDupPixelTripletsInGPUFromMap_workDiv,
-                                    removeDupPixelTripletsInGPUFromMap_kernel,
-                                    *pixelTripletsInGPU,
-                                    false));
+  auto const removeDupPixelTripletsInGPUFromMapTask(alpaka::createTaskKernel<Acc>(
+      removeDupPixelTripletsInGPUFromMap_workDiv, removeDupPixelTripletsInGPUFromMap_kernel, *pixelTripletsInGPU));
 
   alpaka::enqueue(queue, removeDupPixelTripletsInGPUFromMapTask);
   alpaka::wait(queue);
@@ -1158,8 +1155,7 @@ void SDL::Event<SDL::Acc>::createPixelQuintuplets() {
   auto const removeDupPixelQuintupletsInGPUFromMapTask(
       alpaka::createTaskKernel<Acc>(removeDupPixelQuintupletsInGPUFromMap_workDiv,
                                     removeDupPixelQuintupletsInGPUFromMap_kernel,
-                                    *pixelQuintupletsInGPU,
-                                    false));
+                                    *pixelQuintupletsInGPU));
 
   alpaka::enqueue(queue, removeDupPixelQuintupletsInGPUFromMapTask);
 

diff --git a/SDL/Kernels.h b/SDL/Kernels.h
@@ -233,13 +233,8 @@ namespace SDL {
             if (nMatched >= 7) {
               if (score_rphisum1 > score_rphisum2) {
                 rmQuintupletFromMemory(quintupletsInGPU, ix);
-                continue;
               } else if ((score_rphisum1 == score_rphisum2) && (ix < jx)) {
                 rmQuintupletFromMemory(quintupletsInGPU, ix);
-                continue;
-              } else {
-                rmQuintupletFromMemory(quintupletsInGPU, jx);
-                continue;
               }
             }
           }
@@ -276,15 +271,15 @@ namespace SDL {
 
           for (unsigned int ix1 = 0; ix1 < nQuintuplets_lowmod1; ix1 += 1) {
             unsigned int ix = quintupletModuleIndices_lowmod1 + ix1;
-            if (quintupletsInGPU.partOfPT5[ix] || quintupletsInGPU.isDup[ix])
+            if (quintupletsInGPU.partOfPT5[ix])
               continue;
 
             for (unsigned int jx1 = 0; jx1 < nQuintuplets_lowmod2; jx1++) {
               unsigned int jx = quintupletModuleIndices_lowmod2 + jx1;
               if (ix == jx)
                 continue;
 
-              if (quintupletsInGPU.partOfPT5[jx] || quintupletsInGPU.isDup[jx])
+              if (quintupletsInGPU.partOfPT5[jx])
                 continue;
 
               float eta1 = __H2F(quintupletsInGPU.eta[ix]);
@@ -325,9 +320,7 @@ namespace SDL {
 
   struct removeDupPixelTripletsInGPUFromMap {
     template <typename TAcc>
-    ALPAKA_FN_ACC void operator()(TAcc const& acc,
-                                  struct SDL::pixelTriplets pixelTripletsInGPU,
-                                  bool secondPass) const {
+    ALPAKA_FN_ACC void operator()(TAcc const& acc, struct SDL::pixelTriplets pixelTripletsInGPU) const {
       auto const globalThreadIdx = alpaka::getIdx<alpaka::Grid, alpaka::Threads>(acc);
       auto const gridThreadExtent = alpaka::getWorkDiv<alpaka::Grid, alpaka::Threads>(acc);
 
@@ -360,25 +353,17 @@ namespace SDL {
 
   struct removeDupPixelQuintupletsInGPUFromMap {
     template <typename TAcc>
-    ALPAKA_FN_ACC void operator()(TAcc const& acc,
-                                  struct SDL::pixelQuintuplets pixelQuintupletsInGPU,
-                                  bool secondPass) const {
+    ALPAKA_FN_ACC void operator()(TAcc const& acc, struct SDL::pixelQuintuplets pixelQuintupletsInGPU) const {
       auto const globalThreadIdx = alpaka::getIdx<alpaka::Grid, alpaka::Threads>(acc);
       auto const gridThreadExtent = alpaka::getWorkDiv<alpaka::Grid, alpaka::Threads>(acc);
 
       unsigned int nPixelQuintuplets = *pixelQuintupletsInGPU.nPixelQuintuplets;
       for (unsigned int ix = globalThreadIdx[1]; ix < nPixelQuintuplets; ix += gridThreadExtent[1]) {
-        if (secondPass && pixelQuintupletsInGPU.isDup[ix])
-          continue;
-
         float score1 = __H2F(pixelQuintupletsInGPU.score[ix]);
         for (unsigned int jx = globalThreadIdx[2]; jx < nPixelQuintuplets; jx += gridThreadExtent[2]) {
           if (ix == jx)
             continue;
 
-          if (secondPass && pixelQuintupletsInGPU.isDup[jx])
-            continue;
-
           int nMatched = checkHitspT5(ix, jx, pixelQuintupletsInGPU);
           float score2 = __H2F(pixelQuintupletsInGPU.score[jx]);
           if (nMatched >= 7) {